Professional Documents
Culture Documents
Foundations of Inventory Management
Foundations of Inventory Management
INVENTORY
MANAGEMENT
.tlHl
P ';, ,. I '1:_'"
THE IRWINIMCGRAW-HILL SERIES
Operations and Decision Sciences
-
!OIWsIjcs Quantitative Business Analysis: Spread:iheets
Wasserman Casebook
Applied Linear Regression First Edition
First Edition
Models
z •
, .;mg Third Edition
.,.., Case
§!tetistics,
s-nes
,..
.........
... aId
_Tracy
aid
-=
-'
Tools and
__ Statistics
archal
BIlSiness and
FOUNDATIONS OF
INVENTORY
MANAGEMENT
Paul H. Zipkin
Fuqua Schaal ofBusiness
Duke University
Boston Burr Ridge, IL Dubuque, IA Madison, WI New York San Francisco Sl. Louis
Bangkok Bogota Caracas Lisbon London Madrid
Mexico City Milan New Delhi Seoul Singapore Sydney Taipei Toronto
........
II·,
,I ; I,:
Copyright © 2000 by The McGraw-Hill Companies, Inc. All rights reserved. Printed in the United States of
America. Except as permitted under the United States Copyright Act of 1976, no part of this publication
may be reproduced or distributed in any form or by any means, or stored in a database or retrieval system,
without the prior written permission of the publisher.
1234567890DOCroOC09876543210
ISBN 0-256-11379-3
Includes index.
ISBN 0-256-11379-3
http:\\www.mhhe.com
p R E F A c E
iIIiII::d States of During the making of this book I have often reflected on my good fortune-so many fine
""""""'0
-system,
people helped me, directly or indirectly. Here I can express only a small fraction of my
gratitude to them.
My grandparents were four quite different people, but they shared certain qualities-
all were extraordinarily generous and remained cheerful even in harsh circwnstances. They
were very proud of us, their children, and grandchildren, and they would have been proud
to see this book (never mind whether it deserves such pride). My late mother was a source
of strength, warmth, and humor to all around her.
My father has taught me many things, but two especially come to mind. He showed
me the importance of carefully analyzing things, oflooking below surface appearances.
And, he demonstrated the importance of empathy and interest towards all people. I have
not always been the most responsible or responsive son, in these respects among others,
but give the man credit, he tried.
My wife Karen and my children Joe and Leah patiently and graciously endured
countless hours when I occupied myself with the computer screen instead of them. For
that, and for their constant love and support, I am deeply grateful.
How can I adequately thank the great teachers I have learned so much from? The
faculty ofthe lEOR Department atthe University of California, Berkeley, introduced me
to the field of Operations Research. I wish to thank especially C. Roger Glassey, William
Jewell, the late Ronald Shephard, Donald Topkis, and Ronald Wolff for encouraging my
fledgling efforts and tolerating my many mishaps. Later, during my doctoral studies at
Yale University, I was fortunate to receive the guidance and encouragement of Ron
Dembo, Donald Brown, Eric Denardo, and Ward Whitt, among others. lowe special
thanks to Harvey Wagner, whose wisdom played a great role in forming my under
standing of the field. Finally, to my advisor and mentor, Matthew 1. Sobel, who gave so
much of his time and effort to my dissertation work and taught me so much about the
99-40847 tasks and meaning of scholarship, and who kindly put up with all the egregious nonsense
I subjected him to, I offer my heartfelt thanks and appreciation.
vii
viii Foundations ofInventory Mana.gement
I have had wonderful colleagues, first at Columbia University and later at Dnke Uni
• •
versity.lowe a special debt of gratitude to my research collaborator and friendAwi Fed
ergruen. Side by side we fought many battles, intellectual and otherwise, and from them
I learned much about science and life.
This book grew out oflecture notes and assignments for courses I have given at Co
lumbia and Duke over the last 18 years. It represents my view ofthe key concepts of in
ventory management-how they really work and why they really matter. Chapter I sets
forth the goals of the book more fully.
To the students who suffered through those courses, I offer my apologies and thanks.
Their questions and suggestions contributed greatly to my understanding of the subject
and to the coherence of this book. And what superb students I have had! I cannot resist
mentioning in particular Shoshana Anily, Mark Ferguson, Arie Harel, Oded Koenigs
berg, Sang-Bum Lee, Yong-Joo Lee, Agues Pena-Perez, Antony Svoronos, Michal Tsur,
Weirning Zhang, Shaohui Zheng, and Yu-Sheng Zheng. The greatest joy of my profes
sionallife has been to watch you grow to surpass at least this one of your teachers. (My
apologies also to future students, whose wrists and minds this book may strain.)
Very special thanks go to Jing-Sheng (Jeannette) Song, former student and now col
laborator, friend, and true scholar. Over the years she read many drafts and offered count
less suggestions for improvement, and the book is much better than it would have been
without her efforts. Moreover, her continuing encouragement and (mostly) gentle prod
ding helped to counter my habitual sloth. Finally, after learning a few tricks from me, she
has taught me much more, from the technical to the philosophical, by both precept and
example.
Sincere thanks to the reviewers of the book, whose suggestions were so helpful:
Harry Groenevelt, University of Rochester; Ananth V Iyer, Purdue University; Hau L.
Lee, Stanford University; Kamran Moinzadeh, University of Washington; Steven Nah
mias, Santa Clara University; Leroy B. Schwarz, Purdue University; and Robert T.
Sumichrast, Virginia Polytechnic Institute and State University.
Finally, thanks to all the talented, hard-working and patient people at Irwin!
McGraw-Hill who contributed to the creation of the book. Especially hearty thanks to
my editor, Richard Hercher, who kept kindly suggesting that I click the "print" button,
but whose patience never faltered when I procrastinated.
Paul H. Zipkin
Durham
August, 1999
B R I E F C o N T E N T S
I General Introduction I
4 Time-Varying Demands 73
Bibliography 487
Index 503
"'H.Zipkin
DorIIam
~1999
Ix
c o N T E N T S
1 General Introduction 1
Disciplines 12
Practical 14
Engineering 16
Notes 17
2.1 Introduction 19
2.2 Contexts 20
2.2.2 Monitoring 21
xi
xii Foundations ofInventory Management
2.3 Models 23
2.3.1 Overview 23
Notes 27
3.1 Introduction 29
3.5.4 Discussion 57
Notes 67
Problems 67
4 Time-Varying Demands 73
4.1 Introduction 73
4.2.4 Discussion 76
4.3.5 Examples-Opacity 87
4.3.6 Heuristics 90
4.4 Extensions 93
4.4.1 Leadtimes 94
4.4.4 Backorders 95
Notes 103
Problems 104
Notes 168
Problems 170
Contents
xv
6.llntroduction 175
Notes 237
Problems 238
Contents xvii
Notes 289
Problems 290
Notes 359
Problems 361
Notes 427
Problems 428
Notes 437
Notes 444
Contents xxi
Notes 480
BIbliography 487
Index 503
c H A p T E R
1 GENERAL INTRODUCTION
Outline
1.1 Inventories Are Important-So Are 1.2 About the Book 5
Foundations I Notes I7
~
2 Foundations ofInventory Management
Large enterprises too, like the Japanese steel company I visited, need to manage in
ventories. Often, doing that well spells the difference between corporate success and
failure.
None of this is new. InventOly management is hardly a modem innovation. The earli
est humans kept caches of food and stone tools. As for recorded history, well, the whole
business started with inventOly management-writing was invented, it seems, for that very
purpose. Readers familiar with the Bible will recall that Joseph was, among other things, a
remarkably proficient inventory manager. (True, he did enjoy some unusual advantages
intimate access to senior management, not to mention a superb forecasting system.)
Returning to here and now, the person who runs the supply room has a different
viewpoint on inventories. This operation serves about 250 busy, demanding people, and
provides several dozen types and sizes of paper, as well as hundreds of other kinds of of
fice supplies, from paper clips and staples to computer diskettes. It takes a serious, con
scious effort to monitor all these items, to estimate their usages, to order new supplies
from our office-supplies wholesaler at the appropriate times, to decide when to discon
tinue rarely used items and when to add attractive new ones, and so forth. These are the
typical concerns of inventory management at the retail level; the operators ofmost retail
stores face basically the same set of problems.
Behind the supply room and the retail store lies the vast global network of industry,
evelY part of which is permeated by inventories. Consider what has to happen before I
obtain my small stack ofpaper: Wood-products companies maintain inventories ofgrow
ing trees; paper companies keep logs waiting to be processed; the logs are transformed
into wood pulp, which is made into paper at paper mills; finished paper is held in in
ventory at the mills themselves and then at one or more warehouses, until it reaches the
wholesaler; from the wholesaler's stocks the paper is sent to our supply room, where I
pick it up. (As for paper clips, some of that Japanese steel may end up in them.)
Even this picture is simplified: Between these steps there are inventories in transit,
floating on rivers or carried in ships, trains, or trucks. In addition, wood is only one of
the components of pulp and paper; a variety of other inputs (mainly chemicals) are also
needed, all of which are stored as inventories before they are used. Finally, the com
plexities of scope are enormous. There are thousands of different types, grades, sizes,
and colors of paper, for example.
The management of these manufacturing inventories involves many of the same is
sues as a retail operation, but with a different emphasis. One key difference is that pro
ducers have to coordinate inventory-producing and -conswning activities at several re
lated processing stages and locations; logs, chemicals, pulp, and paper must be managed
in concert for the overall system to work.
A similar story could be told about every single item we use at home or at work (in
cluding paper clips). We could not eat or drink or bathe or keep warm, were it not for a
huge, intricate chain of production and distribution activities, all of them fed by diverse
inventories of raw materials, components, spare parts, and finished goods.
Inventories are also critical in the service industries, a fact you may find surprising.
After all, a service, according to one popmar definition, is something you can't drop on
your foot. Haircuts, brain surgery, auto insurance, college teaching-none of these can
be kept in a warehouse until needed. Nevertheless, the provision of any service depends
Chapter 1 General Introduction 3
Now, why does that Japanese steel company need so much coal? Because they ship
it in from Australia. A steel company in Germany, situated near a coal mine, needs a far
smaller inventory. Both are acting reasonably.
When during their programs should students take such a eourse? The nature of the
field makes this question problematic: Inventory theory draws on vITtually all of the ba
sic methodologies of operations research, so ideally students should already be familiar
with at least the rudiments of optimization theory, stochastic processes, and dynamic
programming, On the other hand, the subject is much more than a collection of miscel
laneous illustrations of those methodologies, as explained further below; indeed, it has
its own intellectual core. Also, because its focus is close to applications and practical
managerial concerns, it can serve as a powerful motivator for further learning in ad
vanced methodological courses. Furthermore. the material covered in the course under
lies a good deal of the research literature in operations management and industrial engi
neering broadly construed, not just in the inventory area.
As the theory itself tells us, when faced with contradictory criteria, the best solution
is a compromise. (Well, sometimes.) In my experience the course works best for most
students early in the second year of graduate study; the book is designed primarily for
students with that level of preparation. There are exceptions, however: The course can
also work well for first-year graduate students, and even undergraduates and MBA stu
dents, provided they have solid technical backgrounds and are able to learn some things
on their own. And, of course, it is never too late.
Partly to accommodate students with diverse preparations, and also to provide in
structors with a degree of flexibility in designing the course, I have included fairly ex
tensive appendices to make the book self-contained, or nearly so,
For similar reasons I have adopted a somewhat mixed approach toward mathemati
cal generality and rigor. There are some theorems and proofs in the book, included with
out apologies: One ofits purposes, certainly, is to expose the basic techniques of research
in the field, Also, the proofs often embody insights that cannot be conveyed fully other
wise. I have, nonetheless, tried to motivate these arguments as best I can, and very often
I forego full generality in favor of simplicity, Also, I have tried to identify and encapsu
late the more difficult and technical material, so that instructors can choose to work
around some of it without losing continuity.
In addition to constructing a pedagogical device, I have also aimed to convey cer
tain ideas to my academic and professional colleagues-and to students as future col
leagues-which I hope they will find interesting, useful, or at least provocative, This is
a book with a message, indeed a mission.
First, the book comprises what I view to be the fundamental concepts of inventory
theory as it now stands, the core, the backbone. Besides its function as a textboo~ it is
designed to serve as a reference on the subject for practicing professionals and academic
researchers. While some of these concepts have been around for decades, many are
treated in nonstandard ways or revived here after long neglect, and others are products
of quite recent research.
In forming this selection I have necessarily had to omit much of value and interest.
I have tried to compensate somewhat for these omissions in the notes and the problems
following the chapters, (Still, I remain painfully aware of how many beautiful and sig
nificant results I have been unable to include; I hope my colleagues will accept my pro
fuse apologies on this score.)
Also, the book expresses, in some cases impljcitly, certain judgments and convic
tions concerning the connections between these concepts, the relation of the field to
,,-,-~ .........
oanrre ofthe other disciplines, and especially its role and value in the practice of inventory manage
all ofthe ba ment. I explain these ideas, views, judgments, and convictions a bit more in § 1.2.3.
y be familiar
mddynamic
1.2.2 Structure and Contents
Xl of miscel
indeed, it has
1.2.2,1 Elements DflnventDry Systems
and practical
Inventory systems are extraordinarily diverse, and they differ along many dimen
nning in ad
sions. The overall structure of the book reflects, necessarily, a certain classification or
:nurse under
taxonomy of these systems. Ours is by no means the only way to classify them, but it is
dustrial engi
true that the major divisions of the book correspond to fundamental differences among
real systems. To understand what the various parts of the book are about, and the con
best solution
nections between them, therefore, we need to appreciate some ofthese basic distinctions.
best for most
Let us begin with a general, abstract way of looking at an inventory. We focus first
primarily for
on the inventory of one item only in a single location. Every inventory lies between two
)e course can
activities or processes, which we cal1 the supply process and the demand process. As Fig
IIId~A slU
ure 1.2.1 indicates, the supply process comprises the production, transportation, and/or
• some things
other activities that add new stock to the inventory, while the demand process describes
the various activities that use and thus subtract material from the inventory.
provide in
III
Like any abstraction, this picture omits a great deal. The arrows indicate directions
ded fairly ex
of movement, but otherwise the figure lacks a temporal dimension; it represents struc
ture, not activity. Also, there is nothing here to suggest infonnation and control; we can
:d mathemati
not see who or what directs the supply process to perfonn its tasks or what knowledge
ocluded with
guides these decJsions.
IeS of research
Furthennore, the figure gives no hint as to what lies inside the circles depicting the
>Ii fully other
supply and demand processes. These processes might be quite simple or enonnously com
lOll very often
plex systems themselves. Consider the inventory of one specific type of finished paper at
and encapsu
a paper mill, for example. The supply process here is that extensive chain ofprocessing and
.:JOSe to work
shipping activities that turns trees into paper, including many different inventories along
the way. The demand process too is a complex web, stretching from the mill to all the
iO convey cer
final users ofpaper, and perhaps even further to include certain factors that influence those
3S future col
e:ari\-e. This is
FIGURE 1.2.1
5 of inventory
Inventory-between supply and demand.
textbook, it is
mdacademic Supply Demand
process Inventory process
ies, many are
o v o
s are products
<: <:
e and interest.
I the problems
utiful and sig
oa:ept my pro
FIGURE 1.2.2
Procurement, production, and distribution.
Supply Components Assembly Product Demand
(mixing)
0=\/, ;f\7=0
0=\7=0
O=\7f '\7=0
Products Demands
<;
users (like the general level of the economy). The same supply and demand processes,
moreover, typically affect many other kinds of paper and other goods as well.
Nevertheless, this is a "sefid abstraction, and we shall refer to it repeatedly through
out the book.
For one thing, this basic structure can be used as a building block to compose dia
grams of much larger, more complex systems. Figure 1.2.2 suggests some of the possi
bilities. Here, two difterent components are procured and then assembled or mixed to
form one final producI, which is then shipped to three separate stocking locations.
Indeed, one critical distinction among inventory systems rests on their degrees of
structural complexity: There are some inventories that can be reasonably, plausibly viewed
in the simple terms of Figure 1.2.1 by itself, independently of other items in other places.
There are other cases, however, where multiple items and/or locations are essential, where
the distinct inventories in the system cannot sensibly be considered in isolation.
Even as it stands, Figure l.2.1 expresses a fundamental point about inventories in
general, their '"between-ness": To understand what an inventory does. how it works, why
it is held, the basic functions it performs, and to begin to think seriously about how to
manage it effectively. we must first understand the basic characteristics of the supply and
demand processes that drive it. Indeed, the real subject ofthis bookis supply and demand
and the transactions between them. Inventory is just one aspect of supply-demand inter
action, albeit a crucial one.
To appreciate this point, it is helpful to imagine a situation where an inventory is not
needed and would serve no useful purpose. For this to be so, we would have to find (or
design or adapt) a supply process capable of providing instantly the exact amounts
Chapter J General Introduction 9
needed to fill the demands arising from the demand process. Furthermore, such a per
fect supply process must be not just physically possible, but also reasonably inexpensive
to build and operate.
=and
The possibility and practicality of an ideal supply process, of course, depend in part
on the nature of the demand process. For example, if we could be sure from the begin
J
ning that demands will occur at a regular rate, with only minor variations over time, then
a supply process capable of providing goods at that same rate, and of responding imme
diately to variations within the expected range, would do the job.
No such perfect match exists in the real world, but there are situations that come
J
close: Most of us do not keep an inventory of cold water in our residences. Our sinks and
tubs are connected to a very effective supply process, the network of pipes. pumps, and
reservoirs that make up the public water system, which normally meets virtually all of
our domestic demands. These demands are not regular at all, but the supply process is
J
easily controlled to respond to our demands, by opening and closing the valves we call
taps. (Of course. it is only in the industrialized conntries and only quite recently that such
modem water systems have been created; elsewhere and in earlier times, people do and
did keep inventories of water, to minimize trips to the well.)
This example of water serves as a nice paradigm for the notion of perfectly matched
....
=
supply and demand processes. lfwe can envision an ideal supply process, we can further
imagine attaching it directly to the demand process by means of pipes and valves. We
would then have a system where the product flows directly from the supply system to
meet the demands.
d processes, In most situations, unfortunately, supply and demand processes cannot be matched
perfectly; most products do not flow smoothly in this way. Indeed, most of the important
illy through- functions of inventories can be understood in terms of the various types of mismatches
that arise between supply and demand processes. In other words, inventory serves as a
IXIlpOse dia buffer hetween supply and demand processes that do not fit neatly together, to mitigate
,fthe possi the costly disruptions that would otherwise occur. The most obvious type of disruption
ocmixed to is a shortage, a failure to !Deet demand as it occurs; inventories are often held, then, to
:aliens. prevent or reduce shortages. Most of the other significant potential disruptions are
r degn?es of strains of various kinds on the supply process itself.
sWlyviewed Here is a short (and by no means comprehensive) list of common and important
IJIher places. characteristics of actual supply and demand processes which, alone or in combination.
,.,n,u, where lead to such mismatches.
Ill
Demand processes:
in
ftit=D.tories
t v.urks, why Smooth or lumpy demand
bout how to Variations in demand over time
eutory is not
e to Tmd (or Delays in response-leadtimes
A good part of this book is devoted to Wlderstanding these factors and their interactions,
but for now we explain each one briefly:
Smooth or lumpy demand: When demand can be envisioned as a continuous stream.
drawing down inventory at some rate (constant or variable), then we say demand is
smooth. Lumpy demand, in contrast, occurs in large, discrete chunks. (Of course, these
two are extreme cases; real demands often have both smooth and lumpy components.)
Even smooth demand can lead to a mismatch, depending on the nature of the supply
process, as explained below. Normally, lumpy demand increases the difficulties further.
Variations in demand over time: Seasonal changes in demand, and other temporal
variations, often lead to mismatches, especially when the capacity of the supply process
is severely limited. Then, inventory must be built up during slack periods to have suffi
cient stock to meet later, heavier demands. Even with virtually unlimited capacity, how
ever, the presence of other complicating factors (supply leadtimes and demand uncer
tainties, for example) means that temporal changes must be anticipated somehow.
Unpredictable demand variations: Random fluctuations in demand, in conjunction with
supply leadtimes and capacity limits, are among the major reasons for maintaining wento
ries. In these cases, the supply process is unable to respond quickly to unanticipated demand
surges and declines, so inventory is needed in order to fill demands in a timely manner.
Economies ofscale in supp(V: In many situations the technology and economics of
the supply process favor long production runs or large deliveries, relative to short-tenn
demand. Examples include long production setup or changeover times and shipments
over great distances. Also, if we purchase goods from an independent supplier, the sup
plier may simply charge less per unit for a large order. Thus, even if it is possible to sup
ply small quantities, it may be expensive. This explains why even smooth demand can
lead to a problematic mismatch. In such cases there will be sizable inventories for some
time after a batch is produced or a shipment received.
Capacity limits: Every real supply process has limited capacity, and so has limited
ability to respond to changes or lumps in demand, whether anticipated or not Thus, in
ventories are required to compensate for this inflexibility. Capacity limits are more sig
nificant in some cases than others; often they can be safely ignored for practical pur
poses, but just as often they cannot.
Delays in response-leadtimes: As mentioned above, supply leadtimes and demand
fluctuations together lead to shortages, unless inventories are held to prevent or at least
mitigate them. When the leadtimes themselves are unpredictable, these difficulties be
come even more severe, for the same reasons.
Imperfect quality: Imperfect quality exacerbates many ofthe problems discussed so
far. Defects waste precious capacity, and they add elements of uncertainty to the supply
process, somewhat like random leadtimes.
There are other reasons to hold invcntory besides these supply-demand mismatches.
For example., gold and other precious metals are used as inputs to certain production
processes, but some people own them for purely speculative reasons, hoping their prices
will rise. In retail settings inventories often serve a marketing purpose; customers want
to examine goods before buying. Most of our attention here, however, will be focused
on mismatches such as those described above.
Chapter 1 General Introductiml 11
interactions, Once we understand the characteristics of supply and demand, we can design a good
inventory strategy. But, the job of management does not stop there. We must then step
IOUS stream, back and envision the total system, including supply and demand as well as inventory.
r demand is From that larger perspective we must ask, what can be done to improve the system?
:ourse, these
:xoponents.) 1.2.2.2 OrganizatiDn Df the BDDk
fthe supply These basic characteristics of supply and demand processes, along with differences in
tries further. structural complexity, define the primary divisions between the several parts, chapters,
!'Ply process The distinction between predictable (or deter.mirristic) and unpredictable (or sto
) have suffi chastic) processes is perhaps the single most significant dividing line between different
parity, how systems. One of the crucial functions of inventory, to protect against unforeseen contin
nand uncer gencies, simply does not arise when supplies and demands are (or can be usefully re
oebow. garded as) predictable. Only in the stochastic case must we be concerned with all the is
~onwith sues surrounding information. what we know and when. For just these reasons,
iIing mvento moreover, stochastic models involve different and rather more complex mathematics.
-rodemand The book's major partition is defined by this distinction: After the two introductory
manner. chapters of Part 1, Part II (Chapters 3 to 5) deals with predictable supply and demand,
oooomicsof while Part III (Chapters 6 to 9) treats stochastic models.
Dsbort-ter.m Another key distinction rests on the difference between stationary processes and
d shipments those that change over time. The need to anticipate temporal variations significantly
Iier, the sup complicates the problems of inventory management, as explained above. Models incor
&ible to sup porating time-varying data, accordingly, require their own special techniques of analy
demand can sis and solution. Within Part II, Chapter 3 focuses on stationary data. Models with tem
:ies for some poral variation are introduced in Chapter 4. (Chapter 5 includes both cases.) Similarly,
in Part lll, Chapters 6 to 8 deal mainly with stationary, stochastic models, while the non
) has limited stationary case is taken up in Chapter 9.
lOt. Thus, in It turns out to be convenient to introduce, along with nonstationarity, certain distinct
Ire more Slg approaches to model formulation and analysis. In particular, we model time as discrete
ractical pur in Chapters 4 and 9 (for reasons explained there), whereas time is continuous in most of
the rest of the book. Another difference requires a bit more explanation: We often as
aoddemand sume that a specific type or fonn of control policy will be used; then we try to evaluate
.. or at least difierent policies and to choose a good one, but always within this policy class. Some
lliculties be- times, however, we ask, what form does a truly optimal control policy take among all
possible policies? Both kinds of investigations are valuable, but for different reasons,
discussed so and they involve quite different methods of analysis. Chapters 4 and 9 consider optimal
II> the
supply control, whereas the other chapters mostly focus on particular policy forms.
Regarding stmctural complexity, Parts II and III both begin with one or two chapters
mismatches. on single-item, single-location models, followed by chapters on more complex networks.
• production (Chapter 9 considers both simple and complex structures, but in different sections.)
g their prices Most of the other specific characteristics of supply and demand processes men
ilOIDers want tioned above are dealt with in different sections of the individual chapters. For example,
~ be focused in the context of the stationary, deterministic systems of Chapter 3, limited capacity is
treated in Section 3.4.
12 Foundations ollnventory Management
subjects. In certain cases I have emphasized these connections somewhat more or dif
ferently than usual, however.
¥OWIl gradu
In particular, 1 have highlighted the connection between inventory theory and the
:ussions with
theory of queues. In this respect the book revives a tradition established in the classic
.lleagues and
work of Morse [1958] (or rather continues the revival begtm by Heyman and Sobel
:be selections
[1982] in a somewhat different manner). For queueing theory remains our richest source
of models for what 1 call here supply processes. Recent developments in both fields (for
instance, the flowering of queueing-network models, and the "discovery" of the effects
of batching on congestion and leadtimes) further enhance this interaction_ In addition, 1
ti: inventories find that several of the classic inventory models are most clearly expressed in the lan
:lmiques used guage of queues. The lost-sales model in Chapter 6 is one prominent example.
a motley col On the other hand, the central concerns ofinventory theory are rather different from
bas dynamic those of mainstream queueing theory; neither field subsumes the other. The challenge
for me, then, has been to establish this link without turning the book into a text on queues
131 model for or requiring students to have had a full course on the subject. (The approach here seems
.methods. For to work pretty well in my experience.)
k mirror im There are also strong links with subjects outside operations research, or more pre
:s.. through the cisely, outside typical operations-research curricula. The theory of dynamical systems
y position are and control, in particular, is prominently represented in the book. The systems approach,
deterministic I believe, is a compelling 'metaphor for the functions of management.
:clear the gen Inventory systems are dyuamical systems, as much as any real system is, and the
OODtext of the unified picture of infonnation. estimation and control provided by systems theory is a
I of stochastic valuable one in the context of inventory management. This remains true, even when in
Ill! random de ventory theory cannot, so far at least, follow the entire program exemplified in the analy
Kdown of this sis of the simple linear-quadratic control model. Thus, 1 have included the production
ill difficult.) smoothing model of Holt, Modigliani, Muth, and Simon [1960] (the celebrated HMMS
de\-elopments model). 1 do think this model is worth learning, even though most of inventory theory
dy are slightly and practice have followed other paths. There are reasons the HMMS assumptions are
~'. Likewise, problematic, and these reasons are interesting.
"'" the equally Also, the systems-theoretic framework can be used to pose a rich class of demand
~ presented as process models, which are consistent with standard modern forecasting methods. The
book does not treat forecasting methods per se, beyond infonnal discussions; other
IIIIl:nts, 1 have books and other courses are the places to learn statistical forecasting properly. Still, I be
ilI)"5, to clarify lieve it is important to include demand models that could be specified using modern es
[Im-e revealed timation methods. Then, I have indicated what we know about how to use the informa
tion provided by such models to control and to manage inventories. I have not hesitated
ore, 1 have also to mention rules of thumb, though 1 have tried to be careful to identify them as such.
'" problems. (1 Furthermore, several recent lines of research in inventory theory have exploited the
5 homework.) extraordinary flexibility ofthe linear-systems approach, its capacity to accommodate ex
tensions and elaborations, such as multiple products and stages. In their details these
• models are quite different from the HMMS model, but they are similar in spirit. 1 expect
this style of modeling to become more prominent in the future.
lIlk it does. Finally, this framework has come to playa large role in other areas of management
ory has long and engineering (finance and electrical engineering, to name just two). Indeed, the
ed all of these HMMS model itself has formed the basis ofa good deal of the empirical work on the
14 Foundations offnrentmy Management
,a good thing Today, inventory management encompasses a far wider span ofactivities and issues.
"",~Iedge, not The most striking change has come in our picture of supply activities. We now under
stand much more clearly than before that production and distribution technologies can
improve, and indeed with effort we can drive these improvements. Furthennore, the cu
<tical mulative effects of seemingly mundane reductions in supply leadtimes, production setup
communica times, quality defects, and so forth have a profound impact on the strategic health of the
I lags behind.
organization. This is the central message of the just-in-time (fiT) approach to operations,
perl at a time developed first by Japanese manufacturers, and widely recognized as a major factor in
Ie carried out
their competitive successes.
nanually in a Recently, there has been an explosion of interest in supply-chain management. This
means managing across entire supply networks, even when they cross organizational or
~: I have tried geographical boundaries. This approach offers opportunities for substantial system im
current prac provements, but it also raises new challenges that the traditional inventory-control par
rsrand it, even adigm cannot address.
b1e~
Does this mean that all of our inventory models are obsolete? No. But it does mean
,,,~yious writ that we need to expand and enrich our models in certain directions, and also that we need
ely used, and to analyze, use, and teach even the older models somewhat differently.
ain modes of First, the basic economic fundamentals embodied in these models have not
he",~ Anyone
changed. For example, it has been suggested, frequently and seriously, that the ITT ap
any conceiv proach teaches that every lot size should be one, and thus all the models that calculate
lot sizes should be scrapped. This is utter nonsense. What is true is that, if we succeed in
ldaId. reliable reducing setup costs and leadtimes to negligible values, relative to holding costs and
se methods as other factors, then the lot size shonld be small, just as the models prescribe. Furthermore,
~ time is bet- we should strive to achieve such process improvements, for they lead to tangible overall
benefits, as the models predict. There is no hann in calling attention to these facts with
Ilion of fairly a rhetorical flourish,provided we remember to distinguish flourish from fact. "Inventory
nmching. The is evil;' says another ITT slogan. I say, tell it to tlte squirrels.
here are other Likewise, the kanban system is often promoted as a radically different and superior
substitute for the control methods addressed by the models. This view rests on a misun
~tionspre
derstanding of the kanban system, the earlier methods, or both. In fact, the kanban sys
mat we do to tem is a variant ofcertain older methods, perhaps an improvement, but a variant nonethe
I~ can do is to less, and its use does not fundamentally change the issues of inventory management.
True, the kanban system can be implemented with less centralized short-term in
fonnation and control than we are accustomed to, and this is important, for it has led us
to rethink the whole issue of effective implementation in fresh terms. Nevertheless, the
n-entory man key managerial decisions (setting production rates and numbers of kanbans, for in
im-entory con stance) involve the same logic, the same economic tradeoffs, as traditional systems of
ied in the sup control. (To anyone who thinks these concerns don't matter in Japan, I strongly recom
edietable, then mend reading Shingo [1985], a summary of JlT by one of its inventors. My own vIews
illy. by the en on this and related issues are aired more fully in Zipkin [199Ia].)
im-eotory con Second, prodded in part by the TIT phenomenon, researchers in academia and in
lOoseS to them dustry have been working furiously to extend our models to incorporate such elements
I not an unfair as quality defects and kanbal1-style controls. While this work is by no means complete,
LJe were devel- substantial progress has been made; I have tried to include in the book some of the more
significant and accessible of these recent achievements.
16 Foundations ofInvenfolY Management
Third, in managing a supply chain, it is not enough to know that reducing leadtimes
is a good thing. Even a network of modest size may include thousands of items over
dozens of locations. In each case there may be many possible alternative improvements,
some of them costly, some not. The key questions, then, arc where should improvement
efforts be focused, and what are the anticipated benefits? Similar questions arise con
cerning all sorts of other operational improvement projects. Managers, engineers, and
analysts, whether or not their jobs arc primarily concerned with inventories, are being
called on to answer such questions with increasing frequency.
The models and methods of inventory theory, in my view, constitute the most accu
rate, tractable, thorough, and reliable set of tools now available to address questions of
this nature. To use the tools for this purpose, however, requires a shift of focus. Most of
the models, as noted above, are explicitly addressed to the problems of control, given all
the environmental parameters. We need to ask further, what are the effects of changes in
these parameters? This style of parametric analysis is certaiuly not new; it falls under the
heading of sensitivity analysis. In the pastJ however, sensitivity analysis played a rather
minor role in most expositions of the models. I have placed considerably greater em
phasis on it.
:ing leadtimes typically quite different, after all. One might even say that management and engineering
of items over embody different views of the world, different philosophies. Which of these perspec
mprovements, tives, then, is the most appropriate one from which to fonnulate the conceptual issues
improvement and to handle the practical problems raised by inventories?
oos arIse con In my view the subject of inventories is one place where the arts and sciences of
71gineers, and management and engineering intersect. Both perspectives are necessary to obtain an ad
ies, are being equate grasp of it, whether conceptual or practical. To deal intelligently with inventories
as a professional, a consultant, or a researcher, a person trained in management must
be most accu learn to think like an engineer, while an engineer must become something of a manager.
s questions of The managerial viewpoint is essential, simply because all the major functions of a
0Cus. Most of business, from financial control to sales to production, strongly affect and are strongly
lIrol, given all affected by inventories. Effective inventory management entails at least satisfying cus
,of changes in tomers, keeping capital requirements modest, and avoiding excessive operational costs.
falls under the To balance all of these criteria, one must have some appreciation of where they come
oIayed a rather from and what they mean in a larger context.
Iy greater em- The larger context, the broad view, the big picture, however, are not enongh. The in
ventories maintained by the typical organization and the activities that feed and consume
them compose an enormously complex system. Thousands of items, hundreds of loca
tions, all interrelated, along with records, forecasts, orders and vast amounts of other in
l is commonly fonnation-tms is the usual situation. To cope with such complexity involves a sub
lOry managers stantial task of systems engineering. There is simply no substitute here for the
erween the ac engineering approach to problem solving, the application of basic principles through a
tion ago, and I process of careful, systematic analysis and design. The big pictnre can be misleading if
d practitioners it oversimplifies the complex reality.
Thus, I really do mean the book to be addressed to students of both management and
Ihstanding the engineering. The interplay between the management and engineering points of view is,
mindful of the to my mind at least, one of the fascinations of the whole subject. There is a special chal
..ill someday lenge and reward in forging a conception, a picture, a model that is true to the essential
l\~ tried to ex details of a problem while encompassing the larger whole, particularly when the task in
""l:d from the volves something so critical to all of our economic lives as inventories. If I have con
the theory can veyed some portion of this fascination here, the book will have accomplished one of its
main purposes.
'1 in Chapter 5.
oexposition of
• specific tech
f die model un Notes
s: to structures
•,-antage point,
, it may appear Here is a very brief overview of the development of the theory of inventories. Further
:rpopulartech discussions of specific topics will be found at the end of each chapter.
Modem inventory theory began with the derivation of the EOQ formula by Harris
[1913], an engineer. inventor, and lawyer. Over the next few decades, numerous varia
~eering
tions were elaborated, most printed in popular magazines, written by and for managers.
lis sounds like Many of these results were collected in Raymond [1931], the first published book on in
I engineers are ventory management. The fascinating details of this early history are explored in Er
lenkotter [1989,1990]. Chapter 3 is based on this pioneering work.
18 Foundations qf Inventory Management
These pioneers were certainly aware of the need to plan for uncertainties. However,
the rigorous analysis of models with explicitly stochastic features really began in the
1950s, starting with the seminal papers of Arrow, Harris, and Marschak [1951] and
Dvoretzky, Kiefer, and Wolfowitz [1952]. High points ofthis work are represented in the
books by Whitin [1953], Arrow, Karlin, and Scarf [l958], Morse [1958], and Holt,
Modigliani, Muth, and Simon [1960]. Most of the basic analytical methods of the field
(dynamic programming and stationary analysis, for example) were established during
this period, as were the research-oriented journals in which this work increasingly ap
peared.
The classic, mouumental text of Radley and Whitin [1963], a comprehensive sum
mary of the field to that date, represents the culmination of this period, in a sense, and it
has had a profound effect on all subsequent developments. Appearing shortly afterwards,
the survey of Veinott [1966a] was also influentiaL Much of the work on single-item,
single-location models since then can be seen as elaboration and extension of the ideas
set forth in these works. This is not to say that nothing important has happened; the cu
mulative effect of these elaborations and extensions has been a substantial broadening
and deepening ofthe field.
The study of complex systems with multiple items and stocking locations has pro
gressed somewhat more slowly. Research along these lines began in the late 1950s and
has continued actively ever since. See Scarf, Gilford, and Shelly [1963] for a collection
of influential early papers. The achievements in this area up to about 1970 are summa
rized in Clark [1972], and those of the following decade in Schwarz [1981]. Many im
portant results are quite recent and appear here for the first time in book form.
Good recent surveys of the research literature can be found in Porteus [1990] and
the reviews collected in Graves et aL [1993] and Taynr et aL [1999]. Chilean [1990] has
compiled an extensive bibliography.
c H A p T E R
e5. However,
began in the
, [1951] and
:semed in the
2 SYSTEMS AND MODELS
!]. and Holt,
~ of the field
lisbed during
reasingly ap
bensive SUID
l sense, and it
Iyafterwards,
• single-item,
0. of the ideas
2.1 Introduction
This chapter provides a brief overview of the context of inventory management. It
touches on some oftlle information technologies used in operations, who in the organi
zation should be concerned with inventories and how they should interact with others,
etc. It also presents some observations on models and their uses.
Inventory management employs a wide range of information technologies. These
include databases, cost accounting, and statistical forecasting. It is important that you be
conversant with these technologies, at least enough to be an intelligent consumer of
them. This chapter provides a brief account of what they do, how they relate to inventory
management, and where to get more information about them.
In this respect I am articulating an essentially conservative position. Many books
and articles about inventory management say or imply that you should do your m.. .' ll ac
counting. forecasting, etc. Maybe so, 30 years ago. In those days many companies were
unsophisticated in their uses of these technologies. Since then.. however, the technolo
gies and companies' uses of them have developed considerably. Today, in my view, it
makes no sense to develop, say, your own personal cost accounting system. You are un
likely to do better than the experts, there are organizational and legal dangers in main
taining two sets of books, and you have other tasks to occupy your energies.
19
20 fOllndations ofInventOly Management
2.2 Contexts
2.2.1 Computer Systems
Look at some inventory-control software. Ask to see the systems used in your company
or wllversity. Go to a local retajl store, and ask whether you can see the system used
there. On the World Wide Web (WWW), go to one of the shareware sites, and download
a couple of the inventory programs. Try them out. Do a search for inventory software;
many commercial sites have demonstration packages.
Chances are that what you will find is, essentially, a database. It may be built on a
general-purpose database program (e.g., Access, Paradox, Oracle) or on another type of
application with database functions (e.g., Excel, 1-2-3). Even ifit is a stand-alone pro
gram, however, it will look and act much like a database.
This is the beginning of wisdom about inventory control in practice. An inventory
system is, first andforemost, a database. Such a system is structured as a database, and
much of what it actually does is carried out through database functions.
The simplest and most common inventory systems are structured around a collec
tion (or table) of items. These represent the goods available in a store, or the parts stored
in a repair shop. The items are independent of each other. (Such systems are discussed
in §§5.2 and 8.2. Later, we shall mention more complex databases, where the items are
not independent.) There is a record for each item, containing basic descriptive informa
tion about it (name, code number, etc.). Typically, there are pointers to other structured
information (e.g., a list of suppliers). In addition, there is a transaction history, includ
ing all orders, receipts, and withdrawals for each item; this is analogous to the ledger in
a basic accounting system.
What does the system do? Well, people enter transactions. They add, delete, and
change items. They compile summary reports for management. In sum, they do all the
things that people do with other kinds of databases.
The quality and value ofthe system. moreover, depend to a large extent on how well
it performs its database functions. Is it easy or hard to use, flexible or rigid, dependable
or not? The answers depend entirely on the underlying database. People often complain
about errors in inventory data, but that's true ofall data. Good database programs include
error control capabilities, and these are precisely the teclmiques available to control in
ventory errors, no more and no less.
Nevertheless, an inventory system is special in certain ways, namely, in the nature
ofthe decisions and actions it supports. Specifically, an inventory system triggers orders.
It may do so indirectly, by just signaling when an item's stock is low, leaving the action
to a person. Or, it may place the order automatically. In either case, the system requires
some logic to perform this task. (Other databases can support actions too; for example,
a contact manager may indicate when it is time to phone a certain client.)
This is where models, the subject ofthis book, come in. An inventory model, among
other things, provides the logic required by a computer-based inventory system to trig
ger orders. Although the model typically is a tiny fraction of the code of the system, it is
a very important part. A bad model triggers bad decisions, and bad decisions lead to fu
rious customers, irate suppliers, unhappy bankers, and stressed-out managers.
Chapter 2 Systems and Models 21
By the way, this answers a question that students of inventory theory often raise:
How do you implement an inventory model? In most cases, the answer is simple: Code
it up, and insert the code into an inventory system built on a database. The database pro
gram will take care of most ofthe technical difficulties that implementation entails (nice
our company interface, pretty reports, etc.). (There are other issues in implementation, however, as
system used
discussed below.)
mddownload The answer today is of course very different from what it was 20 years ago. Then,
DrY software;
before database programs were so widely available, it was often necessary to code the
whole system from scratch, and so to worry about interfaces, reports, and all that. (And,
be built on a
since few modelers are good at all those things, the result was a lot of awful systems.)
>Other type of
Today, on the other hand, one must know something about database technology to use it
oo-alone pro- in an intelligent way.
Now, not every existing inventory system will allow you to insert code. Some have
An inventory
the decision logic "hard wired" (that is, buried in an inaccessible place in the program).
database, and
In selecting such a system, therefore, look for one that is sufficiently modular to allow
the decision logic to be revised and replaced. (Be warned, however: No such system is
""'d a collec as flexible as you would like.)
" parts stored Standard inventory systems, as mentioned above, treat the items as independent. In
are discussed
a production setting, however, the relevant items include materials, parts, components,
,the items are
subassemblies, and finished goods. These are not independent. They are related, in that
Jtiye informa
the demand for one is generated by the nceds for others. A different type of software is
Ilcr structured
required for such situations. A material requirements planning (MRP) system is de
istory, includ
signed to capture and use such relationships. Such a system is still a database, but it is
) the ledger in
considerably more complex than an independent-item system. (See Chap. 5 for more in
formation.) There are similar systems for other applications; for example, a distribution
d delete, and
requirements planning (DRP) system applies the same structure to distribution. You can
bey do all the
find a lot ofMRP and DRP software via the www. Try some.
Even more ambitious are enterprise resource planning (ERP) systems. Such sys
Il on how well
tems extend a common database technology to all the information needs of the enter
d dependable
prise. They differ from other databases mainly in scope and scale. Major providers of
lien complain
ERP software include SAP and Baan; many other companies provide related products.
grams include
They are all on the Web too.
, to control 1n
Finally, there are consulting and software companies that provide inventory appli
cations tailored for particular industries.
. . in the nature
Thus, there are a variety of off-the-shelf programs to support inventory manage
'iggers orders.
ment. None may fit a particular business perfectly, but they provide good starting points.
ing the action
There are cases, however, where no existing product adequately captures the real system.
•'Stem requires
Thus, there is still a role for customized systems. Do remember that this is a costly and
,; for example,
hazardous undertaking, and carefully consider the alternatives.
model, among
2.2.2 Monitoring
;ystem to trig
Ie system, it is Inventory is driven by demand and supply. To control an inventory system, it is neces
OIlS lead to fu sary to monitor and record demands and supplies as they occur. Does this sound obvi
gers. ous? Perhaps it is. Still, many companies neglect these basic chores.
22 Foundations oflllventmy Jl,fanagemeflt
Mind you, monitoring can be tricky. Nearly every company adequately records
sales, but demands are something else. A demand is a potential sale. But, if the goods
are unavailable, the customer may go elsewhere. Even if the customer is content to wait,
and purchases the goods when they become available, it is important to record the time
of the original demand as well as that of the later sale. Otherwise, we will have a dis
torted record of the time-pattern of demand.
In some situations it is easy to monitor demands. The customers of a catalogue re
tailer, for example, conununicate with the company only by mail or phone or email. so
it is easy to record each such contact. whether or not it leads to a sale. Likewise. in most
industrial markets, demands arrive in the fann of orders.
In other situations, however. monitoring demands is hard or impossible. Consider a
grocery store. If a customer enters seeking almond-flavored herbal tea, but the store has
none. the customer will leave empty-handed or perhaps buy another flavor. There is no
way to discern the customer's actual demand. There are special statistical methods to es
timate demands from sales data (e.g., Nahmias [1994]), but these are difficult.
There are still other situations where demands can be recorded, but only with some
effort. In a high-end clotlling store, for instance, the sales staff try to talk to every cus
tomer who walks in the door. If a customer wants something that isn't there, it is likely
that a salesperson knows about it. The trick, then, is to train the sales staff to record such
events. Infonnation technologies (e.g., with point-and-click item selection) can make
this process easier.
On the supply side, similarly, it is important to record when each replenishment order
is placed and when it is actually received. (The difference is precisely what we call the lead
time. As you will find throughoutthe book, this is a very important driver ofperformance.)
Obvious? Yes. Does everybody do it faithfully? No. Be sure that your organization does.
Again, modem infonnation technologies make this task much easier and more reli
able than it once was. For example, bar-code scanners have vastly simplified the record
ing of deliveries.
rely records This sort of dilemma confronts most organizations. Indeed, it is inherent in the very
if the goods nature of inventory. Recall. inventories arise at the boundaries between firms, divisions,
llent to wait, or plants, in order to facilitate supply-demand transactions. We should not be surprised
ord the time that all sides want to control them for their own respective purposes. Such contention is
I have a dis natural and inevitable. (As one professional acquaintance put it, inventory management
is part science and part blood sporL)
:atalogue re In truth, there is no universal answer to the question of who should manage inven
or email.so tory. Any of the answers above can work. The key point is, no matter who manages it,
~ise. in most there must be mechanisms in place to ensure that all parties' desires and constraints are
properly acknowledged and balanced.
:. Consider a Many finns recently have created teams to run supply chains, including managing
me store has inventories, comprising people from all the interested groups. Such teams sometimes
~ There is no even span company boundaries. This can be an effective approach. It is important to rec
ethods to es ognize, however, that a team simply localizes the natural contention between groups. The
ult. team must still work out methods to resolve conflicting goals.
~- ",ith some A model typically aims to represent the concerns of different parties in its parame
D every cus ters. The production division's concerns, for example, might be captured through the cost
e. it is likely of setups, while marketing's are represented by a stockout penalty. The model's solution,
'record such then, aims to serve the overall organization's goals by balancing those of the divisions.
n. can make This is good, but remember, compromise is a difficult art. A compromise, even for
the common good, rarely silences all complaints. A model cannot eliminate contention.
shment order On the other hand it can serve as a point of common reference, through which conflicts
call the lead can be worked out.
mormance.)
zarion does.
Dd more reti 2.3 Models
d the record 2.3.1 Overview
Modeling is a wonderful and mysterious art. It underlies much of science and engineer
ing and hence is a basic element of modem life. It is a mode of thinking about and un
derstanding the world and a mode of problem solving. A good model can be beautiful as
;. Pmduction well as useful.
I manage the Still, nobody really understands what a model is. Many have tried to define the term,
none satisfactorily to my mind. Part ofthe problem is that different fields use models in
I::e the goods, different ways. Some ofthe models of physics (e.g., general relativity, quantum electro
ries are there dynamics) are incredibly accurate depictions of the segments of reality they address.
as. So, mar Those models essentially capture all our knowledge oftheir domains. The models of the
ng people al social sciences, in contrast, are far less precise. They aim to capture some important
cuses for not broad features of their domains, while suppressing other less important details. In that
; owners? Fi- setting, the construction of a model requires design decisions-what to include and what
to leave out. Furthennore, engineers and scientists use models for quite different pur
Iseparate de poses. A scientist, whether physical or social, is trying to understand something; an en
KJ<I1d this de gineer is trying to build something that works.
e backwhere Operations research (OR), which includes the subject ofthis book, is more like en
gineering than science, and closer to the social realm than the physical. OR models
24 Foundations ofInventory Management
always leave out more than they include, and their ultimate goal is practical utility, not
pure knowledge. (This point, by the way, explains much of the unfortunate and some
times amusing misunderstanding between operations researchers and economists. The
models they use are quite similar, but their purposes are altogether different.)
Even among operations researchers there is quite a range of styles in making those
design decisions. Some prefer to include as much relevant detail as possible, while oth
ers aim for simplicity above all. Inventory theory tends toward the latter style. Many of
its basic models are radically simple. This approach, then, includes only the most im
portant features of the problem, suppressing everything else. It takes some getting used
to, but it's effective. (I happen to think that these models are beautiful too, but you can
decide for yourself.)
There is another difficulty with the word model. People use it to describe both a piece
of mathematics and a piece of computer software. These are different things. They may be
related, that is, a computer model may implement a mathematical one. But, there are many
ways to implement a given mathematical model, and the software always includes more
than the mathematics. What you will find here in this book are mainly models ofthe math
ematical kind, though many of them have indeed been implemented in software.
What else does the software implementing a model include? A model is a powerful
but crude device. Because it omits much, some deviations between model and reality are
inevitable, and a raw, naked model on its own may behave foolishly. To prevent this, it is
often essential to surround the model with buffers, i.e., various functions to preprocess
the data, postprocess the solution, and so forth. Most successful computer models are
amply buffered in this manner.
Finally, a word about applications. Many have observed that applying a model is
fraught with difficulties. The main reason, I believe, is that a computer model is a piece
of technology. Just like a new type of machine, a computer model requires people to
learn different skills and to work with others in different patterns. Such adjustments are
difficult, costly and time-consuming. It is no wonder that people are conservative in
adopting a new computer model. The model may seem perfectly natural to its creator,
but to the people who have to use it, it's a strange and threatening device. Furthermore,
again like a new machine, a model has political implications; it will increase some peo
ple's power and reduce others'. This is certainly true in the inventory arena, which is in
herently political, as explained above.
Indeed, a vast majority of successful applications have one thing in common: They
have been used elsewhere first. Nobody wants to be the first to try a brand new technol
ogy. Tnlly innovative models are rare.
If you understand these points and undertake applications keeping them in mind,
your chances of success will improve considerably. I'm not saying never innovate. Just
remember that every innovative feature of your model will require more effort in the ap
plication phase. And try to anticipate the technical and political difficulties your users
will face in employing the model.
:aI utility, not come from cost (or managerial) accounting. That discipline is largely concerned with esti
Ie and some mating costs of various kinds, including those used in inventory models. (See, e.g., Zim
oomists. The merman [1997].) This answer, of course, begs the question: What do cost accountants do?
1It.) Well, much of what they do is addition. An item's purchase cost, for example, typi
making those cally consists of several components. There is the actual price charged by the provider,
Ie, while oth taxes, transportation cost, physical handling cost, etc. The accountant's job is to find all
tyle. Many of these sources of cost and add them up. (The total acquisition cost is sometimes called
the most im the "all-in cost," to emphasize that it includes everything, not just the list price.) The
, getting used same goes for the setup cost, holding cost, and so forth.
'. but you can In a system with only one item, this job is straightforward. !t becomes harder, how
ever, when there are htmdreds of thousands of items (not an unusual situation). It is im
e both a piece possible to do a separate study to detennine the cost factors for each item. Moreover, of
They maybe ten many of the cost components are not item-specific, but rather come from shared
Iere are many resources. The truck used for transportation, for example, may carry many items, not just
ocludes more one. Cost accountants have to know division too.
5 of the math For a large, complex system, a cost accountant will try to identifY the key drivers of
.-are. cost, and then to use these to estimate the individual items' costs. Volume (in the geo
is a powerful metric sense), for example, may serve as the driver of transportation cost. That is, deter
00 reality are mine the volume of each item, and divide the total transportation cost among the items
'-ent this, it is by volume. If the results are not precise enough, add another driver, e.g., weight. This
to preprocess approach is sometimes called activity-based costing (ABC). It is entirely analogous to
:r models are regression analysis, in that a few variables (the drivers) are used to estimate many other
quantities (the items' costs).
og a model is There remains a thorny problem. Our models are supposed to represent the effects
del is a piece of alternative actions on the financial condition of the enterprise. In a model where the
res people to order quantity is a variable, for example, the objective function includes the holding cost,
justments are which is supposed to be correct for any possible order quantity. Consequently, the
oservative in holding-cost factor, which multiplies the order quantity, should be a prospective. mar
lO its creator, ginal value; it should measure the per-unit holding cost in the future. The trouble is that
Fmthennore, cost-estimation methods typically provide the historical, average value. They look back
5C some pea ward instead of forward in time. They are based on actual past holding costs, the result
L. "nich is m- of actual past order quantities. The result, as a predictor of future costs, is a pure extrap
olation. Furthennore, such methods allocate the historical costs of shared resources
JIIIIllon: They across items. Even when that allocation is done cleverly, using well-chosen drivers, it
new technol mixes variable and fixed costs together. (Accountants are aware of this problem, and
have tried to devise means to ameliorate it, with modest success.)
bern in mind !tmay be some consolation that we are not the only ones with this problem. Nearly every
innovate. Just sort ofanalysis tries to say something useful about the future based on past data. And so, it is
fort in the ap customary for analysts to complain about accounting numbers (and to hate accountants). But
es your users could we do better? It's hard to see how. The key point is to recognize that cost estimates, even
the most careful ones, are imperfect, and to use them with appropriate caution.
Stockout-penalty costs present a special problem. Whereas inventory holding costs,
say, are internal to the finn, a stockout mainly affects somebody else. One can measure
a holding cost (subject to the caveats above), but it is harder to imagine measuring the
duction setup cost of a stockout to a customer. [n any case, accounting systems typically do not touch
nswer is, they such costs.
26 Foundations ofInventOly Management
Faced with this difficulty, many people simply give up. As we shall see, it is possi
ble to formulate an inventory model without stockout-cost parameters, using constraints
instead, and those people find this approach more comfortable. If you are one of those
people, rest assured, the book includes such fonnulations. My own view, however, is that
this just translates the problem to different terms. The basic problem remains to under
stand what stockouts mean to customers.
This is essentially a problem of market research. Fortunately, marketers have come
to recognize the importance of customer response time and have devised methods to as
sess its importance to customers. (The methods are basically the same as those used to
measure other dimensions of product quality.) We are rapidly passing beyond the point
where the inventory analyst had to pick a number out of the air, whether a penalty cost
or a constraint.
Incidentally, the difficulties of cost estimation have influenced the structure of the
book. There is an emphasis throughout on policy evaluation in physical terms, which
doesn't require cost estimates, before policy optimization, which does. Policy evaluation
is always useful, but especially when reliable cost numbers are hard to get.
e. it is possi This raises another point about monitoring. To assess forecast accuracy, it is essen
g constraints tial to monitor actual demands, as discussed above. But it is equally essential to keep
one of those records of past forecasts. Yon would be amazed how many companies simply discard old
\~,ever, is that forecasts. It is necessary to set up a database to store those old forecasts, but it is well
ins to under worth the effort.
Many discussions of forecasting concentrate on methods, techniques, algorithms.
rs have come "Here are the directions for the exponential-smoothing method...." This emphasis is
"'thods to as misplaced, in my view. It is exactly analogous to learning the technique of regression
those used to analysis without understandiug the basic concept.
ood the poiut The whole purpose ofthe forecasting activity is to construct a concept, or story, or pic
I penalty cost ture, of the world out there, the system that drives demand. The numbers are important, to
be sure, but they make little sense without this broader context. Every forecast is based on
ucture of the a model of demand. That model may be implicit in some cases, but it is there all the same.
terms. which (It is analogous to the linear model which tffiderlies regression analysis.) Every forecast
cyevaluation ing method, moreover, can be interpreted as fitting the parameters of such a model. It is
important to understand the tffiderlying models, because they are quite different. (§C.6.4
in Appendix C discusses the model underlying one popular forecasting method.)
A word about software: If you are actually doiug forecasting, you should get one of
the available packages. This is cheaper and more reliable than writiug your own code.
yofdemand. The manuals and help screens of these packages are, in some cases, good introductions
bout demand to the whole subject offorecastiug.
.nmate. Fore
IJdy statistics.
itics concerns
one particular Notes
:JIems and say
-series data. A The World Wide Web is an excelleut source of information about many of the topics of
o so. the basic this chapter. To find out more about databases, for example. use any of the popular
Take some in search engines with the keyword database.
Itnown some Here are a few web sites that you may find of interest:
Ke.
mber. the best htlp:/lwww.mhhe.com/business/opsci/pom/ This is the Operations Managemeut
IIObably know Center, hosted by our publisher. It has lots of information about operations in general.
.-ise 10 supple http://wwwapics.org/ APICS is the American Production and Inventory Control
as the sample Society. It is a good source for many topics, especially MRP and ERP.
~ mtcrvals. htlp://www.clml.org/ The Cotfficil of Logistics Management, another professional
""e expect de society.
llO look like. htlp://www.injorms.org/ The Institute for Operations Research and the Management
simates, and
http://www.lmi.org/ The Logistics Management Institute, a nonprofit corporation
:De in general,
that advises government agencies on logistics issues.
A CONSTANT
DEMAND RATE
Outline
3.1 Introduction 29 3.5 Quantity Discounts 55
3.2 The Economic-Order-Quantity 3.6 Imperfect Quality 58
Model 30 3.7 The Present-Value Criterion 62
3.3 Planned Backorders 39 Notes 67
3.4 A Finite Production Rate 50 Problems 67
3.1 Introduction
This chapter undertakes a detailed exploration of the meaning and impact of economies
ofscale in supply. Except for this characteristic. the supply and demand processes are as
simple as they could be. The systems here are structurally simple too, involving a single
product at a single location.
Time and the product itself are both continuous, or in other words, infinitely divisi
ble. This is a modeling choice. In reality, products like orange juice, petroleum, wood
pulp and bauxite are infinitely divisible, while airplanes, computers and steel mills are
not. (3.62 gallons of orange juice has a specific value, but a fraction of an airplane is
worthless.) Things like screws, finished paper, and medicine bottles are intemlediate
cases; such items may be thought of as infinitely divisible tor some purposes, even
though they are not really so.
Demand for the item occurs at a known, constant rate, continuously over time; the de
mand over 1.672 years is precisely 1.672 times the annual demand rate. The complicating
features of other demand processes are abstracted away: Demand is smooth, not lumpy;
there are no variations over time, neither predictable nor unpredictable ones; et cetera.
29
30 Foundations ofInventory Management
The supply process is also simple. However, it cannot supply the item at the precise,
constant rate of the demand process. Supply and demand are not perfectly matched, and
they cannot be linked in an ideal flow system. Supply occurs in discrete batches or lots,
because of its nnderlying technology and economics. Each batch, no matter how small,
incurs afixed cost. This is the simplest form of economy of scale. (Much of the chapter
is concerned with how large these lots should be; the topic is sometimes called lot siz
ing.) Also, the supply process responds to replenishment orders only after a fixed lead
time. As we shall see, this feature is not very significant here.
We start with a very basic model in Section 3.2, the economic-order-quantity (EOQ)
model. The key assumption is that demand must be filled immediately; stockouts are for
bidden. The subsequent sections treat important variations of this model.
Section 3.3 relaxes the no-stockout assumption to allow planned backorders.
The next three sections explore variants of the supply system. Section 3.4 focuses
on a production process which produces the item at a constant, finite rate. This model
incorporates a simple capacity limit. Section 3.5 treats more complex economies ofscale
arising from quantity discounts. Section 3.6 examines the effects of imperfect quality.
Section 3.7 refonnulates the EOQ model using a different cost objective. The EOQ
model aims to minimize the long-run average cost over time. This new model is based
instead on the present value of future costs. We then reinterpret the original EOQ model
as an approximation to the new one.
Many other variants of the EOQ model have been developed in response to diverse
practical needs. Some are explored in the problems at the end of the chapter.
The models of this chapter are highly stylized, with all their simplifying assump
tions, but they are nonetheless widely used in practice. Also, the methodology and the
intuition we build up here will provide a solid basis for understanding the more elabo
rate and realistic models presented later.
This chapter requires little mathematics beyond basic calculus.
Before getting started, let us set forth some basic vocabulary: We need two units
of measurement, a unit of time (days, months, years, etc.) and a unit of physical quan
tity (tons, bushels, gallons, etc.). For now we use the abstract tenns "time-units" and
"quantity-units." In a real application, of course, we would choose specific units of
measurement instead.
.,tity (EOQ)
:outs are for
.roers.
i
,E
l 3.4 focuses
This model
nies of scale
~L ~
OCt quality.
,-e. The EOQ
>del is based
EOQmodel ==LJ- L ==j
Time
se to diverse
:r.
ing assump (L does not depend in any way on the time or size of the order. We are assuming, in ef
logy and the fect, that the supplier has unlimited quantities available.) Subsequently, demands further
more elabo deplete the inventory, and we continue to place and receive orders for supplies.
This scenario roughly describes the situation at a retail outlet, which receives fin
ished goods from a supplier and sells them to customers; the leadtime represents order
ed two units processing and transportation time. With minor changes in wording, however, the same
~-sical quan model depicts a simple production process. An "order" becomes a signal to produce a cer
e-units" and tain amount. The entire batch is completed after L time units and then becomes available
ific units of to meet demand. For example, L might represent the time required to procure raw mate
rials and set up a machine, actual production being virtually instantaneous. (This story
presumes no capacity limit. Section 3.4 treats a limited-capacity production process.)
Figure 3.2.1 describes the net effects of supply and demand on inventory. Most of
the time, inventory decreases linearly with slope -t.... When we place an order, nothing
happens until L time units later, when the corresponding batch arrives. At that point, the
inventory jumps up by the amount received.
mand occurs
What would happen if we neglected to order, and inventory dropped to zero? After
that, for a while at least, we would be unable to meet demand. Such stockouts are for
bidden: We always order far enough in advance that stock is available to meet all de
mand. (Section 3.3 drops this restriction.)
A. as stock is
To control an inventory under this scenario means to answer two questions: "U!hen
• accomplish
should an order be placed, and how much should be ordered? (These basic questions
:eific amount
capture the essence of inventory control under a variety of circumstances.)
i:me, denoted
To focus the search for answers, we make two further assumptions. First, subject
to the requirement that all demand be filled immediately, we never order earlier than
32 Foundations oflnventmy Management
necessary. In other words. each order arrives precisely at a moment when we would oth
erwise run out of stock. (lbis is sometimes called the oero-inventory property.) Second,
each order is of the same size. (These rules should seem reasonable. Problems 3.1 and
3.2 provide rigorous support for them.)
Define the following functions (measured in quantity-units) oftime t (measured in
time-units) for t 2: 0:
I(t) ~ inventory at time t
IO(t) ~ inventory on order, the total stock ordered before t but not yet
received by t
IP(t) = inventory position at time t = I(t) + IO(t)
This double-letter notation (f0 and IP) allows us to distinguish different quantities, all re
lated to inventories. By convention all these functions are right-continuous. For example,
if an order arrives at time t, I(t) is the inventory after the arrival, not before. To indicate the
inventory just before the arrival, write I(t-). Figure 3.2.2 depicts these functions.
Also, let
D = AL = demand during a leadtime
Think of time t as now, so the current inventory is I(t). How does it change from t
to t + L1 The new supplies arriving in that interval total exactly IO(t). and the demand
is AL = D. Consequently,
FIGURE 3.2.2
Inventory and inventory position.
o ~ 1(1)
"'I 'J I
Time (r)
Chapter 3 One Item with a Constant Demand Rate 33
OF ~ order frequency
~ limT~~(l/T)(numherof orders in [0, T)}
Notice that in Figure 3.2.2 IP(t) and 1(1) are periodic. We can thus detertnine the
long-run averages I and OF by examining what happens during one cycle, the time in
terval between the receipts of two successive orders. The length of eacb cycle is q/'A, and
there is one order per cycle. Therefore,
1 _.A:
OF = q/'A - q
Also, during a cycle, let) decreases linearly from q to 0, so the average inventory
during the cycle is Xq. Thus,
I = 'Aq
if we pay on ordering, an additional tenn is needed in the total average inventory. This
extra term happens to be D, a constant, independent of q.)
Notice that 7 is increasing in q, while OF is decreasing. These frvo criteria are in
direct conflict: We can exploit economies of scale in the supply process by choosing a
large q, but this leads to large average inventory; we can economize on inventory, but
only at the expense of a high order frequency.
So, how should we select q? There are two basic approaches. Sometimes the world
outside the model imposes a constraint on 7 or OF. Either the inventory or the order fre
quency cannot exceed a prescribed upper limit. In either case it is clear what to do: An
upper bound on 1 implies an upper bound on q. So, to minimize OF, set q as large as
possible, prccisely at its upper bound. Likewise, given a limit on OF, set q so as to hit
the limit precisely.
Without such constraints, the conflict remains. To resolve it, suppose we can trans
late the two criteria to a common scale, monetary cost. That is, we compute the costs en
tailed by orders and inventory, and combine them into an overall cost measure.
To do this, we estimate cost factors for procurement and inventory holding. These
cost factors remain constant over time. We measure all costs in some standard monetary
unit; here, we use the abstract tenn "moneys." Specifically,
k ~ fixed cost to place an order (moneys)
c ~ variable cost to place an order (moneysiquantity-unit)
h = cost to hold one unit in inventory for one unit of time
(moneysi[quantity-unit· time-unit])
The definition of h means, in effect, that at time t an inventory of I(t) causes cost to ac
cumulate at the rate of hI(t).
The estimation of these cost factors is discussed in Chapter 2 and elsewhere in the
book. For now we mention a few basic facts: The fixed cost k represents all order costs
that are independent of the order size. It typically includes administrative order
processing costs as well as transportation and receiving costs. In a production setting k
may include a setup cost. Thus, k reflects economies ofscale in the supply process. The
variable cost c comprises the unit purchase cost as well as any other costs that do depend
on the order size. The total cost per order is thus k + cq.
The holding cost h typically includes two major components. The first comprises all
direct costs associated with inventory itself, including costs for physical handling, insur
ance, refrigeration, and warehouse rental. Denote all these direct costs by II. The second
component is a financing cost (Xc, where a. is an interest rate, reflecting the fact that hold
ing inventory ties up capital. (Section 3.7 discusses this point further.) Thus, Ii = II + etc.
These cost factors, along with the physical performance measures OF and I, deter
mine the average costs per unit time: The long-run average order cost is (k + cq)OF, and
the average inventory-holding cost is hI The overall perfonnance criterion is the sum of
these two quantities:
Chapter 3 One Item with a Constant Demand Rale 35
FIGURE 3.2.3
Total average cost.
\'entory. This
riteria are in
'Y choosing a ",,
,,
.ventory, but ,,
,,
leS the world ,,
the order fre ,,
,
,,
hat to do: An
,,
q as large as
q so as to hit
8 ,, C(q)
,,
,
"'e can trans , Holding cost
,
.the costs en
sure. '"
llding. These Order cost
am monetary
You are responsible for paper supplies in the copying room in a medium-sized office. Paper usage
averages 8 boxes per week. Each box costs $25. The shipping cost per order, regardless of size, is
$50, and receiving an order takes about an hour OfyOUf time, which costs the company $80. Each
order takes a week to arrive. The company figures financing costs at 15%. per year. Keeping the
boxes organized costs about $1.10/week per box. Currently, you order 2 weeks' worth of supplies
with each order. VVhat is the average cost of this policy? Use boxes, weeks, and $ as the units. In
the notation above,
A~8 c ~ 25 k ~ 50 + 80 = 130
0.15
C't ~ - ~ 0.0029
52
= 200 + 65 + 9 ~ 274
rescaling the horizontal axis.) Some of the models of Chapter 5 follow this approach.
Meanwhile, the optimal order interval here can be obtained inilllediately from (3.2.2):
'. Paper usage
~ofsize, is u* ~ Y2kIAh. (3.2.3)
llI)' 580. Each
~ Keeping the
We can implement the optimal policy nsing u* instead of g*: Place the first order L time
th of supplies units before we would otherwise run out of stock. From then on place a new order each
s the units. In u' time units. Under this rule, clearly, the functions IP(t) and 1(1) have precisely the
same values as before. Since the demand rate is constant, we can equivalently monitor
the inventory position or time. (This notion can be expressed in a pleasant little rhyme:
Watch the stock or watch the clock.) In the context of stochastically varying demands,
however, time-based and quantity-based controls are quite different, as we shall see
later.
Substituting (3.2.2) back into I and OF yields
-1 = ~A
-
2h
-OF~ mh
2k
and so
hI=kOF= fi
Thus, the average holding cost and the average (fixed) order cost are equal at the opti
~nowisto mal g ~ g*. By itself this identity is just an odd coincidence, but it leads to a fonnula
L The plan is for the optimal cost:
late it to
zero.
~r. those con
C* ~ C(g*) = CA + y2kAh (3.2.4)
tain C is con
II ~ 0 is nec EXAMPLE 3.2.A, PART 2
:Appendix A
For the paper supply problem above,
For the p~per supply problem above, if >... increases from 8 to 12. then q* increases by the
factor VI ;% 1.22, from 42 to about 51. Returning to A ~ 8, if we can reduce k fioml30 to 100 (by
lowering the shipping cost, not your salary!), then q' becomes (l/W) (42) = 37.
There is another sense in which the EOQ model is robust: Omit the variable pur
chase cost CA from C(q) and C* for the moment. Define the function
E(X) ~l (x + li X>o
2 x/
With a bit of algebra you can show (see Prob. 3.6) that
CCq)
C*
~ E(-'q*L) (3.2.5)
Suppose we use the "wrong" value q instead of q*, because of errors in parameter es
timates, or additional constraints not included in the model, or any other reason. Then,
the relative cost ofthis suboptimal policy, compared to the true optimal cost, depends
only on the relative error in q itself, as indicated by (3.2.5). The formula is entirely in
dependent of the cost and demand parameters. « is sometimes called the EOQ error
function.)
Fonnula (3.2.5) always produces values greater than I (unless of course 'I = '1*).
However, «x) grows slowly as x departs from I, so the relative cost penalty is small, pro
vided q is reasonably close to q*. For example, suppose we overestimate q* by one-third,
so '11'1* ~ 4/3. Then, CCq)/C* = 25/24, a cost penalty of only 1/24 or just over 4%. The
cost penalty is the sarne if 'I underestimates '1* by 25%, so '11'1* ~ 3/4.
Chapter 3 One Item with a Constant Demand Rate 39
This finding has another practical implication: It has been suggested that there are
rate, but advantages to setting lot sizes systematically lower than the EOQ model prescribes.
Lower lot sizes and inventories, the reasoning goes, encourage managers and workers to
the demand
seek opportunities to improve the system. This is one element of the jnst-in-time (JIT)
approach to operations. Now, there is some dispute as to whether this is really an effec
lowerq* by tive way to motivate people. (I'm skeptical; see Zipkin [199Ia].) Even so, the cost
penalty for trying this approach is modest, provided we don't overdo it. For instance, if
ial property we purposely set q to 90% of q*, the cost will increase only slightly. On the other hand,
ely. We must if we set q to 5% of q*, the cost penalty will be substantial indeed.
heless, pro The simple cost formula (3.2.4) also makes it easy to estimate the benefits of oper
A is close to ational improvements in the system. For example, suppose the supply process models a
the relative production activity, and k represents a setup cost. We may be able to reduce k through
es this point engineering changes or training programs. (There is now a fairly well developed tech
'" widely, in nology for reducing setups. This is another element of the JlT approach; see Shingo
[1985], for example.) This will permit us to lower q and hence save on inventory hold
or ofA. Evi ing costs. Assuming we use the appropriate q* both before and after the improvement,
;rohust with we can use (3.2.4) to estimate the overall cost reduction. This approach is discussed fur
ther in § 3.7 and Prohlems 3.7 and 3.27.
It is worth remarking that some parameters have no effects. First, q* does not de
pend at all on the variable-cost parameter c. The intuitive reason is that the long-run
<eases by the
average supply rate must equal the demand rate A, so the average variable order cost
.30 to 100 (by must be cA, a constant independent of q. (However, as mentioned above, h usually de
pends partly on c, so changing c does change q* indirectly.) Also, neither q* nor C*
depends on the leadtime L. The leadtime merely introduces a delay between actions
,Mable pur (orders) and their effects. If the leadtime were to increase, we would need to order
earlier, but this would happen automatically via the trigger quantity D = AL in the or
der rule. All this reflects the fact that, given the current inventory position, we know
exactly when we will run out of stock. Thus, according to this model, there is no ad
vantage to reducing the leadtime. (Of course, such perfect foresight rests on the as
sumption of perfectly known leadtimes and demands. As we shall see later, in a sto
chastic system the leadtime does have an impact, and leadtime reduction is valuable
(3.2.5) indeed.)
arameter es
easOll. Then,
3.3 Planned Backorders
os!, depends
) entirely in 3.3.1 The Setting
: EOQ error Consider the same EOQ system, but relax the requirement that all demand be met from
stock on hand. All demands are ultinlately filled, though perhaps after a delay. That is,
rse q = q*). demands not filled immediately are backordered. Customers are willing to wait, and we
is small, pro are conunitted to meet their demands. (An alternative scenario is lost sales, where de
by one-third, mands not filled inunediately are never filled. We discuss this case below in Section
"'cr 4%. The 3.3.7.) We always use any inventory on hand to fill demands; backorders accumulate
only when we run out of stock entirely.
40 Foundations ofbrfentory Management
= IP(t) - D
This relation is the direct analogue of (3.2.1), with INO replacing I(} It is the
conservation-of-flow law for this system: Between t and t + L, IO(t) gets added to the
net inventory, and D gets subtracted. Thus, IP(t) summarizes all the information needed
to predict the net inventory a leadtirne into the future.
Chapter 3 One Item with a Constant Demand Rate 41
ainIy true of
ensive (or, in
in other mar
e.
i:
II Inventory
ion 3.2:
.~
~
o
Backorders
Time (I)
At any given
, stock to fill
As in the EOQ model, assume that all orders are ofthe same size q > 0. The issue of
when to order is now more complex. We need a second policy variable in addition to q:
r = reorder point (quantity-units)
This variable can take on any real value, positive or negative. Consider the following
policy:
00 they TImc Monitor the inventory position fP(t) constantly. When IP(C) = r, place a new order of size
n1 rate A, re q at time t.
•tt) jumps up
(The EOQ model's policies are special cases with r = D.) In honor of the two variahles,
~ and the rest
a policy of this kind is called a reorder-point/order-quantity or (r, q) policy. Figure 3.3.2
D backorders
illustrates the behavior of IN(t) and IP(t) under such a policy. Note the similarity of Fig
'i(t) is some
ures 3.3.2 and 3.2.2. The graph retains the same sawtooth pattern, but now the pattern
can shift vertically, depending on the choice of r.
(3.3.1)
3.3.3 Performance Criteria
The relevant criteria still include I and OF; but we need more. While the model allows
o. It is the backorders J that does not mean they are welcome. Customers do not like to wait, and if
added to the we hope to please them, we had better not make them wait too often or too long. More
l3tion needed over, just as we do not have to pay for orders prior to delivery, so customers typically de
lay payments until their demands are filled. Sometimes there is even a direct monetary
42 Fou.ndations ofInventory Management
FIGURE 3.3.2
IP(t) and IN(I} under an (r, q) policy.
-
,: '" , IP(t)
,, " ,
, "
: "'"
,S
,
[N(t)
o
Time (t)
cost for backorders, as in a contractual penalty for late delivery. For a variety of reasons,
then, we need to measure and control backorders.
The primary backorder-related performance measure is the following:
]j = long-term average outstanding backorders
~ limT-+~{(l/T)J6B(t} dt}
(The limit here is analogous to the oue defining 7.) There is another criterion which plays
a secondary but still important role:
A(t) ~ 1{IN(t) :5 O}
A = long-run fraction of time out of stock, or stockout frequency
= limT-+~{(lIT)f[,A(t) dt}
(Here, 1 {.} is the indicator function and A(t) is an indicator variable, taking the value I
when IN(t) :5 0 and 0 otherwise.) Clearly, A also measures the fraction ofdemands back
ordered. Thus, the average number of demands backordered per unit time is i\.A, and the
fill rate, the fraction of demands met from stock, is 1 - A. (Terminology alert: We often
call]l the backorders for short. Because A measures the backorder frequency, however,
some writers cali it the backorders) referring to Ii as the time-weighted backorders.)
Let us now compute the criteria as functions ofthe policy variables q and 1: It is con
venient to replace r by the equivalent variable
v ~ safety stock ~ r - D
Chapter 3 One [tern with a Constanf Demand Rate 43
(The phrase safety stock suggests a positive quantity, but v can be negative! Still, this def
inition is consistent with standard usage in stochastic-demand models. There, the safety
stock is usually positive.) Again, call a cycle the time between the receipts of successive
orders, and u the cycle time. Also, define a time-equivalent to V,
. v
y = safety tIme ~ :;,:
FIGURE 3.3.3
The Mo parts ofa cycle.
q+v
[NUl
u+y -y
o
o u
Time (t)
factors k, c, and h defined earlier. Suppose we can also estimate a factor for backorders,
analogous to h:
b ~ penalty
cost for one unit backordered during one time-unit
(moneys/[quantity-unit . time-unit])
This parameter sununarizes all the drawbacks of backorders mentioned above. The total
average cost then becomes
C(v, q) = (k + cq)OF + hI + bE
k>" I h(q + V)2 I by;.
= c>.. + - + - +~-
q 2 q 2 q
(This formulation does not explicitly inciudeA. The backorder cost results, not from the
occurrence of a backorder, but rather from its continuation in time. The reason for this
focus will become clear later.)
JC h(g + v) bv
+-=0
Jv g g
2
JC _ k'A/g 2 + I h(g2 - v ) I bV
Jg 2 i - 27 = 0
Defining the cost ratio
b
w=b+h
g* = rm. (i
Yh"yw (3.3.2)
v* = -(1 - w)g*
The optimal reorder point is r* = D + v*. The optimal cycle time and safety time are
given by
u* ~ fik fI
YM y-;;;
y* ~ -(1 - w)u*
The sensitivity analysis of this model is similar to that of the EOQ model. Again, q*
and C* are robust with respect to changes in A and k, because of the square-root opera
tions. Also, think of w as a relative-cost index, independent of h itself. With the param
eters viewed in this way, the optimal solution is robust with respect to w and h.
Also, fix q to any value, and suppose we set the safety stock optimally, to v ~
-(1 - w)q. Let C(q) denote the cost of this solution, excluding the term CA. One can
show that
kA I
C(q) = - + -hwq (3.3.5)
q 2
This is precisely C(q) in the EOQ model, except for the adjusted inventory cost. The ra
tio C(q)/C* thus reduces to E(q/q*), as in (3.2.5). Consequently, there is little cost penalty
for a modest deviation of q from q*. Now, suppose q = q*, but we choose v ¥- v*. The
cost C(v, q*) is a quadratic function of v. Thus, a small deviation ofv from v* does not
increase the cost much, but a large one does.
lurion: To see the impact ofthis constraint, fix q for now, leaving v variable. The order fre
quency OF is independent of v, so apart from A itself, the only relevant criterion is 7.
(3.3.4) Now, 7 is increasing in v, whileA is decreasing. So, starting from v = 0, we want to re
duce v (thus lowering 7 and raising A) as far as the constraint permits. That is, set v =
xmtrols it any -(I - w_)q, so that A ~ I - w_. Thus, the inequality we began with should be satis
fied as an equality. At this value of v, 7 = Xw':q.
Recall, the solution to the cost-optimization model above yields the similar identity
(3.3.4), A ~ I - w. The cost ratio w plays the same role there as the fill-rate limit w_.
:i This can mean Assuuting we set vas above, the total cost becomes (3.3.5), with w': replacing w. With
bboring copying the same replacement, formula (3.3.2) gives the optimal q.
tae other data are To conclude, there is no fundamental difference between the cost-minimization and
. q* increases to service-constraint approaches. The fraction uL enforces essentially the same effect as a
um249toC* = backorder cost. (Technically, the constrained model is equivalent to a cost-optimization
model with the same h, an inflated fixed cost klw_, and b chosen so that w = bI(b + h) =
UL.) It may be more convenient sometimes to specify uL than b, but each one implies a
>del. Again, q*
value for the other. The basic managerial issue of balancing competing objectives remains
He-root opera
I~nh the param
the sarne, regardless of which approach is used.
andh.
imally, to v ~
3.3.6 Space and Time
m CA. One can
The quantities A and B are related to the service experienced by customers, but it is in
structive to consider a more direct measure:
(3.3.5)
BW = average customer backorder waiting time.
~-cost. The ra Again, assume -q :5 v :5 O. A glance at Figure 3.3.3 should convince you that there is
de cost penalty no waiting in the first part of a cycle, while in the second part, after stock runs out, the
sev-=l=v*.The average wait is X( -y) ~ X( -viA). Thus, the overall average is
m 1"* does not
- q (0) + (-V)[I(-V)]
_BW ~ (q+v) - q 2: T
~ measurement
This is an important conclusion: The average waiting time of a customer demand is
proportional to the average outstanding backorders, with constant of proportionality
OIltrolofback III.., the reciprocal of the demand rate. This is one reason why the measure B is
measures, usu so important. (We shall see later that this same relation holds for a wide variety of
alculation of a models.)
L The fraction
EXAMPLE 3.3.A, PART 2
leteJTI1ined ser
ifferent service For the paper supply problem above, under the optimal policy,B = ~(-3)2/44 = 0.10 boxes, so
BW = 0.1018 = 0.013 weeks.
48 Foundations ofInventory Management
The average inventory I is related similarly to what we call the average stocking
time, denoted IW, the time a unit of physical goods remains in inventory, from the mo
ment its order is received to the instant it is used to fill demand. That is, II).. = [w. (The
reciprocal of the stocking time 1/ IW= 'A/I is called the inventory turnover, or turnover
for short. This measure is widely used in practice.)
These connections suggest an alternative way to think about the tradeoff between
performance criteria: Instead of the physical quantities measured by I and 8, we can fo
cus on the equivalent time measures IWand BW. In these tenns, we can choose to re
duce customer waiting by keeping stock on hand longer. The order frequency OF is just
the reciprocal of the cycle time u, which is already a time-oriented measure.
These are more than mathematical tricks: An inventory or a backlog is the net result
of the interactions between the times of various events. For example, [(t) is the number
of physical units whose arrivals (due to order receipts) precede t, but whose departures
(due to demands) occur after t. In a very real sense, all o.finventory management con
cerns the management oftime. (The time dimension is crucial in other managerial are
nas as well; see Stalk and Hout [1990].) This perspective underlies the specification of
inventory-related cost factors such as h, as discussed in Section 3.7.
There are two cases to consider, depending on the sign of the numerator of the last term:
1. b2 > 2k)"h: The nunleraror is negarive, so to minimize C(q) we want q as small
as possible. But we musr srop ar q ~ blh with v = O. Thus, v* = 0, and so q* is
just the EOQ formula. [More precisely, one can show that any solution (v, q)
with v < 0 has cost at least C(q) > C(blh) > c).. + (2k1..h)1I2.]
2
2. b '" 2k)"h: Here, we want q to be large. But there is no upper limit, so the
model seems to prescribe q* = 00, which is meaningless. What is going on? In
this case, b < C(q) for any q. The average cost of never ordering (b) is less
than the cost of any other policy, so it is optimal never to order.
Chapter 3 One Item with a Constant Demand Rate 49
erage stocking
Perhaps there are situations where this odd conclusion makes sense. In my opinion,
; from the mo
however, the result is an artifact of an odd model: The model's cost is linear inA, even
~ = [w. (The
for A near 0 and I. This is not credible; in real systems, a small stockout frequency is
er; or turnover
usually tolerable, whereas a large one means disaster. The optimization exploits this lin
earity all it can, resulting in the extreme solution above. (In technical terms, A is not a
ideoff between
convex function of v and q, hence neither is C. Strange things can happen when we min
:I B. we can fo
imize a nonconvex function.) In contrast, the model of Section 3.3.5, which constrains
1 choose to re
.fl, yields a perfectly reasonable answer. By restricting A to a narrow range, it avoids the
ncy OF is just
problems above. That model is equivalent to a cost-minimization model, but one based
Jre.
onB, not A.
~ the net result
Moreover, although A is a useful and widely used customer-service measure, it con
Iis the number
veys only limited information. Do customers really care only whether they have to wait,
ose departures
not how long? Rarely. Thus,B captures more of customers' actual experience thanA. For
'Wgement con
all these reasons, we focus more on Ii.
l3ll3.gerial are
(lCCification of
Then, q = ql + Vi :5 q',
The order cost is now k + c(q) ~ k + c(q' + v'). Suppose we incur the penalty b for
each lost sale. (Equivalently, b is the unit cost for the emergency supply source.) Then,
the total average cost is
C(v', q') = [k + c(q' + v')]OF + hI + bAA
~ (k + cq')OF + hI + (b - c)AA
This has exactly the same form as the model in the prior subsection, assuming backorders
but using A to measure service. Consequently, the solution is also the same: Either never
lose sales, or abandon the business. (Problem 3.13 asks you to work out the details.)
q I~~",,""
1i"
'k.
IIflprocess. The
~f a batch is in
f stream of Qut
o pu Time u
_ I )
Ie is inflexible. I~-(l-pq
, "ith demand, 2
Dee production Now, assume that each production star! incurs the fixed cost k, the variable produc
=rnand are per tion cost is c, and h measures the output-inventory-holding cost, as in the EOQ model.
I of the process (There is no holding cost for inputs.) The total average cost is then
k'A I
C(q) = c'A + - + -h(l - p)q (3.4.1)
q 2
inventory hits
3.4.3 The Optimal Policy and Sensitivity Analysis
" EOQ model:
The function ceq) here has precisely the same form as in the EOQ model: The constant
os the time be h there becomes h( I - p). Therefore,
ioti when pra
~
k'A
leis agamu = q* - (3.4.2)
h(l - p)
ength is qlfL =
C* = C(q*) = c'A + V2k'Ah(l - p) (3.4.3)
[mum value of
Observe that q* is larger than in the EOQ model: For any q, the average inventory is
lower here, because stock accumulates gradually. The solution exploits this effect to re
>)q
duce the order cost. The EOQ model is a limiting case of this one with fL ~ 00.
52 Foundations ofInventory Management
Consider an English soft-drink bottler. It measures physical quantities in thousands of cases, time
in days, and money in £. By an amazing coincidence, the data in these units are exactly the same
as in the paper supply problem above, i.e., A = 8, k = 130, and h = 1.173. The bottler's process
ing rate is !-L = 10, so p = 0.8. So,
*~ [ 2(130)(8) ]112 = 94
q (1.173)(0.2)
q*
u*=-=12
l
You may have been thinking of the leadtime L as a setup time, but there is a crucial dif
ference: When the batch size q is small, there may be several orders outstanding at any
given moment; that is, severalleadtimes may overlap. Furthennore, a production run can
begin during the leadtime of a later batch. In short, the leadtime L does not tie up the pro
duction facility. During a setup, in contrast, there can be no output, and certainly setups
cannot overlap.
This distinction is sometimes expressed as the difference between internal and ex
ternal setups. An internal setup is one that ties up the production facility; this is what we
call a "setup." An external setup can take place while the facility is doing other work; in
our terms an external setup is part of the leadtime. Sometimes it is possible in practice
Chapter 3 One Item with a Constant Demand Rate 53
Compare this formula with the optimal cost in the earlier case (3.4.2). You may have
found it odd that previously the optimal cost was increasing in the production rate; ac
cording to that model, faster production eosts more! There, a slower production rate
tends to smooth the inventory, while imposing no delays whatsoever in our capacity to
meet demand. Such delays do occur in the current model, and additional stock is re
quired to cover them. As a result, the total cost is now decreasing in the production rate.
This difference is no minor technical point. The basic production economics are to
tally different in the two cases. It is important to know whether an "improvement," like
an increase in the production rate, will help or hurt cost performance!
EXAMPLE 3.4.A, PART 3
In this case, assuming a negligible setup time 'T, the bottler's optimal solution is
• ~ [ 2(130)(8) ]11' = 3
q (Ll73)lL8) 1
q*
u* = -=4
A
C* ~ 200 + [2(130)(8)(1.1 73)(1.8)]112 = 266
If'T = 3, as above, the lower bound on q is J20, &0 q* = 120. The cost of this solution is c* =
335, a substantial increase over 266.
Chapter 3 One Item with a Constant Demond Rate 55
Improvements in the details ofmaterials handling can have major benefits. Ifwe can
change the system so that output is available immediately, we switch from the current
) meet demand. model to the earlier one with correspondingly reduced costs. (This is true even though
ing production. the switch increases the optimal batch size, contrary to what you might suppose.)
:eo batch is com
:ause some fin-
3.5 Quantity Discounts
eel the demand
l3Ild in this pe 3.5.1 Alternative Specifications ofPurchase Costs
:: maXlIDUffi oc
In the basic EOQ model of Section 3.2, the variable cost c is constant for orders of all
:nt to cover de sizes. However, it is common for suppliers to offer price breaks for large orders. This sec
. the maximum tion shows how to extend the EOQ model to incorporate such quantity discounts. Actu
ally, there are two kinds of discounts, incremental and all-units.
the average of Consider the incremental case first: Suppose the purchase price changes at the
breakpoint X. The variable cost is Co for any amount up to x. For an order larger than X,
the additional amount over Xincurs the rate Cll where Cl < Co. Thus, the total order cost
is k + c(q), where
coq O<q:sx
c(q) ~ ( CoX + Cl(q - X) q>x
Alternatively, define k o = k and k l = k + (co - Cj)X. Then, the order cost is
ko + coq O<q:Sx
k + c(q) ~ ( k + Cjq
j
q> X
_·you may have This function describes more elaborate economies of scale than the original fixed-plus
lCtion rate; ac linear form. In a production setting, it represents costs that depend on the production
rnduction rate quantity in a complex, nonlinear way.
)IJf capacity to
With an all-units discount, if q > X, the entire order is priced at the lower rate Cll so
131 stock is re
roduction rate. coq O<q<x
IOOmics are to c(q) = clq
( q'="x
O\.-ement," like
Figure 3.5.1 illustrates c(q) for both cases.
FIGURE 3.5.1
Order cos! with a quantity discount.
i
Incremental
~
"
All units
Evidently, Cb(q) > Ci(q), so q; < q~. There remain three possible cases:
1. q~ <qr::;X
2. q~ < X < qf
3. X::; q~ < qt
In case 1, for q 2: X,
Co(q) O<q<X
C(q) =
( Cr(q) q 2: X
3.5.4 Discussio1l
The overall effect of a quantity discount, compared to the case where the variable cost c =
ore strictly con
Co is constant, can only be to increase order sizes. Depending on the other parameters, ei
respective min-
ther there is no effect (q* = q ~), or the optimal order size is larger than it would have been
otherwise.
Why then do suppliers offer quantity discounts in practice? The supplier may
achieve economies of scale from large orders. Or, the supplier may hope to attract new
customers with the lower price and/or to induce existing customers to buy more. Re
searchers have investigated these and other mechanisms to explain price breaks, but
there is no clear consensus so far on which are the primary reasons. In fact, it is not easy
to construct a plausible scenario where a quantity discount makes sense.
58 Foundations o/lm'enrOly Management
My own experience suggests that, often, suppliers offer discounts more out of habit
than reasoned analysis. ("It's stupid, but our customers are used to it," one told me.) It
may be possible and mutually beneficial to negotiate a revised cost schedule without
quantity discounts. The supplier and the customer might agree to a constant cost rate c,
say some average of Co and Cj, leaving both parties better off. The model above can help
guide such discussions. (See Problem 3.17 for an illustration.)
nal delivery to Case 2 (immediate detection, no reimbursement). Sometimes, alas, we pay for
defects do oc all units ordered, defective or not. To yield q, we must again order q/~, so the
ouseless. Also, order cost becomes k + (cI~)q. In effect, the cost rate c is inflated by the factor
I/~. Recall, changing c does not directly affect the optimal policy, but it does so
In most cases indirectly through h. The latter includes a direct cost !l and a financing cost ae,
e solution fe so h now becomes
r quality does
ae
lOwever, arise h=h+
- ~
her stochastic
,apter 7 illus Thus, q* is smaller than it would be with 0 = 0, because h is larger. (Again, we
must inflate q* as above to determine the actual order size. The net effect,
. and when are clearly, is an increase in the order size.) Also, the square-root term in C* is
ral places. The larger than before, and of course the tenn CA also increases to (c/~)A. Not
mary itself, as surprisingly, when we pay for defects, they do increase the overall cost.
u. in some 8it Case 3 (delayed detection, no reimbursement). Now, suppose we do not discover
d only at some defects on receipt of a batch. Instead, the defective units are placed in inventory
do matter. It is along with the good ones. The defects are found at the moment of demand.
~er; see Juran
(Imagine, for example, that demands originate from a subsequent processing
1ft just to eval
stage, which is able to detect and reject defective units before using them.)
ection is phys Then, we discard or return the defective units, as before, without
reimbursement. Assume the defects are evenly spread through the inventory, so
nits? If we are at each time 1 it includes exactly 0/(1) of defective product. Thus, inventory is
depletcd by demand at rate A/~ instead of A.
nowhere else. This model too reduces to the EOQ model with revised parameters:
ods the finite Again, q indicates the yield of each batch. Also, reinterpret 1(1) as the
:t. which tends
nondefective stock only. This good inventory is depleted by demand at rate A.
So, 1(1) behaves just as in the EOQ model with no defects. Accordingly,
replace!l by !J/~, since the whole inventory incurs direct handling costs. Also,
divide C and ac by~, as in case 2. Thus, the entire holding cost h is inflated
balch the sup by the factor I/~, i.e.,
,,1 of the batch h~--
h + ae
~
60 Foundations ofInventory Management
(2i)..
q*~...;~p/~)
Otherwise (as in Section 3.4.5). q* decreases:
2kll.
q* =
h(l + P/ ~)
(Again. q* indicates the yield of a production run, so the total run size is q*/~.)
All this assumes that the effective production rate ~I-' remains larger than the de
mand rate lI.. If not, the problem becomes entirely infeasible. Thus, defects can bave
the drastic effect oftransfonning an adequate production process into one that is un
able to meet demand.
Now, suppose there is a positive setup time, 1'. Even if A < ~J-L, so the problem is still
feasible, the optimal policy can change radically. The setup time imposes a lower limit
Chapter 3 One Item with a Constallt Demand Rate 61
are<! to their on q, now TA/(l - p!~). This bound increases in S, and if the original p is near I, it in
min C* are creases sharply. In that case q* and C* increase sharply too.
e inlI. We are thus led to an observation, which in retrospect is obvious, but worth emphasiz
me scenario ing nonetheless: Defects waste production capacity. No such effect was apparent in the con
]y, we pay text of the EOQ model, for its supply process has unlimited capacity. The consequences of
""ed for such waste depend on the details ofthe problem, but in general they are unwelcome indeed.
-e precisely
requires the 3.6.4 Perishable Products
'and the
rot does Now, suppose that rhe supply system operates perfectly, but defects arise while the prod
uct is held in inventory. Actually, this is true of virtually all items to some extent. Boxes
are dropped or mislabeled; containers and foofs leak. The focus here, however, is on per
nely adjusted ishable products, such as food, medicines, certain chemicals, and blood in blood banks.
'lSpection and Such products, because of their physical nature, systematically deteriorate over time.
hough not by The actual pattern of deterioration over time can be quite complex and largely un
s. (It is worth predictable; also, it may be difficult to determine whether a particular unit is spoiled or
c system. The not. Partly for these reasons, and also because the consequences of using spoiled goods
can be catastrophic (think of food poisoning), it is quite common to find a limit on the
time a unit can remain in storage. Here, for practical purposes, all the inventory becomes
defective when this time limit is reached.
It is simple to incorporate such a time limit in the EOQ model. The limit becomes an
1£ themoment upper bound u + on the cycle time u. Equivalently, constrain q:S Au +. To solve this model,
fective goods follow a procedure like that of Section 3.4.4: Compute q* with the EOQ formula (3.2.2),
are produced ignoring the constraint. Check whether q* satisfies the constraint. If not, reset q* = Au +.
When there is no time limit, we need a model to describe spoilage. Here is the sim
Jt have to pay plest model: Inventory decays at a constant rate over time, independent of its age or past
:ss belongs to
experience. That is, redefining S to be rhe decay rate, inventory deteriorates at rate SI(t)
precisely like for all t. For reasons explained shortly, this is sometimes called the exponential decay
,mes p!~. The model. Assume the defective product is detected immediately and discarded, and we pay
eter shifts do for all goods without reimbursement, as in case 2 above.
immediately, Even for this simple case, the analysis and the results are fairly complex. As in the
EOQ model, suppose every order is of size q, placed so as to arrive just when stock
would orherwise run out. So, [(t) is periodic. Assuming time 0 is the beginning of a cy
cle, [(0) = q. Wirhin rhe cycle, taking both spoilage and demand into account, we have
l'(t) = - 3[(t) - A (3.6.1 )
(The prime indicates a derivative.) This is a first-order linear differential equation with
exogenous input and initial condition [(0) = q. Its solution, as you can check directly, is
[(t) = qe-8, - (A/S)(I - e -")
l·!~·)
r rhan the de (The phrase exponential decay refers to rhe e -8t terms.)
eets can have Now, the cycle time u is just that value oft wirh [(t) = O. It is convenientto use u as the
,ne that is un decision variable. Given u, q ~ (A/S)(e8u - I). It is not hard to show (Problem 3.19) that
_ (A/S)(e8u - Su - I)
roblem is still [= (3.6.2)
; a lower limit Su
62 Foulldations ofII/vel/lory Management
tion of muitiplyiug f by the fraction e -a' is called discounting; another name for the
(3.6.3) present value is the discounted value.
Suppose that we can also borrow money at rate cx. If we must pay a cost of 1 now,
we can borrow the money, wait until time t, and then pay the larger amount ert'. Simi
- I)
larly, an obligation to pay fat time t is equivalent to a cost of e -a'[today. This quantity
is called the discounted cost. Thus, a cost works just like a negative cash flow.
L :&xcq has the
Now, suppose we face a sequence of cash flows Vi} at times {tj}' Some of these j';
reasornng may
may be positive and some negative, representing costs. As above, we can replace each
.j To minimize
cash flow by its present value. The sum of these present values LJ exp (-a0JJ is the net
!<s.) This equa
present value of the entire sequence. Given the assumptions above, this provides an un
db. a computer.
ambiguous means of ordering and choosing among alternative sequences.
n: For small &,
This idea extends directly to a continuous cash-flow stream. Letf(t) be the rate at
ned in Section
which cash flow accumulates at time t, positive or negative. The net present value ofthis
os!, forcing q*
stream is JO e -a'[(t) dr.
dial losses.
These assumptions are rather special. In reality, for example, one cannot borrow and
invest at the same interest rate; the borrowing rate is higher. Still, the present-value cri
terion does reflect the fact that a cost today is more burdensome than a cost tomorrow.
This is, I believe, a real advantage over the average-cost measure.
Not everyone agrees. The present-value logic puts little weight on events far into the
e. This seems future. Critics charge that this leads people and companies to make shortsighted decisions.
•licies? A pol Defenders counter that the real problem lies elsewhere. Enough already; you get tile idea.
1Iy capture the For a vivid debate on this issue, see Hayes and Garvin [1982] and Kaplan [1986].
erent perspec
1. and use it to 3.7.3 Formulation and Analysis
can be under Let us reformulate the EOQ model using this new objective. Assume for simplicity that
there is no direct handling cost and no leadtime, i.e.,!l ~ L = O. (Problem 3.25 shows
mancmgcom how to incorporate a positive h..) Also, as before, the order size is the constant q, and the
from the time cycle time is the constant u.
:h gaps, firms We incur the cost k + cq = k + c'Au when we receive each order. We sell goods in
based but not response to demand at price p, so we receive a stream of revenue at the constant rate p'A.
tory underlies Suppose l(O-) ~ 0, so we order at time O. Thus, we order at times t ~ 0, u, 2u, . .. , in
ing in a ware curring cost k + c'Au every time. The present value of all these costs, or the total dis
inancing costs counted cost, is thus
Ie.
Also, the present value of the revenue stream is fo e -a'(pA) dt ~ pAia. The net present
value of all the cash flows is thus pAla - qu). The revenue term is a constant, so to
maximize the net present value, we can equivalently minimize the discounted cost.
~-est money at
To optimize C(u), we solve C'(u) ~ O. (This works; the function C has the right
. now (time 0)
properties. See Problem 3.22.) A bit of algebra reduces this equation to
initial invest
Or- a cash flow f(au) = k
(3.7.1)
M'. The opera- a CA
64 Foundations ofInvent01Y Management
where
f(x) = eX - x-I
Unfortunately, there is no formula for the solution of equation (3.7.1). Fortunately, it is
easy to solve numerically with a computer. It is interesting that the parameters k, c, and
A appear only in the ratio IdcA.
u* = (2k
y~
or
rm
q * = '-/---;;;;
FIGURE 3.7.1
Discounted versus average cost.
6
nmately, it is
* *-*'*
:leIS k, c, and
5
*-*-*'
4
*-*
leI. However,
]' *-*'
"n!,f(x) just
b *'
1~
'B
3
0
2
in (3.7.1) are
_ . . _ Average 12.5%
nes,fla,,) =
____ DiscDunted 125%
- -G - Average 25%
-e- Discounted 25%
o
o 0.5 15 2
Cost-demand ratio
is nearly 2 years. Thus, for items ordered several times per year (i.e., most cases of prac
tical interest), the EOQ approximation is reasonably accurate. Only for an extremely
gilt-hand side low-cost, low-usage item (with small cAl is this not so.
. . Also, remember that the cost penalty for choosing a suboptimal q in the average
mcreasmgm
e u* is always cost model is small. The same is true, it turns out, in the present-value model. So, ifwe
mula. calculate" and q using the "wrong" objective, the resulting cost penalty is usually small.
el ahove. This The EOQ model can be somewhat misleading, however, as the basis for economic
Jrnlula is eas analysis. For example, suppose we can purchase some technology (e.g., production
t on concrete, machinery or communications equipment) to reduce the fixed cost k. This new tech
K"the present- nology requires a one-time expenditure but no ongoing costs. Is the reduction in k
worth the cost?
solution? We To address this question with the EOQ model, one must convert the average cost
tFigure3.7.!. (3.2.4) into an equivalent discounted cost. Imagine, as an approximation, a constant
ilinst the ratio cash-flow stream with this value. The present value of this stream is
od 25% (a =
ed, the differ COCk) = ro e-a'(cA + y2kAh)
- cA J2kAC
dt = - + -
a a
g larger when
where h = ac. This function estimates the operating cost over time for any fixed k. To
at the optimal evaluate the technology, compare its cost to the change in COCk). With the present-value
only when u* criterion, we don't need to imagine or approximate. For each k, compute the optimal u*,
66 Foundations of!nventmy Management
FIGURE 3.7.2
Discounted versus overage cost.
16
***-**-**-*
12
-*-*-**
"8
~
'£'
o
8 _s_f}--S
s_fr-s - fr
!""'- _8- £]-
El-~
plug it into cell). and call the result COCk). Use this function instead of COCk). CThere is
no simple fonnula for it.)
Figure 3.7.2 compares the fimctions COCk) and COCk) for the same two values of lX.
Here, c = A = I, so the ratio k/cA. is now just k. The differences become significant ear
lier, for smaller valnes of k, than in the previous figure. (This is not surprising, because
COCk) > k, while COCk) grows only as the square root of k.) In fact Csee Problem 3.26),
for any k j < k2 ,
C*Ck2 ) - COCk,) > COCk,.) - COCk,)
That is, assuming C* measures the true operating cost, CO always underestimates the
benefits of the technology. An analysis based on CO might reject the investment, even
though it is really worthwhile according to C*.
The rest of this book includes various models based on one or the other objective.
In some places Ce.g., Chapter 9), results for both criteria are presented. Others focus on
just one, but mainly for the sake of brevity. In most cases it is possible to construct and
analyze a model based on either criterion. And the two models usually yield similar re
sults for a small interest rate. In practice, the choice of objective is a matter of modeling
art. These criteria each have their own advantages, as discussed above. The choice is
yours.
--
Notes
Much ofthe material in this chapter was developed in the early twentieth century, as out
lined in the Notes to Chapter 1. Since then, the literature has continued to accumulate.
It is now vast and still growing. I know of no single source that surveys it all. Useful
points of entry include Aggarwal [1974] and Lee and Nahmias [1993].
Recent investigations of quantity discounts include Monahan [1984] and Weng
[1995]. The material on imperfect quality in Section 3.6, straightforward as it is, is
mostly original here. Alternative scenarios are explored by Porleus [1986b] and Lee and
Rosenblatt [1987]. Nahmias [1982] reviews the literature on perishable products.
The discussions of process improvement (e.g., setup reduction, quality improve
ment) are inspired, in spirit ifnot in detail, by Porleus [1985, 1986a, 1986b].
Problems
3.1 In Section 3.2.1 we assume that ao order is placed at time I only if 1(1-) = 0 (the zero-inventory
property). This problem suggests why. For simplicity assume that L ~ 0 and 1(0-) = 10(0-) = 0, so
we must order at time O. Any policy can be described through an increasing sequence of order times
{I,: i 2: I} and a corresponding sequence {q,: i 2: l} of positive order sizes. Consider only feasible
policies, with 1(1) 2: 0, and those with finitely maoy order times I, in every finite interval.
Consider a policy that violates the zero-inventory assumption. Choosej to be the smallest i for
k). (There is which 1(17) > O. Construct a new policy by delaying the order at time 0 until t; ~ mini 0 + I(li)/A,
t;.
lj+ 1 J. Show how to adjust some of the qj if necessary, so that I(t) remains the same as before for t ?:
...'3lues of ct. Argue tliat the new policy is feasible and dominates the old one, i.e., its cost over every finite time
LJificant ear interval is lower.
[fig, because 3.2 This problem justifies another assumption of Section 3.2.1, that all orders are ofthe same size. Consider
"'lem 3.26), a policy with the zero-inventory property. Let OF(I) be the number of orders and C(I) the total cost in
the time interval [0, I). (In the notation of the previous problem, OF(I) = max{i: I, < I }.) Fix I aod
OF(I). Prove that, among all feasible policies with this OF(I), the one that orders equal amounts
q = AlIOF(I) at equal intervals u ~ lIOF(I) minimizes cel). (Set up aod solve an optimization problem
mmates the with variables q, I :0; i :0; OF(I).) Also, show that the minimal value of C(I) is [klu + 'AhAu ]1. Explain
ment, even tliat, for every feasible policy and all I, cel) 2: (2kAh)1/21.
3.3 Consider the function ceq) in Section 3.2.2, the average total cost in the EOQ model. Show that this
:r objective. fimction is strictly convex in q.
~ focus on 3.4 In the EOQ model of Section 3.2, suppose the fixed cost to place an order is $100, demand is
mstruct and 50 tons/year, and the holding cost is $0.015385/pound-week. (There are 2000 pounds/ton and
i similar re 52 weeks/year.) Calculate the optimal order size and the optimal average cost.
>fmodeling 3.5 Consider the following variation on the EOQ model: We own storage space for only h quantity
Ie choice is units. Whenever the inventory exceeds 1+, we have to rent additional space to store the excess. The
cost rate for this rented space is h+ > h. Show that the total average cost now becomes
-
k'A I I h ( , - I') I h( - I )2
C(q)~c'A+-+-hq+{-~-+-- q + ifq>I+l
q 2 2 q 2 q
Show that this function is continuously differentiable and convex. (Be careful: C is not twice
continuously differentiable at q = I+.)
Argue that q* can be computed as follows: First, compute q* as in the EOQ fonnula. If q* -;;
1+. stop. Otherwise, reset
J2k'A - hI~
q* =h +
h+
3.6 VerilY equation (3.2.5) describing the relative cost of a suboptimal q in the EOQ model. Evaluate
this quantity for qlq* = 2 and 1/\12. (Notice. these values are small.)
3.7 Define the function F(x) over x > 0 by
F(x) ~ ax -u + bx~
where a, a, b. and 13 are positive constants. Consider the problem of minimizing F over x. Show that
the unique optimal solution is x* = (aa/R»b)l/(a+~). (We have not assumed 13 :2:: 1) so F is not
necessarily convex!) Show also thatF(x*) ~ (a + 13)[(a/j3)~(bla)"]l/(u+~)
(This problem generalizes the EOQ model and several others in the text; in these cases x = q,
and a = 13 = 1. Here is another application: The optimal average cost C* in the EOQ model is a
simple function of the fixed cost k; setting x = k gives c* ~ bx~, where b = (2'Ah)l/2 and 13~ "
Now, suppose the current fixed cost is Xo, and there is a range of investments we can make to reduce
it. The investment required to achieve any value of x < Xo is ax -a - ax oU.
Thus, the problem of
selecting the amount to invest reduces to the problem above.)
3.8 Throughout this chapter we have assumed that the product is infinitely divisible and demand occurs
continuously. This problem demonstrates that the EOQ model, with a slight modification, also
describes a discrete scenario.
Suppose that demands occur at discrete points in time, spaced 1/'1.. apart, in units of size 1. We
order batches of size q, a positive integer, so that each batch arrives at the latest feasible moment.
Show that OF = Vq, 1 =~(q - I), and
k'A I
C(q) ~ c'A +- +- h(q - I)
q 2
Denote the first difference operator Il.C(q) = C(q + 1) - C(q). The optimal q ~ q* satisfies Il.C(q - I)
< 0 -;; Il.C(q). (This is the analogue of C'(q) = 0 in the continuous setting.) Show that this condition is
equivalent to
2k'A
q(q-I)<--;;q(q+ 1)
h
Let q~ indicate the EOQ fommla, the optimal q for continuous demand. Argue that, if q~ happens to be
an integer, then q* = q~, and otherwise q* is obtained by rounding q~ to an integer, either up or down.
The same idea works, clearly, when the demand size is some number Yother than 1. (Just
redefine the quantity unit to recover unit demands.) Now, return to the continuous scenario, but
suppose that the order interval u is constrained to be integral. Show how to compute u*, adapting
the methods above. Finally, explain what to do when u must be an integer multiple of some
arbitrary base period, IJ..
3.9 In the model with planned backorders of Section 3.3, suppose the parameters are
wice
r. ~ 1000 k= 60 h = 0,75 w = 0,81
"- If q* <;
Compute the optimal policy and the optimal average cost (excluding cr.). Now, suppose we install a
computer system to process orders, reducing k from 60 to 33,75. What are the effects on the
optimal policy and the optimal cost?
3.10 Define the limctions C(x) = h[xt + b[xr and C_(y) ~ C(y - D). Show that the cost function
for the EOQ model with backorders can be expressed as
Evaluate
C(r, q) = cr. + kr. + J;+q C (y) dy
q
3.11 Here is another approach to backorders: Instead of imposing a cost penalty or a constraint, we
explicitly model the effects of waiting time BW on the market for the product. BW affects the price
I. Show that
p that customers are willing to pay, or the amount they are willing to purchase, as expressed in the
5 not
demand rate r., or both. Specifically, p, r., and BW are related by the equation
lSeSX=q, r. = a[p!(BWW~
odel is a
Idll=l{· where a and Il are positive constants with Il > I, and!is an increasing limction withf(O) = 1. So,
ke to reduce as BWgets larger, r. declines or p declines or both. (Think ofp!(BW) as the total per-unit cost to
"'lem of customers.) The per-unit purchase cost c remains fixed.
Use the time variables u and y to describe an inventory policy. Recall, BW = l{/Iu. First, consider
[)3Dd occurs u, y. and hence BW fixed. So, it remains only to choose p. or equivalently r.. Calculate the value of r.
!l,also that maximizes the total average profit P(r.), that is, the average revenue pr. minus all relevant costs.
Now, treat u and y as well as r. as variables. Formulate a function P(u, y, r.) = total average
profit per unit time. Write the first-order necessary conditions for an overall optimal policy. (Do
size L We
not attempt to solve them.)
moment.
3.12 In the planned-backorders model, suppose there is no penalty cost b. but rather an upper limit on
BW, say BW+. The problem then becomes
kr. I h(q + vf
Minimize cr. +- +
q 2 q
:s .1C(q - I)
-q :'S V :'S 0
Argue that the last pair of inequalities can be eliminated, and the remaining constraint reduces to
l{V'lq = f.BW+. Introduce a Lagrange multiplier (dual variable) 'IT for this equation, define an
oppens to be appropriate Lagrangian function, and derive the first-order necessary conditions for an optimum. Use
IIp or down. these to obtain fonnulas for v* and q* in tenns of 1T*, and a simple polynomial equation for 1T itself.
I. (Just Argue that 'IT plays essentially the same role here as b does in a pure cost-optimization model.
Jario, but 3.13 The optimal policy for the lost-sales model of Section 3.3.8 is one of two extremes: Either never
" adapting lose sales or never order. When precisely does each case apply?
70 Foundations ofInventory Management
3.14 Consider the model of Section 3.4.5 with a finite production rate fL, where product becomes
available to meet demand only after a whole batch is complete. Show that, to avoid running out of
stock, we must begin production when I(t) = pq. Also, show that, when a batch is completed, I(t)
= q. Draw a graph of I(t) versus t. Show that Y = ~(I + P)q.
3.15 In the model of Section 3.4 suppose the parameters are
Co = 80 Cj ~ 60 Xo ~ 314
Now, suppose the supplier approaches us with a proposal to set the variable cost C to 62, regardless
of the order size. Determine whether this new arrangement is beneficial for us. Why might the
supplier prefer it? (Hint: Compare the new optimal policy to the earlier one.)
3.18 The EOQ model with imperfect quality of Section 3.6.2 presumes a fixed defect rate. Suppose now
that the defect-generation process depends on the order size. Each increment has its own defect
rate o(x), and the total defect quantity in an order of size q is Ll(q) = f6 o(x) dx. Assume that o(x) is
increasing in x. (Think of the supply system as a production process. As it works on an order, the
system deteriorates, and its defect rate grows.) On the other hand, a defective unit is not entirely
worthless; instead of scrapping it, we can and do fix it at cost cf' where 0 < cf < c. In effect, we
receive the full amount q in each order, but we pay the extra cost c/!.(q).
Write down ceq), the total average cost, and compute its derivative C(q). Argue that the
equation C'(q) = 0 has a unique positive solution, q*' (Do not try to show that C is convex; it need
not be. Instead, reduce the equation to an equivalent one with cl
on the left-hand side and another
function of q on the right-hand side. Assume that o(x) is differentiable, and show that the right
hand side is decreasing in q.) Show that q* is less than the value given by the EOQ formula.
3.19 Verify the expression (3.6.2) for Yin the perishable-product model. Hint: Integrate both sides of
(3.6.1), and use the fact that Y ~ (l/u)f~ I(t) dt.
3.20 For the perishable-product model of Section 3.6.4 show that the average cost ceu) in (3.6.3) is
increasing as a function of 0 for each fixed value of u. Then, argue that the optimal cost C* is
increasing in o.
3.21 This problem generalizes the exponential-decay model of Section 3.6.4. Presuming a cycle begins
at time 0, inventory deteriorates during the cycle at the rate o(t)I(t), where the decay rate o(t) is
some specified function of time, no longer a constant. Set up a differential equation to describe the
,
.... " .. ,
'"'.
....
~,
comes evolution of I(t), analogous to (3.6.1). Define d(t) = fl, S(s) ds. Verify that the following function
"Pleted, I(t)
Use this to express q and 7 in terms of u. Specialize these results to a linear decay rate, that is,
Set) = St for some constant S. (Leave 7 as an integral; don't try to express it in closed form.)
3.22 Consider the function C(u) ~ (k + cAu)/(l - e -au), the objective function of the present-value
the optimal
model of Section 3.7. Show that qu) is strictly convex. Hint: Show that the functiong(x) =
:ches can be
x(1 + e -j - 2(1 - e -j is positive for x> O. Further hint: Clearly, g(O) = 0; show that g'(x) "" O.
:mandrateA
3.23 Again, consider the function qu) of Section 3.7. This problem explores its behavior as a function
>e breakpoint,
of ct, treating II as fixed. First, argue that lil1\..---:oo C(u) = 00 for any II > 0. Next, compute lima---:oo
uqu). (Use I'Hopital's rule from calculus.) Compare the result with the EOQ model.
CI>'" >
c; the highest 3.24 In Seetion 3.7 we assume thatI(O-) = O. Now assume we begin with inventory 1(0-) = I o > O.
~-e Xn-l' Write Argue that the lotal discounted cost now becomes exp [-ufo/A] qu), where qu) is defined as
Ill! policy. before, asslUlling I o = O. Conclude that the policy chosen to minimize qu) is optimal here as well.
mpute the 3.25 In the discounted-cost model of Section 3.7, suppose that, in addition to the order costs, there is a
direct inventory-handling cost rate II, so cost is incurred continuously at rate !lI(t). Show that the
total-cost function C(u) now has the additional term
ctue-au
2
II(Vu )[(uu - I) +" au)]
52, regardless Differentiate q u) to show that the optimal u satisfies the equation
night the
f(uu)=~
Suppose now u A(ne + II)
IOIIl defect
wheref(x) = eX - x-I. [This reduces to equation (3.7.1) when II = 0.] Now, approximatefas in
we that B(x) is
the text to obtain an approximate value for u*. Compare this with the EOQ formula.
III order, the
3.26 This problem explores further the relation between the present-value optimal cost function C*(k)
lOl entirely
and the average-cost function COCk) in Section 3.7. Let u*(k) be the optimal u for given kin the
I effect, we
present-value model. Show that
bat the dC*(k)
JlIVex; it need dk I - e-au*(k)
aiIld another
the right Use this expression, and the fact that e -(Xu> 1 - ctU, II > 0, to show that
<TImla. dC*(k) dCo(k)
oth sides of -->--- k>O
dk dk
Explain that this implies
(3.6.3) is
)Sl C* is C*(k2 ) - C*(kj ) > CO(k2 ) - CO(kil
for k, < k,., as asserted in the text. Also, show that u*(k) --> 00 as k -> 00, so
cycle begins
ateB(t) is dC*(k) =I dCo(k) ~ 0
o describe the limk---:oO'O dk limk-;~ dk
72 Foundations ofIlJ"lientory Management
(So, in the limit, C*(k) grows linearly at rate I, while COCk) becomes essentially flat. Thus, the two
functions behave quite differently fur large k.)
3.27 The present-value model of Section 3.7 forbids backorders, and revenue accrues at the constant
rate ph.. Now, relax this restriction. Suppose we follow an order-quantityJreorder-point policy, as in
Section 3.3, with -q < v < O. We receive revenue only when a demand increment is actually met.
There is no direct backorder-penalty cost.
Derive an expression for the total present value of profit, as a function of u and y. where
y = viA. is the safety time. Differentiate this function to obtain two equations describing u' and y'.
What can you say about these values, compared to the optimal policy in the average-cost case?
(Section 3.7 observes that the usual inventory-holding cost consists of two parts, a physical
handling cost and a financing cost. The present-value model represents the financing cost directly,
through the discounting of order costs. We see here that there is a financing component in the
hackorder-penalty cost too, to account for delayed receipt ofrevenues.)
3.28 Suppose there is a finite production rate fL, as in Section 3.4, and the variable production cost is
incurred continuously at rate qL while production actually occurs. Analyze this system using the
present-value criterion. (Assume there is no direct holding cost, and each unit produced is
immediately available to meet demand.) Show that the total discounted cost is
k + (cA./O'p)(1 - e-·P")
cell) = I - e ."
~ constant
4 TIME-VARYING
: policy, as in
.ctually met. DEMANDS
i\'nere
19 u* andy*.
[)st case?
physical
:ost directly,
rt in the
ion cost is
• using the
rl is
Outline
4.1 Introduction 73 4.4 Extensions 93
4.2 Extreme Cases 74 4.5 Smoothing Models 99
4.3 The Dynamic Economic Lotsize Notes 103
>aldina
represents 4.1 Introduction
overed only at This chapter considers situations where demand changes over time. It also allows for
for defective time-varying purchase costs. Real demand rates and purchase costs do change, of
mat it course, because of seasonal factors, or growing or shrinking markets, or other shifts in
de that. in the economic landscape. The central managerial dilemma, as in Chapter 3, is to find the
~ as if there
right balance between the cost of holding inventory and the cost of ordering, including
scale economies. This issue is complicated here by temporal variations.
In practice, our current picture offuture demand is a forecast. It is important to men
I/; of the
tion that the models ofthis chapter treat such forecasts as perfect. Demands change, but
re produced.
the changes are entirely predictable.
defective or
Section 4.2 explores some simple cases. A crude, heuristic analysis reveals the
ely.)
broad effects oftirne-varying demands. The rest of the chapter is quite different in style.
"em 3.28 as It aims for exact answers to the general problem. This program is more challenging here
"the average-
than in Chapter 3. Partly for this reason, starting in Section 4.3, we model time as dis
crete instead of continuous, with a finite time horizon. We derive algorithms to compute
solutions, but we shall see few simple, appealing formulas. This emphasis on computa
73
-
74 Foundations ofInventory Management
tion continues throughout Section 4.3 and Section 4.4. Numerical examples relate the re
sults back to the intuitive findings of Section 4.2.
Section 4.5 introduces a smoothing model. This model represents diseconomies of
scale, the opposite of economies of scale, where large orders incur substantial cost
penalties. The major virtues of the model-and its primary faults-are certain funda
mental symmetries in its fonnulation. These symmetries are not very realistic, but they
lead to wonderfully simple, interesting results.
Q
e demand rate
:eneral case of '"
~ solutions can
: special cases
3Ilds.
there is some
",1- (This con
cise range for Time (t)
I1eA is almost
I measures the 4.2.3 Slow Changes
eu.C(u)isa
different poli Next, consider the opposite case: Suppose 1..(1) changes slowly. The idea, roughly, is that
od its optimal i\.(t) remain nearly constant over every plausible order intervaL For instance, define
(2k
) demand-rate u(I) ~ "';).ji)h
. Here, the de
e more robust This is just the optimal order interval, computed for each t as if the demand rate were
constant at the current value A(t). One way to define a slow-changing demand rate is to
stipulate that 1..0 change only a little over the interval [I, I + u(t)], for all f. (Evidently,
this condition depends on the cost ratio kill.)
Here is a plausible policy: Order only when stock would run out otherwise. and if I
two cases re is such a moment, order the quantity
J2kA(I)
lin, we define q(l) = -11- (4.2.l)
small interval
xl equal to A. That is, use the EOQ formula, treating the demand rate as constant at its current value.
:: average over Alternatively, order just enough to meet demand until I + u(I).
agmg process Why should this heuristic work? Here is an intuitive argument: Consider a rime pe
riod of intermediate length, covering several order cycles but no large shifts in 1..(1) and
only tiny rip q( I). lrA(l) really were constant, the EOQ model's optimal g* would nearly minimize the
he costs of al average cost over this period. Actually, A(t) does change, but only slightly, and each g(l)
'model as is. is near q*. So, the heuristic policy nearly minimizes the average cost over the period.
76 Foundation.~ afInventory Mmagemellt
But, this is true [or all such periods. Thus, the policy nearly minimizes the overall aver
age cost.
Of course, this is not a rigorous proof, and a good deal of work would be required
to make it so. Nevertheless, the basic idea is clear: The demand rate will change signif
icantly only in the distant future, and we can ignore such changes in the current order I,;y
cleo (This is a myopic heuristic, one that focuses on the near future only. Such methods
are effective in a variety of situations. We shall encounter them again later.)
The same reasoning suggests an approximation C* ofthe optimal average cost: Fol
lowing (3.2.4), define
C*(t) = C),(I) + Y2k),(I)h (4.2.2)
C* = limT-->oo(~)fb C*(I) dl
(Think of C*(I) as the approximate instantaneous cost rate. The integral can be evaluated
exactly, of course, only for certain special forms of),(I).)
This model inherits the robustness properties of the EOQ model. For instance, '1(1)
and C* change only slightly in response to a modest change in Ie, and if the entire func
tion ),(1) doubles. each '1(t) increases only by \12,
as does C*. Moreover, '1(1) is dynami
cally robust as well. That is, consider one fixed function ),(1). If),(I) changes modestly
over some interval of time, then q(t) changes even less over that interval.
For this reason, an even cruder heuristic works fairly weB: Keep the order size (or
the cycle time) constant most ofthe time. Change it only when q(l) (or U(I», as compnted
above, departs too far from the current policy. It is often easier to administer a policy
whose variables change infrequently.
4.2.4 Discussion
To summarize: [fthe changes in A(t) are small and/or rapid, focus on the long-rnn av
erage)" ignoring the current demand rate. If ),(1) evolves slowly, focus on the current
),(1), and ignore longer-run changes. The numerical examples of Section 4.3.5 snpport
these suggestions.
We can even combine these three cases: Suppose A(t) is the sum of three functions,
a small-, a fast-, and a slow-changing one. Thal is, suppose the local average li(f) does
change overtime, but only slowly. Following the arguments above, a reasonable heuris
tic policy is to set q(l) according to (4.2.1), using Mi) instead of ),(1).
This is still a rather extreme case. What about more typical cases, where A(t) in
cludes changes of substantial size that are neither slow nor fast, but somewhere in be
tween? The following crude guideline appears to work well: Suppose ),(1) varies within
a certain range. Compute the corresponding range of values for the EOQ formula. and
set each order quantity within this range.
Of course, when the range of ),(1) is wide, this guideline leaves a great deal of lee
way. As dIscussed below, there seem to be no more precise, simple, reliable guidelines
for finding a truly optimal or even close-to-optimal policy. Algorithmic methods; exact
or approximate, appear to be essential.
---
Chapter 4 Time- Varying Demands 77
dIe overall aver- 4.2.5 A Finite Production Rate
MJld be required Now, suppose there is a finite production rate IJ.., as in Section 3.4. Small or rapid changes
1 change signif in A(t) can again be ignored, following the arguments above. Consider a slow-changing
urrent order cy demand rate. For simplicity, assume X(t) is continuous.
'. Such methods The simplest case is where A(t) < fl. for all t. The heuristic approach above can be
leT.) adapted directly: Define pet) = A(t)/fl., and redefine
'erage cost: Fol
2kA(t)
get) = (4.2.3)
h[1 - pet)]
(4.2.2)
combining (4.2.1) with (3.4.2). Arguing as in Section 4.2.3, the policy nsing the get) as batch
sizes will likely perform reasonably well. (This formula presnmes that stock produced can
be used immediately to meet demand. Otherwise, revise it, replacing I - p(l) by
I + p(l) as in Section 3.4.5.)
an be evaluated
This approach will not work, obvionsly, when A(t) > fl. for some I. In that case, de
mand exceeds capacity, at least in the short run, and we must somehow anticipate such
J< instance, get)
heavy-load periods. Figure 4.2.2 illnstrates the situation. The horizontal line represents
the entire func
the capacity fl., and the curve is A(t). (The notation will be explained in a moment.)
q( t) is dynami
Let D(t) ~ fh A(s) ds denote the cnmulative demand. Feasibility requires D(t) :0; fl.t
mges modestly
for all t. Suppose there is a finite time t+, such that A(t) :0; fl. for t'" t+. There are no ca
pacity shortages after t+, and we can apply (4.2.3) as above; the difficulties come before
e order size (or
1+. Choose 1+ as small as possible. By continnity, A(I+) ~ fl., and for t below but near
)), as computed
t+, A(t) > fl. and
inister a policy
D(t+) - D(t) > fl.(t+ - t)
Reducing I further, this inequality certainly holds as long as A(t) > fl., and it continues
to hold for a while even when A(t) falls below fl.. At some point, however, cumulative
supply catches up with cumulative demand, and
Ielong-nm av
on the current D(I+) - D(t) = fl.(t+ - t)
,4.3.5 support Set I- to be the Imgesl t < t + satisfying this equation. Thus, if we should arrive at time
I- with no inventory, we musl produce continuously at least until t+. The lightly shaded
hree functions, area in Figure 4.2.2 is the total excess demand before t+, which must be anticipated by
:rage Mt) does earlier production, and I- is the point where the darker area equals the light one.
;onable heuris Clearly, A(I-) < fl.. For the moment, suppose A(t) :0; fl. for all I :0; 1-, as in Figure 4.2.2.
In this case the original heuristic fails only during the interval [I-, t+J. This suggests the
where A(t) in following modified heuristic: Apply (4.2.3) untilI-, then produce continuously until t+,
leWhere in be and from then on retnrn to (4.2.3). (Minor adjustments may be needed to hook up the long
) varies within production run in [I-, I+J with the short ones before and after.)
1formula, and Conversely, suppose that some I < I- has A(t) > fl.. Thus, at some time before 1-,
the sitnation looks just like Figure 4.2.2. So, repeat the same approach: Identify a new
",t deal oflee interval like [L, t+], and schedule production around it in the same way. Continue in this
ble guidelines manner to identify such intervals, each requiring a relatively long production cycle to
oethods, exact compensate for its short-term capacity shortage. Outside these intervals, apply (4.2.3) as
above. (This procedure, like everything else in this section, is a heuristic.)
-
F,GURE 4.2.2
Excess demand and capacity.
Capacity
Demand rate
The broad effect of these extended production cycles on performance is clear: The
early part of each cycle builds up inventory in order to meet demand in the later part
when demand is heaviest. Thus, we end up holding larger inventories than we would oth
erwise. (This effect can be quantified; see Problem 4.2.)
The same ideas cover situations where the production rate also changes over time,
i.e., '" ~ ",(t). One special case deserves mention, where ",(t) is proportional to A(t), so
p(t) = A(t)/",(t) is the constant p < 1. Here, capacity expands and contracts in direct re
sponse to demand variations. Equation (4.2.3) reduces to
2kA(t)
q(t) =
h(l - p)
As in the uncapacitated model, q(t) varies as the square root of A(t).
ThlS discussion covers extreme cases only. The general case requires intricate com
putational methods; see Section 4.4.5.
ity, we introduce a very different conception of time. We model time as a finite sequence
of discrete time points. We casually refer to time periods, the intervals between the time
points, but in this view of the world all the significant events occur precisely at the time
points themselves.
There are situations where discrete time describes reality better than continuous
time. For instance, it may well be that, by prior arrangement, we can order only at cer
tain times, say once per week on Mondays at 9:00 A.M. Alternatively, a discrete-time
model can be employed to approximate a continuous-time system. (As in any such dis
cretization, of course, the approximation is more accurate when the time points are close
together.) We may think of the time periods as being of equal length, though this is not
necessary. In fact, it is sometimes better to use short periods near the beginning and
longer ones toward the horizon.
The world of discrete time takes some getting used to; the terminology, notation,
and overall style are rather different from anything we have seen before. Indeed, the con
"
nection to continuous-time models may not be immediately apparent. Rest assured, the
connection exists, as explained later.
The problem, in words, is the following: Demand for a product occurs at each of
several time points. We must meet all demand as it occurs; no backorders or lost sales
are pennitted. We can order or produce supplies at each point, and we can carry inven
tory from one point to the next. Replenishment decisions take effect inuuediately; there
is no order leadtime. Each order incurs a fixed cost regardless of the size of the order,
and also a variable cost proportional to the order quantity. There is a cost to hold inven
tory between any two successive points. These costs as well as the demands may change
over time. The goal is to detennine a feasible ordering plan which minimizes the total
cost over all time points.
:e is clear: The
To forrnnlate this problem, denote
• the later part
I we would oth T = time horizon
I = index for tmoe points, I = 0, ... , T
:ages over time,
The horizon T is finite. Time period I is the time from point I until just before I + I. The
oml to ~(I), so
data of the problem include
cts in direct re
d(l) = demand at tmoe I
These are arbitrary nonnegative constants. The decision variables are
x(l) = inventory at tmoe I
z(l) ~ order size at tmoe I
;intricate com The starting inventory at time t = 0 is the known constant Xo.
We use the notation d(l) instead of ~(I) to emphasize that demand occurs in discrete
chunks, not continuously. Also, we use z(1) for the order size, not q(I). We reserve q(l) to de
scribe a policy, indicating the amount to order if certain conditions are met, whereas z(1)
specifies precisely what to order with no conditions attached. This is the most common us
age in the literature on discrete-tmoe models, perhaps because z(l) looks more like a deci
:Ie the demand sion variable than q(I). For similar reasons we denote inventory by x(l) instead of 1(1). (In
,-el of general Section 4.4, X(I) will have the more general meaning of net inventory or inventory position.)
80 Foundations o/Inventory Management
Initial conditions:
x(O) = Xo (4.3.1)
Dynamics:
x(t + I) ~ x(l) + z(t) - d(t) 1= 0, ... , T - 1 (4.3.2)
Constraints:
X(I) "' 0 t = 1, ... , T (4.3.3 )
Objeclive:
Minimize 2:;;;10 [k{1)8(Z(I)) + c(t)z(t)] + 2:;~1 h(t)x(l) (4.3.4)
This optimization problem is called the dynamic economic lolsize (DEL) model. It is
also called the Wagner· Whilin problem after its inventors, Wagner and Whitin [1958]. (It
is a special instance of a general modeling framework for discrete-time dynamical sys
lems and optimal conlrol. In those terms, X(I) describes the slale ofthe system, z(l) is the
conlrol variable. and d(l) is the exogenous inpul. See Appendix B. The dynamics (4.3.2)
express a conservation-of-flow law for this system.)
There are alternative scenarios, where time is less strictly discrete, which lead to the
same mathematical model. For instance, as explained in Section 4.4.3, holding costs can
accrue continuously throughout the period. The crucial assumption is the restriction of
order decisions to the discrete time points.
Here is some additional notation, used below:
Chapter 4 Time- Varying Demands 81
Thus, the DEL model can be expressed as a linear mixed-integer program. (This refor
mulation is interesting; the same trick can be applied to more complex systems, as in
Chapter 5. But it is not immediately practical, for integer programs are hard to solve.)
Z*(I). This fact is useful: Sometimes il is important that the z(l) be integral, though the
DEL model itself contains no such restrictions. It doesn't need them; they hold auto
(4.3.7) matically. This is even true, it turns out, when Xo is positive but integral.)
PROOF OF THEOREM 4.3.1. For simplicity suppose Xo = 0, all d(l) > 0, and all
{ in brackets is 1«1) > O. For this case, we prove that every optimal solution has the zero-inventory prop
= I + 1. (Prob erty. (Problem 4.4 asks you to prove the result in the general case.) Call an order time
res the smaller I bad if it violates the assertion, that is, if x(l) > O. Consider a feasible solution with a
) need to com bad order time. Let u be the smallest one and I the order time just prior to u. (This I ex
gh to meet de ists, because 0 is a non-bad order time.)
his latter prop- We now construct a new feasible solution, in which u is no longer a bad order time,
and whose cost is less than the original's, by shifting some of z(u) to Z(I) or vice versa.
sent actual or Define
thus replacing c[l, II) == c[l, II) - C(II)
This is the change in the overall costlo increment z(l) by I, reduce z(1/) by I, and adjust
ilolding inven
the inventories between I and " accordingly [as determined by the dynamics (4.3.2)].
oeriod sr') to I. Consider two cases, depending on the sign of this quantity: If c[l, II) -; 0, modifying the
.l1.1) = I for all
solution in this way improves the cost (or at worst leaves it unchanged), and further such
opportunities.
changes are equally beneficial. So, make the largest possible such change; that is, re
ties like fixed
place z(l) by z(I)+z(u) and reduce z(u) to O. so that u is no longer an order time. The to
tal cost change is c[l, u)z(u) -; 0, and we also save the fixed cost 1«1/). On the other hand,
if c[l, 1/) > 0, make the maximal change in the opposite direction. That is, increase z(1/)
by x(1/), reduce z(l) by the same amOlmt, and again adjust the inventories between I and
u; this change drives x(1/) itself to 0, so u is no longer bad. This reduces the cost by
,.to-solve non
c[l, 1/)x(1/) > 0 (or even more if z(l) too is driven to 0).
leI possesses a In constructing this new solution we have created no new bad order times. Indeed,
IDlple solution any remaining ones must all be greater than ll. We can thus repeat this procedure until
all the bad order times are eliminated, reducing the total cost at each step. Certainly, an
:suit describes
optimal solution can have no bad order times.•
:ero-invent01Y
:d so as to run Incidentally, there is another way to demonstrate the zero-inventory property: The
lSe Xo 2:': 0 too, objective is a concave function of the variables over the convex set defined by the con
tive inventory, straints and the dynamics. Its minimum can be found, therefore, at an extreme point of
the feasible region. (See SectionA.3.) It is possible to show that every extreme point has
the zero-inventory property, so the optimal solution does too. We shall not pursue this
ions (4.3.1) to approach here, but it is worth being aware of it. Several extensions of the DEL model
I!y the earliest can be analyzed in this way.
FIGURE 4.3.1
Network/or DEL model.
FIGURE 4.3.2
One possible path
o
0)
o 0
along the path (excluding Titself) identify the order times. Figure 4.3.1 illustrates this
network for a problem with T = 5. Figure 4.3.2 shows one possible path; the order
times are 0, 2, and 3.
Now, consider anyone path and a particular arc (I, u) along the path, Choosing this
arc means that t is an order time, and we order just enough at t to meet the demands at
points t through u - 1. Let k[I, u) denote the total cost incurred by this decision. The dy
namics imply that z(t) ~ D[I, u), x(t + 1) = z(t) - d(t) ~ D[t + I, u), and xes + I) =
xes) - des) ~ D[s + I, u) for t < s < u Thus,
k[t, u) = k(t) + C(t)z(l) + L;~'+l h(s)x(s)
= k(t) + c(t)D[t, u) + L~~HI h(s)D[s, u)
= k(t) + L;=,[ crt, s)d(s) (4.3.8)
Now, add the k[t,u) over all arCs encountered on the path. This sum accounts for the all
costs in all periods-it is the cost of choosing the order times along the path.
Chapter 4 Time-~'iJ,ying Demands 85
::0
d(t) k(l) crt) h(t)
'7
0 10
2
40
40
2
2
2 12 40 2
3 4 40 2
4 14 40 2
5
;I"
o 2
u
3 4 5
3 48 90
4 68
So, compute k[l, u) for every arc (t, u) in the network, and think of k[l, u) as the travel
illustrates this cost along arc (I, u). To determine an optimal sequence of order times, find a palh of
I"'th; the order minimum total cost. In sum, the DEL model reduces to a shortest-path calculation. There
are very efficient solution procedures for such problems. (See Gallo and Pallottino
. Choosing this [1986], for example.) This network is acyclic (there is no path from a node to itself), and
the demands at the nodes are naturally ordered by t, which makes the calculation even easier.
cision. The dy Let us look at one method in detail. It is called forward recursion. The idea is to
mdx(s + I) ~ compute the minimum-cost path from node 0 to every node I, starting with I = I, then
I ~ 2, and continuing until I = T. This method solves the DEL model with horizon I for
each I. Let V"(I) denote the optimal cost of the I-horizon problem, so V'(T) is the opti
mal cost of the original DEL model.
Step I aims to detertnine the lasl order time in the I-horizon problem, denoted S'(I).
(4.3.8) But, assuming s~(t) = s for some fixed s < t, we should set all earlier order times
according to the optimal solution of the s-horizon problem, so the total cost is V"(s) +
IUllts for the all k[s, I) ~ V'(s, I). To find the best such s, the actual s'(f), compute the smallest of the
I"'th. V'(s, I). The following algorithm embodies this idea:
.. --,.,.
~
Algorithm Forward_DEL:
Set V"(O) ~ 0
Forl~I, ... ,T:
V'(s, I) ~ V'(s) + k[s, I) 0 ,,; s < I
V'(I) ~ min. {V'(s, I) : 0 ,,; s < I}
s'(I) = largest s achieving this minimum
Once this calculation is done, recover the optimal order times by tracing the s'(I) back
from T That is, s'(T) is the last order time, s'(s"(T) is the next to last, and so on.
EXAMPLE 4.3.A, PART 2
0 0 0 12
60 0 2 0
2 66 0 0 30
3 114 0 18 0
4 134 0 14 0
5 198 2 0
The solution tells us to plan two orders: Order 12 units at the beginning to eover the demands at
points 0 and 1, and then order 30 at point 2 to cover the remaining demands. The total optima) cost
is 1"(5) ~ 198.
Incideutally. there is another approach to the problem called bacbvard recursion:
Take Figure 4.3.1, tum it upside down, reverse all the arrows, and apply Algorithm
Forward_DEL to this network. In terms of the original problem, tills method works in
the opposite direction, finding the optimal path jrom each I to T The cost of that path,
say V'( I), measures the optimal cost over times I through T starting with X(I) = O. The
optimal cost of the full DEL model is V'(O) = V'(I).
This method is essentially equivalent to Algorithm Forward_DEL, but it is worth
knowing about anyway: People do use it. It recovers not just a solution but a full optimal
policy. Letting q(l) be the first order in the I-to- T subproblem, q(l) prescribes the amount
to order at time I, ifx(l) ~ 0, as in § 4.2.3. Finally, this approach is similar conceptually
to that of Chapter 9 for stochastic models; both are sometimes called dynamic program
ming approaches.
How much effort does Algorithm Forward_DEL require? There are Titerations, and
the tth entails t similar calculations. The total number of calculations is thus proportional
to I, t ~ 'hT(T + I). The quadratic term is the dominant one for large T So, we say the
algorithm requires 0(T 2 ) time. (This uotation has a precise meaning, but this infOlmal
Chapin 4 Time-Varying Demands 87
definition will do here. Setting up the k[s, I) also requires 0(T 2 ) time.) This summarizes
the capability of the algorithm to handle large problems: As T increases, the time re
quired grows roughly as T 2
Quadratic growth is not bad as these things go; other algorithms for other problems
are far more sensitive to their primary size parameters. (See Section 4.4.5, for instance.)
Even so, researchers have recently developed refmed procedures requiring only O[(T) log
(T)] time in general, and only O(T) for problems with no speculation opportunities. These
methods are thus faster than Algorithm Forward_DEL for large T (Federgruen and Tzur
[1991] report that theirs runs about 3 times faster for T = 500 and 70 times faster for
g the s'(t) back T = 5000. Such huge T's can arise, for instance, in a discretization of an originally con
IIld so on. tinuous model.) They require some additional computational overhead, however, and for
small Tthey are a bit slower.
The basic idea is simple: To determine s'(I), Algorithm Forward_DEL searches over
all s < t. Even for large t, it continues to consider even the smallest s's. The new algo
rithms extract information from prior iterations to help guide the search, specifically, to
eliminate some of the s's. [The same idea underlies the streamlined recursion (4.3.7) for
the linear-cost case. See also Problem 4.5a.]
The asymptotic growth rates expressed in the 0(-) notation do not tell the whole
story. If an algorithm took 3 years to solve a problem with T ~ I, it would be little com
fort to know that a problem with T ~ 100 requires "only" 3(1002 ) or even 3(100) years!
The DEL model, fortunately, has no such problem. The individual calculations in
Algorithm Forward_DEL are ve,y simple. The newer methods are nearly as simple. Pro
vided some care is given to efficient programming, most practical-sized problems can
be solved quickly. For example, with T = 26 (half a year ofweekly time periods), the en
tire calculation requires a fraction of a second on any decent personal computer.
One additional fact deserves mention: In some instances of the DEL model, it is
rthe demands at
possible to identify (see PlOblem 4.5b) a planning horizon. This is a time point, say IH ,
otal optimal cost
which essentially separates earlier times from later ones. Specifically, for s < tH -- 1, the
z*(s) in the t}{'horizon plOblem .emain optimal in the I-horizon p.oblem for all 1 > I H.
mr1 recursion: (Terminology alert: Sometimes people .efe. to T as Ihe planning horizon. Usually, it is
'Illy Algorithm clear in the context what "planning horizon" means.)
,mod works in A planning horizon is very convenient: Often in practice, we actually use the solu
st of that path, tion values only for early time points. In that case, we can terminate the algorithm at
u(l) ~ O. The step IH . Also, the z*(s) are entirely independent of the data after I H, so we need not
worry about estimating those data. Thus, a planning horizon justifies a myopic ap
but it is worth proach. Unfortunately, it is hard to predict if and when a planning horizon will occur.
• a full optimal
leS the amount
4.3.5 Examples--{)pacity
II" conceptually
71TIic program- Let us solve a few examples with many periods. Think of these as discretized versions
of originally continuous-time models. They all have T = 200, but we display the solu
iterations, and tion only up to t = 150. All have constant cost factors, indicated in the figures below.
IS proportional First, consider a model with slowly varying demand, shown in Figure 4.3.3. The de
So, we say the mand d(l) is the heavy curve near the bottom, which undulates gently between 2 and 6.
• this informal The lighter curve above shows q(I), the solution to an EOQ model for each 1 assuming
88 Foundations ofInventory Management
FIGURE 4.3.3
Optimal order sizes (k = 50, C = 2, h = 1).
30
25 q(t)
_v 0
~" /4
~~~,
20
///
15
10
5 -v ~ ~e-
t- ~
I--- V
I-- d(t)
o
o 30 60 90 120 150
Time
the demand rate remains constant at the current value d( I). Finally, the vertical lines in
dicate the optimal orders in the DEL model.
It is striking how closely the positive z*(I) track the q(I). The spacing of the lines,
moreover, indicates that orders are more- frequent as well as larger when demand is large.
(The EOQ model exhibits similar behavior.) The correspondence is not perfect, mainly
because of the DEL model's discrete time. This example illustrates one of the special
cases of Section 4.2: When demand changes slowly, the EOQ model using the current
demand rate works well.
Next, Figure 4.3.4 presents a model with fast-changing d(I). The positive z*(I) are
all equal to 32. (The exact equality here is somewhat coincidental. For other values of k,
for example, the z*(I) do change, though not much.) This common value is near the EOQ
model's q* of28.3, based ou the average demand rate of 4. This example illustrates an
other extreme case of Section 4.2: Rapid fluctuations can essentially be ignored.
Figure 4.3.5 shows a model with a highly irregular demand pattern. In this case the
ranges of the orders generated by the DEL and EOQ models are about the same. It is
hard to discern any precise correspondence between them. however. The fine structure
of the optimal solution exploits the demand fluctuations in subtle ways.
Most instances of the DEL model have similarly subtle solutions. In this sense, com
pared to the EOQ model with its siruple formulas, the DEL model is opaque. Except in the
extreme cases of Section 4.2, it is hard to see why the algorithm makes the precise choices
it does. In fact, a tiny change in the model parameters can cause the solution to shift
markedly. We need a clever algorithm to find the optimal solution; no simple formula will
,-
FIGURE 4.3.4
Optimal order sizes (k = 100, C = 2, h = 1).
40 , , - - - - - - - - - - - -
35
q(t)
30
q(t)
25
20
15
10
d(t)
5 d(t)
o
150
o 30 60 90 120 150
Time
ertical1incs in
FIGURE 4.3.5
ng of the lines, Optimal order sizes-k = 50, C = 2, h = 1.
emand is large. 30
(lOdect, mainly
, of the special
25
ing the current
5
l/lV h f/ [\~ I'IJ
r\ "IV ~ 1\ ~I\ \! ~I\! rv
!:tis sense, com
~v
V V d (I)
e. Except in the
precise choices o
,hnion to shift o 30 60 90 120 150
>Ie formula will Time
..
do. On the other hand, the range of order sizes is usually within the range of values indi
cated by the myopic use of the EOQ model, as suggested in Section 4.2.4.
This opacity rules out sensitivity analysis, or at least simple, transparent results like
those in Chapter 3. There do exist some limited sensitivhy results; see Problems 4.6 and
4.7 and Zangwill [1987], for instance.
4.3.6 Heuristics
Even though it is not hard to solve the DEL model exactly, several heuristic methods
have been developed, for reasons explained below. We describe one popular approach,
the Silver-Meal heuristic (Silver and Meal [1973]). Most of the others are similar in
spirit. The idea is to look ahead only a few periods, not all the way to the horizon, to set
each order quantity. (This is a common analytical tactic, sometimes called the greedy or
myopic approach; we saw the idea in Section 4.2.) The heuristic itself determines how
far ahead to look as it works.
Assume Xo = O. Focus on the fIrst order, restricting attention to values consistent
with the zero-inventory property. So, z(O) may cover just the initial demand d(O), or
D[O, 2) ~ d(O) + d(l), or D[O, 3) ~ d(O) + del) + d(2), et cetera. [fit covers D[O, u),
then the total cost over this interval is k(0, u), so the average cost is k[O, u)/u. The heuris
tic aims to make this average cost small: Starting with u = 1, it increments u by 1 as long
as the average cost decreases; once it finds that a further increment would increase the
average cost, it stops. This rule detennines z(O).
If the heuristic stops at ll, setting z(O) = D[O, u), thenx(u) = 0.1t then restarts, treat
ing u as the initial period, and computes z(u) as above. It continues in this manner until
it reaches the horizon.
Reconsider the small example of § 4.3.4. To select z(O), the heuristic computes
k[O, 1) ~ 60 ~ 60
1 1
klO, 2) ~ 66 ~ 33 < 60
2 2
k[2, 4) ~ 76 ~ 38 < 64
2 2
kl2,5) 132
-- ~-~44>38
3 3
.,.,:~."""""""
,,' .. : .
,. I"
u. The heuris ample, in the model of Figure 4.3.4 above with k = 90 instead of 100, the true optimal
uby 1 as long solution remains the same, i.e., z*(t) = 0 or 32, but the heuristic sets the positive z(t) to
d increase the 16, covering only one demand cycle instead of two. The optimal average cost is 32.25,
while the heuristic's is 35.5. about 10% more. Indeed, Axsiiter [1982] demonstrates that
restarts, treat the heuristic can perform awfully: its cost can be arbitrarily large relative to the optimal
;manner until cost. (Bitran et al. [1984] show that another heuristic is better in this sense.)
This negative result is based on a worst-case analysis, which, by design, focuses on
extreme cases. What about more "typical" cases? A good deal of research has been de
voted to empirical performance testing, including Wemmerlov [1982], Blackburn and
Millen [1985], and Baker [1990,1993]. They find that the Silver-Meal heuristic and re
lated methods perfoffil reasonably well over a wide range of problems, on average.
Why bother with a heuristic when the exact algorithms are so fast? Well, the heuristic
is even faster. It requires O(T) time, which certainly beats Algorithm Forward_DEL for
large T The newer algorithms also require O(T) time (for stationary costs). But, simple as
those algorithms are, the heuristic's internal mechanics are even simpler, so it is still faster.
This speed advantage was more important in 1973 when the heuristic was intro
duced than it is now, for computers then were far slower and far less accessible, and the
newer algorithms were unknown. Today, in my view, the exact procedures are fast
enough for most practical purposes.
Apart from the speed issue, there are other reasons that follow to consider the
heuristic.
demands are more-or-Iess imperfect forecasts, and the costs are rough estimates. Typically,
the model is solved repeatedly over time with frequently updated data. (That is, the hori
zon length T is fixed, but the time interval [0, T] shifts ahead as time passes, so that "now"
always means rime point 0.) This is called a rolling-horizon scenario.
Such discrepancies between a model and reality are hardly unique here. Every ap
plied model functions within a far more complex world. This fact is cause for neither
despair nor complacency. A good model is valuable partly because it focuses on sim
ple essentials, ignoring real-world complexities. On the other hand, a good modeler
has to know both the world and the model, and to navigate the gaps betweeu them.
In the context of the DEL model, these general COncerns include a variety of spe
cific modeling and implementation issues. Let us briefly discuss a couple of them:
One issue is the choice of solution method As mentioned above, the optimal solu
tion is usually a few percent better than the Silver-Meal heuristic's, and occasionally
much better. The question is, how meaningful is the difference in the real world? The
heuristic is myopic, ignoring the distant future where the parameter forecasts are least
certain. Does it therefore extract most of the useful infonnation in the data? The optimal
solution cleverly exploits every opportunity embodied in the model parameters to mini
mize cost. But, when the parameters are only estimates, are these opportunities real or
just fleeting mirages?
It would be nice to have a clear-cut answer, but unfortunately there is none, not yet
at"/east This: issue is a contentious one~ and tlIere is eVIdence to support botb SIdes, SOflle
of the empirical studies cited above simulate rolling-horizon scenarios, and in some of
those the heuristics seem to work as well as optimization, or even better. On the other
hand, it is easy to construct a plausible scenario where optimization clearly wins.
The "right" answer, of course, depends a great deal on just how precise the param
eter estimates are. People disagree about this too. In my experience, real-world settings
differ widely in this respect; in some cases the estimates are sharp, while in others they
are crude. So, I sit squarely on the fence; I believe that optimization and heuristics both
have their places.
If the estimates are too noisy, the DEL model itself ceases to be useful, and no com
putational techniqne can save it. The primary issue becomes modeling, not method. One
option is to expand the model to include imperfect forecasts explicitly. Such stochastic
models are discussed in Chapter 9. Some are much more complex than the DEL model,
but others are only little more. The philosophy there is to seek a robust policy, one that
works well for a range of possible future outcomes, instead of the best schedule for one
particular forecast.
Another option is to adjust the DEL model by adding safety stock (sometimes called
buffer stock). This means changing the lower bound on x(t) in the constraints from 0 to
some positive value, say x_(t), for some or all t. This variant of the DEL model, it turns
out, is no harder to solve than the original. (In fact, the new model is equivalent to an in
stance of the DEL model itself, as Problem 4.8 asks you to show.) It is no easy matter,
however, to determine "good" values for the L(t). Most practical methods borrow
guidelines from the analysis of stochastic models, like those of Chapter 9. Thus, the
safety-stock approach tries to mimic a stochastic analysis within a detenninistic model.
Another, related issue is nervousness. This refers to the fact, mentioned above, that
the optimal solution of the DEL model can shift abruptly in response to small parame
Chapter .; Time- Varying Demands 93
ates. Typically, ter changes. In a rolling-horizon scenario, one solution can schedule a large order at
lat is, the hori some future time, and then a subsequent solution, based on updated parameters, can
oso that ''now'' move that entire order to a different time, or combine it with another, or split it into two.
Well, so what? If the data change, the schedule should change, shouldn't it? The dif
ere. Every ap ficulty lies in how the model is used: The DEL model is often used to guide not just the
15e for neither orders inside the model, but also other decisions outside it, such as workforce levels and
~uses on sim maintenance schedules. All these decisions, the orders themselves (as discussed in Sec
good modeler tion 4.4.1) but especially the others, often require substantialleadtimes. [n this context,
ween them. nervousness can be highly disruptive.
"mety of spe This is another contentious issue. There is no consensus on how to think about nerv
, of them: ousness, let alone what to do about it. Some people favor relieving its symptoms with ad
, optimal solu hoc tactics, while others insist on the radical cure of expanding the model.
d occasionally One ad hoc approach focuses on solution technique: The Silver-Meal heuristic tends
"" world? The to be less nervous than the exact algorithms, simply because of the way it processes the
:casts are least data. The danger, as above, is that the heuristic may miss real cost-saving opportunities.
," The optimal An occasional disruption may be tolerable if the savings are sufficient.
Deters to mini Another approach directly suppresses nervousness by partially jreedng the solution.
tunities real or That is, it selects a scheduling horizon, some fairly small T_ < T. Whenever an instance
of the model is solved, the orders z*(t) for t < T_ become final; these =*(t) become fixed
s none, not yet constants in subsequent runs. (Alternatively, fix the order times for t < L, but allow the
dI sides. Some quantities to shift with revisions in de nd forecasts.) This approach too has dangers: If
md in some of T_ is too large, the solution ceases to be responsive to important changes in the envi
r. On the other ronment, and it is hard to know in ad nee what is too large in a particular instance.
iy wins. Yet another tactic adds tenns to th objective to penalize deviations from a previous
i'ie the param solution. [fthis is done in a certain way, the resulting model reduces to an instance ofthe
-world settings DEL model (Problem 4.18). It is not e sy, however, to select appropriate penalties. If
: in others they they are too low, the effect is negligible, but if they are too high, they dominate the rest
heuristics both of the objective and enforce the previous solution. (By the way, adding safety stocks does
nothing to reduce nervousness.)
~, and no com Nervousness can be anticipated and reduced somewhat by using planning-horizon
~ method. One techniques (mentioned in Section 4.3.4). See Federgruen and Tzur [1994], for example.
,uch stochastic These tactics all build on the established technology of the DEL model. They are
oe DEL model, therefore relatively easy to implement; they use the same data structures and computational
",licy, one that methods. Those who favor the radical cure argue that nervousness itself is a symptom of a
hedule for one model under stress. The real ailment is parameter uncertainty; to treat it directly, expand
the model. Don't try to trick the DEL model into doing something it doesn't want to do.
netimes called One final note: Nervousness tends to be most severe when the fixed costs k(t) are
Mts from 0 to large. (It is much more benign in the linear-cost case.) So, reducing the k(t) relieves ner
model, it tums vousness, in addition to its direct cost-saving benefits.
",-alent to an in
lO easy matter, 4.4 Extensions
ethods borrow
' f 9. Thus, the Many extensions of the basic DEL model have been developed. Some are straightfor
ninistic model. ward, while others require extensive analysis. We cover only a few of them here and those
ned above, that only briefly. We discuss them separately, but it should be clear that they can be combined.
small parame- Extensions to multiple products and locations are considered in Chapter 5.
94 Foundatio!l:> q(IllventOl)' Management
4.4.1 Leadtimes
The DEL model has no explicit order leadtime, but it is easy to include one. Let L de
note the leadtime, a fixed, positive integer. An order placed at time I arrives at the end of
period t + L - 1, in time to be counted in the inventory at time t + 1. (By this conven
tion, the original DEL model has leadtime I, not 0.)
Just change the meanings ofthe variables and cost factors: Now, =(I) means the quan
tity ordered at time I - L + I, and k{1) and C(I) describe the costs incurred by this order.
Tbisz(l) still arrives just before time I + I, so the dynamics (4.3.2) are still valid. Tbe ini
tial conditions require a slight adjustment. The z(l) for I < L - I and the x(l) for I < L have
been determined already before time I = 0; they become fixed constants. Thus, as in the
EOQ model, we can account for the leadtime by keeping careful track of the calendar.
There is another, equivalent way to revise the formulation, focusing on physical
stock instead of time: Return =(t) to its original meaning, the order placed at time I, but
redefine x(l) as the inventon' position at step (I) of time t, and Xo as the initial inventory
position. Also, use the symbol.i(l) for the actual inventory at time I. Clearly, these vari
ables are related by the identity
x(1 + L) = x(t) - D[t, I + L).
This is the discrete-time analogue of (3 .2.1). With these definitions the initial conditions
(4.3.1) and the dynamics (4.3.2) remain valid as stated, but x(t) replaces x(t) in the con
straints (4.3.3) and the objective (4.3.4).
Now, use the equations above to eliminate~e xC!). The end result is another instance
of the DEL model. [The lower bound onx(t) i (4.3.3) is now D[t, I + L) > 0 instead of
zero, but that difference is inessential, as Pro lern 4.8 demonstrates. This lower bound
serves as a reorder point. Also, the objective h s an additional constant term.]
The time-based approach above is simpler, ut this stock-based approach is worth
knowing too. We shall use similar transformations in the stochastic-demand models of
Chapter 9.
The objective ofthe DEL model is the total cost from now until the horizon. This is equiv
alent, of course, to the average cost over that time. It is easy to adjust the DEL model to
represent the discrete-time analogue of the discounted-cost criterion of Section 3.7.
How do the present-value calculations work in a discrete-time context? For the mo
ment, suppose that time is really continuous, and the length of each of our periods is ex
actly 1, so the time index t correctly measures time on the continuous scale. Letting ex.
denote the interest rate. the present value of a unit cash flow at time t is e -ut, as before.
Equivalently, defining 'I ~ e -u, the present value is 'I'. If the period length is different
from 1, say w, redefine 'Y = e -UW, so that the present value of a unit cash flow at time
point t is again 'Y t . The fraction 'Y is called the discount/actor.
Now, suppose that time is really discrete. The story in Section 3.7 can be retold en
tirely in discrete-time terms. Here, a is the one-period interest rate (regardless ofthe pe
riod length), and-y = lit I + a). The conclusion is the same: The present value of a unit
cash flow at time point t is 'Y I .
""i
Turning to the DEL model, suppose e(t), k(t), and h(l) are the actual cost factors. To
reformulate the objective (4.3.4), just perform the replacements e(t) <-- '('e(t), k(1) <
one. Let L de
'('k(I), and h(l) <- '('h(I). Then. (4.3.4) measures the total discounted cost, as desired. In
:s althe end of
sum, the discounted-cost DEL model reduces to an instance of the original DEL model
y this conven
with altered cost factors.
As in Chapter 3, the h(l) here should include only physical handling costs, not fi
eans the quan
nancing charges. Indeed, the discounted-cost model is conceptually simpler when the
I by this order.
(original, undiscounted) c(t) change over time: In this case it is not at all clear how to
valid. The ini
specify an appropriate financing cost in the total- or average-cost model; we choose the
I fort < L have
order costs, hence the financing charges, through our decisions. There are ways to re
Thus, as in the
solve this dilemma, but they rely on intricate assumptions and arguments. The
Ie calendar,
discounted-cost objective bypasses the difficulty altogether.
19 on physical
j at time I, but
itial inventory 4.4.3 Continuously Accumulating Costs
dy, these vari
Returning to the total-cost criterion, suppose we use the DEL model to approximate a
continuous-time problem. In the real problem, demand occurs and holding costs accu
mulate continuously over time. For simplicity, scale time so that the time periods are
each ofiength 1.
rial conditions
Given the data of the DEL model, we typically do not know the actual demand and
il) in the con-
holding-cost rates at each instant, nor the precise moment an order is received. Provided
other instance the time periods are short, it is reasonable to asswne that these rates are the constants
d(t) and h(1 + I) over the entire period from I to I + I, and the order z(l) is received at
> 0 instead of
the beginning of the period. Then, for s between I and 1+ 1,
5 lower bound
rIll.] x(s) ~ x(t) + z(I){- (s - l)d(I),
:uach is worth
and the total holding cost over the period is
md models of
h(1 + I)J; + I x(s) ds ~ h(1 + I )[X(\+ z(l) - f;+ I(S - l)d(l) ds]
~ h(1 + I )[x(l) + z(l) - ~d(I)]
.This is equiv Thus, the objective function (4.3.4) correctly accounts for the controllable holding costs
DEL model to in all periods. The actual total cost includes the additional constant y,~, h(1 + 1) d(I).
tion 3.7. If we do know the true demand rate A(I) and retain the other assumptions above, the
t? For the mo end result is the same; only the constant term changes. Alternative assumptions lead to
'periods is ex similar results. In sum, the DEL model requires only slight adjustments to account for
ale. Letting" continuously accwnulating costs.
-m, as before.
:th is different
Ii flow at time
4.4.4 Backorders
Now, suppose that backorders are allowed, as in Section 3.3 of Chapter 3. Redefine the
1 be retold en state variable x(t) to mean the net inventory (inventory minus backorders). The initial con
le:ss of the pe ditions (4.3.1) and the dynamics (4.3.2) remain valid as stated. Omit the constraints
~alue of a unit xlt) ". 0 in (4.3.3), but retainz(t) ". O. In the objective (4.3.4), replace the term h(l)x(l) by
C(I, x(I)), where C(I. x) ~ h(I)[xt + b(I)[xr and b(t) is the backorder-cost rate.
96 Foundations o/Inventory Management
Incidentally, a standard formulation trick recovers a linear objective (apart from the
o(z) terms): Define two new variables for each t, the inventory x +(t) and the backorders
x - (t). Replace x(t) in the dynamics by the difference x +(t) - X -(t) and C(t, xU)) in the
objective by h(t)x+(t) + b(t)x-(t), and add the constraints x +(t) 2: 0 and x -(t) 2: O. (For
this to work x + (t) and x - (t) cannot both be positive. The constraints do not enforce this
condition, but the objective ensures that it holds anyway.)
In the original DEL model, the zero-mventory property implies that each order
covers the entire demand over several consecutive time points. It turns out (see Problem
4.10) that this is true also in the model with backorders. In the original model, however,
the interval covered by an order always begins with the order time itself, whereas here it
may begm earlier. That is, each order (except perhaps the first) covers the demand at its
own time, plus some earlier and later times.
Thus, the overall behavior of the net inventory is much the same as in the continUQllS
time, constant-demand model of Section 3.3. It decreases at most times, increasing only at
order times. Typically, an order raises x( t) from a negative to a positive value. So, a solu
tion divides time into cycles, each consisting of a part with inventory followed by a part
with backorders. (Because time is discrete, however, the inventory or backorder part can
disappear. )
Using this property, one can construct a network whose paths correspond to poten
tial solutions of the problem, as in Section 4.3.4. There is one node for t ~ T, labeled T-,
but two nodes for each t < T, indicated by (- and t+ . The arcs eonnect pluses to minuses
and minuses to pluses. Specifically, there is an arc (t-, u+) for all t ::; u and an arc
(t+, u-) for all t < u. See Fig. 4.4.1. A path from 0- to T - corresponds to a solution as
follows: Tbe nodes t+ encountered along the path describe the order times, while the
nodes t- specify the amounts ordered: Ifthe path contajns the arcs (s-, t+) and (t+, u-)
for s :.s t < u, the order at time t covers the demands ~t time points s through u - 1.
It is possible to assign a cost to each are, as in ~.3 .8), so that the total cost of a so
lution equals the sum ofthe arc costs on the corresp ding path. (Problem 4.11 asks you
to supply the details.) Thus, as in the DEL model, to compute the optimal solution, just
detennine the minimum-cost path in the network.
FIGURE 4.4.1
Network/or DEL model with backorders.
..•••*d"*,
tapart from the Many other extensions of the DEL model can be analyzed by the same logic: First, de
the backorders nve some qualitative characteristic of an optimal solution, analogous to the zero
C(t. x(t)) in the inventory property. Second, use this property to construct a network whose paths correspond
r - (t) 2> O. (For to possible solutions. Third and finally, calculate the minimum-cost path in the network.
lot enforce this
4.4.5 Limited Capacity
hat each order
~ (see Problem Return to the DEL model. but suppose there is an upper limit z+(t) on the order quantity
IOdel, however, at time point t. Thus, the zIt) must respect the upper bounds
i.\-bereas here it z(t) <; z+(t) t ~ 0, ... , T - I (4.4.1)
~ demand at its
in addition to the original constraints (4.3.3). When the zit) represent production quan
be continuous tities, we can interpret the z+(t) as production capacities. (This interpretation is some
:reasing only at what problematic, however, as explained below.) For simplicity, assume that Xo = 0, and
lue. So, a solu to ensure feasibility, assume that l~=o z+(s) ~ L.~=o des) for all t.
owed by a part Incidentally, we can reformulate the model as a mixed-integer program using the bi
'-Order part can nary variables vet), replacing the inequalities in (4.3.5) and (4.4.1) by the single constraint
z(t) <; min {D[t, T), z+(t))v(t) t = 0, ... , T - I
pond to poten
I labeled T- , The qualitative effect of limited capacities is the same as in the continuous-time
tSes to minuses model of Section 4.2.5: The capacity constraints (4.4.1) force us to anticipate large de
; u and an arc mands, i.e., to order (or produce) earlier than we might otherwise. So, capacity con
LO a solution as straints lead to larger inventories. Part of this effect can be quantified as follows: Re
nes, while the cursively define
) and (1+, u-)
x>(T) = 0 (4.4.2)
ighu-l.
fl1 cost of a 80 x>(t) = [d(t) - z+(t) + x>(t + l)t 0 <; t < T
14.11 asks you
This quantity is the excess demand from time t onwar( It is not hard to show that every
I solution, just
feasible solution must have x(t) ~ x>(t). A group o~succeSSive periods with positive
x>(t) is the discrete-time analogue ofan interval [L, t ], in the notation ofSection4.2.5.
Interestingly, it is possible to use these quantitie to derive an equivalent model in
which demand never exceeds capacity. The demands r this new model are d(t) ~ d(t)
+ x>(t + 1) - x>(t), and the initial inventory is Xo = Xo - x>(o); the cost factors are
just the original ones. (See Problem 4.13.) Even in this revised formulation, however, we
cannot ignore the capacity constraints.
~ The analogue of the zero-inventory property here is fairly subtle: Every feasible so
lution divides the periods into what we call inventory cycles. An inventory cycle consists
of successive periods s, with t ::; s < u say, starting with x(t) = 0 but with positive in
ventory at all other times s. (Since u begins the next cycle, we also have x(u) ~ 0.) There
is some optimal solution whose inventory cycles each have the following property: For
every time S except at most one, we produce nothing or at full capacity, i.e., z(s) equals
o or z+(s).
Unfortunately, this property, while interesting, does not always lead to a tractable al
gorithm. The general version of the problem is very hard. (In technical terms it is NP-hard;
the time required by every known algorithm grows exponentially in the problem size.)
98 Foundations afInventory Management
In certain special hut important cases. however, it is possible to solve the problem with
relative ease. In the linear-cost case, the problem is a linear program. (This linear program
has a useful special structure-it can be recast as a minimum-cost network-flow problem,
and this in turn can be expressed as a transportation problem. See Problem 4.14. There are
very fast specialized algorithms to solve such problems; see Chvatal [1983], for example.)
Also, consider the case with positive fixed costs hut equal capacities z+(t) = z+.
Here, the inventory-cycle property can be exploited to devise a reasonably tractable al
gorithm. Here is the basic idea: Focus on any potential inventory cycle, specified by the
pair (I, u). The inventory-cycle property implies that the cumulative production at each
time s can take on only a small number of values. Using this fact, one can compute the
optimal solution within the cycle, assuming X(I) = x(u) = O. (See Problem 4.15.) Call
the cost of this solution k[l, u). Now, use these costs within Algorithm Forward_DEL.
This procedure dctennines the optimal set of inventory cycles, hence the optimal over
all solution. The calculation of the k[l, u) is more involved than in the DEL model, and
the overall algorithm runs in O(T 4 ) time.
The capacitated DEL model is an important one, but it is not a perfect model of a
production process: If there are orders in several consecutive periods, every one incurs
a fixed cost. A typical production system, however, incurs a setup cost only at the be
ginning of a production run. If the time periods are short, and we expect a production
run to cover many of them, then the capacitated DEL model presents a distorted picture
of production-setup costs.
Here is an alternative fonnulation: Say that t hegins aproduction run if z(t) > 0 and
either I ~ 0 or z(t - I) < z+(t - I). The setup cost k(1) is incurred at such time points
and only those. If z(t) > 0 but z(1 - I) = Z+(I - I). an earlier production run continues
through I, so z(l) incurs no setup cost.
In certain cases, this model is easier to solve than the capacitated DEL model. Sup
pose there are no speculation opportunities (S(I) = I), and de,) -<; z+(t) (following the
transfoffilation above, if necessary). One can show that every production run begins with
zero inventory, except perhaps the first. Let k[l, lI) be the cost of beginning a production
run at I and producing a total ofD[I, lI) to meet demand through fell - 1. (This is easy
to compute.) Using these costs within Algorithm Forward_DE solves the problem.
An even simpler model is the all-or-nolhing formulatio : Each z(l) can take only
two possible values, 0 or Z+(I). (Thus, z(l) ~ z+(t)v(I), so the ariables z(t) can be elim
inated.) This is the discrete-time analogue of the finite-capa ity EOQ model of Sec
tion 3.4 and Section 4.2.5. Here, every solution corresponds to a path in a network sim
ilar to Figure 4.4.1. (Node t+ mdicates a possible beginning of a production run, and
node 1- an end.) Thus, a shortest-path calculation determines the optimal solution.
Ie problem with Each c(t, z) is a concave function of =, because the purchase price is smaller for
; linear program larger orders. The objective function (4.3.4) thus remains concave. As noted after the
<-flow problem, proof of Theorem 4.3.1, therefore, the zero-inventory propelty remains valid.
4.14. There are Consequently, Algorithm Forward_DEL again solves the model. The k[f. u) are cal
J, for example.) culated differently, to account properly for the order-cost discounts, but that is the only
ies =+(1) = z+. modification required. (Under an all-units discount, however, c(t, z) is not concave, nor
IJly tractable al even continuous, so this approach does not work.)
;pecified by the
duction at each
an compute the 4.5 Smoothing Models
~em 4.15.) Can
Forward_DEL.
4.5.1 Formulation and Discussion
le optimal over The next model is similar in some ways to the DEL model with backorders, but in oth
IEL model, and ers it is radically different. The demands d(f) have the same meaning as before, as do the
variables =(1) andx(f). (,(I) denotes net inventory, not just inventory.) Again, these are re
feet model of a lated by the initial condition
'-ery one incurs
X(O) ~ Xo (4.5.1)
. only at the be
:ct a production and the dynamics
listorted picture
X(f + I) ~ x(f) + z(l) - d(l) t ~ 0, ... , T - 1. (4.5.2)
°
n ifz(f) > and
;nch time points
However, there are no other constraints on the variables; z(t) as well as x(t) can be neg
ative. The objective function is quite different:
III run continues
Minimize YiL;:~ c(l)='(f) + YiL;~l h(l}<'(f) (4.5.3)
EL model. Sup Here, the c(f) and h(t) are all positive. Call this the basic smoothing model.
) (following the What an odd model! Let us try to understand the story it represents: First, a nega
,nul begins with tive "order" z(f) represents disposal of excess stock. For both actual positive orders and
ing a production disposal, there are diseconomies of scale, expressed by the quadratic term C(f)Z'( f). That
. 1. (This is easy is, the unit ordering cost increases in the amount ordered (and similarly for the disposal
the problem. cost). Moreover, these costs are symmetric; the cost of -z(l) is the same as z(f). Like
f) can take only wise, there are diseconomies of scale for inventory and backo ders, and these costs too
:(f) can be elim are symmetric; the coefficient h(t) measures backorder costs s well as holding costs.
! model of Sec Frankly, this model is not very realistic. It is indeed po sible to dispose of stock in
1 a network sim some situations, and diseconomies of scale do arise in pract ceo The problem lies in the
ruction nul, and symmetry assumptions: It is just not credible that the dispo 'al cost precisely equal the
!l3.1 solution. purchase cost, nor that backorders and inventory have identical cost effects.
Why should we study an unrealistic model? The answer, in a word, is simplicity: The
solution is easy to compute, and it has an appealingly simple form. The model can be ex
tended in many directions while retaining comparable simplicity. Also. although the
Section 3.5.2 of model is unrealistic, the solution accounts for demand variations in a plausible way. The
mits at time t. In details reflect this particular formulation, but the general idea suggests practically use
:ewise linear. (It ful heuristics. Finally, this model helps us to understand others. More realistic models
il term C(f)z(f) in are asymmetric in crucial places; those asymmetries help to explain precisely how and
why such models are more complex.
100 Foundations ofInventory Management
13(1, u) ~ [
e(1 +
LI(I)
I)] 13(t + I, u) t < u< T
Moreover,O < 13(1, u) < I, and 13(t, u) is decreasing in u for each I.•
The actual optimal solution can be recovered recursively from (4.5.4): First, substi
tute Xo for x(O), and setz*(O) ~ q(O). This determines x*(I) through (4.5.2). Next, using
x(l) = x*(1) in (4.5.4), setz*(1) ~ q(1). Continue in this manner.
The coefficients 13(t, u) depend on the cost factors, but nol on the demands. Thus,
q(l) is a linear ftmction of x(l) and the d(II), u 2: t. In this sense the policy prescribes a
linear decision rule.
The fann oftffis linear function, moreover, is intuitively appealing: Given the current
net inventory x(I), q(l) depends only on future demands, not past ones. This relationship is
positive; an increase in any of the d(lI) can only increase q(t). On the other hand, the rela
tionship is stable; an increase in d(u) causes a smaller increase in q(I). Also, q(l) depends
most strongly on the imminent demand d( I); subsequent demands d( u) are decreasingly im
portant. (This result is broadly consistent with the myopic approach above.)
Thus, we can view each =*(t) as a smoothed version of the forecasted demands d(u),
u ~ t. One unusually large demand increases all the prior orders, not just one or a few
nearby ones. Thus, there is typically less variation in the z*(I) than in the d(t) themselves.
(This observation inspires the nanle smoolhing model. Clearly, the DEL model is quali
tatively different in this respect. There, economies of scale induce sharply unequal or
ders, since z*(t) ~ 0 except at order times.) ~
The infinite-horizon version of this model leads to similar results (provideacertain
technical conditions are met), and for some special cases the solution is even simpler:
Suppose e(t) = -y'e and h(t) ~ -y'h, where e and h are positive constants and 0 < -y < 1.
Let 13 denote the (unique) root in the interval (0, I) of the quadratic equation
2
-yx + [I - -y + -ym]X - -ym = 0
Chapter 4 Time-Varying Demands 101
where 1;(1) ~ O.
The proof is by induction: The result is easy to check for f ~ T - I, so suppose it
holds for some f + 1, where f < T - 1. Substitute (4.5.5) and (4.5.4) into (4.5.6). setting
z(f) = q(/), to obtain
~-en the current ~(tlz(f) ~ c(f + 1)2:;':;+1 \3(f + 1, u) d(lI) + [~(f) - c(f)][d(f) - x(f)]
relationship is Solving this for z(l) and resetting q(l) ~ z(f) yields (4.5.4); the 13(1, u) are computed as
hand, the rela in the assertion.
J. q( f) depends Since c(f) < ~ (f), it is clear that 0 < 13(f, f) < 1. Moreover,
x:reasingly im
) 13(f f) > c(1 + I)13(t + 1, f + 1) 13(f f + 1)
demands d(lI), , ~t) ,
t one or a few
Finally, since 13(t + 1, lI) is decreasing in u for u 2> f + I, by assumption, so is \3(1, lI).•
t) themselves.
Jode! is quali
Iy unequal or- 4.5.3 Extensions
J\ided certain Suppose at each time t there is some ideal order quantity z*(t), and we incur ~
e\-en simpler: deviating from it. That is, the term c(f)?(I) in (4.5.2) becomes c(I)[Z(I) - z.(f)f. This
><10<,,< I. model is equivalent to an instance of the original one. Just redefine z(l) <- z(f) - z.(f)
Ion and d(f) <- d(l) - z*(f). Similarly. the model can incorporate a nonzero ideal net inven
tory x.(t).
Also, we can add linear tenns to the objective function. The results are nearly as sim
ple as before. [Formula (4.5.4) now includes an additional constant term.]
102 FOl/ndations qflnventory Management
m of a more re shown in Chapter 9. Even so, stochastic models of this general style, albeit with differ
oodel, but with a ent details, are often used to gain insight into complex systems. (Section 9.4.8.3. tor ex
ample, employs what is essentially a syrrunetry approximation.)
Finally, multiproduct smoothing models can be tormulated and analyzed in the
same spirit. The state and the control become fairly large vectors. But while the scale of
~onomiesof the model grows, its complexity does not; essentially the same methods can he used to
lmodel; let z.(t), solve it. In fact, in some cases, one can aggregate the products in a simple way, thus col
lapsing the large model to the single-product model above. As we shall see later, more
I:mand~~ realistic models cannot be reduced so easily. In some cases, however, a partial, approx
t.") by a qua ~ imate reduction can be achieved; see Section 8.6, for instance.
S'on is precisely
.on accurately
~ is a classic ap
., linearized op
Notes
~ original costs,
~ Assume that Section 4.2: The essential idea here is presented in Daganzo [1991]. A few very special
Jective function, cases can be solved exactly; see Resh et a!. [1976] and Barbosa and Friedman [1978].
ktivities (espe § 4.2.5 on the finite-production-rate model borrows ideas from Shuhnan and Smith
iI called the pro [1992].
lction levels~ Section 4.3: The DEL model was introduced by Wagner and Whitin [1958]; they also de
• "",Il as x(t). To veloped the original shortest-path algorithm. Since then a huge literature has accumu
ilJresent the pre lated. Recent overviews are provided by Salomon [1990] and Baker [1992]. Evans
iIbe smoothing [1985] discusses the details of implementation. Denardo [1982] provides a broad expo
sition of the dynamic-programming approach to modeling and computation. Three
~ the optimality groups of researchers, namely Aggarwal and Park [1993], Federgruen and Tzur [1991],
liar to the basic and Wagelmans et a!. [1992], independently and around the same time, developed as
~
t) is a linear ymptotically faster algorithms. Additional results on planning horizons can be found in
fficients are Eppen et a!. [1969] and Lundin and Morton [1975]. For more on nervousness and its pre
the details.) vention, see Kropp and Carlson [1984] and Federgruen and Tzur [1994].
instances of Section 4.4: Zangwill [1966,1968,1969] extended the DEL model to permit backorders.
tic control Federgruen and Tzur [1993] provide a refined algorithm tor this model also.
described by The DEL model with capacity constraints was posed by Florian and Klem
. 'onal con [1971]. Florian et a!. [1980] and Bitran and Yanasse [1982] demonstrate that the general,
I variables. unequal-capacity problem is NP-hard (indeed several special subclasses are themselves
NP-hard). Solution procedures can be found in Baker et a!. [1978]. Altemative formula
~
model can be tions are discussed by Karmarkat et a!. [1987] and Fleischmann [1990].
ClUTent state, Other extensions not covered here include Sogomonian and Tang [1993], who
inputs. The consider a system where demands are not exogenous, but rather are influenced by pro
motions.
J stochastic de Section 4.5: Smoothing models were introduced by Holt et a!. [1960]. These
jllical to that of were among the earliest discrete-time control models, and they inspired much
iofuJ than it may subsequent work in diverse fields. They have deeply affected our views of the role of in
l The solutions ventory at the macroeconomic level; see Blinder and Maccini's [1991] review article and
Pod reasons, as the historical studies of Blanchard [1983] and Kashyap and Wilcox [1993].
104 Foundations aIlnventory Management
Problems
4.1 Suppose the finite production rate fL in Section 4.2.5 is itself a function oftime, say fL(t)· Both ,\(t)
and fL(t) change slowly. (For instance, we may be able to adjust production capacity to anticipate the
known changes in ,\(t). For now, consider fL(t) as a given, positive, continuous function.) Describe a
reasonable policy, and explain why you expect it to work well.
4.2 Consider the finite-production-rate model of Section 4.2.5 (with constant rate fL). First, assume '\(t)
:5 fL for all t. Modify the definitions of C*(t) and C* in (4.2.2) appropriately to approximate the
optimal average cost. Next, assume '\(t) > fL for some t, but there is only one interval [I-,t+]
requiring an extended production cycle. Estimate the average cost over afinite interval [0, TJ, where
T> t+. Write a simple fommla tor the inventory x(t), I- :5 t:5 t+, assuming x(l-) = x(t+) = 0, and
use it to approximate the average cost.
4.3 Consider the linear-cost case of the DEL model in Section 4.3.2. Argue that the recursion (4.3.7)
correctly computes the minimal costs c[*, t) and the order times s(t).
4.4 The proof ofTheorem 4.3.1 in the text begins with the simplifying assumptions that Xo ~ 0, all
d(t) > 0, and all k(t) > O. Prove the result in the general case, i.e., assuming only that these
quantities are nonnegative.
4.5 In the DEL model suppose there are no speculation opportunities, so for every fixed t, the quantities
c[s, t) are nonincreasing in s. Let s'(t) denote the last order time in the t-horizon problem, as in
Algorithm Forward_DEL.
(a) Using the definition of k[s, t) in (4.3.8), show that s'(t) is nondecreasing in t. (Thus, to
determine s'(t) in iteration t, we need only search over s 2> s'(t - I).)
(b) Suppose that, for some t H, s'(t H) ~ tH - I. Using part (a), prove (by induction) that tH - I is
an order time in the t-horizon problem for all t 2> tHo Then, explain that this implies that tH is a
planning horizon.
4.6 This problem conducts a partial sensitivity analysis ofthe DEL model: Let d denote the entire
sequence of demands {d(t), 0 :5 t < T}, regarded as a nonnegative T-vector, and C*(d) the optimal
cost of the DEL model with demand sequence d, assuming the cost paranleters are fixed. Argue that
C*(d) is a continuous, nondecreasing, concave function of d. (Hint: Let 'IT denote any fixed
sequence of order times, and C(dl'IT) the cost of using the order times 'IT to meet the demand
sequence d. Express C*(d) in terms of the C(dl'IT). Now, derive certain relevant properties of the
C(dl'IT), and use these to arrive at the desired conclusion.)
What does this tell us about C*(2d), assuming we know C*(d)? Compare this finding to the
effect of doubling ,\ in the EOQ model.
4.7 As in the previous problem, let k and c denote T-vectors describing the order-cost parameters, and
COCk, c) the optimal cost, assuming all other parameters are held fixed. Prove that C*(k. c) is
Chapter 4 Time- Varying Demands 105
imations when continuous, nondecreasing, and concave. Can we draw a similar conclusion about C*(k, c, d) or at
e overall spirit least C*(k, d)? Why?
ng [1992], and 4.8 The DEL model includes constraints x(l) 2' 0 prohibitiug negative inventories. Suppose instead
aalman [1978] that there are lower bounds X(I) 2' L(I), where the x_(t) are any nOffilegative constants. Transform
the parameters and the variables of the problem to obtain an instance ofthe original DEL model
which is equivalent to this new oue. Restate the zero-iuventory property in terms of the new model.
4.9 The DEL model envisions a world in which time stops at I ~ T This is a fiction, of course. The
optinIal solution can depend strongly on the choice of T, a phenomenon called the horizon efjeCI,
and there are several ways to reduce it. One is simply to choose a large Twhenever possible. This is
certainly wise, but it is generally hard to know when T is large enough. This problem explores
fJ. Both "-(I) another approach.
anticipate the Suppose the objective (4.3.4) includes an additional term of the form V(T, x(T). Here, Vis a
L) Describe a funclion of x(T), represeuting the costs of alternative terminal inventory levels. Think of Vas
measuring future costs after T Suppose there is a specified, finite set of possible values of x(T).
_assume "-(I) Show how to solve this problem by determining the shortest path in a certain network. (Hinl: If
imate the x(T) is fixed to one particular value, the result is an instance of the DEL model, as in the previous
problem.)
1-,1+]
[0, 1], where 4.10 Consider the DEL model with backorders, as in Section 4.4.4. Prove that each order covers the
~I+) ~ 0, and entire demand for an interval of consecutive periods. (Hinl: Follow the proof ofTheorem 4.3.1. Or,
reformulate the model using the variables x + (I) and x - (I), so that the objective function is concave,
on (4.3.7) and then show that every extreme point ofthe feasible set shares the property above.)
4.11 Consider the network describing the DEL model with backorders, as illustrated in Figure 4.4.1.
= 0, all
Show how to assign a cost to each arc, so that the cost of a solution equals the smn of the arc costs
hese on the corresponding path.
4.12 In the limited-capacity DEL model, verifY that every feasible solution must satisfY x(l) 2' x>(I),
me quantities where the anticipated excess demand x>(I) is defined in (4.4.2).
~ as in 4.13 Consider the model with revised demands d(I), as defined in Section 4.4.5. For any feasible
solution to the original capacitated DEL model, construct a solution to this new model, settiug .(1)
= z(l) and X(I) ~ x(t) - xjl). Argue that this solution is feasible in the revised modeL Also, show
~ to
that the cost of the revised solution ditTers from that of the original by a constant. Thus, the revised
model is equivalent to the originaL Finally, show that d(l) S; z+(t).
at tH - 1 is
4.14 Consider the linear-cost case (all k(1) = 0) of the capacitated DEL model of Section 4.4.5. Explain
5 that tHis a
that the model is a minimum-cost network-flow problem. [The dynamics (4.3.2) provide the flow
conservation constraints.] Now, revise the model to obtain an equivalent transportation problem.
~ entire
(Hinl: The cost coefficients become c[s, IJ.J
the optimal
4.15 Consider the limited-capacity DEL model with equal capacities z +(1) = z +. Given a potential
d. Argue that
ixed inventory cycle specified by the pair of times (I, u), we want to determine the optimal solution
land within the cycle. The total demand over the cycle can be written in the form mz+ + r. where m is a
es ofthe nonnegative integer and r is the remainder, 0 .:::; r < z+; the total production during the cycle must
equal this quantity.
Argue that. for each time s in the cycle, the cumulative production fTom the beginning of the
ding to the
cycle until just before s (excluding z(s) itself) must be of the form 17Z + or 172+ + r, where n is an
integer between 0 and min {m,(s - I + I) l. Now, define a network: Each node indicates a time S
meters, and and a possible value of cumulative production up to s. (There is only one value at t, namely O.
k, c) is
106 Foundations of1111lenlOlY Management
Include a node for time u with production mz+ + ,) Explain exactly which nodes are included and
which arcs connect them. Also, show how to calculate an appropriate cost for each arc. Now, argue
that ti,e shortest path in this network determines the optimal production schedule within the cycle.
4.16 Consider t1le alternative formulation ofthe capacitated DEL model, in which a setup cost is paid
only at the beginning of a production run. Assuming s(l) ~ I and d(t) <; z+(t) for all I, show that
this model enjoys the zero-inventory property. That is, each production run begins with zero
inventory. (Follow the proof ofTheorem 4.3.1.)
4.17 Consider the production-smoothing model of Section 4.5.3. Following t1le proof ofTheorem 4.5.1,
verify that the optimal order-policy variable q(t) is a linear function of the state variables x(l) and
x~(t) and the demands d(u), u ;" I.
4.18 Let zo(l) denote a previous solution to t1le DEL model. Suppose we modify the DEL model to
penalize deviations from this solution, adding the following term to the objective for each I:
e included and
re. Now, argue
bin the cycle.
. cost is paid
5 SEVERAL PRODUCTS
r. show that
it:h zero AND LOCATIONS
beorem4.5.1,
bles x(l) and
model to
each I:
:d cost e + (I),
:.and
"'plain why
Outline
5.1 Introduction 107 5.7 Time-Varying Demand 158
5.2 Independent Items 112 5.8 Material Requirements Planning
5.3 Series Systems 120 (MRP) 163
5.4 Tree Systems 139 Notes 168
5.5 Coordinated Supply: Economies of Problems 170
Scope 150
5.6 Shared Production Capacity:
Economic Lot Scheduling 154
5.1 Introduction
This chapter explores systems with several products, or several locations, or both. The
goal is to nnderstand, both in detail and in hroad terms, what is needed to coordinate di
verse and dispersed activities, so that the overall system functions effectively. This theme
continues in Chapter 8 in the context of stochastic demands. These two chapters. then,
constitnte the core of the theory of supply chains.
This chapter mostly follows the scenario ofChapter 3 (continuous time, constant de
mand rates). Only at the end does it treat time-varying demands.
107
108 Foundations qj'Inventory Management
as well to one as the other. Indeed, a system with multiple products and locations can be
approached in the same way. The differences, such as they are, are mainly dlfferences in
wording.
This parallel should not be too surprising. Demand for one product cannot be met
by supply of another, just as demand in one place cannot be met by supply elsewhere. To
meet the demand, it is necessary to transform the available goods, either by moving them
or by processing them into something else. Production and transportation are both phys
ical transfonnations; both require time and money_ There are differences in detail, cer
tainly, but for our purposes such differences are superficial. So are distinctions among
modes of transportation and types of production processes.
So universal is this likeness that we build on it from the beginning. Instead of prod
ucts or locations, we work with generic entities called items. Items can indicate loca
tions, or products, or product-location pairs. A multi-item model can represent either
geographically separated points, or physically distinct products, or both.
There are a few apparent exceptions, models designed originally to capture specific
production- or transportation-oriented features. Even then, the specificity is more a mat
ter of interpretation than intrinsic structure. Virtually every production-specific feature
has some analogue in the context of transportation, and vice versa.
This is a powerful abstraction. Many innovations in distribution management have
been adnpted from the production sphere, and the other way around. The item approach
allows us to conceptualize, model, and manage production and transportation activities
in a unified manner. It may take a while to get accustomed to this idea, but it is worth the
effort.
iI.
f capture specific
ty is more a mat
~specific feature
\7", /\7
Wlagement have
Ieitem approach V
nation activities
ut it is worth the
\7 \7/
The next simplest structure is an assembly system (Figure 5.1.2). This usually rep
even when each resents production activities. As in a series system, there is only one finished product.
'stems. There may be several raw materials, however, all supplied exogenously. These are
Iy linked, either processed andlor combined ('"assembled") into components, which in tum are assembled
)Cess. Consider further, ultimately forming the final product. Some arcs in the network may represent
purchase goods transportation, moving materials, components, or the final product from one location to
i from the ware
another.
to various com A distribution system looks like a backwards assembly system (Figure 5.1.3). In
production terms, there is one raw material and several final products. The raw material
network, specif is successively specialized or refined as it moves through the production stages. In trans
IICS depict the
portation terms, the first node represents a central warehouse, and the ending nodes are
work structures: retail outlets; the nodes in the middle are intermediate stocking points, such as regional
lIS represent the
warehouses. Of course, a series system is a special case of both an assembly and a dis
lly chain. Each tribution system.
s the next one. A tree system (Figure 5.104) combines the features of an assembly system and a dis
~ last one meets tribution system, roughly in that order.
y and demand A fully general system (Figure 5.1.5) represents still more intricate relationships.
~ories and their Compare Figures 5.104 and 5.1.5: The general system includes distributionlike activities
such a system whose outputs are later combined in assembly operations, while the tree system does not.
This distinction is important. As we shall see, general systems are fundamentally more
110 Foundations a/Inventory Management
FIGURE 5.1.3
Distribution system.
/V
V V V 'V
'" '" V V
FIGURE 5.1.4
Tree system.
V V
V",
-. V
V V~V
FIGURE 5.l.5
General system.
V",
-.
V
:::=: V V
V V~V
Chapter 5 Several Products and Locations 111
complex than tree systems. All these more complex networks, from assembly to general
systems, are discussed in Section 5.4. Again, the main objective is to construct a good
policy and an accurate cost estimate.
We can even think of purely independent items as forming a network, though a triv
ial one with no arcs. In light of the classification above we might call this a parallel sys
tem. This notion is useful conceptually: For instance, a distribution system combines the
features ofthe series and parallel structures.
(Think of CA as a single symbol. Each tenn in the sum, and CA itself, are measured in
costs arise when, moneys per time-unit. Alternatively, think of CA as the product of two quantities, c and
as working capi A, where A = Xj Aj is the total demand rate, and c = ~ (A)A)C) is the demand-weighted
vst drivers in ac average of the cJ" This second interpretation makes sense when all the items have the
f these resources same quantity units, but not otherwise.) Second, let
~rtant measures
oad In this view C.A.)
W = weighted-average order-workload = ~j ( ~: wj
:se aggregates.
l1e perfonnance The product of wand CA is just the sum WCA ~ Xj WjCjA j • Finally, define
I possibly grasp
I monitor the ef (~YW;C;;f
~ enterprise, and
J. = variety index = WCA
'Ppropriate bal There are good reasons for this name: It is not hard to show (Problem 5.1) that
ls measuring or 1 .:o::;J• .:o::;J (5.2.1)
roach, however,
Within this range, J. measures the dispersion among the quantities WjCjAj over j. If
luaI-item EOQ they are very different, then J. is small, while if they are similar, J. is large. As we
resource usage shall see, J* captures the impact of the number of items on performance; in this
se these data to sense it measures the effective variety among them. (This usage may seem back
l1ively tractable wards; different WJCjAj mean less variety. A high-variety system here means a sys
u in practice. tem with many items, even if their parameters are identical, or one that performs
similarly.)
Now we define the performance measures themselves. Even in aggregate, perfor
mance depends on the choice of the policy variables
qj = batch size for item j
114 Foundations oflrrventory Management
Minimize IW(~)
.Iq. J
(5.2.2)
}
subject to 2:jCi~qj) ~ cI
or
(Notice the similarity to the EOQ formula!) Now, substitute these expressions into
the equation and solve for~; use this ~ in (5.2.3) to compute the optimal qj; and sub
q;
stitute the back into the objective function. The end result is the strikingly simple
formula
Chapter 5 Several Products and Locations 115
FIGURE 5.2.1
Inventory-workload tradeoffcurve,
Cl
•
) Given the as
.evant informa
~actual patterns
Itfor capacity
~ore this issue; cI
Thus, we determine the k} and hj in two steps, first measuring the "physical" quantities
Wjand c.i' and then using the factors K and 11 to convert them into cost rates.
Next, use these cost factors in a separate EOQ model for each item. The EOQ for
mula now takes the fonn
.
q;
'2(KW)A
.I .I
j ~ I, ... ,J
11 Cj
Comparing this with (5.2.3) above, the ratio TJ/K corresponds precisely to the dual vari
able ~. Substituting these q} into the sums defining wO and cI yields
wO = Y(~WcAJ.)e) cI~ Y(~wcAJ.)(~) (5.2.6)
Multiplying these two equations eliminates TJ and K entirely, and the end result is (5.2.4).
Thus, this point (cL wO) lies on the tradeoff curve. Moreover, it is easy to check that the
slope of the curve at this point is indeed -TJ/K.
In fact, the individual-item approach can recover the entire tradeoff curve: Just let
the ratio TJ/K range over all positive values. Using (5.2.6), each value generates a point
(wO, cl) lying on the CUlve, and the locus of these points is the curve itself. Thus, these
two approaches are truly equivalent.
Given 11 and K, the optimal total average cost is simply
C* = CA + K(WO) + TJ(cI) (5.2.7)
~ CA + V2(KW)(TJCA)J.
~ CA + y2(KTJ)(WCA)J.
Chapter 5 Several Products alld Locations 117
proportional to Thus, fixing the other parameters (and ueglecting the constant tenn cAl, C* is propor
below in Sec tional to the square root of J•. For systems with identical items, C' is proportional to the
square root of the number of items.
Clearly, the two-step approach to cost estimation depends on the assmnptions above
about Cj and wj • Those assmnptions rarely hold exactly in practice, but they often hold ap
proximately. Otherwise, if the assumptions are seriously 'Violated, aggregate resource us
it cost ofthe ag age cannot be captured in a single measure cI with a uniform cost rate 11, and likewise for
easures the cost wO and K. In such cases it is necessary to work with additional aggregate measures and
rest is certainly cost rates. For instance, it may be possible to identify tlvo key stocking-cost drivers. The
proportional to procedure then becomes the following: Estimate two coefficients for each item, say Cj! and
can then be ex CJ2' and two corresponding cost rates"'1 and 112. and then compute hJ = ll1Cjl + T12Cj2o
Similar cost-estimation approaches are frequently used to cope with large numbers of
(5.2.5) items. The better inventory-control programs include cost-estimation modules ofthis kind.
CA
As these simple examples illustrate, it is not enough to know the change in itself.
We need to know something about how the change is distributed among the items. Re
calling the formulas above, two additional pieces of information are relevant, namely,
the changes in the average workload wand the variety index J*. Now, it often happens
that the wJ are nearly constant, so w changes little in response to demand shifts. This
leaves J", as the remaining detenninant of performance. In the first example above J..
does not change at all, but in the second J. doubles along with cA; this accounts for the
difference in performance in the two cases.
The variety index J* requires a great deal of information about the individual items.
Sometimes it is possible to specify a particular distributional form for the item-specific
data. and thus to compute J* in tenns of a few parameters. See Problem 5.3.
Aggregate-level sensitivity issues are often expressed in terms of turnover. Recal1,
in the single-item context, the turnover is just the ratio 'A/l. The aggregate tunlOver is the
corresponding ratio cA/cI Thus, ifCA increases, we expect turnover to increase,provided
J.. does not increase too much. In general, we need to know how J.. changes, as well as
cA., to predict the effects on turnover.
The name "turnover" is sometimes attached to alternative measures. One is pA/c!,
where
PA = total revenue rate
= "£;P;A;
In this usage, turnover means the ratio of sales to inventory investment. This is a com
plex measure, and it is hard to interpret sensibly. It is helpful to express it as the product
of two more basic ratios:
pI. = (PA)(CA)
el cA el
The second term is the original turnover discussed above. The first ratio p)JcA is the ag
gregate markup. This too can be affected in subtle ways by demand changes. For in
stance, a shift in demand towards high-margin items increases it, and this can occur even
while CA and cI remain constant.
Generally, in working with aggregate pertonnance measures, it is wise to learn pre
cisely how the calculations are perfonned, and to reckon carefully the various factors that
influence them. This may sound obvious, but it is easy to forget.
mge in CA itself. Some of the European subsidiaries had enjoyed rapidly increasing sales in recent
g the items. Re months. The executives were concerned, however, that turnover had actually declined
::levant, namely, somewhat. What was wrong? Had the European managers lost control oftheir inventories?
it often happens First, we looked at the aggregate markup (pA/CA). These figures had fluctuated
and shifts. This over that same period, but in no clear, systematic direction. A more detailed exami
ample above J. nation revealed that there had in fact been a shift toward higher-margin products, but
iCCOunts for the their prices had been reduced also, and these two effects more or less cancelled each
other.
ldividual items. Then, we estimated the variety index (J.). This was not easy: We had no item-spe
Ie item-specific cific workloads (Wi), and the expenditure data (CjA) were incomplete. So, we assumed
5.3. all Wj = W for each subsidiary. Then, with the limited expenditure data and additional as
unover. Recall, sumptions (along the lines of Problem 5.3), we patched together an estimate of J•. (We
~ turnover is the were a bit uncomfortable with this crude approach, but the practical need for a quick an
rease, provided swer overrode our reservations.)
[lges, as well as The results were c1earcut: Most of the subsidiaries' J*'s had increased substantially.
Further investigation provided the explanation: The sales expansion had come mainly
;. One is pAlc!, through opening new markets in widespread geographic regions. The effective number
of items had grown even faster than revenues, so the inventory needed to support them
had increased sharply. (This also explained the price reductions; managers were cutting
prices aggressively to attract new customers.) In such circumstances, declining turnover
was perfectly natural. Far from losing control, the European managers were apparently
doing their jobs well.
This is a com
:as the product
5.2.6 ABCAnalysis
ABC analysis is another tactic for coping with a large number of items. Essentially, it
means dividing the items into a few groups. Commonly, three groups are used, labeled
A, B, and C on the basis of sales volume. The A items have the largest values ofPjAJ' the
'AICA is the ag B items medium values, and the C items the smallest. Normally, the A group includes
I3Ilges. For in only a few items (say, 10%), the B group is larger (30%), and the C group is the largest
::-an OCCUf even (60%). Even so, the A items typically account for the bulk of total sales (often as much
as 80%), while the C items cover only a small fraction, with the B items somewhere in
Ie to learn pre between.
IUS factors that Thus, ABC analysis identifies the most important items, the A group, and the least
important ones, the C group, putting the rest in the B group. This is a preliminary step
to other modes of analysis, such as parameter estimation and modeling. The idea is to
focus effort where it counts the most. Thus, the A items deserve the most intensive data
collection and model-formulation efforts, while relatively crude methods can be used for
"'5: Some years the B items, and even cruder ones for the C items.
oonth they re Some writers offer much more specific advice, recommending particular models
al hundred of for particular groups (e.g., the EOQ model for B items). This is, in my view, too specific.
'. the reviews The "right" model depends, as usual, on many other factors. The B items in one system
) and turnover may be far more important and complex than the B items in another system, and thus re
quire entirely different models and methods.
120 Foundations afInventory Management
The basic notion of dividing items into groups can be usefully applied to any large
data set. It is not just a technique of inventory analysis, but rather much more general.
FIGURE 5.3.1
Series system.
\7- v 9
Chapter j Several Products and Locations 121
lied to any large are distinct from the flU), which measure inventories at the stages. Still, a shipment from
Imore general. j to j hi,
+ 1 essentially consists of stock of item}, so it incurs holding cost at rate just
like Ij(t). (Shipments from the source to stage I incur no such costs.) But, under any
plausible policy, on average, there must be exactly AL;+ 1 units in transit from} to j + 1,
so the total average pipeline-holding cost is ALj<Jh;L}+l' This is a constant. Conse
quently, we ignore these costs from now on. Likewise, we ignore variable purchase and
I, ... , Jfrom shipment costs, since these are constant over all sensible policies.
ges, or stocking Also, as in the EOQ model, we can anticipate leadtimes just by shifting the orders
and an external in time appropriately. as explained below. For simplicity, then, assume each L;= o. So,
ies item 2, item it is possible to order stock from the supplier and pass it through the successive stages,
[l this context is even all the way to stage J, instantaneously. Assume the system starts empty, i.e., 1;(0-)
= 0 for all}.
is a decision to All the proofs in this section are collected at the end, in Section 5.3.8.
)f a prior stage.
and control are 5.3.2 Echelons and Echelon Invelltories
sense to order a
my. (Neverthe Here is a fundamental concept in multiitem systems: The echelon of stage} (or echelon
entralized man- j for short) comprises stage j itself and all downstream stages, i.e., all stages i 2'}. The
echelons of a four-stage system are indicated by rectangles in Figure 5.3.2. This notion
~are economIes captures the supply-demand relationships in a useful manner: First, stage J is its own
11 could operate echelon. The external supplier and all the prior stages can be viewed as stage J's supply
S to find a good process. Likewise, consider echelon J - 1, i.e., the last two stages. This is another sub
system, whose supply process includes the earlier stages i < J - 1. Continuing in this
manner, the entire system can be viewed as a hierarchy of nested subsystems, the eche
lons, each with a clearly defined supply process.
(In military usage the word echelon is a generic term for a group of troops. Thus, an
echelon may refer to a division, or a battalion, or a company. A division includes several
smaIler groups, which include stiIl smaIler groups, and so on.)
Imagine a multistage production process. Item I is a raw material. At each stage the
tee a unit of its material is transformed somehow, or various enhancements are added until the final
11 one lUlit, just product emerges as item J. So, in a sense, a unit of item i > } includes one of item j. The
.d product. For total system inventory of item} thus comprises, not just I;(l), but also the inventories
asured in tons, downstream. To express this idea, define
olin two-tons
~. Accordingly,
FIGURE 5.3.2
Jere are no ca
Echelons.
~tories. These
9- Y7-IY7-~
122 Foundations a/Inventory Management
The original I;(t) is sometimes called the local or installation inventory of item j. The
prime indicates that it is a local quantity.
Also, let
h} = echelon-inventory holding-cost rate for itemj
=hi-h}-l
where h o = O. Assume that each hj > O. (This is usually the case: Suppose Cj is the vari
h;
able order-cost rate for item), and includes financing costs only. To create a unit of
item} incurs all the costs Cj, i <j, so h; = a"l/5,jcj, where a is the interest rate. Thus, the
echelon holding cost is just hj = CiCj, that is, hj rel1ects the value added at stage j. And
cj > 0 inunediately implies hj > O. If h; also includes a physical handling cost ilj, the il;
must also increase, or at least not decrease too fast. This is usually true also; physical
handling tends to be more expensive downstream.)
With these definitions the systemwide inventory cost rate becomes
FIGURE 5.3.3
Local inventories over time.
r of item j. The
I; (/)
.§
;e cj is the vari " 0
~
:reate a unit of
rate. Thus, the
at stage j. And I; (t)
~ cost hj, the hJ
also; physical
o
Time (t)
system just as
ystem. Figure
) behaves like
order epochs,
complex step
FIGURE 5,3.4
.12 (1) = I~(I),
Echelon inventories over time.
_(The dashed
lers but stage I, (t)
:ate A, except
lIl. inventories ,
,
,
, 1
,,
.§ ,
" 0
~
12 (t)
:quent stages
: system, one
time system:
~ the start of
rj < J occur
ts empty and o
me shift, the Time (I)
124 Foundations oflm;entory Management
~(t) are identical in the two systems. Conversely, given a feasible policy for the real sys
tem, we can construct a policy for the zero-leadtime system by the reverse transforma
tion. So, we can and do focus on zero-Ieadtime systems.
gj = h}.
Note, under a This is called the relaxed problem, because its constraints relax those of problem (5.3.2).
ry I;(t) is zero. Thus, its optimal cost C- is a lower bound on the minimal cost of problem (5.3.2), but
, intervals be that's not all:
an others, and
THEOREM 5.3.3. The optimal cost C- of the relaxed problem is a lower bound on the
ler quantities.
average cost of any feasible policy.
lore natural to
~cy. One can The relaxed problem (5.3.3) is obviously simpler than the original (5.3.2). All thc
tually use this variables are continuous, and the constraints are linear inequalities. Moreover, the ob
jective function is strictly convex, so (see PropositionA.3.6 of Appendix A) the optimal
ry, stationary solution is unique. Call it u*.
Thus, (5.3.3) is a standard nonlinear program; one way to solve it is to use a stan
dard nonlinear-progranuning algorithm. But there is a better way. Problem (5.3.3) has a
great deal of special structure. We shall apply an algorithmic strategy (a variant of the
active-set strategy; see, e.g., Luenbcrger [1984], pp. 326-330) that exploits this special
structure to develop a simple, fast method to solve (5.3.3).
The key idea is to distinguish which constraints in (5.3.3) are binding at the optimal
~erefore, ap solution. Let N = {I, ... ,J} denote the set of items. and let A = {I, ... , J - I} index
the constraints. For any subset A= C A, construct a modified version ofthe relaxed prob
lem: Tighten the constraints tor} E A= to equations, and omit the others:
(5.3.l)
Minimize Cluj (5.3.3[A ~])
best such pol-
subject to Uj = Uj + I jEA~
126 Foundations ofImwaory Management
FIGURE 5.3.5
Clusters.
where next(J) = 0, and prev(m) = 0 for the lowest index m. So, for all m, Nm = {prev(m)
+ 1, ... ,m}.
Now, (5.3.3[A~]) separates into INI independent subproblems, one for each cluster.
The subproblem for cluster Nm is
k 1 ]
Minimize ~jENm [ ~ + 2gjUj
subject to Uj = uj + 1 jENm,j<m
or equivalently, using one additional variable u,
Minimize k
.J.. 1 -u ]
+ -zgj
I jENm [U
subject to Uj = U jeNm
For any subset M c::: N denote
k(M) = "i.j<M kj
geM) = "io'M gj
'1T(M) = k(M)/g(M)
Chapter 5 Several Products and Locatiorls 127
~
U 2
But this is the cost function of an EOQ model. Its minimum occurs at " ~ u(m) =
[2k(Nm)/g(Nm)]rl2 = [2'IT(Nm)] 112 Thus, the solution to (5.3.3[A~]) is
u(m) ~ [2'IT(Nm )]1I2 m E N\A~
FIGURE 5.3.6
Cumulative costs.
-"
FIGURE 5.3.7
Concavified cumulative costs.
",
,,
,6
_.0
_-:0---
-"
g
Chapter 5 S(~'eral Products and Locations 129
This partition satisfies the optimality conditions: The slopes of the new func
tion's line segments are precisely the 7r(Nm ). These are nonincreasing, by concavity,
so (5.3.5) holds. Also, to cut any N m means to add a point j E A ~ n N m to the curve,
replacing Nm's line segment by two segments, like the dashed lines in Figure 5.3.7,
one each for N-;'" and N~. That new point lies below (or on) N",'s segment, so the new
segments have nondecreasing slopes, i.e., 'IT(N-;"') ::'S 7r(N~). Thus, the cut satisfies
(5.3.6).
Thus, to detennine an optimal partition N and thereby solve the relaxed problem,
we need only concavify a piecewise-linear function. There are several ways to do this.
The following algorithm, in essence, concavifies the curve through (0, 0) and the nextj
points, first for j ~ I, and then for each successive j up to J:
Algorithm Series_Relaxed:
Set~ ~ UU ~ I,,J
Forj~ I, ... ,J:
This algorithm is easy to implement, and its calculations are simple. (One can show that
it requires O(J) time.) Let N* denote an optimal partition, and u*(m) the optimal u(m)
in (5.3.4).
Initialize:
i = 1:
prev(l) = 0, so no change
130 Foundations ofInventory Management
j ~ 2:
1T(Nll ~ 4 <8 ~ 1T(N2 ), soresetN, ~ {1,2}
prev(2) = 0, so no change
j ~ 3:
The optimal partition is N* = {N 1 , N 4 }, where N 2 = {l, 2}, N 4 = {3, 4]. The optimal solution to
the relaxed problem is
• ~ II,• ~ It•(2)
It, ~
12.12)'" = 3.464
1--
\ 2
. . .
U3=U4=U(4)=
(2.4)'"
-2 =2
and its cost is L := 10.928.
Incidentally, there is an intnguing and useful interpretation of the solution, based on the
duality theory for nonlinear prograrruning. Let ~j be the dual variable for the constraint uj ' "
0;
'~+b and (its optimal value,) < J, and let = 2(~; ~;-l) with ~~ = ~~ = O. The optimal -
ltyconditions for (5.3.3) implythatu;' ~ 2kJ(gj+0;) = 27f(Nm ),jENm . Thus, given the7f(Nm )
0;.
for an optimal partition, we can directly recover the Also, 1; = 21; (~; - ~;-l) ~ O. So, 0;
0;
the essentially reallocate the holding costs, or more precisely the gj' among the items, to re
cover the optimal solution. There is an entirely analogous construction which reallocates the
fIxed costs kj instead. (Just restate the constraints as lluj -;: lIuj +b and proceed as above.)
These are just interesting facts for now, but we shall use them later in Section 5.7.
uo(m))
n(m) ~ closest integer to log2 ( ---;;
(Iflog2(u O (m)/y.) = 3Y. say, which is squarely between 3 and 4, set n(m) = 3. In general,
resolve ties by rounding down.) TIlls policy is feasible: Since the u*(m) are nonincreasing
in m, so are the n(m) and the U + (m), hence the u; are nonincreasing in}. Moreover, each
u; is an integer multiple of u;+
l' (Specifically, u;
is a nonnegative-integer-power-of-two
multiple of u;+ ,,) < J. A policy with this property is called a pawer-qr-twa policy.)
How good is this policy? Let C+ ~ qu+). Recall the EOQ error function E(X) =
ptimal solution to '';(x + l/x) for X> 0, defined in (3.2.5).
THEOREM 5.3.6. C+ 0; E(Vz)e.
A direct calculation reveals that E (v2) = 1.06. So, the cast a{the palicy is, at warst, 6%
more than that ofany other policy. This is quite good performance. The result holds for
any choice of base period M..
EXAMPLE 5.3.A, PART 2
In the example above, let us use 11 ~ 1. IOg2(U*(2)/11) = 1.79, so n(2) ~ 2, and IOg2(U*(4)/11) ~ I
00, based on the ~ n(4). So, u+(2) ~ 22 ~ 4, ,,+(4) ~ 2' ~ 2, and
: constraint uj ~
O. The optimal c+ ~ [¥ +~(2)(4)] +[~+ ~(2)(2)] ~ 11
given the 'IT(N"J
- ~;'l) = O. So, Thus, C+/C- = 11/10.928 = 1.007; that is, the solution is at most 0.7% above optimaL This is
the items, to re much better than the theorem's guarantee of6%.
1 reallocates the
ceed as above.) 5.3,6 Constructing a Feasible Policy with No Base Period
5.7.
Now drop the base-period restriction. The method above seems attractive nevertheless,
so continue to use it. Let C+ (M.) be the cost C+ above. now regarded as a function of H.
Choose 1J. to minimize this function.
This, it turns out, is not terribly difficult. First, observe that ll:. can be restricted to the
, feasible solu fairly small interval [I, 2): Starting with any Y. and then doubling it, each n(m) drops by
I, so u + (m) remains unchanged. Likewise, halving Y. increases each n(m) by I, so again
lch as a day or u + (m) is unchanged.
tre t). We shall
Second, consider what happens to anyone !I +(m) as Y. increases from I to 2. For a
'Y.. (That is, l~ while, n(m) stays constant, so u + (m) is linear in Y.. At some point, n(m) drops by I and
ical reasons to u +(m)jumps to anew value. After that point, 11 +(m) again changes linearly. There is pre
rys, than some ciselyone such breakpoint within (1,2). Thus, each u -I- (m) is a piecewise-linear function
onstruct a fea with two linear pieces. (The exact expression for !I +(m) is given in the proof ofTheorem
m.)
5.3.7 in Section 5.3.8.)
132 Foundations q{InventofY Management
Now, there is typically a different breakpoint for each m. Still, the entire vector u +
is a piecewise-linear function ofg with at most IN] + I pieces. Within each piece C+(g)
is an EOQ-like cost function, so it is easy to optimize. Select the best of these solutions.
(Problem 5.8 asks you to work out the technical details.)
Let C+ .. denote its cost. Of course, this is one possible value of C+, so Theorem
5.3.6 bounds it above. Here is a stronger result:
THEOREM 5.3.7.
C+ + S [In(2~\l2 ]C-
The peculiar constant here is about 1.02. The cost ofthis policy is no more than 2% above
that ofany other policy. This is truly outstanding performance.
EXAMPLE 5.3.A, PART 3
Let x(m) denote the breakpoint for u+(m). By caleulations explained in Section 5.3.8,
li X(4)$li<2
We must evaluate C+(g) over three intervals, [1, X(2), [X(2), X(4)), and [X(4), 2). Over the third
one,
+
C (li) ~ 2"
[12 + 2:12(2li)]+ [4: + 2:12" ]
10 I
= - + -6u
~ 2
By the EOQ formula, the best value of 11 is 11* = (2' to/6)~ = 1.826, and the cost of this solution
is C+(u*) = 10.954. The calculations for the other intervals are similar. The third one turns out to
be best.
The solution is thus u+(2) = 2g* = 3.651, u +(4) = 11* = 1.826, and its eost is C+* = 10.954.
So, c+*/c- = 10.954110.928 = 1.002. This solution is at most 0.2% above optimal, much better
than the guaranteed 2%.
ntire vector u + We spoke at the outset ofthe need to coordinate the items. How exactly does the pol
.ch piece C+ (11) icy do this? The key idea is to coordinate the timing ofevents throughout the system. The
these solutions. stages are partitioned into clusters. Within each cluster, the policy tells us to synchronize
the stages; order for all of them at the same times. Between clusters, the power-of-two
.....,-, so Theorem policy enforces a partial synchronization. Each cluster's orders coincide with some, but
not necessarily all, of those of the next downstream cluster, specifically, either with
every downstream order, or every other order, or every 4, etc. Finally, cluster N/s orders,
and therefore every cluster's orders, occur regularly, at equal time intervals.
Among conceivable coordination mechanisms, this is a very simple one. It is re
markable that it performs so well.
Still, this is a centralized control scheme. It is possible, however, to implement the
than 2% above same policy \'v'ith a minimal degree of central direction: Because the stages within each
cluster all use the same order interval (and order quantity), only the last one actually
holds inventory. Whenever any earlier stage receives a batch, it processes the batch im
mediately and passes it through to the next stage. With this rule in place, we need only
;.3.8, manage the flow of goods between clusters. Imagine that each cluster operates an
EOQ-like policy, based on its own local inventory. The last cluster N J orders q + (J) ~
AU +(J) quantity-units every u +(J) time-units, and transmits these orders upstream to
cluster Nprev(J)' Cluster Nprev(J) views these orders as discrete demands; it fills them
as they occur, as if they were exogenous, by transferring batches downstream to clus
ter N J . Cluster Nprev(J) in turn places orders with its predecessor, cluster Nprev(prev(J»'
In actual physical units, each of Np«v(J/s orders is for the quantity q + (prev(J)) ~
Au+(prev(J)); in terms of the implicit demands Nprev(J) sees, each order is the inte
ger q+(prev(J))/q "'(J) = u+(prev(J))/u+(J). So, cluster Np'~(J) operates just like the
'. Over the third single-item, discrete-demand EOQ model of Problem 3.7. Likewise, every other clus
ter receives orders from its successor, treats these orders as its own demands, and
places orders with its predecessor.
Clearly, this approach correctly implements the poTicy. It is necessary, of course, to
specify the clusters in advance and tell each one its order interval (or order size). In real
time, however, there is no need for centralized information and control. Each cluster
monitors only its own local inventory. The orders passed between them provide suffi
of this solution cient information to coordinate the system as a whole.
)De turns out to This equivalence between local and centralized control is a powerful idea in prac
tice, for a local-control scheme is often much easier to manage. This issue is revisited in
c* ~ 10.954. Section 5.8 and Chapter 8. Also, whether local or centralized, the policy's orders can be
ai, much better driven by time or stock quantities, as in the EOQ model; in certain situations, one may
be more convenient than the other.
Let us now consider the overall performance of the system. The relaxed problem's
cost C- accurately estimates the total optimal cost and the cost of the heuristic policy.
Given an optimal partition N*, this quantity is
em, and then
(5.3.7)
eriod is spec
, error bound We can use this fonnula to address sensitivity issues:
)f the heuris Suppose A changes. Examining (5.3.5) and (5.3.6), we see that the optimal partition
'S. N* remains the same. Each term of C- in (5.3.7) is proportional to y:;..,
so C- itself is
134 Foundations of111ventmy Management
too. Thus, the effect of A on the overall cost (as approximated by L) is the same as in the
EOQmodel.
Also, each u*(m) changes as I/y'I;, so the (relaxed) relative order frequency
u*(m)Iu*(next(m)) is insensitive to A. Consequently. u + (m) is approximately proportional
to lIy'1;, though not exactly. And while the true relative frequency u + (m )Iu + (next(m))
may change with A.• the change is limited. (In fact, this ratio may go up or down by a fac
tor of 2, but no more. and of course it always remains at least I. See Problem 5.10.)
Similarly, cost-factor changes affect C- as in the EOQ model, provided the ky all
change by a conunon multiple, and likewise the hp For instance. if ky = KWj and hj ~ 1)Cj
as in2ection 5.2, and K and 1) change but not the wj and cj , then C- is proportional
to yK1).
In this special case, moreover, one can easily construct an aggregate inventory-work
load tradeoff curve, estimating wO and cI by their values in the relaxed problem. The op
timal partition N* must be determined only once, since it is independent of K and 1). The
tradeoff curve is precisely (5.2.4), the same !orrn as in an independent-item model, with
the aggregate system parameters redefined as follows: For any subset M (;; N let
w(M) ~ lj<Mwj
c(M) = 1;<M cj
Then redefine
CA = 1., C(Nm)A = (1; C)A
W = 1., [ C(Nm)A]
----;:;.:- w(N.,)
J.~
(1 m Yw(Nm)C(N.,)A)2
~~~~~,
WCA
Problem 5.11 asks you to verify these formulas. Also, the optimal cost C- can be writ
ten in terms of these parameters. as in (5.2.6). (The index J. here characterizes effective
variety among the clusters of the optimal partition. This is related to dispersion among
the original items, but indirectly.)
Also, let C- - denote the optimal cost, omitting the constraints of the relaxed prob
lem and optimizing each item separately. Obviously, L ;;" C--. They are equal when
the ratios 1T({i)) ~ k/g; are nonincreasing. Thus, C- is forced above C-- only when a
downstream stage has a larger ratio than an upstream stage.
This observation suggests a crude but useful guideline for fixed-cost (or order
workload or setup-time) reduction efforts: Focus attention on downstream stages. While
reductions anywhere are welcome, downstream fixed costs have the greatest system
wide performance effects.
EXAMPLE S.3.A, PART 4
he same as in the c- = to.928 is larger, because kz. = 8 is more than k1 = 4. (The kj further downstream are both
2, i.e., tess.) So, focusing on the -fIrst two stages, k~ is more critical than k1•
order frequency
Performance evaluation is important also in the context of system design. In logis
!ely proportional
(m)/u+(next(m)) tics, typically, we have some choice over the number and placement of stocking points.
In a production setting there may be several alternative processes, and numerous alter
r down by a fac
'lem 510.) native ways to combine basic operations into stages. To compare alternative designs, of
course, we should estimate their initial costs ofconstruction, equipment, and the like. We
7Vided the k.i all
tH)' and hj = TlCj
should also estimate their ongoing operating costs, and for this purpose we can calculate
C- for each alternative.
is proportional
tIlVentory-work_
5.3.8 Proofs
:oblem. The op
of K and 1]. The
THEOREM 5.3.1. Every non-nested policy is dominated by a nested policy. (The
'Ill model, with
nested policy has lower inventories and no more orders.)
;Nlet
PROOF. We prove the result for a two-stage system. (problem 5.5 asks you to extend it
to general J.) Consider a feasible, non-nested policy. Suppose in particular that, at time
t, stage I orders but stage 2 does not. Let u be the earliest time after t when stage 2 does
order. Thus, all stock arriving at stage 1 at time t must remain there at least until u. Now,
consider an alternative policy, where the stage-l order at t is postponed until u; other
wise (before t and after u), the new policy is identical to the old one. Clearly, 12(s) is the
same in both policies, but I~{s) is lower in the new one, t :5 S < u, and the total numbers
of orders at both stages remain the same. Thus, the new policy dominates the old one.
Clearly, this construction can be repeated to eliminate all non-nested order times t.
At this point, the earliest bad order time, if any, comes after u. We can repeat this
procedure as long as there are any bad order times, pushing the earliest one later and
later, while improving the policy at each step.
THEOREM 5.3.3. The optimal cost C- ofthe relaxed problem is a lower bound on the
average cost of any feasible policy.
(We remarked in Section 5.3.3 that stationary-interval policies dominate all others. In
view of the previous two theorems, then, we already know that this assertion is true. The
following argument, however, uses only the nestedness and zero-inventory properties.)
PROOF. Consider any feasible, zero-inventory policy. Let OFP) denote the total num
ber of orders for item j during the time interval (0, t), aud cet) the total cost. Thus,
k
cet) "" t 'Lj [ .L I
+ -gjU ]
j = tceu)
uj 2
By Theorem 5.3.1 we cau assume the policy is nested, so OFit) :s OFj+1(t) or
uj "" Uj +l' That is, u is feasible in the relaxed problem, so C(u) "" C-. Cousequently,
cet)lt "" C-: The average cost over anyfinite time t is bounded below by C-. Surely, this
inequality holds too when t -> 00.
THEOREM 5.3.4. Conditions (5.3.5) and (5.3.6) ou a partition N are necessary and
sufficient for the solution u given by (5.3.4) to be optimal in the relaxed problem (5.3.3).
PROOF.
Necessity. Fix A~ and N, and let u- be the solution to (5.3.3[A~]), as given by
(5.3.4). Equation (5.3.5) is necessary for feasibility. Assume it holds. Indeed, assume
the inequalities (5.3.5) are all strict. (Unot, one can construct an equivalent partition
with strict inequalities, as remarked in Section 5.3.4 after Theorem 5.3 A.) Now, sup
pose some cut (N-;'" N:) violates (5.3.6). Construct a new partition N' by splitting
N m into the two clusters N~ and N;, let u I be the corresponding solution (5.3A), and
set u' ~ "yu' + (I - "y)u-, 0 < "y < 1. For sufficiently small "y, u' is feasible for the
relaxed problem. (The u} are equal within each cluster ofN', (5.3.5) holds strictly for
N, and 7r(N~) > 7r(N:).) Also, ceu') < ceu-), so ceu') < ceu-), 0 <"y < 1. (This
follows from the strict convexity of C.) That is, u - cannot be optimal for the relaxed
problem.
Chapter 5 St.'\-'eral Products and Locations 137
can repeat this Sufficiency. Suppese A~ and N satisfY (5.3.5) and (5.3.6). Consider the special caseA~
itone later and = A, so there is only one cluster, N itself. Here, (5.3.5) is vacuous, and (5.3.6) says that,
for every cut (N-, N+) of N, -rr(N-) ,.;; -rr(N+). Assume to the contrary that u* '" u
er bound on the
solves the relaxed problem. Defining A: = {j E A : = u; u;+
l}, as above, u* also solves
(5.3.3[A:]). The corresponding partition N' ~ (N:} satisfies (5.3.5) (since u* is feasi
ble for the relaxed problem), and at least one ofthose inequalities is strict (otherwise u*
te all others. In ~ u-), say-rr(N:) > -rr(N;"xttn)' SetN- ~ [j:j";; n} andN+ ~ [j:j> n}. Now,-rr(N-)
Lion is true. The is a weighted average of the -rr(N:,,), m ,.;; n, using weights g(N:,,)/g(N-), and similarly
<y properties.) 7T(N+) is a weighted average of the 7T(N;,), rn > n. Moreover, all the 7T(N;,) in 1i'(N-)
are strictly greater than those in -rr(N+), and so -rr(N-) > -rr(N+). This contradicts (5.3.6),
, the total num
sou*=u-.
:os!. Thus,
Tum to the general case A~ C;; A. By (5.3.5), u- is feasible for the relaxed problem,
so C(u-) 2': C-. Consider the problem
Minimize C(u) (5.3.8)
gnoring all the
'j(l), this quan subject to U.i 2:: Uj + 1 JEA._
ublem 3.2), in
This problem relaxes the constraints of both (5.3.3) and (5.3.3[A~]). It suffices to show
that u- solves (5.3.8), for then C(u-) ,.;; C-, so C(u-) ~ C-. But (5.3.8), like
(5.3.3[A~]), separates into several independent subproblems, one for each cluster N m .
Each subproblem, moreover, is a smaller version of the relaxed problem (5.3.3). Apply
ing the argument above (for the special case A __ ~ A), we see that the optinoal partition
for Nm's subproblem has a single subcluster, Nm itself, and the optimal solution is the
subvector (U7)j.Nm' Thus, u- solves (5.3.8).
THEOREM 5.3.5. Algorithm Series_Relaxed determines an optimal partition, and so
,; 0Fj+I(I) or the corresponding solution (5.3.4) solves the relaxed problem.
Consequently,
-. Surely, this PROOF. It suffiees to show that the algorithm eoncavifies the function illustrated in
Figure 5.3.6. We argue, by induCl10n on), that step j concavifies the firstj points. The
result is true for j = 1. (Since prev(l) ~ 0, step I leaves N[ ~ (I}.) Suppose it is true
llecessary and forj - 1. During step j, N; is reset until either prev(j) = 0 or -rr(Np,~tj) > -rr(N;). At that
oblem (5.3.3). moment, the slopes 1i'(Nm) for m :::; j are decreasing, so the function is concave up to
point j. Also. each time f{i is reset, two line segments are replaced by a single segment
lying on or above both of them. So, all points remain on or below the curve.
. as given by
deed, assume THEOREM 5.3.6. C+ ,.;; E(V2)C.
Jent partition
~.) Now, sup
PROOF. Given an optimal partition N*, Cm[u*(m)] = [2k(Nm)g(Nm )]1/2 By equation
, by splitting (3.2.5) in Chapter 3,
0(5.3.4), and
asible for the + [u+(m)] 1/2
Cm[u (m)] = E u*(m) . [2k(Nm)g(N,,,j]
:Is strictly for
-y < 1. (This Now, the definition ofn(m) implies
IT the relaxed
-15";; n(m) + log2 (li) - log, (u*(m» < +15
138 Foundations ofInventory Management
or
I
ICg) ~ In (2) . g
It is not hard to check that this is a probability density function over [I, 2). So,
C+, oS f1/Cg)C+Cg) dg.
Write C+Cg) ~ lm C:C!!), where C:Cg) is the quantity Cm[u+Cm)], now expressed as a
function of g. Thus,
C+, oS lm fi!Cg)C:Cl1)dg
[
Then, xCm) is the breakpoint in [I, 2) of"+Cm).lndeed,
2+1/2]
fdC!!)
2
e [,,+cm)]
- - du =
u'Cm)
Chapter j Several Products and Locations 139
[ 1]
In(2) .
I ([2+1/2]) (dll\ , ([2-1/2]) (dll))
\ncm)E x(m) Y. I;) + J~(",)<\ x(m)!!. ,;
Change variables in the two integrals; use x ~ [2+ II2 /x(m)]y. in the first, and x ~
~. Consequently, [T 1I2/x(m)]y. in the second. The sum ofthe two (the expression in braces) then reduces to
J"'"
,·w [«X)]
----;- dx _ y2
-
1
1 y2 ]
C+· :;; L m [2k(N",)g(N",)] 112 [ In (2)
~
[Val [1, 2). Cer
K! on C+·. De Ln(;)y2] C
cessor. An item may have several successors. The official definition of a tree s}'stem is a
bit more intricate: Ignoring the directions of the arcs for the moment, the undirected net
work has no circuits; in other words, there is a unique (undirected) path connecting any
two nodes. Provided this condition holds, an item may have several predecessors and
successors. In all these cases the number of arcs is precisely IAI = INI - I = J - 1.
The cost factors ~j and hi have the same meaning as in a series system, as do the lo
cal inventories li(I). Also, let
Aj = demand rate for itemj
x.;
Only an end item may have > O. (This is no real restriction. If some non-end item does
have Ai > 0, add a new item/, to the network with (j, j") E A, kr ~ 0, hi" = hi, and
A'l" = Aj, and reset Aj = O. Usually, every end item has A,f > 0, but this is not strictly
necessary; see Section 5.4.7.)
W
[lon-end item does
I
L' 24 I
--I
L'45
I
L'14
as order leadtime ment. The "project" is the creation of one isolated unit ofthe end product, itemJ ~ 5. The
'Ie system has the arcs represent activities that must be perfonned to complete this project, and the Lif are
their durations. Ofcourse, there are additional activities corresponding to the Lj, but ignore
ude item-specific them for now. Some activities can be perfonned simultaneously, but certain activities re
not be equal over quire others to be completed first; the structure of the network captures such precedence
ponding batch of relations. Evidently, Figure 5.4.1 depicts a feasible schedule for the entire project.
00 sense to speak
We can draw a distribution system in the same way. See Figure 5.4.2. Let Li be the
l Rather, the key sum ofthe arc leadtimes along the path from I to j. Renumber the items, if necessary, so
cs. The inventory that item Jhas the longest suchL:;: and set Li = LJ - Li. (In thefigure,J ~ 6.) Thus, Li
is the horizontal distance from nodej to Jin Figure 5.4.2. Again, shift the events affect
re really positive. ing itemj back in time by Li.
Ie answer, fortu The same idea applies to a tree 'ystem, as illustrated in Figure 5.4.3. The procedure
for drawing the diagram and calculating the time shifts is a bit more intricate in this case,
s for a series sys but the concept is clear from the figure.
(In a series sys This approach does not work for a general system. Figure 5.4.4 illustrates what goes
, a policy for the wrong. There are two paths from I to 5, and their totalleadtimes are different; the dashed
back in time by line indicates the discrepancy. (This discrepancy is analogous to slack time in project
ppose that L ~ ~ scheduling.) We can construct a feasible policy by using the longest path leadtime, ig
= 122. In the real noring arc (I, 3). But then each batch intended for node 3 must sit at node I, accumu
lating cost, for the discrepancy interval. The real holding cost is thus more than the zero
I. Here, each arc leadtime system's. Moreover, it is not clear how to transform a feasible policy for the real
le horizontal dis system into one for the zero-Ieadtime system.
at time t. To de Only under a very special condition, called leadtime balance, do we avoid such dilem
the time axis so mas. In the example of Figure 5.4.4 this means L 13 + L35 = L 14 + L45 . Larger networks
:ation of i on the typically require many more such identities. This condition rarely holds in practice.
nple above, put In truth, we do not yet clearly understand how to coordinate the operations of non
tree systems. Tree systems are fundamentally simpler in this respect.
t is interesting to Next, consider the issue ofquantity units. The set Pre (j) identifies which items are
lIies for the mo- used to make a unit ofj, but not how much. It is possible to transform any tL"ee system,
142 Foundations of[nventmy Management
F,GURE 5.4.2
Distribution system with leadtimes.
W
I
I L'36
W
I I
L'B L'35
W
W
I
L']2
I
W W
I I
L'14
FIGURE 5.4.3
Tree system with leadtimes.
W
I
I L'36
W
I I I
L'35
W
W
I I I I
L'24 L'45
W
I I
L'14
so that every j requires precisely one unit of each i E Pre (j), as in a series system. The
basic idea is similar to the leadtime transfonnation above: Choose itemJ as a reference
point or anchor. Then, follow each (undirected) path leading from J, successively
revising the quantity units ofthe items encountered along the way, until they are all con
sistent with one-for-one usage. Of course, the h j and X. j must be revised accordingly. In
Chapter 5 Several Products and Locations 143
FIGURE 5.4.4
General system with leadtimes.
W
I
I L'36
W
I
L' 13 I
L'35
W
W
I I
L'24 I L'45
W
W
I I
£\4
an assembly system, this means that all items are measured in terms of their usage in
the end product. (A general system cannot be simplified in this manner, again because
of multiple paths between items. The model must include input usages explicitly. Apart
from the extra notation required, however, this is not a serious difficulty like the lead
time-balance issue.)
From now on assume all Lij and L; are zero, and every input requirement is one.
As in Section 5.3, these echelon-level quantities are simpler than local ones. Again,
~(t)declines at a constant rate, now Ai' except at order epochs, when it jumps up by the
order quantity. Also, the total holding-cost rate at time t is again 2,jhi1i(t) ~ 2,i'j~(t).
gj = hJAi
Then, the average cost of policy u is precisely (5.3.1); that is,
C(u) = 2,j [5 + ~
II) _
gjUj ]
FIGURE 5.4.5
Simple distribution system.
\Y
\[J/
'" \7
Chapter 5 Several Producls alld Locations 145
x:al ones. Again, The problem of cboosing the best such policy, analogous to (5.3.2), becomes
jumps up by the
(I) ~ ~AW). Minimize C(u)
subjectto U i = ~UUj
FIGURE 5.4.6
Clusters.
v !( V -v
v v v
o
As in Theorem 5.3.4, these two conditions are necessary and sufficient for the corre
sponding solution to be optimal for (5.4.1).
Here is a useful way to express the optimality condition: Define
N- + k(N:;') +
"'( ( "" N m) ~ (N) - g(N m)
'IT m
"'(N~,. N:) is called the /let capacity of the cut (N;'. N:). A bit of algebra shows that
(5.4.3) is equivalent to "'(N;'. N:) '" O.
There are several ways to detennine an optimal partition. Here is one: It starts with
all items grouped into a single cluster. i.e.• N ~ {N}. and successively cuts it into smaller
ones. using cuts that violate (5.4.3). LetMdenote any possible cluster andB the arcs con
necting its items. so (M, B) is a suhtree of (N, A). The method uses a recursive procedure
that takes any (M, B) as input.
Procedure Tree_Relaxed(M, B)
Else.
For each arc in B, compute "'(M-. M+) for the corresponding cut (Ai. W).
Set"'(* = "'(M;.Mt) ~ min{"'(M-.M+)}.
[f"'(* < O.
Call Procedure Tree_Relaxed(M;. B~)
FIGURE 5.4.7
Expanded nern'ork.
W w
"" :Y
/
w---------- \7
item 1's echelon inventory is the sum of these quantities, i.e., 7 L = Tt2 + 713 = ~(A2UI2
+ A3UI3 ). (Problem 5.18 asks you to work out the details.) Thus, we can write the hold
ing cost for item 1 as 1I(g12uL2 + g13U13), where
panded network. Then, apply the nested-system methodology to this expanded network.
(Unfortunately, the expanded netwolk is not always a tree system; ifnot, its relaxed prob
lem requires a somewhat more complex algorithm than Procedure Tree_Relaxed.)
The end result is a stationary-interval policy for the original system. If the policy
happens to be non-nested however, the order quantities need not be stationary. In the ex
ample above, suppose ui = u t and u; = 2u t. Then, item I's batch sizes alternate be
i
nv-een A2 u + and A2 u + A3 u;. In general, the order quantities are periodic, though the
pattern may be more intricate than simple alternation.
locations could order separately, but it may be advantageous to pool their orders. This
way, especially if the source is far away, we can consolidate the shipments over part of
the distance and thus trim transportation costs. Likewise, we can centralize and sim
plify the administrative tasks of order processing. (And, if the supplier offers quantity
discounts, joint purchasing exploits them better. This sort of scale economy is not the
focus here, however.)
In multiproduct systems, joint replenishment offers similar advantages. When the
products are all supplied by the same source, we can save transportation charges, as
above. In any case it may he simpler to process occasional orders, each including many
products, than frequent single-product orders. In addition, when ordering means pro
duction, we can sometimes pool semp operations.
In such situations the supply process offers economies of scope. Whereas scale
economies enable us to order large quantities cheaply, economies of scope refer to a
large number of diverse activities, in this case the orders for many items.
Here is a model that captures this notion in a simple way: The fixed order cost now
consists of two components. There is an item-specific cost kj , incurred by each order of
item). as before. In addition, there a cost~, incurred on ordering any item or combina
tion of items. For instance, if we order item I alone, the total fixed cost is ko + k j • If we
order items I and 2 together, the total is ko + kj + k2 (not 2ko + k j + ~). In a produc
tion setting. ko is called a major setup cost, and the other '9 minor setup costs. (This
wording suggests that ko is larger than the other '9, but no such assumption is required.)
This model is called the joint-replenishment problem. For the moment assume all lead
times are zero. (We discuss positive leadtimes later on.)
To exploit the economies of scope to the fullest extent, we can choose to order all
the items together. In this case every item uses the same order interval, say u. Define
k ~ ko + "L;~j '9
gj ~ h)\j
_ "J
g - ""J=lgj
Given u, the overall average cost is qu) = klu + Xgu. This is just the cost function of a
single-item EOQ model! Thus, the best order interval u· can be computed in the usual way.
This need not be the best approach, however. There are disadvantages to joint re
plenishment. By forcing all items to confonn to a common order interval, we give up the
flexibility to optimize each one individually. (If ko is tiny compared to the other k;,
clearly, joint ordering makes little sense.) A less rigid approach would be better, one
which recognizes both the advantages and the disadvantages of synchronization.
One reasonable approach is to partition the items into groups. Each group contains
similar items, as measured by the ratios ~;lSj' Then, order each group's items together, as
above, but manage each group independently of the others. This approach retains flexi
bility between groups, while exploiting economies of scope within groups. (It is not that
hard to determine the best partition. Interestingly, the calculation is much like the algo
rithm for solving the DEL model of Chapter 4.)
This approach can miss clear cost-saving opportunities, however. For example, sup
pose that it constructs two groups, and the second group's order interval is precisely 2.01
Chapter 5 Several Products and Locations 151
lIeir orders. This times the first's. Hardly ever do the groups' orders coincide. But, consider the following
ents over part of revised policy: Adjust the order intervals so that the ratio becomes 2, and then alternate
tralize and sim between ordering all items together and just those of the second group. This policy
r offers quantity avoids paying k o to order the first group alone, while only slightly increasing other costs.
lllomy is not the It would be nice to extend the approach to identify and evaluate such opportunities.
As of now, however, it is unclear how to do this systematically.
tages. When the Let us explore a different approach, closer in spirit to those of the last two sections.
tion charges, as This method too partitions the items into groups, each with a common order interval, but
including many it also synchronizes the groups in a controlled manner.
ing means pro- The key idea is to construct a distribution system that is equivalent to the joint
replenishment system. Then, we apply the techniques developed earlier to this distribu
Whereas scale tion system. The resulting policy is thus guaranteed to perform well.
iCope refer to a
i.
5.5.2 An Equivalent Distribution System
lorder cost now
'Y each order of Reconsider for a moment the multilocation system above. The fonnulation of the fixed
mI or combina costs suggests a specific two-step transportation scenario: First, consolidated orders are
is ko+k j • [fwe shipped from the source to some central point at cost k o. From there, each order is
:J. In a produc shipped separately to its designated location. The location-specific costs ~ are incurred
up costs. (This during this second step.
on is required.) Call this central point location O. The overall system is essentially the two-level dis
ssume all lead tribution system shown in Figure 5.5.1. Location 0 is special, however, in that it never
holds inventory. (In practice, such a location is sometimes called a break-bulkpoint or a
ose to order all trans-shipment center.) This requirement forces the distribution system to follow an an
ayu. Define tinested policy, where location 0 orders precisely when one or more original locations
order. Clearly, a multiproduct or a general multi-item system can be interpreted similarly
as a special distribution system of this kind. The original data O'-j, hi' and ~) continue to
describe item j > O. The new item 0 has demand rate "-b ~ 0 and fixed cost ko; since it
never holds inventory, set h o = hb = o.
This antinested system can be analyzed in the same way as a nested system: Con
;t function of a sider the mirror image of Figure 5.5.1 with all arcs reversed but the same ~j and Ai"
1 the usual way. This is an assembly system, and it is equivalent to the original one. The original an
~es to joint re tinestedness condition translates into the usual nestedness property for the assembly
we give up the system. So, just analyze this assembly system as above. This approach determines a
[) the other ~j' policy whose cost is no more than 2% above optimaL If there are positive leadtimes in
be better, one the original system, simply adjust the zero-Ieadtime solution for the distribution sys
ization. tem in the usual way.
~up contains The optimal partition N* happens to be especially simple in this case. Number the
TIS together, as original items so that the ratios '!T(U)) = r,lgj are increasing inj. Then, for some index
1 retains flexi j*, items j ~ j* are clustered with item 0, and each item j > j* forms its own cluster.
I. (It is not that Knowing this, the algorithm for the relaxed problem can be streamlined: Let No denote
I like the algo the large cluster including item O. Starting with all items in one cluster (No ~ N), first
cut off item.!, thenJ - I, and so forth. Keep cutting as long as '!T(NoIU)) < '!T( U)). Stop
example, sup with j* = j the first time this condition fails. (Incidentally, this same procedure can be
precisely 2.0 I used to solve the relaxed problem for any two-stage assembly system.)
152 Foundations afInventory Management
FIGURE 5.5.1
Equivalent distribution system.
9
9/'
'" w
Following the round-off procedure, the final policy has a simple form: Itemsj"; j*
order together, with order interval u + (0). The other items order less often, but always
with the larger group, since their intervals are positive-integer multiples of u + (0).
ecessors) all come from a single source, and a tree system with certain products sharing
setup operations. The same general approach applies even here! (Alas, we haven't room
to cover the details.)
This approach has no simple performance guarantees like the theorems of Sec
tion 5.3. On the other hand, it performs well when the nnmber of locations is large.
(Technically speaking, it is "asymptotically optimal" as J grows.) Also, it has been ex
tensively tested, and it seems to work well.
The same model describes a multiproduct system, where the sctup costs depend on
the production sequence. We explore a similar model in the next section.
~j w{;) 0; wO (5.6.1)
J
The remaining problem is to minimize the aggregate inventory investment, cl. Clearly,
the constraint (5.6.1) is always binding, so we can replace it by the equation
~.w-(~)
'j qj
J
= wO
This problem is the reverse ofproblem (5.2.2), which minimizes wO subject to a con
straint on cI. As discussed in that context, these two problems are essentially equivalent
they generate the sarne tradeoff curve. Moreover, this new problem, like (5.2.2), can be
solved by selecting appropriate aggregate costs K and 'fI, using them to compute item costs
~ and hi' and then solving a separate EOQ model for each item.
The same idea can be applied to supply/demand networks. Consider a series system.
Orders for all stages draw on a common resource, constrained by (5.6.1) as above. Here
is the approach in rough outline: Choose K and 'fI, form the kj and hj accordingly, and
then solve the uncapacitated series system as in Section 5.3. Adjust K and 11, repeating
this process until the constraint (5.6.1) is just barely satisfied.
wsts depend on
c I, (I)
c
~
capacity. Actu
. We begin with ~
(5.6.1)
o
Time (I)
each cycle, for each item}, as in Section 3.4, the machine must produce exactly Xp, so
the run length is precisely P,u, and the average inventory is ~ ~ 15(1 - p)"-;u. The total
production time over all products is pu, and the machine is idle for time (1 - p)u. (This
idle time can be inserted anywhere during the cycle.)
Define
k = l~~l kj
gj = h/l - p)Aj
_ -.:J
g - ~j=l gj
The overall average cost is C(u) ~ klu + Y.gu. Again, compute the optimal cycle time
u* = (2klg)l/2 as in the EOQ model' The end result is markedly similar to the simple
common-cycle approach to the joint-replenishment problem, notwithstanding the radi
cally different characters ofthe two problems.
The rotation-schedule approach certainly is simple, but as in the joint-replenishment
problem, there are disadvantages to imposing a common cycle time on all products. Let
uj ~ (2A/gi 12 betbe optimal cycle time for product} alone, ignoring tbe others. When the
uj are very different, the products' individual economics favor producing some more fre
quently than others.
EXAMPLE 5.6.A. PART 1
j k, Aj ~, hj p, a,
a. u*,
9.61 1.0 5.0 2.500 0.2 2.0 3.1
The cost of this solution is about 36.11, a bit less than that of the rotation schedule above.
5.6.3 Extensions
The setup costs in the basic ELSP are independent ofthe production sequence. Suppose
instead that there is a changeover cost k~; to switch from product i to product}. Also, there
is a cost ko} to start producing} when the machine is idle, and a cost kjo to turn the ma
chine off after producing}. Restrict attention to rotation schedules. The following is a
reasonable heuristic; it is similar to the heuristic for the inventory-routing problem of
These data sug Section 5.5.4: Solve a TSP to find a good product sequence. Calling its cost k, determine
the optimal cycle time u' through the EOQ model above.
~ to a simple Now, consider another system, which combines the features of the ELSP and a se
y simple heuris ries system: There are several products and several production stages. The stages are
arranged in a fixed sequence, and every product requires processing at each ofthe stages
in that order. At each stage there is one machine, capable of processing only one prod
me the uct at a time.
s own cluster, In essence, this system is a sequence of ELSPs in series, where the outputs of one
ben, ROF; is stage become inputs to the next. Put another way, there is an item corresponding to each
product-stage pair. Leaving aside the machine capacities for the moment, the supply
e. Computek demand relationships among these items form several separate series systems, one for
ld u*. each product. The shared machines link these series systems together.
158 Foundations of/memaY\! Management
Thus, this model adds realistic capacity constraints to the series systems of Sec
tion 5.3. Research on this type of model has only just begun. Preliminary results suggest
that a simple heuristic, related to the one above for the ELSP, works quite well.
Consider a supply-demand network like that of Section 5.4, where time is now discrete,
Section 4.2, it
x;(O) = x;o i = 1, ... ,J (5.7.1)
Jrder intervals
Dynamics:
·erage demand
Construct the network ofFigure 4.3.1 using these arc costs. Then, for every pair of nodes
(s, I) withs < I, determine the minimum-cost path froms to I; the cost ofthis path is pre
~TXt) (5.7.4')
cisely 'i(s, I). (One way to do this is to apply Algorithm Forward_DEL for each starting
time s separately. Actually, there are quicker ways.)
i. Can One use this For j ~ 1 we need only compute VI (0, T), the optimal total cost. The nodes on the
Ie answer is no. optimal path are the optimal stage-I order times. Now, take each pair (s, I) of successive
lology of mixed order times, and recover the minimwn-cost path for Vis, t). Those nodes are the opti
o exploit special mal stage-2 order times. Continue working forward in this way to reconstruct the full op
,Ip. Still, the cur timal solution.
The key to this approach is the echelon-cost-accounting scheme: At stage j, choos
lpOrtant one, the ing arc (t, 11) means ordering at time I to COver demand in periods t through u - 1. The
.ch is especially original arc cost ~[I, u) includes the corresponding order and echelonj holding costs;
ODS 5.3 and 5.4. these costs are incurred regardless of what happens at stage j + I and beyond. Now,
tion N* for the Vj + 1(t, u) represents the best way to manage those downstream stages during the same
forces the items time interval, using a nested policy. The augmented arc cost k7[f, u) thus correctly mea
;rep reduces the sures the total cost of choosing arc (I, u).
nstead of items.
:anged in series. 5.7.2.3 Leadtimes
i series system. Returning to a general system, suppose there is a leadtime Lij to transfer item i to j IE Suc
J; his numerical (i). Each Lij is a positive integer. To create item} at time f requires pulling item-i stock
at time t - Lij, for all i E Pre (j). Also, there is an order leadtime L j for each start item
mentioned near j. The original formulation above has Lij ~ I and L;
~ 1.
interpreted as a There are several equivalent ways to adapt the model to incorporate leadtimes, along
rl on series sys the lines of Section 4.4.1. Here is one: Interpretz,(I) as the amount of item i arriving (or
interactions be the production quantity compleled) just before time I + I. With this understanding, re
US! coefficients place the dynamics (5.7.2) by
essed in the dy
;e interactions).
x;(t + I) ~ xi(t) + z,(t) - di(t) - ~.suc(,)zj(1 + Lij - I)
his notion. t= O, ... ,T-I,i= 1, ... ,J (5.7.2')
Also, add appropriate initial conditions to (5.7.1). For example, consider a two-stage
,. O(JT') time.
series system with L: ~ 3 and L 1, ~ 2. The zl(1) for I < L; - I ~ 2 and the z,(I) for
t < L 12 - I = 1 are determined before time t = 0, so they become constants, speci
;eries systems,
factors satisfy fied by initial conditions of the form
multiplies Zj(t) for every t in the objective, but the total item"; orders l,=/t) is the minimal
feasible quantity, and so itselfis a constant. Thus, the extra term in Cj(t) does not affect the
solution, and hence can be omitted.)
There is another, simpler approach for tree systems: Just shift the time axis for each
item, as in Section 5.4. In the two-stage series system above, for instance, interpret xz(t)
and z2(t) as before, but reinterpretxi(t) as inventory at time t - L,'z + I and ZI (t) as pro
duction of item 1 completed just before time t - L,'z + 2. Under this relabeling scheme,
the original dynamics (5.7.2) remains valid. [As in Section 5.4, this idea does /lot work
for a general system, unless the Ieadtimes happen to be balanced; the general case re
quires (5.7.2') in place of (5.7.2), as above.]
Otherwise, borrow the notation of model (5.7.1) to (5.7.4) above. (Omit the primes; e.g.,
write xl(t) for x;(t). Also, D,[t, 1) means the cumulative demand from t onward of item i
only.) Here is the linear mixed-integer programming fonnulation:
Initial conditions:
x,(O) ~ XiO j = 1, ... ,J (5.7.6)
Dynamics:
x,(t + I) = xl(t) + z,(t) - dl(t) t ~ 0, ... , T - 1, i = 1, .. . ,J (5.7.7)
Constraints:
xl(t) 2" 0
=I(t) 2" 0
VI(t)E {O, I}
ZI(t):O; D,[t, 1)v,(t) t = 0, , T- 1, i ~ 1, ... ,J (5.7.8)
l,alz,(t) :0; z+(t) t = 0, ,T- I (5.7.9)
Objective:
Minimize II l;:J [kl(t)vl(t) + cl(t)=,(t)] + li l;~l hi(t)x,(t) (5.7.10)
(t) is the minllnal costs do not reflect continuing production runs, as discussed in Section 4.4.5. It is pos
oes not affect the sible to revise the formulation along the lines suggested there.)
To solve this model too requires integer-programming techniques. The successful
me axis for each approaches to date all exploit the model's special structure: It almost consists of a sepa
:e, interpretx;(t) rate DEL model for each item. Only the capacity constraints (5.7.9) link them together.
andzr<t) as pro Nevertheless, it remains quite difficult to compute the true optimal solution.
abeling scheme, It is straightforward to extend the model to incorporate several limited resources
a does not work (equipment and labor, for instance): Let r = I, ... ,R index the resources. There are now
general case re resource-capacity and resource-usage parameters z+r(t) and air for each resource. The
constraints (5.7.9) now expand to become
l, a"z,(t) <; z+,(t) t = 0, ... , T - 1, r = 1, ... ,R (5.7.11)
One could add capacity limits like (5.7.9) or even (5.7.11) to the network model,
cility of limited (5.7.1) to (5.7.4). The resulting model combines the features of both earlier models. The
sentially that of multiple-resource model has an interesting interpretation: Suppose r indexes distinct
machines. Each item ntilizes only one machine, but several items may share a machine.
So, for each i, there is only one positive air; all the others are zero. Now, were it not for
the supply-demand relationships embodied in the dynamics (5.7.2), the model would
separate into R independent problems, one for each machine. These relationships now
link the machines as well as the items.
be primes; e.g., Tbis combined model is still more difficult to solve than the earlier ones. Nevertheless,
nward of item i it is worth seeing that even such intricate systems can be represented in the language of
mixed-integer linear programming. We learn something about how things work in the act of
formulation itself. When J, R, and T are small, one can indeed obtain the solution-with a
powerful computer and some patience. Otherwise, one must use a heuristic method.
One heuristic approach for this model (and other intricate, large-scale problems)
(5.76) is called hierarchical planning, which determines a solution in two steps. First, it
solves a relatively simple approximate model, called the aggregate model. This model
suppresses the fixed costs ki(t) and aggregates the items into groups. The second step,
.. ,J (5.7.7) called disaggregation, translates the aggregate model's solution into a solution of the
original model. It solves a separate model for each group, which divides the group's
production quantities among its original items.
It is worth mentioning that all the difficulties here are caused by the fixed-cost
terms. Ifthe I<:,(t) = 0, the model becomes a linear program (LP); it is a large and intri
cate Lp, but even those are easy to solve. (Hierarchical planning exploits this fact in its
aggregate model.) Such models are used widely to plan production and distribution. If
(5.7.8) the ki(t) are positive but small, moreover, the model is easier to work with than it would
be otherwise; exact methods work faster, and heuristics perform better. So, reducing the
(5.7.9) fixed costs simplifies analysis, in addition to its direct benefits.
specific set of procedures underlying it. MRP is the de facto standard approach to pro
duction planning and control in the United States, and it is widely used elsewhere too;
the number of companies using it runs easily into the hundreds of thousands.
Our aim in this section is not to give a comprehensive account of:MRP. That would
be impossible. MRP is virtually a managerial subculture with its own extensive vocabu
lary. A small but thriving industry provides MRP software and consulting services; a re
cent directory (APleS [1994]) lists 107 vendors. Rather, we offer a brief overview and
assessment of MRP, focusing on its basic logic and its relation to other approaches.
Initial conditions:
x;(O) = x;o i = 1, ... , J (5.7.1)
Dynamics:
x;(t + 1) CC x;(t) + z,(I) - d;(t) -lj<s",(,)Zj (I + Lij - I)
t = 0, ... , T - 1, i = 1, ... , J (5.7.2')
Constraints:
Jproach to pro Forward_DEL, but others use a heuristic, such as the Silver-Meal heuristic.) Letz~(t) de
elsewhere too; note the resulting order quantities.
mds. Next, MRP focuses on itemJ - I. Now, itemJ - I too may have no successors; in
RP. That would this case MRP treats it just like item J Otherwise, its only successor is J Thus, with the
ensive vocabu zJ(t) fixed atz~(t), its dynamics become
~ services; a re
to constants,
5.8.3 The Larger COiltext
"method for MRP is a great deal more than its core model and the heuristic. It has to be; the environment
1 solution for it operates in is much more complex than the core model. The issues discussed in Sec
Ig upstream. tion 4.3.7 are all relevant here: Typically, MRP is used in a rolling-horizon scenario; the data
::jwhenever of the core model are rough estimates which change often; nervousness is problematic.
:alorder. Like most model-based approaches, MRP includes outer layers as well. Viewed
nation in the from inside, from the model's viewpoint, these outer layers collect and process the data
ms, but only it requires, and send the solution where it needs to go. From an external managerial view
l, is a single point, however, the outer layers are integral parts of the overall approach.
fo solve this For example, consider the d;( t), called demands above. In MRP these are not raw
; Algorithm demand forecasts. Rather, there are one or more layers of analysis between forecasting
166 Foundations o/1nventory Management
and the model. The resulting d;(t) are output targets, collectively called the master pro
duction schedule (MPS). There seems to be no uniform methodology for master pro
duction scheduling; this activity employs a variety of techniques. some fannal, others
informal. Still, the MPS embodies several important functions. Notably, it must plan for
capacity limits, which do not appear in the core model (a point revisited below).
Another crucial layer is buffering. The core model includes no uncertainties in de
mand and supply, but such uncertainties do exist. Buffering refers to adjustments of the
core model's data to account for uncertainties. There are two main buffering techniques,
safety stock and safety time: Safety stock in this context means revising the lower
bounds on the x;(t) in (5.7.3) from 0 to positive values Xi-(t), as discussed in Sec
tion 4.3.7. As mentioned there, it is hard to choose appropriate safety stocks even for the
DEL model, and it is still harder here. Safety time means revising the L; and Lij. upward.
This too is hard, as explained below.
Still further out from the core, many MRP systems provide support for other activ
ities besides production scheduling. Indeed, it is common to find all the finn's control
and data-processing functions, including procurement, accounts payable and payroll,
built around MRP
Partly for this reason, some MRP systems, especially older ones, are slow and clumsy
to work with. Other neVt'cr systems dispense with these extras, concentrating instead on ef
ficient implementation of the core-model heuristic above. This approach is called rapid
MRP, and indeed it is much faster than the traditional, full-featured MRP systems.
Enterprise resource planning (ERP) extends MRP's logic to the entire finn, encom
passing multiple production and distribution facilities. Some versions are even capable of
linking the activities of several finns, using modern communications teclmologies.
the master pro excessive inventories. Even so, occasionally there is more congestion than the Li and L;j
for master pro account for, so the approach falis to meet the output targets of the MPS. (It is possible
: formal, others to augment the leadtimes intelligently, recognizing the potential costs involved. The trou
it must plan for ble is that few users have the time and expertise required to exercise such care.)
below). Many MRP programs include a module called capacity requirements planning (CRP).
::-rtainties in de CRP is a reporting function. After the MRP's heuristic is run, CRP computes and reports
ostments of the the solution's usage of key resources, higWighting those whose capacities are violated. At
ing techniques, that point, some manual intervention is necessary, e.g., fix parts of the solution =f(t) to
sing the lower force feasibility, or adjust the MPS (the d;(t)). Then, rerun the heuristic. Of course, there is
cussed in Sec no guarantee that the revised solution is feasible; several iterations may be necessary.
oks even for the Newer MRP programs provide support for this iterative CRP process, along with
rndLij upward. other enhancements. So different are these programs from the earliest MRP software that
a new name has been coined to describe them, MRP II. The abbreviation MRP here now
for other activ stands for manufacturing resource planning.
lmn's control Thus, MRP II helps managers work around this limitation ofthe core model, but does
'e and payroll, not directly fix it. A totally different approach has gained favor recently. Called
finite-capacity scheduling, it focuses on the resources and their capacities themselves.
[}W and clumsy viewing the demands as jobs to be scheduled on those resources. Most such methods,
~ instead on ef however, suppress or ignore some of the detail of MRP's core model, specifically some
is called rapid ofthe supply-demand relations and/or later time periods. (Still better, of COUl~e, would be
..s tems. an integrated approach, based on a core model with explicit capacity limits and the full
~ lInn, enCOffi
network structure. As indicated in Section 5.7.3, however, while we know how to formu
I'en capable of late such models, we are just beginning to learn how to extract useful results from them.)
ologies. One finite scheduling technique is called optimized production technology (OPT).
OPT's heuristic identifies one or a few bottlenecks, resources whose capacity limits are
most severely strained. Then, to construct a solution, it aims to utilize the bottleneck
resources efficiently. Unfortunately, the actual workings ofthis method remain shrouded
rs a curse. (In in mystery; they are proprietary secrets. (Hyperbole is not the exclusive property of
~ative views MRP; plenty of it surrounds OPT as well.) No one has seriously tested it, to our knowl
)riefly review edge, so at this point we must regard OPT as an intriguing but unproved concept.
Ifview: MRP's approach to uncertainties, as explained above, is a core model that ignores
ained in Sec them, supplemented by a buffering layer to set safety stocks and times. It is not clear how
as networks. well this approach works. (Chapters 8 and 9 examine networks with explicitly uncertain
'roach to em demands and supplies. The approaches that work effectively in that context seem radi
ided the first cally different from MRP's. This fact does not settle the issue, however. Those models
:x supply/de are special in several ways; for instance, most focus on series systems or other special
structures with no fixed costs.)
re model. Of Another weakness stems from the basic philosophy of central control underlying
Nell be illlea MRP Under MRP, like any fully centralized approach (including OPT), all relevant data
'veral altema- flow into one single point, where all key decisions are made: these control directives then
flow out to be implemented at various points in the network. In principle, this is an ideal
Lij to include situation, provided everything works as planned.
i often worse In practice, of course, things do not always work as plaImed: Errors creep into the
, sort. The re files containing current inventory data; machine breakdowns are not reported promptly;
IS generating people misunderstand scheduling decisions; production lots turn out defective; et cetera,
168 Foundafiom' a/Inventory Management
et cetera. MRP is a demanding and fairly rigid planning discipline, and it does not easily
forgive such errors.
Also, as a managerial approach, centralized control is awkward: Generally, it entails
substantial overhead costs (in the form ofa substantial bureaucracy). People on the shop
floor and in the warehouse often experience MRP as a distant, arbitrary master. They
find little opportunity or motivation to take initiative, either to facilitate MRP itself (e.g.,
by reporting problematic conditions), or to improve the actual physical processes.
From this perspective it is easy to appreciate the appeal of a quite different philos
ophy, the just-in-time (ITT) approach. One central tenet of ITT is to decentralize control
as much as possible. Here, demand pulls stock from downstream stages, which in turn
pull stock from upstream stages, with minimal central direction; in contrast. MRP
pushes stock through the system via the MPS and the core-model heuristic. Partly in this
way, but also in others, flT encourages process improvement, through small-scale ini
tiatives as well as large-scale projects. (The ITT approach is discussed further in Sec
tion 88. Caveat. ITT's hyperbole exceeds even MRP's and OPT's.)
Different as their underlying philosophies are, it is nonetheless possible to integrate
MRP and JlT, or at least parts ofthem. Certainly, there is nothing to stop a company from
encouraging process improvements while adhering to MRP Even at the level of control,
MRP can be used to set longer-term output targets, leaving day-to-day or hour-to-hour
control to HI. Alternatively, one can relax central control within certain groups of re
lated stages, but centrally control the flow of goods between groups. These by no means
exhaust the possibilities: there are many others.
This brings us to a final and perhaps decisive issue-adaptability. On this score,
MRP has a mixed record. The MRP industry has been less than eager to listen to ideas
from elsewhere in the operations-management community. This was especially true in
earlier years, but it remains so to some extent. One well-informed observer (Wagner
[1993]) writes, "MRP has become a stern gatekeeper that guards the plant floor from in
cursions by operations research." In my view, this insularity has cost the MRP commu
nity dearly. Had it been more open to alternative formulations, solution techniques, and
control mechanisms, MRP would surely be a more capable discipline today, and fewer
people would regard it as a rigid monolith.
On the other hand, MRP has shown considerable adaptability in certain key directions.
In extending the approach to MRP II and ERP, and incorporating elements of ITT, the in
dustry has responded creatively to real needs of its customers. Based on that experience,
some of its practitioners envision MRP becoming a flexible approach. capable of digest
ing a variety of new methods and ideas. That would be a welcome development indeed.
Notes
Section 5.2: The importance of the inventory-workload tradeoff curve was first recog
nized by Starr and Miller [1962]. See also Gardner and Dannenbring [1979]. De Groote
[1994bJ provides a detailed exploration and interpretation of the variety index.
Chapter 5 Several Produ.cts and Locations 169
does not easily Sections 5.3-5.5.' The overall approach here is attributable to Maxwell and Muckstadt
orally, it entails Schwarz and Schrage [1975], and Jackson et a!. [1985].) Extensions and refinements can
pie on the shop be found in Mitchell [1987], Federgruen and Zheng [1992a,1995], Federgruen,
y master. They Queyranne, andZheng [1992], and Atkins and Sun [1995]. Atkins [1990] and Muckstadt
rocesses. Anily and Federgruen [1990] analyze the inventory-routing problem rigorously and pro
lfferent philos vide extensive numerical results. For recent results and a guide to the literature, see
ltralize control Bramel and Simchi-Levi [1995] and Herer and Roundy [1997]. See also Burns et a!.
:ontrast. MRP Section 5.6: The primary early work on the ELSP was done by Hanssmann [1962] and
:. Partly in this Maxwell [1964]. Ehnaghraby [1978] reviews the research on tbe problem through the
mall-scale ini late 1970s. Significant recent developments include Dobson [1987,1992], Jones and In
Urther in Sec man [1989], Roundy [1989], Gallego [1990], and Zipkin [1991b]. The ELSP can be ex
tended to incorporate planned backorders; see Gallego and Roundy [1992]. See Dobson
~le to integrate and Yano [1990], El-Najdawi [1992], and El-Najdawi and Kleindorfer [1993] for exten
Problems
5.1 Verify that the variety index J. satisfies (5.2.1); i.e., I <; J. <; J. (Hint: Define xj = Wjc)'-j' Think of
the X:i as variables in an appropriate optimization problem,)
5.2 Carry out the steps outlined in the text to solve the optimization problem (5.2.2). Use the solution to
verifY equation (5.2.4). describing the inventory-workload tradeoff curve.
5.3 Consider an independent-item system with a very large number J of items. It may be difficult to
obtain all the individual-item data. This problem suggests a statistical approach to estimating J•.
First, it is often true that the Wj are similar across}. For simplicity, assume that they are equal
(to w). Second, a useful empirical relationship has been observed in many multi-item inventories:
The frequency distribution of the expenditure rates l.:jAj over the items is approximately lognormal.
Y
(If Y is a random variable having the normal distribution with mean iL and variance (]'2, then X = e
has a lognormal distribution with these same two parameters.) Assume that J is so large, the items
effectively fonn a continuwn, and the fraction of items whose expenditure rates lie in the small
interval (x, x + dx) is/ex) dx, where/ex) is the lognormal probability density function (pdf) with
parameters J.1 and 0'2. Also, assume we have estimates of!J- and O'~.
For any positive number E, E[X€J = exp (E!J- + YzE 2 (f2). For instance, in the notation of
Section 5.2, CAll ~ E[X] = exp (fJ- + "<T 2 ). Show that, under these assumptions, J. ~ J. oxp
(_~~O'2), Explain why it makes sense intuitively that J. decreases in (f2.
5.4 Consider the multi-item inventory of Section 5.2, but now suppose that platmed backorders are
allowed. Define a third aggregate perfonnance measure,
= !jc}3j
where Bj is the average backorders for item j. (Actually, a more relevant measure is pB, defined by
using the sales price p) instead of cj to weight Rj . This case is more complex. Assume for now that
the markup ratio p;lc} is the same for all items, so cB expresses the same information as pB.)
Show that the three measures are connected by the following functional relationship:
I
2:(wcA)J.
wO~-------
(vd+ VeB)'
This equation describes a two-dimensional surface in the three-space with coordinates (cl, cB.
wO), analogous to the tradeoff curve of Section 5.2. It is sometimes called the aggregate tradeoff
surface.
Now, drop the assumption above, that the markup ratio pic} is constant, and use pB instead of
cB. There is no simple fonnula for the tradeoff surface in this case, but one can still generate the
surface numerically: Let ~ denote the cost factor for aggregate backorders, analogous to K and 'f),
and w ~ [3/([3 + '1). Derive formulas for cI, cB, and \VO in terms of the cost ratios '1/K and w.
Describe how to use these formulas to calculate points on the tradeoff surface.
5.5 For a series system with any number of stages J, show that a nested policy dominates a non-nested
one, as in Theorem 5.3.1.
Chapter 5 Several Products and Locations 171
5.6 For a series system with any number of stages J, show that a zero-inventory policy dominates any
other, as in Theorem 5.3.2.
5.7 Given the solution u* to problem (5.3.3), define the subset A: (j. A : u; ~ u;+Il. Prove that u*
00
'je/Ai" Think of is the unique solution to problem (5.3.3[A:]). (Use a direct argument by contradiction. Assmning
the contrary, construct a feasible solution to (5.3.3) having lower cost than u*.)
~ the solution to 5.8 To construct a policy according to Section 5.3.6 entails minimizing the function C+ (u) over the
interval [I, 2). The text outlines the approach in general terms. This problem fills in the details.
difficult to The proof of Theorem 5.3.7 (Section 5.3.8) shows how to compute the breakpoints, denoted
timating J•. x(m), and provides a fonnula for 11 (m) as a function ofg. Using this infonnation, and notation of
-j-
ley are equal your invention, write an explicit expression for C+(u). The expression should clearly look like the
inventories: cost function of an EOQ model over each of several subintervals. (For simplicity, assume the x(m)
'Iy loguonnal. are all distinct and strictly larger than I.) Explain how to optimize C+ (g) over each subinterval.
.1, thenX= e Y 5.9 Fill in some of the details of the proof of Theorem 5.3.7: Show that x(m) is indeed the breakpoint
ge, the items of u +(m), and verify the fonnula for u +(m) as a function of u. Show that the sum of the two
the small
integrals in braces is lIV12, as asserted.
I (pdf) with
5.10 In the proof of Theorem 5.3.6 it is shown that
iauof
2-].'2:s u+(m) < 2+ 1/2
oJ. exp
u*(m)
.:orders are This is true, evidently, for all m and any value ofA. Using this fact, argue that, ifA should change from
one value to another, the relative order frequency u+(m)/u +(next (m)) changes by at most a factor of2.
5.11 Consider a series system whose cost factors are specified by kj = K~j' ~j = Tj~i' The optimal
partition N* for the relaxed problem is then independent of K and Tj. Suppose we use the relaxed
problem and N* to estimate wO and cf Argue that the inventory-workload tradeoff curve is
precisely (5.2.4), the same fonn as in an independent-item model. with aggregate system
parameters redefined as in Section 5.3.7.
B. defined by
5.12 Consider a series system like that of Section 5.3, except that a fixed fraction OJ of item) is
for now that
defective. AU defects are discovered immediately and must be scrapped. There is no reimbursement
as pH)
for defective items. (0 1 is the defect fraction for supplies arriving from the outside source. For) >
p:
1, stage} converts one unit of item} - I into (I - 0;) units of usable item} and OJ nnits of scrap.)
Construct an equivalent model with no defects.
5.13 Consider a series system where each stage has a finite production rate IJ.j' as in Section 3.4. Stage
I produces item I gradually, and stage} + I gradually converts item} to item/ + I. Assume that
each increment of item J becomes available to meet demand the instant it is produced. Likewise,
s(cI, cB, item:i production can be used instantaneously in stage} + 1. There are no leadtimes or setup times.
'ale tradeof], So, in principle, a unit can pass through all stages instantaneously.
Assume J.Ll ~ J.L2 ~ ... ~ IJ.J> A., and define Pi = A./IJ.j' Consider a nested, zero-inventory,
,B instead of stationary-interval policy. Argue that any such policy is feasible. Explain that the echelon inventory
merate the 1;(1) follows the same periodic pattern as in a single-item model for each). Then, show that the
to K and Tj, average echelon inventory is ~ )1,(1 - p)u)... Finally, explain how to redefine g; so that the
00
and UJ. function C(u) correctly measures the average cost. (So, from this point on, we can apply the
approach of Section 5.3 as is to find a good policy of this type.)
I non-nested 5.14 As in the previous problem, consider a series system with finite production rates J.Lj' Now, however.
finished goods become available to meet demand only in whole batches, as in Section 3.4.5.
172 Foundations of Inventory Management
Likewise, stock produced during a production run at stage} < J can be used at stage} + I only
when the whole run is complete. After that, the production process at} + I gradually reduces !j(I)
as it increases I;~I(I). Again set Pi ~ IJI'-j' and assume 0 < Pj < I for all). (The f1j need not be
monotonic, however.)
In this context, a policy is nested if stage} + I begins a production run each time stage}
completes one. A zero-inventory policy is one where every production TIm at every stage begins at
the last possible moment consistent with feasibility. The stationary-interval property has the same
meaning as before. Consider only nested, zero-inventory, stationary-interval policies.
(a) Consider a concrete example withJ = 2, A = I, 1'-1 ~ 1'-2 ~ 4, ul ~ 8, and U2 ~ 4. Graph the
1;(1) for 0 "" I "" 16, assuming a production run of item I starts at time 0, and setting li(O) = 0
and 1,(0) = 3. Then, graph the echelon inventories ~(I) over the same interval.
(b) Now, consider the general case: For item J, as in a single-stage model, within each cycle (of
length 11.1,) the production run lasts for time PftlJ. Consequently, the production period must
start with enough inventory to meet demand during this interval, namely PJUJ,.A. Argue that, for
similar reasons, in general, a production run for item j must start with echelon inventory
I/t) ~ (li"j PiUi)A. This is the smallest echelon inventory within each cycle. Then, argue that
the largest is this quantity, plus U;A. Finally, show that the average echelon inventory is l; =
:1(1 + pJlUjA + (li>j PiUi)A. .
(e) Set h} ~ (1 + p)hj + 2pj1i<jhi and redefine g; ~ h }A. Argue that C(u) now correctly
measures the average cost of a policy. (So, we can apply the approach of Section 5.3 to find a
good policy.)
5.15 Consider a series system with the following data: J ~ 4, A = 10, all '9 = 80, and
Ll~ = 6 L24 = 3
L3'5 = 5 Lis = 3
L3'6 = 6
Given a policy for the zero-leadtime system, explain precisely how to construct an equivalent
policy for the real system.
5.17 Consider the distribution system of Figure 5.4.2. Suppose it describes a production process.
Originally, all items are measured in kilograms (kg). To make I kg of end item 6 requires 4 kg of
item 3. Item 5 requires I kg of item 3; I kg of item 3, in turn, requires 2 kg of item 1. Item 4 uses
2 kg of item I, and item 2 uses 0.25 kg of item 1. Revise the items' quantity units, so that each item
uses precisely one Wlit of its (single) predecessor. (Continue to measure item I, the raw material,
in kg.) lfthe original demand rates are
Igej+lonly 5.18 Consider the tluee-item non-nested distribution system discussed in Section 5.4.7. Provide a
lily reduces Ij(t) detailed argument to verify the formulas given there for the average echelon inventories of the
Lj need not be artificial items Ij and of item I itself.
5.19 The basic version of the ELSP in Section 5.6 assumes that stock becomes available to meet
me stagej
demand the instant it is produced. Show how to revise the formulation in the opposite case, where
~ stage begins at
only full, completed batches can be used to meet demand. (Assume a rotation schedule.)
ty has the same
es. 5.20 In the ELSP (assuming output is instantaneously available to meet demand), suppose each product
requires a setup time 'T; in addition to the setup cost kj" Assume a rotation schedule. Define the total
= 4. Graph the
setup time during a cycle T ~ ~ Tj' Using this quantity, show how to adapt the methods of
;etting II (0) = 0
Section 3.4 to determine the optimal cycle length u*
5.21 Consider a series system with a slow-changing demand rate A(t), as in Section 5.7.1. Suggest an
ach cycle (of
period must approximation ofthe total average cost, analogous to (4.2.2). (Hint: Compute something like
C- for each point in time.) Provide the best argument you can to support this approximation.
L Argue that, for
inventory 5.22 Consider the optimization model (5.7.1) to (5.7.4) representing a general system in discrete time.
hen, argue that That model presumes one-for-one input/output usage. Suppose instead that an output unit of item j
rtlory is ~ = requires precisely aij units of item i. for all i € Pre (j). Show how to modify the formulation. (Hint:
Only one change is necessary.)
mectly 5.23 Consider problem (5.7.1) to (5.7.4). Let d,(t), x,(t), and hi(t) denote the echelon-i demand,
"5.3 to find a inventory, and holding cost, respectively, at time t. Give expressions for these quantities in terms of
local data and variables. Write down an equivalent formulation using these echelon-level
quantities. [Notice, the dynamics become simpler than (5.7.2), but the constraints become more
complex than (5.7.3).]
5.24 Consider problem (5.7.1) to (5.7.4) for the special case of a series system. Assume that hi(t) is
"M = 2. What is nondecreasing inj for each t, and k/tJ and cj(t) are nonincreasing in t for eachj. Argue (by
contradiction, as in the proof ofTheorem 5.3.1) that the optimal solution is nested.
Actually, your argument should require only the following weaker condition: For all t < u and
all j, define
lO,
c H A p T E R
6 STOCHASTIC DEMAND:
Outline
6.1 Introduction 175 6.5 Optimization 213
6.2 Policy Evaluation: Poisson Demand 6.6 Lrnnpy Demand 227
178 6.7 Other Extensions 231
6.3 World-Driven Demand 196 Notes 237
6.4 Approximations 205 Problems 238
6.1 Introduction
6.1.1 Major Themes
This chapter addresses some of the truly basic issues of operations management: How
should we control a system whose demand is uncertain? What are the effects of such un
certainties on system performance?
Here, we explore these issues for an otherwise simple system; there is only one item,
the supply process generates constant leadtimes and perfect quality, and demand is sta
tionary. Subsequent chapters study more complex systems with stochastic leadtimes and
imperfect quality (Chapter 7). multiple items (Chapter 8), and time-varying demand
(Chapter 9). Also, this chapter assumes that stockouts are backlogged; Chapter 7 treats
the lost-sales case.
Stochastic demand, in combination with an order leadtime, raises altogether new
difficulties: As a result of the leadtime, there is a delay between the actions we take and
their actual effects. Such lags raise no problems if demand is ce11ain. as we have seen,
for then we know what will happen during the lead time, and we can adjust our current
175
176 Foundations ofIllventOl)l Management
actions to compensate. Likewise, in the extreme (and rare) case of zero leadtirne, even
with random demands, we retain full control of the system. Here, this is no longer true;
we must act in the dark, unable to foresee the ultimate effects of our actions. In particu
lar, no matter what we do, an unexpected surge in demand can exhaust our stock, leav
ing us unable to meet subsequent demands; some degree of stockout risk is inevitable.
Generally, longer leadtimes and greater demand uncertainty both degrade the pre
cision with which we can control the system. And the less precise the control is, the more
inventory we need to serve customers adequately. The primary goal of this chapter is to
probe and elaborate this intuitive insight, to understand precisely how these combined
effects work.
One m,\jor conclusion is that the primary impact of these factors can be captured by
a single summary measure, denoted CT. the standard deviation ofleadtime demand. We
shall explain what this means in due course. For now, think of (T as an index of impreci
sion, a measure of the noise or variation that impedes our efforts to control the system.
(The next chapter shows that 0' also captures the effect of leadtime uncertainty.) Along
with other parameters examined earlier, such as the demand rate and the fixed order cost,
IT is a major determinant of performance. Anything we can do to reduce it-and there
6.1.2 Summary
Here is a brief overview of what lies ahead: Throughout the chapter we model time as
continuous, but demand, inventory, and other quantities as discrete (integer-valued). We
pose several models of demand and supply. In each case we specify a plausible type of
control policy.
The initial goal is performance evaluation, to calculate the key measures of per
formance for any policy in the specified class. Sometimes we obtain exact formulas,
but we also develop useful approximations. We explore some ofthe properties of these
formulas, to learn how performance responds to alternative controls and to changes in
the model parameters. This mode of analysis takes up Sections 6.2 to 6.4 and much of
Sections 6.6 and 6.7.
Later (in Section 6.5 and parts of Sections 6.6 and 6.7), we introduce cost factors.
define an overall cost function, and show how to determine the best policy (within a
given class). Because the models here are so complex, we cannot always express the op
timal solution in closed form. We do obtain such formulas for certain special cases and
approximations. Otherwise, we show how to compute the optimal policy by fairly sim
ple numerical methods. Even without closed-form formulas, moreover! we learn a good
deal about the qualitative behavior of the optimal policy and its performance. It is here
that we see most clearly the impact of the variation index (j.
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 177
ro leadtime, even 6.1.3 Demand Models
is no longer true;
:tions. In particu The basic demand model is the Poisson process, the simplest model of random events
it our stock, leav overtime: Demands occur one unit at a time. In every small time interval, a demand may
isk is inevitable. or may not occur. Each such interval has the same potential to contain a demand, no mat
degrade the pre ter what happens during other intervals. This potential is measured by a positive number
~ntrol is, the more A, the average demand rate. Section 6.2 focuses on this type of demand process.
this chapter is to Section 6.3 introduces a more general model. Again, demands occur one at a time,
, these combined and there is no predictable variation over time. So, there is again a constant demand rate
A. However, there is some exogenous system (conditions in the economy, for example)
"' be captured by whose behavior partly deternoines the demands. The state of this system provides infor
ime demand. We mation about future demands. We call this system the world, and demand a world-driven
ndex of impreci demand process. Because oftheir dependence on the underlying world system, demands
.trol the system. during different time intervals are correlated. (Chapter 9 treats a time-varying world sys
:ertainty.) Along tem, as well as time-varying demand.)
fixed order cost, We study one particular case in detail, where the world is represented by a continuous
ce it-and there time Markov chain. Here, the demand process is called a Markov-chain-driven counting
process, or MCDCprocess. Such a process can be either more or less variable (and thus lead
; virtually never to a higher or lower (T) than a Poisson process, depending on its details.
iated: Too many Section 6.6 considers a model oflumpy demands, a compound-Poisson process. Cus
and zero inven tomers arrive according to a Poisson process, but the amount each customer demands is
nation in a seri random. Such a process generally leads to a larger cr than a Poisson process, because the
x:kout risk, and demand sizes add another source of uncertainty, in addition to the timing of demands.
For all the differences in these models, the end results of their analysis-the
performance fonnulas-nearly all share a common form. Consequently, many impor
tant qualitative properties of the perfonnance measures hold universally. Moreover,
the optimization methods of Section 6.5 apply to a wide variety of systems.
: model time as
~er-valued). We
lausible type of 6.1.4 Policies
What is a reasonable control policy in a stochastic environment? Intuitively, we would like
~ures of per to respond sensibly to demand fluctuations, maintaining sufficient discipline to avoid over
xact fonnulas, reaction. Also, any rule we adopt should be simple, i.e., easy to understand and to imple
lerties of these ment. The policies we explore here meet these requirements, and they are widely used in
I to changes in practice. (Chapter 9 proves that such a policy is optimal among all policies under certain
4 and much of reasonable assumptions. For now, however, we take these policy fonns as given.)
The policies of Chapter 3, assuming detenninistic demand at a constant rate, have
:e cost factors, two pleasing properties: The orders are all of the same size, and they are placed at equal
,!icy (within a intervals of time; in both quantity and timing orders are regular and predictable. With
"'press the op random demands, at least one of these regularities must be foregone. We can still place
~iaI cases and
orders at regular intervals, but then the order quantities will vary. Or we can order equal
by fairly sim amounts, so that the timing of orders becomes unpredictable. Each of these alternatives
e learn a good makes planning more difficult.
!Dce. It is here Here, we focus mainly on the second approach, that is, we base orders on observ
ing the inventory position, not the clock or the calendar. In practice, this means we
178 Foundations ofInventory Management
need an information system capable of tracking the inventory position, which requires
recording demands as they occur. These are sometimes called transaction reporting
systems. (Such systems are fairly common now, because of the advent of electronic
point-of-sale terminals and related communication technologies.) Section 6.7.3 briefly
discusses a periodic-review policy, where orders are placed at regular intervals. Chap
ter 9 treats discrete-time models, in which a regular grid of possible order times is im
posed on the problem from the beginning.
Most of the chapter, assuming demands occur one unit at a time, focuses on the
order-quantitylreorder-point policy, or (r, q) policy, introduced in Section 3.3. An im
portant special case is the base-stock policy, where q = 1.
For the lumpy-demand case (Section 6.6), we consider two plausible extensions of
the (r, q) policies. The first, also called an (r, q) policy, orders integer multiples of q. The
second, called an (I; s) policy, orders enough to raise the inventory position to a fixed tar
get level s. Also, Section 6.7.2 explores a more elaborate policy for world-driven
demand, which bases order decisions on information about the world.
E[·] ~ expectation
V[·] = variance
t = continuous time variable, t 2::: 0
For each t ". 0, let
D(t) = cumulative demand through time r. i.e., demand in the interval (0, t]
~. focuses on the
tion 3.3. An im o
lIe extensions of
llItiples of q. The
on to a fixed tar
br world-driven
L
o
o 5 10 15 20 25 30
Also, D(I, u] has the Poisson distribution with mean A(U - I), which depends on the inter
val (I. u] only through its length, not its location. Finally, D(I) and D(I, u] are independent.
(In technical terms. D has stationary, independent increments. See Section C.2.3.4 and
Section C.5.6 in Appendix C for more about Poisson distributions and processes.)
Figure 6.2.1 illustrates (among other things) some demands generated by a Poisson
process. The demands are quite irregular; in several places, they seem to occur in
bunches, and there are long gaps with no demands at all. This behavior is typical.
A Poisson process is widely used to model demand for several reasons: It is easy to
<val (0, I]
specify; the rate A is its only parameter. In many practical situations. moreover, the
model is fairly accurate; demand really does behave like a Poisson process. (One reason
is that demand often comes from many small, nearly independent sources, e.g., cus
tomers spread over a large region, and a Poisson process approximates such an aggre
the collection gate process reasonably well.) Finally, the mathematical simplicity of this model
smooths the tasks of analysis and calculation.
180 Foundations oflnventory Management
Given a fixed policy. several state variables describe the evolution of the system over
time. For each time t ~ 0,
I(t) = inventory on hand
13 = average backorders
I = average inventory
gI(i)
-
~ limT~ro [(1')S6"
IT
1 {I(t) = ij dt]
I denotes a random variable with this distribution; then, 7 = E[l]. Second, I has a limit
ing distribution. This means that the probability distributions ofthe random variables I(t)
converge to a limit as t ----t co, and this limit does not depend on initial conditions. Let
I = equilibrium inventory, a random variable having the limiting distribution of I
It is relatively easy to determine this distribution. The ergodic property further implies
that the limiting distribution is precisely & above; that is, I has the same distribution as
[. So, 7 = E[I].
Chapter 6 Stochastic Demalld: One Item with Constant Leadtimes 181
)fthe system over The other stochastic processes above are well-behaved in the same senses. In par
ticular, IN is. So. we shall take the following general approach: Detennine the distrihu
tion of
IN ~ equilibriwn net inventory
Use this to detennine the distributions of I and
A = equilibrium stockout indicator
B ~ equilibrium backorders
Then, compute 7 = E[I], A ~ E[A], and B ~ E[B].
To describe Ill/, it turns out, we need the distributions oftwo other random variables.
all t we use bold The first is
c process, as are
t), as above, and IP = equilibrium inventory position
Its distribution is relatively simple. The second random variable is called the leadtime
demand, denoted D. The variation index (J is precisely the standard deviation of D. It is
mainly through D, and especially <J, that tbe demand and supply processes affect the sys
ection 3.3: tem's perfonnance.
'ditions. Let used instead of s; yet another name for a base-stock policy is an (S - I, S) policy.)
ibution of I
6.2.2.2 Analysis
Urther implies Suppose for now that s 2" 0, and IN(O) ~ 1(0) ~ IP(O-) = s. Thus, orders coincide with
futribution as demands. Also, each order remains outstanding for time 1. So, IOtt) (outstanding orders)
Figure 6.2.1 helps to visualize how demands and orders interact. In the lower part
of the figure, each demand is indicated by a vertical line ofheight L. There is a diagonal
line attached to each one with slope - I. So, if there is a demand at I, the diagonal line
meets the time axis at t + L, precisely the moment the corresponding order arrives. Now,
pick an arbitrary time t, not necessarily a demand point, and count the number of diag
anal lines above the time axis at I. The result is just IO(I). The upper part of the figure
graphs lO(I). (This system has A = 1 and L = 3.)
Choose any I > L. Any order placed before I - L must have arrived before I, and so
cannot be included in IO(I); and of course, IO(I) does not include any orders placed after
I. The orders included in lO(l) are precisely those placed in the interval (I _. L, I]. But each
such order corresponds to a demand, so IO(I) is sinlply the demand during (I - L, I], or
lO(l) ~ D(I - L. I]
The net inventory IN(I) is just a translation of IO(I) ~ D(I - L, I]. Also, although we as
sumed IP(O-) = s, (6.2.1) clearly holds regardless of starting conditions for sufficiently
large I (i.e., such that IP(I - L) = s).
Since IO(I) has the same distribution for all large I, certainly 10 has a limiting
distribution, and therefore so does IN. The limiting random variable IO has the Poisson
distribution with mean AL. and
IN=s-lO (6.2.2)
This is the main result. It is convenient to express it in slightly different terms: Call
D(I, I + L] the leadtime demand beginning at I, and let
D ~ leadtime demand, a generic random variable with the Poisson
distribution with mean AL
i.e., the same distribution as IO and D(I, I + L]. (The symbol D stood for the demand
during a leadtime in Chapter 3 also, but there it was the constant AL.) Then,
IN(I + L) = s - D(I, I + L]
IN=s-D (6.2.3)
As shown later (Theorem 6.2.3), this identity also describes the long-run frequency dis
tribution oflN.
We can now compute the key system performance measures: Let g denote the prob
ability mass function (pmf) of D, d' its complementary cumulative distribution (ccdf),
and G L its loss function. (The latter is G1(d) = E[[D - dt], as in Section C.2.2.) From
(6.2.3),
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 183
although we as
; for sufficiently 6.2.2.3 Behavior of Perl'Ormance Measnres
The performance measures above depend on the system parameters A and L only through
Ihas a limiting their product AL, the mean leadtime demand. You might guess that a system with a high
has the Poisson demand rate and a short leadtime would behave quite differently from one with a low de
mand rate and a long leadtime. No: As long as AL is the same for the two systems, their
performance characteristics are identicaL
(6.2.2) The formulas reveal the qualitative effects on performance ofchanges in s: First, E is
ent terms: Call decreasing and convex as a function of s; as the base-stock level increases, backorders de
cline, but at an ever slower rate. Similarly, 1 is increasing and convex in s. As for A, it is
decreasing in s: increasing the base-stock level reduces the stockout probability. (However,
sson A is not generally convex in s. It happens to be convex when AL :-::; 1, but not otherwise.)
Figure 6.2.2 graphs Band 7 as functions ofs for four different values of L with A ~
'Or the demand 10. (Because B and I depend on A and L only through AL. we would obtain the same
en, family of curves were we to fix L at I but increase A to 20, 30, and 40.) The decreasing
curves describe E, and the increasing curves 7. Clearly, all these fimctions are convex.
(Also, for each L, the Band 7 curves cross near s ~ AL, and they seem almost mirror
images of one another; that is, jf we reflect E around the vertical line at S = AL, we ob
(6.2.3)
tain something close to 7.)
frequency dis- Next, compare the perfoffiunce curves for different values of I. As L increases, the
most visible effect is the shifi of both curves to the right. So, as L grows, we must shift
note the prob s accordingly to maintain performance at roughly the same levels. To see the effect of L
bution (ccdf), on the shapes of the curves, consider Figure 6.2.3. Here, we translate all the curves to a
1 C.2.2.) From connnon center, by changing the horizontal axis to s - AL. Viewed in this way, both B
and 7 grow larger as L increases. This effect is strongest near S = AI.
184
Foundations ofInvenrory Management
FIGURE 6.2.2
Peiformance q(base-stock policies.
20
,,
,,
,,
,,
,,
r"
15 ,, 15
,,
,,
,,
~ ,,
'E
~
~
10 ,
\
,,
,,
:'"
..."
! 10 !Ii
, "
\ ,/
\ ,/
5 ,',
"
,, ·L~1
5
,, '" L~2
, -L~3
-L~4
.., ...._:
o o
o 10 20 30 40 50 60
Base-stock level
FIGURE 6.2.3
Performance qfbase-stock policies (centered).
10 10
,y
'l
'/
;/
~(/
"
,,,J"
~ ,,,r/
'E
-'i 5 r/ Ii
N
5 ~
~ ...
//
/ ...... $
/:'
·L~1
.. "- L~2
-L~3
-L~4
.... , .. , ..........
o
",
". o
-10 -5 o 5 10
Base-stock level £[10]
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 185
FIGURE 6.2.4
Stock-service tradeoff.
20
2
L~1
--- L~2
-L~3
15 -L~4
1.5
"
E
.~
~
IO g :&"u ...............
5 !l
"'"
\1
0 ",
"1]
- 5
""""'"
0.5
....................................
o "
o
0.001 0.010 0.100 1.000
Customer waiting time
10
To judge the effect of L on overall pelformance, it is useful to graph the same data
in a different form, as in Figure 6.2.4. Here, we use the rescaled performance measures
BW = average customer backorder waiting time
IW = average stocking time, the time a unit spends in inventory
instead ofB and 7. Section 6.2.4 shows that BW = BIA andlW ~ 7/1... For each s the figure
plots the corresponding values of BWandIW. (The actual values ofs are suppressed.) The
W resulting curves show how much stocking time is necessary to achieve a specified waiting
5 g
time. (The logarithmic scale for waiting time reveals the relationships better.)
5 It is clear from the figure that performance deteriorates as the leadtime grows. To
achieve a given desired waiting time, as L increases, we must increase the stocking time
(by increasing s), or, to maintain a given stocking time, we must accept a longer waiting
time. There is a clear, tangible benefit to reducing the leadtime.
s A I B
The performance target is A .:s 1 - tll2 = 0.2, so the manager should set s = 4. With this policy,
the average inventory will be 2.25.
You might wonder about the various assumptions here. Classical CDs do have roughly Pois
son demands. Many customers, though not all, insist on a particular recording (they will not take,
say, Karajan and the Berlin Philharmonic as a substitute), and are willing to wait a while ifthe item
is not in stock. Popular CDs, in contrast, sometimes have more volatile demands (as discussed in
Section 6.3 below), and stockouts are more likely to result in lost sales (sce Section 7.2).
Another way to resolve the conflict is by cost optimization. We postpone that dis
cussion until Section 6.5.
6.2.3.1 Analysis
Policy evaluation now requires a rather different approach, and more effort, than before.
FIGURE 6.2.5
iible stocking pol Inventory position and net inventory.
'lew York Philhar
Iflth, and the lead
stock.
L ~ (3.6XO.5) = fP(t)
,+q
IN(t)
Time
THEOREM 6.2.3. The limiting distribution of the net inventory is described by the
equation
IN= IP-D
where IP is unifonnly distributed on the integers in the interval [r + I, r + q], D has the
Poisson distribution with mean AL, and the two variables IP and D are independent. This
also describes the stationary and long-run frequency distributions of IN.
Step 4: First, B is precisely the mean of the loug-run frequency distribution of B,
which is in turn a function of IN. It follows thatB = E[B]; similarly, T ~ E[1].
Theorem 6.2.3 tells us that IN is the difference oftwo independent random variables,
IP and D. Condition on IF, say IP ~ s. The distribution of [INIIP = s] is given by (6.2.3).
The (unconditional) distribution of IN is a simple average of these conditional distribu
tions over s = r + 1, ... , r + q. This fact allows us to construct formulas for A, B, and
7using their counterparts (6.2.4) to (6.2.6) for base-stock policies as building blocks. Let
A(s), B(s), and 7(0') denote the performance measures for a base-stock policy, now writ
ten as explicit functions of s. Likewise, let A(I; q), B(I; q), and 7(1; q) denote the per
formance measures for an (,~ q) policy. Then,
A(I; q) ~ Pr {IN"'" O}
Thus, system perfonnance using a policy with q > 1 can be represented as a simple av
erage of the performance of several base-stock policies.
We can write these quantities explicitly in terms of the first- and second-order loss
functions ofD, G 1, and G2 The latter is G2 (d) = 1j>d G1(j) = ~(AL)2 - lO<j";,, G'(j).
From (6.2.8) to (6.2.10) it immediately follows that
B~ (~)[G2(r) - 2
G (r + q)] (6.2.12)
T~ ~(q + 1) + r - U +B (6.2.13)
(Notice the close connection between 1 and B. We observed a similar relation in the case
of detenninistic demand in Chapter 3. The only difference is that here we have a tenn
Chapter 6 Stochastic Demand: One Item with COllstanr Leadtimes 189
described by the iJ(g + I) instead ofiJg. This difference is entirely a result of the demand quantities be
ing discrete, as indicated in Problem 3.8.)
Finally, consider the order frequency OF. Demands occur at rate A, and an order is
placed every g demands, so
r+ g],D has the _ A
ndependem. This OF~
'Ii. g
fistribution of B, as in Chapter 3. We now have expressions for all the performance criteria.
=E[1]. These formulas are correct for all integer values of r, even negative ones. We would
andom variables, never wish to use a policy with r < -q, however, just as we would avoid a base-stock
given by (6.2.3). policy with s < 0: At r = - g, 7 = O. Reducing r below this value leaves 7 = 0 but in
ditional distribu creases backorders. Thus, we can restrict r ~ -q.
~as for A, E, and Incidentally, the functions G t and G2 can be written in terms ofcfl andg:
lding blocks. Let Gt(d) = -(d - AL)cfl(d) + ALg(d) (6.2.14)
>olicy, now writ-
I denote the per
2
G (d) = iJ{[(d - ALf + d]Go(d) - AL(d - AL)g(d)} (6.2.15)
(See Problem 6.6. These expressions exploit the specific form of the Poisson distribu
tion. They are not much easier to compute than the general loss-function definitions
above. However, they give better numerical accuracy for large d, given accurate calcula
tions of GO(d) and g(d).)
(6.2.13)
6.2.3.3 Selecting a Policy to Meet a Service Constraint
Again, there are conflicts between the performance measures. Suppose that q is fixed.
:ion in the case Given the constraint A .: : :; I - w_, it is clear what to do: Set r large enough to meet the
FIGURE 6.2.6
Pe1fonnance of (I, q) policies.
10 10
.q ~ 1
--- q ~ 5
-q~9
_q~13
.a~
w
u
OJ
5
,~-,
{!'
,;«''"
5
!
,,?-,//...
/F'
The record store above must order CDs in batches of size q = 2, for some reason. The data remain
A = 3.6 and L = 0.5, and the seIVice requirement remains .4::s; 1 w_ = 0.2. The performance
of certain (r. q) policies is as follows:
r .4 7
-1 092 0.09
0 0.69 0.40
0.41 1.00
2 0.19 1.81
3 0.08 2.73
4 0.03 3.71
So, the manager should set r = 2. This policy has 1= 1.81. (It is odd that this I is less than the
base-stock policy's in Part 1 above. This is due entirely to the fact that integer policy variables can
not always hit a performance target exactly. Recall that the earlier policy with s = 4 has II = 0.11,
much lower than the required 0.2, while the (r, q) policy here just meets the requirement.)
Chapter 6 Stochastic Demand: Dlle Item with Constant Leadtimes 191
FIGURE 6.2.7
Customer waiting times.
10 10
7
H
5
if
! i
"
6 I I
-g 5
il
o4
3
o
o
Time
intuitively clear that R(I) should not grow systematically as I increases, specifically, that
lim,~~ R(I)l1 ~ O. Therefore,
- . (I)f'
B ~ lImH~t 0 B(u) du
='
hmH~
({P(I)l
I J . {B~}
D(I)
_{R(I)}\
I)
D(I)}.
= limH~ { -1- . lImHoo
{~}
D(I)
~c ABW
(This result is analogous to Little's [1961] formula in the context of queues.) A similar
argument shows that the average stocking time is related to the average inventory by I =
VW.
By the way. this argument is valid for any processes where the averages B, A, and
IW exist. It does not use the assumptions of Poisson demand and constant leadtimes. The
result holds for quite general demand and supply processes, induding those of Sec
tion 6.3 and Chapter 7.
Now, BW directly measures service to customers. The fact that it is proportional to
B is the primary reason that B is important. Also, BW inherits the qualitative properties
ofB discussed above; for example, increasing s (or r) reduces the average customer wait
ing time.
EXAMPLE 6.2.A, PART 3
Reconsider the record store above (with A = 3.6). Under the base-stock policy with s = 4,11 =
0.05, so BW = 0.05/3.6 = 0.014 months, or abont 0.4 days. Under the (r. q) policy with q = r =
2, B ~ 0.11, so BW = 0.11/3.6 ~ 0.029 months, or 0.9 days.
The following theorem is a sharper result. (It does assume Poisson demand and con
stant leadtimes. Actually, the result holds also for some of the more complex supply sys
tems ofChapter 7, but not all; see Section 7.3.8.) LetBW denote the random variable de
scribing the limiting distribution of waiting times, so BW = E[BW]. Also, let D(BW)
indicate demand during the random time BJv, where BW is independent ofD.
THEOREM 6.2.4. Assuming r 2> ·-1, B ~ D(BW).
Thus, not only are the means of B and BWtightly linked, their distributions too are func
tionally dependent. Specifically, the theorem implies the following identity, relating the
z-transform of B and the Laplace transform of B W:
COROLLARY 6.2.5.
V[B] ~ AE[BW]
2
+ A YTBW] (6.2.17)
The first identity is simply E ~ ABW, as above. The second enables us to compute the
variance of BW from that of B.
So, what is V[B]? Consider the quantity l'E[B(B - 1)]. This is called the second bi
nomial momenl of B. For a base-stock policy, following the derivation of (6.2.5), one can
show that
l'E[B(B- I)] ~ des) (6.2.18)
2
(See Problem 6.14. The function G thus has several quite different uses!) For a general
(r, q) policy, analogous to (6.2.9),
""es.) A similar
nventory by T ~
~ E[B(B - I)] =G) :Z;:':;+I G2 (S)
proportional to
6.2.5 Proofs
alive properties
~ customer wait
PROOF OF LEMMA 6.2.1. It is convenient to work with a simple transformation of
IP: Define IPc(t) = (r + q) - IP(I), I 2 0, and IP c = (!PcCl): 12 0]. Thus, IPcCl) mea
sures the distance between IP(t) and its maximum value (r + q). For now, assume IP(O)
:0; r + q, so IPcCl) 2 O.
~ith s = 4,B =
The process IPc is a continuous-time Markov chain. Think of its state space as the
icy with q = r =
integers mod (q). A transition occurs at each demand epoch. At each such epoch I. pro
videdIPjl-) < q - I, IPc(l) jumps up by I. If IPc(t-) = q - I, however, the jump is to
:mand and con o instead of q. reflecting the mod (q) operation. (These are precisely the demands that
,lex supply sys trigger orders according to the policy.) Thus, the process IPc is remarkably simple; it just
lm variable de cycles through its q states, in the same order, forever. (The subscript c is meant to sug
Iso, let D(BW) gest the word cyclic.) Also, all the transition rates equal A. Thus, the generator of IP c is
)fD. the matrix 1.( - I + h), where I is an identity matrix of order q and 1+ is the shift mod
(q) matrix
I
(6.2.16) o
194 Foundations ofltrventory Management
FIGURE 6.2.8
State-transition diagram for [Pc
()_o.. . .o
I \
°\ I8
0-0y.. .O
L- ~
PROOF OF LEMMA 6.2.2. The quantity !O(t) specifies the total amount of supplies
arriving during the interval (t, t + L], since orders placed before tmust arrive before t +
L, while those placed after t arrive after t + L. And, by definition, D(t, t + L] is the de
mand during this same interval. Therefore.
PROOF OF COROLLARY 6.2.5. Letg(zlt) denote the z-transfonn of D(t), that is,
of D(t), that is, At this point you may be worried: Yes, all sorts of factors may affect demand, but
how can we model them all? Do we really need a full-scale model of the weather and/or
the economy? (You may be aware that meteorologists and economists build huge, com
plex models, and even those don't always work very well.) It would take a marketing ge
nius to measure competitive conditions, let alone to understand their dynamics!
Relax. Our purpose here is to understand what happens given a world-demand model
(W, D). As we shall see, the central results of Section 6.2, even the performance formulas,
continue to hold. The only difference is in the leadtirne-demand random vatiahle D. We do
have to compute or estimate its distribution, but no more than that. Moreover, as shown in
Section 6.4, for many practical purposes, it is sufficient to estimate just two parameters,
the average demand rate Aand another one that measures demand variation.
So, to use the results entails some extra work, but we control how much. For a very
imation is rarely
important product, it may be worthwhile to construct a detailed, explicit model of
In-Poisson man (W, D), hut in other cases we can get by with much less, a simpler, approximate model
of (w, D) or just a couple of statistical estimates. Indeed, one can view the model as a
~-ents OCcur that
technically convenient metaphor for complex demand in general. That is, think of Wet)
f these events as
as comprising any and all relevant inforn1ation we have concerning future demand.
I hy a stochastic
We need some rather technical, not overly restrictive, assumptions: W is a Markov
affects the evo
process. Essentially, this just means that W is complete, or self-contained. The variables
, D. The overall
in Wet) constitute the state of the world system; they include all relevant information
ahout the future evolution ofW heyond time t. (Section C.3.2 provides a precise defini
tion, hut that is not critical here.) Also, W is well behaved, like the processes discussed
ther conditions.
in Section 6.2.1 above; it is ergodic and has a limiting distribution. So, it is meaningful
des the relevant
to talk about equilibrium conditions. (So, ifW includes a weather model, it is valid only
for short time intervals, except perhaps in the tropics. Time-varying world models are
conditions in a
discussed in Section 9.7.) Finally, W captures all events that affect D. That is, the joint
biles and office
process (W, D) is itself a Markov process; this process too is well hehaved.
The results allow W to start in any initial state at time O. An important special case
Ie market for a arises when W begins in equilibrium, for then W is stationary and D has stationary in
IS of competing crements (the distribution of D(t, u] depends only on u - t). Define
over time, and
my introducing D ~ leadtime demand, the random variable D(L), assuming W starts in equilibrium
ise, think of a A ~ demand rate
:omputers and
i; there, every = E[D(l)], assuming W starts in equilihrium.
larger or smaller ratio. This is the new demand parameter; it is a fundamental character
istic of the process. So, for large L, V[D] = ",2 AL.
By the way, you might expect that the information embodied in the state W(t) can
be profitably used to help make ordering decisions. Indeed it can. Here, however, we
evaluate an (I; q) policy with fixed rand q, which by definition does not adapt orders to
W(t). (I; q) policies are widely used in practice, because they are so simple to implement.
It is often difficult or impossible to observe W(t) in real time; even when W(t) is ob
servable, it is often hard to process and employ the infonnation. Thus, it is worthwhile
to see what happens under an (I; q) policy, even in settings like this one, where in prin
ciple some more complex policy works better. In the process, we shall see just how ro
bust the techniques above are to the Poisson-demand assumption. (Section 6.7.2 and
Section 9.7 consider more refined policies which do use the information in W(t).)
0
KJllIlegative and Again, W switches back and forth between its two states, and a demand occurs when
megative; and switching from the second to the first. Here, D is a renewal process with Erlang-2 inter
ifferent kinds of demand times.
e is no demand. In general, A = ~Ae. (Both examples above have A = 10.) Also, letting fl be a large
t do generate positive number, one can show (Neuts [1981]) that
and occurs and
V[D] = [A + 2~(A - Al)(fle~ - Q)-lAe]L - 2~A(I - eQ[)(fle~ - Q)-2Ae
I3llds !hat leave
land rates in A. (Odd as it may seem, the result of this calculation does not depend on fl, provided fl is
ices QO and A. sufficiently large.) This may look ugly, but actually it is quite simple. The second term
~-time Markov is a constant, plus some exponential terms which decay to zero as L .------7 00. The first term
cess. is linear in L. Thus,
omplexity, de
f the matrices. 2 A + 2~(A - AI)(fle~ - Q)-lAe
t\J ~
Ig and compu A
and for large L, we can ignore the second term above, using the first (i.e., t\J2 AL) to ap
"fated Poisson
proximate V[D].
.A are all zero
For the MMP process above, t\J2 ~ 1.45, and for the renewal process, t\J2 = 0.5. Fig
IS. During any
ure 6.3.1 graphs V[D] for different values of 1. The dashed lines show the linliting lin
e demand rate
ear functions. For reference, the straight line shows V[D] for a Poisson process with the
200 Foundations ofInventOly Management
r
FIGURE 6.3.1
ffliJrld-driven demand: variance.
1.4
1.2
MMP
0.8
~
'=' 0.6
Poisson
..,.----.
0.4 Renewal
.-~.....
..-:;./(:0;' __
0.2
/;;.
0
same rate, io. = 10. Observe, the MMP process is unifonnly more variable than the Pois
son process, while the renewal process is less variable.
It is not easy to compute the exact distribution of D. This requires solving the differen
tial equations for the transient probabilities of the Markov chain (W, D). Specifically, let
gw(dll) ~ Pr {D(I) = d, I'V(I) ~ w} I :>: 0, d :>: 0
starting with D(O) = 0 and W(O) distributed as t, and let g(dll) be the row-vector
[gw(dll)]w- To compute the g('~I). solve the linear differential equations
g'(Ojl) ~ g(OII)(Q - A)
g(OIO) = t
g(dIO) = 0 d> 0
This is a conceptually straightforward, but practically tedious, numerical task. Then,
g(d) ~ Pr {D = d) = g(dIL)e.
Figure 6.3.2 shows g(d) for the examples above with L = 0.2, so E[D] = io.L ~ 2,
as well as t1le Poisson pmfwitl1 tlIe same mean. It is clear that the MA1.P process gener
.
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 201
FIGURE 6.3.2
Probability mass junctions (world-driven demand).
0.4
0.3
...............
~...-. _MMP
g ..•.. Poisson
{i 0.2 - Renewal
:sson "8
0.
li."\\·al 0.1
....
.....•...
o
o 5 10
0.1
Demand
ates the most variable pmf of the tmee, and the renewal process the least variable, with
: than the Pois-
the Poisson process in between.
Ig the differen
"'ifically, let 6.3.3 Analysis
For simplicity, we state and prove the results assuming that demand is an MCDC process.
They are valid more generally, however. We need a mild technical assumption:
he row-vector
ASSUMPTION. The joint process (!P, W) is irreducible.
This assumption holds for most reasonable models. For example, it is true when D is an MMP
or a renewal process. (See Problem 6.4. Section 6.3.5 discusses the assumption further.)
LEMMA 6.3.1. Under the assumption above, the joint process (IP, W) has a limiting
distribution, IP is uniformly distributed on the integers in the interval [r + I, r + q], and
IP and Ware independent.
In particular, the limiting behavior of the joint process is insensitive to the specification
,I task. Then, of Wand D, except of course for the marginal distribution of W itself. Envision the
result this way: As time passes, IP(t) cycles around its q possible values, and the move
)] ~ AL = 2, ment oflP is determined in part by W. Nevertheless, after a sufficiently long time, the
lfOCess gener position of IP(t) contains negligible information about W(t).
202 Foundations ofInventory Alanagotwnt
IN~IP-D
FrGURE 6.3.3
Irion ofthe net m Performance afbase-stack policies (world-driven demand).
10
l-demand case,
°
case where it does not hold: Suppose W has two states, and 1, and it alternates between
them as time passes. A demand occurs whenever W changes state, but never between
decreasing and transitions. (So, QO = 0, and the diagonal elements of A are all 0. This D is an instance
easing, but not of what is called an alternating renewal process.) Now, suppose q ~ 2, so IP, (defined
ng and convex
)fFigure 6.2.6.
°
in the proof of Lemma 6.2.1) also alternates between the values and 1. Starting with
IPet0) ~ W(O), 1Pett) ~ W(t) for all t 2: 0; more generally, the distance between IPiO)
j model itself.
and W(O) will be preserved forever. Thus, the joint process cannot be irreducible.
two examples It is fair to call this case pathological; it is hard to imagine a real demand process
til with L = 2. with this sort of periodic behavior. There are certain plausible demand models, however,
newal process which also violate the assumption: Suppose the current demand rate depends on the cu
I]. That is, the mulative demand to date; in our terms W(t) is just D(t) itself. Such a model might de
scribe the demand for a new product. The idea here is to model the possible saturation
204 Foundations ofInventory lrfanagement
FIGURE 6.3.4
Stock-sen/ice tradcqff (MMP demand).
2
.. L ~ 1
1.5 ---L~2
-L~J
"
S
.~ -L~4
00
c
'g"
0 '.
'.
~ '.
0
E
~
"
0.5
a
0.001 0.010 0.100 1.000
Customer waiting time
of the market, with the demand rate decreasing in D(t). (See Mahajan and Wind [1986]
for examples.) Clearly, the joint process is reducible.
On the other hand, suppose Wet) increases whenever a demand occurs, but Wet) can
also decrease between demands. Such a construct could model a market with limited ca
pacity to absorb short-term demand surges, provided the market adjusts so that recent
demands are "forgotten" in the long run. It is not hard to check that, indeed, the as
sumption is now satisfied.
PROOF OF LEMMA 6.3.1. Because the joint process (IP, W) has a finite number ofstates,
and in view of the assumption, we need only show that the distribution described in the
lemma is stationary. We shall give two separate arguments, each instructive in its own way.
First, here is a direct algebraic argument. The generator of the joint process (IP c' W)
can be written
10 QO -10 AD + 1+ 0 A
(1+ is the shift mod (q) matrix in the proof of Lemma 6.2.1 above; Section C.S.3 inAp
pendix C explains the Kronecker product notation 0.) The first term corresponds to
transitions ofW without demands, while the second and third together describe demand
events, hence changes in IP C' whether or not W changes at the same time. We wish to
show that the probability row-vector (l/q)(e' 0 !;) is stationary for (IP" W); that is,
when we multiply this vector by the matrix above, the result is zero. But, since e' J = e'l +
~ e' and !;Q = 0, we have
This completes the first argument. For the second, suppose the system begins at time
with [IPiO), W(O)] having the distribution described in the assertion. Now, for any I;=: 0,
°
[!Pil)1 W(I)] = {!PiO) + [D(t) I W(I)]} mod (q)
The expression on the right-hand side consists of a random variable IPiO) distributed
uniformly on the integers mod (q), plus another random variable [D(t)IW(t)] indepen
dent of IPiO), all mod (q). Any such combination, it turns out, is also uniformly dis
tributed. (This can be checked directly, as Problem 6.5 asks you to do; or see Feller
[1971], page 64.) Thus, [IPil)IW(I)] has the uniform distribution on the integers mod
(q); in particular, IPJI) and W(I) are independent. Finally, W(t) itsdfhas the same dis
tribution as W(O), namely, the stationary distribution ofW.
The assumption truly is necessmy for the result. Consider the example above, where
W has only two states and q ~ 2. It is still tme that (l/q)(e' ® ~) is a stationary vector
for the joint process, but there are others as well; the stationaty vector is not unique. Also.
if the transition rates between the two states of W are different, then the limiting distri
bution of IP" depends on the initial state, and typically this distribution is not uniform.
PROOF OF THEOREM 6.3.2. Consider the joint process (IP, DL), as in the proof
of Theorem 6.2.3 above, but expand it to (IP, W, DL). By assumption, D(I, I + L] de
pends on the past (through time t) ofIP, W, and D only through W(I). On the other hand,
(IP, W) has a limiting distribution, in which IP and Ware independent. Thus, the entire
LOOO
process (IP, W. DL) has a limiting distribution, and lP is independent of [W, D]. Fur
thermore,IN(t + L) continues to be described by equation (6.2.7); in particular, IN is a
function of (IP, W, DL). Taking the limit as 1--> 00, the result follows immediately.
d Wind [1986]
We need some basic facts about normal distributions (see Section C.2.5.4): <p de
notes the standard normal probability density function (pdf), and <1>' the standard nor
mal ccdf. The fimctlon
<l>1(z) = S:(x - z)<\>(x) dx (6.4.1)
I
is the standard normal loss function; it is the standard-normal analogue of G . The fol
lowing identities are useful:
<l>1(z) = S;'<I>°(x) dx = -z<l>°(z) + <\>(z) (6.4.2)
Now, following the analysis of Section 6.2.2, using the normal approximation to D
in (6.2.3), we obtain
lJ = 1>'(z)(J" (6.4.5)
of C 1 The fol-
~
g
'" 2 2
!
(6.4.2)
(6.4.3)
>ximation to D
o o
o 5 10 15
Base-stock level
it is fairly good. (Part of the discrepancy, of course, is caused by the approximate value
2
of a .) Again, for larger L, the approximation becomes still more accurate.
The normal approximation makes some people uncomfortable, because it allows D
to assume negative values. (This is not a weakness at all, of course, when customers can
return items. For instance, retailers can usuaIly return unsold CDs to manufacturers.
While such situations are not uncommon, they are not the general rule.) Remember, we
use the approximation, not to describe the dynamics of the system in detail, but rather to
predict performance, and it does that very well in many cases. Still, the approximation
tends to work best when it predicts a small Pr {D < O} = <p o[-v/a),.r1amely, when the
coefficient of variation a/v is small. For Poisson demand, IT/v = lIVAL, and this is one
reason the approximation works poorly for small AL.
(6.4.4)
One note of caution: The nOffilal approximation does capture the broad features of
(6.4.5) performance, but it is not precise when pushed to extremes. Suppose we have a stringent
performance requirement, specifying a tiny value of .4, say 0.001. We know that s must
(6.4.6)
be large. but how large? The normal approximation does not always give the right an
emand case. swer. In technical terms, it does not provide an asymptotically precise estimate ofthe tail
, 10 and two probabilities; the ratio <po(z)/Co(s) does not go to I as s --> 00. (This assumes, of course,
ely accurate. that the model and the data are exactly correct. Errors from those sources usually dom
does the ap- inate those of the approximation.)
In general, the appropriateness ofthe normal approximation depends partly on L be
'ljJ2 = 1.45) ing sufficiently large, but also on the context and purpose ofthe analysis: If we are ana
ore, but still lyzing one particularly important item, and we are unsure whether L is truly large
208 Foundations o/Inventory Management
F,GURE 6.4.2
Performance afbase-stock policies (nomwl versus .MMP).
5 5
. L = 0.5 normal
--- L= 1.0 normal
-L~O.5MMP
4 - L~ 1.0MMP 4
r
00 2
, .II
oj
,'.
,.,.
.1
,.'.
3
2
r
,~
''l,
,,
,,
'.
o '-"""" o
o 5 10 15
Bast'-stock level
enough, then prudence dictates using the exact fonnulas. On the other hand, if we are
conducting a large-scale analysis to support systems design, the approximation is usu
ally quite adequate for any L. III that I..:ase, therefore.. there is no need to model the world
demand system (W, D) explicitly, nor even to estimate the full distribution of D. It is
enough to estimate f... and ~2 directly.
Suppose that we are comfortable with the approximation. Formulas (6.4.4) to (6.4.6)
are quite easy to usc. For instance, to meet a service·level requirement A = ] - w _, find
the value of z such that <l>°(z) = I - OL from a table of the standard nonnal distribution
(or a computerized equivalent), and set s =. v + Z(J. (As mentioned in Section c'Z.5.4,
In [4>°(z)] = - 'IZ2 for large z. So, for I - w_ near 0, z = [-2 In (l - W_)]I/l and s =
v + [- 2 In (I - w _)] 1/2 cr. This is a rough, qualitative indication of how s grows as the
service requirement becomes more stringent. However, as discussed above: the nonnaI
approximation loses accuracy in this range, so this estimate should be used with care.)
This z is sometimes called the safetyfactor. It depends only on w_. It is positive precisely
when (u_ >~. as is usually the case. If v changes but 0" doesu't,just shift s by the same amount
to maintain z. If IT changes, the safety factor indicates the direction and magnitude of the ap
propriate change in s. Following these rules, the average inventory is then proportional to 0".
Thus, the key parameter affecting perfonnance is cr, the standard deviation of Jeadtime de
mand. (The cost-optimization approach of Section 6.5 leads to similar conclusions.)
By the way, some common practices deviate sharply from this approach. Certain
firms follow guidelines of the form, "keep 2 weeks of inventory" or "keep 2 weeks of
---
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 209
safety stock." In practical terms this means setting either s ~ 21e or s = v + 21e (pre
suming time is measured in weeks). So, IT plays no role at alL There is no justification
5 for rules of this kind. They are recipes for trouble.
Now, IT itself is proportional to the square root of 'AL. So, a shorter leadtime im
proves performance; we observed the same phenomenon in Figures 6.2.4 and 6.3.4.
~ Likewise, as the demand rate grows (with \~' fixed), so do 1 and S. However, D' grows
more slowly than 'AL itself; that is, the performance measures display a sort of statistical
economy ofscale. If A doubles, for example, and we adjust s to maintain the same value
3 of z, then 1 and S both increase by only Vi, not 2.
f This pattern of sensitivity is similar to the EOQ model's, though the reason here is
entirely different. When we double A, the new D is the smn of two independent copies of
g" the original. The standard deviation only increases by Vi, essentially because a small
2
value in one copy may cancel a large value in the other. (This is the familiar principle un
derlying statistical estimation; an estimate becomes more precise as the sample size
grows.) The same thing happens if we doubleL, because D has independent increments-
exactly in the case of Poisson demand, approximately for MCDC demand-so V[D] is
linear (exactly or approximately) in L.
Also, (T is proportional to tP, the square root of the variance-to-mean ratio tP 2 . Thus,
o reducing demand variation improves performance. (Problem 6.1 explores another way
to obtain better demand information.)
The approximation also helps to explain more subtle performance phenomena: No
tice that ] here is predse(v B retlected about the value z = 0; the exact measures exhibit
nearly the same symmetry, as shown in Figures 6.2.3 and 6.3.3. Also, it is not hard to
show that
1aIld, if we are
mation is usu as _ al
ion of D. It is
Thus, bothB and 1 are increasing in (T, and their sensitivity to (T is greatest for z = 0, that
.4.4) to (6.4.6) is, for s = v. The figures above show that the exact Sand Ibehave similarly.
. I - w_, find
al distribution 6.4.2 Base-Stock Policies-Other CO/lti/luousApproximations
ction C.2.5.4,
_)]1/2 and s = The normal approximation is one particular continuous approximation. In general, re
r grows as the placing GO by a continuous ccdfF O leads to the following formulas:
e. the normal A ~ FO(s) S ~ FI(s)
d with care.)
l~s-v+S
;itive precisely
~ same amount where F\s) ~ J; F°(x) dx is the continuous loss function. [Clearly, (6.4.4) to (6.4.6)
ude ofthe ap specialize these results to the normal case.] These approximate formulas have the same
)()rtional to IT. qualitative properties as the exact ones: For instance, B is decreasing and convex as a
f leadtime de function of s, while 7 is increasing and convex.
ions.) For example, consider an exponential distribution with mean v = E[D], so FO(x) =
)3.ch. Certain e -.dv. Then, Fl(x) = ve-xlv. (This approximation works well for certain stochastic
p 2 weeks of leadtime systems, as explained in Section 7.3.11.)
..
a. . 0 (z) ~-
1 [ 1- Z J
2 (1 + ~)1I2
It so happens (but it is not really important) that [l0 is the ccdf ofZ ~ where Thas T/V2,
Student's t distribution with 2 degrees of freedom. [l1 is the corresponding loss function,
i.e., [l1(Z) = J; [l°(t) dt. Suppose we approxilnate D by v + o-z. (Unlike the normal ap
proximation, this construction does not match the moments of D. In fact, while Z has
zero mean, its variance is infinite.) The approximate performance measures are similar
in fonn to the nonnal approximation's:
A = [l°(z)
B= [l1(z)<r 7= [l1( -z)<r
Now set
r-v r+q-v
Z ~---
r <r Zr+q
<r
To derive the performance measures, there are two alternative approaches, which lead to
the same results: Follow the analysis leading to the exact formulas (6.2.11) to (6.2.13),
using the continuous approximations in place of the actual IP and D. Or, average the
...
quantities in (6.4.4) to (6.4.6), letting s range over the interval [r, r + qJ. (This second
approach is easier.) Either method yields
as the perfonnance measures, not approximations. They are popular because they are ruI
simpler than the exact formulas. However, as we shall see, they are not much simpler, fIG
and they are not reliably accurate. It is necessary to know about them, because they are tbI
so widely known and used, but 1 do not recommend them.
First, consider the following approximations of A and B: is
pi
- _(I)q
A+ 1 - G 1(r) - (I)q
B+ 1 ~ G2(r) .
III
Compare these formulas to (6.2.11) and (6.2.12). Clearly, A+ 1 is similar to A in (6.2.11),
but it omits the term -(l/q)G 1(r + q). Likewise, B+ 1 omits tbe second term ofB in IIIlj
(6.2.12). Because the omitted terms are negative, /1+ 1 2': A andB+ 1 2: B. These approx
~
I
imations overestimate the exact quantities. The errors can be severe. For example, for the
exponential approximation, it is easy to sbow tbatA +lIA = l3 + /B = 1/(1 - e -qlv). This
can be quite large when qlv is small. That is, these approximations are unreliable, unless
we can be certain that qlv is large.
Second, consider the following approximations:
A+ 2 = GO(r) B+ 2 ~ G\r) 6.5 OPtimi1
Recall, A is the averagc ofthe q numbers GO(s), r -; s < r + q. and A+ 2 replaces tbe
average by the first one. Since CO is nonincreasing, A+ 2 2:': A. Likewise, B+ 2 2:': E. So,
like those above, these approximations overestimate the true quantities. Whereas A + 1
and E+ I are most accurate when q is large,A+ 2 and E +2 work best for small q. These 1
approximations too can give wildly inaccurate results. For the exponential approxima I
tion, A+ 2 IA = B+ 2 /B = (qlv)/(l - e- qIV ), which can be enormous, unless we know in
advance that qlv is small.
By the way, A+ 2 is sometimes proposed as an alternative measure of customer ser
vice. It is the probability of mnning out ofstock during an order leadtime. But, this does
not measure customer service in any meaningful way. Customers might well care about
A andB. but not what happens during one of our leadtimes.
So, neither of these approximations is particularly reliable. As for simplicity, well,
to evaluate a policy, either approximation requires about half the computational effort of
the exact formulas. This is not a compelling improvement.
The approximations do have a slight advantage when it comes to selecting a policy
to meet a specified requirement: For example. suppose q is already deteTIllined some
how, so the remaining problem is to choose r, and there is a given target value I - ClL
for A. Suppose we use the normal approximation for D. Using the exact formula (6.4.9)
for A, we must solve A = I -- ClL for r, a nonlinear equation in one unknown. Using ap
proximation A+ 2 , the equation becomes ct>°(zr) = I - w_.
Now, it is not terribly difficult to solve the exact equation by numerical methods
with a computer. The approximate equation, however, can be solved manually: Just read
z,. ~ (<1>0) -1(1 - 0>_) from a table ofthe standard normal distribution, and set r = v +
Z,.(J'. Likewise, the analogous equation based on A+ 1 requires a single table lookup (in
this case a table of<1>I).
Indeed, these approximations first became popular in the days before computers
were widely available. Armed only with a handbook of statistical tables and a slide
...
because they are rule, an analyst could solve the approximate equation readily, whereas the exact equa
ot much simpler, tion required tedious iterations. Today, however, this is no longer a good reason to use
because they are the approximations.
Also, since A + 1 and A +2 overestimate A, the r obtained from either approximation
is larger than the solution to the exact equation, sometimes much larger. Thus, the ap
proxlmations lead to higher-than-necessary inventory. This is a good reason to avoid
them.
6.5 Optimization
L 2 replaces the 6.5.1 Formulation
e, B +2 ;::: B. So, As we have seen, the fonnulas for perfonnance measures have the same general fonn
s. Whereas A+ 1 for both Poisson and world-driven demand. The demand process affects perfonnance
r small q. These
solely through the distribution of D, the leadtime demand. Also, a continuous approxi
ttial approxima mation of D leads to perfonnance measures of nearly the same fonn.
'ess we know in
hI this section we impose a cost structure on the problem and show how to compute
an optimal policy. We also investigate qualitative characteristics of optimal policies and
)f customer ser optimal system performance. These methods and results apply to all of the specific de
e. But, this does mand models above. (They apply also to many of the more complex supply models of
well care about Chapter 7.)
The cost factors are the same as in the deterministic model of Section 3.3 of
implicity, well, Chapter 3:
Iliona! effort of
k = fixed cost to place an order (moneys)
ecting a policy h = cost to hold one unit in inventory for one unit of time
::rmined some (moneys/[quantity-unit· time-unit])
value 1 - w b ~ penalty cost for one unit backordered for one unit of time (moneys/[quantity
Drmula (6.4.9) unit· time-unit])
>wn. Using ap (Because the average variable cost is the constant cx. in all cases, it plays no role in de
tennining an optimal policy, so for now we ignore it, treating c = 0.) All these factors
:rical methods are positive, unless stated othenvise. The total average cost, then, is
Ially: Just read
Idsetr=v+ C(~ q) ~ kOF(r, q) -I h7(r, q) -I bB(r, q) (6.5.1)
ble lookup (in The goal is to determine (integer) values of rand q that minimize this function.
Similarly, we can [annulate the average cost by using a continuous approximation.
)Ie computers We call this C(r, q) also. Here, it makes sense to treat the policy variables too as contin
'" and a slide uous. As we shall see, this case is easier, both conceptually and computationally.
,~
C10
214 Foundations ofInventory Management
Figure 6.5.1 illustrates this function. For small s, 7(s) is negligible andB(s) is nearly (II
linear with slope -I; thus, C(s) is nearly linear with slope - b in this range. For large s,
B(s) goes to zero. while 7(s) becomes linear with slope + I, so C(s) becomes linear with
slope h. Finally, as noted in Section 6.2.2, 7(s) andB(s) are both convex functions, so C(s)
is too.
I
6.5.2.1 Continuous Approximation: General Case
use the corresponding formulas for B(s) and 7(s) in the cost function (6.5.2), treating s
F,GURE 6.5.1
Average cost function.
Base-stock level, ,~
--_.- - - - ...-...--.
where w is the cost ratio b/(b + h), as in Chapter 3. For simplicity, assume that D has a
positive pdf((x), at least for x '" o. So, FO(x) is strictly decreasing, and (6.5.3) has a
unique solution, say s*. This is the optimal base-stock level. If the approximate D is pos
ben, itive (Fo(x) = I for x sO), then s* is positive.
To detennine s*, then, means to solve the single nonlinear equation (6.5.3), a nu
merical problem for which a variety of standard methods are available. In certain special
cases, the solution can be computed in closed fonn, as shown below.
Recalling that A(s) ~ FO(s), equation (6.5.3) prescribes the following simple rule
:rion 6.4.2, and for detenmining the optimal policy: Set the base-stock revel so that the stockout proba
.5.2), treating s bility equals the critical ratio I - w. This is the same calculation required to meet a spec
ified target level I - w_ for .4.
6.5.2.2 NormalApproximation
Specifically, consider the normal approximation of D. That is, use (6.4.5) and (6.4.6) for
7(5) and Ii(s) in the cost function (6.5.2). In terms of z = (s - v)/IT and the standard nor
mal ccdf <po, equation (6.5.3) now reads
<p°(z) ~ I - w
Indeed, LlC(s) < 0 for all s < s·, and LlC(s) :2: 0 for all s :2: s·.
a scale factor, Now, using (6.2.5) and (6.2.6) and the definition of C', we have
I uni t changed,
Solving (6.5.4) is essentially like solving a nonlinear equation, with minor adjust
ments to account for the integrality of s. It is straightforward to compute s· through a
here z* solves
standard bisection procedure, for example. (In special cases, (jJ can be inverted and s*
expressed in closed form. This is true, for instance, of a geometric distribution, which
arises in the models of Section 7.3.)
Let r* and q* denote the optimal rand q, and c* = CCr*, q*) the optimal cost. Let
qk and Ck denote the optimal batch size and cost in the deterministic model; that is, qk =
(2k)Jhw)l/2 and C k = hwqk = (2k'Ahw)l/2. For the base-stock model, lets* denote the op
timal base-stock level and Crr ~ CCs*) the optimal cost. Also, C a + = (bh)1I2" denotes
the base-stock model's optimal cost under the maximal approximation. Finally, let qrr ~
Cu/hw and qu+ = Cg+/hw.
THEOREM 6,5.1.
20
(6.5.5) ,"0
"
)112 "'" 15
(6.5.6) ~
<
2 10
(6.5.7)
(6.5.8)
5
(6.5.9)
at the interval
o
are bounds on o 2 4 6 8
Dion about D. Sigma
• Cu'
os of both the
One additional property is worth mentioning: Define the function C*(q) = min,
1* and C* are
{C(r, q)}, the optimal cost with q held fixed, and recall the EOQ error function .(x) ~
md C* are ro
,,(x + I Ix), x> O. Then,
'1* and C* are
; before; lead C*(q) "5 .(~) (6.5.10)
lOunt. In sum, c* q*
'he benefits of
lore on sensi (See Problem 6.19.) Chapter 3 presents a similar relation for the deterministic case, but
with equality instead of inequality. Thus, the insensitivity property here is even stronger.
simplest is qk Ifwe choose some other q besides q*, by error or by design, the cost penalty is modest,
tmization pro provided q is not too far away from q*.
S fl"5 I. (The PROOF OF THEOREM 6.5.1. The proof is long, but it requires only basic
Is suggest rea- calculus. Recall thatB(r, q) and I(r, q) are simple averages of B(s) and I(s),
respectively. Therefore, if C(s) denotes the cost of a base-stock policy as in (6.5.2),
mal approxi the objective function can be writren as
;everal values
lit the middle k'A + I;+q C(s) ds
JIVe) and the C(r, q) ~ (6.5.11)
q
19 nearly lin
IUIlds (6.5.8). Let us first prove the bounds on C*. Recall that Cry) = h[y]+ + b[y]-, and define
proximation.
C(s) = C(s - E[D]) = C(s - v)
!3.litative sen
C+(s) ~ Cu + C(s - SO)
_.,
FIGURE 6.5.3
Optimal batch size (exact and bounds).
30
i
25 ~
20
"
N
'f;;
.c 15
B
•
"' 10
o
o 2 4 6 8
Sigma
Thus, C_ and C+ are translated versions of the convex, piecewise-linear function t. See
Figure 6.5.4. We claim that
C_(s):5 C(s):5 C+(s) (6.512)
as in the figure. First, C(s) = E[C(s - D)], so Jensen's inequality (see Section C.2.2)
yields the first inequality in (6.5.12). The second inequality is immediately apparent,
since C+(s*) = Crr = C(s*), and C+(s) is steeper than C(s), both above and below s*.
Now, replace C(s) by C_(s) in the cost function (6.5.11), and call the result C_(/; q).
It is easy to show (Problem 3.10) that this is precisely the cost function ofthe determin
istic model, and (6.5.12) implies C(r, q) 2" C_(r, q) for all (r, q). Thus, C* 2" C" which
is the lower bound in (6.5.7).
The lower bound in (6.5.9) comes from applying a similar argument to the function
C_ +(s) = max {CmC_(s)}. (This is another piecewise-linear function, in this case with
three pieces. It replaces the lowest parts of C _ (s) with a horizontal segment at level Crr-)
Clearly, C_+(s) :5 C(s). So, substituting C_+(s) for C(s) in (6.5.11) yields a lower bound
on C*, and direct calculation shows that this bound is (C; + C~)I12.
Likewise, replace C(s) by C+(s) in (6.5.11), and call the result C+(r, q). Clearly,
C+(r, q) ~ Crr + C_(r', q), where y' = r - s* + v. Thus, the minimal value of C+(r, q)
is Crr + Ck . (The translation of y' leaves the minimum unchanged.) But (6.5.12) implies
C(r, q) :5 C +(r, q) for all (r, q), so C* :5 Crr + C" which is the upper bound in (6.5.9).
For the upper bound in (6.5.7), fix D for the moment, and let V ~ y - D. Direct cal
culation yields
Chapter 6 Stochastic Demand: One Item with Constant Leadtimes 221
FIGURE 6.5.4
Piecewise linear approximations.
C+
Co e-------,-------~--------------------------------
"
function C. See
J;+q C(s - D) ds ~ J~+q C(y) dy
(65.12)
~,b[V" - (q + V)"J V< -q
Section C.2.2)
ately apparent,
j}[h(q + V)2 + bV"] -q"; V.,; °
and below s*. Xh[(q + V)2 - V"] V""O
resnlt C_(I; q).
This quantity, clearly, is.,; j}[h(q + vl + bV 2] for all VTherefore,
fthe detennin
* ~ Ck> which J;+q CCs) ds ~ J;+q E[C(s - D)] ds ~ E[f;+q C(s - D) ds]
.,; XE[h(q + V)" + bV"]
to the function
l this case with ~ X{h(q + vl + bV 2 + (b + h)E[(V - V)2])
nt at level Cu.)
~ X{h(q + vl + bV" + (b + h)<r"}
a lower bound
where V ~ E[V] ~ r - v. Now, C(r, q) = [kA + Xh(q + V)" + XbV 2jlq, so
(r, q). Clearly,
I (b + h)<r2
lue of C+(r, q) CC,; q) .,; C(r, q) + 2: q
,.5.12) implies
nd in (6.5.9). The right-hand side has the same fonn as C_(r, q) itself with kA augmented by
D. Direct cal- X(b + h)<r 2 , so its minimal value is {2[kA + X(b + h)<r2 ]hw}1I2 ~ (2kAhw + bh<r2 )!/2
This, therefore, is an upper bOlUld on C*, as asserted.
a":
c
FIGURE 6.5.5
Optimality condition.
C(s)
H(q)
A(q)
Co
r s* r+q s
Now let us prove the bounds on q*. The optimality conditions for minimizing
C(r, q) in (6.5.11) are aCi ar = 0 and aCI aq ~ O. These equations reduce, respectively, to
Figure 6.5.5 illustrates the idea: (6.5.13) says, for any fixed q, set r to equalize C(r) and
C(r + q). So, on the graph of C(s), these two points have equal heights and can be con
nccted by a horizontal line. (By the convexity of C(s), this implies r ==; s* ==; r + q. which
proves (6.5.5).) Let H = H(q) denote this common height.
With r chosen to satisfy (6.5.13), the shaded area A ~ A(q) is given by
A(q) ~ qC(r + q) - f;+q C(s) ds
Tum now to the exact formulation, where r and q are integer variables. There is no stan
for minimizing dard, tractable method to optimize a function of several integer variables. We now de
,respectively, to velop an algorithm specifically to optimize C(r, q), as defined in (6.5.1). It is based on
properties analogous to those of the continuous model in the proof ofTheorem 6.5.1.
(6.5.13)
First, rewrite C(r, q) in a form analogous to (6.5.11) above:
(6.5.14)
C(r,q ) .k.A_+---,-l:lC;:t:~;c'+CJl,--C::-(c.cS)
~- (6.5.16)
ualize C(r) and q
IIld can be COll
where C(s) is defined in (6.5.2). Think of C(r, q) as the average of the costs of q base
'" r + q, which
stock policies, plus the fixed-cost term kAlq.
Let C 1 denote the smallest value of C(s) as s ranges over all the integers (of course,
by
C 1 ~ C(s*)), C2 the second smallest value, C3 the third smallest, and so on. That is, con
struct the sequence {Clo C2 , . . . } by listing the values C(s) in order, paying no attention
to the order of the s's themselves. Because C(s) is convex, the first q of these numbers
correspond to consecutive values of s, for any q.
lIity condition Fix the value of q for the moment, and let r*(q) denote the best value of r. By
(6.5.16), for any r, C(r, q) is the average of q consecutive values of C(s), plus kAiq. To
determine r*(q), therefore, choose r such that {C(s) : r + 1 '" s '" r + q} comprises the
(6.5.15)
smallest q values of C(s), that is, C lo . . . , Cq .
Ie defining H, Furthermore, suppose we have determined r*(q) in this way. To find r*(q + 1), just
find Cq + 1 and add it to the set C lo ' .. , Cq . Moreover, by convexity, this Cq + 1 is C(s) for
C+(s) instead some s just outside the interval [r*(q) + 1, r*(q) + q]. There are only two possibilities:
ount dq. Then, Either Cq + 1 = C(r*(q)), in which case r*(q + 1) = r*(q) - 1, or Cq + 1 = C(r*(q) + q
_., ._
Step 0 (initialize):
Determine s* as above.
Set q <-- I, r <- s* - l, C* <- klc + C(s*).
Step 1 (determine next smallest C(s)):
Set c. <- min {C(r), C(r + q + Ill.
Step 2 (testfor termination):
If c. ;" C*:
Set q* f----- q, r* (- r.
STOP.
Step 3 (update):
Setq<--q + 1.
Setr<-r- 1.
Go to step 1.
1.73
2 0.91
3 0.62
4 0.65
5 0.83
6 1.06
Step 0:
s*=3,q=1,r=2.
C* ~ 0.54 + 0.62 ~ 1.16.
q < q*, C*(q) > Step 1:
e induction) that c* ~ min (0.91, 0.65} ~ 0.65.
<S, q* is optimal.
Step 2:
"Optimal policy:
c.. = 0.65 < 1.16 = C*, so proceed to
Step 3:
q~I+1~2.
Step 2:
c.. = 0.83 < 0.89 = C*, so proceed to
Step 3:
q ~ 2 + 1 ~ 3.
C* ~ 0.89 - (0.89 - 0.83)/3 ~ 0.87.
rremains 2.
Step 1:
c. ~ min {0.91, 1.06} ~ 0.91.
Step 2:
c. ~ 0.91 > 0.87 ~ Co, so STOP with
r* = 2, q* = 3, C* = 0.87.
6.5.4.2 Discussion
Evidently, Algorithm Optimize_rq starts with g ~ I and successively increments g until
g* is found. It is quite efficient when g* happens to be relatively small, but when g* is
large, it may require a good deal of time. The algorithm can be extended, however, to
start with any policy:
For instance, by analogy to the continuous model, a plausible initial value for g is
g" rounded to an integer. A plausible,. is an integer near s* - ~g, for some ~ between 6.6 LumpJ
\\ and I - w.
Given an initial policy, first find r*(g) for the initial g. This can be done with a sim
ple search procedure based on the ideas above. Then, determine whether to increase or
decrease g by comparing Cq + I to C*(g). If Cq + 1 < C*(g), then g < g*, so proceed with
Algorithm Optimize_rq, beginning with step 3. Otherwise, if Cq + 1 ;" C*(g), then
g ;" g*. It is not hard to modify the algorithm to decrease g until g* is found. (Problem
6.21 asks you to work out the details.)
It is worth mentioning that the results above and Algorithm Optimize_rq do not re
ally require C(s) to be convex. The key property is that -C(s) be unimodal (i.e., C(s) is
nonincreasing for s < s* and nondecreasing for s ~ s*).
On the other hand, the algorithm is not guaranteed to work if there is a penalty cost
on A instead ofB. (As discussed in Section 3.3.7, this is a peculiar objective, and it leads
to peculiar results in the deterministic case.) In this case, the average cost can still be rep
resented in the form (6.5.16), where C(s) is redefined as
C(s) ~ hIes) + bA(s)
= h[s - v + GI(s)] + bGo(s - I)
Now, -C(s) happens to be unimodal when D has a Poisson distribution, and in that case
the algorithm works. In general, however, - C(s) need not be unimodal. (One occasion
ally finds suggestions to the contrary in the literature; beware!) Without that property, it
is much harder to find an optimal policy.
tion 3.3.5, it recovers the optimal policy.) Moreover, as discussed in Section 3.3.7,A is
:rements q until
usually an ad hoc proxy for more fundamental criteria. For these reasons, we believe the
but when q* is
approach is a sound one.
ed, however, to
~ value for q is
>me 13 between 6.6 Lumpy Demand
me with a sim 6.6.1 Discussio1l
r to increase or Up to this point we have assumed that demands occur one unit at a time. Now, consider
o proceed with a system where demand D is a compound-Poisson process: Demand epochs form a Pois
'" C*(q), then son process, denoted D, but at each epoch the amount demanded is a random variable.
""'d. (Problem This model describes demands that arrive in large, unpredictable lumps.
The amounts demanded at different epochs are identically distributed and inde
"_rq do not re pendent of each other as well as of the timing of the epochs. For convenience, assume
Ia! (i.e., C(s) is they are integer-valued. (Similar results hold for continuous quantities.) Let Y be the
generic demand-quantity random variable. For technical reasons, assume gy( I) > 0 and
a penalty cost gy(O) ~ O. The demand-epoch rate is A, so the overall demand rate is AE[Y].
"e, and it leads Each demand increment can be filled separately. For example, if a demand of size
an still be rep Y = 5 occurs at a time t when I(t-) = 2, the customer is willing to take the two units
available, leaving a backlog of B(t) ~ 3. (This is not always true, of course; a customer
may insist on all or nothing. This second scenario leads to a far more complex model,
however, and only rarely to fundamentally different system behavior.) Moreover, the cus
tomer cares when each demand unit is filled; the time to fill the overall demand Y is
>d in that case unimportant. That is, the primary service measure remains B, measured in units, not the
me occasion number of waiting customers.
w property, it What policies make sense in this setting? The (r, q) policy works well for pure Pois
son demand, so it is reasonable to adapt it to compound-Poisson demand. There are sev
eral different adaptations; we shall consider two. In both, a reorder point r triggers or
ders, as before, but the order quantities are determined differently.
First, there is the (r, s) policy. The second parameter s is called the target stock level,
. cost b. Also,
and s 2: r + I. An order is placed whenever IP(t-) drops to r or less. The order size is
udes only the s - IP(t-), which raises IP(t) immediately to s. Thus, each order is for at least s - r units,
ninimize this but some orders may be larger; the order size is random. (Such policies are frequently
called (s, S) policies, with s in place of our r and S for s.)
;fixed q, it is The second type of policy is the (r, q) policy. Here, q 2: I is the basic batch size.
however, re Each order is a positive-integer multiple of q. When IP(t-) <0 r, first order a batch of size
q. If this is not enough to raise IP(t) above r, add another batch of size q to the order.
1>-, replace k Keep adding batches in this way until IP(t) is at least r + I (but no more than r + q).
i above. This This kind of policy is useful when there is some "natural" order size, such as a truck
icy (r, q) has load. (Sometimes this is called an (r, Nq) policy, where N indicates the random number
I which case
of basic batches in an order.)
A bit of reflection should convince you that both types reduce to a standard (r, q)
i>e true cost. policy when the demand size is always one. Also, an (r, s) policy reduces to a base-stock
:ase of Sec policy when s - r = 1, and so does an (r, q) policy when q ~ 1.
·~
As Chapter 9 shows, under staodard cost assunlptions, the best (r, s) policy is in fact
optimal. This is not true of (r, q) policies. As noted above, there are special situations
where ao (r, q) policy is preferred, but otherwise, in general, an (r, s) policy performs bet
ter. However, the best (I; q) policy is often close to optimal.
The program here is the same as before: First, we evaluate the performance mea
sures )f. B, 7, and of. Then, we compute an optimal policy within each of the two pol
icy classes, As we shall see, much of the earlier analysis for pure Poisson demand car
ries over to this model.
(This is the discrete renewal Junction corresponding to gy.) Define q ~ s - r, and let L
IP,(t) = I:een
+Y
if this quantity is < q
otherwise
..
Ipolicy is in fact Again, the fact that g,.(l) > 0 implies that IP, is irreducible, and its stationary vector
pecial situations turns out to be precisely t. (Problem 6.22 asks you to verify this fact.) The distribution
oy perfonns bet of IP itselffollows immediately.
Slep 2: The demand process D again has independent increments. Therefore, for all
rfonnance mea I, the random variables IP(I) and D(I, 1+ L] are independent.
lof the two pol Slep 3: For either policy, (6.2.7) remains valid:
an demand car
IN(I + L) = IP(I) - D(I, 1+ L]
The limit argument in Section 6.2 also remains valid, so
IN~IP-D (6.6.1)
.3: Describe the where IP and D are independent. The leadtime demand D now has a compound-Poisson
mine the distri distribution (Section C.2.3.8). Also, IN describes the long-run frequency distribution of
IN, by arguments parallel to those of Section 6.2.
cribe IP) sepa Thus, the basic qualitative results for the pure Poisson-demand case continue to
; differ. The re- hold. While IP depends on the policy type, the identity (6.6.1) does not.
Slep 4: First, consider an (r, q) policy. Because IP is unifonn and (6.6.1) holds, the
is precisely the perfonnance measures A, B, and 1 are precisely the same as in the Poisson-demand case,
tion IP has the namely (6.2.11) to (6.2.13). The functions d, G', and G2 now represent the actual
compound-Poisson distribution of D. Also, in the formula for Z r - AL now becomes
r) = (r + q) r - XE[Y]L. For a base-stock policy the simpler fonnulas of Section 6.2.2 apply.
s a continuous As for OF, actually an order can be defined in two ways: Each basic batch of size q
ne, then can be viewed as a separate order. Under this interpretation, several orders are sometimes
placed simultaneously. Alternatively, when several basic batches are ordered at once,
they can be viewed as composing a single order. The correct interpretation depends on
istationary for the economics of the situation. The first approach makes sense when each basic batch
proof ofTheo contributes to the workload, as when a separate vehicle must be used for each q, while
ucible. Conse the second applies when the workload is independent of the order size.
mifonn on the Under the first interpretation, where each basic batch is a separate order, the over
all demand rate is AE[Y], so
, here, but still
OF ~ AE[Y]
q
which is analogous to the result for the pure Poisson-demand case. The second interpre
tation requires more care: At a typical demand epoch, an order occurs with probability
- r, and let '"
I;::+,Pr [IP = i} ·Pr [Y'" i - r} = I):~mG~(;)
~ G)[E[Y] - G}(q)]
- IP(I). At a where, consistent with the earlier notation, G~(d) ~ Ij~~d G~(j). Since demand epochs
occur at rate :>t,
OF = (~)IE[Y] - G}(q)]
"] ..
230 Foundations ~fJn\!efltory Management
B = E[[INn
= ~~~'+l Pr (IP = d) . E[[D - dt]
= l~=r+ 1 !'S_dG1(d)
] = E[IP] - E[D] + B
m·
= ~r~l---' +r- AE[Y]E[L] +B
mq
Also, a demand epoch results in an order with probability ~r~l Lq_,c}(i - I) = LO =
l/m q , so
OF=-~
mq
These formulas are not as simple as those for an (r, q) policy, but neither are they overly
complex. Notice., A is a weighted average over q values of GO(d - 1), where d ranges
from r + I to I' + q, but the weights are no longer equal. Likewise, B is an unequally
weighted average of q values of C l
Now, drop the assumption of Poisson D, and suppose that D is a world-driven de
mand process. It turns out that precisely the same results hold. The only difference is in
D. This has a compound distribution involving Yand the new 15.
For an (I; q) policy, and in particular a base-stock policy, the normal and maximal
approximatIOns of Section 6.4 still apply. Using the general formulas for compound dis
tributions of Section C.2.3.8, and letting.)? denote the variance/mean ratio for D,
v = E[D]
= E[D]E[Y]
6.7 Otheri
= AE[Y]L
,,2 = V[D]
2
= E[15]V[y] + V[15]E [Y]
= A(V[Y] + ;j;'E2[Y])L
2
The asymptotic variance-mean ratio for the overall demand process D is 41 = (J2/V =
(V[Y] + ;j;2E2[Y])/E[Y]. This is larger than;j;2, since E[Y] > 1. In the compound-Pois
son case, with;j;2 = I, ,,2 reduces to AE[y2 ]L. and Ij? = E[y2 ]/E[Y] > l.
Thus, ,,2
reflects the uncertainties in both the timing of demands (through ;j;2) and
the demand quantities (through V[Y]). As we have seen, the performance ofthe system
--
Chapter 6 Stochastic Demand: One (tern with Constant LeiJJrimes 231
depends largely on lJ. Thus, the additional demand variation due to uncertain demand
quantities does indeed degrade performance.
6.6.3 Optimization
Now, suppose we impose cost factors k, h, andp on the performance measures VF, 7 and
E, respectively. Denote the total average cost by C(r, q) or C(r, s), depending on the pol
icy class. We wish to minimize C within each class.
First, consider (/; q) policies. Here, B and 1 have the same form as in the Poisson
demand case. Nso, assuming each basic batch counts as an order, OF ~ AE[Y]/q too has
the same form as before. Therefore, C(r, q) is precisely the function in (6.5.1) (with
AE[Y] replacing \.) All the methods, formulas and bounds of Section 6.5 thus apply di
rectly. (The second interpretation of an order, comprising all basic batches ordered to
gether, requires more intricate methods; see Zheng and Chen [1992].)
Turning next to (r, s) policies, things are not quite so straightforward. It is possible to
construct a continuous approxinlation of C(r, s), but this does not help much. (The result
l·(i - I) = '0 = ing cost function is typicaIly 1/ot convex in the now continuous variables (r, s). Standard
nonlinear optimization algorithms thus cannot guarantee a globally optimal solution.) As
for the discrete formulation, there are indeed algorithms available to compute an optinla]
(r, s) policy. The best of these to date is that ofZheng and Federgruen [1991]. This proce
dure is sintilar to Algorithm Optimize_rq ofSection 6.5 and not much harder to implement.
r are they overly Some of the qualitative results of Section 6.5 can be extended directly to (r, s) poli
where d ranges cies: Since the best (r, s) policy costs no more than the best (r, q) policy, the upper bounds
is an unequally on C' in Theorem 6.5.1 remain valid. Also, it is straightforward to prove the lower bound
C* 2:: elF We conclude 1 therefore, that the impact of cr on overall performance is the
,orld-driven de same as before, as described in Section 6.5, regardless of which type of policy is used.
difference is in Useful approximations have been developed; see Ass'ad and Beckmann [1988], for
example. There are comparable results for the discrete-time context, discussed in Chap
al and maximal ter 9. Also, there is a substantial body of empirical work exploring the behavior of opti
compound dis mal policies. See, for example, Archibald and Silver [1978], Naddor [1975], Wagner et
tio for 'D, a1. [1965], andZheng [1991].
over all possible scenarios, and call this the expected present value for the chosen pol
icy. When the cash flows are expressed as costs, as they are here, call the result the ex
pected discounted cost.
In the stochastic context this is not the only possible criterion. (See Singhal [1988],
for instance.) It is the simplest one, however, and it is employed far more than any other.
We restrict attention to the model of Section 6.2.2: Demand is a Poisson process,
and k = 0, so we follow a base-stock policy. As in Section 3.7. a denotes the interest rate.
°
Also, for now, suppose the system starts at time t = with no inventory and no orders
outstanding.
We first evaluate the expected discounted order cost With k = 0 there remains only
the variable order cost, e. Interpret this as a purchase cost, and assume (reasonably) that
we pay for goods when we receive them, after the leadtime, not when the order is placed.
The policy orders s units at time t ~ 0, and we receive them at time t = L. so the dis
counted initial order cost is e-oLcs. From then on we order a llllit each time a demand
occurs. One can show that the expected discounted value of these continuing costs is
e-aLe(A!a).
Next, consider holding and hackorder costs. (The cost factors h and b include only
the real, physical costs!l and Q. not financing costs, as discussed in Section 3.7.) There
is nothing we can do to influence the initial costs during the time interval [0, L], sO ig
nore them. Second, assume the discounted expected costs of this type can be calculated
as follows: For each time t ~ L, determine the expected rate at which cost is incurred;
then, discount this quantity (multiply bye-a,) and integrate over ( (Technically. we are
interchanging an expectation with an integral.) Now, following the derivations in Sec
tion 6.2 and Section 6.5, this expected cost rate is precisely CCs) for t > L The overall
holding-backorder cost is thus rz e -a,C(s) dt = (e -aL!a)CCS).
We should also accollllt for delays in receiving revenues due to backorders, as in Prob
lem 3.27. But this involves complexities that we prefer to skip. (Section 9.4.7 works out
the details for the discrete-time case.) So, assume that each customer pays us at his demand
epoch, even if the demand is hacklogged. (The cost Q above can include a penalty that we
pay continuously to each customer whose demand is backlogged.) Letting p denote the
sales price, the total expected discollllted revenue is pA/a. a constant which we omit.
the chosen pol as in the average-cost model. Also, one can show that crCo(s') is approximately equal to
he resnlt the ex- the optimal average cost.
In sum, while the derivations involve different methods, the two criteria lead to
Singhal [1988], closely related results. We observed the same close cOllIlections in Section 3.7.
, than any other. This approach can be extended to more complex policies, demand processes, etc.
>Oisson process, See Problem 6.23, for instance.
the interest rate.
y and no orders
6.7.2 Using Current Information about Demand
re remains only With world-driven demand, a base-stock or (I; q) policy is unlikely to be optimal if we
reasonably) that actually see the world-system W as it evolves over time. Observing W(t) provides infor
order is placed. mation about future demands, which those policies ignore. We now suggest a new type
~ L. so the dis
ofpolicy which does use such information. (Under standard economic assumptious, this
time a demand form is optimal, as demonstrated in Section 9.7.) Exact calculations now become rather
tinning costs is complex (Section 9.7 includes those too), so we develop a relatively simple approxima
tion. This approach leads to an estimate ofthe value of observing W.
I b include only Suppose there are no scale economies in supply, so a base-stock policy would be at
lion 3.7.)There tractive ({we were unable to observe the driving process. Assuming we do observe it,
al [0, LJ, so ig here is a plausible control rule: Reset the base-stock level s whenever W(t) itself changes.
III be calculated That is, there is a vector s = (sw) offixed parameters, where w ranges over the state space
051 is incurred; ofW. At any time t where W(t) = w, follow a standard base-stock policy with base-stock
mically, we are level Sw' Call this a state-dependent base-stockpolicy.
yations in Sec This policy does what it can to keep IF(t) as close as possible to SW{I)' If SW(I) sud
. L. The overall denly increases so that SW(I) > IF(t-), the policy immediately orders, to bring IF(t) up to
sW(t). If sWU) decreases and sW(I) < IF(t-), it stops ordering until demand reduces IF(t)
lets, as in Prob to SW(t).
/.4.7 works out How might we evaluate such a policy? Note first, the familiar equation (6.2.7) still
5 at his demand applies, i.e.,IN(t + L) = IF(t) - D(t, t + L]. Here,D(t, t + L] depends on W(t). Specif
penalty that we ically, let (Dlw) denote a generic random variable with the distribution of D(t, t + L],
.g p denote the conditional on W(t) ~ w; this is also the distribution of D(L), conditional on W(O) = w.
b we omit. Then, IN(t + L) has the same distribution as IF(t) - (DIW(t)). Thus, the expected cost
rate at time t + L. viewed from time t, is given by CwuPF(t)), where
Cw(y) ~ E[C(y - (Dlw))]
and we again use the function Cry) ~ h[yJ+ + b[y]- defined in Section 6.5.
Ih W(O-) ~ O. Under a state-dependeut base-stock policy, we never observe IF(t) < SWU) (after or
by a constant. dering). Now let us approximate: Ignore the possibility that IF(t) > SW(I)' That is, treat
II.) IF(t) ~ s we'! for all t. The expected cost rate at time t + L then becomes CW(t)(SW(t»).
us approxlma Recall that ~ is the stationary vector ofW, so it specifies the fraction oftime W spends
in each of its states. In particular, ~. is the proportion of time during which we anticipate the
(approximate) cost rate C.(s.) a leadtime into the ftlture. TIlliS, the average expected cost
rate, denoted C(s). is given by C(s) ~ E[Cw (.5w)] = :l:. ~.C•.(s,,). This is in fact (by ergodic
arguments like those of Section 6.2) the overall average cos~ up to the approximation.
t. For small cr, Now, suppose we aim to minimize C(s) over s. But C(sl is separable in its variables.
he same value That is, to minimize it, we need only choose each Sw to optimize the corresponding
-..'
rameters.
Let us examine this ratio: Observe, cr~ is the variance of the leadtime demand as
suming we know W begins with W(O) ~ w. On the other hand, ,,' is the leadtirne-demand
variance presuming nO knowledge of the initial state (beyond ~ itself). Because cr2 in
cludes this additional source of uncertainty. we expect cr to be larger than the cr w' at least
on average. Indeed, decompose cr2 using the conditional-variance formula:
c? ~ V[D]
= EwV[Dlw] + VwE[Dlw]
= L w ~w~ + L w ~w(vw - vf
The two expressions in brackets are variances, in this case the variances of the compo
nents of (J and v, respectively. They are both nonnegative, so ,,' '" (~(J)' and ~(J/" :0; 1.
Chapter 6 Stochastic Demand: One Item with Constant LeadHmes 235
Jroblems can be (This decomposition of a 2 is entirely analogous to the analysis of variance in statistics.
tor of base-stock Here, the "dependent variable" is the leadtime demand, (f2 is its total variance, and (~ ...)2
lzes the expected plays the role of the residual variance.)
::quent costs. Recall, similar ratios played a prominent role in Section 6.5. (We used the notation
",I. It tends to be rr _/rr). There, however, we were interested in improving the physical systems generating
itive to the start supplies and demands, whereas here we are using information more effectively. It makes
perform poorly. intuitive sense that these two should be related, and indeed at this level of analysis we
.s we shall see in see no fundamental difference between them.
r its components For example, reconsider the two-state MMP process of Section 6.3. Recall, A ~ 10.
Suppose L = 0.1. The (exact) overallleadtime-demand parameters are v ~ 1, (f = 1.147.
f any policy, and The state-dependent parameters are v = (0.766,1.234)', ... = (1.013,1.222)'. Since ~ ~
L1: each time t we (0.5,0.5), we obtain~... = 1.118. Thus, ~"'/(f ~ 1.118/1.147 = 0.975.
xl by C Wu). Such In this case the cost saving is (at most) about 2.5%. This is rather small, because W
tends to change state quickly, so knowing the starting state provides only limited infor
ase-stock level), mation about the leadtime demand. ]f L were smaller, the relative savings would be
ost. Thus, C' larger.
nation about W, Recall, the results of Section 6.3 do not require that D be an MCDC process, and
pproximation is that is true here too: Again, think of Wet) as summarizing our current information con
value), but it is cerning future demand. This could be some continuous variable, for instance, instead of
a discrete one. In that case we reinterpret all the vectors above as functions of the vari
ifically, for each able w. In principle, we need to describe all possible w's (the state space ofW), together
late (Dlw). Also, with their relative likelihoods over time (~) and the corresponding leadtime-demand pa
cetion 6.5.2, the rameters (v and ...). In practice, we can use the whole panoply of forecasting tools to es
); a function only timate these quantities. In tillS spirit lV might be some specific demand-related factor,
r.{b + h)<1>(z*). and V w and a w simple functions of].1..; each with a few parameters to be estimated. With
these data we can readily calculate (f, as well as the average standard deviation (~ ...), and
then use them as above. Section 9.7 illustrates these ideas in the discrete-time context.
In this context, IN does not have a true limiting distribution, because of the cyclic
behavior induced by the order policy. It does have a long-run frequency distribution,
however, which we indicate by a random variable ['-[.
Suppose time 0 is a review point, and IP(O-) ,0; s, so [P(O) ~ s. For each time I un
til the next review point (0 ,0; I < u), equation (6.2.1) remains valid, i.e., IN(t + L) =
IP(I) - D(I, I + L], and IP(t) andD(I, 1+ L] are still independent. Moreover, IP(t) isjust
s - D(I), so IN(I + L) = s - D(I + L). Clearly, subsequent cycles follow the same pat
tern. To get the distribution of IN, therefore, we need only average the distribution of
IN(I + L) over t. Let 12 denote the uniform mixture of D(t + T,) over I in the interval
lO, u). Then,
IN=s-12
Notice, INhas the same form as IN in (6.2.3); only the leadtime-demand variable is dif
ferent. The performance measures thus take the same form as in a base-stock policy.
Use g(dlt), GO(d\t), etc. to describe the distribution of 0(1), ie., the Poisson distribu
tion with mean AI. Straightforward calculus leads to an expression for &2' the pmf of 12:
gD(d)
-
~ (!)I~
U
g(d]I + L) dt
= (:JGO(dIL + u) - G"(dIL)l
Consequently,
I '
G12o(d) = ( -)lG1(d + 1(L + u) - Gtcd + IlL)]
AU
I .
1
G12 (d) ~ ( AJd'(d\L + u) - G\dIL)]
7 = S - A(L + .~u) + B
since ElI2l = A(L + )1u). And, of course, OF = lIu. These performance criteria, along
whh the cost factors, can be used to formulate an optimization model, along the lines of
Section 6.5.
The following approximations are widely used:
A'l = (
IAU
~)Gl(SIL + Il)
...
I I )
B+l = \Au G'(sIL + Il)
-
Chapter 6 Stochastic Denumd: One Item with Constant Leadtimes 237
wse of the cyclic Like those of Section 6.4.4, these approximations omit negative tenus from the exact
ncy distribution, values. And, like tbose of Section 6.4.4, tbey are not reliable.
Suppose for tbe moment tbat u is fixed, and we set s optimally. As in the base-stock
1£ each time t nn model, tbe optimal safety stock and cost are determined mainly by [[, the standard devi
Le., IN(t + L) = ation of 12. It is easy to show tbat
over,IP(t) is just
IJW tbe same pal
e distribution of
i" = V[l2] = U + (~)u + (~~)u2
I in tbe interval
This expression reveals the impact on performance of imposing a fixed review period.
For small u, tbe last term is negligible, so the effect is similar to an increase in tbe lead
time from L to L + ;.fu. For large u the effect is greater; fI.2 becomes nearly quadratic in
u, so fI. becomes linear.
d variable is dif
To apply tbe normal approximation (for any world-driven demand process), treat
-stock policy.
D(t) as normal witb mean v, = At and variance cr; ~ All?!. Setting z, ~ (s - v,)!cr, and
I'oisson distnbu
'. tbe pmf of 12:
z;= (s + vJ/ITtgives
A = (~)Ji+u <t>o(z,) dt
= (L)[F'(SIL + u) - F1(sIL)]
2
= (L)[F\SIL + u) - F (sIL)]
7 ~ S - A(L+ Xu) +B
where
Notes
[1959] and Hadley and Whitin [1963]. Concerning the PASTA property of Poisson Cansidol
processes, see Wolff [1982]. Theorem 6.2.4 appears in Svoronos and Zipkin [1991]. 4'mandpno
Section 6.3: The MCDC process of Section 6.3 is a special case of the point proeess in loreIs. .~
troduced by Neuts [1979]. Song and Zipkin [19933, 1996c] discuss various applications. ....r"..._
The main results (for various eases) are attributable to Sabin [1979,1983], Sivazlian Jew'i!!lllj' L'"
[1974], Stidham [1974], Zipkin [1986a], and Browne and Zipkin [1991]. L! This~
Section 6.4: The normal approximation has been in common use for decades; one early ~
exposition is Whitin [1953]. The maximal approximation is attributable to Scarf[1958a]; .,...m." ..,
see also Gallego and Moon [1993]. For more on the inaccuracies of standard perform ua:gus
anee approximations, see Herron [1978] and Zipkin [1986b].
Section 6.5: Most of the material on base-stock policies is adapted from well-known re 6.3
sults on the mathematically equivalent news-vendor problem of Chapter 9. As for (r, q)
policies, Zipkin [1986b] and Zhang [1998] demonstrate the convexity of C(r, q). Theo
rem 6.5.1 and the other results of Section 6.5.3 are attributable to Zheng [1992] and Gal
lego [1998] (which also include more refined results). Algorithm Optimize_rq is attrib
utable to Federgruen and Zheng [1992b]. (Some of the ideas were anticipated by Sahin
[1982].) For the stockout-constraint model ofSeetion 6.5.5, Platt et al. [1997] discuss an
alternative heuristic. 6.4
Section 6.6: The results on (r, q) policies are primarily by Richards [1975] and Zheng
and Chen [1992]. The approach here for (r, s) policies follows Iglehart [1963a] and
6.5
Veinott and Wagner [1965].
Section 6.7: The notion ofa myopie policy (Section 6.7.2) has uses in many areas. (It ap
pears in Chapter 4 and again in Chapter 9 in the context of time-varying parameters.)
Key references include Veinott [1965] and Heyman and Sobel [1984]. The use of it here
6.6
to estimate the value of information appears to be new.
For the periodic-review model of Section 6.7.3, Rao [1994] shows that the av 6.7
erage cost is convex in (u, s) and derives results similar to Theorem 6.5.1. Johnson et al.
[1995] test a variety of approximations; they find that A + I is unreliable.
6.8
Problems
6.1 Consider the model of Section 6.2.2: Demand is a Poisson process, and we follow a base-stock
policy. OUf customers, however, inform us in advance of their demands by the constant time LD ; in
other words, customers order units from us for later delivery. (They will not accept early deliveries.)
useJ~
Thus, the customer orders themselves form a Poisson proeess, and the actual demands describe the
same process shifted forward in time by LD.
First, assume LD ~ 1. Argue that, in this case, we can operate a perfect flow system, meeting
Now, assume L D < L. Let us slightly alter the meaning of a base-stock policy: Supposing we .......l
start with 1(0) ~ s, order a unit whenever a customer order occurs (not an actual demand). Let
Set) denote the cumulative supply, the number of units received through time t. Clearly, IN(t) ~
s + Set) - D(t). ~.,;j
Chapter 6 Stochastic Demand: Dlle Item with COilsfanl Leadtimes 239
rty of Poissoo Consider a conventional system (i.e., with no LD ) with leadtime L - LD , facing the same
<in [199IJ. demand process as the original system, and using a base-stock policy with the same base-stock
int process in level s. Argue that 8(t) and hence IN(t) is the same in the two systems for all t. (Conclude that the
s applications. performance of the original system is identical to that of a standard system with the shorter
83J, Sivazlian leadtime L - L D - The infonnation our customers provide us here is thus valuable indeed.)
6.2 This problem explains the behavior observed in Figure 6.2.6. (Increasing q essentially increases
des; one early backorders.) Supposefis any convex function on the integers. Fix an integer s'. For each odd
'icarf[1958aJ: positive integer q letF(q) denote the average of the quantitiesj(s), taking s over q consecutive
:Ian1 perform integers centered at s'. (For example, for q ~ 3, use s ~ s' - I, s', and s' + 1.) Show that F(q) is
increasing as a function of q.
<oil-known re 6.3 Suppose demand forms a MMP process. The driving process W has three states, which we label
I. As for (r, q)
low, rnedium and high. The corresponding demand rates are A{ < Am < AIr From states I and h
err, q). The<> transitions can occur only to state m; from m W changes to I with probability 113 and to h with
il92J and Gal probability 2/3. Each time W enters a state, it remains there for a random time; the mean times are
e_rq is attnl> 12 weeks for state I, 18 weeks for m, and 19 weeks for h.
ned by Sabin Write down the matrices Q and A, and compute the stationary vector ~. Assuming the demand rates are
17] discuss an Al ~ 35 units/week, Am = 50 units/week, and Ah = 75 units/week, what is the overall mean demand rate A?
6.4 Prove that the assumption of Section 6.3.3 is satisfied when D is a MMP process. (Careful: The
i] and Zheng
demand rates Aww in some states may be O. Some of them must be positive, however, because A is.)
[1963aJ and
6.5 Verify directly the following assertion in the proof of Lemma 6.3.1: Suppose fPc is a random
areas. (It ap variable distributed uniformly on the integers mod (q), and suppose D is an integer-valued random
parameters.) variable, independent of fPc- Then, the random variable [IPc + DJ mod (q) is also uniformly
ttse of it here
distributed on the integers mod (q).
6.6 Verify formulas (6.2.14) and (6.2.15) for the Poisson loss functions. (Prove them by induction.)
; that the av 6.7 Consider the model of Section 6.2, but assume the leadtime L = O. Assume we wish to preclude
>Imson et aI. backorders entirely. What is the smallest value of r we can choose, and why? Assume we do choose
r in this way. Argue that I ~ X(q - I). Ifwe face an order cost k and a holding cost h, argne that
the optimal q can be determined as in Problem 3.8. (In this case, then, the behavior ofthe system
and the optimal policy depend on A, but otherwise the specification of the demand process has no
effect whatsoever.)
6.8 Verify equations (6.4.2) and (6.4.3) describing the standard normal loss function.
6.9 Verify equation (6.4.7) describing the sensitivities of the normal approximations of I andE to 0'.
6.10 Assume that D has mean v and variance (T2, but otherwise we know nothing about its distribution.
-stock
We can write
meLD; in
deliveries.) E ~ E[[D - sJ+J ~ E[X{ID - sl + (D - s)}]
5Cnl>e the
Use Jensen's inequality to argue that
meeting
E[ID - sl] :s {E[(D - S)2]} 1/2
6.11 Suppose we use continuous approximations for D and IF. Argue that 'Q'.~
Ibl~1
B(r, q) = E[[D - (I' + qU)t]
~l_~
where U is a random variable uniform on [0, I], independent of D Then, show that B(r, q) is a ". Suppl
convex function of the continuous variables (r, q). (Since I(r, q) is justB(I; q) plus a linear function ~;
of (I; q), it follows immediately that I too is convex.)
6.12 Verify that the average stocking time IW is given by IIA, following the argument in Section 6.2.4.
6.13 Use equation (6.2.16), relating the transforms ofBand BW,to prove (6.2.17), relating their variances. ~c~
6.14 Verify the expression for XE[B(B - I)] in equation (6.2.18).
6.15 Verify the formulas in Section 6.5.2.3 for z* and C* using the maximal approximation.
6.16 Within the proof of Theorem 6.5.1 there is a geometric argument that 0:0; H'(q) :0; hw. Here, we
substantiate this result more rigorously.
Let r(q) denote the optimal value of I' for fixed q, that is, the r which satisfies (6.5.13). By the
chain rule, H(q) ~ C'(r(q))r'(q). Now, differentiate (6.5.13) with respect to q, and use the implicit IbIS
function theorem, to obtain an expression for r'(q). Then, use the inequalities -b :0; C'(r(q)) :0; 0 m.
and 0 :0; C'(r(q) + q):o; h to argue that 0:0; H(q) :0; hw.
6.17 For the continuous (r, q) model of Section 6.5.3, show that the optimality condition (6.5.13) is
equivalent to A(r, q) ~ I - w. Thus, as in the base-stock model, the optimal policy sets the
stockout frequency to I - w. leI
6.18 Consider the problem of finding an optimal (r, q) policy, using the normal approximation of D. We
investigate what happens to (1'*, q*) and C* as the demand parameters change, keeping the cost
factors fixed. Suppose that (T2 depends linearly on A, specifically, (T2 = AlAI for some constant AI'
Allow v to change independently of A and ~.
Let (rj, ql) denote the optimal policy and C, the optimal cost for A ~ Al (so (T2 ~ I) and v = O.
Show that, for all values of A and v,
*'u. SeI
6.19 This problem asks you to verify the inequality (6.5.10), showing that the optimal cost in the
ca be
continuous (r, q) model is insensitive to suboptimal choices of q.
(a) Use the inequality H(q) ". hwq to show that, for any fixed q, wH(q) - H(wq) is nondecreasing
as a function of tv, for w > O.
(b) Use part (a) to show that, for any fixed q and x > 0, I;q H(y) ,~v :0; X(x - I )qH(q). Hint:
2
-II
q*. Suppose instead we start with a large q, and a test shows that q > q*' Modify the algorithm to
linear function decrease q in increments of 1 until q* is found.
6.22 Section 6.6.2 gives a fonnula for the limiting probability vector L of the process IP, under an (/; s)
Section 6.2.4.
policy. Verify this fonnula.
Jeu variances. 6.23 Consider the discounted-cost model of Section 6.7.1. This problem asks you to derive the
corresponding cost function for an (r, q) policy with fixed order cost k > O.
OIL
(a) Choose any t "" L, and suppose IP(t) = i. Let Tbe the time until the next demand after t, and CT(i)
rw. Here, we the expected discounted holding and backorder cost in the time interval [t + L, t + T + L), as
viewed from time t. Argue that CT(O ~ (l - ,,/)(e-aL/a)C(i), where C(i) is the average cost of a
5.13). By the base-stock policy, as before, and "/ = )"/(),, + a).
ISethe implicit (b) Now let C"(i) be the total discounted expected cost over all time, starting with IP(O) ~ i. Argue
C(r(q»,,; 0 that these quantities satisfy the equations
COO(i) ~ [k + e-aLe(r + q - ill + COO(r + q) i~ r
6.5.13) is
ets the COO(i) ~ CT(i) + ,,/COO(i - 1) r<i:5.r+q
(c) Solve the equations for r ::; i ::; r + q to obtain
ltion of D. We
k + e-aLeq) + ;;;q_ ~q-jCT(r +J')
ogthe cost COO(r) ~ ( ,-I ,
e constant A. I - 1- ,,/q
Compare this result with formula (6.5.16) for the average cost. Verify that this agrees with the
and C* are
in the
• cost of a base-stock policy derived in the text, assuming k = O.
6.24 Suppose the demand process D is a Brownian motion with positive drift rate A. and variance rate
Ij,z)". (So D(t) has a normal distribution with mean )"t and variance Ij,z)"t.) Let v ~ U and a 2 =
Ij,zU, as above. We follow an (/; q) policy.
Set i3 = 2/1j,z, and for any fixed number a define the function h(xla) = I - e -~«-a), x"" a. It
can be shown that IP has the pdf
m.decreasing
G)[h(Xlr) - h(xlr + q)] x~r+q
). Hint: f(x) =
G)h(X1r) r:sx:Sr+q
(IP can go above r + q, because D can have negative increments.) Define the functions
I = j3 + 'f.q + r - v +B
(Thus, the performance measures have the same form as before. So, the methods of
(e) Express the functions F I and F 2 in terms of <1>0, <1>1, and <1>2 What role does IT play in these
fonnulas?
1
7.1 In~
c H A p T E R
7 STOCHASTIC LEADTIMES:
THE STRUCTURE OF THE
SUPPLY SYSTEM
OOsof
IT play in these
Outline
7.1 Introduction 243 7.4 Exogenous, Sequential Supply
7.2 Independent, Stochastic Leadtimes Systems 273
246 7.5 Leadtime-Demand Distributions 279
7.3 Limited-Capacity Supply Systems Notes 289
256 Problems 290
7.1 Introduction
7.1.1 Discussion
The previous chapter examined several stochastic demand processes, but only the sim
plest type of supply process, where every order leadtime is the constant L. In reality,
leadtimes are rarely constant; unpredictable events in the supply system cause unpre
dictable delays.
The impact ofleadtime uncertainty, it turns out, depends on the structure ofthe sup
ply system and how it operates. In certain special cases, surprisingly, leadtime uncer
tainty has essentially no effect and can be ignored Most often, however, leadtime fluc
tuations strongly degrade perfonnance, just as demand uncertainty does. Furthennore,
when the capacity of the supply system is limited, demand uncertainty induces longer
and more variable leadtimes; in other words, limited capacity magnifies the impact of
demand uncertainty.
243
•
•
Unpredictable events in the supply system can also generate quality defects. We
learned in Chapter 3 that imperfect quality can hurt performance even in a completely
deterministic system. Unpredictable qualIty has even more pronounced effects, espe
cially in conjunction with stochastic demands and leadtimes. In limited-capacity sys
tems, moreover, defects waste capacity, thus amplifying the effects of the other sources
of uncertainty.
Even with these additional complicating features, the performance measures in
many cases have the same form as in Chapter 6. Consequently, the optimization meth
ods of Section 6.5 can be applied once again. Also, the impact ofleadtime uncertainty
and imperfect quality, as well as demand uncertainty, can be captured by tbe summary
measure IT, suitably redefined.
7.1.2 Taxonomy
A supply system consists of production and transportation operations. We use the
generic term processors to describe the individual operations. The system may be sim
ple or complex; it may be part of our organization, or a different firm.
We focus on three distinct supply-system structures. These emphasize different
characteristics of real systems. All are abstractions to some degree; real supply systems
are often hybrids of these types, and include additional eomplicating features. To make
sense of more complex systems, however, it is essential to understand these fundamen
tal structures thoroughly.
Section 7.2 explores parallel processing systems: There is an infinite number of
identical processors, which work independently. Each of our orders is handled by a dif
ferent processor. The time a processor requires to process an order is, in general, sto
chastic. So, the order leadtimes are independent, identically distributed (i.i.d.) random
variables. Clearly, such a system has no capacity limits. The constant-Ieadtime system
of Chapter 6 is the special case with deterministie processing times.
It turns out that leadtime uncertainty has relatively little impact in a parallel supply
system. Farone special case (Poisson demand, a base-stock policy), indeed, performance
is entirely insensitive to the leadtime distribution; the system is essentially equivalent to
one with constant leadtimes. This is not SO in general; for world-driven demand and/or
an (1; q) policy, leadtime uncertainty does have an effect. Even ~O, this effect is not as
strong as in other stnlctures.
Section 7.3 treats limited-capacity supply systems. The simplest has a single proces
sor, capable of working on one order at a time. More complex systems have several (but
still finitely many) processors arranged in series, or in parallel, or in more intricate net
works.
Limited capacity in conjunction with stochastic demand leads to congestion, and
this in turn both increases order leadtimes and makes them less predictable. These con
gestion phenomena have dramatic effects on system behavior and. performance, alto
gether missing in a pure parallel system. This impact is strongest and clearest in a se
quential system, which preserves the sequence that orders arrive in. This category
includes single-processor and series systems.
•_~=-_._-_._-----------
---------- --
Chapter 7 Stochastic Leadtimes: The Structure of the Supply System 245
y defects. We Section 7.4 introduces exogenous, sequential supply systems. Such a system is ef
I a completely fectively uncapacitated, but for a different reason than a parallel system: Our orders con
effects, espe stitute a negligible fraction of the system's total workload. Such systems share many of
-capacity sys the characteristics of capacitated sequential systems; in particular, leadtime uncertainty
Olher sources plays a major role in performance. (A constant-leadtime system is also a special case of
this structure.) Section 7.5 focuses on computing performance measures for certain im
~ measures m
portant cases.
lization meth All these systems can be regarded as queueing systems. (We prefer the phrase pro
Ie uncertainty cessing systems. Processing is the function and purpose of these systems, while queues
the summary are unfortunate characteristics of them.) If you have studied the theory of queues, some
will be familiar, others less so.
Processing systems by themselves are commonly used to represent the delivery ofserv
ices. In the production-transportation context, a pure processing system models a service
like product. The outputs are typically customized; each unit has specific features ordered
We use the by the customer. In any case, all processing activities are triggered by customer demands,
maybe sim and each customer waits until his order is completed. TIus is called a make-la-order system.
Here, the processing system is part of a larger make-ta-stock system, which can cre
iize different ate goods and store them in anticipation of demand, in order to reduce customer waiting.
pply systems Processing is triggered by orders, not demands. Clearly, this makes sense only if the
res. To make goods are uniform, or nearly so. (A make-to-order system may have an inventory of units
., fundamen awaiting processing, but these are inputs, not outputs.)
:e number of
Ded by a dif 7.1.3 Preview ofthe Analysis
general, 5tO For the moment, forget about the structure of the supply system; think of it as a black
i.d.) random box. Each ofour orders enters this black box; sometime later the order emerges, at which
hime system point we receive the goods ordered.
Consider a base-stock policy, and assume IN(O) = 1(0) = IP(O) = s ;0. O. As in Sec
rallel supply tion 6.2.2, the process describing our orders is identical to the demand process itself.
perfonnance Thus, equation (6.2.1) continues to apply:
:quivalent to
IN(t) ~ s - IO(t)
mand and/or
ect is not as Also, IO(t) is precisely the number of orders in the supply system at time t, that is, its
current occupancy. This is entirely independent of s; indeed, with s = 0, we have a pure
ngle proces (make-to-order) processing system.
several (but Now, consider some particular supply system. The input to the system is a copy of
ntricate net the demand process. Suppose that the system with this input is sufficiently well behaved
that it is ergodic, its equilibrium occupancy 10 exists, and we can explicitly calculate the
gestion, and distribution of 10. Then, just as in (6.2.2),
. These con
IN~s-IO
Dance, alto
rest in a se From this we can calculate all the relevant performance measures, as in Section 6.2.2.
lis category Indeed, if g denotes the pmfof 10, GO its ccdf, and G I its loss function, the formulas for
fl, B, and I are virtually identical to (6.2.4) to (6.2.6):
246 Foundations a/Inventory MUllagement
A ~ GO(s - I) li = GI(s)
]= S - E[IO] +li
And, of course, OF = A.
Consequently, many ofthe qualitative results ofChapter 6 remain valid. For instance,
Jj is decreasing and convex in s, while Tis increasing and convex, In a cost-minimization
setting, therefore, the optimality condition (6.5.4) remains valid, and the optimal base
stock level essentially sets the backorder probability GO(s) to the critical ratio I - w.
To understand the overall inventory system, therefore, we need to understand the
supply system, no more and no less. For this purpose, we can and do apply the results of
. queueing theory. However, we need to understand the system quite thoroughly; the mean
occupancy E[10] is not enough. As the results of Chapter 6 suggest, we need either the
full distribution of10, or perhaps an approximation based on V[10] as well as E[10].
Next. consider an (r, q) policy with q > 1. Conceptually at least, the same basic idea
applies. However, the input to the supply system is no longer a copy of the demand
process, but rather a filtered version ofit. Consequently, the calculation of10 can be far
more difficult. We shall see some cases where 10 can be determined exactly and others
where it can be approximated accurately.
This is thc analytical approach used in most of the next two sections. Section 7.4,
however, uses a different approach, more reminiscent of Section 6.2.3. There the focus
shifts to the time an order spends in the supply system, instead of the occupancy.
A personal-computer repair service provides "loaners" to its customers. That is, when a customer
brings in a broken computer, the company lends the customer a rcplacement to use until the ma
There is an in chine is fixed. The company provides excellent services, and its customers will go nowhere else,
Iby a different even when no loaner is available. Still, the company views the loaner program as an integral part
essed simulta of its overall servicc package, and so aims to keep enough loaners available that 95l!/~ of customers
rder is, in gen can get one.
ributed (i.i.d.) The company employs a large number oftechnicians, each fully qualified to diagnose and re
mon distribu pair virtually any computer problem. Each arriving job is assigned to one technician, who handles
the cntire job. Rarely does ajob havc to wait for an available technician. (When the workload gets
1 infinite nUffi
unusually large, the company can expand its capacity by calling in regular technicians from days
nd the . means off or vacation, or occasionalIy part-time freelancers.)
Customer arrivals form a Poisson process. Ou average, 16 jobs arrive per day, and the aver
lbe model. the
age job requires 1.25 days to complete. There is a lot of variation among jobs. Many can be fin
o the demand ished within hours, but a few require weeks. The company's owner wonders whether this variation
should affect the number of loaners it maintains.
Ie model is an Since each job brings its own characteristic problems, it is reasonable to assume that the lead
'illite, parallel times (job-completion times) are Li.d. Therefore, variation makes no difference. Given A. = 16 and
,a fairly accu E[L] ~ 1.25,10 has the Poisson distribution with mean (16)(1.25) ~ 20. The 95% availahility tar
=So get means A ::;; 1 - w_ = 0.05.
•
s A Jj I
21 0.44 134 2.34
So, to meet the performance target, the company should set s = 28. With this policy, the average
inventory of loaners will be about 8. The normal approximation prescribes
s = 27.36. Rounding this up to 28, the approximation predicts A = 0.037 and! = 8.07, quite close
to the exact values.
By the way, while leadtime variation itself does not affect the result, the company can stilt
gain by reducing the times of the longest jobs, for that will reduce E[L].
The calculations above, and those of the other examples in this chapter, were performed in a
simple spreadsheet, using the formulas in the tcxt You might want to set up your own spreadsheet
to verify the results and cxplore the effects of parameter changes.
Here, til is the asymptotic variance-to-mean ratio of the demand process, and L[l] indi
cates a random variable, the minimum of two independent copies of L. That is, L[2] =
min {L~ L 1 }, where L 1 and L 1 are ij.d. random variables, each with the same distribu
tion as L. We shall say more about L[2] in a moment.
For a Poisson process, of course, til = 1, so (f2 reduces to the correct value, AE[L].
For 1Ji2 ,.. I, think of ,,2 the base value )Ji[L] plus a correction factor. The sign of this
correction factor depends on (1Ji2 - I), which is a characteristic of the demand process;
it does not depend on L. Thus, according to the approximation, 10 is more variable than
in the Poisson case «(fl > v), precisely when demand is more variable than a Poisson
process. The magnitude of the correction factor depends on the deviation of 1Ji2 from I,
but also on E[L[2]]' which reflects the distribution of L.
For constant L, E[L[2]] ~ E[L] ~ L, so ,,2 ~ 1Ji2v,. The approximation is thus con
sistent with that of Section 6.4. In general, 0 < E[L[21] s E[L], and it is not hard to show
that
E[L] - E[L[2]] = »E[ ILl - L 21 ]
Thus, E[L] - E[L[2]] is a measure of the variation of L. If the distribution of L is spread
out widely, this quantity is large. For example, if L has a unifonn distribution on the in
terval E[L] ± u, for some constant U with 0 < u s E[L], thenE[L] - E[L[2]] = u13. (See
Problem 7.1.)
;Y, the average The case 1Ji2 > I leads to a seemingly paradoxical result: Greater variability in L ac
o prescribes
07, quite close
tually reduces ,,2 In general, ,,2
always lies between )Ji[L] and 1Ji2)Ji[L] , as if the lead
time were fixed but the demand process were mixed with a Poisson process. This is the
essential effect of stochastic, parallel processing.
lp3D.y can still
The (approximate) optimal base-stock level s* can again be determined as in Sec
monned in a tion 6.5.2.2. Also, the optimal cost is (approximately) proportional to" and thus to the
spreadsheet
iD. square root of A...
Figure 7.2.1 illustrates the accuracy ofthe approximation. Demand is the two-state re
newal process of Section 6.3.2, so}. ~ 10 and 1Ji2 ~ ». The leadtimes have an exponential
distribution, in which case E[L[2]] = »E[L]. The figure shows two values of E[L], »and I.
there was the Evidently, the approximation works well in this case, even though )Ji[L] is rather small.
ess, for a par-
FIGURE 7.2.1
Performance ofBase-Slock Policies (Normal versus Exact).
5 5
E[L] ~ 0.5 Normal
--- E[L] = 1.0 Normal
- E[L] = 0.5 Exact
4 - ElL] = 1.0 Exact 4
j:J \ A / [:1
o o
o 5 10 15
Base-stock level
The equilibrium behavior ofthis system too is quite simple: Letting v ~ AE[L], the
pmf of10 is given by
(vdldL
glO(d) = I'_ (vii}!)
J-D
g(d)
d= O, ... ,s
G(s)
where g denotes the Poisson pmf with mean v and G the corresponding cdf. So, glO is
just a truncated version ofg. Again, 10 depends on the distribution of L only through its
mean.
The stockout probability is A = gIO(S), so the rate at which lost sales occur is
'AgIO(s), and OF = 'A(l - A) = 'A[l - gIO(S)]. The average inventory is
1 ~ Ij=o (s - 1)glO(J) ~ S - v + vglO(s)
These criteria have fundamentally different forms than in the backorders model.
Still, A is decreasing and convex in s, and lis increasing and convex. (See Problem 7.2.)
So, given a service-level target or cost factors on .Ii and 1, it is not hard to determine the
optimal base-stock level.
EXAMPLE 7.2.B, PART 2
Suppose that next door to the classical record store is a popular CD outlet. One of its venerable
Beatles albums is requested consistently at the rate of3.6 per mouth (the same as Beethoven's Sev~
_ _IL- . .....;.:-
enth!), according to a Poisson process. The mean leadtime is 2 weeks (also the same). The store
aims to fill about 80% of demands directly from stock. Customers who cannot find the album in
stock, however, walk out angrily and purchase it at a competing store across town.
Again, the average leadtime demand is n = (3.6)(0.5) = 1.8. Direct calculation yields the per
formance of alternative base-stock policies:
s A 7
0 1.00 0.00
0.64 0.36
2 0.37 0.86
3 0.18 1.52
4 0.08 2.34
5 0.03 3.25
6 0.01 4.21
(Comparing this to the table in Example 6.2.A, Part 1, for each value ofs, A is smaller and-j larger
than in the backlog case.) With a performance target of A :::;; 1 - w_ = 0.2, the store should set
s = 3. (Recall, in the backlog case, the sameperlormance target requireds = 4.) This policy leads
to an average inventory of 1.52.
By the way, you might question the 80% service target. A lost sale usually (very likely in this
case) has a greater economic impact than a backorder. So, perhaps the popular store should aim
= "E[L], the
higher, say at 95%.
First, focus on the supply system: It treats each order as a distinct unit. LetIO(t) =
IO(t)/q denote the supply system's occupancy in these units, i.e., the ml111ber of orders
outstanding. An order is placed every q demands, so the input to the supply system is a
renewal process, where each interdemand time has an Erlang-q distribution. This is an
MCDC process; IP plays the role ofW. Clearly, the input rate is:i. = A/q, and it is not
hard to show that the variance-mean ratio is ~2 ~ I/q. So, as in Section 7.2.2, for large
A, the limiting occupancy TO is approximately normal, with parameters
£[10] ~ :I.£[L]
nit. LetlO(t) = The same ideas can be applied when demand itself is a world-driven process.
rmber of orders Here, ~2 ~ '1?!q, where 1jJ2 is the original variance-mean ratio of the demand process.
'Ply system is a A bit of algebra shows that (7.2.2) becomes
rion. This is an (T2 = AE[L] + A(1jJ2 - I)E[L[2]] + A(q - I)(E[L] - E[L[2]D
q. and it is not
7.2.2, for large [It is not entirely clear what the analogue of the refined fonmula (7.2.3) should be.]
Incidentally, it is not hard to extend this approach to situations where the order size
does influence the leadtime: Just specifY L, and hence E[L] and E[L[2]]' as functions of
q, and apply the results above.
Call L the nominal leadtime, the time required for the supply system to deliver a
unit, regardless ofquality. Define L I to be the time needed to deliver a nondefective unit,
the actual effective leadtime. The effective leadtimes for different orders are indepen
dent, so the system operates just like one with perfect quality, but with L' replacing L.
For simplicity, suppose also that the nominalleadtimes within an effective leadtime
are independent. The random variable L' thus has a compound distribution, based on that
of L. But, by the results of Section 7.2.1, only the mean counts, and this is given by
E[L'] ~ E[L]
~
where ~ ~ I - o. The effect of imperfect quality, then, is an inflation of the mean ef
fective leadtime. Under the TIonnal or maximal approximation, (J" grows by a factor of
I/~, and performance deteriorates accordingly.
Imperfect quality also affects the cost parameters, depending on the payment
arrangements. The required adjustments are the same as in Section 3.6. (It is possible,
but much harder, to analyze the second inspection scheme of Section 3.6, where defects
are discovered only when units are given to customers in response to their demands. Sec
tion 7.3.5.3 considers this scheme in the context of a different supply system.)
Finally, assume a constant leadtirneL, and consider a general (r, q) policy. Here, an ar
riving batch may include both defective and nondefective units. Thus, the effects of poor
quality cannot be captured exactly by an inflated leadtirne, as above. That approach seems
to work fairly well as an approximation, however (see Moinzadeh and Lee [1987]): That is,
replace L by the constant L' = LI~, and then apply the calculations of Section 6.2.3.
;tern to deliver a fixed probability, and exits the supply system (arrives in our inventory) with some other
ondefective unit, probability. So, anyone order might encOlmter the same node several times, and other
ers are indepen nodes not at all. Assume that every order does exit eventually. (In queueing-theory par
L' replacing L. lance, a network with this property is called open. There are other models called closed
Ifective leadtime networks, which we do not consider here.)
}n, based on that Specifically, let rOi be the probability that the initial node isj, rij the probability of
; is given by moving from i to j, and riO the probability of exiting the system from i. These are called
routing probabilities. The movements of orders in the network are independent of the
processing times, and each such event is independent of the others.
These routing probabilities can be interpreted as quality defect rates. Suppose that
of the mean ef the normal processing sequence, when all goes well, is nodes 1,2, ... , J, as in the se
o by a factor of ries system above. That is, when processing is successful, an order moves from ito i +
I for i < J, and from J it leaves the network. However, node i may induce a defect. In
n the payment
that case, the order may have to be reworked at node i (with probability r ii), or worse, re
. (It is possible,
turn to an earlier node (with probability rij' j < i). If the defect is serious enough, we
i, where defects
may have to scrap the order and start a new one, which means returning to node 1. The
.demands. Sec
imperfect-quality model of Section 7.2.6 is the special case with J = I and & ~ rll' (In
stern.)
terms of network structure, this is called a series system with feedback.)
icy. Here, an ar
Again, assume Poisson demand and a base-stock policy. Beginning orders arrive at
effects of poor
nodej at rate ArOf Let ~ denote the row-vector (Aro;);. Let R denote the rOl/ting matrix
'PProach seems
(rij)'j.,<o, Also, let ~ be the total arrival rate oforders to nodej, and A the row-vector (~)i'
1987]): 'That is,
(We don't know these yet.) Thus, orders move from ito j at rate A,rij' For the A; to be de
JIJ. 6.2.3.
fined consistently, then, requires 'A.;" = Aroj + ~i Airij' or in matrix notation,
A=AO+AR
Now, assume that I - R is a nonsingular matrix. This condition, in fact, ensures that each
ICh a parallel
order leaves the network eventually. Then, the equations above have the unique solution
:lCh order must
nged in series,
A=~(I-R)-I (7.2.4)
"'Pter 5, how
(In the series system above, each ~ = A.) The '-i are always nonnegative.
tories of semi
It turns out that each [OJ has the Poisson distribution with mean AiE[Lj ] , just as if
ntemal inven
node j were an MlG/oo system in isolation, and the ID.i are independent. Therefore, the
ese times are
total occupancy [0 has the Poisson distribution with mean v ~ kj AiE[LJ The per
fonnance of the system is thus identical to that of a constant-Ieadtime system with aver
Ii order's lead
age leadtime demand v. (In the one-node, imperfect-quality model, [ - R ~ (I - &) ~
L reduces to a
(s) and rOI ~ I, so Al = A/s, and v = AE[L]/s, consistent with the earlier result.)
Here is another way to express this result: Assume that each Ai ~ A(as is always true
i and a base
for a series system with feedback), and let Si ~ }JAi' Define the mean effective leadtime
,has the Pois
E[L;] ~ E[Lj]/si, and set E[L'] = :£.i E[L;]. Then, v = AE[L']. Thus, v can be interpreted
Klent. (This is
in terms of the total input rates '-i, or equivalently the mean effective leadtimes E[L;].
'oisson distri The effects of quality on performance depend on the structure of R in an intricate
way. Roughly, as the defect rates increase, the '-i (or E[L;]) become larger, hence per
rocessing se
fonnance deteriorates. Also, the impact ofr ji is greater for larger i and smallerj. A seri
node. Subse
ous defect sends an order backward a long way, and that requires repeating the opera
~j with some
tions at all nodes in between. That is, r ii affects all the Ak' for j s k s i.
256 Foundations oflnvenrOlY Management
stable. The code for this processing system is MlMII. (Both Ms stand for Markovian; the
second one refers to the assumptions above about processing.)
Markovian processing means a great deal of variation. The exponential distribu
tion puts substantial probability on very long times relative to its mean; its variance
rners. The theory is (l/fL)2, the mean squared. It describes, for instance, the lengths of telephone calls,
serve as the sup
the run times of computer programs, and complex equipment repairs (as in Example
lilly a few impor
7.2.A). Most production processes are less variable. A processor that breaks down or
its main results. produces defects frequently, however, may exhibit nearly Markovian behavior (see
nd variation to
Section 7.3.5).
ss work at some In this case (see Section C.5.4), the equilibrium occupancy IOhas the geometric dis
During a period
tribution with parameter p:
shortage, and a
e to work off the g(i)~(I_p)pi i?-O
ion. Variation in
Consequently, E[IO] = p/(1 - p), and
:2Ses congestion
l1e to process its A ~ GO(s - I) ~ p' (7.3.1)
FIGURE 7.3.1
Peljormance ofbase-stock policies (sequential versus parallel supply).
20 - , . , - - - - - - - - - - - - - - - - - - - -
15
~
"80
10
~ - - Sequential
- Parallel
"•
u
'"
5
o
o 10 20 30 40 50 60
Base-stock level
FIGURE 7.3.2
Stock-service tradeo.!f (MI1"'f!] system).
6
5
·L~I
--- L~2
4 -L=3
-L~4
.~"
j" 3
u
a
2
............
o
0.4 2.0
Waitin,g: time
Chapter 7 Stochastic Leadtimes: 17,e Structure of the SJlpply System 259
under 0.06 days, so that E[10] == 20 as before. The performance of selected base-stock policies is
as follows:
s A B T
58 0.059 1.18 39.18
To meet the requirement of A ::; 0.05, the company should set s = 62, far more than the 28 for
the original parallel system. This policy has average inventory about 43, compared to 8 for the par
allel system.
Now, suppose the processing times are still i.i.d., but have an arbitrary distribution. The
code for this processing system is M/GIl. Let S denote the generic processing-time ran
dom variable. Assume that E[S], E[S2], and E[S3] are finite. Again, denote I-' ~ IIE[S]
There is no simple formula for g, the pmf of 10, but it can be computed by a recur
sive procedure: Let gD(S) denote the pmf of D(S), the demand during a random time in
terval S, and GZ(S) its ccdf. (Usually, gDiS) can be derived or computed easily.) Then,
g(O) = I - p
1
E[LQ] = ;; (l + ,@\.I ~ P.)/1)
1
I-
\1-'
(7.3.2)
~ E [L ~ p)(~)'
.0
2
V[Ld Q] + G)(1-'3E[SJnC (7.3.3)
260 Foundations ofInveni01Y Management
(1jJ~ is the square of the coefficient of variation of S; i.e., V[S]IE [S] ~ fJ.1V[S].) These
2
are known as the Polleczek-Khintchine formulas. The total order leadtime L. then, has
mean and variance
E[L] = E[L Q ] + E[S]
V[L] ~ V[LQl + V[S]
It turns out (see Section 7.3.8), moreover, that
E[IO] ~ >-E[L]
V[IO] ~ >-E[L] + A1V[L]
These moments can be used to approximate the distribution of 10.
The formulas clearly reveal the impact of the system parameters on L Q and hence
on Land 10: In (7.3.2), the term 1/fJ. = E[S] is a scale factor; the rest ofthe formula ex
presses E[LQl as a multiple of E[S]. The utilization p again plays a dominant role through
p/(l - p); as p approaches 1, E[LQl explodes. Processing variation has a linear impact
via the term Y:(I + 1jJ~). These factors also affect V[LQl in (7.3.3) through the E 1[LQl
term. The additional component of V[L Q ] is similar in fonn, but it expresses processing
variation by the third-moment factor iOfJ. 3E[S3].
In the fortunate case of constant processing times, 1jJ~~ 0 and fJ.3 E[ S3] ~ I, and the
fonnulas above reduce to
E[LQl ~ ~ C~ p)(~)
V[L
Q
] = E 2 [L
Q
] +~ (_p_)(1)2
31-pfJ.
When Sis exponential, 1jJ~= I and fJ.3E [S3] = 6, and we recover the results above for the
MIMII system. Most production processes have values of 1jJ1 between 0 and 1.
Now, suppose that S has a CPR distribution (Section C.5.5) with parameters ("" M),
where,..... is a row-vector and M a square matrix. Assume .....e = 1, so the unit processing
time is positive with probability I. Recall, M- I 2" 0, and E[S] = 1/fJ. = ",/r.1 'e.
In this case, L too has a CPR distribution. Its parameters are (v, N), where
v = ",[(I - p)I + Mil] N ~ M - Ae",
(Actually, N need not have all the required properties for a CPR distribution. Neverthe
less, the distribution of L follows the CPR form.) Also, for P ~ A(AI + N) - I and
'IT = vp, 10 has a DPR distribution (Section Co404) with parameters ('IT, Pl. Therefore,
A ~ eO(s - I) ~ vP'e
Jj ~ el(s) ~ vP(1 - p)-IP'e (7.3.4)
>..
P, ~ - i> 0
J..Lf
l L Q and hence
he formula ex Ri = TIj:S;iPj i ;:;:: 0
Olt role through R = l~=oRi
il linear impact
2
Igh the E [Ld The system is stable provided R is finite. (This is true, for instance, when PI is
ses processing bounded above by some constant < 1 for sufJiciently large i.) In that case, the pmf of
10 is given by
'1 = I, and the
R,
g(i) = R i;:;:: 0
From this, GO and G' can be computed directly. Problem 7.4 presents a special case.
Old 1. 10 has precisely the form above with 1'-, ~ ifL. In this case, IO has the Poisson distribu
nit processing This agrees with the earlier result. What is interesting is the following alternative in
L_W--1e. terpretation: Instead of infinitely many processors 1 each with a constant processing rate,
O\nere there is a single processor, whose processing rate grows linearly with IO(t). The process
10 is precisely the same as before, so 10 is also, even though the order leadtimes are cer
tainly not independent. Thus, the result of Section 7.2.1 depends less on the indepen
on. Neverthe dence assumption than on the expansion of processing capacity with the workload. (Un
+ N)-' and der either interpretation the supply system has unlimited capacity.)
). Therefore, Next. suppose there is a finite number m of independent processors, each with con
stant processing rate jL. So, f-li = min {i, In} f-l. The code for this system is MlM/m. Set
v = >../fL and p = vim CC>..lmfL, and assume p < 1. Then,
(7.3.4)
vi/it
0:$ i:$ m
R
ays the same
ion of calcu-
,UF 1 vm/m!pl-m
i> m
R
i111.
"'I~
where
R='L7-o-+
Vi -
- i!
(vm!m
)( p )
--
1- p
So, g(i) looks like a Poisson pmf for i :s m, and a geometric poo for i > m. The system
combines features of the MIMlI and MlMloo systems, and so does its solution.
EXAMPLE 7.3.A, PART 2
Suppose the computer-repair company has 15 technicians. The processing rate is now !-L = 1.1731,
so that v = 13.639, p = 0.9093, andE[IO] is about 20 as before. Theperfonnance of selected base
stock policies is as fcHows:
, A Jj I
39 0.065 1.30 20.30
To meet the requirement of A ~ 0.05, the company should set s = 42, between the values (28 and
62) of the pure parallel and sequential systems. This policy has average inventory about 23, also
between the values for those two systems.
A(n +s - i)
A, = ~ s<i<n +s
n
0 i ~ n +s
,. I
!! I I
We leave it as an exercise for you (Problem 7.10) to determine the distribution of TO and
formulas for performance measures.
In modeling a spare-parts inventory, often the demand rate is treated as constant.
, m. The system
This is an approximation ofthe model above. It is reasonable when (but only when) n is
)lution. large.
. (I - p)p'
gra(l) = I ,+1 i = 0, ... , S
-p
For p < I, groCi) = g(i)/G(s), where g and G are the pmf and cdf of the geometric dis
Ie values
(28 and
tribution with parameter p. That is, glo truncates the solution ofthe corresponding back
y about 23, also
log model, just as in Section 7.2.3.
Again, Ii = gra(s), so the lost sales rate is Agra(s), and OF ~ A[I - gra(s)]. For
p = I, I = :;'s, while for p ". 1,
s a finite num _ P (s + l)p'+1
I = s - -- + "----'-'-;~
:kordered, that I - p I _ p'+ I
s are pieces of
itical part fails Several related systems can be analyzed in the same way. For example, for a general
:ess is a repair processing-time distribution (Section 7.3.2.2) with p < I, first calculate g for the back
e cannot oper orders case, and then truncate it at s.
on the current
7.3.5 Imperfect Quality
Ii demand rate
mand rate be- 7.3.5.1 Markovian Processing
Return to the MIMII system with backorders. Suppose now that each unit, on complet
ing processing, turns out to be defective with probability o. Aside from the different sup
ply system, the scenario is the same as in Section 7.2.6. This model is a stochastic ana
logue of the finite-capacity deterministic model of Section 3.6.3 in Chapter 3. For now,
assume we inspect each unit on receipt and discard the defectives. Again, ~ = I - o.
While the processor is working, the output of nondefective units is a thinned Pois
son process, itself a Poisson process with rate ~f" (Section C.S.6). Call this the effective
."
n
processing rate. The system is thus identical to one with perfect quality but this slower
processing rate. AlI the results of Section 7.3.2.1 apply, with p = il./(l;fL).
Just as in the deterministic model, defects waste production capacity. If the defect
rate is too high, an otherwise stable system can be rendered lffistable. Within the range
of stability, however, the effects of poor quality are more significant than in the deter
ministic case. For example, the quantity p/(1 - p) now becomes il./(l;fL - iI.), which is
very sensitive even to small 8 if J.L is near A.. The stochastic elements of the system, in
cluding quality itself, amplify the impact of bad quality. Also, the effects are stronger
here than in the uncapacitated system of Section 7.2.6. Poor quality has the greatest im
pact in the presence of both uncertainty and limited capacity.
E[S'] ~ E[S]
I;
~ V[S] + E [S]&
2
V[S']
I; i;'
.1 2
,IS' = ~' + (I - &)",i
Referring to (7.3.2), & affects fL and p just as in the MlMII case, but there is an addi
tional effect through ",1,: Ifthe nominal processing time S is relatively predictable, with
",1 < I, then ",1, > "'i.
In particular, if S is constant (",1 ~ 0), then ",1, ~ &. In sum, de
fects can induce variability in the processing times. This effect is absent in the MIMI!
model above, precisely because "'i
~ I there. (Problem 7.11 obtains sharper results for
the case where S has a CPH distribution.)
- (p)
B(s)
I - P
~p+; - 7(5) ~ C! pJ(l - p~)
Chapter 7 Stochastic Leadtimes: The Structure ofthe Supp~v System 265
ty but this slower Observe,B(s) has precisely the same geometric form as in the immediate-inspection sce
L). nario, and the constant p/(l - p) ~ li(o) is the same in both cases. Here, however, the
city. If the defect geometric-decay rate p+ is larger than p, soB(s) decreases more slowly as s increases.
Within the range Intuitively, postponing inspection makes the system harder to manage, for it creates yet
ban in the deter another source of uncertainty. When there is inventory on hand, some of it may be de
L - A), which is fective, and we do not know how much. So, we need more inventory to protect against
If the system, in backorders.
ects are stronger
;the greatest im
7.3.6 Processing Networks
7.3.6.1 Markovian Processing
Now, suppose the supply system is a network of J nodes. The initial node and movements
e it not for qual between nodes are governed by routing probabilities '"ij' just as in Section 7.2.7. Again,
nsider an MlGIl the Tij can be interpreted as defect rates. Here, however. each node consists ofjust a sin
has the same ef gle processor that operates like the ·/MII system above. Let fLj be the processing rate for
g the same logic node j. (The imperfect-quality system of Section 7.3.5.1 with immediate inspection is
Lble SI, where the special case with J = 1.) This model is called a Jackson network, after its inventor,
Jackson [1957,1963],
Assume that the matrix 1 - R is nonsingular, so the network is open! and every or
der ultimately leaves it. Equation (7.2.4) then gives the total input rates Aj . Set Pj = A/fLJ'
and assume too that Pj < I for all}, so the capacity of each node is adequate to process
its input.
As in Section 7.2.7, the equilibrium occupancies 10) are independent. Here. how
ever, each 10) has the geometric distribution with parameter Pi' as if node j were an
MlMII system operating in isolation. Now, [0 is the sum of the [OJ' that is, the sum of
tlere is an addi independent, geometrically distributed random variables. Consequently, [0 has a DPH
redictable, with distribution, and we can compute the performance measures using the matrix-algebraic
= O. In sum, de formulas for GO and G I , as in (7.3.4).
It in the MlMII This distribution can be constructed directly, but there is a convenient alternative
trper results for method: Let L ~ = lf~ I Li . where Lj is an exponentially distributed random variable with
paraDJeter vj = fLj - Aj ~ fLj(l - p), and the Lj are independent. (Thus, Lj would be the
leadtime at node), if node} did operate in isolation.) So, L ~ has the CPH distribution
with parameters (v, N)! where v is the unit J-vector e l ', and N is the J X J matrix
JofSection3.6
ustomers; units
urs, withdraw a
VI -VI
w another unit,
V2 -V2
which case the
N~
I a unit arrives
-VJ-!
nore elaborate
: Setting p + = VJ
Then, IO = D(L ~), which has the DPH distribution with parameters P = A(lJ + N)-I
and 'IT = vP
266 Foundations a/Inventory Management
Now, consider a series system where the processing time S,i at node} is the constant l/f-l;
Let S* = max {S; : I <; j <; J], and consider an MlGIl system with processing time S·
and input rate A. 'It turns out that the total order-waiting time L Q is precisely the same in
the series system and in this MIGll system. So, L here is simply the sum of all the pro
cessing times S ~ 4j Sj ~ 4j (lIfL), plus L Q in the MlG/I system. Section 7.3.8 explains
Letj* indicate any node for which Sj' = S·. We callj* a bottleneck. The bottlenecks
determine L Q. [fwe remove one of the other nodes, L is reduced by that node's process
ing time, but L Q remains unaffected. Only if we remove all bottlenecks is L Q itself
reduced. Such behavior is quite different from the Markovian network above, where each
node's contribution to L is independent of the others, and the removal of any node re
duces waiting as well as processing time.
Certain more general networks also have simple solutions, for instance, when each node
has several identical Markovian processors working in parallel, as in Section 7.3.3.2. Un
fortunately. other important models do not. Even for a series system with nonexponen
tial and nonconstant processing times, it is generally impossible to obtain the exact dis
More generally, suppose 10+ has a DPH distribution with parameters ('IT+, P +),
e constant 11 f.I:; where 'IT + = V +P + for some vector v + with v +e = 1. (Several models above have dis
::essing time S* tributions oftllis form.) Then, 10 also has a DPH distribution with parameters ('IT, P),
:ely the same in
where
I of all the pro
17.3.8 explains P~ 6P+[I- (1- 6)p+]-1 'IT=v+P (7.3.5)
(See Problem 7.5.) Formulas (7.3.4) then yield the performance measures.
i"he bottlenecks
oode's process
'ks is L Q itself 7.3.8 Customer Waiting Time
we, where each
,f any node re Many of the supply systems above are sequential; they complete orders in the same se
quence the orders are placed. The single~processor systems of Section 7.3.2 are sequen
tial, as are the series networks of Section 7.3.6, but not a network with feedback. Sup
pose also that the total leadtime of an order is entirely unaffected by the arrivals and
ihen each node processing times of subsequent orders. This rules out the systems of Section 7.3.3 with
on 7.3.3.2. Un load-dependent processing rates. Continue to assume Poisson demand, a base-stock
b. nonexponen policy, and backorders.
1 the exact dis Under these conditions, Theorem 6.2.4 of Section 6.2.4 applies. (If you examine its
proof carefully, you will find that it works.) That is, the backorders random variable B
-stem: Imagine has the same distribution as D(BW), the demand during a customer's waiting time. Set
"'Pacity. When ting s = 0 yields [0 ~ D(L) as a special case. These facts have several interesting and
cess any units, useful implications:
~al processing First, consider the Markovian processing network of Section 7.3.6.1. There, we con~
one its perfor structed a random variable L ~ and found that 10 ~ D(L ~). Therefore, for a series sys
'S.) tem (with no feedback), L ~ is identical to L, the actualleadtime experienced by units in
the supply system. (With feedback, however, these two random variables have different
distributions.)
Next, consider the series system with constant processing times of Section 7.3.6.2.
"", the supply Recall, S* denotes the maximal processing time, and S the total processing time over the
ith rate AI, in nodes. Here,IO is the snm oftwo independentrandom variables: The first is D(S - S*),
·time distribu which has a Poisson distribution with mean A.(S - S*). The second is the occupancy of
the MlGIl system with processing time S*; we can compute or approximate its distri
atel,+ = A. + bution as in Section 7.3.2.2. The full distribution of [0 is the convolution of these two.
~ system, 10+_ In general, the mean and variance of 10 can be obtained easily from those ofL:
one of our or E[/O] ~ AE[L]
ional distribu
g on 10 yields V[IO] ~ AE[L] + A. 2 V[L]
Notice the crucial difference between a sequential system and the paraJlel~processing
system of Section 7.2.2, both assuming Poisson demand: There, V[IO] = E[/O] =
AE[L]; the second term in V[IO] above is absent. In a sequential system, however, the
variance of the leadtime does play an important role. (Here, of course, L is not specified
lmetric distri with the data of the model; it is determined by the interaction of the demand and supply
ows that 10 is processes.) In all the sequential systems above, moreover, V[L] is positive, even in the
c )J(fL - A.+). M1G/1 model with constant processing times. And, given a sequential and a parallel
268 Foundations afInventory Management
model with equal values of E[L], the sequential model will always produce larger values
of lJ, and hence J, for all s. (See Problem 7.6. Figure 7.3.1 above illustrates this fact. Sec
tion 7.5 further explores the effects of V[L] on performance.)
A similar relationship holds between the moments ofBand BW
E[B] ~ '\E[BW]
Notice, (7.3.2) is identical to (7.3.6) with ~1 ~ I, so this approximation agrees with the
exact result for the M/OIl system. In general, the term y,(~1 + ~}) represents an over
all variation factor, combining the effects of demand and supply variability. The formula
also works for lumpy demand, assuming each unit is processed individually, as in Sec
tion 7.2.5.
The approximation works especially well when p is large. In fact, this is a heavy
traffic approximation. The true value and (7.3.6) agree in the limit (in a certain sense) as
p --> 1. Even for moderate p, the approximation often works well, though not always.
(Researchers have developed a host ofrefinements to improve its accuracy; see Heyman
and Sobel [1982], for example.)
Why do we use similar notation for ~1 and ~}? Let S indicate the supply process,
that is, Set) is the cumulative output through time t, assuming the processor works con
tinually, never stopping. Because the processing times S are ij.d., S is a renewal process.
It so happens that~} is precisely the long-run variancelmean ratio for this process. So,
Chapter 7 Stochastic Leadtimes: The Structure o.lthe Supply System 269
nce larger values ll'~ and t\Jheally do measure similar quantities. In fact, (7.3.6) applies more generally to
l.es this fact. Sec a world-driven supply process S, i.e., one with the same properties as D discussed in Sec
tion 6.3. In that setting l\J~ is defined as the variance/mean ratio.
The heavy-traffic theory also shows that, as p approaches I, the distribution of L be
comes exponential with E[L] given by (7.3.6). (Of course, E[L] includes E[S] as well as
E[Lg], but E[L Q ] dominates for large p.) This suggests using the exponential distribution
to approximate L
ntain important How does this approximation work for the MlG/1 system? It is inexact, but close:
e series network The variance of an exponential variable is the square of its mean. For the M/GIl system.
the first term in V[Lg] is the square ofthe first tenn in E[Lg], and when p is near I these
terms dominate the others.
As for 10, it remains true that E[IO] ~ AE[L], and (7.3.6) estimates E[L]. Actually,
the following approximation works a bit better:
Jillpossibilities:
for instance, for E[1O] ~ ~(t\J~ + t\J~)C ~ p) (7.3.7)
[ogle Markovian
(This agrees with the result for the MIMII system.) Moreover, by yet another result of
able to compute
heavy-traffic theory, the distribution of 10 itself is approximately exponential. Sec
mtlMCDClMlI
tion 7.3.11 explores this approximation fmther.
sting and useful,
ate performance
7.3.10 Larger Batches
<ler class of sys
7.3.10.1 Order-Driven Control
;ystem behavior,
Things become considerably more difficult for an (r, q) policy with q > I, even with
ern and a world
Poisson demand. As mentioned in Section 7.2.4, the supply system then becomes a bulk
Uowing is a use
queue, which is difficult to analyze.
The approximation scheme there is plausible here too: Approximate fa by an ex
ponential distribution as in Section 7.3.9 above with t\J~ = l/q, and then use the result to
(7.3.6)
approximate 10 and IN as in Section 7.2.4. This approach should work well for a heav
ily loaded system (p near I), but to our knowledge it has not been tested.
agrees with the
resents an over 7.3.10.2 Direct Processor Control
ity. The lonnula In the models above, we control the supply system indirectly through the stream of or
!lally, as in Sec ders we send it. The system works when it finds orders to work on, and it shuts off when
the orders are exhausted. This is the scenario outlined in Section 7.1 and followed
this is a heavy throughout most of the chapter.
:ertain sense) as Now, suppose we have a direct communication link to the supply system: Instead of
Jgh not always. placing orders, we can tell the system precisely what to do. For simplicity, suppose there
:y; see Heyman is but a single processor with two possible states, on and off, so our instructions to it are
signals to start or stop producing.
supply process, First, suppose we follow the analogue of a base-stock policy: When the net
iSOr works con inventory falls below some fixed value s. start the processor, and stop it when the
~ewal process.
net inventory is s or more. A moment's thought should convince you that this system
lis process. So, is identical to the order-driven one. Given the same demand sequence and unit
270 Foundations ofInventOlY Management
processing times, the processor turns on and off at the same times, and IN is the same
in the two cases. The order stream carries essentially the same information as the di
rect production signals.
Now, consider a more general policy: There are two policy variables, r and s, where
r < s. Suppose the system starts at time 0 with IN(O) ~ s and the processor off. Nothing
happens until IN(t) hits r. At that moment we tum the processor on. When IN(I) hits s,
we tum the processor off. Now the system is back at the starting point, and we continue
in the same manner.
This policy makes sense when there is a setup cost to start (and/or a shutdown cost
to stop) production, or simply a cost to send a signal of either kind. Increasing the range
s - r means that such costs are incurred less frequently. In this sense the policy is sim
ilar in spint to an (r, q) policy; it exploits scale economies. (Ifs ~ r + I, clearly, the pol
icy reduces to a base-stock policy.) As we shall see, however, this direct-control system
is considerably easier to analyze than a standard order-transfer system.
Call this an (r, s) policy. (The same notation is used for a related policy in Sec
tion 6.6.) The model is a stochastic version of the detenninistic limited-capacity models
of Chapter 3, Section 3.4.
As in Section 7.1, define 10(1) = s - IN(I), and call 10 the occupancy of the sup
ply system. Again, 10 does not depend on the particular values of rand s, but it does de
pend on the difference s - r, denoted by q. (Of course, q is not the batch size here. In
deed, q is the minimal number of units produced during a run; more units may be needed
to compensate for demands occurring during the run.) Also, let X (capital X) be the sto
chastic process indicating whether the processor is on or off; X(t) = I means on, while
X( I) = 0 means off.
Now, assume that in all other ways the system functions just like the MIM/l system
above: Demand is Poisson, and the unit processing times are exponentially distributed.
Problem 7.7 asks you to demonstrate that the joint process (10, X) is a continuous-time
Markov chain and to derive its stationary distribution. Specifically, letting gU, X) ~ Pr
{lO=I, X=X}, X E {O, I}, I;': 0,
g(l,. 0) -- ~ 0 <0 l.
_ <q
q
= (l)[G
\q
2
(r) - d(r + q)]
Chapter 7 Stochastic L<!adtimes: The Strll('(urt' ofthe Supply System 271
d IN is the same where G" is the second-order loss function for the geometric distribution. Likewise,
nation as the di
7~s-E[1O] +B
:5,rand s, where 1
iSor off. Nothing ~ -(q + l) +r - ( -p-) + B
_
2 I - P
'hen IN(t) hits s,
and we continue (Problem 7.7 also asks you to verify these expressions.) These formulas are precisely
analogous to those for an (r, q) policy in Section 6.2.3, namely (6.2.12) to (6.2.13), but
a shutdown cost they use a geometric instead of a Poisson distribution. Finally, interpret OF as the fre
easing the range quency of production starts. Production commences at a transition from state (q - 1,0)
Ie policy is sim to (q, I), and this event occurs at rate
, clearly, the pol
t-control system OF~'Ag(q-I,O) (1- p)'A
q
I policy in Sec This measure too has the same form as before, except for the constant term (1 - p). (In
capacity models deed, the average run length is q' ~ q/(I - p) units, and OF = 'A/q'.)
More generally, suppose the unit processing times have a CPH distribution, as in
mcy of the sup Section 7.3.2.2. The results have the same form as those above:
r. but it does de
. I - p
:h size here. In g (I, 0) ~ O<5i<q
;may be needed q
at xl be the sto g(i, I) ~ vP(1 - P)e
[leans on, while l::'5i$q
q
OF~ (I - p)'A
q
where G" is the second-order loss function of the DPH distribution with parameters
(v, P). (An explicit formula for G" is given in Section 7.5 and Section Co404.)
Because the performance measures have the same fonn as in the constant-Ieadtime
model, the qualitative results of Section 6.2.3 still apply; e.g., B is decreasing and con
.2:: 0, 80S.2:: q.
vex in I: Also, the algorithm of Section 6.504 finds a cost-minimizing policy, and, up to
a continuous approximation (see below), the bounds of Section 6.5.3 remain valid.
section, it is evident that the exponential distribution plays a similar role for limited
capacity systems:
several places, and a geometric distribution has essentially the same shape as an expo
nential. In other places, 10 has a DPH distributIon, and the tail of a DPH distribution is
= p'+l for large i, where p ~ max {p).) Finally, for the (1; s) policy of Section 7.3.10.2,
the performance formulas look like those of Chapter 6 for an (r, q) policy, but using a
limited-capacity system. That distribution has only one parameter, its mean v = E[!Oj.
To estimate it, we need some information about the supply and demand processes, but
not all that much. For a single-processor system, for instance, we may use (7.3.7). In
most cases, v is nearly proportional to 11(1 - pl. Also, the standard deviation is a = v.
The performance measures for a base-stock policy are given in Section 6.4.2, but
So, to meet a target value 1 - w_ for A, set s ~ [-In (l - w_)]v. In that case, I ~
-----..J
[-In (l - w_) - w_]v. For average-cost minimization, following the approach of Sec
tion 6.5.2,
s* ~ -In(l- w)v
C* = hs* ~ h[-ln(l- w)]v
Let us examine the behavior of these formulas and compare them with those for the
normal approximation. Since 0" = v, it is clear that 0" plays essentially the same role in both
cases. It is a scale factor in the ratio s/O", and it multipliesB, 7, s*, and C*. So again! 0" is a
key determinant of system performance; the original system parameters (p, !jib. !jIl, etc.)
affect performance through 0".
However, a reflects limited capacity, mainly through the explosive factor 1/(l - p),
an effect absent in a parallel-processing system. Also, given two systems with Poisson
demand and equal v, one parallel and one sequential, 0" is different in the two cases;
a = '.(v for the parallel system, but a ~ v in the sequential case. In general, the expo
nential approximation and the finite-capacity supply systems it describes do not exhibit
statistical economies ofscale. For instance, if (7.3.7) is used, 0" includes a term propor
tional to the demand-variance ratio 4Jb. not its square root. Likewise, 0" grows at least as
fast as A, not \lA, even ignoring the factor 1/(1 - pl. That factor, of course, grows hy
perbolically in A, indicating strong diseconomies of scale.
In other words, while the benefits of system-improvement efforts can still be esti
mated through a, those benefits are quite different here, and usually greater, than with a
",,_._-
role for limited parallel system. ]n particular, one way to improve the system is to increase its capacity
and thus lower p. Obviously, this has no meaning for an infinite-capacity system.
7.3.9, the distri Furthermore, sand 7 are more sensitive to w_ than their normal-approximation
MItions appear in cOlmterparts.]n the normal case) sand lincrease only as the square root of - In (1 - w_)
13pe as an expo as '"_ approaches I, but here they are linear in this quantity. (s* and C* depend on '" in
H distribution is the same way.) This finding coincides with the discussion of Figure 7.3.2 in Sec
ck, in fact, GO(i) tion 7.3.2.1. A capacity limit implies that much more inventory is required to achieve
;"ction 7.3.10.2, good service.
'!icy, but using a Virtually every real supply system has elements ofbotlJ models. No system truly has
unlimited capacity, but capacity can often be adjusted in response to workload varia
e behavior of a tions, at least to some extent. Given the choice, then, which is the "right" form, the nor
oean v ~ E[IO]. mal or the exponential? There is no conclusive answer. Ifhard capacity limits are likely
d processes, but to be encountered frequently, then the exponential approximation is probably a better
~ use (7.3.7). In choice, but the normal is superior when capacity is relatively flexible.
~tion is IT = v. There are more refined approximations, intermediate between these two. A gamma
'Clion 6.4.2, but distribution, for instance, has two parameters, allowing (indeed requiring) a separate es
timate of V[IO] as well as E[IO]. It shares many of the qualitative characteristics of an
exponential, but for large values of its shape parameter (n), it looks nearly nom,a!. This
approach is discussed further in Section 7.5 below. (Problem 7.4 explores a system
wlJere capacity is somewhat but not totally flexible. There, 10 has a negative-binomial
distribution, which can be fairly well approximated by a gamma.)
n that case, 1=
pproach of Sec 7.4 Exogenous, Sequential Supply Systems
7.4,1 Description
7.4.1.1 Examples and Summary
This section presents a different type of supply system. Here are some examples to mo
ith those for the Imagine that we operate a small retail store and order goods from a large manufacturer.
anoe role in both Think ofthe supplier's production and transportation activities as a processing network. For
So again, (T is a simplicity, suppose this network is sequential; it delivers our orders in the same sequence as
(p, lJJi"lJJ1,
etc.) they are placed. Having read Section 7.3.6, we could create a network model ofthe kind dis
cussed there to represent the supplier. The manufacturer doubtless has many other cus
actor 1/(1 - p), tomers, so we would adjust the model accordingly, along the lines of Section 7.3.7.
!IS with Poisson Something seems out of balance: The great virtue of the models of Section 7.3 is
l the two cases;
that they capture the dynamic effects of orders on the supply system. Each order's lead
oeral, the expo time depends on the congestion it finds in the supply system, which in turn depends on
~ do not exhibit prior orders. The supply system is endogenous in this sense.
; a term propor This is important in other contexts, but here such effects are negligible. While the
7UWS at least as supplier's overall workload, and hence the leadtimes we experience, may fluctuate over
urse, grows hy time, our demands and orders contribute little to these fluctuations. We can thus reason
ably approximate the supplier's network as exogenous, ignoring the effect of our orders.
:an still be esti As we shall see, this means we can do without much of the fine detail of a processing
lter, than with a network mode!.
274 Foundations oflnventOlY Management
Here is another example: Suppose leadtimes represent travel times. When we place
an order, the corresponding goods are loaded onto a vehicle, say a truck, a train, or a ship.
The vehicle then travels to us, over a highway, a track, or an ocean. The vehicles are iden
tical for all orders, and they all follow the same route. The transit times can vary, how
ever, because of local traffic eonditions, switch failures, the weather, and so forth. Now,
suppose that the traffic our vehicles add to the overall load on the transport system is
negligible. Thus, we can again think of the supply system as exogenous. Also, suppose
that none of our vehicles ever passes another, so the system is sequential.
These examples emphasize the two key properties of the systems considered here:
The starting point is an exogenous supply system, whose evolution is independent of our
demands and orders; the operation of this system determines OUf leadtimes. This system
is sequential, so our orders do not cross in time. In practical tenus, the difference be
tween this scenario and the prior section's is one of scale. We are assuming, in effect, that
the supply system is large relative to our operations. Of course, no real supply system is
truly exogenous; like the parallel supply system of Section 7.2, this model is a limiting
case, but a useful one.
EXAMPLE 7.4.A
A Japanese trading company purchases hardwoods from an intemationallumber company and dis
tributes them to furniture makers throughout Japan. The supplier harvests logs in Southeast Asia,
processes them into lumber, and ships them all over the world. For each of its major markets, in
cluding Japan, it fills orders in sequence. Although the trading company is large, it is only one
among many customers of the lumber company. Thus, from the trading company's viewpoint, the
supply system is essentially exogenous.
The advantage of this approach is that it allows us to apply essentially the same
methodology used in Section 6.3 for the case of constant leadtimes. In particular, we can
easily evaluate perfonnance for any world-driven demand process using any (r, q) pol
icy. This is in sharp contrast to the other stochastic-Ieadtime systems above, where non
Poisson demands or large batch sizes complicate matters significantly. Furthennore, the
optimization methods and results of Section 6.5 all apply.
7.4.1.2 Formulation
We now describe the supply system in fonnal tenns: The core of the model is a Markov
process Z ~ {Z(t) : t 2: A}, which describes the state of the supply system. In addition,
way: For each t, given Z(t), the distribution of L(t) depends on Z(u) only for u 2: t. not
u < t. That is, conditional on Z(t), L(t) is independent of the past of Z, but certain com
mon events may affectL(t) and the future ofZ. Moreover, the joint process (Z, L) is well
behaved, in the sense of Section 6.2.1; it is ergodic and has a limiting distribution, rep
For each t, L(t) is a scalar with 0 :=; L(t) < 00. We call L(t) the virtualleadtime attime
t. If an order is placed at time t, its leadtime is L(t); however, L(t) is defined for all t.
Two additional assrunptions are needed. First, (Z, L) is entirely independent of D;
this is what exogenous means. Second, t + L(t) is nondecreasing in t, which ensures that
orders do not cross in time, i.e., the supply system is sequential.
Chapter 7 Stochastic Leadtfmes: The Structure ofthe Supply System 275
o 5 10 15 20 25 30
Time
::ompany and dis
n Southeast Asia,
.ajor markets, in Figure 7.4.1 illustrates the interactions between demands, leadtimes, and orders, as
ge, it is only one suming a base-stock policy. The format is the same as Figure 6.2.1 's. In the lower part of
(s viewpoint, the
the figure, L(t) is the dotted curve. Each demand is indicated by a vertical line of height
L(t), so the attached diagonal line meets the time axis at t + L(t), when the correspon
tially the same ding order arrives. The upper part of the figure graphs 100t).
rticular, we can For example, consider a sequential processing network, like the manufacturer's sys
: any (r, q) pol tem above, and assume its input is a Poisson process. (This is not D; remember, the sup
"'-e, where non ply system is exogenous.) Also, assume Markovian processing at each node. The state
nthermore, the variable Z(t) describes the current occupancy of each node (the exogenous analogue of
the vector [I0/tn in Section 7.3.6). This Z is indeed a Markov process, specifically, a
continuous-time Markov chain.
Let L(t) be the time until all units now present depart the network; ifthe network is
lei is a Markov empty, L(t) = O. Thus, each of Our orders is a "virtual customer"; it rides along with one
m. In addition, of the actual units, the one destined to leave the network last, adding nothing to that
quasi-Markov unit's processing times. Notice that L is dependent on Z in the manner above; L(t) de
'foru~t,not pends on Z(I) and also the future processing times of units currently in the network,
It certain com which also affect Z(u), u 2: t. Under standard assumptions (e.g., stability), clearly, (Z, L)
;(Z, L) is well satisfies the conditions above.
otribution, rep Here is an alternative fonnulation ofthe same system: Imagine that, when a unit en
ters the network, it brings all the information needed to determine its processing times
'adtime at time at the nodes. Now, Z(t) includes, in addition to the occupancies, each unit's remaining
ed for all t. processing times. This is sometimes called the "workload formulation." Given Z(t), one
pendent of D; can determine the entire future evolution of the network, insofar as it affects units al
:h ensures that ready present at time t. Again, L(t) is the time until all units now present depart. In this
construction, L(I) is a deterministic function of Z(t).
---
276 Foundations ofInv€ntOf.Y Management
These models may seem complex. Fortunately, we need them for conceptual pur
poses only In the end, as we shan see, the performance of the system depends only on
the distribution of L. For instance, although the original (Z, L) above and the workload
formulation are different stochastic processes, they have the same equilibrium Ieadtime
L, and that's what counts. (Incidentally, the workload formulation can be extended to a
more general network, where the processing times have general distributions, and these
times are independent over units.)
Here is a slight variation: Suppose Z(r) includes, in addition to the information
above, the procesl'ing times of the next unit to arrive. So, when a unit arrives, it inherits
those processing times, and determines the following unit's times. Let L(t) be the time
until that next unit would depart, ifit arrived at time t. 0'le may call this L(I) the virtual
sojourn time.) Here, provided the processing times are always positive, so arc L(t) and
1. This is more realistic in most situations. Assuming markovian processing, in fact, L
has the same distribution as L= in Section 7.3.6, i.e., a simple CPH distribution.
By the way, you might expect that observation of Z(t) should affect our order deci
sions. Indeed it should. Tlie slory here is parallel to Section 6.3, where a world variable
Wet) drives demand: Assume that we do not have, or cannot use? current infonnation
about the supply system in the replenishment policy. When the supply syslem is some
large network whose operations we cannot see, as in the examples above, this assump
tion makes perfect sense. So, it is reasonable to use a simple base-stock or (r, q) policy,
depending on scale economies in supply, as before.
7.4.2 Allalysis
First. \\'e prove an analogue (Q Lemma 6.2.2:
PROOF. All and only those orders placed at time t or before arrive at or before time
t + L(t).
Now, consider a world-driven demand process satisfying the assumption of Sec
tion n..n . Thus, Lemma 6.3.1 describes the limiting distribution of (IP, Wi, and for any
fixed leadtime, Theorem 6.3.2 describes the limiting distribution of IN in terms of the
leadtimc demand.
For the actual stochastic-leadtime system, define the leadtime demand as the ran
dom variable D = D(L). That is, start the processes (W, D) and (Z, L) in equilibrium at
time 0, and observe the demand over the random time L = 1.(0). Let g(dlt) denote the
pmf of D( t), i.e., the leadtime demand for a constant leadtime t. Then, the pmf of Dis
g(d) = ELf g(diL)] d2>O
That is, D is a mixture of the D(t) over t, Idting t be distributed as L.
We now show that the analogue ofTheorem 6.3.2 describes IN in this case:
~
conceptual pnr THEOREM 7.4.2. Under the assnrnptions above, the limiting distribution of the net
Iepends only on inventory is described by the equation
ld the workload
Ibrinrn Ieadtime IN~ IP - D
~ extended to a
nons, and these
where IP is uuifonnly disttibuted on the integers in the interval [r + I, r + q], D is the
leadtime demand defined above, and IP and D are independent.
:he information
rives, it inherits PROOF. We give an intuitive argument (which can easily be fonnalized): First, it is
[.(t) be the time not hard to show that IN(t + L(t)) has the same limit as IN(t) itself, namely IN So con
L(t) the virtual sider the random variable IP(t) - D(t, t + L(t)] in the right-hand side of (7.4.1). As
so are L( t) and t --j 00, as in Lemma 6.3.1, IP(t) goes to IF, a uuiform random variable, independent of
iSing, in fact, L W. Also, since there is no interaction between the processes (w, D) and (Z, L), it is clear
ibution. that D(t, t + L(t)] --j D. Combining these two limits yields the result.
our order deci This theorem is the exact analogue of the earlier results for constant leadtimes. It
world variable follows immediately that the perfonnance criteria have exactly the same fonn. Let GO
nt infonnation denote the ccdf of D, G 1 and G'- its loss functions, and v = E[D] = AE[L].
;ystem is some
,_ this assnrnp COROLLARY 7.4,3,
[)f (r, q) policy,
Ii ~ G)[G'-(r) - 2
G (r + q)] (7.4.3)
_ I _
I = 2(q + I) +r - v +B (7.4.4)
(7.4.1) Consequently, the average cost C(r, q) has the sarne fonn as in Section 6.5, so the results
Jf before time there apply directly.
It is worth noting that, in the Poisson-demand, base-stock-policy case, the results co
incide with those for an endogenous, limited-capacity system. The leadtime demand D
plion of Sec here and 10 there play the sarne role in perfonnance, and 10 ~ D(L) ~ D.
), and for any Here, of course, L is quite independent of the demand process, whereas there L is
I tenns of the
strongly affected by D, Still, if the real supply system is not entirely exogenous, and we
wish to study changes in A, say, there is nothing to prevent us from specifying a differ
uJ as the ran ent L for each Ato represent congestion, using what we know about endogenous systems.
quilibrium at Strictly speaking, this is cheating, since the exogenous-supply model rules out such in
it) denote the teractions. As an approximation, however, it is quite reasonable.
pmfof Dis Many different processes (Z, L) can give rise to the sarne L. It is interesting and use
ful to see that only the distribution of L matters in the end, not finer details of the supply
system. On the other hand, D reflects the entire distribution of L, not just its mean, uu
like 10 in the case of independent leadtinles. As Section 7.5.3 shows, this has major im
case: plications for the perfonnance of the system.
•
278 Foundations afInventory Management
The same approach works for lumpy demand, say a compoWld-Poisson process. The re
sults are identical to those of Section 6.6, depending on the policy, with D = D(L), as above.
Consider a lost-sales system. Reviewing the analysis above, troubles begin immediately:
Lemma 7.4.1 no longer holds, because lost demands do not reduce the net inventory. The
demand process effectively ceases to operate when IN(t) ~ O. Since the rest of the analy
sis builds on the lemma, the entire approach breaks down. (The same is true of any sys
tem whose effective demand process, describing the demands actually filled, depends on
the current value of IN(t), including the finite-source demand model of Section 7.3.3.3.)
Here is an approximation, inspired by the results ofthe previous two sections: First, con
sider a base-stock policy. Let g and G indicate the distribution of D, as above, and estimate
/I(s) ~ g(s)
G(s)
OP(s) = 1..[1 - /I(s)]
l(r, q) ~ (~)-S::':;+J(S)
Here is an alternative approximation. These formulas assume that there is never
more than one order outstanding. This assumption truly holds when 0 :5 r < q, so in this
case the formulas are exact.
_ G\r - I)
A(r, q) q + G\r)
_ A
OF(r, q) = q + G1(r)
process. The re (The derivation involves rather intricate methods and calculations, which we omit. See
D(L), as above. Hadley and Whitin [1963], pages 197-204.)
Which ofthese approximations works better? We know of no concrete evidence. We
expect the first to be more accurate when q is small relative to r, and the second when q
is relatively large.
Ib.e approach is Anyway, the (r, q) policy is a heuristic. See Section 9.6.5 of Chapter 9 for a more re
~ and imperfect
fined heuristic in the discrete-time setting.
Return to the backorders case, but suppose the supply system sends us defective units,
in immediately:
I inventory. The which we inspect and discover immediately. (If the supply system itself detects and cor
rects its own defects, simply include the resulting delays within L. based on the results
led, depends on
<:ction 7.3.3.3.) As in the lost-sales case, Lemma 7.4.1 breaks down, but for a different reason. The
nons: First, con definition of IP(t) becomes problematic: The inventory on hand consists of good
e. and estimate quality units, and backorders must be filled by such units. Some of the inventory on or
der, however, may ultimately tum out to be defective, and we do not know how much.
Thus, the two components IN(t) and IO(t) of IP(t) measure different things. We canl/ot
simply add them to form IP(t).
We do not yet know how to resolve this dilemma. For the constant-Ieadtime system,
Section 7.2 provides exact (for q = I) and approximate (for q > I) methods. Chapter 9
presents a quite different approach. Perhaps these techniques can be adapted to cover the
system here.
g(z) = (I-P)"
I - pz
where p ~ AI(A + fl.). Consulting Section C.2.3.6, we see that the leadtime demand has
the negative-binomial distribution with parameters p and n. Section Co2.3.6 also shows
how to compute the pmf and the loss functions.
EXAMPLE 7.5.A
The hardwood distributor of Examplc 7.4.A estimates the average leadtime for shipments ofma
hogany to beE[L] = 1.5 months, with a variance of V[L] = 0.75 months2 . Demand for mahogany
is a Poisson process of rate A = 8 (hundred board-meters/month). The supplier accepts orders only
in units of5 (hundred board-meters). The cost of holding inventory is h = 1 (thousand ¥/[(hoo
dred board-meters)-monthD, and the backorder penalty cost is b = 9 (in the same units). Since q
= 5 is fixed, there is no need to include the order cost.
Use a gamma distribution to approximate L. Thus, fJ, ~ E[L]IV[L] ~ 1.5/0.75 ~ 2 and n ~
~[L] = 2(1.5) = 3. Thus, D has the negative-binomial distribution with parameters n = 0.75 and
p = 8/(8 + 2) = 0.8. Using the negative-binomial formulas in Section C.1.3.6 within the per
formance measures of Section 7.4.2, we obtain the costs (in thousand ¥/month) of alternative (r,
q) policies for q = 5:
,. e(r, q)
15 18.55
16 17.73
17 17.16
18 16.80
19 16.63
20 16.63
21 16.77
22 17.04
23 17.41
24 17.88
25 18.42
FIGURE 7.5.1
Performance o.fbase-stock policies (effect o.fleadtime variance).
10 25
\\ '. .<> . 15 _
shipments of ma
~ \~ . :.:
. . . :?:.;//
... ~
md for mahogany
:cepts orders only
bollSand ¥/[(hun
]
g
co
5
10
!
rte units). Since q
~// ... ;..../' .......•.
/\+ = /\ + /\' the total arrival rate. So, p is more sensitive thanp to /\. But, if ~ and /\'
are large relative to /\, this interaction effect is slight. The exogenous system is thus a
limiting case of the multiple-source system, as the discussion of Section 7.4.1 suggests.
THEOREM 7.5.1: If demand is a Poisson process, and L has a CPH distribution, then
D has a discrete phase-type (OPR) distribution (Section C.4.4) with parameters ('IT, P),
where
P = A(H + M)-I 'IT ~ ...p (7.5.2)
(Problem 7.16 asks you to show that 'IT and P have the required properties.)
The loss functions are given by
GI(i) = 'IT(1- p)-IP'e (7.5.3)
(1)
Section C.4.4 explains how to evaluate such quantities. For a base-stock policy with
q ~ I and s = r + I, these formulas reduce to
A = ...P'e
lJ = 'IT(I-P)-lre
We encountered essentially the same formulas in (7.3.4).
Furthennore, using Theorem 6.2.4, we can fully describe the distribution of CliS
tomerwaiting times, BW' First, consider a base-stock policy with s ~ O. Given the form
t
ofD, it follows immediately that B = [D - s has a OPH distribution with parameters
('IT?", Pl. Now, reverse the argument of Theorem 7.5.1, with B playing the role of D, to
conclude that BWhas a CPH distribution with parameters ( ... p', M). Thus, the/orm of
BW is precisely the same as that of L, and they even share the same matrix parameter M;
only their vector parameters differ.
For example, suppose L has an exponential distribution with parameter /-L; here,
... ~ (I) and M oc (11). Then, BWhas a delayed exponential distribution, in particular,
F2w(t) = p'; e- "'. r > 0, where p = AI(A + 11).
Similarly, for an (r, q) policy with r 2 0, BWhas a CPH distribution with matrix pa
rameter M and initial vector (l/q) ...P(1 - p)-I(I - pq)P'.
es.) THEOREM 7.5.2. If demand is a MCDC process, and L has the CPH distribution
above, then D has the DPH distribution with parameters ('IT, Pl.
(7.5.3) The expressions for P and 1T generalize (7.5.2) for the Poisson-demand case. (There, Q
(7.5.4) becomes the I X 1 matrix (0), ~ ~ (1), and A ~ (A).) Thus, fonnulas (7.5.5) to (7.5.6)
for A andB apply immediately to this model. (Under a base-stock policy, BW again has
a CPH distribution with matrix parameter M. The initial vector now involves powers of
a larger matrix like R See Song and Zipkin [1992].)
(7.5.5) Evidently, the effort required to compute the performance measures depends on the
complexity of the demand and supply models, through the sizes of the matrices Q and
M. It is worth noting that, in the special case where demand is a renewal process, the cal
(7.5.6)
culations simpliry somewhat: Assume that the interdemand times have a CPH distribu
tion with parameters (K, K). So, Q = - K(f - eK) and A = KeK. Therefore, A ® f ~
lCk policy with (Ke ® f)(K ® f), so
GO(d) ~ 'ITpde ~ 'ITRP~e d2>O
where
PR = (K®1)(A®I+ _QEbM)-l(Ke®l)
M~ (~ -~)
is driven by a Demand is the MMP process of Section 6.3.1 above. The matrices above are thus
D specifies the
0.5
~ L is constant 0.5 )
A®f =
'" = 1. With a ( 19.5
19.5
284 Foundations ofbrventory Management
-Q@I~ (W," )
-20
20
20
-20
-20 20
~
-6
3
;0M (:
6
0
-:)
Combine these as above to compute P. (The actual numbers in P are not especially illu
minating.) Also, ~ @ JL ~ (Y" 0, Y" 0). Suppose we use a base-stock policy. Figure 7.5.2
compares the performance of this system to one with Poisson demand. Evidently, the
more variable MMP process leads to higher backorder" but only slightly higher. (We ex
plain why in a moment.)
7.5.3 Approximations
As indicated above, it is common practice to approximate the distribution of L using a
convenient fonn, on the basis of estimates of its moments or a model of the supply sys
tem. An alternative method, also common practice, is to approximate D directly. Sup-
FIGURE 7.5.2
Pe'fomtance afbase-stock policies (effiet of demand variation).
5 16
14
4
,
- Poisson 12
-MMP
.," 3
I ~ 10
,
."
0
'"•"
8 ""
<
:'.
0
"
'<
'" 2
6
o 0
o 5 10 15 20
Base-stock level
I:!
pose that we estimate !I., ll? = <V1, E[L] and V[L]. Then, the mean and (approximate) wui
ance of Dare
v ~ E[D] = !l.E[L]
c? ~ V[D] = <v2 !1.E[L] + !l. 2 V[L] <7.5.7)
as in (6.3.1).
We can fit these estimates to several distributional forms. One is a negative
binomial distribution. (This requires V[D] > E[D].) Another is a gamma distribution
(for D itself, not L). This is another continuous approximation. It can be viewed as a re
finement of the exponential approximation of Section 7.3.11. (It allows any positive
specially illu
E[D] and V[D].) The loss functions are given in Section C.2.5.3 and Problem 7.17. Also,
.. Figure 7.5.2
for relatively small V[L], we can even use the nomlal approximation. Finally, the maxi
ovidently, the
mal approximation provides a universal upper bound.
gher. (We ex-
Once more, the optimal cost is nearly proportional to (T. So, V[L] does affect sys
tem performance for a sequential system. unlike a parallel one. Also, (T is no longer
proportional to the square root of A; rather, for large A, (T is roughly linear in A. Thus,
the system does not enjoy the statistical scale economies discussed in Section 6.4.1.
Iof L using a In the numerical example above, E[L] = ~, V[L] = (Yo)' + (~)' ~ 0.139, A ~ 10,
e supply sys and <v' ~ 1.45. Thus, v ~ 5, (T2 ~ 21.1, and (T ~ 4.6. For Poisson demand, again v = 5,
Iirectly. Sup- but <V 2 = I, so (T' = 18.9 and (T ~ 4.3. Evidently, the values of (T in the two models are
nearly the same; the second term in (7.5.7) dominates the first. This fact helps explain
why the two curves in Figure 7.5.2 above are so similar.
I _ p)n
= ( 1- zp
Next, suppose 15 has the OPR distribution with parameters (fr,P), and Y itself has
a OPR distribution with parameters (0, 8). (Since gyrO) ~ 0, Oe ~ 1.) Then, D also has
a OPR distribution with parameters ('IT, P), where
'IT=fr®O P = f® 8 + l' ® [(I - 8)eO]
(See Problem 7.20.) For example, suppose that Yhas a geometric distribution, shifted up
by I so thatgy(O) = O. So, 0 and 8 are both I X I arrays, 0 = (I), and 8 = (8) for some
scalar 8 with 0 < 8 < 1. In this case,
17=fi' P ~ 8f + (I - 8)1'
The loss functions are given by (7.5.3) and (7.5.4). These can be used within the formu
las of Section 6.6 for ii, JJ, and J, depending on the type of policy used. For an (r, q) pol
icy, A andE are given by (7.5.5) and (7.5.6). (For an (r, s) policy, it is possible to com
pute ,explicitly.)
In general, we can adapt the approximation of Section 7.5.3. Assuming compound
Poisson demand,
E[l5] = AE[L]
The term A(V[Y] + ~2E2[y]) in (7.5.9) represents the overall demand variation, as dis
cussed in Section 6.6.2; this quantity is scaled by the average leadtime. The term V[L]
represents supply variation, which is scaled by the (squared) overall demand rate. The
total variation in the system, V[D], summarizes the contributions from all these sources.
7.5.5 Proofs
PROOF OF THEOREM 7.5.1. From (7.5.1) and the Laplace transform of L in
Section C.5.5,
g(z) ~ J.L[(I - z)AI + Mr'Me (7.5.10)
.......
and Y itself has Before proceeding further, let us derive (7.5.10) by another method. (We shall use
hen, D also has this approach again later on.) Let U be the Markov chain whose time to absorption de
termines the distribution of L. So, the initial distribution of U is specified by the vector
J.L, and the transition rates ofU, restricted to its transient (nonabsorbing) states, are given
by - M (Note, U is not the same as Z above, which describes the supply system.)
Ilion, shifted up Consider the joint Markov chain (D, U). Define
= (8) for some
g",Adjt) ~ Pr{D(t) = d, U(t) = u' I U(O) ~ u}
G(dlt) ~ [g"Adlt)lu"'
(This includes only transient states u and u'.) By denoting G' ~ aGlat, the dynamics of
Ibin the formu
the joint process can be represented through the matrix differential equation
Dr an (r, q) pol
:>ssible to com G(010) = I
(7.5.10)
[(I - z)H + ]\,fj-I M ~ (I - P)(I - Zp)-I (7.5.12)
,' ...
This is tbe z-transfonn of a DPH distribution with parameters ('IT, P), where 'IT ~ ",P, as
claimed.
(The matrix G(dll) has rows indexed by w and u, and columns indexed by w' and u'.
Specifically, to index the rows, group together states with the same value of w, letting u
vary over all the transient states of U; the columns are indexed similarly, by grouping
Notes
states with a common Wi. As before, only transient states u and u' are included.)
Now, W and U are independent Markov chains, so the generator of (W, U), restricted
to the transient states of U, can be written in the form
QEB(-M) ~ Q®I+I®(-M) = -(-QEBM)
The dynamics of G(dll) can thus be represented as follows:
G(OIO) = I® 1
G(dIO) ~ 0 ®0 d> 0
Hence)
: second demon
g(=) ~ (1; 0 JL)(I - P)(I - zp)-l(e 0 e)
-(0) ~ u} This is the z-transform of the DPH distribution with parameters (IT, P), where Tr ~
(1; 0 JL)P' as claimed.
~ by Wi and u'.
Je of w, letting u
r1y, by grouping Notes
.eluded.)
W, U), restricted Section 7.1: The use of a queueing system as a part of a make-to-stock system seems to
have been suggested first by Karush [1957], Morse [1958]. and Scarf [1958b]. Recent
books on the theory of queues include Gross and Harris [1985], Hall [1991], Heyman
and Sobel [1982], and Walrand [1988].
Section 7.2: The same paper by Scarf [1958b1introduces the use of Palm's theorem in
this context. Many extensions ofthis model are reviewed in Nalrmias [1981]. The distri
bution of customer waiting times is discussed by Berg and Posner [1990], Higa et at.
[1975], and Sherbrooke [1975].
Exact (but intricate) results for certain world-driven demand processes (Section 7.2.2)
are provided by Ramaswami and Neuts [1980] and Scarf[1958b]. The normal approxi
d>O mation and the formula for (T are discussed by Whitt [1982,1992], Glynn and Whitt
[1991], and the references therein. The paradoxical effects of leadtime variation in this
context are discussed further by Wolff [1977].
The lost-sales model of Section 7.2.3 is by Karush [1957]. Smith [1977] presents an ap
proximation technique. Johansen and Thorstenson [1993] show how to compute the op
the probability timal policy for the case of exponentialleadtimes. For the (/; q) policies of Section 7.2.4,
Galliher et aJ. [1959] and Sahin [1983] derive exact (but intricate) results. Song and Zip
kin [1996a] evaluate the approximation (7.2.3). Yano and Lee [1995] review the litera
ture on lot-sizing models with imperfect quality.
Section 7.3: Models with state-dependent processing rates (as in Section 7.3.3) can be
found in Gross and Harris [1971]. See Whitt [1983] and Harrison and Nguyen [1990] as
well as Walrand [1988] for approximations of processing networks. Baker [1992] and
Me~ e0Me.) Dallery and Gershwin [1992] provide overviews of finite-buffer systems. The direct
processor-control system (Section 7.3.10.2) is studied in Baker [1973] and Sobel [1969].
Refinements of the exponential approximation can be fowld in Abate et aJ. [1995].
Section 7.4: The concept of an exogenous supply system and the results here come from
M)(e 0 e) Zipkin [1986a]. drawing on earlier work by Ehrhardt [1984], Kaplan [1970], and Nah
mias [1979]. Song and Zipkin [1996b] investigate policies that do use current informa
tion about the supply system.
290 Foundations a/Inventory Management
Section 7.5: The derivations involving CPH distributions use ideas from Neuts [1981] 7.4 Conside."
and Zipkin [1988]. Explicit formulas have been derived for many other distributions; sorne~
Bagchi, Hayya, and Ord [1984] and Bagchi [1987] provide reviews. See Bagchi et a!. sma11, wi
[1986] for further discussion ofthe effects ofleadtime variation. Feeney and Sherbrooke intermeel
[1966] study the system with independent Ieadtimes and compound-Poisson demand. Trearill
p~
7.5 Considell
distri~
Problems equari~
7.6 As PO'
7.1 Consider the random variable L[2J introduced in Section 7.2, the minimum of two independent Use this
copies of L. Argue that its ccdf is given by the square of Ft, the ccdf of L itself. Use this fact to parnllel
show that, if L has a uniform distribution on the interval E[L] :+: lJ, then E[L[2]] = E[L] - lJ/3. 7.7
7.2 Consider the lost-sales model of Section 7.2.3, assunting Poisson demand and independent (a)
leadtimes. Show that A is decreasing and convex as a function ofs for s 2::: o.
Hint: First, verify that A(s) satisfies the recursion
7.8
A(s) = /I(s - I)
A(s - I) + s/v
7.9
or
- - - s
_ s + 1 _ _ _ /I(s)
[ A(s) +--],1A(s) ~ [1 - A(s)]6A(s - I) -
v v
Now, show directly that ,1A(O) = -11(1 + v) < 0, and then argue by induction that ,1A(s) < O.
- s + 2 2
[ A(s + 1) + --],1 A(s)
v
- 2- - - I
~ [1 - A(s)],1 A(s - I) - 2,1A(s)[,1A(s) + -]
v
Use another induction to argue that ,1A(s) > - I/v and,12 A(s) > O.
7.3 Consider the M/G/l system of Section 7.3.2.2. This problem concerns the modeling of the unit
production time S. when the processor breaks down occasionally. Suppose the machine is always
up whenever a unit begins production, and during production the machine completes units at the
rate fLo, as in the MIMII system. While the machine is producing, however, it tends to break down
at the constant rate 13d' The repair time is exponentially distributed with mean 1113".
Model the distribution of S, including time waiting for repair, as a CPH distribution. Compute its
mean and variance in terms ofthe parameters specified above. Also, describe the distribution of L,
•:4(s) < o. 7.11 Consider the MlG/l system, where each unit is defective with probability 0, as in Section 7.3.5.2.
Assume units are inspected when they are received and defects discarded. For the case where the
original ("nominal") processing time S has a CPH distribution, show that the effective processing
time S' also has a CPH distribution, and compute its parameters.
7.12 Consider the MIMII system with defect rate 0 under the inspection scheme of Section 7.3.5.3,
where units are inspected and defects discovered only at demand epochs. The process 10 is a
Markov chain. Its dynamics depend on s.
(a) For fixed s draw the state-transition diagram for 10. Argue that the transition rates qij have the
following form:
>fthe unit A(l - o)&-H
I
i:Ss,i<j:Ss
ne is always
units at the Ao S - i
ioSs,j=s+1
) breakdown
A i>s,j=i+1
11.Compute its
ribution of L,
• f"
f"(l - 0)
0< i oS
i>s,j=i-I
s,j = i-I
292 Foundations of bl~'NltO,.y lv!a1!agement
(b) Let p = )J(~",) and p+ ~ I - (I - p)~. Show that the following prohahilities satisfY the 7.17 Suppose F
balance equations, and so describe the distrihution of /0: integrariao
"0 ~ I - p
- (I
'IT) - - P+)PP+j - l l$j:Ss
Use this 11
"j ~ (I - PJP~~-' j>s
7.18 Suppose.
(e) Show that, as asserted in the text, and Sectil
Section 6.
B(s) = (-p_)p~
I - P
specifical
qualitari"1l
mathe~
I(s) = s - ( P )(1 - p~)
I - P+
7.19 suppo~.
sequen
7.13 Consider the MiM/I system of Section 7.3.2.1, modified to describe a perishable product. Each
andb =
unit of inventory may spoil at any time, according to a Poisson process of rate 8, where 8 > O. (The
(a) Vi
unit spoils at the first event of this Poisson process, if that event occurs before the unit leaves the
(b) Use
inventory to meet a demand.) These Poisson processes are independent of the demand and supply
processes and of each other. stodi
Explain that 10 is a birth-death process, and determine its transition rates Al and ~i' Express the (c) U~
cost.
P, in terms of the ratios P ~ AI", and T] = S/A. Show that, for i 2'" s, R, ~ p'II;~l(1 + T]J)' Also,
compute the R, for i < s, and R itself. (d) N~
7.14 Several situations are descrihed helow. The issue in each case is to decide which type of supply pool
system model to use: parallel, limited-capacity, or exogenous-sequential. Specify which kind you 7.20 Consi~
think is most appropriate, and explain why. (There need not be a single right answer. If you are pararnetl
also~
unsure, describe what additional information you would seek to help decide the question.)
(a) You operate a small machine shop which makes spare parts used throughout a larger factory.
One essential part is demanded frequently, and making it consumes a considerable fraction of
your capacity. 7.21 Considll
(b) You order the raw materials used in a chemical process. One of the materials is supplied by each.~
many different chemical companies, and you routinely spread your orders among these servl~
sce~
companies. It is not uncommon for your orders to cross; that is, an order may be received from
one company before another order placed later with a different company. There is no descnlJl
systematic, predictable difference between the companies' order-response times, however. proces:!l
(c) You operate a small retail clothing store. 'tou obtain one of your largest-selling lines from a
inanm
hi~~
single manufacturer. This supplier is large, with annual sales of several hundred million dollars.
You are quite sure this company's president has never heard your name.
7.15 Suppose D is a Poisson process with A ~ 3, and L has a ganlllla distribution with E[L] ~ 2, V[L 1 =
2. Determine the parameters of Land D. Suppose we wish to follow a base-stock policy with A(s)
aSbe~
conte
::; 0.05. What is the smallest value of s that achieves this performance requirement? (You can do cong
the calculation by hand or by computer, using a spreadsheet program.)
7.16 Suppose M is the transition matrix of a CPH distribution. That is, M is a square, nonsingular
matrix, whose otT-diagonal entries are nonpositive, and Ale 2:: O. Show that JvI1 2:: O.
Define P = A(X/ + M) -I. Show that P is the transition matrix of a DPH distribution: P 2'" 0,
atisfY the 7.17 Suppose FO is the ccdf of a gamma distribution and.fits pdf. Defining F J as in the text, use
integration by parts to show that
19u1ar
I:P~O,
C H A P T E R
Outline
8.1 Introduction 295 8.6 Distribution Systems: Central
8.2 Independent Items 298 Control 339
8.3 Series Systems 302 8.7 Shared Supply Processes 348
8.4 Assembly Systems 323 8.8 Kanban Systems 352
8.5 Distribution Systems: Local Control Notes 359
331 Problems 361
8.1 Introduction
The last two chapters focus on a single product at a single location. Now we shift atten
tion to structurally complex systems, as in Chapter 5, where several items (products
and/or locations) are linked to form a network. In contrast to Chapter 5 (but as in Chap
ters 6 and 7) the demand process and the supply system may include random factors.
Such a network is sometimes called a supply chain. There has been a surge of in
terest recently in supply-chain management, as many companies have realized substan
tial benefits by restructuring their supply chains. To do this, one must understand how
alternative control schemes and system designs work. This chapter summarizes the best
analytical tools available to support supply-chain management at this conceptual level.
A supply chain is a type of business process. (This phrase also encompasses pure
information-processing systems, where physical goods play a marginal role.)
Business-process reengineering is a careful, systematic approach to network restruc
turing in order to enhance performance. Again, this chapter provides tools to support
such endeavors.
295
---------
-
We begin in Section 8.2 with a system composed of many independent items, each
facing stochastic demand and supply conditions. The results generalize those of Sec
tion 5.2 of Chapter 5: We characterize overall system performance with the aid of a few
simple aggregate statistics. These include the variety index J. and a systemwide index
of variation (the cost-weighted sum ofthe items' individual <T's). These results allow us
to measure precisely and to understand intuitively the benefits of item consolidation.
The rest of the chapter treats systems whose items are linked, mainly through
supply-demand relationships. We focus on a few special structures, namely, series,
assembly, and distribution systems. Also, we discuss shared supply processes.
One key distinction among such systems concerns the flow of information and con
trol. (We touched on this issue in Chapter 5, but here it becomes even more crucial.)
When a structurally complex system is driven partly by random events, we need to un
derstand clearly just who knows what and when, and how that information is used to con
trol the various parts of the network.
The simplest case, conceptually, is a centralized system. Here, all relevant information
in the network flows to a single point, where all decisions are made. These decisions are then
transmitted throughout the network to be implemented. In a sense this situation is ideal;
what could be better than a fully informed decision Illilker with full control over the system?
In practice, however (as mentioned in Section 5.8), centralization may be impracti
calor even dysfunctional: It requires a fast, reliable communication system, a powerful
infonnation-processing capability. and an organization willing and able to act in syn
chronized fashion. For a host of technical. economic, and cultural reasons, these ele
ments are often lacking. To impose centralized control where it cannot be supported
properly can have perverse effects.
For this reason we also consider relatively decentralized systems, where infonnation
and control are distributed throughout the network. We say "relatively" because, while
the inforllliltional requirements of such systems are less formidable than those of a fully
centralized system, they are fairly stiff all the same. There are many ways to decentral
ize, but not all of them work. For one thing, lacking a central authority, there must be
some alternative mechanism to avoid chaos. All the control policies we consider are
carefully sculpted to achieve reasonable degrees of coordination.
Sometimes, there is a clear reason to centralize infonnation and control. Central
ization allows us to balance the items' inventories in situations where otherwise, under
purely local control, imbalances would arise. The exact meaning of balance depends on
the context, but this general notion is broadly valid.
We begin with series systems in Section 8.3, focusing on those with no scale
economies in supply. The root problem here, just as in a single-stage system, is the com
bination of demand uncertainty and supply leadtimes; again, this leads to a tradeoff be
tween inventory and backorders. But now we have several possible stocking points, of
fering different degrees of stockout protection at different costs. We must decide not just
how much inventory to keep, but also where to position it.
Interestingly, the distinction between centralized and decentralized systems disap
pears here. We start with a simple, plausible local control scheme, using a base-stock
policy at each stage, and show that it is equivalent to a fully centralized system. Thus,
centralized control can be implemented in a decentralized fashion.
ChapleT 8 St.....'eral Items with Stochastic Demands 297
~t items, each We develop algorithms to evaluate any such policy and to select the best one. For
those of Sec the two-stage case we derive simple bounds on the optimal policy variables and the op
Ie aid of a few timal cost. These results, along with some illustrative examples, help us understand in
emwide index tuitively the main drivers of effective control and system performance.
~ul ts allow us We obtain these results first for a basic model with Poisson demand and constant
O5olidation. leadtimes. Later, we find that the same approach works, exactly or approximately, for
ainly through compound-Poisson demand and all the supply systems of Chapter 7.
amely, series, Then, we consider a system with fix.ed order costs. We start with a simple case,
cesses. where external orders incur fixed costs, but internal shipments do not. Most of the ear
arion and con lier results extend directly to this system. Here, the first stage uses an (r, q) policy, while
1Il0re crucial.) the others use base-stock policies. Again, centralized and decentralized control are
re need to un equivalent. We then tum to the general case, where internal shipments too incur fixed
is used to con costs. This system is more complex than the earlier ones. In particular, centralized and
decentralized control are not the same. (Actually, centralized control is more general.)
nt information We develop a policy-evaluation method which covers both cases and a partial optimiza
isions are then tion algorithm, to compute the best reorder points given fixed order quantities.
l3tion is ideal; Section 8.4 extends these ideas to assembly systems. Remarkably, an assembly sys
erthe system? tem can be reduced to an equivalent series system. Thus, we can use the methods of Sec
~ be impracti tion 8.3 for policy evaluation and optimization. The policy we end up with. however,
Ill, a powerful cannot be implemented with local infonnation alone. We can control each item locally,
£0 act in syn but we require explicit information about others, in order to balance the items' invento
05, these ele ries properly. Of course, it is possible to use a decentralized control scheme in tlus con
be supported text, say a base-stock policy, though this approach is more costly. We show how to eval
uate such a policy, again using series-systems techniques.
e information Next, we study a two-level distribution s.l/stem, where one location (called the
ecause, while warehouse) receives goods from the outside source, and supplies in tum several oth
lose of a fully ers (the retailers). In structure this system is a hybrid of a two-stage series system and
to decentral an independent-item (or parallel) system. This viewpoint is a fruitful one-we can in
here must be terpret the operation and performance of the system in similar terms, as combining el
consider are ements of series and parallel systems.
Here, even without fixed costs, the centralization-decentralization issue does mat
trol. Central ter. We treat local control in Section 8.5. Each location follows a base-stock policy, and
~ise, under the warehouse fills retailer orders on a first in, first out (FIFO) basis. We show how to
e depends on evaluate such a policy and to determine the best one. This approach extends directly to
arbitrary distribution networks and to systems with fixed external order costs.
,;th uo scale The analysis reveals that warehouse inventory serves two distinct economic func
~ is the com tions. It offers a low-cost fonn of stockout protection and a means of partially consoli
I tradeoff be dating the retailers, i.e., of pooling their leadtime-demand uncertainties.
19 points, of Section 8.6 discusses a distribution system operating under centralized control. We
;x:ide not just develop an approximation aud a heuristic policy. One part of the policy is myopic allo
cation, a simple technique to divide outbound shipments from the warehouse among the
stems disap retailers. The balanced approximation reduces the model to a two-stage series system,
a base-stock which we solve with the methods of Section 8.3. This yields both an estimate ofthe op
~"Stem. Thus, timal cost and the remaining elements ofthe heuristic policy. The approx.imation and the
policy work quite well, especially for systems with high demand volumes.
298 Foundations ofInventory Management
We find that, in general, a system with centralized control needs less warehouse in
ventory than a local-control system. The centralized system is able to substitute infor
mation for inventory to achieve some degree of retailer consolidation.
Section 8.7 discusses two kinds of shared supply processes. First, we solve a simple
joint-replenishment problem. Then, we turn to systems with shared production capacity.
These are similar, but not quite identical, to distribution systems. The section concludes
with a general discussion of setup costs in systems with limited reSQurces.
Finally, Section 8.8 explores variants of the base-stock policy for networks with
limited-capacity processors. These policies are based on the celebrated kanban system.
We find that, although these policies are more intricate than a base-stock policy, the sim
ple models of Section 8.3 capture the fundamental economics of the system. Then, we
discuss these ideas in the broader context of the just-in-time approach to operations.
cA = ~:.J c·A·
J .I
=;ci~
pB = aggregate sales-value of backorders
= 'iiPjBi
where 1.J andBj denote average inventory and backorders of item j.
Chapter 8 Several Items with Stochastic Demands 299
i warehouse in To study the relationship between cI and pB, we adapt the individual-item approach
ubstitute infor of Section 5.2.3. That is, specify two positive constants, 'Tl and 13, and use these along
with the coefficients in cI and pB to define each item's cost factors:
: solve a simple
hj = TlCj bj = 13Pj,j ~ I, ... ,J (8.2.1)
Iction capacity.
:tion concludes Then. compute the optimal base-stock policy for each item separately. This determines
:So the ~ andEj and thus eI andpB. This pair (el, pB) is efficient; that is, there is no smaller
networks with pR consistent with cI Vary the ratio i31'Tl and repeat this procedure to obtain other effi
ranban system. cient pairs. The locus of such pairs describes an aggregate inventory-backorders trade
(IOlicy, the sim off curve. This construct is very useful to managers, in the same ways as the tradeoff
item. Then, we curve of Section 5.2.
operations. This approach works, but it offers no broad insights into system behavior; in gen
eral, it cannot be reduced to an elegant formula like (5.2.4). For that, we must impose ad
ditional assumptions: First, suppose the markup ratio p)cj is the same for all items. Then
(as in Problem 5.4), we can use
:em, as in Sec
apters 6 and 7: cR = aggregate cost-value of backorders
"'Pply system. = lj cjIfj
:Jerformance in
Wich explains as the aggregate backorders measure, for it captures the same information as pB, and
respecify the backorder costs as bj = 13e)' (The meaning of 13 has changed now.) Also,
suppose that each item's leadtime demand is normally distributed (approximately, as in
Section 6.4).
Following the approach of Section 6.5.2.2 leads to
w.=_J_=
b. 13e.
j
13
--==:w
J bj+hj 13 cj + 'Tlej 13 + 'Tl
for all items. Consequently, all items share the optimal standardized base-stock level
z* = (<p 0)-l(l - w). Thus, ~ = crj<P'( -z*), Ej ~ crj<P\z*), and
where
cO' = ~ CfIj
IlCk policy for
(The maximal approximation leads to similar results.)
:age inventory
To compute the tradeoff curve, then, we can bypass the specification of'Tl and 13; just
measures:
let z* range over all real values, and directly calculate (eI, eR) via (8.2.2). Notice, the re
lationship depends on the system data only through the aggregate parameter ecr, and the
dependence is homogeneous; that is, cO" acts as a scale factor for both eI and cB.
Figure 8.2.1 shows one such tradeoff curve, computed using the normal approxi
mation. Here, cO" is about 2.2. Figure 8.2.2 plots the same data, using a logarithmic scale
for eR, to show more clearly the impact of reducing eR to a very small value. This ap
proach is similar to the stock-service tradeoff figures of Chapter 6, but at an aggregate
level.
300 Foundations ofInventory Management
FIGURE 8.2.1
Aggregate inventory-backorders tradeoff.
8
';J 4
o
o 2 3 4
cB
FIGURE 8.2.2
Aggregate inventory-backorders tradeoff (logarithmic scale).
8
';J 4
o mm TTTTTTT
Aside from the constant purchase cost cx., the total cost also is proportional to CIT.
The aggregate parameter CIT, then, expresses the essential effect of multiple items,
something like the variety index J. of Section 5.2. Specifically, we cannot predict per
formance solely on the basis of an aggregate-demand statistic such as cA. For example,
if each item's demand process is Poisson, and all items share the common fixed leadtime
L. then u j ~ V¥, so cu = ('{i)'j;jc)~. To estimate performance, then, requires in
formation about the distribution of the 'Aj over the items. Again, because of the concav
ity of the square-root function, we expect a system with few items haviog large Aj to cost
less than one with many low-demand items. (Remember, nevertheless, there may be
solid marketing arguments for a differentiated product line.)
This point is clearest in the case of identical items. Suppose we keep the total de
mand rate A constant, so cA is just the product of A and the (common) cost rate c. Each
A} = )JJ. so u; = yUIJ and cu = cy JAIL. Thus, apart from the constant term CA, the
total cost c* IS proportional to the square root of the number of items.
4
8,2,2 General (,; q) Policies
Now, suppose there are scale economies in supply, and we follow an (r, q) policy for each
item. We need another aggregate performance measure (first introduced in Section 5.2):
wO ~ aggregate average workload
~ 2ywpFj
Let us extend the individual-item approach above: Specify K, the unit cost of the re
sources measured in wO, and the cost factors
k; = KW; j ~ 1, ... ,J
io addition to 'T], [3, and the hj and bj defined in (8.2.1). Then, find the optimal (,; q) pol
icy for each item separately and combine the results to detennine an efficient point (cI,
pB, wO). By repeatiog this process for different values of K/'T] and [3/'T], we can trace out
an aggregate tradeoff surface, a two-dimensional surface in the three-space with coor
dinates (cI, pB, wO). This surface describes the best (smallest) possible value of anyone
of the measures, given the values of the other two.
Under the simplifying assumptions above, we obtain simpler results: Use the upper
bound in (6.5.9) to approximate the optimal cost for each item. The (approximate) total
cost becomes
C* = I j [CjAj + C oj + CkJ
where
IT
Suppose that each stage starts with inventory si and an empty supply system. Then,
each customer demand at stage J immediately triggers a demand at stage J - I, which
(8.24)
in turn generates a demand at stage J - 2, and so on. In this way demands propagate
backward through the system, all the way to the external source. Thus, every stage ex
rmance. The ag periences the original demand process.
1 the base-stock Our goal is to evaluate such a policy. Define for t ;=: 0
nninistic model
the constant CAl li(t) = local inventory at stage)
BJ(t) ~ local backorders at stage)
INi(t) = local net inventory at stage)
= li(t) - B;(t)
10j (t) ~ Inventory on order at stage)
10Pi (t) ~ local inventory-order position at stage)
= INJ(t) + lOit)
, Leadtimes
I1j(t) ~ inventory in transit to stage) (number of units in stage)'s supply
e numbering of system)
;ide source. All ITPi (t) ~ local inventory-transit position at stage)
Each stage has ~ INJ(t) + I1j(t)
wrd stage j (or
The "orders" in the definitions of IOit) and 10P;(t) refer to those generated by the lo
:h stage j's sup cal base-stock policy. The other variables describe actuaL physical quantities. Notice the
is stochastic; it
difference between stock on order 10j (t) and stock in transit I1j(t): [O/t) includes all
outstanding stage:i orders, but I1j(t) includes only those that stage) - 1 has been able
[)isson process
to fill already; i.e., those that have begun their passage through stage j's supply system.
IDes, Lj. There
The difference is precisely the local backorders at) - 1. Of course, [T L(t) = 10 L(t), be
ateh orders or
cause the outside source responds inunediately to orders from stage 1. Thus, setting Bb(t)
= 0, for all) and t,
ingle stage. In
lO/t) - I1j(t) ~ lOP; (t) - ITPi (t) ~ Bi- L(t)
chelan. As we
oolext. Each stage's inventory-order position remains constant at IOPj(t) = sj. but the inventory
ch stage mon transit position ITP;(t) changes over time.
ld responds to There are several possible approaches to performance evaluation. One, parallel to
e-stock policy those of Sections 6.2.2. 7.2, and 7.3 in earlier chapters, focuses on the supply-system oc
cupancies I1j(t). The definitions above imply
Bi(t) = [-INi(t)] +
:Sj)j=l'
= [10j (0 - 10Pi(0]+
£tion. It expe
ion operating = [Bi-l(t) + I1j(t) - si]+
J - I treats
So, given the I1j(t), the random variables B;(t) satisfy the recursion
:nts, to arrive
'S to be filled Bb(t) =0 (8.3.1)
mrine the or
ods, et cetera.
Bj(t) = [B;_ L(t) + I1j(t) - s; t
o they arrive Now, it is difficult to characterize the I1j(t) directly. So, we shall utilize a different ap
proach here. [There will be other uses for (8.3.1) later on.]
...
.
LetLo ~ 0 and
Lj = Lt-:s;,jL;
This is the backward echelon leadtime for stage j (not the forward echelon leadtime used
in Chapter 5). Rewrite (8.3.2) at shifted times:
The interval (t + Li - h t + LJ has length Lj, and the intervals are disjoint. Therefore,
D(t + Lj - l , t + L,] in (8.3.3) has the Poisson distribution with mean ALj, and these ran
dom variables are independent. This distribution does not depend on t, so the Bj(t + L)
are also stationary.
To describe the equilibrium behavior of the system, therefore, we can simply omit
the time indicators in (8.3.3). Let
Dj = leadtime demand for stage j, a generic random variable having the Poisson
distribution with mean XL;.. these Dj are independent
Then,
Bo~ 0
;ervation-of-flow We know how to compute E[B;] and V[B{], since D] has a Poisson distribution. Suppose
we have estin>ates of E[B}_I] and V[B}_l] forj '" 2. Since B}-l is independent of Dp
(8.3.2) E[B}_I + DJ ~ E[B}_d + E[Dj ]
V[Bi-l + Dj ] ~ V[Bi-l] + V[Dj ]
Using these quantities, approximate BJ-l + Dj with a negative-binomial distribution, and
then compute E[B}] and V[B}] through (8.3.4) and the methods of Section 7.5.1.1. Con
tinue in this manner to j = J C'Ne can use the negative-binomial distribution here, because
every V[B}] '" E[B}]; see Problem 8.1.) This approximation, it turns out, is quite accurate.
>n leadtime used We also need E[I1j]. On average, we send A units into stage j's supply system per
unittime, and each unit stays there for tin>eLj, soE[I1j] ~ AL}. Notice, this is also E[DJ
To evaluate a policy in economic tenns, we specify cost factors:
hJ =inventory holding-cost rate at stage j
b = backorder cost rate at stage J
(There may also be a variable order cost c, but the average order cost is then just cA, a
constant, so we ignore it. Likewise, we ignore constants representing the variable costs
I~ (8.3.3) of shipments between stages. On the other hand, we apply the holding cost hJ not just to
actual inventory at j but also to stock in transit to j + I; on average this "pipeline hold
ioint. Therefore,
ing cost" is just the constant h}E[ITj + d ~ hiALj+ I') The total average cost is thus
;, and these ran
;0 the B}(t + L) E [lJ~lh}(Ij + 11j+l) + bBj] (8.3.5)
where ITJ + 1 ~ O.
:an simply omit
Let us review: The main result so far is (8.3.4). This recursion concisely expresses
the linkages between stages. Each stage j operates much like an isolated single-stage sys
g the Poisson tem. The random variable Bj-l + D j plays the role of the leadtime demand; the local
base-stock level s; reduces stage-j backorders while increasing inventory in the usual
way. What links the stage to its predecessors is the term B;-I' which augments its "own"
leadtime demand Dj •
Reflect for a moment on the managerial issues here: Customers see only the final
stage backorders B}, and only those incur cost directly in (8.3.5). Only final-stage inven
(8.3.4)
tory provides direct protection against these backorders. The inventories at prior stages
have indirect impacts; reducing B; reduces B;+l' which reduces B;+z, and so on through
Bj. (Problem 8.2 asks you to formalize this notion, specifically, to show that each E[B}] is
nonincreasing and convex in all the sr) Although direct stockout protection is more effec
tive, it costs more; typically, the h} are increasing inj (as explained in Section 5.3.2).
To manage the system effectively, then, we must address some rather subtle trade
rimary interest. offs. The overall issue remains the balance of inventory and backorders, but there are
!)a Poisson dis now alternative places to hold inventory with different costs and benefits. Clearly, we
,f D I • Next, we s;
cannot choose the independently, for if we keep more stock at one stage, we need less
~ to get B2. The at the others.
As we shall see now, there is another way to look at the system's dynamics and costs,
I and V[Dj ], and which expresses these interactions in simpler terms. This approach sets the stage for the
I ~ V[Bo] = O. policy-optimization method developed later.
306 Foundations oflnventOlY Management
= Bj(l)
ITPI(I) ~ SI
By (8.3.7) this algorithm does compute the functions correctly. The average cost (8.3.8)
is precisely C(s) ~ CL(slls).
At termination, set s* = (sj) and C* = C1(s!). We prove below that these quantities de
scribe the optimal policy and the optimal cost.
Evidently, this recursion is nearly identical to the policy-evaluation algorithm above,
but there the Sj are predetermined constants, while here the sj emerge within the calcu
lation. (If the minimum over C/y) is not unique, select the smallest minimizing value.
Also, c.;
may be a nonincreasing function; this happens when hj = O. If so, set sj = 00,
As discussed above, an infinite S.i does have a sensible meaning. In that case {;.j = Cj-)
The construction of C j from CJ is illustrated in Figure 8.3.1.
In general. the optima]~policy vector s* need not be nonincreasing. As in Sec
tion 8.3.2, however, it is equivalent to a nonincreasing vector s-*, where sJ--:* =
min'''j {sn. This policy, and hence the corresponding local policy s", can be deter
mined only when the entire algorithm is completed. Only then can we see which stages
hold stock and which do not.
Now, let us prove that the algorithm does find the best policy:
LEMMA 8.3.1. For allj, Cj and £:j are convex fimctions. Also, tor any fixed echelon
base-stock policy s, CAls) ;", CAl and £:,('Is) ;", CA)·
PROOF. We argue by induction onj. Certainly, £:J+ Lis convex and.(1+ 1('ls) = CJ+I'
So, suppose that C j + I is convex and £:j+ I (-Is) ;", C j + Lfor any j ;", 1. First, Cj is clearly
convex, and the expectation defining Cj preserves convexity, so Cj is convex. And since
Cj is convex, so is £:}" (This is clear from Figure 8.3.1: C;Cv) is decreasing and convex
for y < sj and constant for y ;", sJ.) Also, from C J + I ( 'Is) ;", £:j+lO, it follows immedi
ately that C;(-Is) ;", C;('), and this implies
(The last inequality is also evident from Figure 8.3.1: If we use some other value of
Sj besides sj to construct Cj , we end up with a function that is greater than Cj itself.)
This completes the induction.
Chapter 8 Several Items with Stochastic DernaudY 309
(8.3.9)
~ quantities de
(J (y)
gorithm above,
thin the calcu
imizing value.
;0,set sj = 00,
=e {;j ~ Cj .)
g. As in Sec
where s)* = ,
"J
can be deter
, which stages
THEOREM 8.3.2. The policy determined by s* is optimal among all base-stock policies.
fIXed echelon
PROOF. Comparing the algorithms in this and the previous subsections, it is apparent
that C I (st) is in fact C(s*), the average cost ofthe policy s*. And, the lemma tells us that
,('!s) = {;J+l' no other policy has lower cost.
t, (;j is clearly
ex. And since Recursion (8.3.9) deserves to be called the fundamental equation of supply-chain
g and convex theory. It captures the basic dynamics and economics of serial systems.
lows immedi It is remarkable that such a simple recursive scheme works at all: To determine sj.
we ignore nearly everything about upstream stages. We need to know the prior slage's
holding cost hi-l to compute hj' but that is all; sJ is completely unaffected by the Db
i < J. (Still, we don't know what Sf means in local terms until we compute all the sf.)
We can think of each C; as the average-cost function of a single-stage system. This is
literally true for j ~ J; the cost rates may seem odd (hJ plays the role of the holding cost
and b + hi-l the penalty cost), but the form is correct. For j < J the term involving {;j+ I
replaces the usual penalty cost; 1:..;+ 1 is sometimes called the implicitpenalty-cost/unction.
ther value of How hard is it to implement the algorithm? It is helpful to know that every Cjis con
an {;; itself.) vex, so it is relatively easy to search for sJ. The evaluation of C/y) = E[C/y - D i )] for
j < J, however, is hard. It requires a direct numerical calculation.
310 Foundations ofInventory MUllagement
of demand can be filled separately. All the arguments and results above remain valid.
Here, each Dj has a eompound-Poisson distribution, but that is the only difference.
Suppose that each stage has an exogenous, sequential supply system, as in Section 7.4.
Let Lj(t) denote the virtualleadtime for stage j; a shipment to j begun at time t arrives at
t + Li(t). Assume these systems are independent of one another, as well as of the de
mand process, so the Lj(t) are independent over}. Let Li be the equilibrium leadtime
o
B = 0
BJ = [BJ-I + Dj - sJJ+
The Dj have new meanings, however. Here, Dj has the distribution of D(LiJ, the demand
over a random interval of time Li, like D in Section 7.4. For Poisson demand,
2
E[Dj ] = hE[Lj] V[Dj ] = hE[L;] + X V[Lj]
as in (7.5.7). Moreover, the D j are independent. Similarly, (8.3.7) evaluates an echelon
base-stock policy directly, and therefore algorithm (8.3.9) determines the best such pol
icy. In sum, policy evaluation and optimization can be carried out just as in the fixed
leadtime case, using the new Dj in place of the old ones.
When each Li has a continuous phase-type (CPH) distribution, the recursion (8.3.4)
can be performed exactly by using matrix-vector operations. See Problem 8.4.
Chapter 8 Several Items with Stochastic Demollds 311
in Section 7.4.
B; = [Bj_l + !T.J - sJ]+ (8.3.10)
IN; ~ ITPj - I~
•.4); that is,
ITPj + 1 = min {S:i+b IN)
Moreover, according to the approximation, the IT; are independent, just like the Dj . So,
using Dj to stand for the approximate I~, we can apply as is the methods of policy eval
), the demand uation and optimization developed above. In particular, algorithm (8.3.9) finds the (ap
Old, proximately) best base-stock policy.
We remark that a base-stock policy need not be optimal here, in contrast to the
constant-leadtime model. A base-stock policy often works well, but the true optimal
~ an echelon policy is generally much more complex. We consider a more general (though still sub
lest such pol optimal) type of policy in Section 8.8.
in the fixed
8.3.4.4 Independent Leadtimes
rrsion (8.3.4) The same approach works well when each supply system consists of multiple identical
8.4. processors in parallel, so the leadtimes are independent, identically distributed random
--
variables. The recursion (8.3.10) remains valid. As mentioned in Section 7.2.7, when
s = s' = 0, the I1j are independent, and IIj has the Poisson distribution with mean
E[llj] ~ AE[LjJ. We use this same distribution as an approximation in the general case.
According to this approximation, then, the perfonnance of any policy is the same as
in a constant-leadtime system; only the E[Lj] matter. As in Chapter 7, this is very dif
ferent from the behavior of a sequential supply system, where the leadtime variances
contribute significantly to performance.
8.3.5 Illustrations
This subsection presents some numerical examples, to provide insight into the behavior
of the optimal policy.
We assume Poisson demand and constant leadtimes. Without loss of generality, we fix
the time scale so that the totalleadtime is L ~ I, and the monetary unit so that the last
stage's holding cost is h; = 1. The stages are spaced symmetrically, so each stagej's lead
L;
time is = 1/J. We consider four numbers of stages, J = 1,4, 16, 64; two demand rates,
A ~ 16,64; and two penalty costs, b ~ 9,39 (corresponding to fill rates of90%, 97.5%).
We consider several forms of holding costs hJ, depicted in Figure 8.3.2. The sim
plest form has constant holding costs, where all hJ = 1. Here, there is no cost added from
source to customer. This is a rather unrealistic scenario, but it is a useful starting point
to help understand other forms. The linear holding-cost form has h; = j1J, or hj = 1/J.
Here, cost is incurred at a constant rate as the product moves from source to customer.
This is quite realistic. Affine holding costs, where hi ~ a + (I - a)jlJ for some a E (0,
t), are even more realistic. Here, the material at the source has some positive cost, and
the system then adds cost at a constant rate. This fonn is a combination of the constant
and linear forms. Here, 0' = 0.75.
Chapter 8 Several Items with Stochastic Demands 313
'""
0
u
;tage i's supply
mediately and ~'" 0.5
; like this with
3.6.) Here, we
..
-0
.c
u
0
,..l
Kink
m implicit de
0.25
filled demand
network with
I"words, com o
. lbis approx o 16 32 48 64
Stagej
ut internal in
d parallel sup
",city sequen
The last two fonns represent deviations from linearity. The kink fonn is piecewise lin
ear with two pieces. The system incurs cost at a constant rate for a while, but at some point
shifts to a different rate, which remains constant from then on. Here, the kink occurs halfWay
through the process, at stage J12. So, for some {H (-I, I), hj ~ (l - a)IJ,) ~ J12, and hj
• the behavior = (l + a)IJ,) > J12. Again, we set a = 0.75. Finally, in the}ump fonn, cost is incurred at
a constant rate, except for one stage with a large cost. Here, the jump occurs just after stage
JI2. So, hj ~ a + (l - a)IJ,) ~ JI2 + I, andhj ~ (l - a)/J otherwise,for some a E (0, I).
We can view this as linear cost before JI2 and affine cost after. Here again, a = 0.75.
:rality, we fix
) that the last 8.3.5.2 Optimal Policy
~ge j's lead For constant holding costs the optimal policy is simple: For} < J, sJ* ~ 0; only the last
Iemand rates, stage carries inventory. Stage J, in effect, becomes a single-stage system with leadtime
10%,97.5%). L. The optimal policy is the same for all J This is also the optimal policy for J = I un
1.2. The sim der any other holding-cost form.
tadded from Figure 8.3.3a shows the optimal policy s* for linear holding costs, J ~ 64, and two
;tarting point values each of A and b. Several observations are worth noting: The curves are smooth
: or hj = l/J and nearly linear,' the optimal policy does not lump inventory in a few stages, but rather
to customer. spreads it quite evenly. The departures from linearity are interesting too: The curves are
some a e (0, concave. Thus, the policy focuses safety stock at stages nearest the customer.
rve cost, and Figure 8.3.3b shows the optimal policy for affine holding costs. For} > 1, the curves
the constant follow the same pattern as in the previous figure. (Indeed, the curves for b ~ 9 here are
identical to those for linear costs and b = 39, because these two cases have identical
...
FIGURE 8.3.3
Optimal policy. (a) Linear holding costs; (b) affine holding costs; (c) kink holding costs;
(d) jump holding costs.
90
80
b~9
b~ 39
70
" 60
~ 1. ~ 64
~
-"0 50
i~ 40
~
"E
".p
30
0-
20
R::
1. ~ 16
10
(a) Stage)
90
80
~
- b~9
70 { - b~39
•
" 60
""~ I ~ ......... 1. ~ 64
-"0 50
ii/,
.
~
~
.jj 30
40
0- 1. ~ 16
20
'J::::-::::
10
0
(b) Stage}
',11111,'1'111111111"'1'1'
g costs;
- •
90
80
"
"0
60
.§
oj
0
50
l '-, l.~64
~
..
~
.D
S
= 30
40
8
~ l. ~ 16
20
10
I
0 J
(c) Stagej
90
80
~ ~
_ b~9
70 - b=39
•
~ 60
.§ I '-- "' l.~ 64
1 50
."
] 40
8s 30
20
10
~
-- l. ~ 16
0
(d) Stagej
".
ratios h)(b + h;),j > 1.) However, the curves break down sharply atj ~ I (because hI
is large). Therefore, the equivalent policy s- is flat for small j, and so the policy holds
no inventory at early stages. This solution is intermediate between those for constant and
linear costs. As ex. increases and the costs move upward, stocks shift toward the customer.
The total system stock decreases slightly. But, perhaps surprisingly, stocks near the cus
tomer actually increase.
Figure 8.3.3c displays s* for kink holding costs. Downstream from the kink [before
algorithm (8.3.9) encounters it], the curves exhibit the same pattern as in the linear case.
Upstream from the kink, the policy again follows the linear pattern, ahilost as if the kink
were the last stage. The net result is substantial stock at and just before the kink, where
holding costs are low relative to later stages.
Finally, Figure 8.3.3d displays s* for jump holding costs. From the jump on, the pol
icy behaves much as in the affine case-smooth, concave decrease beyond the jump, but
a sharp break downwards at the jump. Upstream from the jump, the policy again follows
the pattern of the linear case. Thus, there is substantial stock just before the jump and
none just after it.
Ie kink [before 70
the linear case. •0
;t as if the kink 60
oe kink, where "-"
>
-"
u 50
.9
~
np on, the pol "~ 40
~ --J~64
l the jump, but -;;
. again follows 6
"R 30 --J~16
the jump and 0
20 - -J~4
10 -J~l
:osts, A = 64,
the restricted 0
)Oint is nearly 0 16 32 48 64
stem stock is Stage)
wte slowly, as
;t is even less
This is the cost function of a single-stage system, so it is easy to optimize. The stage-I
, sigoificaotly
calculations are harder and less transparent. We shall derive simple bounds on Cj (y), sf.
ckiog point at
and the optimal cost C*, aod use them to gain additional insight into stock positioning.
isplays a sim
Suppose we forbid stage I to hold inventory. That is, we restrict the feasible poli
I the kink (or
cies, fixing si = O. In echelon terms, we force 82 = 00, leaving S1 free. Therefore) 2 (x) .c
becomes Cz(x) itself. Makiog this substitution in CI and then C I , we obtain the ap
rock position
proximation
ocking points
C~(y) = E[hb D I) + Cz(y D I )]
forA = 16in
= hIE[DzJ + h~(y - E[D]) + (b + h~)E[[y - Dr]
proportional
Ie cumulative where D = D] + D2 . This is the cost function of a single-stage system. Let s i denote
,for A = 16. the optimal value ofy, and C+* = C~(s~).
Since we restricted the feasible policies, C+* is an upper bound on C*. Also, .6..C2 (Y)
2: ilC 2(y), so ilC~(y) 2: ilCI(y) for ally. Consequently, s ~ provides a lower bound on sf.
Now, assume that hz > 0, so s! is finite. Observe, C~Cv) coincides with Cj(y) for y
::::; s~. So, suppose it turns out that s r:: :;
sl Then, st = s ~ ::::; sl In this case, the corre
sponding local policy has s;* = 0, i.e., stage 1 holds no stock. Otherwise, s~ < s i ::;
st,
so si * > 0, and stage I does hold stock.
Thus. l'Fe can determine whether or not stage 1 should carry inventory by solving
two single-stage systems, one each for s~ and s ~. If not, we have actually solved the
I"
FIGURE 8.3.5
Optimal cost. (aJ Linear holding costs; (b) kink holding costs.
20
I
16
l.~64
•U 12
·0
0
l"
0 & l. 16
4
b~ 39
=b~9
o ---1
o 4 16 64
(a) Number of stages J
20
16
•u 12 ~ ~ l.~64
·0
0
"a.§ &
0
l.~ 16
4 J _b~39
-b~9
0 .J
4 16 64
(b) Number of stages J
.·III'ITIII
,
"'!:'!!li~I~~~I~
original model. If so, we have at least a lower bound on the amount of stock stage I
should keep.
Some simple sensitivity analysis reveals how the distinction works: First, fix all pa
rameters, including h 2, but let h j vary between 0 and h2(so h2 = h z - hj). As h j changes,
so does s!. and in the same direction, but s~ remains fixed. Indeed, as hI ---7 h;, s1 ---7 00,
so s ~ :5 s1 for sufficiently large hI' Conversely, as hI --70, h z --7 h 2, and since D is (sto
r
chastically) larger than D 2 , s > s! for sufficiently small hj.
Thus, stage 1 holds inventory when (but only when) it is significantly cheaper to
hold it there than at stage 2. The required difference, of course, depends on the other
parameters.
Next, focus on the constant-Ieadtime case. Fix all parameters except L;. Only s 7 is
t
affected by changes in Li, not s!. When L; is small, D is near D z, so S :5 s1. Conversely,
r
for sufficiently large Li, s > s!. (This is true for any value of L 2, but of course the dis
L;
tinction depends on L;. One can show that, as L 2 grows, the value of required to make
--- r
s 2 s1 grows also.)
That is, stage 1 holds inventory when (but only wheni its leadtime is substantial rel
ative to stage 2 s. Otherwise, if its leadtime is negligible, inventory there serves no use
ful function. (A similar conclusion describes stochastic-leadtime models.)
---
Finally, applying the normal approximation to Cr,we see that the optimal cost is
bounded above by a linear function of (T, where (T2 = V[D]. Thus, the system parame
64 ters affect performance much as in a single-stage system.
Clearly, stage I's (echelon) inventory-order position lOP I behaves just like that of
a single-stage system, so ITP I = IOP I is distributed uniformly over the integers I' + I
through I' + q. Also, it is not hard to show that recursion (8.3.7) again describes the ITPj
andI~. From this it follows (Problem 8.8) that each ITPj is a mixture ofthe ITPj(s), with
s = I' + I, ... , I' + q. using the equal mixing weights l/q, and I~ IS a similar mixture
of the I~(s). Consequently, E[I~] is just the simple average of E[I~(s)] over these s's,
and E[B] is the average of the E[B(s)].
To compute tbe average cost of the policy, we can use these quantities directly. Al
ternatively, compute the functions Cj(yls) above. The actual average cost is then
k A + 2;,:,:q C ( Is)
C(r, q, s) = I y-. + I I Y (8.3.11)
q
These facts allow us to determine the best policy through a slight modification of
algorithm (8.3.9). Compute the functions C;,j ~ I, and the s;,j > I. These sJ are in fact
optimaL Moreover, the average cost of a policy using these base-stock levels and any
(I; q) at stage 1 is just
s just like that of Now, consider an echelon (r, q) policy. This works just likc a local (r, q) policy, except
te integers r + 1 that each stage monitors its echelon inventory-order position and bases its orders on that
escribes the ITPj quantity. The policy variables are (rj • q). Assume that the policy is quasi-nested; that is, qj
the ITPj(s), with is an integer multiple of qj+ l' (This makes sense, for the reasons mentioned above. We can
I similar mixture and do set each IJ(O) to an integer multiple of qj+ L, so INJ(t) remains an integer multiple
rJ] over these s's, of qj+ 1 for all t.) As for the rJ' however, we require only that rj + qj ~ o.
Given a local (r. q) policy, there is an equivalent echelon (r. q) policy: Use the same
ities directly. AI CJ.j, and set sJ = 1/ + 9;, ~~; = li~jsi: and I; = s} - 9; (so that S.i = rj + q). Clearly, the
1St is then state variables have precisely the same values at all times under these two policies. Con
versely, starting with an echelon-based policy, there is an equivalent nested local policy,
(8.3.11) if each rj - rj + 1 is an integer multiple of qj+ l' (The argument parallels the earlier one
for base-stock policies.) Otherwise, thete is no equivalent local policy. Although it may
modification of not be obvious, the full freedom to choose the rJ in echelon (r. q) policies, beyond the re
ese s_f are in fact stricted values oflocal policies, can indeed be valuable.
(: levels and allY Intuitively, the relation between local and echelon policies can be seen by compar
ing the behavior of lOP; and IOPi . Both increase by qj whenever stage j orders. But,
lOP; decreases only when stage j + I orders, while lOP} decreases with each actual cus
tomer demand. That is, under the echelon policy stage j learns about customer demands
(8.3.12)
instantaneously, whereas under the local policy there is a delay in transmitting this in
fonmation until enough demands accumulate to trigger an order by stage j + 1. If the
~ I is convex, Al
echelon policy parameters satisfy the condition above, so that there is an equivalent lo
e minimal solu cal policy, then this delay is immaterial; the extra information is not used. Otherwise,
there is a real difference. (We call the echelon policy "centralized," because it requires
izations of per more information than stage-ta-stage orders. Given that information, however, the con
n Section 8.3.6. trol decisions are still taken locally.)
ingle-stage cost We now present a recursive method to evaluate an echelon (1; q) policy, analogous
in the no-fixed to (8.3.7). First, ITP, ~ lOP! IS distributed unifonmly over the integers r, + I through
"I to the square r 1 + q 1· Second, it remains true that
INj = ITPj - Dj
Finally, recall that, for a base-stock policy,ITPj + 1 = min {Sj+l' IN), which can be writ
gs now become ten as Sj + 1 - ITPj + 1 = [Sj+ I - I~] + . The analogue here, it turns out, is
. We describe a
Sj+l - ITPj+1 = 7r(sj+l-IN.,. qj+d (8.3.13)
entralized con
It turns out that whete, for any positive integer q, the function 1f(', q) is given by
sent evaluation
1f(x, q) = max {x, (x) mod (q)}
~beme.
I. where rj + qj x x2:0
ders batches of
Is. Assume that
of qj+ 1- This is
I«) mod (q) x <
This recursion is numerically intricate but conceptually straightforward.
0
kes sense to or To evaluate the shipment frequencies, we must determine just how shipment costs
) = rj+ qjand work. The policy may sometimes dictate shipping several batches of size qj to stage j si
t. And, the pol- multaneously. Does such a shipment incur only one fixed cost kj or several, one for each
batch? We encountered the same issue in Section 6.6, and again the correct answer
..
depends on the actual situation. The analysis is much simpler, though the cost is greater,
when each batch incurs its own cost, for then the average shipment cost to stage j is just
8.4 Asse....
ki,lqj' The other case can he analyzed too, but only with more effort; we shall skip it.
This technique can evaluate any echelon (r, q) policy. What about optimization? Given
fixed q, we can find the optimal r (or equivalently s = r + q) by an algorithm analogous to
(8.3.9). Set £:J+l(X) = (b + h;)[x]-. Forj ~ J, J-I, ... , 1, given £:/+1> compute
- j (y)'
C = (~)
q. "q,-'
~z=o
C·(y
J
- ")
...
}
sJ = argmin {C;(y))
Ie cost is greater,
t to stage} is just 8.4 Assembly Systems
'" shall skip it. 8.4.1 The Model
imization? Given
thm analogous to Recall, in the classification scheme of Chapter 5, an assembly system is a supply
:ompute demand network with a single end item J, such that every other item} < J has just one
successor. In contrast to a series system, however, an item j may have several predeces
sors, which compose the set Pre (j). To make a unit ofitemj requires one unit each of
all of its predecessors i E Pre (j).
Assume that customers demand only the end item, according to a Poisson or
compound-Poisson process. There are no scale economies. Each item has its own con
stant leadtime, Lj. (Recall, Chapter 5 allowed predecessor-specific leadtimes Lij. Here
we assume that, for i E Pre (j), the Lij are equal to Lj.) For instance, Lj might represent
a production activity that begins with an assembly operation, requiring all the compo
nents at the outset. So, it makes sense to take predecessor units from inventory and send
them toward} only in complete kits, consisting of one unit each ofevery i E Pre (j), and
, the methods of we assume this from now on.
ilial q. Chen and Figure 8.4.1 depicts such a system. The vertical lines show assembly operations, and
lSCussed below), the length of each arrow indicates the corresponding leadtime. Thus. the supply-system
illgorithm above dynamics correspond to units moving at constant speed along the arrows. (We shall ex
plain later the Ljat the bottom.)
c (problem 8.11
Ie best) echelon
"JSsible altema
FIGURE 8.4.1
:Ion (r, q) policy
rew percent. We Assembly system.
~ do not have a
nains a pressing 9
1
I ,
L'.
\7I
j
w_ on the frac
\ in the spirit of
policy. Among
en the other sj.
L' 4
W
I near I - '>L.
ion is approxi
L' I
\7
!ered.
:minimizes the
ISO, and the ap
~L" I L".,• 3 L" 4 L" ,
.....
As in Section 5.4.2, we can interpret this picture as a project schedule, and this view
point provides some insight into the meaning of effective control. Suppose we operate
the system entirely without inventories. When a demand occurs, we must order all the
components and assemble them from scratch. Clearly, we want to schedule these events
as indicated in the figure. This approach fills the demand as soon as possible, but given
that requirement, it performs all activities as late as possible. Specifically, we order item
1 immediately, but wait a while before ordering item 2, so that the two units arrive to
gether to be assembled into item 4. We order item 3 even later (but before the arrival of
1 and 2), so it arrives just when item 4 is ready.
Can a local control scheme realize this tight schedule? Consider what happens un
der a local base-stock policy with base-stock levels zero. A demand propagates back
ward through the network, so we order all raw materials (items 1,2, and 3) immediately.
Items 2 and 3 thus arrive too soon, before they can be used. Of course, we could modify
this policy, either by delaying the demand signals or by instructing each item to wait be
fore acting on its signals. But such modifications violate the spirit of pure local controL
The same issues arise when we do maintain inventories. Purely local information is
inadequate to operate the system intelligently. (And the simple modifications above do
not suffice in genera1.) We need global information.
Nevertheless, as we shall see, an assembly system can be reduced to an equivalent
series system. Once we construct this equivalent system, we can use the methods of Sec
tion 8.3 for policy evaluation and optimization. The best policy. it turns out, is much like
an echelon base-stock policy. To manage each item, however, it utilizes more than just
locally available information.
As explained in Section 8.1, there are situations that require a local policy, where
centralized control is expensive or impossible to implement. As shown in Section 8.4.5
below, series-system methods can also be used to evaluate a local base-stock policy_
(We do not yet know how to extend these results to stochastic supply systems. Nor
do we know how to incorporate fixed costs, except in one very special case, discussed
in Problem 8.13.)
e, and this view The total average cost can again be written as (8.3.8); i.e.,
~se we operate
E [X;~l hJN.i + (b + hj)B] (84.1)
US( order all the
FIGURE 8.4.2
Equal forward echelon leadtimes.
L'2 W
L' 3
1 I L' ,
\7.I
L' 4
9
L' 1
\Y
f-I-----------
L" ] , L" L" 3 L" 4 L". ,
there is no reason to order more ofitem 4. Otherwise, if rrp4(1) < IN; (f), we may wish to
raise ITP4 (1), but not beyond IN;(I).
We can visualize the idea here by drawing a vertical slice through the network, like
the dashed line ofFigure 8.4.2. To the right ofthe slice, downstream, is a subnetwork in
cluding all of item 4's supply system, but only part of item 3's. The points where the line
crosses the leadtime arrows represent the same totalleadtime to the end item. The ech
elon inventories downstream from these points are precisely IN;(I) and ITP4(1), and we
want to equate these values as nearly as possible.
We can draw a slice anywhere in the network. Such a slice typically intersects sev
eral items' leadtimes. These intersection points are equidistant from the end item in
tenus of total leadtime. Now, consider the echelon inventory downstream from each in
tersection point, i.e., the total inventory along the path from the point to the end item.
Ideally, we would like all these echelon inventories to be equal.
It is clear, then, that we cannot hope to operate the system intelligently using local
information only. In the example above, the slice relates two items, 3 and 4, which hap
pen to be close neighbors, but of course that need not be true in general. We may well
wish to balance the inventories of items that are quite far apart in the network. In other
words, we must coordinate actions explicitly across the entire system.
Fortunately, the requirements for effective coordination are less fonnidable than
they seem at first glance. To see why, suppose some slice indicates balancing the inven
tories of seven different items. To achieve the necessary coordination, it makes sense to
---
Chapter 8 Sewrol Items with Stochastic Demands 327
consider the items in pairs, that is, to balance the first with the second, the second with
the third, and so on. This, it turns out, is sufficient; to control each item, we must pay at
tention to one other item. But which one? The new item-numbering scheme above pro
vides the answer: For each item} > I we must take into account the status ofitem} - 1.
THEOREM 8.4.1. Assume the long-run balance condition (8.4.3) holds for all t.
Then, for all) > I and i. Pre (j). IN;(I) "" IN;_l(t).
<:we postpone the proof until Section 8.4.6.) Thus, we never have to worry about the
predecessor inventories; they are guaranteed to be sufficient. We need only compare
ITPj(t) to mm {Sj,INj_1 (t)}.
(Actually, from any initial conditions, if the system runs for a while, it will attain
(8.4.2) at some finite time t. And once the system enters this state, it will never leave.
So, the theorem applies and the policy simplifies for sufficiently large t. To simplify the
discussion, assmne the special starting conditions above, so the theorem holds as stated.)
Consequently, (8.4.2) in fact holds as an equality:
ITP;(t) = min {Sf' IN;_I(t))
Notice the similarity to (8.3.6)! Also,
IN;(t + Lj') = ITPj(t) - D(t, I + Lj']
Following the approach of Section 8.3.2, rewrite these equations using a simple time shift:
ITP1(1) = SI
IN;(t + L) = ITPj(t + Lj - 1) - D(l + Li -" I + Ltl
ITP;+I(t + L;l ~ min {si+" INj(t + L;)}
In equilibrium this becomes
ITP j ~ s,
INj = ITPj - Dj (8.4.4)
Suppose we apply the policy-evaluation algorithm of Section 8.3.2 to this series sys
tem, using the same policy parameters s. The average cost, in the assembly system's
(8.4.2) terms, is
: same amount.) E[Lf~1 h;IN; + (b + hj)[IN;n (804.5)
This is not the same as (8.4.1), the average cost ofthe assembly system. But LJ = Lj,
(8.4.3) so IN; = INJ and [IN;J- = B. Also, by a simple conservation argument, E[ITj] ~
E[D(Li - L;')], so
E[IN;] = E[IN;] + E[ITj] = E[IN;] + E[D(Li - L;')]
holds for all t.
Consequently, the computed cost (8.4.5) exceeds the true cost (804.1) by
Lf~lhjE[D(Li - Lj')] (8.4.6)
l"orry about the
I only compare This is a constant, independent of the policy. In sum, to evaluate a balanced base-stock
policy in the assembly system, we need only compute the average cost (804.5) ofthe cor
e, it will attain responding policy in the equivalent series system and then subtract the constant (804.6).
ill never leave. It follows immediately that, if we apply the optimization algorithm of Section 8.3.3
To simplify the to the equivalent series system, the resulting policy vector s* is optimal also for the orig
IOlds as stated.) inal assembly system. We're done!
Well, almost; there remains one loose end: What happens when two or more of the
L.} are equal? We can resolve ties arbitrarily in renumbering the items. But then some of
the Lj' are 0, and the corresponding Dj ~ O. What happens in the equivalent series sys
tem (or any series system) in this case? If D; = 0, intuitively, inventories at stagesj - 1
andj perform the same function, but stagej's is more expensive, so there is no reason to
hold stock there. Indeed, C/y) = hjy + C;+ I(Y), and so (since hj "'" 0) sj :5 sj+1- which
'PIe time shift: means stagej never holds inventory.
In the assembly system too, we choose sj :s; Sjt-l, but here this condition has a dif
ferent meaning: lfj E Pre (j + I), we never hold inventory of itemj, but otherwise we
may. For instauce, consider the system of Figure 804.1, but suppose items I and 2 have
t,
the same actualleadtime L so L Z ~ O. This implies INrCt) = ITP I(t). The policy thus
dictates that ITP,(t) = min {s!,ITPI(t)) ~ ITPI(t) ~ Sl' Consequently, we always main
tain the same inventories of items 1 and 2. This is just what common sense suggests.
the supplier for each raw material. Consequently, JOPj(l) = sf' and IN;(I + Lj) = sJ
D(I, I + Lj],j < J. It is clear, moreover, that
Define the L} and Lj as above. Let us evaluate the material backorders at time t +
L J- J. Notice, (I + L J- J) - Lj = I + Lj _ j • Also, suppose the s; are nonincreasing for
j<J(whichisreasonable), and sets7_1 = SJ-h andsj= s; - s;+j, 1 5 j 5 J - 2. Then,
IN'j(t) = + L - lej ]
ITP,(I) - D(I, I
~ ITP,(I) - D(I, I + Lj _ 1 - L,_ tl
Chaptel" 8 Several Items with Stochastic Demallds 331
i) so (since L7-1 +- L i = L. i - L)
INij!I):5 ITP,_I(I - L;'_I) - D(I - L7-1> I + lo, - loj ]
j<Jj = IN,_I,j(t-L7_1)'
-----
Denote
~ = uemand rate for retailer j, j =
1, ... , J
Next, consider retailer j. The logic of Section 8.3.1 leads to the following analogue of
(8.3.4):
(The second equation uses the conditional-variance formula from Section C.2.2 in
twork of loca Appendix C.)
ehouse is lim It is possible to define echelon inventories and related quantities. Indeed, one can
""'" the prod show that local and echelon base-stock policies are equivalent. This approach offers no
ependent, and special advantages here, so we postpone it until the next section, where we make good
odels are fre use of it. (A base-stock policy, local or echelon, need not be optimal bere. We know lit
tle about the true optimal policy.)
live heuristic,
Ltion manages
: tries to com 8.5.3 Optimization
:ach location,
MuctS. This Consider the problem of choosing a base-stock policy to minimize the long-run avemge
equately cap-
h;
cost. There is a holding-cost rate for each location; each retailer has its own backorder
cost rate bj . (We use the notation Cj in a different way from Section 8.3.)
The total cost has several components: First, define
s;.
This is the average cost at retailer j. It is a function of So as well as Next, the average
to (8.3.2) to inventory in transit from the warehouse to retailer) is the constant 'AjL;; as above, we in
cur cost at rate hoon this stock. As for the warehouse itself, the only relevant cost is the
inventory holding cost. Sum these holding costs to obtain
(8.5.1) Co(.,M ~ ho(E[Io] + lf~ 1 'A;L;)
. ..
.,
Let s' denote the vector of retailer policy variables (S;)f~ I. The total average cost is then
C(so, s') ~ Corso) + lf~1 C;(so, sf)
Unfortunately, there is no simple, elegant procedure to optimize this function. (The
recursive algorithm for series systems cannot be extended to distribution systems.) The
most popular approach is a projection method: The main loop of the algorithm searches
over possible values of So to find the best. For each So, an inner loop detennines the best
values of the other s}.
The advantage of this approach is that, with So temporarily fixed, the objective func
tion C(5(), s') separates by j, so the inner loop can optimize over each sJ separately. More
over, as a function of sJ only, each of the Cj(soJ sf) has the same fonn as the single-location
cost function of Chapter 6 (in particular, it is convex), so it is easy to minllnize. Further
more, let s;(so) denote the optimal value ofs; for fixed sQ. It is intuitively clear (and not hard
to show) that each s}(so) is nonincreasing in so; as we add more stock at the warehouse, we
need less at the retailers. So, if the search process increases sO sequentially, we need only
consider ever smaller values of the s}.
We are still left with the problem of finding the optimal So,
that is, of choosing So to
minimize C(so, s'(so)). Unfortunately, this is not a convex function of So' Exhaustive
search is the only sure method. (On the other hand, people who work with such models
have found that a local search nearly always finds the global optimum.)
The normal approximation can be applied to this system. Let IJ.-j and a} denote the
mean and variance of Dj . j 2:: O. The normal approximation at the warehouse yields
E[B o] = <lJI(zo)"o
E[Bo(B o - I)] = 2<lJ2(ZO)"~
E[lo] = <lJI( -zo)"o
where Zo = (so - i-Lo)/CTo. Moreover,
E[B oj] = ejE[Bb]
V[B oj] = e)1 - e;)E[Bb] + e/V[Bb] j>O
Now, approximate each BOj + Dj by a normal distribution with the same mean and variance,
j!,; = E[B oj ] + ~j
,,2
-J
= VIB O] .] + ,,2]
The normal approximation for retailer j yields
E[B;] = <lJI(Z)ZJ
E[I)] = <lJI( -z)!!j
where Zj = (s; -j!)/!!j" We now have all the elements needed to evaluate the cost C of
any policy.
Furthermore, if we set s) optimally, given So,
then the cost for retailer j becomes
Cj (so) = (bj + h)<lJ(zJ)!!j
'1"",,""'11
Tage cost is then where zj solves <l>°(z) ~ !lAbj + h). (The function Cj (so) depends on So through ([}")
The total cost thus reduces to a function of one variable,
is function. (The
C(so) o o] + Lj>o 0(so)
~ h E[l
f choosing So to Thus, if z* represents the (common) optimal value of z for all retailers,
. ~' Exhaustive 9so, sj(so)) = !Lib + h;.)<I>(z*) (8.5.4)
th such models
e(so, s'(so)) = ho(E[lo] + AoL;) + ([,(b + h;.)<I>(z*)
i rrJ denote the where!I..J is the standard deviation of B Oj + D1 , and
'use yields
!I.. r = Ij!Ij = J!Ij (8.5.5)
Let us examine the joint impact of So and J on the total cost (8.5.4). First, fix
So = O. In this case the retailers operate independently, as in the parallel system of
Section 8.2, each with total effective leadtime L o + L;. Indeed, V[B o] = E[B o] =
AoL o, so ([, ~ [JAo(Lo + L;.)] 1/2 Apart from the constant hoAoL;., (8.5.4) reduces to
the same form as (8.2.3); for fixed total demand 1.. 0 , the total cost is proportional to
Vi. Now, raise So to any positive value. This increases E[Iol and reduces V[Bbl, of
course, but those effects are independent of 1. Also, raising So reduces one of the fac
n and variance,
tors of J in!L, (E[BoD, leaving the other (AoL;) constant. In sum, for any so, the over
all cost (8.5.4) depends roughly on the square root of the number of retailers, but the
strength of this dependence declines as So increases.
Thus, warehouse stock performs two distinct functions: It may just be cheaper than
retailer inventory, as in a series system. In addition, it serves to pool some ofthe retail
ers'demand uncertainties; it can be used to fill an order from any retailer, and this flex
ibility softens the impact ofthe warehouse's own leadtime, Lb. Warehouse stock thus al
lows us to enjoy some of the advantages of consolidation, while still maintaining
multiple retail locations.
, the cost C of
EXAMPLE 8.5.A, PART 1
jbecomes
A Brazilian company makes doce de leite (literally, milk sv.>eets, a wonderful concoction, something
like caramel sauce only better) and ships it from its plant in the countryside to several locations in and
336 Foundations ~f [nn;lJ!ory Management
around Sao Paulo. It is considering leasing a warehouse in the outskirts of the city and supplying its
outlets from there.
Tbere are J = 16 locations, all identical. Each faces Poisson demand at a rate of A = 0.625
cases per day. The totalleadtime to each location is L = 4.4 days. The cost rates are b = R$10 and
h = R$O.62 per case-day. (The currency in Brazil is the real, abbreviated R$.) There are no scale
economies.
The current setup treats the locations as independent items (as in Section 8.2). The optimal
base-stock level for each location is s* = 5. with total cost over all locations (omitting the pipeline
holding cost) ofR$38 per day.
If the company leases the warehouse, the Teadtimes will be La = 2.8 and L; = 1.6. The back
order cost will remain b = 10, and the holding costs will be hb = h~ = 0.62. (Notice, the holding
cost at the warehouse is the same as at the loeations, and the totalleadtime remains 4.4.) The best
local base-stock policy is s~ = 28, and each s) = 3, with total eost C = R$26.4. Evidently, een
tralized inventory offers substantial savings in this case.
8.5.4 Extensions
Most of the extensions of the basic series-system model work here as well.
demands. The warehouse itself is no problem: Suppose Y; indicates the demand-size ran
dom variable for retailer j. The total demand at the warehouse, then, is also a compound
Poisson process; the demand-epoch rate is AD, and the demand size Yo is a mixture ofthe y;,
using the mIxing weights aj" Knowing this, we can determine the distribution of Do> and then
The difficulty appears when we try to analyze the retailers. The binomial-disaggregation
method breaks down, and the waiting time BWh becomes quite complex. Exact analysis is
therefore much harder here than in the pure Poisson-demand case.
We can extend the two-moment approximation, by invoking another rather crude
approximation step: Suppose all retailers have the same demand-size distribution, so
Yo = .f;. = Y Now, assume that the backorders B oalways comprise an integral mimber
of demand batches, say Nh. (This is true for So ~ 0, but not otherwise.) Under this ap
proximation, Bohas a compound distribution, built from those of No and Y Given E[Bo]
and V[Bol, E[NhJ and V[NhJ can be obtained from the identities
E[BhJ = E[NhlE[Yl
V[Bol = E[Nh]v[y] + V[Nh]E 2 [Y]
Next, the number Nh j ofthese batches originating at retailer j can be detennined through
binomial disaggregation, as in (8.5.3):
E[NhjJ= ajE[NQl
V[Nh) = 97V[Nhl + ail - a)E[Nhl
III"""""III'I'I'IIII'~
eU.
Suppose each supply system is exogenous and sequential. The warehouse's system is in
dependent of the retailers'. (The retailers' supply systems. however, need not be inde
pendent of one another. For instance, every retailer's supply system may include loading
:=mand-size ran Combining the ideas of Section 8.5.1 and Section 8.3.4.2 leads immediately to
iO a compound (8.5.1) and (8.5.2). Dj now means location}'s demand over its (stochastic) limiting lead
Iixture ofthe lj, time. Binomial disaggregation still works to derive Bb j from B~.
Lof Do, and then
Itegral number L;
as exogenous, with distributed exponentially with parameter (l - PJ)~;'
Suppose each supply system generates independent leadtimes. As in Section 8.3.4.4, ap
AjE[L;l
We can combine this approach with the two-moment approximation above. An even
simpler approximation is sometimes used in this context: Compute E[Bb;J as in (8.5.3),
nined through
but treat B~; as Poisson-distributed with this mean (and skip the calculation of its vari
ance). Equivalently, replace BWb by the constant E[BW~]. Or, again equivalently, treat
the warehouse waiting times for successive demands as independent; thus, each retailer
faces independent total order leadtimes. Under any of these three interpretations, Bbj +
rrj now has a Poisson distribution.
'11
.;;::
Now, suppose evcry location has its own fixed cost kJ for shipments. As in Section 8.3.7,
we begin with the special case where only external orders (here, from the warehouse)
entail scale economies, so k o > 0 and the other ki = O. Accordingly, suppose the retail
ers all use base-stock policies, but the warehouse uses an (r, q) policy. (These are aIllo
cal policies.)
This system is quite tractable: The warehouse operates like a single-location system
with Poisson demand, so again IGPois uniformly distributed. Fmthermore, conditional
on IGP = So = s, B and the B; have the same distributions as under a pure base-stock
o o
policy with parameters (s, s'). Consequently, the total cost can be written as a simple av
erage, analogous to (8.3.11):
, koAo + I;~;+ 1 C(s, s')
C(r,q,S)= (8.5.7)
q
Again, to optimize this function, use a projection algorithm: For any fixed (r, q), com
pute the best sJ separately for each} by minimizing a convex function. Using this tech
nique as a subroutine, search for the optimal (r*, q*).
The general case, where there are also scale economies for internal shipments, so
any of the '9 can be positive, is far more difficult. We briefly summarize the state of the
art: As in Section 8.3.7.2, it is plausible to use a an (r, q) policy at each location. Such a
policy may use either local or echelon information. (The difference appears only at the
warehouse, of course.) It is no longer the case, however, that local (r, q) policies reduce
to echelon (r, q) policies; these are two distinct policy classes. Neither one, moreover, al
ways performs better than the other. There are methods to evaluate any such policy ex
, :'I"'I:"'' :IL!II'IIIII'' "
,'I" '11,1I1,111,111,lilllllllulllilli
...
Iemand. That is, actly. These techniques are intricate and computationally demanding, however, and be
J'f the arguments come intractable for large J There are also simpler approximations, which seem reason
the (compound ably accurate. (Section 8.7.2.2 sketches one such approach.)
Hil
otroduced in the
the two-moment
8.6 Distribution Systems: Central Control
The local base-stock policy of the last section is appealing in many ways. It seems rea
sonable, in particular, for the warehouse to fill demands sequentially. Duning periods
when the warehouse has stock on hand, this just means filling demands as they OCCUf.
etwork of items But consider what happens when the warehouse runs out of stock (a common event if it
, and item 0 has keeps little or no inventory): Suppose that, first, a single demand arrives from retailer 1.
Retailer 2 generates the next 45 demands. At that point the warehouse receives a unit
IDS. In the two
from its supply system. Where should it send this unit? Under the FIFO rule, the unit
, well as the re must go to retailer 1, because its demand occurred first, even though retailer 2 clearly
independent of needs it morc.
:he warehouse's The weakness of the local-control scheme, then, is that it relies on history rather
and base-stock than the current status of the system to make crucial decisions. This section describes an
il one. alternative approach, which requires and exploits fully centralized information. This ap
proach seems to work well, especially for high-volume goods, specifically, when the to
tal demand volume is large compared to the individual retailers' demand fluctuations.
This is just the kind of situation where a capable information system can be justified eco
nomically. (For low-volume items the local-control mechanism suffices, because the re
I Section 8.3.7, tailers rarely compete for shipments, and major discrepancies like the example above are
he warehouse) even rarer.)
[lOse the retail For now assume the retailers have independent Poisson demands, and allieadtimes
nese are all 10 are constant. We point out later that these assumptions can be relaxed considerably.
>eation system
re, conditional 8.6.1 Echelon Cost Accounting
lire base-stock
IS a simple av
This approach uses echelon-level information. Define
Bt(t) ~ B;(t)
W) = I;(t)
Ie state of the ho = hb
'Ilion. Such a
Notice,ITPr(t) sums the net inventories at, and stocks in transit to, all the retailers; so
rs only at the
INo(t) includes all stock at, and downstream from, the warehouse.
,licies reduce
The total cost rate at time t is given by
moreover, al
eh policy ex- ho[/o(t) + '5;;~, ITJ(t)] + '5;;~1 [hil;(t) + bjBi(t)]
340 Foulidatiow; ojInventory Management
= E[C(y
J
- D)]
J
Minimize l j 9S j) (8.6.4)
subject to ~Sj = Sr
IOthing we can
Sj 2 Sf j = I, ... ,J
lrifting certain
;: t + rj. Thus, Suppose we approximate the Dj by continuous random variables and treat the Sj as
continuous variables. The model (8.6.4) then becomes a simple nonlinear program. The
first-order optimality (or Karush-Kuhn-Tucker) conditions are as follows: Letting' be
(8.6.1) the dual variable corresponding to the equation, we require
) justifies it in C;(Sj) ~ , j = I, ... , J (8.6.5)
Also, (8.6.5) is tight (holds as an equality) for allj with Sj > si.
(8.6.2) Let us sketch what is involved in solving (8.6.4): Imagine that we start with each
Sj = si and then allocate in small increments, maintaining (8.6.5) throughout, until the
equation in (8.6.4) is satisfied, i.e., until the entire quantity withdrawn has been allo
cated. First, identify the smallest of the Cils), and set' to that value. Slowly increase
this Sj' thus increasing e;ls) (by convexity), keeping' = Ci(s). until' equals the sec
ond smallest of the e;I.>). Next, increase both ofthese '>j' keeping their Cj(s)'s equal, and
, equal to this common value, until' hits the third smallest Ci(s). Continue in this man
ner, stopping when the equation is satisfied exactly. The end result, clearly, is a feasible
solution satisfying (8.6.5), hence an optimal solution.
(8.6.3) [This is just the continuous analogue ofan algorithm that solves the discrete ver
sion of (8.6.4), called marginal allocation: Starting with sJ = sJC' allocate one unit at
a time, always choosing that j with the smallest current value of the first difference
,lC/sJ]
;sible to com There are more efficient ways to implement the procedure; the detailed mechanics
::b that seems need not concern us here. It is important to understand conceptually what solving (8.6.4)
means: The optimal solution tries to equalize the marginal costs C;(s) to the extent pos
lishelpfulto sible, subject to the limits imposed by the inequalities Sj ;::: sj. If Sr is sufficiently large
K)()se a quan (compared to ~. Si), so we allocate something to every j, then all the marginal costs will
i>OO to all the be equal. Otherwise, some j's will receive no allocations, namely those with large initial
IS. inventory-transit positions s j. In this sense myopic allocation tries to balance the retailer
l1ere maybe inventories. (This concept of balance is similar in spirit, though not in detail, to the no
IlP~ denote tion of balance in assembly systems.)
>lIS. We then In the case of identical retailers, the cost functions Cj are identical, so myopic allo
. to a highel cation reduces to a simpler rule: Allocate stock to the retailers with the lowest inventory
: the individ transit positions. In other words, try to equalize the ITPAt), specifically, make the small
est ITPJit) as large as possible.
ninimizethe Let us now explore an approximation of the model as a whole: We continue to use
cation, for it (8.6.3) to measure the cost rate. Imagine that, at every point in time, we can costlessly
lOring future and instantaneously redistribute inventory, backorders, and shipments in transit among
the retailers. That is, we have a total value ITPrV), which remains at its true value, but
IS: Set sj = we can divide it among the individual ITPJit) as we wish.
I these terms With these new rules in place, we know we can shift stock among the retailers at any
time in the future, so we can and should focus solely on the current cost rate (8.6.3). That
342 Foundations ofInventory Management
is, given this scenario, myopic allocation is truly optimal. In this setting, myopic alloca
tion means solving a model like (8.6.4), but without the inequality constraints:
Minimize ~qSj) (8.6.6)
subject to 1) Sj = Sr j ~ 1, ... ,J
Let C~(s,) denote the optimal objective value as a function of s,. The optimality condi
tions now become the simple equations
C;(S) ~ ~ j = 1, ... ,J
instead of (8.6.5). The solution to this model thus equates aU the marginal costs C;(Sj)'
The approximation thus assumes that the retailers' inventories are perfectly balanced at
an points in time. For this reason we call it the balanced approximation.
Now, let us also apply the normal approximation: We have
C(S)
JJ
= hJ - (bJ + h')<l>°(z)
J J
=~ (8.6.7)
or
<l>°(z)_hj-~
j - ~bL+-h"c,
] J
where Zj = (Sj - v)lo/ The ratio involving ~ plays the same role here as I - w in a single
location model. The task, then, is to adjust ~ until the corresponding values of Sj
satisfy ~j S,i = Sr'
Consider the equal-cost case, where all hj = h, and all bj = b. Here, (8.6.7) implies
that all the Zj must be equal, say to z,. Indeed,
Sr = lj Sj = lj (Vi + lJjZr)
SO zr = (S,. - vr)/(J'n where
, myopic alloca and Dr represents a normally distributed random variable with mean V r and standard de
traints: viation rr,. Thus, we can write the optimal cost of problem (8.6.6) in closed fonn; that
form. moreover, is precisely the cost of a single-location model.
(8.6.6)
(The maximal approximation leads to similar results. The equal-fractile solution is
again optimal. The random variable Dr now has the same form as the approximate Dj ,
)timality condi i.e., D, = v, + rr, (TIV2), where T has Student's t distribution with 2 degrees of free
dom. Likewise, if we approximate the D j by exponential random variables. Dr also has
an exponential distribution with mean v'r and standard deviation a r = Vf" More gener
ally, we can use any family of distributions detennined, like the nonnals, by a translation
oal costs Cl(s,). and a scale parameter, and obtain comparable results.)
:tIy balanced at With lUlequal costs, it turns out that a function of this same fonn approximates
C;(sr) with reasonable accuracy, at least when the cost rates are not too different. To do
this, set hand b to the smallest retailer cost rates. i.e.,
nearly so. If so, the distortions induced by the balanced approximation will be minor.
This is especially likely in high-volume systems, where there is ample stock flowing
through the warehouse to correct imbalances in the retailer inventories, and relatively
minor variation in the demands themselves.
There is strong empirical evidence that the approach does work well in such situa
tions. Typically, the cost of the heuristic policy (estimated by computer simulation) and
the balanced approximation's cost lie within a few percent of each other, and hence both
are close to the true optimal cost.
8.6.3 Discussion
Let us reflect for a moment on the mechanics of the heuristic policy: It bases warehouse
orders and withdrawals entirely on the aggregate state variables ITPo(t) and ITP,(t), ig
noring the distribution of ITPlt) among the retailers. The fact that the policy performs well
implies that these quantities capture the most crucial system characteristics for decisions
at that level. The individual ITP/tJ are important, ofcourse, althe allocation level. It is only
because myopic allocation manages them successfully that they can be ignored elsewhere.
Also, although the heuristic policy seems quite different from a local base-stock pol
icy, the two operate similarly when the warehouse actually has inventory on hand. To see
this, consider the identical-retailer case, assume that s~ > si> 0, and suppose we start
at time 0 with inventory st at the warehouse and none at the retailers. We immediately
withdraw s~ and allocate it myopically, so that each ITPj(O) = sJ = s~/J. Notice, the
C;(s')') are equal. Now, suppose a demand occurs. According to the policy, we withdraw
an equal amount ofstock from the warehouse, if that much is available, or else all the re
maining warehouse stock. We allocate that stock, clearly, to the retailer where the de
mand occurred. because its Cj(ITPj(t)) is now smaller than the others. The net result is
precisely the same as under a base-stock policy. This pattern continues until the ware
house runs out of stock. Only in response to warehouse backorders do the two policies
differ. The local policy allocates using the FIFO rule, while the centralized policy uses
myopic allocation, which better reflects current conditions at the retailers.
Because the balanced approximation is accurate, the bounds of Section 8.3.6 and
their interpretations all apply here. In particular, the performance of the system is
strongly influenced by the parameters of Do and D" especially "0 and"r'Now, Do is the
total system demand over the warehouse's leadtime, and cr~ is its variance. [t is tempting
to interpret Dr too as measuring systemwide leadtime demand. But D, is not ~'j Dj • It has
the same mean, v" but ,,; is larger than V[~j DJ (They would be equal if, but only if,
the Dj were perfectly correlated.) For instance, consider the identical-retailer case, where
in particular the Dj have equal variances. Then, since the Dj are independent, we have
V[~i Di] = J"y, but,,~ ~ J "J
2
This difference reflects the fact that, in the real system, retailer demands and inven
tories are not consolidated in a single location. (Even the balanced approximation does
not actually pool the retailers. It aggregates the ITPj(t), not the I~(t). In other words, it
begins with the cost rate (8.6.3), which already reflects the separation of the retailers.)
We enjoy some of the economies of scale that such consolidation would provide, as re
flected in "0' but not all.
"""l:'II'II"il"",,',I'I~
,ill be minor. By the way, there is one sense in which the balanced approximation distorts the real
toek flowing system: In the identical-retailer case, suppose h = O. The optimal policy for the two
md relatively stage series system, of course, tells us not to hold inventory at the warehouse. In other
words, according to the balanced approximation, warehouse inventory has only one
n such situa function, like internal inventory in a series system, to help save holding costs. We saw in
IUlation) and Section 8.5 that, under local control, warehouse stock also helps to pool the retailers'
d hence both demand uncertainties, but here that second function seems to disappear. This is not quite
true. Simulation studies have shown that, even with h = 0, a small amount of warehouse
stock can sometimes improve performance. The optimal warehouse stock is smaller than
in a local-control system, though, and the performance improvement is slight. In sum,
warehouse inventory still plays a risk-pooling role under centralized control, but the ef
s warehouse fect is weaker than in a local-control system.
I I7P,(t), ig
monns well EXAMPLE 8.S.A, PART 2
Or decisions
.,,1 It is only
The Brazilian maker of doce de leite, having decided to lease the warehouse, is now considering
a centralized control scheme, supported by a computer-based system.
d elsewhere. The balanced approximation yields So = 0, Sr = 54, with total cost C = 20.8. The bestheuris
e--stock pol tic policy (found using simulation) has So = 6, Sy = 48 (so eachsj = 3) with total cost C = R$21.6.
land. To see This policy uses much less warehouse inventory than the local-eontrol scheme above (so = 6
ose we start versus 28), and its cost is 20~/o less (R$21.6 versus R$26.4). The cost of this policy, moreover, is
mmediately very dose to the lower bound of R$20.8, so the policy is nearly optimal.
Notice, the
., withdraw
;e all the re 8.6.4 Coupled Systems
oere the de Consider the special case where the warehouse cannot or does not hold stock. Under lo
let resnlt is cal control this means, in effect, that the retailers operate independently; each places its
I the ware own orders with the supplier, and the effective leadtimes are theLJ = Lo+ Li. Section 8.2
wo policies shows how to evaluate such systems.
I'Olicy uses Here, however, the warehouse does playa key role, as a transshipment or break-bulk
center. Orders are placed with the supplier from a central point. On receipt of an order
18.3.6 and at the warehouse, the order is immediately withdrawn and allocated among the retailers.
system is The key advantage of centralized control is the ability to wait until that moment to allo
Ji,D o is the cate the order. The streamlined materials-handling techniques that make this possible are
5 tempting sometimes called cross-docking, because goods are transferred quickly from the in
:;Dj'Ithas bound loading dock across the warehouse to outbound ones. This mode of operation is
JUt only if, becoming increasingly popular. See Rosenfield and Pendrock [1980] for a general dis
a.se. where cussion of its managerial advantages. For instance, Wal-Mart runs its central ordering
~ we have and distribution facilities in this manner, and this capability is a major element of its
overall corporate strategy; see Stalk et aJ. [1992].
rod inven In other situations the warehouse need not be a physical location at all, but rather a
<Irion does control function. For instance, our arrangements with the supplier may allow us to order
r words, it in bulk for all the retailers, but decide the disposition of the order only later, say after the
retailers.) supplier's production leadtime.
ide, as re- Consider the two-stage system we obtain through the balanced approximation. This
system performs precisely as in the approximation of Section 8.3.6. Under the normal
346 Foundations a/In.ventory Management
approximation, therefore, we can describe the best policy and its perfonnance in closed
form:
v = Va + v,. 0-
2
= O"~ + 0-:
b
"'0 - b + h'
<I>(z~) = "'0
s~=v+az~
C· = hovr + (b + h')<I>(z~)()"
Let us examine the key perfonnance index 0" in the identical-retailer case:
cr2 22
~ ()" 0 + ()" r (8.6.8)
= "o(L a + JL;)
To understand this formula, consider two extreme cases: When Lb = 0, the retailers
operate independently, in effect, and each contributes its own cost to the total. In this
case ()" = V"oJL;. Keeping the total demand rate "0
fixed, the overall cost is propor
tional to the square root of the number of retailers, as in the indvendent-item systems
of Section 8.2. When L;.
= 0, on the other hand, we have (T = A.oL~; the total cost is
independent of the number of retailers, as if their demands were consolidated at one
location.
The general case, where both Loand L; are positive, lies between these extremes.
For a fixed totalleadtime La + L;, performance improves as La grows (and L; shrinks).
Let us compare (8.6.8) to (8.5.5), the formula for the local-control system under similar
assumptions. The first term V[BaJ there corresponds to "oLa here (since we carry no
warehouse stock), and both expressions have the same last tenn, representing the retail
ers' leadtime-demand variances. The middle term (J - I )E[Ba] in (8.5.5), however, is
entirely absent in (8.6.8). This term expresses retailer imbalances, which the balanced
approximation approximates by zero. (As we have seen, this estimate is roughly correct,
though it is a bit overoptimistic.) Thus, even a coupled system under centralized control
achieves some, but not all, of the advantages of retailer consolidation. Centralized in
formation (here) and warehouse stock (in a local-control system) have somewhat simi
lar effects on perfonnance. In other words, a capable infonnation system serves as a par
tial substitute for inventory.
Lee et al. [1993] describe an interesting application ofthese ideas to product design.
The model describes Hewlett-Packard's logistics network for distributing printers to
markets in Europe. The printers are produced in the United States, shipped to a central
point in Europe (corresponding to the warehouse), and then distributed to the several
countries (retailers). The printers destined for different countries require slightly differ
ent features (e.g., power supplies), and initially these differences were built in during the
production process. In effect, each retailer operated independently. Its leadtime Li was
1 I "I IIIHIIIIII!III!IIII.
1111
mance in closed the production time plus the full shipment time from factory to market, and La was es
sentially zero.
Then, a new approach was suggested: Make a common printer for all countries, and
insert the country-specific features only at the warehouse. This approach would allow
HP to decide the destination of a particular printer only on its arrival in Europe, not at
La
the American factory. Clearly, much of the originalleadtinJe could be included in and
deducted from Lj. An analysis along the lines above demonstrated substantial savings
with the new approach, and partly for this reason the proposal was adopted.
This story illustrates a general design principle, sometimes called delayed special
ization: Wait as long as possible before adding customer-specific features to products.
T case:
8.6.5 Extensions
(8.6.8)
The approach above extends immediately to the case of compound-Poisson demands.
(Of course, the formulas for special cases must be adjusted appropriately. Recall, this
case is problematic under local control, but under central control it is not.) However, the
additional demand variation here may cause greater imbalance among the retailer in
ventories, so the balanced approximation may lose accuracy. The issue, again, is indi
0, the retailers vidual-retailer variation versus aggregate volume. Provided the systemwide demand vol
I>e total. In this ume is large compared to the retailers' demand fluctuations, the overall approach still
cost is propor works well.
rt-item systems Also, the approach does not actually require the demands to be independent across
be total cost is retailers; nothing changes if we have positively or negatively correlated demands. Of
olidated at one course, the distribution of Do must be adjusted appropriately; D, remains the same. (This
is another relative strength of this centralized approach; we have no idea how to analyze
hese extrcmes. local-control policies without the independence assumption.) In the case of a coupled
Old L; shrinks). system, moreover, one can incorporate intertemporal correlation, as in the world-driven
[l under similar demand process of Section 6.3: The balanced approximation works as above, and the re
:e we carry no sult is a model similar to that of Section 6.7.2.
Iring the retail The approach should perform well also for the other types of supply systems. (We
5), however, is are most comfortable making this assertion for exogenous-sequential supply systems,
h the balanced for the approach can be applied in that case with no additional approxinJations. We are
Jughly correct, less sure about the other types, because they do require approximation of the D j . The em
ralized control pirical evidence to date is encouraging, but limited.) The crucial step in the analysis re
:entralized in mains the same---add the standard deviations of the Dj to get (J,.
Jmewhat simi Now, suppose there are economies of scale for external shipments, Le., a fixcd cost
;erves as a par- ko for warehouse orders. If we apply the balanced approximation as above, we end up
with a two-stage series system like that of Section 8.3.6, and we know how to solve that
Ifoduct design. model. The solution, it turns out, provides a good estinJate of the optimal cost and a good
wg printers to policy for warehouse orders and withdrawals. To make the overall approach work, how
cd to a central ever, it is necessary to adjust the allocation policy.
to the several To see why, consider a coupled system with Poisson demands and equal retailer
slightly differ costs. We use an (r, q) policy for warehouse orders here, so there may be a long interval
It in during the between successive arriving orders. When an order arrives at the warehouse, it is with
adtime L;
was drawn and allocated immediately. If we allocate myopically, we optimally balance the
..
inventories now, but they may well slide out of balance in the future, long before the next
order arrives.
Specifically, call the current time 0, and let us look ahead time I 2> 0. Denote vj(t) =
aJ(t) = E[Dj(L; + I)] = I-;(L; + t). Ifwe allocate Sj to retailer), we would like the ftactiles
z/I) ~ [Sj - v/I)]!a/tJ to be eqnal for all t, but that may not be possible. Myopic alloca
tion tries to equalize the z;(O), but this may systematically induce unequal =N) for f > 0.
In the identical-retailer case, happily, there is no such difficulty. Equal 2/0) implies
equalz/t) for all t. (See Problem 8.18.) Myopic allocation continues to work well in this
special case.
Otherwise, we must choose between balance now and balance later. For what time
t :2: 0 wonld balanced inventories be most valuable? Consider the quantity s,. - v - "-of
(In case the Li are equal (to L;), this is just the expected total retailer net inventory at time
L; + t, assuming no further orders arrive in the interim. In general this is E[l: (ITP;(f)
DJi].) If this quantity is large, the system will have ample stock at tin,e f, and it does not
matter where the resulting holding costs are incurred. Likewise, if s,. - v - Aot is small,
the system will be plagued with backorders at t, and it is of no consequence whether they
are evenly or unevenly spread among the locations. Imbalance hurts the most at the mo
ment the system runs out of stock, when this quantity hits zero, for then some retailers in
cur penalty costs at the same time that others pay holding costs. In short, we care most
about imbalance for t ~ (s, - v)!l- o. (Actually, we need to revise this t slightly: Ifs,. < v,
so twould otherwise be negative, reset f = 0. Also, there is no need to look beyond the next
order arrival. So, if there is an outstanding order dne to arrive before t, reset t to that arrival
time; if not.. compute the expected next arrival time, taking into account the number of de
mands needed to trigger the next order, and proceed accordingly.)
Then, use this f to redefine the myopic-allocation problem (8.6.4): Just replace Dj
by D;(LJ + t) in the cost functions Cj . It makes sense to call this near-myopic allocation.
With this adjustment the overall approach performs well. (Different adjustments are
needed for other cases, when the warehouse does hold stock or when the retailer costs
are unequal! but the key idea remains the same: Balance the retailer inventories at some
appropriate future time t.)
What about fixed costs for internal shipments? Alas, we know ofno satisfactory ap
proach to this problem.
~ before the next with rates Aj . There is a constant leadtimeLj for the units ofitemj within each order. (In
many situations, all the items within an order are shipped and received together, in which
. Denote viet) ~ case the Lj are equal. Here, we allow for the possibility that certain items require special
like the mctiles processing, and so have longer leadtimes.)
. Myopic alloca As in Section 5.5.2, construct an equivalent two-level distribution system: The items
I =it) for t > 0. correspond to retailers. We add a fictitious item 0 representing the warehouse. The fixed
ill! z/o) implies
'·ork well in this
°
cost k o ~ k and leadtime are assigned to item 0; the original items have no fixed costs
and leadtimes Lj. We never hold inventory of item 0.
Here is a plausible policy for this system, in the spirit of Section 8.5.5: Use a base
.. For what time stock policy for the original items, and an (r, q) policy for item 0. SpecifY base-stock lev
:y Sr - V - Aot. els s} for the retailers (items), and start with 1;(0) = s}. Also, fix r = -q, to ensure that
Iventory at time the warehouse holds no stock, so JOPo(O) ~ INo(O) ~ 10(0) ~ r + q ~ 0. Thus, there is
EI2 (ITP;(t) only one policy parameter for item 0, namely q.
and it does not In terms of the original system, the policy works as follows: We place an order each
- ''QI is small, time q demands accumulate over all the items. The composition of the order precisely
oe whether they re!lects the demands since the last order.
nost at the mo This is just a special case of the system of Section 8.5.5, so the cost of any such pol
me retailers in icy can be expressed in the fotrn (8.5.7). (Actually, it is simpler to evaluate a policy di
t we care most rectly: Given q, Bb = - IN~ = - IOP~, so Bb is uniformly distributed on the integers 0
ghtly: If s,. < v, through q - I. We can then proceed as in Section 8.5.2. Binomial disaggregation yields
l>eyond the next the distribution of the B o), and we can apply (8.5.2) to obtain the BJ.) Likewise, we can
i I to that arrival apply the straightforward policy-optimization approach outlined in Section 8.5.5.
,number of de- Models of this basic !lavor have been widely used for several decades by large oil
companies, to supply filling stations with incidental products, such as motor oil and
Just replace Dj transmission !luid. These products are delivered by truck. Typically, a truck visits a sin
lJ1ic allocation. gle station, carrying all the products the station needs at the same time. The model rep
djustments are resents that one station's problem. (In teality, of course, the trucks draw from centralized
e retailer costs inventoties, so the model includes only part of a larger distribution system.) The fixed
nones at some cost k measures the cost of a delivery. Sometimes, the order size q is fixed to the capac
ity of a truck.
atisfactory ap
8.7.2 Shared Productioll Capacity
Next, consider a system with several items sharing a production facility of limited ca
pacity. In a detetrninistic world this situation leads to the ELSP of Section 5.6.2. Here,
results briskly, the items' demands are independent Poisson processes, and we model the production fa
cility by one of the lintited-capacity supply systems of Section 7.3.
This is essentially a special kind of distribution system. The items play the roles of
tetailers, but with zero leadtimes (because goods leaving the production facility move
immediately into inventory). The processing system, including its queue, corresponds to
eraI items are the warehouse's supply system. The warehouse itself, of COUlse, cannot hold stock.
lier costs. We We say "essentially," because the analogy is not quite perfect: In a true distribu
~dless of the tion system the physical units in the warehouse's supply system are indistinguishable.
jor setup cost, Here, this is true only under two very special symmetry conditions, namely, (I) the
son processes items have the same processing requirements, i.e., the same unit processing times,
,,
and (2) a unit in process acquires its specific item identity only at the completion of
processing. So, the results of Section 8.5 and Section 8.6 apply only under these con
ditions. Nevertheless, many of the basic ideas in those earlier sections can be fruit
fully adapted here.
First, suppose there are no economies of scale in the system-no setup times or costs.
Suppose we use the analogue of the local-control scheme of Section 8.5: Each item fol
lows a base-stock policy and sends orders to the production facility. The facility responds
to these orders using the first come, first served (FCFS) discipline.
For the moment. suppose the first symmetry condition above holds, so the items
have identical processing requirements. (Because of the policy, the second condition is
unnecessary.) The processing system itself has Poisson input, so it is relatively simple to
analyze, as in Section 7.3. Then. binomially disaggregate its occupancy to obtain the out
standing orders for each item. as in Section 7.3.7. Finally, optimize each item's base
stock level separately. The identical-item case leads to a simple performance formula
like (8.5.5); see Problem 8.19.
This same approach can be used even without the symmetry condition, at least in
some cases. For instance, suppose there is a single processor. The processing times Sj for
itemj are i.i.d. random variables, but the S; may differ overj. It turns out that the pro
cessing system behaves like the M/G/l system of Section 7.3.2.2. Its overall processing
e
time S is just the j weighted mixture ofthe S;. (In particular, E[S], E[S2], andE[S3] are
the corresponding weighted averages of the E[Sj], E[SJ], and E[S;], respectively.) So,
suppose we compute or approximate its occupancy, fO Dividing fO among the items re
quires a bit more care than before: Given fO = n > 0, each of the n - I units in the
e
processor's queue has the same probability j of being an order for item j, so binomially
disaggregate these units as before. The unit in process, however, has probability pip of
belonging to itemj, where Pj ~ AjE[S;l and p = L; Pj' With this slight revision we can
obtain the individual items' occupancies, and then proceed as above.
Instead of this local approach, we may prefer a centralized control scheme, along
the lines of Section 8.6. Let us focus on the single-processor case. When both symme
try conditions above apply, we must allocate each unit of finished product as it emerges
from the processor; otherwise, we allocate not stock but rather processor time. In both
cases, analogues of myopic allocation seem to work well. With identical products, again,
myopic allocation reduces to a simpler rule: When it comes time to allocate, whether
stock or processing time, choose that product with the lowest net inventory. (In fact, this
rule is truly optimal.)
There remains the decision when to operate the processor and when to shut it down.
the analogue here of the warehouse's ordering decision. The following simple heuristic,
essentially a base-stock policy, seems to work well: Monitor some aggregate measure of
inventory, and turn the processor on when that measure falls below some fixed base
stock level; likewise, turn the processor offwhen the measure equals or exceeds the base
stock level. (lu the identical-product case, the appropriate measure is total net inventory.)
Also, one can determine a good value of the base-stock level and global performance
measures, in the spirit of the balanced approximation.
Chapter 8 Several Items with Stochastic Demallds 351
completion of 8.7.2.2 Setup Costs and Times
Ider these con Now, suppose there are economies ofscale. Specifically, there is a cost and/or a time de
s can be fruit- lay to switch production from one item to another. This system is very complex, and we
We sketch one approach, which extends the local-control policy above: Each prod
uct uses an (/; q) policy to place orders, and the facility responds to those orders on a
times or costs. FCFS basis. (Many firms use control schemes ofthis general flavor, especially for batch
Each item fol production processes, as found fOT instance in the chemical, petroleum, pharmaceutical,
ciJity responds and food industries.)
For simplicity, suppose the products are identical, so each one has demand rate AU,
i, so the items and they all use the same policy variables (r, q). Also, there is a single processor, and the
Id condition is processing time for an order consists of two parts, a setup time 1" and a processing time
ively simple to IIIL for each unit in the order, both constants. Thus, the total processing time takes the
obtain the out form S = T + qllL.
h item's base Now, the orders for each item form a renev.ral process; the times between orders have
l3.Ilce fonnula an Erlang distribution with parameters AU and q. Taking all the products' orders together,
the overall input is a superposition of renewal processes. Approximate this by a Poisson
on, at least in process. (If J is large, this approximation is reasonably accurate, as mentioned in Sec
19 times S; for tion C.5.6.) The total order rate is,\. = J(>JJ)/q = Alq. The processing system then becomes
t that the pro
an MlGII system with constant processing time S and utilization i! = 't"S = ATlq + p, where
all processing p = A/IL. The formulas of Section 7.3.2.2 then describe the order leadtime L.
andE[S3] are Now, make another approximation: For each item, treat the processing system as ex
oectively.) So, ogenous. (Again, this is reasonable for large .I.) Then, analyze each item by itself, as a
~ the items re single-item system with the stochastic leadtime L. For instance, use a two-moment ap
I units in the proximation (normal, negative binomial, etc.), using E[L] and V[L] to compute the re
;0 binomially 2
quired parameters v and 0- .
Ibility pip of The lot size q thus has two distinct effects here: Given L, q along with r determines
-ision we can the perfonnance ofeach item in the usual way. However, q also affects ~ and S, and these
determine L itself. Indeed, for the overall system to be stable, we must have i! < I, which
'heme, along is equivalent to q > AT(I - pl. And, when q just barely satisfies this lower bound, i! is
lOth symme near I, so E[L] and V[L] are large and overall performance suffers. Thus, even in the ab
as it emerges sence of a setup cost, there is a strong reason to avoid small values of q. (Consequently,
we. In both there is a strong incentive to reduce the setup time T.)
ducts, again, We see here in an explicit and extreme form the scenario outlined in Section 5.2 and
ate, whether Section 8.2, where each item's lot size determines the workload it imposes on some com
(In fact. this mon resource. Here, the resource is the production facility, and its capacity is strictly fi
nite. We carmot account for that limited resource, however, by a simple capacity con
;hut it down, straint like (5.6.1). As the overall workload approaches the limit. the interaction of the
tie heuristic, system's uncertainties with its capacity causes congestion, and performance deteriorates
~measure of
rapidly.
fIXed base This congestion effect is not just an artifact of this particular model, with all its sim
:<Is the base plifying assumptions and approximations. The effect is real. Indeed, it reflects the true
l inventory.) impact of lot sizes in many situations better than simple fixed costs do.
erfonnance Now, the explicit model above may be too hard to work with in practice. There may
be no alternative to treating the items as independent, as in Section 8.2. (In those terms
352 Foundations o/Imwlfvry Alanagemel7.t
o .v
Ibat is. if_
~ up,."3Id
D
Box
~
ods bet.....
1'5Ie'IIL ~
ringT~--oa·,
011( i:o.Iuw: , Units
D·' !
Cards
<$n==
a IIlO£e" gat
IIJU:l<t of"
~"SleID. ro. the previous stage. When that happens, the card is attached to the incoming unit. It re
Ile at a t:iJnre.. mains attached during the entire time the unit remains in the stage, i.e., throughout the
,idual uniIs.. waiting time in the processor's queue, the actual processing time, and the time spent in
ores. and ... inventory.
'irtue is i6 There is one difference: Units leave stage J to fill customer demands. At stage
"""Uofa j < 1; however, a demand occurs, in effect, only when one of stage j + 1's cards enters
"" 1I>Odds..' its box. For instance, suppose there is no inventory at stage J, and all cards are attached
iIing m., . . to units in the processor and its queue. If a customer demand occurs, that demand is
ISOIldJuMs backlogged; no card moves into the box. In this situation, then, a customer demand does
not trigger a demand at stage J - 1. Only later, when the processor completes a unit
III m., .... (which then immediately leaves to fill a backordered demand), does a card enter the box,
,-~ thus generating a demand at stage J ~ 1.
The kanban system, then, is a particular type of policy, a rule specifying what to do
~.• §O . . under various circwnstances. It is similar in many ways to a base-stock policy: Both are
I>lXUI5.. ,.., pull systems: Customer demands trigger all other events, directly or indirectly; demand
lScanI_ information propagates backwards through the network, stage by stage, from the last to
ibIe.1IIo"" the first. In both cases there is one policy variable for each stage, the base-stock level or
m m.,,,,," the nwnber of cards, which determines that stage's maximal inventory.
'jlIOlXS§U The differences are noteworthy also: Under a base-stock policy a demand at one
it..lIM
JIll stage always triggers a demand at the prior stage. The kanban system, in contrast,
Oocepro generates a demand upstream when the stage fills one of its own demands. Thus, the
Iu\~. demand-propagation mechanisms are the same when (but only when) the stage has
inventory on hand; backlogged demands do not propagate upstream under the kanban
[inIm~ system until later. Consequently, the number of cards has yet another role in the
....-bed .",: kanban system-it limits the occupancy of the stage's supply system. Under a base
[Zan fr-.:t.::: stock policy, by contrast, the occupancy is unlimited.
....I
354 Foundations of [m'enf01y Management
This is the primary strength of the kanban policy. For instance, consider a two-stage
system, where the first stage has a fast processor, but the second's processor is slower.
Under a base-stock policy the queue at the second stage grows rapidly. No matter how
large it gets) the policy instructs stage 1 to continue producing units and transmitting
them to stage 2, as subsequent demands occur. The kanban policy, however, blocks the
flow ofunits into stage 2 when it becomes too congested; equivalently, the policy blocks
the transmission of demands to stage 1, and this ultimately stops stage 1's processor. In
this way the kanban approach avoids excessive queues.
It should be clear that the basic logic here does not require the use of cards. We can
use any other physical tokens instead. Some companies implement the policy with
painted circles on the floor, marking the legal spaces in the queue; here, every unit in the
queue must rest on one of these circles. Other firms embed the idea in a computer pro
gram; the logical equivalents of cards are sometimes called electronic kanbans.
(Suggestion: To make sure you understand the kanban mechanism, simulate manu
ally the operation of a two- or three-stage system. Use pennies to represent physical
units, and make cards out of paper, using a different color for each stage. You will also
need markers, say dimes, to represent backlogged customer demands. Variation: Play
this game with several friends. Choose one player to simulate customer demands; each
ofthe others operates a single stage.)
idcr a two-stage to customer demands, while each of the other stages responds to the demands from its
:essor is sloWe£. suceessor. The demand-transmission mechanism ensures that each supply system's oc
~o matter hoa
cupancy remains within its limit, i.e., ITj(t) <; nj . Otherwise, the policy operates just like
od transmitting a base-stock policy. That is, each stage fills its demands when it can and otherwise back
O\'er, blocks !be orders them; also, each processor operates when it has work to do, i.e., when [TiCl) > O.
Ie policy blocks
Specifically, a demand at stage j immediately triggers a demand at j - I unless it is
~s processor. In
blocked. Demands are blocked when
• 00" .. stage I
s;
is or Ilj' [Condition (8.8,2) holds when there is no inventory, and blocked demands cor
respond to stagej' backorders.] Also, a base-stock policy is a limiting case with each
~'atstage2
n) = 00 (so demands are never blocked). Another special case, where all s; = 0, is a pro
be occupan<:y cessing network with finite buffers (mentioned briefly in Section 7.3.6.3).
g.. even..me. It is possible to implement a generalized kanban policy using physical tokens, anal
, its demand<.
ogous to kanbans. Figure 8.8.2 illustrates one such scheme: We now follow the flow of
I idleness. ...,
demands, or tokens representing them, in addition to units and cards. The cards playa
r become too somewhat different role here. Also, there are two boxes, a delay box and a block box.
risely, while There are Ilj cards, which circulate through the subsystem; at time t = 0 they all star! in
it is needed the block box.
When a demand arrives from the subsequent stage, it first goes to the block box, If
as strengths.. it finds a card there, the two move together to the delay box, and the demand proceeds
"" and base to the prior stage; otherwise, the demand is blocked, and it remains in the block box un
LRecall. iota til a card arrives. The delay box works like a kanban system's box; a card there is a sig
~ Iimirior: nal to move a unit from the prior stage's inventory into the queue. (So, B}_l(t) is the num
'P"Iate Ibe!r ber of cards in the delay box.) The card remains with the unit during its queueing time
1aJJ.:m.l and processing time. After processing the card does not accompany the unit to inventory,
but rather returns to the block box to await another demand.
lUry,at~
The generalized kanban policy provides considerable freedom to modulate per
formance. Each sJ has essentially the same impact as in a base-stock policy; as we in
crease it, we obtain better protection against backorders at the cost oflarger inventories.
Each Ilj too affects inventory and backorders, but less directly. It mediates the backward
iIJo, transfer of demands. By increasing nj , we reduce the chance of forced idleness at stage
-pob..,'s j, but increase its occupancy. (These properties should seem plausible intuitively; they
: J n:spe-.ri;
can be proved rigorously, but we shall not do so.)
356 Foundations o(InventOly Management
D o v h, b
=~61=='4)= D==
0.5
80
10
20
40
80
Units Cards Demands
.-: (b) SlowS_
" 10
0.8 20
(A moment's thought should convince you that 111 can be set arbitrarily. That is, pro
vided n\ 2:: 1, the choice Of11\ has no effect on IN{(t) or on events at downstream stages.
: j
This is because the external source always has stock available, so the stage-l processor 10 i
is never idle for lack of input. Thus, V,re can fix n j = 1; a larger 111 does no harm, how 0.5 20 1
40 I
ever, since £T,(t) incurs no holding cost.)
80
To illustrate, consider a m'o-stage system with Poisson demand and limited-capacity
supply systems with exponential processing times, as in Section 8.3.4.3. Fix A = 1 and (c) Slow Stage ,
h; = 1. Table 8.8.1 compares the hest hase-stock, kanhan, and generalized kanban policies. 10 I
The table has three parts, (a) to (e). The systems descrihed in the three parts differ in
their processing rates ILr Part (a) describes a base case with ILj ~ 2, so Pi = ~. Parts (b) 0.8 2O!
and (e) explore the impacts of slower processing rates and thns larger Pi" Each part exam 40 I
ines two values ofh 1 and four values of b. ~
It is striking to observe that, in nearly all cases, the base-stock levels sJ are identical 10 I
across the three policy types; the occasional differences are slight. The overall costs too
are qnite close. [The greatest differences appear in part (b) ofthe table, descrihing a slow
stage-2 processor. This is precisely where we expect an occupancy limit to help the most.
s;
Even then, the differ by at most 1, and the costs only hy about 7%.]
Thus, the simple model of Section 8.3 seems to capture the fundamental economics
ofthe system. The base-stock levels play the same roles in all three policies, with or with
out occupancy limits. This observation suggests using algorithm (8.3.9) as a heuristic
method to determine the sJ in a kanban or a generalized kanban policy.
Notice also that, in the generalized kanban policy, 112 and s~ are quite different in
many cases; nz is mnch smaller in part (b), but larger in part (e). Moreover, they respond
to parameter changes differently_ For instance, s~ is very sensitive to the backorder cost
b, but n2 is not; conversely, n2 is much more sensitive than s~ to !-Lt-
Chapter 8 Several Items with Stochastic Demands 357
-
TABLE 8.8.1
Base-Stock, Kanban, and Generalized Kanban Policies
•
40 6 2 7.16 6 9 2 7.12 6 6 2 7.13
80 7 2 8.22 7 10 2 8.19 7 7 2 8.19
=mds
=~ (b) Slow Stage 2: 1-12 = 1.2
10 14 0 17.36 14 7 0 16.38 13 13 0 16.98
0.8 20 17 0 20.92 18 8 0 20.02 17 17 0 20.70
~'. That is, pro- 40 21 0 24.57 21 8 0 23.74 21 21 0 24.46
l5tream stages. 80 25 0 28.31 25 8 0 27.51 25 25 0 28.26
~e-l processor 10 14 0 15.86 14 8 0 15.33 14 14 0 15.62
)() harm, how- 0.5 20 17 0 19.42 18 8 0 18.94 17 17 0 19.28
40 21 0 23.07 21 9 0 22.62 21 21 0 23.00
nited-eapacity 80 25 0 26.81 25 9 0 26.38 25 25 0 26.78
'Lx.le ~ 1 and
(c) Slow Stage 1: fl.l = 1.5
mban policies.
parts differ in 10 5 2 7.24 5 19 2 7.21 5 5 2 7.26
•
e different in policy types constrain n2 in very different ways, yet their perfonnance is nearly identi
""". respond cal. In general, we must take care not to set n2 too small, to avoid excessive forced idle
lCk:order COSt ness. (In these examples, all the policies easily satisfy this requirement, even the kanban
Sz
policy, but that is because s~ is always comfortably large. If were smaller, say because
358 Foundations ofInventory .Management
of a smaller b, we might well prefer a larger n2') If we err on the high side, the cost
penalty is modest.
Caveat: This is a small test bed, and the conclusions should be regarded accordingly
as tentative.
8.8.3 Analysis
There are a few tractable approximations for policy evaluation in the research literature.
(See the Notes at the end of the chapter.) These are based on methods for finite-buffer
processing networks.
h side, the cost Consider this scenario: There are 17 production stages. Stage II happens to be the
bottleneck; 1t has the longest processing times. Now, the sales department learns ofa de
ied accordingly mand surge upcoming in a month, Must we wait until then, relying on the kanban mech
anism to transmit demands backward after they actually occur?
Of course not. In practice, typically, there is a higher-level planning-control func
tion between the physical production activities and the market. The "demands" seen by
the production process are not actual customer demands, but rather signals generated at
:::arch literature. this higher level. The planners do respond to real demands, but also to forecasts; also,
or finite-buffer they often try to damp demand fluctuations, in order to smooth the load on the produc
tion facility. Moreover, to anticipate longer-term changes in demand, they may increase
or decrease the numbers of cards at some or all stages, In the scenario above, the plan
ners may choose to generate demand signals before the demand surge occurs, or perhaps
to increase the number ofcards at the critical stage II. ThIs is one of the ways that MRP
or MRP-like logic can be integrated with elements of flT.
system. All de JIT also addresses the issue of managerial incentives: Many companies continue to
ion, moreover, judge managers along dimensions like throughput (total output per unit time) and uti
s). The mecha lization (fraction of time a facility is operating). lfthese are your objectives, it is clear
itS)' to describe
what to do-keep all equipment working as hard as possible. It is equally clear that this
e-a worker or rule leads to disaster: Inventory grows without limit. Sooner or later, something gives;
either the plant breaks down by itself or somebody shuts it down. Under this de facto
Ie policy is also control policy, then, the system oscillates between massive overproduction and layoffs.
al (r, q) policy. Now, all of the policies we have studied are smarter than this. Why not just impose
celebrated suc one of them? Doesn't that settle the issue? No: When incentives systematically conflict
mportant these with official policy, however refmed, people often find ways to subvert the policy. The
of centralized main message of JIT on this score is to look at existing incentives and to eliminate those
that, however well intentioned, lead to perverse behavior. We should not have needed nT
llions manage to tell us this, of course, but we did. (This negative insight is crucial, but one would like
lleSe word that more-a positive, constructive theory of incentives, a set of criteria designed to support
ent and proce intelligent operations. Academic research and business practice are just beginning to ad
up times, pro dress this question.)
IIDd vastly im
mrkers whose
, closest to the
hers.
men operates Notes
is psycbologi
is transparent, Section 8.1: Overviews of the subject are provided by Clark [1972] and the collections
forces kaizen, of Schwarz [1981], Axsiiter et al. [1986], and Graves et al. [1993]. Davis [1993] eluci
L local control dates the roles of uncertainties and inventories in complex networks from a managerial
:ntives, not to viewpoint. A good general discussion of centralized versus decentralized systems can be
:Iegree ofcen found in Rosenfield and Pendrock [1980]. Lee and Billington [1993] provide a cogent
discussion of decentralized control in practice. Business process reengineering is de
ically; indeed, scribed in Hammer [1990] and Davenport [1993].
xl more com- Section 8.2: The notion ofan aggregate tradeoff surface is discussed extensively in Gard
ner and Dannenbring [1979].
~-
Section 8.3: The study of series systems with stochastic demand and constant leadtimes
was initiated by Clark and Scarf [1960] and later refined by Federgrueu and Zipkin
[1984c], Rosling [1989], and Chen and Zheug [1994b]. The two-moment approximation
is by Graves [1985].
The extension of the policy-evaluation scheme in Section 8.3.4 to exogenous, sto
chastic leadtimes is attributable to Svoronos and Zipkin [1991]. Buzacott et a1. [1992],
Lee and Zipkin [1992,1995] and Zipkin [1995] develop the approximation for limited
capacity supply systems; Graves [1985] and Sherbrooke [1986] develop the independ
ent-Ieadtime approximation. Gallego and Zipkin [1999a] observe that the optimization
algorithm can be used in these contexts. Song and Zipkin [1992] extend the policy-eval
uation method to an MCDC demand process. --I
The numerical results of Section 8.3.5 are from Gallego and Zipkin [1999a]. The ProbleDlli
JUStifY~1
methods for (r, q) policies are attributable to Chen and Zheng [1994a] and Chen [1999].
Section 8.4: Rosling [1989] originated this approach to assembly systems, and Chen and
8.1
Zheng [1994b] refined it. V[Bj] 2:
Seclion 8.5: The original METRIC model was developed by Sherbrooke [1968]. Nah
8.2 Consi
mias [1981] and Axsater [1993a] review the literature. See Ahmed et a1. [1992], Axsater
[1990], Schneeweiss and Schroder [1992], Sherbrooke [1992], and Tripp et a1. [1991]
for recent enhancements and applications. Muckstadt and Thomas [1980] demonstrate
the inadequacy of the alternative heuristic discussed in Section 8.5.1. Recent results on From~
exact and approximate methods for (r, q) policies can be found in Axsater [1993b,1997] s~as~
and Chen and Zheng [1998]. 8.3 Consi~
There is an interesting new stream of research, which unfortunately we have been
unable to include because of space constraints, on systems under truly decentralized
management. There, not only is the policy implemented locally, but also the policy vari Use this ~
ables are chosen by local managers, according to their own local incentives. See Cachon 8.4 Considelj
[1999] for a review. demand~
Section 8.6: This approach to centralized control of distribution systems was initiated by has a DI1
Miller [1974] and Eppen and Schrage [1981], elaborated by Federgruen and Zipkin Section 1
[1984a--{;], Jackson [1988]. and Erkip et at. [1990], and reviewed in Federgruen [1993].
Section 8.7: The joint-replenishment problem of Section 8.7.1 is attributable to Renherg
and Planche [1967]. Centralized control schemes for the shared-production-capacity sys
tems of Section 8.7.2.1 have inspired considerable research activity recently. See Zheng
and Zipkin [1990], Menich and Serfozo [1991], Wein [1992], Zipkin [1995], Ha [1997],
and Pena-Perez and Zipkin [1997]. TI,e heuristic approach for the system with setups in
Section 8.7.2.2 is explored by Williams [1984], Zipkin [1986c], and Kamlarkar [1987h];
see Karrnarkar [1993] for a review. Karmarkm' and Rummel [1990] discuss in detail the
role of congestion effects in fixed costs. There are other. more centralized approaches to
8.5
this system, inspired by the ELSP; see Leachman and Gascon [1988], Gallego [1990],
Katalan [1995], and Markowitz et a1. [1995].
Section 8.8: For a general introduction to the kanban system and JIT see Schonberger
[1982] or Shingo [1985,1989]. Zipkin [199Ia] provides a critical guide to the popular
literature. Berkley [1992] and Groenevelt [1993] review the research literature.
The generalized kanban policy of Section 8.8.2 is discussed by Buzacott [1989], 8.6
Zipkin [1989], and Frein et a1. [1995]. Even more general policies are analyzed by Buza
Chapter 8 Several Items with Stochastic Demands 361
I3nt leadtimes cott and Shanthikumat [1992] and Cheng and Yao [1993]. Approximate policy-evalua
n and Zipkin tion methods ate developed by Buzacott and Shanthikumat [1992], Di Mascolo et al.
pproximation [1996], and Mitra and Mitrani [1990,1991]. Even these generalized kanban policies need
not be optimal. Veatch and Wein [1994] show that the true optimal policy can be quite
ogenous. 510 complex, even in a two-stage system.
e! al. (1992]. There is a considerable body of simulation-based research on more complex net
D for limited works than those we treat in this chapter. For instance, Vatgas and Dear [1990] study the
he independ effectiveness of alternative buffering methods in a setting much like MRP.
optimization
, policy-eyal-
~.~-s
. See Zheng Ttj = [7j-l> (1 - 7j_,e)TtJJ
. Ha [1997J. P.- 1 (I - p._I)eTt')
tt.setupsin
p. ~
(
J J J J' >I
J 0 P~
J
<ar [1987b]:
in deIail the Prove by induction that B; has the DPH distribution with paratneters (7j , P). Then, write formulas
proacbes to for E[BJ] andE[I;] in terms of these atrays.
ego [1990]. 8.5 Consider the same system as in the previous problem, and focus on the case J = 2. Show how to
implement the optimization algorithm (83.9) using the arrays above: First, write down an explicit
chonberger formula for C2 (Y) in terms of the data (Tt~, P~). Then, given s~write down a formula for C1(y) in
thepopuiar each of two cases, namely,y = SI ::; s~ andy = S1 > s~. (Use the equivalence between local and
""_ echelon base-stock policies.)
ttl [1989]. 8.6 For a two-stage series system, use the maximal approximation to obtain an explicit upper bound on
,j by Buza- C', using the results of Section 83.6. What does this tell us about the impact of IT on performance?
~_._._~~-~._._"'.~'._- --------
•
362 Foundations oflnventOl), Management
8.7 Show how to exteud the hound of Section 8.3.6 to a three-stage system: Restrict the permissible schedule.
policies to those with 52 = 53 = aJ. Argue that this leads to bounds on 51 and C*. Express these andgiYeD'
bounds in closed form using the normal approximation. Now.•
8.8 Consider a series system with fixed costs, where all k i ~ 0 except k j • Suppose we follow a policy response.
of the form proposed in the text, namely, an (r, g) policy at stage I and base-stock policies 8.13 Consider.
elsewhere. Show that the recursion (8.3.7) describes ITFj and INi' Then, argue that each ITFj is a aflXed~
mixture of the ITFj(s), with s = r + I, ... , r + g, using the equal mixing weights l/g, and IN; is a items. ",tj
similar mixture ofthe IN; (s). Conclude that E[lN;] is just the simple average of E[~(s)] over these system. ai
8.9
s's, and E[B] is the average of the E[B(s)].
For the model of Section 8.3.7.2, verify equation (8.3.13); i.e., ~~,
is, k]A.q
~j+l - ITPj + 1 = TI(s/+ I -INj , qj+l) there is
Hint: Recall that IN; = !! + ITPj + l' Consider the two cases Sj+ I - IN.; 2: 0 and Sj+ I - IN.; < 0 8.14 Show'
separately. tecbni~
8.10 Consider a two-stage series system with a fixed cost k j at stage I, but k2 = O. Replace CI (y) in
(8.3.11) by ctCY) from Section 8.3.6. Then apply Theorem 6.5.1 in Chapter 6 to obtain a bound on
the optimal cost. What does this result tell us about the determinants of performance? (Discuss at
8.11 Consider a two-stage series system with fixed costs at both stages. This problem shows how to
calculate a lower bound on the true optimal cost. Construct a pair of new systems as follows:
The~l
set s~ :S
System I also has two stages; system 2 has only one stage, which we call stage 2. We use calc
superscripts to distinguish these systems' parameters: 8.15 Consi~
kl ~ k] ki = 0 k';, ~ Ie" base-staj
equati~
D] = DI Di =D2 D';, ~D2 8.16 Consi
hl = h] j>1
. 1 ., 1" 1 2
De
For the stage-2 holdmg costs and backorder costs we select any h2 , h2, b , and b-, such that h2 + h2
pr'
~ h 2 and b j + b' ~ b.
with
Now, consider any policy for the original system. Describe a policy for each of the new
0..
systems, such that, if all three systems face the same demands, the stages receive the same orders
(in quantity and time) in all three systems. Argue that the sum of the costs in the two new systems
These policies are not necessarily optimal for the new systems, of course. Explain why it is
relatively simple to optimize each of these systems separately. Then, argue that the sum of these
optimal costs provides a lower bound on the true optimal cost of the original system. (To find the
Z
best such bound, we may search over possible values of hi, h';" b\ and b , subject to hi + h';, = h2
8.12 Consider the five-item assembly system shown in Figure 8.4.1. Suppose we follow a balanced
echelon base-stock policy. For simplicity assume that the ~j are nonincreasing in j, so s" = s, and
we initialize the system at time t ~ 0 as explained in the text, so that IN;(O) = s}'
First set S = o. Suppose a single demand occurs at time t. Describe the response of the system
according to the policy; that is, determine exactly when units of all items are shipped and received,
l~
,...
and explain why (in terms of the policy's rules). Specifically, show that the policy realizes the ideal
Chapter 8 Sel-w"alltems with Stochastic Demands 363
>ermissible schedule mentioned in Section 8.4.1. That is, demand is filled at t + L J , the earliest possible time,
Jress these
and given that as a requirement, the policy ships all nnits at the latest feasible times.
Now, suppose the Sj are strictly positive and strictly decreasing inj. Again, describe in detail the
Iowa policy response of the system to a single demand at time t.
ticies 8.13 Consider an assembly system like that of Section 8.4, with Poisson demand at rate 11., but now there is
chlTP; is a a fixed cost k} for external orders of item 1. There are no economies of scale (kj = 0) for the other
q, and~is a items, even those obtained externally (i.e., those with no predecessors). Show how to analyze this
(s)] over these system, combining the ideas of Sections 8.4 and 8.3.7.1. Specifically, suppose that item I follows an
echelon (r, q) policy, and the others follow a balanced echelon base-stock policy. (This type of poliey
is indeed optimal.) Argue that the average cost of such a policy can be computed as in (8.3.11), that
is, kl)Jq plus a simple average of the costs of certain (balanced) base-stock policies. Conclude that
there is an equivalent series system with the same fixed cost k} for external supplies.
-~<o 8.14 Show how to evaluate a local base-stock policy for the assembly system of Figure 8.4.1, using the
techniques of Section 8.4.5. Specifically, explain that
,C,(y) in
B2(t + L,) ~ max {a, (D z + D,) - s~, (DI + D z + D,) - s;}
in a bound on
(Discuss at B;,(t + L 4) ~ max {O,D4 - s~, (D, + D 4) - S3' (Dz + D, +D.) - (s~ + s~),
(D, + D z + D, + D 4) - (si + s~)}
rs how to Then show how to compute B's in terms of B~. Explain (in words) why it is intuitively reasonable to
ilIlows: sct s4 :s s~ :s s~ + s~ :s si + S4' Finally, describe an equivalent series system, with which we can
use calculate E[B s]'
8.15 Consider a two-level distribution system with compound-Poisson demands, operating under a local
base-stock policy. The text suggests an approach to two-moment approximation, leading to the
equations (8.5.6). Verify these formulas.
8.16 Consider a general distribution system (a network of items indexedj ~ 1, ... , J. where each item
j > I has a unique predecessor. and item I has none). Let Sue (i) be the set of successors of item i.
Demands occur only for end items (those with no successors). these are independent Poisson
1 that h~ + h~ processes with rates 11.;, and the leadtimes are the constants Lj. We follow a local base-stock policy
with policy variables sJ .
new
Our aim here is to extend the policy-evaluation techniques of Section 8.5.2 to this system,
arne orders
beginning with the exact approach. Suppose we know the distribution of B; for some item i. For
ew systems
j e Sue (i), let Bi;· denote that portion of B; due to demands for itemj. Argue that we can obtain BI.;
wy it is
through binomial disaggregation from B:.
What are the appropriate weights eJ" Theu, show how to
detennine Bj. We now have a recursive scheme to obtain the backorders for all items. Show how to
" of these apply the two-moment approximation here. (Clearly, these methods can be used also for the other
To find the
supply processes of Section 8.5.4.)
,+ h~ = h2 8.17 To construct the balanced approximation in the unequal-cost case in Section 8.6.2, we used the
smallest retailer cost rates for hand b in Cr Call the resulting cost function C; for now. Here, we
Ilanced explore alternative approximations of q, the optimal-cost function of problem (8.6.6). (We shall need
·=s,and to use the facts that C~ is strictly convex, and its derivative ~:(s) is precisely the dual variable 0
(a) Suppose we use instead the <Trweighted averages of the original cost rates; i.e., h = 2:; (<T/u,)hj
the system and b ~ 2:j (u/u,)bj" Call the resulting fimction C;. Show that C,+(s,) measures the cost of the
Ild received, (feasible) equal-fractile solution to (8.6.6). Conclude that C,+ 2: C~. (This approach sometimes
res the ideal works better.)
364 Foundations oflnventOlY Management
(b) Consider the special case where the retailers have proportional cost rates: There exist constants 8.21 Consider d
Tj, Tj', 13, and ej' such that hj ~ Tjej and bj + h; = (13 + Tj')eJor all j. (We have seen similar all times r:
constructions in Chapter 5 and Section 8.2.) Letting e + ~ I j (cr/cr,)ej' we have
C;(s) ~ e+E[Tj(s - D,) + (13 + Tj')[D, - st]
(Argue thai
Let s~ be the value of s,. that minimizes G; (so the corresponding ~ ~ 0). Show that C: (s~) ~ the variabll
C:(s~). (So, this approximation agrees with the true value at least at one point. Notice, this argue that I
same value must also minimize C;.)
(c) Continuing with the proportional-cost case, let e - ~ min {ei]' We have C;:(s) = (e- Ie+)C;(s).
Show that G;(s) 2 C;:(s) + [C:(s~) - C;:(s~)] = C;:(s) + (e+ - e-)cril3 + Tj')<I>(z:), where
z~ solves <I>(z) ~ w, ~ (13 + Tjo)/(13 + Tj') and Tjo = Tj' - Tj. (Hint: Let C = C,:-'(s), and set
s- to solve G;'(s-) = C. Thus, s- is a function ofs. Show that, ifs < s~, then C < 0 ands
2 s, so ~ -s: C; while if s > s~, then C > 0 and s- -s: s, so ~ 2 C.) Adding this constant to
C; thus provides a stronger lower bound on C~J while retaining the same simple fonn.
8.18 Section 8.6.5 discusses a coupled system with identical retailer costs and a fixed cost for external
orders. In this setting. the myopic allocation rule must be revised, in general. When the retailers are
identical, however, no revision is necessary. Explain why. That is, argue first that, if the Zj(O) are
identicaloverj, so are the Zj(t) for all t. Secoud, suppose we revise problem (8.6.4), replacing Dj by
DJ (L; + t) in the cost functions Ci Show that the new problem is equivalent to the original.
Extend this Idea to the case of compound-Poisson demands. Now, vj(t) ~ AjE[lj](L; + t) and
crJ(t) ~ AjE[Y;](L; + t). Show that the results above remain true for identical retailers, and more
2
generally when the L; and the ratios E[YJ]/Aj E [lj] are equal over j.
8.19 Consider the shared-production-capacity system of Section 8.7.2.1 with no economies of scale,
local control, and identical products. Suppose we are able to determine the distribution of fO, the
overall occupancy of the processor, and we then apply binomial disaggregation to obtain fOJ , the
occupancy of item j's orders. Finally, approximate fOj with a normal distribution. Show that the
total optimal cost is proportional to q:, where q:2 = V[fO] + (J - I)E[fO). Compare this formula
with (8.5.5).
8.20 This problem extends the modeling approach of Section 8.7.2.2 to the case of nonidentical
products and stochastic processing times: Item j has Poisson demand with rate Aj and follows an
(r, q) policy with variables (Ij, q). Its processing time Sj includes a random setup time Tj and a
random unit processing time Pj; all these times are independent, even for the units within a batch.
Approximate each item"s orders by a POiSSOU process. Now, as explained in Section 8.7.2.1, the
production facility becomes an MlGII system.
(a) Write down equations for 2 E[Sjl and V[Sj] in terms of E[Tj], E[P ], V[Tj], V[P ], and qj' Also,
. i. .i
use these to compute E[ Sf ).
(b) Let~j denote the order rate for itemj, Pj ~ AjE[Pi ], J,!j ~ £l,jE[Sj], ~ ~ Ii £l,j' P ~ Ii Pi' and
J,! ~ Ii J,!j' Write the simplest expression you can for l! in terms of the data and the variables qj'
(Use the summary parameter p.)
(c) Let L Q denote the time spent waiting in the queue itself (the total time L less the actual
processing time). In our terms equation (7.3.2) can be written as E[L Q] = l'£l,E[S2]/(l - l!). So,
write the simplest expression you can for £l,E[s'], again using the data and the qj' (It is possible
to obtain a similar formula for V[L Q ). Instead, approximate V[L Q ] by E 2[L Q ].)
(Ii) Finally, letting Lj be the totalleadtime for itemj, express E[L,) and V[q in terms of E[L Q ] and
the data.
,I
J
.111111
exist constants 8.21 Consider tbe generalized kanban policy of Section 8.8.2. Sbow that the following identity bolds for
een similar all times t:
(c-/c+)C;(s).
1'I>(z:'), where
'(s), and set
- <Oands
constantto
fOfID.
for external
tie retailers are
Ie Zj(O) are
:placing Dj 1»'
ginaI.
Lj + t) and
i,. and more
i of scale,
.ofl0, the
IinIOj • the
... that the
his formula
lical
"'Uows an
1j and a
IJin a batch.
r2.1, the
1Qr Also,
EjPj. and
.'3riables Qr
tuaI
(I - Q). So.
:is posstble
fEIL Q] and
C H A p T E R
9 TIME-VARYING,
STOCHASTIC DEMAND:
POLICY OPTIMIZATION
Outline
9.1 Introduction 367 9.6 Leadtimes 404
9.2 Extreme Cases 368 9.7 World-Driven Demand 415
9.3 Discrete-Time Formulation 371 9.8 Series Systems 420
9.4 Linear Order Costs 374 Notes 427
9.5 Fixed-Plus-Linear Order Costs 395 Problems 428
9.1 Introduction
Tbis chapter explores situations where demand is subject to both predictable and unpre
dictable variations. The scenario thus combines features that we studied separately in
Chapters 4 and 6.
We begin in Section 9.2 with a heuristic analysis ofa few extreme cases, in the same
spirit as Section 4.2. Apart from that initial section, the primary goal in this chapter is to
prove rigorously that the optimal policy takes on a certain form, depending on the de
tails ofthe system. We have done nothing of the kind until now. In Chapter 6, for exam
ple, we argued that a base-stock policy is plausible under certain conditions, but here we
show that, under similar conditions, it is truly the best among all possible policies.
To achieve this ambitious aim, along with the greater generality of time-dependent
demand, we adopt a discrete-time formulation, as in Chapter 4, where all important
events occur at prespecified time points. Also, we introduce the concept of dynamic pro
gramming, a powerful mode of formulation, analysis, and computation.
367
-- ~_.~----- --
Here is the plan: Section 9.3 describes the basic setup. The costs in the model are
just like those assumed previously-fixed and variable order costs, a holding cost for in
ventory, and a penalty cost for backorders. The demand at each time point is a random
variable, whose distribution is a function oftime. The demands at different points are in
dependent (until Section 9.7).
Section 9.4 treats the special case without economies of scale, that is, with a linear
order cost. The optimal policy is indeed a base-stock policy. It uses a distinct base-stock
level for each time point. Otherwise, it works just like the base-stock policy of
Chapter 6. It is not easy to calculate the optimal base-stoek levels, so we explore a sim
ple approximation, called the myopic policy, which often works well. This myopic ap
proach is an important one, and we use it throughout the chapter. When the data are sta
tionary, the optimal base-stock level too becomes stationary, as in Chapter 6.
The same policy structure remains optimal for certain variations of the basic model,
namely, for lost sales instead of backorders, and for the broader objective of profit in
stead of cost. For imperfect supply quality, however, a base-stock policy is not optimal.
We need to consider a larger class of rules called inflated base-stock policies.
In Section 9.5 the order cost includes a fixed component as well as a linear one.
Here, an (r, s) policy is optimal. With time-varying data, both rand s change over time.
In the infinite-horizon, stationary setting, a stationary (r, s) policy is optimal.
The models of Sections 9.3 to 9.5 have a minimalleadtime. Section 9.6 shows that the
results remain valid for any constant leadtime., and also with sequential stochastic leadtimes.
Here, finally. we have full analogues in discrete time of the continuous-time models of
Chapters 6 and 7. We point out in Section 9.6.4 that simple structured policies are optimal
in the continuous-time setting also. (These results do not extend to lost sales, however.)
The remainder of the chapter extends these results in various directions. Section 9.7
examines a system with world-driven demand, the discrete-time analogue of the com
plex demand process of Section 6.3, where the distribution ofdemand in each period de
pends on the current state of the world, The earlier results remain valid with minimal ad
justments. With a linear order cost, for example, a base-stock policy is still optimal, but
the optimal base-stock level reflects the current state of the world as well as time.
Section 9.8 treats a series system in discrete time, With linear shipment costs, an
echelon-base-stock policy is optimal.
tion 4.2.1 and Section 4.2.2. So, it makes sense to ignore the current value ofA(t) and
,Iding cost for in
Next, suppose A(t) changes slowly. Let D(tIL) ~ D(t + L) - D(t) be the leadtime de
stinct base-stock
mand, a Poisson-distributed random variable with meanE[D(tIL)] ~ J:+L A(1') dv. Defin
srock policy of
ing ,vt) ~ (IIL)J:+L A(v) dv, we can write E[D(tIL)] = ,vt)L. Since A(t) changes slowly,
'I: explore a sim
For now, suppose there is no fixed order cost. It is plausible, then, to use a base-stock
rhe data are sta
lter 6.
policy, with a base-stock level set) that itself depends on time. Specifically, define the
function
rhe basic model,
r is not optimal.
whereC(y) = h[yt + b[y]- as before, and set s(l) ~ s +(t) to minimize CU, s) over s.
licies.
That is, specify set) as if ,vI) were really constant. This approach is parallel to that of
as a linear one.
Section 4.2.3. Because it looks ahead only a Ieadtime into rhe future and ignores subse
mnge over time.
quent costs, it is called rhe myopic approach. Also, estimate rhe optimal average cost by
timal.
C*, where
6 shows rhat the
IJastic leadtimes.
~ limT...,oo(~)J6 C*(t) dl
-time models of
I as time.
where
ment costs) an
1'(1) ~ rrOCI) ~ ,v1)L
Deticn of time.
C* ~ CA + (b + h)<I>(z*)rr_
t mean E[ D(I)]
where A is the average demand rate and rr _ rhe average of rr(l) over I. The maximal ap
s and the lead
proximation leads to similar results.
,ty and heuris Here, as in Chapters 6 and 7, system performance depends on the standard devia
tion of leadtime demand; the mean vet) has no impact. Notice, u: measures residual,
unpredictable demand variation, not total variation. Raw observations ofdemand also in
clude predictable variation, due to changes in v(I), which does not belong in an estimate
of u:.
" rhe changes
An even cruder heuristic often works well: Notice, rr(l) depends on rhe square root
lOdels of Sec-
of ,vI), and so is robust with respect to demand changes; the mean 1'(1) is linear in MI),
370 Foundations ofInventory Managemerl!
so it is more sensitive. Consider some interval over which 1.(t) changes only a modest
amount, and let rr_ be the average of rr(t) over that interval. Then, fix the safety stock to
the constant rr _z* and set set) = vet) + rr _zoo Use (9.2.2) with this rr _ to estimate per
formance. Adjust rr _ only rarely, when .Mt) departs significantly from its original value.
Under this heuristic the base-stock level moves in parallel with the mean leadtime
demand. This is the simplest way to account for demand changes, and it is used widely
in practiee. Moreover, apart from the order-cost term cl.(t), the cost rate C*(t) remains
constant.
Now, suppose there is a positive fixed cost k > O. The analogue of the myopic ap
proach above is this: Use an (r, q) policy with time-dependent policy variables ret) and q(t).
Choose these variables to minimize a cost function that combines (6.5.16) with (9.2.1):
This method should work well provided the demand rate changes slowly, for the same
reasons as in Section 4.2.3.
Or, set ret) and q(t) using the simpler heuristics mentioned in Section 6.5.3. For in
stance, set q(t) ~ [2kl.(t)/hw] 1/2 by analogy to (4.2.1), and ret) = s + (t) - I3q(t) for some
reasonable fraction 13.
Or, even more crudely, fix q(t) to some average value q, use some average rr _ of
rr(t), approximate set) ~ vet) + rr _z* as above, and set ret) = set) - I3q(t) = vet) +
(rr _z* - I3q). Adjust q and rr _ only in response to a substantial demand change. Here,
the reorder point moves in parallel with the mean leadttme demand. This simple tech
nique too is popular in practice.
Here is one way to estimate the average cost: Approximate C*(t) = qt, ret), q(t)]
for each t separately, using the bounds of Theorem 6.5.1 together with the normal or
maximal approximation. Then, average over time as in (9.2.2).
,
9.2.3 Other Supply Processes 9.3 DiSCI
The same ideas extend directly to systems with stochastic leadtimes, generated by either an
independent-parallel or an exogenous-sequential supply process. Again, small and/or fast
demand changes can be ignored, while slow changes can be handled by a myopic approach.
In the exogenous-sequential case, for instance, redefine D(tIL) to be the leadtime de
mand starting at time t, i.e., the demand between t and t + L, where now L is the mar
ginalleadtime, a random variable. In general, D(tIL) is no longer Poisson, nor does it
have any other any simple form, but we can approximate: Compute its mean and vari
ance vet) and rr\t) (Problem 9.1), and fit a standard distributional form, such as a nor
mal or a gamma. Then, proceed with the methods above.
Recall, this supply system has unlimited capacity, and that makes it tractable. A
finite-capacity supply process requires more care: Small and/or fast demand-rate
changes can again be ignored. Slow changes too can be treated as above, but only when
the changes are very slow, and the demand rate stays well below the system's capacity.
For instance, consider an MIMI! system with processing rate J.L. The direct analogue
ofthe myopic approach sets pet) = I.(t)/"" and replaces D(tIL) in (9.2.1) by a geometric
,"" ',I"
, only a modest random variable with parameter pet). That is, pretend that the supply system reaches its
, safety stock to equilibrium state at every point in time. (This is sometimes called the pointwise station
to estimate per ary approximation. For more on this and related techniques see Green and Kolesar
;original value. [1991] and Whitt [1991].)
mean leadtime Clearly, this approach breaks down if A(t) > I.L for any I. Even when demand never
is used widely exceeds capacity, the supply system itselfresponds slowly to changes in its input. For ex
, 0(1) remains ample, if A(I) is near I.L for a while, the occupancy JO(i) grows large, and this congestion
will persist for some time even after ACI) falls lower. Thus, the equilibrium approxima
!he myopic al> tion is valid only when demand-rate shifts are slow enough for the supply system to re
Ies r(l) and q(I). cover quickly. This condition in turn depends on p(t) itself; the supply system responds
with (9.2.1): quicker when p(t) remains well below 1.
These are stringent conditions. What if they do not apply, ifA(l) sometimes exceeds
capacity or changes faster than the supply system? First, if the base-stock level s(t) is
held constant, then the supply system behaves like a queue with nonstationary inpnt.
r, for the same Methods to analyze such queues have been devised, e.g., by Rothkopf and Oren [1979].
However, it seems more sensible to build up inventory in advance of demand surges, as
.6.5.3. For in in Section 4.2.5, by increasing s(t). But when and how much? We know no good, simple
Jlq(I) for some answers. The only way to get any answers, given the current state of the art, is by algo
rithmic methods (as outlined in Problem 9.4).
mverage (1_ of Sometimes it is possible to adjust the processing rate in response to the demand rate.
iq(1) ~ V(I) + Consider the following special case: We can and do set [L(I) proportional to A(t), so the
change. Here, ratio A(I)/[L(I) is the constant p < I. It turns out that this system is equivalent to one with
s simple tech stationary J\ and (.1, observed on a clock of variable speed. Thus, there is no reason to
change the base-stock level. Assuming we do keep it constant, the analysis of Sec
ql. r(I), q(t)] tion 7.3.2.1 applies directly.
!he nonnal or In sum, for a system with limited capacity and nonstationary demand., the current
state of the art provides no single methOd to handle all cases. Some cases are easy, some
are hard.
At the horizon Twe observe xCI), but that is all; there is no order z(T) or demand d(I).
This model is usually described in the literature as having a leadtime of zero. We
haven't said exactly when orders arrive, and we don't really need to; (9.3.1) remains valid
as long as each order arrives during its own period. If z(t) arrives at the very end of pe
riod t, however, it makes sense to call the leadtime 1, not O. This usage is consistent with
Section 9.6, where we introduce an explicit leadtime. Whatever we call it, the leadtime
is certainly short. Remember, a period can be very short, especially when the discrete
time model is constructed to approximate a continuous one. So, the model is quite spe
cial. Fortunately, virtually all the results carry over to more realistic models with longer
leadtimes, as shown in Section 9.6.
The nature of the decision problem is quite different from that of Chapter 4. There,
because the d(t) were known in advance, we could determine the z(t) all at once. Here,
we choose each z(t) at point t, based on what we know at step 2. At that moment, we know
x(t), which reflects all past demands, but not d(t) or any later demands. It makes no sense
to specify z(t) any earlier, for that would ignore useful information.
What we want, then, is not a set of fixed decisions, but rather an order policy, a
rule for making decisions based on current information. In contrast to Chapter 6, we
place no advance restrictions on the form of the policy. One of our primary goals is to
determine which type o.fpolicy is best, given the economics of the situation. Accord
ingly, we bypass the analysis of physical performance measures, such as average back
orders. The costs associated with inventory, backorders, etc., enter the formulation
from the beginning.
9.3.2 Costs
The cost parameters are
Future costs are discounted at rate "I, as in Section 4.4.2. The cost factors indicate the
original, undiscounted costs; unlike Section 4.4.2, we do not replace k(t) f- -ytk(t), etc.
Chapter 9 Time- Jiuying, Stochastic Demand· Policy Optimization 373
Letting 8(z) denote the Heaviside function (I whenz > 0,0 otherwise), the total or
der cost at time t is k(t)8(z(t» + c(t)z(t). The inventory-holding or backorder-penalty
cost is assessed onx(t) at step I oftime t. This cost can be written compactly as Crt, x(t»,
where
i\t, x) = h(t)[x]+ + b(t)[x]
Actually, it is convenient to count this cost just before time t, at the end of period t - I.
z(t) arrives. The
The corresponding cost at the end of period t is Crt + I, x(t + I». This convention bet
I:
ter reveals the links between causes and effects; the decision z(t) affects Crt + 1,
(9.3.1) x(t + 1»but not (\t, x(t». In particular, ignore the constant C(O, x(O» = C(O, xo).
Denote
... demand d(T).
me of zero. We yet) ~ x(t) + z(t)
I) remains valid
This IS the inventory position just after ordering at step 2. Thus, (9.3.1) is equivalent to
'l:T)' end of pc
consistent with x(t + I) ~ yet) - d(t) t ~ 0, ... , T - 1
it, the leadtime
Define the function
en the discrete
lei is quite spe
ql, y) ~ E[C(I + I,y - d(t))]
leis with longer Then, Crt, y(I» measures the expecled inventory-backorder cost, as viewed from step 2.
We call C(t, ') the one-period cosl function.
l3pter 4. There, (This setup differs from the most common one in the literature, but the difference is
I at once. Here, purely cosmetic. There, h(l) appears in place ofour h(1 + I), and b(l) replaces our bet + 1).)
men!, we know We have seen functions like C and C in earlier chapters. As before, for each I.
mkesnosense C(I, y) and ql, y) are convex as functions ofy. Such convexity properties were impor
tant in Chapters 6 to 8, and they are crucial here.
onIer policy, a The strict discrete-time scenario can be relaxed somewhat, as in Section 4.4.3. De
Chaptet 6, we mands can occur and holding-backorder costs can accumulate throughout the period.
my goals is to The result is a slightly different one-period cost function qt, y). The new function, how
iltion. Accord ever, has the same form as the one above.
average back
le formulation
9.3.3 Salvage Value
Sometimes, when T is finite, it is convenient to include a special mechanism to "settle
accounts" at the end. Specifically, when x(T) < 0, we must purchase stock to fill the re
maining backorders -x(T) at the unit purchase price crT). (Remember, the last real or
der is at time T - I, so the meaning of crT) is unambiguous. There is no fixed cost for
such purchases, i.e., k(T) ~ 0.) Also, if x(T) > 0, we can sell the leftover stock at the
same price; thus, we receive total revenue c(T)x(T), or equivalently, we pay a cost of
-c(T)x(T). In sum, regardless of the sign ofx(T), there is a tenninal cost -c(T)x(T).
The tenninal-cost factor crT) is called the salvage value. Actually, we can do with
out it; we achieve the same effect by adding crT) to beT) and subtracting the same
amount from h(T). (Indeed, we can represent asymmetric terminal costs by adjusting
"S indicate the beT) and h(T) appropriately.) Nevertheless, the salvage-value construction is useful in
<- -y'k(t), etc. certain situations. If there is no such mechanism. just set crT) = O.
"'> I
This is easy to solve for any given x. To detennine an optimal policy, however, we must
solve (9.4.1) for every value of x.
Define
The optimal solution to (9.4.1) depends ou the relation between x and s*. Ifx :-;; s*,
then y ~ s* is feasible and hence optimal. If x > s*, then the optimal solution is y ~ x,
the linear fimc because H(y) is nondecreasing over the entire feasible range y 2:: x. This is precisely a
base-stock policy with base-stock level s*. If the initial inventory is above s*, don't or
st, we analyze a der; if it is below s*, order the difference, i.e., just enough to raise the inventory to s*.
1, we extend the To compute s*, we must solve H'(y) ~ O. But C(y) has the same fonn as the
zon. In all these average-cost function of Section 6.5.2. Specifically, C'(y) ~ [h - (b + h)F°(y)], where
FO is the ccdf of d (~ d(O». Consequently,
Ies, profit max
H'(y) ~ (e+ + h) - (b + h)Fo(y)
.- explore other
and s* solves
c++ h
FO(y) ~ b + h (9.4.2)
ision at time O. This is entirely analogous to equation (6.5.3). Only the cost ratio differs. The demand d
I the inventory plays the same role here as the leadtime demand in a continuous-time model. If we use
It's all. Because a nonnal approximation for dwith v ~ E[d] and <T2 ~ V[d], and setz* to solve q,o(z) ~
we abbreviate (e+ + h)/(b + h), thens* = v + <Tz*, andH(s*) ~ ev + <T(b + h)<j>(z*).
, and C(O, y) = Of course, (9.4.2) makes sense only when e + < b. If e + ;" b, i.e., e ;" b + -ye(l),
"Y. then H(y) is nondecreasing, so set s* = -00. The optimal policy tells us never to order,
lrays situations i.e., to get out ofthe business. It is cheaper to incur the penalty cost b plus the discounted
nagazines, and tenninal cost -ye(l) than to purchase a unit at cost e. Likewise, assume c++ h > O. Oth
oc newsvendor erwise, H(y) is nonincreasing; indeed, it is strictly decreasing (except in degenerate
ikely 0, but we cases), and s* = +00, Here, we can make unlimited profit by purchasing units at cost c
and later selling them at price e(I), even after deducting the holding cost h. Ruling out
the inventory these bizarre cases, the fraction in (9.4.2) is strictly between 0 and 1.
the discounted Let Vex) denote the optimal cost of problem (9.4.1), regarded as a function ofx.
Vex) = -ex + V+(x)
where
~.:
FIGURE 9.4.1
Optimal costjUnction (linear order cost).
H(x)
V1-(x)
s' x
how can we intelligently select yeO)? And then there is y(2), which should probably de
pend on y( I) and yeO), or maybe it's the other way around.... We seem to be trapped!
Fortlmately, there is a way out. Tbis is precisely the sort of dilemma addressed by
dynamic programming. The key idea is to think backward in time.
Consider time point T - 1. Once we arrive there, no matter what may have happened
in the past, we sball face a single-period problem like the one above. We know bow to
solve that problem. The optimal base-stock policy tells us the best yeT - I) for any
x(T - I).
Next, suppose we find ourselves at time T - 2. We face a two-period problem. To
account for the future, assume tbat we shall act optimally at T - I, according to the pol
icy obtained already. With this infonnation, as we shall see, we can detennine the best
yeT - 2) for all x(T - 2).
At time T - 3, assume that we shall act optimally at bOlh future points. Again, we
can use this information to optimize yeT - 3). Continue working backward in this way,
all the way to point I ~ O. At that point, we can detennine the best y(O) for the actual
x(O) ~ Xo.
The fact that this approach works is sometimes called the principal ofoptimality. It
should be intuitively clear, and it can be proven rigorously.
Here is a precise statement of the approach: Define the following functions for
t = 0, ... , T:
We compute these functions recursively. In the process, we obtain the optimal policy.
First (or rather, last),
V(T, x) = -c(T)x (9.4.3)
Suppose we have determined Vet + I, x), along with the optimal policies for points
I + I through T - I. To address the prohlem at time t. given x(l) ~ x. define
H(I, y) = c(l)y + CCI, y) + "fE[V(1 + l,y - d(I))] (9.4.4)
This quantity measures all the relevant costs if we choose y(l) ~ y. The first two terms
(minus the constant C(I)x) represent the costs at time I itself. The last term is the expected
value of all future costs, assuming we act optimally in the future. since x(t + 1) =
y - d(I). The problem at point t, then, is to choose y to minimize H(I, y), subject to
y ~ x. The solution to this problem for each x gives the optimal policy at time I. Then set
V(t, x) ~ -c(t)x + min {H(I, y) : y ~ x} (9.4.5)
Now, using V(t, .), we can proceed to compute V(t - 1, .) in the same way, then V(I - 2, '),
and so on.
Equations (9.4.4) and (9.4.5), together with the boundary or terminal condition
(9.4.3), describe a recursive scheme to compute the optimal-cost functions for all time
points, as well as the optimal policy. They arefimctional equations, i.e., equations whose
I unknowns are the functions V(I, .) and H(I, -). This model is often called the dynamic
programmingfonnulalion of the problem. [Dynamic progranuning and oplimal control
I probably de are, essentially, alternative names for the same subject. We can just as well call (9.4.3)
I be trapped!
to (9.4.5) an optimal-control mode1.]
addressed by We're done, in principle. In practice, however, two difficulties arise:
First, x andy are continuous variables, and so, for each t, (9.4.4) prescribes an infi
ave happened nite number of expectations, and (9.4.5) an infinite number of optimizations. To solve
knowhow to functional equations of this kind typically requires numerical approximation techniques.
- I) for any One approach, the simplest conceptually, is to discretize and tnmcate the state space: Se
lect a finite set of values for the x(t), and similarly restrict the d(t) and y(I). This selec
I problem. To tion must be done carefully, of course, to obtain accurate results.
IIg to the pol Second, the full optimal policy (the best y for every x and I, written out, say, in a
nine the best large table) would be hard to understand and tedious to implement. Fortunately, it has a
simple, appealing structure:
ts.Again, "" THEOREM 9.4,1, For all t:
I in this way, (a) H(I, y) is a convex function ofy.
or the actual (b) A base-stock policy is optimal. The optimal base-stock level s*(t) is the smallest
value ofy minimizing H(t, y).
>ptimality. It (c) V(I, x) is a convex function ofx.
Unctions foe PROOF. By induction on t: We verified the result for t = T - I in Section 9.4.1. Sup
pose it is true for any t + I. The first two terms of H(t, y) are convex in y, as in the
single-period problem. And, because Vet + I, x) is convex (by the induction hypothesis),
assuming so is the third term of H(t, y). Thus, H(t, y) is convex iny; we have established part (a).
Parts (b) and (c) follow immediately, just as in the one-period model.
378 Foundations ofbwentory Management
By the way, one can also show that H(I, y) and V(I, y) are continuously differentiable in
y. SO, S*(I) solves H'(t, y) = O. (Here, H' is the derivative with respect to y,)
In sum, the optimal policy is characterized by a single parameter, the base-stock
level S*(I), for each time point. In general, however, it is not easy to compute the s*(I),
Even koowing the form ofthe optimal policy, we still must perform the recursive calcu
lations indicated in (9.4.3) to (9.4.5) using numerical techniques.
V+(T, x) ~ 0 (9.4.6)
(problem 9.2 asks you to verify the alternative expression for H in (9.4.7).) Evidently,
V+(t, x) plays the same role here as V+ (x) in the single-period problem, Indeed, we can
write
V+(I, x) ~ H(I, max {s*(t),x}) (9.4.9)
Now, C+ plays the role of the current period's cost in this transformed model. Let
s+(I) be the smallest value ofy that minimizes C+(I, y). The corresponding base-stock
policy minimizes the current cost while ignoring the future, so we call it the myopic pol
icy. Because C+ (t, y) is just the actual one-period cost C(I, y) plus a linear function of y
and a constant, it is convex, and so easy to optimize. In particular, s +(t) solves an equa
tion of the same form as (9.4,2):
° c+(I) + h(1 + 1)
(9.4.10)
FdV)(y) = b(1 + 1) + h(1 + I)
This is essentially the same approach outlined in Section 9.2.2 in the continuous
time setting with a slow-changing demand rate. s +(1) reflects the data for period I in a
simple manner, like s* in a one-period model. For instance, suppose the cost factors are
stationary, we approximate d(l) using a normal distribution, denote V(I) ~ E[d(I)] and
0'\1) = V[d(I)], and set z* to solve q,o(z) ~ (c++ h)/(b + h). Then, s +(1) ~ v(l) +
O'(I)z*, and C+(I, s+(I» = cV(I) + O'(I)(b + h)q,(z*).
The myopic and optimal policies are closely related:
Chapter 9 Time-Hlrying, Stochastic Demand: Policy Optimization 379
EXAMPLE 9.4.A
We shall solve fOUf models. In all cases, T = 5, 'Y = 1, and the holding and penalty costs are con
stant at h(l) = 1, bet) = 9. The demands d(t) have either Poisson or geometric distributions.
First, consider a pair of systems with constant C(l), so every c+(t) = O. The mean demand
E[d(t)] stays constant at 40 until time 3, when it abruptly drops to 2, and remains eonstant there
after. The systems differ only in their demand distributions, Poisson or geometric. Figure 9.4.2 dis
plays the optimal and myopic policies.
In the Poisson-demand model, s*(t) = S + (l) for all t; the myopic policy is optimal. In the geo
metric case, we see that s*(3) = s +(3) = s*(4) = s + (4), as expected, but s*(2) < s \2) and even
s*(I) < s + (l), anticipating the drop in demand at time 3. It is remarkable, however, that the dif
ferences are so small. These anticipation effects die out at time 0, where again s*(O) = s + (0).
Next, consider a pair of systems with constant mean demand E[d(t)] = 30. The order costs
e(t) change, however; they follow a general downward trend with fluctuations along the way. Fig
ure 9.4.3 shows the results.
We see that s + (t) is large when we anticipate a higher cost in the next period, i.e.. c(t) <
c(t + 1), so c+(t) < 0, and small in the opposite case. The s*(t) follow the same pattern. Indeed
only for t = 0 and 2 do we observe s*(t) < s + (t). Even then, the differences are slight, except in
the geometric-demand case for t = O.
In these four cases, the myopic policy is almost always nearly optimal. Such a close relation
is typical. The myopic policy can be far from optimal, but only when the problem data change rap
idly over time.
FIGURE 9.4.2
Optimal and myopic base-stock policies (stationary costs).
100
-->1<-"(1)
~"
60
I Poisson demands
"
40
20
o
o 2 3 4
Chapter 9 Time- furying, Stochastic Demand: Policy Optimization 381
FIGURE 9.4.3
Optimal and myopic base-stock policies (stationary demands).
~ costs are con
stributions. 140, ::>I'< 8
.e mean demand
s constant there -B- c(t)
120
Figure 9.4.2 dis ---*- s+(t)
--.- 8*(1) 6
inIal. In the geo 100
: 5 - (2) and even
'\·er, that the dif 80
(0) ~ s +(0). .~ 4
n
o
'0
The order costs
IIlg the way. Fig-
" 60 '"
nod, i.e., c(t) < 40
2
pattern. Indeed,
Poisson demands
slight, except in 20
a close relation
:lata change rap 0 o
0 2 3 4 5
The myopic approach also provides a simple lower bound on the optimal cost: Let
7T denote any policy. The total expected cost starting with x(O) = x using policy 7T can
be written
Eachy(t) here is determined by the policy andx(t). The demands play an implicit role,
through x(t + I) = yet) - d(t). Using this identity to eliminate thex(t) for t > 0, we get
V(O,X]7T) = E[ l;~J'YtC+(t,y(t))] - c(Olx
Now, s +(t) minimizes C+(t, y). Consequently, V+(O, XI7T) '" V_CO) for all x, where
V_CO) ~ l;~htc+(t, s +(t))
This is true, in particular, for the optimal policy; i.e., V+ (0, x) '" V_(0). With stationary
cost parameters, llllder the nonnal approximation, the lower bound becomes
o c+ + h
Fd(y) = b+h (9.4.11)
Assume that the cost ratio lies strictly between 0 and 1, so s* is finite. Thus, s* is the
(stationary) myopic base-stock level.
THEOREM 9.4.3. The statiouary base-stock policy with base-stock level s* is optimal.
The myopic policy is optimal!
Because s* remains constant over time, we can describe the policy as a demand
replacement nile: Assume that Xo oS 0'*. The policy sets y (0) ~ 0'*, sox(l) oS so, so y(l) =
s*, and so on. Each yet) = s*, so x(t + I) = s* - d(t), so z(t + 1) = dU). Thus, each order
just replaces the prior period's demand. (If Xo > s*. wait until demand depletes the inven
tory so that x(1) oS so; from then on the demand-replacement rule takes effect.) This is the
way we interpreted a base-stock policy in earlier chapters. (When s*(t) changes over time,
of course, this interpretation breaks down; a time-varying base-stock policy is not a
demand-replacement rule.)
It is no harder to solve the infinite-horizon problem than the single-period problem;
equation (9.4.11) is identical to (9.4.2). When -y = I (the average-cost case), the cost ra
tio in (9.4.11) becomes h/(b + h) = I - w,just as in (6.5.3). (As in the continuous-time
model, the optimal policy does not reflect the order-cost rate c, because under any rea
sonable policy the average order cost is tbe constant cE[ d].)
The optimal cost is easy to compute also: The optimal average cost (for the case
-y = 1) is given by
v* ~ C+(0'*) = cE[d] + C(s*)
For the discounted-cost case (-y < I), let V*(x) be the optimal discounted expected cost
over the infinite horizon, and V+*(x) ~ V*(x) + (x. Then,
C+ (s*)
V+*(x) ~- x S s*
I - -y
The remaining values take a bit more work. It is necessary to solve the functional equation
V+(x) = C+(x) + -yE[V+(x - d)] x> s*
,,11.11,.-..
This calculation requires numerical teclmiques. Under the normal approximation the ba
sic quantity C+ (s*) becomes
C+(s*) ~ cv + arb + h)<jJ(z*)
IS and the de
l) allhave the
, assumptions 9.4.4.2 Proof
Before justifying this result, we need to formulate the problem precisely. We distinguish
Ie: Derme between the two cases 'I < I and 'I = 1. When 'I < I there is some hope that the total
expected cost is finite over an infinite horizon, while if'Y = 1 the total cost is certainly
infinite under every policy, so we need a different criterion. (This is true, not just of in
ventory models, but of dynamic programs in generaL)
In the discounted-eo5t case ('I < I) suppose we ehoose a particular policy, say 7T,
an ordering rule for each of an infinite number of periods. Let
V(XI7T) ~ total expected discounted cost using policy 7T, starting withxo = x
(9.4.111
We can write
bus. 5* is the V(XI7T) ~ E[~~~o "Y'{e(y(t) - x(l)) + C(y(t»} I x(O) ~ xl (9.4.12)
The expectation is over all possible demand sequences. We want to choose a policy 'TT*
r*isoptimaL that minimizes V(XI7T) for all x. Let V*(x) denote the optimal cost.
We have negleeted some fine points in this formulation: Each sample path of the
demand process contributes an infinite series to (9.4.12), and it is fair to ask whether
ISa demand sueh series converge finitely. One can show that, for any reasonable policy 1i', almost all
s*.soy(l)=
the series do converge finitely (i.e., with probability 1). Moreover, the expectation is well
IS. each order
defined and finite.
Ii'5 the inveo
Turning to the average-cost case ('I = 1), we pose a different objective:
L I This is the
~ O\'er time. V(XI7T) ~ expected long-run average cost using policy 7T, starting with Xo ~ x
licy is not a
= E[limT~ro (~)t~;':-J {c(y(t) - x(I» + C(y(t»} - cx(T) I x(O) ~ x] (9.4.13)
iod problem:
~ the cost ta The goal is to find 7T' minimizing V(X]7T) for all x.
bnuous-time Again, we have neglected eertain technical issues. Everything works out satisfacto
Ide.- any rea- rily for our problem. For other dynamic programs, however, these issues require more
care. Indeed, the limit in (9.4.13) need not exist, and it is common to reformulate the cri
Ifor the case terion as
V(XI7T) = lim SUPT~ro (~)E[~;':-J {e(y(t) - x(l)) + C(y(t») - cx(T) I x(O) ~ x]
For our problem, this objective is equivalent to (9.4.13).
On the other hand, for reasonable problems (like ours) and reasonable policies,
q>ected cost
V(XI7T) is the same number for all x. In the long run, the effect of initial conditions dis
appears. So, we can denote the cost of policy 7T by V(7T) and the optimal cost by v'.
Now that we have formulated the model, let us verify the solution outlined above.
PROOF OF THEOREM 9.4.3. We prove the result for 'I < I; the case 'I ~ I is virtu
oal equation ally identical. Substitute x(t + I) = y(l) - d(t) into (9.4.12) to obtain
Now, for every t and every possible x(t), the base-stock policy with base-stock level s'
sets yet) to minimize C+(y(t», subject to yet) '" x(t). (The argument is the same as in
the single-period model.) Moreover, ifx(t) 0; s' and we continue to apply this policy, we
shall have yet + I) = s',y(t + 2) ~ s', etcetera. On the other hand, ifx(t) > s', the pol
icy sets yet) as small as possible, to x(I). So, whatever value of d(t) happens to occur,
x(t + I) will be as small as possible, and the feasible range for yet + I) as large as pos
sible. The same is true for subsequent times. In sum, this policy minimizes C+(y(t»for
all t andfor every demand sequence. It certainly minimizes the expected discounted sum
of these costs. and so is optimal.
This argument is quite easy and ilhuninating. It exploits certain special characteris
tics of the model. Other dynamic programs, unfortunately, cannot be analyzed in this
way, including the fixed-cost model ofthe next section. We now sketch (omitting certain
details) an alternative proof ofTheorem 9.4.3, which illustrates the general approach re
quired by more complex models. Again, we focus on the case -y < I. (There is an argu
ment along these lines for the average-cost case, but it is more intricate.)
This approach analyzes the sequence of finite-horizon models for increasing values
of T. To identify a time point in this context, what matters is not its distance from time 0
but rather its distance from the horizon. Accordingly, we utilize a different indexing
scheme for time. Specifically, set
n ~ T- I
So, n refers to that point in time when there are n periods remaining in the problem. Since
the data are stationary, V(n, .) ~ VeT - I, 'J, s'(n) ~ s'(T - t), etc., have the same val
ues for different T, provided that t is shifted accordingly to keep n constant. Equations
(9.4.3) to (9.4.5) now become
There is another functional equation that plays a key role in the infinite-horizon dy
namic program:
The pair (9.4.17) and (9.4.18) is called the oplimality equation. This is just (9.4.15) and
(9.4.16) with the time index n removed. That difference is important, however, for the
same functions H and V appear on both sides; these are unknowns to be solved for. With
out further analysis, we do not know whether a solution exists, nor, if so, whether it is
umque.
We plan to apply (but not prove) a general theorem about dynamic programs with
stationary data. Stated in our terms, the result is this:
Chapter 9 Time-Varying, Stochastic Demand: Policy Optimi~ation 385
in t, and the demands d(t) be limited similarly.) Then, one can show that a base-stock po~'
is optimal, and Theorem 9.4.2 continues to relate the myopic and optimal policies.
In practice, it is useful to combine the finite- and infinite-horizon models, to avoid
misleading horizon effects: It is often possible to specify the data with some confidence
for a relatively short horizon T, but not for time points further into the future. However.
the system is expected to continue operating well past T. Setting VeT, x) = -c(T)x may
thus distort the results.
Here is an alternative approach for the case 'Y < 1: Treat the data after T as station
ary, estimate them as well as possible, and solve an infinite horizon model with the data;
specifically, compute the optimal cost V'(x). Then, set VeT, x) = V'(x), and solve the
finite-horizon modeL
This same idea can be used with all the more elaborate models below.
°
Imagine that, if d(t) > yet), the lost sales d(t) - yet) occur andx(t + I) is set to just at
time t + I. Reinterpret b(t + I) as the unit cost oflost sales. So, the one-period cost func
tion remains C(t, y) as defined earlier. (Usually, bet + 1) includes the unit sales price, so
C includes revenue lost to shortages. This specification is consistent with the model of
Section 9.4.5.2 below, which treats revenue explicitly.)
Also, assume that bet + 1) + h(t + 1) "" '(c(t + 1). This condition is nearly always
true in practice; the sales price normally exceeds the unit cost. [n theory, moreover, it
should be true: Think of bet + I) as the unit cost charged by a special, emergency source,
which delivers instantaneously at time t + 1, in time to satisfy those prior-period de
mands [yet) - d(t)]- that would be lost otllerwise. Given that possibility, why not ac
quire extra emergency stock, hold it in inventory at cost h(t + 1), and use it at the end of
period t + I? That transaction is equivalent to a regular order, and if it is cheaper, i.e., if
bet + I) + h(t + I) < '(c(t + I), tllen we should always use it instead of the regular sup
ply channel. In that case, we may as well replace e(t + I) by [bet + l)+h(t + 1)]/'(.
The dynamic-programming formulation is almost identical to (9.4.3) to (9.4.5), but
(9.4.4) now becomes
H(t,y) = c(t»)' + C(t,)') + '(E[V{t + 1, [y - d(tW}]
It is convenient to work directly with the myopic formulation analogous to (9.4.6) to
(9.4.8). Redefine
C+ (t, y) ~ e(t»)' - '(c(t + I )E[[y - d( t)]+] + C(t, y)
'1I"IIIIII,il:I~I!I!~~
lDOreover, it
When this is compared to (9.4.11), it appears that the cost ratio is larger than in the back
""'Y source, orders case, so s* is smaller. But, the cost factor b has different meanings in the two mod
r-periOO de
els; it is likely to be bigger here, for a lost sale is worse than a delayed one. The larger b
"by not ac reduces the ratio and increases s*.
~theendof
aper, i.e., if
>:gular sup
9.4.7 Profit Maximization
71)Yy· The models above focus on operational costs; revenues are outside their scope. In many
(9.4.5), but situations, however, the primary rationale for inventory is to support the exchange of
goods for money. Fortunately, it is possible to extend the models to include revenues ex
plicitly. The overall objective now broadens from cost to profit.
Suppose that demands are filled and revenues received just at the end of each pe
} (9.4.6) to riod. Let
pet) ~ unit revenue (sales price) at the end of period t - I
These are fixed, nonnegative constants.
~
"I~,
""~ "
"~
Consider the lost-sales case first. The nnrnber of units sold at the end of period t is
precisely
min (y(t), d(t)} ~ d(t) - [d(t) - y(t)]+
Viewed from step (2) of time t, the expected revenue is given by R(t, yet)), where
R(t, y) = pet + l){E[d(t)] - E[[d(t) - Jr]}
The relevant costs are the same as those in the earlier models. We can combine them
with R to set up a model to maximize profit. Let us equivalently minimize the negative
ofprofit, i.e., cost minus revenue. (This is a peculiar objective, but it is perfectly valid,
and it will allow us to use the methods developed earlier.) Define the one-period profit
function pet, y) ~ R(t, y) - C(t, y). To formulate the model, we need only replace
Cabove by -P ~ C - R.
Observe thatR(t, y) is a concave function ofy (i.e., -R is convex). Therefore, -Plt, y)
is convex in y (y,Ie actually need only - p+ ~ C+ - R to be convex, and this is true under
the reasonable conditionp(t) + b(t) + h(t) 2> -yc(t).) The analysis of the original (cost
minimization) model above thus remains valid; in particular, a base-stockpolicy is optimal.
Notice that subtracting R from C leaves the form of the function unchanged. There
is a constant term -pet + I)E[d(t)], and the penalty-cost coefficient bet + 1) is incre
meuted by pet + 1). This makes sense: Wben a sale is lost, we lose the corresponding
revenue.
In the stationary infinite-horizon case, s* satisfies
F~(y) ~ (c+ + h)
(p+b+h--yc)
Now, suppose nnfilled demands are backlogged. This case is harder, conceptually
as well as technically. Suppose the price p(t) changes over time. If a demand occurs in
period t - 2, but can only be filled later, at the end of period t - I, which price applies,
pet - I) or p(t)? IfpU) > p(t - I), it seems unfair to charge the higher price. Fair or not,
if we do charge p(t), why should customers be willing to wait?
Nevertheless, assume for simplicity that indeed pet) is the price for all demands and
backorders filled at time t, even those due to earlier demands. To make this assumption
more palatable, and to facilitate the analysis, we restrict the price parameters, requiring that
pet) - '(pet + I) + bet) + h(t) 2> 0 (9.4.19)
This condition holds when pet) is constant or decreasing. It allows price increases, but
only modest ones.
Ifx(t) 2> 0, thenR(t, y) above still measures the expected revenue. Ifx(t) < 0, how
ever, we can also collect revenue by filling outstanding backorders. Imagine that demand
occurs before the order arrives. So, we could sell as many as d(t) - x(t) units. Of these,
we cannot fill more than Z(I) ~ Ylt) - x(t). Thus, the number of units sold when
x(t) < 0 is
min {yet) - x(t), d(t) - x(t)} = d(t) - [d(t) - y(t)]+ - x(t)
In general, for any x(t), the nmnber of 'mits sold is
d(t) - [d(t) - y(t)r + [x(tW = d(t) - [x(t + IW + [x(tW
""",,,,,,,,,,,,,,,,,"11" 1,1 , 1.1Ii.
(9A,19J By (9.4.19) this is convex in y. Thus, we can invoke Theorem 9.4.1: Here too, a base
stock policy is optimal.
x:reases. but In the stationary infinite-horizon case, s* satisfies
tl < 0.00-.'· o c+ +h
that demand Fd(y)=p++b+h
its, Of these. c+ +h
s sold when
(p+ - c+ + b) + (c+ + h)
For y ~ I the cost ratio again reduces to I - w. Fot'( < I, the effect ofincluding revenue
in the model is to increment the backorder cost by p + . Alternatively, since the factors hand
b represent physical costs only, we can interpret c + as the financing cost for inventory and
p + - C + as the corresponding financing cost for backorders,
390 Foundations oflnvenrory Management
This entire discussion presumes that the prices p(l) are given constants. Problem 9.8
explores a scenario where they are decision variables. (The problem explores just that
one scenario, however. The general case is difficult.)
There are Iwo sources of uncertainty about x(t + I), the demand and the yield.
What happens if we follow a base-stock policy? Consider a stationary setting for sim
plicity, with base-stock level s(l) ~ s, starting with x(O) :0; s. We order z(O) = s - x(O), so
x(I) = s - [~(z(O)) + d(O)]. The defects make x(l) less than it would be otherwise. We do
make up the difference at time I, orderingz(J) = s - x(l) ~ ~[z(O)] + d(O). But then x(2)
= s - [~(z(l)) + d( I)], which is likely to be even smaller, because of the large z(l). This
pattern continues: At each time I, z(l) replaces period (I - I)'s defects as well as demand.
Even so, period I'S defects lead to a lower x(1 + I) than we would otherwise expect. Know
ing this, we most likely want to choose s higher than in the no-defect model.
How might we do better? Instead of ordering just to correct for past defects, we
might aim to anticipate defects in the current period.
Consider the deterministic-yield case: There is a constant yield rate 1;, or a constant
defecI rale 8, as in Section 3.6. The aetual yield S(z) is]ust the function (;z. The solution
here is straightforward: Solve the no-defect model. A base-stock policy is optimal there;
the optimal order size is [s* - X(I)]+. In the real system, inflale this quantity to antici
pate defects; that is, set z(l) ~ 0 if x(l) '" s* and z(t) ~ [s* - x(1)]/(; otherwise, so that
x(l) + S(z(l)) ~ max {s*, x(I)}. The actual net inventories x(l) and the costs are precisely
the same as in the no-defect model. (This idea works even with nonstationary data.)
With random yields, of course, we can no longer anticipate perfectly. If we set
z(l) > 0 so that x(t) + z(l) > s*, and the yield turns out higher than expected, we may find
x(t + 1) > s*. This is an unwelcome outcome. It is not clear just how to anticipate
properly.
Here is the plan: First, we consider a special case. A dynamic-programming analy
sis shows that, indeed, we should modify the base-stock rule to anticipate defects: We
can decide when to order by comparing x(l) to a certain base-stock level s*(t). When x(l)
< s*(I), the actual amount to order is more than s*(I) - x(I). Unfortunately, the analysis
cannot tell us just how much more; that requires a detailed calculation in each period.
Chapter 9 Time-Varying, Stochastic Demand: Policy Optimization 391
ols. Problem 9.8 Then, we explore a certain heuristic policy, called a linear inflation rule, inspired by the
<plores just that exact analysis but simpler to implement than the true optimal policy. We perform an ap
proximate steady-state performance evaluation and, on the basis of that analysis, select
the best such policy.
tits. If we order The collection {2(z) : z ;", 0 j is a family of random variables, distinguished by the pa
ft. z), a random rameter z. How shall we model it? There are many possibilities. Let us focus on one sim
I is the quantity ple case, that ofproportional yield: The yield rate ~ ~ ~(t) in each period is a random
we pay for the variable between 0 and I, independentofz. and 2(z) ~ ~z. Given the yield rate, the yield
he actual order is linear in the order size. (The yield rates in different periods are i.i.d.) Equivalently,
nts. i.e., the de there is a random defect rate 0 = oCt), independent ofz(t).
III time, but for This proportional-yield model is very special. Observe that E[2(z)] ~ E[~]z and
2-). V[2(z)] = V[~]Z2, the mean is linear in the order quantity, but the variance is quadratic.
og on: The dy This is a worst case. Imagine dividing the order into several equal portions. Then. each
portion has the same yield; their yields are perfectly correlated. In effect, the order is
(9.4.20) thoroughly mixed, so that every portion is defective to the same degree. This says a great
yield. deal about the physical process generating the defects. Defects occur because of sys
letting for sim tematic causes only, which influence the entire order uniformly. Some physical processes
= s - x(O), so do work this way, at least approximately, but others do not. There is no room here for in
>erWise. We do dependence, partial or total, among the portions. (If the portions were independent,
,. But thenx(2) V[2(o)] would be linear in z, like the mean.) Still, we focus on this model because of its
mgez(l). This simplicity.
el! as demand. The family of yields {2(z) : z ;", OJ has four appealing and important properties:
expect. Know 1. 2(z) is bounded. specifically, 0 s; 2(z) s; z. We never receive more than we
t ask for.
st defects, we
2. 2(z) is smooth. That is, for any scalar function/('), E[/(2(z»] preserves the
smoothness properties of/(·). For instance, if/O is continuously
, or a constant
differentiable, so is E[/(2(:))]. This condition is natural in a world of
: The solution
continuous quantities.
optimal there;
3. 2(z) is stochastically increasing. That is, if/(·) is nondecreasing, E[/(2(z»] is
ltity to antici
a nondecreasing function of z. Also, z - 2(:) = lI.(z) is stochastically
I"\\oise, so that
j are precisely
increasing. If we order more, we are likely to receive more nondefective and
defective goods.
I3l)' data.)
dy. If we set 4. 2(z) is stochastically convex. That is, for any convex/('), E[/(2(z»] is convex
.wemayfmd in z. Moreover, if/ex. y) is jointly convex, then E[/(x. 2(z))] is jointly convex
to anticipate in (x, z).
Now let us analyze the inventory system: Let Crt. y) be the same function defined
mning analy before, the one-period cost in the no-defect model, and C+ (t, y) the myopic cost func
, defects: We tion. Define
t). Whenx(t)
•the analysis Crt, x, z) ~ E[C(t, x + 2(z»]
each period. C+(t, x, z) ~c E[C+(t, x + 2(z»]
392 Foundations ofIrwentory Management
By (9.4.20) the actual one-period cost is now Cit, x(t), zit»~. The stochastic-convexity
property ensures that C(t, x, z) and C+ (t, X. z) are jointly convex in (x, z) for each t.
Let us proceed directly to the myopic formulation of the dynamic program:
V+(1; x) = 0
H(f. x. z) = C+(t, x, z) + 'YE[f··+(t + 1, x + 8(z) - d(t»]
V+(f. x) = min {H(f, x, z) : ="" OJ
Notice that H can be written as H(t, x, z) = + 8(z»], where
E[H(t, x
H(t, y) = C+(t, y) + 'YE[f·+(t + l,y -d(t»]
In the one-period model we ean drop the time index. The problem is to choose
z"" 0 to minimize H(x, z) ~ C+(x, z), where C+(x, z) = E[ C+(x + 8(z»]. What does
the optimal policy look like? Let s' mininrize C+ (y). so s' is the optimal base-stock
level for the no-defects problem. Let z' = z'(x) denote the true optimal order size, and
y'(x) ~ x + z'(x).
THEOREM 9.4.8. In the one-period problem
(a) z'(x) > 0 if and only if x < so, and in generaly'(x) "" sO.
(b) For x"" so, y'(x) is nonincreasing in x.
(e) The optimal cost V+(x) is convex inx.
Any rule with properties (a) and (b) we call an inflated base-stoekpoliey. For x above s',
order nothing. When x drops below s*, order enough to bring y = y*(x) above s*, to an
ticipate defects. Moreover, as x decreases further, y*(x) increases further above s*.
PROOF,
(a) First, suppose x "" s" Fix z > O. For every realization of 1;, 8(z) ~ I;z "" 0, so
C+(x + 8(z» "" C+(x). Consequently, C+(x, z) = E[C+(x + 8(z»] "" C+(x) =
C+(x, 0), so z'(x) = O. Conversely, suppose x < so. Fix z < s' - x. Over the range
x"" y "" so, C+(y) is nonincreasing, so for any realization of 1;, C+(x + 8(z» ~
C+(x + I;z) "" C+(x + I;(s' - x)). Therefore, C+(x, z) = E[C+(x + 8(z»] ""
E[C+(x + 8(s' - x»] ~ C+(x, s' - x), soz'(x) "" s' - x.
(b) Choose x] < x z ::; s*, and Iet': J = =*(x 1) andz2 = z*(x 2 ). Setz' = X 2 - Xl + Z20 and
consider z' as a potential value of.: t We have
>
= E[(1z)c+(X + B(z»]
= E[£C+'(x + £=)]
Since C+1(o) is nondecreasing, we have for every realization of ~
£C+'(x) + i;z') ~ £C+'(x, + £Z2 - 0(=' - =,»
"" £C+'( x, + £Z2)
'11:11,:: !I
We call this a linear-inflation mle. Evidently, it is one particular inflated base-stock pol
icy. It embodies the main ideas of the theoretical analysis in a simple, unambiguous way_
An ordinary base-stock policy is a linear-inflation rule with ~ ~ I. The optimal policy
for the deterministic-yield model is too; it uses ~ ~ ~.
How should we pick ~ and the S(I)? Focus on the infinite-horizon model with sta
tionary data and constant s(t) = s. Define
u(l) ~ s - x(l)
,I'I!.·""I
Our goal is to approximate thc mean and variance of U(f) in equilibrium. The steady-state
random variable u = s -- x plays the same role as d in a no-defects model, and we shall
approximate it with a normal distribution. Given these approximations, we can set sand
[3 optimally.
We continue to assume proportional yield. (In fact, more general yield models can
be analyzed similarly; see Problem 9.9.) For simplicity, set ~l ~ E[~] and ~2 = V[~].
The key step in the approximation is to ignore the possibility ofu(l) < 0, setting z(l)
E[~(I);(I)] ~ (~)E[U(I)]
so
13* = ~i + ~2 E[e]
> E[~]
Notice, 13* < I, so the resulting policy is a linear-inflation rule. In the deterministic
yield case 13* ~ ~I ~ ~, the correct value. Otherwise. 13* > ~I' The best inflation factor
is larger than in the deterministic case. We do anticipate defects, but only somewhat.
Oddly, when the yield is highly variable, 13* is near 1, so the rule comes close to a base
stock policy.
Substituting 13* into E[u] and V[u], we get
E[u] = (I + ~PE[d]
Section C.6.3,
V[u] = (l + ~;){(~;)E2[d] + V[d])
~l ~l
Call these parameters 11 and a 2 , and use them in the normal approximation: Set s* =
) u(t)]}
Y
v + <Tz* as usual. The optimal cost C+ (s*) is proportional to <T = V[ u]. When the yield
is really random (~2 > 0), <T is larger than the original <T = YV[d]. Yield uncertainty
hurts performance. That uncertainty is expressed here by the ratio ~2/~i = V[~]/E2[~],
the squared coefficient of variation of~.
As noted at the beginning, this is an approximate analysis. Still, the linear-inflation
rule using 13* seems to work fairly well. The even simpler rule with 13 = E[~] works well
also.
Again, we proceed in three major steps, depending on the horizon length, and then
discuss extensions.
V+(x) =
IH(s*) + k x:5 r*
(9.5.3)
IH(x) x> r*
See Figures 9.5.1 and 9.5.2.
These functions are not convex. This does not affect the solution above, of course.
Thinking ahead to the multiperiod problem, however, it is clear that we are going to have
difficulties. In the linear-order-cost model the convexity of V was essential to the analy
sis. To prepare for later developments, then, we now reanalyze the single-period prob
lem. As we shall see, the results above remain valid under a weaker condition on H than
convexity. This property, called k-convexity, is preserved by V. It will playa central role
in the subsequent analysis.
Letf(x) be a function and m a nonnegative nmnber. We say thatfis m-collvex if, for
all x and all positive nmnbers ~ and D,
M+ tiM-~-~:5~+e+m
D (9.5.4)
Iffis differentiable, the following condition is equivalent: For all x and all positive ~,
f(x) + U'(x) :5 f(x + ~) + m (9.5.4')
Chapter 9 Time-Varying, Stochastic Demand: Policy Optimi::ation 397
mlation is par
need to solve
H(s"')! + k
(95.1)
II ~ s*. How
ll:r COsL If_
to sety = 5*, H(s·)I· . ; ' ~
then, we must
, tbat value of
(952) s· y
"
on >r*(bm
>ptimalpolicy
" taIget stock
FIGURE 9.5.2
Jere now Optimal cost function (fixed-plus-linear order cost).
(953) H(x)
lie, of course.
going to have
I to the anaIy_
-period pro/>
V+(x)
on on Htban
• central role
(9.5-4)
positive ~,
(9.5.4') s' x
"
398 Foundations ~fInventory Management
Let us try to understand this property intuitively: Iffis actually convex, then (9.5.4)
holds for m = 0; O-convexity is equivalent to convexity itself. Also, the condition becomes
weaker as m grows; that is, if/is 1n,-convex and m2 > m lo then/is also mrconvex. The
expression on the left of (9.5.4) or (9.5.4') can be viewed as a linear approximation of
f(x + ~) for fixed x. When m ~ 0 andfis convex, we recover the familiar fact that the lin
ear approximation underestimates the true valuef(x + ~). When m > 0, the approximation
may overestimatef(x + ~), but the envr is bounded above by m. Figure 9.5.3 illustrates
this property. The dotted line is the linear approximation.
We state some additional facts in the following lemma; the proofis left to Problem 9.11.
LEMMA 9.5.1.
(a) Iff(x) is m-convex, then so is the shifted functionf(x + '1'), for any fixed 'I' (posi
tive or negative).
(b) If.ft is ml-convex,j2 is m2-convex, and at and 02 are positive numbers, then the
functionf= alii + f1.2h is m-convex, where m = aIm, + (l2m2
(c) IfJ is m-convex, andf(x) = E[J(x - d)], thenfis m-convex.
Now, let us return to our problem. Since H above is convex, it is certainly k-convex.
THEOREM 9.5.2. LetHbe any continuous, k-convex function in problem (9.5.1). The
optimal policy is an (r, s) policy with parameters (r', s·), where s· is the smallesty min
imizing H(y), and r· is the largest x -<; s· satisfying (9.5.2).
FIGURE 9.5.3
An m-convexfunctiofl.
........
...·······1·······
x-v x x+~
Chapter 9 Time-Varying, Stochastic Demand: Policy Optimization 399
rex, then (9.5.4) PROOF. Clearly, it is optimal not to order for x .2::: s*. Also, for r* < x < s*, we have
ldition becomes H(x) < H(s*) + k, by the definition ofr*, so again it is optimal not to order. So, we need
m2-eonvex. The only show that it is optimal to order up to s*, i.e., H(x) ;,. H(s*) + k. for all x :0; r*. But,
lpfOximation of suppose this inequality is violated at x = r* - v for some 11 > 0, and set ~ = s* - r*.
fact that the lin Then, H(r*) > H(r* - v), so
, approximation
H(r*) + HH(r*) - H(r* - v)]
9.5.3 illturtrates =----'--="--"------=-="------"" > H(r*) ~ H(r* + ~) +k
v
o Problem 9.11. This violates the k-convexity of Hat r*.
Next, we show that k-convexity is preserved by V+(x).
'lXed '!' (posi
THEOREM 9.5.3. Let Hbe any continuous, k-convex function, and define V+ in terms
ubers, then the of H, s*, and r*, according to (9.5.3). Then, V+ is continuous and k-convex.
PROOF. Continuity is obvious. For k-convexity we need to consider three cases:
Case (i). x + ~ :0; r*. Here, V+(w) is the constant H(s*) + k for all w:O; x + ~, so the
UnIy k-convex. k-convexity condition (9.5.4) is immediate.
"" (9.5.1).The Case (ii). x - v > r*. In this case f"+(w) ~ H(w), w;,. x - v, and the result follows
;ma1]est y min
from the k-convexity of H.
If V+ (x) > H(s') + k, then x > s*, and V+ (x) ~ H(x). Also, setting v' =x - s*, we
have Vi < u. Thus,
/
V+(x) - V+(x - v)
-~~ ~
11
COROLLARY 9.5.4. Define V+ as in the theorem, and set V(x) ~ V+(x) - ex. Then,
Vex) is continuous and k-convex.
PROOF. Again, continuity is obvious. Also, V is the sum of a linear function and the k
convex function V+, and so is k-convex by part (b) of Lenuna 9.5.1.
400 Foundariolls of Inventory Management
Again, these equations describe a recursive scheme to compute an optimal policy. The
form of the optimal policy is characterized in part (b) of the following theorem, the di
rect analogue ofTheorem 9.4.1:
PROOF. The proofis parallel to that ofTheorem 9.4.1: Use induction on I, starting with
1= T - 1 and working backward. Assuming that V(I + I,x) is k-convex, Lemma 9.5.I(e)
implies that EV[(I + I,y - d(I))] is too, hence so is H(I, y). For parts (b) and (e), use the
results for the one-period model.
Next, we derive relatively simple bounds on the optimal policy variables, analogous
to the myopic policy of Section 9.4. We shall work with a transformed recursion, analo
gous to (9.4.6) to (9.4.8):
V+(T, x) = 0 (9.5.8)
Again, let s + (I) denote the myopic base-stock level, the smallest value of y that globally
minimizes C+ (I, y). Also, set s ++ (I) to the value ofy > s +(I) that solves
These quantities directly reflect the data for period t. For example, if d(t) shifts up or
h period. We as down by a simple translation, then s + (t), s + + (t), and 1'+(t) all shift by the same amount.
!em 9.12 for the Also, as k grows, both s + + (t) and 1'+(t) move further away from s + (t) and each other.
D, analogous to
THEOREM 9.5.6. For each t. s*(t) :5 s ++(t) and r*(t) :5 1'+ (t).
(955) PROOF. The argument uses the following preliminary results: First, from the defini
tion and k-convexity of V+, one can show that, for all x and all ~ > 0,
(9.5.61
or H(I,y) < H(t, s*(t)) + k. So,y cannot satisfy the equation for r*(t): thus, r*(t):5 r+(t).
Ies, analogous
Second, assunle s*(I) > s +(t). For y with s +(t) < y:5 s*(t)
~amIo-
H(t, s*(I» - H(I. y) 2: C+(t, S*(I» - C+(I, y) - ~k
(95.81 >O-k~ -k
(9.5.91 while fory with r+(t) < y:5 S +(t) we can use the same argmnent as in the first case above.
(9.5.10) In sum, H(t, y) < H(t, s*(t» + kfor all y with r+(t) < Y :5 s*(t), so again r*(t) :5 r+(t).
Example 9.5.A illustrates the optimal (r, s) policy.
'tba1 globally
EXAMPLE 9.5.A
As in Example 9.4.A, we consider four systems with 'Y = 1, h(t) = 1, and bet) = 9. The d(t) have
either Poisson or geometric distributions. In every case k = 5.
Figure 9.5.4 shows the optimal policy for a pair of models with stationary e(t). The E[d(t)]
change, but not too much; they oscillate between 20 and 40. Observe, in both the Poisson- and
geometrie-demand models, the difference q"'(t) = s"'(t) - r"'(t) remains nearly eonstant over t. In
402 Foundations ofInventory Management
the Poisson-demand case. moreover, the changes in r*(t) are nearly parallel to the movements of
E[d(t)]. This behavior is consistent with the heuristic proposed in Seetion 9.2. In the geometric
demand case too, r*(t) changes in the same direction as E[d(t)], but the changes are more pro
nounced; indeed, r*(t) is nearly proportional to E[d(l)]. (Here, the standard deviation of d(t) is no
longer nearly constant; instead, it is close to E[d(t)].
Next, we examine two systems with E[d(l)] fixed at 30 but changing e(t), as in Figure 9.4.3.
The results are displayed in Figure 9.5.5. Both r*(t) and s*(t) anticipate changes in e(t), much like
the optimal base-stock level in the linear-cost case. But r*(t) changes somewhat less than s*(t), so
the difference q*(t) = s*(t) - r*(t) grows and shrinks considerably as the data change, especially
in the geometric-demand case. ill contrast to demand changes, cost changes seem to have strong
effects on all the policy variables.
FIGURE 9.5.4
Optimal (r, s) policies (stationary costs).
100
--B Eld(n]
...•... r*(f)
80
........... s*(t)
.................
g 60
.'
.'
.'
.£ ......*.
.i~."
40 .... ....
20
o 2 4 6 8
"II!llillil'IIII'IIIIII~_
120
--a-- c(t)
in Figure 9.4.3. ...+.., 1'*(1)
I c(t), much like
....... s'(t) 6
"" than s*(t), so 100
lIlIge, especially
I to have strnog
80
~ * 4 ~
01 case haoe '" 60
umerical tesls
ptimaI 40
2
20 Poisson demands
Iogue ofThe-
';:
focy < I aod 0 o
0 2 3 4 5
It is quite difficult to prove this fact, however. Nothing like the simple, direct proof
of Theorem 9.4.3 works here.
The alternative approach, using Theorem 9.4.4, can be adapted to this modeL The
argument is essentially the same as that of Section 9.4.3. Certain details, however, are
more problematic here. For instance, it is not generally true that the optimal policy vari
ables r*(n) and s*(n) are constant, as in Lemma 9.4.5(b), nor even monotonic. However,
they are bounded, so the sequence {[r*(n), s*(n)]} has a limit point. This, it turns out, is
enough. Any such limit point describes an optimal policy.
To compute an optimal (r*, s*), requires an algorithm. The situation is essentially
the same as in the continuous-time setting: In the special case where the demand d is
either 0 or I, an (r, s) policy reduces to an (r, q) policy. To determine the best such pol
icy, use Algorithm Optimize_rq of Section 6.5.4, with minor adjustments to account
for discrete time. Otherwise, use the similar method mentioned in Section 6.6, simi
larly adjusted.
Alternatively, employ one ofthe algorithms for general dynamic programs. The sim
plest one, conceptually, is called successive approximations: Perform the finite-horizon
recursion (9.4.14) to (9.4.16) (modified to include the fixed cost k, of course) for in
creasingly large n. until an appropriate termination condition is met. The temlinal pol
icy [r*(n), s*(n)] estimates (r*, SO).
The upper bounds of Theorem 9.5.6 now become stationary; denote them by s ++
and r+. Thus,s*(n)::; s++ andr*(n)::; r+ foralln, andsos*::; s++ andr*::; r+. These
bounds directly reflect the data of the problem, as discussed above.
~
404 Foundations q{ Inventory Management
q ~ 1.30vo494(k/b)o.506[1 +(<Tlv)2]O.116
0= [(ql<T)(h/b)]1I2
r = 0.973v + (0.183/0 + 1.063 - 2.1920)<T
s=r+q
These formulas were obtained by solving a wide range of models, choosing a plausible
functional form, and fitting the parameters with regression analysis. (The approximation
works well provided qlv is not too small, say above 1.5. Otherwise, use the following re
vision: Compute the optimal base-stock level, s +. If s + < s, then replace s by s + and
rby s+ - q.)
Notice that the expression for q closely resembles the EOQ formula. The initial con
stant 1.30 is not far from {i, and the exponents of v and klh are close to X. The last term
is an inflation factor to account for demand variance. In broad terms, therefore, the for
mula is consistent with the bounds and numerical example of Section 6.5.3. This re
semblance is remarkable, for the context there differs in several ways (e.g., continuous
time). In the formula for r, the coefficient of 11 is nearly 1; the reorder point moves in tan
dem with the mean demand.
9.5.4 Extensions
What happens if we extend the model as in Sections 9.4.6 to 9.4.8?
In the lost-sales case it is still true that an (r, s) policy is optimal. The r*(t) and s*(t)
should be non-negative, of course, and we impose that as a constraint. In the one-period
problem, for example, it is always true that s* 2: 0, but ifno x 2: 0 satisfies (9.5.2), set
r* = O.
With the profit objective of Section 9.4.7, everything works fine, as above.
For the imperfect-qnality model of Section 9.4.8 we do not know what the optimal
policy looks like; as yet there are no results comparable to Theorem 9.4.8.
An analogue of the linear-inflation heuristic seems to work well for the stationary
infinite-horizon model: Choose (r, s) and a constant [3 somehow. Then, if x .::::; r, order
z ~ (s - x)ll3. We do not know how to extend the equilibrium analysis of Section 9.4.8.3
to help select the parameters. It seems quite effective, however, simply to treat the yield as
deterministic. That is, choose the optimal (r *, s *) in the no-defects model, and 13 = E[~].
9.6 Leadtimes
9.6.1 Constallt Leadtime
Suppose there is a constant, positive-integer leadtime 1. An order placed at t arrives by
the end of period t + L - 1, in time to be counted in the net inventory at time t + L. As
Chapter 9 Time-Varying, Stochastic Demand: Policy Optimization 405
Ie investigated we shall see, this leadtime aIters some ofthe data of the model, but not its overall form.
alion: Assume Al/ the qualitative resuIts of the last two sections remain valid.
In these terms the original model has leadtimeL ~ 1. As mentioned in Section 9.3.1,
some writers define "Ieadtime" a bit differently, so that the original model's leadtime is
0, and the leadtime here is L - I. This is a difference in labeling, not substance.
We slightly alter the timing conventions of Section 9.3: The cost ('(t, ') now occurs
precisely at time t, not before. Also, assume that we pay the order cost.. not when we
place an order, but rather when we receive it, i.e., time L later. So, the effective order cost
is y-[k(I)8(z(t» + C(t)Z(I)]. (These changes make the results come out a bit neater. They
matter, of course, only when"y < 1. In the original model with L ~ I the only effect is
ng a plausible
to multiply every cost factor by"y, so the solution remains unchanged.) By the way, pur
rpproximatioo
chase costs commonly are paid on receipt of goods, not when the order is initiated. There
, following Ie
are exceptions, e.g.. some international transactions. It is easy to revise the results for
esbys+and
those cases.
The leadtime imposes a lag between our actions and their effects. There is nothing
he initial con
.The last term
we can do to influence the net inventory, and hence the holding-penalty costs, at times
1hrough L - I. So, it is reasonable to omit those costs from the problem. The cost at time
°
efore, the for
L, moreover, can be affected only by the initial order z(O); later orders, including z(L) it
•.5.3. This Ie
self: take effect only after 1. There is nothing lost, therefore, if we imagine that time L's
~~ continuous
inventory-backorder cost occurs at time O.
moves in tan
In general, at time t, the current order Z(I) directly affects the net inventory at time
t + L. and no later order does so. It makes sense, then, to reassign the corresponding in
ventory-backorder cost to point t. (This is an intuitive rationale; Section 9.6.2 presents a
rigorous one.)
To implement this new accounting scheme, redefine the state and control variables:
Thus, the expected inventory-backorder cost at step (2) oftime tis C(t, yet)), where we I
4
redefine
CCt, y) = -fEr e(t + L, y - D[t, t + L»)]
For finite T this approach makes sense for t "" T - L, but what about times near the
horizon? For simplicity, imagine that the horizon extends to time T + L - I; demands
continue to occur and inventory-backorder costs continue to accrue until then. So, the
same approach works for all t < T Again, the last order is placed at time T - I, and the
cost under the revised accounting scheme, excluding the holding-penalty costs before
time t + L, starting at time t with inventory position x. With the new definitions of Vand
C. ti,e functional equations (9.4.3) to (9.4.5) or (9.5.5) to (9.5.7) remain valid as stated,
We now have a model o{precisely the sameform as the original. The symbols have
new meanings, but that is the only difference. The new C(t, y) is still convex iny. Thus,
the main results of Section 9.4 and Section 9.5 still hold: In the linear-order-eost case,
a base-stock policy is optimal; with afixed order cost, an (r, s) policy is optimal. The
myopic policy for the linear-cost model optimizes the function C+(t, y), constructed
Of course, the actual policies reflect the new data. The one-period cost C(t, y) now
involves an expectation over demand in L periods, not just one. For instance, in the
linear-order-cost model, s + (t) solves an equation like (9.4.10), but the ccdf of the lead
time demand D[t, t + L) plays the role of F~(y). Thus, the myopic policy is the discrete
time analogue ofthe heuristic approach of Section 9.2.2. For the reasons given there and
In the infinite-horizon problem with stationary data and k = 0, the myopic policy is
optimal. and s* solves
(e+ + h)
FJ,(y) = (b + h)
as in (9.4.13), where D = D[O, L). Interestingly, only the left-hand side depends on L,
not the cost ratio. When 'Y = 1, the cost ratio again becomes 1 - w. The solution here is
essentially the same as that of equation (6.5.3).
The optimal cost retains the same form as before: For 'Y = 1 J
C+ (s*)
.x :::; s*
V+*(x) = I - "I
[
C+(x) + -yE[V+*(x - d)] x> s*
Chapter 9 Time- Valying, Stochastic Demand: Policy Optimization 407
t. yet»~, where we Under the normal approx@ation C+ (s*) again transparently reflects the system param
eters: Letting v = E[D] and a 2 ~ V[D],
C+(s*) ~ YlycE[d] + c+v + a(b + h)<I>(z*)]
lilt times near the (The first two terms reduce to cv for L ~ I, because v ~ E[d], but not for L > 1.)
L - I; demands
,til then. So, the
IeT-I,andthe 9.6.2 Constant Leadtime: Rigorous Argument
'+L -1.AIso, We now develop a rigorous justification for the cost-assignment scheme and the simple
dynamic-programming fonnulation above.
:ans the optimal Suppose we find ourselves at time t. What is the current state of the system? That
LIly costs before is, what useful information about the past and the present is available? We need to track
nitions of V and outstanding orders, in addition to the current net inventory. Specifically, for L > 1, let
I valid as stated,
z,(t) ~ order placed I periods ago, i.e., at time t - I. I ~ I, ... , L - I
Ie symbols have z(t) ~ [z,(t)]t~/.
ElVex in y. Thus, The state is thus the pair [x(t), z(t)]. (The symbol z(t) without a subscript continues to
ftler-cost case, mean the order at t@e t. It is not part of the vector z(t). When L ~ I, as in the original
is optimal. The model, there are no outstanding orders, and z(t) is vacuous,) The dynamics are
~,), constructed
x(t + 1) = x(l) + ZL_\(t) - d(t)
ost qt. y) now Z(I + I) = [z(I), z\(t) , ... , ZL_2(t)]
<lSlance, in the
:df of the lead In these terms the inventory position is x(l) = x(t) + I, >"z ,(t). Also, define the functions
is the discrete C'(I. y) = ·y'E[C(1 + I, y - D[I, I + I»]
:n-en there and
Notice that C"(t. x(t» ~ G(I, X(I» is the actual (not reassigned) inventory-backorder cost
"'Pic policy is at time t. and CL(t, y) is C(t. y) defined above.
Define the functions
Vet, x. z) = optimal cost from t@e t onward, starting in state x(t) = x and z(t) = z
G(t.x. z) = c"(t.x) + C'(t.x + ZL_\) + C2(t.x + ZL-l + ZL-2) + ' .. + cL-\t, x)
The latter represents all holding and penalty costs before time I + L. At time T, there are
Iqlends 011 L no further orders to place, but costs continue to accrue until time T + L - 1. These re
IIoIion bere is maining costs are given by G( T, x. z), and we set
VeT, x. z) ~ G(T, x, z) (9.6.1)
This dynamic program is very complex. We now show that it is equivalent to the
x
simpler one above with the single state variable x = + !I>OZI and optimal-cost func
tion Vet, x).
THEOREM 9.6.1.
Vet, x, z) = Crt, x, z) + Vet, x) 0-5d-5T
Moreover, the optimal order quantity Z in (9.6.2) is the same one that achieves the min
imum in Vet, x).
PROOF. We argue by induction on t: The result is true for t = T, by VeT, x) = 0 and
(9.6.1). Suppose it is true for t + 1. Using this induction hypothesis, and the fact that
Now, suppose the leadtime is a random variable. That is, an order placed at time twill
arrive at some future time t + L(t), but L(t) changes randomly over t. We allow L(t) = 0;
in this case z(t) arrives immediately, and is eounted in time t's net inventory x(t). Assume
that the L(t) arise from a discrete-time version of the exogenous, sequential supply sys
tem of Section 7.4. That is, t + L(t) -5 (t + I) + L(t + I), so orders never cross in time.
Also, we can do nothing to influence the L(t), and they are independent of the demands.
Finally, the L(t) are stationary; each has the same (marginal) distribution as a generic
leadtime variable L.
Assume too that at each time t we have no information to help us predict L(t). That
is, even conditional on what we know at t, L(t) has the same distribution as L. This can
mean one of two things: Either L(t) is independent of all past events, so there really is no
such information; or, while such information could be extracted in principle, we cannot
obtain or use it. (In case you are wondering, there are systems where L(t) is independent
of the past. Here is one: There is a sequence ofi.i.d. non-negative-integer random vari
ables A(t). At each time t, all orders that have been outstanding for time A(t) or more ar
rive immediately; that is, of the orders still outstanding, we receive those placed at
t - A(t) or earlier. Thus, L(t) ~ min {u '" t: A(u) -5 u - t}. The constant-leadtime model
is the special case where all A(t) = L.)
Let us adaptthe cost-accounting scheme above to this stochastic setting: If we knew
that L(t) ~ I, it would make sense to assign to time t the holding-penalty cost at time
t + I, by the argument for the constant-Ieadtime case. So, we assign afraction ofthis cost
1'1"'111
....
to t, the fraction Pr {L ~ l}, and we do the same for every I 2: 0. (Again, the rationale
-~ here is intuitive. The idea can be supported rigorously, but we shall not do so.)
The actual cost at each fixed time u is distributed among earlier times u - 1with these
same weights. Its cost is left unassigned with probability Pr {L > u}. This is exactly the
probability that the order at time 0 arrives after u. Thus, this scheme, like the original one,
omits those costs we cannot i.t.fluence. As for tenninal effects, assume that demands con
=---- tinue to occur and costs to accrue tmtil the end of the last leadtime, i.e., until tin,e (T - I)
+ L(T - I). (This last assmnption is not really necessary, but it simplifies the notation.)
......
1)=0"
We still pay for orders on receipt, so the effective order cost is E[·y'][k(t)o(z(t)) +
c(t)z(t)]. The expected inventory-backorder cost at time tnow becomes C(t, yet)), where
qt, y) ~ E[ -IC(t + L, Y - D[t, t + L))]
The expectation is over L as well as the demands. This function qt, y) is convex in y, for
it is a positive-weighted average of convex functions. With this new C. the dynamic
programming fonnulation is precisely the same as before. So, again, all the previous re
sults on optimal and myopic policies remain valid.
.,..
'/.41) = ~
and the optimal base-stock level s* satisfies
Fg(y) = I - w
The random variable D is again D[O, L), the demand over L periods, but now L is ran
~~
dom. This is precisely analogous to the leadtime-demand random variable of Section 7.4.
""'I!'S'"
• • 1iIIIIII:..
The optimal cost cE[d] + C(s*) now reflects the leadtime's variance as well as its mean,
as in Chapter 7. (When'Y < 1, s* again solves a simple equation, but its fonn is a bit dif
I
ferent from before, because of the random discountterm"yL See Problem 9.13.)
.~
_iIF
-->
lem 9.4.) Of course, s*(t) is different from, typically larger than, that of the original un
capacitated model.
This result remains valid for several extensions of the model: We can have limited
IiaI::rd • capacity and a constant leadtime. Or, z+(t) can be random, realized at the same time as
"'..-.:I d(t); here, if we order z(t), we actually receive ntin {z(t), z+(t)}. In the fixed-order-cost
..,..,..
model, however, it is not always true that an (r, s) policy is optimal.
Also, a base-stock policy is optimal for the infinite-horizon model with stationary data.
t . timt
(Certain ntild regularity conditions are needed; for instance, in the case"y = 1, we must as
this CUit
sume that E[d] < =+.) In general, exact policy evaluation and optimization are difficult.
~ ~
410 FOllndations afInventory Management
There is a good approximation, however: Assume "y = I. Let u(t) = s - x(t), and let
u be the corresponding steady-state random variable. This quantity plays the same role as
D in the unlimited-capacity model, namely, C(y) = E[C(y - u)], and s* solves F~(y) =
I - w. (In other words, u(t) is the occupancy ofa related discrete-time queue.) The tail of
u, it turns out, is approximately exponential. That is, there are certain constants K and e,
such that F~(y) = Ke -."1' for large y. (There are tractable numerical methods to estimate
these constants; see Glasserman [1997].) Consequently,
s* = O[ln (K) - In (I - w)]
E[u]
C(s*) = Oh[(l - -0-) + In (K) - In (I - w)]
This approximation is very accurate when w is near 1. This is the discrete-time analogue
of the exponential approximation of Section 7.3.11.
, - x(t), and let tially the same as constructing a discrete-time Markov chain to represent a continuous-time
he same role as one, as discussed in Section C.5 ofAppendix C. The continuous-time model is one instance
;elves F~(y) = of a general class called semi-Markov decision processes or continuous-time dynamic pro
ue.) The tail of grams. The embedding technique is a standard one in that arena.) This diserete-time model
ilants K and e, has the same form as those above, so a base-stock policy is optimal.
00s to estimate The second and third approaches can be applied to other models above and below.
(With rare exceptions, however, the first cannot.) With a fixed cost k. for instance, an (r. s)
policy is optimal. In the Poisson-demand case, of course, this reduces to an (r. q) policy.
In sum, the policies we chose to evaluate in Chapter 6 really are optimal, at least
for Poisson and compound-Poisson demand. Therefore, the optimization methods ofSec
tion 6.5 find the best policy overall.
Return to the discrete-time setting for the rest of the chapter.
rime analogue
Ie. Thus, the x(t + I + I) ~ [x(t + I) + ZL-l-/ - d(t + 1)]+ O-s;I<L (9.6.4)
= lP(t) = }'. x
where X(t) = for I = 0 and Zo = z for I = L - I. (9.6.4) recursively defines the random
lJlational ad variables .x(t + I). The expected cost at time t + L as viewed from time t, then, is
...
--.-::iiI--'
""'11-'
This recursion looks much like (9.6.2). Unfortunately, because the dynamics and hence
the one-period cost C are so complex, we cannot reduce the state to the single variable x.
In principle, once again, the recursion prescribes a computational procedure to find
the optimal policy. However, the state variable is now an L-vector. [t is a sad fact that, in
general, dynamic programs with high-dimensional states are VelY hard, indeed virtually
impossible, to solve. (This pomt is vividly expressed by the phrase, the curse ofdimen
sionality. If we discretize and truncate the state vector, using n values for each compo
nent, say, then the total number of states is n L . The computational effort required is at
least proportional to this number, which grows exponentially in L.)
Thus, in practice, there is no way to find the true optimal policy. Moreover, little is
known about its structure; it is probably very complex and hard to implement. If we can't
optimize, are there heuristic approaches that work well? Fortunately, the answer is yes.
For simplicity, focus on the stationary, infinite-horizon model.
The key idea is to adapt the myopic approach to this system. Let t = 0 indicate the
current time, so (9.6.4) becomes
ics and hence set z = 0, and otherwise find z > 0 such that
g1e variable x.
mUTe to find c+ +h
Pr {i(L) = O}= b + h - -yc (9.6.6)
od fact that, in
ked virtually
This relation generalizes the exact optimality condition for L = 1 in Section 9.4.6.
rse ofdim en
It turns out that the myopic policy leads the system to an "equilibrium" where
each compo
(9.6.6) holds in every period. That is, perhaps after some initial periods with no orders,
required is at
the system arrives at a state where (9.6.6) holds, and from then on (9.6.6) continues to
hold. (See Problem 9.14.) Thus, the myopic policy aims to maintain a constant stockout
~over, little is
probability, just as a base-stock policy does in a backlog system.
IlL Ifwe can't
Let s~ solve F~(y) ~ (e + + h)/(b + h - -ye). The myopic policy is related but not
lIlSWer is yes.
identical to the base-stock policy with base-stock level s~. Letting x be the inventory po
) indicate the sition and y ~ x + z, it is clear from (9.6.5) that i(L) :2: Y - D, so
Pr {i(L) = O} oS Pr {y - D oS O} = ~(y)
(9.6.5) Like the base-stock policy, the myopic policy orders nothing when x :2: s~, and if
x < s~, the new y never exceeds s~. However, y may be less than s~. Thus, the myopic
policy aims to keep the inventory position at or below s;.
The two policies difler mainly in how they respond to unusually large or small de
mands. Recall, in a backlog system, a stationary base-stock policy can be viewed as a
demand-replacement rule. The same is true in a lost-sales system, except that lost sales
are not counted; the base-stock policy replaces filled demands. When the prior demand
is large. the myopic policy typically orders less than the base-stock policy. That demand
leS the efleet will likely lead to lost sales, now or in the near future, and once lost they are gone for
is subtle; we good; a large order arriving at t + L cannot help. Conversely, when the prior demand is
'»YJpic policy very small, the base-stock policy orders that same small amount, but the myopic policy
",urks quite typically orders more. In sum, the sequence of orders generated by the myopic policy
tends to be smoother than the filled demands themselves. (This property makes inhritive
Jog case. At sense. It is probably true of the optimal policy too, but no one knows for sure.)
Jine the best There are simpler heuristics than the myopic policy. The simplest is a base-stock pol
ual:eL, we icy. Unfortunately, except for special cases, we do not know how to evaluate such poli
n.i(L). StiR cies. (We have seen one case, the continuous-time model of Section 7.2.3', where base
stock policies ean be evaluated and the best one selected readily.) Also, it is possible to
~Tite L in approximate the myopic policy using simpler calculations. This approach is intermediate
in complexity between a base-stock policy and the myopic policy. See Problem 9.15.
)1 What about the overall performance of lost-sales systems? In the backorder case we
understand the eflects ofthe leadtime, the demand variance, and other basic parameters.
Unfortunately, to our knowledge, there are no comparable results for the lost-sales case.
:::. So. ooce It is a fair guess that these same parameters affect performance in similar ways, but this
• fairly IOU is just a guess.
, 0. Eqtma- This entire discussion presumes a constant leadtime. Certain limited-capacity sup
ply systems are actually easier to comprehend. Consider the continuous-time model of
Section 7.3.4 with Poisson demand, a single processor, no scale economies, and lost
sales. Assume that we have direct control over the processor, as in Section 7.3.10.2, so
we decide at each instant whether to tum it on or off. It turns out that a simple base-stock
policy is optimal: For some s* keep the processor on when inventory is less than s* and
off otherwise. The methods of Section 7.3.4 can thus be used to evaluate alternative poli~
cies and to find the best.
~ d(t +L
min(X(t+L-1)+z,d(t+L-I)}
- I) - [d(t + L - 1) - x(t + L - I) - z]+
9.7 Worl~
"I
demand-driving factors that change systematically over time, such as the weather over
several seasons.
Also, many standard forecasting techniques embody a world-driven demand model,
explicitly or implicitly. The most familiar techniques work in discrete time. For exam
ple, the popular exponential-smoothing method essentially views the world as scalar
variable W( t), representing the current demand forecast itself or some simple transfor
mation of it, which evolves according to a Markov process. More elaborate techniques,
designated by the acronym ARMA, model Wet) as a vector. See Section C.6.4 of Ap
pendix C.
Here is a related model: Demand has a stationary distribution of some particular,
known form, but with an unknown parameter. For example, we may know that demand
has a Poisson distribution, but not its actual mean. Here, Wet) represents our current es
timate of the parameter (more precisely, a sufficient statistic for it). The dynamics ofW
include some mechanism for updating the estimate (usually Bayes' rule) to incorporate
each additional demand observation.
As we shall see, the qualitative results of earlier sections remain valid. For instance,
in the linear-order-cost case, a base-stock policy is optimal. However,just as the optimal
base-stock level depends on t in the original model, so here it reflects all relevant infor
mation about the future, namely, the current world state 11-' as well as t.
The expectation in (9.7.2) is over the demand d(t) Gndthe next state Wet + I) = w', both
conditional on W(I) ~ w. Again, H(I, w, y) measures all the relevant costs, present and
"""I"il,"IIII1IIIIIIIIII~
weather over future, if we choose y(t) = y at step (2) of time I. These equations describe a recursive
scheme to compute an optimal policy.
=andmodel, Virtually the same argument as in Theorem 9.4.1 demonstrates that a (state
Ie. For exam dependent) base-stock policy is optimal: In the induction, assume that V(t + I, w. x) is
orld as scalar convex in x for each w. Write the expectation in (9.7.2) in two steps, as
.ple transfor
Ew,{Ed(O[V(t + I, w', y - d(t)) I W(t) = W, W(t + I) == w'J)
!e techniques,
C.6.4 of Ap The inner expectation is convex in y, as usual. The outer expectation is just a weighted
average of convex functions, so it too is convex in y. Thus, the entire expectation pre
ne particular, serves the convexity of V(t + I, w, x), and so H(t, w, y) is convex in v. Consequently, a
. that demand base-stock policy is optimal. The optimal base-stock level s'(t, w) is the smallest value
t.rr current es ofy that minimizes H(t, w, y). Because H depends on w as well as t. so does s* Also,
1laJl1icsofW V(t, w, x) is convex in x for all w, completing the induction.
[) incorporate Next, consider the infinite-horizon problem with stationary data. A base-stock pol
icy is again optimal. The optimal base-stock level is s' = s't w), which depends on the
For instance~ current state l-v, but is stationary in time. (The proof is similar to the second argument for
s the optimal Theorem 9.4.3, the harder one involving Theorem 9.4.4.)
'levant infor In the fixed-order-cost model (k > 0), an (r, s) policy is optimal. With finite hori
zon Tthe optimal policy variables are r'(t, w) and s'(t, w). Both depend on w as well as
t. The proof follows closely that of Theorem 9.5.5. For infinite T and stationary data,
r' = r'(w) and s' = s'(w).
By the way, all these results remain valid if we replace c(t) by c(t, w) in the fOlIDU
:non 9.6, the lation above. Here, the order cost depends on the current world state W(t) = was well
as t. (We still pay for goods on receipt, but the cost is determined when the order is
placed.) Likewise, the fixed cost can be k == k(t, w) (subject to certain restrictions, along
the lines of Problem 9.12.) Thus, we have a full theory of world-driven economics and
ime demand demand.
he effects of Also, the results remain valid when the leadtime L is stochastic, as in Sec
the form of tion 9.6.3.1. The calculation of C(t, "" y) is different, and the -0/ factors above become
o to approx E[Y'l
[model, and To summarize, the form of the optimal policy, whether a base-stock or an (1; s) pol
emedmore icy, depends on the nature of the order cost, specifically on the presence or absence of a
C( I, ,,; y) is fixed cost. Regardless of the policy fonn, the optimal values of the policy variables re
flect information about the present and future. That infonnation includes the time index
link through t, when, but only when, t is relevant. [n an infinite-horizon setting with stationary data,
." at the fol t is irrelevant, for the time points all look alike, so the policy variables are stationary, in
dependent of t. Otherwise, t is relevant; it provides infonnation about time-varying data
or horizon effects. Likewise, when there is other relevant information, expressed in the
(9.7.11
world state w, the policy variables depend on it. (Viewed in this way, the time index is
=..-] 19.7.11 one possibly relevant item ofinfonnation, but there is nothing special about it.)
To compute the optimal policy is not easy, however. As mentioned above, it is hard
(9.-:-31
to calculate C(t, w, y). Even given the one-period cost, the state (w, x) now has at least
I ==
..-'. bodJ. one additional dimension besides x, so the recursion (9.7.1) to (9.7.3) requires more
present and work than before for each t. The curse of dimensionality appears once again.
~
418 Foundations a/Inventory Management
In modeling the world-demand system (W, D), then, we confront the classic trade
off between accuracy and tractability. It is tempting to include everything we know about
the drivers of demand in the world W. When we know a lot, however, the resulting model
may be too complex to solve. We must use our best judgment and/or analytical skills to
V-l
identify the really important factors and include only those.
Things do not become substantially simpler in the infinite-horizon model. The most
tion 9.5.3, entails solving a sequence of finite-horizon models with increasingly long
horizons.
Theorem 9.4.2(a) thats + (t, w) is an upper bound ons*(I, w). Analogues of the other parts
of Theorem 9.4.2 can also be derived. See Problem 9.17. (The same idea works with
world-dependent order costs e(l, w). The function C+ involves a cost factor e+(I, w)
Moreover, the myopic approach yields a relatively simple lower bound on the opti
mal cost. Reasoning as in Section 9.4.3, V+(O, w, x)'" V_CO, w), where
The expectation in term I is over [W(I) I W(O) ~ w]. So, to compute this lower bound, we
Once again, s +(t, w) reflects the problem data in a simple way. Assume the cost fac
In the exponential-smoothing model of Problem 9.16, the demand data are also sta
v(w) takes the form j3w + dL, a linear function ofw, and <T(I, w) = <T(w) is a constant
For -y ~ I, Wet) goes to W regardless of initial conditions, so the optimal average cost is
bounded by v* ~ v_, where
..n-e <1>°1=1 =
. (1)"T.l
v. = hmT~oo T ""~l E[C + {Wet), s + (W(t))) I W(O) ~ w]
= E[C+{W, s+UV)}]
Under the normal approximation the lower bounds for -y < 1 and -y = I, respectively,
become
/flO, = ,,-j -yL[(-yC + c+L)E[d] + tJ .(b + h)<\>(z*)]
V. I - -y
aarealsoSl3
rdil. *. "., = v. ~ E[c(d I W) + tJ(W)(b + h)<\>(z*)]
., is a coustaa
~ cE[d] + tJ.(b + h)<\>(z*)
where E[d] is the overall mean demand and tJ. is the average of tJ(w).
These are discrete·time analognes of the cost approximation of Section 6.7.2. As
1IL-* •.:. emphasized there, the variation index ( j _ represents the residual variance of leadtime
IIl1l-'"
,,,
demand, that is, the conditional variance given the information embodied in the current
world-state w. This is typically less than the totalleadtime-demand variance.
Specializing to the exponential-smoothing model, O'(w) is just 0'_ itself and E[d] ~ 11
so
9.8.1 Formulation
The formulation extends that of Section 9.6.2. The following state variables describe the
system:
ij(t) ~ (local) net inventory at stage j at time t
=j1(t) = shipment in transit to stage j sent I periods ago, i.e., at time t - I, 1= 1,
... ,Lj - 1
zit) ~ (zjl(f)),
(Forj ~ I, ii(t) is just the local inventory, so ii(t) "" 0.) The decision variables are
zit) = shipment sent to stage j at time t
We call z;(t) a shipment, not an order. The distinction is not important for stage I, but it
does matter for stage 2. Under centralized control there is no place for stage-2 orders that
stage I cannot fill immediately. Only actual, feasible shipments count. Thus, we must
constrain Z2(t) to ensure that we ship from stage I no more than what is available there.
Specifically, the constraints are
I in the CUITeDI The shipment-cost rates are C/f), the local holding-cost factors are h;ef), and the
:ICe. backorder-cost rate is b(f). The total shipment cost at time f is c\ (f)z\ (f) + C2(f)Z2(f). For
'andEld) = it simplicity, assume that this cost is paid at time f, not when the shipments arrive. (This
departs from the convention of Section 9.6. If these costs actually occur later, replace
cJ (f) by the appropriately discounted value. If stage-I shipments are outside purchases
which we pay for upon arrival, for example, replace c\(t) by ·-/'C\(f). But, stage-2 ship
ments are internal transfers, and those costs may occur at any time.) The total inventory
backorder cost is given by
." can obtain h;(f)[x:(f) + L/>oz21(f») + h~(f)[.<;(f»)+ + b(f)[x;(tW (9.8.2)
This expression is analogous to (8.3.5). Here too, shipments in transit to stage 2 are
charged at stage l's holding-cost rate.
As in Chapter 8, it is convenient to work with echelon-level quantities. Define
I\=ience. but
Xi (f) = echelon net inventory at stage j at time f
orce to Sla"oe I xi (f) ~ echelon inventory transit-position at stage j at time f
e after a lead Specifically,
The condition X;(f) 2: 0 now becomes x,(t) 2: x,(t), and the constraints can be written
'describe !be
Z;(f) 2: 0 xit) + Z2(t) "" .1'\(1)
Also, let
abies are
h/t) = echelon holding cost at stage j at time f
That is,
tIge I. but ir
2 orders thar h,(f) =h;(f) h2(f) ~ h;(f) - h;(tj
lIS. VIo-.;: DllISl
Then, the inventory-backorder cost (9.8.2) becomes
IiIabIe there.
h\(t)x\(t) + hzCMzCt) + [bet) + h,(f)][x,(fW
[This is analogous to (8.3.8).] Equivalently, the cost can be written as h\(t)x\(f) +
C2(t,x2(f», where
~
422 Foundations a/Inventory Management
zl(T - I); later shipments cannot arrive in time to meet demands. (These assumptions
are not really necessary, but they make the fonnulation neater.)
We are now prepared to fonnulate the dynamic program: Let
Vet, X" z" Xz, Zz) = optimal cost from time t onward, starting with x)t) ~ xj and
Zt(t) = Zt
This is defined for states satisfying Xl ~ Xz, where the stage-2 inventory position Xl is
the sum over (xz. zz), as above. We have
V(T+L r +L z - I,x"z"xz,zz) = 0
and for t < T + L I + L z - I, the Vsatisfy the recursion
Vet, X" z" Xz, Zz) ~min {Cr(t)ZI + Cz(t)zz + hl(t)XI + Cz(t, Xz)
+ "(E[V(t + 1, XI+ ZI,L,-I - d(t), (z" ZI"'" ,ZI,I,-Z),
Xz + zZ,L,-1 - d(I), (zz, ZZ" ... , ZZ.I,-Z))]
: Zj ~ 0, Xl + Zz :::; id (9.8.3)
(We further constrain Zz = 0 for t:2: T + L r andz l = 0 for t:2: T. Actually, provided the
cost rates are all positive, these choices are optimal.)
Evidently, this is a very complex dynamic program with a large state space. Fortu
nately. it can be simplified radically. (The dreaded curse of dimensionality can be lifted
sometimes!) We shall describe the solution and then prove that it really is optimal.
e assumptions and set Cj(t, y) to Ci(t, y) with I ~ L j. Define the functions Vj through another dynamic
program:
Vj(T, x) = 0
(t) = x and
j
Hj(t, y) ~ Cj(t)y + C,(t, y) + 'YE[V,(t + 1, Y - d(t))] (9.8.5)
With an infinite horizon and stationary data, the myopic approach yields the opti
mal policy. (This can be proven by using Theorem 9.4.4.) Specifically, set c;= (I - '!)cj ,
D, = D[O, L,), and D 2 = D[L" L I + L 2 ). Define
C2 (Y) = '!iE[C2 (y - D 2 )]
C?(y) ~ ,!c2E[d] + c?y + C2(y)
and let s! minimize C?(y) over y. Then, define
1:?(y) = C?(min {s!,y)) - C?(s!)
9.8.3 Proof
We now argue rigorously that the policy determined by the recursions above is indeed
optimal.
Let us obtain an explicit expression for VeT + L" " " " -). Recall that, in (9.8.3),
Z2 ~ 0 for I '" T + L" and ZI ~ 0 for I '" T. Also, hl(t) ~ 0 for I '" T + L I. Therefore,
the expected cost at time I = T + L I is
VeT + L" X" Z" X2, Z2) = C 2(T + L" X2, Z2)
where
C2(1, X2> Z2) = Cg(l, X2) + C~(t, X2 + z2L, ,-I) + ... + ci,-I(I, X2)
We simplifY (9.8.3) in three steps: (I) We apply the logic ofTheorem 9.6.1 to reduce
the vector (X2> Z2) to the scalar X2' This step yields a dynamic program involving functions
ofthe form V(t, x" z" x2)' (2) We show that these functions are separable, specifically, that
V(t, x" z" x2) = VI (I, x" ZI) + V2(1, x 2) for some"'. The function V2 is the one defined
above in (9.8.4). (3) We apply Theorem 9.6.1 once more to reduce t\(I, x" z,) to V,(x l ),
where V, is the function defined in (9.8.5).
Slep 1. Observe thatz2 (1) affects the stage-2 costs measured by C2 (-,·) only at time
1+ L 2. So, we can reassign C 2(t + L 2, .) to time I. More precisely, define V(t, x" z" x2)
for loST + L I by the recursion
V(T + L1, Xl' Zb X2) = 0
V(I, x" Z" x,) ~ min {c,(t)z, + c2(t)(Y2 - x2) + hl(t)x I + C 2(t, Y2)
Chapter 9 Time- Varying, Stochastic Demand: Policy Optimization 425
yields the opti + 'fE[V(t + 1,'\ + Z1./.,-1 - d(I), (Z"zlI,··' ,ZI,L,-2),)', - d(I))]
cj= (1- 'f)cp
: Z1 2:: 0, Xl ~)'2 S id
Then, by ao inductive argument following that of Theorem 9.6. I, one can show that, for
t ~ T + Lt.
V(t, X" z" X20 Z2) ~ (:,(1, x2' z,) + V(t, x" z" xJ
Moreover. the values of Zl aod)'2 that optimize Vare truly optimal.
Slep 2. Now define the functions VI(I, X" ZI) for I <0 T + L 1 by the recursion
V ,(T + L10 XI' z,) = 0
Iat,in (9.8.3). V(t, X" z" X2) = min (c,(t)z, + h 1(1)i1 (9.8.7)
[,. Therefore. + 'fE[V,(1 + l,x I + ZlL, ,_, - d(I), (Z"ZII,.·· ,ZI,£ -,))] ,
+ c,(I)()', - x,) + Cil, )',) + 'fE[V,(1 + 1')'2 - d(l))]
: ZI "" OJ
Recalling that CI(t, Xl) ~ hI(t)XI + H2 (t, Xl), we see that the minimum here is precisely
VIet, X), ZI)' The induction is complete.
I
Step 3. Examine the recursion (9.8.6) defining VI' This has precisely the fonn of
the dynamic program (9.6.2) for a single-stage system, so we can apply Theorem 9.6.1
directly to simplify it. Specifically, defining
Notes II
CI(t, Xl' ZI) ~ C1(t, Xl) + cl(t, Xl + ZI L , -1) + ... cf' -I(t, Xl)
. I
and recalling the definition of VI in (9.8.5), we have
VIet, X), ZI) ~ CI(t, X), ZI) + VIet, Xj) O:s;t:s;T
Moreover, the base-stock policy that solves (9.8.5) yields the optimal stage-I shipments.
9.8.4 Extensions
It should be clear that the same logic applies to longer series systems (J> 2): Work back
ward from J to 1. For each stage, reduce the state vector (as in step 3 above), and then
separate the optimal cost (as in step 2). In conclusion, an echelon base-stock policy is
optimal.
Returning to the two-stage system, suppose there is an external fixed cost kj (t) =
k I > 0, but the internal fixed cost k 2 (t) = O. The entire argument above remains valid as
stated. In the dynamic programs involvingzj [(9.8.3), (9.8.5), and several others], the or
der cost includes a tenn kjO(ZI)' So, the optimal policy is a base-stock policy for stage
2, computed using (9.8.4), and an echelon (r, s) policy for stage I, obtained by solving
(9.8.5).
A positive internal fixed cost k2 (t), however, destroys the argument. (Step 2 breaks
down.) We can say little about the optimal policy here. An echelon (I; s) policy (as de
scribed in Section 8.3) is certainly a plausible heuristic.
Now, consider an assembly system with linear shipment costs. In general, the opti
mal policy is quite complex. However, consider the infinite-horizon model with station
ary data. Provided the initial state satisfies an analogue of the long-run balance condi
tion (8.4.3), a balanced echelon base-stock policy is optimal. The best such policy,
moreover, can be found by analyzing an equivalent series system.
What about other starting states? If we do use that policy, sooner or later the system
will achieve the balance condition, and from then on, the policy is optimal. That suffices
Chapter 9 Time- Varying, Stochastic Demand: Policy Optimization 427
for the average-cost objective ("y ~ I); the best balanced echelon base-stock policy is
,1(. II)' truly optimal. For the discounted-cost case ("y < I), however, during that initial period
the policy is suboptimal; the true optimal policy also reaches the balance condition, and
e brooght 0lIl it does so at lower cost. But, the optimal policy is very intricate; it is hard to calculate
and hard to implement. It is reasonable, therefore, to use the balanced echelon base-stock
policy as a heuristic for all starting states.
This seems to be as far as one can go in the direction of structural complexity, at
. ·=I.L,-~lJl least for now. For a distribution system, for example, it seems impossible to compute or
even to describe the true optimal policy. (The curse of dimensionality once again asserts
its power!) We must turn to approximations. The techniques of Section 8.6, for instance,
Ie is preriseIy
can be adapted to the discrete-time system to yield a heuristic policy and a lower bound
on the optimal cost.
[I' the fonn of
b<orem 9.6.1
" , Notes
Section 9.3: The discrete-time dynamic-programming formulations of this and the next
two sections were introduced in the papers of Arrow et al. [1951] and Dvoretzky et al.
[1952]. Since then, a massive literature has accumulated, extending the basic model in
-I sbipmems.
many directions. Porteus [1990] provides a thorough review.
Section 9.4: Karlin [1958] analyzes the linear-order-cost model. The myopic policy and
its properties are attributable to Karlin [1960] and Veinol! [1965]. See Morton and Pen
tico [1995] for empirical evidence on its performance and Lovejoy [1992] for an analyt
~ 'loOIi bacl< ical performance bound. One source for Theorem 9.4.4 is Bertsekas and Shreve [1978].
oe ~ and t:b<n The imperfect-quality model of Section 9.4.8 is by Henig and Gerchak [1990], and
od~·is the linear-inflation rule is by Ehrhardt and Taube [1987] and Baker and Ehrhardt [1995].
The equilibrium approach of Section 9.4.8.3 borrows from Tang [1990] and Denardo and
I COSI kl(t) = Tang [1992], who use similar ideas to analyze more complex systems. See Bollapragada
mns ,-alid as and Morton [1994b] for a promising alternative approach. Yano and Lee [1995] review
Io:n], the O£ the related literature.
il:yfor~ Section 9.5: Scarf [1960] introduced the notion of k-convexity to prove the optimality of
d~'sohiog an (r, s) policy. See Veinol! [1966b] for an alternative approach; this paper also develops
the bounds of Theorem 9.5.6. The infinite-horizon model was first tackled by Iglehart
Ii:p 2 breaks [I 963a,b]. Zheng [1991] provides a relatively simple proof. The power approximation is
otic). (as de- attributable to Ehrhardt and Mosier [1984]; Ehrhardt [1984,1985] extends the idea.
Section 9.6: Karlin and Scarf[1958] showed how to extend the basic model to include a
GlI. the opti constant leadtime. As for stochastic leadtimes, Kaplan [1970], Nahmias [1979],
IriJh SIatioo Ehrhardt [1984], and Song and Zipkin [1996b] discuss exogenous supply systems, and
Ianre coodi Federgruen and Zipkin [1986] explore limited-capacity systems. The myopic approach
such poli0: to the lost-sales system of Section 9.6.5.1 is by Morton [1971] and Nahmias [1979]. See
also Johansen and Thorstenson [1996].
rtheS)o= Section 9.7: The world-driven demand model was introduced in Iglehart and Karlin
bat suffices [1962] and elaborated by Johnson and Thompson [1975], Lovejoy [1992], Song and
1I.1~!;i
._------
Zipkin [1993a,1996c], and Sethi and Cheng [1997]. See Azoury [1985] for the case of
Section 9.8: The basic results for series systems are attributable to Clark and Scarf
[1960] (as are many of the central concepts in the field, such as echelon inventory).
Schmidt and Nahmias [1985] and Rosling [1989] extend the ideas to assembly systems.
Chen and Zheng [1994b] provide elegant proofs for the infinite-horizon case. lida
[1998] shows that the myopic policy is effective when the data change slowly.
Eppen and Schrage [1981] and Federgruen and Zipkin [1984a,b,c] analyze distribution (a)
systems along similar lines. Bessler and Veinott [1966] and Karmarkar [1981,1987a] de proh
velop qualitative results for these and other structurally complex systems. (b) Pro,..,
(As you
makes
statio
Adapt the analysis of Section 9.4.8.3 to cover !his case. Show !hat, in equilibrium, Now. 51
that, for ..
E[u] = (:')E[d]
c+ +h
P'bh1(y)
b + h
9.14 Consider the lost-sales model of Section 9.6.5.1 with an infinite horizon, stationary data, a linear where ~
order cost, and L > I. Define Z[/, L) to be the total supply arriving from time I until just before L. linear fil
That is, Z[L -- I,L) = O,Z[O,L) = x, and
Now, suppose the myopic policy chooses z ~ z(O) at time 0 so that (9.6.6) holds exactly. Argue
that, for any realization of d(O),
,
Pr {x(L + I) -'. 0 I d(O)}I,(l)~o
• >~~
- b h )
( + - 'Y e
Td]j so (9.6.6) will again hold at time t = 1. (Therefore, (9.6.6) will continue to hold for all t > 0.)
9.15 Consider again the lost-sales model of the previous problem. Using the representation above of
an .f(L), argue that
Pr {x(L) ~ OJ oS min {Pr (X[l, L) oS O}: 0 oS I < L}
d the yield
I indicates Let us use this to devise an approximation to the myopic. policy: Let s 7solve Fg[I,L/Y) = (c++
, distributed. s;
h)/(b + h - 'Yc). (This is consistent with the definition of in the text.) Set = z7 [s7-
Z[I, L)]+.
md (This is the order size for the base-stock policy, using base-stock level s 7and treating Z[I, L) as the
be event 7
inventory position.) Finally, set z + = min {z :0 oS 1 < L}. The new policy uses z = z + as the order
his, 7,
size. Argue that, if Z[I, L) oS s then
c+ + h
min {Pr {X[!, L) oS 0] : 0 oS 1 < L} 20 - - - -
b+h-'Yc
ISIs .I:(t) do
Y is optimal [From tbis we can't conclude anything aboutPr {x(L) ~ O} itself, but it probably comes close to
d ..ily is it solving (9.6.6).]
9.16 Consider the wOrld-driven-demand model of Section 9.7. Suppose that W is a stable, first-order
iooary data. autoregressive process. That is, the dynamics of Ware given by
=(0 =2 a)')L
l:it'm)j, 9.17 Consider the world-driven-demand model of Section 9.7 with linear order costs. Assume for
convenience that W has a discrete state space. Argue that s'(t, w) oS s +(t, w), as in Theorem 9.4.2(a).
432 Foundations ofInvenfOlY Management
Also, prove the following analogue ofTheorem 9.4.2(b): Suppose that s + (t, w) S s*(t + I, w') for all
states w that are immediately reachable from w', i.e., such that Pr {W(t + I) ~ w'l Wet) = w} > O.
Then,s*(t, w) ~ s+(t, w).
9.18 Considerthe world-driven-demand model of Section 9.7.2 with linear order costs crt, w) that
depend on the current world state. Show how to compute c +(t, w) and C+ (t, W, y). That is, set
things up so that we can rewrite the dynamic program (9.7.1) to (9.7.3) in a form analogous to
(9.4.6) to (9.4.8), i.e., in terms offunctionsH(t, w,y), C+(t, w,y), and V+(t, w, x).
9.19 Consider the series system of Section 9.8. Justify the dynamics describing the echelon-inventory
measures x;(t) and x/t) in Section 9.8.1. Also, demonstrate that the echelon-based inventory
backorder cost hj(t)xj(t) + C2(t. x2(t)) is equivalent to the original cost (9.8.2).
9.20 Consider the two-stage series system of Section 9.8, specifically, the myopic policy constructed in
Section 9.8.2. Prove that s](t) s s ;(t). [This is easy forj ~ 2; just follow the proof of Theorem
9.4.2(u). Likewise, letting sJ(t) minimize
I ··)fo...n
= } > o. ApPENDlXA
OPTIMIZATION AND CONVEXITY
..) !bat
t is, set
'llIJIIS to Outline
A.I Introduction 433 A.3 Convexity 434
-inveolory
A.2 Optimization 433 Notes 437
Dory
l>1locted in
A.I Introduction
Ibeon:m
This appendix provides a brief outline of selected eoncepts in the theory of mathematical opti
mization. These are essential tools in the analysis ofinventory systems. The references cited in the
Notes below provide more complete treatments of the subject.
(I).]
A.2 Optimization
A.2.t Definitions
LetXbe a set, and! = I(x) a real-valued function defined on X. We say that x* E X minimizes f
on X; or is a (global) minimizer of/on X if/Cx·) "$. leX), x E X The value/(x"'} is the mi1limum
of/ouX Wewritef(x*) = min {/(x): x EX}, and x* = argmin (/(x): x EX). Likewise, x* E
X maximizes f on X aod/(x*) is the maximum off on X. if lex") ~ /(X), x E X Clearly, xli'
minimizes f, if and only if it maximizes -f
Not every function has a minimizer and a minimum. (Consider X = Hi: (the real numbers), and
f(x) = eX.) Every function does, however, have an injimum. the greatest lower bound of the range
(I(x) : x EX}. written inf (I(x) : x EX}. Thus, inf {e' : x E :ll ) ~ O. even though there is uo
x with eX = O. Also, inf {x: x e ffi } = -00. The corresponding concept for maximization is the
supremum, written sup {f(x) : x EX}.
Suppose X C. mil. x* is a local minimizer andf(x*) a local minimum ofIon X ifj[x*) ~ j{x)
for all x eXin some neighborhood ofx*.
SupposeXC. fi" is an open set andfa real-valued h.mction defined on X Assume thatJis con~
tinuously differentiable. The gradient off, denoted VJ(x), is the n-vector of partial derivatives of
J evaluated at x; that is,
vf(x) CC [ a{~:)]j~l
Next, assume thatJis twice continuously differentiable. The Hessian off, denoted HJ(x), is the
n X n matrix of second partial derivatives ofJevaluated at x; that is,
2f
Hf(x) = [ a (X)]n
Jx.Jx lJ=1
, j
433
....
;1lI
Vf(x*) ~ 0 (A2.1)
This result suggests a method to find x*: Solve the system ofn equations in n unknowns (A.2.1).
There is no guarantee, however, that a solution, if one exists, really solves the problem. To resolve
that issue, we need the concept of convexity, introduced in the next section.
Meanwhile, there are corresponding results for constrained optimization. We shall focus on the
linear case: Suppose A is an m X n matrix with In < nand b an m-vector. XI} is an open sct in ffi:".
lis a real-valued, continuously differentiable funetion on X m and we want to minimize lover Xo ,
subjcct to the constraints
Ax =b (A2.2)
(So, the actual set X is X o n {x : Ax = b}.) Let A' denote the transpose of A.
PROPOSITION A.2.2. Under these assumptions, ifx* is a local minimizer of/over X; then there
exists an m-vector t*, such that
The components oft* are called dual variables or Lagrange multipliers. Here, (A.2.2) and (A.2.3)
comprise n + In equations in the n + m unknowns x and t.
A.3 Convexity
A.3.t Definitions and Properties
The set X ~ ffi:" is convex if, for every pair of vectors Xl, x 2 E X and every scalar 'A E [0, 1],
The expression on the right is an approximation of/ex), obtained by linear interpolation between
the values at Xl and x 2 . So, lis eonvex when the true value/(x) always lies at or below this ap
proximation. In other words, the graph of/lies below the line segment connecting [xl,/(x l )] and
[x'J(x')] inffin +!. ForX(;; ffil the graph offin ffi' has a U-like shape; for X (;; ffi'the graph of
I in ffi:3 is shaped like a bowl.
Appendix A Optimization and COllvex;/y 435
Here are some related definitions:/is strictly convex if the inequality CA.3.1) holds strictly
(with < replacing:=;;) whenevef)::] ~ x · and A E (0, l).fis concave when -Jis convex, i.e., when
J
cooditions help
<JnIcr necessary CA.3.!) holds with the inequality reversed. Likewise,Jis strictfl' concave when -fis strictly
convex.
Examples. A linear function is both convex and concave, but not strictly. For X = the positive
tX Ifx* is a 10
real numbers, and m a scalar constant with m > 1 or m < 0, the function x'" is strictly convex. If
o < m < 1,:i" is strictly concave, as is In (x). For X = 91, the functions [x]+ = max {D, x} and
[xr = max to, -x} are both convex.
(AZ_I)
Convexity is preserved under certain transformations:
kDowns (AZ_I ~ PROPOSITION A.3.I.
*"'- To resoh"
(a) IfJis convex onX and y is a constant vector, then J(x + y) is convex on the translated set X +
IalI focus on lhr y.
I open set in ~_ (b) Iffandg are convex functions on.¥, then so are the functionsj(x) + g(x) and max {f(x), g(X)}.
-fo=x". (c) IfJis convex and ~ is a scalar with ~ ~ 0, then ~J(x) is convex.
(d) If g is convex onX andJis convex and nondecreasing on m, thenJ[g(x)] is convex on X
(e) If {Jm} is a sequence of convex functions that converges pointwise to a function J(i.e.,Jm(x)
(All) --7 J(x) for each x), thenJis convex also.
From these basic laws, various other properties follow. Here is an important one, used extensively
...-X tben there in Chapters 6 and 9:
COROLLARY A.3.2. Suppose C is a convex function on X = m", and d is a random variable
taking values in m:n. Then, the function/(x) = E[C (x - d)] is convex.
(A23)
mJique. PROOF: First, suppose that d has a finite range of values, say {d,}, with Pr {d = d i } = pj' For
each i, C(x - d,) is convex in x [by part (a) ofthe Proposition], so PiC(X - d i ) is too [by part (c)].
Ulmd(A231 Consequently,f(x) ~ L;p/':(X - d;) is convex [by part (b)].
Next, consider the general case. The function Jean be represented as the limit of a sequence
{Jm}, wherej~,(x) = E[C(x - d m )], and dm is a random variable of finite range. (The expectation
is an integral, and an integral is a limit of this sort.) EachJ,.., is convex (by the previous argument),
sofis too [by part (e)].
.E[O.I]. There are equivalent characterizations of convexity whenJ is smooth to various degrees:
PROPOSITION A.3.3. Suppose Xis open and convex, and the functionJ:X --7 H1 is continuously
differentiable. Then,Jis convex if and only if, for all x and yinA',
; aJIM:L In :ll"
o:::L (Such. lEI f(x) <-fry) + (x - y) V fey) (A.3.2)
cc:D1:LIf X&; Also,Jis strictly Convex if and only if (A3.2) holds strictly for y =1= x.
isCXllM1
.. -.
"'"air """y
'
(A_HI
The right-hand-side of (A.3.2) is the first-order linear approximation ofJ(x) centered at the point y.
So,Jis convex when the UUe value j{x) lies at or above the approximation. In the one-dimensional
case, (A.3.2) becomesf(x) <- fry) + (x - y)f'(y). An equivalent condition in this case is tbatf'(x)
be nondccreasing in x.
A square matrix Q is nonnegative dejinite if, for all x, x'Qx ~ O. Also, Q is positive dejinite
-""'
_lbis. when x'Qx > 0 for all x =1= O.
[,,'-iII 'I] . PROPOSITION A.3.4. Suppose Jis twice continuously differentiable. Then,fis convex if and only
flhrgnpha( if the Hessian HJ(x) is nonnegative definite for all x. Also, ifHJ(x) is positive definite for all x, then
Jis strictly convex.
436 Foundations qj1nve/ltOlY Management
In the one-dimensional case, the condition for convexity reduces to f"{x) ::2: 0, wbilef"(x) > 0
implies strict convexity. In two dimensions, the convexity condition becomes c,z!(x)lax1 .2: 0,
;If(x)laX~::2: 0, and 1!V(x)1 ~ 0; ifall these inequalities are strict, thenfis strictly convex.
One can similarly define convexity for a function/whose domain X is an interval of integers:
fis convex if (A.3.1) holds whenever x itself is an integer. EquivalentlY,Jis convex if the first dif
ference a/(x) = [(x + 1) - f(x) is nondecreasmg inx. Also,fis convex if the second difference
f1'f(x) ~ !1[!1f(x)] '" 0 for all x.fis s"iclly convex if these relations hold strictly. (This theory
does not extend, however, to functions of several integer variables.)
A.3.2 Implications
Why do we care about convexity? The main reason is that it simplifies optimization. Assume
that X is a convex set. The key results are the following:
PROPOSITION A.3.5. Suppose/is convex onX Ifx* is a local minimizer ofJon X then x* is Notes
also a global minimizer.
PROPOSITION A.3.6. Suppose/is strictly convex onX Then,/has at most one, unique local
(and hence global) minimizer x*.
Thus, if we want to minimize!, and we know that/is convex, we need only search for a loci31 min
imum. And, ifJis strictly convex, we know that the optimal point is unique, provided one exists.
Convexity also provides necessary and sufficient conditions for loei31 optimality itself. First,
some definitions: A functionJis locally convex at x ifit is convex on some neighborhood ofx.
PROPOSITION A.3.7. Ifx* is a local minimizer off, thenJis locally convex at x*.
In the last result, if/is in fact convex everywhere, then x* is a true global minimizer, by Proposi
tion A.3.5. The first-order necessary condition (A.2.t) thus becomes sufficient also. CNe use this
result in Chapter 3 and elsewhere.) There is a comparable result for linearly constrained opti
mization:
PROPOSITION A.3.9. Under the assumptions ofPropositionA.2.2, if/is loeally eonvex at x*,
and (x*, ~*) satisfies (A.2.2) and (A.2.3), then x* is a local minimizer of/on X.
PROPOSITION A.3.10. Suppose X and Yarc convex sets, and/ = lex, y) is convex on X X Y.
Define g(x) to be the minimi31 value of/for fixed x, i.e., g(x) = min {lex, y) : y E Y}. Then, g is
convexonX
PROPOSITION A.3.l1. If/is concave onX and there exists a global minimizer, then some ex
treme point of X is a global minimizer.
IIIIIIIIIIIIII!IIIIIIIIIIIII_
while f"(x) >0 Thus, to find the global minimum, we need only examine the extreme points. Even this is hard, in
;;2f(x)la~ '" 0, general, but it is easy in certain cases, depending on the structure ofX (This result is used in Chap
, convex. ter 4).
Pial of integers: There are analogues of Propositions A.3.5 and A.3.6 for a function! defined on an interval of
I: if the :first dif integers: A local minimizer here is a value x such that I(x - 1) ~ f(x) and f(x + 1) ~ lex), or
cond difference equivalently, fj"f(x - 1) ~ 0 ~ fj"f(x). Iffis convex, andx* is a local minimizer, thenx* is also a
!y. (This theory global minimizer. IfJis strictly convex, there is at most one unique local and global minimizer.
(There are no such simple connections between local and global optimization, however, for TImc
tions of several integer variables.)
zariOll. Assume
00 X then x* is Notes
Ie, unique local There are several accessible introductions to classical optimization, the main subject here, includ
ing Luenberger [1984] and Bazaraa and Shetty [1993]. For convex sets and functions the standard
reference remains Rockafcllar [1970].
liJ<. local min We have entirely omitted the important topic ofmathematical progrannning, including linear
led one exists.. and integer programming. For good introductions see Chvatal [1983] and Nemhauser and Wolsey
iIy itself. FUst, [1988]. The collection of survey articles compiled by Nemhauser et a1. [1989] provides a useful
orboodofL overview.
r. To solve a real optimization model, there are several options: All the leading spreadsheet pro
grams include optimization modules, which are quite adequate for small, routine models. For
heavier-duty work (like the big models of Chapter 5), try one ofthe dedicated modeling languages
uwex at x· and
with built-in links to powerful optimization codes, such as GAMS (Brooke et a1. [1988]) and
AMPL (Fourer et.1. [1993]).
OL by Proposi-
"- (v.e use this
lIStr.Iined opli
raJO\~ at.·,
IlCI. 00. X x r
Y}. Then. g is
~ is an ertreIW
Iaodr:;:fL
I""""" !JOIl.'
Outline
B.l Definitions and Examples 439 B.3 Continuous-Time Systems 444
B.2 Discrete-Time Systems 442 Notes 444
The solar system. The sun and the planets, plus the moons, comets, asteroids, et cetera.
These bodies are related mainly by gravity.
connected by wires.
Here, of course, we are mainly interested in this third type of system, but the other two provide
useful analogies.
These are real systems, but we are also interested in mode's of systems, primarily graphical
and mathematical models. Both are called "systems" in common usage. This appendix explores
deterministic systems exclusivcly; stochastic systems, whose behavior is affected by random
events, are discussed in AppendiX C.
No real system exists entirely alone (except the entire universe). So, to discuss a system entails
a conceptual focus, implicit or explicit, on one part of the world, leaving out the rest. The bound
my of a system distinguishes which elements are part of it and which are not. In most practical sit
uations, it is obvious what the system is. that is, where its boundary belongs.
But not always. People can and do argue over the "right" scope ofthe system. Of course, this
depends on the purpose of the discussion. For some purposes, the solar system can be regarded as
gravitationally independent of the rest of the universe, but for others it can't. Likewise, sometimes
it is usefullo discuss a firm's own production-distribution network by itself, but sometimes it is
essential 10 inel ude customers and suppliers.
A .wbsystem is a system that is part of some larger system. TIle earth and its moon, for instance,
compose a subsystem of tile solar system. One particular item's inventory composes a subsystem
of a logistics system. Typically, a subsystem operates somewhat independeutly of the rest of the
system, so it is meaningful to focus attention on it.
R.LI Vocabulary
There is a general vocabulary for describing systems. We begin with four fundamental concepts,
input, output, state, and control.
The inpuf of a system is a stimulus or driving foree, which enters the system from outside it<;
boundary. The input may be the flow of a physical substance, or infonnation from the environment.
439
440 Foundations ofInventOlY Management
For an electrical network, a battery might provide part of the input, in the fann of a voltage
source. If the circuit is designed to operate an electric motor, the input might also include signals
indicating the desired speed of the motor. The inputs to a production system include the raw ma
terials and infonnation about customer demands for finished goods.
The output of a system measures the response or result of its operation. The definition of out
put depends on what aspects of the system's behavior interest us, or the purpose it is designed for.
Also, the output must be something we can actually observe.
For the motor-driving circuit above, the output might be the actual speed of the motor. A pro
duction system's output may simply be its production rate. We may also wish to observe, say, the
fill rate (the fraction of demands satisfied immediately).
The input and output, then, describe how a system interacts with its environment. The state of
a system, in contrast, describes the internal workings of the system itself. The state is a eomplete
description of the system's elements at a particular point in time. The key word here is complete.
The state summarizes all available information about the system, past and present, that is useful
in predicting its future evolution.
In a cireuit, for example, the state describes the voltage at each node and the current on each
arc ofthe network. The state of the solar system comprises the positions and velocities of the plan
ets. In a production-distribution system the state includes, at least, the inventory level of each item
at each location.
Sometimes. it is not obvious how to specify the state of a system; it may take some work to ar
rive at a satisfactory fonnulation. The "right" state description depends on the system's dynamics
(discussed below). Astronomers care mainly about the positions ofthe planets, not their velocities.
The velocities are necessary also, however, to predict future positions, because of the way gravity
works.
Likewise. the appropriate state of a production-transportation system depends on how the sys
rem operates. If it is hard to change a production rate, then thc state must include the current rate,
in addition to the inventory levels. In a distribution network with delivery lags, thc state must de
seribe all shipments in transit, since those affect future inventory levels.
Finally, a control is some action taken in order to modify the behavior of the system. Some
times, the control comes from outside the system, in which case it serves to modify the input. In
other situations the control is intrinsic to the system itself, based on observing the actual output.
For example, the circuit above may include a device to compare the desired and actual speeds
of the motor, and to adjust the input voltage accordingly. The control of a logistics system might
order materials and adjust production rates, based on current inventory levels.
To complete the description of a system, we need some additional terminology, initial condi
tions, dynamics, constraints, and objective.
The initial conditions specify the starting state of the system. FJGUREB.L
The dynamics of a system describes how the state and the output evolve over time. in response Block diagram:
to the input and the control. This is the fundamental core of a system.
The dynamics ofthe solar system embody Newton's (or Einstein's) laws ofmotion. Other phys
icallaws (such as Ohm's and Kirchhoff's) govern the dynamics of an electric circuit. A logistics
system's dynamics include stock-conservation laws: Ifwe start this week with certain inventories
(the current state), then decide specific amounts to order and produce (controls), and then de
mands oecm (inputs), the dynamics speeify the inventories that will be left next wcek.
Constraints place limits on the control actions or on the resulting values of the state. In most
production systems a negative production quantity is meaningless, and positive values arc limited
by production capacity; constraints codify these facts. Also, while negative inventory does have a
meaning (it can be interpreted as a backlog, as explained in Chapter 3), we may wish to forbid it Input
nonetheless, and we can do this by imposing additional constraints.
Appendix B Dynamical Sysfems 441
Finally, the objective is a summary measure of the system '5 performance. In the case of a
production-distribution system, for example, we may be interested in the total cost over a long time
interval, or perhaps the average cost over time. The performance of the circuit might be measured
as the average squared difference between the desired and actual motor speeds. The solar system
has no objective (as far as we know).
FIGURE Rl.l
Block diagram: simplest.
,-------------,
Input Output
FIGURE B.l.2
Block diagram: more detail.
..
--
--.
. _ill
-....,.~.
lnpul Control
I
State
C Output
Finally, the left-most block tells us that the control depends on the input and also the output. Again,
the diagram by itself does not tell us what happens inside the blocks, that is, the mechanisms of
the corresponding transformations. (Some systems laek eertain afthe bloeks and arrows shown in
the figure. For example, as mentioned above, the solar system has no real input, output or control,
only dynamics.)
There are yet more detailed diagrams, which show just how the blocks work. Alternatively, one
can use a hierarchical diagram. This is actually a collection of diagrams, starting with figures like
those above. It also includes additional diagrams, one for each of the blocks in Figure 8.1.2, show
ing the details ofthe block's transformation. And, some blocks in these additional diagrams may
be supplemented by yet more diagrams showing still greater detail.
Here is a fairly general mathematical model of a system. We model time as a scquencc of discrete
time points. (This notion is explained further in Chapter 4.) There is no explicit output, or rather,
T = timc horizon
t = index for time points, t = 0, ... , T
X(I) ~ state
z(l) = control
e(l) ~ inpnt
These last three are vectors. They need not have the same dimensions.
Initial conditions:
X(O) ~ Xo (B.2.1)
Dynamics:
x(l) eX t = 0, ,T (B.2.3)
Z(I) e Z t = 0, , T- 1
Objective:
Here, :!.t) is a given vector. X and Z are subsets, g is a vector function of its argwnents, and fis a
scalar fimetion of the same arguments. (For t = T,fis a function only ofx(T).)
This formulation describes ajirsf-order system, in that the next state x(t + I) depends only on
the current state, control, and input, i.e., their values at time t, not at earlier times. Some systems,
however, have higher-order dependencc. Fortunately, it is possible to modify the formulation, to
reduce a higher-order system to a first-order one. For example, suppose x(t + I) depends on
X(I - 1) as well x(f), Z(I), and .(f). Redefine the state to be Y(I) ~ [X(I), X(I - 1)]. Now, y(1 + 1)
depends only on CUTrcnt values of the state, control, and input.
.:rows sbown in Vet, x) = minimal cost from time t to time T, starting at x(t) = x
CIlIIpla 0< rootrol H(t, x,z) = f[l, x,z,e(t)] + V{I + 1, g[t, x,z,e(t)]}
Continue all the way to t = O. The optimal value of the objective is given by V(O, xo). To construct
the optimal solution, work forward in time. Set z(O) to the optimal z in (B.2.5) for t = 0 with x =
Xu, and setx(l) to the corresponding g[O, X, z, e(O)]. Sct z(l) to the optimal z in (B.2.5) for t = 1
with x = x(l), and set x(2) to the corresponding g[l, x, z, e(l)]. Continue in this manner all the
waytot=T
This method works, in principle. The fact that it does so is called the optimality principle ofd.v
namic programming.
In practice, the exact recursive solution of (B.2.5) can be a very demanding eomputation. It is
easy to make up problems which are essentially impossible to solve. Notice, (8.2.5) requires, for
each t, solving a separate optimization problem for each x. It is possible. however, to solve some
important problems exaetly, and various numerical approximation techniques ean be applied to
others.
Chapter 9 extends this idea to stochastie systems.
Here is a special but important case. It has linear dynamics, quadratic objective, and no constraints.
Initial conditions:
(B2.1, x(o) = Xo
Dynamics:
(B.2l)
x(t + 1) ~ Ax(t) + BZ(I) + Celt) t ~ 0, ... , T- 1
°Necfive:
(B2.3)
Minimize L;~oX[X'(t)Qx(t) + z'(I)Rz(t)]
Here, A, B, C, Q and R are matrices of appropriate dimensions. (It is also possible for them to de
pend on t, but we'll keep them stationary for simplicity.)
(B2.4)
Let t(l) be a vector of dual variables for the dynamic equations, and t(l) = O. The optimality
conditions are
---
amples above) and the steering mechanisms of vehicles. There. both positive and negative values
of the state variables and controls are meaningful (they just express opposite directions), and the
goal of the system is to control deviations, which the quadratic objective captures.
Chapter 4 discusses the application of this approaeh to an inventory system.
",.
444 Foundations ofInventory Management
x(O) ~ xo
Dynamics:
Constraints:
x(t) eX t e [0, 1]
z(t) e Z Ie [0, 1]
Objective:
Minimize fif[t, x(I), z(t), e(l)] dt
Notes
There are many books on systems. I have found especially useful Luenberger [1979] and Bertsekas
[1995].
ApPENDIXC
PROBABILITY AND STOCHASTIC PROCESSES
e. This approach
Outline
C.l Introduction 445 C.5 Markov Chains: Continuous Time
C.2 Probability and Random Variables 470
445 C.6 Stochastic Linear Systems in
C.3 Stochastic Processes 460 Discrete Time 476
C.4 Markov Chains: Discrete Time 463 Notes 480
C.I Introduction
This appendix is a review of probability and stochastic processes. It also introduces nomencla
ture, conventions, and notation used in the text. The material is mostly self-contained, but it is
both selective and informal: Many important subjects are left untouched, and many definitions
and arguments are intuitive rather than rigorous. The references in the Notes at the end provide
fuller expositions.
The concept of probability is truly one of the great and beantiful ideas of science and human
culture. It provides a clear, precise vocabulary for describing uncertainty and imprecision. Uncer
tainty and imprecision are central elements of inventory systems.
I and Bertsekas
C.2 Probability and Random Variables
C.2.1 Discussion
Intuitively, a random variable is some quantity which, as of now, is not precisely known, but which
will become known in the future. (There is a precise, technical definition, which we omit.) For ex
445
,
,... .,"
, ., .''': ,.~~~! 1'_
The realization of a random variable is the value that actually occurs, when all uncertainties
have been resolved. For instance, D above describes a demand quantity bf{ore the period in ques
tion. .~tierthat period, we might observe a demand of 103; this is the realization of D.
Sometimes we consider several related quantities, all uncertain. We can describe these as a 001
leetion of random variables or as a single, vector-valued random variable. (A stochastic process,
defined later, is a special kind of colleetion of random variables.)
gx(x) = Pr {X = x} X € S
Thus, gx(x) 2: 0, X € S, and ~x~ gx(x) = 1. It is sometimes useful to extend the function gx to all
the integers, setting gx(x) = °
for x rl S. When the symbol X is understood, we write g for gx.
(Some people refer to gxas the probability density function; this phrase has a different but related
meaning in the context of continuous random variables.)
A mode of gxis a value x at which gx(.:t) is a local maximum. We say gx is unimodal when it
has only one mode (or more generally when all the modes are contiguous).
The probability distribution or cumulative distn'butionfunction (abbreviated cdf) of X is
E[X] ~ ~x.sxg(x)
(When S is infinite, the series need not converge, in which case E[X] does not exist; even when
E[X] exists, it may be infinite. We nearly always assume that E[X] exists and is finite. The same
goes for similar series below.) More generally, if h is some function defined on S, the expectation
of II(X) is
is called the nth moment of X E[.X] itself is the first moment. The variance of X is
A related quantity is the second binomial moment of;r, Z;E[A:"(A:" - 1)]. The standard deviation of
X is the square root of its variance. For a random variable X with E[X] positive, the coefficient of
variation is the ratio of its standard deviation to its mean, i.e., (V[X])"~/E[X].
The z-tran.~torm of X (also called the probability generatingfimction, or pgf) is
g(z) ~ E[zx]
(This function is defined for those =such that the expectation exists. Such z can be complex num
bers as well as real ones.) This function has a number of important uses. For example, E[X] =
~
,~ ...1I.lI._.... ,.
For X ~ 0, G (0)
2
= Yill[X(X - I)], so
(The subscripts here indicate which variable the operator E or V applies to. For instance, in the first
~ tenn, think of Vx[XIY] as a function of Y = y; the operator Eydirects us to compute the mean of
""""'- EI.\1 ~ this function.)
.,
i ir!t~':llil~fir/l~~
,,'ff I i'!1 !!!r'li~~"'~':
When X and Yare independent, V[X; Y] = O. (However, VEX; Y] can be 0, even when X and Yare
not independent. When VEX; Y] = 0, we say that X and Yare uncorollated.) For Y = X; VEX X]
The convolution of two independent random variables X and Y is their sum W = X +r The
gw(w) = 2,xgx(x)gy(w - x)
In tenns ofz-transforms,
g.-ce) ~ gx(z)gy(z)
SupposeX;, i = 1,2, ... are independent random variables with the common pmf gx The n-fold
Next, we review some important types of discrete probability distributions. (Section C.4.4 dis
In this case S = {a, I}. A random variable X of this kind represents a choice between yes and no,
or on and off, or success and failure, etc. For instance, X = 1 might mean that a particular unit pro
Thus,
E[X] ~ P V[X] ~ p(l - p) g(z) ~ (l - p) + pz
This type has two parameters, 11 and p, and S = {O, 1, ... , n}. Here, X can be viewed as the sum
ofn independent random variables, eaeh having a Bernoulli distribution with parameter p. For in
stance, X might indicate the number of defectives out of 11 total units produced.
.. ~
Appendix C Probability and Stochastic Processes 449
Here,
iben X and r 2R Here, there are two parameters a and b. both integers, with a < b. The range is S = {a + 1, a +
T=X /1-L11 2, ... , b - 1, b}. We say Xhas the uniform distribution on S, whenXhas equal probability of oc
I
g(x) ~ b - a XES
FIGURE C.2.!
Probability mass fUnctions (binomial: n = 20).
r&o-lk~ 0.25
p ~ 0.25
--·p~O.50
~ p~0.75
_C-4.......
0.2 ..,....•.
\\ ... '",
o 0.15 ,, ,
, . -
,/
''
''
_!ID'.-.I: -..
...... --~ ~
~
£ 0.1
., .
. ,
,
,
.~ /
'
.
'
'
\,
,f
,,.
.
,
,,,
.! ,,
,"
.,,,
ft.
0.05 ,
,/
...-r,._
,,
i •
-, ',
o -'
/'
'.
fW . . o 5 10 15 20
x
450 Foundations oj Inlienrory Mallagement
)<,,"e_ A
g(,)~~ x~o
and
E[X] ~ VEX] ~ A
Also,
g(z) ~ e-A(I-.)
There is no closed form expression for G. The simplest way to compute it is to sum terms ofg di
rectly. (This approach is fast and accurate enough for most purposes, provided x is not too large.
For large x, sophisticated numerieal methods are required; see, e.g., Kniisel [1986] and Fox and
Glynn [1988].) The loss functions can be written in terms of GO as
GI(x) ~ -(x - A)Go(x) + Ag(x)
G'(x) ~ X{[(x - A)' + x]Go(x) - A(x - A)g(x)}
Here, S is the non-negative integers, and there is a single parameter p with 0 < P < 1. Consider
an infinite sequence of independent variables, each having the Bernoulli distribution with param
eter p. (This is called a Bernoulli sequence or Bernoulli process.) Then, Xis the number of l's be
Thc geometric distribution is commonly used to model the time until some event occurs, like
the failure of a machine, where there is a constant probability of failure, regardless of the age of
the machine. It arises also in Chapter 7, describing demand over a stochastic leadtime.
We have
P p
E[X] ~ 1 _ p VEX] ~ (1 _ p)'
Appendix C Probability and Stochastic Processes 451
FIGURE C.2.2
1Ii:gerS. The pmfis
Probability mass fUnctions (poisson).
0.7
0.6
mean = 0.5
---~ mean=4
0.5 _ mean=8
_ mean = 12
.£' 0.4
~
A:: 0.3
_ &:rnti 0(g . .
0.2 .-."
,/
,:,.1 "-,
~6"iOO~
""'l-Fm_
0.1
.
I
/;
.II ~
:
/'" " ,
o
o 5 10 15 20
x
~B .Mi"..y ..
-~
.... «. tJ!;IRS I-p
g(z) ~ I - pz
-,mj-.I
_ _ I k . -..
.
G1(x) ~ (-p)
I-p
p" G'(x) ~ (-p)'
I-p
p" x:=::: 0
~afiKd.~
A slightly more general class of distributions arises in certain applications: Suppose the first
Bernoulli random variable in the sequence has probability 'IT of being 1, but all the rest have prob~
ability p. Then, Xhas a delayed geometric distribution:
---.
-
- ........ ik
E[Xl~l_p
1T VEX] ~ 1T(1 +P
(1 - p)'
- 1T)
The same formula applies even when n is not an integer if the binomial coefficient is interpreted
as
(
" - I + x) ~ f(n + x)
x f(n)f(x + 1)
where r is the gamma function. The geometric distribution is the special case with n = 1.
We have
np
E[X]~~ VEX]
1- p (1 - p)'
ge,) ~ ( 1 - pz
I - p)"
Here is a useful reeursion for the pmf:
g(O) ~ (l - p)"
g(x) ~
(n-l+X) pg(x -
X 1) x 2:::: 1
(Thus, even when n is not an integer, we never actually have to compute the function r.) There is
no closed-form formula for cP, G 1 , or G2 • The loss functions can be expressed in telms of GO and
g, however. Let ~ ~ pl(l - p), so E[X] ~ n~.
Figure C.2.3 shows several negative-binomial pmf's, for p = 0.75 and four different values of
n. (The case n = I is a geometric pmf.) Evidently, as n increases, the mean increases, and the dis
tribution seems to become more spread out. Figure C.2.4 again examines four values of n, but now
p is adjusted to keep the mean fixed at 9. Here, increasing n serves to concentrate the distribution;
the variance decreases with n. In all cases, g is unimodaL Figure 7.5.1 illustrates the loss function.
The negative-binomial distribution is widely used to model demand over a (stochastic) lead
time, as in Chapter 7.
Consider a sequence of negative-binomial distributions with n increasing to GO, butp adjusted
to maintain a constant mean A. This sequence converges to a Poisson distribution with the same
mean. Thus, the Poisson distribution is a limiting case of the negativewbinomiaL (We mentioned a
similar property above for the ordinary binomial distribution.)
Appendix C Probability and Stochastic Processes 453
F,GURE C.2.3
O<p<laod Probability mass/unctions (negative binomial: p = 0.75).
0.25 I
.. random \1Iri
iixe n l's 0CCtI£
iDes fail In this
0.2 .
n= 1
• n=2
11=4
_ n=8
• is inrerpreu:d
g 0.15
i 0.1
f
..
~ ~"--'
\ '--.
,
I "~
" '.
, ~
._ = L
0.05
' .
., .... ~ -- -:...---- ..
o - ·•·· ····.- a
o 5 10 15 20 25 30 35 40
x
While it is not obvious, E[X] and VEX] are both increasing in p; they increase to 00 as p ----7 1. As
~ and "'" dis
oC..lu . . P -> 0, E[X] -> 1 and VEX] -> O. Figure C.2.S shows the pmffor four values of p. Evidently, g is
:[ iI··· monotone decreasing in all cases.
The logarithmic distribution is sometimes used to model the demand size, when an arriving
.....
_,iom customer may request more than one unit.
.... ""'
.• .m. • .....-J ...
A compound distribution is built up from two other distributions. The following example illus
trates the idea: During some time interval, several customers arrive, each demanding a certain
amount of a product. The number of customers is a random variable N, and the amount demanded
"
FIGURE C.Z.4
Probability massfimctions (negative binomial: mean = 9).
0.1 ----t-
:3
11
£ 0.04
...............
0.02
by customer n is another random variable Y". Each Y" has the same distribution as a generic ran
dom variable Y, and the Y" are independent of each other and of N We are interested in the total
demand,X= I~=l Y".
The properties of X can be derived from those of Nand y.
gXIN(x I n) ~ g/"'(x)
E[X] ~ E[N]E[Y]
[The second identity comes from the conditional-variance fonnula (C.2.l). Here, gZ[y] means
(E[Y])'.]
For example, suppose N has the Poisson distribution with mean ErN] = A. We say X has a
compound-Poisson distribution. In this case, the formulas above simplify to
iix(z) ~ exp (- A[I - !ty(z)])
The range S is now some interval of real numbers. Most ofthe basic properties of continuous ran
dom variables are similar to those in the discrete case, but there are some notable differences.
~
.S:1il_ at 7
FIGURE C.2.5
Probability mass functions (logarithmic).
0.8
l
I
0.6 -
I
If :
"I:
...•... p = 1/2
____ -p=3/4
___ p=7/S
_p=15/16
g "I:
:E
•
.<;
,I;
0.4
1
~i1:"
" \: "
"
~ ,
1\
0.2
aa:a
o
\... '~~~.:.-:.::!" --
T --r T
-lO
o 5 10 15 20
x
ia generic r.m The probabilities can be described in terms ofthe cumulative distributionfunction (cdf), which
Ol:d in the rotaI has the same meaning as in the discrete case:
Fx(x) ~ Pr{X';x} XES
The complementary cumulative distn"bution fimction (ccdf) is again
Ff(x) ~ Pr (X>x) ~ 1 - Fx(X) XES
The analogue of the probability mass function here is the probability densifyjimction (or pdf). This
is a function Jx, such that
Fx(x) ~ J~oofx(x) dx
The pdf exists only when Fx is smooth, which we assume. In particular, ifFx is differentiable, then
fx{x) ~ F{(x) .
.E 11] moan< When X is understood, we drop the subscript.x; as in the discrete case.
Expectations involving continuous random variables are defined as in the discrete case, by us
IC~-X""a ing integrals in place of sums. Thus, the mean of X is
E[X] ~ Jsxf(x) dx
and in general
E[h(X)] ~ J s h(x)f(x) dx
(Other transfonns are used for random variables defined on the full real1ine.)
The loss function has the same meaning as before:
Flex) ~ E[[X - xt] ~ J;(y - x)f(y) dy ~ J; FO(y) dy
The second-order loss function has a slightly different meaning:
F'(x) ~ Y,E[([X - xt)'] ~ y,J;(y - x)2f (y) dy ~ J; Fl(y) dy
The joint cdf of two random variables X and Y is again given by
Fuex,y) ~ Pr{X';x, Y,;y}
The joint pdfis
'
,Ax, y) ~
a2F
X,y(x, y)
J ayax
The conditional pdf. independence, and covarianee are defined in terms of the joint pdf, exactly
as in the discrete case, and their properties are the same also.
Let X denote a column vector (X;)7=1 ofn random variables with joint pdf! Some special no
tation and results ean be used in this case: The mean of X is the column vector E[X] = (E[Xa)7=lo
The covariance manix of X is the n X n matrix V[X] = (V[X;., ~·n~.j=" This matrix is always
symmetric and non-negative-definite.
Now, suppose m is any positive integer, and A is an 111 X n matrix. Define a new vector of ran
dom variables Y by the linear transformation
Y~AX
(That is, if we observe the realization of X, say X = x, then the realization ofY is y = Ax.) Then,
(b a)2
E[X] ~ Y,(b + a) V[X] ~ "'---:c~
12
]l.s) = I' + s
e- fl.:>: e -""
F'l.x) = --;;:- F'l.x) = -, x;;::.: 0
I'
Analogous to a delayed geometric distribution, a delayed exponential distribution has FO(x) =
> 0, for some eonstant 'IT, O::s;; 'IT::S;; 1.
'TI"e-tJ..r, x
Again, S is the non-negative real numbers. There are two positive parameters, nand 1-1-. Just as the
""'"' >peciaJ ,... negative-binomial distribution generalizes the geometric, so the gamma generalizes the exponen
~l ~ (£I.\](~:
tial. In particular, a gamma distribution with n = 1 reduees to an exponential.
marrix is~~
When n is a positive integer, X is the sum of n independent random variables, each having the
exponential distribution with parameter j.1. In this case, we say Xhas an Erlang (or Erlang-n) dis
=--'~ofa.-
tribution.
For general n > 0,
]l.s) = ( _ I ')"
_ _Sa I' +s
There is no closed-form expression for the edf F, in general, but it is available in many scien
tific and statistical software packages; also see Kniisel [1986]. In the Erlang case, where n is a pos
-.. .....
itive integer,
(This is also the Poisson ccdfwith mean j.1X, evaluated at n - l.) The loss functions can be ex
pressed in terms of F and f
Q
Figure C.2.6 shows four gamma pdf's with a cornmon mean of 1, but different values ofn (so
I..L =n). Evidently, the distribution becomes more concentrated near the mean for larger n.
1 ) exp
"'(z) ~ ( .Ifi,; (-2:1 z 2) ZE S
The corresponding edf is the standard nonnal cd}; denoted <I>(z), and the ccdfis <l>°(z). Thus,
q,o(z) ~ J~ ",(x) dx
This integral cannot be evaluated in closed form, but it is available in many software packages; ap
proximations can be found in Abramowitz and Stegun [1964J, Chapter 26. The standard normal
loss junctions are
FIGURE C.2.6
Probability density junctions (gamma: mean = 1).
2.5
2
n=l
---- n=2
- - n=4
.,
Q
1.5 _ _ n=8
0.5
"
"_...... _- ---..
o
o 2 3 4
x
Appendix C Probability and Stochastic Processes 459
i:n:nt \-alues ofn (so These functions have important symmetry properties:
• fiJ< laIger n.
<p( -z) ~ <p(z)
<po( -z) ~ [ - <po(z) ~ <P(z) (CZA)
With these relations, the loss functions can be written in terms of <P and <po as
boJary and notarioB
aJ distriburio•. lb< <p1(z) ~ -z<po(z) + <p(z)
<p\z) ~ ~[(i' + [)<po(z) -z<p(z)]
Thus, although there is no forom!a for <po(z), for largez, <po(z) = <Plz)!z, and so In [<p0(z)] = _~Z2
...." packa.,oes; ap
" stmrdanJ _ _ By comparison, for an exponential distribution, In [Fo(x)] = -IJ-X, and even for a gamma distri
bution one can show that In [Fo(x)] = -!-LX for large x. Thus, the tail probability epo(z) decays to
overy fast as z gets large, faster than an exponential or a gamma tail. The loss functions epl(Z) and
ep2(z) also decay to 0 relatively fast.
Now, for any v and cr > 0, letX = v + crz. Then, X has the normal distribution with parame
ters v and cr. Conversely, starting with X; Z = (.X - v)lcr has a standard normal distribution. Thus,
if z = (x - v)/cr is the standardized value or thefractile of x,
The central limit theorem says roughly tbe following: Suppose X is the sum of many inde
pendent random variables. Then, the distribution of Xis approximately normal. For example, sup
pose that demand for an item over some time period comes from many independent sources. Then,
the total demand is approximately normal.
There is also the multivariate normal distribution for a vector of random variables. We begin
with a special case: Let Z = (Z;)7=1 be a column vector of independent random variables, each
having the standard normal distribution abovc. Thus,
E[Z] ~ 0 V[Z] ~ J
Now, let l' denote an 11- vector of constants and I a symmetric, nonsingular n X n matrix, and
define the vector
~ X=v+:l:Z
~ Then,
We say X has the (nwdimensional) multivariate normal distribution with parameters v and:$. The
pdf of X is given by
f(.) ~ (1~1}1>"(X-l(. - v»
where IIi is the determinant of'$..
Y~AX
where A is an m X n matrix. For simplicity, suppose m ~ n and A has full rank (its rows are lin
early independent). Then, Y too has a multivariate normal distribution with parameters given by
(C.Z.Z) and (C.Z.3) above:
E[Y] ~ AE[X] ~ Av
(When m > n, or more generally A is not offull rank, then Y has a singular normal distribution.
It has a normal distribntion on its range, the subspace generated by A.)
C.3.2 Classification
C.3.2.t Time
The time parameter f can be continuous (real) or discrete (integer). Depending on this choice, we
say X is a continuous-time process or a discrete-time process.
'.
C.3.2.2 Space
Regardless of how we model time, the random variables X(t) themselves may be discrete or con
t (lis n:J8i'S are . . tinuous; in these cases, respectively, X is called a discrete-state or continuous-state process. With
_an&:iS gAtxa ~ two ways each to model time and space, then, therc arc four possible combinations; all four are
possible and usefuL
A continuous-state process X ean model the quantity (e.g., demand or inventory) of an infi~
nitely divisible produet. such as jet fueL A diserete-state process, in contrast, models a discrete
item, e.g., a jet. (These are not hard and fast rules, but rather modeling ehoices, just as with time,
as discussed in Chapter 3.)
r-u~
An important distinction, both in continuous and in discrete time, is whether Dr not a process
rI.... lIIaj<r dis changes predictably over time. Ifso, the process is nonstationary; if not, it is stationary. More pre
cisely, a process X is stationary if a shift in the time axis leaves its probability law unchanged. That
is, first, the X(t) all have the same (marginal) distribution, as viewed just before time O. Second,
for any fixed u > 0, the pair [X(t), X(t + u)] has the same distribution for all t. Likewise, the dis
tribution of any group of three or more variables rcmains constant over time. In particular, a sta
tionary process has £[X(I)] ~ £[X(O)], t 2: O. (Of course, on any given sample path, the quantity
X(t) typically changcs with t. Stationarity is a property of distributions, not sample paths.)
, .... cboice. "'" Here is a related definition: For any fixed u > 0, let X;it) = XU + u) - X(t), and Xli =
{XlI (t) : t 2: O}. Xit) is called the increment of X over the interval Ct, t + u]. The process X has
.~
sfationalY increments if Xu is stationary for all u > O. This is the natural concept of time
invariance for an accumulation process. It implies that E[X(t)J = E[X(I)]t, t 2: O. For example,
if the (cumulative) demand process has stationary increments, then demand during any week has
the same distribution. (Sometimes, we loosely describe a process as "stationary" when we rc
ally mean stationary increments; the context usually makes the intended meaning clear.)
Among stationary processes X, some have an important property, ergodicity; a process with this
property is ergodic. The precise definition is rather technical, and we omit it, but its major conse
quence is worth mentioning: Consider any sample path. and calculate the long-run frequency dis
tribution of X(t) on that path. For example, if X(t) is integer-valued, compute the long-run fraction
of time X(t) spends on each of the integers. IfX is ergodic, then this frequency distribution is iden
tical to the probability distribution of X(O). (Actually, this identity holds for "almost all" sample
paths, i.e., with probability 1.) In particular, the average of X(I) over I is E[X(O)].
Consider a stationary, discrete-time process X, and suppose that the X(t) are mutually inde
pendent. Then, the property above holds; this is precisely the strong law of large numbers. An er
godic process, then, is one to which the strong law applies, whether or not the X(t) are indepen
dent. (For example, let Y(I) ~ X(I) + XU + 1), where the XU) are independent. It is not hard to
show that Y is ergodic, even though the yet) are not independent.)
Ergodicity is a remarkable and useful property. The distribution of X(O) embodies information
on the likelihoods of all possible sample paths. The frequency distribution, in contrast, describes
the behavior of X on one sample path. It is a striking fact that these two distributions are identical
for certain (namely, ergodic) processes.
To appreciate this property, consider a stationary X that is not ergodic: Time is discrete, and thc
state space is S ~ {O, I}. For X(O), each value has probability li. Then, for all I > 0, X(I) ~ X(O).
Thus, cach sample path is either {D, 0, 0, ... } or {I, 1, 1, ... }. So, the proportion of D's is never
l1, which is Pr{X(O) = D}. (Herc, of course, the X(t) are not independent.) Such things are impos~
sible for an ergodic process.
Turning now to a different property, a (discretc- or continuous-time) process X has indepen
dent increments ifits increments over disjoint time intervals are independent. This implies thatX(t)
andX,lt) arc independent, for all t and u. That is, the current value provides no information about
future increments. This property is especially relevant for an accumulation process, e.g., demand.
In this case, the demands during different weeks are independent random variables.
coocept of time- More fonnally, consider the two conditional random variables
~ 0_ For example.
--=--
current state. For a Markov process X, these variables have the same distribution, for all t and u.
(There are yet more formal definitions, but this will do.)
.......
I ....... ~
.5 I ::0 __
The dynamies of a Markov process can be expressed in relatively simple terms. For instance,
in the discrete-time case, it suffices to specify the conditional distributions of [X(t + 1) IX(t)] for
-
,
r
7
...........
...
_-\a CF
F~ ~
~
all t. The initial conditions give the distribution of X(O).
Sometimes, starting with a non-Markov process X, there is a way to refonnulate the model as
a Markov process. This usually involves adding auxiliary variables. That is, we identify another
process Y, such that the joint process (X, Y) is a Markov process. (This is entirely analogous to in
eluding velocities along with positions in a planetary system.) Such a refonnulation is often crit
ieal in analysis.
lin .., -...
- 01'"",. A Markov process X is time-homogeneous if the conditional distribution of [X(t + u) IX(t)] re
j
,a: ........
mains constant over t, for all fixed u. That is, the rules governing changes from present to future, i.e.,
the dynamics of the process, remain constant over time. Notice, X can be time-homogeneous with
---'.....
out being stationary. (A stationary Markov process, however, is nearly always time-homogeneous.)
65::
5
-..4 .... The simplest kinds of Markov processes are called Markov chains. These are tbe subjects of
: : : :
the next two sections.
,
...][".-111'-........
_-
... .IIn=._
. . . .SSWSlL!r CA.l Definition
A Markov chain is a discrete-time, discrete-state, time-homogeneous Markov process. Thus, the
state space S is countable. Such models are simple but also flexible. A great variety of economic
and physical phenomena ean be represented in this framework.
~~ The phrase Markov chain sometimes refers to other models. The state space S may be some
more general set, such as the real numbers; we shall see examples in Section C.6. Or, the dynam
..
,
L~ __ ics may be time-dependent. There is even a related continuous-time model, discussed in Sec
....
tion C.S. For now, Markov chain has tbe restricted meaning above .
As time passes, X jumps from state to state within S; such jumps are called transitions. The
transition probabilities are the numbers,
' -j -
Pij ~ Pr {X(t + I) ~j I X(I) ~ i} i, j ES
"a-rpr Because X is time-homogeneous, these do not depend on t. By the Markov property, the Pi) fully
" £ jaw-.-: specify the dynamics ofX. By definition. the Pij satisfy
L_-*_
~we~'"
-.,~
2.jES P i.i = 1 i E S
..........
ia . . . . __
!'• ..-. • •
Let us collect these probabilities in a matrix P = (P,). Each row i represents the current state,
and the columns j correspond to the possible subsequent states. (If S is infinite, then P is an infi
nite matrix.) We call P the transition probability matrix ofX. If e denotes a colunm vector of ones,
the conditions above can be written
~.-lli
r
...,.
p,=,o Pe = e
II!·!I ii"
'i.'. I IIlIS til
where the matrix pi is the lith power of P That matrix thus specifies the u-step transition proba
bilities.
To complete the specification of X, we need only initial conditions, i.e., the starting stateX(O),
or more generally, the distribution of X(O). Often, however, we are interested in the family of
proccsses with a given Sand P, covering all possible initial distributions; it is common to refer to
the whole family as a Markov chain.
To aid modeling and intuition, it is useful to rcpresent Sand P by a directed graph ealled the
state-transition diagram: There is a node for each state i E S. There is an arc from node i to nodc
j, whenpij > O. Thus, the arcs indicate which transitions can occur. (Sometimes the Pi) are attached
to the arcs, but sometimes they are omitted to avoid clutter.)
For example, consider a three-state chain with S = {l, 2, 3} and
P~ (I
0
2
0
1
2
J)
Figure C.4.1 shows the state~transition diagram. (There is an arc from node I to itself, because
p" > 0.)
When S is infinite, of course, we cannot draw the whole graph. Still, there may be a pattern or
structure in P that can be indicated graphically. For example, suppose S = (0, 1,2, ... }, and for
all i
Pi,i+l = X Pi+2,i = ~
Pij = 0 otherwise
Figure C.4.2 shows the diagram. Fortunately, most Markov chains encountered in practice do have
fairly simple structures, reflecting the dynamics of the systems they model. (This faet is important
in analysis too, as explained later.)
1T,(I) ~ Pc (X(I) ~ i}
We call the 1r(t) probability vectors. Evidently, net) expresses the distribution of X(t) in the form
of a vector. The initial vector nCO) is part ofthe specification of X. For t > 0, net) describes X(t)
as viewed from just before time O.
The definitions of the 1f;(t) and Pi} imply
1T,(1 + 1) ~ lj<SPij1Tj(l) i eS
or
~
. 'll;;~;.""lirt:iti'!'"t" 'r m
FIGURE C.4.1
State-transition diagram (three states).
J transition proba n
iIalting state X(O).
1 in the family of 8
/~
OOIIDOn to refer to
0' 8
FIGURE C.4.2
to
Iy
itself because
be a pattern m
.2. ... J. and f<r
--
State~transition diagram
/ A A---X "'"
(infinite state space).
Ii ... Ii
8-8- 8-8-8· • •
-., ,---
'
Here, I is the identity matrix. Thus, the n(t) satisfy a discrete-time linear system. (It is the nota
tional convcntion for Markov chains to use row vectors.) Thus, while X(t) itself jumps unpre
dictably from state to state, thc probabilities describing this behavior evolve in an orderly, pre
dictable fashion.
Here are some key definitions:
DEFINITION. The probability vector n - is stationary if it satisfies the system of linear equa
tions
.,---
• 7 ... .lIn
n = nP (C.4.2)
If the process starts with nCO) = n -, where n - is stationary, then net) = n -, f ~ D. In this case,
and only then, X is a stationary process. Equations (CA.2) are called the balance equations.
DEFINITION. The probability vector n°O is limiting if, for every initial vector 1T(D),
limHoo 1T(t) = n ""
DEFINITION. State j is reachable from state i if, in the state-transition diagram, there is a
path from i toj. States i and} communicate if each is reachable from the other.
In terms of P, j is reachable from i if (PU)ij > 0 for some u > o. Clearly, we can partition S into
subsets Sk) such that all the states within each subset communicate, but states in different subsets
do not.
In Figure CA. I, aU three statcs communicate, so there is only one subset, S itself. Figure CA.3
shows a seven-state chain with three subsets; the subsets are indicated by rectangles and indexed
by letters.
Also, we say that one subset is reachable from another, if any state in the second subset is reach
able from any state in the first. (In this case, clearly, every state in the second is reachable from
every state in the first.) However, two subsets eannot communicate, or else wc would combine
them into one. In Figure CA.3, subsets So and Sc are reachable from SA'
DEFINITION. A subset Sk is transient if some other subsct is reachable from it. Otherwise, Sk
is recurrent.
A transient subset is literally transient; if X starts in the subset, sooner or later X must leave it,
never to return. Conversely, if X starts in a recurrent SUbset, it stays there forever. There is always
at least one recurrent subset. ill Figure CA.3, SB and Sc are recurrent, while SA is transient. (The
states in a transient subset are themselves called transient: likewise, there are recurrent states.)
DEFINITION. A Markov chain is reducible ifthere is more than one recurrent subset. Otherwise,
it is irreducible.
Thc long-tenn behavior of a reducible chain thus depends crucially on its initial conditions; start
ing in one recurrent subset, X can never reach another one. The chain in Figure CA.3 is reducible,
while the one in Figure CA.I is irreducible.
PROPOSITION CA.2. An irreducible chain has a unique stationary probability vector, 'iT-.
Conversely, a reducible chain has several stationary vectors.
PROPOSITION C.4.3. In an irreducible chain, the stationary vcctor 11'- describes the long
run frequency distribution of X for almost every sample path, regardless of 11'(0).
Next, consider the chains in Figure CAA. The first cycles indefinitely through three states. The
second is less predictable, but still it alternates between the groups of stales {I, 2} and {3, 4}.
(Both are irreducible.) Such chains are called periodic. We shall skip the formal definition; the in
tuitive idea should be clear. A chain without such systematic cycles is called aperiodic.
PROPOSITION C.4.4. An irreducible, aperiodic chain has a (unique) limiting probability
vector, 'iTe<>. Furthermore, 'iTe<> = 'iT -.
A chain with all these properties is called ergodic. Recall, Section C.3.2.5 discusses this term
in the context of a general stochastic process. Here, the usage is a bit different. An irreducible chain
with'iT(D) = 1T - has the key property of an ergodic process in the original sense, by Proposition
~
Appendix C Probability and Stochastic Processes 467
FIGURE CA.3
A reducible chain.
/~ /~
.. partition S imD
I diffi""m subsets
Now we turn to the general case of countable S = {O, 1, 2, ... }. The definitions of reachability
~. '-ector.. 'ii and communication among states remain the same as above. We need to refine the notions oftran
Suppose X(O) = i. We say that state i is transient if there is a positive probability that X never
..,.""'~ returns to i in the future, i.e., for t > 0; otherwise, i is recurrent. It turns out that either all states
in a subset are transient or all are recurrent, so we can again refer to subsets as transient or recur
_-.lk rent. (In the case of finite S this definition is consistent with the earlier one.)
It may be the case now that all subsets are transient. For example, consider Figure CA.5. Here,
.._ _ :3~.4:_
2.i'" Ik __ each state has its own subset (no two states communicate), and each is transient. Also, reconsider
-
.-s_a:a
. ~I'q
Here, all states communicate, so there is a single subset, but it is clearly transient; X(t) almost al
ways increases, and only rarely jumps down. In such cases we say that X itself is transient. Here,
there is no stationary vector; Proposition CA. 1 fails .
."~.-
~
....
FIGURE C.4.4
Periodic chains.
0)
I ,
0-0
0) o
o 8
FIGURE C.4.5
Periodic chains.
8-8-0-0-8" .
Given this new definition of recurrent subsets, the notions of reducible and irreducible chains
remain the same as before; that is, a nontransient chain is irreducible if it has only one recurrent
subset.
There are now two distinct kinds of recurrent subsets: Suppose X{O) = i, where i is recurrent.
So, X will return to i in the future, with probability 1. Consider the first-return time, that is. the
earliest time t > 0 withX(t) = i. If the mean of the firstwretum time is finite, i is positive recur·
rent; othenvise, if the mean is infinite, i is null recurrent. It turns out that either all states in are·
current subset are positive recurrent or all are null recurrent, so these terms characterize subsets.
(For finite S, every recurrent subset is positive recurrent.) Null-recurrent subsets appear only
rarely; we shall see an example in the next section.
We now make an assumption:
ASSUMPTION. There is at least one positive-recurrent subset.
(Recall, this is always true when S is fInite.) It turns out that, under this condition, the results above
for the finite-state case remain valid: There is a stationary vector (as in Proposition CA.l). Ifthe
chain is irreducible, the stationary vector is unique (Proposition C.4.2), and it characterizes the
long-run frequency distribution of X (Proposition C.4.3). If the chain is also aperiodic, there is a
limiting vector, which equals the stationary vector (Proposition CA.4).
Appendix C Probability and Stochastic Processes 469
There is no direct general method to check positive recurrenee. However, it turns out that the
assumption above holds, if and only if a stationary vector exists. One (and usually the easiest) way
to check the assumption, then, is to try to solve the balance and normalization equations (CA.2)
and (CA.3); success means the assumption holds. Unfortunately, there is now an infinite number
of equations in an infinite number of variables; there is no general procedure to solve snch sys
tems. Fortunately. many chains have speeial structure, and it is often possible to exploit that struc
ture to solve the system.
Let us now define an important class of Markov chains. The state space S can be finite or infinite.
An absorbing state i is one having Pii = 1, so Pu = O,j ::;6 i. Thus, once X arrives in an ab
sorbing state, it never leaves. Clearly, every absorbing state forms its own subset, and each is pos
itive recurrent. An absorbing Markov chain is one where every state is either transient or absorb
ing; therefore, X will enter an absorbing state sooner or later, and then stay put.
An absorbing chain is irreducible, then, when it has precisely one absorbing state. In this case,
thc long-run behavior of the chain is obvious and uninteresting. (The absorbing state has 'IT -; = 1.)
Its short-term ("transient") behavior, however, ean be interesting indeed, as we see next.
Next, we describe a type of probability distribution defined in terms of an absorbing chain. Con
sider a finite-state, absorbing, irreducible chain X, i.e., with a single absorbing state. Renumber
We revise the notation slightly: Use P to denote, not the full transition matrix, but rather only
the submatrix corresponding to transient states, leaving out the last row and column; let Pall de
note the full matrix. So,
_(P0 (1-1p)e)
Pall -
(Here, f and e have the dimensions of P, not P uf{. The 0 is a row-vector of O's.) The chain is ab
A i 4- c::hIiIIIIi
..
sorbing, it turns out, if and only if 1- Pis nonsingular.
-
Iy c-= ittaRa Similarly, let 'ITau(t) denote the full probability vector, and 'IT(t) the subvector corresponding to
transient states. Abbreviating 'IT = '7T(O), we have 'ITal/(O) = ('IT, 1 - 'ITe). It is easy to check that
n: i is iaaiC~_
..... ....
'posi:liw ~-
P~Il= ( 0
P' (I-jP')e)
• SI::a5 iB a Jr
.-:i2e~
so
In particular, the '7TU) satisfy a linear recursion of the same fonn as (CA. I):
'IT(f + I) ~ 'IT(f)P
~1C5IIB_ This system is stable (the eigenvalues of P all lie inside the unit disk), so 1im/-t",'IT(t) = O.
•
!
C.U~If_
i__
Let T be the first time the chain enters the absorbing state, that is, the time until absorption.
Observe, for each t ~ 0, the event {T > t} is equivalent to {X(t) is in a transient state}. Conse w
iIIIIIt-~•• quently,
GO(f) ~ 'IT(f)e ~ 'ITp'e «('.4.4)
)
- - - _..............._-~
470 Foundations ofInventory Management
The random variable Thas a discrete phase-type distribution with parameters ('1T, P), or for short,
T - DPH(1T,P). Any fmite row-veetor1T and matrix P of the same dimensions speeify such a dis
tribution, under the following conditions:
'IT ~ 0 'lie::::;; 1
p?o Pe:S e 1- Pis nonsingular
By using (CA.4), it is not hard to show that
g(O) ~ 1 - 'lie g(1) ~ 'li(I - p)p'-I e I> 0
To evaluate these quantities requires no more, but no less, than standard numerical tactics for
solving linear systems. To compute GO(I), for instance, start with the eolumn-vector e. Premulti
ply by Pto obtainPe; premultiply again to obtainp2 e ; continue in this manner to obtain pte. Then,
multiply by the row-vector '1T to yield GO(I). To compute Gl(t), follow the same procedure, but at
the last stcp use the row-vector 'IT(l - P)-t instead of 'IT.
This approach works well for small t. For large t, something else is required. Here are two ap
proaches: First, compute the power-of-two powers of P (P, p2, p4, p8, etc.) by direct multiplica
tion. Then, following thc binary representation of t, premultiply e by the appropriate subset of
thcse matrices to obtain pte, and proceed as above. (For example, p I8 e = pI6p2 e .)
Alternatively, start by computing the spcctral decomposition of P. That is, express P in the form
p = ADA-l, where D is a "simple" matrix. (Often, D is a diagonal matrix.) Then, pt = ADfA- i .
Because D is simple, it is easy to compute D t directly for any t.
Some of the formulas become simpler in the special ease where 'IT can be written as 'IT = lfIP
for some probability vector lfI with \fie = 1. Then,
For example, a geometric distribution has precisely this form; there is only onc transient state,
P = (p), and \fI = (1). Thus, this subclass ofDPH distributions generalizes the geometric distri
butions; the matrix P plays the same role as the scalar p. Similarly, a general DPH distribution is
the matrix analogue of a delayed geometric distribution (see Section C.2.3.5). Here, 'IT is analo
gous to 'IT, and P to p.
Phase-type distributions are quite flexible: Consider the sum, or a mixture, or the minimum, or
the maximum, of several independent random variables, each with a DPH distribution. Each such
combination also has a DPH distribution. Also, any distribution on the nonnegative integers can
be approximated as closely as desired hy a DPH distribution.
Markov process. (1bis is identical to the earlier definition of a Markov chain, except for the con
•• rl II
I,". Pl. or fuc sOOn. Here is one way to specify such a process: Suppose we are given
,specify such a dis
• A discrete-time Markov chain X E with state space S and transition matrix P, where each
Pi; =0
• A vector () = (Eli)i"S' where each Eli > 0
Construet the process X = {X(t) : t 2:: O} as follows: The state space of X too is S. The sequence
of values is precisely that of the discrete-time chain X E . The time X spends in state i on each visit
there, however, is a random variable, distributed exponentially with parameter Eli- These times are
independent of X E and of each other.
So, on entering state i, X stays there for an exponential amount of time with mean lIEl r Then,
a transition occurs to another state, according to the probabilities in the matrix P The discrete
time chain X E is said to be embedded in X (which explains the superscript E). This X, then, is
a continuous-time Markov chain.
Here is another, equivalent way to specify the data and to construct X, without reference to X E :
Define the matrix Q = (q i)' where
ID:ricaI tactics fa<
a:mr- e. Pren:mIti IO,p" j7'=i
.obiainP'e.Tben. Qij=
,JlI'lC"lIure. bul • I -OJ j=i
Notice that Pe = e implies Qe = O. Q is called the (infinitesimal) generator ofX. Then, X oper
. Bi:re are "'" lip ates as follows: Suppose X is now in state i. Over a small increment of time At, the probability of
mrea mnl:tipIica jumping to j 7'= i is approximately %(At), and that of staying at i approximately 1 - 8,(At). More
qJriaIe subset of over, these probabilities are independent of how X arrived at i and how long it has been there.
!e,.}
Thus, as long as X remains in state i, qij measures the potential of a jump to j,j 7'= i; it is called
n::ssP:i:n the fomt the transition rate from i to j.
a.P'=A..DxA- t _ The generator Q thus fully describes the behavior ofX. To fonnulate X, we can specify Q di
rectly, bypassing P and 6. (Indeed, starting with a square matrix Q, having negative diagonal and
~as'iO=~ non-negative off-diagonal entries, such that Qe = 0, we can easily recover P and 6.)
The state-transition diagram for a continuous-time chain has nodes and arcs defined as in the
discrete-time case: There is a node for each state, and an arc from i to j when qij > 0, j 7'= i. (There
is never an arc from i to itself, since Pit = 0.) The transition rates qif are used as labels on the arcs,
not the P,p
For example, suppose S = {I, 2, 3} and
e transian S13le..
~distti -1 o
H disrribmioo is
Icn:. '" is anal<>
Q~ ( ~ -6
4 -ID
be miuimum O£ Figure C.5.1 shows the state-transition diagram.
bIlL Eac!I sud> Just as in the discrete-time case, define 'lTi(t) = Pr {X(t) = i} and the row-vector '7T(t) = ['IT/t)];,
:iwe imegers can t;:::: O. Let '7T'(t) denote the row-vector [d'lTJt)/dtL the (componentwise) derivative of '7T(t) with re
spect to t. One can show that '7T(t) satisfies the system of linear differential equations
This is analogous to the difference equation (CA.I) for the discrete-time case.
C.S.2 Analysis
:-botoogt:uaus
A stationary vector '7T - in this context is one satisfying the balance equations
qJt for the con-
1TQ ~ 0
'~.'
, t"'
'~
.....
'
FIGURE C.S.1
State-transition diagram (three states, continuous time).
0)
/% 4 (0
4
or
E
11"7 11" ,~
e e,
Appendix C Probability and Stochastic Processes 473
When additionally X E (and hence X) is irreducible, this is the unique stationary vector; in that
case, 'IT-also describes the long-run frequency distribution. Under these conditions, furthermore,
X has a limiting vector 'IT = 'IT - •
00
Consider two independent continuous-time Markov chains Xl and X", with finite but possibly dif
ferent state spaces 8 1 and S2' and generators QI and Ql' Then, the joint process X = (Xl> Xl) =
{[XI (t), Xl(t)} : t ~ O} is also a continuous-time Markov chain. But, what is its generator?
Here is some useful notation: Suppose A = (a l;)!} and B are finite, square matrices, where A is
of dimensions n X nand B is m X m. The Kronecker product of A and B is
aiIB
"1)3 )
A@B~ :
(
anlB a"fi
"'-
The Kronecker sum of A and B
A(:f>B~A@I+I@B
The first I here has the dimensions of B, and the second the dimensions of A. Thus, A ® Band
~_1"_I_ A 81 B are both of dimcnsions (mn) X (mn).
•• c ., It hlms out that the generator of X is precisely Ql EB Q2'
Let us sketch an explanation: Suppose SI = {I, ... , n} and Sz = {I, , m}. List the states
---~ of the joint process X in the order (1, 1), (I, 2), ... , (I, m); (2, 1), (2, 2), , (2, m); ... ; (n. I),
~_ss-. (n, 2), ... , (n, m). Now, at each transition of X, only one ofthe component processes Xl and Xl
..
_ _ 1"..
changes; the other remains as it was. So, consider the transition rates for each type of change: If
Xl were fixed, with Ql replaced by 0, the generator of X would be n copies of Q2' arranged as
blocks along the diagonal. This is precisely I ® Qz. Likewise, if Xl were fixed, the generator of
X would be Ql ® 1. The actual generator is the sum of these two. namely Ql E8 Q2'
.7 __
I
aJ ___
logue to a continuous process.
Such chains are used to model populations of people or animals. A transition upward repre
sents a birth, and a transition downward a death (hence the name birth-death process). There are
many other applications, e.g., to supply systems, as discussed below.
A birth-death process has an infinite S, but still is highly structured. Using this structure, we
can detennine whether the chain has a stationary vector and, if so, compute it:
Denote
For a birth-death process, all the other qi, = 0, i oF j. Assume all the j.L, are positive; Ihis implies X
is irreducible. For now, assume all the Ai are positive also.
, ,......;fJ:Wio,t41l;~
Define Po = 1, and
~(
Pi = - i> 0
fLi
Rj = II;S:i Pi i~0
R = I~=oR,
'iTt = Rj 7ff) i ~ 0
_
'IT; = 'IT 1 = ItR i ~ 0
and 1T - = ('iT ~)i is the unique stationary and limiting vector. Otherwise, if R = 00, there is no so
lution, and the chain is either transient or null-recurrent.
For example, suppose all the Ai = A and the I-li = I-l, where A. and I-l are positive constants. (This
case describes a simple queue, the MIMI! system.) Then, Pi = P = A11-l, i > 0, so R; = pi, i ~ o.
Therefore, R is finite, if and only if p < 1, in which case R = 1/(1 - p). Here, the stationary vec
tor describes a geometric distribution with parameter p, that is, 'lTi - = (l - p)pi, i ~ O. If p ~ 1,
there is no stationary vector. It turns out that X is transient when p > 1, and null-recurrent when
p ~ 1.
Now, consider the case where some of the Ai = O. Let s + 1 be the smallest such i. Now, all
states i > s are transient. Ignoring these transient states, we can think of X as describing afinite
state chain on the state space {O, 1, ... , s}. This model is called afinite birth-death process. In
this case, R is always finite (because Ri = 0, i > s), and the stationary vector is computed as above.
In the special case where Ai = A and l-li = J.L, i::s: s (a finite MIMI! queue), there is a station
ary vector for any value of p. For p ¥- 1, R = (1 - ps-t-l)/(l - p), so 'IT "7 = (1 - p)pi/(l - ps+l), i
::s: s. For p = 1, 'IT - describes a unifonn distribution: 'IT; = 1/(s + 1), i::s: s.
Next, we introduce a continuous probability distribution, analogous to the discrete phase-type dis
tribution of Section C.4.4. Just as the latter is defined in terms of a discrete-time Markov chain,
Suppose X is a continuous-time Markov chain with generator Qall and a specified initial vec
tor ....alf. The last state is an absorbing state, and we partition the data as follows:
M~)
-M
Qoll~ ( 0
Again, let T be the time until absorption. Arguing as in Section CA.4, one can show that the
cedf of T is
F~(t) = ....e -Mte
That is, the subvector net) of'il'aU(t) solves the system of linear differential equations
with initial conditions nCO) = I-L, and F~(t) = 1T( t)e. We say that T has a continuous phase-type
distribution with parameters (I-L, M), or T ~ CPH(I-L, M).
If J1e < I. then Pr {T = O} > 0, so FTis not purely continuous. Nevertheless, we include this
case among the CPR distributions. Also, - M is more commonly used as the matrix parameter than
M itself; we find M more convenient.
One can show that Pr (T ~ 0 j = 1 - I'e, and fort> 0, T has pdf
x. there is no so
E[T] = !LM-Ie E[T'] = !LM-'e
F}(t) = I-LM-1e-Mte F~(t) = ....M- 2 e- Mte
=... IS"'' s. (This
!iORi = pi. j ~ O. The Laplace transform and the moments can thus be computed through matrix inversion. To cal
culate the cedf FO, we must solve the differential equations (C.5.2). While this is not a trivial task,
Il: stariooa:I). 'u
. j~O_Ifp~L there are standard numerical methods, whieh are widely available in scientific software packages.
Just as the DPH distribution generalizes the geometric, the CPH distribution can be viewed as
I-tu:::w lent "ilen
the matrix analogue of the exponential distribution. Indeed, an exponential distribution with pa
SId1 i XO"\-. an rameter j..L is a CPH distribution; there is one transient state, J.L = (1), and M = (j.1).
The CPH distributions are flexible in the same senses their discrete cousins are: They ean
II:ribing afinite
loom process. In model widely diverse phenomena, and they can approximate any distribution on the nonnegative
reals.
opmmasabo<".
is a sw:ion
II:re
)pin -pJ-\i C.S.6 Poisson Processes and Renewal Processes
A Poisson process is a counting process with stationary, independent increments. It has one pa
rameter,)., > 0, called its rale. Let D ~ (D(I): 120 OJ deoole lhe process. Then,)., ~ E[D(l)], so
E[D(t)J = At. This construct is widely used to model demands, among many other things; there,
DCt) is the cumulative demand through time t. The basic facts are set forth in Section 6.2.1 of Chap
~l"'dis
ter 6. Here, we mention some additional properties.
~'chain.
A Poisson process is a special continuous-time Markov chain. The state-transition diagram is
preeisely Figure CA.5; the only possible transitions are from i to i + 1, i ~ 0, and the transition
Ii<d initial '''>7
rate on every such are is A. Thus, regardless of the current state i. the probability of a jump to state
j + 1 (e.g., a demand occurring) during a short time At is about AfAt). (This chain is transient.)
For each t > 0, the random variable D(t) has the Poisson distribution with mean At. Also, let
T J denote the time of the first demand, and T" the time between the (1'1 - l)st and nth demands,
1'1 ~ 1. Then, the Tn are independent random variables, each having the exponential distribution
with mean VA.
These processes have some important conservation properties: Consider two independent Pois
:s I. A. liJ< .ll son processes D t and D 2 with rates At and A2' respectively. Then, the superposition D, where D(t)
boo..Jlis .... = Dt(t) + D~(t), is also a Poisson process with rate A = At + A2' Clearly, the superposition of
more than two such processes is also Poisson.
.....
476 Foundations ojInventory Management
Conversely, consider a Poisson process D with rate A and a Bernoulli sequence {Xn : n :=:: I}
(Section C.2.3.5) with parameter p, independent ofD. Now, construet a pair of processes D 1 and
D2 as follows: At the instant the nth point ofD occurs, examineXn- If..:(, = 1, assign this point to
D t , that is, incrementDJ(t) by one at the same instant; otherwise, ifXn = 0, assign the point to the
other process D z- The pair (D b D 2 ) is called a splitting ofD. It turns out that both D 1 and D 2 are
Poisson processes, with respective rates:\l = p"A and A2 = (1 - p)"A, and they are independent.
Sometimes, we are interested only in D t ; in effect, we "throwaway" the points assigned to D2 .
The process D 1 by itself is called a thinning ofD. A thinning of a Poisson process, then, is itself a
Poisson process.
Furthermore, consider a large number of independent, stationary counting processes, not nec
essarily Poisson processes, and consider the superposition of all of them. Provided certain regu
larity conditions are satisfied, this superposition behaves approximately like a Poisson process, at
least over a relatively short time interval. (This fact is sometimes referred to as the central limit
theorem for counting processes. It explains why so many actual, observed counting processes ap
pear much like Poisson processes.)
A renewalprocess is a counting process whose interevent times Tn are independent, identically
distributed random variables. The Tn can have any distribution. A Poisson process is the special
case where the distribution of the T" is exponential.
All vectors are column vectors. The e(t) represent pure noise. Specifically, the vectors e(t) are in
dependent over t, and for all t,
E[e(I)] = 0 V[e(I)] ~ I
Assume that each e(t) has a normal distribution. (This assumption is unnecessary for many of the
results, but it helps to sharpen the discussion.) Thus, each e(t) is a vector of independent random
variables, each with a standard normal distribution. The initial value Xo may be a fixed constant or
more generally a random variable with a given normal distribution.
The system operates according to the following equations:
Initial conditions:
x(O) ~ Xo
Dynamics:
X(I + 1) = Ax(t) + Se(l) t = 0, ...
~J.:.2::J: Such models are used for many purposes. For instance, many standard demand-forecasting
.
~-.;~ methods essentially fit a model ofthis form. We shall see some examples later on. Here, typieally,
...........
"'-"-.
~-.,
demand in period t is one ofthe components of x(t) or some other simple function ofx(t). There
are powerful methods to estimate the required coefficients. These and related methods fall under
the heading of time-series ana(ysis.
~-,.
• 2 7.~
c'6,2 Analysis
L a
Because Xo and the e(t) are all normally distributed, so are the x(t). By (C.2.2) and (C.2.3), the
mean and covariance matrix ofthe x(t) satisfy the following recursions:
..,....._-.
W __
~_8=£
E[x(t + Ii] = AE[x(t)]
Consequently,
~
E[x(t)] ~ A'E[Xol t = 0, ...
...........
..... 7 ""!
It is not possible to write V[x(t)J in closed form, in general, but it can be computed easily using
the recursion above.
What does system stability mean in this context? Here is a reasonable definition: The system
is stable when X has a limiting distribution. A necessary and sufficient condition for stability, it
turns out, is AI ~ 0. This is, of course, the same condition as in the deterministic case. Thus, all
-. ... the eigenvalues of A lie inside the unit disk. In this case, letting x denote the limiting random vari
able, then x has a normal distribution, where
E[x] ~ 0
and VIx] is the solution to the following matrix equation in the unknown v.
v ~ Afjj' + SS'
..
(The equation does have a solution; there are numerical methods to compute it.) IfXo has this dis
tribution, then X is stationary and ergodic.
--'~ The ease where x(t) and e(t) are both scalars is instructive. Here, X is called a first-order
autoregressive process. Write x(t) ~ x(t), e(t) ~ e(l), A ~ (0) and S = (s). So, the dynamics
,.. ....
become
.....
5 7
,
_
x(1
E[x(1 +
+ 1) ~ ax(t) + se(l)
< 1. Presuming stability, we have
1)] ~ aE[x(t)] ~a>+IE[xo]
E[x] ~ 0
k 7 _~=
V[X]~C~02)S2
...
.
. .
. ~
Here, x(t) is in general not normally distributed, but we can compute its moments: Since a(t) is
independentofx(~,
E[x(1 + 1)] = E[a(I)]E[x(I)] ~ E'+I[a]E[x o]
and by (C.2.!),
E[x] ~ 0
~
...
_ _ In its pIa<z rent information, which means that e(t) is independent ofl(t), and therefore of all past deviations.
Thus, the input sequence {eel)} has (or should have) two of the key properties assumed above,
""'" becooJe namely, E[e(t)J = 0 and independence. Now, assume that the deviations are stationary, so that
V[e(t)] is a constant V[ e], and redefine s ~ o:(V[e ])112 and e(t) ~ (o:/s)e('), Then, V[e(l)] ~ 1, and
the system dynamics are x(t + 1) = x(t) + se(t), as above.
..... dJe .. ,~As- Exponential smoothing is a popular technique in practice, perhaps too popular, in my opinion.
It is certainly simple, and that is a virtue. And, it works well when the underlying model fits, i.e.,
.-.: Sio:e .." io when the real demand is generated by something like the model above. Othcrwise, it may not. Un
fortunately, it is sometimes presented as an all-purpose tool, appropriate for all occasions, free of
model assumptions and restrictions. It is not. Also. numerous (and conflicting) rules of thumb are
cheerfully proposed to seleet the smoothing constant. But rules of any kind are inappropriate here.
The smoothing constant is a model parameter, and it should be estimated like any other. This does
not mean that fancy methods are necessary; there is a role for simple heuristics in statistical esti
mation, provided they are understood as such.
Notice too that the system here is unstable, since a = 1. In particular,
_.. ~
x(t + I) ~ ax(t) + seCt)
where 0 < a < 1, so the system is stable. In these terms the forecast becomes [CO) = 4 and J(t)
-.....
~ g' + E[x(t + I) I x(t)] ~ g' + ax(l) ~ (1 - a)g' + ad(t - I),' > O. The two parameters g' and a
....... 'CaI . .
are estimated by statistical methods.
~ An ARM"A process is a more general type of model, also used in forecasting. (ARMA stands
-~ for autoregressive maving average.) Again, there is a base demand d, and we model the deviation
x(t + 1). Now, however, x(t + I) depends not just on x(t), but also on the values at certain earlier
times, say x(t - 1) and x(t - 2). Also, it depends on sealar noise factors over several time periods,
say e(t) and e(t - 1). So, the stochastic process {x(t) : t ~ O} is certainly not a Markov proeess.
~.---=
,_tIC.. Such a process can be written as a lincar system in the form above by formulating thc state vec
tor x(t) appropriately, treating past quantities as auxiliary variables. In the example above, sct
,_"" > 0. Jro x(t) ~ [x(t), x(l- I), x(t - 2), e(l- 1)]', a vector with four components. Also, e(t) ~ e(l). Clearly,
~.'*'~- x(t + I) depends only on x(t) and e(t), so X is indeed a Markov process.
Some of the dynamic equations are now definitional identities, for example, xit + 1) = X3(t).
R~.an.
SO, some elements of A and S are fixed. Still, there are more parameters here than in the first
B-.r.=l'" order autoregressive process. This is good, in that the model is more flexible, but bad, in that more
__ a.._ data and effort are required to obtain good parameter estimates.
Most forecasting techniques are best viewed as statistical methods to fit dynamical systems
Ilk 7 •
of this kind. That is true even ofmore intricate techniques, which account for seasonal and trend
l~ • • . - . factors.
480 Foundan'ons ofInventory Management
Notes
Good introductory textbooks on probability and stochastic processes include Cox and
Miller [1965], Heyman and Sobel [1982,1984], Karlin and Taylor [1975, 1981], and Ross
[1980]. The volumes by Feller [1957,1971] are classics in this area. For a collection of
recent surveys ofvarious topics see Heyman and Sobel [1990].
These books all discuss Markov chains, but Kemeny and Snell [1960] provide a fo
cused treatment of the subject. On (both discrete and continuous) phase-type distribu
tions, see Neuts [1981].
For introductions to time series analysis and forecasting! see Box and Jenkins
[1976], Brown [1963], and Makridakis et aJ. [1983].
.,.. 1J "" 'I
AI'PENDlXD
NOTATIONAL CONVENTIONS
481
..
. ," 'H ;""',,1' """." 11111'1 d j .1
There may remain other inconsistencies that my old, jaded eyes have missed. Ifso, I apologize
again.
(Section 5.4)
A(I) Stockout indicator (at time I) ~ 1 (I(I) ~ O}
A Stoekout frequency
b Backorder penalty-cost rate
B Backorders (equilibrium random variable)
B Backorders (stochastie process)
B(I) Backorders (at time t)
11 Backorders (average)
BW Customer waiting time (equilibrium random variable)
BW Customer waiting time (average)
c Unit variable order cost
1'[1, tI) Unit cost to order and hold (from time t until u, Chapter 4)
c+ (l - 1)c (Chapter 9)
cB Aggregate backorders value (Section 5.2, Section 8.2)
cI Aggregate inventory value (Section 5.2, Section 8.2)
C(-) Total cost (e.g., average total cost, function of policy variables)
CO Current holding-backorder-cost function; C(x) = h[x]+ + b[xr
1:0 Truncated cost (Section 8.3.3)
c+O Augmented cost for myopie formulation (Section 9.4)
d(l) Demand (at time t, discrete time)
D Leadtime demand
D Demand process
D(I) Cumulative demand through time t
D[I, u) Demand in the interval [t, u)
e Base of natural logarithms
e Column-vector of ones
E['] Expectation
f Probability density function (Pdf); arbitrary function
/ Laplace transform
F Cumulative distribution function (or cdf, for continuous random variable)
FO Complementary cumulative distribution function (or ccdf, for continuous
random variable)
~
•
Appendix D Notational Conventions 483
5D.. I ape- !tiN
Fl Loss funetion (for continuous random variable)
F' Second-order loss function (for continuous random variable)
- -
g Probability mass function (pmf)
it z-transform
G Cumulative distribution function (or cdf, for discrete random variable)
GO Complementary cumulative distribution function (or ccdf, for discrete
random variable)
".a~
- G1
G'
h
!l
H(I, y)
Loss function (for discrete random variable)
Second-order loss function (for discrete random variable)
Inventory holding-cost rate
Direet handling cost rate
Total expected cost from t onward (Chapter 9)
Index for items, stages
I Inventory (equilibrium random variable); identity matrix
I Inventory (stochastic process)
I(t) Inventory (at time t)
I Inventory (average)
IN Net inventory
10 lnventory on order
IP Inventory position
IPc r+q-IP
IT Inventory in transit
ITP Inventory-transit position
IW Average stocking time
~
r I j
J
J.
Index for items, stages
Number of items, stages
Variety index (Section 5.2)
k Fixed order cost
k(t, u) Arc cost (Section 4.3)
L Leadtime (constant or random variable)
L(t) Virtualleadtime (at time t)
m Number of processors (Section 7.3)
mj Renewal funetion (discrete) (Seetion 6.6)
M Matrix parameter of CPR distribution
n Shape parameter of gamma, negative-binomial distributions: customer
index; reversed time index (Section 9.4)
-
N Nodes in a nernrork (Section 5.4)
OF Order frequency (average)
• -FF I P
P
Unit sales price; probability (parameter of geometric distribution, etc.)
Transition probability matrix
PO Total profit
: i'
Pr {-} Probability
Pre (j) Predecessors (ofitemj, Section 5.4)
q Order or batch size
Q Generator matrix (of continuous-time Markov chain)
r Reorder point
rij Routing probability for network
R Routing-probability matrix
RO Total revenue
s Base-stock level; target stock level for (r, s) policy; time index
S Processing time (random variable); arbitrary set
Sue (j) Successors (ofitemj, Section 5.4)
1 Time variable
T Time horizon
u Time between orders, or cycle time; time index
u(t) S x(l) (Section 9.4.8)
v Safety stock
1'(1) Indicator variable for ordering (discrete time)
V(t, x) Optimal expected cost from t onward (Chapter 9)
V' (I, xl V(I, x) + c(l)x (Chapter 9)
V[] Variance
V[,j Covariance
w State of the world (Section 6.3, Section 9.7); workload for an order
(Section 5.2)
wO Aggregate workload (Section 5.2)
W World, a stochastic process driving demands (Section 6.3, Section 9.7)
x(l) Inventory position (at time t, discrete time)
x(t) Net inventory (at time t, discrete time)
y(t) Inventory position after ordering (at time t, discrete time)
y Demand size (random variable)
z Standardized base~stock level; argument of z-transform
z(t) Order quantity (at time t, discrete time)
z+(t) Capacity (at time t, discrete time)
a Interest rate (continuous time)
~ Backorder-cost multiplier (Section 5.2 and Section 8.2); order-inflation
factor (Section 9.4.8)
'Y Discount factor (discrete time)
r Gamma function
8 Defect rate; Heaviside function (Chapter 4)
Ll. Defects (Section 9.4.8): difference operator
EO EOQ error function, E(l") = ~(x + 1Ix)
~ Dual variable
'1 Holding-cost multiplier (Section 5.2, Section 8.2)
~
Appendix D Notational Conventions 485
, r'~
BmLIOGRAPHY
Abate, 1., G. Choudhury, and W Whitt [1995]. Exponential approximations for tail probabilities
in queues, I: Waiting times. Operations Research 43, 885-901.
Abramowitz, M., and I. Stegun [1964]. Handbook ofMathematical Functions, National Bureau
of Standards, Washington, DC.
Afentakis, P:, B. Gavish, and U. Kannarkar [1984]. Computationally efficient optimal solutions to
the lot-sizing problem in multistage assembly systems. Management Science 30, 222-239.
Aggarwal, A., and J. Park [1993]. Improved algorithms for economic lot-size problems.
Operations Research 41, 549-571.
Aggarwal, S. [1974]. A review of current inventory theory and its applications. International
Joumal a/Production Research 12,443-472.
_ _ [1985]. MRP, JIT, OPT, FMS? Harvard Business Review 63 (September-October), 8-16.
Ahmed, M., D. Gross, and D. Miller [1992]. Control variate models for estimating transient
perfonnance measures in repairable item systems. Management Science 38,388-399.
Akella, R., S. Rajagopalan, and M. Singh [1992]. Part dispatch in random yield multistage
flexible test systems for printed circuit boards. Operations Research 40, 776-789.
Anily, S., and A. Federgruen [1990]. One warehouse multiple retailer systems with vehiele
routing costs. Management Science 36, 92-114.
APICS [1994J. MRP II software/vendor directory. APICS-T7te Petformance Advantage 4, 9
(September),5ll-58.
Archibald, B., and E. Silver [1978]. (s, S) policies under continuous review and discrete
compound Poisson demands. Management Science 24, 899-908.
Arrow, K. [1958]. Historieal background. Chapter 1 mArrow et al. [1958].
Arrow, K., T. Harris, and 1. Marschak [1951]. Optimal inventory policy. Econometrica 19,
25ll-272.
Arrow, K., S. Karlin, and H. Scarf(eds.) [1958]. Studies in the Mathematical Theory oj
Inventory and ProdlLction, Stanford University, Stanford, CA.
_ _ [1962]. Studies in Applied Probability and Management Science, Stanford University,
Stanford, CA.
Arrow, K., S. Karlin, and P. Suppes (eds.) [1960]. Mathematical Methods in the Social Sciences,
Stanford University, Stanford, CA
Askin; R. [1981]. A procedure for production lot sizing with probabilistic dynamic demand.
AIlE Transactions 13,132-137.
Ass'ad, M., and M. Beckmann [1988]. Approximate solution of inventory problems with
Poisson demand, continuous review, and fixed delivery time. In Chikan, A, and M. Lovell
(eds.) [1988]. Economics ofInventory Management. Elsevier, Amsterdam.
Astrom, K. [1970]. Introduction to Stochastic Control Theory, Academie Press, New York, NY.
Atkins, D. [1990]. A survey of lower bounding methodologies for production/inventory models.
Annals ofOperations Research 26, 9-28.
Atkins, D., and D. Sun [1995]. 98%-effective lot sizing for series inventory systems with
backlogging. Operations Research 43, 335-345.
Axsater, S. [1982]. Worst case perlormance for lot sizing heuristics. European Journal of
Operations Research 9, 339-343.
487
· Ii l'I ••i'f~~!IlI".R!I •
Bibliography 489
problems. Bitran, G., T. Magnanti, and H. Vanasse [1984]. Approximation methods for the uncapacitated
dynamic lot size problem. Mmlagement Science 30, 1121-1140.
ith stoehastic Bitran, G., and H. Vanasse [1982]. Computational complexity of the capacitated lot size
problem. Management Science 28, 1174-1186.
rtwo-Ievel Blackburn, 1. (ed.) [1991]. TI"me Based Competition, Irwin, Homewood, lL.
Blackburn, J., and R. Millen [1982]. Improved heuristics for multi-stage requirements planning.
I\'entory systems. Management Science 28, 44--56.
___ [1985]. A methodology for predicting single-stage lot-sizing perfonnance. Journal of
ultilevel Operations Management 5, 433-438.
Blanchard, 0. [1983]. The production and inventory behavior of the American automobile
'I Planning and industry. Journal ofPolitical Economy 91, 365-400.
Blinder, A., and L. Maccini [1991]. Taking stock: A critical assessment of recent research on
) demand inventories. Journal ofEconomic Perspectives 5, 73-96.
Bollapragada, S., and 1. Morton [1994a]. A simple heuristic for computing non-stationary (s, S)
lead time. Naval policies. Working paper, Carnegie-Mellon University, Pittsburgh.
_ _ [1994b]. Myopic heuristics for the random yield problem. Working paper, Camegie
case of Mellon University, Pittsburgh.
Box, G., and G. Jenkins (1976J. Time Series Analysis (revised ed.), Holden-Day, San Francisco.
rision Sciences Bramel, 1., and D. Simchi-Levi [1995]. A location-based heuristic for general routing problems.
Operations Research 43, 649-660.
Jlenishment Brooke, A., D. .Kendrick, and A. Meeraus [1988]. GAMS, Scientific Press, Redwood City, CA.
Brown, R. [1963]. Smoothing, Forecasting Wid Prediction ofDiscrete Time Series, Prentice
and inventory Hall, Englewood Cliffs, NJ.
Browne, S., and P. Zipkin [1991]. Inventory models with continuous, stochastic demands.
•tile literature. Annals ofApplied Probability 1,419-435.
Bums, L., R. Hall, D. Blumenfeld, and C. Daganzo [1985]. Distribution strategies that minimize
1JI5. Journal of transportation and inventory costs. Operations Research 33. 469-490.
Buzacott,1. [1989]. Queueing models ofkanban and MRP controlled produetion systems.
Engineerhlg Costs and Production Economics 17, 3-20.
IIiUIlic lot size Buzacott, J., S. Price, and 1. Shanthikurnar [1992]. Service level in multistage MRP and base
~e24, stock controlled production systems. In G. Fandel,1. Gulledge, and A. Jones (eds.), Nell"
Directions for Operations Research in Manufacturing, Springer, Berlin.
it geoeraI fO()( Buzacott, J., and G. Shanthill1II1ar [1992]. Stochastic Models ofManufacturing Systems,
Prentiee-Hall, Englewood Cliffs, N1.
Cachon, G. [1999]. Competitive supply ehain inventory management. In Tayur et a1. [1999].
mDurions of Chase, R., and N. Aquilano [1992]. Production and Operations Management (6th ed.), Irwin,
Homewood, lL.
P"reS Chaudhry, M., and 1. Templeton [1983]. A First Course in Bulk Queues, Wiley, New York.
Chan, L., A. Federgruen, and D. Simchi-Levi [1998]. Probabilistic analyses and practical
!liaR aIId algorithms for inventory-routing models. Operations Research 46, 96--106.
Chen, F. [1999]. Optimal polieies for multi-echelon inventory problems with batch ordering.
Iific. &imooL Operations Research (in press).
Chen, E, and Y. Zhcng [1994aJ. Evaluating eehelon stock (r, nq) policies in serial
_e-. production/inventory systems with stochastic demand. Management Science 40,
1262-1275.
.._a
oauy _ _ [1994b]. Lower bounds for multi-echelon stochastic inventory systems. Management
Science 40, 1426-1443.
_ _ [1997J. One-warehouse, multi-retailer systems with centralized stock information.
Operations Research 45, 275-287.
mII PI
'SIemS. _ _ [1985]. Easily computed approximations for (s, S) inventory system operating
characteristics. Naval Research Logistics Quarterly 32, 347-359.
oodel and Ehrhardt, R., and C. Mosier [1984]. A revision of the power approximation for eomputing (s, S)
policies. Management Science 30, 618-622.
, Inventory Ehrhardt, R., and L. Taube [1987]. An inventory model with random replenishment quantities.
International Journal afProduction Research 25, 1795-1804.
Elmagbraby, S. [1978]. The economic lot scheduling problem (ELSP): Review and extensions.
..earr:h Management Science 24,578-598.
EI-Najdawi, M. [1992]. A compact heuristic for common cycle lot-size scheduling in multi
obIem. stage, multi-product production processes. Intemational Journal ofProduction Economics
27,29-42.
urioo systems: EI-Najdawi, M., and P. Kleindorfer [1993]. Common cycle lot-size scheduling for multi-product,
multi-stage production. Management Science 39, 872-885.
On Eppcn, G. [1979]. Effects of centralization on expected costs in a multi-location newsboy
iaIlbIJ.. problem. Management Science 25, 498-501.
Eppen, G., F. Gould, and B. Pashigian [1969]. Extensions of the planning horizon theorem in the
dynamic lot size model. Jl..fanagement Science 15, 268-277.
>f models and Eppen, G., and R. Martin [1987]. Solving multi-item capacitated lot-sizing problems using
variable redefinition. Operations Research 35, 832-848.
L ~f..\.. Eppen, G., and L. Schrage [1981]. Centralized ordering policies in a multi-warehouse system
34,4 with leadtimes and random demand. Chapter 3 in Schwarz [1981].
Erickson, R., C, Moruna, andA. Veinott [1987]. Send-and-split method for minimum-concave
<nIalioNll cost network flows. Mathematics ofOpeTations Research 12, 634-664.
Erkip, N., W. Hausman, and S. Nahmias [1990]. Optimal centralized ordering policies in multi
"..jof echelon inventory systems with correlated demands. Jl..fanagement Science 36,381-392.
Erlenkotter, D. [1989]. An early classic misplaced: Ford W. Harris' economic order quantity
model of 1915. Management Science 35, 898-900.
~ _ _ [1990]. Ford Whitman Harris and the economic order quantity model. Operations
Research 38, 937-946.
pr1JIIdl fir Evans, 1. [1985].An efficient implementation of the Wagner-Whitin algorithm for dynamic lot
sizing. Journal ofOperations Afanagement 5, 229-235.
odeImgand Faaland, B., and 1~ Schmitt [1993]. Cost-based scheduling of workers and equipment in a
*'mce. fabrication and assembly shop. Operations Research 41 1 253-268.
_.......
Fcdergruen, A. [1993J. Centralized planning models for multi-echelon inventory systems under
-.: uncertainty. Chapter 3 in Graves et al. [1993].
~Lo
.~. Federgruen. A., M. Queyranne, and Y. Zheng [1992]. Simple power of two pOlicies are close to
optimal in a general class of production/distribution networks with general joint setup
costs. Mathematics ofOperations Research 17, 951-963.
~ Federgruen, A., and M. Tzur [1991]. A simple forward algorithm to solve general dynamic lot
sizing models with n periods in O(n log n) or O(n) time. Management Science 37, 909-925.
~~ _ _ [1993]. The dynamic lot sizing model with backlogging; A simple O(n log n) algorithm
<:10_ .....
and minimal forecast horizon procedure. Naval Research Logistics 40, 459-478.
_ _ [1994]. Minimal forecast horizons and a new planning procedure for the general dynamic
_.
A
~
,
--
L~_
_ _ [1999]. A single-item inventory model for a nonstationary demand process.
Manufacturing & Service Operations Management I, 5Q--61.
Graves, S., A. Rinnooy Kan, and P. Zipkin (eds.) [1993]. Logistics ofProduction and Inventory,
Handbooks in Operations Research and Management Science, Vol. 4, Elsevier (North~
......
Holland), Amsterdam.
II.... L ~ Green, L., and P. Kolesar [1991]. The pointwise stationary approximation for queues with
nonstationary arrivals. Management Science 37,84-97.
--
Groenevelt, H. [1993]. The just-in-time system. Chapter 13 in Graves et al. [1993].
Gross, D., and C. Harris [1971]. On one-for-one-ordering inventory policies with state~
dependent leadtimes. Operations Research 19,735-760.
_ _ [1985]. Fundamentals ofQueueing Theory (2nd ed.), Wiley, New York.
~ Ha, A. [1997]. Optimal dynamic scheduling policy for a make-to-stock produetion system.
Operations Research 45, 42-53.
-.-.CA. Hadley, G., and T. Whitin [1963]. Analysis ofInventory Systems, Prentice-Hall, Englewood
r"'~3L Clifts, NJ.
-
....
IS--Iai
Hall, R. [1991]. Queueing Jfethods for Services and }"fanufacturing, Prentice-Hall, Englewood
Cliffs, NJ.
Hammer, M. [1990J. Reengineering work. Harvard Business Review 68 (July-August), 104-112.
.......
Hanssmann, F. [1962]. Operations Research in Production and Inventory Control, Wiley, New
York.
....
Harris, F. [1913]. How many parts to make at once. Factory. The Magazine ofManagement 10,
135-136,152.
. .Z8-Zll. Harrison, 1., and V Nguyen [1990]. The QNET method for two-moment analysis of open
queueing networks. QueHeing Systems 6, 1-32.
Hax,A., and D. Candea [1984]. Production and ImJentory Management, Prentice-Hall,
Englewood Cliffs, NJ.
- Hayes, R., and 0. Garvin [1982]. Managing as if tomorrow mattered. Harvard Business Review
60 (May-June), 70-79.
~L
.......
--
Heinrich, C., and C. Schneeweiss. [1986]. Multi-stage lot-sizing for general production systems.
Chapter 9 in Axs.ter et al. [1986].
Henig, M., and Y. Gerchak [1990]. The structure of periodic review policies in the presence of
...
_ _ [1984]. Stochastic Models in Operations Research, Vol. II, MeGraw-Hill, New York.
_ _ (eds.) [1990]. Stochastic Models, Handbooks in Operations Research and Management
Science, vol. 2, Elsevier (North-Holland), Amsterdam.
Higa, I., A. Gegerheim, and A. Machado [1975]. Waiting time in an (S - I, S) inventory system.
Operations Research 23, 674-680.
Holt, C., F. Modigliani, 1. Muth, and H. Simon [1960]. Planning Production, Inventories and
Work Force, Prentiee-Hall, Englewood Cliffs, NJ.
Iglehart. D. [1963a]. Dynamic programming and stationary analysis in inventory problems.
Chapter I in Searf et al. [1963].
_ _ [1963b]. Optimality of(s, S) policies in the infinite horiwn dynamic inventory problem.
Management Science 9, 259-267.
IgLehart, D., and S. Karlin [1962]. Optimal poliey for dynamic inventory process with
nonstationary stochastic demands. Chapter 8 in Arrow et al. [1962].
lida, T. [1998]. The infInite horizon non-stationary stochastic multi-echelon inventory problem
and near myopie policies. Working paper, Tokyo Institute of Technology, Tokyo.
Jackson, 1. [1957]. Networks of waiting lines. Operations Research 5, 518-521.
_ _ [1963]. Jobshop-like queueing systems. Management Science 10, 131-142.
Jaekson, P. [1988]. Stoek allocation in a two-echelon distribution system, or "what to do until
your ship eomes in." Management Science 34, 880-895.
Jackson, P., W. Maxwell, and 1. Muckstadt [1985]. The joint replenishment problem with powers
of two restrictions. AIlE Transactions 17, 25-32.
_ _ [1988]. Detennining optimal reorder intervals in capacitated production-distribution
systems. Management Science 34, 938-958.
Jacobs, F. [1983]. The OPT schedUling system. Production and Inventory Management (3d
quarter), 47-51.
Johansen, S., andA. Thorstenson [1993]. Optimal and approximate (Q, r) inventory polieies
with lost sales and gamma-distributed lead time. International Journal ofProduction
Economics 30,179-194.
_ _ [1996]. Optimal (r, Q) inventory policies with Poisson demands and lost sales:
Discounted and undiscounted cases. International Journal ofProduction Economics 46,
359-371.
Johnson, G., and H. Thompson [1975]. Optimality of myopic inventory policies for certain
dependent demand processes. Management Science 21, 1303-1307.
Johnson, M., H. Lee, T. Davis, and R. Hall [1995]. Expressions for item fill rates in periodic
inventory systems. Naval Research Logistics 42, 57-80.
Jones, P., and R. Inman [1989]. When is the economic lot scheduling problem easy? IIE
Transactions 21,11-20.
Juran, J. [1988]. The Quality Control Handbook (4th edition), McGraw-Hill, New York.
Kaplan, R. [1970]. A dynamic inventory model with stochastic lead times. Management Science
16,491-507.
_ _ [1986]. Must elM be justified by faith alone? Harvard Business Revie»' 64 (March
April), 87-97.
Karlin, S. [1958]. Optimal inventory policy for the Arrow-Harris-Marschak dynamic model.
Chapter 9 in Arrow et a1. [1958].
_ _ [1960]. Dynamic inventory policy with varying stochastic demands. Management
Science 6, 231-258.
Karlin, S., and H. Scarf[1958]. Inventory models of the Arrow-Harris-Marschak type with time
lag. Chapter 10 in Arrow el al. [1958].
~
Bibliography 495
....bk. Karlin, S., and H. Taylor [1975]. A First Course in Stochastic Processes (2nd ed.), Aeademic
• L • '. Press, New York.
_ _ [1981]. A Second Course in Stochastic Processes (2nd ed.), Academic Press, New York.
--..
-".,-.. Kannarkar, U. [1981]. Policy structure in multi-state produetionlinventory problems: An
application of eonvex analysis. Chapter 16 in Schwarz [1981].
_ _ [1987a]. The multilocation multiperiod inventory problem: Bounds and approximations.
Management Science 33, 86-94.
_ _ [1987b]. Lot sizes, lead times and in-process inventories. Management Science 33,
409-418.
.,.~ _ _ [1989]. Getting control ofjust-in-time. Harvard Business Review 67 (Septernber
... October),122-131.
_ _ [1993]. Manufacturing lead times, order release and capacity loading. Chapter 7 in
Graves et a1. [1993].
1IIy~ Karmarkar, U, and 1. Rummel [1990]. The basis for costs in batching decisions. Journal of
Manufacturing and Operations Management 3, 153-176.
._
~
Karmarkar, U, S. Kekre. and S. Kekre [1987]. The dynarnic lot-sizing problem with startup and
reservation costs. Operations Research 35, 389-398.
._
Karush, W [1957]. A queueing rnodel for an inventory problem. Operations Research 5,
693-703.
Kashyap, A., and D. Wilcox [1993]. Production and inventory control at the General Motors
Corporation during the 1920s and 1930s. American Economic Review 83, 383--401.
iI Katalan, Z. [1995]. Production and service management under setup times and uncertainties.
Ph.n Dissertation, Columbia University, New York.
_01 Kelly, F. [1979]. Reversibility and Stochastic Networks, Wiley, New York.
Kemeny, 1., and 1. Snell [1960]. Finite Markov Chains, Van Nostrand Reinhold, Princeton, N1.
r fIIIIIir::ios Kimball, G. [1988]. General principles of inventory controL Journal ofManufacturing and
~ Operations Management 1, 119-130.
..
Kling, R., and Iacono, S. [1984]. The control of information systerns development after
-
,
Ik irnplernentation. Communications ofthe ACM 27, 1218-1226.
_ _ [1989]. The institutional character of eomputerized information systems. qllice:
Technology and People 5, 7-28.
Kniisel, L. [1986]. Computation of the chi-square and Poisson distribution. SIAM Joumal on
Scientific and Statistical Computation 7, 1022-1036.
'.........
Krajewski, L., B. King, L. Ritzman, and D. Wong [1987J. Kanban, MRP, and shaping tbe
manufacturing environment. Management Science 33,39-57.
r. Kropp, n, and R. Carlson [1984]. A lot-sizing algorithm for reducing nervousness in MRP
-
systems. Management Science 30, 240-244.
Lasdon, L., and R. TeIjung [1971]. An efficient algorithm for multi-item scheduling. Operations
_s.-.. Research 19, 946-969.
Lawler, E., 1. Lenstra, A. Rinnooy Kan, and n Shmoys (eds.) [1985]. The Traveling Salesman
~ Problem: A Guided Tom ofCombinatorial Optimization, Wiley, New York.
Leachman, R., and A. Gascon [1988]. A heuristic scheduling policy for multi-item, single
maehine production systems with time-varying, stochastic demands. Management Science
~-
-
~--
34,377-390.
Lee, H., and C. Billington [1993]. Material management in decentralized supply ehains.
Operations Research 41,835-847.
Lee, H.. C. Billington, and B. Carter [1993]. Hewlett-Packard gains control of inventory and
serviee through design for localization. Intel/aces 23, 4 (July-August), 1-11.
..........
Lee, H., and S. Nahmias [1993]. Single-product, single-location models. Chapter 1 in Graves et
al. [1993].
Lee, R., and M. Rosenblatt (1987]. Simultaneous determination of production cycle and
inspeetion schedules in a production system. Management Science 33, 1125-1136.
Lee, Y, and P. Zipkin [1992]. Tandem queues with planned inventories. Operations Research 40,
936-947.
_ _ [1995]. Processing nehvorks wilh planned inventories: Sequential refinement systems.
Operations Research 43, 1025-1036.
Lippman, S., and K. McCardle [J 977]. The competitive newsboy. Operations Research 45,
54-{j5.
Little, J. [1961]. A proof of the queueing formula L = X-W Operations Research 9, 383-387.
Love, S. [1972]. A facilities in series model with nested schedules. Management Science 18,
327-338.
Lovejoy, W [1992]. Stopped myopic policies in some inventory models with generalized
demand processes. Management Science 38, 688-707.
Luenberger,o. [1979]. Introduction to Dynamic Systems: Theory, Models, and Applications,
Wiley, New York.
_ _ [1984]. Linear and Nonlinear Programming (2nd ed.), Addison-Wesley, Reading, MA.
Lundin, R., and T. Morton [1975]. Planning horizons for the dynamic lot size model. Operations
Research 23, 711-734.
Maes, 1., and L. Van Wassenhove [1988]. Multi-item single-level capacitated dynamic lot-sizing
heuristics: A general review. Journal ofthe Operational Research Society 39, 991-1004.
Mahajan, v., and Y. Wind [1986]. Innovation Diffusion ModelsIor New Product Acceptance,
Ballinger, Cambridge, MA.
Makridakis, S., S. Wheelwright, and V. McGhee [1983]. Forecasting: Methods andApplications
(2nd ed.), Wiley, New York.
Markowitz, D., M. Reiman. and L. Weill [1995]. The stochastic economic lot scheduling
problem: Heavy traffic analysis of dynamic cyclic policies. Working paper, MIT,
Cambridge, MA.
Maxwell, W [1964]. The scheduling of economic lot sizes. Naval Research Logistics Quarterly
11,89-124.
Maxwell, W, and 1. Muckstadt [1985]. Establishing consistent and realistic reorder intervals in
production-distribution systems. Operations Research 33,1316-1341.
Menich, R., and R. Serfozo [l991]. Optimality of routing and servicing in dependent parallel
proeessing systems. Queueing Systems 9, 403-418.
Miller, B. [1974]. Dispatehing from depot repair in a recoverable item inventory system.
Management Science 21, 316-325.
Mitchell, 1. [1987]. 98°/'O-effective lot-sizing for one-warehouse, multi-retailer inventory systems
with backlogging. Operations Research 35, 399-404.
Mitra, D., and 1. Mitrani [1990]. Analysis of a kanban discipline for eell eoordination in
production lines, I. Management Science 36,1548-1566.
_ _ [1991]. Analysis ofa kanban discipline for cell coordination in production lines, II:
Stochastic demands. Operations Research 39, 807-823.
Moinzadeh, K., and H. Lcc [1987]. A continuous-review inventory model with constant
resupply time and defective items. Naval Research Logistics 34, 457--467.
Monahan,1. [1984]. A quantity discount prieing model to increase vendor profits. Management
Science 30, 720--726.
Montgomery, D., and L. Johnson [1976]. Forecasting and Time Series Analysis, McGraw-Hill,
New York.
~
Bibliography 497
I (inIo:s l:I Morse, P. {1958]. Queues, Inventories and Maintenance, Wiley, New York.
-
Morton, T. [1971]. The near-myopic nature of the lagged-proportional-cost inventory problem
-' with lost sales. Operations Research 19, 1708-1716.
36. Morton, T., S. Lawrence, S. Rajagopolan, and S. Kekre [1988]. SCHED-STAR: A price-based
.... shop scheduling modu1c. Journal ofManufacturing and Operations Management 1,
131-181.
.,-.,... Morton, T., and D. Pentico {I 995]. The finite-horizon non-stationary stochastic inventory
problem: Near-myopic bounds, heuristics, testing. Management Science 41, 334-343.
'*45, Muckstadt, 1., and R. Roundy [1993]. Analysis of multistage production systems. Chapter 3 in
Graves et aL [1993].
l3--3lrL Muckstadt, 1., and L. Thomas {1980]. Are multi-echelon inventory methods worth implementing
-..11, in systems with low-demand-rate items? Management Science 26, 483--494.
..... Naddor, E. [1975]. Optimal and heuristic decisions in single and multi-item inventory systems.
-
Management Science 21, 1234-1249.
Nahmias, S. {1979]. Simple approximations for a variety of dynamic Icadtime lost-sales
inventory models. Operations Research 27, 904-924.
_ _ [1981]. Managing repairable item inventory systems: A review. Chapter 13 in Schwarz
~MA. [1981].
q ., _ _ [1982J. Perishable inventory theory: A review. Operations Research 30, 680-708.
_ _ {1994]. Demand estimation in lost sales inventory systems. Naval Research Logistics 41,
:~ 739-757.
'1-1_. Nemhauser, G., and L. Wolsey [1988J. Integer and Combinatodal Optimization, Wiley, New York.
-... Nemhauser, G., A. Rinnooy Kan, and M. Todd (eds.) [1989]. Optimizatian, Handbooks in
Neuts, M. [1979]. A versatile Markovian point process. Journal ofApplied Probability 16,
•
764-779.
--..
~
I""'*f
Peiia-Perez, A., and P. Zipkin [1997]. Dynamic scheduling rules for a multiproduct make-to
Platt, D., L. Robinson, and R. Freund [1997]. Tractable (Q. R) heuristic models for constrained
Porteus, E. [1985]. Investing in reduced setups in the EOQ model. Management Science 31,
998-1010.
_ _ {1986a]. Investing in new parameter values in the discounted EOQ model. Naval
Research Logistics Quarterly 33, 39-48.
l}'~ _ _ [1986b]. Optimal lot sizing, process quality improvement and setup eost reduction.
-
Operations Research 34, 137-144.
_ _ [1990]. Stochastic inventory theory. Chapter 12 in Heyman and Sobel [1990].
Ramaswami, v., and M. Neuts [1980]. Some explicit fonnulas and computational methods for
.. II: infinite-server queues with phase-type arrivals. Journal ofApplied Probability 17, 498-514.
Rao, A. [1989]. A survey ofMRP II software suppliers' trends in support ofilT. Production and
• Inventory Management 30 (3d quarter), 14-17.
Rao, U. [1994]. Convexity and sensitivity properties ofthe (R, T) inventory control policy for
" stochastic demand models. Working paper, Cornell University, Ithaca, NY.
Raymond, R. [1931]. Quantity and Economy in Manufacture, McGraw-Hill, New York.
_.M.
Renberg, B., and R. Planche {1967]. Un modele pour la gestion simultanee des n articles d'un
stock. Revue d'Informatique efde Recherche Operationelle 1, 6, 47-59.
...
498 Foundations ofIrrventory Management
Resh, M., M. Friedman, and L. Barbosa [1976]. On a geeeral solutioo ofthe detenninistic lot
size problem with time-proportional demand. Operations Research 24, 718-725.
Richards, F. [1975]. Comments on the distribution of inventory position in a continuous-review
(s, S) inventory system. Operations Research 23, 366-371.
Rockafellar, R. [1970]. Convex Analysis, Princeton University, Princeton, NJ.
Rosenblatt, M., and H. Lee [1986). Economie production cyeles with imperfeet produetion
processes. lIE Transactio1LY 18, 48-55.
Rosenfield, D., and M. Pendrock [1980]. The effects of warehouse eonfiguration design on
inventory levels and holding costs. Sloan Management Review 21, 4, 21-33.
Rosting, K. [1986]. Optimal lot-sizing for dynamic assembly systems. Chapter 7 in Axsater et
a!. [1986].
_ _ [1989]. Optimal inventory policies for assembly systems under random demands.
Operations Research 37, 565-579.
_ _ [1993]. 94%-effective lotsizing: User-oriented algorithms and proofs. Working paper,
Linkoping Institute of Technology, Linkfiping, Sweden.
Ross, S. [1980]. Introduction to Probability Models (2nd ed.), Academic Press, New York.
Rothkopf, M., and S. Oren [1979]. A closure approximation for the nonstationary MlM/s queue.
Management Science 25, 522-534.
Roundy, R. [1985]. 98% effective integer-ratio lot-sizing for one warehouse multi-relailer
systems. Management Science 31,1416--1430.
_ _ [1986]. A 98% effective lot-sizing rule for a multi-product, multi-stage production
inventory system. Mathematics a/Operations Research 11, 699-727.
_ _ [1989]. Rounding off to powers of two in continuous relaxations of capacitated lot sizing
problems. Management Science 35, 1433-1442.
_ _ [1993J. Efficient, effective lot sizing for multistage produetion systems. Operations
Research 41, 371~385.
Sahin, I. [1979J. On the stationary analysis of continuous review (s, S) inventory systems with
constant lead times. Operations Research 27, 717-729.
_ _ [1982]. On the objective function behavior in (s, S) inventory models. Operations
Research 30, 709-725.
_~ [1983]. On the continuous review (s, S) inventory model under compound renewal
demand and random lead times. Journal a/Applied Probability 20,213-219.
Salomon, M. [1990J. Deterministic Lotsizing Models/or Production Planning. Ph.D.
Dissertation, Erasmus University, Rotterdam.
Searf, H. [1958a]. A min-max solution of an inventory problem. Chapter 12 in Arrow et a1.
[1958].
_ _ [19 58b]. Stationary operating charaeteristies of an inventory model with time lag.
Chapter 16 in Arrow et a!. [1958].
_~ [1960]. The optimality of (s, S) polieies in the dynamie inventory problem. Chapter 13 in
Arrowetal. [1960].
Scarf, H., D. Guilford, and M. Shelly (eds.) [1963]. Multistage Irrventory Models and
Techniques, Stanford University, Stanford, CA.
Schmidt, c., and S. Nahmias [1985]. Optimal policy for a two-stage assembly system under
random demand. Operotions Research 33, 1130-1145.
Sehneeweiss, C. [1974]. Optimal production smoothing and safety inventory. Management
Science 20,1122-1130.
Schneeweiss, C., and H. SehrOder [1992]. Planning and scheduling the repair shops ofthe
Deutsehe Lufthansa AG: A hierarehieal approaeh. Production and Operations Management
1,22-33.
Bibliography 499
_lor Schonberger, R. [1982]. Japanese Manufacturing Techniques, Free Press, New York.
!5. Schwarz, L. [1973]. A simple continuous-review deterministic one-warehouse N-retailer
«view inventory problem. Management Science 19, 555-566.
_ _ (ed.) [1981]. !YIlllti-Level Production/Inventory Control Systems: Theory and Practice,
.
.ADiIo.-cr
production/inventory assembly systems. Management Science 21, 1285-1294.
Sethi, S., and F. Cheng (1997]. Optimality of(s, S) policies in inventory models with Markovian
demand. Operations Research 45, 931-939.
Sherbrooke, C. [1968]. METRIC: A multi-echelon technique for recoverable item control.
--
Operations Research 16, 122-141.
_ _ [1975]. Waiting time in an (S -1, S) inventory system-----<:onstant service time case.
--....
1Ib-. Shingo, S. [1985]. A Revolution in Manufacturing: The S/J.,fED System, Productivity Press,
Cambridge, MA.
_ _ [1989]. A Study ofthe Toyota Production System from an Industrial Engineering
Viewpoint (revised ed.), Productivity Press, Cambridge, MA.
Shulman, R., and D. Smith [1992]. Scheduling production to minimize holding cost<;. Working
paper, AT&T Bell Laboratories, Holmdel, NJ.
---
~ Silver, E., and R. Meal [1973]. A heuristic for selecting lot size quantities for the case ofa
deterministic time-varying demand rate and discrete opportunities for replenishment.
Production and Inventory Management (2nd quarter), 64-74.
Silver, E., and R. Peterson [1985]. Decision Systems for Inventory Management and Production
Planning (2nd ed.), Wiley, New York, NY.
---
l
Simon, R. [1971]. Stationary properties ofa two~echelon inventory model for low-demand
items. Operations Research 19, 761-773.
Singhal, V [1988]. Inventories, risk and the value of the firm. Journal ofManufacturing and
Operations Management 1, 4-43.
Sivazlian, B. [1974J. A continuous-review (s, S) inventory system with arbitrary interarrival
distribution between unit demand. Operations Research 22, 65-71.
Smith, S. [1977]. Optimal inventories for an (S -1. S) system with no backorders. Management
-
••• Science 23, 522-528 .
Sobel, M. [1969]. Optimal average-cost policy for a queue with start-up and shut-down eosts.
Operations Research 17, 145-162.
Sogomonian, A., and C. Tang [1993]. A modeling framework for coordinating promotion and
...-u. production decisions within a fUID. Management Science 39,191-203 .
Song, 1., and P. Zipkin [1992]. Evaluation of base-stock policies in multiechelon inventory
r systems with state-dependent demands. Naval Research Logistics 39, 715-728.
--
......
_ _ [1993a]. Inventory control in a fluetuating demand environment. Operations Research 41,
351-370.
... ...
_ _ (l996a]. The joint effect of leadtime variance and lot size in a parallel processing
environment. Management Science 42, 1352-1363 .
_ _ [1996b]. Inventory control with information about supply conditions. Management
Science 42,1409-1419.
_ _ [1996c]. Managing inventory with the prospect of obsolescence. Operations Research 44,
215-222.
...........
Stalk, G., and T. Hout [1990]. Competing against Time, Free Press, New York.
Stalk, G., P. Evans, and L. Shulman. [1992]. Competing on capabilities: The new rules of
corporate strategy. Harvard Business Review 70 (March-April), 57-69.
Starr, M., and D. Miller [l962J. [llvenlO1Y Control: Theory and Practice, Prentice-Hall,
Englewood Cliff" N1.
Stidham, S. [1974]. Stochastic clearing systems. Stochastic Processes and Their Applications 2,
85-113.
Svoronos, A., and P. Zipkin [1991]. Evaluation of one-for-one replenishment policies for
multiechclon inventory systems. Management Science 37, 68-83.
Swain, 1. [1994]. Cruncbing numbers. ORiMS Today 21 (October), 48--{)!.
Tang, C. [1990]. The impact of uncertainty on a production line. Management Science 36,
1518-153!.
Tayur, S., R. Ganeshan, and M. Magazine (eds.) [1999]. Quantitative Models for Supply Chain
Management, Kluwer, Norwell, MA.
Tripp, R., et al. [1991]. A decision support system for assessing and controlling the effectiveness
of multi-echelon logistics actions. Interfaces 21~ 4 (July-August), 11-25.
Vargas, G., and R. Dear [1990]. Buffering against multiple uncertainty sources in assembly
manufacturing. Journal ofManufacturing and Operations Management 3, 309-334.
Veatch, M., and L. Wein [1994]. Optimal control of a two-station tandem production/inventory
system. Operations Research 42, 337-350.
Veinott, A. [1965]. Optimal policy for a multi-product, dynamic, nonstationary inventory
problem. Management Science 12, 206-222.
_ _ [1966a]. The status of mathematical inventory theory. Management Science 12~ 745-777.
_ _ [1966b]. On the optimality of(s, S) inventory policies: New conditions and a new proof.
SIAM Journal on Applied Mathematics 14,1067-1083.
_ _ [1969]. Minimum concave cost solution ofLeontiefsubstitution models of multi-facility
inventory systems. Operations Research 17, 262-299.
Veinott, A., and H. Wagner [1965]. Computing optimal (s, S) inventory policies. Management
Science 11, 525-552.
VoUmann, T., W Berry, and D. Whybark [1992]. Mam~facturing Planning and Control Systems
(3d ed.), Irwin, Homewood, IL.
Wagelmans, A., S. Van Hoesel, and A. Kolen [1992]. Economic lot sizing: An O(n log n)
algorithm that runs in linear time in the Wagner-Whitin case. Operations Research 40,
SI45-S156.
Wagner, H. [1993]. Plus ,a cbaage.... Interfaces 23, 5 (September-October), 2~25.
Wagner, H., and T. Whitin [1958]. Dynamic version of the economic lot size model.
Management Science S, 89-96.
Wagner, H., M. O'Hagan, and B. Lundh [1965]. An empirical study of exact and approximately
optimal inventory policies. Management Science 11, 690-723.
Walrand, 1. [1988]. An Introduction to Queueing Networks, Prentice Hall, Englewood Cliffs, N1.
Wein, L. [1992]. Dynamic scheduling of a multiclass make-to-stock queue. Operations Research
40, 724-735.
Wemmerlov, U. [1982]. A comparison of discrete, single stage lot-sizing heuristics with special
emphasis on rules based on the marginal cost principle. Engineering Costs and Production
Economics 7, 45-53.
Weng, Z. [1995]. Channel coordination and quantity discounts. Management Science 41,
1509-1522.
Whitin, T. [1953]. The The01Y ofInventory Management, Princeton University, Princeton, N1.
,- i'l I
Bibliography 501
Whitt, W [1982]. On the heavy-traffic limit theorem for GI/G/oo queues. Advances in Applied
noIes of Probability 14, 171-190.
_ _ [1983]. The queueing network analyzer. Bell System Technical Journal 62, 2779-2815.
Hall _ _ [1991]. The pointwise stationary approximation for MjM/s queues is asymptotically
correet as the rates increase. Management Science 37, 307-314.
rf"i -1, _ _ [1992]. Understanding the efficiency of multi-server service systems. Management
Science 38, 708-723.
:ic5 lOr Williams, T. [1984]. Special products and uncertainty in production/inventory systems.
European Journal o.fOperations Research 15,46-54.
Wolff, R. [1977J. The effect of service time regularity on system performanace. In K. Chandy
.... 36, and M. Reiser (eds.) Computer Performance, North-Holland, Amsterdam.
_ _ [1982]. Poisson arrivals see time averages. Operations Research 30, 223-231.
iot¥.>. arm. Yano, c., and H. Lee [1995]. Lot-sizing with random yields: A review. Operations Research 43,
331-334.
~ di;;ai.Uit55 Yurkiewicz,1. [1993]. Forecasting software survey, OR/MS Today 20 (February), 64-75.
Zangwill, W [1966]. A dctenninistic multiproduct, multifacility production and inventory
>bIy model. Operations Research 14,486-507.
,...334.
_iz .M,
_ _ [1968]. Minimwn concave cost flows in certain networks. Management &ience 14,429-450.
_ _ [1969]. A backlogging model and a multi-echelon model of a dynamic economic lot size
-s:
Io!-'
-41,
" Logistics 39, 285-305.
Zheng, Y, andA. Federgruen [1991]. Finding optimal (s, S) policies is about as simple as
evaluating a single policy. Operations Research 39, 654-665.
Zheng, Y, and P Zipkin [1990]. A queueing model to analyze thc value of centralized inventory
infonnation. Operations Research 38, 296-307.
Zimmerman, 1. [1997]. Accountingfor Decision Making and Control (2d ed.), Irwin,
Chicago, IL.
!S. Zipkin, P [1982]. Exact and approximate cost functions for product aggregates. Management
l Science 28, 1002-1012.
I _ _ [1986a]. Stochastic leadtimes in continuous-time inventory models. Naval Research
b Logistics Quarterly 33, 763-774.
.
77
,...-
"..
systems. Operations Research 34, 91-104.
_ _ [1988]. The use of phase-type distributions in inventory-control models. Naval Research
Logistics Quarterly 35, 247-257.
_ _ [1989]. A kanban-like production control system: Analysis of simple models. Working
~ paper, Columbia University, New York.
_ _ [1991a]. Does manufacturing need a JIT revolution? Harvard Business Review 69
_"I. (January- February), 40-50.
~
..
_ _ [1995]. Processing networks with planned inventories: Tandem queues with feedback.
~
"
"'R
; r
e
; ; ·
.• e'
i
,
I,
r i J:
~ ~
',' ;
'I
I N D E x
m.
fPrd>ack.
Abate, 1., 289, 487
Backorders; see Planned backorders
Backward recursion, 86
Accounting numbers, 25
Baker, R, 427, 488
Algorithm Forward_PEL, 86
Basic smoothing modeL 99
All-or~nothing formulation, 98
Beckmann, M., 231, 487, 488
All-units discounts, 57
Berg, M., 289, 488
Allocation, 340
Berkley, R, 360, 488
American Production and Inventory Control Society (APICS), Bernoulli distribution, 448
27
Bernoulli process, 450
Asymmetries, 99
Book, overview, 5, 11
503
,~--
504 Index
Buffer stoek, 92
gamma distribution, 457
Buffering, 166
general properties, 454-456
Cards, 352
Convexity,434-437
Clusters, 126
Cyclic schedule, 155
Communicate, 466
Daganzo, C., 103, 169,489,490
Compound~Poissondemand
Dannenbring, D., 168,359,492
Compound-Poisson process, 227; see also Lumpy demand Dear, R., 361, 500
Concave, 435
Decentralized systems, 296
Constraints, 440
Continuity,461
discounted costs, 94, 95
--.
~
.J'.u........ ' 7
Index 505
DEL model-Cont.
Distribntion systems-Cont.
extensions, 93-99
leadtimes, and, 141, 142
heuristics, 90, 91
local control
leadtimes, 94
basic model, 331, 332
nervousness, 92, 93
general distribution networks, 338
opacity,87-90
limited-capacity supply systems, 337
Demand process, 7
Dual variables, 434
Demand uncertainty, 26
Dynamic economic lotsize model; see DEL model
Dimensionality, 411-414
continuous-time systems, 444
Discount factor, 94
discrete-time systems, 442, 443
Discounted cost, 63
vocabulary, 439-441
Discounted value, 63
Discounting, 63
Diseconomies of scale, 99
inventory-routing problem, 153, 154
general control
EI-Najdawi, M., 169,491
..
506 Index
Engineering viewpoint, 17
one item (constant demand rate), 50-55
EOQ,36
Finite statc space, 466, 467
EOQ formula, 36
Fixed-plus-linear order costs, 396-404
Generator, 471
Federgruen,A, 93,103,169,231,238,360,427,428,487,
Glassennan, P., 493
489,491,492,501
Glynn,~,289,450,492,493
Feeney, G., 492
Gould, F., 491
Fill rate, 42
Graves, S., IS, 104, 169,359,360,493
Filtration, 460
Greedy (myopic), 76, 90
Finite-horizon problem
Gross, D., 289, 360, 487, 493
~
L: :
Index 507
Hall, R, 169,238,289,489,493,494
processing networks, 254-256
Hanssman, E, 169,493
Infimum, 433
Harris, E, 17,493
fixed-plus linear order costs, 403, 404
Harrison, 1, 290,493
linear order costs (stationary data), 382-385
Hax,A,169,493
hunan,R., 169,494
Hayes, R., 63, 493
Input, 439
Heaviside function, 80
Institute for Operations Research and the Management
Hessian, 433
Inventory theory, 5
19 Heymwn,D., 13,238,269,289,480,493,497
Inventory-workload tradeoff curve, 114-116
Historical costs, 25
HMMS model, 13
Holding·cost factor, 25
Jackson, 1., 265, 494
(Li.d.) random variables, 246; see also Independent, stochastic llT approach, 15,39,168,358,359
leadtimes
Job·replenishment problem
Imperfect quality
Johnson, G., 427, 494
Increment, 461
Just-in-time (ITT) approach, 15, 39, 168,358,359
Independence, 462
508 Index
Kelly, E, 495
Lee, Y, 360, 496
Leadtime-demand distributions
Linear holding costs, 312, 314, 315
approximation, 285
Linear-inflation heuristics, 393-396
proofs, 287-289
finite-horizon problem (dynamic programming), 375-378
Leadtimes
imperfect quality, 390-396
DEL model, 94
infinite-horizon problem (stationary data), 382-385
~
'+""""1'''0. wi" U~1t 11'''
Index 509
Lost sales
countable state space, 467-469
Lot sizing, 30
MaIkmvitz, D" 360, 496
Luenberger, 0.,125,437,444,496
Maximal approximation, 210
opti.mization, 231
Mean,446,455,456
policy evaluation, 228-231
Meeraus, A., 437, 489
Maccini, L, 104,489
Mismatches, 9, 10
Managerial accounting, 25
Monma, c., 169,491
Managerial viewpoint, 17
Monotonicity, 461
~--
-- Market research, 26
Markov chains
continuous time
analysis, 471-473
MPS, 166
510 Index
Myopic, 76, 90
OPT, 167
N,l26
Optimization
N.Juni", S., 22, 67, 289, 360, 427, 428, 491, 496, 497, 498
formulation, 213, 214
le81!lju ...
Negative-binomial distribution, 452
general (r, q) model (continuous approximation), 217-223
r
Nemhauser, G., 437, 497
general (r, q) model (discrete formulation), 223-226
Nervousness, 92, 93
lumpy dcmand, and, 231
Newsvendorproblem, 374
Optimized production technology (OPT), 167
Overview of book, 5, 11
Objective, 441
Off-the-shelf programs, 21
Pallotino, S., 492
approximations, 205-213
Pentico, D., 427, 497
optimization, 213-227
Planche, R" 360,498
Opacity, 87-90
performance criteria, 41-44
...'11""
",0".-''-'' j
1
Index 511
Planning horizon, 87
(r, s) policy, 227
pmf,446
Ramaswami, v., 289, 497
Point estimates, 26
Random variables, 445, 446
Rate, 475
~ preliminaries, 178-181
Realization, 446
proofs, 193-196
Recurrent subset, 466
•
Poisson process, 475, 476
Predecessor. 139
Renewal process, 199,476
Predictable processes
Reorder-pointlorder-quantity policy
Probability, 445
Richards,E, 238,498
Ritzman, L, 170,495
Processors, 244
Ross, S., 480, 498
Rummel, 1.,169,490,495
Quantity-units, 30
planned backorders, 43
512 Index.
Scale diseconomies, 99
optimal policy. 313-316
ScaIT, H., 18, 238, 239, 360, 427, 428, 487, 490, 494, 498
service-level constraint, 322
Scheduling horizon, 93
lWo-stage system, 316-319
Schrage, 1.,169,360,428,491,499
distribution systems (cenlral control), 339-348; see also
Sensitivity
shared supply processes, 348-352
Sensitivity analysis
aggregate performance measures, 112-] 14
systems
inventory routing problem, 153, 154
coordination/sensitivity, 132-135
series systems, 120-139; see also Series systems
proofs, 135-139
Shared supply processes, 348-352
302-322
Shulman, H., 103,499
illustration, 312-316
Simond, M., 238, 239, 492
Index 513
Speculation opportunities, 82
Supply-demand mismatches, 9,10
-
Speculative motive, 82
Supply process, 7
Splitting, 476
Supply variation, 286
--
It State space, 460
Target stock level, 227
Theorem, 476
Thinning, 476
Stochastic processes
Time-varying demands, 73-106
continuity/monotonicity,461
dynamic economic lotsize model; see DEL model
defIned, 460
fast changes, 74
ergodicity/independence, 462
finite production rate, 77, 78
space, 461
smoothing models, 99~103
514 Index
stochastic demand)
Wagner, H., 80, 103, 168, 169,231,238,500
Time-weighted hackorders, 42
Wemmerlov, u., 91, 500
Transitions, 463
Withdrawal, 340
coordination/sensitivity, 149
Wong, 0., 170, 495
TSP, 153
Wosley, L., 437, 497
Thmover, 118
Unimodal, 446
Zheng, Y, 169,231,238,322,360,427,428,489,491,492,
Variance, 446,455
Zipkin, P., 15,39, 104, 169, 170,238, 283, 289, 290, 335, 360,