Paper Business Mathematics and Informatics

Stochastic Programming in Health Care Planning

Student name:
Student number:

J.M. de Rue
1163493

Supervisor:

S. Bhulai

BMI paper J.M. de Rue

Preface
This BMI paper is the last phase of the study Business Mathematics and
Informatics at the VU University in Amsterdam. This paper has the goal to use
the knowledge gained from the study to address a problem, formulate this in a
clear way, and to provide a solution.
Operations Research is a widely applied mathematical technique. It is used for
many different business problems. However, Operations Research has its
shortcomings. In this paper, I will give an example of these shortcomings and
describe several approaches to solve these shortcomings. I will apply one of
these approaches to an example in which the major shortcoming is solved by
recourse models.
I would hereby like to thank Sandjai Bhulai. Without his time, help, and
catching enthusiasm, this paper would not have been possible. Furthermore, I
would like to thank Maarten van der Vlerk for sending me his lecture notes on
Stochastic Programming. These lecture notes enabled me to learn a great deal on
the subject of Stochastic Programming.

Jaap de Rue
Haarlem, August 2007

Page 1

BMI paper J.M. de Rue

Page 2

BMI paper J.M. de Rue

Management summary
Linear programming is one of the many specializations of Operations Research
and it is one of the most applied mathematical techniques. The many different
algorithms of linear programming can be used for the optimization of a wide
range of problems. However, there are some drawbacks to using these
algorithms. The major drawback is that all coefficients (i.e., the quantity of
resources, time, and distance) need to be deterministic and known in advance.
However, in many practical situations one can not be certain of the true value of
these coefficients. These shortcomings can be solved by stochastic linear
programming, which uses random variables to approach the values of these
coefficients.
This paper describes several approaches that can be applied when the model
contains uncertainty on some of the parameters. In particular, recourse structures
are seen as the most important class within Stochastic Programming. This paper
therefore concentrates on the application of recourse structures. An example is
given of a model which contains uncertainty on some of the parameters. Adding
a recourse structure to this model shows that this leads to a model in which the
risk is explicitly taken care of.
Operations Research is already one of the mostly applied mathematical
techniques. Combining Stochastics and Mathematical Programming increases
the applicability enormously and may thus account for an even greater number
of application areas.

Page 3

BMI paper J.M. de Rue

Page 4

BMI paper J.M. de Rue

Table of contents
PREFACE............................................................................................................ 1
MANAGEMENT SUMMARY .......................................................................... 3
TABLE OF CONTENTS ................................................................................... 5
1

INTRODUCTION...................................................................................... 7

OPERATIONS RESEARCH .................................................................... 9

2.1
2.2

2.2.1
2.2.2
2.3
3

STOCHASTIC PROGRAMMING ........................................................ 17

3.1
3.2
3.3

LINEAR PROGRAMMING ..................................................................... 10

TECHNIQUES ...................................................................................... 11
Simplex method ................................................................................. 11
Column generation............................................................................ 13
DRAWBACKS ...................................................................................... 14
APPROACHES ..................................................................................... 17
RECOURSE MODELS ............................................................................ 19
MODELLING ASPECTS ......................................................................... 21

THE MODEL ........................................................................................... 23

4.1
4.2

MODELLING ....................................................................................... 23
EXTENDING THE MODEL ..................................................................... 27

CONCLUSIONS ...................................................................................... 31

BIBLIOGRAPHY .................................................................................... 33

Page 5

BMI paper J.M. de Rue

Page 6

BMI paper J.M. de Rue

1 Introduction

Operations research, operational research, or simply OR, is an interdisciplinary

science which uses scientific methods like mathematical modelling, statistics,
and algorithms to help with decision making in complex real-world problems
which are concerned with coordination and execution of the operations within
an organization (Wikipedia.org). It is one form of applied mathematics that is
being used more and more often.
Linear programming is one of the many methods of Operations Research. It is
one form of applied mathematics that is being used more and more often,
especially after the introduction of the personal computer. However, linear
programming has some shortcomings. One of the biggest shortcomings is that
the model builder needs to provide numerical values for each of the coefficients,
for instance, the run times of a production plant or the number of different
resources. But, in most cases the model builder can not be completely certain
about the true values of these coefficients.
Stochastic linear programming is a framework for modelling optimization
problems that take into account this uncertainty. Whereas combinatorial
optimization problems are formulated with deterministic known parameters, real
world problems almost invariably include some unknown parameters. Stochastic
linear programming interprets the unknown parameters as realizations of
random variables. Hereby, they use the framework for the quantitative analysis
of uncertainty that is provided by probability theory.
In this thesis, the most important class of stochastic linear programming will be
given and an example will be given where this concept will be used to take into
account the uncertainty of the values of some of the parameters of the linear
programming model.

Page 7

BMI paper J.M. de Rue

Page 8

BMI paper J.M. de Rue

2 Operations Research
The field of operations research arose during World War II, during which the
team of Patrick Blackett made numerous crucial analyses to aid the war effort.
An example of one of these analyses is the number of boats in a convoy. The
idea of ships travelling in convoys was introduced by Britain, this way warships
could be used to accompany and protect merchant ships. Since these convoys
travel at the speed of the slowest member, the question arose whether convoys
should be large or small. Since a small convoy will be faster, it will be harder to
detect by a U-boat. Large convoys may be slower, but they can deploy more
warships against such an attack.
Another example of the work of Blacketts team is given by a report they
analysed. This report was carried out by RAF Bomber Command, for which
they inspected all bombers returning from bombing raids in Germany. This
report contained a note of all the damages on the plane inflicted by Germanys
air defence. The RAF concluded from this report that the bombers should be
armoured more heavily on the damaged areas. However, they made the
remarkable conclusion that the armour should be placed on the areas that were
not damaged. Since only the planes that came back from Germany were
included in the survey, his team reasoned that the untouched areas were
apparently the most vital areas.
After the war, people realized that the planning methods that were developed for
the military could also be used in the profit- and non-profit sector. The study
after operations research flourished after the war and not long after that the
famous simplex-method was invented by George Bernard Dantzig that still is of
great practical use nowadays.
Dantzigs original example of finding the best assignment of 70 people to 70
jobs still explains its success. The computing power required to scan all the
permutations to select the best assignment is vast and impossible. He observed
that it takes only a moment to find the optimum solution using the simplex
method, which is effectively observing that a solution exists in the extreme
points of the polygon described by the equations formed from the given
constraints.
Two other founders are John van Neumann, who developed the theory of duality
in the same year, and Leonid Kantorovich, a Russian mathematician who used
similar techniques in economics before Dantzig and won the Nobel Prize in
1975 in economics. The linear programming problem was first shown to be
solvable in polynomial time by Leonid Khachiyan in 1979, but a larger major

Page 9

BMI paper J.M. de Rue

theoretical and practical breakthrough in the field came in 1984 when Narendra
Karmakar introduced a new interior point method for solving linear
programming problems.
The possibilities of applying linear programming only became bigger and bigger
after the introduction of the personal computer and nowadays it is one of the
most applied mathematical techniques.

2.1 Linear Programming

In mathematics, linear programming involves the optimization of a linear
objective function that is subject to (multiple) linear equality and inequality
constraints. The goal is to find a feasible point for which the objective function
assumes the smallest or largest value. Such points may not exist, but if they do,
searching through the vertices of the model is guaranteed to find at least one of
them.
The standard (canonical) form of a linear program is the following:
Maximize
Subject to

cT x
Ax = b
x 0,

where x represents the vector of variables, where c and b are vectors of

coefficients, and A is a matrix of coefficients. The expression c T x to be
maximized (or minimized) is called the objective function. The equations
Ax = b are the constraints which specify a convex polyhedron over which the
objective function is to be optimized. Linear programming can be applied to
various fields of study. Most extensively it is used in business and economic
situations, but can also be utilized for some engineering problems. Some
industries that use linear programming models include transportation, energy,
telecommunications, and manufacturing. It has proved useful in modelling
diverse types of problems in planning, routing, scheduling, assignment, and
design.
There are many practical problems in operations research that can be expressed
as linear programming problems, of which certain cases are considered
important enough to have generated much research on specialized algorithms for
their solution. Linear programming is, for example, heavily used in
microeconomics and business management where it is used to minimize costs or

Page 10

BMI paper J.M. de Rue

maximize the income. Here one could think of, among others, inventory
management, resource allocation for human and machine resources, and
planning. Next to that, many of the central concepts of optimization theory are
based on ideas from linear programming. It can thus be said that linear
programming is an important field of optimization.

2.2 Techniques
This paragraph will give some of the most frequently used techniques to solve
linear programming problems.

2.2.1

Simplex method

The simplex method is a technique that solves a linear programming problem in

a finite number of steps, or it determines the insolvability of the problem. The
name originates from the fact that a simplex (a convex hull of n + 1 points) is
formed by the equations of the problem.
Given the following linear programming problem:

cT x
Ax = b
x 0,

Maximize
Subject to

(1)

with n variables and m restrictions and for which holds that m<n. In the simplex
algorithm (an iterative procedure) bases {B} are formed with corresponding
basis solutions {x}. A basis B is a part of the matrix A consisting of m linear
independent columns of the matrix A. The remaining columns of A form the
matrix R. The basis solution x for j = 1,2,...., n is then defined by:

x j = B 1b j ,

for j B ,

xj = 0,

for j R .

The reduced costs are defined by:

Page 11

BMI paper J.M. de Rue

c = c BT B 1 A

c,

and the corresponding criterion value of a solution x = ( x B , x R ) given a basis B

is:

z ( x ) = c BT B 1b c j x j .
jR

The linear programming problem (1) is called the primary problem and the dual
problem is defined as follows:
bT y
AT y c.

Minimize
Subject to

In the primary problem it is taken care of that for every iteration it holds that x is
allowed for the primary problem, i.e., x j 0 j and a new iteration is started

until x is also feasible for the dual problem, i.e., c j 0 j .

If j R : c j < 0 , a non-basis column a s R is chosen with negative reduced
costs ( c s < 0 ) to be entered into the basis B. A ratio-test has to point out which
column al has to leave the basis B, with:

(
(

) : (B
)

B 1b
l = arg min 1
jB
B a s

as

> 0 .

0.
This test ensures that for all new basis variables it holds that x new
j

In the dual problem it is taken care of that in every iteration x is feasible for the
dual problem, i.e., c j 0 j and a new iteration is started until x is also
feasible for the primary problem, i.e., x j 0 j .
If j B : x j < 0 , a basis column al B with a negative value

(x = (B b)
1

< 0 is chosen to leave the basis. A ratio-test has to point out which
column a s enters the basis B, with:
l

Page 12

BMI paper J.M. de Rue

cj

1
s = arg min 1
: (B a j )l < 0 .
jR
(B a j )l

This test ensures that for all new reduced costs it holds that c new
0 , where c new
j
j
is defined by:

new
j

= c j

(B

(B

1
1

)
c
a )

aj

for j = 1,2,..., n .

s l

If a basis solution corresponding to a base of the primary problem is both

feasible for the primary problem and for the dual problem, then this solution is
optimal.
However, in some cases the matrix A can become so large that it is almost
impossible to solve because of enormous calculating times. A technique that can
be used when matrix A becomes too large is column generation. This technique
is described in the next sub-paragraph.

2.2.2

Column generation

Column generation can most easily be explained by considering a cutting stock

problem. This problem is defined as a company producing iron bars of a given
length L. Each week the company receives orders of customers, so the company
has to produce bi bars of length li (i = 1,...., m ) which have to be cut of the large
bars. The question then is: How to cut these bars in such a way that as less iron
bars as possible are used? This problem can be solved by applying the technique
of column generation.
This technique is based on taking a part of the total collection of cutting
patterns, while the rest of these cutting patterns is (not yet) taken into account.
The linear programming problem for the chosen part of the collection is solved
and studied for optimality for the linear programming problem for all cutting
patterns. If this is not the case, one or more cutting patterns are added to the
collection that is used. This iteration is repeated until the linear programming
problem for the total collection of cutting patterns is optimal.
To describe this technique in mathematical terms, suppose we have the
following linear programming problem:

Page 13

BMI paper J.M. de Rue

Maximize
Subject to

cT x
Ax = b
x 0,

with n variables and m restrictions for which holds that m<n. Now, let
N 1 {1,...., n}; k = 1 ; and the matrix D = ( Ai : i N k ) be a part of the matrix A
for which holds that rank(D)=m, i.e., an optimal solution is possible. The vector
c N k and x N k are formed in a similar way. Then, the limited linear programming
problem is defined by:
Maximize
Subject to

c TN k x N k

Dx N k = b
x N k 0.

This limited linear programming problem is solved with the simplex method. If
an optimal solution is found for this limited linear programming problem, this
solution is also optimal for the original problem if all reduced costs are nonnegative, which are defined as follows:
c j = y T A j c j ,

for j = 1,...., n .

If the solution is not optimal for the original linear programming problem, the
matrix D is changed, i.e., N k is replaced with N k +1 ; k = k + 1 and the whole
process is repeated.

2.3 Drawbacks
It is obvious that linear programming is very popular, since so many practical
problems can be modelled as a linear program, or at least an approximation can
be made. When a problem is modelled as a linear program, powerful software is
available to solve these problems. However, the linear programming approach
does have some drawbacks.
One of these drawbacks is that numerical values have to be provided for each of
the coefficients. But, in most situations one can not be completely certain about
the true values of these coefficients. There are some approaches to take this
uncertainty into account when building the model. For instance, one could use

Page 14

BMI paper J.M. de Rue

ranges or solve the linear programming model for every possible realization of
the coefficients.
However, if the optimal solution depends heavily on the value of, some of, these
coefficients, this uncertainty will have to be taken into account in a more
fundamental way. Stochastic programming is characterized by interpreting the
uncertain coefficients as realizations of random variables. Hereby the framework
for the quantitative analysis of uncertainty provided by probability theory is
used.

Page 15

BMI paper J.M. de Rue

Page 16

BMI paper J.M. de Rue

3 Stochastic Programming
Stochastic programming is a framework for modelling optimization problems
that involve uncertainty (stoprog.org). Many practical problems can be modelled
with linear programming, in case these problems include known parameters.
But, real-world problems almost invariably include some parameters of which
the true values are unknown. Stochastic programming interprets these unknown
parameters by taking advantage of the fact that probability distributions
governing the data are known or can be estimated.
When a linear programming problem has been completely specified with the
exception of some parameters, which are assumed to be random variables with a
known joint distribution, two types of models are considered in stochastic
programming. These two types of models are the wait-and-see model and the
here-and-now model.
In the first case, the model builder is assumed to be able to wait for the
realization of the random parameters. In the second case, the model builder will
have to decide upon the value of these parameters without knowledge of the
realizations. So, in the here-and-now model some parameters are undetermined,
which makes the feasibility and/or optimality of the solution useless. Because of
this, additional specification is needed to deal with questions of risk and risk
aversion.
The following paragraph will describe some approaches to deal with these
questions of risk and risk aversion.

3.1 Approaches
Given is the following linear programming problem with random parameters in
the constraints:
Minimize
Subject to

cT x
Ax ~ b
Tx ~ h
x 0.

(2),

where the relational symbol ~ denotes =, , or .

Page 17

BMI paper J.M. de Rue

Then assume that the real value of (T , h ) is not known, i.e., it is not known
which instance of the model occurs. Furthermore, assume that the uncertainty is
expressed by a probability distribution, e.g., so-called scenarios:

)}

Pr (T , h ) = T s , h s = ps ,

s = 1,2,...., S .

In addition to this, assume that the probability distribution is known, e.g., by

data, or experts, and that a deterministic linear program is a degenerate case.
By stochastic linear programming it is possible to decide on x here and now,
without knowing the real value of (T , h ) , but only by knowing its probability
distribution. This is done by interpreting Tx h as a goal constraint, which is to
be specified more precisely.
There are several approaches which can be taken:
-

Fat solution
The idea is to replace Tx h by T s x h s , s = 1,2,...., S . This constraint is to
be satisfied in all scenarios. The advantage this approach has is that the
problem is deterministic again. But on the other hand, it is very conservative
and expensive. In addition to this, it results in a lot of constraints, which is
the reason this approach is called the fat solution.

Expected value
_

The idea is to replace Tx h by T x h , with T = s psT s and

_

h = s ps h s . The advantage is again that this results in a deterministic

linear program. But, the risk is not addressed in this model. The constraint
T s x h s only holds for some of the scenarios. This could be partly solved
_

by using more conservative values for T and h , or by applying sensitivity

analysis. But, this would still lead to a poor model of a decision under
uncertainty.
-

Scenario analysis

The idea in scenario analysis is that for every scenario T s , h s , s = 1,2,...., S

the following linear programming problem is solved:
Minimize

cT x

Page 18

BMI paper J.M. de Rue

Ax ~ b
Subject to

T s x ~ hs .
x0

The solutions for the corresponding scenarios are given by x s . An overall

solution is then found by examining the different scenario solutions, i.e.,
what solutions are there and what are the probabilities p s of the
corresponding scenarios occurring. This approach is an improvement over
the expected value approach and the advantage is that each scenario problem
is a linear programming problem. The disadvantage, however, is that flexible
solutions do not show up.
-

Chance constraint
With chance constraints, Tx h is replaced by Pr (Tx h ) for some
prescribed reliability level 1 ,1 , where is to be determined by the
2
problem owner. The advantage this approach has is that the risk is explicitly
taken care of, i.e., risk := Pr{Tx / h}. So the maximum accepted risk is 1 .
The disadvantage this approach has is that a discrete probability distribution
leads to a mixed-integer linear programming problem.

( )

The final approach is called recourse actions. Since recourse actions are seen as
the most important class of stochastic programming, this concept will be
discussed in more detail in the following paragraph.

3.2 Recourse models

Given is the following linear program with random parameters in the
constraints:
Minimize
Subject to

cT x
Ax ~ b
T ( )x ~ h( )
x 0.

(3).

Here Ax ~ b represents m1 inequality constraints. The m random inequality

constraints are represented by T ( )x ~ h( ) , where T ( ) is an m n matrix,

Page 19

BMI paper J.M. de Rue

and h( ) is an m 1 vector. Both the matrix and the vector are dependent on a
random vector r . It is possible that the whole of T and h is random, but
more often than not only a restricted number of matrix and vector elements is
random. We assume that the joint distribution of the random vector is known.
Now consider the linear programming problem given by (3) and suppose this is
a so-called here-and-now problem. This problem needs a decision on x before
the real value of is known, i.e., only the joint distribution of on is
known. Moreover, in recourse models the random constraints of (3) are
reformulated as soft constraints, i.e., violation of the constraints is accepted but
not at any price. To describe how violated constraints are dealt with a secondstage linear programming model is introduced. This second stage model
contains second stage variables y p . The name recourse can now be
explained by the second stage, as its decisions are made after observations of the
value of .
A recourse structure is formally specified by a triple (Y , q,W ) , which is defined
as follows:
Y = {y p : y 0}, which describes the feasible set of recourse
actions y,
q is a 1 p vector of unit recourse costs,
W is an m p matrix, the recourse (technology) matrix.

When the recourse structure is applied to the linear programming problem (3),
the following decision problem can be defined:
Minimize

Subject to

c T x + min q T y
Ax
T ( )x

x X

first-stage
decisions

Wy

]
~b
~ h( ) ,

y Y

second-stage
decisions

first stage
constraints
second stage
constraints

This representation makes clear how the introduction of the recourse structure is
based on a relaxation of the constraints T ( ) ~ h( ) .

Page 20

BMI paper J.M. de Rue

3.3 Modelling aspects

Given a linear programming problem of which some parameters are unknown,
the question is how to transform this problem into a well-specified linear
programming problem with a recourse structure. There are three possibilities to
insert the recourse structure (Y , q,W ) into the linear programming problem.
-

The first possibility is not to model recourse actions. Instead, deviations

from the goals are penalized. This means that the recourse variables only
represent surpluses or shortages with respect to the goals. These recourse
variables are then used in the objective function to penalize these surpluses
or shortages.

The second possibility is to introduce recourse variables which represent

corrective actions. These corrective actions are to be taken after realizations
of if the goals are not reached.

The third possibility is to split the vector of decision variables in the original
linear programming problem in two parts. One set has to be determined
before obtaining , and the remaining set of variables may depend on the
value of . The constraints are also split into two parts. The first part of the
constraints does not include , and the second part of the constraints does.

In the next chapter an example will be given of a linear program which is

transformed into a well-specified linear programming problem with a recourse
structure.

Page 21

BMI paper J.M. de Rue

Page 22

BMI paper J.M. de Rue

4 The model
In this chapter an example will be worked out in which the theory of the
previous chapters will be applied. A linear programming model will be built to
schedule operations with a given duration over multiple days and operating
rooms in such a way that as much of the available time will be used for these
operations. First a simple model will be built, after which the model will be
expanded to deal with multiple operating rooms.
Operations almost always take more or less time than anticipated beforehand by
the model builder. Therefore, recourse actions will be applied to deal with a
random surplus or shortage of the duration of these operations.
For this first model, numerous assumptions have to be made regarding the
values of the parameters such as the duration of an operation, the number of
hours the operating staff works on one day, and more of such assumptions will
be made along the path of building the model.

4.1 Modelling
The problem that has to be modelled is how to plan operations such that the staff
and equipment are being scheduled most efficiently. In other words, if the staff
and equipment are available for eight hours per day, operations have to be
planned such that the staff and equipment are scheduled to be occupied during
these eight hours.
First the decision variable is defined:
xij = starting time of operation j on day i.
The assumption is thus made that the staff is expected to work eight hours per
day. Then assume that there are n operations, that are being planned over m
days, whereby operation j has a duration of dj hours. The decision variable is
equal to zero if operation j is not scheduled on day i and it is equal to or greater
than > 0 if operation j is scheduled on day i. Hereby, one could think of a
starting time for the staff of eight oclock in the morning, which would result in
= 8.

Page 23

BMI paper J.M. de Rue

First, the assumption is made that only one operating room is available to the
hospital. This assumption keeps the model somewhat simpler, but will
eventually be dropped later on. This assumption means that only one operation
at the time can be carried out. To make sure that an operation j that is scheduled
for day i is not scheduled again on a different day, a dummy-variable is needed.
This dummy-variable z ij is a 0-1 variable, which means it will only have the
values 1, if operation j is scheduled on day i, and the value 0 in all other cases.
Now, the following constraints can be formulated:
xij Nz ij ,

i, j ,

xij z ij ,

i, j ,

z
i =1

ij

= 1,

j ,

whereby N has a sufficiently large value, such that it always holds that xij N

i, j and where represents the starting time of the operating staff.

So, if operation j is to be scheduled on day i, the variable z ij will be assigned the
value of 1, which will result in xij 8 and xij N . If operation j is not to be
scheduled on day i, the variable z ij will be assigned the value of 0, which will
result in xij 0 and xij 0 .
Next to this, it has to hold that for every pair of operations j and k, x ij + d j x ik ,
or xik + d k xij i, j and k > j . In other words, if operation j is scheduled to
start before operation k, the starting time of operation k has to be greater or equal
to the starting time of operation j plus the duration of this operation so that
operation k does not start until operation j is finished. In addition to this, this
also holds for the case when operation k is scheduled to start before operation j
respectively.
To make this work, another dummy-variable y jk is needed, which again is a 0-1
variable. This dummy-variable has the value 1 if operation j is scheduled to start
before operation k, and has the value 0 in all other cases. This results in the
following constraints:
x ij + d j xik + M (1 y jk ) ,

i, j and k > j ,

xik + d k xij + My jk ,

i, j and k > j ,

Page 24

BMI paper J.M. de Rue

where M has a sufficiently large value so that xij + d j M holds i, j .

So, if operation j is scheduled to start before operation k, the variable y jk will be
assigned the value of 1, which would result in xij + d j xik and
xik + d k xij + M . If operation j is not scheduled to start before operation k, the
variable y jk will be assigned the value of 0, which would result in xik + d k xij
and xij + d j xik + M .
As is assumed before, the operating staff and equipment is available for eight
hours per day. So, given the duration of the operations, some additional
constraints are needed to make sure no operations are scheduled with a
cumulative duration of over eight hours. To take this into account in the model,
the previously declared variable z ij will be used. This variable will be assigned
the value of 1 when operation j is scheduled on day i and the value of 0 in all
other cases. So, when this variable is multiplied with the duration of the
corresponding operations, the total duration of all operations scheduled on day i
will be known. This leads to the following constraint:
n

z
j =1

ij

d j 8 , i .

This constraint will make sure that the total duration of all operations scheduled
on day i will be eight hours at maximum.
What remains to be done is the objective function. To schedule the operations as
efficient as possible, the difference 8 z ij d j needs to be minimized. This is the
difference between the total duration of all operations planned on day i and the
total available time (eight hours) of the operating staff and equipment. So, to
optimize the schedule, the maximum difference has to be found. In other words,
the biggest difference needs to be as small as possible. The biggest difference is
given by the next expression:
n
n
n

max 8 z1 j d j ,8 z 2 j d j ,...,8 z mj d j .
j =1
j =1
j =1

To minimize this maximum difference, an additional variable q is needed, such

that:

Page 25

BMI paper J.M. de Rue

n
n
n

q = max 8 z1 j d j ,8 z 2 j d j ,...,8 z mj d j .
j =1
j =1
j =1

Now, the linear programming problem becomes:

Minimize

Subject to

q 8 z ij d j , for i = 1,..., m

j =1

xij Nz ij ,

i, j ,

xij z ij ,

i, j ,

z
i =1

ij

= 1,

j ,

x ij + d j xik + M (1 y jk ) ,

i, j and k > j ,

xik + d k xij + My jk ,

i, j and k > j ,

z
j =1

ij

d j 8,

xij 0 ,

y jk {0,1},

z ij {0,1} ,

i ,

i, j ,
j, k ,

i, j .

Suppose the model needs to be extended in order to be able to schedule

operations in multiple operating rooms. Then, the decision variable needs to be
adjusted first so it shows in which operating room which operation is scheduled
on which day. The decision variable is defined as follows:
xijl = starting time of operation j on day i at operating room/location l.
Second, additional constraints are needed to prevent one operation being
scheduled in more than one operating room. To prevent this from happening, the
dummy-variable z ij is used again, but with an additional index. The dummyvariable z ijl will have the value of 1 when operation j is scheduled at operating
room/location l on day i, and will have the value of 0 in all other cases. The
additional constraints that are needed then become the following:
xijl Oz ijl ,

i, j , l ,

xijl z ijl ,

i, j , l ,

Page 26

BMI paper J.M. de Rue

z
i =1 l =1

ijl

= 1,

j ,

where O has a sufficiently large value so that xijl O i, j , l always holds.

The new linear programming problem then becomes:

Minimize

Subject to

q 8 z ijl d j ,

i, l ,

xijl Oz ijl ,

i, j , l ,

xijl z ijl ,

i, j , l ,

j =1

z
i =1 l =1

ijl

= 1,

j ,

xijl + d j xikl + M (1 y jkl ) ,

i, j , l and k > j ,

xikl + d k xijl + My jkl ,

i, j , l and k > j ,

z
j =1

ijl

d j 8,

xijl 0 ,

y jkl {0,1},

zijl {0,1},

i, l ,
i, j , l ,
j , k , l ,

i, j , l .

4.2 Extending the model

The model as built in the last paragraph will optimize a schedule such that
personnel and equipment are planned as efficiently as possible. This model and
resulting schedule are heavily dependent on the parameters, among others the
duration of the operations. These durations will have to be known upon building
the model. The drawback this has is that this is often not the case. Operations
often take more, sometimes less, time than anticipated beforehand.
To be able to deal with the surpluses or shortages in the duration of the
operations, and thus the random factor in these parameters, the model will be
expanded with a recourse structure.

Page 27

BMI paper J.M. de Rue

First, the following additional variables are needed:

j = planned decrease in duration of operation j,

+j = planned increase in duration of operation j,

j = random demand of extra time for operation j,

c +j = unit cost for the increase in duration of operation j,
c j = unit cost for the decrease in duration of operation j.

If an operation runs over time it means that personnel will have to work
overtime. Obviously, this leads to additional costs. If an operation takes less
time than anticipated, personnel is scheduled less efficiently than would have
been possible afterwards. This will also lead to additional costs.
Taking into account the aforementioned additional variables, the objective
function needs to be adapted. The adapted objective function is as follows:
c+ + + c .

Minimize

Next to the objective function, the following constraint also needs to be adapted:
n

z
j =1

ijl

d j + j = 8 j + +j ,

i, l .

Notice that this constraint is a second stage constraint, as described in paragraph

3.2. Lastly, the following constraints need to be added to the model:

j d j ,

j ,

+j , j 0 ,

j .

These two constraints give an upper and a lower bound to the increase and
decrease in duration of operation j.
Suppose that for the operations hold that the random demand of extra time is
assumed to be normally distributed. Then for every operation, the mean and
standard deviation of the random deviation j per operation has to be given. In

Page 28

BMI paper J.M. de Rue

addition to this, the different costs for an increase or decrease in duration need to
be specified.
The linear programming problem with the recourse structure then is as follows:
Minimize

c+ + + c
n

Subject to

z
j =1

ijl

d j + j = 8 j + +j ,

i, l ,

xijl Oz ijl ,

i, j , l ,

xijl z ijl ,

i, j , l ,

z
i =1 l =1

ijl

= 1,

j ,

xijl + d j xikl + M (1 y jkl ) ,

i, j , l and k > j ,

xikl + d k xijl + My jkl ,

i, j , l and k > j ,

xijl 0 ,

i, j , l ,

y jk {0,1},

j , k ,

z ijl {0,1} ,

i, j , l ,

j d j ,

j ,

+j , j 0 ,

j .

Page 29

BMI paper J.M. de Rue

Page 30

BMI paper J.M. de Rue

5 Conclusions
Linear programming is one of the many specializations of Operations Research
and is one of the most applied mathematical methods. The many different
algorithms of linear programming can be used for the optimization of a wide
range of problems.
The example that is worked out in this paper showed that there are drawbacks to
Linear Programming. If there are parameters of which the values are uncertain,
the measure of optimality is questionable. There are several approaches to take
this uncertainty into account in building the model. However, it shows that these
approaches lead to models in which the risk is not taken into account, or to
models that are very restrictive.
We have shown that adding a recourse structure to a well-defined linear
programming problem is relatively easy to do. Furthermore, adding a recourse
structure leads to a model in which the risk is explicitly taken care of. In
addition to this, it leads to a large scale Linear Programming problem. However,
it does have one disadvantage. The disadvantage is that the model may become
too large to solve.

Page 31

BMI paper J.M. de Rue

Page 32

BMI paper J.M. de Rue

6 Bibliography
H. Tijms
Operationele analyse, 2004
H. Tijms, A.A.N. Ridder
Mathematische Programmering, 2002
W.K. Klein Haneveld, M.H. van der Vlerk
Stochastic Programming, 2006
Wikipedia
http://www.wikipedia.org
Stoprog
http://stoprog.org
Google
http://www.google.com
Mathworld
http://www.mathworld.com

Page 33