Many Criteria Optimization and Decision Analysis State of The Art Present Challenges and Future Perspectives Dimo Brockhoff (Editor)
Many Criteria Optimization and Decision Analysis State of The Art Present Challenges and Future Perspectives Dimo Brockhoff (Editor)
https://ebookmeta.com/product/multi-criteria-decision-analysis-
case-studies-in-engineering-and-the-environment-2nd-edition-igor-
linkov/
https://ebookmeta.com/product/government-transparency-state-of-
the-art-and-new-perspectives-gregory-porumbescu/
https://ebookmeta.com/product/breast-mri-state-of-the-art-and-
future-directions-katja-pinker-editor/
https://ebookmeta.com/product/an-archaeology-of-greece-the-
present-state-and-future-scope-of-a-discipline-anthony-m-
snodgrass/
Criminalization of Activism Historical Present and
Future Perspectives 1st Edition Valeria Vegh Weis
Editor
https://ebookmeta.com/product/criminalization-of-activism-
historical-present-and-future-perspectives-1st-edition-valeria-
vegh-weis-editor/
https://ebookmeta.com/product/confronting-corruption-past-
concerns-present-challenges-and-future-strategies-fritz-heimann-
mark-pieth/
https://ebookmeta.com/product/cancer-diagnostics-and-
therapeutics-current-trends-challenges-and-future-perspectives-s-
k-basu/
https://ebookmeta.com/product/national-health-services-of-
western-europe-challenges-reforms-and-future-perspectives-1st-
edition-guido-giarelli-editor/
https://ebookmeta.com/product/the-climate-demon-past-present-and-
future-of-climate-prediction-saravanan/
Natural Computing Series
Series Editors
Thomas Bä ck
Natural Computing Group–LIACS, Leiden University, Leiden, The
Netherlands
Lila Kari
School of Computer Science, University of Waterloo, Waterloo, ON,
Canada
Susan Stepney
Department of Computer Science, University of York, York, UK
Founding Editor
Grzegorz Rozenberg
Leiden University, Leiden, The Netherlands, and, University of Colorado,
Boulder, USA
Scope
covers theory, experiment, and implementations at the intersection of
computation and natural systems. This includes: The Natural
Computing book series
Computation inspired by Nature: Paradigms, algorithms, and
theories inspired by natural phenomena. Examples include cellular
automata, simulated annealing, neural computation, evolutionary
computation, swarm intelligence, and membrane computing.
Computing using Nature-inspired novel substrates: Examples
include biomolecular (DNA) computing, quantum computing,
chemical computing, synthetic biology, soft robotics, and artificial
life.
Computational analysis of Nature: Understanding nature through a
computational lens. Examples include systems biology,
computational neuroscience, quantum information processing.
Editors
Dimo Brockhoff, Michael Emmerich, Boris Naujoks and
Robin Purshouse
Michael Emmerich
Leiden Institute of Advanced Computer Science, Leiden University,
Leiden, The Netherlands
Boris Naujoks
Informatik und Ingenieurwissenschaften, TH Kö ln, Gummersbach,
Germany
Robin Purshouse
Automatic Control and Systems Engineering, University of Sheffield,
Sheffield, UK
The publisher, the authors, and the editors are safe to assume that the
advice and information in this book are believed to be true and accurate
at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the
material contained herein or for any errors or omissions that may have
been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Thanks
Because the book would not have been written without the support of
the Lorentz Center, we warmly thank everybody who was involved in
the MACODA workshop, especially Michelle Grandia-Jonkers, Aimée
Reinards and Henriette Jensenius. To point us to the open space idea
and assisting in using it was a big plus for the workshop and thus for
the success of this book. Many thanks also go to Mio Sugino and Ronan
Nugent at Springer for their support and assistance with the technical
issues. But most of all, we thank all workshop participants and all
authors for their input, ideas and hard work to make this book possible.
Richard Allmendinger
Alliance Manchester Business School, The University of Manchester,
Manchester, UK
Vitor Basto-Fernandes
Instituto Universitá rio de Lisboa (ISCTE-IUL), University Institute of
Lisbon, ISTAR-IUL, Av. das Forças Armadas, Lisboa, Portugal
School of Computer Science and Informatics, Faculty of Technology, De
Montfort University, Leicester, United Kingdom
Leiden Institute of Advanced Computer Science, Faculty of Science,
Leiden University, Leiden, The Netherlands
Mickaël Binois
LJAD, CNRS, Inria, Université Cô te d’Azur, Nice, France
Dimo Brockhoff
Inria, Ecole Polytechnique, IP, Paris, France
Tinkle Chugh
University of Exeter, Exeter, UK
Kalyanmoy Deb
Electrical and Computer Engineering, Michigan State University, East
Lansing, MI, USA
Timo M. Deist
Centrum Wiskunde & Informatica, Amsterdam, The Netherlands
Andre H. Deutz
Leiden University, Leiden, The Netherlands
João A. Duro
University of Sheffield, Sheffield, UK
Michael Emmerich
Leiden University, Leiden, The Netherlands
Jonathan E. Fieldsend
Department of Computer Science, University of Exeter, Exeter, UK
Peter Fleming
Department of Automatic Control and Systems Engineering, The
University of Sheffield, Sheffield, UK
Carlos M. Fonseca
Department of Informatics Engineering, Polo II, University of Coimbra,
CISUC, Coimbra, Portugal
Antonio Gaspar-Cunha
University of Minho, Braga, Portugal
David Gold
Department of Civil and Environmental Engineering, Cornell University,
Ithaca, NY, USA
Andreia P. Guerreiro
INESC-ID, Rua Alves Redol, Lisbon, Portugal
Abderrahmane Habbal
LJAD, UMR 7351, CNRS, Inria, Université Cô te d’Azur, Nice, France
Jussi Hakanen
University of Jyvaskyla, Faculty of Information Technology, Jyvaskyla,
Finland
Dani Irawan
TH Kö ln–University of Applied Sciences, Cologne, Germany
Yaochu Jin
Department of Computer Science, University of Surrey, Guildford,
Surrey, UK
Kathrin Klamroth
University of Wuppertal, Wuppertal, Germany
Joshua Knowles
Alliance Manchester Business School, The University of Manchester,
Manchester, UK
Schlumberger Cambridge Research, Cambridge, UK
Mariapia Marchi
ESTECO SpA, Trieste, Italy
Kaisa Miettinen
Faculty of Information Technology, University of Jyvaskyla, Jyvaskyla,
Finland
Boris Naujoks
TH Kö ln–University of Applied Sciences, Cologne, Gummersbach,
Germany
Yusuke Nojima
Osaka Metropolitan University, Sakai, Osaka, Japan
Daniel C. Oara
University of Sheffield, Sheffield, UK
Akira Oyama
Japan Aerospace Exploration Agency, Sagamihara, Japan
Victor Picheny
Secondmind, Cambridge, UK
Robin Purshouse
University of Sheffield, Sheffield, UK
Alma Rahat
Swansea University, Swansea, UK
Patrick M. Reed
Department of Civil and Environmental Engineering, Cornell University,
Ithaca, NY, USA
Diana Salvador
Instituto Universitá rio de Lisboa (ISCTE-IUL), University Institute of
Lisbon, ISTAR-IUL, Av. das Forças Armadas, Lisboa, Portugal
School of Computer Science and Informatics, Faculty of Technology, De
Montfort University, Leicester, United Kingdom
Leiden Institute of Advanced Computer Science, Faculty of Science,
Leiden University, Leiden, The Netherlands
Hiroyuki Sato
The University of Electro-Communications, Chofu, Tokyo, Japan
Tea Tušar
Jožef Stefan Institute, Ljubljana, Slovenia
Vanessa Volz
modl.ai, Copenhagen, Denmark
Hao Wang
Leiden Institute of Advanced Computer Science, Leiden, CA, The
Netherlands
Yali Wang
Leiden University, Leiden, The Netherlands
Kaifeng Yang
University of Applied Sciences Upper Austria, Hagenberg, Austria
Iryna Yevseyeva
Instituto Universitá rio de Lisboa (ISCTE-IUL), University Institute of
Lisbon, ISTAR-IUL, Av. das Forças Armadas, Lisboa, Portugal
School of Computer Science and Informatics, Faculty of Technology, De
Montfort University, Leicester, United Kingdom
Leiden Institute of Advanced Computer Science, Faculty of Science,
Leiden University, Leiden, The Netherlands
Part I
Key Research Topics
© Springer Nature Switzerland AG 2023
D. Brockhoff et al. (eds.), Many-Criteria Optimization and Decision Analysis, Natural
Computing Series
https://doi.org/10.1007/978-3-031-25263-1_1
1. Introduction to Many-Criteria
Optimization and Decision Analysis
Dimo Brockhoff1 , Michael Emmerich2, Boris Naujoks3 and
Robin Purshouse4
(1) Inria, Ecole Polytechnique, IP, Paris, France
(2) Leiden University, Leiden, The Netherlands
(3) TH Kö ln, Cologne, Germany
(4) University of Sheffield, Sheffield, UK
Dimo Brockhoff
Email: [email protected]
Abstract
Many-objective optimization problems (MaOPs) are problems that
feature four or more objectives, criteria or attributes that must be
considered simultaneously. MaOPs often arise in real-world situations
and the development of algorithms for solving MaOPs has become one
of the hot topics in the field of evolutionary multi-criteria optimization
(EMO). However, much of this energy devoted to MaOP research is
arguably detached from the challenges of, and decision analysis
requirements for, MaOPs. Motivated by this gap, the authors of this
chapter organized a Lorentz Center workshop in 2019 entitled Many-
Criteria Optimization and Decision Analysis—MACODA—bringing
researchers and practitioners together to reflect on the challenges in
many-objective optimization and analysis, and to develop a vision for
the next decade of MACODA research. From the workshop arose the
MACODA book, for which this chapter forms the introduction. The
chapter describes the organizers’ perspectives on the challenges of
MaOP. It introduces the history of MaOP principally from the
perspective of EMO, from where the terminology originated, but
drawing important connections to pre-existing work in the field of
multi-criteria decision-making (MCDM) which was the source or
inspiration for many EMO ideas. The chapter then offers a brief review
of the present state of MACODA research, covering major algorithms,
scalarization approaches, objective-space reduction, order extensions
to Pareto dominance, preference elicitation, wider decision-maker
interaction methods and visualization. In drawing together the vision
for MACODA in 2030, the chapter provides synopses of the unique and
varied contributions that comprise the MACODA book and identifies
further under-explored topics worthy of consideration by researchers
over the next decade and beyond.
1.1 Motivation
Myriad decision problems require the simultaneous consideration of
multiple performance criteria, objectives, or solution attributes.
Oftentimes, there are many such criteria that need to be weighed in the
balance when deciding upon a solution. Several research communities
have been established that seek to provide theory, methods and tools to
support decision-makers in solving problems of this type. The principal
two communities are multiple criteria decision-making (MCDM) and
evolutionary multi-criteria optimization (EMO), although multi-criteria
problem-solving is also a topic in the more general fields of operational
research and optimization theory.
The types of MCDM approaches vary according to whether the
specific solution options to be considered are small in number and
already known, or whether the set of potential candidate solutions is
much larger and yet to be determined. The former situations belong to
the field of multi-criteria decision analysis (MCDA), whilst the latter
belong to the field of multi-objective optimization. Excellent
introductions to each field can be found in the texts of Goodwin and
Wright [45] and Miettinen [66], respectively. The EMO community also
focuses almost exclusively on multi-objective optimization problems—
the main difference from classical MCDM being the use of population-
based randomized search heuristics rather than analytical or
deterministic methods as the instruments used in identifying optimal
candidate solutions. However, over the past two decades, the
distinction between the MCDM and EMO communities has become
purposefully blurred, due to successful initiatives such as a programme
of shared Dagstuhl seminars [9]. Excellent introductions to EMO and
MCDM can be found in the texts of Deb [23] and Coello Coello, Lamont
and Van Veldhuizen [20] as well as Miettinen [66], Ehrgott [30], Steuer
[77], Belton and Stewart [7], Zopounidis and Pardalos [89] and Greco et
al. [46], respectively.
Historically, the EMO community has focused on intensive algorithm
development driven by artificial bi-objective benchmarking problems,
with the aim of identifying high-quality sample-based representations
of Pareto fronts and the associated Pareto sets. A plethora of methods
and tools are now available to support decision-makers in identifying
Pareto fronts for such problems—and these have been demonstrably
useful across a broad range of applications. More recently, as will be
summarized in Sect. 1.2.2 of this chapter, the EMO community began to
recognize that the algorithms developed for bi-objective problems did
not scale well to problems with many more objectives—an important
issue, since many real-world problems possess this characteristic. This
realization has triggered a new phase of intensive algorithm
development on artificial benchmarking problems featuring many
objectives, but using the same conceptual framework as the earlier
methods. Where these research efforts are claiming successful results,
the studies use performance metrics designed for low numbers of
criteria. Since there is no good understanding of how these metrics
work in many-criteria spaces, it is not clear whether the findings from
these now copious studies are meaningful. Whilst this rather
pathological characterization of EMO research is a generalization that
some researchers would undoubtedly raise objection to, it does raise
concerns about the true usefulness of a large corpus of contemporary
EMO research for real-world decision-making.
As four researchers who had been working on a variety of many-
criteria optimization topics over the past decade or more, we reached a
joint view that recent EMO research on many-criteria optimization had
become rather uncoordinated and would benefit from redirecting its
efforts from algorithmic development towards developing a better
understanding of the special context presented by many-objective
optimization problems (MaOPs).1 We argued that, once this new
context is understood, the aims of solving MaOPs have been articulated
and fundamental questions posed by MaOPs have been addressed, the
EMO community would be much better positioned to make a beneficial
contribution to designing methods and tools for MaOPs, including for
real-world problems. To this end, we proposed the Many-Criteria
Optimization and Decision Analysis (MACODA) workshop at the Lorentz
Center in Leiden, the Netherlands in 2019. In this week-long invitation-
only workshop, experts from EMO and other disciplines came together
to discuss and develop new concepts to tackle the challenging problems
in many-criteria optimization and also the wider decision aid systems
needed for practical applications of EMO algorithms. Whilst we
organizers developed an initial set of research questions, participants
got the opportunity to propose alternative directions when registering
for the workshop; these suggestions were the basis for 34 extensive
self-organized “open space” sessions in which the participants
organized themselves around the central topic of the workshop and
many aspects of many-criteria optimization were discussed [49].
During the MACODA workshop, a collective decision was made by
participants to compile and extend the many outcomes of the workshop
within a single book. We are happy that this plan worked, and that we
have collected both the state-of-the-art as well as open research
directions in many-criteria optimization. This book can help the reader
to get a good overview of the field, to serve as a reference and to
provide guidelines of where and how to start contributing to the field.
Each of the following eleven chapters thereby focuses on a certain
aspect of many-criteria optimization, such as benchmarking,
visualization, or theoretical aspects. An ontology of the field rounds up
the book. Although most of the authors who have contributed to the
present book would arguably identify themselves foremost as being
part of the EMO community, the topics covered are relevant to a
broader community of researchers, in particular in the MCDM
community, and the book also features contributions from this field.
Before the specific chapters, we provide the fundamentals and a
broader overview with this introductory chapter. We specifically
explain what is many-criteria optimization in a more formal way in
Sect. 1.2—including a discussion on the additional challenges a MaOP
poses compared to a problem with fewer criteria. Section 1.3 gives a
more detailed overview of what had been achieved collectively by the
time of the MACODA workshop, whilst Sect. 1.4 discusses the open
questions and the organizers’ vision for where the research field should
focus the efforts. Finally, Sect. 1.5 briefly introduces the constitutive
chapters of this book and their main contributions.
Note that this definition is not known in classical research fields such as
operations research or multiple criteria decision-making, where
distinctions might be made between bi-objective ( ) and multi-
objective problems, but not for a larger number of objective
functions. Note further that some authors consider many-objective
problems to be a subset of multi-objective problems, whilst others now
refer to multi-objective as meaning problems with exclusively two or
three objectives, i.e. . This terminological issue has not yet
been resolved by the community and examples of both kinds of use can
be found in the chapters of this book.
In the following, we review the main differences between many-
criteria problems and multi-criteria problems with fewer objective
functions (Sect. 1.2.1) and survey the history of the research field of
many-criteria optimization (Sect. 1.2.2) which arose due to these
additional differences.
1.3.1.1 MOEA/D
Building on the MSOPS concept and earlier MCDM foundations, the use
of decomposition-based approaches was popularized by the Multi-
objective Evolutionary Algorithm based on Decomposition (MOEA/D)
algorithm [87]. The method decomposes the objective space by
defining reference vectors pointing towards the true Pareto front. The
task for solutions is to identify promising, close directions and follow
these as fast as possible towards the Pareto front.
Within MOEA/D different decomposition approaches can be
involved. The most prominent ones are weighted sum, Chebyshev and
boundary intersection. All lead to the original problem being
decomposed into scalar aggregation functions that are then optimized
by an evolutionary algorithm (EA) simultaneously. The EA basically
assigns one incumbent solution to each scalarization function, which
then undergoes reproduction, variation and selection.
This approach provides an easy way to introduce decomposition
into MOEA. Even more decomposition schemes can be incorporated
easily and used interchangeably. In fact, boundary intersection and its
variants became a quasi-standard here. As a consequence of the
approach, general scalar optimization problems need to be addressed
in contrast to optimizing the original MOP directly. This makes fitness
assignment as well as diversity maintenance much easier. A further
advantage is the low computational complexity at each generation,
compared with Pareto and set-based indicator approaches. The
disadvantage of MOEA/D—and other scalarizing methods—is the
requirement to normalize what may be non-commensurable objectives.
The quality of normalization can substantially influence the efficacy of
the algorithm [58].
1.3.1.2 NSGA-III
Comparing MOEA/D to a second prominent approach in many-criteria
optimization, the Non-dominated Sorting Genetic Algorithm-III (NSGA-
III, cf. [24, 57]), there are similarities and differences. One major
similarity is that NSGA-III also decomposes the original problem into
scalar optimization subproblems. However, the way this is performed
as well as the way these subproblems are handled are completely
different.
In general, NSGA-III follows a very similar approach as its precursor
NSGA-II [25]. Only the secondary selection criterion within the
selection operator underwent significant changes. Instead of crowding
distance, a new criterion is applied that is based on reference points
and, thus, on lines from the reference points towards the origin of a
transformed/normalized search space. A hyperplane is defined in each
generation and, regularly, the reference points are equally distributed
on the hyperplane, e.g. by the method proposed by Das and Dennis [22].
Each population member is then associated with the nearest reference
line, i.e. the line connecting a reference point with the adapted origin of
the objective space. The secondary selection criterion then prefers
members that are associated to lines with a smaller number of
members associated with it. If ties exist, first the distance to the
corresponding line is considered. If that does not help, remaining ties
are broken randomly.
1.3.1.3 HypE
The third method regularly applied to MaOPs is a hypervolume
estimation (HypE, cf. [4]) based algorithm similar to the one proposed
by Zitzler and Kü nzli [88] or SMS-EMOA [8]. The underlying algorithm
is NSGA-II with its selection scheme and the secondary ranking
criterion replaced again. This time, the crowding distance of a solution
is replaced by a fitness that is inversely proportional to the lost
hypervolume indicator value when the solution is removed. To cope
with the high computational complexity of the hypervolume
calculations in high-dimensional objective spaces, HypE approximates
these hypervolume contributions using Monte Carlo sampling.
1.3.1.4 Discussion
Comparing results of the three methods mentioned, there is no clear
winner. The results depend on the specific problem, its search and
objective space dimension as well as on the parameterization of the
specific algorithm. Each algorithm has its pros and cons. What they
have in common is that the objective space dimension has a severe
influence on its computational performance. This holds for HypE with
respect to the number of points needed to adequately approximate the
hypervolume, for NSGA-III with respect to the number of reference
points needed to receive an adequate distribution on the considered
hyperplane, and for MOEA/D with respect to the number of
subproblems, i.e. the number of scalar aggregation functions needed to
achieve well-distributed results. Thus, all three methods still undergo
the curse of dimensionality.
Due to the limitations of the presented approaches above, there is
an ongoing challenge to derive new/improved methods for many-
objective optimization. When working on such methods, however, it
will be valuable to pay attention to salient topics such as performance
assessment, visualization, alternative problem formulations and also
some difficulties in reaching convergence that result from the high
dimensionality of the objective space.
1.5 Synopsis
The topics in this book are structured into lead chapters addressing
major themes in MACODA (Part I) and chapters addressing emerging
and more specialized topics (Part II).
1.5.1 Key Topics
In the first contributed chapter of the book, Chap. 2, Peter Fleming and
colleagues introduce the real-world context for many-objective
optimization. In addition to providing examples of real-world
applications—highlighting the efficacy of many-objective techniques—
the authors set their sights on seven key issues for the successful
resolution of real-world many-objective problems. In so doing, the
chapter covers, for example, issues like the problem formulation, the
selection of adequate algorithms, and how to handle uncertainties. All
such issues are addressed taking real-world problems into
consideration and often from the viewpoint of a domain expert, not
necessarily a many-objective optimization expert. This way, the chapter
bridges the gaps from domain to optimization experts as well as from
the application-oriented approach to the more formal methods
following in the remaining chapters of the book.
In Chap. 3, Koen van der Blom and colleagues present the findings
from a questionnaire among practitioners on the characteristics of real-
world problems. Surprisingly, little is known about the general
characteristics of such problems, or their relationships to the prolific
field of artificial benchmarking in the research community. Whilst the
findings will be of interest to algorithm designers faced with single and
bi-objective problems, there are particular points of interest to those
designing for many-objective problems, including a notable prevalence
of positively correlated objectives and the existence of preference
information to guide the search. In just over two fifths of cases, the
respondent was able to identify that the Pareto front was convex—
which presents efficiency opportunities for the configuration of
scalarization functions in decomposition-based algorithms, particularly
in many-objective settings.
In Chap. 4, André Deutz, Michael Emmerich and Yali Wang discuss
extensions of the classical Pareto dominance relation to better cope
with multi-criteria problems. After introducing the basic concepts of
binary relations, (partial) orders, cone orders, and their properties, the
authors discuss the idea of order extensions—a new relation that does
not contradict the previous one but is allowed to introduce additional
dominance relationships among objective vectors. Based on this
definition, a large amount of previously suggested dominance relation
alternatives from the literature are introduced, discussed and
compared. A list of open questions and ideas for future research
directions concludes the chapter.
In Chap. 5, Bekir Afsar and colleagues consider the impact of many
objective problem features on the quality measures, or indicators, that
are often used to assess the performance of MOEAs and sometimes
used within optimizers to direct the search. Moreover, the authors
discuss indicators developed for preference-based MOEAs, where
preference information is incorporated into the indicators. In addition
to the computational demands of computing the indicators in the
presence of many objectives, the authors highlight concerns associated
with the relationship between indicator parameters (e.g. reference
points) and the optimal distribution of an approximation set of a given
size. These findings lead to strong recommendations for authors to
publish full details of indicator parameter settings in reported studies,
particularly for many-objective analyses. The authors also argue the
case for retaining the use of Pareto-compliant indicators for MaOPs.
The chapter further, and uniquely, highlights the many-objective
challenges associated with stochastic objective functions,
accompanying robustness metrics and quality measures for interactive
methods—areas for fruitful investigation in future MACODA research.
In Chap. 6, Vanessa Volz and colleagues give an overview of existing
benchmark suites for multi- and many-objective optimization and
discuss their shortcomings. They also extend on more general aspects
of benchmarking like common pitfalls when it comes to performing
numerical benchmarking experiments and presenting their results.
Finally, they make suggestions on how to avoid those pitfalls by
following guidelines and a concrete checklist for benchmarking
algorithms.
In Chap. 7, Jussi Hakanen and colleagues explore the topic of
visualization of (sets of) objective vectors and solutions within decision
support processes for optimization problems with many objectives.
Visualization is important as is an understanding of the different ways
of utilizing visualization in many-objective applications. Guidance is
provided for choosing and applying visualization techniques including
recommendations from the field of visual analytics. This is illustrated
through a complex real-world decision problem having ten objectives.
In Chap. 8 Andreia Guerreiro, Kathrin Klamroth and Carlos Fonseca
explore theoretical aspects and challenges related to multi- and many-
objective optimization problems. The aim of this chapter is to explore
the connections between set-quality indicators and scalarizations by
considering to what extent indicators can be seen as a generalization of
scalarizations. This serves as a motivation for the discussion of the
corresponding theoretical properties, including monotonicity,
independence/covariance (e.g. with respect to translation and/or
scaling), as well as theoretical aspects related to the specification of
parameters, such as weights and reference points, and possibly its
connection to DM preferences. The authors also discuss indicator-based
subset selection in the context of the classical problems in MCDM
(choice, ranking and sorting), as subset selection may be seen as a
generalization of choice that is different from ranking, itself another
generalization of choice.
In the next chapter, Chap. 9, a special emphasis is put on the
correlation between objectives. Tinkle Chugh and colleagues provide an
insight into solving multi-objective optimization problems by
considering the correlation among objective functions. After six
methods and approaches which have been used to find correlations
from data were reviewed, more light is shed on the conflict and
harmony between objectives and how these can be compared using
correlation measures. The use of correlations in fields such as data
mining, innovization and objective reduction, is described. The chapter
finishes with an overview on problems—benchmarks and one real-
world problem—with correlated objectives and reviews articles which
focus on implicit as well as explicit control of corresponding
correlations.
"Vai huoli se sinut", sanoo hän. "Minusta nähden ei olisi ollut väliä,
vaikka olisit tehnyt lapsesi yksin."
*****
Tällaisina saapuvat puheet Helmin korviin. Ei hän niistä välitä,
tiesihän hän, että puhumaan tultaisiin. Hän vain odottaa joka
hermollaan, eikö puheisiin ala sekaantua muuta, mutta mitään ei
kuulu.
"Tyttö-parka, tyttö-parka!"
Ei muuta.
Eikä hän voi olla avaamatta ovea. "Kyllä se niin on, appi-parka",
sanoo hän rauhallisesti, "ettei niistä teidän Heikin hommista olisi
mitään tullut. Minä sen otin ja sillä hyvä."
Helmi nauttii äärettömästi nähdessään Suontaan ukon siinä
istuvan häkeltyneenä ja kuin nuijalla päähän lyötynä.
"Minä kai saan omassa kodissani olla oven edessä tai takana.
Tämä ei ole Suontaa."
Mutta vaikka Helmin unen laita on niin ja näin, ovat yöt kuitenkin
rauhallisimmat. Ajatukset tosin harhailevat omia teitään ja pelko ja
vavistuskin värisyttävät ruumista. Mutta Heikin selän takana on
lämmin ja turvallinen olla niinkuin olisi kaukana kaikkien ulottuvilta.
Iltaisin, maata mentyä, Heikki vielä lueskelee sanomalehtiä. Hän
nostaa pöytälampun viereensä tuolille, kääntyy kyljelleen ja lehti
alkaa rapista hänen käsissään.
"Kun ei nukuta."
"Hm."
Vaan entä jos Heikki saa tietää ja kun hän saa tietää…
*****
Ei hän nyyhkytä eikä edes huokaile, hän vain luulee niin. Hän vain
katselee kirjaimia, joiden takaa ruskeat, uhkaavat silmät kalpeitten
kasvojen keskeltä iskeytyvät häneen.
"Ei, Helmi, et sinä niin huono ollut, olit vain heikko ja pelkäsit
liiaksi… Katso, minä tulen ja vien sinut mukanani…"
Ei helvettiin, ei…
*****
Aurinko paistaa heleästi, lintujen laulu soi, ulkona on kevät.
Puheen sorinaa kuuluu pihamaalta, joku helistelee avaimia.
Nikolai, älä katso enää. Kuulun sinulle enkä voi kuulua. Unohda ja
ellet voi unohtaa, niin opi vihaamaan. Sinä kärsisit meistä kuitenkin
vähimmän. Ja sitäpaitsi sinä olet mies.
Ettetkö voi vihata? Ah, minä en ole sinun rakkautesi arvoinen enkä
muidenkaan. Sinä teet minulle suurimman palveluksen, jos vihaat
minua, minun on silloin helpompi.
"Hyvä Niku!
"Hyvin, hyvin kai siinä lopuksi käy", vastaa isäntä. "Niin se ainakin
sanoi, tohtori."
"Mistä?"
*****
"Enkös mä sanonut?"
*****
Taas hän on pitkän aikaa vaiti eikä voi ajatella. Kun hän on
virkeämpi, on hän melkein kiitollinen siitä, että on niin avuton,
ajatukseton ja voimaton. Hänelle saattaa nyt, hänen voimatta
kohottaa sormeaankaan, tehdä mitä ikinä tahtoo.