Many Criteria Optimization and

Decision Analysis State of the Art

Present Challenges and Future
Perspectives Dimo Brockhoff (Editor)
Visit to download the full and correct content document:
https://ebookmeta.com/product/many-criteria-optimization-and-decision-analysis-state
-of-the-art-present-challenges-and-future-perspectives-dimo-brockhoff-editor/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Multi-criteria decision analysis : case studies in

engineering and the environment 2nd Edition Igor Linkov

https://ebookmeta.com/product/multi-criteria-decision-analysis-
case-studies-in-engineering-and-the-environment-2nd-edition-igor-
linkov/

Government Transparency State of the Art and New

Perspectives Gregory Porumbescu

https://ebookmeta.com/product/government-transparency-state-of-
the-art-and-new-perspectives-gregory-porumbescu/

Breast MRI: State of the Art and Future Directions

Katja Pinker (Editor)

https://ebookmeta.com/product/breast-mri-state-of-the-art-and-
future-directions-katja-pinker-editor/

An Archaeology of Greece The Present State and Future

Scope of a Discipline Anthony M. Snodgrass

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/

Confronting Corruption Past Concerns Present Challenges

and Future Strategies Fritz Heimann Mark Pieth

https://ebookmeta.com/product/confronting-corruption-past-
concerns-present-challenges-and-future-strategies-fritz-heimann-
mark-pieth/

Cancer Diagnostics and Therapeutics Current Trends

Challenges and Future Perspectives S.K. Basu

https://ebookmeta.com/product/cancer-diagnostics-and-
therapeutics-current-trends-challenges-and-future-perspectives-s-
k-basu/

National Health Services of Western Europe: Challenges,

Reforms and Future Perspectives 1st Edition Guido
Giarelli (Editor)

https://ebookmeta.com/product/national-health-services-of-
western-europe-challenges-reforms-and-future-perspectives-1st-
edition-guido-giarelli-editor/

The Climate Demon Past Present and Future of Climate

Prediction Saravanan

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

Many-Criteria Optimization and

Decision Analysis
State-of-the-Art, Present Challenges, and Future
Perspectives
Editors
Dimo Brockhoff
Inria Saclay-Île-de-France and CMAP, É cole Polytechnique, Palaiseau,
France

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

ISSN 1619-7127 e-ISSN 2627-6461

Natural Computing Series
ISBN 978-3-031-25262-4 e-ISBN 978-3-031-25263-1
https://doi.org/10.1007/978-3-031-25263-1

© Springer Nature Switzerland AG 2023

This work is subject to copyright. All rights are reserved by the

Publisher, whether the whole or part of the material is concerned,
specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other
physical way, and transmission or information storage and retrieval,
electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks,

service marks, etc. in this publication does not imply, even in the
absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general
use.

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.

This Springer imprint is published by the registered company Springer

Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham,
Switzerland
Preface
Optimization problems must be tackled in many areas of our modern
lives. More often than not, we have to optimize several conflicting
criteria simultaneously: performance and cost, risk and return, or
quality, cost and environmental issues. The availability of fast
computing resources, more complex models and paradigm shifts such
as towards sustainable solutions in recent years results in more and
more of those criteria in real-world optimization formulations. The
field of many-criteria optimization aims at developing optimization
algorithms for such kind of problems and at shedding light at the
mathematical foundations of such many-criteria problems.

Why This Book?

Starting from the appearance of the term many-criteria or many-
objective optimization back in 2002, the field of many-criteria
optimization has become one of the most active subfields of multi-
criteria optimization with many independent research groups working
on various aspects. In our opinion, previous research on many-criteria
optimization was uncoordinated and often rather incremental and the
aspects that strongly distinguish optimization with a very small
number of objectives (typically two or three) from that with larger
numbers of objective functions were often neglected. For example,
more and more algorithms for many-criteria optimization have been
proposed in recent years without addressing fundamental issues such
as proper performance assessment/benchmarking and theoretical
aspects. Hence, we were convinced that it was important to organize a
workshop around the topic of many-criteria optimization to change
this. Finally, in September 2019, the ‘Many-Criteria Optimisation and
Decision Analysis (MACODA)’ workshop was held at the Lorentz Center
in Leiden, The Netherlands, jointly co-organized by the four of us.
Fortunately, we were able to gather more than 50 leading experts and
young researchers on MACODA and related fields of science for a week-
long workshop to discuss the current status of MACODA, to initiate
collaborations, and, more importantly, to propose a joint, coordinated
research agenda for the years to come.
 The original goal of the workshop was to compile a comprehensive
list of important topics to work on and to define the possible path of the
research field as a whole in a joint journal publication by the workshop
participants. As suggested by the Lorentz Center staff, we thereby
followed the idea of an ‘open space’ environment that allowed the
participants to discuss freely the many different aspects of MACODA
without a pre-specified schedule. Soon into the workshop days with the
open space arrangement, it became apparent that a single article was
not sufficient to summarize the workshop output. Instead of the
original plan to write a single joint journal article, the participants
instead resolved to write this book, to be edited by the MACODA
workshop organizers. It should not only serve as a summary of the
state of the art in many-criteria optimization and of the workshop
discussions in particular but also provide a starting point for future
contributions to the advancements of the field. To address the latter
aspect, the contributing chapters explicitly discuss important open
research questions.

How Did We Organize the Writing?

Based on the most important open questions, identified at the
workshop, the editors approached a few lead authors to organize the
writing of five key chapters of this book. Those lead authors joined
forces with other authors, interested in these topics to write the
chapters. Right after the choice of the main topics of the book, we asked
every participant in the workshop to contribute additional book
chapters of interest. Also, other experts in the research field, who have
not been able to participate in the workshop, have been approached to
give the widest possible view on the field to date. The reviewing and
proofreading of the single chapters involved authors of related chapters
to increase consistency and reduce repetitions throughout the book
and to not forget to mention the most important aspects of many-
criteria optimization. Further reviewers, who did not participate in the
original workshop, gave additional feedback. We are very grateful to the
following reviewers for the time they spent in helping ensure the
scientific rigour of the MACODA book: Richard Allmendinger, Vitor
Basto-Fernandes, Jü rgen Branke, Tinkle Chugh, Carlos Coello Coello,
Antó nio Gaspar Cunha, Joã o Duro, Jonathan Fieldsend, Bogdan Filipič,
Peter Fleming, Carlos M. Fonseca, Hisao Ishibuchi, Yaochu Jin, Pascal
Kerschke, Kathrin Klamroth, Joshua Knowles, Rodica Lung, Frank
Neumann, Alma Rahat, Gü nter Rudolph, Pradyumn Shukla, Ricardo
Takahashi, Tea Tušar, Hao Wang, Margaret Wieczek and Iryna
Yevseyeva.

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.

A Peek into the Book

We eventually have 13 chapters to offer, including an ontology, plus an
appendix with a glossary. We start with a general introduction and two
chapters about real-world applications, followed by chapters on
fundamental aspects of many-criteria optimization, namely on order
relations, quality measures, benchmarking, visualization and theory
before more specialized chapters on correlated objectives,
heterogeneous objectives, Bayesian optimization and game theory.
We wish you a nice read and hope that you will be able to apply the
information of the book in your own application or that you even
contribute to the advancements of the field of many-criteria
optimization as intended by the book.
Boris Naujoks
Dimo Brockhoff
Michael Emmerich
Robin Purshouse
Gummersbach, Germany
 Palaiseau, France
Leiden, The Netherlands
Sheffield, UK
Contents
Part I Key Research Topics
1 Introduction to Many-Criteria Optimization and Decision
Analysis
Dimo Brockhoff, Michael Emmerich, Boris Naujoks and
Robin Purshouse
2 Key Issues in Real-World Applications of Many-Objective
Optimisation and Decision Analysis
Kalyanmoy Deb, Peter Fleming, Yaochu Jin, Kaisa Miettinen and
Patrick M. Reed
3 Identifying Properties of Real-World Optimisation Problems
Through a Questionnaire
Koen van der Blom, Timo M. Deist, Vanessa Volz, Mariapia Marchi,
Yusuke Nojima, Boris Naujoks, Akira Oyama and Tea Tušar
4 Many-Criteria Dominance Relations
Andre H. Deutz, Michael Emmerich and Yali Wang
5 Many-Objective Quality Measures
Bekir Afsar, Jonathan E. Fieldsend, Andreia P. Guerreiro,
Kaisa Miettinen, Sebastian Rojas Gonzalez and Hiroyuki Sato
6 Benchmarking
Vanessa Volz, Dani Irawan, Koen van der Blom and Boris Naujoks
7 Visualisation for Decision Support in Many-Objective
Optimisation:​State-of-the-art, Guidance and Future Directions
Jussi Hakanen, David Gold, Kaisa Miettinen and Patrick M. Reed
8 Theoretical Aspects of Subset Selection in Multi-Objective
Optimisation
Andreia P. Guerreiro, Kathrin Klamroth and Carlos M. Fonseca
9 Identifying Correlations in Understanding and Solving Many-
Objective Optimisation Problems
T. Chugh, A. Gaspar-Cunha, A. H. Deutz, J. A. Duro, D. C. Oara and
A. Rahat
Part II Emerging Topics
10 Bayesian Optimization
Hao Wang and Kaifeng Yang
11 A Game Theoretic Perspective on Bayesian Many-Objective
Optimization
Mickaël Binois, Abderrahmane Habbal and Victor Picheny
12 Heterogeneous Objectives:​State-of-the-Art and Future
Research
Richard Allmendinger and Joshua Knowles
13 Many-Criteria Optimisation and Decision Analysis Ontology and
Knowledge Management
Vitor Basto-Fernandes, Diana Salvador, Iryna Yevseyeva and
Michael Emmerich
Glossary
Contributors
Bekir Afsar
Faculty of Information Technology, University of Jyvaskyla, Jyvaskyla,
Finland

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

Koen van der Blom

Leiden University, Leiden, The Netherlands
Sorbonne Université, CNRS, LIP6, Paris, France
Leiden Institute of Advanced Computer Science, Leiden, Netherlands

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

Sebastian Rojas Gonzalez

Department of Information Technology, Ghent University, Gent, Belgium
Data Science Institute, University of Hasselt, Diepenbeek, Belgium

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.

1.2 What is Many-Criteria Optimization?

Many practical optimization problems involve multiple, conflicting
objective functions or criteria such as risk, sustainability and return in
portfolio optimization, performance and energy efficiency in chip
design or cost, reliability, comfort, fuel consumption and the reduction
of tailpipe emissions in the design of cars or planes. Other application
areas include urban planning, staff scheduling, embedded systems
design, forest and environmental management, medical therapy design
and drug discovery [19]. Note that both methodological and application
works tend to use the terms ‘criterion’ and ‘objective’ synonymously
and interchangeably, which we will continue to do throughout this
chapter and is also the case across other chapters in the book.
From an abstract point of view, the evaluation of a system, policy, or
schedule can be viewed as the evaluation of a mathematical function
that represents the performance of the system. Formally, we are
interested in minimizing a vector-valued objective function

We denote as the search or decision space and the image

as the objective space of the m-criteria problem.
Depending on the context, we will use the terms criteria, objectives and
objective functions interchangeably with a preference for the term
objective function(s) when we address the actual above-defined
mathematical function(s). When , the problem has a single
objective, whilst denotes a multi-objective problem. Problems
with objectives are sometimes known as bi-objective problems.
Common to multi-criteria problems is the typical conflict among the
objective functions, meaning that not a single optimal solution exists,
but that without the knowledge of preferences from decision-maker(s)
(DMs), we aim at finding a set of trade-off solutions. Depending on
when a decision-maker is involved in the optimization process, we talk
about a priori, a posteriori, or interactive approaches. In a priori
approaches, a single-criterion optimization problem is created before
the optimization process begins, based on the preferences of the
decision-maker. In a posteriori approaches, the decision-maker bases
her decisions on the knowledge gained from an optimization process
that identifies a set of candidate solutions, offering trade-offs between
the different criteria. In an interactive approach, search and decision-
making are intertwined before a final solution is obtained—refining the
search iteratively towards solutions interesting to the decision-maker.
Note that the above formalization does not include constraints
explicitly but that, in many practical problems, constraints have to be
dealt with in addition to the trade-offs—as is also the case for other
typical challenges of optimization problems such as large search space
dimension, mixed-type variables, nonlinearities, ill-conditioning or
multi-modality.
With the introduced conflict and the corresponding trade-offs, basic
properties of optimization problems change when moving from single-
criteria to multi-criteria problems. Most prominently, non-dominance
can occur, i.e. two solutions do not need to be necessarily comparable
anymore.
We say that one solution dominates another solution
if for all objective functions, x is not worse than y and if for at least one
objective, x is better. We denote such a dominance of x over y as .
For minimization, this means if and only if
and . We call
the set of solutions that are not dominated by any other
solution in the efficient or Pareto(-optimal) set and its image
in objective space the efficient frontier or Pareto(-optimal) front. For
additional notations related to multi-criteria problems used later in this
book, we refer to the Glossary 13.4.
Interestingly, there is another important change when the number
of objective functions rises beyond three. In this case, our spatial
cognition and intuition, by which we are used to visualizing and
comparing solutions, does not support us well anymore. Besides, in
higher-dimensional spaces many, often counter-intuitive, phenomena
occur and the qualitative and quantitative characteristics of dominance
orders and related concepts change quite drastically. Following the
historical terminology (see also Sect. 1.2.2), we therefore denote

many-criteria optimization problems are problems with four or

more objective functions.

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.

1.2.1 Salient Challenges in Many-Criteria

Optimization
Formally, many-criteria optimization problems differ from classical 2-
and 3-criteria and even single-criteria optimization problems only in
their increased number of objective functions. Practically, however, the
higher number of objectives results in additional properties and
challenges that are not prevalent or even existing when the number of
objective functions is lower.
The Curse(s) of Dimensionality
Coined by Richard E. Bellman [6], the term “curse of
dimensionality” describes all difficulties in high-dimensional spaces,
related to optimization, sampling, sparse data, etc. In our context of
many-objective problems, we refer to this term with respect to the
objective space. The number of solutions to cover the Pareto front or
Pareto set with a given precision typically increases exponentially with
the number of objectives. The Pareto dominance relation provides less
and less guidance towards the Pareto front the larger the number of
objectives gets (because in uniformly distributed point sets, solutions
tend to become mutually non-dominated at an exponential rate,
see [55]). As a consequence, locally, the area of improvement is
decreasing exponentially. More precisely, it was shown that the
difficulty to obtain dominating solutions using isotropic mutation
decreases exponentially fast [2, 60].
All those aspects of the curse of dimensionality are of high
importance in algorithm design and all related analyses of many-
criteria data (for example in performance assessment) alike.
Computational Aspects
Related to the curse of dimensionality, the optimization of many-
objective problems also often means a high computation time. Despite
the obvious calculation of more objective function values, several
algorithmic components of multi-objective optimization algorithms
have a runtime that grows with the number of objectives, sometimes
even exponentially. Prominent examples are non-dominated
sorting [25] with a polynomially bounded runtime in the number of
objectives or the calculation of the hypervolume indicator which is even
#P hard in the number of objectives [12] and other hypervolume-
related questions [83].
Correlation Structure Among the Objectives
With more objective functions, also more structure can be observed
among the objectives. Two objective functions can be everything from
negatively correlated, through uncorrelated, to positively correlated (up
to the point that the resulting problem is actually, in mathematical
terms, a single-objective one). But with more objective functions,
correlations as well as redundant objectives can be observed. In
addition, it gets more and more likely that not all objective functions
are equally important in practice when more and more objectives are
introduced in the problem formulation. To efficiently optimize, such
structural properties of the objectives might be identified and
explained by a solver.
 Cognitive Challenges
Finally, novel cognitive challenges appear with many-objective
problems—especially related to visualization. Approximations of
Pareto sets and any solution sets in general are still easily visualizable
with two or three objectives, but starting with four objectives, objective
vectors can only be visualized indirectly (by projections, linear, or
nonlinear mappings, etc. embedding them into the real 2D or 3D space).
See for example the work of Tušar [82, 83] for an extensive overview of
visualization techniques for multi-objective optimization.
Another cognitive challenge with many objective functions is the
articulation of preferences. Already difficult when only two objective
functions are involved, preference articulation becomes even harder
when more objectives have to be taken into account.
Related Scientific Challenges
The above mentioned rather generic challenges of many-objective
problems relate to more concrete research areas in many-objective
optimization such as
Algorithm Design: algorithmic strategies that work well with two or
three objectives might not anymore work well in higher dimensions
due to the curse of dimensionality, computational aspects of internal
procedures, or the presence of correlation structures (which could be
exploited). Hence, new concepts need to be introduced to optimize
efficiently in high-dimensional objective spaces.
Performance Assessment and Benchmarking: performance
assessment and algorithm benchmarking becomes more challenging
not only because of the absence of valid and efficient performance
indicators but also because the choice of algorithmic and quality
indicator parameters becomes more crucial for the algorithms’
behaviour, for example, when choosing (a few) reference points(s) in
the higher-dimensional space. The lack of approaches for easy
visualization of large, many-dimensional sets plays an important role
here as well. Eventually, the objective space dimension as a new
adjustable parameter asks for scalability assessment of algorithms in
both search and objective space. Naturally, more objective functions,
thus, mean more computation time in benchmarking experiments
and also the inclusion of decision-makers in the performance
assessment gets more involved. Automated tuning of algorithms,
 automated benchmarking and parallelization aspects play an
important role here.
Preference Articulation/User Interaction: once we go beyond the
optimization process itself by putting a decision-maker or a group of
decision-makers into the loop (for example after the optimization or
within interactive methods), additional challenges appear in terms of
visualization and the formulation of preferences. Navigation in high-
dimensional objective spaces becomes more difficult than with two
or three objectives. We cannot see all information at once anymore
and additional preference information to focus on small parts of a
huge Pareto front is needed in the decision-making process. The
articulation of preferences itself also becomes more complex with
more objective functions and might even involve more different
decision-makers than if only a few objectives are present (for
example in a multidisciplinary approach). Finally, the larger amount
of incomparabilities among solutions affects the decision finding
process as well: not only is it more difficult to find agreements among
the decision-makers but it also makes the involvement of the
decision-maker(s) during the optimization process more important.
The latter is done in order to increase the probability of actual
improvements within the optimization.
Objective Space Reduction: As a direct consequence of the presence
of many objective functions and potential correlations among them,
the question arises whether all defined objective functions are
necessary to describe the optimization problem at hand or whether
some of the objectives are redundant and can therefore be removed
in a reformulation of the problem with smaller objective space
dimension. This not only has obvious applications in the visualization
of solution sets after the optimization but also in reducing the actual
optimization time of algorithms.
A good portion of the mentioned research topics will be discussed
in more detail in the other chapters of this book. On an abstract level,
we will detail later in Sect. 1.3 where the research field is now in terms
of (some of) the mentioned topics as well as where we would like to be
in the mid-term future in Sect. 1.4. But before, let us review the
beginnings of the many-objective optimization research field.
1.2.2 History of Many-Criteria Optimization
1.2.2.1 Origins of the Terminology
To our knowledge, the terminology many-criteria and many-objective
were both introduced by Farina and Amato [34] in the title and abstract
of a paper in the Proceedings of the 2002 Annual Meeting of the North
American Fuzzy Information Processing Society that proposed
replacing the conventional Pareto dominance and optimality definitions
with fuzzy alternatives. These authors did not subsequently define
either term in their paper, but the concluding sentence, “The Pareto
optimality definition is unsatisfactory when considering multi-criteria
search problems with more than three objectives.” implied that the
terms related to problems with greater than three objectives. This
definition was clarified in the authors’ subsequent journal paper [35].
The terminology was first adopted by the EMO field in 2003 by
Purshouse and Fleming [69, 72, 73]. These authors were motivated by
the observation that methodological papers in the field were typically
developed in the context of bi-objective (and, occasionally, three-
objective) benchmark problems for a posteriori optimization, whilst
applications of these methods were typically grappling with problems
with more than three criteria. The term evolutionary many-objective
optimization was introduced to emphasize this distinction, being
exemplified as problems featuring four to 20 objectives [37], and to
highlight the need for more methodological research set in the context
of many-objective problems. Sometime after this, ‘many-objective’
became terminology that was particular to the EMO field. It was not
adopted by other research communities in the field of multi-objective
optimization, such as MCDM. Indeed, many of the methods
subsequently developed for many-objective problems in the EMO field
have reused, been inspired by, or simply rediscovered ideas already
developed within MCDM for handling large numbers of objectives. The
historical review that follows focuses on developments in the EMO
community, but aims to draws attention to those conceptual links with
MCDM.

1.2.2.2 Pre-terminology Works: 1995–2003

It is important to note that early applications work in the EMO field, in
selected cases, featured many-objective problem formulations. The rich
survey of applications in [20] identifies a good number of such works
across a range of domains. Here, we focus on journal examples that aim
to identify trade-off surfaces (rather than adopting a single cost
function based on a priori weighting). In an early example, Flynn and
Sherman presented a helicopter panel design problem with four
objectives [38]. Chipperfield and Fleming presented a gas turbine
engine control problem with nine objectives [18]. Hinchliffe and
colleagues developed a four objective formulation for modelling
chemical processes [50]. Yu considered a four objective formulation of
radiotherapy treatment planning [86]. Reynolds and Ford worked with
a 10 objective calibration problem for ecology models. Kumar and
Rockett framed classification problems using seven objectives [61].
Fonseca and Fleming had already identified potential issues with
Pareto-based approaches by this point, demonstrating the high
proportion of non-dominated solutions in their analysis of a seven-
objective gas turbine engine control problem [39]. These authors
proposed the use of decision-maker preferences to better refine the
ranking induced by Pareto dominance. The result was the seminal
preferability relation, that explicitly unified a range of European school
MCDM preference approaches into a single operator [40]. Alternative
operators to modify the dominance relation were also subsequently
proposed by other researchers [29, 35, 78].
In his seminal book introducing MOEAs [23], Deb highlighted
scalability in the objective space as one of the “salient issues” for the
research field. Reinforcing Fonseca and Fleming’s applied analysis, Deb
identified the issue of a large increase in the number of non-dominated
solutions in an approximation set for steadily increasing numbers of
objectives.

... as the number of objective functions increases, more and more

solutions tend to lie in the first non-dominated front ... this may
cause a difficulty to most multi-objective evolutionary
algorithms.
A major step towards progress in this area came with the subsequent
publication of the so-called DTLZ problems by Deb and colleagues [28].
Crucially, the DTLZ problems were scalable in the number of objectives
(although this scalability also increased the number of decision
variables and was therefore subject to potential confounding).
Empirical analysis of the then state-of-the-art EMO algorithms using
the DTLZ test suite was first published by Khare and colleagues in 2003
[59]. This was a broad study across algorithms and problems and did
not identify mechanisms responsible for variations in performance. In a
narrower, but more focused study, Purshouse and Fleming [72] took a
component-level approach [62] and a specific test problem (DTLZ2) to
explore scalability issues for different configurations of Pareto
optimizer. This study identified that:
the sweet spot of algorithm configurations (controlling the balance
between exploration and exploitation) that produce effective
convergence to the Pareto front reduced as the number of objectives
increased;
conventional parameterizations for variation operators that had
arisen from bi-objective studies were no longer located in the sweet
spot for problems with greater numbers of objectives;
poor convergence (little or no better than random search) was
associated with inability to find new solutions that dominate existing
solutions (related to the notion of dominance resistance [53]);
reverse convergence, in which the optimizer performed worse than a
random search, was driven by over-promotion of diversity in the
presence of dominance resistance (since diverse solutions can be
found in poor performing areas of the objective space).
Purshouse and Fleming suggested, but did not demonstrate, that
these findings were generalizable to the wider class of EMO algorithms
that relied on dominance and diversity mechanisms to drive the search
and to other many-objective problems that exhibited conflict across
four or more objectives.

1.2.2.3 Early Post-terminology Works: 2003–2008

Subsequently, a number of key works took up the challenge of
evolutionary many-objective optimization. Hughes developed
important insights in a many-objective empirical analysis of his Multiple
Single Objective Pareto Sampling (MSOPS) [51] algorithm [52]. Hughes
examined two decomposition approaches based on the “traditional”
weighted min-max (i.e. Chebyshev) scalarization function from the
MCDM community: (i) individual single objective optimization runs for
each scalarization function; (ii) a population-based approach (MSOPS),
in which all the scalarizations are considered together. Hughes found
that both these scalarization approaches outperformed a Pareto-based
optimizer (NSGA-II, [25]) on four-objective and seven-objective
problems. In a prelude to the success of the later MOEA/D algorithm
[87], the study showed that considering the scalarizations
simultaneously also provided a more effective search than a set of
isolated scalarization runs.
These findings were reinforced in further empirical studies by
Corne and Knowles [21] and Wagner et al. [84]. The former study
considered five, 10, 15 and 20-objective instances of travelling
salesman problems [21], finding that a scalarization function based on
the weighted average ranking for each objective performed better than
Pareto-based approaches. The latter study extended Hughes’ earlier
findings to DTLZ problems and also identified that indicator-based
algorithms (using set-based fitness metrics) also showed promising
performance on many-objective problems—although the high time-
complexity of computing indicator values or contributions when the
number of objectives is large should be noted.
Whilst most of the works in evolutionary many-objective
optimization have assumed, either explicitly or implicitly, that the
objectives were all in conflict with each other, other problem
formulations are possible. Purshouse and Fleming proposed a typology
consisting of three types of relationship in many-objective problems
[71]: conflict, harmony (in which objectives worked in synergy) and
independence (in which the objectives were unrelated). The concept of
harmony drew explicitly on existing ideas in the MCDM field on what
was variously called ‘redundant’ [42], ‘supportive’ [17] or ‘nonessential’
[41] objectives—objectives for which the Pareto set is unchanged when
those objectives are removed from the problem formulation.
Whilst independence has received little attention [70], harmony has
been considered more extensively by the EMO community, alongside
parallel developments in MCDM [64, 81]. The early focus here was on
methods for reducing the dimensionality of the objective space by
automatically identifying and removing harmonious (or redundant)
objectives during the optimization process. Brockhoff and Zitzler [14,
16] proposed a mathematical framework for objective reduction, based
on the extent to which the dominance relations considering the full set
of objectives would be preserved. The authors also proposed
algorithms to identify minimum objective subsets under specific
preservation conditions. In a separate approach, Ló pez Jaimes and
colleagues used feature detection to identify candidates for objective
reduction [56], also based on minimum subset definitions. Deb and
Saxena investigated principal component analysis as a method for
reducing the dimension of the objective space [26], and later provided
an extension for nonlinear geometries [74, 75]. In [15], the authors
investigated how the reduction of the number of objectives can reduce
the internal computation time of a hypervolume-based algorithm.
Further fundamental studies considered the effect of removing or
adding objectives to a problem formulation [13, 48].
The culmination of this early phase of many-criteria optimization
research was the review paper by Ishibuchi and colleagues [55].
Combining new experiments with reviews of the existing literature,
these authors identified that scalarization approaches and Pareto
dominance relations modified to include preference information both
offered promise for many-criteria optimization going forward. The
authors also identified further work with the MCDM community as a
promising route to progress.

1.3 Where are We Now? MACODA by the Time of

the 2019 Lorentz Center Workshop
The number of publications in many-objective optimization has
increased rapidly over the decade since the Ishibuchi review [55], and
several ideas and techniques that have already been published deserve
further attention. This section provides a brief overview of state-of-the-
art methods and salient topics.

1.3.1 Algorithmic Aspects

In the early period of EMO research, the differences to many-objective
optimization problems were almost ignored and it was thought that
similar algorithms can be applied to problems featuring both up to
three and more than three objectives. The development,
implementation, and testing of methods tailored towards problems
with four or more objectives began in the early 2010s. Now, in the very
early days of the 2020s, we have several methods well established.
Three of these deserve a closer look.

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.

1.3.2 Salient Topics

Looking at the existing literature on many-objective optimization the
question of how to design algorithms for many-objective optimization
or how to generalize existing algorithmic concepts for multi-objective
optimization has been a major topic. Already the previously mentioned
MOEAs show that the motivations for adaptations of existing
algorithms are diverse and relate to various topics, such as diversity
measures and performance indicator properties in high-dimensional
spaces. Moreover, there are several proposals for algorithms that
exploit structures which are not prevalent in low-dimensional solution
spaces—first and foremost algorithms that make use of correlations
between objectives in order to reduce the problem complexity. There
are alternative approaches available which might be worth taking a
look, cf. also [55].
1.3.2.1 Use of Scalarization Functions
One major point in two of the methods described above is the use of
scalarization functions like the weighted sum. Here, the methods are
rather open with respect to which scalarization function to use, and
how these are parameterized and integrated. This leaves room for
several improvements.
A huge advantage of the use of scalarization functions is their
computational efficiency. They are rather simple and easy to calculate
yielding a rather small computational effort to do so. However, if the
number of scalarization functions becomes too large, this advantage
diminishes. This number of functions can depend on the population
size of the appointed algorithm or the number of objectives of the
(many-objective) optimization problem. In both cases, this number may
become too large to efficiently use the corresponding algorithmic
approach. More research is needed to identify the best way to integrate
such functions in the overall algorithmic context, which scalarizations
to be used for which optimization problems, and how to parameterize
the functions adequately.
In addition, the apparent empirical success of decomposition-based
methods for MaOPs, in comparison to Pareto-based methods, has been
questioned by some researchers. Giagkiozis and Fleming noted the
equivalence of Chebyshev scalarization to Pareto dominance in terms of
convergence trajectories, and argued that the performance of
decomposition approaches based on Chebyshev scalarization could not
be reasoned in terms of improved probability of identifying better
solutions [43]. Further analysis by Giagkiozis et al. suggested that the
performance advantage may arise from the stability offered by
retaining fixed scalarization directions during the optimization process
(also cautioning against the use of adaptive weighting schemes) [44]. A
recent analysis by Takahashi has demonstrated convergence problems
with Chebyshev-based decomposition at a finite distance from the
Pareto front [79].
If the Pareto front is convex, weighted-sum scalarization functions
can be used, which provide effective comparability between solutions
and consequent convergence efficiencies. More generally, it is
important to consider the relationship between objective space and
scalarization function geometries in achieving success with
decomposition-based methods [1].

1.3.2.2 Objective Space Reduction

The use of dimensionality reduction techniques in many-objective
optimization has been an active topic of research in the past ten years
and there are various techniques available now. We refer to the
discussion in Sect. 1.2.2.3. Still challenges might remain when the
number of objective functions is very large and when there are different
types of relationships between objectives that, besides correlation, have
to be considered.

1.3.2.3 Order Extensions to Pareto Dominance

The study of alternative order relations to the Pareto order is another
promising research direction in many-objective optimization. Order
relations, such as the Pareto dominance order, can be viewed as special
cases of binary relations, i.e. a set of ordered pairs. An extension of an
order relation is an order relation that includes all ordered pairs of the
original order relation and includes additional ordered pairs. As a
consequence, it reduces the number of indifferent and/or incomparable
solution pairs.
The number of incomparable and, thus, also the number of non-
dominated solutions using the Pareto dominance relation tends to be
very high in many-criteria optimization. This effect has been analysed
in much detail in a recent paper by Allmendinger et al. [2], and it is
inherent to problems with many objectives, in particular, if there is no
strong correlation between them. Hence, the introduction of order
extensions to the Pareto dominance relation has been proposed as a
remedy to the undesirable situation of having too many non-dominated
solutions to choose from.
The Pareto order and the total order introduced by scalarization can
be viewed as two extreme cases, in the sense that, in the first case,
there is a maximum number of incomparable solutions (given the
general problem definition) and, in the second case, all solutions are
mutually comparable. In other words, the Pareto dominance can be
seen as a minimalistic order relation that requires no additional
preferences than a specification of which objectives are to be
minimized and to be maximized. On the other side, scalarizations such
as a weighted sum of objectives, introduce a total order on the set of
candidate solutions, where every pair of two solutions is comparable.
In many-criteria optimization, extensions of the Pareto dominance
relations that fall in between these extremes and form partial, non-total
orders, have been proposed recently. Examples are k-optimality [35], l-
optimality [90], and winning score [65]. To give an example, the k-
optimality considers in how many objectives a solution is better than
another solution.
A common idea is also to use the information about the trade-off
between objectives when comparing two solutions. This can be done by
using dominance cones that expand the Pareto dominance cone [85].
Although initial analyses of these alternative orderings and comparing
them to the Pareto order exist, work on comparing these alternative
order relations with each other is missing. Chapter 4 of this book is
devoted to an in-depth discussion of these various partial order
relations.

1.3.2.4 Preference Elicitation and Interactive

Methods
Besides using general ways of extending orderings it is also possible to
ask the decision-maker for additional preference information. This can
be done a priori or in a progressive interactive way, e.g. by asking for
additional information at certain moments of the optimization process.
For many-objective optimization, both approaches have been
exemplified. Preference elicitation is defined as the process of asking
the decision-maker simple questions from which the algorithm infers
utility functions or preference relations. A common method for
preference elicitation is ordinal regression [10], where the decision-
maker is asked for pairwise comparisons of a subset of all pairs in
and based on this information a compatible utility function is
inferred that can be used to rank all solutions in . Other ways of
preference elicitation can be to ask for reference points, target regions
on the Pareto front, or information on the relative importance of
objective functions. These methods will be discussed in the next
subsection on interactive methods.

1.3.2.5 Navigation and Interactive Methods

It was mentioned before in this introduction, that when facing the
problem of visualizing high-dimensional manifolds, our spatial
cognition does not very well support us. Instead of computing and
trying to visualize the entire non-dominated set, navigation methods
focus on gradual changes that occur when moving from one solution to
another. By repeatedly asking the decision-maker for preferable moves,
they enable the decision-maker to navigate along a path through higher
dimensional objective space.
Classical Pareto navigation methods navigate across a set of non-
dominated solutions, where each move entails losses in some objectives
and gains in others. The so-called Nautilus method tries to avoid
trading off: starting from a very bad, dominated solution, they follow a
path of gradual improvement until they hit a solution on the Pareto
front [67]. This way, the decision-maker has the positive experience
that the solutions are always improving in some objectives without
losses in other objectives. The decision-maker has control on which
objectives to improve.
In general, navigation methods are well suited for many-criteria
optimization and by working in a path-oriented manner they do not
have to face the curse-of-dimensionality. Within the MCDM community,
there has been some work on navigation methods [31, 32], but the
methods remain underexploited by the plethora of approaches that
seek to cover the Pareto set in its entirety.
In interactive methods [66, 68], the decision-maker is not only
providing additional information but s/he is also learning about
possibilities, limitations and/or trade-offs and in the light of new
information can adjust preferences. Reference vectors and their
components (called aspiration levels) are a common way to formulate
preferences—representing desirable objective function values.
Reference points can also be processed and visualized in many-
objective optimization. Interactive multi-objective optimization has
been an active topic of research (see, e.g. [68] and references therein).
1.3.2.6 Visualization Tools
In addition, to better support the user in recognizing the structure of
the non-dominated set, recently, methods from visual analytics that can
deal with high-dimensional multi-variate data sets have been discussed
[47, 82]. For a very large number of objectives, even graph theoretical
methods that visualize networks of (compatible, conflicting) objective
functions have been proposed. Last but not least, the so-called Pareto
scanner is an approach where constraints or weights can be defined on
objective function values by using sliders, and this enables the fast
interactive exploration of the set of non-dominated solutions [5, 80].
Another direction might be to describe the set of non-dominated
solutions by human understandable rule sets that describe trade-offs,
as it is proposed for instance in so-called ‘innovization’ techniques [27].

1.4 What Remains to be Done? A Vision for

MACODA in 2030
In this introduction, we have so far covered the history and current
state of the art in MACODA. But what remains to be done? When setting
up an agenda for research several salient topics are to be addressed.
First and foremost, in order to proceed in a scientific way, besides a
collection of ideas, that is perhaps already available, we require means
to measure progress in implementing these ideas. In this context, it will
be paramount to define scalable performance indicators. Scalability
should be achieved in two ways: firstly, in terms of computational
resources, which rules out some attractive ideas for performance
indicators for low dimensional spaces such as the hypervolume
indicator, and also coverage of manifolds will be more difficult to
achieve, due to the curse of dimensionality. Secondly, the indicators
should assess the performance of solution sets in a way that is
meaningful in many-objective optimization. Even when the goal
remains to compute an evenly spaced coverage of the non-dominated
set, often performance indicators suffer from biases that become more
prominent in high dimensions and we are only beginning to understand
how to counteract these biases [3, 54, 76]. There are, however, diversity
indicators that favour uniformly distributed sets and that can be scaled
in dimension, such as the Riesz s-Energy [33]. Future work could
investigate how they can be best integrated into MaOP frameworks.
The visualization and interactive exploration of solution sets in high
dimensions is another active topic of research: how can we make
techniques of visual analytics more adapted to the specific structure of
Pareto non-dominated point sets or non-dominated point sets of other
orders proposed for many-criteria optimization? How to assess the
performance of interactive and path-oriented methods, such as
navigation methods is largely uncharted terrain. Perhaps it requires a
radical shift of mindset to assess performance of interactive algorithms,
and first ideas in this direction come from the recent efforts to create
artificial decision-makers [63]. One might also consider the
minimization of cognitive load, such as number of questions,
complexity of questions and number of interaction steps [11, 36].
Once performance metrics are established and agreed upon,
algorithm performance can be measured and improved. Also, the
creation of benchmarks that resemble characteristics of real-world
problems and typical difficulties in many-objective optimization will
then become an important topic. Sub-classes of many-objective
optimization may become independent research areas, such as
problems with a large number of homogeneous objectives or with a
hierarchical structure.
As compared to classical multi-criteria optimization with a small
number of objectives, the amount of different concepts and varieties of
general solution approaches is much larger in many-criteria
optimization. It makes it a challenge to establish a terminology and to
define sub-fields for research in an ordered manner. One attempt to
proceed into this direction is to define an ontology of concepts and link
this ontology to the existing work. This book makes a first attempt in
this direction by initiating the MyCODA ontology, as a curated effort to
define taxonomies and sort existing work.

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.

1.5.2 Emerging Topics

In Chap. 10, authors Hao Wang and Kaifeng Yang provide an overview
of Bayesian optimization generalizations to multi-objective and many-
objective optimization problems. These modern methods provide a
blend of machine learning and optimization technology and are
distinguished by acquisition functions. A list of real-world applications
is also provided, which is mainly on problems with costly or time-
expensive function approximations. These play an important role in
Another random document with
no related content on Scribd:
 *****

Kun Helmi tulee kotiinsa, menee hän suoraan äitinsä kamariin.

"Menemme Heikin kanssa lähiaikoina vihille", sanoo hän tylysti.

Äiti ei käänny häneen päinkään, katselee vain johonkin pisteeseen

vastapäisellä seinällä.

"Niin ettette puuhaile mitään häitä", jatkaa tytär samaan tapaan

kuin äskenkin.

Saaren emäntä ei vieläkään vastaa mitään. Huulet vain vetäytyvät

tiukemmin kiinni.

"Kuuluttamisestakin on jo huoli pidetty", lopettaa Helmi ja kääntyy

lähteäkseen.

Nyt Saaren emäntä rävähdyttää silmäluomiaan.

"Vai huoli se sinut", sanoo hän. "Minusta nähden ei olisi ollut väliä,
vaikka olisit tehnyt lapsesi yksin."

"Miten niin?" kysyy Helmi kuivasti.

"Eikös sen isäkin ollut yksin tehty?"

"Ei", vastaa tytär hyvin sävyisästi. "Kaksihan sellaisiin hommiin

aina tarvitaan. Pitäisihän äidin se tietää."

Kääntää selkänsä ja menee

XX

Pitäjäläiset ovat odotelleet häitä, joita Saaressa kuuluu valmistellun,

mutta niitä ei kuulukaan. Sensijaan saavat he jonakin päivänä tietää,
että Saaren tytär ja Suontaan poika ovat hiljaisuudessa
pistääntyneet vihillä ja että Suontaan Heikki on muuttanut Saareen.

Yleinen ällistys. Mitä tämä nyt on ja mitä siinä on takana? Pitäjän

isännät ja emännät heräävät kuin unesta ja hierovat silmiään.

Vihdoin saadaan selvityskin, Helmin asiat eivät ole oikein

mallillaan.

Ällistys kasvaa yhä suuremmaksi. Siitäpä asiasta ei ole halaistua

tavua kuulunut ja mahtaneeko siinä edes olla perääkään. Helmi!
Joka aina on ollut vähin niinkuin enkelin kuva.

Mutta — ja nyt puheen sävy muuttuu — joskin nyt näin on käynyt,

niin kylläpähän Heikin niinkuin Helminkin kunnia näkyy sen
kestäneen. Itse ovat saaneet vastata ja hyvin ovat näyttäneet
asiansa järjestäneen. Vaikka eihän sitä, sen puolesta, mitenkään
olisi uskonut…

*****
 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.

Joulu on jo mennyt vanhoine tunnelmineen ja päivät alkavat

valjeta. Miten tämä joulun aika oikein on mennyt, sitä ei Helmi voi
käsittää, jos hän sitä oikein muistaakaan. Jotakin siinä on ollut
vanhaa, turvallista ja kodikasta, mutta siihen oli koko ajan
sekaantunut aavistusta peloittavasta, joka armottomana pilkisti esiin
alkavien arkien takaa. Ei mitään tietoa Nikulta. Helmi melkein iloitsi
ajatuksesta, että Niku ehkä muiden takia jo oli hänet unohtanut. Niin
olisi helpompi ja paras. Mutta ehkä isä oli vain yksinkertaisesti kortin
tai kirjeen hävittänyt. Sillä Niku ei voinut häntä unohtaa, hän vain teki
työtä ja uurasti. Suuri rakkaus ei puhkea valitteleviksi sanoiksi. Jos
sen on kärsittävä, niin se kärsii hiljaisesti.

Saaren Helmi havahtuu. Mitä hän nyt tässä taas istuu

mietiskelemässä.

Heikki tulla tömistelee sisälle suurine saappaineen ja tyytyväisine,

vilpittömine kasvoineen. Mutta ovelta hän jo kääntyy takaisin ja palaa
uudelleen porstuaan harjaamaan pois liiat lumet. Hänen liikkumis- ja
esiintymistavassaan on itsetietoisuutta ja varmuutta, niinkuin hän
aina olisi ollut talossa, eikä Helmi saata olla sitä itsekseen
ihailematta. Vaan kun ajatus siitä, kuinka olisi voinut olla, tunkeutuu
hänen mieleensä, karkoittaa hän sen väkivoimalla pois. Jotakin hän
silloin menettää itsestään, mutta sen kai onkin mentävä.

Ei hän myöskään rupea muistelemaan, kuinka tämä muutos on

tapahtunut. Ja mitä muistelemista siinä onkaan: hän oli siihen jo
ennakolta niin väkevästi totuttautunut ja se oli tapahtunut hiljaa ja
huomaamatta, aivan kuin unta nähden. Hän seisoo rinnakkain Heikin
kanssa papin virkahuoneessa. Ulkona on jo hämärä ja hämärä on
huonekin, lämmin, pehmoinen ja hämärä. Siinä tulevat sitten sisään
papin rouva, vanhahko, pyöreä nainen, ja hänen hiljainen, vanha
renkinsä. Yhä vaan hiljaisempaa ja hämärämpää. Helmi pelkää
itsekseen, että jokohan se sytyttää valot, mutta pappi siinä vain
kävelee muutaman kerran edes takaisin, sanoo harvakseen jonkun
hiljaisen sanan, joka hukkuu nurkkien pimeyteen, pysähtyy lopuksi
heidän eteensä ja huokaa: "Niin no…" Siten se käy, hiljaa, niinkuin
varkain, hän vastaa jotakin tahdottomasti ja koneellisesti, ja kun
tullaan ulos, ollaan vihitty aviopari.

Sen kummempaa ei ollut kotonakaan. Isä istuu tavallisella

paikallaan ja panee lehden sivuun, kun he tulevat. Juodaan kuppi
kahvia niinkuin ainakin, kun tullaan kauempaa, ja Heikki
valmistautuu lähtemään kotiinsa. Saattaahan hän huomenna tulla
hakemaan Helmin. "Niin sitten."

Mutta Saaren Juhani miettii. "Minnepä tästä enää lähdet", sanoo

hän,… "tarvitaan kyllä tässäkin."

Enempää hän ei sanonut ja siihen se sai jäädä. Eikä Saaren

isäntä millään tavalla sanomiaan perustellut. Hän muuttaa emännän
kanssa eteläpäätyyn, siellä on kolme kamaria, jotka sopivat aivan
parahiksi. Tässä jää tilaa yllin kyllin. "Joo, niin se on. Hyvää yötä."
Niin alkavat päivät kulua. Heti huomenista saa Suontaan Heikki
seurata mukana ja käydä töihin. On ihmeellistä, miten ihminen näin
yhtäkkiä voikin perehtyä ventovieraaseen paikkaan. Vaikka onhan
kyllä niin, että paikkahan se on Suontaakin. Ehkäpä perehtyminen ja
sitä seuraava toiminta onkin veressä.

Emännän puheille ei Saaren isäntä tytärtään eikä vävyään vie.

 "Kun se nyt kävi noin", sanoo hän ja katselee muualle.

Hän puristaa tapansa mukaan huulensa yhteen, mutta Helmiä

vihlaisee, kun hän näkee, miten Heikki syyllisyydentuntoisena
punastuu ja miten hän on aikeissa sanoa jotakin, muttei saa sanaa
suuhunsa. Vasta isän mentyä Heikki kuin vihoissaan itselleen
mutisee:

"Kyllä se olikin kurjaa, ettei sitä ajoissa huomattu."

Ja Helmi kuulee tämän, mutta hän ei anna kasvojensa värähtää

eikä päästä kyyneleitä silmiinsä, ennenkuin Heikkikin on ulkosalla.
Hän vain huokaa, mutta se huokaus tulee raskaasta rinnasta ja on
aitoa. Ah, kunpa senkin oppisi hillitsemään ja oppisi pian!

Kuitenkin Helmin on tavattava äitinsä ja hän on siihen

tapaamiseen valmistautunutkin. Miksikä enää paaduttaa itseänsä, on
vain nöyryydellä ja tyynesti otettava vastaan se, minkä on
ansainnutkin. Sieltä voi tulla tylyä ja ankaraa, mutta mitä sieltä muuta
voisikaan tulla. Äiti on ankara ja äiti on rehellinen, mutta suurinta
onnettomuutta hän tuskin kykenee tekemään vielä suuremmaksi. Ei
suinkaan hän suin päin syökse tytärtään turmioon ja häpeään. Helmi
pysähtyy tähän ja puree hampaansa yhteen.

Silloin hän näkeekin äitinsä edessään. Hän on menossa

sisähuoneisiin, mutta toisen kamarin ovella hän peräytyy. Ne kasvot,
jotka häneen katsovat, kellertävät ja kalpeat, ovat äidin kasvot
eivätkä kuitenkaan hänen. "Tule sisään vaan!" kuuluu ääni sanovan.
Mutta Helmi ei liikahda paikaltaan. Ääni on matala ja masentunut ja
tulee kuin maan alta. Silmät, jotka häneen katsovat ovat
menettäneet välkkeensä ja niistä katselee vain autius ja väsymys.
Käsi lepää valkeana ja verettömänä tuolin käsinojalla. "Tulin tähän
hiukan istahtamaan", sanoo äiti taas hiljaisesti, "ainainen
makaaminenkin väsyttää. Älä ollenkaan pelkää, vaan tule
peremmälle."

Ja kun tytär astuu lähemmäksi, sivelee äiti hänen käsivarttaan ja

sanoo hellästi:

"Tyttö-parka, tyttö-parka!"

Ei muuta.

"Mitä sinä nyyhkit?" kysyy Heikki tuvassa.

"Minä tapasin äidin", vastaa Helmi.

"Yhä vain julmistunut? Kyllä se oli surkeaa…"

Mutta Helmi ei vastaa, vaan vaipuu yhä hillittömämpään itkuun.

Kun hän viimein nostaa katseensa, on miehenkin silmissä vesiä.

Tämä kaikki on jo takanapäin, tosin vielä aivan lähellä, mutta yhä

hän vain istuu tässä ja ajattelee sitä ja muutakin — häntä, joka on
kaukana. Kirjoissa puhutaan, että voidaan vetää paksu, musta viiva
menneen yli. Eikö hän voi sitä tehdä? Joskus tuntuu, niinkuin hän
olisi sen tehnyt, yhtä usein, ettei voi sitä tehdä. Olihan hänelläkin
omatunto ja ilman sitäkin, olihan hän rakastanut.

Heikki on jo tarkastanut, ettei tuvassa ole väkeä ja kaatanut

itselleen kahvia.

"Helmi", sanoo hän sitten, istuutuen vaimonsa viereen, "minä olen

kuullut, että monet ovat miettimällä tulleet hulluiksi."

"Mutta enhän minä miettinyt. Istuin tässä vain ja odottelin sinua."

 "Istuit vain ja odottelit minua!" kertasi mies melkein äkäisesti. "Sinä
olisit voinut puuhata jotakin pientä sen ohessa. Monella on suurempi
suru kuin sinulla, jolla ei oikeastaan enää olekaan mitään surua, eikä
heidänkään sovi ruveta sitä hautomaan. Sitä tehdessään he
kuolisivat nälkään. Joo, mutta kun he tekevät työtä, niin surukin
lieventyy."

"Puhutko tätä itsestäsi vai ovatko muut sen sinulle sanoneet?"

"En minä tiedä, mutta pirukaan ei saa tehtyä tekemättömäksi ja

kyllä elämä lopultakin on otettava sellaisenaan."

Mutta hän lievensi heti sanojensa vaikutusta, tarttui vaimoaan

vyötäisiin ja kuiskasi hellästi toruen:

"Istut siinä silmät haileina ja joskus vetisteletkin. Ei sellainen sovi.

Ison talon emäntä. Saattavat piiatkin pian nähdä."
XXI

Päivät ovat Helmille vaikeimmat. Askareet antavat tilaa ajatuksille ja

on aikoja, jolloin hän ei voi ajatuksiaan käskeä. Silloin hänestä
tuntuu kuin väki katselisi häntä erikoisella tavalla, hän säikkyy ja
luulee nurkista huonon omantunnon katselevan itseään. Mutta on
toisia aikoja, pitkiäkin, jolloin uhma palaa entistäkin jäykempänä.
Niinkuin kerrankin, kun hän kuulee Suontaan äijän oven takana
vinkuvan jollekin toiselle vieraalle, jonka kanssa ovat sattumalta
jääneet kahden:

"Se pelkäsi, se meidän poika, vuosikausia, että saako se tätä

emäntäänsä vai eikö saa… Minä vain kiusasin ja yllytin, että kyllä
sinä kans olet mies, sinäkin… Vaan näkyy se poika osaavan asiansa
järjestää niin, että saakin…"

"Voi, hyvä Luoja", ajattelee Helmi, "kyllä tunnut olevan hyvä

sinäkin."

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.

"Äitiisi tullut, näemmä", yrittää hän kääntää leikiksi, "oven takana

leikkipuheita kuuntelet."

"Minä kai saan omassa kodissani olla oven edessä tai takana.
Tämä ei ole Suontaa."

Suontaan ukko hätääntyy silminnähtävästi. Hän ei enää yritäkään

mitään.
Hän lähtee käydä köpittämään niinkuin on huvikseen tullutkin, ja
vasta
myöhemmin Helmi huomaa, että olisihan hänet voinut kyydityttääkin
Suontaahan.

Tällaiset puuskat pitävät jonkun aikaa pään pystyssä, mutta pian

on kaikki taas entisellään.

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.

"Luenko mä sulle jotakin ääneen?"

"Anna olla. Lue vaan itse."

Joskus Helmi menee puoliuneen ennenkuin Heikki alkaa nukkua.

Kun hän havahtuu, on huoneessa jo pimeää. Toinen hengittää
rauhallisesti ja raskaasti ja liikahtelee unissaan, sanoo välistä jonkun
katkonaisen sanankin. Helmi ei voi olla tunnustelematta hänen
kasvojaan ja Heikki herää.

"No, mitä sinä?" kysyy hän unissaan.

"En mitään, minä vain valvon."

"Ummista silmäsi ja nuku."

"Kun ei nukuta."

"Hm."

Mies nukahtaa taas. Ihosta huokuu lämpöä ja omituinen lemu.

Helmi tulee ajatelleeksi, että ne kai ne juuri ovatkin, jotka tuovat tuon
turvallisuudentunteen, melkein onnea lähentelevän. Ja hän makaa
taas pitkät kotvat katsellen ikkunaverhon läpi kuvastuvan kuutamon
leikkiä vastapäisellä seinällä.

Rakastaako hän Heikkiä vai rakastuuko hän häneen vähitellen,

mutta varmasti. Helmi ei uskalla tätä asiaa ajatella. Toiset kasvot, eri
näköiset, mutta yhtä luotettavat, astuvat silloin rinnalle ja hän alkaa
vavista. Mutta hän oppii päivä päivältä yhä enemmän mieheensä
luottamaan, ja kun hän on poissa, tuntee hän itsensä turvattomaksi
ja yksinäiseksi ja kaipaa häntä.

Vaan entä jos Heikki saa tietää ja kun hän saa tietää…

Helmin täytyy pimeässäkin sulkea silmänsä eikä hän pääse

mielikuvansa loppuun. Hän painautuu kuin turvaa etsien mieheensä
kiinni. Silloin hän tulee ajatelleeksi, että kaikki lopuksi on Jumalan
kädessä ja ettei hän ole muuta kuin suuri ja kauhistuttava syntinen,
joka ei uskalla rikostansa tunnustaa, vaan päivä päivältä paaduttaa
itsensä.

Ja sillä hetkellä hän miestään kohden tuntee tunteen, jossa on

rakkauttakin, mutta enemmän sentään kunnioitusta.
XXII

Nikolai kirjoittaa aivan yksinkertaisesti, että nyt on koti kunnossa ja

rahaakin on sikäli säästössä, ettei ensi hädässä tule pulaa. Eihän se
koti tietenkään niin erinomainen ole, jollaiseen Helmi on tottunut,
mutta kyllä siinä reilusti asuu. Sitten hän kysyy, miksei Helmi ole
pitkiin aikoihin kirjoittanut, — vaikka ehkä niin onkin ollut parempi.
Mutta viisainta on, ettei Helmi nyt vastaa ollenkaan, vaan lähtee
liikkeelle, keritäänhän sitten puhella asioista, jotka ovat menneet.
Nythän, Nikun laskujen mukaan, alkaa Helmin aikakin jo olla
käsissä. Lopussa Niku valittaa, että Helmi näin on saanut odotella,
mutta hyvä on, että viimeinkin päästään yhteen, ettei enää tarvitse
olla peloissaan ja epävarmuudessa eikä enää koskaan erota.

*****

Sellainen oli kirje, jotenkin lyhyt ja sanoiltaan vähäinen.

Maa viheriöitsee jo, koivut ovat hiirenkorvalla ja väki kevättöissä.

Aurinko paistaa kirkkaana ja keltaisena ja äsken kulki Heikki pihan
ylitse jyvämakasiiniin antamaan rengeille siemeniä.

Avara tupa on hulveillaan valoa ja tomuhiutaleet karkeloivat

auringon valossa. Hiillos pesässä on hiipumassa, sekin pitäisi saada
palamaan, mutta joku piioista tullee pian.

Helmi, Saaren Helmi, Suontaan Heikin emäntä, istuu

pöydänpäärahilla kädet sylissään. Ja käsissä on eloton kirje, jossa
puolitoista sivua on täynnä suuria, mustia kirjaimia. Eivätkö nuo, nuo
samanlaiset ja saman käden piirtelemät kirjeet ennen uhkuneet
valoa, lämpöä ja rakkautta. Nyt ne ovat siinä mustina, totisina ja
vaativina. Aivan niinkuin jokaisen kirjaimen takaa tuijottelisi silmä,
kysyvä, vaativa ja uhkaava silmä.

Suontaan Heikin emäntä tahtoisi laskea kirjeen kokoon, mutta käsi

ei halua tehdä tehtäväänsä. Ja silmät tuijottavat herkeämättä
kankeita, totisia rivejä.

"Voi, Nikolai, voi…"

Ei hän nyyhkytä eikä edes huokaile, hän vain luulee niin. Hän vain
katselee kirjaimia, joiden takaa ruskeat, uhkaavat silmät kalpeitten
kasvojen keskeltä iskeytyvät häneen.

Mies on rakastanut ja tehnyt työtä, hänen vartensa on ehkä käynyt

kumaraksi ja hänen hiuksensa harvenneet. Ja hän on uskonut ja
luottanut itseensä ja Helmiin. Yötkin on hän muuttanut päiviksi, jotta
Helmin olisi hyvä ja jotta Helmi tulisi onnelliseksi. Mies on pitänyt
lupauksensa, nyt olisi sitä myöten valmista. Lähde! Ei mitään eleitä,
ei mitään vetoomisia, — näin minä nyt vain olen koettanut laittaa.
Autereiset, onnelliset päivät ovat takana, niin myös lemmenkuumat
yöt, joita silloin seurasi kaihoisa, itkettävä aamu. Niin, ja mies on
lupauksensa täyttänyt, täsmälleen niinkuin pitikin. Hän, Helmi, on
myöskin vannonut jotakin.

"Voi, Niku… rakas, näin sen nyt piti käydä…"

 Mutta Helmi ei sano mitään, ei hiiskahdakaan, ja jos niin olisikin,
niin kuka häntä voisi kuulla. Ja jos kuulisikin, niin entäpä sitten, —
hän, kaksinkertainen valapatto… Kuinka yksinkertaista sentään, kun
nyt jälkeenpäin kuvittelee, olisi ollut pitää sanansa.

Ruskeat silmät kirjainten takana kadottavat surunvoittoisuutensa,

kysymys ja sitä seuraava uhkaus tulevat yhä pelottavammiksi.

"Nikolai, Nikolai, minä olin huono ihminen ja minä pelkäsin… Anna

anteeksi…"

Kirkastuuko tuo katse, tuleeko se lähemmäksi.

"Ei, Helmi, et sinä niin huono ollut, olit vain heikko ja pelkäsit
liiaksi… Katso, minä tulen ja vien sinut mukanani…"

Helmin silmäluomet värähtävät. Hyvä Jumala, ei, se ei saa

tapahtua, ei vaikka kuolema tulisi. Viimeinen villitys tulisi
ensimmäistä pahemmaksi. Ei kukaan takaa, vaikka veri vuotaisi,
viaton veri.

Saaren Helmi, Suontaan Heikin emäntä, vavahtaa ja ikäänkuin

herää: eikö olisi parasta, että hän katoaisi, menisi pois…

"Niin, helvettiin jo ajoissa, varmuuden vuoksi…"

Kuka puhui? Ei kukaan. Äänestä kuulosti kuin olisi se ollut äidin

ääni, mutta hän on toisessa päässä rakennusta, ei pääse
liikkumaankaan.

Ei helvettiin, ei…

*****
 Aurinko paistaa heleästi, lintujen laulu soi, ulkona on kevät.
Puheen sorinaa kuuluu pihamaalta, joku helistelee avaimia.

Ja pitkät hetket kuluvat.

Katsovatko nuo silmät taas häneen kysyvästi, kiiruhtavasti ja

rakastavasti.

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.

Ettetkö voi! Katso, piruhan minussa asuu ja minä voin opettaa

sinut vihaamaankin. Näin vain!

Ja Helmi kirjoittaa Sarkan Nikolaille näin:

"Hyvä Niku!

Kyllähän itse ymmärrät, että se kaikki silloinen oli vain

lapsellisuutta ja suurta erehdystä. Ja ymmärrät senkin, etteihän
meistä olisi voinut mitään tulla. Minulla on jo mies, joka sopii
minulle paremmin. Meidät vihittiin jo kuukausi sitten. Kyllä sinäkin
sopivamman löydät. Älä kiusaa itseäsi enää äläkä kirjoita minulle.
Hyvästi. Älä muistele minua pahalla.

Nyt, Niku, nyt on elämäsi reilu ja vapaa. Unohda tai vihaa, se on

paras, mitä voit Helmillesi tehdä ja Helmi kiittää sinua siitä
haudassaankin."
 Suontaan Heikin emäntä on entisessä yliskamarissaan. Ei hän
muista, miten hän sinne on tullut eikä sitäkään, missä välissä hän on
saanut kirjeensä kirjoitetuksi.

Niin kuluu pitkiä hetkiä tiedottomuudessa, kunnes hän herää

lattialta.
XXIII

Saaren rakennus on tosin tilava ja laaja, ääni ei suljettujen ovien

takaa kantaudu talon keskustasta päätyihin, mutta siitä huolimatta
liikutaan talossa hiiviskellen ja varpaillaan. Huoneet on pesty ja
puhdistettu, ilmassa on lysoolin- tai karboolin hajua.

"Kuinka siellä nyt on?" kysyy Saaren vanhaemäntä mieheltään.

"Hyvin, hyvin kai siinä lopuksi käy", vastaa isäntä. "Niin se ainakin
sanoi, tohtori."

Saaren vanhaemäntä nousisi kovin kernaasti tytärtään katsomaan,

muttei pääse vuoteeltaan. Hän huokaa hiljaa ja ristii kätensä. Viimein
hän kysyy:

"Mitä luulet ihmisten sanovan?"

"Mistä?"

"Tästä kun se tuli näin aikaisin."

Isäntä sytytti piippunsa ja mietti hetken.

"Sitäkö se sitten ihmisiin kuuluu", vastasi hän vihdoin.

 Niinkuin ei kuulukaan. Ei ole heiltä pyytämistä ellei antamistakaan.
Hoidelkoot he vain omat asiansa.

*****

Myöskin palvelusväki kuiskuttelee touhunsa ja hiiviskelemisensä

lomassa, kuiskuttelee, mutta puhuu jo ääneenkin navetassa ja
pihamaalla.

"Enkös mä sanonut?"

"Että mitä sanoit?"

"Sitä vain että vahinko se tulee viisaallekin."

"Niinkuin ei muilla nyt olisi silmiä päässä niinkuin sinullakin."

"Silmiä päässä… tietenkin. Mutta sanopa, olisitko ensi sanomalla

uskonut, että Helmi — emäntä…"

"Helmi kun Helmi ja emäntä kun emäntä. Ihminen se on niinkuin

muutkin ja syyhyttää sillä niinkuin muillakin."

"Anna olla ne meikäläisten syyhytykset, mutta sanos, missä

välissä nuori isäntä ja Helmi sellaisissa muhinoissa! Oletkos nähnyt
tai kuullut?"

"Kolusihan se nuori isäntä siellä syystalvesta…"

"Jaa, se on syystalvesta. Mutta se ei passaa rätinkeihin."

"Eikös se vanha isäntä sitten karhannut sitä pois Helmin lutin

ovelta syksyllä."
 "Taisi. Mutta ennen sen on täytynyt tapahtua, — eikä ole puheita
kuulunut."

"Osaavathan ne järjestää. Ja paljonko tuollainen niinkuin Helmikin,

joka ei tee työtä, tarvitsee. Ei muuta kuin vähän viereen istahtaa, niin
heti paksuna."

Lyhyt, tukahdutettu naurahdus. Sitten:

"Vaan entäs tällainen palvelusihminen. Saa siinä käydä yksi ja

toinenkin, vaan vahinkoa ei tule kuin vahingossa."

"Itse itsesi parhaiten tiedät."

"Tiedän kyllä, mutten taida sellaisia neuvoa, joiden pitäisi itse

osata konstinsa."

"Niin aina. Ja kun ei olisi vaan nuoren isännän ja Helmin puolelta

edeltäkäsin harkittu koko homma."

*****

Kesän tuntu on jo voittava. Vielä ei kesä varsinaisesti ole astunut

esiin, mutta viileys on ilmasta kadonnut ja ruoho on kasvanut.

Helmi on valkea ja veretön ja hänen huulensa pysyvät visusti

suljettuina. Synnytys on ollut repivä ja tuskallinen, hän muistaa
jotakin huudelleensa ja hourailleensa, muttei muista, mitä. Milloin
mies töiltään ehtii, tulee hän vaimonsa vuoteen viereen ja kätilö
menee toisiin huoneisiin.

"Kuinka sinun nyt on?" kysyy mies.

 Menee aikoja, ennenkuin Helmi vastaa, eikä hän silloinkaan avaa
silmiään.

"Kiitos, onhan se nyt parempi."

Mies huokaa itsekseen:

"Kyllä se olikin kamalaa… en olisi sitä sellaiseksi uskonut."

Helmi pitää silmänsä edelleenkin suljettuina, niinkuin tahtoisi hän

nukkua, mutta hän kuulee jokaisen sanan.

"Taisin olla niinkuin mieletön silloin", kuiskasi hän.

"En minä saattanut kaikkea nähdä enkä kuulla… Et sinä ollenkaan

tiennyt, kuka olit."

"Niin kai… tuskat panevat sekaisin."

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.

"Minkä näköinen se poika on?" kuiskaa hän taas hetkisen kuluttua.

"Kyllä se komea mies on", toteaa mies reippaasti ja kysyy kuin

innostuen: "Entä, kuule, mikäs me sille pannaan niinkuin nimeksi?"

Helmi puristaa huulensa yhteen ja hänestä tuntuu niinkuin

viimeisetkin voimat nyt olisivat lähdössä. Mutta kun mies katsahtaa
häneen vastausta odottaen, huomaa hän, että vaimo on avannut
silmänsä ja että ne katsovat häneen kuin jostakin kaukaa, mutta
ankaroina ja totisina.