NONLINEAR MULTIOBJECTIVE

OPTIMIZATION
INTERNATIONAL SERIES IN
OPERATIONS RESEARCH & MANAGEMENT SCIENCE

Frederick S. Hillier, Series Editor

Department of Engineering-Economic Systems and
Operations Research
Stanford University
Stanford, California

Saigal, Romesh
LINEAR PROGRAMMING: A Modern Integrated Analysis

Nagumey, Annal Zhang, Ding

PROJECTED DYNAMICAL SYSTEMS AND VARIATIONAL INEQUALITIES WITH
APPLICATIONS

Padberg, Manfredi Rijal, Minendra P.

LOCATION, SCHEDULING, DESIGN AND INTEGER
PROGRAMMING

Vanderbei, Robert J.
LINEAR PROGRAMMING: Foundations and Extensions

Jaiswal, N.K.
MILITARY OPERATIONS RESEARCH: Quantitative D~cision Making

Gal, Tomas / Greenberg, Harvey J.

ADVANCES IN SENSITIVITY ANALYSIS AND PARAMETRIC PROGRAMMING

Prabhu, N. U.
FOUNDATIONS OF QUEUEING THEORY

Fang, S.-C./Rajasekera, J.R.! Tsao, H.-S.J.

ENTROPY OPTIMIZATION AND MATHEMATICAL PROGRAMMING

Yu. Gang
OPERATIONS RESEARCH IN THE AIRLINE INDUSTRY

Ho, Teck-Hua / Tang, Christopher S.

PRODUCT VARIETY MANAGEMENT

El-Taha, Muhammad / Stidham, Jr., Shaler

SAMPLE-PATH ANALYSIS OF QUEUEING SYSTEMS
NONLINEAR MULT/OBJECTIVE
OPTIMIZATION

by
Kaisa Miettinen, PhD
University ofJyvaskyla

.....
"
Springer Science+Business Media, LLC
Library of Congress Cataloging-in-Publication Data

Miettinen, Kaisa.
Non1inear multiobjective optimization / by Kaisa Miettinen.
p. cm. -- (International ser ies in operations research &
management science : 12)
lncludes bibliographical references and index.
ISBN 978-1-4613-7544-9 ISBN 978-1-4615-5563-6 (eBook)
DOI 10.1007/978-1-4615-5563-6
1. Multiple criteria decis ion making. 2. Non1inear programming.
1. Title. II. Series.
T57.95.M52 1999
658.4 '03--dc21 98-37888
CIP
Copyright © 1998 by Springer Science+Business Media New York. Fourth Printing 2004.
Origina11y published by Kluwer Academic Publishers in 1998
Softcover reprint of the hardcover 1st edition 1998
AU rights reserved. No part of this publication may be reproduced, stored in a
retrieval system or transmitted in any form or by any means, mechanica1, photo-
copying, recording, or otherwise, without the prior written permission of the
publisher, Springer Science+Business Media, LLC.

Printed on acid-free paper.

This printing is a digital duplication of the original edition.

 To my parents
Anna-Liisa and Kauko
with love and respect
 CONTENTS

PREFACE xiii
ACKNOWLEDGEMENTS xix
NOTATION AND SYMBOLS xxi

Part I TERMINOLOGY AND THEORY

1. INTRODUCTION 3
2. CONCEPTS 5
2.1. Problem Setting and General Notation 5
2.1.1. Multiobjective Optimization Problem 5
2.1.2. Background Concepts 6
2.2. Pareto Optimality 10
2.3. Decision Maker 14
2.4. Ranges of the Pareto Optimal Set 15
2.4.1. Ideal Objective Vector 15
2.4.2. Nadir Objective Vector 16
2.4.3. Related Topics 18
2.5. Weak Pareto Optimality 19
2.6. Value Function 21
2.7. Efficiency 23
2.8. From One Solution to Another 25
2.8.1. '!fade-Offs 26
2.8.2. Marginal Rate of Substitution 27
2.9. Proper Pareto Optimality 29
2.10. Pareto Optimality Tests with Existence Results 33
viii Contents

3. THEORETICAL BACKGROUND 37
3.1. Differentiable Optimality Conditions 37
3.1.1. First-Order Conditions 37
3.1.2. Second-Order Conditions 42
3.1.3. Conditions for Proper Pareto Optimality 43
3.2. Nondifferentiable Optimality Conditions 45
3.2.1. First-Order Conditions 47
3.2.2. Second-Order Conditions 52
3.3. More Optimality Conditions 54
3.4. Sensitivity Analysis and Duality 56

Part II METHODS

1. INTRODUCTION 61
2. NO-PREFERENCE METHODS 67
2.1. Method of the Global Criterion 67
2.1.1. Different Metrics 67
2.1.2. Theoretical Results 69
2.1.3. Concluding Remarks 71
2.2. Multiobjective Proximal Bundle Method 71
2.2.1. Introduction 71
2.2.2. MPB Algorithm 73
2.2.3. Theoretical Results 75
2.2.4. Concluding Remarks 75

3. A POSTERIORI METHODS 77
3.1. Weighting Method 78
3.1.1. Theoretical Results 78
3.1.2. Applications and Extensions 82
3.1.3. Weighting Method as an A Priori Method 83
3.1.4. Concluding Remarks 84
3.2. c-Constraint Method 85
3.2.l. Theoretical Results on Weak and Pareto Optimality 85
3.2.2. Connections with the Weighting Method 88
3.2.3. Theoretical Results on Proper Pareto Optimality 89
3.2.4. Connections with Trade-Off Rates 92
3.2.5. Applications and Extensions 94
3.2.6. Concluding Remarks 95
3.3. Hybrid Method 96
 Contents ix

3.4. Method of Weighted Metrics 97

3.4.1. Introduction 97
3.4.2. Theoretical Results 98
3.4.3. Comments 99
3.4.4. Connections with Trade-Off Rates 100
3.4.5. Variants of the Weighted Tchebycheff Problem 100
3.4.6. Connections with Global Trade-Offs 103
3.4.7. Applications and Extensions 106
3.4.8. Concluding Remarks 106
3.5. Achievement Scalarizing Function Approach 107
3.5.1. Introduction 107
3.5.2. Theoretical Results 108
3.5.3. Comments 110
3.5.4. Concluding Remarks 112
3.6. Other A Posteriori Methods 112

4. A PRIORI METHODS 115

4.1. Value Function Method 115
4.1.1. Introduction 115
4.1.2. Comments 116
4.1.3. Concluding Remarks 117
4.2. Lexicographic Ordering 118
4.2.1. Introduction 118
4.2.2. Comments 120
4.2.3. Concluding Remarks 120
4.3. Goal Programming 121
4.3.1. Introduction 121
4.3.2. Different Approaches 122
4.3.3. Comments 126
4.3.4. Applications and Extensions 127
4.3.5. Concluding Remarks 129

5. INTERACTIVE METHODS 131

5.1. Interactive Surrogate Worth Trade-Off Method 136
5.1.1. Introduction 136
5.1.2. ISWT Algorithm 137
5.1.3. Comments 140
5.1.4. Concluding Remarks 141
5.2. Geoffrion-Dyer-Feinberg Method 141
5.2.1. Introduction 141
5.2.2. GDF Algorithm 143
5.2.3. Comments 146
5.2.4. Applications and Extensions 146
5.2.5. Concluding Remarks 148
x Contents

5.3. Sequential Proxy Optimization Technique 149

5.3.l. Introduction 149
5.3.2. SPOT Algorithm 151
5.3.3. Comments 152
5.3.4. Applications and Extensions 153
5.3.5. Concluding Remarks 153
5.4. Tchebycheff Method 154
5.4. I. Introduction 154
5.4.2. Tchebycheff Algorithm 158
5.4.3. Comments 159
5.4.4. Concluding Remarks 160
5.5. Step Method 161
5.5.l. Introduction 161
5.5.2. STEM Algorithm 162
5.5.3. Comments 163
5.5.4. Concluding Remarks 164
5.6. Reference Point Method 164
5.6.1. Introduction 165
5.6.2. Reference Point Algorithm 165
5.6.3. Comments 167
5.6.4. Implementation 167
5.6.5. Applications and Extensions 169
5.6.6. Concluding Remarks 170
5.7. GUESS Method 170
5.7.l. Introduction 171
5.7.2. GUESS Algorithm 173
5.7.3. Comments 173
5.7.4. Concluding Remarks 174
5.8. Satisficing Trade-Off Method 174
5.8.1. Introduction 174
5.8.2. STOM Algorithm 176
5.8.3. Comments 177
5.8.4. Implementation 178
5.8.5. Applications and Extensions 178
5.8.6. Concluding Remarks 179
5.9. Light Beam Search 179
5.9.1. Introduction 180
5.9.2. Light Beam Algorithm 182
5.9.3. Comments 183
5.9.4. Concluding Remarks 184
5.10. Reference Direction Approach 184
5.10.1. Introduction 184
5.10.2. Reference Direction Approach Algorithm 185
5.10.3. Comments 187
5.10.4. Concluding Remarks 189
 Contents xi

5.11. Reference Direction Method 190

5.11.1. Introduction 190
5.11.2. RD Algorithm 192
5.11.3. Comments 193
5.11.4. Concluding Remarks 193
5.12. NIMBUS Method 195
5.12.1. Introduction 195
5.12.2. Vector Subproblem 197
5.12.3. Scalar Subproblem 198
5.12.4. NIMBUS Algorithm 198
5.12.5. Optimality Results 201
5.12.6. Comparison of the Two Versions 203
5.12.7. Comments 205
5.12.8. Implementations 205
5.12.9. Applications 206
5.12.10. Concluding Remarks 207
5.13. Other Interactive Methods 208
5.13.1. Methods Based on Goal Programming 208
5.13.2. Methods Based on Weighted Metrics 209
5.13.3. Methods Based on Reference Points 210
5.13.4. Methods Based on Miscellaneous Ideas 211

Part III RELATED ISSUES

1. COMPARING METHODS 217

1.1. Comparative Table of Interactive Methods Presented 218
1.2. Comparisons Available in the Literature 219
1.2.1. Introduction 220
1.2.2. Noninteractive Tests 221
1.2.3. Interactive Tests with Human Decision Makers 221
1.2.4. Interactive Tests with Value Functions 225
1.2.5. Comparisons Based on Intuition 226
1.3. Selecting a Method 227
1.3.1. General Guidelines 227
1.3.2. Method Selection Tools 228
1.3.3. Decision Tree 229

2. SOFTWARE 233
2.1. Introduction 233
2.2. Review 235

3. GRAPHICAL ILLUSTRATION 239

3.1. Introduction 239
3.2. Illustrating the Pareto Optimal Set 240
xii Contents

3.3. Illustrating a Set of Alternatives 240

3.3.1. Value Path 240
3.3.2. Bar Chart 242
3.3.3. Star Coordinate System 243
3.3.4. Spider-Web Chart 243
3.3.5. Petal Diagram 244
3.3.6. Scatterplot Matrix 245
3.3.7. Other Illustrative Means 246
3.3.8. General Remarks 247

4. FUTURE DIRECTIONS 251

5. EPILOGUE 255
REFERENCES 257
INDEX 293
 PREFACE

Life inevitably involves decision making, choices and searching for compro-
mises. It is only natural to want all of these to be as good as possible, in other
words, optimal. The difficulty here lies in the (at least partial) conflict between
our various objectives and goals. Most everyday decisions and compromises are
made on the basis of intuition, common sense, chance or all of these. However,
there are areas where mathematical modelling and programming are needed,
such as engineering and economics. Here, the problems to be solved vary from
designing spacecraft, bridges, robots or camera lenses to blending sausages,
planning and pricing production systems or managing pollution problems in
environmental control. Many phenomena are of a nonlinear nature, which is
why we need tools for nonlinear programming capable of handling several con-
flicting or incommensurable objectives. In this case, methods of traditional
single objective optimization are not enough; we need new ways of thinking,
new concepts, and new methods - nonlinear multiobjective optimization.
Problems with multiple objectives and criteria are generally known as mul-
tiple criteria optimization or multiple criteria decision-making (MCDM) prob-
lems. The area of multiple criteria decision making has developed rapidly, as
the statistics collected in Steuer et al. (1996) demonstrate. For example, by the
year 1994, a number of 144 conferences had been held and over 200 books and
proceedings volumes had appeared on the topic. Moreover, some 1216 refereed
journal articles were published between 1987 and 1992.
The MCDM field is so extensive that there is good reason to classify prob-
lems on the basis of their characteristics. They can be divided into two distinct
types (in accordance with MacCrimmon (1973)). Depending on the properties
of the feasible solutions, we distinguish multiattribute decision analysis and
muItiobjective optimization. In multiattribute decision analysis, the set of fea-
sible alternatives is discrete, predetermined and finite. Specific examples are
the selection of the locations of power plants and dumping sites or the pur-
chase of cars and houses. In multiobjective optimization problems, the feasible
alternatives are not explicitly known in advance. An infinite number of them
exists and they are represented by decision variables restricted by constraint
functions. These problems can be called continuous. In these cases, one has to
generate the alternatives before they can be valuated.
As far as multiattribute decision analysis is concerned, we refer to the mono-
graphs by Hwang and Yo on (1981) and Keeney and Raiffa (1976). More ref-
XIV Preface

erences, together with 17 major methods in the area accompanied by simple

examples, can be found in the latter monograph. A more recent summary of
the methodology is given in Yo on and Hwang (1995). A brief historical account,
including the basic ideas behind both multiobjective optimization and multiat-
tribute decision analysis together with suggestions for further reading, can be
found in Dyer et al. (1992) and Zionts (1992). (The latter also handles multi-
attribute utility theory and negotiation.) In addition, a review of the research
in both of these problem classes accompanied by future directions appears in
Korhonen et al. (1992a). It contains short descriptions of many concepts and
areas in multiple criteria optimization and decision making not included here.
In this book we concentrate solely on continuous multiobjective optimiza-
tion. This does not mean that some of the methods presented cannot be applied
to multiattribute decision analysis. Nevertheless, most of the methods have
been designed only for one or other of the problem types, exploiting certain
special characteristics.
The importance of multiobjective optimization can be seen from the large
variety of applications presented in the literature. Some idea of its possibilities
can be gained from the fact that over 500 papers describing different applica-
tions (between the years 1955 and 1986) are listed in White (1990). They cover,
for example, problems concerning agriculture, banking, the health service, en-
ergy, industry, water and wildlife.
Even though we have restricted ourselves to handling only multiobjective
optimization problems, it nonetheless remains a broad area of research and we
are therefore obliged to omit several topics to be able to give a uniform presen-
tation. We shall restrict the treatment to deterministic problems. Nevertheless,
a few words and further references are in order in relation to problems involv-
ing uncertainties. These can be divided into stochastic and fuzzy problems. In
stochastic programming it is usually assumed that uncertainty is due to a lack
of information about prevailing states, and that this uncertainty only concerns
the occurrence of the states and not the definition of the states, results or cri-
teria themselves. A problem containing random variables as coefficients on a
certain probability space is called a stochastic programming problem (treated,
for example, in the monographs of Guddat et al. (1985) and Stancu-Minasian
(1984)). When decision making takes place in an environment where the goals,
constraints and consequences of possible actions are not precisely known, it
is called decision making in fuzzy environments (handled, for example, in the
proceedings of Kacprzyk and Orlovski (1987)). Fuzzy coefficients may also be
involved in the problem formulation. Both stochastic and fuzzy multiobjective
optimization (for linear problems) are dealt with and compared in the proceed-
ings of Slowinski and Teghem (1990). Let us stress once again that here we
assume the problems to be deterministic; that is, the outcome of any feasible
solution is known for certain.
Solving problems with several conflicting objectives usually requires the
participation of a human decision maker who can express preference relations
 Preface xv

between alternative solutions and who continues from the point where math-
ematical tools end. Here we assume that a single decision maker is involved.
With several decision makers, the whole question of problem setting is very dif-
ferent. In addition to the mathematical side of the solution process, there is also
the aspect of negotiation and consensus striving between the decision makers.
The number of decision makers affects the means of approaching and solving
the problem significantly. A summary of group decision making is given in the
monograph of Hwang and Lin (1987). Here we settle for one decision maker.
A number of specific problem types requires special handling (not included
here). Among these are problems in which the feasible solutions must have inte-
ger values or 0-1 values, multiobjective trajectory optimization problems, where
the multiple objectives have multiple observation points, multiobjective net-
works or transportation networks and multiobjective dynamic programming.
Here we shall not go into these areas but adhere to standard methods.
Thus far we have outlined our interest here as being in deterministic contin-
uous multiobjective optimization with a single decision maker. This definition
still contains two broad areas, namely linear and nonlinear cases. Because lin-
ear programming utilizes the special characteristics of the problem, its methods
are not usually applicable to nonlinear problems. Further, linear multiobjective
optimization theory and methodology have been extensively treated in the lit-
erature, so there is no reason to repeat them here. One of the best presentations
focusing mainly on linear problems is Steuer (1986). However, the methodol-
ogy of nonlinear multiobjective optimization has not been drawn together since
Hwang and Masud (1979) (currently out of print). One more fact to notice is
that improved computational capacity enables problems to be handled with-
out linearizations and simplifications. Finally, linear problems are a subset of
nonlinear problems and that is why nonlinear methods can be used in both
cases. For these reasons, this book concentrates on nonlinear multiobjective
optimization.
The aim here is to provide an up-to-date, self-contained and consistent
survey and review of the literature and the state of the art on nonlinear (de-
terministic) multiobjective optimization starting with basic results.
The amount of literature on multiobjective optimization is immense. The
treatment in this book is based on about 1500 publications in English printed
mainly after the year 1980. Almost 700 of them are cited and listed in the bib-
liography. This extensive list of references supplements the contents regarding
areas not covered.
Problems related to real-life applications often contain irregularities and
nonsmoothnesses. The treatment of nondifferentiable multiobjective optimiza-
tion in the literature is rather rare. For this reason we also include in this book
material about the possibilities, background, theory and methods of nondiffer-
entiable multiobjective optimization.
Theory and methods for multiobjective optimization have been developed
chiefly during the last four decades. Here we do not go into the history as
xvi Preface

the origin and the achievements in this field of research from 1776 to 1960
are widely treated in Stadler (1979). A brief summary of the history is also
given in Gal and Hanne (1997). There it is demonstrated that multiobjective
optimization has its foundations in utility theory and economics, game theory,
mathematical research on order relations and vector norms, linear production
theory, and nonlinear programming.
Let us mention some further readings. The monographs of Chankong and
Haimes (1983b), Cohon (1978), Hwang and Masud (1979), Osyczka (1984),
Sawaragi et al. (1985), Steuer (1986) and Yu (1985) provide an extensive
overview of the area of multiobjective optimization. Further noteworthy mono-
graphs on the topic are those of Rietveld (1980), Vincke (1992) and Zeleny
(1974, 1982). A significant part of Vincke (1992) deals, however, with multiat-
tribute decision analysis. The behavioural aspects of multiobjective optimiza-
tion are mostly treated in Ringuest (1992), whereas the theoretical aspects are
extensively handled in the monographs by Jahn (1986a) and Luc (1989).
As far as this book is concerned, the contents are divided into three parts.
Part I provides the theoretical background. Chapter 1 leads into the topic and
Chapter 2 presents important notation, concepts and definitions in multiob-
jective optimization with some illustrative figures. Various theoretical aspects
appear in Chapter 3. For example, analogous optimality conditions for dif-
ferentiable and nondifferentiable problems are considered. A solid, conceptual
basis and foundation for the remainder of the book is laid. Throughout the
book we keep to problems involving only finite-dimensional Euclidean spaces.
(Dauer and Stadler (1986) provide a survey on multiobjective optimization in
infinite-dimensional spaces.)
The methodology is handled in Part II. Methods are divided into four classes
in Chapter 1 according to the role of a (single) decision maker in the solution
process. The state of the art in method development is portrayed by describing
a number of different methods accompanied by their theoretical background in
Chapters 2 to 5. For ease of comparison, all the methods are presented using a
uniform notation. The good and the weak properties of the methods are also in-
troduced with references to extensions and applications. The class of interactive
methods in Chapter 5 contains most of the methods, and it is the most exten-
sively handled. Linear problems and methods are only occasionally touched
on. In addition to describing solution methods, we introduce some implemen-
tations. In connection with every method described, some author's comments
appear in the concluding remarks. Some of the methods are depicted in more
detail and some only mentioned. Appropriate references to the literature are
always included.
Part III is Related Issues. After the presentation of a set of different so-
lution methods, some comparison is appropriate in Chapter 1. Naturally, no
absolute order of superiority can be given, but some points can be raised. A
table comparing some of the features of the interactive methods described is
included. In addition, we present brief summaries of some of the comparisons
 Preface xvii

available in the literature. Moreover, we suggest some outlines regarding the

important question of selecting an appropriate method. Method selection itself
is a problem with multiple objectives. Nevertheless, in addition to considering
some significant factors, we present a decision tree to aid selection. This tree
contains all the interactive methods previously described in some detail. It is
based on some of the fundamental assumptions underlying the methods and
different ways of exchanging information between the method and its user.
Compared with the plethora of methods, only a relatively few computer
implementations are widely known and available. However, some implement a-
tional aspects are touched on and some software mentioned in Chapter 2.
As computers and monitors have developed, graphical illustration has in-
creased in importance and has also become easier to produce. Hence graphical
illustration of alternative solutions together with related matters are featured
in Chapter 3. The potential and restrictions of graphics are treated and some
clarifying figures are enclosed.
We conclude with comments on future directions in Chapter 4 and an epi-
logue in Chapter 5.
This book is intended both for researchers and students in the areas such as
(applied) mathematics, engineering, economics, operations research and man-
agement science; it is meant both for professionals and practitioners in many
different fields of application. For beginners, this book provides an introduc-
tion to the theory and methodology of nonlinear multiobjective optimization.
For other readers, it offers an extensive reference to many related results and
methods. Obviously it is not possible in a single book to include all the aspects
and methods of nonlinear multiobjective optimization. However, the intention
has been to provide a consistent summary using a uniform notation leading
to further references. The uniform style of presentation may help in selecting
an appropriate method for the problem to be solved. It is hoped the extensive
bibliography will be of value to researchers.
The book gives sufficient theoretical background to allow those interested to
follow the derivation of the featured methods. However, the theoretical treat-
ment in Chapter 3 of Part I, for example, is not essential for the continua-
tion. For both theoretically and practically oriented readers, the algorithms
are described in a consistent manner with some implementational remarks and
software information also presented. Because, however, this is not an actual
textbook, no exercises or illustrative examples have been included.
 NOTATION AND SYMBOLS

Rn n-dimensional Euclidean space

Ii objective function page 5
k number of objective functions page 5
x decision (variable) vector page 5
n number of decision variables page 5
S feasible region page 5
Z feasible objective region page 5
f{x), z objective vector page 5
Rn nonnegative orthant of R n page 6
+
IIxll Euclidean norm page 6
dist{x, S) Euclidean distance function page 6
B{x,6) open ball page 6
intS interior of set S page 6
convS convex hull of set S page 6
V' fi{X) gradient of Ii at x Definition 2.1.4, page 7
!Ui.SE. partial derivative of Ii subject to Xj Definition 2.1.4, page 7
OZj
RnE blunt cone in R n page 9
V'2 fi{X) Hessian matrix of Ii at x Definition 2.1.11, page 9
81i{x) subdifferential of Ii at x Definition 2.1.14, page 10
{ subgradient Definition 2.1.14, page 10
Zi aspiration level Definition 2.3.1, page 14
Ii reference point Definition 2.3.1, page 14
z* ideal objective vector Definition 2.4.1, page 16
z** utopian objective vector Definition 2.4.2, page 16
znad nadir objective vector page 16
U value function Definition 2.6.1, page 21
D ordering cone page 23
Aij trade-off rate Definition 2.8.4, page 27
Af; global trade-off Definition 2.8.5, page 27
mij marginal rate of substitution Definition 2.8.6, page 28
A,p Karush-Kuhn-Tucker multipliers page 38
8. achievement (scalarizing) function page 108
P number of alternative objective vectors
 ACKNOWLEDGEMENTS

I wish to express my gratitude to several individuals. First, I must thank

Professor Pekka Neittaanmaki, who originally proposed multiobjective opti-
mization to me as a research subject. I also want to take this opportunity
to express a special thank you to Professor Pekka Korhonen whose ideas and
contacts helped me in publishing this book.
I am indebted to lecturers Michael Freeman, Ari Lehtonen and Jukka-Pekka
Santanen as well as to my colleague Marko Makela for their efforts in reading
the manuscript and suggesting improvements. In addition, I wish to thank
Marko for availing me freely of his expertise in nondifferentiable optimization.
I am grateful to my colleague Timo Mannikk6 for his helpful TEXnical
hints. Further, I wish to express my appreciation to Markku K6nkk6la as well
as Tapani Tarvainen for their help in numerous practical and technical matters
in the course of the preparation of this book, not to mention Markku's efforts
with some of the figures. I also want to thank my students for their assistance
in the implementations and Marja-Leena Rantalainen for some revisions.
On this occasion a special vote of thanks should be extended to all those
software developers who have given their programs to me for test purposes.
My thanks go to the Academy of Finland (grant number 22346) for financial
support and to the University of Jyvaskyla for providing me with the facilities
in which to work.
Finally, I want to truly acknowledge the support and love of my dear par-
ents, Anna-Liisa and Kauko, who laid the foundations for my education.

Kaisa Miettinen
 Part I

TERMINOLOGY AND THEORY

 1. INTRODUCTION

We begin by laying a conceptual and theoretical basis for the continuation

and restrict our treatment to finite-dimensional Euclidean spaces. First, we
present the deterministic, continuous problem formulation to be handled and
some general notation. Then we introduce several concepts and definitions of
multiobjective optimization as well as their interconnections. The concepts and
terms used in the field of multiobjective optimization are not completely fixed.
The terminology used here is occasionally slightly different from that in gen-
eral use. In some cases, only one of the existing terms is employed. Somewhat
different definitions of concepts are presented, for example, in Zionts (1989).
To deepen the theoretical basis, we treat optimality conditions for differ-
entiable and non differentiable multiobjective optimization problems. We also
briefly touch on the topics of sensitivity analysis, stability and duality.
Throughout the book, even some simple results are proved, for the conve-
nience of the reader (with possible appropriate references), in order to lay firm
cornerstones for the continuation. However, to keep the text to a reasonable
length, some proofs have been omitted if they can directly be found as such
elsewhere. In those cases, appropriate references in the literature are indicated.
Multiobjective optimization problems are usually solved by scalarization.
Scalarization means that the problem is converted into a single (scalar) or a
family of single objective optimization problems. In this way the new problem
has a real-valued objective function, possibly depending on some parameters.
After the multiobjective optimization problem has been scalarized, the widely
developed theory and methods for single objective optimization can be used.
Even though multiobjective optimization methods are presented in Part II,
we emphasize here at the outset that the methods and the theory of single
objective optimization are presumed to be known.
 2. CONCEPTS

This chapter introduces the basic concepts of (nonlinear) multiobjective

optimization and the notations used in the continuation.

2.1. Problem Setting and General Notation

We begin by defining the problem to be handled.

2.1.1. Multiobjective Optimization Problem

We study a multiobjective optimization problem of the form

minimize {II (x), 12 (x), ... , fk (x)}
(2.1.1)
subject to XES,

where we have k (~ 2) objective functions fi: Rn -+ R. We denote the vector

of objective functions by f(x) = (II (x), hex), ... , fk (x))T. The decision (vari-
able) vectors x = (Xl,X2, ... ,xn )T belong to the (nonempty) feasible region
(set) S, which is a subset of the decision variable space Rn. We do not yet fix
the form of the constraint functions forming S, but refer to S in general.
The word 'minimize' means that we want to minimize all the objective func-
tions simultaneously. If there is no conflict between the objective functions, then
a solution can be found where every objective function attains its optimum. In
this case, no special methods are needed. To avoid such trivial cases we assume
that there does not exist a single solution that is optimal with respect to every
objective function. This means that the objective functions are at least partly
conflicting. They may also be incommensurable (i.e., in different units).
In the following, we denote the image of the feasible region by Z (= f(S))
and call it a feasible objective region. It is a subset of the objective space R k. The
elements of Z are called objective (function) vectors or criterion vectors and
denoted by f(x) or z = (Zl, Z2, ... , Zk)T, where Zi = fi(x) for all i = 1, ... , k are
objective (function) values or criterion values. The words in the parentheses
above are usually omitted for short.
K. Miettinen, Nonlinear Multiobjective Optimization
© Springer Science+Business Media New York 1998
6 Part I ~ 2. Concepts

For clarity and simplicity of the treatment we assume that all the objective
functions are to be minimized. If an objective function Ji is to be maximized,
it is equivalent to minimize the function - Ii-
In what follows, whenever we refer to a multiobjective optimization prob-
lem, it is problem (2.1.1) unless stated otherwise. Finding a solution to (2.1.1)
in one way or another is called a solution pmcess in the continuation.

2.1.2. Background Concepts

First, we present some general concepts and notations. We use bold face
and superscripts for vectors, for example, Xl, and subscripts for components
of vectors, for example, Xl. All the vectors here are assumed to be column
vectors. For two vectors, x and x· ERn, the notation x T x· denotes their
scalar pmduct and the vector inequality x ~ x* means that Xi ~ xi for all
i = 1, ... , n. Correspondingly x < x· stands for Xi < xi for all i = 1, ... ,n.
The nonnegative orthant of R n is denoted by R+. In other words, R+ =
{x E R n I Xi ~ 0 for i = 1, ... , n}. The Euclidean n01'm of a vector x E Rn
is denoted by Ilxll = (E~=l xT) 1/2. The Euclidean distance Junction between a
point x* and a set S is denoted by dist(x*, S) = infxEs IIx· - xII. The symbol
B(x*,6) denotes an open ball with a centre x· and a radius 15 > 0, B(x·, 15) =
{x ERn IlIx· - xII < c5}. The notation int S stands for the interior of a set S.
The vectors Xi, i = 1, ... , m, are linearly independent if the only weighting
coefficients (3i for which E::1 (3ixi = 0 are (3i = 0, i = 1, ... , m. The sum
E:: 1 (3ixi is called a convex combination of the vectors Xl, x , ... ,x
2 m E S, if

(3i ~ 0 for all i and E::1 (3i = 1. The convex hull of a set S eRn, denoted by
conv S, is the set of all convex combinations of vectors in S,
A set 8 eRn is a cone if (3x = ((3x1, ... , (3xm)T E S whenever x E 8 and
(3 ~ O. The negative of a cone is -8 = {-x ERn I x E 8}. A cone 8 is said to
be pointed if it satisfies S n -S = {OJ. A cone -S transformed to x* ERn is
denoted by x* - S = {x ERn I x = x* + d, where d E -S},
It is said that d ERn is a feasible direction emanating from xES if there
exists a* > 0 such that x + ad E S for 0 ~ a ~ a*.
In some connections we assume that the feasible region is formed of inequal-
ity constraints, that is, 8 = {x E R n I g(x) = (gl(X),g2(X), ... ,gm(x))T ~ OJ.
An inequality constraint gj is said to be active at a point x· if gj(x*) = 0,
and the set of active constraints at x* is denoted by J(x*) = {j E {l, ... , m} I
gj(x*) = OJ.
Different types of multiobjective optimization problems can be defined.

Definition 2.1.1. When all the objective functions and the constraint func-
tions forming the feasible region are linear, then the multiobjective optimiza-
tion problem is called linear. In brief, it is an MOLP (multiobjective linear
programming) problem.
If at least one of the objective or the constraint functions is nonlinear, the
problem is called a nonlinear multiobjective optimization problem.
 2.1. Problem Setting and General Notation 7

A large variety of solution techniques have been created as to enable the

special characteristics of MOLP problems to be taken into account. Here we
concentrate on cases where nonlinear functions are included and thus methods
for nonlinear problems are needed. Methods and details of MOLP problems are
mentioned only incidentally.
Before we define convex multiobjective optimization problems, we briefly
write down the definitions of convex functions and convex sets.

Definition 2.1.2. A function J;: R n -+ R is convex if for all Xl, x 2 ERn is

valid that h({3x l + (1 - (3)x 2) ~ (3fi(X I ) + (1 - (3)fi(X 2) for all 0 ~ {3 ~ 1.
A set S c Rn is convex if Xl, x 2 E S implies that {3x l + (1 - (3)x 2 E S for
allO${3:5l.

Definition 2.1.3. The multiobjective optimization problem is convex if all the

objective functions and the feasible region are convex.

A convex multiobjective optimization problem is an important concept in

the continuation. We shall also need related generalized concepts, quasiconvex
and pseudoconvex functions. The pseudo convexity of a function calls for dif-
ferentiability. For completeness, we write down the definitions of differentiable
and continuously differentiable functions.

Definition 2.1.4. A function h: R n -+ R is differentiable at x* if

J;(x· + d) - fi(X*) = V li(x*)T d + IIdll c(x·, d),

where Vfi(X*) is the gradient of J; at x· and c(x*,d) -+ 0 as IIdll-+ O.
In addition, h is continuously differentiable at x* if all of its partial deriva-
tives 8J~i~') (j = 1, ... , n), that is, all the components of thp I!radient are
1
continuous at x*.

The gradient of Ii at x* can also be denoted by V xli(x*) to emphasize that

the derivation is carried out subject to x.
Now we can define quasiconvex and pseudoconvex functions.

Definition 2.1.5. A function Ii: Rn -+ R is quasiconvex if fi({3X I + (1 -

(3)x 2) $ max [1i(x I ),/i(x 2)] for all 0 $ (3 $ 1 and for all xI,x 2 ERn.
Let Ii be differentiable at every x ERn. Then it is pseudoconvex if for all
xl,X2 ERn such that Vfi(x l )T(x 2 - Xl);::: 0, we have fi(x 2);::: fi(X 1).

As far as the relations of quasiconvex and pseudoconvex functions are con-

cerned, every pseudoconvex function is also quasiconvex.
The definition of convex functions can be modified for concave functions by
replacing '$' by';:::'. Correspondingly, the definition of quasiconvex functions
becomes appropriate for quasiconcave functions by the exchange of '$' to ';:::'
and 'max' to 'min'. In the definition of pseudo convex functions we replace
 8 Part I ~ 2. Concepts

'~' by':::;' to get the definition for pseudoconcave functions. Notice that if a
function J; is quasiconvex, all of its level sets {x ERn I Ii (x) :::; a} are convex
and if it is quasi concave, all of its level sets {x ERn I J; (x) ~ a} are convex
(see, for example, Mangasarian (1969, pp. 133-134)).
Sometimes we also need strict definitions.

Definition 2.1.6. A function J;: Rn -+ R is strictly convex if li((Jx l + (1 -

(J)x 2 ) < (Jli(x l ) + (1 - (J)/i(x 2 ) and strictly quasiconvex if li((Jx l + (1 -
fJ)x 2 ) < max[J;(x l ),J;(x 2)] for all 0 < (J < 1 and for all xl,X2 ERn, where
h(x l ) i-h(x 2 ).

Notice that strict convexity of a function implies convexity and convexity

implies both strict quasi convexity and quasiconvexity. If differentiability is as-
sumed, convexity implies pseudoconvexity which implies strict quasiconvexity.
See Bazaraa et al. (1993, pp. 78-118) or Mangasarian (1969, pp. 131-147) for
the details of the relations. The corresponding results are valid for concave
functions and their generalizations. It is worth pointing out that convexity,
concavity and related concepts can be defined in a convex set S c R n as well
as in Rn.
We also need other function types. The first of these are related to mono-
tonicity.

. Definition 2.1.7. A function Ii: Rn -+ R is increasing iffor Xl and x 2 E Rn

x}:::;x; for all j=I, ... ,n imply h(x l ):::;/i(x 2 ).

Correspondingly, the function Ii is decreasing if J; (x 1 ) ~ J; (x 2 ).

A function is monotonic (or order preserving) if it is either increasing or

decreasing. Monotonicity can be tightened up in several ways.

Definition 2.1.8. A function Ii: Rn -+ R is strictly increasing if for xl and

x 2 ERn

x} <x; for all j=I, ... ,n imply h(x l ) <J;(x 2 ).

Definition 2.1.9. A function Ii: Rn -+ R is strongly increasing if for Xl and

x2 E R n
x} :::; Xl for all j := 1, ... , n and xl < x; for some 1 imply J;(x l ) < li(x 2 ).
Correspondingly, the function J; is strongly decreasing if J;(x l ) > J;(x 2 ).

Notice that if a function is strongly decreasing and differentiable, all of its

partial derivatives have to be (strictly) negative.
In the next definition we need a subset R~ of R n. It is defined as
 2.1. Problem Setting and General Notation 9

R~ = {x E R n I dist (x, R~) :::; cllxll}·

Definition 2.1.10. A function Ii: Rn -t R is c-strongly increasing if for Xl

and x 2 E Rn

For the convenience of the reader we define twice differentiable functions

and some related concepts.

Definition 2.1.11. A function Ii: Rn -t R is twice-differentiable at x· if

where '\7 fi(X·) is the gradient, the symmetric nxn matrix '\7 2 h(x·) is a Hessian
matrix of h at x· and c(x·, d) -t 0 as Ildll -t O. The Hessian matrix of a
twice-differentiable function consists of second-order partial derivatives 8;!;c:,,~)
(j, 1 = 1, ... , n). In other words,

~)
8x l 8x n

8 2/i\X*)
8xa

In addition, Ii is twice continuously differentiable at x· if all of its second-

order partial derivatives are continuous at x·.

A symmetric n x n matrix M is called positive definite, if xl'Mx > 0 for

allO::j:.xERn.
We shall also handle nondifferentiable multiobjective optimization prob-
lems. For that reason we define locally Lipschitzian functions (see Clarke (1983,
pp. 9-11) and MiikeUi and Neittaanmiiki (1992, pp. 5-10)).

Definition 2.1.12. A function Ii: Rn --+ R is locally Lipschitzian at a point

x· ERn if there exist scalars K > 0 and J > 0 such that

Notice that a convex function h: Rn --+ R is for any point x E R n locally

Lipschitzian at x.
In what follows, a function is called nondifferentiable if it is locally Lip-
schitzian (and not necessarily continuously differentiable).
10 Part I - 2. Concepts

Definition 2.1.13. The multiobjective optimization problem is nondifferen-

tiable if some of the objective functions or the constraint functions forming the
feasible region are nondifferentiable.

According to Rademacher's Theorem (see, e.g., Federer (1969)), we know

that a locally Lipschitzian function, defined in an open set, is differentiable
almost everywhere in that set. A set where a function fi is not differentiable
is denoted here by ilIi' In the sequel, we employ the concept sub differential
as defined in Clarke (1983). It corresponds to the gradient in the differentiable
case.

Definition 2.1.14. Let the function fi: Rn -+ R be locally Lipschitzian at a

point x' ERn. The set

8fi(X*) = conv {{ ERn I{= lim V' h(xl ); xl -+ x', xl ERn \ ill.}
l-too

is called a subdifferential of the function h evaluated at the point x*. In addi-

tion, the vectors { E 8 fi (x *) are called subgmdient.~.

We end with a special type of upper semidifferentiable function (see Wang

(1989)).

Definition 2.1.15. Let the function fi: R n -+ R be locally Lipschitzian at a

point x* ERn. Then it is upper semidifferentiable at x* if for every d ERn,
any sequence {tj}~l with tj -+ 0 and sequence {{i}, wheree E 8fi(X* +tjd)
for every j, we have

· . f fi(X* + tjd) - h(x*) <

IImm _ I"Imsup (~j)Td
... .
J-too tj j-too

Special properties of nondifferentiable functions are introduced in Section

3.2, in the context where nondifferentiability is handled.
After these general definitions and concepts we can continue with multiob-
jective optimization terminology.

2.2. Pareto Optimality

In this section, we handle a crucial concept in optimization, namely op-

timality. In single objective optimization problems, the main focus is on the
decision variable space. In the multiobjective context we are often more inter-
ested in the objective space. For one thing, it is usually of a lower dimension
than the decision variable space. Further, objective values are used below in
defining optimality.
 2.2. Pareto Optimality 11

Because of the contradiction and possible incommensurability of the objec-

tive functions, it is not possible to find a single solution that would be optimal
for all the objectives simultaneously. Multiobjective optimization problems are
in a sense ill-defined. There is no natural ordering in the objective space be-
cause it is only partially ordered (meaning that, for example, (1,I)T can be
said to be less than (3, 3)T, but how to compare (1,3)T and (3, I)T). This is
always the case when vectors are compared in real spaces (see also Chankong
and Haimes (1983b, pp. 64-67».
Anyway, some of the objective vectors can be extracted for examination.
Such vectors are those where none of the components can be improved without
deterioration to at least one of the other components. Edgeworth (1987) pre-
sented this definition in 1881. However, the definition is usually called Pareto
optimality after the French-Italian economist and sociologist Vilfredo Pareto,
who in 1896 developed it further (see Pareto (1964, 1971». However, in some
connections, like in Stadler (1988b), the term Edgeworth-Pareto optimality is
used for the above-mentioned reason. Koopmans was one of the first to employ
in 1951 the concept of Pareto optimality in Koopmans (1971). A more formal
definition of Pareto optimality is the following:

Definition 2.2.1. A decision vector x" E S is Pareto optimal if there does not
exist another decision vector xES such that fi(X) ~ fi(X") for all i = 1, ... , k
and fJ(x) < fJ(x*) for at least one index j.
An objective vector z* E Z is Pareto optimal if there does not exist another
objective vector z E Z such that Zi ~ z; for all i = 1, ... , k and Zj < z] for
at least one index jj or equivalently, z* is Pareto optimal if the decision vector
corresponding to it is Pareto optimal.

In Figure 2.2.1, a feasible region S C R3 and its image, a feasible objective

region Z C R2, are illustrated. The fat line contains all the Pareto optimal
objective vectors. The vector z" is an example of them.

Figure 2.2.1. The sets Sand Z and the Pareto optimal objective vectors.
12 Part I - 2. Concepts

There are usually a lot (infinite number) of Pareto optimal solutions. We

can speak about a set of Pareto optimal solutions or a Pareto optimal set. This
set can be nonconvex and nonconnected.
In addition to Pareto optimality, several other terms are sometimes used for
the optimality concept described above. These terms are, for example, nonin-
fer'iority, efficiency and nondominance. At variance with this practice, a more
general meaning is given to efficiency later. In general, Pareto optimality is
here used as a concept of optimality, unless stated otherwise.
Definition 2.2.1 introduces global Pareto optimality. Another important con-
cept is local Pareto optimality.

Definition 2.2.2. A decision vector x* E S is locally Pareto optimal if there

exists /j > 0 such that x* is Pareto optimal in S n B(x*, /j).
An objective V(dor z* E Z is locally Pareto optimal if the decision vector
corresponding to it is locally Pareto optimal.

Naturally, any globally Pareto optimal solution is locally Pareto optimal.

The converse is valid for convex multiobjective optimization problems. (For
this result, see e.g., Censor (1977).)

Theorem 2.2.3. Let the multiobjective optimization problem be convex.

Then every locally Pareto optimal solution is also globally Pareto optimal.

Proof. Let x* E S be locally Pareto optimal. Thus there exist some /j > 0 and
a neighbourhood B(x*, /j) of x* such that there is no xES n B(x*, /j) for which
h(x) ~ fi(X*) for all i = 1, ... , k and for at least one index j is Ji(x) < Ji(x*).
Let us assume that x* is not globally Pareto optimal. In this case, there
exists some other point XO E S such that

(2.2.1) fi(XO) ~ fi(X*) for all i = 1, ... , k and Ji(XO) < Ji(x*) for some j.
Let us define x = (3xo + (1 - (3)x*, where 0 < (3 < 1 is selected such that
x E B(x*, 6). The convexity of S implies that xES.
By the convexity of the objective functions and employing (2.2.1), we obtain
fi(X) ~ (3fi(XO) + (1 - (3)fi(X*) ~ (3fi(X*) + (1 - (3)h(x*) = f;(x*) for every
i = 1, ... , k. Because x* is locally Pareto optimal and x E B(x*, /j), we must
have fi(X) = fi(X*) for all i.
Further, h(x*) ~ (3h(xO) + (1 - (3)h(x*) for every i = 1, ... , k. Because
(3 > 0, we can divide by it and obtain f;(x*) ~ fi(XO) for all i. According
to assumption (2.2.1), we have Ji(x*) > fj(xO) for some j. Here we have a
contradiction. Thus, x* is globally Pareto optimal. 0

We can establish the above-mentioned result with somewhat weaker as-

sumptions. It is sufficient to assume that all the objective functions are qua-
siconvex and strictly quasiconvex. This result has been treated, for example,
 2.2. Pareto Optimality 13

in Ruiz-Canales and Rufian-Lizana (1995). These assumptions can be further

relaxed according to Luc and Schaible (1997).

Theorem 2.2.4. Let the multiobjective optimization problem have a convex

feasible region and quasiconvex objective functions with at least one strictly
quasiconvex objective function. Then every locally Pareto optimal solution is
also globally Pareto optimal.

Proof. Let x* E S be locally Pareto optimal. Thus there exist some J > 0 and
a neighbourhood B(x*, J) of x* such that there is no x E SnB(x*, J) for which
J;(x) :::; J;(x*) for all i = 1, ... , k and for at least one index j is hex) < h(x*).
Let us assume that x· is not globally Pareto optimal. In this case, there
exists some other point XO E S such that

(2.2.2) fi(XO):::; J;(x*) for all i = 1, ... , k and h(xO) < h(x*) for some j.
Let us define x = f3x + (1 - f3)x*, where 0 < f3 <
o 1 is selected such that
x E B(x*, 8). The convexity of S implies that XES.
Employing (2.2.2) and by the quasiconvexity of the objective functions,
respectively, for each index i such that fi(XO) = J;(x·), we obtain

and for each index j such that h(xO) < fj(x*), we have

Because at least one of the objective functions is strictly quasiconvex, at least

one of the inequalities above is strict. Here we have a contradiction with the
local Pareto optimality of x·. Thus, x* is globally Pareto optimal. 0

For the sake of brevity, we shall usually speak only about Pareto optimality
in the sequel. In practice, however, we only have locally Pareto optimal solu-
tions computationally available, unless some additional requirement, such as
convexity, is fulfilled.
Usually, we are interested in Pareto optimal solutions and can forget the
other feasible solutions. Exceptions to this practice are problems where one of
the objective functions is an approximation of an unknown function or there
are underlying unexpressed objective functions involved. Then, the real Pareto
optimal set is unknown.
According to the definition of Pareto optimality, moving from one Pareto
optimal solution to another necessitates trading off. This is one of the basic
concepts in multiobjective optimization. Let us, however, mention that the
idea of trading off can be called into question, as suggested, for example, in
Zeleny (1997). It is not perhaps always necessary to trade off in order to attain
improved results. One can argue that it has been possible to produce things
14 Part 1 - 2. Concepts

both at lower cost and with higher quality. Changing the way of approaching the
problem and its formulation may produce better results than simply trading off
in the old formulation. (This can also be regarded as an example of expanding
habitual domains, to be introduced in Section 2.3.) Zeleny goes so far as to
claiming that trade-offs are properties of inadequately designed systems. For
that reason one can claim that we should aim at designing systems better.

2.3. Decision Maker

Mathematically, every Pareto optimal point is an equally acceptable solu-

tion of the multiobjective optimization problem. However, it is generally desir-
able to obtain one point as a solution. Selecting one out of the set of Pareto
optimal solutions calls for information that is not contained in the objective
functions. This is why - compared to single objective optimization - a new
element is added in multiobjective optimization.
We need a decision maker to make the selection. The decision maker is a
person (or a group of persons) who is supposed to have better insight into the
problem and who can express preference relations between different solutions.
Usually, the decision maker is responsible for the final solution.
Solving a multiobjective optimization problem calls for the co-operation
of the decision maker and an analyst. By an analyst we here mean a person
or a computer program responsible for the mathematical side of the solution
process. The analyst generates information for the decision maker to consider
and the solution is selected according to the preferences of the decision maker.
It is assumed in the following that we have a single decision maker or a
unanimous group of decision makers. Generally, group decision making is a
world of its own. It calls for negotiations and specific methods when searching
for compromises between different interest groups (see, for example, Hwang
and Lin (1987) and Yu (1973)).
In Part II, solution methods are classified according to the role of the deci-
sion maker in the solution process. In some methods, various assumptions are
made concerning the preference structure and behaviour of the decision maker.
Note that assuming a single decision maker does not exclude the possibility
that there may be others involved infiuencing the decision maker (as stressed
in Zionts (1997a, b)).
During solution processes, various kinds of information are solicited from the
decision maker. Such items of information may include, for example, desirable
or acceptable levels in the values of the objective functions. These objective
values (whether feasible or not) are of special interest and importance to the
decision maker.

Definition 2.3.1. Objective function values that are satisfactory or desirable

to the decision maker are called aspiration levels and denoted by Zi, i = 1, ... , k.
The vector z E R k, consisting of aspiration levels, is called a reference point.
 2.4. Ranges of the Pareto Optimal Set 15

By solving a multiobjective optimization problem we here mean finding a

feasible decision vector such that it is Pareto optimal and satisfies the needs
and the requirements of the decision maker. Assuming such a solution exists,
it is called a final solution. However, as stressed in Zionts (1997a, b), it may
be difficult for the decision maker to distinguish between good and optimal
solutions in real problems. If this is the case, the emphasis should be on finding
good solutions (and sometimes, only, on finding solutions).
We do not focus here on the problems of decision making, which is a research
area of its own. Interesting topics in this area are, for instance, decision making
with incomplete information, validity of the problem formulation and habitual
domains. The first of these matter is treated, for example, in Weber (1987).
Reasons for incomplete information include lack of knowledge, pressure of time,
fear of commitment and matters related to the future.
We usually assume that decision makers are only interested in Pareto op-
timal points and the rest can be excluded. However, this is not the case if the
problem has not been formulated well enough. As already emphasized, non-
Pareto optimal solutions may be important if there are some unformulated or
hidden objective functions in the mind of the decision maker or some of the
objective functions are simply proxies of the objective functions proper (see,
for example, Zionts (1997a, b». In such cases, the Pareto optimal sets of the
problem handled and the actual problem which should be solved, do not co-
incide. Here we assume the mathematical model to be accurate and static so
that we can mainly concentrate on Pareto optimal solutions.
A habitual domain is defined in Yu (1991) as a set of ways of thinking, judg-
ing and responding, as well as the knowledge and experience on which they are
based. Yu emphasizes that in order to make effective decisions it is important
to expand and enrich the habitual domains of the decision makers. Several ways
of carrying this out are presented in Yu (1991, 1995). Understanding, expand-
ing and enriching the domains of thinking is also stressed, for example, in Yu
(1994) and Yu and Liu (1997).

2.4. Ranges of the Pareto Optimal Set

Let us for a while investigate the ranges of the set of Pareto optimal so-
lutions. We assume that the objective functions are bounded over the feasible
region S.

2.4.1. Ideal Objective Vector

An objective vector minimizing each of the objective functions is called an

ideal (or perfect) objective vector.
16 Part I - 2. Concepts

Definition 2.4.1. The components zi of the ideal objective vector z* E R k

are obtained by minimizing each of the objective functions individually subject
to the constraints, that is, by solving

minimize 1; (x)
subject to XES,

for i = 1, ... , k.

It is obvious that if the ideal objective vector were feasible (that is, z* E Z),
it would be the solution of the multiobjective optimization problem (and the
Pareto optimal set would be reduced to it). This is not possible in general since
there is some conflict among the objectives. Even though the ideal objective
vector is not attainable, it can be considered a reference point, something to go
for. From the ideal objective vector we obtain the lower bounds of the Pareto
optimal set for each objective function.
Note that in practice some caution is in order with nonconvex problems.
The definition of the ideal objective vector assumes that we know the global
minima of the individual objective functions. Guaranteeing global optimality
in numerical calculations is not that simple. This must be kept in mind with
practical problems. Properties of ideal objective vectors, for example, their
uniqueness, are treated in Skulimowski (1992).
Sometimes we also need a vector that is strictly better than, in other words,
strictly dominates, every Pareto optimal solution.

Definition 2.4.2. A utopian objective vector z** E Rk is an infeasible objec-

tive vector whose components are formed by

for all i = 1, ... , k, where zi is a component of the ideal objective vector and
Ci > 0 is a relatively small but computationally significant scalar.

2.4.2. Nadir Objective Vector

The upper bounds of the Pareto optimal set, that is, the components of
a nadir objective vector (or imperfect objective vector) znact, are much more
difficult to obtain. However, they can be estimated from a payoff table.
A payoff table is formed by using the decision vectors obtained when calcu-
lating the ideal objective vector. Row i of the payoff table displays the values of
all the objective functions calculated at the point where Ii obtained its minimal
value. Hence, zi is at the main diagonal of the table. The maximal value of the
column i in the payoff table can be selected as an estimate of the upper bound
of the objective Ii for i = 1, ... , k over the Pareto optimal set.
The black points in Figure 2.4.1 represent ideal objective vectors, and the
grey ones are nadir objective vectors. The nadir objective vector may be feasible
 2.4. Ranges of the Pareto Optimal Set 17

or not, as illustrated in Figure 2.4.1. The Pareto optimal set is represented by

the bold lines.

z 2

•.............
ideal
Z I

Figure 2.4.1. Ideal objective vectors and nadir objective vectors.

Note that the objective vectors in the rows of the payoff table are Pareto
optimal if they are unique. In other words, if the individual objective functions
have alternative optima, the obtained objective vector may not be Pareto op-
timal. This fact can weaken the approach and it can happen in linear as well
as in nonlinear problems.
It is important to note that the estimates based on the payoff table are
not necessarily equal to the real components of the nadir objective vector as
demonstrated, for example, in Korhonen et al. (1997) and Weistroffer (1985).
Instead of being correct, the nadir objective value approximate may be either
far too low or too high.
The difference between the complete Pareto optimal set and the subset of
the Pareto optimal set bounded by the ideal objective vector and the upper
bounds obtained from the payoff table in linear cases is explored in Reeves and
Reid (1988). It is proposed that relaxing (i.e., increasing) the approximated up-
per bounds by a relatively small tolerance should improve the approximation,
although it is ad hoc in nature. However, small tolerances may not necessar-
ily help because the error between the correct and the approximated nadir
objective value may be significant.
For nonlinear problems, there is no constructive method for calculating the
nadir objective vector. That is why we here mention some treatments for MOLP
problems. Isermann and Steuer (1988) include an examination of how many of
the Pareto optimal extreme solutions of some MOLP problems are above the
upper bounds obtained from the payoff table. Three methods for determining
the exact nadir objective vector in a linear case are also suggested. None of
them is especially economical computationally. In Dessouky et al. (1986), three
heuristics are presented for calculating the nadir objective vector when the
18 Part I - 2. Concepts

problem is linear. A heuristic for MOLP problems is also described in Korho-

nen et al. (1997). It is demonstrated how much better are the approximations
the heuristic can provide. Heuristics are usually able to improve the approxi-
mations obtained from the payoff table even though they may not always find
the correct nadir objective values. Heuristics are often computationally much
less demanding than exact procedures.
Nonetheless, the payoff table may be used as a rough estimate as long
as its robustness is kept in mind. Because of the above-described difficulty
of calculating the actual nadir objective vector, we shall usually refer to the
approximate nadir objective vector as znad.

2.4.3. Related Topics

In many occasions it is advisable to rescale, that is, normalize the objective

functions so that their objective values are of approximately the same mag-
nitude. If the ideal objective vector and a good enough approximation to the
nadir objective vector are known, we can replace each objective function fi(x)
(i = 1, ... , k) by the function
fi(x) - zt
zi ad - zt .

In this case, the range of each new objective function is [0,1].

Another related possibility is to use a range equalization factor, as suggested
in Steuer (1986). The range R; of each objective function is first estimated by
the difference between the (possibly approximated) nadir objective vector and
the ideal objective vector. Then, constants
11
Ki = R. ,\,k 1
1 i.-Ij=l Rj

are defined for every i = 1, ... , k, and finally each objective function is multi-
plied by K i .
A simple alternative for normalizing the objective function values is to di-
vide each objective function by its (nonzero) ideal objective value. This has
been suggested, for example, in Osyczka (1984, 1992). This is not as exact as
the previous methods but does not necessitate information about the nadir
objective vector.
It is usually advisable to use normalized objective values only in calcu-
lations and to display restored objective values in the original scales to the
decision maker. In this way the different scales do not confuse computation
and significant objective values are offered to the decision maker.
It is possible that (some) objective functions are unbounded, for instance,
from below. In this case some caution is in order. In multiobjective optimiza-
tion problems this does not necessarily mean that the problem is formulated
 2.5. Weak Pareto Optimality 19

incorrectly. There may still exist Pareto optimal solutions. However, if, for in-
stance, some component of the ideal objective vector is unbounded and it is
replaced by a small but finite number, methods utilizing the ideal objective
vector may not be able to overcome the replacement.
Finally, let us look at some examples of the problem of optimizing a function
over the Pareto optimal set of a multiobjective optimization problem. This is a
more general problem than just looking for the ranges of the Pareto optimal set.
In Benson and Sayin (1994), the authors deal with the maximization of a linear
function over the Pareto optimal set of an MOLP problem. A general function
is minimized over the Pareto optimal set of an MOLP problem in Dauer and
Fosnaugh (1995), and a convex function is optimized over the Pareto optimal set
of linear objective functions and a convex feasible region by duality techniques
in Thach et al. (1996). Maximization of a function over the Pareto optimal set
is also considered in Horst and Thoai (1997).

2.5. Weak Pareto Optimality

In addition to Pareto optimality, other related concepts are widely used.

These are weak and proper Pareto optimality. The relationship between these
concepts is that the properly Pareto optimal set is a subset of the Pareto
optimal set which is a subset of the weakly Pareto optimal set.
A vector is weakly Pareto optimal if there does not exist any other vector
for which all the components are better. More formally it means the following:

Definition 2.5.1. A decision vector x* E S is weakly Pareto optimal if there

does not exist another decision vector XES such that /i(X) < /i(x*) for all
i = 1, ... , k.
An objective vector z* E Z is weakly Pareto optimal if there does not
exist another objective vector z E Z such that Zi < zi for all i = 1, ... , kj or
equivalently, if the decision vector corresponding to it is weakly Pareto optimal.

The bold line in Figure 2.5.1 represents the set of weakly Pareto optimal
objective vectors. The fact that the Pareto optimal set is a subset of the weakly
Pareto optimal set can also be seen in the figure. The Pareto optimal objective
vectors are situated along the line between the dots.
Similarly to Pareto optimality, local weak Pareto optimality can be defined
in addition to the global weak Pareto optimality of Definition 2.5.1. It must
still be kept in mind that usually only locally weakly Pareto optimal solutions
are computationally available. Nevertheless, for the sake of brevity, we shall
usually refer only to weak Pareto optimality.
Let us state as a curiosity that if the feasible region is convex and the objec-
tive functions are quasiconvex with at least one strictly quasiconvex function,
the set of locally Pareto optimal solutions is a subset of the set of weakly Pareto
20 Part I - 2. Concepts

weakly Pareto
..-- optimal set
I Pareto optimal
setl

Z I

Figure 2.5.1. Weakly Pareto optimal vectors.

optimal solutions. This result is an immediate corollary of Theorem 2.2.4, where

we proved that under the above-mentioned assumptions all the locally Pareto
optimal solutions are also globally Pareto optimal.
The connectedness of the sets of Pareto optimal and weakly Pareto optimal
solutions has not been widely treated. Yet, this is an important feature because
it is often useful to know how well one can move continuously from one (weakly)
Pareto optimal solution to another.
The Pareto optimal set of an MOLP problem is proved to be connected in
Steuer (1986, pp. 158,220). It is stated in Warburton (1983), that the Pareto
optimal set is connected in convex multiobjective optimization problems. In
addition, Warburton shows that if the feasible region is convex and compact
and the objective functions are quasiconvex, then the set of weakly Pareto
optimal solutions is connected. The connectedness of the Pareto optimal set is
guaranteed for a certain subclass of quasiconvex functions. A noncompact case
is also studied in Warburton (1983).
The structure, including connectedness, of the sets of weakly, properly or
Pareto optimal solutions for nonconvex problems with two objective functions
is investigated in Tenhuisen and Wiecek (1996). A review of connectedness
results for Pareto optimality is given in Benoist (1998). Benoist also proves
that the Pareto optimal set is connected for continuous, strictly quasiconvex
objective functions (when transformed for minimization problems) defined on
a convex and compact set.
Although weakly Pareto optimal solutions are important for theoretical
considerations, they are not always useful in practice, because of the large
size of the weakly Pareto optimal set. However, they are often relevant from
a technical point of view because they are sometimes easier to generate than
Pareto optimal points. A more restrictive concept than Pareto optimality is
proper Pareto optimality (to be defined in Section 2.9).
 2.6. Value Function 21

2.6. Value Function

It is often assumed that the decision maker makes decisions on the basis of
an underlying function of some kind. This function is called a value function.

Definition 2.6.1. A function U: Rk -+ R representing the preferences of the

decision maker among the objective vectors is called a value function.

Let Zl and z2 E Z be two different objective vectors. If U(zl) > U(Z2), then
the decision maker prefers Zl to z2. If U(zl) = U(Z2), then the decision maker
finds the objective vectors equally desirable, that is, they are indifferent.
It must be pointed out that the value function is totally a decision maker-
dependent concept. Different decision makers may have different value functions
for the same problem.
Sometimes the term utility function is used instead of the value function.
Here we follow the common way of referring to value functions in deterministic
problems. The term utility function is reserved for stochastic problems (not to
be handled here). See Keeney and Raiffa (1976) for a more extended discussion
of both terms.
If we had at our disposal the mathematical expression of the decision
maker's value function, it would be easy to solve the multiobjective optimiza-
tion problem. The value function would simply be maximized by some method
of single objective optimization. The value function would offer a total (com-
plete) ordering of the objective vectors. However, there are several reasons why
this seemingly easy way is not generally used in practice. The most important
reason is that it is extremely difficult, if not impossible, for a decision maker
to specify mathematically the function behind her or his preferences. Secondly,
even if the function were known, it could be difficult to optimize because of its
possible complicated nature. An example of such situations is the nonconcavity
of the value function. In this case, only a local maximum may be found instead
of the global one. In addition, as pointed out in Steuer and Gardiner (1991), it
is not necessarily all to the good that optimizing the value function results in
a single solution. After specifying the value function, the decision maker may
have doubts about its validity. This is why (s)he may want to explore different
alternatives before selecting the final solution.
One more thing to keep in mind about value functions is that their existence
is not necessarily guaranteed. At least it may be restricting to assume that a
fixed and stable function can explain the behaviour and the preferences of the
decision maker.
Even though value functions are seldom explicitly used in solving multi-
objective optimization problems, they are very important in the development
of solution methods and as a theoretical background. In many multiobjective
optimization methods, the value function is assumed to be known implicitly
and the decision maker is assumed to make selections on this basis. In several
22 Part I - 2. Concepts

methods, convergence results are obtained by making certain assumptions, for

example, quasiconcavity about the implicit value function. In all, we can say
that value functions are usually more important to the analyst than to the
decision maker (see Zionts (1997a, b)).
Generally, the value function is assumed to be strongly decreasing. This
means that the preference of the decision maker will increase if the value of
an objective function decreases while all the other objective values remain un-
changed (i.e., less is preferred to more). This assumption is justified by Rosen-
thal (1985), who stresses that "Clearly, under the monotonicity assumption a
rational decision maker would never deliberately select a dominated point. This
is probably the only important statement in multiobjective optimization that
can be made without the possibility of generating some disagreement."
However, there are exceptions to this situation. Rosenthal mentions as an
(maximization) example the deer population, where more deer are usually pre-
ferred to fewer for aesthetic and recreational reasons, but not in the case when
the deer population is large enough to remove all the forest undergrowth.
The following theorem presents an important result concerning the solutions
of strongly decreasing value functions.

Theorem 2.6.2. Let the value function U: R k -t R be strongly decreasing.

Let U attain its maximum at z* E Z. Then z* is Pareto optimal.

Proof. Let z· E Z be a maximal solution of a strongly decreasing value func-

tion U. Let us assume that z· is not Pareto optimal. Then there exists an
objective vector z E Z such that Zi :s
zi for all i = 1, ... , k and Zj < z; for
at least one index j. Because U is strongly decreasing, we have U(z) > U(z*).
Thus U does not attain its maximum at z*. This contradiction implies that z*
is Pareto optimal. 0

Different properties and forms of value functions are widely treated in Hem-
ming (1978). Some references handling the existence of value functions are listed
in Stadler (1979) where different value functions are also presented.
The way a final solution was earlier defined means that a solution is final if
it maximizes the decision maker's value function. Sometimes another concept,
that of the satisficing solution, is distinguished.
Satisficing solutions are connected with so-called satisficing decision making.
Satisficing decision making means that the decision maker does not intend to
maximize any general value function but tries to achieve certain aspirations.
A solution which satisfies all the aspirations of the decision maker is called
a satisficing solution. In the most extreme case, one can define a solution to
be satisficing independent of whether it is Pareto optimal or not. Here we,
however, always assume that a satisficing solution is Pareto optimal or at least
weakly Pareto optimal.
 2.7. Efficiency 23

It is important to realize that regardless of the existence of an underlying

value function, a general assumption still is that that less is preferred to more
by the decision maker, that is, lower objective function values are preferred to
higher. This assumption is usually made even in methods not involving value
functions in any way. Thus, assuming that less is preferred to more is a more
general assumption than assuming a strongly decreasing value function,

2.7. Efficiency

It is possible to define optimality in a multiobjective context in more general

ways than by Pareto or weak Pareto optimality. Let us have a pointed convex
cone D defined in R k. This cone D is called an ordering cone and it is used to
induce a partial ordering on Z. Let us have two objective vectors, Zl and Z2 E Z.
An objective vector Zl dominates Z2, denoted by zl '5.D Z2, if Z2 - Zl E D and
Zl -I- Z2, that is, Z2 - Zl E D \ {OJ. The same can also be written as Z2 E Zl + D
and zl -I- z2, that is, z2 E Zl + D \ {OJ as illustrated in Figure 2.7.1.

Figure 2.7.1. Domination induced by a cone D.

We can now present a definition of optimality based on domination, which

is an alternative to the definitions previously given. When an ordering cone
is used in defining optimality, then the term efficiency will be used in what
follows.

Definition 2.7.1. Let D be a pointed convex cone. A decision vector x* E S

is efficient (with respect to D) if there does not exist another decision vector
xES such that rex) '5.D r(x·).
An objective vector z* E Z is efficient if there does not exist another objec-
tive vector z E Z such that z '5.D z*.
24 Part I -- 2. Concepts

This definition means that a vector is efficient (nondominated) if it is not

dominated by any other feasible vector. The definition above can be formulated
in many ways. If we substitute ~D for its definition, we have the condition in
the form 0 f:. z* - zED or z* - zED \ {O} (see Corley (1980)).
Other equivalent formulations are, for instance, z* E Z is efficient if (Z -
z*) n (-D) = {O} (see Pascoletti and Serafini (1984) and Weidner (1988)), if
(z* - D \ {O}) n Z = 0 (see Tapia and Murtagh (1989) and Wierzbicki (1986b))
or if (z* - D) n Z = z* (see Chen (1984) and Jahn (1987)).
Let us give an alternative formulation to Definition 2.7.1 using one of the
equivalent representations.

Definition 2.7.2. Let D be a pointed convex cone. A decision vector x* E S

is efficient (with respect to D) if there does not exist another decision vector
xES such that f(x*) E f(x) + D \ {O}, that is, (f(x*) - D \ {O}) n Z = 0.
An objective vector z* E Z is efficient if there does not exist another objec-
tive vector z E Z such that z* E z + D \ {O}, that is, (z* - D \ {O}) n Z = 0.

Different notions of efficiency are collected in Ester and Tr6ltzsch (1986).

They provide several auxiliary problems in the interests of obtaining efficient
solutions.

Remark 2.7.3. The above definitions are equivalent to Pareto optimality if

D = Rt
(see Figure 2.7.2).

Figure 2.7.2. Pareto optimality with the help of cone Ri.

When Pareto optimality or efficiency is defined with the help of ordering
cones, it is trivial to verify that Pareto optimal or efficient objective vectors
always lie on the boundary of the feasible objective region Z.
Instead of a cone D, which is constant for every objective vector, we can use
a point-to-set map D from Z into Rk to represent the domination structure. In
this case domination is dependent on the current objective vector. For details
 2.8. From One Solution to Another 25

of ordering cones, see Sawaragi et al. (1985, pp. 25-31) and Yu (1974, 1985,
pp. 163-209).
Theorem 2.6.2 gives a relationship between Pareto optimal solutions and
value functions. Relations can also be established between efficient solutions and
value functions. To give an idea of them, let us consider a pseudoconcave value
function U. According to pseudoconcavity whenever '\7U(Zl V(Z2 - Zl) ~ 0, we
have U(Z2) ~ U(Zl). We can now define an ordering cone as a map D(z) =
{d E R k I '\7 U (z) T d ~ o}. This ordering cone can be used to determine efficient
solutions. Note that if we have a value function, we can derive its domination
structure, but not generally vice versa. See Yu (1974) for an example.
Weakly efficient decision and objective vectors can be defined in a corre-
sponding fashion to efficient ones. If the set Z of objective vectors is ordered
by an ordering cone D, weakly efficient vectors may be characterized in the
following way (see Jahn (1987) and Wierzbicki (1986b»:

Definition 2.7.4. Let D be a pointed convex cone. A decision vector x' E S

is weakly efficient (with respect to D) if there does not exist another decision
vector XES such that f(x*) E f(x) + int D, that is, (f(x*) - int D) n Z = 0.
An objective vector z* E Z is weakly efficient if there does not exist another
objective vector z E Z such that z* E z + int D, that is, (z* - int D) n Z = 0.

An alternative formulation is that an objective vector z* E Z is weakly

efficient if (Z - z*) n (-int D) = 0 (see Sawaragi et al. (1985, pp. 33-34».
Connectedness of the sets of weakly efficient and efficient points is studied
in Helbig (1990) whereas Luc (1989, pp. 148-154) treats particularly weakly
efficient sets in convex problems where the objective functions are quasiconvex.
In addition, connectedness results for efficient points in multiobjective combi-
natorial problems are given in Ehrgott and Klamroth (1997).
In the following, we mostly settle for treating Pareto optimality. Some ex-
tensions related to efficiency are only mentioned in passing.
Thus far, we have defined Pareto and weak Pareto optimality and more
general efficiency and weak efficiency. Proper Pareto optimality and proper
efficiency are yet to be introduced. To clarify their practical meaning and for
other further purposes we must first, however, define trade-off's and marginal
rates of substitution.

2.8. From One Solution to Another

Thade-offs and marginal rates of substitution are related to changes in the

objective values when we move from one solution to another. Trade-offs are
defined mathematically whereas marginal rates of substitution depend on the
decision maker.
26 Part I - 2. Concepts

2.8.1. Trade-Offs

We have several concepts involved in trading off. A trade-off reflects the

ratio of change in the values of the objective functions concerning the increment
of one objective function that occurs when the value of some other objective
function decreases. In the following definitions we have i, j = 1, ... , k, i =I- j.

Definition 2.8.1. (From Chankong and Haimes (1983b)) Let Xl and x 2 E S

be two decision vectors and let f(Xl) and f(x 2 ) be the corresponding objective
vectors, respectively. We denote the ratio of change between the functions h
and Ji by
A A (I 2) h(x l ) - li(x 2 )
ij = ij X,X = Ji(xl) _ Ji(x 2 )'
where Ji (Xl) - fj (x 2 ) =I- O.
Now, Aij is called a partial trade-off, involving hand Ji between Xl and
x 2 if !I(XI) = II(x2) for alll = 1, ... ,k, 1 =I- i,j. If !I(xl) =I- fl(x2) for at least
one 1 = l, ... ,k, and 1 =I- i,j, then Aij is called a total trade-off, involving h
and fj between Xl and x 2 .

Note that in the case of two objective functions there is no difference be-
tween partial and total trade-offs. If partial trade-offs are presented to the de-
cision maker, (s)he can compare changes in two objective functions at a time.
This is usually a more comfortable procedure than comparing several objec-
tives. If the points Xl and x 2 are Pareto optimal, then there always exist some
objective functions hand Ij for which the trade-off is negative. A concept
related to the trade-off is the trade-off rate.

Definition 2.8.2. (From Chankong and Haimes (1983b» Let x* E S be a

decision vector and let d* be a feasible direction emanating from x*. The total
trade-off rate at x*, involving Ii and Ii along the direction d*, is given by
Aij = Aij(X*,d*) = 0'-t0+
lim Aij(X* +ad*,x*).

If d* is a feasible direction so that there exists 0> 0 satisfying !I(x* + ad*) =

II(x*) for alll = 1, ... , k, 1 =I- i,j and for all 0::; a ::; 0, then the corresponding
.Aij is called a partial trade-off rate.

Remark 2.8.3. If the objective functions are continuously differentiable, then

\7fieX*)Td*
Aij = \7fj(x*)Td*'
where the denominator differs from zero.

For continuously differentiable objective functions we can alternatively give

the following definition.
 2.8. From One Solution to Another 27

Definition 2.8.4. Let the objective functions be continuously differentiable

at a decision vector x' E S. Then a partial trade-off rate at x*, involving fi
and fJ, is given by

Differing from the idea of the definitions above, a so-called global trade-
off is defined in Kaliszewski and Michalowski (1995, 1997). A global trade-off
involves two objective functions and one decision vector which does not have to
be Pareto optimal. It is the largest pairwise trade-off of two objective functions
for one decision vector. Let us consider x* E S and modify the definitions for
minimization problems. We define a subset of the feasible decision vectors in
the form

Sj(x*) = {x E S I fJ(x) > fJ(x*), fi(X) ~ j;(x*), for i = 1, ... , k, if:. n.

Now we can introduce global trade-offs.

Definition 2.8.5. (From Kaliszewski and Michalowski (1995,1997)) Let x* E

S be a decision vector. We denote a global trade-off between the functions fi
and fJ by
G G * Ji(x*) - fi(X)
Aij = Aij(X ) = sup f·() f.( *).
XESJ(x*) J x - J X

If Sj(x*) = 0, then Af;(x*) = -00 for every i = 1, ... , k, i f:. j.

A generalized definition of trade-offs in terms of tangent cones, meaning

feasible directions, in the objective space is presented in Henig and Buchanan
(1994, 1997). These generalized trade-off directions can be used for calculat-
ing trade-off rates at every Pareto optimal point of a convex multiobjective
optimization problem.
Note that trade-offs are defined mathematically and the decision maker
cannot affect them. If we take into consideration the opinions of the decision
maker, we can define indifference curves and marginal rates of substitution.

2.8.2. Marginal Rate of Substitution

It is said that two feasible solutions are situated on the same indifference
curve (or isopreference curve) if the decision maker finds them equally desir-
able, that is, neither of them is preferred to the other one. This means that
indifference curves are contours of the underlying value function. There may
also be a 'wider' indifference band. In this case we do not have any well-defined
boundary between preferences, but a band where indifference occurs. This con-
cept is studied in Passy and Levanon (1984).
28 Part I - 2. Concepts

For any two solutions on the same indifference curve there is a trade-off
involving a certain increment in the value of one objective function (/j) that
the decision maker is willing to tolerate in exchange for a certain amount of
decrement in some other objective function (h) while the preferences of the
two solutions remain the same. This is called the marginal rate of substitution.
This kind of trading between different solutions is characteristic of multiobjec-
tive optimization problems when moving from one Pareto optimal solution to
another. The marginal rate of substitution (sometimes also called indifference
trade-off) is the negative of the slope of the tangent to the indifference curve
at a certain point.

Definition 2.8.6. A marginal rate of substitution mij = mij (x*) represents

the preferences of the decision maker at a decision vector x· E S. It is the
amount of decrement in the value of the objective function h that compensates
the decision maker for the one-unit increment in the value of the objective
function /j, while the values of all the other objectives remain unaltered.

Note that in the definition the starting and the resulting objective vectors lie
on the same indifference curve and i, j = 1, ... , k, i i- j.
It can be stated that the final solution of a multiobjective optimization
problem is a Pareto optimal solution where the indifference curve is tangent to
the Pareto optimal set. This tangency condition means finding an indifference
curve intersecting the feasible objective region that is farthest to the southwest.
This property is illustrated in Figure 2.8.1.

indifference curve

Figure 2.8.1. The final solution.

 2.9. Proper Pareto Optimality 29

Remark 2.8.7. If the partial derivatives exist, then

.. ( *) _ aU(f(x*)) /aU(f(X*))
mtJ x - afJ ali'

If the Pareto optimal set is smooth (that is, at every Pareto optimal point
there exists a unique tangent), we have the following result. When one examines
the definition of a trade-off rate at some point, one sees that it is the slope of
the tangent of the Pareto optimal set at that point. We can also define that
when a Pareto optimal solution is a final solution, then the tangents of the
indifference curve and the Pareto optimal set coincide at it, that is,

(2.8.1) -mij = Aij for all i,j = 1, ... , k, i t= j.

Thus, with the help of the negative of the marginal rate of substitution and
the trade-off rate one can get a local linear approximation of the indifference
curve and the Pareto optimal set, respectively.
Usually, one of the objective functions is selected as a reference function
when trade-offs and marginal rates of substitution are treated. The trade-offs
and the marginal rates of substitution are generated with respect to it. In the
notations above, Ii is the reference function. When co-operating with decision
makers, it is important to select the reference function in a meaningful way. An
important criterion in the selection is, for example, that the reference function
is in familiar units or that it is dominant.

2.9. Proper Pareto Optimality

Kuhn and Thcker were the first to note that some of the Pareto optimal solu-
tions had undesirable properties (see Kuhn and Tucker (1951)). To avoid such
properties, they introduced properly Pareto optimal solutions and suggested
that Pareto optimal solutions be divided into properly and improperly Pareto
optimal ones. The idea of properly Pareto optimal solutions is that unbounded
trade-offs between objectives are not allowed. Practically, a properly Pareto op-
timal solution with very high or very low trade-offs does not essentially differ
from a weakly Pareto optimal solution for a human decision maker.
There exist several definitions for proper Pareto optimality. The idea is
easiest to understand from the following definition.

Definition 2.9.1. (From Geoffrion (1968)) A decision vector x* E S is prop-

erly Pareto optimal (in the sense of Geoffrion) if it is Pareto optimal and if
there is some real number M > 0 such that for each Ii and each xES sat-
isfying fi(x) < Ii(x*), there exists at least one Ij such that fJ(x*) < fJ(x)
and
30 Part I - 2. Concepts

fi(X*) - fi(X) < M.

fj(x) - fJ(x*) -
An objective vector z* E Z is properly Pareto optimal if the decision vector
corresponding to it is properly Pareto optimal.

In other words, a solution is properly Pareto optimal if there is at least one

pair of objectives for which a finite decrement in one objective is possible only
at the expense of some reasonable increment in the other objective.
Geoffrion's definition can be generalized 80 that the upper bound is a pos-
itive function M(x) instead of a constant (see Mishra (1996) and Mishra and
Mukherjee (1995)). This leads to the definition of conditional proper Pareto
optimality.
A method for obtaining all the properly Pareto optimal solutions satisfy-
ing prescribed marginal rates of substitution in the convex case is proposed
in Geromel and Ferreira (1991). Upper estimates for properly Pareto optimal
solutions are given as well.
Durier (1988) studies the relationships between Pareto optimal and properly
Pareto optimal sets in a convex case. One of the results is that if the set of
properly Pareto optimal solutions is closed, then the two sets are equal. A
property called a locally flat surface, which guarantees the very same equality
in convex and differentiable problems, is presented in Zhou et al. (1993).
Results concerning Pareto optimal and properly Pareto optimal solutions
are collected in Gal (1986). In Chew and Choo (1984), it is proved that every
Pareto optimal solution is also properly Pareto optimal for a nonlinear prob-
lem involving. only pseudolinear functions (i.e., differentiable functions which
are both pseudo convex and pseudoconcave). The results of Chew and Choo can
be considered special cases of more general results presented in Weir (1990). In
Gulati and Islam (1990), it is shown that the preceding result can be generalized
by assuming quasiconvexity of the active constraints (of the form g(x) ~ 0)
with some regularity properties. Pseudolinearity is extended by defining semilo-
cally pseudolinear functions in Kaul et al. (1988).
We shall present some results concerning the relationships between Pareto
optimal, weakly and properly Pareto optimal solutions in the context of solution
methods in Part II.
Next, we introduce €-proper Pareto optimality, which is easy to illustrate
graphically.

Definition 2.9.2. (From Wierzbicki (1980b» A decision vector x* E Sand

the corresponding objective vector z· E Z are €-properly Pareto optimal if
(z* - R! \ {O}) n Z = 0,

where R; = {z E Rk I dist (z, R~) :5 €llzll} or R; = {z E Rk I maxi=l ..... k Zi +

€ L~=l Zi ~ O} and € > 0 is a predetermined scalar.
 2.9. Proper Pareto Optimality 31

Note that this definition differs from that of Pareto optimality so that a
larger set R~ is used instead of the set Ri. The set of e-properly Pareto optimal
solutions is depicted in Figure 2.9.1 and denoted by a bold line. The solutions
are obtained by intersecting the feasible objective region with a blunt cone.
The end points of the Pareto optimal set, Zl and Z2, have also been marked to
ease the comparison.
An alternative formulation of Definition 2.9.2 is that a decision vector x* E
S and the corresponding z* E Z are e-properly Pareto optimal if (z* -R~)nZ =
z*. (The definition can be generalized into proper efficiency by using a convex
cone D such that Ri C int D U {o}.)

Z2

z*- R~
z*- R! -------

Z I

Figure 2.9.1. The set of e-properly Pareto optimal solutions.

An interesting aspect of e-properly Pareto optimal solutions is that the

trade-offs are bounded bye and lie (see Wierzbicki (1986a, b». We return to
this concept in Section 3.5 of Part II.
Before we continue with the original definition of Kuhn and Tucker, we
should mention briefly another way of decreasing the set of Pareto optimal
solutions according to Liu (1996)_ There, z* E Z is called f-Pareto optimal if
(z* - (Ri +(;) \ {OJ) n Z = 0, where (; E Ri.
Let us for a while assume that the feasible region is defined with the
help of inequality constraints. In other words, S = {x E R n I g(x) =
(gl(X),g2(X), ... ,gm(x»T :::; OJ. In addition, all the objective and the con-
straint functions are assumed to be continuously differentiable at every point
xES. Thus the next definition is not applicable to nondifferentiable multiob-
jective optimization problems.

Definition 2.9.3. (From Kuhn and Tucker (1951» A decision vector x* E S

is properly Pareto optimal (in the sense of Kuhn and Tucker) if it is Pareto
optimal and if there does not exist any vector d ERn such that
32 Part 1 - 2, Concepts

for all i = 1, ... , k, for some j

and
Vy'(X*)T d ~ 0
for alll satisfying YI(X*) = 0, that is, for all active constraints at x*.
An objective vector z* E Z is properly Pareto optimal if the decision vector
corresponding to it is properly Pareto optimal.

Kuhn and Tucker also derived necessary and sufficient conditions for proper
Pareto optimality in Kuhn and Tucker (1951), Those conditions will be pre-
sented in the next section,
A comparison of the definitions of Kuhn and Tucker and Geoffrion is pre-
sented in Geoffrion (1968), For example, in convex cases the definition of Kuhn
and Thcker implies the, definition of Geoffrion. The reverse result is valid if
the so-called Kuhn-Tucker constraint qualification (see Definition 3,1.3) is sat-
isfied. The relationships of these two definitions are also treated, for example,
in Sawaragi et al. (1985, pp, 42-46), Several practical examples are given in
Tamura and Arai (1982) to illustrate the fact that properly Pareto optimal so-
lutions according to the definitions of Kuhn and Tucker and Geoffrion (and one
more definition by Klinger; see Klinger (1967)) are not necessarily consistent.
Conditions under which (local) proper Pareto optimality in the sense of Kuhn
and Tucker implies (local) proper Pareto optimality in the sense of Geoffrion
are proved as well. More mathematical results concerning the properties and
the relationships of the definitions of Kuhn and Tucker, Geoffrion and Klinger
are given in White (1983a).
Borwein (1977) and Benson (1979a) have both defined proper efficiency
when a closed, convex cone D is used as an ordering cone. Borwein's definition
is based on tangent cones and Benson's on so-called projecting cones. Let us
mention that proper efficiency according to Benson's definition implies proper
efficiency in the sense of Borwein, (The reverse is valid in convex cases.) These
two definitions are generalized in Henig (1982b) using convex ordering cones.
The ordering cone D used in defining efficiency is utilized in the following,

Definition 2.9.4. (From Henig (1982b)) Let D be a pointed convex cone,

A decision vector x· E S is properly efficient (in the sense of Henig) (with
respect to D) if there does not exist another decision vector xES such that
f(x*) E f(x) + E \ {O}, that is, (f(x*) - E \ {O}) n Z = 0 for some convex cone
E such that D \ {O} C int E,
An objective vector z* E Z is properly efficient if there does not exist
another objective vector z E Z such that z* E z + E \ {O}, in other words,
(z· - E \ {O}) n Z = 0 with E as above.
 2.10. Pareto Optimality Tests with Existence Results 33

The desirable property is valid also here: if a point is properly efficient, it

is also efficient. Notice that Definition 2.9.4 is related to Definition 2.9.2 when
we set D = Ri.
As pointed out, different definitions of proper efficiency (and proper Pareto
optimality) are not equivalent with each other but they have connections. The
relationships between the definitions in the sense of Kuhn and Tucker, Geof-
frion, Borwein, Benson and Henig are analysed in Sawaragi et al. (1985, pp. 39-
44). For instance, Geoffrion's and Benson's definitions are equal when D = Ri
(see also Benson (1983)). On the other hand, Definition 2.9.4 is equivalent to
Benson's definition if the ordering cone D is closed and its closure is pointed.
For further analysis we refer to Sawaragi et al. (1985, pp. 39-44).
In Henig (1982b), necessary and sufficient conditions for the existence of
properly efficient solutions are given.
Let us finally mention that a new kind of proper efficiency, called super
efficiency, is suggested in Borwein and Zhuang (1991, 1993).
In the following, proper Pareto optimality is understood in the sense of
Geoffrion unless stated otherwise.

2.10. Pareto Optimality Tests with Existence Results

Let us have a look at how the Pareto optimality of feasible decision vectors
can be tested. The procedures presented can also be used to find an initial
Pareto optimal solution for (interactive) solution methods or to examine the
existence of Pareto optimal and properly Pareto optimal solutions.
Specific results for MOLP problems are presented in Ecker and Kouada
(1975). They are generalized for nonlinear problems with the help of duality
theory in Wendell and Lee (1977). The treatment is based on an auxiliary
problem
k
minimize Lli(x)
(2.10.1) i=l
subject to li(x) ~ h(fc.) for all i = 1, ... , k,
xES,

where fc. is any vector in S. Let us denote the optimal objective function value
by ¢(fc.).

Theorem 2.10.1. Let a decision vector x* E S be given. The vector x*

is Pareto optimal if and only if it is a solution of problem (2.10.1) so that
¢(x*) = 2::=1 li(X*).
On the other hand, let x* E S be a solution of problem (2.10.1). Then x*
is Pareto optimal and li(x*) ~ li(fc.) for all i = 1, ... , k.
34 Part I -- 2. Concepts

Proof. See Wendell and Lee (1977).

Theorem 2.lO.1 means that if problem (2.lO.1) has an optimal solution for
some x E 5, then either x is Pareto optimal or the optimal solution of (2.10.1)
is.
When studying the (primal) problem (2.lO.1) and its dual, a duality gap is
said to occur if the optimal value of the primal problem is not equivalent to
the optimal value of the dual problem.

Theorem 2.10.2. Let a decision vector x

E 5 be given and assume that
¢(x) = -00. Then some x* E 5 is Pareto optimal only if there is a duality gap
between the primal (2.lO.1) and its dual problem at x*. If such a gap exists,
the optimal solution of (2.10.1) is Pareto optimal.

Proof. See Wendell and Lee (1977).

The significance of Theorem 2.10.2 is that precluding duality gaps the

nonexistence of Pareto optimal points is characterized by the condition that
¢(x) = -00 for some x E 5. It can also be proved that if a multiobjective
optimization problem is convex and if ¢(x) = -00 for some x E 5, then no
properly Pareto optimal solutions exist. See the details in Wendell and Lee
(1977).
Tests for Pareto optimality and the existence of Pareto optimal and properly
Pareto optimal solutions are also investigated in Benson (1978). The results can
be combined into the following theorem.

Theorem 2.10.3. Let a decision vector x* E 5 be given. Solve the problem

k
maximize '2:: Ci

(2.10.2) subject to J;(x) + Ci = j;(x*) for all i = 1, ... , k,

ci~O for all i=l,... ,k,
x E 5,

where both x ERn and E E Ri are variables. Then the following results are
valid.
(1) The vector x* is Pareto optimal if and only if problem (2.10.2) has an
optimal objective function value of zero.
(2) If problem (2.10.2) has a finite nonzero optimal objective function value
obtained at a point x, then x is Pareto optimal.
(3) If the multiobjective optimization problem is convex and if problem
(2.10.2) does not have a finite optimal objective function value, then
the set of properly Pareto optimal solutions is empty.
 2.10. Pareto Optimality Tests with Existence Results 35

(4) If in addition to the conditions in (3), the set {z E Rk I z :S f(x) for

some XES} is closed, then the Pareto optimal set is empty.

Proof. See Benson (1978) or Chankong and Haimes (1983b, pp. 151-152).

Problem (2.10.2) is a popular way of checking Pareto optimality and of

generating Pareto optimal solutions. However, sometimes equality constraints
cause computational difficulties. Therefore it is useful to note that the equalities
in (2.10.2) can be replaced with inequalities fi(X) + Ci :'S fi(X*) for all i =
1, ... , k without affecting the generality of the results presented.
Two simple tests are suggested in Brosowski and da Silva (1994) for deter-
mining whether a given point is (locally) Pareto optimal or not. The tests are
not based on any scalarizing functions but linear systems of equations. There
are, however, several limitations. The objective functions are assumed to be
continuously differentiable and their number has to be strictly larger than the
number of variables. Further, no constraints can be included. Finally, the tests
may also fail as demonstrated in Brosowski and da Silva (1994).
It is proved in Sawaragi et al. (1985, p. 59), that Pareto optimal solutions
exist to multiobjective optimization problems where all the objective functions
are lower semicontinuous (more general than continuity) and the feasible region
is compact. Several ways of determining the Pareto optimality of a particular
point in an MOLP problem are presented in Eiselt et al. (1987). They all apply
to special situations. Further, the existence of Pareto optimal solutions when
there is an infinite number of objective functions is considered in Alekseichik
and Naumov (1981).
The existence of weakly Pareto optimal solutions in convex differentiable
multiobjective optimization problems is treated in Deng (1998a). In addition,
the compactness of the weakly Pareto optimal set is considered. The nonempti-
ness of the Pareto optimal and the weakly Pareto optimal sets in convex prob-
lems is also characterized in Deng (1998b).
As mentioned, auxiliary problems (2.10.1) and (2.10.2) can be used to pro-
duce Pareto optimal solutions, for example, from weakly Pareto optimal solu-
tions. However, in some practical problems it is very expensive to carry out
these additional optimizations. An alternative is suggested in Helbig (1991). If
optimality is defined by an ordering cone, efficient solutions can be generated
by perturbing this cone. In other words, using a method producing weakly effi-
cient solutions with respect to the perturbed cone gives results that are efficient
vis a vis the original problem.
The existence and the characterization of efficient solutions with respect
to ordering cones are studied in Henig (1982a), and the existence of efficient
solutions in linear spaces is treated in Borwein (1983). In addition, the existence
of weakly and properly efficient (in the sense of Borwein) and efficient solutions
in the presence of ordering cones is studied in Jahn (1986b). The existence of
36 Part 1 - 2. Concepts

efficient solutions is also treated in Cambini and Martein (1994) by introducing

so-called quasi-D-bounded sets.
A phenomenon called complete efficiency occurs when every feasible deci-
sion vector of a multiobjective optimization problem is Pareto optimal. Tests
are presented in Benson (1991) to check for complete efficiency in linear and
nonlinear cases. A significant saving of computational efforts can be attained if
the problem is tested for complete efficiency before it is solved. If the problem is
completely efficient, no time, effort and special machinery for generating some
or all of the Pareto optimal solutions is needed. Anyway, no solution algorithm
exists which first checks for complete efficiency. The frequency of completely ef-
ficient problems among multiobjective optimization problems deserves further
study. It may be more common than is generally thought, especially with spe-
cial problem types, for example, when the feasible region S has no interior, as
Benson points out. Transportation problems feature in this category. Complete
efficiency is also treated in Weidner (1990).
One further area of research concerns the domination property. It refers to
the situation where there always exists an efficient solution that is superior to
any nonefficient solution, that is, for each xES and corresponding z E Z there
exists an efficient point x· and corresponding z' such that z - z* ED, where
D is the ordering cone. Validity conditions for the domination property are
examined in Benson (1983). The results of Benson are corrected and necessary
and sufficient conditions for the domination property to hold are supplied in
Luc (1984a). The domination property and its sufficient conditions are also
treated in Henig (1986). Further, it is demonstrated that the existence of an
efficient solution, the existence of a properly efficient solution, and the dom-
ination property are equivalent in solving convex problems. The domination
property in infinite-dimensional spaces and for the sum of two sets is handled
in Luc (1990).
The last concept to be mentioned here is the redundancy of objective func-
tions. In MOLP cases this can be understood as linear dependency. In other
words, an objective function is redundant if it does not affect the Pareto optimal
set (see Gal and Leberling (1977)). This is not necessarily valid for nonlinear
problems or in connection with interactive methods. For both of these, it is
important to define redundancy on the basis of conflict between the objectives,
which is why in Agrell (1997), an objective function is defined as redundant if
it is not in conflict with any other objective function. Agrell suggests a proba-
bilistic Monte-Carlo simulation-based redundancy test for nonlinear problems
where the correlation of the objective function is observed. Redundancy checks
are important because it may ease the burden of the decision maker if redun-
dant objectives are eliminated.
 3. THEORETICAL
BACKGROUND

We present a set of optimality conditions for multiobjective optimization

problems. Because the conditions are different for differentiable and non differ-
entiable problem, they are handled separately.

3.1. Differentiable Optimality Conditions

Optimality conditions are an important sector in optimization. As else-

where, we restrict the treatment also here to finite-dimensional Euclidean
spaces. We consider problems of the form

(3.1.1)
minimize {II(x), h(x), ... , fk(x)}
subject to xES = {x E R n I g(x) = (gl(X),g2(X), ... ,gm(x)f ~ o}.
We denote the set of active constraints at a point x* by

l(x·) = {j E {1, ... ,m} I gj(x*) = O}.

We assume in this section that the objective and the constraint functions are
continuously differentiable. In Section 3.2 we treat non differentiable functions.
Similar optimality results are also handled, for example, in Da Cunha and
Polak (1967), Kuhn and Tucker (1951), Marusciac (1982), Simon (1986) and
Yu (1985, pp. 35-38, 49-50). In order to highlight the ideas, the theorems are
here presented in a simplified form as compared to the general practice. For
this reason, the proofs have been modified.

3.1.1. First-Order Conditions

We begin with a necessary condition of the Fritz John type.

Theorem 3.1.1. (Fr'itz John necessary condition for Pareto optimality) Let
the objective and the constraint functions of problem (3.1.1) be continuously
differentiable at a decision vector x* E S. A necessary condition for x* to be

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
38 Part I - 3. Theoretical Background

Pareto optimal is that there exist vectors 0 ::; ~ E Rk and 0 ::; pERm for
which (>",p) =1= (0,0) such that
k m
(1) LAiV'fi(X*) + LJ.ljV'gj(x*) = 0
i=1 j=1

(2) J.ljgj(x*) =0 for all j = 1, ... , m.

Proof. See, for instance, Da Cunha and Polak (1967).

We do not present the proof here because it is quite extensive. The theorem
can be considered a special case of the corresponding theorem for nondifferen-
tiable problems, which is proved in Subsection 3.2.1. For convex problems, nec-
essary optimality conditions can be derived by using separating hyperplanes.
This is realized, for example, in Zadeh (1963). A separation theorem is also
employed in the proof of the general case in Da Cunha and Polak (1967).

Corollary 3.1.2. (Fritz John necessary condition for weak Pareto optimality)
The condition of Theorem 3.1.1 is also necessary for a decision vector x* E S
to be weakly Pareto optimal.

The difference between Fritz John type and Karush-Kuhn-Tucker type op-
timality conditions in single objective optimization is that the multiplier (A) of
the objective function is assumed to be positive in the latter case. This elimi-
nates degeneracy since it implies that the objective function plays its important
role in the optimality conditions. To guarantee the positivity of A, some regu-
larity has to be assumed in the problem. Different regularity conditions exist
and they are called constraint qualifications.
In the multiobjective case it is equally important that all the multipliers
of the objective functions are not equal to zero. Sometimes the multipliers
connected to Karush-Kuhn-Tucker optimality conditions are called Karush-
K uhn- Tucker multipliers. This concept will be used later.
In order to present the Karush-Kuhn-Tucker optimality conditions we must
formulate some constraint qualification. From among several different alterna-
tives we here present the so-called Kuhn-Tucker constraint qualification.

Definition 3.1.3. Let the constraint functions gj of problem (3.1.1) be con-

tinuously differentiable at X* E S. The problem satisfies the K uhn- Tucker con-
straint qualification at x* if for any d E Rn such that V'gj(x*)Td ::; 0 for
all j E J(x*), there exists a function a: [0,1) --7 Rn which is continuously
differentiable at 0, and some real scalar a > 0, such that

a(O) = x*, g(a(t))::; 0 for all 0::; t::; 1 and a'(O) = ad.
 3.1. Differentiable Optimality Conditions 39

Before we can continue, we write down the so-called Motzkin's theorem

of the alternative. It will be needed in the proof of the following necessary
condition.

Theorem 3.1.4. (Motzkin's theorem) Let A and C be given matrices. Then

either the system of inequalities

Ax < 0, Cx ~ 0

has a solution x, or the system

has a solution (A., 1'), but never both.

Proof. See, for example, Mangasarian (1969, pp. 28-29).

Now we can formulate the Karush-Kuhn-Tucker necessary condition for

Pareto optimality.

Theorem 3.1.5. (Karush-Kuhn-Tucker necessary condition for Pareto opti-

mality) Let the assumptions of Theorem 3.1.1 be satisfied by the K uhn-Tucker
constraint qualification. Theorem 3.1.1 is then valid with the addition that
A.~ O.

Proof. Let x' E S be Pareto optimal. The idea of this proof is to apply
Theorem 3.1.4. For this reason we prove that there does not exist any d E R n
such that
V' f;(x*)T d < 0 for all i = 1, ... , k, and
(3.1.2)
V'gj(X,)T d :::; 0 for all j E J(x").

Let us on the contrary assume that there exists some d* E Rn satisfying

(3.1.2). Then from the Kuhn-Tucker constraint qualification we know that there
exists a function a: [0, 1J -+ R n which is continuously differentiable at 0 and
some real scalar a > 0 such that a(O) = x*, g(a(t)) :::; 0 for all 0 :::; t :::; 1 and
a/CO) = ad*.
Because the functions J; are continuously differentiable, we can approximate
fi(a(t» linearly as
fi(a(t» = f;(x') + V' fi(x*)T(a(t) - x*) + lIa(t) - x'II!p(a(t), x*)
= J;(x") + V'J;(x*f(a(t) - a(O» + Ila(t) - a(O)II!p(a(t),a(O»

= J;(x*) + tV'J;(x") T (a(o + t~ - a(o») + Ila(t) - a(O) 11!p(a(t), a(O»,

where !p(a(t) , a(O» -+ 0 as Ila(t) - a(O)1I -+ O. As t -+ 0 tends Ila(t) - a(O)11 to

zero and (a(O + t) - a(O»jt -+ a/CO) = ad'.
40 Part I - 3. Theoretical Background

After utilizing the assumption V' Ji(X*)T d* < 0 for all i = 1, ... , k (and
t ~ 0), we have Ji(a(t)) < fi(x·) for all i = 1, ... , k for a sufficiently small t.
This contradicts the Pareto optimality of x·.
Thus we have proved statement (3.1.2). Now we conclude from Theorem
3.1.4 that there exist multipliers Ai ~ 0 for i = 1, ... , k, .\ =1= 0, and Jlj ~ 0
for j E l(x·) such that E~=l AiV' Ji(x·) + EjEJ(x+) JljV'gj(x*) = O. We obtain
statement (1) of Theorem 3.1.1 by setting Jlj = 0 for all j E {I, ... , m} \ l(x·).
If gj(x·) < 0 for some j = 1, ... ,m, then according to the above setting
Jlj = 0 and equalities (2) of Theorem 3.1.1 follow. 0

A proof basically similar but different in realization is presented in Marus-

ciac (1982).

Corollary 3.1.6. (Karush-Kuhn-Tucker necessary condition for weak Pareto

optimality) The condition of Theorem 3.1.5 is also necessary for a decision
vector x· E S to be weakly Pareto optimal.

Constraint qualifications based on the linear independence of gradient vec-

tors are stated in Da Cunha and Polak (1967). Other constraint qualifications
are collected in Simon (1986). In addition, a new constraint qualification for
convex problems is introduced in Zhou et al. (1993).
If the multiobjective optimization problem is convex, then we can state
a sufficient condition for Pareto optimality. Let us first recall the sufficient
condition of optimality in the single objective case.

Theorem 3.1.7. (Karush-Kllhn-Tllcker sufficient condition for optimality)

A sufficient condition for a point x· ERn to be a (global) minimum of the
problem
minimize Ji (x)
subject to g(x) = (gl (x), g2(X), . .. , gm(x)f :s 0,
where the objective function fi: Rn -+ R and the constraints gj: R n -+ R,
j = 1, ... , m, are convex and continuously differentiable at x·, is that there
exist multipliers 0 :S pERm such that
m

j=1

(2) Jljgj(x*) =0 for all j = 1, ... , m.

Proof. See, for example, Simon (1986).

Now we can extend Theorem 3.1. 7 for the multiobjective case.

Theorem 3.1.8. (Karush-Kuhn-Tucker sufficient condition for Pareto opti-

mality) Let the objective and the constraint functions of problem (3.1.1) be
 3.1. Differentiable Optimality Conditions 41

convex and continuously differentiable at a decision vector x* E S. A sufficient

condition for x* to be Pareto optimal is that there exist multipliers 0 < .\ E R k
and 0 ~"E R m such that
k m
(1) L AiV'h(x*) + L f.ljV'gj(X*) = 0
i=1 j=1

(2) /1jgj(x*) =0 for all j = 1, ... , m.

Proof. Let the vectors.\ and" be such that the conditions stated are satisfied.
We define a function F: Rn ---+ R as F(x) = 2:7=1 Adi(X), where XES.
Trivially F is convex because all the functions fi are and we have .\ > o.
Now from statements (1) and (2), we obtain V'F(x*) + 2:;:1 f.ljV'gj(X*) = 0
and f.ljgj (x*) = 0 for all j = 1, ... , m. Thus, according to Theorem 3.1. 7, the
sufficient condition for F to attain its minimum at x* is satisfied. So F(x*) ~
F(x) for all xES. In other words,
k k
(3.1.3) L Adi(X*) ~ L Adi(X)
i=1 i=1

for all xES.

Let us assume that x* is not Pareto-optimal. Then there exists some point
xES such that h(x) ~ fi(X*) for all i = 1, ... , k and for at least one index
j is h(x) < fj(x*). Because every Ai was assumed to be positive, we have
2::=1 Adi(X) < 2::=1 Adi(X*). This is a contradiction with inequality (3.1.3)
and x· is thus Pareto optimal. 0

Note that because the multiobjective optimization problem is assumed to

be convex, Theorem 3.1.8 provides a sufficient condition for global Pareto op-
timality. This was stated in Theorem 2.2.3.

Theorem 3.1.9. (Karush-Kuhn-Tucker sufficient condition for weak Pareto

optimality) The condition in Theorem 3.1.8 is sufficient for a decision vector
x* E S to be weakly Pareto optimal for 0 ~.\ E Rk with.\ f- O.

Proof. The proof is a straightforward modification of the proof of Theorem

3.1.8.

The convexity assumption in Theorem 3.1.8 can be relaxed. The stated

sufficient condition is also valid if the objective functions are pseudoconvex
and the constraint functions are quasiconvex. This extension is handled, for
example, in Majumdar (1997), Marusciac (1982) and Simon (1986).
If an ordering cone D is used in defining efficiency, then the optimality con-
ditions are similar to those presented above except for the multipliers Ai. Now
42 Part I - 3. Theoretical Background

they are not only nonnegative real scalars but belong to a dual cone D* , where
D* = {l E Rk IlTy ~ 0 for all y ED}. Because of the close resemblance, we
do not here handle optimality conditions separately for efficiency. For details
see, for example, Chen (1984) and Luc (1989, pp. 74-79).

3.1.2. Second-Order Conditions

Second-order optimality conditions (presuming twice continuously differen-

tiable objective and constraint functions) have been examined substantially
less than first-order optimality conditions. Second-order optimality conditions
provide a means of reducing the set of candidate solutions produced by the
first-order conditions but at the same time tighten the assumptions set to the
regularity of the problem.
Second-order optimality conditions for (local) Pareto optimality are treated,
for example, in Wan (1975). For completeness, we here present examples ofnec-
essary and sufficient second-order optimality conditions following Wang (1991).
First we need one more constraint qualification, namely the regularity of
decision vectors.

Definition 3.1.10. A point x* E S is said to be a regular point if the gradients

of the active constraints at x' are linearly independent.

Theorem 3.1.11. (Second-order necessary condition for Pareto optimality)

Let the objective and the constraint functions of problem (3.1.1) be twice con-
tinuously differentiable at a regular decision vector x' E S. A necessary condi-
tion for x' to be Pareto optimal is that there exist vectors 0 ::; l E Rk, l::j:. 0,
and 0 ::; " E R m such that
k m
(1) L:Ai\7fi(x*) + L:llj\7gj(x*) = 0
i=l j=l

(2) Iljgj(x*) =0 for all j = 1, ... , m

(3) dT (t Ai\7 2 fi{x*) + ~ Ilj\72 gj (X*)) d 2: 0

for all d E {O =f. d E Rn I \7 fi{X*)T d ::; 0 for all i = 1, ... , k, \7gj(x*)T d =

o for all jE J(x*)}.

Proof. See Wang (1991).

Note that when second-order optimality conditions are concerned, we need

some kind of second-order constraint qualifications even if we do not obtain a
result satisfying l =f. O. In Theorem 3.1.11, the regularity, that is, the linear
independence of the gradients of the active constraints at the point considered,
 3.1. Differentiable Optimality Conditions 43

is the kind of a second-order constraint qualification that can guarantee an

even stronger result where not all the A-coefficients can vanish.
The following theorem gives two second-order sufficient optimality condi-
tions. The difference lies in the sets of search directions.

Theorem 3.1.12. (Second-order sufficient condition for Pareto optimality)

Let the objective and the constraint functions of problem (3.1.1) be twice con-
tinuously differentiable at a decision vector x· E S. A sufficient condition for
x· to be Pareto optimal is that there exist vectors 0 ~ .\ E R k and 0 ~ pERm
for which (.\,p) =I- (0,0) such that
k m
(1) LA;V'J;(x*) + LJLj\7g j (x*) = 0
i=l j=l

(2) JLjgj(x*) =0 for all j = 1, ... , m

(3) dT(~Ai'V2J;(x*)+ ~JLj\72gj(X'»)d>O
for either all d E {O =I- dE Rn I 'V fi(x*)T d ~ 0 for all i = 1, ... , k, \7gj(x*)T
d ~ 0 for all j E J(x*)} or all d E {O =I- d E Rn I \7gj(x*)T d = 0 for all j E
J+(x*), \7gj(x*)T d ~ 0 for all j E J(x*) \ J+(x*)}, where J+(x*) = {j E
J(x*) I JLj > OJ.
Proof. See Wang (1991).

Second-order sufficient conditions for Pareto optimality are also treated in

Simon (1986), and more necessary and sufficient conditions for Pareto and
weakly Pareto optimal solutions are presented in Wang (1991).

3.1.3. Conditions for Proper Pareto Optimality

For completeness we also present the original necessary optimality condition

formulated for proper Pareto optimality in the sense of Kuhn and Tucker (see
Definition 2.9.3) as stated by Kuhn and Tucker (1951). To begin with, we write
down 'IlIcker's theorem of the alternative, which will be utilized in the proof.

Theorem 3.1.13. (Tucker's theorem) Let A and C be given matrices. Then

either the system of inequalities

Ax ~ 0, Ax =I- 0, Cx ~ 0

has a solution x, or the system

has a solution (.\,p), but never both.

44 Part 1 - 3. Theoretical Background

Proof. The proof is similar to the proof of Theorem 3.1.4.

We can now present the necessary condition for proper Pareto optimality.

Theorem 3.1.14. (Kuhn-Tucker necessary condition for proper' Pareto opti-

mality) Let the objective and the constraint functions of problem (3.1.1) be
continuously differentiable at a decision vector x· E S. A necessary condition
for x' to be properly Pareto optimal (in the sense of Kuhn and Tucker) is that
there exist vect.ors 0 < l E R k and 0 :::; " E R m such that
k m
(1) LAi\7j;(X*) + L/Lj\7Yj(x*) = 0
i=1 j=1
(2) P'jYj(x*) =0 for all j=l,... ,rn.

Proof. Let x· be properly Pareto optimal (in the sense of Kuhn and Tucker).
From the definition we know that no vector dE R n exists such that \7 fi(x*)T d
:::; 0 for all i = 1, ... , k, \7 fj(x*)T d < 0 for some indexj, and \791(X*)T d :::; 0 for
alii E J(x*). Then, from Theorem 3.1.13 we know t.hat there exist multipliers
Ai > 0 for i = 1, ... ,k and /Lj 2:: 0 for j E J(x*) such that 2:::=1 Ai\7fi(x*) +
LjEJ(x') /Lj \7 gj (x*) = O. We obtain statement (1) by setting /Lj = 0 for all
j E {l, ... ,rn} \ J(x*).
If gj(x*) < 0 for some j, then according to the above setting /Lj = 0 and
equalities (2) follow. 0

It is proved in Geoffrion (1968) and Sawaragi et al. (1985, p. 90), that if the
Kuhn-Tucker constraint qualification (Definition 3.1.3) is satisfied at a decision
vector x· E S, then the condition in Theorem 3.1.14 is also necessary for x* to
be properly Pareto optimal in the sense of Geoffrion. Finally, we write down
the sufficient condition for proper Pareto optimality.

Theorem 3.1.15. (Kuhn-Tucker sufficient condition for proper Pareto opti-

mality) If problem (3.1.1) is convex, then the condition in Theorem 3.1.14 is
also sufficient for a decision vector x* E S to be properly Pareto optimal (in
the sense of Kuhn and Tucker).

Proof. See Sawaragi et al. (1985, p. 90) or Shimizu et al. (1997, p. 112).

Let us finally mention that necessary and sufficient conditions for proper
Pareto optimality in the sense of Geoffrion are presented in Gulati and Islam
(1990) for pseudolinear objective (i.e., differentiable functions that are both
pseudo convex and pseudoconcave) and quasiconvex constraint functions.
 3.2. Nondifferentiable Optimality Conditions 45

3.2. Nondifferentiable Optimality Conditions

In this section, we no longer necessitate differentiability but put forward

non differentiable counterparts for the optimality conditions presented in Sec-
tion 3.1. Usually, when the assumption of continuous differentiability is given
up, functions are assumed to be locally Lipschitzian (see Definition 2.1.12).
Remember that a function is here called nondifferentiable if it is locally Lip-
schitzian (and not necessarily continuously differentiable).
In every other way the multiobjective optimization problem to be solved is
still of the form

(3.2.1)
minimize {!1 (x), hex), ... ,!k (x)}
subject to xES = {x E R n I g(x) = (91(X),92(X), ... ,9m(x»T ~ OJ.

We first briefly present some properties of subdifferentials (see Definition

2.1.14) without any proofs.

Theorem 3.2.1. Let the functions Ii: Rn -+ R, i = 1, ... , k, be locally

Lipschitzian at a point x· ERn. Then, for weights Wi E R we have

The two sets are equal if at least k - 1 of the functions Ii are continuously
differentiable, or if the functions are convex and the weights are positive.

Proof. See, for example, Makela and Neittaanmaki (1992, p. 39) and Clarke
(1983, pp. 38-39).

Theorem 3.2.2. Let the functions Ii: R n -+ R, i = 1, ... , k, be locally

Lipschitzian at a point x* E R". Then the function I: R n -+ R

I(x) = max li(X)

i=l, ... ,k

is also locally Lipschitzian at x*. In addition,

al(x*) c conv {ali(x·) liE l(x·)},

where l(x") C {I, ... ,k} denotes the set of indices i for which I(x·) = li(X*).

Proof. See, for example, Makela and Neittaanmaki (1992, pp. 47-49).

Theorem 3.2.3. Let the function Ii: R n -+ R be locally Lipschitzian at a

point x* E Rn and attain its (local) minimum at x*, then
46 Part I - 3. Theoretical Background

o E 8J;(x*).
If the function Ii is convex, then the condition is also sufficient and the mini-
mum is global.

Proof. See, for example, Makela and Neittaanmaki (1992, pp. 70-71).

Before moving on to the optimality conditions of the Fritz John and Karush-
Kuhn-Thcker type we should point out the following. If a single objective func-
tion is defined on a set, the counterpart of the condition in Theorem 3.2.3
says that zero belongs to the algebraic sum of two sets formed at the point
considered. The sets are the sub differential of the objective function and the
normal cone of the feasible region. This result is adapted for convex multiobjec-
tive optimization problems involving continuous objective functions and closed
feasible regions in Plastria and Carrizosa (1996). The necessary and sufficient
condition for weak Pareto optimality is that zero belongs to the sum of the
union of the sub differentials of the objective functions and the normal cone
of the feasible region. Note that the functions do not have to be even locally
Lipschitzian. According to Clarke (1983, pp. 230-231), the same condition is
necessary for weak Pareto optimality in general problems as well. We do not
treat these results more thoroughly here. Instead, we present one more result
for single objective nondifferentiable optimization.

Theorem 3.2.4. (Fritz John necessary condition 101' optimality) A necessary

condition for a point x* ERn to be a local minimum of the problem
minimize J; (x)
subject to g(x) = (91(X),92(X), ... ,9m(x)f :::: 0,
where the objective function Ii: R n --t R and the constraints 9j: R n --t R,
j = 1, ... , rn, are locally Lipschitzian at x*, is that there exist multipliers
0:::: A E Rand 0 :::: pERm for which (A,p) f:. (0,0) such that
m
(1) 0 E A8Ii(x*) +L J-lj 89j(X*)
j=l

(2) J-lj9j(X*) = 0 for all j = 1, ... , rn.

Proof. See, for example, Clarke (1983, pp. 228-230) or Kiwiel (1985c, p. 16).

Now that we have assembled a set of tools, we are in a position to handle

the actual optimality conditions. More information can be found, for example,
in Craven (1989), Dolezal (1985), Minami (1980-81, 1981, 1983), Shimizu et
al. (1997, pp. 322-325) and Wang (1984). The theorems are presented in a
simplified form here compared to the general practice so as to emphasize the
ideas. For this reason, the proofs have been modified.
 3.2. Nondifferentiable Optimality Conditions 47

3.2.1. First-Order Conditions

The first result to be presented is a necessary condition of the Fritz John

type for Pareto optimality.

Theorem 3.2.5. (Fritz John necessary condition lor Pareto optimality) Let
the objective and the constraint functions of problem (3.2.1) be locally Lip-
schitzian at a point x* E S. A necessary condition for the point x* to be
Pareto optimal is that there exist multipliers 0 :::; ..\ E R k and 0 :=:; pERm for
which (..\,p) =I (0,0) such that
k m
(1) 0 E :~::>'i8h(x*) +L fj j 8g j (x*)
i=l j=l
(2) J.Lj9j(X*) =0 for all j = 1, ... , m.

Proof. Because it is assumed that (..\,p) =I (0,0), we can normalize the mul-
tipliers to sum up to one. We shall here prove a stronger condition, where
L:~=1 Ai + L:;:1 fjj = l.
Let x* E S be Pareto optimal. At first we define an additional function
F: Rn -t R by

F(x) = max [Ji(X) -h(X*),9j(X) Ii = 1, .. . ,k, j = 1, . .. ,m]

and show that for all x ERn is

(3.2.2) F(x) ~ O.

Let us on the contrary assume that for some XO E Rn is F(xO) < O. Then
9j(XO) < 0 for all j = 1, ... ,m and the point XO is thus feasible in problem
(3.2.1). In addition, h(xO) < li(x*) for all i = 1, ... ,k, which contradicts the
Pareto optimality of x*. Thus (3.2.2) must be true.
Noting that the point x* is feasible in problem (3.2.1), we obtain g(x*) :::; O.
This implies F(x*) = O. Combining this fact with property (3.2.2), we know
that F attains its (global) minimum at x*. As all the functions Ii and 9j are
locally Lipschitzian at x*, likewise F (according to Theorem 3.2.2). We deduce
from Theorem 3.2.3 that 0 E 8F(x*).
Note that

(3.2.3)

applying Theorem 3.2.l.

We designate the set of indices j for which F(x*) = 9j(X*) by J(x*) C
{I, ... , m}. Now we can employ Theorem 3.2.2 and (3.2.3) and obtain

o E conv {8h(x*), 89j(X*) Ii = 1, ... , k, j E J(x*)}.

48 Part I ~ 3. Theoretical Background

Employing the definition of a convex hull, we know that there exist vectors
.\ and p of real multipliers for which Ai ::::: 0 for all i = 1, ... , k, /1j ::::: 0 for all
j E J(x*) and 2:~=1 Ai + 2: jEJ (x*) /1j = 1, such that
k

oE L A/J/i(x*) + L /1j8g j (x*).

i=l jEJ(x*)

Now we can set /1j = 0 for all j E {I, ... , m} \ J(x*). Statement (1) follows
from this setting.
Part (2) is trivial. If gj(x*) < 0 for some j, then j E {I, ... , m} \ J(x*) and
we have /1j = O. This completes the proof. 0

Now we can define decision vectors called substationary points.

Definition 3.2.6. A decision vector x* E S is called a substationary point if

it satisfies the (necessary) optimality condition presented in Theorem 3.2.5.

Theorem 3.2.5 can also be proved by first employing a scalarization method

and then Theorem 3.2.4 for the resulting single objective optimization problem
(see, e.g., Dolezal (1985)).

Corollary 3.2.7. (Fritz John necessary condition for weak Pareto optimality)
The condition of Theorem 3.2.5 is also necessary for a decision vector x* E S
to be weakly Pareto optimal.

Next, we examine some constraint qualifications. It is obvious that they

differ from the differentiable case.
Note that when the necessary optimality conditions are derived with the
help of a scalarizing function, it is easy to generalize the constraint qualifica-
tions from single objective optimization to the multiobjective case. One simply
assumes that both the original constraints and the possible additional con-
straints satisfy a constraint qualification. This is expressed in Dolezal (1985).
The constraint qualifications used there are those of calmness and Mangasarian-
Fromovitz.
The so-called Cottle constraint qualification is used in the following theo-
rem. Other constraint qualifications are presented, for example, in Ishizuka. and
Shimizu (1984).

Definition 3.2.8. Let the objective and the constraint functions of problem
(3.2.1) be locally Lipschitzian at a point x* E S. Problem (3.2.1) satisfies the
Cottle constraint qualification at x* if either gj (x*) < 0 for all j = I, ... ,m or
Of/. conv{8gj (x*) I gj(x*) = OJ.
Assuming the Cottle constraint qualification, we obtain the Karush-Kuhn-
Tucker necessary condition for Pareto optimality.
 3.2. Nondifferentiable Optimality Conditions 49

Theorem 3.2.9. (Karush-Kulm-Tucker necessary condition for Pareto opti-

mality) Let the assumptions of Theorem 3.2.5 be satisfied by the Cottle con-
straint qualification. Theorem 3.2.5 is then valid with the addition that ..\:f; O.

Proof. The proof of Theorem 3.2.5 is here valid up to the observation 0 E

8F(x*) and result (3.2.3). We prove also this theorem in a stronger form,
where the multipliers sum up to one.
From the definition of F we know that

F(x*) = o.
We continue by first assuming that gj(x*) < 0 for all j = 1, ... , m. In this case,
F(x*) > gj(x*) for all j. Now we can apply Theorem 3.2.2 and equation (3.2.3)
and obtain
o E conv{8fi(X*) Ii = 1, ... ,k}.
From the definition of a convex hull we know that there exists a vector
o ~..\ E Rk
of multipliers for which 2:~=1 Ai = 1 (thus..\:f; 0) such that
k
oE L Ai 8 fi(X*).
i=1

We obtain the statement to be proved (denoted by (1) in Theorem 3.2.5) by

setting /-tj = 0 for all j = 1, ... ,m.
On the other hand, if there exists some index j such that gj(x*) = 0, we
denote the set of such indices by J(x*). By the Cottle constraint qualification
we know that

(3.2.4) o ~ conv {8g j (x*) I j E J(x*)}.

In this case, we deduce from Theorem 3.2.2 and result (3.2.3) that

o E conv {8fi(X*), 8g j (x*) Ii = 1, ... , k, j E J(x*)}.

Applying the definition of a convex hull, we know that there exist multipliers
Ai ~ 0, i = 1, ... ,k, and /-tj ~ 0, j E J(x*), for which 2:~=1 Ai + 2: jE J(x*) /-tj =
1, and by assumption (3.2.4), especially ..\:f; 0, such that
k
oE L Ai8 fi(x*) + L /-tj8g j (x*).
i=1 jEJ(x*)

Again, we obtain the statement to be proved by setting /-tj = 0 for all j E

{1, ... , m} \ J(x*).
The proof of part (2) is the same as in Theorem 3.2.5. 0
50 Part I - 3. Theoretical Background

Corollary 3.2.10. (K arush-K uhn- Tucker necessary condition for weak Pareto
optimality) The condition of Theorem 3.2.9 is also necessary for a decision
vector x' E S to be weakly Pareto optimal.

If we assume the multiobjective optimization problem to be convex and l >

0, we get a Karush-Kuhn-Tucker sufficient condition for Pareto optimality. Note
that the convexity of a function implies that the function is locally Lipschitzian
at any point xES.

Theorem 3.2.11. (Karush-Kuhn-Tucker sufficient condition for Pareto op-

timality) Let problem (3.2.1) be convex. A sufficient condition for a decision
vector x* E S to be Pareto optimal is that there exist multipliers 0 < l E Rk
and 0 ::; pERm such that
k m
(1) 0E L Ai 8 fi(x*) + L Jl 8g j j (x*)
i=1 j=1
(2) Jljgj(X*) =0 for all j = 1, ... , m.

Proof. To start with, we define an additional function F: R n -* R by F(x) =

2:~=1 Adi(X) + 2:7=1 Jljgj(x), where the multipliers Ai and Jlj satisfy the above
assumptions. Because the functions fi and gj are convex, l > 0 and p ~ 0,
then F too is convex, and 8F(x) = L:7=1 Ai8fi(x) + L:j~1 Jl j 8g j (x) (as stated
in Theorem 3.2.1).
From assumption (1) we know that 0 E 8F(x*), and, according to Theorem
3.2.3, the point x* is a (global) minimum of F. This implies that for any
XO ERn, especially any XO satisfying g(XO) ::; 0, the following is valid:

k m k m
= L Ad;(xO) + L Jljgj(XO) - L '\di(X*) - L Jljgj(x*).
;=1 j=1 i=1 j=1
Employing assumption (2), the fact that g(XO) ::; 0 and p ~ 0, we obtain
k k
(3.2.5) L Adi(X*) :s :L Adi(XO)
i=1 i=1

for any XO E S.
Let us assume that x* is not Pareto-optimal. Then there exists some feasible
x such that j;(x) ::; j;(x*) for all i = 1, ... , k and for at least one index
j is fj(x) < fJ(x*), Because every Ai was assumed to be positive, we have
L:~=1 Adi(X) < L:~=1 Ajj;(X*). This contradicts inequality (3.2.5) and X* is
thus Pareto optimal. 0
 3.2. Nondifferentiable Optimality Conditions 51

Theorem 3.2.12. (Karush-Kuhn-Tucker sufficient condition for weak Pareto

optimality) The condition stated in Theorem 3.2.11 is sufficient for a decision
vector x* E S to be weakly Pareto optimal for 0 ::; A E R k with A i= o.

Proof. The proof is a trivial modification of the proof of Theorem 3.2.11.

Finally, we introduce one more constraint qualification. It can only be ap-

plied to convex problems and it will also be needed in Part II (Section 2.2) in
connection with the multiobjective proximal bundle method. It is called the
Slater constraint qualification. It is independent of the differentiability of the
functions involved.

Definition 3.2.13. Let problem (3.2.1) be convex. Problem (3.2.1) satisfies

the Slater constraint qualification if there exists some x with gj(x) < 0 for all
j=l,... ,m.

Theorem 3.2.9 and Corollary 3.2.10 can now be reformulated for convex
problems assuming the Slater constraint qualification. Remember that convex-
ity implies that functions are locally Lipschitzian at any point in the feasible
region.

Theorem 3.2.14. (Karush-Kuhn-Tucker necessary condition for (weak) Pareto

optimality) Let problem (3.2.1) be convex, satisfying the Slater constraint quali-
fication. A necessary condition for a point x* E S to be (weakly) Pareto optimal
is that there exist multipliers 0 ::; A E Rk with A i= 0 and 0 ::; pERm such
that
k m
(1) 0 E L Ai 8 fi(x*) +L Ilj8gj (x*)
i=l j=1

(2) Jljgj(x*) = 0 for all j = 1, ... , m.

Proof. The proof is a trivial modification of the proof of Theorem 3.2.9 when
we note the following. In case the set J(x*) is nonempty we denote g(x) =
max[gj(x) I j = l, ... ,mj. Now g(x*) = gj(x*) for j E J(x*). By the Slater
constraint qualification there exists some XO such that gj(XO) < 0 for all j.
Thus, x* cannot be the global minimum of the convex function g. According
to Theorem 3.2.3 we derive

o ¢ conv{8gj (x*) Ij E J(x*)}.

The proof of Theorem 3.2.9 can now be applied. o

Necessary optimality conditions for Pareto optimality in those nondiffer-
entiable problems where the objective functions are fractions of convex and
52 Part I - 3. Theoretical Background

concave functions are formulated in Bhatia and Datta (1985). In addition, nec-
essary Fritz John and Karush-Kuhn-Tucker type optimality conditions for weak
Pareto optimality involving so-called semidifferentiable pre-invex functions are
treated in Preda (1996).
If an ordering cone D is used in defining efficiency, then the optimality
conditions are similar to those presented above, except for the multipliers Ai
(simply as in the differentiable case). The multipliers belong to the dual cone
D* = {A E Rk I >7y 2: 0 for all y ED}. Because of the similarity, we do
not present here separate optimality conditions for efficiency. Necessary and
sufficient conditions for efficiency and weak efficiency are handled, for example,
in Wang (1984). Furthermore, in Craven (1989) and EI Abdouni and Thibault
(1992), necessary conditions for weak efficiency in normed spaces and Banach
spaces, respectively, are presented. The objective and the constraint functions
are still assumed to be locally Lipschitzian.
Direct counterparts of optimality conditions for proper Pareto optimality
in the sense of Kuhn and Tucker, presented in Section 3.1, cannot be stated
in the nondifferentiable case. The reason is that the definition of Kuhn and
Tucker assumes continuous differentiability. However, a sufficient condition for
proper Pareto optimality in the sense of Geoffrion, when the objective and the
constraint functions are compositions of convex, locally Lipschitzian functions,
is formulated in Jeyakumar and Yang (1993). This treatment naturally includes
ordinary convex, locally Lipschitzian functions. The authors also present nec-
essary conditions for weak Pareto optimality and sufficient conditions of their
own for Pareto optimality in problems with convex composite functions. A nec-
essary and sufficient condition for proper efficiency (in the sense of Henig) is
derived in Henig and Buchanan (1994, 1997) for convex problems.

3.2.2. Second-Order Conditions

At the end of this section we shall say a few words about the case where the
functions involved are continuously differentiable and their gradients are locally
Lipschitzian. Such functions are called Cl,l-functions. Second-order optimality
conditions for multiobjective problems with CI,I-functions are handled in Liu
(1991). Here we briefly state the main results. First we must introduce one
concept according to Liu.

Definition 3.2.15. Let the function j;: S ~ R be a CI,I-function at the point

x* E S. The set

a; Ji(x*)(d, d) = {¢i E R Ithere exists a sequence {tj }~l'

¢i = lim
tj-to+
t~j (Ji(X* + tjd) - Ji(X*) - t/~7 j;(X*)T d)}
 3.2. Nondifferentiable Optimality Conditions 53

is called a generalized second-order directional derivative of the function /i

evaluated at x· in the direction d ERn.

The set 8;/i(x*)(d, d) is nonempty according to Liu and Krizek (1997).

The following second-order necessary and sufficient conditions are presented
in Liu (1991), for both the case where all the objective functions are C l ,l_
functions and all the constraint functions are twice continuously differentiable
and the case where all the objective functions are twice continuously differen-
tiable and all the constraint functions are Cl,l-functions. Here, we formulate
only the first case, while the only difference in the conditions is the reversed
roles of the Hessian matrices and generalized second-order directional deriva-
tives.
Let us again denote the set of active constraints at x· E S by J(x·).

Theorem 3.2.16. (Second-order necessary condition for Pareto optimality)

Let the objective functions of problem (3.2.1) be CI,l-functions and its con-
straint functions twice continuously differentiable at a decision vector x· E S.
A necessary condition for x· to be Pareto optimal is that there exist vectors
0:::; A E Rk and 0 :::; I' E R m for which (A,p) =/: (0,0) such that
k m
(1) LAiV'/i(X*)+ LJLjV'gj(x*) =0
i=l j=l

(2) JLjgj(x*) =0 for all j = 1, ... , m

k m
(3) L Ai V' /i(x*)T d = 0, LJLjV'gj(x*fd = 0
i=l j=l

(4) t.).;~; + dT(t,~;V2g;(X'))d" 0

for all d E {O =/: d E R n I V' /i(X*)T d :::; 0 for all i = 1, ... , k, V'gj(x*)T d :::;
o for all jE J(x*)} and £Pi E 8;/i(x*)(d,d).

Proof. See Liu (1991).

Finally, we present a second-order sufficient optimality condition for prob-

lems involving CI,I-functions.

Theorem 3.2.17. (Second-order sufficient condition for Pareto optimality)

Let the objective functions of problem (3.2.1) be CI,I-functions and its con-
straint functions twice continuously differentiable at a decision vector x* E S.
A sufficient condition for x* to be Pareto optimal is that there exist vectors
0:::; A E Rk and 0 :::; I' E Rm for which (A,I') =/: (0,0) such that
54 Part I - 3. Theoretical Background

k m
(1) LAiVfi(X*)+LJljVgj(x*)=O
i=l j=1

(2) Jljgj(X*) =0 for all j=l, ... ,m

k m
(3) LAiV/;(x*fd=o, LJljVgj(x*)Td=O
i=1 j=1

(4) ~Ai¢i+dT(~J.LjV2gj(X*))d>O
for all d E {O oj:. d E Rn I V /;(x*f d ::; 0 for all i = 1, ... , k, Vgj(x*)T d ::;
o for all jE J(x*)} and ¢i E aUi(x*)(d, d).

Proof. See Liu (1991).

Actually, the results in Liu (1991) are given in a more general form for
efficient solutions and for problems where the constraint functions belong to a
polyhedral convex cone.

3.3. More Optimality Conditions

Many necessary and sufficient conditions for weak, proper or Pareto opti-
mality (or efficiency) have been suggested in the literature. They are based on
different kinds of assumptions as to the properties and form of the problem.
Many of them are based on a scalarization of the original problem and con-
ditions are se.t to both the original functions and the scalarization parameters
(some such conditions are presented in Part II in connection with the scalariza-
tion methods). In this book, we settle for a closer handling of the Fritz John and
the Karush-K uhn-Tucker type conditions, presented in the two earlier sections.
For the interested reader we list some other references.
Necessary conditions for proper and improper Pareto optimality in the sense
of Kuhn and Tucker are derived with the help of cones in Tamura and Arai
(1982). Geoffrion (1968) was the first to give the basic characterization of prop-
erly Pareto optimal solutions in terms of a scalar problem, called a weighting
problem (see Section 3.1 of Part II). He extended the results by a compre-
hensive theorem into necessary and sufficient conditions for local and global
proper Pareto optimality. Geoffrion's treatment is closely followed in Chou et
al. (1985), where properly Pareto optimal solutions are characterized for multi-
objective optimization problems with set-valued functions. In addition, neces-
sary and sufficient Karush-Kuhn-Tucker type optimality conditions for (-Pareto
optimality in convex problems using the weighting method for the objectives
and exact penalty functions for the constraints are handled in Liu (1996).
 3.3. More Optimality Conditions 55

In Chankong and Haimes (1982), the Karush-Kuhn-'lUcker optimality con-

ditions for Pareto optimality are modified for use in connection with certain
solution methods (the e:-constraint method and the jth Lagrangian problem;
see Section 3.2 of Part II). Chankong and Haimes also propose optimality
conditions for proper Pareto optimality (in the sense of Geoffrion) with the
c-constraint method. Further, in Benson and Morin (1977), necessary and suf-
ficient conditions are given for a Pareto optimal solution to be properly Pareto
optimal. This is done with the help of the jth Lagrangian problem. Necessary
and sufficient conditions for Pareto optimality with convex and differentiable
functions partly based on the e:-constraint problem are proved in Zlobec (1984).
Necessary and sufficient conditions for Pareto optimality and proper Pareto
optimality are proved with the help of duality theory and auxiliary problem
(2.1O.1) (presented in Section 2.1O) in Wendell and Lee (1977). However, it
is stated that nonlinear problems do not generally satisfy the conditions de-
veloped. In such cases Pareto optimal solutions have to be tested for proper
Pareto optimality on a point-by-point basis.
In Gulati and Islam (1988), linear fractional objective functions and general-
ized convex constraints are handled. Necessary conditions of the Karush-Kuhn-
Tucker type are presented for Pareto optimal solutions, and the conditions
under which Pareto optimal solutions are properly Pareto optimal are stated.
Necessary and sufficient conditions for Pareto optimality in problems with non-
linear fractional objective functions and nonlinear constraints are proved in Lee
(1992). In addition, necessary optimality conditions for fractional multiobjec-
tive optimization problems with square root terms are given in Egudo (1991).
In Benson (1979b), a necessary and sufficient condition is given for a point to
be Pareto optimal when there are two concave objective functions (problem of
maximization) and a convex feasible set.
The following references deal with conditions for efficiency, where the ob-
jective space is ordered by an ordering cone.
In Zubiri (1988), necessary and sufficient conditions are proved for weak
efficiency in Banach spaces with the help of a weighted Loo-metric (see Sec-
tion 3.4 of Part II). Several necessary and sufficient conditions for efficient,
weakly efficient and properly efficient solutions (in the sense of Borwein) in
real topological linear spaces are collected in Jahn (1985). Necessary and suf-
ficient optimality conditions of the Karush-Kuhn-'lUcker type are derived in
Hazen (1988), for cases where preferences are and are not representable by
cones.
Let us finally briefly mention some further references handling nondiffer-
entiable cases. Necessary and sufficient conditions for Pareto optimality and
proper Pareto optimality are derived in Bhatia and Aggarwal (1992), by the
weighting method (see Section 3.1 of Part II) and Dini derivatives. The func-
tions in the problem are assumed to be nondifferentiable such that the objec-
tive functions are pseudoconcave and the constraint functions are quasiconvex.
Some duality results are provided as well.
56 Part I - 3. Theoretical Background

Optimality conditions based on the optimization theory of Dubovitskii and

Milyutin presuming certain convexity assumptions are presented in Censor
(1977) for Pareto optimality in Rn and in Minami (1981) for weak Pareto
optimality in a linear topological space. No differentiability assumptions are
needed. Necessary and sufficient conditions for weak, proper and Pareto opti-
mality in finite-dimensional normed spaces are presented in Staib (1991) under
different assumptions. In Shimizu et al. (1997, pp. 319-322), nondifferentiable
optimality conditions assuming constraint qualifications based on directional
derivatives are derived.

3.4. Sensitivity Analysis and Duality

The last topics to be mentioned in this chapter are sensitivity analysis,

stability and duality. Sensitivity analysis studies situations when the input
parameters defining the multiobjective optimization problem change or contain
errors. In sensitivity analysis, an answer is sought to the question of how much
the parameters can be altered and varied without affecting the solution. More
justification for sensitivity analysis is provided in Rarig and Haimes (1983).
Given a family of parametrized multiobjective optimization problems, a
set-valued perturbation function is defined in Tanino (1990), such that it as-
sociates with each parameter value the set of Pareto optimal points of the
perturbed feasible region. The behaviour of the perturbation function is ana-
lyzed both qualitatively and quantitatively. In this context stability means the
study of various continuity properties of the perturbation function of a family
of parametrized optimization problems, that is, qualitative analysis. Sensitiv-
ity means the study of the derivatives of the perturbation function, that is,
quantitative analysis.
In general rnultiobjective optimization problems, considerable attention has
been paid to the stability of the preference structure of the decision maker. In
these cases, it is usually assumed that the partial ordering of the objective
space is induced by an ordering cone.
However, mathematical stability and sensitivity analysis are broad areas of
research, and we do not intend to go into details here. Instead, we refer, for
example, to BalMs and Guerra (1996), Craven (1988), Ester (1984), Gal and
Wolf (1986), Kuk et al. (1996), Luc (1989), Lucchetti (1985), Papageorgiou
(1985), Tanino (1988a, b, 1990) and Tanino and Sawaragi (1980), for further
analysis.
Let us still mention that stability is not an unambiguous notion. As stressed,
for example, in Dauer and Liu (1997), the terms and results connected to stabil-
ity and sensitivity analysis are not universally defined in the literature. Different
types of stability can be defined and measured in many ways. Often stability is
associated with worst case performance and analysing how fast a solution de-
grades to a certain still acceptable level. Thus, analysis of stability is important
 3.4. Sensitivity Analysis and Duality 57

in implementing solutions in practice. Regardless if its significance, stability has

been widely ignored in the multiple criteria decision-making context thus far.
A review of sensitivity analysis results for both linear and nonlinear prob-
lems is given in Dauer and Liu (1997). In addition, they study sensitivity analy-
sis for MOLP problems in the objective space and deal with priority structures
in goal programming. Sensitivity analysis for MOLP problems is also treated
in Gal (1995).
Changes that occur in the solution of an MOLP problem, if the number of
objective functions, the number of variables or the number of constraint func-
tions is altered, are examined in Eiselt et al. (1987). This is also an interesting
topic for nonlinear problems, as, for example, an objective function may have
been left out of the model, and it would be useful to know how this can affect
the solution obtained. For example, if a convex objective function is added to
a convex multiobjective optimization problem, all weakly Pareto optimal solu-
tions remain weakly Pareto optimal (see Lowe et al. (1984)). The corresponding
result is not always valid for Pareto optimal solutions. A counterexample can
be found in Steuer (1986, p. 179). A result regarding the generation of the
weakly Pareto optimal set of a convex problem as a union of such Pareto op-
timal sets where subsets of the objective functions are used, is proved in Lowe
et al. (1984).
An overview is presented of duality theory for linear and nonlinear cases
in Nakayama (1985c). Duality theory for nonlinear multiobjective optimiza-
tion problems is also presented, for example, in Bitran (1981), Gopfert (1986),
Luc (1984b, 1987, 1989), Nakayama (1984, 1985b, 1996), Singh et al. (1996)
and Weir (1987); for convex problems in Jahn (1983) and Martfnez-Legaz and
Singer (1987); for more general convex-like problems in Das and Nanda (1997),
Preda (1992, 1996) and Wang and Li (1992); for nonconvex problems in Luc
and Jahn (1992);' and for nonconvex nondifferentiable problems in Preda and
Stancu-Minasian (1997). Some regularity results for multiobjective optimiza-
tion problems are presented in Martein (1989). On the other hand, duality
theory designed for a decision maker determining preferred solutions in convex
multiobjective optimization problems is derived in Tarvainen (1996).
Finally, we state that an excellent account of stability and duality in mul-
tiobjective optimization can be found in Sawaragi et al. (1985). More than a
third of the contents of the monograph addresses these topics.
 Part II

METHODS
 1. INTRODUCTION

Generating Pareto optimal solutions plays an important role in multiobjec-

tive optimization, and mathematically the problem is considered to be solved
when the Pareto optimal set is found. The term vector optimization is some-
times used to denote the problem of identifying the Pareto optimal set. How-
ever, this is not always enough. We want to obtain only one solution. This
means that we must find a way to put the Pareto optimal solutions in a com-
plete order. This is why we need a decision maker and her or his preference
structure. Here in Part II, we present several methods for solving multiobjective
optimization problems. Usually, this means finding the Pareto optimal solution
that best satisfies the decision maker.
We are not here going to interfere with the formulation of a real-life phe-
nomenon as a mathematically well-defined problem. We merely stress that a
proper formulation is important. Let us emphasize that in real-life problems
inaccuracy in some form is often present. Remember that we exclude the han-
dling of stochastic or fuzzy problems in this context. Even when the problems
are modelled in a deterministic form, restricting the treatment to Pareto opti-
mal solutions only may be misleading. For example, forgetting or misspelling
an objective function may affect the Pareto optimal set. If it is impossible to
model the practical problem in an explicit and precise mathematical form, we
cannot automatically leave non-Pareto optimal solutions out of consideration.
For example, imprecision of the data, the measurement or the objective func-
tions means that the Pareto optimal set available is only an approximation of
the real one. Here we have a gap between theory and practice.
Several crucial issues to bear in mind in the formulation of problems are
treated in Haimes (1985) and Nijkamp et al. (1988). Among these are risk
assessment, sufficient representativeness of the objective functions and precision
of information. In many complicated, practical cases it may be impossible to
give a correct formulation to the problem before it is solved. This means that
the modelling and the solution phases should not be undertaken separately,
which is generally the case nowadays. In other words, the modelling phase may
require interaction with the solution phase. A parallel idea of approaching the
modelling phase by including the decision maker in the modelling is suggested
in Brans (1996). The goal is to give more freedom to the decision maker and
not to limit her or his way of thinking to a prespecified model and its concepts.
62 Part II - 1. Introduction

In most methods we are interested in the objective space instead of the de-
cision variable space. One reason for this is that the dimension of the objective
space is usually considerably smaller than the dimension of the decision vari-
able space. Another reason is that decision makers are often more interested
in the objective values. However, calculation still takes place in the decision
variable space because we do not usually know the explicit form of the feasi-
ble objective region. In brief, decision makers usually handle objective values
whereas mathematical programming takes place in the decision variable space.
In general, multiobjective optimization problems are solved by scalarization.
The most important exceptions to. this are MOLP problems, which are not
to be dealt with here, where some simplex-based solution methods can find
Pareto optimal extreme points or, in some cases, the whole Pareto optimal
set. Another exception, which is presented here, is the multiobjective proximal
bundle method for nondifferentiable problems. It is not based on scalarization
in the traditional sense.
As mentioned in Part I, scalarization means converting the problem into a
single or a family of single objective optimization problems with a real-valued
objective function, termed the scalarizing junction, depending possibly on some
parameters. This enables the use of the theory and the methods of scalar opti-
mization, that is, nonlinear programming. Of fundamental importance is that
the optimal solutions of multiobjective optimization problems can be charac-
terized as solutions of certain single objective optimization problems. Because
scalarizing functions usually depend on certain auxiliary parameters, some nu-
merical difficulties may appear if the single objective optimization problem has
feasible solutions only with very few parameter values or it is not solvable with
all the parameter values. Thus the seemingly promising idea of simplifying the
problem into single objective optimizations has also its weaknesses. In what
follows, we assume that solutions to scalarizing functions exist.
In Sawaragi et al. (1985), three requirements are set for a scalarizing func-
tion:
(1) It can cover any Pareto optimal solution.
(2) Every solution is Pareto optimal.
If the scalarizing function is based on aspiration levels, then, in addition
(3) Its solution is satisficing if the aspiration levels used are feasible.
Unfortunately, there is no scalarizing function that can satisfy all three require-
ments.
An important fact to keep in mind is that standard routines for single objec-
tive optimization problems can only find local optima. This is why only locally
Pareto optimal solutions are usually obtained and handled when dealing with
scalarizing functions. Global Pareto optimality can be guaranteed, for exam-
ple, if the objective functions and the feasible. region are convex (as stated in
Theorem 2.2.3 of Part I) or quasi convex and convex, respectively (see Theorem
 1. Introduction 63

2.2.4 of Part I). An alternative is to employ global single objective optimiz-

ers. In the following, however, the solutions are understood to be local, unless
stated otherwise.
Another matter to consider is the possibility of the scalarizing function
having several alternative optimal solutions. In this case, the objective vector
produced depends on the solution chosen. This may affect in an uncontrolled
way the direction in which the solution process proceeds. This fact has not been
taken into account in most method developments. Ideas for handling alternative
optima in MOLP problems are presented in Sarma and Merouani (1995).
There is a large variety of methods for accomplishing multiobjective opti-
mization. None of them can be said to be generally superior to all the others.
When selecting a solution method, the specific features of the problem to be
solved must be taken into consideration. In addition, the opinions of the de-
cision maker are important. It is not enough that the analyst simply prefers
some method. It may happen that the decision maker cannot or does not want
to use it. The decision maker may be busy or mathematically ignorant. One
can say that selecting an appropriate multiobjective optimization method itself
is a problem with multiple objectives! We shall return to the method selection
problem in Section 1.3 of Part III.
Methods of multiobjective optimization can be classified in many ways ac-
cording to different criteria. In Cohon (1985), they are categorized into two
relatively distinct subsets: generating methods and preference-based methods.
In generating methods, the set of Pareto optimal (or efficient) solutions is gen-
erated for the decision maker, who then chooses one of the alternatives. In
preference-based methods, the preferences of the decision maker are taken into
consideration as the solution process goes on, and the solution that best satisfies
the decision maker's preferences is selected.
Rosenthal (1985) suggests three classes of solution methods: partial gen-
eration of the Pareto optimal set, explicit value function maximization and
interactive implicit value function maximization. In Carmichael (1981), meth-
ods are classified according to whether a composite single objective function,
a single objective function with constraints, or many single objective functions
are the basis for the approach. One more rough division could be made into
interactive and noninteractive methods. These classes can be further divided
in many ways.
Here we apply the classification presented in Hwang and Masud (1979). This
classification is followed, for instance, in Buchanan (1986), Hwang et al. (1980)
and Lieberman (1991a, b). Hwang and Masud clasSify the methods according
to the participation of the decision maker in the solution process. The classes
are:
1) methods where no articulation of preference information is used (no-
preference methods),
2) methods where a posteriori articulation of preference information is used
(a posteriori methods),
64 Part II - 1. Introduction

3) methods where a priori articulation of preference information is used (a

priori methods), and
4) methods where progressive articulation of preference information is used
(inter'active methods).
For short, the names in the parentheses are used in the following.
However, no classification can be complete, as demonstrated, for example,
in Despontin et al. (1983). Thus, one must bear in mind that the classifications
are not absolute. Overlapping and combinations of classes are possible and some
methods can be considered to belong to more than one class. The presented
grouping is for guidance only.
In addition to the role of the decision maker we consider an alternative way
of classification into ad hoc and non ad hoc methods. This division, suggested
by Steuer and Gardiner (1991), is mainly intended for interactive methods, but
can be applied to some other methods as well. It is based on the existence of an
underlying value function. The common feature of ad hoc methods is that even
if one knew the decision maker's value function, one would not exactly know
how to respond to the questions posed by the algorithm. On the other hand,
in non ad hoc methods the responses can be determined or at least confidently
simulated if the decision maker's value function is known.
It should be pointed out that several concepts and assumptions underlying
methods and solution processes can be questioned. For example, ten myths of
multiobjective optimization and decision making are discussed and called into
question in Michalowski (1997) and Zionts (1997a, b). Among them are con-
cepts of well-defined decisions, isolated decision makers, optimal solutions, the
value of Pareto optimal solutions, value functions, static decisions, preference of
sophistication, mathematical convergence and technical assumptions. Here we
do not go into details of these myths but refer to the presentations mentioned.
A further aspect concerns the relative importance of objective functions. As
emphasized in Roy and Mousseau (1996), such a notion is more complex than
is usually recognized.
Before presenting the methods, we mention several references for further
information. In Hwang and Masud (1979), a large number of methods is pre-
sented and illustrated by solving numerical examples in detail. A similar but
shortened presentation is given in Hwang et al. (1980). The detailed solution
process descriptions are intended to help in selecting solution methods.
Extensive surveys of concepts and methods for multiobjective optimization
are provided in the monographs Chankong and Haimes (1983b) and Steuer
(1986). Similar matters are studied briefly in Buchanan (1986), Chankong and
Haimes (1983a), Chankong et al. (1985), Dyer and Sarin (1981), Rosenthal
(1985), Steuer (1989b), Steuer and Gardiner (1990), Stewart (1992) and Van-
derpooten (1990). An overview is given in Evans (1984) and several methods
are also presented in Cohon (1985) and Osyczka (1984).
A set of methods developed up to the year 1973 for both multiattribute
decision analysis and multiobjective optimization is collected in MacCrimmon
 1. Introduction 65

(1973). A wide collection of methods available (up to the year 1983) is assembled
also in Despontin et al. (1983). Almost 100 methods for both multiobjective
and multiattribute cases are included.
As far as different nationalities are concerned, overviews of multiobjective
optimization methods in the former Soviet Union are presented in Lieberman
(1991a, b) and of theory and applications in China in Hu (1990). Nine multi-
objective optimization methods developed in Germany are briefly introduced
in Ester and Holzmiiller (1986).
A great number of interactive multiobjective optimization methods is col-
lected in Shin and Ravindran (1991) and Vanderpooten and Vincke (1989).
Interactive methods are also presented in Narula and Weistroffer (1989a) and
White (1983b). Information about applications of the methods is also reported.
Some literature on interactive multiobjective optimization between the years
1965 and 1988 is gathered in Aksoy (1990). A set of scalarizing functions is out-
lined in Wierzbicki (1986b) with special attention to whether weakly, properly
or Pareto optimal solutions are produced.
As to different problem types, an overview of methods for MOLP problems
can be found in Zionts (1980, 1989). Methods for hierarchical multiobjective
optimization problems are reviewed in Haimes and Li (1988). Such methods are
needed in large-scale problems. A wide survey on the literature of hierarchical
multiobjective analysis is also provided.
Methods with applications to large-scale systems and industry are presented
in the monographs Haimes et al. (1990) and Tabucanon (1988), respectively.
Several groups of methods applicable to computer-aided design systems are
presented briefly in Eiduks (1983). Methods for applications in structural op-
timization are reported in Eschenauer (1987), Jendo (1986), Koski and Silven-
noinen (1987) and Osyczka and Koski (1989). The collections of papers edited
by Eschenauer et al. (1990a) and Stadler (1988a) contain mainly applications
in engineering.
In the following, we present several methods (in four classes) for multiob-
jective optimization. Some of them will be described in more detail and some
only briefly mentioned. It must be kept in mind that the existing methodology
is very wide. We do not intend to cover every existing method but to introduce
several philosophies and ways of approaching multiobjective optimization prob-
lem solving. Where possible we try to link references to some of the applications
and extensions available in the literature with the methods presented here. The
description of each method ends with concluding remarks by the author taking
up important aspects of the method. Unless stated otherwise, we assume that
we solve problem (2.1.1) defined in Part I.
In connection with methods, a mention is made only of such implementa-
tions as have been made available to the author for testing purposes. By a
user we mean either a decision maker or an analyst who uses the solution pro-
gram. If the user is a decision maker, it is usually assumed that the problem
has been formulated earlier (and perhaps loaded in the system) so that the
decision maker can concentrate on the actual solution process.
 2. NO-PREFERENCE METHODS

In no-preference methods, where the opinions of the decision maker are

not taken into consideration, the multiobjective optimization problem is solved
using some relatively simple method and the solution obtained is presented to
the decision maker. The decision maker may either accept or reject the solution.
It seems quite unlikely that the solution best satisfying the decision maker could
be found with these methods. That is why no-preference methods are suitable
for situations where the decision maker does not have any special expectations
of the solution and (s)he is satisfied simply with some optimal solution. The
working order here is: 1) analyst, 2) none.
As examples of this class we present the method of the global criterion and
the multiobjective proximal bundle method.

2.1. Method of the Global Criterion

The method of the global criterion is also sometimes called compromise

programming (see Yu (1973) and Zeleny (1973». In this method, the distance
between some reference point and the feasible objective region is minimized.
The analyst has to select the reference point and the metric for measuring the
distances. All the objective functions are thought to be equally important.

2.1.1. Different Metrics

Here we examine the method where the ideal objective vector is used as a
reference point and Lp-metrics are used for measuring. In this case, the Lp-
problem to be solved is

(2.1.1)
minimize (t 1J;(x) - z:I') 'I,
subject to xES.
From the definition of the ideal objective vector z* we know that fi(X) ~ zt
for all i = 1, ... ,k and all xES. This is why no absolute values are needed if
we know the global ideal objective vector. If the global ideal objective vector
K. Miettinen, Nonlinear Multiobjective Optimization
© Springer Science+Business Media New York 1998
68 Part II - 2. No-Preference Methods

is not known, the method does not necessarily work as it should. In order to
emphasize this fact, we keep the absolute value signs in the notations when
introducing the method.
If the ideal objective vector is replaced by some other vector, it must be se-
lected carefully. Pessimistic reference points must be avoided since the method
cannot find solutions better than the reference point.
The exponent lip may be dropped. Problems with or without the exponent
lip are equivalent for 1 ::; p < 00, since Lp-problem (2.1.1) is an increasing
function of the corresponding problem without the exponent.
If p = 00, the metric is also called a Tchebycheff metric and the Loo- or
the Tchebycheff problem is of the form

minimize
(2.1.2)
subject to XES.

Notice that problem (2.1.2) is nondifIerentiable even in the absence of absolute

values. In this case, it can, however, be transformed into a differentiable form
if the objective and the constraint functions are differentiable. Then, instead
of problem (2.1.2), the problem

minimize a:
subject to a: 2: Ji (x) - zt for all i = 1, ... , k,
xES,

is solved, where both x ERn and a: E R are variables.

The solution obtained depends greatly on the value chosen for p. Widely
used choices are p = 1,2 or 00. In Figure 2.1.1, the contours of these three
different metrics are shown. The black point is the ideal objective vector and
the bold line represents the Pareto optimal set. It is worth noticing that if the
original problem is linear, the choice p = 1 maintains the linearity. As the value
of p increases, the nonlinear minimization problem becomes more difficult and
badly conditioned to solve.
For linear problems, the solutions obtained by the Lp-problems where
1 < p < 00 are situated between the solutions obtained by the L 1 - and Loo-
problems. It is illustrated in Zeleny (1973) that this set of solutions is a part
of the Pareto optimal set, but only a substantially small part.
Instead of the terms Ih(x) - ztl, denominators may be added to prob-
lems (2.1.1) and (2.1.2) to normalize the components, that is, to use the terms
Ih(x) - ztl Ilztl instead. Some other denominators, like Izyad - ztl, can also
be used. The reason for employing denominators is that sometimes it is worth-
while to use relative distances in the calculations. For example, using the com-
ponents of z* forms the contour of the metric to reflect better the location of
the ideal objective vector. Naturally, the denominators zt cannot be used if
some of them equals zero.
The objective functions may also be normalized by
 2.1. Method of the Global Criterion 69

, .,'

, . I ",

/.... ideal
.... ~ criterion vector
.: I •
• ••••1

r ..
'"" .',.f
,, ". .... .' ,
,
,,' I
, ". ,,'

.... L 1 - metric

D Lao - metric

Figure 2.1.1. Different metrics.

(2.1.3) !i(X) = h(x} - zt

maxxES fi(X) - zt
before the distance is minimized. In this case, the range of the new objec-
tive functions is [0,1]. This normalizing is possible only if the objectives are
bounded. However, it is usually better to employ the ranges of the Pareto opti-
mal set and replace the max term by the component of the approximated nadir
zr
objective vector ad in (2.1.3).
A variation of the Tchebycheff problem is suggested in Osyczka (1989a,
1992), where the problem to be solved is

(2.1.4)
minimize
i~~"~.k [max [ lfi(X)-ztllh(X)-ztll]
zt ' hex)
subject to xES.

2.1.2. Theoretical Results

Next, we present some theoretical results concerning the method of the

global criterion. We assume that we know the global ideal objective vector and
can, thus, leave the absolute values.
70 Part II - 2. No-Preference Methods

Theorem 2.1.1. The solution of Lp-problem (2.1.1) (where 1 ::; p < 00) is
Pareto optimal.

Proof. Let x· E S be a solution of problem (2.1.1) with 1 ::; p < 00. Let us
suppose that x· is not Pareto optimal. Then there exists a point xES such
that fi(X) ::; fi(X*) for all i = 1, ... , k and fJ(x) < fJ(x*) for at least one j.
Now (h(x) - zt)p ::; (h(x*) - zt)P for all i and (fJ(x) - zr)p < (fJ(x*) - zr)P.
From this we obtain
k k
L (h(x) - zt)P < L (h(x*) - zt)p.
i=1 i=1

When both sides of the inequality are raised into the power 1/P we have a
contradiction to the assumption that x· is a solution of problem (2.1.1). This
completes the proof. D

Yu has pointed out in Yu (1973) that if Z is a convex set, then for 1 < p < 00
the solution of problem (2.1.1) is unique.

Theorem 2.1.2. The solution of Tchebycheff problem (2.1.2) is weakly Pareto

optimal.

Proof. Let x· E S be a solution of problem (2.1.2). Let us suppose that x* is

not weakly Pareto optimal. In this case, there exists a point xES such that
fi(X) < J;(x*) for all i = 1, ... , k. It means that, fi(X) - zt < h(x*) - zt
for all i. Thus, x· cannot be a solution of problem (2.1.2). Here we have a
contradiction which completes the proof. D

Theorem 2.1.3. Tchebycheff problem (2.1.2) has at least one Pareto optimal
solution.

Proof. Let us suppose that none of the optimal solutions of problem (2.1.2) is
Pareto optimal. Let x* E S be an optimal solution of problem (2.1.2). Since we
assume that it is not Pareto optimal, there must exist a solution xES which is
not optimal for problem (2.1.2) but for which J;(x) ::;: J;(x·) for all i = 1, ... , k
and fj(x) < fJ(x*) for at least one j.
We have now fi(X) - zt ::; J;(x·) - zt for all i with the strict inequality
holding for at least one index j, and further maxi[Ji(X)-Zt] ::;: maxdfi(x*)-ztl.
Because x* is an optimal solution of problem (2.1.2), x has to be an optimal
solution, as well. This contradiction completes the proof. D

Corollary 2.1.4. If Tchebycheff problem (2.1.2) has a unique solution, it is

Pareto optimal.

A linear numerical application example of the method is given in Hwang and

Masud (1979, pp. 23-29). Sufficient conditions for the solution of an Lp-problem
 2.2. Multiobjective Proximal Bundle Method 71

to be stable with respect to changes in the feasible region S are presented in

Jurkiewicz (1983). Reference points more general than the ideal objective vector
guaranteeing Pareto optimal results are handled in Skulimowski (1996).

2.1.3. Concluding Remarks

The method of the global criterion is a simple method to use if the aim
is simply to obtain a solution where no special hopes are set. The properties
of the metrics imply that if the objective functions are not normalized in any
way, then an objective function whose ideal objective value is situated nearer
the feasible objective region receives more importance.
The solution obtained with the Lp-metric (1 ~ p < 00) is guaranteed to be
Pareto optimal. If the Tchebycheff metric is used, the solution may be weakly
Pareto optimal. In the latter case, for instance, problem (2.10.2) of Part I can
be used to produce Pareto optimal solutions. It is up to the analyst to select
an appropriate metric.

2.2. Multiobjective Proximal Bundle Method

The multiobjective proximal bundle (MPB) method is an extension of

single-objective bundle-based methods of nondifferentiable optimization into
the multiobjective case. It is derived in Makela (1993) and Miettinen and
Makela (1995, 1996a) according to the ideas of Kiwiel (1984, 1985a, b) and
Wang (1989). The underlying proximal bundle method, presented in Kiwiel
(1990), is an advanced version of the bundle family for convex, unconstrained
non differentiable single objective optimization. It is generalized for nonconvex
and constrained problems in Makela and Neittaanmaki (1992, pp. 112~137).
The idea of the MPB method, in brief, is to move in a direction where the
values of all the objective functions improve simultaneously. Here we describe
features of the MPB method from an implementational viewpoint. For details
see Makela (1993).

2.2.1. Introduction

The MPB method is capable of solving problems with nonlinear (possi-

bly nondifferentiable) functions. It is assumed that all the objective and the
constraint functions are locally Lipschitzian.
The MPB method is not like the other scalarization methods. Ordinary
scalarization methods transform the problem into one with a single objective
function. This new problem can then be solved with any appropriate method
for nonlinear programming. In the MPB method the scalarizing function lies
inside a special (nondifferentiable) optimizer, which is why its philosophy is so
72 Part II - 2. No-Preference Methods

different from that of the other methods described here, and why implemen-
tational aspects that can be forgotten with other methods have to be touched
on.
Let us now suppose that the feasible region is of the form

As mentioned above, the MPB method is not directly based on employing

any scalarizing function. Some kind of scalarization is, however, needed in de-
riving the minimization method for all the objective functions. Theoretically,
we utilize an unconstrained improvement function H: R n x R n -t R defined
by

Let us first prove a result about the optimal solutions of improvement func-
tions. The sufficient condition necessitates the Slater constraint qualification
(Definition 3.2.13 in Part J).

Theorem 2.2.1. A necessary condition for a point x* E R n to be weakly

Pareto optimal is that the improvement function H (x, x·) attains its minimum
at x'. If the multiobjective optimization problem is convex and the Slater
constraint qualification is satisfied, then the condition above is also sufficient.

Proof. The necessity follows immediately from the proof of Theorem 3.2.5
(and Corollary 3.2.7) in Section 3.2 of Part I.
As to the sufficiency component, let the assumptions stated be valid and
let x· E Rn be a minimal solution of H(x, x*). Let us assume that x· is not
weakly Pareto optimal. Then, there exists some x E R n such that gj(x) ::; 0
for all j = 1, ... , m and Ji(x) < fi(x*) for all i = 1, ... , k. If gj(x) < 0 for all
j = 1, ... ,m, then
H(x, x*) < 0 = H(x·, x·),
which contradicts the assumption that H(x,x·) attains its minimum at x*.
Otherwise, that is, if gj(x) = 0 for some index j, it follows from the Slater
constraint qualification that there exists some x E R n such that gj (x) < 0 for
all j = 1, ... , m. If fi(X) < fi(X*) for all i = 1, ... , k, then
H(x, x*) < 0 = H(x·, x*)

and we have a contradiction with x· minimizing H.

Otherwise, we define 10 C {I, ... , k} such that fi(X) ~ fi(X*) > Ji(x) for
all i E 10 . Let us denote y = AX + (1 - A)X for 0 < A < 1. Then the convexity
of the constraint and the objective functions implies that for all 0 < >. < 1

(2.2.1)

for all j = 1, ... ,m, and

 2.2. Multiobjective Proximal Bundle Method 73

hey) ~ >"h(x) + (1 - >")h(x) < >..fi(X*) + (1- >")h(x*) = h(x*)

for all i E {I, ... ,k} \ 10 , If 10 is nonempty, we choose

>.. = min [fi(X*) -h(~) I ~ E 10 ] _ c,

max [hex) -hex) I Z E 101
where c > 0 is small enough so that 0 < >.. < 1. By the convexity of the
objective functions, we have for all i E 10

(2.2.2)
h(x*) - fi(X) _ ( fi(X*) -h(x»)
hex) _ fi(X) fi(X) + hex) -hex) hex) -
A _ A

fi(Y) ~ 1- c(fi(X) - fi(X»

= h(x*) - c(fi(X) - fi(X» < fi(X*),
Then, combining the results (2.2.1) and (2.2.2), we obtain

H(y,x*) < 0 = H(x*,x*),

which is again a contradiction with the condition that x* minimized H. Thus,
x" is weakly Pareto optimal. 0

2.2.2. MPB Algorithm

In the following, we take a look at the MPB method. We do not describe

the method completely but present its idea roughly. The reason is that the
structure of the method is highly connected to the underlying nondifferentiable
proximal bundle method.
In the MPB method, the solution is looked for iteratively, until some stop-
ping criterion is fulfilled. The iteration counter h refers to the inner iterations
of the MPB method. Let xh be the current approximation to the solution of
the multiobjective optimization problem at the iteration h. Then, by Theorem
2.2.1, we seek for the search direction d h as a solution of the unconstrained
optimization problem
minimize H(x h + d, xh)
(2.2.3)
subject to dE Rn.

Since problem (2.2.3) is still nondifferentiable, we must approximate it some-

how.
Let us assume for a moment that the problem is convex. We suppose that,
at the iteration h in addition to the iteration point xh, we have some auxiliary
points yi ERn from past iterations and subgradients{}. E Ofi(yi) for j E Jh,
i = 1, ... , k, and~, E og!(yi) for j E Jh, l = 1, ... , m, where Jh is a nonempty
subset of {I, ... , h}. We linearize the objective and the constraint functions at
the point yi and denote
74 Part II - 2. No-Preference Methods

fi,j(X) = fi(yj) + ({jif(x - yj) for all i = 1, ... ,k, j E Jh and

91,j(X) = gl(yj) + ({~,)T(x - yJ) for all l = 1, ... ,m, j E J h .

We can now define a convex piecewise linear approximation to the improve-

ment function by

jjh(X) = max [h,j(x) - j;(x h ), 91,j(X) 1i = 1, ... , k, l = 1, ... , m, j E Jh]

and we get an approximation to (2.2.3) by

minimize iIh(Xh + d) + tuhlldl12

(2.2.4)
subject to dE Rn,

where u h > 0 is an inner parameter to be updated automatically. The penalty

term tuhlldll2 is added to guarantee that there exists a solution to problem
(2.2.4) and to keep the approximation local enough.
Notice that (2.2.4) is still a nondifferentiable problem, but due to its min-
max-nature it is equivalent to the following differentiable quadratic problem
with d ERn and v ERas variables:
minimize v + tuhl/dl/2
v 2: -o.I"j + ({Ii) d,
h j T h
(2.2.5) subject to i = 1, . .. ,k; j EJ
v 2: -o.;"j + ({~,)T d, l = 1, ... ,m, j E J h ,
where

o.",j = fi(xh) - fi,j(xh), i = 1, ... , k, j E Jh and

o.;"j = =
-lh,j(x h ), l 1, ... , m, j E Jh
are so-called linearization errors.
In the nonconvex case, we replace the linearization errors by so-called sub-
gradient locality measures:

f3j"j = max [Io.tjl ,I'J;/Ixh - y i l/2]

f3~"j = max [Io.;"jl ,1'9' II x" - yjl/2],
where 1'1, 2: 0 for i = 1, ... ,k and 1'9' 2: 0 for l = 1, ... ,m are so-called distance
measure parameters (/'/, = 0 if fi is convex and 1'9' = 0 if gl is convex).
Let (dh, v h ) be a solution of problem (2.2.5). Then the two-point line search
strategy is carried out. It detects discontinuities in the gradients of the objective
functions. Roughly speaking, we try to find a step-size 0 < t h ~ 1 such that
H(x h + thd", xh) is minimal when xh + thd h E S.
A line search algorithm (in MakeHi and Neittaanmaki (1992, pp. 126-130))
is used to produce the step-sizes. The iteration is terminated when a pre-
determined accuracy is reached. The subgradient aggregation strategy due to
Kiwiel (1985c) is used to bound the storage requirements (i.e., the size of the
 2.2. Multiobjective Proximal Bundle Method 75

index set Jh} and a modification of the weight-updating algorithm described in

Kiwiel (1990) is used to update the weight u h • For details, see Miettinen and
Makela (1995, 1998a).
This is roughly the MPB method. Next, some words about optimality are
in order.

2.2.3. Theoretical Results

According to Theorem 2.2.1 we, on the one hand, know that minimizing an
improvement function produces weakly Pareto optimal solutions. On the other
hand, any weakly Pareto optimal solution of a convex problem can be found
under minor conditions. While we do not optimize the improvement function
but its approximation, the optimality results of the MPB method are somewhat
different. Here we only present some results without proofs, since giving these
would necessitate explicit expression of the MPB algorithm.

Theorem 2.2.2. Let the multiobjective optimization problem be convex and

the Slater constraint qualification be satisfied. If the MPB method stops with
a finite number of iterations, then the solution is weakly Pareto optimal. On
the other hand, any accumulation point of the infinite sequence of solutions
generated by the MPB method is weakly Pareto optimal.

Proof. See Kiwiel (1985a) or Wang (1989).

If the convexity assumption is not satisfied, we obtain somewhat weaker

results about substationary points (See Definition 3.2.6 of Part I). This result
involves upper semidifferentiable functions (see Definition 2.1.15 of Part I).

Theorem 2.2.3. Let the objective and the constraint functions of the multi-
objective optimization problem be upper semidifferentiable at every XES. If
the MPB method stops with a finite number of iterations, then the solution is a
substationary point. On the other hand, any accumulation point of an infinite
sequence of solutions generated by the MPB method is a substationary point.

Proof. See Wang (1989) and references therein.

Note that only the substationarity of the solutions of the MPB routine is
guaranteed for general multiobjective optimization problems.

2.2.4. Concluding Remarks

The MPB method can be used as a method where no opinions of the decision
maker are sought. In this case, we must select the starting point so that it is not
(weakly) Pareto optimal but that every component of the objective vector can
76 Part II - 2. No-Preference Methods

be improved. The method can also handle other than nonlinear constraints, but
they have not been included here for the sake of the clarity of the presentation.
The MPB routine can also be used as a black-box optimized within inter-
active multiobjective optimization methods. This is the case with the vector
version of NIMBUS (see Section 5.12).
The accuracy of the computation in the MPB method is an interesting mat-
ter. Accuracy can be considered in a more extensive meaning as a separating
factor between ordinary scalarizing functions and inner scalarizing function, as
in the MPB method. If some ordinary scalarizing function is employed, then it
is the accuracy of that additional function that can be followed along with the
solution process. It may happen that when the accuracy of the scalarizing func-
tion has reached the desired level, the values of the actual objective functions
could still change considerably.
Many scalarizing functions have positive features whose importance is not to
be underestimated, such as producing only Pareto optimal solutions. However,
employing some scalarizing function usually brings along extra parameters and
the difficulty of specifying their values. This causes additional stability concern.
To put it briefly, scalarizing functions add extra characteristics to the problem.
Scalarization cannot completely be avoided even in the MPB routine. How-
ever, the scalarization is carried out under the surface, invisible to the user.
Whatever additional parameters or phases are needed, they cannot be seen
and the user does not have to be bothered with them. The weakness of the
MPB routin'e is that the Pareto optimality of the solutions obtained cannot
be guaranteed. In theory, only the substationarity of the solutions is certain.
In practice, it is, however, very likely that the solutions are at least weakly
Pareto optimal. As a matter of fact, in the numerical experiments performed,
the final solutions obtained have usually proved to be Pareto optimal at the
final testing.
For problems with nondifferentiable functions the MPB routine represents
an efficient proximal bundle-based solution approach. The implementation of
the MPB routine (called MPBNGC) is described in Makela (1993). It calls a
quadratic solver derived in Kiwiel (1986).
 3. A POSTERIORI METHODS

A posteriori methods could also be called methods for generating Pareto

optimal solutions. After the Pareto optimal set (or a part of it) has been gen-
erated, it is presented to the decision maker, who selects the most preferred
among the alternatives. The inconveniences here are that the generation process
is usually computationally expensive and sometimes in part, at least, difficult.
On the other hand, it is hard for the decision maker to select from a large set
of alternatives. One more important question is how to present or display the
alternatives to the decision maker in an effective way. The working order in
these methods is: 1) analyst, 2) decision maker.
If there are only two objective functions, the Pareto optimal set can be gen-
erated parametrically {see, for example, Benson (1979b) and Gass and Saaty
(1955». When there are more objectives, the problem becomes more compli-
cated.
Let us briefly mention that in solving MOLP problems the methods can be
divided into two subclasses. In the first are the methods that can find all the
Pareto optimal solutions and in the second are the methods that can find only
all the Pareto optimal extreme solutions. In the latter case, edges connecting
Pareto optimal extreme points may be Pareto optimal or not. In nonlinear
problems, the distinction lies between convex and non convex problems. In other
words, some methods can only generate Pareto optimal solutions of convex
problems.
The methods presented in detail here are called basic methods, since they
are used frequently in practical problems, and they are also used as elements
of more developed methods. Basic methods are the weighting method and the
€-constraint method. After them, we give a limited overview of a method com-
bining features of both the weighting and the €-constraint method. Then we in-
troduce two more basic methods. The method of weighted metrics is a weighted
extension of the method of the global criterion. It is followed by the handling
of achievement scalarizing functions. Finally, some other methods in this class
are briefly mentioned.

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
78 Part II - 3. A Posteriori Methods

3.1. Weighting Method

In the weighting method, presented, for example, in Gass and Saaty (1955)
and Zadeh (1963), the idea is to associate each objective function with a weight-
ing coefficient and minimize the weighted sum of the objectives. In this way, the
multiple objective functions are transformed into a single objective function.
We suppose that the weighting coefficients Wi are real numbers such that Wi ~ 0
for all i = 1, ... , k. It is also usually supposed that the weights are normalized,
that is, L~=l Wi = 1. To be more exact, the multiobjective optimization prob-
lem is modified into the following problem, to be called a weighting problem:

minimize
(3.1.1)
i=l
subject to xES,

where Wi ~ 0 for all i = 1, ... , k and L~=l Wi = 1.

3.1.1. Theoretical Results

In the following, several theoretical results are presented concerning the

weighting method.

Theorem 3.1.1. The solution of weighting problem (3.1.1) is weakly Pareto

optimal.

Proof. Let x* E S be a solution of the weighting problem. Let us suppose

that it is not weakly Pareto optimal. In this case, there exists a solution XES
such that j;(x) < j;(x*) for all i = 1, ... , k. According to the assumptions
set to the weighting coefficients, Wj > 0 for at least one j. Thus we have
L~=l Wdi(X) < L~=l Wdi(X*). This is a contradiction to the assumption that
x* is a solution of the weighting problem. Thus x* is weakly Pareto optimal.
o
Theorem 3.1.2. The solution of weighting problem (3.1.1) is Pareto optimal
if the weighting coefficients are positive, that is Wi > 0 for all i = 1, ... , k.

Proof. Let x* E S be a solution of the weighting problem with positive weight-

ing coefficients. Let us suppose that it is not Pareto optimal. This means that
there exists a solution xES such that /i(X) :s J;(x*) for all i = 1, ... , k and
/j(x) < /j(x*) for at least one j. Since Wi > 0 for all i = 1, ... , k, we have
L~=l Wdi(X) < L~=l Wdi(X*). This contradicts the assumption that x' is a
solution of the weighting problem and, thus, x' must be Pareto optimal. 0
 3.1. Weighting Method 79

Theorem 3.1.3. The unique solution of weighting problem (3.1.1) is Pareto

optimal.

Proof. Let X* E S be a unique solution of the weighting problem. Let us

suppose that it is not Pareto optimal. In this case, there exists a solution
xES such that fi(X) ~ h(x*) for all i = 1, ... , k and h(x) < h(x*) for
at least one j. Because all the weighting coefficients Wi are nonnegative, we
have E~=l Wdi(X) ~ E~l Wdi(X*). On the other hand, the uniqueness of x*
means that L~=l Wdi(X*) < E~=l wih(x) for all XES. The two inequalities
above are contradictory and, thus, X* must be Pareto optimal. 0

As Theorems 3.1.2 and 3.1.3 state, the solution of the weighting method is
always Pareto optimal if the weighting coefficients are all positive or if the solu-
tion is unique, without any further assumptions. The weakness of the weighting
method is that not all of the Pareto optimal solutions can be found unless the
problem is convex. This feature can be relaxed to some extent by convexifying
the nonconvex Pareto optimal set as suggested in Li (1996). The convexifica-
tion is realized by raising the objective functions to a high enough power under
certain assumptions. However, the main result is the following:

Theorem 3.1.4. Let the multiobjective optimization problem be convex. If

x* E S is Pareto optimal, then there exists a weighting vector w (Wi ~ 0,
i = 1, ... , k, E~=l Wi = 1) such that X* is a solution of weighting problem
(3.1.1).

Proof. The proof is presented after Theorem 3.2.6.

According to Theorem 3.1.4 any Pareto optimal solution of a convex mul-

tiobjective optimization problem can be found by the weighting method. Note
that the weighting vector is not necessarily unique. The contents of Theorem
3.1.4 is illustrated in Figure 3.1.1. On the left, every Pareto optimal solution
along the bold line can be obtained by altering the weighting coefficients. On
the right, it is not possible to obtain the Pareto optimal sol.utions in the 'hol-
low.'
An equivalent trigonometric formulation to the weighting problem with two
objective functions is presented in Das and Dennis (1997). This formulation can
be used in illustrating geometrically why not all the Pareto optimal solutions
of nonconvex problems can be found.

Remark 3.1.5. According to Theorem 3.1.4, all the Pareto optimal solutions
of MOLP problems can be found by the weighting method.

Let us have a look at linear cases for a while. In practice, Remark 3.1.5 is
not quite true. The single objective optimization routines for linear problems
80 Part II -~~ 3. A Posteriori Methods

Z I Z I

Figure 3.1.1. Weighting method with convex and non convex problems.

usually find only extreme point solutions. Thus, if some facet of the feasible re-
gion is Pareto optimal, then the infinity of Pareto optimal non-extreme points
must be described in terms of linear combinations of the Pareto optimal ex-
treme solutions. On the other hand, note that if two adjacent Pareto optimal
extreme points for an MOLP problem are found, the edge connecting them is
not necessarily Pareto optimal.
The conditions under which the whole Pareto optimal set can be gener-
ated by the weighting method with positive weighting coefficients are presented
in Censor (1977). The solutions that it is possible to reach by the weighting
method with positive weighting coefficients are characterized in Belkeziz and
Pirlot (1991). Some generalized results are also given. More relations between
nonnegative and positive weighting coefficients, convexity of Sand Z, and
Pareto optimality are studied in Lin (1976b).
If the weighting coefficients in the weighting method are all positive, we
can say more about the solutions than that they are Pareto optimal. The fol-
lowing results concerning proper Pareto optimality were originally presented in
Geoffrion (1968).

Theorem 3.1.6. The solution of weighting problem (3.1.1) is properly Pareto

optimal if all the weighting coefficients are positive (sufficient condition).

Proof. Let X* E S be a solution of the weighting problem with positive

weighting coefficients. In Theorem 3.1.2 we showed that the solution is Pareto
optimal. We shall now show that x* is properly Pareto optimal with M =
(k - 1) maXi,j(Wj/Wi).
Let us on the contrary suppose that x* is not properly Pareto optimal. Then
for some i (which we fix) and for xES such that fi(X*) > fi(X) we have

j;(x*) - j;(x) > M(h(x) -h(x*))

for all j such that h(x*) < fj(x). We can now write
 3.1. Weighting Method 81

After multiplying both sides by wd(k - 1) > 0, we get

k ~ 1 (Ji(X") -/i(x» °
> wi(h(x) - h(x"» ( > ~ WI(JI(X) - fl(x"») ,
where l differs from the fixed index i and the indices j, which were specified
earlier. After this reasoning we can sum over all j :f:. i and obtain
k
w;(/i(x") -/i(x» > L(Wi(Jj(X) - h(x"»),
j=l
Noi
which means
k k
L wih(x*) > L wih(x).
j=1 i=l
Here we have a contradiction to the assumption that x" is a solution of the
weighting problem. Thus, x" has to be properly Pareto optimal. 0

Theorem 3.1.7. If the multiobjective optimization problem is convex, then

the condition in Theorem 3.1.6 is also necessary.

Proof. See Geoffrion (1968) or Chou et al. (1985).

Corollary 3.1.8. A necessary and sufficient condition for a point to be a

properly Pareto optimal solution of an MOLP problem is that it is a solution
of a weighting problem with all the weighting coefficients being positive.

The ratio of the weighting coefficients gives an upper bound to global trade-
offs.

Theorem 3.1.9. Let x* be a solution of weighting problem (3.1.1), when all

the weighting coefficients Wi, i = 1, ... , k, are positive. Then
G W·
A. .(x") < max .-l..
'J - i=1, ... ,k Wi
i#-i
for every i, j = 1, ... ,k, i ::f. j.

Proof. See Kaliszewski (1994, p. 93).

Some results concerning weak, proper and Pareto optimality of the solutions
obtained by the weighting method are combined in Wierzbicki (1986b). Proper
Pareto optimality and the weighting method are also discussed in Belkeziz and
Pirlot (1991) and Luc (1995). The weighting method is used in Isermann (1974)
82 Part II - 3. A Posteriori Methods

in proving that for linear multiobjective optimization problems all the Pareto
optimal solutions are also properly Pareto optimal. More results concerning
MOLP problems and the weighting method are assembled in Chankong and
Haimes (1983a, b, pp. 153-159).

3.1.2. Applications and Extensions

As far as applications are concerned, the weighting method is used to gen-

erate Pareto optimal solutions in Sadek et al. (1988-89) in solving a problem of
the optimal control of a damped beam. The weighting method is also applied in
Weck and Fortsch (1988) to structural systems in the optimization of a spindle
bearing system and in the optimization of a table, as well as in ReVelle (1988),
where reductions in strategic nuclear weapons for the two superpowers are ex-
amined. Furthermore, Pareto optimal solutions are generated for an anti-lock
brake system control problem by the weighting method in Athan and Papalam-
bros (1997). However, no attention is paid to the possible nonconvexity of the
problem. In addition, the weighting method is an essential component in the
determination of the optimal size of a batch system in Friedman and Mehrez
(1992) and a fuzzy optimal design problem concerning a bridge is solved in
Ohkubo et al. (1998).
Linear problems with two objective functions are studied in Gass and Saaty
(1955). The systematic generation of the Pareto optimal set is possible in these
problems by parametric optimization. The Pareto optimal set of multiobjective
optimization problems with convex quadratic objective functions and linear
equality constraints is characterized analytically in Goh and Yang (1996). The
characterization involves the weighting method and active set methods.
Systematic ways of perturbing the weights to obtain different Pareto opti-
mal solutions are suggested in Chankong and Haimes (1983a, b, pp. 234-236).
In addition, an algorithm for generating different weighting coefficients auto-
matically for convex (nonlinear) problems to produce an approximation of the
Pareto optimal set is proposed in Caballero et al. (1997).
A method for reducing the Pareto optimal set (of an MOLP problem) before
it is presented to the decision maker is suggested in Soloveychik (1983). Pareto
optimal solutions are first generated by the weighting method. Then, statistical
analysis (factor analysis) is used to group and partition the Pareto optimal set
into groups of relatively homogeneous elements. Finally, typical solutions from
the groups are chosen and presented to the decision maker.
It is suggested in Koski and Silvennoinen (1987) that the weighting method
can be used to reduce the number of the objective functions before the actual
solution process. The original objective functions are divided into groups such
that a linear combination of the objective functions in each group forms a new
objective function, and these new objective functions form a new multiobjective
optimization problem. It is stated that every Pareto optimal solution of the
new problem is also a Pareto optimal solution of the original problem, but the
reverse result is not generally true.
 3.1. Weighting Method 83

As mentioned earlier, the weighting vector that produces a certain Pareto

optimal solution is not necessarily unique. This is particularly true for linear
problems. A method is presented in Steuer (1986, pp. 183-187) for determining
ranges for weighting vectors that produce the same solution. Note that some
weighting vectors may produce unbounded single objective optimization prob-
lems. This does not mean that the problem may not have feasible solutions
with some other weighting vectors.
A property related to producing different Pareto optimal solutions by al-
tering the weighting coefficients is the weak stability of the system. On the one
hand, a small change in the weighting coefficients may cause big changes in
the objective vectors. On the other hand, dramatically different weighting co-
efficients may produce nearly similar objective vectors. The reason is that the
weighting problem is not a Lipschitzian function of the weighting coefficients.
In addition, it is emphasized and illustrated in Das and Dennis (1997) that
an evenly distributed set of weighting vectors does not necessarily produce an
evenly distributed representation of the Pareto optimal set, even if the problem
is convex. Further, Das and Dennis demonstrate how an even spread of Pareto
optimal solutions is obtained only for very special shapes of Pareto optimal
sets. The treatment concerns two objective functions.
An entropy-based formulation of the weighting method is suggested in Sul-
tan and Templeman (1996). The entropy-based objective function to be op-
timized has only one parameter no matter what the number of the original
objective functions is. A representation of the Pareto optimal set can be gen-
erated by varying the value of the single parameter. The properties of the
suggested method are the same as those of the weighting method, for example,
all the Pareto optimal solutions of nonconvex problems cannot be found.

3.1.3. Weighting Method as an A Priori Method

The weighting method can be used so that the decision maker specifies a
weighting vector representing her or his preference information. In this case, the
weighting problem can be considered (a negative of) a value function (remember
that value functions are maximized). Note that according to Remark 2.8.7
of Part I, the weighting coefficients provided by the decision maker are now
nothing but marginal rates of substitution (mii = Wj/Wi). When the weighting
method is used in this fashion, it can be considered to belong to the class of a
priori methods. Related to this, a method for assisting in the determination of
the weighting coefficients is presented in Batishchev et al. (1991). This method
can also be extended into an interactive form by letting the decision maker
modify the weighting vectors after each iteration.
The objective functions should be normalized or scaled so that their ob-
jective values are approximately of the same magnitude (see Subsection 2.4.3
in Part I). Only in this way can one control and manoeuvre the method to
produce solutions of a desirable nature in proportion to the ranges of the ob-
84 Part II - 3. A Posteriori Methods

jective functions. Otherwise the role of the weighting coefficients may be greatly
misleading.
If the weighting method is used as an a priori method one can ask what
the weighting coefficients in fact represent. Often they are said to reflect the
relative importance of the objective functions. However, it is not at all clear
what underlies this notion, as discussed in Roy and Mousseau (1996). It is
remarked in Hobbs (1986) that instead of relative importance, the weighting
coefficients should represent the rate at which the decision maker is willing to
trade off values of the objective functions.
It must be noted that if some of the objective functions correlate with each
other, then seemingly 'good' weighting vectors may produce poor results and
seemingly 'bad' weighting vectors may produce useful results (see Steuer (1986,
pp. 198-199) for an illustrative example).
On the basis of practical experience it is emphasized in Wierzbicki (1986b)
that weighting coefficients are not easy to interpret and understand for average
decision makers.

3.1.4. Concluding Remarks

The weighting method is a simple way to generate different Pareto opti-

mal solutions. Pareto optimality is guaranteed if the weighting coefficients are
positive or the solution is unique.
Applying Theorem 3.1.2, we know t.hat different Pareto optimal solutions
can be obtained by the weighting method by altering the positive weighting
coefficients. However, in practical calculations the condition Wi 2: E, where
t: > 0, must be used instead of the condition Wi > 0 for all i = 1, ... , k.
This necessitates a correct choice as to the value of E. All the Pareto optimal
solutions in some convex problems may be found if t: is small enough. But the
concept of 'small enough' is problem-dependent and for this reason difficult to
specify in advance, as pointed out in Korhonen and Wallenius (1989a).
As observed before, the weakness of the weighting method is that all of the
Pareto optimal points cannot be found if the problem is nonconvex. If this is
the case, a duality gap is said to occur (according to duality theory). The same
weakness may also occur in problems with discontinuous objective functions as
demonstrated in Kitagawa et al. (1982).
Sometimes, the results concerning the weighting method are presented in a
simpler form, assuming that zeros are not accepted as weighting coefficients. It
may seem that the weighting coefficient zero makes no sense. It means that we
have included in the problem some objective function that has no significance
at all. Nevertheless, zero values have here been included to make the presen-
tation more general. On the other hand, by also allowing zeros as weighting
coefficients, it is easy to explore how solutions change when some objective
fUIlction is dropped.
Employing the weighting method as an a priori method presumes that the
decision maker's underlying value function is or can be approximated by a linear
 3.2. c:-Constraint Method 85

function (see Section 2.6 in Part I). This is in many cases a rather simplifying
assumption. In addition, it must be noted that altering the weighting vectors
linearly does not have to mean that the values of the objective functions also
change linearly. It is, moreover, difficult to control the direction of the solutions
by the weighting coefficients, as illustrated in Nakayama (1995).

3.2. e-Constraint Method

In the c:-constraint method, introduced in Haimes et al. (1971), one of the

objective functions is selected to be optimized and all the other objective func-
tions are converted into constraints by setting an upper bound to each of them.
The problem to be solved is now of the form
minimize It (x)
(3.2.1) subject to J; (x) ~ C:j for all j = 1, ... ,k, j:f; t,
XES,

where t E {1, ... ,k}. Problem (3.2.1) is called an c:-constraint problem.

An alternative formulation is proposed in Lin (1976a, b), where proper
equality constraints are used instead of the above-mentioned inequalities. The
solutions obtained by this proper equality method are Pareto optimal under
certain assumptions. Here, however, we concentrate on formulation (3.2.1).

3.2.1. Theoretical Results on Weak and Pareto Optimality

We begin by proving a result concerning weak Pareto optimality.

Theorem 3.2.1. The solution of c:-constraint problem (3.2.1) is weakly Pareto

optimal.

Proof. Let x· E S be a solution of the c:-constraint problem. Let us assume

that x· is not weakly Pareto optimal. In this case, there exists some other
xES such that fi(X) < Ji(x·) for all i = 1, ... , k.
This means that J;(x) < J;(x·) ~ C:j for all j = 1, ... ,k, j :f; t. Thus x
is feasible with respect to the c:-constraint problem. While in addition /L(x) <
!l(x·), we have a contradiction to the assumption that x· is a solution of the
c:-constraint problem. Thus, x· has to be weakly Pareto optimal. 0

Next, we handle Pareto optimality and the c:-constraint method.

Theorem 3.2.2. A decision vector x· E S is Pareto optimal if and only if

it is a solution of c:-constraint problem (3.2.1) for every t = 1, ... , k, where
C:j = J;(x·) for j = 1, ... , k, j :f; t.
86 Part II - 3. A Posteriori Methods

Proof. Necessity: Let x· E S be Pareto optimal. Let us assume that it does not
solve the €-constraint problem for some e where €j = h(x·) for j = 1, ... , k,
j =f. e. Then there exists a solution XES such that hex) < h(x·) and
hex) :s: Ij(x*) when j =f. e. This contradicts the Pareto optimality of x·. In
other words, x* has to solve the problem for any objective function.
Sufficiency: Since x* E S is by assumption a solution of the €-constraint
problem for every e = 1, ... , k, there is no xES such that hex) < h(x·)
and hex) :s: /j(x*) when j =f. e. This is the definition of Pareto optimality for
x·. 0

Note that according to the necessity component of Theorem 3.2.2, it is

possible to find every Pareto optimal solution of any multiobjective optimiza-
tion problem by the €-constraint method (regardless of the convexity of the
problem).

Theorem 3.2.3. A point x* E S is Pareto optimal if it is a unique solution

of €-constraint problem (3.2.1) for some e with €j = h(x*) for j = 1, ... , k,
j=f.f..

Proof. Let x· E S be a unique solution of the €-constraint problem for some e.

Let us assume that it is not Pareto optimal. In other words, there exists some
point XO E S such that h(xO) :s: h(x·) for all i = 1, ... , k and for at least one
index j is valid /j (XO) < Ij(x·). The uniqueness of x* means that for all XES
such that hex) :s: h(x·), i =f. e, is h(x·) < hex). Here we have a contradiction
with the preceding inequalities, and x· must be Pareto optimal. 0

The following theorem is a straightforward extension of Theorem 3.2.3.

Theorem 3.2.4. The unique solution of E-constraint problem (3.2.1) is Pareto

optimal for any given upper bound vector E = (El' ... ,Ee-l, El+l, ... ,Ek)T.

Proof. Let x· E S be a unique solution of the E-constraint problem. This

means that h(x*) < fe(x) for all xES when Ij(x*) :s: Ej for every j = 1, ... , k,
j =f. f. Let us assume that x* is not Pareto optimal. In this case, there exists a
vector XO E S such that J;(XO) :s: li(X*) for all i = 1, ... , k and the inequality
is strict for at least one index j.
If j = f, this means that fe(XO) < fe(x·) and J;(XO) :s: li(X·) :s: Ej for all
i =f. e. Here we have a contradiction with the fact that x· is a solution of the
E-constraint problem.
On the other hand, if j :f. e, then /j(XO) < Ij(x*) :s: Ej, li(XO) :s: li(X*) :s: Ei
for all i =f. j and e, and fe(XO) :s: fe(x·). This is in contradiction to x* as a
unique solution of the E-constraint problem, and x· has to be Pareto optimal.
o
In Figure 3.2.1, different upper bounds for the objective function hare
given, while the function 11 is to be minimized. The Pareto optimal set is
 3.2. e:-Constraint Method 87

shown by a bold line. The upper bound level 6"1 is too tight and so the feasible
region is empty. On the other hand, the level 6"4 does not restrict the region at
all. If it is used as the upper bound, the point Z4 is obtained as a solution. It
is Pareto optimal according to Theorem 3.2.4. Correspondingly, for the upper
bound 6"3 the point z3 is obtained as a Pareto optimal solution. The point Z2
is the optimal solution for the upper bound 6"2. Its Pareto optimality can be
proved according to Theorem 3.2.3. Theorem 3.2.2 can be applied as well.

------------------------------

Figure 3.2.1. Different upper bounds for the 6"-constraint method.

To ensure that a solution produced by the 6"-constraint method is Pareto

optimal, we have to either solve k different problems or obtain a unique solution.
In general, uniqueness is not necessarily easy to verify. However, if for example,
the problem is convex and the function It to be minimized is strictly convex,
we know that the solution is unique without further checking (see Chankong
and Haimes (1983b, p. 131)).
According to Theorem 3.2.1 we know that the 6"-constraint method pro-
duces weakly Pareto optimal solutions without any additional assumptions.
We can show that any weakly Pareto optimal solution can be found with the
6"-constraint method (for some objective function to be minimized) if the feasi-
ble region is convex and all the objective functions are quasiconvex and strictly
quasiconvex. This result is derived in Ruiz-Canales and Rufhin-Lizana (1995)
and a shortened proof is also given in Luc and Schaible (1997). The value
of this result is somewhat questionable because usually we are interested in
Pareto optimal solutions and we know that any of them can be found with the
e-constraint method.
88 Part II - 3. A Posteriori Methods

3.2.2. Connections with the Weighting Method

The relationships between the weighting method and the E-constraint

method are presented in the following theorems.

Theorem 3.2.5. Let x* E S be a solution of weighting problem (3.1.1) and

o ~ w E R k be the corresponding weighting vector. Then
(1) if We > 0, x* is a solution of the c:-constraint problem for fe as the
objective function and Ej = h(x') for j = 1, ... , k, j f:- i!; or
(2) if x· is a unique solution of weighting problem (3.1.1), then x' is a
solution of the E-constraint problem when Ej = hex") for j = 1, ... , k,
j f:- I! and for every fe, I! = 1, ... , k, as the objective function.

Proof. Let x" E S be a solution of the weighting problem for some weighting
vector 0 ~ w E Rk.
(1) We assume that We > O. We have
k k
(3.2.2) L wd;(x) 2 L wd;(x')
;=1 i=l

for all xES.

Let us suppose that x· is not a solution of the E-constraint problem. Then
there exists a point xES such that fe(x) < fe(x*) and iJ(x) ~ h(x') when
j = 1, ... ,k, j f:- f. We assumed that We > 0 and Wi ~ 0 when i f:- e. Now we
have
k
0> Wt(fl(X) - fe(x")) + L Wi (fi (x) -h(x·)),
i=l
i#i
which is a contradiction with inequality (3.2.2). Thus x" is a solution of the
E-constraint problem.
(2) If x" is a unique solution of the weighting problem, then for all xES
k k
(3.2.3) L Wdi(X") < L wd;(x).
;=1 i=l
Ifthere is some objective function fe such that x" does not solve the E-constraint
problem when fe is to be minimized, then we can find a solution xES such
that fe(x) < fe(x') and hex) ~ hex") when j f:- e. This means that for any
w 2 0 is I:~=1 wd;(x) ~ I:~=1 wd;(x·). This contradicts inequality (3.2.3).
Thus x· is a solution of the E-constraint problem for all fe to be minimized. 0

The next theorem is valid for convex problems.

Theorem 3.2.6. Let the multiobjective optimization problem be convex. If

x· E S is a solution of E-constraint problem (3.2.1) for any given fe to be
 3.2. E-Constraint Method 89

minimized and Cj = h (x") for j = 1, ... ,k, j f. l, then there exists a weighting
vector 0 ::; W E Rk, E~=1 Wi = I, such that x" is also a solution of weighting
problem (3.1.1).

Proof. The proof needs a so-called generalized Gordan theorem. See Chankong
and Haimes (1983b, p. 121) and references therein.

We have now appropriate tools for proving Theorem 3.1.4 from the previous
section.

Proof. (Proof of Theorem 3.1.4) Since x· is Pareto optimal, it is by Theorem

3.2.2 a solution of the c-constraint problem for every objective function /t to
be minimized. The proof is completed with the aid of the convexity assumption
and Theorem 3.2.6. 0

A diagram representing several results concerning the characterization

of Pareto optimal solutions and the optimality conditions of the weighting
method, the c-constraint method and a so-called jth Lagrangian method, their
relations and connections is presented in Chankong and Haimes (1982, 1983b,
pp. 119). The jth Lagrangian method, presented in Benson and Morin (1977),
means solving the problem
k
minimize h(x) +L uili(x)
(3.2.4) i=1
i¥j

subject to x E 8,

where u = (U1, ••. ,Uj_1,Uj+1, ... ,Uk)T and Ui :? 0 for all if. j. The jth
Lagrangian method is from a computational viewpoint almost equal to the
weighting method. This is why it is not studied more closely here. Chankong
and Haimes have treated the problems separately to emphasize two ways of
arriving at the same point.

3.2.3. Theoretical Results on Proper Pareto Optimality

Let us now return to the c-constraint problem and the proper Pareto op-
timality of its solutions. In Benson and Morin (1977), an auxiliary function,
called the perturbation /unction, v: Rk-1 --+ R associated with the c-constraint
problem is defined in the form (modified here for the minimization problem)

v(y) = inf {/t(x) I h(x) -

xES
Cj ::; Yj for all j = 1, ... ,k, j f. l}.
Naturally, the optimal value of the objective function of the c-constraint prob-
lem is v(O). We can now define the stability of c-constraint problems.
90 Part II -- 3. A Posteriori Methods

Definition 3.2.7. The £-constraint problem (3.2.1) is said to be stable when

v(O) is finite and there exists a scalar R > 0 such that, for all 0 i y E R k-l

v(O) - v(y) < R

Ilyll -'

After this, a theorem concerning the proper Pareto optimality of the solu-
tions of the £-constraint problem can be presented.

Theorem 3.2.8. Let the multiobjective optimization problem be convex and

let x* E S be Pareto optimal. Then x* is properly Pareto optimal if and only if
£-constraint problem (3.2.1) is stable for each f = 1, ... , k, where £j = fj(x*)
for all j = 1, ... ,k, j if.

Proof. See Benson and Morin (1977) or Sawaragi et al. (1985, p. 88).

Let us now suppose that the feasible region is of the form

S = {x ERn I g(x) = (g1 (x), g2(X), . .. , gm(x»T ::; O}.

The £-constraint problem is a constrained single objective optimization
problem and it can be converted into an unconstrained problem by formulating
a Lagrange function of the form
k m
hex) + I: Aj(/j(X) -€j) + I:/-Ligi(X)
j=1 i=l
#i

to be minimized. Setting some assumptions on the Lagrange multipliers -\ E

R k-l and pERm, we can derive more conditions for proper Pareto optimality.
In the following, we need the constraint qualification of Definition 3.1.10
of Part I, that is the definition of a regular point applied to the £-constraint
problem. In other words, a point x* E S is regular if the gradients of the active
constraints of the £-constraint problem at x* are linearly independent.
For clarity, we shall now formulate the classical Karush-Kuhn-Tucker nec-
essary condition for optimality (see Kuhn and Tucker (1951» applied to the
€-constraint problem. The proof for general nonlinear problems is presented,
for example, in Kuhn and Tucker (1951) and Luenberger (1984, p. 315). The
condition can also be derived from the optimality conditions for multiobjective
optimization problems, presented in Section 3.1 of Part 1.

Note 3.2.9. (Karush-Kuhn-Tucker necessary optimality condition applied to

the €-constraint problem) Let the objective and the constraint functions be
continuously differentiable at x* E S which is a regular point of the constraints
of the €-constraint problem. A necessary condition for x* to be a solution of the
 3.2. c-Constraint Method 91

6-constraint problem is that there exist vectors 0 ::;;\ E R k - 1 and 0 ::; I' E Rm
such that
k m
(1) "V1t(x*) + :~=>.j"V(/j(x*) -C:j) + Ll1i"Vgi(X*) = 0
j=l i=l
j#
(2) Aj(/j(X*)-€j) =0 for all j=lf, l1;g;(X*) =0 for all i=1, ... ,m.

Note that the (Lagrange) multipliers .\ are in what follows called Karush-
Kuhn-Tucker multipliers, when they are associated with the Karush-Kuhn-
Tucker optimality condition. The condition states, for example, that if the
constraint concerning /j is not active, the corresponding multiplier Aj must be
equal to zero.
We can now present the following theorem.

Theorem 3.2.10. Let all the objective and the constraint functions be con-
tinuously differentiable at x* E S which is a regular point of the constraints of
the €-constraint problem. Then the following is valid.
(1) If x* is properly Pareto optimal, then x· solves the €-constraint prob-
lem for some It being minimized and 6j = /j(x*) (for j = 1, ... , k,
j =I f) with all the Karush-Kuhn-Tucker multipliers associated with the
constraints /j(x) ::; 6j for j = 1, ... , k, j =I f, being positive.
(2) If the multiobjective optimization problem is convex, then x* is prop-
erly Pareto optimal if it is a solution of the c:-constraint problem with
the Karush-Kuhn-Tucker multipliers associated with the constraints
fJ(x) S 6j for j = 1, ... , k, j =I e, being positive.

Proof. See Chankong and Haimes (1983b, p. 143).

It can also be proved that if some solution is improperly Pareto optimal

and the problem is convex, then some of the associated Karush-Kuhn-Tucker
multipliers equal zero. On the other hand, if some of the Karush-Kuhn-Tucker
multipliers equal zero, then the solution of the c:-constraint problem is improp-
erly Pareto optimal (see, e.g., Chankong and Haimes (1983a)).
According to Theorem 3.2.10, we can say that if a multiobjective optimiza-
tion problem is solved by the c:-constraint method, proper Pareto optimality
can be checked by employing the Lagrange function. In the previous section in
connection with the weighting method, we also presented some conditions for
proper Pareto optimality. Let us mention that proper Pareto optimality is char-
acterized with the help of jth Lagrangian problem (3.2.4) in Benson and Morin
(1977). There are, however, many methods where proper Pareto optimality is
difficult to guarantee algorithmically.
92 Part II - 3. A Posteriori Methods

3.2.4. Connections with Trade-Off Rates

The relationships between Karush-Kuhn-Tucker multipliers and trade-off

rates are studied in Chankong and Haimes (1983b, pp. 159-165) and Haimes
and Chankong (1979). Indeed, under certain conditions to be presented in the
following, the Karush-Kuhn-Tucker multipliers of the Lagrange problem are
equivalent to trade-off rates.
For notational convenience we state the second-order sufficient condition
applied to the c-constraint problem. See Chankong and Haimes (1983b, p. 58)
for details.

Note 3.2.11. (Second-order sufficient condition for optimality applied to the

c-constraint problem) Let the objective and the constraint functions be twice
continuously differentiable at x" E S which is a regular point of the constraints
of the c-constraint problem. A sufficient condition for x* to be a solution of the
c-constraint problem is that there exist vectors 0 :::;.\. E Rk-l and 0 :::; I' E R m
such that the optimality condition of Note 3.2.9 is satisfied and the Hessian
matrix of the corresponding Lagrange function
k m
V 2 h(x*) + L Aj V2(fj(X") - Cj) + L JLi V2 gi(X*)
j=1 i=1
#i

is positive definite on the set {d E Rn Vgi(x,,)T d = 0 for all i such that

J.ti > OJ.
A connection between Karush-Kuhn-Tucker multipliers and trade-off rates
is presented in the following theorem. The upper bound vector is denoted by
eO E Rk-l and it is assumed to be chosen so that feasible solutions exist.

Theorem 3.2.12. Let x* E S be a solution of c-constraint problem (3.2.1)

for some It, i = 1, ... , k, to be minimized. Let Aj = Aij, for j = 1, ... , k,
j -:j:. i, be the corresponding Karush-Kuhn-Tucker multipliers associated with
the constraints /j{x) :::; cj for j -:j:. i. If
(1) x" is a regular point of the constraints of the c-constraint problem,
(2) the second-order sufficient condition of Note 3.2.11 is satisfied at x",
and
(3) there are no degenerate constraints at x" (i.e., the Karush-Kuhn-'fucker
multipliers of all the constraints are strictly positive),

then Aij = - af~(x") for all j = 1, ... , k, j -:j:. i.

Cj

Proof. The proof is based on the implicit function theorem, see Luenberger
(1984, p. 313).
 3.2. c:-Constraint Method 93

From the assumption A;(f;(X*) - Cj) = 0 for all j = 1, ... , k, j :f. l of the
Karush-Kuhn-Tucker necessary optimality condition and the nondegeneracy of
the constraints we know that I;(x*) = Cj for all j =I i. Thus, from Theorem
3.2.12 we have the trade-off rates

Al; = - oJt(x*)
01;
.
for all J =I i.
An important result concerning the relationship between Karush-Kuhn-
Tucker multipliers and trade-off rates in a more general situation, where zero-
valued multipliers also are accepted, is presented in the following. For notational
simplicity we now suppose that the function to be minimized in the c-constraint
problem is fk (Le., we set It = fk). In addition, the upper bounds e" E Rk-l
are assumed to be chosen so that feasible solutions exist. This does not lose
any generality. For details and a more extensive form of the theorem we refer
to Chankong and Haimes (1983b, pp. 161-163).
Let Ak; be the Karush-Kuhn-Tucker multipliers associated with the con-
straints I;(x) $ c'J, j = 1, ... , k - 1. Without loss of generality we can as-
sume that the first p (1 $ p :::; k - 1) of the multipliers are strictly positive
(i.e., Ak; > 0 for j = 1, ... ,p) and the rest k - 1 - p multipliers equal zero
(i.e., Akj = 0 for j = p + 1, ... , k - 1). We denote the objective vector corre-
sponding to x* by z* E Z.

Theorem 3.2.13. Let x* E S be a solution of c-constraint problem (3.2.1)

(when fir. is minimized) such that
(1) x* is a regular point of the constraints of the c-constraint problem,
(2) the second-order sufficient condition of Note 3.2.11 is satisfied at x*,
and
(3) all the active constraints at x* are nondegenerate.
Then we have the following.
1) Ifp = k - 1, that is, all the multipliers Air.; are strictly positive, then the
Pareto optimal surface in the feasible objective region in the neighbour-
hood of z* can be represented by a continuously differentiable function
fir. such that for each (Zl. Z2, ..• , Zk)T in the neighbourhood of z· is
Zk = fk(zl,z2, ... ,Zk-l). Moreover, for alII $ j $ p = k -1 is

Thus Ak; represents the partial trade-off rate between !k and I; at x· .

2) If 1 :::; p < k -1, that is, some of the multipliers Ak; equal zero, then the
Pareto optimal surface in the feasible objective region in the neighbour-
hood of z* can be represented by continuously differentiable functions
z; = /,(Zl. ... , zp,c~+l! ... ,Ck-l) for j = p + 1, ... , k. Moreover, for all
l$i$pis
94 Part II - 3. A Posteriori Methods

where d; is the direction of 8X(EO)/8f:i. In addition, for all p + 1 ~ j ~

k - 1 is

Thus Aki represents the total trade-off rate between fk and j; at x* in

the direction of 8X(EO)/8e i .

Proof. See Chankong and Haimes (1983b, pp. 163-165).

Let us now consider the contents of Theorem 3.2.13. Part 1) says that under
the given conditions there are exactly k - 1 degrees of freedom in specifying
a point on the (locally) Pareto optimal surface in the objective space in the
neighbourhood of z*. In other words, when the values for Zl, Z2, •.. , Zk-l have
been chosen from the neighbourhood of z* , then the value for Zk can be calcu-
lated from the given function and the resulting point z will lie on the (locally)
Pareto optimal surface in the objective space.
Part 2) of Theorem 3.2.13 extends the result of part 1) by relaxing the
assumption that all the constraints iJ(x) ~ f:j, j = 1, ... ,k - 1, should be
active and nondegenerate, that is, Akj > 0 for all j = 1, ... , k - 1. When
the number of nondegenerate constraints is p « k - 1), then the degree of
freedom in specifying a point on the (locally) Pareto optimal surface in the
objective space in the neighbourhood of z* is the number of non degenerate
active constraints (p). The results of Theorem 3.2.13 will be needed in Section
5.1 when the f:-constraint method is used as a part of an interactive method.

3.2.5. Applications and Extensions

Systematic ways of perturbing the upper bounds to obtain different Pareto

optimal solutions are suggested in Chankong and Haimes (1983a, b, pp. 283-
295). The f:-constraint method is used for generating Pareto optimal solutions
in Osman and Ragab (1986b). Then the solutions are clustered and a global
Pareto optimum is located.
Sensitivity analysis of the f:-constraint method is dealt with in Rarig and
Haimes (1983). An index is defined approximating the standard deviation of the
optimal solution. The objective and the constraint functions are not supposed
to be known for a certainty.
Now that we have introduced two basic methods it is worthwhile to men-
tion a method for nonlinear problems presented in Osman and Ragab (1986a).
It combines features from both the weighting method and the f:-constraint
method. The nonconvex feasible objective region is divided into convex and
 3.2. c-Constraint Method 95

nonconvex parts. The positive feature of the weighting method that the feasi-
ble region is not disturbed in the solution process is utilized in the convex parts,
and the capability of the €-constraint method to find all the Pareto optimal
solutions is utilized in the non convex parts. Therefore, merits of both these
basic methods are exploited.
A method related to the €-constraint method is presented in Youness (1995).
It generates the Pareto optimal set for problems with quasiconvex (and lower
semi continuous) objective functions. The method is based on level sets. If we
consider an objective vector zh and level sets Li(zf) = {x E S I Ii(x) :::; zf}
for i = 1, ... , k, and if we have nf=l Li(zf) = {zh}, then the vector zh is Pareto
optimal.
An entropy-based formulation of the €-constraint method is suggested in
Sultan and Templeman (1996). The entropy-based objective function to be op-
timized has only one parameter no matter what the number of the original
objective functions is. A representation of the Pareto optimal set can be gener-
ated by varying the value of the single parameter. The entropy-based function
contains logarithms and exponential functions.

3.2.6. Concluding Remarks

Theoretically, every Pareto optimal solution of any multiobjective optimiza-

tion problem can be found by the €-constraint method by altering the upper
bounds and the function to be minimized. It must be stressed that even duality
gaps in nonconvex problems (see, e.g., Section 2.10 of Part I and Chankong and
Haimes (I983b, pp. 135-136)) do not disturb the functioning of the €-constraint
method. However, computationally, the conditions set by Theorems 3.2.2, 3.2.3
and 3.2.4 are not always very practical. For example, according to Theorem
3.2.2, the E-constraint problem needs to be solved k times for all Ie as objec-
tive functions in order to generate one Pareto optimal solution. On the other
hand, the uniqueness of the solution demanded in the other theorems is not
always too easy to check either.
Computationally, the €-constraint method is more laborious than the weight-
ing method because the number of constraints increases. It may be difficult to
specify appropriate upper bounds for the objective functions. The components
of the ideal objective vector can be used to help in the specification. Then we
can set €j = zj + ej for j = 1, ... , k, j ::J e, where ej is some relatively small
positive real number that can be altered.
The €-constraint method can also be used as an a priori method, where the
decision maker specifies Ie and the upper bounds. Then it can be characterized
as an ad hoc method. It means that one can never be completely sure how
to select the objective function and the upper bounds to obtain a desirable
solution. This is a common weakness with the a priori weighting method.
96 Part II - 3. A Posteriori Methods

3.3. Hybrid Method

At this point it is worth mentioning a method combining the weighting

method and the C'-constraint method. This method is described in Corley
(1980) and Wendell and Lee (1977) in slightly different forms. The name hybrid
method is introduced in Chankong and Haimes (1983a, b).
The hybrid problem to be solved is

minimize
(3.3.1) i=l
subject to Ji(x) '5. C'j for all j = 1, ... , k,
x E 5,
where Wi> 0 for all i = 1, ... ,k.
Notice that problem (3.3.1) is equivalent to problem (2.10.1) in Part I if we
set Wi = 1 for every i = 1, ... , k. In Corley (1980), the problem is formulated in
a more general setting with a pointed convex ordering cone defining efficiency.
Optimality results were already handled in Section 2.10 of Part I. Nevertheless,
we write them down here as well.

Theorem 3.3.1. The solution of hybrid problem (3.3.1) is Pareto optimal for
any upper bound vector E E R k. On the other hand, if x* E 5 is Pareto optimal,
then it is a solution of problem (3.3.1) for E = f(x*).

Proof. See Corley (1980) or Wendell and Lee (1977).

The set of Pareto optimal solutions can be found by solving problem (3.3.1)
with methods for parametric constraints (where the parameter is the vector of
upper bounds E), see, for example, Rao (1984, pp. 418-421). This means that
the weighting coefficients do not have to be altered.
Optimality conditions for the solutions of problem (3.3.1) to be properly
Pareto optimal are presented in Wendell and Lee (1977).
We can say that the positive features of the weighting method and the C'-
constraint method are combined in the hybrid method. Namely, any Pareto
optimal solution can be found independently of the convexity of the problem
and one does not have to solve several problems or think about uniqueness
to guarantee the Pareto optimality of the solutions. On the other hand, the
specification of the parameter values may still be difficult. Computationally,
the hybrid method is similar to the C'-constraint method (with the increased
number of constraint functions).
 3.4. Method of Weighted Metrics 97

3.4. Method of Weighted Metrics

In the method of the global criterion, introduced in Section 2.1, Lp- and
Loo-metrics were used to generate (weakly) Pareto optimal solutions. These
metrics can also be weighted in order to produce different (weakly) Pareto
optimal solutions. The weighted approach is also sometimes called compromise
programming (see Zeleny (1973». Here we use the term the method of weighted
metrics.

3.4.1. Introduction

We assume that Wi ~ 0 for all i = 1, ... , k and L~=l Wi = 1. We obtain

different solutions by altering the weighting coefficients Wi in the weighted Lp-
and Tchebycheff metrics. The weighted Lp-problem for minimizing distances is
now of the form

minimize
(3.4.1)
subject to xES
for 1 :::; p < 00. The weighted Tchebycheff problem is of the form
minimize . max [wilh(x) - ztl)
(3 ..
4 2) I=l •...• k
subject to XES.
Problem (3.4.2) was originally introduced in Bowman (1976). Again, denomina-
tors may be included. Further, the absolute value signs can be dropped because
of the definition of the ideal objective vector, if it is known globally. Weighting
vectors can also be used in connection with problems of form (2.1.4).
Ifp = 1, the sum of weighted deviations is minimized and the problem to be
solved is equal to the weighting problem except for a constant (if z* is known
globally). If p = 2, we have a method of least squares. When p gets larger,
the minimization of the largest deviation becomes more and more important.
Finally, when p = 00, the only thing that matters is the weighted relative
deviation of one objective function.
Problem (3.4.2) is nondifferentiable like its unweighted counterpart. Corre-
spondingly, it can be solved in a differentiable form as long as the objective
and the constraint functions are differentiable and z* is known globally. In this
case, instead of problem (3.4.2), the problem
minimize a
(3.4.3) subject to a ~ Wi (ji(X) - z;) for all i = 1, ... , k,
XES,
is solved, where both x ERn and a E R are variables. This formulation will
be utilized later.
98 Part II - 3. A Posteriori Methods

3.4.2. Theoretical Results

In the following, we present some results concerning the weighted metrics.

Most of the proofs are so closely reminiscent of those presented earlier in Section
2.1 that there is no reason to repeat them. We assume that the global ideal
objective vector is known.

Theorem 3.4.1. The solution of weighted Lp-problem (3.4.1) (when 1 :::; p <
00) is Pareto optimal if either
(i) the solution is unique or
(ii) all the weighting coefficients are positive.

Proof. The proof is not presented here since it follows directly from the proofs
of Theorems 3.1.2, 3.1.3 and 2.1.1. See Chankong and Haimes (1983b, p. 144)
or Yu (1973).

Theorem 3.4.2. The solution of weighted Tchebycheff problem (3.4.2) is

weakly Pareto optimal if all the weighting coefficients are positive.

Proof. The proof is a straightforward modification of the proof of Theorem

2.1.2.

Theorem 3.4.3. Weighted Tchebycheff problem (3.4.2) has at least one Pareto
optimal solution.

Proof. The proof follows directly from the proof of Theorem 2.1.3.

Corollary 3.4.4. If weighted Tchebycheff problem (3.4.2) has a unique solu-

tion, it is Pareto optimal.

Convexity of the multiobjective optimization problem is needed in order to

guarantee that every Pareto optimal solution can be found by the weighted
Lp-problem (see Sawaragi et al. (1985, p. 81)). The following theorem shows
that, on the other hand, every Pareto optimal solution can be found by the
weighted Tchebycheff problem.

Theorem 3.4.5. Let x' E S be Pareto optimal. Then there exists a weighting
vector 0 < w E R k such that x' is a solution of weighted Tchebycheff problem
(3.4.2), where the reference point is the utopian objective vector z**.

Proof. Let x· E S be Pareto optimal. Let us assume that there does not exist
a weighting vector w > 0 such that x' is a solution of the weighted Tchebycheff
problem. We know that J;(x) > zt* for all i = 1, ... , k and for all xES. Now
we choose Wi = fJ/(fi(X') - zt*) for all i = 1, ... , k, where fJ > 0 is some
normalizing factor.
 3.4. Method of Weighted Metrics 99

If x* is not a solution of the weighted Tchebycheff problem, there exists

another point XO E S that is a solution of the weighted Tchebycheff problem,
meaning that

Thus wi(fi(XO) - zi*) < f3 for all i = 1, ... , k. This means that

h(X*~ _ zi* (h(xO) - z;*) < {3,

and after simplifying the expression we have

for all i = 1, ... ,k. Here we have a contradiction with the Pareto optimality of
x*, which completes the proof. 0

A theorem, parallel to Theorem 3.4.5, is proved in Kaliszewski (1995).

3.4.3. Comments

Theorem 3.4.5 above sounds quite promising for the weighted Tchebycheff
problem. Unfortunately, this is not the whole truth. In addition to the fact that
every Pareto optimal solution can be found, weakly Pareto optimal solutions
may also be included. Auxiliary calculation is needed in order to identify the
weak ones. Remember that as far as the weighted Lp-problem (1 ~ p < 00) is
concerned, it produces Pareto optimal solutions but does not necessarily find
all of them.
Selecting the value for the exponent p is treated in Ballestero (1997b) from
the point of view of risk aversion. The conclusion is that for greater risk aversion
we should use greater values for p. Another guideline is that for a smaller
number of objective functions we should select greater p values.
More results concerning the properties of the Lp-metrics (1 ~ P ~ 00) with
and without the weighting coefficients can be found, for example, in Bowman
(1976), Chankong and Haimes (1983b, pp. 144-146), Koski and Silvennoinen
(1987), Nakayama (1985a) and Yu (1973), the first of these treating especially
the Tchebycheff metric. Some results concerning the proper efficiency (in the
sense of Henig) of the solutions of the weighted Lp-problem are presented briefly
in Wierzbicki (1986b).
100 Part II - 3. A Posteriori Methods

3.4.4. Connections with Trade-Off Rates

Useful results concerning trade-off rates and the weighted Tchebycheff prob-
lem are proved in Yano and Sakawa (1987). The approach is closely related to
what was presented in Subsection 3.2.4 in connection with the e-constraint
problem.
Let us once again suppose that the feasible region is of the form

All the objective and the constraint functions are assumed to be twice con-
tinuously differentiable, which is why problem (3.4.3) is the one to be dealt
with.
Problem (3.4.3) is first formulated as an unconstrained minimization prob-
lem with one objective function, the Lagrange function, of the form
k m
(3.4.4) 0+ 2: Ai (Wi(!i(X) - zi) - 0) + 2: l'igi(X),
i=1 ;=1

where l E Rk and I' E R m are Karush-Kuhn-Tucker multipliers. The decision

variable vector being (0, x) ERn+!, let us denote the minimal solution of
function (3.4.4) by y* ERn+!.
It is assumed that the assumptions in Theorem 3.2.13 when applied to
problem (3.4.3) are satisfied. This means that y* is a regular point of the
constraints of (3.4.3), the second-order sufficient condition of Note 3.2.11 ap-
plied to problem (3.4.3) is satisfied at y* and all the active constraints at y*
are nondegenerate. (The last assumption means that the Karush-K uhn-Tucker
multipliers of all the active constraints are positive.)
If all the constraints connected to the objective functions are active, we
have
A" _ AjWj
lJ - AiWi'
Notice that the weighting coefficients have an essential role in these trade-off
rates, unlike those related to the e-constraint method. The procedure is not
treated here in more detail because of its resemblance to what was presented
in Subsection 3.2.4. For details, see Yano and Sakawa (1987).

3.4.5. Variants of the Weighted Tchebycheff Problem

Thus far, it has been proved that the weighted Tchebycheffproblem can find
any Pareto optimal solution. According to Corollary 3.4.4, the unique solution
of the weighted Tchebycheff problem is Pareto optimal. If the solution is not
unique or the uniqueness is difficult to guarantee, the weakness of the problem
is that it may produce weakly Pareto optimal solutions as well. This weakness
can be overcome in different ways. One possibility is to solve some additional
 3.4. Method of Weighted Metrics 101

problem, for example problem (2.10.1) or problem (2.10.2), given in Part I.

Problem (2.10.1) can also be modified so that instead ofthe sum of the objective
functions, the sum of objective functions minus utopian objective values is
minimized. Such an approach is handled in Section 5.4. It is also discussed, for
example, in Korhonen (1997). One more way is to use lexicographic ordering
(to be introduced in Section 4.2).
If additional optimizations must be avoided, another possibility is to vary
the metric. Weakly Pareto optimal solutions can be avoided by giving a slight
slope to the contour of the metric. The price to be paid is that in some cases
it may be impossible to find every Pareto optimal solution. For that reason,
properly Pareto optimal solutions are of interest here. Note that the utopian
objective vector is used as a reference point as in Theorem 3.4.5.
It is suggested in Steuer (1986) and Steuer and Choo (1983) that the
weighted Tchebycheff problem be varied by an augmentation term. In this
case, the distance between the utopian objective vector and the feasible ob-
jective region is measured by an augmented weighted Tchebycheff metric. The
augmented weighted Tchebycheff problem is of the form
k

(3.4.5)
minimize . max
,,==l, ... ,k
[wilfi(X) - zt*I] + p L
.
Ifi(X) - zt*1
• =1
subject to XES,
where p is a sufficiently small positive scalar.
A slightly different modified weighted Tchebycheff metric is used in the mod-
ified weighted Tchebycheff problem

t; Ifi(X) - zt*I)]
k

(3.4.6)
minimize i=r1f~k [Wi (11i(x) - zt*1 + p
subject to XES,

where p is a sufficiently small positive scalar. It is shown in Kaliszewski (1987)

that problem (3.4.6) is equivalent (up to scalar multiplication) to that proposed
in Choo and Atkins (1983).
The difference between the augmented and the modified weighted Tcheby-
cheff problems is in the way the slope takes place in the metrics, as illustrated
in Figure 3.4.1. In the augmented weighted Tchebycheff problem the slope is a
function of the weighting coefficients and the parameter p. In other words, the
slope may be different for each objective function, that is, for each coordinate
of the objective space. As far as the modified weighted Tchebycheff problem is
concerned, the slope is a function of the parameter p and, thus, constant for
all the objective functions. In Figure 3.4.1 we have

f3i = arctan p and f3 = arctan -p-.

1- Wi + P 1+P
To ease the comparison, the dotted lines represent the slope of the weighted
Tchebycheff metric. See Kaliszewski (1986, 1987) for details.
102 Part II - 3. A Posteriori Methods

Figure 3.4.1. Slopes of two metrics.

The augmented weighted Tchebycheff problem is illustrated in Figure 3.4.2,

where the dotted lines represent the contours of the augmented metric. The con-
tour of the weighted Tchebycheff metric (continuous line) has only been added
to ease the comparison. Sensitivity analysis of the augmented and the modified
weighted Tchebycheff metrics is handled in Kaliszewski (1994, pp. 113-119).

Figure 3.4.2. Augmented weighted Tchebycheff problem.

It is valid for both the augmented and the modified weighted Tchebycheff
problem that they generate only properly Pareto optimal solutions and any
properly Pareto optimal solution can be found. In what follows, the symbol M
is the scalar from Definition 2.9.1 of proper Pareto optimality in Part I.
 3.4. Method of Weighted Metrics 103

Theorem 3.4.6. A decision vector x· E S is properly Pareto optimal if and

only if there exist a weighting vector w E R k with Wi > 0 for all i = 1, ... , k and
a number p > 0 such that x· is a solution of augmented weighted Tchebycheff
problem (3.4.5).
In addition, for each properly Pareto optimal solution x· E S there exists
a weighting vector 0 < w E R k such that x· is a unique solution of problem
(3.4.5) for every p > 0 satisfying M :::; mini=l, ... ,k w;/((k - 1)p).
Further, the inequality
1
M :::; - . max Wi + (k - 1)p
p t=l, ... ,k
is valid for every solution x· of problem (3.4.5).

Proof. See Kaliszewski (1994, pp. 51-53).

The proof in Kaliszewski (1994) is based on a cone separation technique.

The necessity and the sufficiency components are also proved in Kaliszewski
(1985).
The theorem for the modified weighted Tchebycheff problem is almost sim-
ilar.

Theorem 3.4.7. A decision vector x· E S is properly Pareto optimal if and

only if there exist a weighting vector w E R k with Wi > 0 for all i = 1, ... , k
and a number p > 0 such that x· is an optimal solution of modified weighted
Tchebycheff problem (3.4.6).
In addition, for each properly Pareto optimal solution x· E S there exists
a weighting vector 0 < w E Rk such that x· is a unique solution of problem
(3.4.6) for every p > 0 satisfying M :::; l/((k - 1)p).
Further, the inequality M:::; (1 + (k - 1)p)/p is valid for every solution x·
of problem (3.4.6).

Proof. See Kaliszewski (1994, pp. 48-50).

This necessity and sufficiency formulation is an extension of the original

theorem in Choo and Atkins (1983). The necessary conditions in Theorems
3.4.6 and 3.4.7 are also proved in Kaliszewski (1995).

3.4.6. Connections with Global Trade-Offs

Some metrics for measuring the distance between the utopian objective
vector and the feasible objective region can be formed in such a way that they
produce solutions with selectively bounded global trade-offs. This is in reverse
to the general way where trade-offs are calculated only after solutions have
been generated.
104 Part II - 3. A Posteriori Methods

For the simplicity of notations we here assume the global ideal objective
vector and, thus, the global utopian objective vector to be known. This implies
that we can drop the absolute value signs.
All the properly Pareto optimal solutions produced with modified weighted
Tchebycheff problem (3.4.6) have bounded global trade-offs. Further, we have
a common bound for every global trade-off involved.

Theorem 3.4.8. Let x· E S be a solution of modified weighted Tchebycheff

problem (3.4.6) for some weighting vector 0 < W E Rk and p > O. In this case,
the global trade-offs are bounded, that is

A~(x.) < 1 + P
'J - P
for every i, j = 1, ... , k, i f:. j.
Proof. See Kaliszewski (1994, pp. 94-95).

Corresponding results can be proved for other types of problems, see Kaliszewski
(1994, pp. 82-113).
Sometimes the decision maker may wish to set a priori bounds on some
specific global trade-offs. Such a request calls for a scalarizing function of a
special form. These topics are treated in Kaliszewski and Michalowski (1995,
1997). Thus far, the additional term multiplied with p was added to guarantee
the proper Pareto optimality of the solutions. If we leave it out, we obtain
weighted Tchebycheff problem (3.4.2) and, thus, weakly Pareto optimal solu-
tions. In what follows, we use metrics without modification or augmentation
terms but use other parameters (7 and O'i > 0 to control the bounds of the
global trade-offs involved. Thus, the following results deal with weak Pareto
optimality.
The next theorem handles a case where we wish to set a priori bounds for
a group of selected global trade-offs. Let us choose a subset of the objective
functions 10 C I = {I, ... ,k} and define I(i) = {j I j E 10, j f:. i}.

Theorem 3.4.9. A decision vector x· E S is weakly Pareto optimal and

(3.4.7) A~(x*) ~ 1 + 0' for all i E 10 and all j E I(i)

0'

if and only if there exist a weighting vector 0 <W E Rk and a number 0' >0
such that x· is a solution of the problem

minimize max [IJi:fC [Wi (1 + O')(/i(x) - z:*) + 0' L (iJ(x) - zj*))] ,

o iE/(i)

max [Wi(/i(X) -
iE/\Io
z:*)]]
subject to xES.
 3.4. Method of Weighted Metrics 105

Proof. See Kaliszewski and Michalowski (1995).

Result (3.4.7) of Theorem 3.4.9 is utilized so that the decision maker is asked
to specify upper bounds for the selected global trade-offs. These values are set as
upper bounds to (1+0')/0'. A lower bound for the parameter 0' is obtained from
these inequalities. By using the calculated 0' value, we receive different weakly
Pareto optimal solutions satisfying the global trade-off bounds by altering the
weighting coefficients. In other words, we avoid generating solutions exceeding
the specified bounds for global trade-off.
In Theorem 3.4.9 we have a common bound for the selected set of global
trade-offs. This can further be generalized by using several different parameters
0'.

Theorem 3.4.10. A decision vector x* E S is weakly Pareto optimal and for

each i E 10 and each j E 10, A~(x*) is bounded from above by a positive finite
number if and only if there exist a weighting vector 0 < w E R k and numbers
O'j > 0 for j E 10 , such that x* is a solution of the problem

minimize max [ n~; [Wi (Ii (x) - zt* + L O'j (iJ (x) - zj*) ) ] ,
jE/o
(3.4.8)
max (wi(li(x) - zt*)]]
iE/\Io
subject to xES.

Further, for any solution x* of problem (3.4.8), we have

for i, j E 10 and j i:- i.

Proof. See Kaliszewski and Michalowski (1997).

Theorem 3.4.10 is applied in the following way. Ifwe want to generate weakly
Pareto optimal solutions such that certain global trade-offs are bounded (each
global trade-off with an individual bound), we form a system of equations from
the global trade-off information. That is, we set (1 +O'i) /O'j equal to the specified
upper bound, where desired. If the system is consistent, we solve it and obtain
values for the parameters O'j. If the system is inconsistent, some equation(s)
must be dropped in order to form a consistent system. In this way, parameters
O'j are used to control the bounds of the selected global trade-offs.

Let us return to proper Pareto optimality. Theorem 3.4.8 can be modified

to handle individual global trade-offs. This means that we solve a problem
106 Part II - 3. A Posteriori Methods

(3.4.9)
minimize i=IIf,a.X,k [Wi (fi(X) - Zt* + LPi(fi(X) - Zt*))]
t=l
subject to xES.

It is proved in Kaliszewski and Michalowski (1997) that Theorem 3.4.7

is valid with Pi > 0, i = 1, ... , k. In other words, a decision vector x' E S is
properly Pareto optimal if and only if there exist a weighting vector 0 < W E R k
and numbers Pi > 0 for all i = 1, ... , k such that x' is a solution of problem
(3.4.9). Further, for the solution x· of problem (3.4.9) the upper bounds of the
global trade-offs are of the form

A(](x*) < 1 + Pi
J' - Pj

for i, j = 1, ... , k and i # j.

3.4.7. Applications and Extensions

A shape optimization problem of a spillway profile is solved by the weighted

L 2 -metric (with denominators) in Wang and Zhou (1990). Further, an extension
of the method of weighted metrics called composite programming is presented in
Bardossy et al. (1985). The Lp-metric (p < (0) is divided into nested parts with
different exponents. In particular, this approach can be applied to problems
where objective functions consist of several components. The method is applied
to problems of multiobjective watershed management and observation network
design.
The weighted Tchebycheff metric is used in Kostreva et al. (1995) in de-
riving an integral approach for solving nonlinear problems with discontinuous
objective functions and a discontinuous feasible region. The close connections
between the weighted Tchebycheff metric and the weighting method for con-
vex problems are handled in Dauer and Osman (1985). Karush-Kuhn-Tucker
optimality conditions for these two methods are treated as well.
One more possibility of avoiding weakly Pareto optimal solutions is sug-
gested in Helbig (1991). Helbig defines optimality with ordering cones. The
idea is to perturb the ordering cone so that even though the method produces
weakly efficient solutions, they are efficient to the original problem.
One way to apply the weighted Tchebycheff metric and its augmented ver-
sion successfully will be introduced in Section 5.4.

3.4.8. Concluding Remarks

Particularly the method of weighted Tchebycheff metric and its variants are
popular methods for generating Pareto optimal solutions. They work for convex
as well as non convex problems (unlike the weighting method) and alteration of
the parameters is easier than in the €-constraint method.
 3.5. Achievement Scalarizing Function Approach 107

As far as the weighted Lp-metrics are concerned, the solutions produced

are Pareto optimal but not necessarily all of them are found (depending on the
degree of nonconvexity of the problem). The weighted Tchebycheff metric can
generate any Pareto optimal solution to any type of a problem. The drawback
of also generating weakly Pareto optimal solutions can be overcome by aug-
menting or modifying the metric or by solving another optimization problem
after minimizing the distance. These alternatives will be further addressed in
Section 5.4. For nonconvex problems the success of the method of weighted
metrics depends on whether the global ideal objective vector is known or not.
An interesting feature related to some variants of the weighted Tchebycheff
metric is the ability to produce solutions with a priori-specified bounds for
global trade-offs. In this way, a subset of weakly or properly Pareto optimal
solutions can be generated satisfying given fixed bounds.

3.5. Achievement Scalarizing Function Approach

The approach to be presented is related to that of weighted metrics. Namely,

we handle special types of scalarizing functions, termed achievement scalarizing
functions. They have been introduced by Wierzbicki, for example, in Wierzbicki
(1981, 1982, 1986a, b) (and are also handled in Wierzbicki (1977, 1980a, b)).
Somewhat similar results for scalarizing functions are also presented in Jahn
(1984) and Luc (1986).

3.5.1. Introduction

In the method of weighted Lp-metric or the weighted Tchebycheff metric,

the distance is minimized between the ideal objective vector and the feasible
objective region. If the global ideal objective vector is unknown, we may fail
in producing (weakly) Pareto optimal solutions. In other words, if the refer-
ence point used is an objective vector inside the feasible objective region, the
minimal distance to it is zero and we do not obtain a (weakly) Pareto optimal
solution. We can overcome this weakness by replacing metrics with achievement
scalarizing functions.
For example, weakly Pareto optimal solutions can be generated with any
reference point Z E R k by solving the problem

minimize . max [Wi(fi(X) - Zi) 1

t=l, ... ,k
(3.5.1)
subject to xES.

It differs from weighted Tchebycheff problem (3.4.2) only in that the abso-
lute value signs are missing. This change ensures that weakly Pareto optimal
solutions are produced independently of the feasibility or infeasibility of the
reference point.
108 Part II - 3. A Posteriori Methods

Scalarizing functions of a special type are called achievement scalarizing

functions. Problem (3.5.1) is one example of them. Let us next handle achieve-
ment scalarizing functions in general.
It is shown, for instance, in Wierzbicki (1980b, 1986a, b) that Pareto opti-
mal solutions can be characterized by achievement scalarizing functions if the
functions satisfy certain requirements. An achievement scalarizing function is a
function Sz; Z -t R, where z E R k is an arbitrary reference point of aspiration
levels (see Definition 2.3.1 in Part I). In the following, we shorten the name to
an achievement function.
Because we do not know the feasible objective region Z explicitly, in practice
we minimize the function sz(f(x)) subject to xES (see, e.g., Figure 2.2.1 in
Part I). Thus, we deal with the feasible region in the decision variable space.
For notational convenience, we, however, present the problem here as if it were
solved in the feasible objective region.

3.5.2. Theoretical Results

We need to apply some of the general properties introduced in Part I to

an achievement function sz, namely the definitions of strictly increasing (Def-
inition 2.1.8), strongly increasing (Definition 2.1.9) and c-strongly increasing
(Definition 2.1.10) functions. In the last-mentioned concept the definition of
the set R: is the same as in connection with c-proper Pareto optimality (see
Definition 2.9.2 in Part I).
Next we can define order-representing and order-approximating achieve-
ment functions.

Definition 3.5.1. A continuous achievement function Sz: Z -t R is order-

representing if it is strictly increasing as a function of z E Z for any z E R k
and if
{z E Rk I sz(z) < O} = z - intR~
(for all Z E Rk).

Definition 3.5.2. A continuous achievement function Sj: Z -t R is order-

approximating if it is strongly increasing as a function of z E Z for any z E R k
and if
z - R~ C {z E R k I Sj (z) ::; O} C Z - R~
(for all z E Rk) with c > €::::: O.
Remark 3.5.3. For a continuous order-representing or order-approximating
achievement function Sz; Z -t R we have

Sz(z) = O.
 3.5. Achievement Scalarizing Function Approach 109

We can now present some optimality results concerning different types of

achievement functions according to Wierzbicki (1986a, b). The achievement
problem to be solved is
minimize Sz (z )
(3.5.2)
subject to z E Z.

Theorem 3.5.4. If the achievement function Sz: Z ~ R is strictly increasing,

then the solution of achievement problem (3.5.2) is weakly Pareto optimal. If
the achievement function Sz: Z ~ R is strongly increasing, then the solution of
problem (3.5.2) is Pareto optimal. Finally, if the achievement function Sz: Z ~
R is c:-strongly increasing, then the solution of problem (3.5.2) is c:-properly
Pareto optimal.

Proof. Here we only prove the second statement because of the similarity of
the proofs. We assume that Sz is strongly increasing. Let z* E Z be a solution
of the achievement problem. Let us suppose that it is not Pareto optimal. In
this case, there exists an objective vector z E Z such that Zi ~ zi for all
i = 1, ... , k and Zj < zj for some j. Because Sz is strongly increasing, we know
that sz(z) < sz(z*), which contradicts the assumption that z* minimizes Sz.
Thus z* is Pareto optimal. 0

The results of Theorem 3.5.4 can be augmented by the following theorem.

Theorem 3.5.5. If the achievement function Sz: Z ~ R is increasing and the

solution of achievement problem (3.5.2) is unique, then it is Pareto optimal.

Proof. The proof corresponds to the proof of Theorem 3.5.4.

Note that Theorems 3.5.4 and 3.5.5 are valid for any scalarizing function.
Thus, the Pareto optimality and the weak Pareto optimality results proved for
the weighting method, the c;-constraint method and the method of weighted
metrics are explained by the 'monotonicity properties of the scalarizing func-
tions in question (see, e.g., Vanderpooten (1990».
We can now rewrite Theorem 3.5.4 so as to be able to characterize Pareto
optimal solutions with the help of order-representing and order-approximating
achievement functions. The proof follows from the proof of Theorem 3.5.4.

Corollary 3.5.6. If the achievement function Sz: Z ~ R is order-representing,

then, for any z E R k, the solution of achievement problem (3.5.2) is weakly
Pareto optimal. If the achievement function Sz: Z ~ R is order-approximating
with some c: and t as in Definition 3.5.2, then, for any z E Rk, the solution
of problem (3.5.2) is Pareto optimal. If Sz in addition is t-strongly increasing,
then the solution of problem (3.5.2) is t-properly Pareto optimal.
110 Part II - 3. A Posteriori Methods

The preceding corollary states the sufficient conditions for a solution of an

achievement function to be weakly, e:-properly, or Pareto optimal. The following
theorem gives the corresponding necessary conditions.

Theorem 3.5.7. If the achievement function Sz: Z -t R is order-representing

and z' E Z is weakly Pareto optimal or Pareto optimal, then it is a solution
of achievement problem (3.5.2) with z = z* and the value of the achievement
function is zero. If the achievement function Sz: Z -t R is order-approximating
and z* E Z is e:-properly Pareto optimal, then it is a solution of problem (3.5.2)
with z = z* and the value of the achievement function is zero.

Proof. Here, we only prove the statement for Pareto optimality. The proofs
of the other statements are very similar. (The proof of the necessary condition
for e:-proper Pareto optimality can be found in Wierzbicki (1986a).)
Let z· E Z be Pareto optimal. This means that there does not exist any
other point z E Z such that Zi ~ z; z;
for all i = 1, ... , k and Zj < for some
j. Let us assume that z* is not a solution of the achievement problem when
z = z*. In this case there exists some vector ZO E Z such that sz(ZO) < sz(z*) =
sz(z) = 0 and ZO =f. Z·. Since Sz was assumed to be order-representing, we have
ZO E z - int Rt = z* - int Rt. This means that zi < z; for all i = 1, ... , k,
which contradicts the assumption that z* is Pareto optimal. Thus, z* is a
solution of the achievement problem. D

Remark 3.5.8. Aided by the results in Theorem 3.5.7 a certain point can be
confirmed not to be weakly, e:-properly or Pareto optimal (if the optimal value
of the achievement function differs from zero).

We are now able to completely characterize the set of weakly Pareto optimal
solutions with the help of order-representing achievement functions. The sets
of Pareto optimal and €-properly Pareto optimal solutions are characterized al-
most completely (if the closure of the sets of solutions of achievement problem
(3.5.2) for an order-approximating achievement function is taken as f -t 0).
If the solutions of achievement problem (3.5.2) are assumed to be unique, the
theorems above render the characterization of Pareto optimal solutions com-
plete.

3.5.3. Comments

If the reference point is feasible, or to be more exact z E Z + Rt, then

the minimization of the achievement function sz(z) subject to z E Z must
produce a solution that maximizes the distance to the Pareto optimal set. If
the reference point is infeasible, that is, z rt. Z + Rt, then the minimization of
the achievement function sz(z) subject to z E Z must produce a solution that
minimizes the distance to the Pareto optimal set.
 3.5. Achievement Scalarizing Function Approach 111

The advantage of the achievement functions is that any arbitrary weakly

Pareto optimal or Pareto optimal (or at least e-properly Pareto optimal) so-
lution can be obtained by moving the reference point only. It is shown in
Wierzbicki (1986a) that the solution of the achievement function depends
Lipschitz-continuously on the reference point.
There are many achievement functions satisfying the above-presented condi-
tions. An example of order-representing functions was given in problem (3.5.1).
The corresponding achievement function is
sz(z) = . max [Wi(Zi - Zi)],
,=l, ... ,k

where w is some fixed positive weighting vector. Let us briefly convince our-
selves that the above-mentioned function really is order-representing. The con-
tinuity of the function is obvious. If we have ZI and Z2 E Z such that zt < z;
for all i = 1, ... , k, then sz(ZI) = maxdwi(zt - Zi)J < maxi[Wi(Z; - zdJ =
sz(Z2) and thus the function is strictly increasing. If the inequality sz(z) =
maxi [Wi (Zi - Zi)] < 0 holds, then we must have Zi < Zi for all i = 1, ... , k, that
is, z E z - intR~.
An example of order-approximating achievement functions is
k
(3.5.3) sz(z) = t=l,
.max [Wi(Zi - zd 1+ P L Wi(Zi - Zi),
... ,k .
1=1

where w is some fixed positive weighting vector and p > 0 is sufficiently small
when compared with c and large when compared with e. The weighting coeffi-
cients can also be dropped from the latter part. This function is also e-strongly
increasing. FUnction (3.5.3) is related to augmented weighted Tchebycheff prob-
lem (3.4.5) and, thus, it can be called an augmented weighted achievement
function.
An example of a so-called penalty scalarizing function is

where fl > 1 is a scalar penalty coefficient and (z - z)+ is a vector with com-
ponents max [0, Zi - Zi]. This function is strictly increasing, strongly increas-
ing for all the metrics in Rk except for the Tchebycheff metric and order-
approximating with e ~ 1/ (} (see Wierzbicki (1980a, 1982». More examples of
order-representing and order-approximating functions are presented, for exam-
ple, in Wierzbicki (1980b, 1986a, b).
In cases when there exists a weighting vector such that the solution of
weighting problem (3.1.1) is equal to the solution of the achievement problem,
the weighting vector can be obtained from partial derivative information of the
achievement function. See Wierzbicki (1982) for details.
Let us finally mention a subset of reference points, termed dominating
points, considered in Skulimowski (1989). A point is called a dominating point
112 Part II - 3. A Posteriori Methods

if it is not dominated by any feasible point and it dominates at least one of the
feasible points.

3.5.4. Concluding Remarks

Achievement scalarizing functions are a set of general scalarizing functions

satisfying certain requirements. In general, achievement functions are concep-
tually very appealing for generating weakly, E-properly or Pareto optimal so-
lutions. They overcome most of the difficulties arising with other methods in
this class.
The results concerning achievement functions will be utilized, for example,
in Section 5.6 when deriving an interactive method. For interactive methods,
the idea of moving the reference point instead of the weighting coefficient seems
more natural and easier for the decision maker. A fact favouring achievement
scalarizing functions against weighted metrics is that the global ideal objective
vector does not have to be known. Thus, the method is more reliable.

3.6. Other A Posteriori Methods

Finally, we briefly mention some other methods of the a posteriori type. For
more detailed information, see the references cited.
The so-called hyperplane method is introduced in Yano and Sakawa (1989)
for generating Pareto optimal or properly Pareto optimal solutions. It is
shown that the weighting method, the E-constraint method and the method
of weighted metrics can be viewed as special cases of the hyperplane method.
A theory concerning trade-off rates in the hyperplane method is provided in
Sakawa and Yano (1990). A generalized hyperplane method for generating all
the efficient solutions (with respect to some ordering cone) is presented in
Sakawa and Yano (1992).
Another method for a general characterization of the Pareto optimal set is
suggested in Soland (1979). For example, the weighting method, the method of
weighted metrics and goal programming (see Section 4.3) can be seen as special
cases of the general scalar problem of Soland. Further, the weighting method
and the E-constraint method are utilized in a so-called envelope approach for
determining Pareto optimal solutions in Li and Haimes (1987). An application
to dynamic multiobjective programming is also treated.
The non inferior (meaning here Pareto optimality) set estimation (NISE)
method for MOLP problems can also be considered to belong to this class
of a posteriori methods. It is a technique for generating the Pareto optimal
set of two objective functions (see Cohon (1978)). It can be generalized for
convex problems with two objective functions (see, for example, Chankong and
Haimes (1983b, pp. 268-274)). In Balachandran and Cero (1985), the method
is extended to problems with three objective functions. The weighting method
is the basis of the NISE method.
 3.6. Other A Posteriori Methods 113

Multiobjective optimization problems with polynomial objective and con-

straint functions are treated in Kostreva et al. (1992). The method for deter-
mining Pareto optimal solutions is based on problem (2.10.2) of Part I and a
so-called homotopy continuation. Note that problems with polynomial func-
tions are highly nonlinear, non convex and nonconcave.
A scalarization method for multiobjective optimization problems, where
optimality is defined through ordering cones, is suggested in Pas coletti and
Serafini (1984). By varying the parameters of the scalar problem it is possible
to find all the efficient solutions. A further investigation is conducted in Sterna-
Karwat (1987).
It is suggested in Benson and Sayin (1997) that instead of trying to generate
the whole Pareto optimal set one should aim at finding a truly global repre-
sentation of it. This would decrease both the burden of the decision maker and
computational costs. Benson and Sayin introduce a global shooting procedure
to meet this need. In Armann (1989), a method is presented for generating
a dispersed subset of the Pareto optimal set, which is then presented to the
decision maker.
One more method for generating an evenly distributed set of Pareto opti-
mal solutions to a differentiable nonlinear multiobjective optimization problem
is suggested in Das and Dennis (1998). The approach is called the normal
boundary intersection (NBI) method. The idea in broad outline is to intersect
the feasible objective region with a normal to the convex combinations of the
columns of the payoff matrix. Evenly distributed parameters, that is, the coef-
ficients in the convex combinations, produce evenly distributed solutions. The
weakness of the approach is the fact that it may produce non-Pareto optimal
solutions ~o nonconvex problems.
The difficulty of illustrating the set of Pareto optimal solutions to the deci-
sion maker is treated in Bushenkov et al. (1994, 1995) and Lotov et al. (1992,
1997). An extension of the Pareto optimal set, a so-called Pareto optimal hull,
is approximated by polyhedral sets (see Lotov (1995, 1996» using convolution-
based algorithms (see Bushenkov et al. (1995) and Lotov (1996». Different
decision maps are generated in this way (see Chapter 3 of Part III). Specific
approaches exist for linear, convex and non convex cases but they are all based
on the same so-called generalized reachable sets method. An implementation
of the generalized reachable sets method is available (see Section 2.2 in Part
III).
 4. A PRIORI METHODS

In the case of a priori methods, the decision maker must specify her or
his preferences, hopes and opinions before the solution process. The difficulty
is that the decision maker does not necessarily know beforehand what it is
possible to attain in the problem and how realistic her or his expectations are.
The working order in these methods is: 1) decision maker, 2) analyst.
Below, we handle three a priori methods. First, we give a short presentation
of the value function method. Then we introduce lexicographic ordering and
goal programming.

4.1. Value Function Method

The value function optimization approach was already mentioned earlier.

Here we present it again briefly.

4.1.1. Introduction

In the value function method, the decision maker must be able to give an
accurate and explicit mathematical form of the value function U: R k -+- R that
represents her or his preferences globally. This function provides a complete
ordering in the objective space. Then the value function problem
maximize U(f(x))
(4.1.1)
subject to xES
is ready to be solved by some method for single objective optimization as
illustrated in Figure 4.1.1. The bold line represents the Pareto optimal set.
Remember Theorem 2.6.2 of Part I, which says that the solution of problem
(4.1.1) is Pareto optimal if the value function is strongly decreasing.
The value function method seems to be a very simple method, but the diffi-
culty lies in specifying the mathematical expression of the value function. The
inability to encode the decision maker's underlying value function reliably is
demonstrated by experiments in de Neufville and McCord (1984). It is shown
that encoding methods that should theoretically produce identical value func-
tions fail: the functions may differ from each other by more than 50 %. It is also
K. Miettinen, Nonlinear Multiobjective Optimization
© Springer Science+Business Media New York 1998
116 Part II - 4. A Priori Methods

" ,
\

z1
contours of U

Figure 4.1.1. Contours of the value function.

pointed out that there is no actual analysis of the accuracy of the value function
assessment. The consistency checks, that is, whether decision makers provide
consistent answers to similar questions, are not adequate: a biased instrument
can provide consistent data.
On the other hand, even if it were possible for the decision maker to express
her or his preferences globally, the resulting preference structure might be too
simple, since value functions cannot represent intransitivity or incomparability
(see Rosinger (1985». More features and weaknesses were presented in connec-
tion with the definition of the value function (Definition 2.6.1) in Section 2.6
of Part I.

4.1.2. Comments

The value function method could be called an 'optimal' way of solving mul-
tiobjective optimization problems if the decision maker could reliably express
the value function. The use of the value function method is restricted in prac-
tice to multiattribute decision analysis problems with a discrete set of feasible
alternatives. The theory of value and utility functions for multiattribute prob-
lems is examined broadly in Keeney and Raiffa (1976). But, it is believed, for
example, in Rosenthal (1985), that these experiences can also be utilized in
continuous cases.
Important results concerning value functions and the conditions for their ex-
istence are collected in Dyer and Sarin (1981). Two general classes of value func-
tions, additive and multiplicative forms, are presented extensively in Keeney
and Raiffa (1976) and briefly in Rosenthal (1985). The existence of value func-
tions and the nature of additive decreasing value functions are handled in Starn
et al. (1985). These topics and the construction of value functions are pre-
sented more widely in Yu (1985, pp. 95-161). General properties and some
desirable features of certain types of value functions (e.g., additive, max-min,
 4.1. Value Function Method 117

min-sum and exponential forms) are stated in Bell (1986), Harrison and Rosen-
thal (1988), Soland (1979) and Sounderpandian (1991). More examples of value
functions are given in Tell and Wallenius (1979). Utility compatible measures
of risk are deduced in Bell (1995). Relations between value functions, ordering
cones and (proper) efficiency are studied in Henig (1990).
In some interactive methods, it is assumed that the underlying value func-
tion is of some particular (e.g., additive or exponential) form, after which, its
parameters are fitted according to the decision maker's preferences. Such meth-
ods are presented, for example, in Rothermel and Schilling (1986) and Sakawa
and Seo (1980, 1982a, b) (see Section 5.3).
Three kinds of conditions for value functions under which it is not possi-
ble to exclude any Pareto optimal or properly Pareto optimal solution from
consideration a priori are identified in Soland (1979).
The convergence properties of additive value functions (assuming prefer-
ential independence of the objective functions) are investigated by simulation
experiments in Stewart (1997). One observation is that piecewise linear value
functions perform dramatically better than linear ones.
Relationships between the method of weighted metrics and the value func-
tion method are reported in Ballestero and Romero (1991). It might be imag-
ined that the two methods have nothing in common, since a value function
represents the opinions of the decision maker and the method of weighted met-
rics does not take the decision maker into consideration. However, conditions
can be set on the value function to guarantee that its optimum belongs to the
solution set obtainable by the method of weighted metrics. More relationships
between these two methods, when the value functions are of a certain type,
are presented in Ballestero (1997a). It is demonstrated in Moron et al. (1996)
that there are large families of such well-behaved value functions for bi-criteria
problems where the connection is valid.

4.1.3. Concluding Remarks

The value function method is an excellent method if the decision maker

happens to know an explicit mathematical formulation for the value function
and if that function represents wholly the preferences of the decision maker.
These two crucial preconditions are the difficulties of the approach.
There are certain conditions that the decision maker's preferences must
satisfy so that a value function can be defined on them. The decision maker
must, for instance, be able to specify consistent (implying transitive) prefer-
ences. Thus, there may not necessarily exist a value function that will impose
a total order in the set of feasible objective vectors. The assumption of a total
order is often contrary to our intuitive aims and hence is quite likely to lead to
less than ideal selections, as Polak and Payne (1976) remind us. This fact must
be kept in mind below, when several methods which assume the existence of a
value function (at least implicitly) are introduced.
118 Part II ~ 4. A Priori Methods

One important thing to take into account in practice is that the aspirations
of the decision maker may change during the solution process. Possible expla-
nations of such behaviour are pondered in Steuer and Gardiner (1990). Is it
possible that the decision maker's value function will change considerably over
a short period of time and thus be unstable? Another alternative is that it is
difficult for the decision maker to know the real value function without getting
to know the problem better, that is, without interaction with the solution pro-
cess. More open questions concerning value functions are listed in Nijkamp et
a!. (1988).
The weighting method may be regarded as a special case of a value function
where the utilities are linear and additive. If the underlying value function is
assumed to be linear, this means that the marginal rates of substitution of the
decision maker are constant for every solution. See comments on this feature
in Section 4.3.

4.2. Lexicographic Ordering

Lexicographic ordering was mentioned earlier as a tool for producing Pareto

optimal solutions from weakly Pareto optimal ones. It can also be used as an
a priori solution method.

4.2.1. Introduction

In lexicographic ordering the decision maker must arrange the objective

functions according to their absolute importance. This ordering means that a
more important objective is infinitely more important than a less important
objective. After ordering, the most important objective function is minimized
subject to the original constraints. If this problem has a unique solution, it is
the solution of the whole multiobjective optimization problem. Otherwise, the
second most important objective function is minimized. Now, in addition to
the original constraints, a new constraint is added. This new constraint is there
to guarantee that the most important objective function preserves its optimal
value. If this problem has a unique solution, it is the solution of the original
problem. Otherwise, the process goes on as above. Lexicographic orders and
utilities are widely examined in Fishburn (1974).
An example of lexicographic ordering is presented in Figure 4.2.1. There
are two objective functions of which the first is the most important. After
minimizing the first objective, t.here are two points left and after minimizing
the second objective, the point Zl is obtained as the final solution. The bold
line represents the Pareto optimal set in the figure. This example is somewhat
too positive since all the objective functions have their effect on the solution
process.
 4.2. Lexicographic Ordering 119

Figure 4.2.1. Lexicographic ordering.

Let the objective functions be arranged according to the lexicographic order

from the most important h to the least important /k' We write the lexico-
graphic problem as
lex minimize h (x), hex), ... ,fk(X)
(4.2.1)
subject to XES.

We can now present the following result concerning the Pareto optimality of
the solutions.

Theorem 4.2.1. The solution of lexicographic problem (4.2.1) is Pareto op-

timal.

Proof. Let x· E S be a solution of the lexicographic problem. Let us assume

that it is not Pareto optimal. In this case, there exists a point xES such that
hex) $ h(x*) for all i = 1, ... ,k and for at least one j the inequality is strict,
that is, hex) < h(x·).
Let be i = 1. From the definition of lexicographic ordering we know that h
attains its minimum at x*. Since also h (x) $ h (x*), it is only possible that
hex) = h(x").
There are two possibilities in determining the lexicographic optimum. Either
a unique solution is found during the optimization process, or optimizations are
performed for every i = 1, ... , k. In the latter case, where i = 2, we also have
hex) = h(x·) and with similar reasoning we have that fi(X) = h(x·) for every
i = 1, ... , k. This contradicts the assumption of at least one strict inequality.
Thus, x· is Pareto optimal.
On the other hand, if lexicographic ordering stops before every objective
function has been examined, this means that a unique solution x· has been
120 Part II - 4. A Priori Methods

obtained for Ii. The assumption h(x) ~ h(x*) implies that fi(X) = h(x*),
which is a contradiction. Thus, x* is Pareto optimal. D

4.2.2. Comments

Numerical application examples of the method are given in Hwang and

Masud (1979, pp. 49-55). In Ben-Tal (1980), Pareto and lexicographic optima
are characterized in convex problems. Duality theory for convex problems with
the help of lexicographic ordering is developed in Martinez-Legaz (1988).
A modification of lexicographic ordering, called hierarchical optimization,
is applied to a vehicle design problem of mechanical engineering in Bestle and
Eberhard (1997). In hierarchical optimization the upper bounds obtained when
minimizing more important objective functions are relaxed by so-called wors-
ening factors. These factors are specified by the decision maker.
Lexicographic ordering corresponds to the weighting method when the
weighting coefficients are of very different magnitude. The question whether
there exist weighting vectors such that the optimal solution of the weighting
method is identical to the solution obtained by lexicographic ordering is con-
sidered in Sherali (1982) and Sherali and Soyster (1983). The answer is positive
for linear problems and several discrete problems. In practice, this means that
the problem of lexicographic ordering can be solved as a weighting problem
with standard optimizers.
The notion absolute importance of objective functions is discussed in Roy
and Mousseau (1996). Roy and Mousseau also consider under what kind of
conditions one can say that one objective function is more important than
another.

4.2.3. Concluding Remarks

The justification for using lexicographic ordering is its simplicity and the
fact that people usually make decisions successively. However, this method has
several drawbacks. The decision maker may have difficulties in putting the
objective functions into an absolute order of importance. On the other hand,
the method is usually robust. It is very likely that the less important objective
functions are not taken into consideration at all. Ifthe most important objective
function has a unique solution, the other objectives do not have any influence
on the solution. And even if the most important objective had alternative
optima and it was possible to use the second most important objective, it is
very unlikely that this problem would have alternative optima, and the third
or other less important objectives could be used.
Note that lexicographic ordering does not allow a small increment of an
important objective function to be traded off with a great decrement of a less
important objective function. Yet, this kind of trading might often be appealing
to the decision maker.
 4.3. Goal Programming 121

Lexicographic ordering may be used as a part of the following solution

method, called goal programming.

4.3. Goal Programming

The ideas of goal programming were originally introduced in Charnes et

al. (1955), but the term goal programming was fixed in Charnes and Cooper
(1961). It is one of the first methods expressly created for multiobjective op-
timization. Among more recent papers, an easy-to-understand presentation of
goal programming is given in Ignizio (1983a, 1985). Goal programming was
originally developed for MOLP problems, and this background is very evident
in the formulation.

4.3.1. Introduction

The basic idea in goal programming is that the decision maker specifies
(optimistic) aspiration levels for the objective functions and any deviations
from these aspiration levels are minimized. An objective function jointly with
an aspiration level forms a goal. We can say that, for example, minimizing the
price of a product is an objective function, but if we want the price to be less
than 500 dollars, it is a goal (and if the price must be less than 500 dollars, it
is a constraint). We denote the aspiration level of the objective function Ii by
Zi for i = 1, ... ,k.
For minimization problems, goals are of the form h(x) ::; Zi (and of the
form h(x) ~ Zi for maximization problems). Goals may also be represented
as equalities or ranges (for the latter, see Charnes and Cooper (1977)). The
aspiration levels are assumed to be selected so that they are not achievable
simultaneously.
It is worth noticing that the goals are of the same form as the constraints
of the problem. This is why the constraints may be regarded as a subset of the
goals. This way of formulating the problem is called generalized goal program-
ming. In this case, the goals can be thought of as being divided into flexible and
inflexible goals, where the constraints are the inflexible (or rigid) ones. More
detailed presentations and practical applications of generalized goal program-
ming are given, for example, in Ignizio (1983a) and Korhonen (1991a). See also
Section 5.lD.
After the aspiration levels have been specified, the following task is to min-
imize the under- and overachievements of the objective function values with
respect to the aspiration levels. It is sufficient to study the deviational vari-
ables Oi = Zi - Ii(X). The deviational variable Oi may have positive or negative
values, depending on the problem. We can present it as the difference of two
positive variables, that is, Oi = 0; - 0;. We can now investigate how well each
of the aspiration levels is attained by studying the deviational variables. We
122 Part II - 4. A Priori Methods

can write Ii (x) + t5j - 151

= Zi for all i = 1, ... , k, where t5j is a negative devia-
tion or underachievement and 151
is a positive deviation or overachievement in
relation to the aspiration level. It is valid that t5j .151 = 0 for all i = 1, ... , k.
We now have the multiobjective optimization problem in a form where
we can minimize the deviational variables. For minimization problems it is
sufficient to minimize the k variables 151. If the ith goal is in the form of an
equality, we minimize t5j + 151.

4.3.2. Different Approaches

Thus far, we have only formulated the multiobjective optimization problem

in an equivalent form, where we have deviational variables as the objective
functions. There are several ways to proceed from this point. Here we present a
weighted (also called Archimedian) and a lexicographic (also called preemptive)
approach. More methods are handled in Ignizio (1983a) and some formulations
are explored in de Kluyver (1979).
In the weighted approach, see Charnes and Cooper (1977), the weighted
sum of the deviational variables is minimized. This means that in addition to
the aspiration levels, the decision maker must specify information about the
importance of attaining the aspiration levels in the form of weighting coeffi-
cients. The weighting coefficients are assumed to be positive and sum up to
one. The bigger the weighting coefficient is, the more important is the attain-
ment of that aspiration level. (Sometimes negative weighting coefficients are
used to represent a premium instead of a penalty.)
To put the introduction presented above into mathematical form and to
reason about the usage of the deviation variables, we can say that the problem
k

(4.3.1)
minimize L wil.f~(x) - zil
i=1
subject to xES

is converted into a new form by adding the overachievement variables

151 = max [0, Ji(X) - z;] or 151 = ~ [Iz; - Ji(x)1 + J;(x) - Zi]

and underachievement variables

This means that the absolute value signs can be dropped from problem (4.3.1)
by introducing the underachievement and the overachievement variables. The
resulting weighted goal programming problem is
 4.3. Goal Programming 123

k
minimize ~)wic5i + wtc5t)
.=1
(4.3.2) subject to Ii (x) + c5i - cSt = Zi for all i = 1, ... , k,
cSi,cSt 2:0 for all i=l,... ,k,
xES,
where we give separate weighting coefficients for underachievements and over-
achievements, and x ERn, cSi and c5t ' i = 1, ... , k, are the variables. If all the
goals are in the form li(x) ~ Zi, we can leave the underachievement variables
and write the problem in the form
k
minimize '" wT cST
L..J
i=1 ••
(4.3.3) subject to li(x) - cSt ~ Zi for all i = 1, ... ,k,
c5t 2: 0 for all i = 1, ... ,k,
xES,
where x E Rn and cSt, i = 1, ... , k, are the variables.
Figure 4.3.1 portrays how problem (4.3.3) is solved. The black spot is the
reference point of the aspiration levels. Every weighting vector produces differ-
ent contours by which the feasible objective region is to be intersected. Thus,
different solutions can be obtained by altering the weights. Contours with two
weighting vectors have been depicted in the figure. The bold line illustrates the
Pareto optimal set.

reference

•
point

Figure 4.3.1. Contours with different weighting vectors.

Even though the constraints cSi . cSt = 0 for all i = 1, ... ,k are not usually
included in the problem formulations, some attention must be paid to guar-
124 Part II - 4. A Priori Methods

antee that they are valid (see details in Rosenthal (1983)). An example of the
required conditions is given in Sawaragi et al. (1985, p. 253). The weighted goal
programming problem may be solved by standard single objective optimization
methods. If the original problem is linear, then the corresponding weighted goal
programming problem is also linear. The close connection between goal pro-
gramming and MOLP problems explains why the above-mentioned constraint
is usually absent from the problem formulation (it would make the problem
nonlinear) .
Note that weighted goal programming is closely related to the method of
weighted metrics or compromise programming. This can be seen particularly
well in formulation (4.3.1). Instead of the ideal objective vector, the reference
point of the decision maker is used in goal programming. The distances can be
measured by metrics other than the L1-metric. The L1-metric is widely used in
connection with goal programming because of the origin of the method in linear
programming. (This metric maintains the linearity of the problem.) If some
other Lp-metric is used there is another problem in determining an appropriate
value for p. Note, however, that if we have appropriate solvers available, we can
solve problem (4.3.1) directly without any deviational variables and using any
metric.
In the lexicographic approach, the decision maker must specify a lexico-
graphic order for the goals in addition to the aspiration levels. The goal at the
highest priority level is supposed to be infinitely more important than the goal
at the second priority level, etc. This means that no matter how large a mul-
tiplier is selected, a lower priority goal multiplied by it can never be made as
important as a higher priority goal. After the lexicographic ordering, the prob-
lem with the deviational variables as objective functions and the constraints as
in (4.3.2) is solved as explained in Section 4.2. In order to be able to use the
lexicographic approach, the decision maker's preference order for the objectives
must be definite and rigid.
A combination of the weighted and the lexicographic approaches, to be
called a combined approach, is quite popular. In this case, several objective
functions may belong to the same class of importance in the lexicographic
order. In each priority class, a weighted sum of the deviational variables is
minimized. The same weaknesses presented in connection with lexicographic
ordering are valid for this and the lexicographic approach.
It is not necessary to include the original constraints (x E S) in the lex-
icographic optimization problem in the normal way. They can be considered
to belong to the first priority level. In this way, they are taken into account
before any objective function is optimized and the feasibility of the solutions
is guaranteed by the nature of the lexicographic ordering.
Next, we prove a result concerning the Pareto optimality of the solutions of
goal programming.
 4.3. Goal Programming 125

Theorem 4.3.1. The solution of a weighted or a lexicographic goal program-

ming problem is Pareto optimal if either the aspiration levels form a Pareto
optimal reference point or all the deviational variables rSt for functions to be
minimized and 8i for functions to be maximized have positive values at the
optimum.

Proof. For the lexicographic approach, the proof corresponds to that of The-
orem 4.2.1. Here, we only present a proof for the weighted approach. For sim-
plicity of notation, we assume that the problem is of the form (4.3.3). A more
general case is straightforward.
Let x· E S be a solution of the weighted goal programming problem, where
the deviational variables (denoted here for clarity by rS;) are positive. Let us
assume that x· is not Pareto optimal. In this case, there exists a vector XO E S
such that fi(XO) $ h(x·) for all i = 1, ... , k and h(xO) < h(x·) for at least
one index j.
°
We denote h(x·) - h(xO) = f3 > 0. Then we set rSf = 8; > for i f- j and
8j = max [0, 8; - f3j ~ 0, where rSf is the deviational variable corresponding to
XO for i = 1, ... , k.
We have now fi(XO) - rSf $ h(x·) - 8; $ Zi for all i f- j. If 8; - f3 > 0,
then fj(xO) - 8j = h(xO) - rS; + h(x·) - fj(xO) ~ Zj, and if 8; - f3 ~ 0, then
h(xO) - rSj = h(xO) + h(x·) - h(x·) = h(x*) - f3 $ h(x·) - rS; $ Zj.

° °
This means that XO satisfies the constraints of problem (4.3.3). We have
8j < 8; (this is also valid if rSj = since 8; > for all i), and rSf $ rS; for all
if- j. As the weighting coefficients are positive, we have EwtrSi < EwtrS;,
which contradicts the fact that x' is a solution of weighted goal programming
problem (4.3.3).
For aspiration levels forming a Pareto optimal point the proof is self-evident.
o

Let us briefly mention one more form of goal programming, min-max goal
programming (suggested in Flavell (1976». It is not as widely used as the
weighted and the lexicographic approaches. For minimization problems the
min-max goal programming problem to be solved is

minimize . max 8t
t=l, ... ,k
(4.3.4) subject to hex) - 8t $ Zi for all i = 1, ... ,k,
xES,

where x E R n and rSt, i = 1, ... , k, are the variables.

126 Part II - 4. A Priori Methods

4.3.3. Comments

If the optimal objective function value of the goal programming problem

equals zero, some caution is in order, since the solution obtained may not
be Pareto optimal. The reason is that if the aspiration levels are all feasible,
then the value zero for all the deviational variables gives the minimum value
(zero) for the goal programming objective function. Thus the solution is equal
to the reference point, and normally there exist many feasible points that are
not Pareto optimal. If the solutions are intended to be Pareto optimal despite
the selection of the aspiration levels, then we must maximize the distance if
the aspiration levels are feasible and minimize the distance if the aspiration
levels are infeasible. This is the case with achievement scalarizing functions, as
explained in Section 3.5.
It is shown in Caballero et al. (1996) that the solution of min-max goal
programming problem (4.3.4) is Pareto optimal if it is unique. Other tests for
Pareto optimality in goal programming are provided in Romero (1991).
It is pointed out in Romero (1997) that lexicographic orderings imply dis-
continuous preferences. This means that lexicographic goal programming is in-
compatible with ordering the decision maker's preferences by a decreasing value
function. Thus, Romero recommends caution with lexicographic goal program-
ming because it is applicable only for problems with discontinuous preferences.
In Romero (1997), the importance of knowing the different preferential logics
underlying each goal programming approach is also emphasized. Different log-
ics are introduced and the min-max approach is slightly favoured, but mixtures
of the approaches are recommended.
The lexicographic goal programming approach can be modified so that it
can even take into account goals with lower priority, as suggested in Caballero
et al. (1996, 1997). It is proved in Caballero et al. (1996) that the solution
of the modified lexicographic goal programming problem is Pareto optimal if
the solution of the lowest priority level (Le., the last optimization problem) is
unique.
As pointed out in Dyer and Sarin (1981), although it is not readily apparent,
goal programming implicitly assumes that there is a measurable, additive and
rigid piecewise linear underlying value function. Rosenthal stresses, in Rosen-
thal (1983), that weighted goal programming problem (4.3.2) is equivalent to
the value function maximization problem where

8U(f(x)) = { wi if I;(x) < Zi,

81i -wi if I;(x) > Zi,
which means that the marginal utility is constant on either side of the aspiration
level. This is contrary to the economic idea that a decision maker considers the
next unit of decrease of Ii more important when Ii is plentiful than when Ii is
scarce.
 4.3. Goal Programming 127

This idea is even more evident when we look at the marginal rates of
substitution in goal programming problems. In Remark 2.8.7 of Part I it
was mentioned that the marginal rates of substitution may be defined as
mij (x) = aU~~~x)) / au~~~x)). Thus, goal programming does not take into con-
sideration the possibility that it is easier for the decision maker to let some-
thing increase a little if (s)he has got little of it than if (s)he has got much of
it. The reason for this is that goal programming implicitly assumes that the
marginal rates of substitution are piecewise constant. This critique also applies
to the lexicographic approach (see details in Rosenthal (1983, 1985)). More
critical observations about goal programming are presented in Romero (1991)
and Rosenthal (1983).

4.3.4. Applications and Extensions

A comprehensive presentation on goal programming and its extensions is

given in Ignizio (1976), and a summary of different variations of goal program-
ming is provided in Charnes and Cooper (1977). In addition, a wide survey of
the literature around goal programming up to the year 1983 is presented in
Soyibo (1985). Several modifications and improvements as well as applications
are reviewed. A survey of goal programming is also given in Kornbluth (1973)
and the weighted and the lexicographic approaches are applied to problems
with fractional objective functions. Further, a broad collection of journal pa-
pers and books on goal programming is assembled in Schniederjans (1995a).
References in nine broad areas of application are also included.
In the literature, goal programming is the most widely used solution method
in terms of practical applications. Weighted goal programming with equal
weighting coefficients is employed in the planning of public works in Yoshikawa
et al. (1982). Weighted goal programming with sensitivity analysis is also used
for portfolio selection in Tamiz et al. (1996).
Lexicographic goal programming is applied in Benito-Alonso and Devaux
(1981) to a problem concerning the location and size of day nurseries, in Sinha
et al. (1988) to storage problems in agriculture and in Mitra and Patankar
(1990) to aid manufacturers in selecting the price and the warranty time of
their products. Lexicographic goal programming is also applied in Kumar et
al. (1991) to nonlinear multi-stage decision problems in manufacturing systems,
in Ng (1992) to aircraft loading and in Brauer and Naadimuthu (1992) to solve a
mixed integer MOLP problem involving inventory and distribution planning. In
Hemaida and Kwak (1994) a linear trans-shipment problem and in Current and
Storbeck (1994) a location model are solved by lexicographic goal programming,
and in Giannikos et al. (1995) it is applied in an integer allocation problem. In
Berbel and Zamora (1996) lexicographic goal programming is applied in wildlife
management and in Kim et al. (1997) in solving a linear problem of military
budget planning. An implementing decision support system is also described.
A numerical application example of combined goal programming is given in
Hwang and Masud (1979, pp. 79-95). Combined goal programming is applied
128 Part II - 4. A Priori Methods

in Levary (1986) to problems of optimal control, in Giokas and Vassiloglou

(1991) to the (linear) management of the assets and liabilities of a Greek bank
and in Ghosh et al. (1992) to the resource planning of university management.
In Sankaran (1990), the combined approach is also used to solve an integer
MOLP problem in cell formation, and in Schniederjans and Hoffman (1992),
combined zero-one goal programming is applied to a problem concerning inter-
national business expansion analysis. The ideas of combined goal programming
are adapted in Miyaji et al. (1988) in solving a transportation problem-type
problem of dividing students into groups. In addition, the combined goal pro-
gramming approach is applied in fund and portfolio management in Powell and
Premachandra (1998).
The applications mentioned here are only a few of the existing ones. The
popularity of goal programming is well affirmed by the fact that in a bibli-
ography collected in White (1990) on multiobjective optimization applications
(covering the years from 1955 to 1986) more than a half involved goal program-
ming.
Four different goal interpretations in multiobjective optimization are pre-
sented in Dinkelbach (1980). Goal programming is adapted to multiobjective
generalized networks for integer problems in Ignizio (1983b). In Inuiguchi and
Kume (1991), goal programming is extended to linear problems where the co-
efficients and the aspiration levels are given as intervals. The aspiration level
intervals do not there represent regions within which the decision maker is sat-
isfied, but regions where the aspiration levels may vary. A generalization of
goal programming through the theory of variational inequalities is presented in
Thore et al. (1992).
An extension of goal programming to MOLP problems is given in Martel
and Aouni (1990). Instead of the deviational variables, some functions describ-
ing the wishes of the decision maker about attaining the goals set are used
in the weighted approach. An illustrative example is also provided. Technical
improvements to the preference modelling method of Martel and Aouni are
presented in Tamiz and Jones (1995). This approach is extended in Martel and
Aouni (1998) by allowing goals to be intervals instead of exact numbers. This
means that indifference thresholds (see Subsection 5.9.1) are used in modelling
the imprecision of the goals. (Even though we have mentioned some interesting
MOLP extensions and solution methods, we skip most of them here.)
An adaptation of lexicographic goal programming for convex problems is
provided in Caballero et al. (1996). The idea is to produce satisfying solutions
by solving the hybrid problem (in Section 3.3) with the components of the goal
programming solution as upper bounds. Varying the weights produces different
solutions.
Lexicographic goal programming is modified significantly in Caballero et
al. (1997). No deviational variables are used and the objective function of each
priority level is optimized at each iteration. The approach is valid for convex
problems.
 4.3. Goal Programming 129

A solution method for lexicographic goal programming problems where ob-

jective functions are fractions of linear or nonlinear functions is described in
Pal and Basu (1995). More than one objective function can then belong to the
same priority class. The method has characteristics of dynamic programming.
A generalized reduced gradient (GRG) method-based solution algorithm for
lexicographic and weighted nonlinear goal programming problems is introduced
in Saber and Ravindran (1996). This partitioning technique is demonstrated
to be reliable and robust. Several aspects to take into account when aiming at
the efficient implementation of goal programming approaches are collected in
Tamiz and Jones (1996).
Goal programming can be expanded in an interactive direction in different
ways. One can systematically modify the weighting vectors or the lexicographic
order of the objective functions or ask for new aspiration levels from the decision
maker. These topics are considered in Tamiz and Jones (1997a, b).

4.3.5. Concluding Remarks

Goal programming is a very widely used and popular solution method for
practical multiobjective optimization problems. One of the reasons is its age.
Another reason is that goal-setting is an understandable and easy way of mak-
ing decisions. The specification of the weighting coefficients or the lexicographic
ordering may be more difficult. The weights do not have so direct an effect on
the solution obtained as in the a priori weighting method. However, they are
relative to each other. This means that only the relations of the weighting co-
efficient matter, not the weights themselves. It may be difficult to specify the
weights because they have no direct physical meaning. It is demonstrated in
Nakayama (1995) that desirable solutions are very difficult to obtain by ad-
justing the weighting coefficients in the weighted goal programming problem.
Anyway, it is as advisable as in the weighting method to normalize the objective
functions when weighting coefficients are used.
One must be careful with the selection of the aspiration levels so that the
Pareto optimality of the solutions can be guaranteed. The correct selection may
be difficult for a decision maker who does not know what the feasible region
looks like. Presenting the ranges of the Pareto optimal set, or at least the ideal
objective vector, to the decision maker may help in the selection.
Goal programming is not an appropriate method to use if it is desired to
obtain trade-offs. Another restricting property is the underlying assumption of
a piecewise linear value function and thus piecewise constant marginal rates of
substitution.
Assuming that goal programming follows a traditional product life cycle,
it is inferred in Schniederjans (1995b) that the current stage of productivity
is in decline. It is pointed out that the number of goal programming papers
has been on the decrease for several years. One of the reasons suggested is the
aging of the few active contributors to goal programming.
 Part III

RELATED ISSUES
 5. INTERACTIVE METHODS

The class of interactive methods is the most developed of the four classes
of methods presented here. The interest devoted to this class can be explained
by the fact that assuming the decision maker has enough time and capabilities
for co-operation, interactive methods can be presumed to produce the most
satisfactory results. Many of the weak points of the methods in the other three
classes are overcome. Namely, only part of the Pareto optimal points has to be
generated and evaluated, and the decision maker can specify and correct her
or his preferences and selections as the solution process continues and (s)he
gets to know the problem and its potentialities better. This also means that
the decision maker does not have to know any global preference structure. In
addition, the decision maker can be assumed to have more confidence in the
final solution since (s)he is involved throughout the solution process.
In interactive methods, the decision maker works together with an analyst
or an interactive computer program. One can say that the analyst tries to de-
termine the preference structure of the decision maker in an interactive way.
A solution pattern is formed and repeated several times. After every iteration,
some information is given to the decision maker and (s)he is asked to answer
some questions or provide some other type of information. The working or-
der in these methods is: 1) analyst, 2) decision maker, 3) analyst, 4) decision
maker, etc. After a reasonable (finite) number of iterations every interactive
method should yield a solution that the decision maker can be satisfied with
and convinced that no considerably better solution exists. The basic steps in
interactive algorithms can be expressed as
a) find an initial feasible solution,
b) interact with the decision maker, and
c) obtain a new solution (or a set of new solutions). If the new solution
(or one of them) or one of the previous solutions is acceptable to the
decision maker, stop. Otherwise, go to step b).
Interactive methods differ from each other by the form in which information
is given to the decision maker, by the form in which information is provided by
the decision maker, and how the problem is transformed into a single objective
optimization problem. One problem to be solved when designing an interactive
method is what kind of data one should use to interact with the decision maker.
It should be meaningful and easy for the decision maker to comprehend. The

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
132 Part II - 5. Interactive Methods

decision maker should understand the meaning of the parameters for which
(s)he is asked to supply values. On the other hand, the data provided to the
decision maker should be easily obtainable by the analyst and contain infor-
mation about the system. Too much information should not be used and the
information obtained from the decision maker should be utilized efficiently. To
ensure that the greatest possible benefit can be obtained from the interactive
method, the decision maker must find the method worthwhile and acceptable,
and (s)he must be able to use it properly. This usually means that the method
must be understandable and sufficiently easy to use. This aim calls for research
in understanding the underlying decision processes and how decisions are made.
As stressed in Kok (1986), experiments in psychology indicate that the
amount of information provided to the decision maker has a crucial role. Ifmore
information is given to the decision maker, the percentage of the information
used decreases. In other words, more information is not necessarily better than
less information. More information may increase the confidence of the decision
maker in the solution obtained but the quality of the solution may nonetheless
be worse.
In addition to the fact that the decision maker has an essential role in
interactive methods, the analyst should not be forgotten either. The analyst
can support the decision maker in many ways and, in the best possible case,
explain the behaviour of the problem to the decision maker. Thus, the analyst
may play a meaningful role in the learning process of the decision maker.
Interactive methods have been classified in many ways, mainly according to
their solution approaches. Here we do not follow any of those classifications. Let
us, however, mention two different conceptions regarding interactive approaches
according to Vanderpooten (1989a, b, 1992). The approaches are searching
and learning. In searching-oriented methods a converging sequence of solution
proposals is presented to the decision maker. It is assumed that the decision
maker provides consistent preference information. In learning-oriented methods
a free exploration of alternatives is possible allowing trial and error. The latter
does not guide the decision maker and convergence is not guaranteed. The best
procedure would be a combination of these two approaches, drawing on their
positive features. Such an approach would support the learning of preferences,
while it would also include guiding properties.
Before we present any methods, some critical comments are in order. Re-
peatedly, it has been and will be assumed that the decision maker makes consis-
tent decisions or that (s)he has an underlying (implicitly known) value function
upon which her or his decisions are made. The purpose is not to go deeply into
the theories of decision making. However, it is worth mentioning that those
assumptions can be called into question because they are difficult to verify.
Consistency of the responses of the decision maker is one of the most im-
portant factors guaranteeing the success of many interactive solution methods.
Because of the subjectivity of the decision makers, different starting points,
different types of questions or interaction styles may lead to different final so-
 5. Interactive Methods 133

lutions. Some methods are more sensitive with respect to consistency than oth-
ers. The handling of inconsistency with respect to several interactive methods is
treated in Shin and Ravindran (1991). In general, inconsistency can be reduced
by consistency tests during the solution process or by minimizing the decision
maker's cognitive burden. In other words, interactive methods assuming con-
sistent answers should have built-in mechanisms to deal with inconsistencies.
This is one of the motivations in developing new methods for multiobjective
optimization.
Further, once the existence of an underlying, implicit value function is sup-
posed, several assumptions are set on it. How can one guarantee and verify,
for example, the pseudoconcavity of a function that is not explicitly known?
Naturally, something can be concluded if we find out enough about the deci-
sion maker's preference structure. Steps in that direction are, however, very
laborious and in any case the results are likely to be controversial.
In solving practical problems, knowledge about decision processes and deci-
sion analysis is needed to guarantee fruitful co-operation between the decision
maker and the analyst. An understanding of the behaviour of the decision
maker is important in both developing and applying interactive methods. This
fact has been somewhat underestimated, as emphasized in Korhonen and Wal-
lenius (1996, 1997). Korhonen and Wallenius also handle several behavioural
issues related to interactive methods. Among them are the learning process
of the decision maker, her or his wish to control the search process, and the
permissibility of cyclic behaviour or making errors. Perhaps the behavioural
sciences should be taken more widely into account when designing interactive
methods. A critique of the assumptions underlying interactive methods is also
presented in French (1984). The primary concern is that the assumptions should
be supported by empirical research from the behavioural sciences.
One noteworthy aspect is that it is unrealistic to assume that decision mak-
ers can provide precise information and inputs. After studying 86 reported
applications of decision analysis in the literature, it is concluded in Corner and
Corner (1995) that the methods should become more user-friendly and descrip-
tive in dealing with the input of the decision maker. In Wierzbicki (1997a), it
is stressed that intuition plays an essential role in decision making. Wierzbicki
defines intuitive decisions as "quasiconscious and subconscious information pro-
cessing, leading to an action, utilizing aggregated experience and training and
performed (most probably) by a specialized part of the human mind." To pro-
voke intuitive decision making, analysts should provide information in rich and
multidimensional graphic terms and avoid requiring consistency.
Decision making is appositely described in Zeleny (1989) as "searching for
harmony in chaos." One can criticize the way decision makers are forced into a
priori formulas, patterns or contexts (like wandering around the Pareto optimal
set). Instead, the decision maker should be guided through her or his own
creative search process since decision making can be regarded as a process of
continuous redefinition of the problem.
134 Part II - 5. Interactive Methods

Decision analysis is not handled here in more detail. The above-mentioned

aspects are only a few examples of the issues involved.
One more interesting concept is the convergence of an interactive method.
One can understand several different features as convergence. On the one hand,
it may be said that the method converges into Pareto optimal points if the fi-
nal solution can be proved to be Pareto optimal. One can also say that the
method converges into a satisficing solution, if the final solution is satisficing.
Finding the best Pareto optimal compromise solution may be understood as
convergence. On the other hand, convergence may mean that the final solution
is optimal for an underlying value function. This kind of mathematical conver-
gence result necessitates certain assumptions about the underlying value func-
tion. In this case, the observations mentioned above are valid. If the method is
not based on the assumption on any underlying value function, this conception
of convergence cannot always be applied.
To sum up, it is not unequivocal what convergence means and how it should
be proved. For this reason, it is difficult to provide convergence results for the
different methods under consideration. It can also be claimed that mathemat-
ical convergence is neither necessary nor sufficient to indicate the practical va-
lidity of a method, as stated, for example, in Stewart (1997). The same idea is
also expressed in Gardiner and Vanderpooten (1997) and Zionts (1997a, b). On
the grounds of the above-mentioned statements, the mathematical convergence
properties have been relegated to a secondary position in what follows.
As far as the structure of the methods is concerned, one can require that
interactive procedures should converge well immediately in the few initial iter-
ations. This is concluded, for example, in Korhonen et al. (1990) after experi-
mental tests with interactive methods. Decision makers are not willing to wait
for progress for a long time.
Stopping criteria are related to the convergence of interactive methods.
There are three main stopping criteria. Either the decision maker gets tired of
the solution process, some algorithmic stopping (convergence) rule is fulfilled or
the decision maker finds a desirable solution and wants to stop. It is difficult to
define precisely when a solution is desirable enough to become a final solution.
The convergence of the method has sometimes been considered to be an
important factor when selecting a method. However, as stated in Vanderpooten
and Vincke (1989), the solution process should not be stopped because of any
convergence test. The only practical stopping criterion is the satisfaction of
the decision maker with the solution obtained. This usually means that the
decision maker must feel that (s)he has received enough information about the
problem to be solved.
The current view is that a solution is a final solution if the decision maker
is convinced that it is preferred to all the other Pareto optimal solutions (see
Korhonen and Wallenius (1996, 1997)). This means that the decision maker
must have sufficient evidence that no significantly better solutions exist (see
Gardiner and Vanderpooten (1997)). Gardiner and Vanderpooten have studied
the interactive solution processes reported in the literature. They point out
 5. Interactive Methods 135

that the median number of iterations has been between three and eight. One
can ask whether such rapid convergence is the result of getting tired or whether
it is due to some other reason. Possibly the decision makers did not know how
to continue the solution process.
An important factor when using interactive solution methods is the selection
of the starting point. Particularly for nonconvex problems where the objective
functions may have several local optima, the starting point affects greatly the
solutions generated. If the starting point is somehow biased, it may anchor the
desires and the preferences of the decision maker. It is not desirable that the
final solution is affected by the starting point. In general, the starting point
should provide a useful basis for the decision maker in exploring the Pareto
optimal set. The starting point can, for example, be generated by some of the
noninteractive methods.
Nonconvexity is a mathematical aspect. Another aspect related to starting
points from the point of view of human judgment and decision making is the
above-mentioned anchoring. To be more exact, anchoring means that the deci-
sion maker fixes her or his thinking on some (possible irrelevant) information,
like the starting point, and fails to sufficiently adjust and move away from
that anchor. In other words, the decision maker is unable to move far from
the starting point. This kind of behavioural perspective on interactive decision
making is handled in Buchanan and Corner (1997). On the basis of a number
of experiments it is argued that anchoring effects are connected more to di-
rected and structured solution methods than to methods based on free search.
Buchanan and Corner conclude that whenever an anchoring bias is possible, it
is important that the starting point reflects the initial pref€rences of the deci-
sion maker. The reasoning is that since any starting point is likely to bias the
decision maker, it is best to bias her or him in the right direction.
Even though interactive methods can be regarded as most promising so-
lution methods for multiobjective optimization problems, there are still cases
where these methods are not practicable regardless of the availability of the de-
cision maker. Such problems include, for instance, many engineering problems
that require extensive and expensive calculations (like large-scale finite element
approximations). One must, however, remember that computational facilities
have developed greatly during the last few years. Thus, the number of problems
that cannot be solved by interactive methods has decreased. See Osyczka and
Zajac (1990) for a suggestion of handling computationally expensive functions.
On the other hand, the large number of objective functions may make interac-
tive methods impractical. In this case, it may be difficult for the decision maker
to absorb the information provided and to give consistent answers in order to
direct the solution process.
Below, we present several interactive methods. Some of them are relatively
old and much tested and developed, whereas some others are new and deserve
further refinement. The methods to be described are the interactive surrogate
worth trade-off method, the Geoffrion-Dyer-Feinberg method, the sequential
136 Part II - 5. Interactive Methods

proxy optimization technique, the Tchebycheff method, the step method, the
reference point method, the GUESS method, the satisficing trade-off method,
the light beam search, the reference direction approach, the reference direction
method and the NIMBUS method. The first three methods are based on the
existence of an underlying value function, whereas the last eight use reference
points and the classification of the objectives. (In developing the last of these,
attempts have been made to overcome some of the drawbacks observed in the
other methods.)
All the methods to be presented are based on generating mainly weakly,
properly or Pareto optimal solutions. In each method, it is assumed that less
is preferred to more by the decision maker. The same notion could be formu-
lated to require that the underlying value function is strongly decreasing. The
reason for avoiding this wording is that an underlying value function is not
always assumed to exist. The assumption only concerns the form of the general
preference structure of the decision maker.
In connection with the methods, some applications reported in the literature
are mentioned. However, let us keep in mind that the impressions obtained from
such applications may be biased because unsuccessful applications are hardly
ever published. In addition, we give references for extensions and modifications
of the methods. We also indicate whether the methods belong to the class of ad
hoc or non ad hoc methods. (These classes were introduced at the beginning
of this part in Chapter 1.)
Throughout the book the iteration counter is denoted by h and the deci-
sion variable vector at the current iteration by xh. In addition, the number of
alternative objective vectors presented to the decision maker is denoted by P.

5.1. Interactive Surrogate Worth Trade-Off Method

The interactive surrogate worth trade-off (ISWT) method, put forward in

Chankong and Haimes (1978, 1983b, pp. 371-379), is an extension of the surro-
gate worth trade-off (SWT) method presented in Haimes and Hall (1974) and
Haimes et al. (1975). We do not go into details of the SWT method here, but
present directly the interactive version. The motivation for including the ISWT
method in this book is that it is of theoretical interest.

5.1.1. Introduction

The .s-constraint method, introduced in Section 3.2, is a fundamental el-

ement of the ISWT method. The idea is to maximize an underlying (implic-
itly known) value function. The opinions of the decision maker concerning the
trade-off rates at the current solution point determine a search direction. The
step-size to be taken in the search direction is determined by solving several
 5.1. Interactive Surrogate Worth Trade-Off Method 137

c-constraint problems and asking the decision maker to select the most satis-
factory solution for the continuation. In what follows, appropriate assumptions
are assumed to be valid so that the solutions produced by the c-constraint
method are Pareto optimal (see Section 3.2).
It is assumed that
1. The underlying value function U: R k -t R exists and is implicitly known
to the decision maker. In addition, U is continuously differentiable and
strongly decreasing.
2. The objective and the constraint functions are twice continuously differ-
entiable.
3. The feasible region S is compact (so that a finite solution exists for every
feasible c-constraint problem).
4. The assumptions in Theorem 3.2.13 are satisfied.

5.1.2. ISWT Algorithm

The main features of the ISWT method can be presented cursorily with
four steps.
(1) Select the reference function It to be minimized and give upper bounds
to the other objective functions. Set h = l.
(2) Solve the current c-constraint problem to get a Pareto optimal solution
xh. Trade-off rate information is obtained from the connected Karush-
Kuhn-Tucker multipliers.
(3) Ask the opinions of the decision maker with respect to the trade-off
rates at zh corresponding to xh.
(4) If some stopping criterion is satisfied, stop with xh as the final solu-
tion. Otherwise update the upper bounds of the objective functions
with the help of the answers obtained in step (3) and solve several c-
constraint problems (to determine an appropriate step-size). Let the
decision maker choose the most preferred alternative. Denote the cor-
responding decision vector by xh+1 and set h = h + l. Go to step (3).
First, we examine how trade-off rate information is obtained from Karush-
Kuhn-Tucker multipliers. As noted in Theorem 3.2.13 of Section 3.2, the
Karush-Kuhn-Tucker multipliers represent trade-off rates under the specified
assumptions.
Let xh E S be a solution of the c-constraint problem at the iteration h,
where It is the function to be minimized and the upper bounds are c7 for
i = 1, ... , k, i :f; P. We suppose that xh satisfies the assumptions specified in
Theorem 3.2.13. If the Karush-Kuhn-Tucker mUltipliers Aii associated with the
constraints I;(x) ~ c7 are strictly positive for all i = 1, ... , k, i :f; P, then Aii
represents the partial trade-off rate at xh between It and Ii. In other words, if
the multiplier Aii corresponding to the constraint involving /; is positive, this
particular constraint is active and binds the optimum.
138 Part II - 5. Interactive Methods

We know now that to move from xh to some other (locally) Pareto optimal
solution in the neighbourhood of xh, the value of the function It. decreases by
A~i units for every unit of increment in the value of the function h (or vice
versa), while the values of all the other objective functions remain unaltered.
The opinion of the decision maker with regard to this kind of trade-off rate for
all i = 1, ... ,k, i "I f, is found out by posing the following question.
Let an objective vector f (xh) = zh be given. If the value of ie is decreased
by '\~i units, then the value of fi is increased by one unit (or vice versa)
and the other objective values remain unaltered. How desirable do you find
this trade-off?
If the situation is not so convenient as presented above, that is, some of the
Karush-Kuhn-Tucker multipliers '\~i equal zero, then another type of question is
needed. Let us suppose that A~i > 0 for i E N> and A~j = 0 for j E N=, where
N> U N= = {i I i = 1, ... ,k, i "I f}. As noted in Theorem 3.2.13, increasing
the value of fi' where i E N> decreases the value of It. and in addition, the
values of all fj also change, where j E N=. The question to the decision maker
for all i E N> is now of the form
Let an objective vector f(xh) = zh be given. If the value of It. is decreased
by A~i units, then the value of h is increased by one unit (or vice versa) and
the values of h for j E N= change by 'Vh(xh)Ta~;ih) units. How desirable
do you find these trade-offs?
A problem with the question above is that the values of a~~h) for i E N>
are unknown. One of the ways suggested in Chankong and H~imes (1983b)
for coping with this is that the values can be approximated by solving the
c:-constraint problem with a slightly modified upper bound vector as ~(i) =
(c:~, ... ,c:LI' C:~+l' ... ,c:? + E, ... ,c:~), where E "lOis a scalar with a small
absolute value. Let the solution of this s-constraint problem be x(~(i)). We
obtain now an approximation by

8x(~) x(~(i)) - Xh
&;-~ E

Note that the decision maker's opinions are asked respecting certain amounts
of change in the values of the objective functions, and not of changes in general.
The following problem to be handled is the form of the answers expected from
the decision maker. It is suggested in Chankong and Haimes (1978, 1983b) that
the decision maker must specify an integer between 10 and -10 to indicate her
or his degree of preference. If the decision maker is completely satisfied with the
trade-off suggested, the answer is 10. Positive numbers less that 10 indicate the
degree of satisfaction (less than complete). Correspondingly, negative answers
reflect the decision maker's satisfaction with the trade-off which is converse to
that in the question. The answer 0 means that the decision maker is indifferent
to the given trade-off.
 5.1. Interactive Surrogate Worth Trade-Off Method 139

In Tarvainen (1984), it is suggested that far fewer choices are given to the
decision maker. The possible answers are integers from 2 to -2 and their mean-
ing corresponds to that presented above. The justification is that it is easier for
the decision maker to give an answer and maintain some kind of consistency
when there are fewer alternatives. These five alternatives are enough to rep-
resent the direction and rough degree of the decision maker's preferences and
satisfaction.
Regardless of the scale selected, the response of the decision maker is called
a surrogate worth of the trade-off rate between It and Ii at xh and denoted by
Wl~' At each point xh, a number of k - 1 (or less, if N= :j:. 0) questions of the
previously described form are presented to the decision maker and the values
for Wl~ (i = 1, ... ,k, i :j:. £) are obtained.
According to Theorem 3.2.13, there exists a Pareto optimal solution in the
neighbourhood of xh when the values of the objective functions are changed
according to the information given in the trade-off rates. The problem is how
much the values of the objective functions can be changed in order to remain on
the Pareto optimal surface and obtain the best possible solution. We must find
a way to update the upper bounds of the objective functions in an appropriate
way.
How to proceed from this point depends on the scale chosen for the surrogate
worth values. The idea is to obtain an estimate for the gradient of the underlying
value function with the help of the surrogate worth values. Then a steepest
ascent-type formulation is used to maximize the value function. The upper
bounds of the c:-constraint problem are revised and a new solution is obtained.
It is assumed to satisfy the preferences of the decision maker indicated by the
surrogate worth values as well as possible.
In the original version by Chankong and Haimes, it is suggested that the
upper bounds are updated from iteration h to h + 1 by

E~+l = c~ + t(WI~Ij;(xh)l)
for i E N> and

for j E N=, where i E N> and t is a step-size to be determined. For details,

see Chankong and Haimes (1978) and references therein.
For simplicity, it is assumed in Tarvainen (1984) that the Karush-Kuhn-
Tucker multipliers are all strictly positive. The decision maker is asked to spec-
ify small and meaningful amounts 11j; for all i = 1, ... , k, i :j:. £. The scalar
iJ.j; represents the amount of change in the value of Ii that is relevant to the
decision maker. The upper bounds are now updated by

c:~+1 = c:~ + t(WI~l1J;)

for i = 1, ... , k, i :j:. £, where t denotes the step-size.
140 Part II - 5. Interactive Methods

Several discrete values may be given to the step-size t in each updating

formula. Then the £-constraint problem is solved for every value. The resulting
objective vectors are presented to the decision maker, who is asked to choose
the most preferred one. A graphical representation of the alternatives may be
helpful. This topic is handled in Chapter 3 of Part III. After choosing the new
solution (and thus an appropriate step-size), trade-off rate information at that
solution is obtained from the corresponding Karush-Kuhn-'Thcker multipliers
(as earlier). The procedure continues by asking the decision maker for the
surrogate worth values.

5.1.3. Comments

In practice, when the decision maker is asked to express her or his prefer-
ences concerning the trade-off rates, (s)he is implicitly asked to compare the
trade-off rates with her or his marginal rates of substitution. (Naturally, the
decision maker does not have to be able to specify the marginal rates of sub-
stitution explicitly.) If mu < Au, then the surrogate worth value is positive
(and the contrary respectively). If mu = Ali for all i = 1, ... , k, i 'l-i, meaning
Wei = 0, then the stopping criterion (2.8.1) introduced in Subsection 2.8.2 of
Part I is valid. Thus, the condition Wl~ = 0 for all i 'l-i is a common stopping
criterion for the algorithm. Another possible stopping situation is that the de-
cision maker wants to proceed, but only in an infeasible direction. The latter
condition is more difficult to check.
The ISWT method can be classified as non ad hoc in nature. If the value
function is known, then the trade-off rates are easy to compare with the
marginal rates of substitution. Further, when comparing alternatives, it is easy
to select the one with the highest value function value.
The convergence rate of the ISWT method greatly depends on the accuracy
and the consistency of the answers of the decision maker. It was pointed out in
Section 2.8 of Part I that it is important to select the reference function care-
fully. This comment is also valid when considering the convergence properties.
If there is a sharp limit in the values of the reference function where there is a
change in satisfaction from 'very satisfactory' to 'very unsatisfactory,' the so-
lution procedure may stop too early. Further references are cited in Chankong
and Haimes (1978) for convergence results.
A method related to the ISWT method is presented in Chen and Wang
(1984). The method is an interactive version of the SWT method, where new
solution alternatives are generated by Lin's proper equality method (see Section
3.2), and the decision maker has to specify only the sign of the surrogate worth
values.
There are many other modifications of the SWT method in the literature.
Among others, it is generalized for multiple decision makers in Chankong and
Haimes (1983b, pp. 359-366), Haimes (1980) and Hall and Haimes (1976). The
first two handle also the SWT method in stochastic problems.
 5.2. Geoffrion-Dyer-Feinberg Method 141

5.1.4. Concluding Remarks

The role of the decision maker is quite easy to understand in the ISWT
method. (S)he is provided with one solution and has to specify the surrogate
worth values. The complicatedness of giving the answers depends on how ex-
perienced the decision maker is in such specification and which variation of the
method is employed. The set of 21 different alternatives as surrogate worth
values in the original version is quite a lot to select from. It may be difficult for
the decision maker to provide consistent answers throughout the decision pro-
cess. In addition, if there is a large number of objective functions, the decision
maker has to specify a lot of surrogate worth values at each iteration. At least
for some decision makers it may be easier to maintain consistency when there
are fewer alternative values for the surrogate worth available (as suggested by
Tarvainen (1984».
Trade-off rates play an important role in the ISWT method, and that is
why the decision maker has to understand the concept of trade-off properly.
Attention must also be paid to the ease of understanding and careful formula-
tion of the questions concerning the trade-off rates. Careless formulation may,
for example, cause the sign of the surrogate worth value to be changed.
It is a virtue that all the alternatives during the solution process are Pareto
optimal. Thus, the decision maker is not bothered with any other kind of solu-
tions.
A negative feature is that there are a lot of different assumptions to be
satisfied to guarantee that the algorithm works. It may be difficult (and at
least laborious) in many practical problems to ensure that the assumptions are
satisfied. One can argue that the validity of the assumptions is not always that
important in practice. However, for example, the correctness of the trade-off
rates is crucial for the success of the ISWT method.

5.2. Geoffrion-Dyer-Feinberg Method

The Geoffrion-Dyer-Feinberg (GDF) method, proposed in Geoffrion et

aJ. (1972)' is an interactive method based in principle on the same idea as
the ISWT method; maximization of the underlying (implicitly known) value
function. The realization is quite different, though. The GDF method is one of
the most well-known interactive methods.

5.2.1. Introduction

The basic idea behind the GDF and the ISWT methods is the same. At
each iteration, a local approximation of an underlying value function is gener-
ated and maximized. In the GDF method, the idea is somewhat more clearly
visible. Marginal rates of substitution specified by the decision maker are used
142 Part II - 5. Interactive Methods

to approximate the direction of steepest ascent of the value function. Then the
value function is maximized by a gradient-based method. A gradient method
of Frank and Wolfe (FW) (see Frank and Wolfe (1956)) has been selected for
optimization because of its simplicity and robust convergence (rapid initial con-
vergence) properties. The GDF method is also sometimes called an interactive
Frank-Wolfe method, because it has been constructed on the basis of the FW
method.
The problem to be solved here is
maximize u(x) = U(f(x))
(5.2.1)
subject to xES.

It is assumed that
1. The underlying value function U; R k --+ R exists and is implicitly known
to the decision maker. In addition, u; R n --+ R is a continuously differen-
tiable and concave function on S (sufficient conditions for the concavity
are, for example, that U is a concave decreasing function and the objec-
tive functions are convex; or U is concave and the objective functions
are linear), and U is strongly decreasing with respect to the reference
function (denoted here by It) so that au~~~x)) < o.
2. The objective functions are continuously differentiable.
3. The feasible region S is compact and convex.
Let us begin by presenting the main principles of the FW method. Let a
point xh E S be given. The idea of the FW method is that when maximizing
some objective function u: Rn --+ R subject to constraints XES, instead of u,
a linear approximation of it at some point xh E S is optimized. If the solution
obtained is yh, then the direction d h = yh - xh is a promising direction in
which to seek an increased value for the objective function u.
At any feasible point x", a linear approximation to u(y) is

The maximization of this linear approximation, after excluding constant terms,

is equivalent to the problem
maximize \7xu(xh)7'y
(5.2.2)
subject to yES,

where xh is fixed and y ERn is the variable. Let yh E S be the solution.

A well-known condition for Xh to be an optimal solution of problem (5.2.1)
is that \7 xu(xh)7' d ~ 0 for all dES. Therefore, if after solving problem (5.2.2)
is yh = xh, then we know that 0 = \7x u (x h )T(yh - xh) ~ \7x u (x h )T(y - xh)
for all YES, and, thus, the optimality condition is fulfilled at Xh.
If yh "I- xh, then we set d h = yh - Xh. The points yh and xh are feasible
and, because of the convexity assumption of S, any new point X h + 1 = xh +td h
 5.2. Geoffrion-Dyer-Feinberg Method 143

where 0 ~ t ~ 1 is feasible. Finally, we must determine an appropriate step-size

in the direction d h by maximizing u(x h + td h ) subject to 0 ~ t ~ 1.

5.2.2. GDF Algorithm

Below, we shall show that even though we do not know the value function
explicitly, we can obtain a local linear approximation for it or to be more exact,
its gradient, with the help of marginal rates of substitution. This is enough to
permit the FW method to be applied. Before going into details we present the
basic phases of the GDF algorithm.
(1) Ask the decision maker to specify a reference function It. Choose a
feasible starting point Xl. Set h = 1.
(2) Ask the decision maker to specify marginal rates of substitution between
it and the other objectives at the current solution point xh.
(3) Solve problem (5.2.3), where the approximation of the value function
is maximized. Denote the solution by yh E S. Set the direction d h =
yh _ xh. If d h = 0, go to step (6).
(4) Determine with the help of the decision maker the appropriate step-size
t h to be taken in the direction d h . Denote the corresponding solution
by xhH = xh + thd h .
(5) Set h = h + 1. If the decision maker wants to continue, go to step (2).
(6) Stop. The final solution is xh.
In the algorithm above we need a local linear approximation of the value
function at the point xh. As explained earlier, we only need to know the gradient
of the value function at xh. According to the chain rule, we know that the
gradient of the objective function of problem (5.2.1) at the point xh E Scan
be written in the form

In assumption 1 we supposed that aU~(;h» < 0, where it is the reference

function. Positive scaling does not affect t~e direction of the gradient, so we
can divide the gradient of the value function by a positive scalar _ au(!}:"».
We have now the direction of the gradient of the value function at the point
xh in the form
k
L -m?'V x/i(xh),
i=l

h
were m ih = aU(f(xh»
ali / aU(f(xh»
alt lor a11'z = 1, ... , k, Z. ...J,
c b
-r e. The num ers m i
h

(= mli) represent the marginal rates of substitution at xh between Ie and Ii

(see Remark 2.8.7 of Part I).
144 Part II - 5. Interactive Methods

The role of the reference function is significant, because marginal rates of

substitution are generated with respect to it. The decision maker must be asked
to specify the reference function so that the marginal rates of substitution
are sensible. Note that if the underlying value function is linear, then only
one iteration is needed to achieve the final solution (the marginal rates of
substitution are constant).
It may be difficult for the decision maker to specify the marginal rates
of substitution directly (or straight away). If this is the case, some auxiliary
procedures may be brought in to assist. One such procedure is presented in
Dyer (1973a). The idea there is to determine (at the point f(xh)) small amounts
of ft and Ii, denoted by Lift and Lif., respectively, such that an increase in
the value of Ii by Lif. is matched for the decision maker by a compensatory
decrease by dft in the value of ft, while the values of all the other objective
functions remain unaltered. In other words, the vectors (II (x h ), ... , fk(xh))Y
and (II (x h), ... , ft(x h) - Lift"'" Ii(x h ) + df.,···, Jk(xh))T are indifferent to
the decision maker. We obtain now
m h,...., Lilt
i,....,~'
~f.

where the approximation becomes arbitrarily exact when the Li-amounts of

change approach O. Note that m~ = 1.
The approximation of marginal rates of substitution is illustrated in Figure
5.2.1. The bold curve is a contour of the value function and the continuous line
its tangent at zh. The marginal rate of substitution at Zh is the negative of
the slope of that tangent. The slope of the approximating broken line is quite
different.

contourofU

Figure 5.2.1. An approximation of the marginal rate of substitution.

In practice, the Li-amounts of change cannot be made arbitrarily small near

0, as emphasized in Sawaragi et al. (1985, pp. 259-260). The reason is that
human beings cannot recognize small changes beyond a certain point. This
threshold of human recognition is called a just noticeable difference. That is
why the marginal rates of substitution are always approximations of the correct
 5.2. Geoffrion-Dyer-Feinberg Method 145

values. An example of the effects of the just noticeable difference is given in

Nakayama (1985a) by illustrating how the solution process may terminate at
a wrong solution. For this reason one may have doubts about the adequacy of
marginal rates of substitution as a means of providing preference information.
They seem to be difficult for the decision maker to specify and their accuracy
is questionable.
However, we must now assume that the marginal rates of substitution are
provided accurately enough. According to the FW method, the maximization
of the linear approximation of U is equivalent to the problem
k T

(5.2.3)
maximize (~-m7V,Ji(Xh») y

subject to yES

with y E R n being the variable. The solution is denoted by yh. The existence
of the optimal solution is ensured by the compactness of S and the continuity
of all the functions.
The search direction is now d h = yh - Xh. Provided that the marginal rates
of substitution are reasonably accurate, the search direction should be usable.
Let us mention that a scaling idea presented in Clinton and Troutt (1988) can
be included in the method. Heterogeneous objective functions can be scaled to
have equal effect in problem (5.2.3) by adjusting the norms of the gradients of
the objective functions with scalar coefficients.
The following problem is to find an appropriate step-size for going in the
search direction. The only variable is the step-size. The decision maker can
be offered objective vectors, where Zi = j;(xh + td h ) for i = 1, ... , k, and t
varies stepwise between 0 and 1 (e.g., t = ~-=-~ where j = 1, ... , P, and P is the
number of the alternative objective vectors to be presented). Another possibility
is to draw the objective values as a function of t, provided no serious scaling
problems exist. An example of the graphical presentation is given in Hwang
and Masud (1979, p. 109). Graphical illustration of the alternative objective
vectors is handled in Chapter 3 of Part III. Note that the alternatives are not
necessarily Pareto optimal. From the information given to the decision maker
(s)he selects the most preferred objective vector and the corresponding value of
t is selected as th. It is obvious that the task of selection becomes more difficult
for the decision maker as the number of objective functions increases.
The opinions of the decision maker and the situation yh = xh are used here
as stopping criteria. Other possible criteria are presented in Hwang and Masud
(1979, pp. 108-110) and Yu (1985, p. 327).
146 Part II - 5. Interactive Methods

5.2.3. Comments

The GDF method can be characterized to be a non ad hoc method. If one

knows the value function, it is easy to specify the marginal rates of substitution
and select the best alternative. The mathematical convergence properties of
the GDF method are closely related to the convergence properties of the FW
method. The convergence of the FW algorithm under the assumptions provided
at the beginning of this section, is proved in Zangwill (1969). However, it must
be kept in mind that the correctness of the marginal rates of substitution and
the step-sizes affects the convergence considerably. If it is assumed that the
answers of the decision maker become ever more exact as the solution process
continues, it is asserted in Geoffrion et al. (1972) that infinite convergence
holds.
More important than infinite convergence in an interactive procedure like
this is the initial rate of convergence, since a satisfactory solution should
be found in a reasonable number of iterations. It is claimed in Geoffrion et
al. (1972) that the error in the objective function values is at least halved at
each of the first H iterations (H is unknown). The convergence becomes slower
near the optimum because of the zig-zag-phenomenon.
The effects of errors in estimating the gradient of the value function are
investigated in Dyer (1974). The result is that even if the answers of the decision
maker are not strictly consistent and the just noticeable difference affects the
marginal rates of substitution, the method is stable and converges (only slower)
under certain assumptions.

5.2.4. Applications and Extensions

The GDF method is applied in Geoffrion et al. (1972) to the operation

of an academic department. Numerical examples are also given, for example,
in Hwang and Masud (1979, pp. 111-121) and Steuer (1986, pp. 377-379). A
time-sharing computer program implementing the GDF algorithm is suggested
in Dyer (1973a). The GDF method is implemented for convex problems by a
so-called projection-relaxation procedure in the objective space in Ferreira and
Machado (1996). An application in water resources allocation is also given. The
GDF method is adapted for continuous equilibrium network design problems
in Friesz (1981).
In Dyer (1972), a method called interactive goal programming is presented.
It is a combination of the GDF method and goal programming. The vector yh
is obtained by the means of weighted goal programming with the marginal rates
of substitution as weights. Some convergence results are also given. The GDF
method and the interactive goal programming method are applied in Jedrze-
jowicz and Rosicka (1983) to multiobjective reliability optimization problems
appearing in multiple classes of system failures.
The GDF method has been a subject of many modifications in the lit-
erature. New versions have been mainly developed to overcome some of the
 5.2. Geoffrion-Dyer-Feinberg Method 147

weaknesses of the GDF method. In Hemming (1981), a simplex-based direction-

finding problem is proposed for MOLP problemA to avoid the specification of
the marginal rates of substitution. It is stressed that the convergence proper-
ties may be impaired, but the cognitive burden placed on the decision maker
is diminished. A revised step-size problem is also presented to produce Pareto
optimal solutions. In addition, the GDF method is modified for MOLP prob-
lems in Winkels and Meilm (1984) so that when determining the step-size at
each iteration, the objective vectors are projected with a so-called efficiency
projection onto the Pareto optimal set. This is done by solving a parametric
linear programming problem.
The GDF method is altered in Rosinger (1981) by constructing a wide fam-
ily of possible inquiry patterns to lead into the determination of the marginal
rates of substitution. The decision maker can choose the form of the inquiry at
each iteration. The convergence of the method is also proved.
The so-called proxy approach is introduced in Oppenheimer (1978). The
value function is no longer approximated linearly. The idea is to give a local
proxy to the value function at each iteration. A sum-of-powers or a sum-of-
exponentials proxy is fitted locally by specifying the parameters connected to
the problem. Now, direction finding and step-size determination problems are
replaced by the maximization of the proxy function. The proxy is not a valid
approximation globally, but when used locally, it gives a higher convergence
rate than the original GDF method. Even this method does not guarantee
the Pareto optimality of the solutions. Oppenheimer does not establish any
systematic procedure for maximizing the proxy function. A method improving
on and utilizing Oppenheimer's ideas is presented in Section 5.3.
Several modifications of the GDF method are presented in Sadagopan and
Ravindran (1986). First, the FW method is replaced by a generalized reduced
gradient method. Then, the role of the decision maker is facilitated by asking
for intervals for the marginal rates of substitution instead of exact values. The
step-size is computed with the help of upper and lower bounds for the objective
functions without the decision maker.
In Musselman and Talavage (1980), the idea of the adaptation is to reduce
the feasible region according to the marginal rates of substitution specified by
the decision maker. Solutions with lower values of the value function than the
current solution are dropped. The method permits sensitivity analysis of the
decision maker's inputs.
Ideas of the GDF method are applied in the interactive integrated approach
for quasiconcave value functions in Al-alvani et al. (1992). A large set of Pareto
optimal solutions is first generated and the form of the underlying implicit
value function is deduced with pairwise comparisons. The gradient of the value
function is also obtained from the comparisons and, thus, marginal rates of
substitution are not needed. Solutions along the search direction are projected
by the weighted Tchebycheff metric. The stopping criterion is based on trade-off
information.
148 Part II - 5. Interactive Methods

Finally, we mention a modification of the GDF method, known as the sub-

gradient GDF method, for non differentiable multiobjective optimization prob-
lems, presented in Miettinen (1994) and Miettinen and Makela (1991, 1993,
1994). The twice continuous differentiability of the objective functions is re-
laxed and they are assumed to be locally Lipschitzian, but the value function
has to still be continuously differentiable. The FW method is replaced by the
subgradient method (see Shor (1985)) in optimizing the approximated value
function.
In addition to being able to handle nondifferentiable functions, the modifi-
cation has another advantage. It produces only Pareto optimal solutions, unlike
the original GDF method. Each calculated solution is set as a reference point
to an order-approximating achievement scalarizing function. We know from
Section 3.5 that the solutions of such achievement functions are always Pareto
optimal. Naturally, additional optimizations increase the computational bur-
den but this is the price to be paid for the certainty that the decision maker
does not have to handle non-Pareto optimal solutions. (A strongly decreasing
value function implies that less is preferred to more in the mind of the decision
maker.)
Some applications solved with the subgradient GDF method are presented
in Miettinen (1994). The subgradient GDF method is used in solving an optimal
control problem concerning an elastic string in Miettinen and Makela (1993)
and continuous casting of steel in Miettinen and Makela (1994).

5.2.5. Concluding Remarks

In the GDF method the decision maker is first given one solution where
(s)he has to specify the marginal rates of substitution. After that the decision
maker must select the most preferred solution from a set of alternatives. Thus,
the ways of interaction are versatile.
In spite of the plausible theoretical foundation of the GDF method, it is not
so convincing and powerful in practice. The most important difficulty for the
decision maker is the determining of the k - 1 marginal rates of substitution
at each iteration. Even more difficult is to give consistent and correct marginal
rates of substitution at every iteration. The difficulties of the decision maker in
determining the marginal rates of substitution are demonstrated, for example,
in Wallen ius (1975) by comparative tests. The same point can be illustrated
by an example from Hemming (1981) where a politician is asked to specify the
exact marginal rate of substitution between unemployment and a decrease of
1 % in the inflation rate.
A drawback ofthe GDF method is that the final solution obtained is not nec-
essarily Pareto optimal. Naturally, it can always be projected onto the Pareto
optimal set with an auxiliary problem. A more serious objection is that when
several alternatives are given to the decision maker from which to select the
step-size, it is likely that many of them are not Pareto optimal. They can also
be projected onto the Pareto optimal set before presentation to the decision
 5.3. Sequential Proxy Optimization Technique 149

maker, but this necessitates extra effort. The projection may be done, for in-
stance, by lexicographic ordering or by the means presented in Section 2.10 of
Part I. The use of achievement functions is demonstrated in the subgradient
GDF method. The weakness in the projection is that the computational burden
increases. It is for the analyst and the decision maker to decide which of the
two shortcomings is less inconvenient.
Theoretically, the Pareto optimality of the final solution is guaranteed if the
value function is strongly decreasing (by Theorem 2.6.2 of Part I). In any case,
marginal rates of substitution are crucial in approximating the value function,
and for many decision makers they are difficult and troublesome to specify.
For many people it is easier to think of desired changes in the objective
function values than to specify indifference relations. This may, especially, be
the case if the objective vector at which the marginal rates of substitution are
to be specified is not particularly desirable. Then it may be frustrating to think
of indifferent solutions instead of the improvements sought.
The Frank-Wolfe gradient method has been selected as the maximization
algorithm for its fast initial convergence. In some cases, other gradient-based
methods may be more appropriate. For example, the subgradient method is
employed in the subgradient GDF method.
There are a lot of assumptions that the problem to be solved must satisfy
in order the method to work and converge. Several sufficient conditions on the
decision maker's preferences are presented in Sawaragi et al. (1985, pp. 258-
259) to guarantee the differentiability and the concavity of the value function.
Even these conditions are not very easy to check. For more critical discussion
concerning the GDF method, see Sawaragi et al. (1985, pp.257-261).

5.3. Sequential Proxy Optimization Technique

Like the two previous methods, the sequential proxy optimization technique
(SPOT), presented in Sakawa (1982), is based on the idea of maximizing the
decision maker's underlying value function, which is once again assumed to be
known implicitly. SPOT includes some properties of the ISWT and the GDF
methods, and that is why we describe it here briefly.

5.3.1. Introduction

As in the two interactive methods presented thus far, the search direction
in SPOT is obtained by approximating locally the gradient of the underlying
value function, and the step-size is determined according to the preferences
of the decision maker. Here, both marginal rates of substitution and trade-off
rates are used in approximating the value function.
 150 Part II - 5. Interactive Methods

It is assumed that
1. The underlying value function U: R k -t R exists and is implicitly known
to the decision maker. In addition, U is a continuously differentiable,
strongly decreasing and concave function on the subset of Z where the
points are Pareto optimal.
2. The objective and the constraint functions are convex and twice contin-
uously differentiable.
3. The feasible region S is compact and convex (and there exist some upper
bounds for the c-constraint problem so that the solution is finite).
4. The assumptions in Theorem 3.2.13 are satisfied.
The c-constraint problem is used to generate Pareto optimal solutions. The
solution of c-constraint problem (3.2.1) is denoted by xh. It is assumed to be
unique so that Pareto optimality is guaranteed. Throughout this section it is
assumed that all the upper bound constraints are active at the optimum. (If
this is not the case, then the upper bounds must be slightly modified.) Then,
Ii (xh) = cJ for all j = 1, ... ,k, j f. £. The optimal value of h, that is, h(x h ), is
denoted by zf. It is also assumed that all the Karush-Kuhn-Tucker multipliers
associated with the active constraints are strictly positive. The conditions of
Theorem 3.2.13 are assumed to be satisfied so that trade-off rate information
can be obtained from the Karush-Kuhn-Tucker multipliers.
Here, the value function is not maximized in form (4.1.1) as before. Instead,
the set of feasible alternatives is restricted to the Pareto optimal set. According
to the assumption above stating that h(x h ) = cJ for all j = 1, ... , k, j i- £,
we have a new formulation:

(5.3.1) ..
maximize U( cI"'"
hhhh
c£_I' Z£, C(+I"'" Ckh) .
No constraints are needed here since the formulation includes the original con-
straints. The optimization is now carried out in the objective space Rk-l, where
the upper bounds cJ are the variables.
It is proved in Sakawa (1982) that the new function is concave with respect
to those E E R k-l for which the upper bound constraints are all active. Sakawa
also claims that the partial derivative of (5.3.1) with respect to c:J, j = 1, ... , k,
j i- £, is equivalent to 8~j;) (m1j - ..\1j ) , where m1 j is the marginal rate of substi-
tution between hand fj at xh (obtained from the decision maker, see Section
5.2) and ..\1j is the partial trade-off rate between hand hat xh (obtained from
the Karush-Kuhn-Tucker multipliers, see Sections 3.2 and 5.1).
Because it was assumed that the value function is strongly decreasing, we
know that 8~j;) < 0 and we can divide by it. We denote now

Llc:J = -(m1j - ..\1j)

for j = 1, ... ,k, j i- £, and it represents the direction of steepest ascent of the
value function (5.3.1) at the current point xh for j i- £. According to Sakawa,
the £th component of the direction is
 5.3. Sequential Proxy Optimization Technique 151

k k
L.xfj(mfj - .xfj ) = L -.xtilcJ
j=1 j=1
j#f j#l

denoted by ilzf.
After obtaining the search direction, we have to find the step-size t which
in theory maximizes the function

(5.3.2)

The step-size could be determined as earlier by presenting different objective

vectors to the decision maker with different values of t and by letting the
decision maker choose the most preferred one. The problem with alternative
objective vectors of this kind is that they are not necessarily Pareto optimal.
However, the step-size is not determined by asking the decision maker's
opinion. Nonetheless, different alternatives are generated with different step-
sizes. Their Pareto optimality is guaranteed by solving the c-constraint problem
with the upper bounds cj + tilcj for j = 1, ... , k, j "I f, still assuming that
the constraints are active. This increases the number of calculations since the
c-constraint problem must be solved for several values of t.
The best alternative is selected employing local proxy preference functions
p (in the same spirit as the proxy approach presented in Section 5.2 in con-
nection with the GDF method). The proxy function replaces function (5.3.2)
and the alternative with the highest proxy function value is selected for the
continuation.
According to the preference structure of the decision maker a sum-of-
exponentials, sum-of-powers or sum-of-Iogarithms proxy function of the form
k k k
-L aie-w;fi(X) , - L ai(ni + fi(X»"'i or L ai In(ni - fi(X»,
i=1 i=1 i=1

respectively, is used. The constants ai, Wi, ni and D:i are used to tune the proxy
functions so that they represent the current problem and the preferences of
the decision maker better, and they are derived from the marginal rates of
substitution; see, for example, Sakawa (1982) and Sakawa and Seo (1982b)
for further details. This kind of proxy function is very restrictive globally but
reasonable when assumed locally.

5.3.2. SPOT Algorithm

We can now present the basic ideas of the SPOT algorithm.

(1) Choose a reference function It and upper bounds el E R k - 1 for which
all the constraints of the c-constraint problem are active. Set h = 1.
152 Part II - 5. Interactive Methods

(2) Solve the current (active) c-constraint problem for eh to obtain a solu-
tion xh.
(3) Denote the Pareto optimal objective vector corresponding to xh by Zh
and the corresponding Karush-Kuhn-Tucker multipliers by j , j = Ai
1, ... , k, j I- E.
(4) Ask the decision maker for the marginal rates of substitution mi j for
j = 1, ... , k, j I- E, at Xh. Test the consistency of the marginal rates of
substitution and ask the decision maker to respecify them if necessary.
(5) If Imij - A:jl < (), where () is a prespecified positive tolerance, then stop
with xh as the final solution. Otherwise, determine the components
..:1cJ, j I- E, of the search direction vector.
(6) Select an appropriate form of the proxy function and calculate its pa-
rameters. If the obtained proxy function is not strongly decreasing and
concave, then ask the decision maker to specify new marginal rates of
substitution.
(7) Determine the step-size by solving the c-constraint problem with the
upper bounds cJ + t..:1cJ, j = 1, ... , k, j I- E, for different values of t.
Denote the optimal value of the objective function by z~(t). A step-size
t h maximizing the proxy function is selected. If the new objective vector
(cf + t h ..:1cf , ... , z~ (th), ... , c~ + t h ..:1c~) T is preferred to zh, denote the
corresponding decision vector by xh+1, set h = h + 1 and go to step
(3). If the decision maker prefers zh to the new solution, reduce t h to
be !th, ith , ... until an improvement is achieved.
The maximum of the proxy function is determined by altering the step-size
t, calculating the corresponding Pareto optimal solution and searching for three
t values, t 1, t h and t2 so that tl < t h < t2 and p( tt} < p( t h ) > p( t2), where p is
the proxy function. When the condition above is satisfied, the local maximum
of the proxy function pet) is in the neighbourhood of th.
Under assumptions 1-4 (in Subsection 5.3.1), the optimality condition for
problem (5.3.1) at eh is that the gradient equals zero at that point. This means
that mi
j = Aii
for j = 1, ... k, j I- E. This is the background of the absolute
value checking at step (4) (see also (2.8.1) in Part I).

5.3.3. Comments

The consistency of the marginal rates of substitution is checked because it

is important for the successful convergence of the algorithm. The consistency
at a single point is tested by the chain rule and by limiting the discrepancy (the
formula is given in Sakawa (1982» by a given tolerance level. The consistency
at successive points is tested by checking the concavity and monotonicity of
the proxy function (the proxy function must fulfill the same assumptions as
the value function). A theorem giving conditions for different types of proxy
functions is presented in Sakawa (1982).
 5.3. Sequential Proxy Optimization Technique 153

To ensure the convergence of the algorithm it must, at each iteration, be

checked that a sufficient improvement is obtained. If the decision maker prefers
the new solution, the procedure may continue. Otherwise, a new step-size must
be estimated.
It is remarked in Sakawa (1982) that the SPOT algorithm is nothing but
a feasible direction method as for the convergence rate. The convergence can
be demonstrated by the convergence of the modified feasible direction method.
For this statement to be true, an ideal (i.e., consistent with correct answers)
decision maker must be assumed.
SPOT can be classified among methods of a non ad hoc nature. If the value
function is known, the marginal rates of substitution can be computed directly
and the step-size is easy to calculate.

5.3.4. Applications and Extensions

The functioning of the SPOT algorithm is demonstrated in Sakawa (1982)

by an academic example. It is shown that even though the marginal rates of
substitution are only approximations, this does not necessarily worsen the re-
sults remarkably. A problem concerning industrial pollution in Osaka City in
Japan is solved by SPOT in Sakawa and Seo (1980, 1982a, b). The problem is
defined as a large-scale problem in Sakawa and Seo (1980) and a dual decom-
position method is used to solve the c:-constraint problems.
A fuzzy SPOT is presented in Sakawa and Yano (1985). The decision maker
is assumed to assess the marginal rates of substitution in a fuzzy form. In
Sakawa and Mori (1983), a new method for nonconvex problems is proposed,
where the weighted Tchebycheff problem is used to generate Pareto optimal
solutions instead of the c:-constraint method, and trade-off rates are not used.
A method related to the preceding one is presented in Sakawa and Mori (1984).
The difference is a penalty scalarizing function used in generating Pareto op-
timal solutions (see Section 3.5). This method is also applicable to nonconvex
problems.

5.3.5. Concluding Remarks

Ideas from several methods are combined in SPOT and several concepts
are utilized. As far as the role of the decision maker is concerned, (s)he is only
required to determine the marginal rates of substitution. Difficulties related
to this determination were mentioned in Section 5.2 and they are still valid.
However, the consistency of the marginal rates of substitution in SPOT is
even more important than in the GDF method. This is a very demanding
requirement.
A positive feature of SPOT when compared to the GDF method is that only
Pareto optimal solutions are handled. Because the multiobjective optimization
154 Part II - 5. Interactive Methods

problem was assumed to be convex, globally Pareto optimal solutions are ob-
tained. The burden on the decision maker is decreased by employing a proxy
function when selecting the step-size.
Many assumptions are set to guarantee the proper functioning of the al-
gorithm. Some of these are quite difficult to check in practice (see concluding
remarks concerning the GDF method in Subsection 5.2.5).

5.4. Tchebycheff Method

The Tchebycheffmethod, proposed in Steuer (1986, pp. 419-450) and Steuer

and Choo (1983) and refined in Steuer (1989a), is an interactive weighting vec-
tor space reduction method. Originally, it was called the interactive weighted
Tchebycheff procedure. A notable difference when compared to the methods de-
scribed thus far is that a value function is not used in the Tchebycheff method.
In addition, the role of the decision maker is different and somewhat simpler.
Here, we introduce the Tchebycheff algorithm according to the refined version
but modified for minimization problems.

5.4.1. Introduction

The Tchebycheff method has been designed to be user-friendly for the deci-
sion maker, and, thus, complicated information is not required. To start with,
a utopian objective vector below the ideal objective vector is established. Then
the distance from the utopian objective vector to the feasible objective region,
measured by a weighted Tchebycheff metric, is minimized. Different solutions
are obtained with different weighting vectors in the metric, as introduced in
Section 3.4. The solution space is reduced by working with sequences of smaller
and smaller subsets of the weighting vector space. Thus, the idea is to develop a
sequence of progressively smaller subsets of the Pareto optimal set until a final
solution is located. At each iteration, different alternative objective vectors are
presented to the decision maker and (s)he is asked to select the most preferred
of them. The feasible region is then reduced and alternatives from the reduced
space are presented to the decision maker for selection.
Contrary to the previous interactive methods for multiobjective optimiza-
tion, the Tchebycheff method does not presume many assumptions regarding
the problem to be solved. It is assumed that
1. Less is preferred to more by the decision maker.
2. The objective functions are bounded (from below) over the feasible region
S.
In what follows we assume that the global ideal objective vector and, thus,
the global utopian objective vector are known, and we can leave the absolute
value signs from the metrics. The metric to be used for measuring the distances
 5.4. Tchebycheff Method 155

to a utopian objective vector is the weighted Tchebycheff metric (see Section

3.4). That is, the function to be minimized is
(5.4.1)

where w E W = {w E Rk I 0 < Wi < 1, E~=l Wi = I}. We have a family

of metrics since w E W can vary widely. This nondifferentiable problem can
be solved as a differentiable weighted Tchebycheff problem (3.4.3) (where the
ideal objective vector is replaced by the utopian objective vector).
According to Theorem 3.4.5, we know that every Pareto optimal solution of
any multiobjective optimization problem can be found by solving the weighted
Tchebycheff problem with z**. The negative aspect with this problem is that
some ofthe solutions may be weakly Pareto optimal. This weakness was handled
in Subsection 3.4.5. Producing weakly Pareto optimal solutions is overcome in
the Tchebycheff method by formulating the distance minimization problem as
a lexicographic weighted Tchebycheff problem:
k
lex minimize . max [Wi(fi(X) - zi*)], L(fi(X) - zr)
(5.4.2) l=l,... ,k i=l

subject to xES.

The functioning of problem (5.4.2) is described in Figure 5.4.1 by a problem

with two objective functions. The bold line illustrates the Pareto optimal set.
The weighted Tchebycheff problem has L-shaped contours (the thin continu-
ous line) whose vertices lie along the line emanating from z** in the direction
(I/W1, l/w2,' .. , l/wk)' When minimizing the distance, a contour is determined
which is closest to z** and intersects Z. If this problem does not have a unique
solution, that is, there are several feasible points on the optimal contour in-
tersecting Z, then some of them may not be Pareto optimal. In practice, the
uniqueness is usually difficult to check, and, to be on the safe side, the fol-
lowing step must be taken. In this case, the sum term is minimized subject to
the obtained points to determine which of them is closest to z** according to
the Lrmetric (the dotted line). Thus a Pareto optimal solution (see Theorem
3.4.1) is obtained.
The following theorems formulate the connection between the lexicographic
weighted Tchebycheff problem and Pareto optimal solutions.

Theorem 5.4.1. The solution oflexicographic weighted Tchebycheffproblem

(5.4.2) is Pareto optimal.

Proof. Let x* E S be a solution of problem (5.4.2). Let us assume that it is not

Pareto optimal. In this case there exists some XO E S such that f;(xO) ~ f;(x*)
for all i = 1, ... , k and h (XO) < h (x*) for at least one j. This and the positivity
of the weights implies that wi(fi(XO) - zr) ~ Wi (fi(X*) - zt*) for every i and
thus maxi[Ji(XO) - zrJ ~ max;[Ji(x*) - zt*J.
156 Part II - 5. Interactive Methods

.........

•
utopian ...... second minimization
vector

" "
..... z1

first minimization

Figure 5.4.1. Lexicographic weighted Tchebycheff problem.

On the other hand, h(xO) - zt* ~ h(x*) - zt* for all i = 1, ... , k and at
least one of the inequalities is strict. That is why we have E~=1 (fi(XO) - zr) <
E~=l (fi(X*) - zr)· Here we have a contradiction with x* being a solution of
(5.4.2). Thus, x* is Pareto optimaL 0

Theorem 5.4.2. Let x* E S be Pareto optimaL Then there exists a weighting

vector 0 < w E R k such that x* is a unique solution of lexicographic weighted
Tchebycheff problem (5.4.2).

Proof. Let x* E S be Pareto optimal. Let us assume that there exists no

weighting vector w > 0 such that x* is a unique solution of problem (5.4.2).
We know that h(x) > zr for all i = 1, ... , k and for all xES. That is why
we can choose for all i = 1, ... ,k

Wi
1
= h(x*) - zt*
(
£; h(x*)1- zr )-1
k

If x* is not a unique solution of (5.4.2), there exists another point XO E S

that is a solution of this lexicographic weighted Tchebycheff problem. This
implies that XO must be a solution of the weighted Tchebycheff problem. This
means that
 5.4. Tchebycheff Method 157

i!:?~.~k [hex*)I_ zr (~ hex*)I_ zr) -1 (fi(XO) - z;*) 1

-::; i!:?~,k [ f;(x*)I_ zr (~ h(x*)l_ zt* ) -1 (fi(X*) - z;*) 1
k 1 )-1
= ~
(
fi(X*) - z:*

After simplifying the expression we have

for every i = 1, ... , k. Because X* is Pareto optimal, we must have fi(XO) =

h(x*) for all i. In other words, the weighted Tchebycheff problem, and thus
also the lexicographic weighted Tchebycheff problem, has a unique solution.
o
An alternative proof of Theorems 5.4.1 and 5.4.2 is given in Steuer (1986,
p. 445) and Steuer and Choo (1983). Now we know that the lexicographic
weighted Tchebycheff problem produces Pareto optimal solutions and any
Pareto optimal solution can be found.
In the Tchebycheff method, different Pareto optimal solutions are obtained
by altering the weighting vector. At each iteration h, the weighting vector space
W h = {w h E Rk I[hz < w th < u z' h
L....z==l w t = I} is reduced
",k h
~
to W h +1 , where
W h +1 C W h • With a sequence of progressively smaller subsets of the weighting
vector space, a sequence of smaller subsets of the Pareto optimal set is sampled.
At the first iteration, a sample of the whole Pareto optimal set is generated
by solving the lexicographic weighted Tchebycheff problem with well dispersed
z:
weighting vectors from W = WI (with = 0 and u} = 1). The reduction of Wh
is done by tightening the upper and the lower bounds for the weighting vectors.
Let zh be the objective vector that the decision maker chooses from the sample
at the iteration h and let w h be the corresponding weighting vector in problem
(5.4.2). Now a concentrated group of weighting vectors centred around w h is
formed. In this way, a sample of Pareto optimal solutions centred about zh is
obtained. It is advised to use normalized objective functions in the calculations.
The number of the alternative objective vectors to be presented to the de-
cision maker is denoted by P. The number is usually specified by the decision
maker. It may be fixed or different at each iteration. The algorithm becomes
more reliable, if as many alternatives as possible can be evaluated effectively
at each iteration. Human capabilities are yet limited, and some kind of a com-
promise is desirable.
When reducing the weighting vector space at each iteration, a reduction
factor r is needed. The larger the reduction factor is, the faster the weighting
vector space is reduced and the smaller are the decision maker's possibilities
158 Part II - 5. Interactive Methods

for making errors and changing her or his mind concerning her or his desires
during the process. The correct selection of r is thus important. It is suggested
in Steuer (1986) and Steuer and Choo (1983) that
(1/ p)l/k ~ r ~ v1/(H-l),

where v is the final interval length of the weighting vectors with lk ~ v ~ 23k'
H is the number of iterations to be carried out and ~ stands for 'approximately
equal or less.'

5.4.2. Tchebycheff Algorithm

We can now present the main features of the Tchebycheff algorithm.

(1) Specify values for the set size P(~ k), a reduction factor r < 1 and
an approximation for the number of iterations H(~ k). Set If = 0 and
ui = 1 for all i = 1, .. " k. Construct the utopian objective vector. Set
h=l.
(2) Form the weighting vector space W h = {w h E R k I If < wf <
h
Ui'
",k h -
L...i=l wi -
I} .
(3) Generate 2P dispersed weighting vectors w h E Wh.
(4) Solve lexicographic weighted Tchebycheff problem (5.4.2) for each of
the 2P weighting vectors.
(5) Present the P most different of the resulting objective vectors to the
decision maker and let her or him choose the most preferred among
them, denoting it by zh.
(6) If h = H go to step (8). Otherwise, modify, if necessary, the weight-
ing vector corresponding to Zh such that if problem (5.4.2) was solved
again, zh would be a uniquely generated solution at the vertex of the
intersecting new contour.
(7) Specify l~+1 and u~+1 for the reduced weighting vector space W h +1 ,
set h = h + 1 and go to step (2).
(8) The final solution is xh corresponding to zh.
Dispersed weighting vectors are generated from W h in step (3). In practice,
this can be realized by generating randomly a large set (e.g., 50k) of weighting
vectors. Then the vectors are filtered (see Steuer (1986, pp. 311-326)) or clus-
tered. The clustering is practical since subroutines for it are available in many
subroutine libraries (such as IMSL). While we want to obtain 2P well dispersed
weighting vectors, we form 2P clusters and choose one candidate from each of
them either arbitrarily or near the centre.
Computationally, the following algorithm can be used to obtain random
weighting vectors in W h . We omit the index h for clarity.
For i = 1, ... , k set
 5.4. Tchebycheff Method 159

where ra is a random number between 0 and 1. Calculate the sums we

2::7=1 we;, l = 2::7=1[; and u = 2::7=1 Ui' Finally, set for each i = 1, ... , k
we; +we (li _ wei)
I[ - if we> 1,
-we
W· - { 1- we
•- wei + ---CUi - wei) if we < 1,
u-we
we; otherwise.
The lexicographic weighted Tchebycheff problem is solved 2P times for 2P
weighting vectors (instead of P) to overcome the case that the same or a very
similar solution is obtained with different weighting vectors. The 2P (or less)
objective vectors are again filtered or clustered to obtain the P most different.
In this way, it is guaranteed that P different objective vectors can be presented
to the decision maker. For graphical illustration of the alternatives, see Chapter
3 in Part III.
The modification of the weighting vector in step (6) can be conducted by

(t.
setting

wf = j;(xh)1_ z7* !i(Xh; _ Z7*) -I

for all i = 1, ... ,k as in the proof of Theorem 5.4.2. The modification is relevant
because the weighting vector space is reduced with respect to w h . It is useful
to have an unbiased basis for the reduction.
Several possibilities for reducing the weighting vector space have been sug-
gested. It is proposed in Steuer (1986) to set

if w'-'• 2 --, rh <0

1 Wi
'f h
+ T 2:: 1 ,
rh

otherwise,

where rh means raising r to the power h. In Steuer (1989a), an auxiliary scalar

w is determined so that the ratio of the volumes of W h +1 and W h is r. Then
w is used in the reduction instead of the term !rh.
The predetermined number of iterations is not necessarily conclusive. The
decision maker can stop iterating when (s)he obtains a satisfactory solution or
continue the solution process longer if necessary.

5.4.3. Comments

It is suggested in Steuer (1986, 1989a) that the sampling of the Pareto

optimal set works in the most unbiased way if the ranges of the objective
function values over the Pareto optimal set are approximately the same. This
can be accomplished by re-scaling the objective functions in a way similar to
that presented in Subsection 2.4.3 of Part I, when necessary. It is advisable
160 Part II - 5. Interactive Methods

to use the scaling only in the calculations and present the alternatives to the
decision maker in the original form. More suggestions for modifications of the
algorithm are presented in Steuer (1989a).
The convergence rate of the Tchebycheff method is very difficult to estab-
lish. It is stressed in Steuer (1989a) that the Tchebycheff method is able to
converge to any Pareto optimal solution. The reduction factor r is compre-
hended as a convergence factor because it determines how fast the reduction
takes place. The weighting vector space is reduced until a solution is obtained
that is satisfactory enough to be a final solution (see Steuer and Choo (1983)).
The Tchebycheff method can be characterized as a non ad hoc method. If
the value function is known, it is easy to select from the set of P alternatives
the one maximizing the value function.
We do not here go into details of the alternative version of the Tchebycheff
method. We only mention that the possibility of getting weakly Pareto optimal
solutions may be overcome by using augmented weighted Tchebycheff problem
(3.4.5) (see Figure 3.4.2). This means that properly Pareto optimal solutions
are handled instead of Pareto optimal ones (see Theorem 3.4.6). In this way, the
lexicographic optimization is avoided, but the Tchebycheff algorithm is more
complicated in other ways. For example, the determination of the correct value
for the augmentation parameter p brings additional problems. It is proved in
Steuer (1986, pp. 440-444) and Steuer and Choo (1983) that the augmented
weighted Tchebycheff problem can be used to characterize Pareto optimal solu-
tions if the feasible region is finite or all the constraints are linear. A numerical
illustration of the algorithm is presented in Steuer (1986, pp. 468-472).
Implementing the Tchebycheff method in a spreadsheet (Excel) environ-
ment is suggested in Steuer (1997). The Tchebycheff method in its augmented
form is applied in Wood et al. (1982) to water allocation problems of a river
basin and in Silverman et al. (1988) to manpower supply forecasting. The aug-
mented method form is also used in Agrell et al. (1998) when solving an MOLP
problem of reservoir management. In Olson (1993), the Tchebycheff method is
applied to a sausage blending problem and in Kaliszewski (1987) it is proposed
that modified weighted Tchebycheff problem (3.4.6) is used to minimize the
distances in the Tchebycheff method.

5.4.4. Concluding Remarks

A positive feature of the Tchebycheff method is that the role of the decision
maker is quite easy to understand. (S)he does not need to realize new concepts
or specify numerical answers as, for example, in the ISWT and the GDF meth-
ods. All (s)he has to do is to compare several alternative objective vectors and
select the most preferred one. The ease of the comparison depends on the mag-
nitude of P and on the number of objective functions. The personal capabilities
of the decision makers also play an important role. It is also positive that all
the alternatives are Pareto optimal.
 5.5. Step Method 161

The flexibility of the method is reduced by the fact that the discarded parts
of the weighting vector space cannot be restored if the decision maker changes
her or his mind. Thus, some consistency is required.
The weakness of the Tchebycheff method is that a great deal of calculation
is needed at each iteration and many of the results are discarded. For large
and complex problems, where the evaluation of the values of the objective
functions may be laborious, the Tchebycheff method is not a realistic choice.
On the other hand, it is possible to utilize parallel computing since all the
lexicographic problems can be solved independently.
Although no absolute superiority can be attributed, it is worth mentioning
that the Tchebycheff method performed best in the comparative evaluation of
four methods (the ZW, the SWT, the Tchebycheff and the GUESS methods) in
Buchanan and Daellenbach (1987) (see Subsection 1.2.3 of Part III). However, a
difficulty was encountered in comprehending the information provided. The test
example had only three objective functions and six alternatives were presented
at each iteration. And the cognitive burden only becomes larger when the
number of the objective functions is increased.

5.5. Step Method

The step method (STEM), presented in Benayoun et a1. (1971), contains el-
ements somewhat similar to the Tchebycheff method, but is based on a different
idea. STEM is one of the first interactive methods developed for multiobjec-
tive optimization problems. It was originally designed for the maximization of
MOLP problems but can be extended for nonlinear problems, as described,
for example, in Eschenauer et al. (1990b) and Sawaragi et al. (1985, pp. 268-
269). It can be considered to aspire at finding satisfactory solutions instead
of optimizing an underlying value function. We describe the method for the
minimization of nonlinear problems.

5.5.1. Introduction

It is assumed in STEM that at a certain Pareto optimal objective vector

the decision maker can indicate both functions that have acceptable values and
those whose values are too high. The latter can be said to be unacceptable. The
decision maker is now assumed to allow the values of some acceptable objective
functions to increase so that the unacceptable functions can have lower values.
In other words, (s)he must give up a little in the value(s) of some objective
function(s) Ii (i E P) in order to improve the values of some other objective
functions Ii (i E [<) such that [> U [< = {I, ... , k}.
STEM uses the weighted Tchebycheff problem (3.4.2) to generate new so-
lutions. The ideal objective vector z* is used as a reference point in the calcu-
lations. According to Theorem 3.4.2 the solutions are weakly Pareto optimal.
162 Part II - 5. Interactive Methods

It is assumed that
1. Less is preferred to more by the decision maker.
2. The objective functions are bounded over the feasible region S.
Information concerning the ranges of the Pareto optimal set is needed in
determining the weighting vector for the metric. The idea is to make the scales
of all the objective functions similar with the help of the weighting coefficients.
The nadir objective vector znad is approximated from the payoff table as
explained in Subsection 2.4.2 of Part I. Thus, the maximal element of the
column i is called ziad. The weighting vector is calculated by the formula

Wi = k
ei
,i = 1, ... , k,
Ej=l ej

where for every i = 1, ... , k

1 znad - z~
e·• ---• •z~ z!,ad
••
as suggested in Eschenauer et al. (1990b), or
z!,ad - z'!<
e· -• •
•- max [Iziadl, Iztl]
as suggested in Vanderpooten and Vincke (1989). (The denominators are not
allowed to be zero.) The weight is larger for those objective functions that are
far from their ideal objective vector component.

5.5.2. STEM Algorithm

The basic phases of the STEM algorithm are the following:

(1) Calculate the ideal and the nadir objective vectors and the weighting
coefficients. Set h = 1. Solve weighted Tchebycheff problem (3.4.2) with
the calculated weights. Denote the solution by xh E S and the corre-
sponding objective vector by zh E Z.
(2) Ask the decision maker to classify the objective functions at zh into
satisfactory J> and unsatisfactory ones J<. If the latter class is empty,
go to step (4). Otherwise, ask the decision maker to specify relaxed
upper bounds €i for the satisfactory objective functions.
(3) Solve problem (5.5.1), where the upper bounds are taken into account.
Denote the solution by xh+l E S and the corresponding objective vector
by Zh+1 E Z. Set h = h + 1. Go to step (2).
(4) Stop. The final solution is xh.
In the first step the distance between the ideal objective vector and the
feasible objective region is minimized by the weighted Tchebycheff metric (the
weighting coefficients specified as above). The solution obtained is presented to
 5.5. Step Method 163

the decision maker. Then the decision maker is asked to specify those objective
function(s) whose value(s) (s)he is willing to relax (Le., weaken) to decrease
the values of some other objective functions. (S)he must also determine the
amount(s) of acceptable relaxation. Ways of helping the decision maker in this
phase are presented in Benayoun et al. (1971).
The feasible region is restricted according to the information of the decision
maker and the weights of the relaxed objective functions are set equal to zero,
that is Wi = 0 for i E J>. Then a new distance minimization problem
minimize . max [wilfi(X) -
t=l, ... ,k
z71]
(5.5.1) subject to hex) ~ Ci for all i E [>,
hex) ~ h(x h ) for all i E [<,
xES

is solved. The first new constraint set allows the relaxed (acceptable) objective
function values to increase up till the specified level and the second new con-
straint set makes sure that the unsatisfactory objective function values do not
increase, that is, get worse. The procedure continues until the decision maker
does not want to change any component of the current objective vector. If the
decision maker is not satisfied with any of the components, then the procedure
must also be stopped. In this case, STEM fails to find a satisfactory solution.
Different versions of the method vary in the formulation of the constraint
set. In some versions, a new constraint set is generated at every iteration and
in some other versions new constraints are included to accompany the old ones.
In the latter model the decision maker must be somewhat consistent in her or
his actions because it is not possible to withdraw the restrictions set on the
feasible region.

5.5.3. Comments

STEM does not assume the existence of an underlying value function. Even
if one were available, it would not help in answering the questions. Thus STEM
can be characterized as an ad hoc method. Naturally, nothing can be said
about the convergence of STEM with respect to a value function. However, the
developers of the method mention that the algorithm produces a final solution
fast if the new constraints constructed during the solution process become
ineligible for further relaxations.
A linear numerical application example of STEM is given in Hwang and
Masud (1979, pp. 174-182). The properties of the solution set of STEM are
studied in Crama (1983). A so-called exterior branching algorithm is presented
in Aubin and Naslund (1972). It is another kind of extension of STEM into
nonlinear problems. There are several differences when compared with the orig-
inal method. For example, the decision maker does not need to specify any
amounts of change and an implicit value function is assumed to exist. Some
164 Part II - 5. Interactive Methods

extensions and modifications of STEM are also mentioned in Chankong and

Haimes (1983b, p. 329).

5.5.4. Concluding Remarks

Because we are moving around the (weakly) Pareto optimal set, a decrement
in some objective function values can be achieved only by paying the price of
an increment in some other objective function values. The idea of specifying
objective functions whose values should be decreased or can be increased seems
quite simple and appealing. However, it may be difficult to estimate appropriate
amounts of increment that would allow the desired amount of improvement in
those functions whose values should be decreased. In other words, the control
of the solution is somewhat indirect. On the other hand, a positive feature is
that the information handled is easy to understand. No complicated concepts
are introduced to the decision maker.
According to the results presented in Section 3.4, the solutions of STEM
are not necessarily Pareto optimal, but weakly Pareto optimal solutions may
be obtained. It must also be kept in mind that the global ideal objective vector
has to be known.
STEM was the first interactive method to be based on the classification
idea. Numerous other methods adapting this idea in one way or the other have
appeared since. In what follows, we present several methods where the decision
maker can specify both the amounts of relaxation and desirable aspiration
levels. In this way the decision maker can control the solution process in a
more direct way than in STEM.

5.6. Reference Point Method

As its name suggests, the reference point method, presented in Wierzbicki

(1980a, b, 1981, 1982), is based on a reference point of aspiration levels. The
reference point is a feasible or infeasible point in the objective space which is
reasonable or desirable to the decision maker. The reference point is used to
derive achievement scalarizing functions having minimal solutions at weakly, c;-
properly or Pareto optimal points as introduced in Section 3.5. In this method,
generating Pareto optimal solutions is based on reference points, not on value
functions or weighting vectors. No specific assumptions are set on the problem
to be solved. The reference point idea has been utilized in several methods
in different ways. Wierzbicki's reference point method was among the first of
them.
 5.6. Reference Point Method 165

5.6.1. Introduction

The basic idea behind the reference point method is to reconsider how
decision makers make decisions. It is doubted in Wierzbicki (1980a, b) that
individuals make everyday decisions by maximizing a certain value function.
Instead, Wierzbicki claims that decision makers want to attain certain aspi-
ration levels (e.g., when making purchases according to a shopping list). He
suggests that, while thousands of consumers may behave on the average as
if they were maximizing a value function, no individual behaves in that way.
The basic idea is satisficing (introduced in Section 2.6 of Part I) rather than
optimizing. In addition, reference points are intuitive and easy for the decision
maker to specify and their consistency is not an essential requirement.
Classifying the objective functions into acceptable and unacceptable ones
(at a current objective vector) was mentioned in connection with STEM. Spec-
ifying a reference point can be considered a way of classifying the objective
functions. If the aspiration level is lower than the current objective value, that
objective function is currently unacceptable, and if the aspiration level is equal
to or higher than the current objective value, that function is acceptable. The
difference here is that the reference point can be infeasible in every component.
In other words, where the set of acceptable objective functions is empty, the
reference point-based approach can still be utilized. Naturally, this does not
mean that all the objective values could be decreased but a different solution
can be generated.
Further information concerning the matters addressed in this section can be
found in Wierzbicki (1977, 1980b, 1981, 1982, 1986a, b). By a reference point
method we here mean that of Wierzbicki's. The reference point method relies
heavily on the properties of achievement functions, which were dealt with in
Section 3.5. Of particular interest are Corollary 3.5.6 and Theorem 3.5.7. As
far as the preference structure of the decision maker is concerned, it is assumed
that
1. Less is preferred to more by the decision maker.

5.6.2. Reference Point Algorithm

The interactive multiobjective optimization technique of Wierzbicki is very

simple and practical. Before the solution process starts, some information is
given to the decision maker about the problem. If possible, the ideal objective
vector and the (approximated) nadir objective vector are presented to illus-
trate the ranges of the Pareto optimal set. Another possibility is to minimize
and maximize the objective functions individually in the feasible region (if it
is bounded). Both decision variable and objective values are presented. An
appropriate form for the achievement function must also be selected.
The basic steps of the reference point method are the following:
(1) Present information about the problem to the decision maker. Set h = 1.
166 Part II - 5. Interactive Methods

(2) Ask the decision maker to specify a reference point zh E Rk (an aspi-
ration level for every objective function).
(3) Minimize the achievement function and obtain a (weakly, c-properly or)
Pareto optimal solution xh and the corresponding zh. Present zh to the
decision maker.
(4) Calculate a number of k other (weakly, c-properly or) Pareto optimal
solutions by minimizing the achievement function with perturbed ref-
erence points
z(i) = Zh +dhei ,
where d h == IIzh - zhll and e i is the ith unit vector for i = 1, . .. , k.
(5) Present the alternatives to the decision maker. If (s)he finds one of the
k + 1 solutions satisfactory, the corresponding Xh is the final solution.
Otherwise, ask the decision maker to specify a new reference point Zh+l .
Set h = h + 1 and go to step (3).
The reason for writing the words weakly or c-properly in parentheses in the
algorithm is that it depends on the achievement function selected whether the
solutions are weakly, c-properly or Pareto optimal.
The advantage of perturbing the reference point in step (4) is that the de-
cision maker gets a better conception of the possible solutions. If the reference
point is far from the Pareto optimal set, the decision maker gets a wider de-
scription of the Pareto optimal set and if the reference point is near the Pareto
optimal set, then a finer description of the Pareto optimal set is given. The
effects of the perturbation and close and distant reference points are illustrated
in Figure 5.6.1.

z1

Figure 5.6.1. Altering the reference points.

 5.6. Reference Point Method 167

5.6.3. Comments

As to the infinite convergence of the algorithm, the following result is stated

in Wierzbicki (1980b).

Theorem 5.6.1. If the solutions of the achievement function in the algorithm

are unique and if the minimal value of liz - ill subject to Pareto optimal
objective vectors is equal to the minimal value of the achievement function 8.
subject to Z for i f/. Z + Ri, then for any metric in Rk, the solution procedure
is convergent. In other words, limh~oo Ilzh - zh+111 = o.

Proof. See references in Wierzbicki (1980b).

A modification of the algorithm guaranteeing the convergence is presented in

Wierzbicki (1980b).
The reference point method can be characterized as an ad hoc method or a
method having both non ad hoc and ad hoc features. Alternatives are easy to
compare if the value function is known. On the other hand, a reference point
cannot be directly defined with the help of the value function. However, it is
possible to test whether a new reference point has a higher value function value
than the earlier solutions.
A different way of generating new reference points is suggested in Wierzbicki
(1997b). It is a way of realizing the idea of a reference ball where a set of
additional reference points in a ball of a fixed radius centered on the current
solution is produced.
An appendix to the reference point method is suggested in Wierzbicki
(1997b). After the decision maker has found a final solution (s)he can check
whether more satisfactory solutions exist by a so-called outranking trials
method. In the spirit of outranking methods of (discrete) multiattribute de-
cision analysis, the decision maker is asked to specify preference, indifference
and veto thresholds (see Subsection 5.9.1) for each objective function. Differ-
ent states of outranking relations are established and a sequential questioning
procedure is gone through with the decision maker to check whether there
exist objective vectors whose components outrank the current final solution.
This procedure may involve a lot of questions. The convergence of many other
interactive methods can be investigated by the outranking trials method, as
well.

5.6.4. Implementation

A software family called DIDAS (Dynamic Interactive Decision Analysis

and Support) has been developed on the basis of the reference point ideas of
Wierzbicki. The nonlinear version of DIDAS has been created and developed
in several phases. For example, the International Institute for Applied Systems
168 Part II - 5. Interactive Methods

Analysis (IIASA) in Austria and the Warsaw Technical University have been
involved. The latest version is called IAC-DIDASN++. There is a lot of lit-
erature describing the various phases in the development work (see Granat et
al. (1994a, b), Grauer (1983a, b), Grauer et al. (1984), Kreglewski (1989), Kre-
glewski et al. (1987, 1991), Lewandowski and Grauer (1982), Lewandowski et
al. (1987) and Rogowski et al. (1987».
DIDAS is a dynamic decision support system which aims at helping to
achieve better decisions. The ideology has been extended from the reference
point method with reservation levels. Reservation levels Zi are objective func-
tion values the user wants to avoid. For the objective functions to be minimized
they must be above the aspiration levels forming the reference point z. In DI-
DAS, the user is asked to specify both aspiration and reservation levels for each
objective function. The achievement function has to be reformulated to take
the reservation levels into account. Several achievement functions have been
suggested in different versions of the system.
The user can easily obtain different Pareto optimal solutions by changing
the aspiration levels and the reservation levels. The objective functions are
scaled and the user is assumed to specify aspiration levels between the ideal
objective vector and the nadir objective vector. In this setting, the user can
implicitly attach more importance to attaining a particular aspiration level
by placing it near the ideal objective value. In that case, the corresponding
objective function is weighted stronger in the achievement function.
We give an example of achievement functions, including both aspiration
and reservation levels. If all the objective functions are to be minimized, an
order-approximating achievement function to be maximized can be of the form

zt
where are components of the ideal objective vector, p > 0 is an augmentation
term and
. Zi - z;
V = mm -v--_-.
i=l ..... k Zi - Zi

Achievement functions may be computationally complicated. Further, they are

typically nondifferentiable. However, their simplified and differentiable coun-
terparts are generally used (see, e.g., Granat et al. (1994a) and Kreglewski et
al. (1991».
In further developing DIDAS, attention has been paid to computational
efficiency in both achievement functions and their nonlinear single objective
solvers. Gradient-based solvers are efficient and robust enough to be employed
in interactive decision support systems. However, it is not advisable to ask for
gradient information from the user of the system. Firstly, the formulation of the
 5.6. Reference Point Method 169

derivatives is a time-consuming and laborious task, and secondly, errors and

mistakes are likely to occur. (Mistakes have been found to be a main reason for
the failure of nonlinear optimization methods in convergence.) The difficulties
can be overcome with symbolic differentiation. This is briefly handled in Kre-
glewski (1989). One more alternative is to use automatic differentiation (see,
e.g., RaIl (1981». However, no results of doing this have been reported.
DIDAS is general and can thus handle objective functions needing to be
minimized, maximized or stabilized (Le., the objective function should have a
value as close to the given level as possible). Different objective function types
imply changes in the achievement function used (see, for example, Granat et
al. (1994a».

5.6.5. Applications and Extensions

The reference point method is applied to econometric models in Olbrisch

(1986). Some experiences in applying DIDAS to macroeconomics planning are
reported in Grauer et al. (1984). DIDAS is used in empirical tests in Bischoff
(1985) to experiment with different scalarizing functions. A problem of de-
termining the optimal temperature in a greenhouse is solved by DIDAS in
Udink ten Cate (1985). In Starn et al. (1992), DIDAS is used in analysing
the acid rain problem in Europe. A trajectory-oriented extension of DIDAS is
described and applied in Lewandowski et al. (1985a, b). Three applications of
IAC-DIDASN++ for engineering design are reported in Wierzbicki and Granat
(1997). They handle the design of a spur gear transmission unit, ship navigation
support and automatic control.
The reference point idea is modified in Mocci and Primicerio (1997) to
better handle non convex problems and avoid local optima. At each iteration,
the achievement function is minimized in a reduced feasible region determined
by the decision maker. This modified method is applied to a problem of ring
network design.
The reference point method is generalized for several decision makers or
several reference points in Song and Cheng (1988) and Vetschera (1991a). The
reference point method is also essential in a group decision support system,
described in Vetschera (1991b), where each group member uses the reference
point method.
A so-called combined procedure combining the Tchebycheff method and the
reference point method is introduced in Steuer et al. (1993). There the deci-
sion maker is asked to specify both reference points and the most satisfactory
solution among the alternatives produced by the means of the two methods.
So-called reference sets, extensions of reference points, are the basis of an in-
teractive procedure described in Skulimowski (1996).
An extension of the reference point method, called the preemptive reference
point method, is introduced in Ogryczak (1997a). The approach formulates ref-
erence point problems in the form of goal programming. Instead of considering
170 Part II - 5. Interactive Methods

all the deviations to be equally important (which is the case in the reference
point method) predefined priorities between the goals are also handled. The
reference point method is modified for problems with homogeneous and anony-
mous objective functions in Ogryczak (1997b). Here, anonymity stands for sym-
metry with respect to permutations of the objective functions.

5.6.6. Concluding Remarks

Wierzbicki's reference point method is quite easy for the decision maker to
understand. The decision maker only has to specify appropriate aspiration lev-
els and compare objective vectors. What has been said about the comparison of
alternatives in connection with the previous methods is also valid here. The so-
lutions are weakly, c:-properly or Pareto optimal depending on the achievement
function employed.
The freedom of the decision maker has both positive and negative aspects.
The decision maker can direct the solution process and is free to change her or
his mind during the process. However, the convergence is not necessarily fast
if the decision maker is not purposeful. There is no clear strategy to produce
the final solution since the method does not help the decision maker to find
improved solutions.
Wierzbicki's method can be regarded as a generalization of goal program-
ming. Aspiration levels are central in both methods, but unlike goal program-
ming Wierzbicki's method is able to handle both feasible and infeasible aspi-
ration levels.
Methods based on reference points are widely regarded efficient for the
solution of practical problems. They are easy to understand and to implement.
Further, they do not necessitate consistency from the decision maker. One can
say that controlling a method with reference points is a more direct and a more
explicit way than, for example, with weighting coefficients.

5.7. GUESS Method

The GUESS method is a simple interactive method related to the reference

point method. The method is also sometimes called a naive method and it
is presented in Buchanan (1997). It will be referred to in Subsection 1.2.3
of Part III when describing method comparisons available in the literature.
For this reason we present the method here briefly (modified for minimization
problems).
 5.7. GUESS Method 171

5.7.1. Introduction

The GUESS method does not involve any special assumptions. The only
requirement is that the ideal objective vector z* and the nadir objective vector
znad are available. Thus, it is assumed that

1. Less is preferred to more by the decision maker.

2. The objective functions are bounded over the feasible region S.
The method proceeds as follows. The decision maker specifies a reference
point (or a guess) zh and a solution with equal proportional achievements is
generated. Then the decision maker specifies a new reference point and the iter-
ation continues until the decision maker is satisfied with the solution produced.
The search procedure is not assisted in any other way.
The scales of the objective functions are normalized with denominators
zr ad - zt
for every i = 1, ... , k. The general idea is to maximize the minimum
weighted deviation from the nadir objective vector. Thus, the idea is oppo-
site to, for example, the weighted Tchebycheff problem where the maximum
weighted deviation from the ideal objective vector is minimized.
We can put the same reasoning in other words. We can say that the objective
functions are rescaled so that they all have the range [0,1]. This means that
each objective function !i(X) is replaced by a normalized function

zr -hex)
ad .
for all z = 1, ... , k.
zr zi
a
d
-

Let us once again emphasize that the global ideal objective vector and the nadir
objective vector are assumed to be known.
The weighted max-min problem to be solved is

maximize ._
ml' n [~zrad - !i(X)]
. --"--n-ad-:--'---''-*'':'"
(5.7.1) ,_l, ... ,k W, zi - zi

subject to xES,
where the weighting coefficients Wi, i = 1, ... , k, are positive and the denomi-
nators must not equal zero.
We have the following result.

Theorem 5.7.1. The solution of weighted max-min problem (5.7.1) is weakly

Pareto optimal.

Proof. Let x* E S be a solution of the weighted max-min problem. Let us

suppose that x* is not weakly Pareto optimal. In this case, there exists a
point xES such that hex) < h(x*) for every i = 1, ... , k. This means that,
zr ad -hex) > zr ad - !i(X*) for all i. While we have Wi > 0 and zr ad - zi > 0,
we can write
172 Part II - 5. Interactive Methods

This implies that

min [..!.. -,zi;..;.a_d.-....::f;..;.i..:...(X-'-)]

i=l ..... k Wi zr ad - zt
Thus, x* cannot be a solution of the weighted max-min problem. This contra-
diction completes the proof and x* is weakly Pareto optimal. 0

The weighting coefficients are not any positive numbers whatsoever, but
normalized aspiration levels. In other words, we have
z!lad - z!t
w? =
zr
Z d
a -
'for all i = 1, ... ,k.
zt
With the specified weighting coefficients we can write the problem to be
solved in the form

maximize . mm
. [zrad-li(x)]
nad - zi-h
(5.7.2) .=l ..... k Zi

subject to xES.
Notice that the aspiration levels specified by the decision maker have to be
strictly lower than the nadir objective vector, that is, zh < znad. If all the
objective functions are differentiable, problem (5.7.2) can be written in a dif-
ferentiable form with the help of an additional variable, whereas the nondiffer-
entiable formulation can be solved with appropriate single objective optimizers.
We can prove that any Pareto optimal solution can be found with problem
(5.7.2).

Theorem 5.7.2. If x* E S is Pareto optimal, then it is a solution of problem

(5.7.2) with z = f(x*).

Proof. Let x* E S be Pareto optimal and let us suppose that it is not a solution
of (5.7.2) with z = f(x*). In this case there exist another XO E S such that

mm
. [z!lad
'
- Ji(X > m m [z!lad
'
-1i(X*)]
O
)] •
= 1.
i=l, ... ,k zr ad - Ji(X*) i=l, ... ,k zr ad -li(x*)
This means that fi(XO) < h(x*) for every i = 1, ... , k which is a contradiction
with the Pareto optimality of x*. In other words, x* must be a solution of
(5.7.2). 0

According to Theorems 5.7.1 and 5.7.2 we know that all the solutions gener-
ated are weakly Pareto optimal and any Pareto optimal solution can be found.
 5.7. GUESS Method 173

5.7.2. GUESS Algorithm

The GUESS method has five basic steps.

(1) Calculate the ideal objective vector and the nadir objective vector and
present them to the decision maker. Set h = l.
(2) Let the decision maker specify upper or lower bounds to the objective
functions if (s)he so desires. Update problem (5.7.2), if necessary.
(3) Ask the decision maker to specify a reference point zh E Rk between
the ideal and the nadir objective vectors.
(4) Solve problem (5.7.2) and obtain a weakly Pareto optimal solution xh.
Present the corresponding objective vector zh to the decision maker.
(5) If the decision maker is satisfied with Zh, set xh as a final solution and
stop. Otherwise, set h = h + 1 and go to step (2).
In step (2) the specification of upper or lower bounds means adding con-
straints to problem (5.7.2). Nevertheless, the components of the ideal or the
nadir objective vector needed in the calculation elsewhere are not changed or
affected.
The only stopping rule is the satisfaction of the decision maker. No guidance
is given to the decision maker in setting new aspiration levels. This is typical
of many reference point-based methods.

5.7.3. Comments

The GUESS method is based on trial and error. The decision maker can
examine what kind of an effect her or his input has on the solution obtained and
then modify the input, if necessary. The system does not provide any additional
or supporting information about the problem to be solved.
As long as no additional constraints are included in the problem, the com-
ponents of the solution obtained are in equal proportion with the components
of the reference point specified. In other words, when the solution obtained
and the corresponding reference point are normalized, the quotients of their
component are the same for each component. The reason for this behaviour is
that the reference point is contained in the weighting vector.
The GUESS method is an ad hoc method. The existence of a value function
would not help in determining new reference points or upper or lower bounds
for the objective functions.
An interesting practical observation is mentioned in Buchanan (1997).
Namely, decision makers are easily satisfied if there is a small difference between
the reference point and the solution obtained. Somehow they feel a need to be
satisfied when they have almost achieved what they wanted. In this case they
may stop iterating 'too early.' The decision maker is naturally allowed to stop
the solution process if the solution really is satisfactory. But, the coincidence
of setting the reference point near an attainable solution may unnecessarily
increase the decision maker's satisfaction.
174 Part II - 5. Interactive Methods

5.7.4. Concluding Remarks

The GUESS method is simple to use and does not set any specific assump-
tions on the behaviour or the preference structure of the decision maker. The
decision maker can change her or his mind since no consistency is required.
The only information required from the decision maker is a reference point and
possible upper and lower bounds.
The method has been compared to several other interactive methods in dif-
ferent comparative evaluations (to be described in Subsection 1.2.3 of Part III).
It has been received relatively well in the experiments reported. The reasons
may be its Simplicity and flexibility.
The optional upper or lower bounds specified by the decision maker are
not checked in any way in the method. Inappropriate lower bounds may lead
into solutions that are not weakly Pareto optimal. In other words, additional
constraints may invalidate the result of Theorem 5.7.1. This can be avoided,
for example, by allowing only upper bounds.
The weakness of the GUESS method is its heavy reliance on the availabil-
ity of the nadir objective vector. As mentioned in Subsection 2.4.2 of Part I,
the nadir objective vector is not easy to determine and it is usually only an
approximation.

5.B. Satisficing Trade-Off Method

The satisficing trade-off method (STOM), presented in Nakayama (1989,

1995), Nakayama and Furukawa (1985), Nakayama and Sawaragi (1984) and
Nakayama et al. (1986), is based on ideas similar to the reference point method
of Wierzbicki and the GUESS method. The differentiating factor is the trade-
off information utilized. The method is here presented according to Nakayama
(1995).

5.S.1. Introduction

STOM originates from classification and aspiration levels. It is based on

satisficing decision making, as can be deduced from its name.
The functioning of STOM is the following. After a weakly or a properly
Pareto optimal solution has been obtained by optimizing a scalarizing function,
it is presented to the decision maker. On the basis of this information (s)he is
asked to classify the objective functions into three classes. The classes are the
unacceptable objective functions whose values (s)he wants to improve (1<),
the acceptable objective functions whose values (s)he agrees to relax (impair)
(J» and the acceptable objective functions whose values (s)he accepts as they
are (1=) (such that [< u [> U [= = {1, ... , k}). Trade-off rate information is
utilized so that the decision maker only has to specify aspiration levels for the
 5.8. Satisficing Trade-Off Method 175

functions in J<. Next, a modified scalarizing function is minimized and the

decision maker is asked to classify the objective functions at the new solution.
The theoretical derivation of the method is based on the three requirements
for scalarizing functions presented at the beginning of this part in Chapter 1.
As was then stated, there does not exist a scalarizing function that can satisfy
all three requirements. Nonetheless, a rather promising approach is to use the
weighted Tchebycheff problem that can find any Pareto optimal solution. Hence
it and its augmented variant are used in STOM.
Different forms of scalarizing functions have been suggested for use in
STOM. In the original formulation the weighting coefficients are set as

(5.8.1) wf = -h 1 for every i = 1, ... , k,

Zi - zt*
where zh is a reference point and z** is a utopian objective vector so that
zh > z**, and the scalarizing function to be minimized is once again (5.4.1).
If weakly Pareto optimal solutions are to be avoided, the scalarizing function
used is
k
(5.8.2) . max [Wf(fi(X) - zt*) 1+ P
1.=l, ... ,k
L wf hex),
.
t=l

where p is some sufficiently small positive scalar, for example, of the order
10- 6 . Both these scalarizing functions presume that the ideal objective vector
and, thus, the utopian objective vector are known globally. However, if some
objective function fJ is not bounded from below in S, then some small scalar
value can be selected as zj*.
If the problem is bounded, then the solutions obtained by function (5.4.1)
are guaranteed to be weakly Pareto optimal (see Theorem 3.4.2) and every
Pareto optimal solution can be found (see Theorem 3.4.5). Further, it is proved
in Nakayama (1985a) and Sawaragi et al. (1985, pp. 271-272) that the solution
obtained is satisficing (Le., Ji(X*) ::; zf for all i = 1, ... , k) if the reference point
is feasible and weighting coefficients (5.8.1) are employed. For function (5.8.2)
all the solutions are properly Pareto optimal and any properly Pareto optimal
solution can be found. Even though the formulation slightly differs from (3.4.5),
the results of Theorem 3.4.6 are still valid. Unfortunately, function (5.8.2) does
not satisfy the third requirement concerning satisficing decision making (see
Nakayama (1985a)).
Other forms of weighting coefficients can also be used. The selection affects
the results obtained. This is demonstrated in Nakayama (1995). The reference
point method-type achievement functions can be used as well. This means that
the utopian objective vector is replaced by the reference point.
Both the scalarizing functions mentioned are nondifferentiable but they can
be written in a differentiable form assuming the differentiability of the functions
involved. This is carried out by introducing a scalar variable 0: as in (3.4.3).
176 Part II - 5. Interactive Methods

In what follows, we refer to the differentiable form where all the objective
functions have been transformed into constraints.
As mentioned in Subsection 3.4.4, trade-off rate information can be obtained
with the help of differentiable formulation (3.4.3). Both weighting coefficients
and Karush-Kuhn-Thcker multipliers are then utilized. That is why it must be
assumed that
1. Less is preferred to more by the decision maker.
2. The objective and the constraint functions are twice continuously differ-
entiable.
The availability of trade-off rates also necessitates the fulfillment of other
assumptions mentioned in Subsection 3.4.4. They are parallel to those in The-
orem 3.2.13; see also Yano and Sakawa (1987). This fact has not earlier been
sufficiently emphasized when introducing the method.

5.8.2. STOM Algorithm

Let us now write down the steps of the algorithm.

(1) Calculate the utopian objective vector z**. Set h = 1.
(2) Ask the decision maker to specify a reference point zh E R k such that
if > zi* for every i = 1, ... , k.
(3) Minimize the scalarizing function used. Denote the solution by xh. Let
the corresponding objective vector be zh. Present it to the decision
maker.
(4) Ask the decision maker to classify the objective functions into the classes
J<, J> and J=. If J< = 0, go to step (6). Otherwise, ask the decision
maker to specify new aspiration levels i~+l for the functions in J<. Set
z~+1 = ft(x h ) for i E r.
(5) Let A" E Rk be the Karush-Kuhn-Tucker multipliers connected to Xh.
Use automatic trade-off to obtain new levels (upper bounds) if+! for
the functions in J>. Set h = h + 1 and go to step (3).
(6) Stop with the final solution Xh.
It is naturally possible that the decision maker will also specify new levels
for those objective functions whose values (s)he agrees to relax (i.e., increase).
But particularly when the number of objective functions is great, the decision
maker may appreciate the automatic trade-off feature when (s)he does not have
to specify new aspiration levels for all the functions at each iteration. Naturally,
the decision maker can modify the calculated aspiration levels if they are not
agreeable.
As long as trade-off rates are obtainable from the Karush-Kuhn-Thcker
multipliers and the weighting coefficients, the burden set on the decision maker
can be decreased by employing automatic trade-off in specifying the aspira-
tion levels (upper bounds) for the functions to be relaxed. They are derived
 5.8. Satisficing Trade-Off Method 177

from sensitivity analysis on the basis of staying in the Pareto optimal set (see
Nakayama (1991b, 1992a, 1995». We set for each i E J>

-h+l
Zi = ! i (X h) + N()"~ +1 )w h "~ ()..hj + P)Wjh(!i (h)
X-
-h+l) ,
Zi
• P 'jE/<

where N is the number of the objective functions in the class J>. If no aug-
mentation term is used in the scalarizing function, we set P = 0 in the formula
above. Automatic trade-off increases all the objective functions in J> in the
equal proportion to ()..7 + p)w7· If the amounts of change are large or the prob-
lem is nonlinear, the aspiration levels produced by automatic trade-off may
not be large enough to allow the desired improvements to the other objective
functions (see Nakayama (1992b».

5.8.3. Comments

If the problem is linear or quadratic, we can go even further than the au-
tomatic trade-off. In this case parametric optimization is used in generating
so-called exact trade-off. This means that we can calculate exactly how much
the objective function values must be relaxed in order to stay in the Pareto
optimal set. Thus, we get a new Pareto optimal solution without having to
re-optimize the scalarizing function (see Nakayama (1991b, 1992a, b».
Trade-off information can also be used to check the feasibility of the refer-
ence point specified by the decision maker. If it is not feasible, the number of
minimizations of the scalarizing function can be reduced by directly specify-
ing higher aspiration levels (remember that satisficing solutions are obtained
when the reference point is feasible in scalarizing function (5.4.1». See details
in Nakayama (1985a, 1989), Nakayama and FUrukawa (1985) and Nakayama
and Sawaragi (1984).
Trade-off information is valuable even if some Karush-Kuhn-Tucker multi-
pliers are equal to zero. For example, if all the Karush-Kuhn-Tucker multipliers
of the functions to be relaxed equal zero, we know that it is not possible to
improve the desired objective function values with this classification. In other
words, the functions to be relaxed cannot compensate for the improvement de-
sired. The reason is that the objective functions to be relaxed are positively
affected by other objective function(s) to be improved (see Nakayama (1995».
Note that STOM can be used even in the absence of trade-off rate informa-
tion. This may be the case if all the differentiability and the regularity assump-
tions are not satisfied. If trade-off rates are not used, no special assumptions
need to be set on the problem to be solved. In this form STOM is almost the
same as the GUESS method - only the achievement function used is different.
Because no specific assumptions are set on the underlying value function,
convergence results based on it are not available. Even if a value function
existed, it could not be directly used to determine the functions to be decreased
178 Part II - 5. Interactive Methods

and increased or the amounts of change. Thus the method is characterized as

an ad hoc method. On the other hand, one must remember that the aim of the
method is particularly in satisficing rather than optimizing some value function.
It is pointed out in Nakayama (1995) and Nakayama et al. (1995) that the
roles of the objective and the constraint functions can easily be interchanged
when the nondifferentiable scalarizing function is solved in differentiable form
(3.4.3). This is carried out by adding constant multipliers {3i to the artificial
variable a in each additional constraint. The objective function Ii becomes a
constraint function by changing the value of (3i from one to zero. For further
details, see Nakayama et al. (1995). The idea of interchanging the roles of the
functions is also handled in Subsections 5.10.3 and 5.12.7.

5.8.4. Implementation

A STOM implementation has been carried out in Bulgaria. The software is

called MONP-16j see Vassilev et al. (1990). This program has been developed
for nonlinear multiobjective optimization problems. The system suggests new
aspiration levels for all the objective functions, and the user can freely change
them. The feasibility of the aspiration levels is checked in the sense of linear
approximations. If it is impossible to satisfy all the specified aspiration levels,
the user can either modify the levels or go ahead and optimize anyway. In
the latter case the solution will be weakly Pareto optimal but not satisficing.
MONP-16 can handle objective functions both to be minimized and to be
maximized.

5.8.5. Applications and Extensions

Some theoretical specifications concerning the STOM algorithm are pre-

sented in Nakayama and Furukawa (1985). The method is also applied to the
aseismic design of a tower-pier system for a long span suspension bridge. Soft-
ware implementing STOM for interactive construction accuracy control sys-
tems of cable-stayed bridges is introduced and applications are described in
Nakayama et al. (1995, 1997). A linear diet problem and the erection of cable-
stayed bridges are mentioned as STOM applications in Nakayama (1994). The
latter problem is also handled in Nakayama (1995). An application to a water
quality control problem of a river basin is presented in Nakayama (1985a) and
Nakayama and Sawaragi (1984). In Olson (1993), the method is employed to
solve a sausage blending problem and, in Nakayama et al. (1986), to solve a
blending problem of industrial plastic materials. STOM is adapted for linear
fractional objective functions in Nakayama (1991a) with an application con-
cerning material blending in cement production. A blending problem in feed
formulation for live stock is described in Nakayama (1995) as well as an inter-
active support system for bond trading. In addition, STOM is applied in a diet
planning problem in Mitani and Nakayama (1997).
 5.9. Light Beam Search 179

Sensitivity analysis of the STOM algorithm for linear problems is investi-

gated in Nakayama (1989). Further, the relationship between STOM and fuzzy
mathematical programming is handled in Nakayama (1995).
The ISWT method and STOM are combined,in the method suggested in
Wang (1992). The decision maker can choose the form of the interactive ques-
tions (surrogate worth values or new aspiration levels) at each iteration. The
Pareto optimality of the solutions produced by the c:-constraint method is guar-
anteed by solving auxiliary problem (2.1O.1) of Part I.

5.8.6. Concluding Remarks

STOM contains identical elements with STEM, the reference point method
and the GUESS method. Therefore, the comments given there are not repeated
here. The role of the decision maker is easy to understand. STOM requires even
less input from the decision maker than the above-mentioned methods because
only a part of the aspiration levels need to be given. The solutions obtained are
properly Pareto optimal or weakly Pareto optimal depending on the scalarizing
function used.
As said before, in practice, classifying the objective functions into three
classes and specifying the amounts of increment and decrement for their values
is a subset of specifying a new reference point. A new reference point is implic-
itly formed. Either the new aspiration levels are larger, smaller, or the same as
in the current solution. Thus the same outcome can be obtained with different
reasoning. A positive differentiating feature in STOM when compared to other
classification-based methods is the automatic or exact trade-off. This decreases
the amount of information inquired from the decision maker. STOM is in a
sense opposite to STEM. In STOM, only desired improvements are specified,
whereas only amounts of relaxation are used in STEM.
Because the method is based on satisficing decision making, the decision
maker can freely search for a satisficing solution and change her or his mind,
if necessary. No convergence based on value functions has even been intended.

5.9. Light Beam Search

Light beam search, described in Jaszkiewicz and Slowinski (1994, 1995),

combines the reference point idea and tools of multiattribute decision analysis.
That is why it is an interesting method for inclusion here to represent how the
benefits of different problem solving areas can be put to use. Here we modify
the original method for minimization problems.
180 Part II - 5. Interactive Methods

5.9.1. Introduction

The basic setting in the light beam search is identical to the reference point
method of Wierzbicki in the spirit of satisficing decision making. The achieve-
ment function to be minimized is function (3.5.3) where weighting coefficients
are used only in the maximum part. They take into account the ideal and the
nadir objective values. This achievement function means that c-properly Pareto
optimal solutions are generated. The reference point is here assumed to be an
infeasible objective vector.
It is assumed that
1. Less is preferred to more by the decision maker.
2. The objective and the constraint functions are continuously differen-
tiable.
3. The objective functions are bounded over the feasible region S.
4. None of the objective functions is more important than all the others
together.
Assumption 3 is needed in order to have the ideal and the nadir objec-
tive vectors available. The other assumptions are related to the generation of
alternative solutions.
In the light beam search it is acknowledged that reference points provide a
practical and an easy way for the decision maker to direct the solution process.
However, the learning process of the decision maker is supported better if the
decision maker receives additional information about the Pareto optimal set
at each iteration. This means that other solutions in the neighbourhood of
the current solution (based on the reference point) are displayed. Thus far,
the motivation is the same as in the reference point method. But what if the
comparison of even a small number of alternative solutions is difficult for the
decision maker? Or what if all the alternatives provided are indifferent to the
decision maker? In such cases the decision maker may even stop the solution
process and never get as far as the satisfactory solutions.
An attempt is made to avoid frustration on the part of the decision maker
in the light beam search by the help of concepts used in multiattribute decision
analysis and particularly in ELECTRE methods (see, for example, Roy (1990)
and Vincke (1992, pp. 56-69)). The idea is to establish outranking relations
between alternatives. It is said that the alternative Zl outranks the alternative
Z2, denoted by ZlSZ2, if Zl is at least as good as Z2. In the light beam search,
additional alternatives near the current solution are generated so that they
outrank the current one. Incomparable or indifferent alternatives are not shown
to the decision maker.
To be able to compare alternatives and to define outranking relations, we
need several thresholds from the decision maker. Assumption 4 is related to
this. Because of the just noticeable difference or for some other reasons it is
not always possible for the decision maker to distinguish between different
alternatives. This means that there is an interval where indifference prevails.
 5.9. Light Beam Search 181

For this reason the decision maker is asked to provide indifference thresholds
qi for each objective function (i = 1, ... , k). In fact the thresholds should be
functions of the objective values, that is qi(Zi), but in the light beam search
they are assumed to provide only local information and are thus constants.
The line between indifference and preference does not have to be sharp
either. The hesitation between indifference and preference can be expressed by
preference thresholds Pi for i = 1, ... ,k. Applying the same reasoning as above,
we assume here that Pi is not a function of the values of the objective function
but constant. In addition, we must have Pi ~ qi ~ 0 for i = 1, ... , k.
Given these thresholds we can distinguish three preference relations between
pairs of alternative objective vectors (Zl and Z2) for each component, that is,
each objective function. We can say that as far as the ith components (i =
1, ... ,k) of the two objective vectors are concerned,
Zl and Z2 are indifferent if Izl- zll ~ qi
2 1
Zl is weakly preferred to Z2 if qi < Zi - Zi < Pi
Zl is preferred to z2 if zl- zl ~ Pi·
One more type of threshold, namely a veto threshold Vi for i = 1, ... , k can
be defined. It prevents a good performance in some components from compen-
sating for poor values on some other components. As earlier, we assume the
threshold to be constant and have the relation Vi ~ Pi for i = 1, ... , k. In this
case z2 cannot be preferred to zl if Z[ - zl ~ Vi.
We can now define outranking relations on the basis of for how many com-
ponents indifference, weak preference or preference is valid or preference cannot
be valid. Let us compare the objective vector of the current iteration zh and
some other objective vector z. Below, #i denotes the number of components,
that is, objective functions, for which the condition mentioned holds. We define
ms(z,zh) as #i where z is indifferent, weakly preferred or preferred to zh,
mq(zh,z) as #i where zh is weakly preferred to z,
mp(zh,z) as #i where zh is preferred to z,
mv(zh, z) as #i where z cannot be preferred to zh.
The outranking relations are defined according to the numbers above. If
the decision maker has specified all the thresholds, that is the indifference, the
preference and the veto thresholds, it is proposed in Jaszkiewicz and Slowinski
(1994, 1995) that
zSZh if mv(zh,z) = 0, mp(zh,z) ~ 1 and mq(zh,z) +mp(zh,z) ~ ms(z,zh)
be defined. This definition must be modified if no veto thresholds are available.
In this case
182 Part II - 5. Interactive Methods

If no preference thresholds have been specified, the definition is

zS zh if mv(zh, z) =0 and mq(zh, z) ~ 1.
Finally, if only indifference thresholds are available, the outranking relation is
defined by

5.9.2. Light Beam Algorithm

Let us now outline the light beam algorithm.

(1) If the decision maker wants to or can specify the best and the worst
values for each objective function, denote the corresponding vectors by
z* and znad, respectively. Alternatively calculate z* and znad. Set h = 1
and the reference point Zh = z*. Initialize the set of saved solutions
as B = 0. Ask the decision maker to specify an indifference threshold
for every objective function. If desired, (s)he can also specify preference
and veto thresholds for them.
(2) Calculate a current solution xh and the corresponding zh by minimizing
the achievement function with zh.
(3) Present zh to the decision maker. Calculate k Pareto optimal character-
istic neighbours of zh and present them as well to the decision maker.
If the decision maker wants to see alternatives between any two of the
k + 1 alternatives displayed, set their difference as a search direction,
take different steps in this direction and project them onto the Pareto
optimal set before showing them to the decision maker. If the decision
maker wants to save the current solution zh, set B = B U {Zh}.
(4) If desired, the decision maker can revise the thresholds. If this is the
case, set zh = zh+l, h = h + 1 and go then to step (3). Otherwise,
if the decision maker wants to give another reference point, denote it
by zh+1, set h = h + 1 and go to step (2). If, on the other hand, the
decision maker wants to select one of the alternatives displayed or one
solution in B as a current solution, set it as zh+1, set h = h + 1 and
go to step (3). Finally, if one of the alternatives is satisfactory, set the
corresponding decision vector to be xh+1 , set h = h + 1 and go to step
(5).
(5) Stop with xh as the final solution.
The option of saving desirable solutions in set B increases the flexibility
of the method. The decision maker can explore different directions and select
the best among different trials. The possibility of having a look at solutions
between any two alternatives is related to the GDF method. The same idea
will also be handled in Section 5.12. The alternative solutions in step (3) can
be projected by minimizing the achievement function.
Let us consider how the Pareto optimal characteristic neighbours of zh
are generated. The thresholds specified by the decision maker are needed in
 5.9. Light Beam Search 183

defining outranking relations as described in Subsection 5.9.1. Characteristic

neighbours are new alternative objective functions that outrank the current
solution. The number of characteristic neighbours z(i) is equal to the number
of objective functions. For each i = 1, ... , k, the neighbour z(i) is the point
in the outranking neighbourhood of zh with maximal distance from zh in the
direction where the value of the ith component locally improves most.
The neighbours are determined by projecting the gradient of one objective
function at a time onto the linear approximation of those constraints that are
active in zh with gradient projection methods (this necessitates differentiability
and, thus, assumption 2); see Jaszkiewicz and Slowinski (1994, 1995) for details.
The feasible direction in the objective space offering the greatest improvement
for the ith component of zh is denoted by d i . The outranking characteristic
neighbour in that direction is obtained with the problem
maximize 0:

(5.9.1) subjectto z(i)Szh where z(i)=zh+o:d i ,

0: ~ O.
After solving problem (5.9.1) for each i = 1, ... , k, we have k characteristic
neighbours. Each z(i) is projected onto the Pareto optimal set before being
displayed to the decision maker. This can be carried out, for example, by min-
imizing the achievement function with each neighbour as a reference point.

5.9.3. Comments

The idea of the light beam search is analogous to projecting a focused beam
of light from the reference point onto the Pareto optimal set. The lighted part
of the Pareto optimal set changes if the location of the spotlight, that is, the
reference point or the point of interest in the Pareto optimal set are changed.
This connection explains the name of the method. An implementation of the
light beam search is available from its developers (see Section 2.2 in Part III).
The light beam search can be characterized as an ad hoc method. If a value
function were available, it could not directly determine new reference points. It
could, however, be used in comparing the set of alternatives. Yet, the thresholds
are important in the method and they must come from the decision maker.
This method combines elements of multiobjective optimization and multi-
attribute decision analysis in an interesting way. An extension is suggested in
Wierzbicki (1997b), where both aspiration levels forming a reference point and
reservation levels (to be avoided) are used. In this case the reference point still
determines the source of light but the reservation levels are used to generate a
cone of light. Some convergence ideas are put forward in Wierzbicki (1997b) as
well.
184 Part II - 5. Interactive Methods

5.9.4. Concluding Remarks

The light beam search is a rather versatile solution method where the deci-
sion maker can specify reference points, compare a set of alternatives and affect
the set of alternatives in different ways. Thresholds are used to try to make
sure that the alternatives generated are not worse than the current solution.
In addition, they are different enough to be compared and comparable on the
whole. This should decrease the burden on the decision maker.
Specifying different thresholds is a new aspect when compared to the meth-
ods presented earlier. This may be demanding for the decision maker. Anyway,
it is positive that the thresholds are not assumed to be global but can be al-
tered at any time. In other words, outranking relations based on the threshold
values are only used as local preference models in the neighbourhood of the
current solution.
The idea of combining strengths from different areas certainly deserves fur-
ther study. Nevertheless, this approach also has its weaknesses. As noted in
Jaszkiewicz and Slowinski (1994, 1995), it may be computationally rather de-
manding to find the exact characteristic neighbours in a general case. Parallel
computing is one solution. Ifthis is not possible, one can at least present differ-
ent neighbours as soon as they are calculated instead of waiting till all of them
have been generated. The visualization of alternatives is handled in Chapter 3
of Part III.

5.10. Reference Direction Approach

The reference direction approach was introduced in Korhonen and Laakso

(1984, 1985, 1986a) by the name visual interactive approach. It contains
ideas from, for example, the GDF method and the reference point method of
Wierzbicki. However, more information is provided to the decision maker. The
algorithm works best for MOLP problems if it is desired to check the optimal-
ity of the final solution. Otherwise, the algorithm can be applied to nonlinear
problems as well. The algorithm was originally designed for the maximization
of problems but here it is presented in the form of minimization. The reference
direction approach and its extensions are also briefly described in Korhonen
(1997).

5.10.1. Introduction

In reference point-based methods, a reference point consisting of aspiration

levels for each objective function is projected onto the Pareto optimal set by
an achievement function. This idea is extended here so that a whole so-called
reference direction is projected onto the Pareto optimal set. The reference di-
rection is a vector from the current iteration point to the reference point. After
 5.10. Reference Direction Approach 185

the projection the decision maker can examine this Pareto optimal curve or a
representation of it by the means of computer graphics.
An interesting feature in the reference direction approach is that no ex-
plicit knowledge is assumed about the properties of the value function during
the solution process. However, sufficient conditions for optimality can be es-
tablished for the termination point of the algorithm, if the decision maker's
underlying value function is assumed to be pseudo concave (and differentiable)
at that point (and several other assumptions to be listed later are fulfilled).
The optimality conditions are necessary only for MOLP problems.

5.10.2. Reference Direction Approach Algorithm

The algorithm is as follows. Once again, in the notation we employ objective

vectors for simplicity. Naturally, the actual calculations are performed in the
decision variable space.
(I) Find an arbitrary starting objective vector zl E R k. Set h = 1.
(2) Ask the decision maker to specify a reference point zh E R k and set
d h +1 = zh - zh as a new reference direction.
(3) Find the set Zh+l of (weakly) Pareto optimal solutions z that solve the
problem
minimize Si/J,w(z)
subject to z= zh + td h +1 ,
z E Z is Pareto optimal,
where sz,w is an achievement function, w is a weighting vector and t
has different discrete nonnegative values.
(4) Ask the decision maker to select the most preferred solution Zh+l in
Zh+l.
(5) If Zh -:f. zh+l, set h = h + 1 and go to step (2). Ot.herwise, check the
optimality conditions. If the conditions are satisfied, stop with xh+1
corresponding to Zh+l as the final solution. Otherwise, set h = h + 1
and set d h +1 to be a new search direction identified by the optimality
checking procedure. Go to step (3).
The setting of the algorithm makes it possible for the starting point to be
any point in the objective space. It does not have to be feasible, much less
Pareto optimal, since it is projected onto the (weakly) Pareto optimal set in
step (3). As a weighting vector one can use, for example, the reference point
specified by the decision maker.
The straight line from the current iteration point zh (or its Pareto optimal
projection at the first iteration) to the boundary of the Pareto optimal set is
discretized and projected onto the set of Pareto optimal points. The discretiza-
tion means using several different values for t. For linear problems parametric
linear programming can be used to obtain a Pareto optimal curve when the pa-
rameter t has values from zero to infinity. The idea is to plot the obtained values
186 Part II - 5. Interactive Methods

of the objective functions on a computer screen as value paths (see Subsection

3.3.1 in Part III) with different colours illustrating each of the objectives. The
decision maker can move the cursor back and forth and see the corresponding
numerical values at each point.
The achievement function Si,W(Z) is ofthe same form as presented in Section
3.5, namely

(5.10.1) Siilw(Z)
,
= max---,
iEI
Zi - Zi
Wi

where I = {i I Wi > O} C {I, ... , k}, W is a weighting vector, Z E Z is an

objective vector to be searched for and Z E Rk is a reference point. Note that
if it is desired to avoid weakly Pareto optimal solutions, then an augmentation
term can be added to the achievement function as, for example, in Subsection
3.4.5 or Section 3.5 (see also Steuer (1986, pp. 422-431)). An alternative is
suggested in Korhonen (1997) and Korhonen and Halme (1996), where lexico-
graphic ordering is used to guarantee the Pareto optimality of the solutions.
However, originally such kind of actions were not considered to be necessary be-
cause the purpose was simply to produce different solutions effectively. Distance
measure (5.10.1) has been chosen to facilitate parametric linear programming
(even though the solutions are only guaranteed to be weakly Pareto optimal).
The minimization problem of Sl,w is nondifferentiable but it can be trans-
formed into an equivalent, differentiable form assuming the differentiability of
the functions involved. Let us, for clarity, formulate the problem in the decision
variable space where it is solved as
minimize a
subject to h(x) - aWi ~ zf + td~H for all i E I,
xES,
with x ERn and a ERas variables.
Checking the optimality conditions in step (5) is the most complicated part
of the algorithm. Thus far, no specific assumptions have been set on the value
function. It may change during the solution process or it may not even exist at
all. It is only assumed that
1. Less is preferred to more by the decision maker.
However, we can check whether a given objective vector ZhH (and the cor-
responding decision variable x hH ) is optimal for the value function, assuming
that
2. The underlying value function U: Rk --t R exists and is pseudoconcave
on Z.
3. The feasible region S is convex and compact.
4. The constraint functions are differentiable.
Let the feasible directions at zh+ 1 be denoted by dU), j = 1, ... ,p. We
define a cone C containing all those feasible directions by
 5.10. Reference Direction Approach 187

(5.10.2) c= {Z E Z I z = Zh+ 1 + fJ3 j d(j),

)=1
f3j;::: o}.
If we have zh = zh+1 in step (5), we know that the projection of d h in
the weakly Pareto optimal set is not a direction of improvement. Then we can
apply the following sufficient condition for optimality.

Theorem 5.10.1. Let assumptions 2-4 be satisfied. Let Zh+1 E Z and let C
be a cone containing all the feasible directions at zh+1 (as in (5.10.2)). Let us
assume that

U(Zh+1) ~ U(zh+1 + f3 j d(j)) for all f3j ~ 0 and j = 1, ... ,p.

Then zh+1 is a globally optimal solution (with respect to U).

Proof. See Korhonen and Laakso (1986a).

For MOLP problems we know that if the current solution is not optimal,
then one of the feasible directions of cone C must be a direction of improvement.
This direction is then used as a new reference direction in step (3). In other
words, to be able to apply Theorem 5.10.1 at a certain point, the decision
maker must first check every feasible direction at that point for improvement.
This increases both the computational costs and the burden on the decision
maker. It is demonstrated in Halme and Korhonen (1989) and Korhonen and
Laakso (1986a, b) how the number of search directions can be reduced. For
nonlinear problems the cone containing all the feasible directions may consist
of an infinite number of generators. In this case, the optimality cannot be
checked in practice (an infinite number of checks would be needed).

5.10.3. Comments

Note that the termination condition of Theorem 5.10.1 is analogous to the

Karush-Kuhn-Tucker optimality conditions. This is proved in Halme and Kor-
honen (1989). If the value function is known, it is easy to compare alternative
objective vectors. However, what was said concerning the difficulty in deter-
mining new reference points in connection with the reference point method in
Section 5.6 is also valid here. Thus the reference direction approach can be
characterized as an ad hoc method.
The graphical illustration of the alternatives has been an important as-
pect in the development of multiobjective optimization methods that seek to
improve and facilitate the co-operation between the decision maker and the
analyst (computer). That is why graphical illustration of the alternatives is
here emphasized.
The computation time for large problems can be reduced by presenting one
piece at a time of the weakly Pareto optimal curve or its representation to the
188 Part II - 5. Interactive Methods

decision maker. If (s}he finds the end point to be the most satisfactory one,
then the next piece can be presented. If the number of objective functions is
large, the quality of graphical illustration suffers. For this reason, it is advisable
not to have more than ten objective functions at a time.
If it is not desired to check the optimality of the final result, the problem
to be solved does not have to satisfy any special assumptions. This means that
the reference direction approach can be applied to more general problems. The
reverse is valid as well. If the assumptions set are not satisfied, the optimality
cannot be checked, but the method can, of course, be used in any other way.
A similar interactive line search algorithm for MOLP problems is presented
in Benson and Aksoy (1991). The procedure generates only Pareto optimal
points and is able to automatically correct possible errors in the decision
maker's judgement.
The ideas of the reference direction approach are adapted to the goal pro-
gramming environment in Korhonen and Laakso (1986b). The intention is to
relax the predetermined roles of the objective functions and the constraints,
that is, to enable the roles to be interchanged. For that reason, the problem to
be solved is now assumed to be in the generalized goal programming form (see
Section 4.3). The objective functions are considered to be flexible goals and the
constraint functions inflexible goals. At each iteration, the decision maker can
easily convert flexible goals into inflexible ones and vice versa. This increases
the freedom of the decision maker. Combining achievement functions into goal
programming also eliminates the problems caused by feasible aspiration levels
(see Section 4.3).
The idea of changing the roles of the functions is refined in Korhonen and
Narula (1993). A systematic way of changing the roles of the objective functions
and the constraints is described therein. The presentation examines where and
how the changes can be carried out. This systematic handling concerns MOLP
.
problems, but the idea can in principle be generalized to other problems .
A dynamic user interface to the reference direction approach and its adapta-
tion to generalized goal programming is introduced in Korhonen and Wallenius
(1988). This method has been designed for MOLP problems and is called the
Pareto race. The software system implementing the Pareto race is called VIG
(Visual Interactive Goal programming) and it is described in Korhonen (1987,
1990, 1991a) and Korhonen and Wallenius (1989c, 1990). VIG is a dynamic,
visual and interactive solution system for MOLP problems with the emphasis
on graphical illustration.
The Pareto race develops reference directions in a dynamic way. In VIG,
the reference directions and the step-sizes are updated according to the actions
of the decision maker who can thus feel that (s)he is in control. The decision
maker can travel around the (weakly) Pareto optimal set as if driving a car.
The pioneering ideas of realizing user interfaces in VIG are supported by a
comparison of five MOLP programs in Korhonen and Wallenius (1989b). VIG
 5.10. Reference Direction Approach 189

was found to be superior. The main reason was that the decision makers found
the aspiration levels to be a comfortable way of expressing preference relations.
The Pareto race is extended into a computer graphics-based decision sup-
port system in Korhonen et al. (1992b). The new method is especially useful
for large-scale MOLP problems.

5.10.4. Concluding Remarks

In the reference direction approach the role of the decision maker is reminis-
cent of the reference point method. (S)he has to both specify reference points
and select the most preferred alternatives. In the reference point methods, how-
ever, there are fewer choices to select from. Ifthe problem is set in a generalized
goal programming form, the decision maker can also interchange the roles of
the objective and the constraint functions. By the reference direction approach,
the decision maker can explore a wider part of the weakly Pareto optimal set
than by the reference point method, even by providing similar reference point
information. This possibility brings the task of comparing the alternatives and
selecting the most preferred of them.
The reference direction approach works best for MOLP problems, as it has
basically been designed for them. It is interesting that the method requires no
additional assumptions about the problem and the underlying value function
until the optimality of the final solution is to be examined. The optimality can
be guaranteed under certain assumptions and with some effort.
The performance of the method depends greatly on how well the decision
maker manages to specify the reference directions that lead to improved solu-
tions. Korhonen and Laakso (1986a) mention that particularly when the num-
ber of objective functions is large, the specification of reference points may
be quite laborious for the decision maker. In this case, they suggest that ran-
dom directions in conjunction with decision maker-defined reference directions
should be used. See Korhonen and Laakso (1986a) for a discussion concerning
other ways of specifying the reference directions. Naturally, the choice of the
weighting coefficients affects the direction of the projection even though the
selection of their values has not been stressed here.
The consistency of the decision maker's answers is not important and it is
not checked in the algorithm. Thus the algorithm may cycle. This can also be
seen as a positive feature, since the decision maker is able to return to such
parts that (s)he already has examined, if (s)he changes her or his mind.
190 Part II - 5. Interactive Methods

5.11. Reference Direction Method

The reference direction (RD) method, introduced in Narula et al. (1994a, b)

is closely related to the reference direction approach. As its name suggests it is
also based on reference directions. To avoid confusion between these two meth-
ods with very similar names we use the name RD method when referring to the
reference direction method in what follows. The RD method has been designed
for nonlinear maximization problems but here we revise it for minimization and
generalize it. In addition, we relax the original convexity assumption and settle
for local optima.

5.11.1. Introduction

In the RD method, objective function values Zh calculated at a point xh are

presented to the decision maker and (s)he is asked to specify a reference point
zh consisting of desired levels for the objective functions. Once again, we move
around the weakly Pareto optimal set, which is why some objective functions
must be allowed to increase in order to attain lower values for some other
objective functions. In other words, some components of the reference point
have to be lower and some others higher or equal when compared to the current
solution. Allowing the set of higher values to be empty is a generalization of the
original form of the method. (Weakly Pareto optimal solutions can be made
Pareto optimal.)
As mentioned earlier, specifying a reference point is equivalent to an implicit
classification using three classes and indicating those objective functions whose
values should be decreased till they reach some acceptable aspiration level,
those whose values are satisfactory at the moment, and those whose values are
allowed to increase to some upper bound. Let us denote the sets of functions by
J<, J= and J> , respectively. We denote the components of the reference point
corresponding to the set J> by ef because we have upper bounds in ques-
tion. To put it briefly, a reference point is here sensible and the corresponding
classification feasible if J< f:- 0 and [> U J= f:- 0.
It is once again assumed that
1. Less is preferred to more by the decision maker.
The reference direction Zh - Zh is a fundamental element in the RD method.
The decision maker specifies a priori the number of steps to be taken in the
reference direction. The idea is to move step by step as long as the decision
maker wants to. In other words, extra computation is avoided by calculating
only those alternatives the decision maker wants to see.
 5.11. Reference Direction Method 191

The alternatives are produced by solving the RD problem

minimize max
zf
fi(X) -
iEI< zf - if
(5.11.1) subject to hex) ::; Ef + a(zf - Ef) for all i E J>,
hex) ::; zf for all i E r,
XES,

where zh is the current solution, 0 ::; a < 1 is the step-size in the reference
direction, if < zf for i E J< and Ef
> zf for i E [>. The problem is nondiffer-
entiable but it can be transformed into a differentiable form by introducing an
additional variable as described earlier (see, e.g., problem (3.4.3». If some of
the objective or the constraint functions are nondifferentiable, a single objective
solver applicable to nondifferentiable problems is needed.
The RD problem produces weakly Pareto optimal solutions.

Theorem 5.11.1. The solution of RD problem (5.11.1) is weakly Pareto op-

timal for every 0 ::; a < 1.

Proof. Let x· E S be a solution of the RD problem for some 0 ::; a < 1. Let
us assume that it is not weakly Pareto optimal. In this case there exists some
point XO E S such that fi(XO) < fi(X·) for every i = 1, ... , k.
Because x· is feasible in problem (5.11.1), xo, being weakly Pareto optimal,
must also be feasible. In addition, zf - if > 0 for every i E [< and that is why

fi(XO) -
"--":-"--~:'"
zf <
fi(X*) - zf for every i E J<.
Zh - zh zh _ zh
l z t t

This implies that

and, thus, X* cannot be a solution of problem (5.11.1). This contradiction

completes the proof and x· is weakly Pareto optimal. 0

A result concerning the opposite direction and Pareto optimality can also
be established.

Theorem 5.11.2. Let X* E S be Pareto optimal. Then there exists a reference

point and a real number 0 ::; a < 1 such that x· is a solution of RD problem
(5.11.1).

Proof. Let x· E S be Pareto optimal. Let us assume that there does not exist
z and a such that x· is a solution of the RD problem. Let us suppose that
192 Part II - 5. Interactive Methods

we have the current solution of the RD problem xh and the corresponding zh

available.
Let us choose r(x·) as a reference point. This means that we set Zi = h(x*)
for those indices where z~ > Ii (x*) and we denote this index set by [<. Further,
we set ti = /;(x·) for indices i E J> satisfying zf < /;(x·). Finally, the set
of indices where zf = li(X*) is valid is denoted by [=. This setting is possible
because x· is assumed to be Pareto optimal and xh is weakly Pareto optimal
according to Theorem 5.11.1. That is why [< "10 and [=uJ> "10. In addition
we set a: = O.
Because x· is not a solution of the RD problem, there exists another point
XO E S that is a solution of the RD problem, meaning that

h(xO) - zf h(x·) - zf 1
rna h <rna h =-,
iEI< zi -h(x*) iEJ< Zi - li(X*)

Thus h(xO) - zf < -(zf -h(x*)), that is, h(xO) < h(x*) for all i E [<,
Because XO is a solution of problem (5.11.1), it must be feasible. In other
words, we have li(XO) ~ li(X*) + 0 for i E J> and h(xO) ~ li(X*) for i E [=.
Here we have a contradiction to the assumption that x· is Pareto optimal. This
completes the proof and x* must be a solution of the RD problem. 0

According to Theorem 5.11.2 we know that any Pareto optimal solution can
be found with an appropriate classification.
An augmented formulation of the RD problem is presented in Narula et
al. (1994a, b) in order to produce only Pareto optimal solutions.

5.11.2. RD Algorithm

The steps of the RD algorithm are the following:

(1) Calculate a starting solution xl by solving auxiliary problem (5.11.2).
Show the corresponding objective vector zl to the decision maker. If
(s)he wants to stop, go to step (5). Otherwise, set h = 1.
(2) If the decision maker does not want to decrease any component of zh,
go to step (5). Otherwise, ask the decision maker to specify a reference
point zh, where some of the components are lower and some higher or
equal when compared to those of zh. If there are no higher values, set
P = r = 1 and go to step (3). Otherwise ask the decision maker also to
specify the maximum number of alternatives P (s)he wants to see. Set
r = 1.
(3) Set a: = 1 - r / P. Solve RD problem (5.11.1) to obtain a solution xh(r)
and the corresponding zh(r). Set r = r + 1.
(4) Show zh(r) to the decision maker, If (s)he is satisfied, go to step (5).
If r ~ P and the decision maker wants to see another solution, go to
step (3). Otherwise, if r > P or the decision maker wants to change the
reference point, set zh+1 = zh(r), h = h + 1 and go to step (2).
 5.11. Reference Direction Method 193

(5) Stop with xh corresponding to zh as the final solution.

The starting solution is calculated by solving the problem
minimize . max h(x)
.=l,... ,k
(5.11.2)
subject to xES.
Naturally, this problem can be formulated as a differentiable problem with, if
necessary and possible, the help of an additional variable (as earlier).

5.11.3. Comments

The RD method is an ad hoc method. The existence of a value function

would not help in specifying reference points or the numbers of steps to be
taken. It could not even help in selecting the most preferred alternative. The
reason is that one must decide for one point at a time whether to calculate new
alternatives or not. If the new alternative turned out to be less preferred than
its predecessor, one could not go back anyway.
In Miettinen and Makela (1998a), a water quality management problem
is solved by the RD method. A modification of the RD method for convex,
nonlinear integer problems is introduced in Gouljashki et al. (1997).

5.11.4. Concluding Remarks

The RD method can be considered an interactive classification-based meth-

od. It does not require artificial or complicated information from the decision
maker; only reference points and the number of intermediate solutions are used.
The decision maker is not asked to compare several different alternatives but
only to decide whether another alternative is to be generated or not.
The decision maker must a priori determine the number of steps to be taken
in the reference direction, and then intermediate solutions are calculated one
by one as long as the decision maker wants to. This can be seen as a benefit as
well as a weakness. On the one hand, it is computationally efficient since it may
be unnecessary to calculate all the intermediate solutions. On the other hand,
the decision maker is unable to return to a solution once it has been discarded,
which may be a disadvantage. Further, the number of steps to be taken cannot
be changed.
194 Part II - 5. Interactive Methods

Thus far, we have described several different methods for multiobjective

optimization. The question of differentiability has not been emphasized. How-
ever, non differentiability and many kinds of irregularities and discontinuities
are characteristic of real-world optimization problems, for example, in eco-
nomics and engineering. For this reason, it is important to have methods that
are able to solve nondifferentiable problems. We shall now handle nondifferen-
tiability aspects. Remember that by nondifferentiability we mean that all the
objective and the constraint functions are locally Lipschitzian.
One way of handling nondifferentiabilities is to regularize them into a dif-
ferentiable form by utilizing some smoothing techniques. However, regulariza-
tion simplifies the problem and causes errors in the model. Depending on the
smoothing parameters, the regularized problem is numerically either unreliable
or unstable. That is why it is important to be able to solve nondifferentiable
problems as they are, without simplifications.
For example, the area of optimal control has complex problems of a multiob-
jective nature containing nondifferentiable functions. Traditionally, they have
been solved (e.g., in Haslinger and Neittaanmiiki (1988)) by first scalarizing the
multiple objective functions into one by some simple method (like the weighting
method) and then regularizing the nondifferentiabilities. After discretization,
the problems have been solved by traditional, differentiable single objective
optimization methods. The drawbacks of regularization were mentioned above.
In scalarization, the appropriate weighting coefficients are difficult to specify.
If some of the objective functions originate in technological constraints, the
weighting method may bring about inaccuracies and the solution may be irrel-
evant in a technological sense. For this reason, it is important to use interactive
methods, where the user can direct the solution process in a desirable direction.
All the noninteractive methods presented in Chapters 2 to 4 can be em-
ployed with non differentiable problems whenever the single objective solvers
utilized can handle nondifferentiable functions. It is to be noted that the MPB
method in Section 2.2 has particularly been designed for nondifferentiable prob-
lems.
The case is different with interactive methods. Many of the interactive
methods described thus far assume that the functions involved are differen-
tiable. This is true especially for methods based on maximizing the underlying
value function. Most classification- and reference point-based methods do not
guide the decision maker, nor do they assume differentiability. Of the interac-
tive methods described, the Tchebycheff method, STEM, the reference point
method, the GUESS method and the RD method can all be used to solve non-
differentiable problems assuming that a non differentiable single objective solver
is available.
Few methods especially designed to handle nondifferentiable problems have
been proposed. In the sequel, we present one such method, known as NIMBUS.
Another aim in developing NIMBUS has been in trying to overcome some of
 5.12. NIMBUS Method 195

the weaknesses detected in the older methods. Most of the methods previously
described have had an effect on the development of NIMBUS. Either they have
offered useful ideas to adopt or unsatisfactory properties to avoid.
Trade-off rate information cannot be exploited in non differentiable prob-
lems in the way it is used in the ISWT method and in SPOT and STOM.
The natural reason is that obtaining trade-off information from the Karush-
Kuhn-Tucker mUltipliers necessitates that the functions are twice continuous
differentiable. How to obtain trade-off information in nondifferentiable cases
needs and deserves more research.
The ideas of reference points and satisficing decision making seem to be
generalizable to nondifferentiable problems. We can adopt the ideas of classify-
ing the objective functions and reference points and mix them with some ideas
from nondifferentiable analysis. The outcome is described in the next section.

5.12. NIMBUS Method

NIMBUS (Nondifferentiable Interactive Multiobjective BUndle-based opti-

mization System), presented in Miettinen (1994) and Miettinen and Makela
(1995, 1997), is an interactive multiobjective optimization method designed
especially to be able to handle nondifferentiable functions efficiently. For this
reason, it is capable of solving complicated real-world problems. We introduce
two versions of NIMBUS. They have differences in both their theoretical and
computational aspects. Theoretically, the versions differ in handling the infor-
mation requested from the user. Numerical experiments indicate differences in
the computational efficiency and controllability of the solution processes.

5.12.1. Introduction

The starting point in developing the NIMBUS method has been somewhat
the opposite to theoretical soundness. Emphasizing theoretical aspects may
lead to difficulties on the decision maker's side and more or less instable results,
not to mention higher computational costs. In the NIMBUS method, the idea
has been to overcome the difficulties encountered with many other interactive
methods. The most important aspects have appeared to be the effectiveness and
the comfortableness of the decision maker. Thus, the interaction phase has been
aimed at being comparatively simple and easy to understand for the decision
maker. NIMBUS offers flexible ways of performing interactive evaluation of
the problem and determining the preferences of the decision maker during
the solution process. At each iteration of the interactive solution process the
decision maker can direct the search according to her or his wishes.
Aspiration levels and classification have been selected as the means of inter-
action between the decision maker and the algorithm. It has been emphasized
on several occasions (e.g., in Nakayama (1995)) that an aspiration level-based
196 Part II - 5. Interactive Methods

approach is effective in practical fields. Among the validating facts for this
statement are the following. Aspiration levels do not require consistency from
the decision maker and they reflect her or his wishes well. In addition, they are
easy to implement. Using aspiration levels as a way of receiving information
from the decision maker means avoiding difficult and artificial concepts.
It is assumed that
1. Less is preferred to more by the decision maker.
2. The objective and the constraint functions are locally Lipschitzian.
3. The objective functions are bounded (from below) over the feasible region
S.
The second assumption comes from nondifferentiable analysis, and the third
assumption from the requirement of having the ideal objective vector available.
In the classification of the objective functions, the decision maker can easily
indicate what kind of improvements are desirable and what kind of impairments
are tolerable. The idea is that the decision maker examines at every iteration h
the values of the objective functions calculated at the current solution xh and
divides the objective functions into up to five classes. The classes are functions
Ii whose values
o should be decreased (i E [<),
o should be decreased to a certain aspiration level if
(i E [-:5.),
o are satisfactory at the moment (i E [=),
o are allowed to increase to a certain upper bound E? (i E I»~, and
o are allowed to change freely (i E r),

where [< u [-:5. U [= U [> u r r

= {I, ... , k}, [< u [-:5. =I- 0 and [= U [> u =I- 0.
In addition to the classification, the decision maker is asked to specify the
aspiration levels if
for i E [-:5. satisfying if < fi(xh) and the upper bounds E?
for i E [> such that E? > fi(X h ). Notice that the two somewhat parallel classes
[< and [-:5. are available. The difference between them is that the functions in
[< are to be minimized as far as possible but the functions in [-:5. only as far as
the aspiration level. Thus the functions in the latter class are called aspiration
functions.
The classification is the core of NIMBUS. However, the decision maker can
tune the order of importance inside classes J< and [-:5. with optional positive
weighting coefficients w? summing up to one (for numerical stability). If the
decision maker does not want to specify any weighting coefficients, they are set
equal to one. Note that the weighting coefficients do not change the primary
orientation specified in the classification phase.
NIMBUS has more classes than STEM, STOM or the RD method. In this
way the decision maker has more freedom in specifying the desired changes in
the objective values and (s)he can select a class reflecting her or his desires
best. The class r is new when compared to the methods described thus far.
In practice, it means that not all the objective functions have to be classified
at all. Naturally all the classes do not have to be employed.
 5.12. NIMBUS Method 197

After the decision maker has classified the objective functions, one of the two
alternative subproblems, called vector and scalar subproblems, is formed. Thus,
the original multiobjective optimization problem is transformed into either a
new multiobjective or a single objective optimization problem, accordingly.
The subproblems lead to two different versions of NIMBUS, to be called vector
version and scalar version. We first introduce the older, that is, the vector
version.

5.12.2. Vector Subproblem

According to the classification and the connected information, a vector sub-

problem

minimize {fi(X) (iEI<), ~~[max [Ji(x)-z7, oJ]}

(5.12.1) subject to Ji(x)::; Ji(x h ) for all i E [=,

/i{x) ::; €? for all i E [> ,

xES
is formed (see Miettinen (1994) and Miettinen and MakeUi (1995».
The vector subproblem seems to be even more complicated than the original
problem. Nonetheless, the advantage of this formulation is the fact that the
opinions and the hopes of the decision maker are taken carefully into account.
Notice that if ['.5. =j:. 0, we have a non differentiable problem regardless of the
differentiability of the original problem. This fact does not bring any additional
difficulties since we are in any case prepared for handling nondifferentiabilities.
The vector version is quite general. The classification of the objective func-
tions can be performed as if the €-constraint method, the weighting method,
lexicographic ordering or goal programming were used to produce new solu-
tions.
In order to be able to solve the vector subproblem, we need the MPB method
(introduced in Section 2.2). If the constraints are inequalities, that is S =
{x E Rn I g(x) = (gdx), g2(X), ... , gm(x»T ::; O}, the improvement function
H: Rn x Rn -+ R applied to problem (5.12.1) is of the form

H(x 1,X 2) = max {fi(xl)/w~ - fi(X2)/W~, (i E 1<),

~~? [max [Ji(x1)/wJ - zJ, OJ] - ~~? [max [fJ(x 2)/wJ - zj, 0]),
fi(Xl) - fi(Xh), (i E J=),
h(x 1 ) - €7, (i E J»,
gl(X 1), (l=l, ... ,m)}.

Notice that the weighting coefficients wr,

i E 1< UI'.5., are not included in multi-
objective problem (5.12.1), but are taken into account inside the MPB method.
198 Part II - 5. Interactive Methods

As explained in Section 2.2, the minimization of the improvement function takes

place iteratively. For details, see also Makela (1993) and Miettinen and Makela
(1995,1998a).

5.12.3. Scalar Subproblem

In the scalar version of NIMBUS, after the classification, a scalar subproblem

minimize max [WfUi(X) -

iEI<
zn, w h max [h(x) -
J zJ\ 01]
jEIS,

(5.12.2) subject to J;(x) ~ /i(X h) for all i E J< U Js. U J=,

/i(X) ~ E7 for all i E J>,
xES

is formed (see also Miettinen and Makela (1996b, 1998a) and Miettinen et
al. (1996b)), where zt
for i E J< are components of the ideal objective vector
(assumed to be known globally).
Notice that problem (5.12.2) is nondifferentiable but has one objective func-
tion. It can be solved by any method for nondifferentiable single objective
optimization, for example, by efficient bundle methods (see MiikeUi and Neit-
taanmiiki (1992, pp. 112-137)).
Scalar subproblem (5.12.2) can be formulated in an alternative form:

minimize max [wfmax[J;(x),zfl]

iEI< u/S,

(5.12.3) subject to J;(x) ~ J;(x h ) for all i E J< U Js. U J=,

/i(X) ~ E7 for all i E J>,
XES,

where the aspiration level is constant zf = zt

for i E J<. This formulation seems
somewhat simpler but the idea is the same. Subproblems (5.12.2) and (5.12.3)
do not, however, produce identical results because of scaling differences. In the
following, we refer to problem (5.12.2) as the scalar subproblem but problem
(5.12.3) could equally be used instead.
We prove in Subsection 5.12.5 that the solutions of the vector and the scalar
subproblem are weakly Pareto optimal under certain conditions.

5.12.4. NIMBUS Algorithm

The solution of vector subproblem (5.12.1) or scalar subproblem (5.12.2) is

denoted by xh. If the decision maker does not like the objective vector f(xh) for
some reason, (s)he can explore other solutions between xh and Xh. This means
that we calculate a search direction d h = xh - Xh and provide more solutions
 5.12. NIMBUS Method 199

by taking steps of different sizes in this direction. The step-size is determined

by the decision maker as in the GDF method. Objective vectors f(xh + td h ) are
calculated with different values of t (0 :s; t :s; 1). Their weakly Pareto optimal
counterparts are presented to the decision maker, who then selects the most
satisfying solution among the alternatives.
A detailed algorithm of the NIMBUS method is given below. The same
algorithm is valid for both of the NIMBUS versions. Note that the decision
maker must be ready to give up something in order to attain improvement
for some other objective functions. The search procedure stops if the decision
maker does not want to improve any objective function value.
(1) Select subproblem (5.12.1) or (5.12.2) to be used in the continuation.
Choose a starting point x E R n and project it onto the feasible region
by solving auxiliary problem (5.12.4). Denote the new point by xo.
Calculate its weakly Pareto optimal counterpart xl by setting J< =
{I, ... , k} and by solving the selected subproblem. Set the iteration
counter h = 1.
(2) Ask the decision maker to divide the objective functions into the
classes 1<, IS., 1=, I>, and 10 at the point zh = f(Xh) such that
[= u J> u 1 0 f= 0 and J< U IS. f= 0. If either of the unions is empty,
go to step (9). Ask the decision maker for the aspiration levels z~ for
i E IS. and the upper bounds e7 for i E J>. Ask also for the optional
weighting coefficients wf > 0 for i E 1< U IS., summing up to one.
(3) Calculate j{h by solving the subproblem. If j{h = xh, ask the decision
maker whether (s)he wants to try another classification. If yes, set
xh+1 = xh, h= h + 1, and go to step (2); if no, go to step (9).
(4) Now j{h is a new solution. Let us denote zh = f(j{h). Present zh and
zh to the decision maker. If the decision maker wants to see different
alternatives between zh and zh, set d h = xh - xh and go to step (6).
If the decision maker prefers Zh, set Xh+1 = xh and h= h + 1, and go
to step (2).
(5) The decision maker wants now to continue from zh. If [< f= 0, set
xh+l = Xh, h = h + 1, and go to step (2). Otherwise (I< = 0), the
weak Pareto optimality must be guaranteed by setting [< = {I, ... , k}
and solving the subproblem. Let the solution be xh. Set X h + 1 = xh
and h = h + 1, and go to step (2).
(6) Ask the decision maker to specify the desired number of alternatives P
and calculate vectors f(x h + tjd h ), j = 2, ... , P - 1, where tj = t-~.
(7) Produce weakly Pareto optimal objective vectors from the vectors
above by solving auxiliary problem (5.12.5).
(8) Present the P alternatives to the decision maker and let her or him
choose the most preferred one among them. Denote the corresponding
decision vector by xh+1 and set h = h + 1. If the decision maker wants
to continue, go to step (2).
(9) Check the Pareto optimality of xh by solving auxiliary problem
(2.10.2) of Part I with xh as x*. Let the solution be (x,E).
200 Part II- 5. Interactive Methods

(10) Stop with the final solution x.

The projection in step (1) is connected to the fact that most solvers ne-
cessitate feasible starting points. Thus, it is a more implementational than
algorithmic matter. Let us mention that, for example, if the feasible region
consists of inequality constraints gi(X) ::; 0 for i = 1, ... , m, any starting point
can be projected onto the feasible region by solving the auxiliary problem
minimize max [0, gl (x), g2(X), ... , gm(x)]
(5.12.4)
subject to x ERn.

The justification of step (5) is given in the optimality results of Subsection

5.12.5. If we only employ the class [75: to minimize functions (and the class [<
is empty) we do not necessarily stay within the weakly Pareto optimal set. In
this case we project the obtained result onto the weakly Pareto optimal set.
This is acceptable according to assumption l.
The intermediate solutions between zh and zh are not necessarily weakly
Pareto optimaL That is why they have to be projected onto the weakly Pareto
optimal set. A practical way of doing so is to employ the results of Corollary
3.5.6 and solve the auxiliary problem

minimize
(5.12.5)
subject to xES

for every j = 2, ... , P-l. This treatment works for convex as well as nonconvex
problems. An alternative method can be applied if the vector subproblem is
used and the problem is convex (see Miettinen and Makela (1995)). In this case
weak Pareto optimality can be guaranteed by solving the vector subproblem
with [< = {I, ... ,k} starting from each intermediate solution.
Since the Pareto optimality of the solutions produced cannot be guaranteed
(see Subsection 5.12.5), we check the final solution in the end by solving an
additional problem introduced in Theorem 2.10.3 of Part I. As the decision
maker was assumed to prefer less to more, we can presume that (s)he is satisfied
with the Pareto optimal final solution even where it was not her or his choice.
For clarity of notation, it is not stated in the algorithm that the decision maker
may check Pareto optimality at any time during the solution process. Then,
problem (2.10.2) of Part I is solved with the current solution as x·.
Note that, if scalar subproblem (5.12.2) is employed in the algorithm, we
have to calculate the components of the ideal objective vector z* in the first
step. However, presenting z* to the decision maker gives valuable information
about the problem in both NIMBUS versions.
We must remember that we cannot guarantee global optimality. If the solu-
tion obtained is not completely satisfactory, one can always solve the problem
again from a different starting point. This action is also advised if the decision
maker has to stop the solution process with xh = xh after step (3).
 5.12. NIMBUS Method 201

It is also possible to improve the algorithm in step (3) to avoid the case
xh = xh. If the upper bounds specified by the decision maker are too tight,
one can use them as a reference point and project them with (5.12.5) onto
the (weakly) Pareto optimal set. Showing the new solution to the decision
maker provides her or him with information concerning the possibilities and
the limitations of the problem, and some dead ends can be avoided as well.
Unlike some other methods based on classification, the success of the solu-
tion process does not depend entirely on how well the decision maker manages
in specifying the classification and the appropriate parameter values. It is im-
portant that the classification is not irreversible. Thus, no irrevocable damage
is caused in NIMBUS if the solution f(xh) is not what was expected. The
decision maker is free to go back or explore intermediate points. (S)he can eas-
ily get to know the problem and its possibilities by specifying, for example,
loose upper bounds and examining intermediate solutions. NIMBUS is indeed
learning-oriented.

5.12.5. Optimality Results

First, we state two theoretical results concerning the optimality of the so-
lutions of vector subproblem (5.12.1) and scalar subproblem (5.12.2).

Theorem 5.12.1. The Pareto optimal solution of vector subproblem (5.12.1)

is weakly Pareto optimal (to the original multiobjective optimization problem)
if the set J< is nonempty.

Proof. Let us denote the feasible region of vector subproblem (5.12.1) by S.

Let x* E S be a Pareto optimal solution of the vector subproblem with some
sets J<, J5:, J=, J> and r, where J< =f. 0. In other words, there does not
exist another decision vector xES such that li(X) ~ /;(x*) for all i E J< and
maxjEI$ [max (fAx) - zj, oJ] ~ maXjEI$ [max [Ii(x*) - zj, oJ] and at least
one of the inequalities is strict.
Let us assume that x* is not weakly Pareto optimal for the original problem.
This means that there exists a decision vector XO E S such that li(XO) < li(X*)
for all i = 1, ... , k.
Because x* is a feasible solution of problem (5.12.1), we have /;(XO) <
/;(x*) ~ /;(xh) for i E J= and 1;(xO) < li(X*) ~ c~ fori E J>. Thus, also
XO E S.
For all i E J5: is valid /; (XO) - zf < Ii (x*) - zf. It implies that max [/; (XO)-
zf, 0] ~ max [/;(x*) - zf, 0] for all i E J5:, and, further,

While, in addition,
202 Part II - 5. Interactive Methods

for all i E J< f. 0, the point x" cannot be a Pareto optimal solution of the
vector subproblem. This contradiction implies that x* must be weakly Pareto
optimal. The proof is also valid if some of the classes J$, J=, J> or [0 are
empty as long as J> U J O f. 0. 0

Theorem 5.12.2. The solution of scalar subproblem (5.12.2) is weakly Pareto

optimal if the set J< is nonempty.

Proof. Let us denote the objective function of scalar subproblem (5.12.2) by

f(x) to be minimized and the feasible region by S. Let x" E S be a solution of
the scalar subproblem with some sets [<, [$, [=, [>, and JO, where [< f. 0.
In other words, f(x'") ~ f(x) for all xES.
Let us assume that x" is not weakly Pareto optimal. This means that there
exists a vector XO E S such that h(xO) < h(x*) for all i = 1, ... ,k.
Note in the following that all the weighting coefficients are strictly positive.
Because x" E S, we have h(xO) < fi(X*) ~ h(x h ) for i E J< U [$ U [= and
h(xO) < fi(X·) ~ Ci for i E J>. Thus, also XO E S.
Since zt ~ h(xO) < h(x·) for all i E [< f. 0, we have h(x*) - zt
> 0 for
all i E [<, With Wi> 0 we also have wi(h(x·) - zt) > 0 for all i E [<.
Let us consider
f(xO) = max
iEI<
[WiUi(XO) - zt),Wj max [/j(XO) - :ZJ~' OJ].
jEI5.

The maximum can be attained either in the class [< or in [$ (or, naturally, in
both of them). In the first case we have
f(xO) = WiUi(XO) - zi) < WiUi(X*) - zi) :::; I(x*)
for some i E [<. The latter case has two different alternatives. Firstly,
f(xO) = wjmax[/j(xO) - :zj, OJ = 0 < wi(h(x*) - zt):::; f(x")
for some j E [$ and for all i E [<. Secondly,
l(xO) = Wj max [/j(XO) -:zj, OJ = Wj(/j(XO) - :zj) < Wj(/j(x·) - :zj) :::; I(x*)
for some j E [$,
In conclusion, we can state that the point x* cannot be a solution of the
scalar subproblem. This contradiction implies that x* must be weakly Pareto
optimal. The proof is also valid if some of the classes [$, [=, [> or JO are
empty as long as J> U [0 f. 0. 0

The following optimality result is common for the scalar and the vector
subproblem (even the proofs can be combined).

Theorem 5.12.3. Any Pareto optimal solution can be found with an appro-
priate classification in problems (5.12.1) and (5.12.2).
 5.12. NIMBUS Method 203

Proof. Let x· E S be Pareto optimal. Let us assume that there does not
exist a classification such that x* is a solution of the vector or the scalar
subproblem. Let us suppose that we have the current NIMBUS solution xh
and the corresponding zh available.
Let us choose f(x") as a reference point. This means that we choose Zi =
j;(x") for those indices where zf > fi(X*) and set i E ]<:;. Further, we set
/Oi = j;(x*) for indices i E J> satisfying zf < j;(x*). Finally, the set J=
consists of indices where zf = fi (x*). This setting is possible because x* is
assumed to be Pareto optimal and xh is weakly Pareto optimal according to
Theorems 5.12.1 and 5.12.2 and the structure of the NIMBUS algorithm. That
is why J<:; =f. 0 and J= U J> =f. 0. In addition, we set Wi = 1 for i E ]<:;.
Because x" is not a solution of the vector or the scalar subproblem, there
exists another point XO E S that is a solution, meaning that

~~~ [ max [Jj(XO) - iJ(x*), OJ] < ~~~ [max [fj(x*) - iJ(x"), OJ] = o.

Thus, max [iJ(XO) - hex"), OJ < 0 for every j E ]<:;. In other words, we have
iJ(XO) < h(x*) for every j E J<:;. Because XO is a solution of problems 5.12.1
and 5.12.2, it must be feasible. In other words, we have fi(XO) ~ fi(X*) for
i E J= U J>. Here we have a contradiction to the assumption that x* is Pareto
optimal. This completes the proof and x· must be a solution of the vector and
the scalar subproblem. 0

The MPB method is an essential element of the vector version of NIMBUS.

It can also be used for solving the scalar version of NIMBUS. The performances
of the two versions are comparable if the same solver is employed in both of
them. Thus, the optimality of the solutions produced by the MPB method is an
important fact to consider. The optimality of the solutions of the MPB method
was handled in Subsection 2.2.3. Accordingly, in theory only the substationarity
of the solutions of the MPB method is guaranteed for general multiobjective
optimization problems. For fluency, we have thus far referred and shall continue
referring to the solutions as weakly Pareto optimal. Note, on the other hand,
that a global single objective optimizer can be employed with the scalar version
to produce globally (weakly) Pareto optimal solutions.

5.12.6. Comparison of the Two Versions

The vector and the scalar versions of NIMBUS differ in the form of the
subproblem used. The origin of the development of the scalar version lies in
the drawbacks discovered in the vector version.
Theoretically, the solution of the vector subproblem has to be Pareto opti-
mal in order to guarantee weakly Pareto optimal solutions to the original mul-
tiobjective optimization problem. This assumption is quite demanding. With
the scalar subproblem we do not have problems of this kind. Further, the vector
version needs a special solution tool - MPB. In addition to this limitation, the
204 Part II - 5. Interactive Methods

role of the weighting coefficients is not commensurable between the classes [<
and [~. This implies that the controllability of the method suffers.
The advantage of having a single objective function in the scalar version
is that we can employ any efficient optimization routine of nondifferentiable
optimization. This gives more generality and applicability to the method. Fur-
thermore, in the scalar subproblem, we treat the functions in [< and IS. in a
consistent way and, thus, the roles of the weighting coefficients are also iden-
tical. In all, this means that the decision maker can better direct the solution
process.
Notice that in addition to the difference in the objective functions of the
subproblems, there is also deviation in the constraint part. Due to the goal of
the classes [< and [s., we have to make sure that the values of these func-
tions do not increase. This is the reason for modifying the constraints of scalar
subproblem (5.12.2).
In the vector subproblem, the MPB method does not allow increment in
[<. However, there is no guarantee that the values of the functions in IS. could
not increase. It is clear that including additional constraints in an optimization
problem increases its computational complexity. Because the increasing feature
occurs very rarely in the vector version, no additional constraints have been
used in order to emphasize computational efficiency. Thus, either we pay the
price of additional computational costs or take the risk of increment (depending
on the classification).
On the one hand, the calculation of the ideal objective vector used in the
scalar version also needs computational effort. On the other hand, the ideal
objective vector can provide supporting information for the decision maker in
any kind of multiobjective solution process.
A numerical comparison of the two versions of NIMBUS is reported in Miet-
tinen and Makela (1996b, 1998a) with versatile multiobjective optimization
problems. The standards of comparison chosen are computational efficiency
and the opinion of the decision maker concerning the controllability of the dif-
ferent versions. The efficiency can be measured by the number of times the
subroutine containing the objective functions is called. The controllability side
must be elicited from the decision maker. It is measured in the form of a rating
(between 1 and 5).
The numerical tests indicate that the scalar version obeys the decision maker
better, whereas the vector version is computationally more efficient. However,
it is important to note that the classifications employed affect considerably
the performance of the NIMBUS versions. In any case, the user has to choose
between controllability and computational efficiency when selecting a solution
method.
 5.12. NIMBUS Method 205

5.12.7. Comments

The NIMBUS method has not been developed to converge in the traditional
sense. While the method does not assume the existence of any underlying value
function, no explicit convergence results can be put forward on the basis of the
assumptions about the properties of the function. In particular, the intention
has been to release the decision maker from the assumption of an underlying
value function. What is important is that the method satisfies two desirable
properties of interactive methods: not to place too demanding assumptions
on the decision maker or the information exchanged, and to be able to find
(weakly) Pareto optimal solutions quickly and efficiently.
The aim has been to formulate a method where the decision maker can
easily explore the (weakly) Pareto optimal set. When the decision maker no
longer wants to change any objective function value and the solution process
stops, the solution is then optimal.
An important factor is that the final solution is always Pareto optimal
because of the structure of the algorithm. In addition, all the intermediate
points are at least substationary points and they can be projected onto the
Pareto optimal set, if so desired.
The method is ad hoc in nature, since the existence of a value function
would not directly advise the decision maker how to act to attain her or his
desires. A value function could only be used to compare different alternatives.
The possibility of interchanging the roles of the objective and the constraint
functions has been mentioned thus far in connection with some methods. This
is easy to carry out also in NIMBUS because the class J> is nothing but
constraints with upper bounds. One can even go that far as to formulate all
the constraint functions as objective functions and modify their upper bounds
or roles during the solution process from one iteration to the other.

5.12.8. Implementations

The NIMBUS algorithm was originally implemented in the mainframe en-

vironment at the University of Jyvaskyla, Finland. This approach is suitable
for even large-scale problems, but the lack of a flexible user interface decreases
its usability. It is evident that the user interface plays a crucial role in realizing
interactive algorithms.
An alternative is to use microcomputers to develop a functional user inter-
face by paying the price of reduced computational capacity. However, both the
mainframe and the microcomputer environment share weaknesses in common
from the viewpoint of both the user and the developer of the implementation.
As far as the user is concerned, the system and some specific compilers have to
be installed. It is limiting that the programs are appropriate only for certain
computer environments and operating systems. For the developer the delivery
of the software updates is laborious. Implementing and keeping up separate
versions for different environments requires also extra effort.
206 Part II - 5. Interactive Methods

The strengths of the mainframe and the microcomputer environment are

the computational efficiency and graphical user interface, respectively. On the
one hand, the readiness to combine them and, on the other hand, the rapid
expansion in the use of the Internet and the World-Wide Web (WWW) have
motivated the development of a WWW-NIMBUS system in the Internet (see
Miettinen and Makela (1998b».
Via the Internet we can centralize the computing to one server computer
(at the University of Jyviiskyla) and the WWW is a way of distributing the
interface to the computers of each individual user. The WWW enables the
realization of a graphical user interface. Thus, it is possible to implement an
interactive optimization algorithm by the means of the WWW. The idea is not
to demand high computing capacity or special compilers from the computer
of the user. Instead, this is left to one efficient server computer. The same
computer takes care of the visualization and the problem data management.
This centralizing offers benefits to both the user and the software producer.
The user always has the latest version of the NIMBUS method available, and
the producer has only to update and develop one version. In addition, the user
saves the trouble of installing the software.
The most important aspect of WWW-NIMBUS is that it is easily accessible
and available to any Internet user (http://nimbus.math.jyu.fi/). No special
tools, compilers or software besides a WWW browser are needed. The system
is independent of the computer and the operating system used.
The WWW environment enables the possibility of graphical classification
and graphical visualization of the alternatives. They both support the decision
maker in getting to know the problem and finding better solutions.

5.12.9. Applications

In Miettinen (1994), two academic problems and a state-constrained optimal

control problem concerning an elastic string are solved by the vector version of
NIMBUS. The vector version is applied to solve an academic nondifferentiable
test problem and a river pollution problem in Miettinen and Makela (1995). A
non differentiable version of the pollution problem is dealt with in Miettinen and
MiikeUi (1997). A structural design problem is solved by the vector version of
NIMBUS in Miettinen et al. (1996a). In Miettinen and MiikeHi. (1996b, 1998a),
a water quality management problem is solved by both the vector and the
scalar versions.
An optimal control problem related to the continuous casting of steel is
solved by the vector version in Miettinen (1994) and by the scalar version in
Miettinen et al. (1996b). This problem is an example of the case where the
modelling phase ends up with an empty feasible region. The so-called techno-
logical constraints are so tight that there does not exist any feasible solution.
When this happens, the constraints can be treated as objective functions with
the original objective function(s) thus forming a multiobjective optimization
 5.12. NIMBUS Method 207

problem. One of the goals is then to find a solution as close to the feasible
region as possible.

5.12.10. Concluding Remarks

NIMBUS is one of the few efficient, interactive methods especially devel-

oped for solving nondifferentiable multiobjective optimization problems. Nat-
urally, differentiable problems can be solved as well. In two different versions
of NIMBUS, the decision maker moves around the weakly Pareto optimal set
and expresses iteratively her or his desires by specifying those objectives whose
values should improve and those whose values are allowed to deteriorate with
the help of five available classes. The selection of the most preferred alternative
from a given set is also possible. The questions posed to the decision maker are
not demanding. The method aims at being flexible and the decision maker can
select to what extent (s)he exploits the versatile possibilities of the method.
The calculations are not too massive, either. The use of efficient bundle meth-
ods as the underlying nondifferentiable optimizers is recommended (see MakeHi
and Neittaanmaki (1992, pp. 138-143».
In NIMBUS, the decision maker is free to explore the (weakly) Pareto op-
timal set and also to change her or his mind if necessary. Previous acts do not
limit the movements. The decision maker can also extract undesirable solu-
tions from further consideration. Naturally, the decision maker does not have
to employ all of the five classes if (s)he feels uncomfortable with some of them.
However, it is important to provide the decision maker with alternative courses
of action.
The classification of the functions and the specification of the appropriate
parameter information does not necessarily have to succeed as well as in some
other classification-based methods. The reason is that intermediate solutions
can be examined and, thus, more information about the problem obtained. This
makes the method more flexible. In addition, the decision maker can cancel
any classification step because nothing is irreversible. A further advantage is
that not all the objective functions have to be classified. This is not possible
in other classification-based methods. In NIMBUS, the decision maker can set
some objective function free and examine what happens to the other objectives.
Even though the Pareto optimality of the solutions produced cannot be
guaranteed, at least the final solution is Pareto optimal. If it is important to
the user that the intermediate solutions are Pareto optimal, they can be pro-
jected. However, this adds to the computational costs. The fact that the Pareto
optimality of the solutions produced is not guaranteed automatically is at least
partly compensated for by the computational efficiency of the underlying opti-
mizer.
From numerical experiments we can conclude that of the two NIMBUS
versions available the scalar version is more controllable and dirigible to the
decision maker. The advantage of the vector version is its computational effi-
208 Part II - 5. Interactive Methods

ciency. Thus, the user has to choose between these aspects when selecting a
solution method.
Eventually, it is up to the user interface to make the most of the possibilities
of the method and provide them to the user. When the first WWW version of
the NIMBUS algorithm was implemented in 1995 it was a pioneering interactive
optimization system in the Internet. The realization is based on the ideas of
centralized computing and a distributed interface.
Naturally, there are many challenges in the further development of the NIM-
BUS method and its implementations. One of the challenges, applicable to
software development in general, is to create illustrative and easy-to-use user
interfaces. If the interface is able to adapt to the decision maker's style of mak-
ing decisions and is of help in analyzing the alternatives and results, and can
perhaps give suggestions or advice, then the interface may even overcome some
of the deficiencies of the method itself.

5.13. Other Interactive Methods

The number of existing interactive methods is large. That is why it is neither

the purpose nor practical nor possible to discuss all of them here. Neverthe-
less, in addition to those presented in the previous sections, some methods are
listed below. Only the basic concepts and ideas of the methods are mentioned
together with references. The methods are roughly divided according to their
basis in goal programming, in weighted metrics, in reference points, and in
miscellaneous ideas. Some methods for linear problems are included because of
their interesting basic ideas or because they are referred to in connection with
method comparisons in Section 1.2 of Part III.

5.13.1. Methods Based on Goal Programming

A rather straightforward extension of goal programming into an interactive

form is presented in Masud and Hwang (1981). The method is called the in-
teractive sequential goal programming (ISGP). The interactive multiple goal
programming (IMGP) method, described in Nijkamp and Spronk (1980) and
Spronk (1990), has also been created to combine the flexibility of goal pro-
gramming and the robustness of interactive approaches. The decision maker
indicates which objective value(s) should be improved and either revises the
aspiration levels of the corresponding goals or the problem is automatically
modified with additional constraints.
The sequential multiobjective problem solving (SEMOPS) technique is
briefly outlined in Monarchi et al. (1973). Five types of goal specifications in
the form of points and intervals are allowed. A different measure of deviation is
utilized for each type. (For example, if the goal is of the form Ji(x) ~ Zi, then
the corresponding measure of deviation is lSi = fi(X)/Zi.) At each iteration, a
 5.13. Other Interactive Methods 209

subset of deviations is summed up and then minimized. The decision maker

may change that subset and specify new aspiration levels. Unfortunately, the
solutions are not guaranteed to be Pareto optimal. A related method, called the
sequential information generator for multiple objective problems (SIGMOP), is
introduced in Monarchi et al. (1976). SIGMOP is a flexible method where the
decision maker can alter aspiration levels and weighting coefficients as (s)he
separates attainable solutions from among the desired ones. As an applica-
tion, a pollution problem in water resources is solved by the SEMOPS and the
SIGMOP methods in the references mentioned.
The ideas of goal programming, the oS-constraint method and trade-offs are
combined in the direction-searching method proposed in Masud and Zheng
(1989). The method aims at reducing the cognitive burden on the decision
maker while not increasing computational complexity. The algorithm is illus-
trated by a numerical example. The properties of the method are also compared
with those of several other interactive methods.
The general purpose interactive goal programming algorithm is suggested
in Tamiz and Jones (1997b). An interactive goal programming algorithm for
nonlinear problems based on different norms and updating the aspiration levels
is presented in Weistroffer (1983).

5.13.2. Methods Based on Weighted Metrics

The idea of the method in Moldavskiy (1981) is to form a grid in the space
of the weighting vectors and to map this grid onto the Pareto optimal set.
Weighted Lp-metrics are used as scalarizing functions to produce a represen-
tation of the Pareto optimal set. The decision maker can contract the space of
the weighting vectors until the most satisfactory solution is obtained.
A method based on sensitivity analysis and the weighted Tchebycheff metric
is presents in Diaz (1987), where the effects of changing aspiration levels are
studied by sensitivity analysis. The method in Sunaga et al. (1988) utilizes
also the weighted Tchebycheff metric. It transforms the constrained min-max
problem into a series of (differentiable) unconstrained problems by penalty
functions.
The interactive cutting-plane algorithm is presented in Loganathan and
Sherali (1987) with applications. The idea is to maximize the underlying value
function. The weighted Tchebycheff metric is utilized with marginal rates of
substitution as weighting coefficients. The convergence of the algorithm is also
treated.
The method proposed in M'silti and Tolla (1993) combines features from
the oS-constraint method and the augmented weighted Tchebycheff metric. The
global Pareto optimality of the solutions obtained is checked.
The method of the displaced ideal for MOLP problems, described in Zeleny
(1973, 1974, 1976), can be characterized as an interactive extension of the
method of weighted metrics. A subset of the Pareto optimal set is obtained
by minimizing the distance between the ideal objective vector and the feasible
210 Part II - 5. Interactive Methods

objective region by the weighted Lp-metrics with altered exponents p. The

subset is reduced by moving the reference point towards the feasible objective
region until the subset of the Pareto optimal solutions is small enough for the
decision maker to select the most preferred solution. The method is based on
empirical studies of the decision maker's behaviour.
The distance to the utopian objective vector is minimized by weighted Lp-
metrics in K6ksalan and Moskowitz (1994). This interactive method is based
on determining the weighting coefficients according to the preferences of the
decision maker. The preference information is obtained from pairwise compar-
isons.
Ways of approaching discrete multiattribute decision analysis problems have
been included in the method introduced in Kok and Lootsma (1985). The ideal
objective vector is used as a reference point. Pairwise comparison methods are
applied between the reference point and the (possibly approximated) nadir ob-
jective vector. The distances are measured by solving the augmented weighted
Tchebycheff problem.
An interesting method is suggested in Kaliszewski et al. (1997). In this hy-
brid interactive decision making technique, the decision maker can select what
kind of information to specify. (S)he can either classify the objective functions
or specify upper bounds on global trade-offs. New properly Pareto optimal
solutions are generated by modified weighted Tchebycheff problem (3.4.9) ac-
cording to the results derived in Subsection 3.4.6. This is the way of taking the
bounds on the global trade-offs into account.

5.13.3. Methods Based on Reference Points

Multiple reference points and a gradient projected method are bases of

the method of Costa and Climaco (1994) for MOLP problems. The method is
related to Pareto race (in Subsection 5.10.3) but it utilizes parallel processing
when handling several reference points simultaneously.
The method of Wierzbicki forms the basis of the interactive reference point
methods introduced in Bogetoft et al. (1988). The multiobjective optimization
problem is assumed to be convex. Karush-Kuhn-Tucker multiplier information
is presented to the decision maker to guide the specification of new reference
points. Several different modifications are also presented and their convergence
properties are studied.
In the method presented in Tapia and Murtagh (1989), the decision maker
is asked to express preferential desires to attain her or his reference point. So-
called preference criteria are formed from this information. These preference
criteria are then used as a reference point in the achievement function. The
authors also report some encouraging numerical experiments.
A method combining the ideas of reference points and measuring distances
is suggested in Hallefjord and J6rnsten (1986). After the decision maker has
specified the reference point, the distance between it and the feasible objective
 5.13. Other Interactive Methods 211

region is minimized by an entropy function. The mathematical background of

the method is widely handled in the reference.
The method presented in Weistroffer (1982) assumes that the decision maker
specifies required values or maximum-achievement levels. The surplus is then
maximized to the Pareto optimal set. Further, the methods in Narula and
Weistroffer (1989b) and Weistroffer (1984, 1987) expect the decision maker to
provide both the required and desired values for every objective function. Then
an achievement function is optimized. The required and the desired values are
modified until the most preferred solution is obtained. Some convergence results
are also dealt with.
The bi-reference procedure presented in Michalowski and Szapiro (1992)
has been developed for MOLP problems. The decision maker is asked to specify
the worst acceptable objective vector, and a search direction is obtained as the
difference between the worst and the ideal objective vectors. As long a step as
possible is taken in that direction and the decision maker is asked to divide the
objective functions into three classes (to be improved, to be kept unchanged
and to be relaxed). Then the worst and the ideal objective vectors are replaced
and the procedure continues until no significant improvement is achieved. The
performance of the bi-reference procedure is compared with other interactive
procedures by solving some test examples from the literature. At least in those
examples the procedure manages quite well.
An extreme point method for MOLP problems is described in Kirilov and
Vassilev (1997). It is based on the reduction of the weighting space and is related
to the Tchebycheff method. The decision maker can compare different solution
alternatives and guide the solution process by specifying reference points.

5.13.4. Methods Based on Miscellaneous Ideas

An interactive extension of the weighting method is presented in Steuer

(1986, pp. 394-399). Many of its ideas are related to those of the Tchebycheff
method. The set of the weighting vectors is reduced according to the choices of
the decision maker. Weighting vectors are generated randomly from the reduced
space and filtered to obtain a well dispersed set. Below, this approach will be
called the method of Steuer.
Another interactive method based on the weighting method is introduced
in Hussein and El-Ghaffar (1996). It can handle convex problems and is based
on solving systems of equations formed according to the Karush-Kuhn-Tucker
type optimality conditions.
Two different interactive relaxation methods are put forward in Nakayama
et aI. (1980) and Lazimy (1986b). The latter is applicable to both continu-
ous and integer problems. The methods are based on the maximization of an
underlying value function in a new but equivalent form with additional con-
straints. Marginal rates of substitution and other estimates of the value func-
tion are required from the decision maker. Similar ingredients are utilized in
212 Part II - 5. Interactive Methods

the decomposition method presented in Lazimy (1986a). It is based on the du-

ality theory for nonlinear programming. The original problem is decomposed
into a series of linear sUbproblems and two-attribute problems. In addition,
a relaxation-projection technique, especially for bi-objective problems, is pro-
posed in Ferreira and Geromel (1990). An application to scheduling is also
handled.
An interactive algorithm with several alternative subproblems is proposed
in Mukai (1980). The subproblems generate feasible directions in which the val-
ues of all the objective functions improve. The decision maker can then indicate
what objective functions to improve at the expense of others, and a new direc-
tion is generated. Tools for extending the applicability of Mukai's algorithms
to nondifferentiable objective functions are presented in Kiwiel (1984, 1985a,
b) and Wang (1989). The ideas were applied in the MPB method in Section
2.2.
The method introduced in Roy and Mackin (1991) is hased on a sequence of
pairwise questions and it tries to approximate the parameters of a proxy value
function.
An example of including ideas from other research areas in interactive mul-
tiobjective optimization is presented in Tapia and Murtagh (1992). The pref-
erences of a decision maker are analyzed in MOLP problems with Markovian
processes. The cardinal priority ranking of Pareto optimal solutions is a part
of the method proposed.
Methods for discrete and continuous problems are combined in Slowinski
(1991). A finite set of Pareto optimal points is generated and then ordinal
regression is applied. The method is intended to be practical in situations where
the decision maker wants to focus on a subset of Pareto optimal points at early
stages of the process. Similar ideas are utilized in Bard (1986). A set of Pareto
optimal solutions is generated by the e-constraint method and ranked by means
of multiattribute decision analysis. The method is demonstrated by an example
concerning the selection of automation options for an upcoming space station.
The interactive step trade-off method combining ideas from the SWT
method and STEM is presented in Yang et al. (1990). It utilizes trade-off rates
and the division of objectives into those to be improved, those that should
maintain their values and those to be impaired. In Yang and Sen (1996), the
interactive step trade-off method is extended to a two-phase algorithm to in-
clude the estimation of piecewise linear value functions based on pairwise com-
parisons.
Possibilities of multiobjective optimization in structural mechanics are pre-
sented in Eschenauer et al. (1989). Two interactive methods are briefly de-
scribed and applied to the optimization of a conical shell.
The Zionts-Wallenius (ZW) method for MOLP problems (see Wallenius and
Zionts (1977) and Zionts and Wallenius (1976, 1983)) based on the weighting
method and optimizing the underlying (implicitly known) value function can
be generalized to convex objective functions by considering piecewise lineariza-
tions. The ZW method is extended for concave (maximization case) objective
 5.13. Other Interactive Methods 213

and value functions and convex feasible regions in Roy and Wallenius (1992).
A more general case of nonlinear objective functions, nonconvex feasible re-
gions and concave value functions is also discussed. This approach uses the
generalized reduced gradient method instead of the original simplex.
The method in Kim and Gal {1993} is intended for MOLP problems. It is
based on the concept of a maximally changeable dominance cone and marginal
rates of substitution. The effectiveness of the method is illustrated by a numer-
ical example.
Ideas for reducing the burden on the decision maker in interactive methods
are introduced in Korhonen et al. (1984) and further developed in Ramesh
et al. (1988). An underlying quasiconcave value function is assumed to exist.
Convex cones are formed according to the preference relations of the decision
maker. The cones are formed so that the solutions in them can be dropped from
further consideration, because they are dominated by some other solutions.
Thus, fewer questions have to be put to the decision maker in charting the
preferences. These ideas concerning convex cones can be applied equally to
multiobjective optimization as to multiattribute decision analysis. The ideas
are utilized, for example, in Ramesh et al. (1989a, b).
A method for complex problems with high dimensionality is proposed in
Baba et al. (1988). The method uses a random optimization method and is
also applicable to nondifferentiable objective functions.
The parameter space investigation (PSI) method is described briefly in
Lieberman (1991b) and in more detail in Statnikov and Matusov (1996) and
Steuer and Sun (1995). It has been developed for complicated nonlinear prob-
lems involving possible differential equations. Such problems occur, for exam-
ple, in engineering. The method is very simple and intended to be applicable
to problems where more sophisticated methods are useless. The PSI method is
a naIve sampling technique rather than an optimization method. Both the con-
straint functions and variables are assumed to have upper and lower bounds.
Thus, the feasible region is a parallelepiped. The Pareto optimal set is approx-
imated by generating randomly uniformly distributed points between the vari-
able bounds. Infeasible solutions are dropped as well as solutions not satisfying
the upper bounds specified by the decision maker. Pareto optimal solutions are
selected from this set. The sample size can be altered and the decision maker
can adjust the upper bounds. The method does not assume differentiability.
It works for nonconvex problems since its structure enables global search. The
method contains a random number generator of its own, but it is claimed in
Steuer and Sun (1995) that any generator can equally well be used. Conver-
gence properties and the accuracy of the approximation of the Pareto optimal
set assuming general Lipschitz conditions are handled in Sobol' and Levitan
(1997) and Statnikov and Matusov (1996). The PSI method has its origins in
the former Soviet Union, which is why most of the information about it has
been published in Russian. It is said to have been applied in many fields of the
national economy in Russia. One engineering application is described in Sobol
(1992).
 1. COMPARING METHODS

As has been stressed many times thus far, a large variety of methods exists
for multiobjective optimization problems and none of them can be claimed to be
superior to the others in every aspect. Selecting a multiobjective optimization
method is a problem with multiple objectives itself. Thus some matters of
comparison and selection between the methods are worth considering.
The theoretical properties of the methods can rather easily be compared. We
summarize some of the features of the interactive methods treated in this book
in a comparative table at the beginning of this chapter. However, in addition
to theoretical properties, practical applicability also plays an important role
in the selection of an appropriate method for the problem to be solved. The
difficulty is that practical applicability is hard to determine without experience
and experimentation.
More fruitful information relating to the question of method selection would
likely emerge if computational applications were more extensively reported. Un-
fortunately, not too many actual computational applications of multiobjective
optimization techniques have been published. Instead, methods have mainly
been presented without computational experiences or with simple academic
test problems. As it is aptly remarked in Bischoff (1986), most of the applica-
tions presented are merely proposals for applications or they deal with highly
idealized problems. For most interactive methods a natural reason is the diffi-
culty (in finding and) in testing with real decision makers. A complicating fact
is also the enormous diversity of decision makers.
One more thing to keep in mind is that for the most part only successful
applications are published. This means that we cannot draw a complete picture
of the applicability of a method on the basis of the experiences reported.
The evident lack of benchmark-type test problems for nonlinear multiobjec-
tive optimization complicates the comparison of different methods. Naturally,
some methods are useful for some problems and other methods for other types
of problems. However, benchmark problems could be used to point out such
behaviour.
In this section we outline some comparisons of methods reported in the
literature. We also consider selected issues in deciding upon a method, including
a decision tree.

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
218 Part III - 1. Comparing the Methods

1.1. Comparative Table of Interactive Methods

Presented

In Figure 1.1.1, we present a comparative table of the twelve interactive

multiobjective optimization methods described in Chapter 5 of Part II. This
can be regarded as a brief summary of these methods. However, a sceptical
attitude should always be taken towards such attempts to compress matters
to an extreme. The table is subjective and there is no point in even trying to
deny it.
Different problems arise when one tries to put together a table of this kind.
Among them are, for example, deciding what property is important enough
to be included, how it should be formulated, and whether it is a positive or a
negative one.
Figure 1.1.1 presents some of the properties described when the methods
were introduced. They are related to the general features of the methods and
their solutions. Properties concerning the nature of the assumptions set in
relation to the problem to be solved have not for the most part been included.
They will be handled in a decision tree in Figure 1.3.1.
The table is by no means self-contained. However, we do not explain the
table here in detail but refer to the corresponding sections were the methods
were presented.
The properties of the methods have been grouped into four categories ac-
cording to different aspects. The first group is related to the general character-
istics of the methods. The style of interaction is treated in the second group.
The third group describes what kind of information is solicited from the de-
cision maker. Finally we indicate those methods whose implementations have
been mentioned in connection with the descriptions of the methods.
For clarity, explanatory comments on some of the properties listed are in
order. As far as the property 'final solution Pareto optimal' is concerned, paren-
theses indicate methods where Pareto optimality depends on the scalarizing
function used.
The distinction between 'classification of objective functions used' and 'ref-
erence points used' must be emphasized. As stated earlier, the classification
and the reference points are parallel. However, classification is somehow more
demanding because it necessitates pointing out both the objective functions to
be relaxed and those to be improved when compared to the current solution.
In Figure 1.1.1, specifying a reference point is understood to be independent
of the current solution in this sense.
The implementation of the reference direction approach is in parentheses
because the program can solve linear problems only.
Some of the properties have been subjectively classified into positive or
negative ones. The other properties are mainly matters of taste.
 1.2. Comparisons Available in the Literature 219

..c
~
80..
~ ~
~ ~ "
0
-g
~ a -g ~1;l '0
-5
"8
~ :t:: "8. -5 ~ ~
1 ~ ~ g
e
~
'" ::E
§
II
g ~
e ~
'";:J
~ ~'"
f-< ~
~
!!l § ~ ~
~
~ 0
f-< i
:::l
~
"
p: ~
~
z
'"
{.;)
'"
+ final solution
Pareto optimal
X X X eX) (x) X ex) ex) X

+ final solution weakly X X X X X X X X X X X

Pareto optimal

+ suitable also for X X X X X X

nondifferentiable problems

ad hoc nature X X X X X X X X

objecti ve functions X X X X X X X X X
assumed to be bounded
- sensitive needing X X X X (X)
consistent answers
- computationally X X
expensive
- difficult questions X X X
posed
+ trade-off rates X X X
provided
comparison of X X X X X X X
alternatives used
classification of X X X X
objective functions used
reference points X X X X
used
marginal rates of X X
substitution used

thresholds used X

implementation X X X
mentioned
eX)

Figure 1.1.1. Properties of interactive methods.

1.2. Comparisons Available in the Literature

Here we briefly mention some of the comparisons available together with

a few results and some conclusions. For more detailed information, see the
references cited.
220 Part III ~ 1. Comparing the Methods

1.2.1. Introduction

The comparisons have been carried out with respect to a variety of crite-
ria. Among them are ease of use and confidence in both the solution obtained
and the method used from the viewpoint of the decision maker. The rapidity
of convergence and CPU time are among the criteria from the mathematical
point of view. The number of Pareto optimal solutions needed to solve a prob-
lem could also serve as a comparison criterion, as pointed out in Ferreira and
Machado (1996). However, such a measure of effectiveness has not generally
been reported in the comparative evaluations available.
Some caution is in order when trying to judge something from the com-
parisons. The comparisons have been performed according to different criteria
and under varied circumstances. Thus they are not fully proportional. Which
method is the most suitable for a certain problem depends highly on the per-
sonality of the decision maker and on the problem to be solved.
Practical experience is especially important in evaluating the techniques
with respect to criteria related to the decision maker. It is important to com-
pare a method under a variety of circumstances so that the conclusions can be
generalized. As emphasized, for example, in Hobbs et al. (1992), the appropri-
ateness, ease of use and validity of a method must be tested with real decision
makers.
Using a human decision maker does not, however, mean that the practi-
cal applicability of the method has been fully investigated. Unfortunately, few
experiments have been reported with problem-related decision makers. Most
of the comparative evaluations with human decision makers have involved stu-
dents as the decision makers. This is understandable for practical reasons. How-
ever, this kind of a setting can be called into question. The results might have
been different with real decision makers who are actually responsible for the
solution obtained. Another aspect is the wide range of different problem ar-
eas and their different decision makers. Obvious examples are business-related
problems that typically involve less than ten objective functions and engineer-
ing design problems with hundreds of objective functions. Further aspects to be
kept in mind when testing several methods with human decision makers are the
effects of learning and anchoring. Learning is related both to getting to know
the problem- better and to the order of the methods used whereas anchoring is
related to the selection of starting points.
Instead of a human decision maker one can sometimes employ value func-
tions in the comparisons. Value functions may be useful in evaluating theo-
retical performance, but such tests do not fully reflect the real usefulness of
the methods. One can try to compensate for the lack of a real decision maker
by employing several different value functions. If, for example, marginal rates
of substitution are desired, the inconsistency and inaccurate responses of a
decision maker can be imitated by multiplying them with different random
numbers. These means are employed in Shin and Ravindran (1992). On the
other hand, value functions cannot really help in testing ad hoc methods.
 1.2. Comparisons Available in the Literature 221

One crucial factor that can affect the performance of the methods in the
comparisons is the user interface. Nothing is usually mentioned concerning the
realization of the user interface in the comparisons reported. It is important
to remember that one can spoil a 'good' method with a poor user interface or
support a 'poor' method with a good interface. In addition to the illustration
of the (intermediate) results, a good user interface also means a clear and
intelligible input phase.
It is interesting to observe that most of the multiobjective optimization
problems solved when testing the methods (and reported in the literature)
have been linear. It is true that complex nonlinear functions cause difficulties of
their own and the characteristics of the solution methods may be disturbed. On
the other hand, features concerning nonlinear problems may remain unnoticed
with MOLP problems. On the whole, the comparisons available are not of too
much help if one is looking for a method for a nonlinear problem, and more
contributions in this area are needed. Nevertheless, we review some of the
comparisons published.

1.2.2. Noninteractive Tests

An MOLP problem for determining the most economical combination of

grape growing and wine production in Hungary is solved by the weighting
method, the e--constraint method, lexicographic ordering and the weighted L 1 -
and L 2 -metrics with normalized objective functions in Szidarovszky and Szen-
teleki (1987). It is observed that different solutions are obtained with each
method. It is also stated that the weighted L1 - and L 2 -metrics with normal-
ized objective functions produce the most uniform distribution of objective
vectors. Finally, the weighted L1 -metric is seen as the most convenient way for
generating Pareto optimal solutions in large-scale MOLP problems.
A linear problem in the mining industry is solved in Peterson (1984) by the
weighting method, the e--constraint method, the method of weighted metrics
with and without denominators, and by lexicographic ordering. The solutions
obtained from the other methods are utilized in the method of weighted metrics
and all the solutions are analysed. The conclusion is that solution methods
should be applied so that they complement each other.

1.2.3. Interactive Tests with Human Decision Makers

No interactive methods were included in the comparisons mentioned thus

far. The following comparisons involve interactive methods.
It is described in Dyer (1973b) how nine (student) decision makers were
presented with an MOLP problem involving choosing an engine for a car. They
were first asked to suggest an approach and then compare it with the GDF
method and a trial-and-error procedure. In the trial-and-error procedure the
decision maker was simply asked to enter an objective vector and the procedure
222 Part III - 1. Comparing the Methods

stated whether it was feasible or not. The decision makers were assumed to ex-
plore the feasible objective region until they were unable to find more preferred
solutions.
The criteria in the evaluation were the ease of using the method and the
confidence in the solution obtained. The results obtained favoured the GDF
method. Thus, a conclusion could be drawn that the GDF method can success-
fully be used by untrained decision makers.
The performance of the GDF method, STEM and the trial-and-error pro-
cedure (the same as that used by Dyer) is compared from the point of view of
a decision maker in Wallenius (1975). A total of 36 business school students
and managers from industry were employed as decision makers. The following
aspects of the methods were compared: the decision maker's confidence in the
solution obtained, ease of use and understanding of the method, usefulness of
the information provided, and rapidity of convergence. The linear management
problem to be solved contained three objective functions.
The results are analysed statistically in Wallenius (1975). One interesting
conclusion was how well the trial-and-error procedure competed with the more
sophisticated methods. Nevertheless, Wallen ius points out that its performance
might be weakened if the problems were more complex. Difficulties in estimating
the marginal rates of substitution weakened the overall performance of the GDF
method. Thus, Wallenius suggests that research should be directed to finding
ways of better adjusting methods to suit the characteristics of a human decision
maker.
The results of Dyer and Wallenius concerning the GDF method differ re-
markably. Some trials analysing the reasons are presented in Wallenius (1975).
The capabilities of the ZW, the SWT, the Tchebycheff and the GUESS
methods (without the upper and lower bounds) are compared in Buchanan and
Daellenbach (1987) from the point of view of the user in solving a linear three-
objective optimization problem. The problem concerned the production of the
electrical components of lamps. A total of 24 decision makers (students and
academic staff) were employed. The criteria in the comparison were partly the
same as those used by Wallenius. In addition to confidence in the final solution,
ease of use and ease of understanding the logic of the method, CPU and elapsed
time were compared. The most important criterion was the relative preference
for using each method. The conclusions are that the Tchebycheff method was
clearly preferred to the other methods and the ZW method came out the worst
in relation to the first four criteria. The SWT method was in the middle. The
GUESS method performed surprisingly well. On the basis of this experiment
one can say that decision makers seem to prefer solution methods where they
can feel that they are in control.
Experimental evaluations of interactive methods with 24 decision makers
(students) and two three-objective MOLP problems are reported in Buchanan
(1994). The methods involved were the Tchebycheff method, the GUESS
method and the simplified interactive multiple objective linear programming
 1.2. Comparisons Available in the Literature 223

(SIMOLP) method (by Reeves and Franz) based on fitting a hyperplane

through a set of extreme point solutions. The criteria of the evaluation were
basically the same as in Buchanan and Daellenbach (1987).
A two-stage solution approach was tested where an introductory phase was
performed before employing the actual solution methods. In the introductory
phase a set of Pareto optimal solutions was generated and the decision makers
were asked to provide a rating for each solution. A linear regression model was
fitted according to the ratings and the decision maker was asked for feedback
as to the correctness of the preference structure formed.
The motivation for using two phases was the following. It has been suggested
at times that the opportunity of familiarizing the decision maker with the
problem to be solved should improve the actual solution process. However, the
experiment did not support this assumption. Either the form of the preference
structure to be fitted was incorrect or the ratings were too difficult to provide.
The methods tested had remarkable philosophical differences. SIMOLP was
the most structured and GUESS the most unstructured method. In general, the
GUESS method was the most favoured of the three. Added to this, the fact that
SIMOLP was the least popular method, there is clearly a preference for less
structure in the solution method. The SIMOLP method required the decision
maker to select the least preferred solution from a given set. This proved out to
be difficult. It seemed that it is easier to select the best rather than the worst
alternative. (This observation must be related to the nature of the problem to
be solved. The opposite may be true in some other cases.)
It is interesting that the Tchebycheff method was rated better than the
GUESS method in Buchanan and Daellenbach (1987), whereas GUESS was
better than the Tchebycheff method in Buchanan (1994). In the latter test,
the GUESS method had been supplemented by allowing the decision maker
to specify upper and lower bounds for the objective functions. This may have
improved the functionality of the method. Once again, the decision makers
liked being in control of the solution process. However, they would have liked
to receive pairwise trade-off information to support the process.
The decision makers were also asked to state the solution method they
would prefer to use next time. The most important element in the preference
proved to be familiarity with the method.
The method of Steuer and the ZW method (in Subsection 5.13.4 of Part II)
are compared in Michalowski (1987). Five decision makers from the planning
department of a factory were employed. A linear production planning problem
with three objective functions was solved and the evaluation criteria were not
fixed in advance, although the main interest was in the decision phase. The
decision makers had critical comments concerning both the methods, and each
of them obtained a different final solution. One can say that the decision pro-
cesses by the ZW method terminated slightly faster than those by the method
of Steuer.
224 Part III - 1. Comparing the Methods

The method of Steuer and STEM are tested in Brockhoff (1985). A total of
147 decision makers were employed to solve six problems involving purchasing
cars. The results and progress are analysed according to several criteria, with
the method of Steuer emerging with the best outcomes on the average.
An experiment on the differences in the philosophies of methods for con-
tinuous compared to discrete problems is presented in Corner and Buchanan
(1995,1997). In Corner and Buchanan (1997), the continuous GUESS method,
a modified ZW method and a discrete SMART method (based on construct-
ing a value function) were used to solve a production planning problem with
three objective functions by 84 undergraduate students as decision makers. The
problem was nonlinear and had continuous variables. The main interest was to
determine the ability of the methods to capture the preferences of the decision
maker. In other words, how well the methods were able to find desirable solu-
tions and how much the decision makers liked the methods. The time spent on
each solution process was also recorded.
One of the conclusions is that the continuous methods were better and
faster than the discrete method. The GUESS method was rated easiest to use
and to understand. All the methods produced different solutions of which the
one generated by the GUESS method was ranked the best. The order of the
methods used was found to have no effects on the results. The exception was
the case when SMART was used first. Then the solutions obtained with the
other methods were statistically the same. In addition, it was observed that a
weighted additive value function explaining their ranking behaviour could be
found for most decision makers.
Another experimental test involving the ZW method and the GUESS
method is reported in Buchanan and Corner (1997). The emphasis was in test-
ing whether any anchoring effect can be explained by the structure of the
solution method. A number of 84 students acted as decision makers and solved
a nonlinear problem with three objective functions. The conclusion was that
an anchoring effect could be seen with the structured ZW method but less so
with the free search method GUESS. Thus, it can be deduced that the selection
of the starting point is even more crucial with more structured methods than
with less structured methods.
Some comparisons of continuous and discrete methods are also presented
in Korhonen and Wallenius (1989b). A continuous MOLP problem with five
objective functions concerning the allocation of a student's time between study,
work and leisure was solved by 65 student decision makers. The five methods
compared were all based on the reference direction approach. Only the speci-
fication of the reference direction varied. The original way of using aspiration
levels was found to be clearly superior to the others.
A more detailed review of the above-described and some other empirical
studies involving real decision makers is given in Olson (1992). However, no
final conclusions can be drawn from the experiments. The reason is that the
test settings and the samples are not similar enough.
 1.2. Comparisons Available in the Literature 225

1.2.4. Interactive Tests with Value Functions

Next we review some comparisons utilizing value functions to replace deci-
sion makers.
An MOLP problem with three objective functions concerning operations
planning in the natural gas business is solved by several methods in Mote et
al. (1988). A nonlinear value function was employed instead of human decision
makers. The problem was solved by the GDF, the SWT and the ZW methods,
STEM, goal programming, and the method of Steuer. Only standard LP codes
were utilized in the calculations. No single technique was shown to be superior.
The methods had differences concerning the burden upon the decision maker
and ad hoc and non ad hoc properties.
The method of Steuer and the method of Franz (an interactive adapta-
tion of weighted and lexicographic goal programming) are compared in Gibson
et al. (1987) in solving several randomly generated MOLP problems. Different
value functions were used to replace the decision maker. The aim of the compar-
ison was to investigate the applicability of the methods to different situations
with the help of statistical tests. The number of iterations was also recorded.
The conclusion is that, for example, the number of iterations and whether all
the objective functions are of relatively equal importance or not are important
in the selection of a method. These criteria lead to different recommendations
respecting methods.
The Tchebycheff method and the SIMOLP method (by Reeves and Franz)
are compared in 15 MOLP test problems with four objective functions using
both linear and nonlinear value functions to replace the decision maker in
Reeves and Gonzalez (1989). The comparison criteria were the quality of the
solution (how far the best solution found was from the best extreme point),
user-friendliness, computational requirements, whether nonextreme solutions
could be found, number of iterations needed and flexibility. The Tchebycheff
method was used in the comparison because of its promising performance in
the test reported in Buchanan and Daellenbach (1987). The main difference
between the two methods is that the SIMOLP method moves away from the
least preferred alternative whereas the Tchebycheff method moves toward the
most preferred one. Thus, the SIMOLP method is more flexible and it is easier
for the decision maker to change her or his mind. Further, the SIMOLP method
needs much less calculation.
The SIMOLP method was able to find slightly better solutions at less com-
putational cost in most problems even with the nonlinear value function. The
fact that the SIMOLP method is limited to Pareto optimal extreme points
did not seem important in the tests. However, Reeves and Gonzalez (1989)
suggest combining the advantages of both methods. Either the decision maker
can choose at each iteration which method to utilize for the next iteration, or
the flexibility of the SIMOLP method can be used first and the ability of the
Tchebycheff method to produce nonextreme solutions can be used in the last
iterations.
226 Part III - 1. Comparing the Methods

Another summary of comparisons published is given in Aksoy et al. (1996).

The presentation summarizes six comparisons with human decision makers and
fourteen studies utilizing value functions. The aspects treated are the compari-
son criteria, types of decision makers, nature of the test problems to be solved,
form of value functions, starting solutions, stopping criteria, and ordering of
methods and test problems. The obvious conclusion from this is that there is
an urgent need for further comparative evaluations of nonlinear multiobjective
optimization methods.

1.2.5. Comparisons Based on Intuition

A characteristic shared by the evaluations to be described next is that they

are based on intuition and insight rather than practical experiences and tests.
The comparative table in Section 1.1 could equally have been included here.
A collection of the features of five nonlinear interactive methods is presented
in Masud and Zheng (1989). The methods are compared with regard to eleven
items, for example, the certainty of obtaining a Pareto optimal solution, the
optimization technique used, the type of information required from the deci-
sion maker, computational complexity compared to the GDF method, and the
number of iterations needed with input from the decision maker compared to
the GDF method. A similar table comparing the decision maker's burden, ease
in actual use, effectiveness and handling of inconsistency is collected in Shin
and Ravindran (1991) for ten methods. A further classification and evaluation
of methods according to 21 criteria is given in Rietveld (1980).
The number of items a decision maker has to assess simultaneously and
per iteration for eight different methods in a medium size linear problem are
tabulated in Kok (1985). It is concluded that the method of displaced ideal,
the interactive multiple goal programming method and STEM are promising
because their presumptions are realistic. In Kok (1986), the learning effects,
information load, effort of technical support and group decision capabilities are
evaluated for five methods: ZW, interactive multiple goal programming, ISWT,
STOM and pairwise-comparisons. No strict preference can be expressed.
A total of 19 interactive methods for MOLP problems are listed according
to three characteristics in Larichev et al. (1987). The characteristics are the
reliability of the way information is elicited from the decision maker, insignif-
icant sensitivity to random errors on the part of the decision maker and good
speed of convergence. The basic principles of the methods are also introduced.
In addition, the features of STEM, the GDF and the Tchebycheff methods, the
reference point method and the reference direction approach, among others,
are tabulated in Vanderpooten and Vincke (1989) and Vincke (1992, p. 105).
The criteria are, for instance, assumptions of the existence of a value function,
applicability, trial and error support, mathematical convergence, the number
of questions posed and the computational burden.
The bi-reference procedure is compared to STEM, the GDF method, the
ZW method and the reference direction approach in Michalowski and Szapiro
 1.3. Selecting a Method 227

(1992). The idea is to compare the performance of the bi-reference procedure

to the published results of the other methods.
Finally, we mention some other comparative studies. Characteristic values
in optimizing the multiobjective layout of a conical shell by the GDF method,
STEM and three other methods are reported in Eschenauer et al. (1990b).
As far as the relative performance of STOM and the Tchebycheff method is
concerned in finding a solution to a linear sausage blending problem in Olson
(1993), the main intention is to emphasize the power of the weighted Tcheby-
cheff metric in multiobjective optimization.

1.3. Selecting a Method

Choosing an appropriate solution method for a certain multiobjective op-

timization problem is not easy, as has been made abundantly clear. None of
the existing methods can be labelled as the best for every situation, since there
is a multiplicity of aspects to consider and many of the comparison criteria
are of a somewhat fuzzy character. The features of the problem to be solved
and the capabilities and the type of the decision maker have to be charted be-
fore a solution method can be chosen. Some methods may suit some problems
and some decision makers better than others. Let us sum up by offering some
general guidelines and a decision tree.

1.3.1. General Guidelines

Several different comparison criteria were already mentioned in Section 1.2

in connection with the tests reported. Some of the criteria to consider when
evaluating methods are also collected in Hobbs (1986). These selection criteria
are appropriateness, ease of use, validity and the sensitivity of the results to the
choice of method. Appropriateness means that the method is appropriate to the
problem to be solved, to the people who are to use it and to the institutional
setting where it is to be implemented. Ease of use refers to the effort and the
knowledge required from the analyst and the decision maker. Validity means
that the method measures what it is supposed to and the assumptions set are
consistent with reality. The sensitivity of the results to the choice of method
expresses the desire that solutions obtained by the method do not significantly
differ from those of other methods. If the method chosen has a significant
effect on decisions, then the relative validity of different methods should be
considered. If the form of the method does not matter, then the most important
criteria are ease of use and appropriateness.
The number of crucial criteria in selecting a solution method is reduced to
three in Stewart (1992). The input required from the decision maker must be
meaningful and unequivocal, the method must be as transparent as possible
and it must be simple and efficient.
228 Part III - 1. Comparing the Methods

The role of the decision maker is important and should be taken seriously.
Many experiments have shown that decision makers prefer simpler methods
because they can more easily understand such methods and they feel more in
control. The valuation placed on some methods may increase if the decision
makers can practice using them or obtain advice. An important fact to keep
in mind is that theoretically irrelevant aspects, such as question phrasing, may
affect the confidence that the decision maker feels in the method. The concept
of the decision maker's confidence is analysed further in Bischoff (1986).
Other important criteria for the decision maker in selecting the solution
method are, for example, the simplicity of the concepts involved, possibilities of
interaction, the ease with which the results can be interpreted and the chances
of choosing the most preferred solution from a wide enough set of alternatives.
The method must also fit the decision maker's way of thinking. The language of
communication between the decision maker and the method (solution system)
must be understandable to the decision maker. (S)he wants also to see that the
information (s)he provides has a (desirable) effect on the solutions obtained.
One more element, not mentioned thus far, in the selection of a method is
how well the decision maker knows the problem to be solved. If (s)he does not
know its limitations, possibilities and potentialities well, (s)he needs a method
that can provide support in getting acquainted with the problem. In the oppo-
site case, a method that makes it possible to focus directly on some interesting
sector is advisable. Ways of identifying appropriate methods for different types
of decision makers are needed.

1.3.2. Method Selection Tools

Few universally applicable guidelines have been given for the method se-
lection problem in the literature. Let us mention some of them including even
approaches for discrete problems.
An attempt to assist in the selection of a solution method is presented in
Gershon and Duckstein (1983). The selection problem is modelled as a multiob-
jective optimization problem. A set of 28 criteria for the selection are suggested
and they are divided into four groups. Only the criteria in the last group have to
be considered every time the selection algorithm is applied. The criteria take
into account the characteristics of the problem, the decision maker and the
methods. Many types of problems are taken into consideration in the criteria
(e.g., discrete and continuous variables). The model contains 13 solution meth-
ods from which to select. The set of methods can naturally be modified. The
number of selection criteria can also be varied to include only those relevant
to the problem to be solved. Finally, after the methods have been evaluated
according to the selection criteria, the resulting multiobjective optimization
problem is solved by the method of the global criterion (e.g., L1-metric).
A related procedure is suggested in Tecle and Duckstein (1992). There, a
set of 15 methods is evaluated with respect to 24 criteria in four classes. The
weighted Lp-metric is used in each class and another weighted Lp-metric is used
 1.3. Selecting a Method 229

to combine the classes and obtain the best method. For the example problem
provided, the weighted Lp-metric turns out to be the best method. One may
wonder whether the weighted Lp-metric favours itself or whether this is a mere
coincidence. Some critique of the approach is also expressed in Romero (1997).
Different decision trees and rules for providing assistance in selecting a
method for multiattribute decision analysis problems are described in Hwang
and Yoon (1981) and Teghem et al. (1989). However, as criticized in Ozernoy
(1992a), to design a comprehensive and versatile decision tree usually results
in an explosion in the number of nodes. Another problem with decision tree
diagrams is what to do when the user answers 'I do not know.'
An expert system for advising in the selection of solution methods for prob-
lems with discrete alternatives in proposed in Jelassi and Ozernoy (1989). Steps
in the development of another expert system for selecting the most appropri-
ate method for discrete problems are described in Ozernoy (1992a, b). The
questions posed by the system are based on if/then rules. They lead to recom-
mending a method or stating that no method can be recommended. The user
of the system can also always ask why a particular question is posed.

1.3.3. Decision Tree

Little advice exists for selecting a method for nonlinear and continuous prob-
lems. Therefore, despite the above-mentioned pitfalls and faults in decision-tree
diagrams, we nevertheless present one in Figure 1.3.1. The tree has primarily
been created on the basis of plain theoretical facts concerning the assumptions
imposed by the methods on the problem to be solved and secondarily according
to the preferences of the decision maker. Because of space limitations it has not
been possible to include all the properties.
The decision tree includes the twelve interactive methods described in Part
II. Only those methods are included that have been presented in more detail
or whose main features have been introduced. Remember that in practice, the
functioning of a method may not always require that all its technical assump-
tions are satisfied (as stated, for example, in Zionts (1997a, b)). Or it may even
be impossible to verify all the assumptions. If some of the assumptions are
not valid, some of the results may be incorrect, but this does not necessarily
mean that the method will not work in some contexts. The results may still be
adequate for practical purposes. This must be kept in mind when studying the
decision tree. Nonetheless, the assumptions provide some guidelines to follow.
The starting node is situated on the left. The tree diagram has been created
in such a way that only the answers 'yes' or 'no' are possible. Whenever the
immediate answer is 'I do not know,' the answer 'no' can be given. In order to
avoid confusing the picture any further, the words yes and no have been replaced
by arrows of different types. Continuous lines represent positive answers and
broken lines stand for negative ones. In addition, 'no' arrows always leave a
node to the right of the 'yes' arrows.
 to.:>
objective functions marginal rates c:.>
o
continuously ----1
"'...... of substitution
differentiable ~ to be specified
"'0
NIMBUS method
®'" opinions ~
'. ' available about .....
.....
_ ... _._ .... __ ._ ' ... -.~ trade-off rates .....
y , ~

t. (~) , rreference points or ~

alternatives classification to be used o
functions desired for ...
~
continuously comparison
S
'0
decreasing,
... IU OrongO differentiable
J. .,'\.
oq
r ~ .
e:S·
r:: continuously Oq
~
('D differentiable ...::r
I-' II>
~ s::::
I-' II>
S U exists, ...
::r
o
C-
T is
~
('D A oo
(1) implicitly steps in reference direction
R points or desired, also aspiration
c. known HIRDmethod I
T classification
i»'
()q
... ,
-" -..,,
\ \
objective
functions
to be used
levels to be specified

;teps in reference direction desired, RD methOd~

~ bounded also aspiration levels to be specified .....
(§J'
\\ reference point method ~ISTEMI
alternatives
functions twice
desired for
continuously no amouns of relaxation to
comparison
differentiable be specified, no steps in
...... reference direction desired
.,....... , - - - - - - - - . ,
for comparison
I~~s ._. ._~I
 1.3. Selecting a Method 231

The nodes containing only capital letters are used in two different cases.
The first is to avoid repetition. In the second, no method can be found along
the path followed. In that case, one can try another path. The aim has been to
allow as many previous answers as possible to be exploited. Thus, some dead
ends may be avoided.
The same method may be reached by following different paths. In this case
varying questions may be needed in order to separate the methods. That is
why either general or more detailed questions leading to the same method are
used.
As already repeated several times, selecting the solution method is a difficult
and important task. After describing each method (see Part II), we have tried
to indicate what it has in its favour and what its drawbacks are. These matters,
of course, are always more or less subjective.
 2. SOFTWARE

The development of computers and the improvement in the speed, stor-

age capacities and flexibility of computing facilities have made it possible to
produce more sophisticated and demanding software for solving multiobjec-
tive optimization problems. Efficient computers enable, for example, the im-
plementation of interactive algorithms, since they can produce sufficiently fast
responses for the decision maker without the user getting frustrated waiting.
Nevertheless, taking into account the multiplicity of methods developed for
solving nonlinear multiobjective optimization problems, the number of widely
tested and user-friendly computer programs that are generally available is
small. At least they are difficult to find. Most implementations are done for
academic testing purposes and their existence is not advertised. In other words,
there is a real need for functional and reliable software for solving nonlinear
multiobjective optimization problems.

2.1. Introduction

Most of the software packages developed for multiobjective optimization

problems can be termed multiobjective decision support systems, and they
form one class of decision support systems. Decision support systems (DSSs)
can be defined as interactive computer-based systems designed for helping and
assisting in the decision-making process. Their main objects are to help decision
makers in solving problems more efficiently and making better decisions.
The main components of a decision support system are a model, an op-
timizer (solver) and an interface between the model, the optimizer and the
user. By an interface we mean the input language and style, exchange of in-
formation and presentation of the results. It should be remembered that the
human-computer interface must be designed with at least as much care and
effort as the other components of the system.
Another, similar characterization of a decision support system is that it
consists of a model, by which is meant a mathematical algorithm, data man-
agement and interface. It is essential that the model reflects the preference
structure of the user. That is why it is important that users take an active role
in developing decision support systems.

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
234 Part III - 2. Software

The role and the requirements of the model, the optimizer and the interface
in the multiobjective optimization environment are outlined, for example, in
Jelassi et al. (1985). It is useful to have capabilities of self-learning and model
updating in a decision support system. The interface is an important factor in
relation to the user-friendliness of the system.
One can state that developing software for multiobjective optimization
problems is once again a multiobjective optimization problem in itself, and
proper planning is essential. Several (conflicting) objectives to be taken into
consideration in multiobjective software design and realization are collected
in Olkucu (1989). Among them are a short development time, long product
life, easy and cheap maintenance, reliable implementation of the algorithm,
an efficient user interface and a large number of potential users. Of no mi-
nor importance in this regard are the selection of the realization environment
(including the operating system) and development tools.
Features to be taken into consideration when designing decision support
systems are also handled in Lewandowski (1986). Different definitions of user-
friendliness and rules for dialogue design are given.
It should be pointed out that while a great deal of effort has gone into
developing the methodological and computational aspects of the systems, the
interface between the system and its user is often of poor quality. This is a
serious weakness, since no matter how brilliant the methodology and its imple-
mentation are, it will be discarded if the interface does not suit the user. In any
case, the algorithms must be implemented in such a manner that computer-
technical requirements do not overshadow the real problem and non-skilled
persons can also use the programs. One way to try to improve the situation
is to provide different interface possibilities for the same system for computer
specialists, trained users and average users.
An effort of measuring the effectiveness of decision support systems is de-
scribed in Sainfort et al. (1990). Even though it is widely assumed that decision
support systems really do help in decision making and problem solving, research
results in this important area are few. Added to this is the fact that there is
currently no general theory about problem solving because of its complexity.
Group decision support systems are mostly handled in Sainfort et al. (1990),
but the conclusions favouring decision support systems are general. It is demon-
strated that decision support systems increase the understanding of the prob-
lem, reduce frustration in the problem solving and contribute to progress in
the solution process.
To put if briefly, a decision support system should be easy to use, it should
capture the thinking procedure of the decision maker, it should support differ-
ent decision styles and it should help the decision maker to structure different
situations. Other desirable characteristics of decision support systems are listed
in Weistroffer and Narula (1997).
 2.2. Review 235

The user of the decision support system may need guidance and training
to be able fully to make the most of it. Such a step may increase usability and
render the system more user-friendly than before.
An interesting question is raised in Verkama and Heiskanen (1996) about
how research concerning both methodology and decision support system soft-
ware should be reported in the literature. Verkama and Heiskanen suggest that
numerical examples should accompany both the algorithms proposed and the
software described. This would enable any interested reader to use the example
to understand details not presented in the paper.

2.2. Review

Existing software packages up to the year 1980 are listed in Hwang et

al. (1980). These programs were mainly developed for linear and goal pro-
gramming problems. They are rather primitive when compared with modern
computer facilities. A somewhat more up-to-date list of decision support sys-
tems developed to aid in multiobjective optimization and multiattribute deci-
sion analysis problems up to the year 1988 is collected in Eom (1989). However,
the presentation is only cursory. In addition, a classification of the system ap-
plications is provided. Some software implementations are also mentioned in
Weistroffer and Narula (1991), whereas the overview in Buede (1996) handles
software for discrete problems only.
The latest state of decision support systems for multiple criteria decision-
making problems up to the year 1997 is presented in Weistroffer and Narula
(1997). Systems for both continuous and discrete problems are listed with in-
formation about where they can be obtained. Unfortunately, from among the
seventeen continuous products mentioned only six are applicable to general
nonlinear multiobjective optimization problems. Of these GRS generates and
illustrates the Pareto optimal set implementing the generalized reachable sets
method (see Section 3.6 in Part II). The actual solvers are CAMOS, DIDAS,
LBS, MONP-16 and NIMBUS, where LPS is an implementation of the light
beam search (see Section 5.9 of Part II). The others have already been in-
troduced in Part II except for CAMOS (see the end of this section). DIDAS,
the implementation of the reference point method, was described in Subsection
5.6.4, MONP-16 implementing STOM was mentioned in Subsection 5.8.4 and
implementations of NIMBUS were handled in Subsection 5.12.8.
The current state of the software development can be inquired from A. Lo-
tov, Russia (GRS), A. Osyczka, Poland/Japan (CAMOS), J. Granat, Poland
(DIDAS), A. Jaszkiewicz, Poland (LBS), L. Kirilov, Bulgaria (MONP-16) and
K. Miettinen, Finland (NIMBUS). The WWW-NIMBUS system is available at
http://nimbus.math.jyu.fi/.
Software for discrete problems is more easily available than software for con-
tinuous problems. For example, several discrete systems have been developed
236 Part III - 2. Software

into commercial products. Yet, as emphasized in Buede (1996), even those soft-
ware developers concentrate too closely on features of analysis at the expense
of user-friendliness, as mentioned earlier.
Among software products for solving MOLP problems is VIG by P. Kor-
honen, Finland (see Subsection 5.10.3 in Part II). Let us also mention a pack-
age of subroutines, called ADBASE, by R.E. Steuer, USA (see Steuer (1986,
pp. 254-267)). ADBASE contains, for example, tools for generating Pareto op-
timal extreme points. These are examples of the generally available products
for linear problems.
The situation is worse for continuous nonlinear problems. Most of the soft-
ware implementing the extensive amount of existing multiobjective optimiza-
tion methods is neither commonly available nor widely known.
One explanation sometimes mentioned is the lack of a free and reliable
nonlinear solver that could be integrated and distributed with the software.
Most software products have been implemented for academic testing purposes
and have not been updated along with the development of computer facilities.
Consequently, their existence is not advertised. Simply designing and realizing
a functional user interface is demanding. One must assume that the need has
not been large enough to motivate the work, or the need has not been realized
because good solution tools have not been available.
However, some implementations were mentioned in connection with the
method descriptions in Part II. No detailed information was given, since the
implementations are under continuous development and the details may be
out-of-date at any moment.
Software comparisons reported in the literature mainly concern programs
for multiattribute decision analysis. We simply mention that seven microcom-
puter implementations are presented and compared in Colson and de Bruyn
(1987). Five of them are intended for multiattribute decision analysis. An im-
plementation of STEM is also reported. In addition, the main features and
requirement.s of eight microcomputer software packages are introduced in Lotfi
and Teich (1989). One of them is VIG and the other seven are for discrete
al ternati ves.
There exist several software packages for general single objective optimiza-
tion problems that also contain some possibilities for noninteractive multiobjec-
tive optimization. Let us briefly indicate some of them. The implementation of
the MPB method (see Section 2.2 in Part II), called MPBNGC by M. MakeUi,
Finland, is among them.
CAMOS (Computer Aided Multicriterion Optimization System) has been
developed to treat especially nonlinear computer aided optimal design problems
(see Osyczka (1989b, 1992) and for an earlier version Osyczka (1984)). CAM OS
produces Pareto optimal solutions with different generating methods.
The methods for identifying (weakly) Pareto optimal solutions are the
weighting method with or without normalizing the objective function, the
c-constraint method, the method of the global criterion and the method of
 2.2. Review 237

weighted Tchebycheff metric. Problem (2.1.4) of Part II is also used. For more
details, see, for example, Osyczka (1984, 1992). Different underlying single ob-
jective optimization algorithms may be used.
The functioning of CAMOS is illustrated by two practical problems in Osy-
czka (1992, pp. 93-125). They are the optimal design of multiple clutch brakes
and the optimal counterweight balancing of robot arms.
NOA, a collection of subroutines for minimizing nondifferentiable functions
subject to linear and nonlinear (nondifferentiable) constraints, is described in
Kiwiel and Stachurski (1989). NOA is applicable to multiobjective optimization
problems since the single objective function to be minimized is assumed to be a
maximum of several functions. Thus, for example, some achievement functions
can be optimized.
Let us finally mention the optimization toolbox of the MATLAB system
including the weighting method, the E-constraint method and a modification
of goal programming. Naturally, other multiobjective optimization algorithms
may be coded within the MATLAB environment, taking advantage of the pow-
erful single objective solvers and graphics available.
 3. GRAPHICAL ILLUSTRATION

3.1. Introduction

Graphical illustration plays an essential role when designing modern soft-

ware user interfaces. Graphics may be used to describe the problem, to assist
the decision maker in specifying values for problem parameters or to illustrate
the contents and the meaning of questions posed by the algorithms. In such
realizations, the upper limit lies in one's imagination.
In spite of the more general possibilities, we restrict our treatment in this
chapter. By graphical illustration we here mean the ways of presenting several
alternative objective vectors to the decision maker. To be convinced of the need
for such illustration one has only to examine the interactive methods described
in Part II. Good graphical illustration helps the decision maker to gain a better
insight into the problem and the different alternatives generated.
As computers have developed, more attention has been paid towards the
role and the possibilities of computer graphics in building human-computer in-
terfaces. Nevertheless, utilizing graphical illustration does not mean that the
limits on human information processing capacity are transcended. This means
that there is no sense in trying to offer too many objective vectors for evalua-
tion, no matter how clear the illustrations are.
Several psychological tests are summarized in Miller (1956) to prove that
the span of absolute judgment and the span of immediate memory in human
beings is rather limited. We cannot receive, process or remember large amounts
of information. The magical number seven plus or minus two appears in several
tests and in several ways. However, no number can be regarded as an absolute
limit. Everything depends on the circumstances. Still, the findings of Miller are
to be kept in mind when deciding the number of alternatives to be presented to
the decision maker or the number of objective functions to be treated (if these
can be affected). Miller's findings must also be remembered when expecting
exact information from the decision maker. Let us mention that, for example,
seven ways of decreasing the number of alternatives are presented in Graves et
al. (1992).
Naturally, many different ways for illustrating objective vectors can be
thought of. However, elegant graphics must not be an end in itself. The graph-
ics must be easy to comprehend by the decision maker. On the one hand, not
K. Miettinen, Nonlinear Multiobjective Optimization
© Springer Science+Business Media New York 1998
240 Part III - 3. Graphical Illustration

too much information should be allowed to be lost and, on the other hand, no
extra unintentional information should be included in the presentation.

3.2. Illustrating the Pareto Optimal Set

In the case of two objective functions, graphical illustration of the objective

space is effective. The feasible objective region and, especially, its Pareto opti-
mal subspace can be sketched on a plane. If this is not possible, the available
objective vectors can be plotted in the objective space. As far as three objective
functions are concerned, the Pareto optimal set can be expressed by three pro-
jections on a plane, as suggested, for example, in Meisel (1973). However, the
interpretation of such information is far more difficult for the decision maker.
Another way of illustrating the Pareto optimal set of three objective func-
tions is to draw a two-dimensional plot with fixed values assigned to the third
objective function. There is a resemblance here with topographic maps. Such
an approach is handled in Bushenkov et al. (1995) and Lotov et al. (1997),
where so-called decision maps are used. Several level sets of the third objective
function are drawn in the picture of the Pareto optimal hull of the first two
objective functions. These sets are called efficiency frontiers. If there are more
than three objective functions, several different pictures can be drawn each
having fixed values for the other objective functions. For example, in the case
of five objective functions a matrix of decision maps may be displayed. There,
the fourth objective function has the same fixed value in every picture in each
row and the fifth objective function has the same fixed value in every picture
in each column. In addition, scroll-bars and animations can be used. According
to its developers, this approach works for up till seven objective functions.

3.3. Illustrating a Set of Alternatives

Below, we present some ways of illustrating a set of alternative objective

vectors graphically. Some of the ways are clarified by applying them to an
example of three alternative objective vectors of a problem with three objective
functions.

3.3.1. Value Path

A widely used way of representing sets of objective vectors is to use value

paths, as suggested, for example, in Geoffrion et al. (1972) and Schilling et
al. (1983). This means that horizontal lines of different colours or of different
line styles represent the values of the objective functions at different alter-
natives. In other words, one line displays one alternative. This is depicted in
Figure 3.3.1. The bars in the figure show the ranges of the objective functions
 3.3. Illustrating a Set of Alternatives 241

in the Pareto optimal set. If the ranges are known, they give additional infor-
mation about the possibilities and limitations of the objective functions. Note
that each objective function can have a scale of its own. Examples are suggested
in Torn (1983) of how to display the scales of the objective functions.
Value paths are a recommendable method of illustration because they are
easy to interpret. For example, it is easy to distinguish non-Pareto optimal
alternatives if they are included. Further, even a large number of objective
functions or alternatives causes no problems. Value paths are used, for example,
in WWW-NIMBUS (see Subsection 5.12.8 of Part II) and the visual interactive
sensitivity analysis system VISA, see Belton and Vickers (1990).

Z! z2 z3
70

, , , ".............
,,
60
,
50
,,
, .,'.
40
,,
;,: '"
30
.'

20

to

alternative I - - - - alternative 2 alternative 3

Figure 3.3.1. Value paths.

In value path illustrations the roles of the lines and the bars can also be
interchanged. Then bars denote alternatives and lines represent objective func-
tions. In this case, possible different scales of the objective functions have to
be interpreted differently (see, e.g., Hwang and Masud (1979, p. 109». This
reversal of roles has been utilized, for instance, in the first implementations of
the reference direction approach (described in Section 5.10 of Part II), and its
counterpart for discrete problems, called VIMDA, see Korhonen (1986, 1991a).
The idea in VIMDA is that when the user horizontally moves the cursor to a
bar representing an alternative, the corresponding numerical objective values
are presented.
242 Part III - 3. Graphical Illustration

3.3.2. Bar Chart

Value paths are an effective means of presenting information without over-

loading the decision maker. Another general mode of illustration is to use bar
charts. This means that a group of bars represents the alternative values of a
single objective function, as in Figure 3.3.2. The bars of the same colour indi-
cate one alternative. Separate ranges for objective functions are possible as well.
Parallel ideas have been realized, for example, in DIDAS and WWW-NIMBUS,
treated in Subsections 5.6.4 and 5.12.8 of Part II, respectively.

z1 z2 z3
70

60

50

40

30

20

10

alternative I . . alternative 2 . . alternative 3

Figure 3.3.2. Bar chart.

Naturally the roles of the alternatives and the objective functions can be
interchanged in bar charts as well as in value paths. This, of course, means
that the order of the bars is altered. This is possible, for example, in WWW-
NIMBUS.
An alternative to using separate ranges for the objective functions is to
provide bar charts and value paths using both absolute and relative scales.
This is advisable in particular if the ranges of the objective functions vary
widely. This option is also available in WWW-NIMBUS.
 3.3. Illustrating a Set of Alternatives 243

3.3.3. Star Coordinate System

It is suggested in Manas (1982) that objective vectors can be represented

in a star coordinate system. For example, an alternative of three objective
functions is represented as an irregular triangle. This requires the ideal objective
vector and the (possibly approximated) nadir objective vector to be known.
An example is given in Figure 3.3.3. Each circle represents one alternative
objective vector. The ideal objective value is at the centre and the component
of the nadir objective vector is at the circumference. Each ray represents one
objective function. The area of the star depicts each alternative. See details in
Manas (1982).

70
~

Figure 3.3.3. Star coordinate system.

One can say that in the star coordinate system an alternative is better the
smaller the area of the star. If the order of the objective functions is altered,
the shape of the star changes. This can be considered a weakness of the system,
as stated in Tan and Fraser (1998). Tan and Fraser also suggest a modified star
graph to include the weight information of the decision maker (if available) in
the same display with the objective values.

3.3.4. Spider-Web Chart

Ideas similar to the star coordinate system are exploited in Kasanen et

al. (1991). An example is presented in Figure 3.3.4. This form of illustration can
be called a spider-web because of its shape. Sometimes it is also called a radar
chart. Each apex represents one objective function. The outer triangle shows the
(possibly approximated) nadir objective vector, the inner triangle (the darkest
one) stands for the ideal objective vector and the middle triangle (the grey
one) presents one alternative objective vector. Thus, only the middle triangle
is different in the different alternatives. These ideas are further developed in
Kasanen et al. (1991).
244 Part III - 3. Graphical Illustration

Figure 3.3.4. Spider-web chart.

3.3.5. Petal Diagram

Somewhat parallel ideas to the two previous representations are utilized in

Angehrn (1990a, b) when illustrating discrete alternatives in a program called
Triple C (Circular Criteria Comparison). A circle is divided into k (the number
of objective functions) equal sectors. The size (radius) of each slice indicates the
magnitude of the objective value. Here we have one circle for each alternative
objective vector. The same idea is suggested in Tan and Fraser (1998) and it
is called a petal diagram. Each segment of the diagram, that is, each objective
function can be associated with a different colour, as in Figure 3.3.5. Notice
that the order of the objective functions has no effect on the shape of the
diagram. The relations of the different segments are clearly shown. A way of
connecting weighting information in the petal figures is suggested in Tan and
Fraser (1998). In this case the segments are not of equal size but reflect the
weighting coefficients.

Figure 3.3.5. Petal diagram.

It is mainly a matter of taste in the star coordinate system, the spider-web

chart and the petal diagram how the ideal objective vector is situated. One
may think that when minimizing the objective functions it is logical to have
the ideal area as small as possible. However, the roles can be interchanged so
that the ideal objective value is located on the circumference and the nadir
 3.3. Illustrating a Set of Alternatives 245

objective value at the centre. In this case the larger the area the better. If this
is the setting, the ideal objective values can be replaced with, for example,
average objective values. This means that the figures can extend beyond the
circumference, stressing values better than the average.

3.3.6. Scatterplot Matrix

The scatterplot matrix described in Cleveland (1994) can be adapted for

visualizing different alternatives. The scatterplot matrix consists of panels each
representing one objective function pair. The dimension of the square matrix
is the number of objective functions. Different alternatives can be denoted
by different symbols or colours. As can be seen in Figure 3.3.6, each pair is
graphed twice with the scales interchanged. This means that either the lower or
the upper triangle could be dropper without losing any information. However,
displaying the whole matrix makes it easier to compare the objective function
values. One can measure the performance of one objective function against the
other objectives by having a look at one column or one row. Each objective
function can naturally have a range of its own in the panels, as in Figure 3.3.6.

30 40 50 60
II II 1111111111 IIIIII
65 - 0 f-
0
55 -
• •
f-
45 - f-

35 - X X ZI I- o alternative I
25 - I- X alternative 2
15 -

[J
I-
• alternative 3

--=
70
X X ~
I-
f-
: l-

• •
f-
45 -- I-
I-
- I-
25 - 0 0 I-
-
•
l- 60

- X X· I- 50
0 0
- Z3 I- 40

- l- 30
I III 1111111111 111111
25 45 70 15 25 35 45 55 65

Figure 3.3.6. Scatterplot matrix.

246 Part III - 3. Graphical Illustration

3.3.7. Other Illustrative Means

A graphical display system called GRADS is introduced in Klimberg (1992).

GRADS is dynamic and can be applied to problems with about five to twelve
objective functions. The decision maker is first asked to indicate two objective
functions whose values in the different alternatives are drawn as points in a
plane. In this space, the adjacent alternatives are connected with lines. In
other words, we have one value path. The decision maker obtains information
about the other objective values for one alternative at a time by indicating
that point with a mouse. Then, the other objective values are depicted as lines
originating from the point considered. The lengths of the lines are proportional
to the objective values. The end points of the lines are connected, thus forming
triangles of a different colour. The percentage achievements of the alternative
in question are also displayed. They are calculated as the difference between the
nadir objective value and the current objective value divided by the range. The
decision maker can change the alternative considered and the two objectives
whose value paths form the base of the display.
Different ideas of graphical illustration are also handled in Korhonen
(1991b). One of the ideas is Chernoff's faces, originally developed to illustrate
numerical information. The idea is to represent the values of up to 18 objective
functions as the characteristics of a face. In other words, the values of each
objective function are parametrized to represent some feature of an icon. The
icon used must be such that the user can see the icon becoming 'better' as the
value of the objective function improves. This is why concepts like symmetry
and harmony are important. An icon that people have been used to seeing
in a harmonious and symmetrical form is a house. Thus, Korhonen suggests
so-called harmonious houses to be used as icons. Objective functions are asso-
ciated with the corner points of the house, the door, the windows or the roof.
The aim is that when the values of the objective functions are close to the ideal
objective vector, the house is quite harmonious and symmetrical. This type of
illustration has especially been intended for pairwise comparison.
Literature describing the graphic presentation of data is summarized in Le-
wandowski and Granat (1991). It can be concluded that the research done does
not provide clear answers regarding what types of data presentation to favour
in the decision-making context. Lewandowski and Granat suggest a technique
called BIPLOT for the graphical presentation of matrices of rank 2. The set of
Pareto optimal objective vectors forms a matrix. This matrix is factorized into
a product of two matrices. The vectors in the two matrices are of order two
and can be plotted on a plane giving a representation of the original objective
vectors. Dynamic BIPLOT in aspiration level-based decision support systems
is also described. Another question is how much experience one must have to
be able to interpret representations like these.
Several tools for use in creating illustrations are also summarized in Klim-
berg (1992). One of them is to transform objective vectors into two-dimensional
 3.3. Illustrating a Set of Alternatives 247

curves with the aid of Fourier series. In this way all the vectors can be plotted
on the same coordinate system for comparison.
Other proposals for the graphical illustration of alternatives are given, for
example, in Vetschera (1992). They are based on indifference regions and linear
underlying value functions.
Let us finally mention a projection idea called GAIA (Geometrical Anal-
ysis for Interactive Aid). It is a part of the discrete multiattribute decision
analysis method PROMETHEE and it is described, for example, in Brans and
Mareschal (1990) and Mareschal and Brans (1988). The objective functions are
first modified to include some preference information of the decision maker and
then normalized. These objective functions have some benefits when compared
to the original ones. Namely, they are in the same scales, big differences in the
objective values are emphasized and small differences are lessened.
Principal component analysis is used in order to find a plane (two dimen-
sions) in which the new objective functions can be projected. The idea is to lose
as little information and variation as possible. In other words, the two largest
principal components are selected to form the projection plane. The weakness
here is that if the objective functions have nonlinear relations, principal com-
ponent analysis cannot find it.
If selecting the plane is managed well enough, the relations between the new
objective functions and the alternative solutions can be seen in their projec-
tions. Objective functions are depicted as vectors and alternatives as points on
the plane. For example, if two objective functions are highly conflicting, their
vectors go in opposite directions, whereas independent objective functions are
orthogonal and similar objective functions are oriented approximately in the
same direction. From the location of the alternatives one can see how well they
perform with respect to each objective function, that is, how near or far they
are from each other.
It seems that this GAIA plan ideology is a rather clear method of illustra-
tion. However, it has two main limitation. Firstly, the plane contains only a
part of the information available. Secondly, the conflict characteristics of the
objective functions are not absolute but depend on the alternatives considered.

3.3.8. General Remarks

The problem of how we can determine a priori whether the graphical formats
used will aid rather than hinder decision making is examined in Jarvenpaa
(1989) by comparative studies. The conclusion is that knowledge concerning
the relationship between the presentation format and the decision strategy can
facilitate the selection of the presentation format. Special attention is given to
the benefits of bar charts and grouped bar charts.
Similar matters are handled in connection with visual interactive simulation
in Bell and O'Keefe (1995). For example, it is concluded that the use of visual
displays generates solutions that are demonstrably better than those that make
248 Part III - 3. Graphical Illustration

limited use of such displays. This means that different levels of usage of spe-
cific displays have an impact on the quality of the solutions generated. In the
experiments, bar charts were the most favoured visual displays.
Several existing studies on the applicability of graphs versus tables are anal-
ysed in Vessey (1991). According to the theory developed, it is concluded that
tables perform better in information acquisition tasks in both time and accu-
racy of performance. Thus tables are in order when specific data values must
be extracted, since they represent discrete data values. If information must be
viewed at a glance, evaluated or relationships in the data are of interest, graphs
are recommendable. Thus, graphs and tables emphasize different characteristics
of the same data.
Using colours in illustrations has advantages and disadvantages. Above all,
the colours must be easy to discriminate. Another important issue is that some
colours may have specific connotations to the user. Such colours should be
avoided as far as possible.
An experimental evaluation of graphical and colour-enhanced information
presentation is given in Benbasat and Dexter (1985). Colours improve the read-
ability and understandability of both symbolic and graphical displays. Colours
make it easier for the decision maker to associate visually information belong-
ing to the same context or unit since such data are coded in the same colour.
Encouraging results with multi-colour reports are mentioned by Benbasat and
Dexter (1985), who also stress that tabular representation is the best when
a simple retrieval of data is important and a graphical representation is the
best when relationships among the data have to be shown. Graphs are visu-
ally appealing but sometimes tables are easier to read since they provide exact
values.
A recommended way of presenting information to the decision maker is to
offer the same data in different forms. In this way, the decision maker can choose
the most illustrative and informative representations. The illustrations may also
supplement each other. The decision maker can change her or his attention
from one figure to another and possibly skip undesirable alternatives before
making the final selection. A simple tabular format may be one of the figures.
Corresponding ideas are suggested, for instance, in Silverman et al. (1985) and
Steuer (1986, pp. 520-522).
An interesting alternative to graphics and numerical values is suggested
in Matos and Borges (1997). The idea is to illustrate alternatives in natural
language phrases. In this approach, fuzzy membership functions are formed
for every objective function defining fuzzy bounds, for example, 'very little',
'little', 'medium', 'very' and 'most' values. An example of alternatives in a
washing machine selection problem could be 'most cheap, medium power saver
and little fast.' The decision maker is asked for some descriptive information
as the basis of the membership functions before the solution process. This is a
promising way of illustrating data and it deserves further development.
 3.3. Illustrating a Set of Alternatives 249

Finally, one must concede that where a great number of alternatives co-
exists, the decision maker may get confused no matter how the alternatives
are illustrated. In this case, statistical tools, for example, principal component
analysis, may be useful.
 4. FUTURE DIRECTIONS

In this chapter, we outline some challenging topics for the future devel-
opment of multiobjective optimization, mainly from a mathematical point of
view. In addition, we give examples of promising ideas for research where the
first steps have been taken but further work is needed. All the issues mentioned
and many others merit further research and examination.
Multiobjective optimization is important, and improved solution methods
can bring about change in many areas and aspects of life. Even though mul-
tiobjective optimization methods have been applied to solving a variety of
problems in many areas of life, such as design problems in engineering, produc-
tion problems in economics, and environmental control problems in ecology,
there continue to exist many new problem types which could benefit highly
from multiobjective optimization. Particularly challenging in this respect are
real-life problems. There is clearly a need for more contributions reporting on
practical applications (making good use of more developed methods).
One interesting type of problems is so-called multidisciplinary re-engineer-
ing. It means that old engineering problems, for example, in optimal design,
whose solutions have been revised one feature at a time over the course of
years, are solved again from the very beginning, taking various aspirations and
aspects into consideration at the same time. Obviously this requires tools of
multiobjective optimization.
An important challenge for the developers of interactive methods and ap-
proaches is how to approach the decision maker. For example, a real experi-
ment with problem-related decision makers in Hobbs et al. (1992) shows that
the decision makers were sceptical of the value of multiobjective optimization
methods and they in some cases preferred unaided decision making. This means
that the methods should not only be user-friendly but also of real help to deci-
sion makers. Combining knowledge from the behavioural sciences with method
development could usefully serve in this direction.
The methodology of multiobjective optimization must be improved. This
means, for example, creating computationally efficient ways of generating trade-
off information for more general problem types under less restricting assump-
tions than those employed thus far. Another aspect is the structure of the
methods. On the one hand, providing the decision maker with the opportunity
for free search is important. On the other hand, guidance and support must

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
252 Part III - 4. Future Directions

be available, if desired. This necessitates developing mechanisms for dealing

with inconsistencies. In addition, ways of identifying appropriate methods for
different problems and different types of decision makers are certainly needed.
It is important not only to develop general methods but also to create
algorithms specifically tuned to certain problem types and areas of application.
An example is the monograph by Janssen (1992), where methods and decision
support tools for environmental management are dealt with.
An alternative to creating new methods is to use different methods in differ-
ent phases of the solution process. In this way, the positive features of various
methods can be exploited to their best advantage in appropriate phases of
the solution process. In addition, it may be possible to overcome some of the
weaknesses of the existing methods.
An example of the combination of several methods is a meta algorithm en-
deavouring at consolidating different methods of multiobjective optimization.
This is proposed in Steuer and Whisman (1986). The idea is that the same
meta program can be transformed into different methods by varying its control
parameters. The GDF, the Tchebycheff and the reference point methods with
the reference direction approach, STEM, the €-constraint method and two in-
teractive versions of the weighting method are available. This idea is further
developed in Steuer and Gardiner (1990). An important fact to consider, when
switching from one method to another in the middle of the solution process,
that is, how to maintain the convergence properties, needs further investiga-
tion. In Gardiner and Steuer (1994a, b), the meta algorithm is extended into a
unified algorithm containing thirteen different interactive methods. A vital ele-
ment of the algorithm is a matrix describing what kinds of switches are allowed
between the methods.
Similar ideas of combining several methods are proposed in CHmaco and
Antunes (1991). The system (only for MOLP problems) contains, for exam-
ple, the ZW method, STEM and VIG. Only problems with three objective
functions can be handled. The system has also been implemented. A further
developed implementation of the above-mentioned ideas is described in Antunes
et al. (1992a) and CHmaco and Antunes (1994). The method base package has
been named TOMMIX. Suitable means to support the decision maker in de-
ciding when and how to change from one method to another have still to be
explicated. TOMMIX is further extended into SOMMIX for more than three
(linear) objective functions in CHmaco et al. (1997).
Another approach to be elaborated is combining methods for continuous and
discrete problems. It may, for example, be that a set of solutions is generated
for the continuous problem and then ranked by means of discrete methods.
Examples of this are presented in Bard (1986), Kok and Lootsma (1985) and
Slowinski (1991). One example of these methods, the light beam search, was
described in Section 5.9 of Part II.
One can also combine methods of global optimization with multiobjective
optimization methods. In this way, one can aim at being able to handle globally
 4. Future Directions 253

Pareto optimal solutions, instead of locally Pareto optimal ones, in nonconvex

problems as well. Ideas of global multiobjective optimization on the basis of
clustering are proposed in Torn (1983).
Stochastic global optimization methods, like genetic algorithms, can also be
applied in multiobjective optimization. An example of this approach is given
in Osyczka and Kundu (1995). Another possibility for avoiding jamming into
locally Pareto optimal solutions is to use simulated annealing or tabu search
as an underlying solver.
In Arbel and Korhonen (1996a, 1997a), a new aspiration level-based method
is developed in the spirit of interior point methods (of linear programming) for
MOLP problems. The idea is to wander in the interior of the feasible objec-
tive region and only at the end to ascend to the Pareto optimal surface. Here,
the generally adopted idea that decision makers should handle only (weakly)
Pareto optimal solutions is called into question. One can justify such an ap-
proach by the fact that the decision maker can see some improvement in each
objective function instead of having to trade off all the time. The interior point
method used is an affine-scaling primal algorithm (also treated in Arbel (1993,
1994b, c)). The same idea is implemented by using an interior point method
called the primal-dual algorithm in Arbel and Korhonen (1996b, 1997b) (also
treated in Arbel (1994a, 1995)). Another modification of interior point meth-
ods for MOLP problems is described in Arbel and Oren (1994, 1996). In this
method, the gradient of an implicitly known value function is approximated
and a method of multiattribute decision analysis (namely AHP) is employed
in comparing alternatives. The gradients of an implicitly known value function
are also approximated and primal-dual linear methods used in Arbel (1997).
Results from other fields of research, for example, game theory, can also
be used in the solution processes. Among others, Rao studies the relationship
between Pareto optimal solutions and game theory in Rao (1987). He also
applies his results to structural optimization.
Another important area of development is software designed to implement
different methods and, especially, the user interface. As has been demonstrated,
few well-known software products exist for nonlinear multiobjective optimiza-
tion problems. As more and more advanced computers and graphical devices
are created, more tools become available in the quest for ease and even en-
joyment of use. This in turn involves new ideas for representing information,
such as illustrating alternatives in natural language phrases or using new kinds
of symbols. If the interface is able to adapt to the decision maker's style of
making decisions and is of help in analyzing the alternatives and results, and
can perhaps give suggestions or advice, then the interface may even overcome
some of the deficiencies of the method itself.
As far as large-scale problems are concerned, the possibilities of parallel
computing are worth examining in making the solution processes more efficient.
Multimedia possibilities in decision support systems are reviewed, for example,
in Grauer and Merten (1995).
254 Part III - 4. Future Directions

One potentiality not to be forgotten is the utilization of expert systems.

These can be applied to both suggesting a solution method according to the
properties of the problem and the preferences of the decision maker, and in the
solution process itself through supporting the decision maker. As an example,
interactive MOLP methods and expert system techniques are integrated in
Antunes et al. (1992b). The system described includes five methods, among
them, STEM and the ZW method. When the user of the system expresses her or
his hopes for further actions (such as a wish to get to know the neighbourhood
of the current solution), the system suggests one of the available interactive
methods. Computer graphics are also available. There are many features that
deserve further research and development, but this is certainly an interesting
path to follow.
A way of utilizing artificial neural networks in developing interactive multi-
objective optimization methods in proposed in Sun et al. (1996). The decision
maker is asked to articulate preference information over representative samples
of the Pareto optimal set and the neural network is trained to represent this
preference structure. The neural network is then used to generate improved
solutions. The preference information can be specified by a preference value for
each alternative or by pairwise comparisons between the alternatives. Possibili-
ties of artificial intelligence and neural networks in multiobjective optimization
are also charted in Gal and Hanne (1997).
The possibilities of several new technologies in computer science are re-
viewed in Antunes and Tsoukias (1997). The topics handled are fuzzy sets,
multimedia, distributed computing, expert systems, object-oriented program-
ming, neural networks, and the World-Wide Web. For example, the possibilities
of the World-Wide Web in implementing interactive methods and making them
easily available were dealt with in Subsection 5.12.8 of Part II. The example
given was WWW-NIMBUS.
One more thing to mention are spreadsheets. They are widely used and thus
provide a familiar environment for implementing inter activity in the methods.
This idea is realized in Steuer (1997) but it deserves further examination.
Flexibility in the mathematical modelling of the problem is often desirable.
Flexibility includes the possibility of interchanging the roles of objective and
constraint functions and updating the model if necessary. The decision maker
may, for example, wish to relax some constraints in order to be able to attain
certain aspirations. This means that integrating the modelling and the solution
processes deserves more attention.
It is not to be forgotten that dealing with incomplete information or un-
certainty is a part of solving real-life applications. This area is important even
though it has not been included in this book.
 5. EPILOGUE

We have presented a self-contained survey of the state of the art of nonlinear

multiobjective optimization together with a great number of further references.
After treating several important concepts and their relations, we have consid-
ered some theoretical results and connections.
We have demonstrated the methodology of multiobjective optimization by
describing several methods and by giving references in respect of a large number
of other methods. Methods have been classified into four groups according to
the contribution of the decision maker in the solution process. Because the
group of interactive methods has been developed most, it has received the
main emphasis. We have endeavoured to characterize the methods by some
comments on their positive and negative features.
Some of the features of the interactive methods dealt with have been col-
lected in a comparative table. Selected experiences and comparative observa-
tions of the methods have also been presented. In addition, some attempts
to aid in the selection of a solution method have been made. A decision tree
containing interactive methods has been suggested.
Some software packages have been mentioned. As far as software is con-
cerned, several possibilities of graphical illustrations of alternative solutions
have been introduced.

In general, one can say that the theory and the methods of multiobjective
optimization have been extensively developed during the past couple of decades.
Software implementations are considerably less in evidence. There is also a
lack of documentation in solving real-life multiobjective optimization problems
(using more developed methods). The reasons for this may be ignorance of
the full range of possibilities contained in existing methods as well as the lack
of suitable methods. For our part, we have filled a gap in the literature by
collecting several nonlinear multiobjective optimization methods between the
same covers.
In the development of methods the obvious conclusion is that it is important
to continue in the direction of user-friendliness. Methods must be even better
able to correspond to the characteristics of the decision maker. Ifthe aspirations
of the decision maker change during the solution process, the algorithm must
be able to cope with this situation. Computational tests have confirmed the
idea that decision makers want to feel in control of the solution process, and

K. Miettinen, Nonlinear Multiobjective Optimization

© Springer Science+Business Media New York 1998
256 Part III - 5. Epilogue

consequently they must understand what is happening. However, sometimes

the decision maker simply needs support, and this should be available as well.
Thus, the aim is to have methods that support learning so that guidance is
given whenever necessary. The decision maker must be the basis in developing
new interactive methods. Specific methods for different areas of application
that take into account the characteristics of the problems are also important.
 REFERENCES

Agrell P.J., On Redundancy in Multi Criteria Decision Making, European Journal of Oper-
ational Research 98, No.3 (1997), 571-586.
Agrell P.J., Lence B.J., Starn A., An Interactive Multicriteria Decision Model for Mul-
tipurpose Reservoir Management: the Shellmouth Reservoir, Journal of Multi-Criteria
Decision Analysis 1, No.2 (1998), 61-86.
Aksoy Y., Interactive Multiple Objective Decision Making: A Bibliography (1965-1988),
Management Research News 13, No.2 (1990), 1-8.
Aksoy Y., Butler T.W., Minor E.D. Ill, Comparative Studies in Interactive Multiple Objective
Mathematical Programming, European Journal of Operational Research 89, No.2 (1996),
408-422.
Al-alvani J.E., Hobbs B.F., Malakooti B., An Interactive Integrated Multiobjective Opti-
mization Approach for Quasiconcave/Quasiconvex Utility Functions, MUltiple Criteria
Decision Making: Proceedings of the Ninth International Conference: Theory and Appli-
cations in Business, Industry, and Government, Edited by A. Goicoechea, L. Duckstein,
S. Zionts, Springer-Verlag, New York, 1992, pp. 45-60.
Alekseichik M.l., Naumov G.E., Existence of Efficient Points in Infinite-Dimensional Vector
Optimization Problem, Automation and Remote Control 42, No.5, Part 2 (1981), 666-
670.
Angehrn A.A., Supporting Multicriteria Decision Making: New Perspectives (3 New Systems,
Working Paper, INSEAD, Fontainebleau (1990a).
Angehrn A.A., "Triple C" A Visual Interactive MCDSS, Working Paper, INSEAD, Fontain-
ebleau (1990b).
Antunes C.H., Tsoukias A., Against Fashion: A Travel Survival Kit in "Modern" MCDA,
Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997,
pp. 378-389.
Antunes C.H., Alves M.J., Silva A.L., Climaco J.N., An Integrated MOLP Method Base
Package - A Guided Tour of TOMMIX, Computers & Operations Research 19, No.7
(1992a), 609-625.
Antunes C.H., Melo M.P., Climaco J.N., On the Integration of an Interactive MOLP Pro-
cedure Base and Expert System Technique, European Journal of Operational Research
61, No. 1-2 (1992b), 135-144.
Arbel A., Generating Interior Search Directions for Muitiobjective Linear Programming Us-
ing Approximate Gradients and Efficient Anchoring Points, Optimization 28, No.2
(1993), 149-164.
Arbel A., An Interior Multiobjective Primal-Dual Linear Programming Algorithm Using
Approximated Gradients and Sequential Generation of Anchor Points, Optimization 30,
No.2 (1994a), 137-150.
Arbel A., Anchoring Points and Cones of Opportunities in Interior Mu/tiobjective Linear
Programming, Journal of the Operational Research Society 45, No.1 (1994b), 83-96.
Arbel A., Using Efficient Anchoring Points for Generating Search Directions in Interior
Multiobjective Linear Programming, Journal of the Operational Research Society 45,
No.3 (l994c), 330-344.
258 References

Arbel A., An Interior Multiple Objective Primal-Dual Linear Programming Algorithm Us-
ing Efficient Anchoring Points, Journal of the Operational Research Society 46, No.9
(1995), 1121-1132.
Arbel A., An Interior Multiobjective Primal-Dual Linear Programming Algorithm Based
on Approximated Gradients and Efficient Anchoring Points, Computers & Operations
Research 24, No.4 (1997), 353-365.
Arbel A., Korhonen P., Using Aspiration Levels in an Interactive Interior Multiobjective
Linear Programming Algorithm, European Journal of Operational Research 89, No.1
(1996a), 193- 20l.
Arbel A., Korhonen P., Using Aspiration Levels in an Interior Primal-Dual Multiobjective
Linear Programming Algorithm, Journal of Multi-Criteria Decision Analysis 5, No. 1
(1996b), 61-71.
Arbel A., Korhonen P., Generating Interior Search Directions in Multiple Objective Lin-
ear Programming Problems Using Aspiration Levels, Multicriteria Analysis, Edited by
J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997a, pp. 146-155.
Arbel A., Korhonen P., An Interior Multiobjective Linear Programming Algorithm Using
Aspirations, Multiple Criteria Decision Making: Proceedings of the Twelfth International
Conference, Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics
and Mathematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997b, pp. 245-254.
Arbel A., Oren S.S., A Modification of Karmarkar's Algorithm to Multiple Objective Linear
Programming Problems, Multiple Criteria Decision Making - Proceedings of the Tenth
International Conference: Expand and Enrich the Domains of Thinking and Application,
Edited by G.H. Tzeng, H.F. Wand, U .P. Wen, P.L. Yu, Springer-Verlag, New York, 1994,
pp.37-46.
Arbel A., Oren S.S., Using Approximate Gradients in Developing an Interactive Interior
Primal-Dual Mu/tiobjective Linear Programming Algorithm, European Journal of Oper-
ational Research 89, No.1 (1996), 202-21l.
Armann R., Solving Mu/tiobjective Programming Problems by Discrete Representation, Op-
timization 20, No.4 (1989), 483-492.
Athan T.W., Papalambros P.Y., Multicriteria Optimization of ABS Control Algorithms Us-
ing a Quasi-Monte Carlo Method, Multiple Criteria Decision Making: Proceedings of
the Twelfth International Conference, Hagen (Germany), Edited by G. Fandel, T. Gal,
Lecture Notes in Economics and Mathematical Systems 448, Springer-Verlag, Berlin,
Heidelberg, 1997, pp. 457-466.
Aubin J.-P., Niislund B., An Exterior Branching Algorithm, Working Paper 72-42, European
Institute for Advanced Studies in Management, Brussels (1972).
Baba N., Takeda H., Miyake T., Interactive Multi-Objective Programming Technique Us-
ing Random Optimization Method, International Journal of Systems Science 19, No.1
(1988), 151-159.
Balachandran M., Gero J.S., The Noninferior Set Estimation (NISE) Method for Three
Objective Problem, Engineering Optimization 9, No.2 (1985), 77-88.
Balbas A., Guerra P ..l., Sensitivity Analysis for Convex Multiobjective Programming in Ab-
stract Spaces, Journal of Mathematical Analysis and Applications 202, No.2 (1996),
645-658.
Ballestero E., Utility Functions: A Compromise Programming Approach to Specification and
Optimization, Journal of Multi-Criteria Decision Analysis 6, No.1 (1997a), 11-16.
Ballestero E., Selecting the CP Metric: A Risk Aversion Approach, European Journal of
Operational Research 97, No.3 (1997b), 593-596.
Ballestero E., Romero C., A Theorem Connecting Utility Function Optimization and Com-
promise Programming, Operations Research Letters 10, No.7 (1991), 421-427.
Bard J.F., A Multiobjective Methodology for Selecting Subsystem Automation Options, Man-
agement Science 32, No. 12 (1986), 1628-1641.
 References 259

Bardossy A., Bogardi 1., Duckstein L., Composite Programming as an Extension oj Compro-
mise Programming, Mathematics of Multi Objective Optimization, Edited by P. Serafini,
Springer-Verlag, Wien, New York, 1985, pp. 375-408.
Batishchev D.l., Anuchin V.F., Shaposhnikov D.E., The Use oj the Qualitative InJormation
on the Importance oj Particular Criteria for the Computation of Weighting Coefficients,
Multiobjective Problems of Mathematical Programming, Edited by A. Lewandowski, V.
Volkovich, Lecture Notes in Economics and Mathematical Systems 351, Springer-Verlag,
1991, pp. 2-7.
Bazaraa M.S., Sherali H.D., Shetty C.M., Nonlinear Programming: Theory and Algorithms,
Second Edition, John Wiley & Sons, Inc., New York, 1993.
Belkeziz K., Pi riot M., Proper Efficiency in Nonconvex Vector-Maximization-Problems, Eu-
ropean Journal of Operational Research 54, No.1 (1991), 74-80.
Bell D.E., Double-Exponential Utility Functions, Mathematics of Operations Research 11,
No.2 (1986), 351-361.
Bell D.E., Risk, Return, and Utility, Management Science 41, No.1 (1995), 23-30.
Bell P.C., O'Keefe R.M., An Experimental Investigation into the Efficacy of Visual Interac-
tive Simulation, Management Science 41, No.6 (1995), 1018-1038.
Belton V., Vickers S., Use of a Simple Multi-Attribute Value Function Incorporating Vi-
sual Interactive Sensitivity Analysis for Multiple Criteria Decision Making, Readings in
Multiple Criteria Decision Aid, Edited by C.A. Bana e Costa, Springer-Verlag, Berlin,
Heidelberg, 1990, pp. 319-334.
Benayoun R., de Montgolfier J., Tergny J., Laritchev 0., Linear Programming with Multiple
Objective Functions: Step Method (STEM), Mathematical Programming 1, No.3 (1971),
366-375.
Benbasat 1., Dexter A.S., An Experimental Evaluation of Graphical and Color-Enhanced
Information Presentation, Management Science 31, No. 11 (1985), 1348-1364.
Benito-Alonso M.A., Devaux P., Location and Size of Day Nurseries - a Multiple Goal
Approach, European Journal of Operational Research 6, No.2 (1981), 195-198.
Benoist J., Connectedness of the Efficient Set for Strictly Quasiconcave Sets, Journal of
Optimization Theory and Applications 96, No.3 (1998), 627-654.
Benson H.P., Existence of Efficient Solutions for Vector Maximization Problems, Journal of
Optimization Theory and Applications 26, No.4 (1978), 569-580.
Benson H.P., An Improved Definition of Proper Efficiency for Vector Maximization with
Respect to Cones, Journal of Mathematical Analysis and Applications 71, No.1 (1979a),
232-241.
Benson H.P., Vector Maximization with Two Objective Functions, Journal of Optimization
Theory and Applications 28, No.2 (1979b), 253-257.
Benson H.P., On a Domination Property for Vector Maximization with Respect to Cones,
Journal of Optimization Theory and Applications 39, No.1 (1983), 125-132; Errata
Corrige, Journal of Optimization Theory and Applications 43, No.3 (1984), 477-479.
Benson H.P., Complete Efficiency and the Initialization of Algorithms for Multiple Objective
Programming, Operations Research Letters 10, No.8 (1991), 481-487.
Benson H.P., Aksoy Y., Using Efficient Feasible Directions in Interactive Multiple Objective
Linear Programming, Operations Research Letters 10, No.4 (1991), 203-209.
Benson H.P., Morin T.L., The Vector Maximization Problem: Proper Efficiency and Stability,
SIAM Journal on Applied Mathematics 32, No.1 (1977), 64-72.
Benson H.P., Sayin S., Optimization over the Efficient Set: Four Special Cases, Journal of
Optimization Theory and Applications 80, No.1 (1994), 3-18.
Benson H.P., Sayin S., Towards Finding Global Representations oj the Efficient Set in Mul-
tiple Objective Mathematical Programming, Naval Research Logistics 44, No.1 (1997),
47-67.
260 References

Ben-Tal A., Characterization of Pareto and Lexicographic Optimal Solutions, Multiple Cri-
teria Decision Making Theory and Application, Edited by G. Fandel, T. Gal, Lecture
Notes in Economics and Mathematical Systems 177, Springer-Verlag, Berlin, Heidelberg,
1980, pp. 1-11.
Berbel J., Zamora R., Application of MOP and GP to Wildlife Management (Deer). A
Case in the Mediterranean Ecosystem, Multi-Objective Programming and Goal Pro-
gramming: Theories and Applications, Edited by M. Tamiz, Lecture Notes in Economics
and Mathematical Systems 432, Springer-Verlag, Berlin, Heidelberg, 1996, pp. 300-308.
Bestle D., Eberhard P., Dynamic System Design via Multicriteria Optimization, Multiple
Criteria Decision Making: Proceedings of the Twelfth International Conference, Hagen
(Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical
Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 467-478.
Bhatia D., Aggarwal S., Optimality and Duality for Multiobjective Nonsmooth Programming,
European Journal of Operational Research 57, No.3 (1992), 360-367.
Bhatia D., Datta N., Necessary Conditions and Subgradient Duality for Non-Differentiable
and Non-Convex Multi-Objective Programming Problem, Cahiers du Centre d'Etudes de
Recherche Operationelle 27, No. 1-2 (1985), 131-139.
Bischoff E., Two Empirical Tests with Approaches to Multiple-Criteria Decision Making, Plu-
ral Rationality and Interactive Decision Processes, Edited by M. Grauer, M. Thompson,
A.P. Wierzbicki, Lecture Notes in Economics and Mathematical Systems 248, Springer-
Verlag, Berlin, Heidelberg, 1985, pp. 344-347.
Bischoff E., Multi-Objective Decision Analysis - The Right Objectives?, Large-Scale Mod-
elling and Interactive Decision Analysis, Edited by G. Fandel, M. Grauer, A. Kurzhanski,
A.P. Wierzbicki, Lecture Notes in Economics and Mathematical Systems 273, Springer-
Verlag, 1986, pp. 155-160.
Bitran G.R., Duality for Nonlinear Multiple-Criteria Optimization Problems, Journal of
Optimization Theory and Applications 35, No.3 (1981), 367-401.
Bogetoft P., HalJefjord A., Kok M., On the Convergence of Reference Point Methods in Mul-
tiobjective Programming, European Journal of Operational Research 34, No.1 (1988),
56-68.
Borwein J.M., Proper Efficient Points for Maximizations with Respect to Cones, SIAM Jour-
nal on Control and Application 15, No.1 (1977), 57-63.
Borwein J.M., On the Existence of Pareto Efficient Points, Mathematics of Operations Re-
search 8, No.1 (1983), 64-73.
Borwein J.M., Zhuang n.M., Super Efficiency in Convex Vector Optimi:llation, ZOR - Meth-
ods and Models of Operations Research 35, No.3 (1991), 175-184.
Borwein J .M., Zhuang D.M., Super Efficiency in Vector Optimization, Transactions of the
American Mathematical Society 338, No.1 (1993), 105-122.
Bowman V.J. Jr., On the Relationship of the Tchebycheff Norm and the Efficient Frontier of
Multiple-Criteria Objectives, Multiple Criteria Decision Making, Edited by H. Thiriez,
S. Zionts, Lecture Notes in Economics and Mathematical Systems 130, Springer-Verlag,
Berlin, Heidelberg, 1976, pp. 76-85.
Brans J.-P., The Space of Freedom of the Decision Maker - Modelling the Human Brain,
European Journal of Operational Research 92, No.3 (1996), 593-602.
Brans J.P., Mareschal B., The PROMETHEE Methods for MCDM, the PROMCALC, GAIA
and BANKADVISER Software, Readings in Multiple Criteria Decision Aid, Edited by
C.A. Bana e Costa, Springer-Verlag, Berlin, Heidelberg, 1990, pp. 216-252.
Brauer D.C., Naadimuthu G., A Goal Programming Model for Aggregate Inventory and
Distribution Planning, Mathematical and Computer Modelling 16, No.3 (1992), 81-90.
Brockhoff K., Experimental Test of MCDM Algorithms in a Modular Approach, European
Journal of Operational Research 22, No.2 (1985), 159-166.
Brosowski B., da Silva A.R., Simple Tests for Multi-Criteria Optimality, OR Spektrum 16,
No.4 (1994), 243-247.
 References 261

Buchanan J.T., Multiple Objective Mathematical Progmmming: A Review, New Zealand

Operational Research 14, No.1 (1986), 1-27.
Buchanan J .T., An Experimental Evaluation of Intemctive MCDM Methods and the Decision
Making Process, Journal of the Operational Research Society 45, No.9 (1994), 1050-
1059.
Buchanan J.T., A Nai've Approach for Solving MCDM Problems: The GUESS Method, Jour-
nal of the Operational Research Society 48, No.2 (1997), 202-206.
Buchanan J .T., Corner J .L., The Effects of Anchoring in Intemctive MCDM Solution Meth-
ods, Computers & Operations Research 24, No. 10 (1997), 907-918.
Buchanan J.T., Daellenbach H.G., A Compamtive Evaluation of Intemctive Solution Meth-
ods for Multiple Objective Decision Models, European Journal of Operational Research
29, No.3 (1987), 353-359.
Buede D., Second Overview of the MCDA Software Market, Journal of Multi-Criteria Deci-
sion Analysis 5, No.4 (1996), 312-316.
Bushenkov V., Kaitala V., Lotov A., Pohjola M., Decision and Negotiation Support for
Transboundary Air Pollution Control between Finland, Russia and Estonia, Finnish
Economic Papers 7, No.1 (1994).
Bushenkov V.A., Chernykh O.L., Kamenev G.K., Lotov A.V., Multidimensional Images
Given by Mappings: Construction and Visualization, Pattern Recognition and Image
Analysis 5, No.1 (1995), 35-56.
Caballero R., Rey L., Ruiz F., Determination of Satisfying and Efficient Solutions in Convex
Multi-Objective Progmmming, Optimization 37, No.2 (1996), 125-137.
Caballero R., Rey L., Ruiz F., Gonzalez M., An Algorithmic Package for the Resolution
and Analysis of Convex Multiple Objective Problems, Multiple Criteria Decision Mak-
ing: Proceedings of the Twelfth International Conference, Hagen (Germany), Edited by
G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical Systems 448, Springer-
Verlag, Berlin, Heidelberg, 1997, pp. 275-284.
Cambini A., Martein L., On the Existence of Efficient Points, Optimization 28, No. 3-4
(1994), 283-290.
Carmichael D.G., Structuml Modelling and Optimization, Ellis Horwood Ltd., 1981.
Censor Y., Pareto Optimality in Multiobjective Problems, Applied Mathematics and Opti-
mization 4, No.1 (1977), 41-59.
Chankong V., Haimes Y.Y., The lntemctive Surrogate Worth Trade-Off (ISWT) Method for
Multiobjective Decision-Making, Multiple Criteria Problem Solving, Edited by S. Zionts,
Lecture Notes in Economics and Mathematical Systems 155, Springer-Verlag, Berlin,
Heidelberg, 1978, pp. 42-67.
Chankong V., Haimes Y.Y., On the Chamcterization of Noninferior Solutions of the Vector
Optimization Problem, Automatica 18, No.6 (1982), 697-707.
Chankong V., Haimes Y.Y., Optimization-Based Methods for Multiobjective Decision-Mak-
ing: An Overview, Large Scale Systems 5, No.1 (1983a), 1-33.
Chankong V., Haimes Y.Y., Multiobjective Decision Making Theory and Methodology, Else-
vier Science Publishing Co., Inc., New York, 1983b.
Chankong V., Haimes Y.Y., Thadathil J., Zionts S., Multiple Criteria Optimization: A State
of the Art Review, Decision Making with Multiple Objectives, Lecture Notes in Eco-
nomics and Mathematical Systems 242, Edited by Y.Y. Haimes, V. Chankong, Springer-
Verlag, 1985, pp. 36-90.
Charnes A., Cooper W.W., Management Models and Industrial Applications of Linear Pro-
gmmming, vol. 1, John Wiley & Sons, Inc., New York, 1961.
Charnes A., Cooper W.W., Goal Progmmming and Multiple Objectitle Optimization; Part
1, European Journal of Operational Research 1, No.1 (1977),39-54.
Charnes A., Cooper W.W., Ferguson R.O., Optimal Estimation of Executive Compensation
by Linear Progmmming, Management Science 1, No.2 (1955), 138-15l.
Chen G.-Y., Necessary Conditions of Nondominated Solutions in Multicriteria Decision
Making, Journal of Mathematical Analysis and Applications 104, No.1 (1984), 38-46.
262 References

Chen T., Wang B.-I., An Interactive Method for Multiobjective Decision Making, Large Scale
Systems: Theory and Applications 1983, Edited by A. Straszak, Pergamon Press, 1984,
pp. 277-282.
Chew K.L., Choo E.U., Pseudolinearity and Efficiency, Mathematical Programming 28, No.
2 (1984), 226-239.
Choo E.U., Atkins D.R., Proper Efficiency in Nonconvex Multicriteria Programming, Math-
ematics of Operations Research 8, No.3 (1983), 467-470.
Chou J.H., Hsia W.-S., Lee T.-Y., On MUltiple Objective Programming Problems with Set
Functions, Journal of Mathematical Analysis and Applications 105, No.2 (1985), 383-
394.
Clarke F.H., Optimization and Nonsmooth Analysis, John Wiley & Sons, Inc., 1983.
Cleveland W.S., The Elements of Graphing Data, Revised Edition, AT&T Bell Laboratories,
Murray Hill, New Jersey, 1994.
Climaco J.N., Antunes C.H., Flexible Method Bases and Man-Machine Interfaces as Key
Features in Interactive MOLP Approaches, Multiple Criteria Decision Support, Edited
by P. Korhonen, A. Lewandowski, J. Wallenius, Lecture Notes in Economics and Math-
ematical Systems 356, Springer-Verlag, 1991, pp. 207-216.
Climaco J.N., Antunes C.H., Man-Machine Interfacing in MCDA, Multiple Criteria Decision
Making - Proceedings of the Tenth International Conference: Expand and Enrich the
Domains of Thinking and Application, Edited by G.H. Tzeng, H.F. Wand, V.P. Wen,
P.L. Yu, Springer-Verlag, New York, 1994, pp. 239-253.
Climaco J.N., Antunes C.H., Alves M.J., From TRIMAP to SOMMIX - Building Effective
Interactive MOLP Computational Tools, Multiple Criteria Decision Making: Proceedings
of the Twelfth International Conference, Hagen (Germany), Edited by G. Fandel, T. Gal,
Lecture Notes in Economics and Mathematical Systems 448, Springer-Verlag, Berlin,
Heidelberg, 1997, pp. 285-296.
Clinton R.J., 'Troutt M.D., Toward Optimal Scaling in the Method of Abstract Forces for
Interactive Multiple Criteria Optimization, Mathematical and Computer Modelling 11
(1988), 589-594.
Cohon J.L., Multiobjective Programming and Planning, Academic Press, Inc., New York,
1978. •
Cohon J.L., Multicriteria Programming: Brief Review and Application, Design Optimization,
Edited by J.S. Gero, Academic Press, Inc., 1985, pp. 163-191.
Colson G., de Bruyn C., Multiple Criteria Supports on Micro-Computers, Belgian Journal
of Operations Research, Statistics and Computer Science 27, No.1 (1987), 29-58.
Corley H.W., A New Scalar Equivalence for Pareto Optimization, IEEE 'Transactions on
Automatic Control 25, No.4 (1980), 829-830.
Corner J.L., Buchanan J.T., Experimental Consideration of Preference in Decision Making
under Certainty, Journal of Multi-Criteria Decision Analysis 4, No.2 (1995), 107-121.
Corner J.L., Buchanan J.T., Capturing Decision Maker Preference: Experimental Compar-
ison of Decision Analysis and MCDM Techniques, European Journal of Operational
Research 98, No.1 (1997), 85-97.
Corner J.L., Corner P.D., Characteristics of Decisions in Decision Analysis Practice, Journal
of the Operational Research Society 46, No.3 (1995), 304-314.
Costa J.P., Clfmaco J.N., A Multiple Reference Point Parallel Approach in MCDM, Multiple
Criteria Decision Making - Proceedings of the Tenth International Conference: Expand
and Enrich the Domains of Thinking and Application, Edited by G.H. Tzeng, H.F. Wand,
V.P. Wen, P.L. Yu, Springer-Verlag, New York, 1994, pp. 255-263.
Crama Y., Analysis of STEM-Like Solutions to Multi-Objective Programming Problems,
Multi-Objective Decision Making, Edited by S. French, R. Hartley, L.C. Thomas, D.J.
White, Academic Press, 1983, pp. 208-213.
Craven B.D., On Sensitivity Analysis for Multicriteria Optimization, Optimization 19, No.
4 (1988), 513-523.
 References 263

Craven B.D., Nonsmooth Multiobjective Programming, Numerical Functional Analysis and

Optimization 10, No. 1-2 (1989), 49-64.
Current J.R., Storbeck J.E., A Multiobjective Approach to Design Franchise Outlet Networks,
Journal of the Operational Research Society 45, No.1 (1994), 71·-81.
Da Cunha N.O., Polak E., Constrained Minimization under Vector- Valued Criteria in Fi-
nite Dimensional Spaces, Journal of Mathematical Analysis and Applications 19, No.1
(1967), 103-124.
Das I., Dennis J.E., A Closer Look at Drawbacks of Minimizing Weighted Sums of Ob-
jectives for Pareto Set Generation in Multicriteria Optimization Problems, Structural
Optimization 14, No.1 (1997), 63-69.
Das I., Dennis J.E., Normal-Boundary Intersection: A New Method for Generating the
Pareto Surface in Nonlinear Multicriteria Optimization Problems, SIAM Journal on
Optimization 8, No.3 (1998), 631-657.
Das L.N., Nanda S., Symmetric Dual Multiobjective Programming, European Journal of
Operational Research 97, No.1 (1997), 167-17l.
Dauer J.P., Fosnaugh T.A., Optimization over the Efficient Set, Journal of Global Optimiza-
tion 7, No.3 (1995), 261-277.
Dauer J., Liu Y.-H., Multi-Criteria and Goal Programming, Advances in Sensitivity Analysis
and Parametric Programming, Edited by T. Gal. H.J. Greenberg, Kluwer Academic
Publishers, Boston, 1997, pp. 11-1-11-3l.
Dauer J.P., Osman M.S.A., Decomposition of the Parametric Space in Multiobjective Convex
Programs Using the Generalized TchebychefJ Norm, Journal of Mathematical Analysis
and Applications 107, No.1 (1985), 156-166.
Dauer J.P., Stadler W., A Survey of Vector Optimization in Infinite-Dimensional Spaces,
Part 2, Journal of Optimization Theory and Applications 51, No.2 (1986), 205-24l.
de Kluyver C.A., An Exploration of Various Goal Programming Formulations - with Appli-
cation to Advertising Media Selection, Journal of the Operational Research Society 30,
No.2 (1979),167-171.
de Neufville R., McCord M., Unreliable Measurement of Utility: Significant Problems for
Decision Analysis, Operational Research '84, Edited by J.P. Brans, Elsevier Science
Publishers B.V. (North-Holland), Amsterdam, 1984, pp. 464-476.
Deng S., Characterizations of the Nonemptiness and Compactness of Solution Sets in Con-
vex Vector Optimization, Journal of Optimization Theory and Applications 96, No.1
(1998a), 123-131.
Deng S., On Efficient Solutions in Vector Optimization, Journal of Optimization Theory and
Applications 96, No.1 (1998b), 201-209.
Despontin M., Moscarola J., Spronk J., A User-Oriented Listing of Multiple Criteria Deci-
sion Methods, Revue BeIge de Statistique, d'Informatique et de Recherche Operationnelle
23, No.4 (1983),4-110.
Dessouky M.l., Ghiassi M., Davis W.J., Estimates of the Minimum Nondominated Crite-
rion Values in Multiple-Criteria Decision-Making, Engineering Costs and Production
Economics 10 (1986), 95-104.
Diaz A., Interactive Solution to Multiobjective Optimization Problems, International Journal
for Numerical Methods in Engineering 24, No. 10 (1987), 1865-1877.
Dinkelbach W., Multicriteria Decision Models with Specified Goal Le1leis, Multiple Criteria
Decision Making Theory and Application, Edited by G. Fandel, T. Gal, Lecture Notes
in Economics and Mathematical Systems 177, Springer-Verlag, Berlin, Heidelberg, 1980,
pp.52-59.
Dolezal J., Necessary Conditions for Pareto Optimality in NondifJerentiable Problems, Prob-
lems of Control and Information Theory 14, No.2 (1985), 131-140.
Durier R., Weighting Factor Results in Vector Optimization, Journal of Optimization Theory
and Applications 58, No.3 (1988), 411-430.
Dyer J.S., Interactive Goal Programming, Management Science 19, No.1 (1972), 62-70.
264 References

Dyer J.S., A Time-Sharing Computer Program for the Solution of the Multiple Criteria
Problem, Management Science 19, No. 12 (1973a), 1379-1383.
Dyer J.8., An Empirical Investigation of a Man-Machine Interactive Approach to the So-
lution of the Multiple Criteria Problem, Multiple Criteria Decision Making, Edited by
J.L. Cochrane, M. Zeleny, University of South Carolina Press, Columbia, South Carolina,
1973b, pp. 202-216.
Dyer J.S., The Effects of Errors in the Estimation of the Gradient on the Frank-Wolfe
Algorithm, with Implications for Interactive Programming, Operations Research 22, No.
1 (1974), 160-174.
Dyer J.S., Sarin R.K., Multicriteria Decision Making, Mathematical Programming for Op-
erations Researchers and Computer Scientists, Edited by A.G. Holzman, Marcel Dekker,
Inc., 1981, pp. 123-148.
Dyer J.S., Fishburn P.C., Steuer R.E., Wallenius J., Zionts S., Multiple Criteria Decision
Making, Multiattribute Utility Theory: The Next Ten Years, Management Science 38,
No.5 (1992), 645-654.
Ecker J .G., Kouada I.A., Finding Efficient Points for Linear Multiple Objective Programs,
Mathematical Programming 8, No.3 (1975), 375-377.
Edgeworth F.Y., Mathematical Psychics, University Microfilms International (Out-of-Print
Books on Demand), 1987 (the original edition in 1881).
Egudo R.R., A Duality Theorem for a Fractional Multiobjective Problem with Square Root
Terms, Recent Developments in Mathematical Programming, Edited by S. Kumar, Gor-
don and Breach Science Publishers, 1991, pp. 101-113.
Ehrgott M., Klamroth K., Connectedness of Efficient Solutions in Multiple Criteria Com-
binatorial Optimization, European Journal of Operational Research 97, No.1 (1997),
159-166.
Eiduks J., Control of Complicated Systems in Case of Vectorial Indicator of Quality, Fourth
Formator Symposium on Mathematical Methods for the Analysis of Large-Scale Systems,
Edited by J. Bebes, L. Bakule, Academia, Prague, 1983, pp. 165-180.
Eiselt H.A., Pederzoli G., Sandblom C.-L., Continuous Optimization Models, Operations
Research: Theory, Techniques, Applications, Walter de Gruyter & Co., Berlin, New York,
1987.
El Abdouni B., Thibault L., Lagrange Multipliers for Pareto Nonsmooth Programming Prob-
lems in Banach Spaces, Optimization 26, No. 3-4 (1992), 277-285.
Eom H.B., The Current State of Multiple Criteria Decision Support Systems, Human Sys-
tems Management 8 (1989), 113-119.
Eschenauer H.A., Multicriteria Optimization Procedures in Application on Structural Me-
chanics Systems, Recent Advances and Historical Developments of Vector Optimization,
Edited by J. Jahn, W. Krabs, Lecture Notes in Economics and Mathematical Systems
294, Springer-Verlag, Berlin, Heidelberg, 1987, pp. 345-376.
Eschenauer H.A., Schafer E., Bernau H., Application of Interactive Vector Optimization
Methods with Regard to Problems in Structural Mechanics, Discretization Methods and
Structural Optimization - Procedures and Applications, Edited by H.A. Eschenauer, G.
Thierauf, Lecture Notes in Engineering 42, Springer-Verlag, Berlin, Heidelberg, 1989,
pp. 95-101.
Eschenauer H., Koski J., Osyczka A. (Eds.), Multicriteria Design Optimization Procedures
and Applications, Springer-Verlag, Berlin, Heidelberg, 1990a.
Eschenauer H.A., Osyczka A., Schafer E., Interactive Multicriteria Optimization in Design
Process, Multicriteria Design Optimization Procedures and Applications, Edited by H.
Eschenauer, J. Koski, A. Osyczka, Springer-Verlag, Berlin, Heidelberg, 1990b, pp. 71-
114.
Ester J., About the Stability of the Solution of Multiple Objective Decision Making Problems
with Generalized Dominating Sets, Systems Analysis Modelling Simulation 1, No.4
(1984), 287-291.
 References 265

Ester J., Holzmiiller R., About Some Methods and Applications in Multicriteria Decision
Making, Systems Analysis Modelling Simulation 3, No.5 (1986), 425-438.
Ester J., Tr5ltzsch F., On Generalized Notions of Efficiency in Multicriteria Decision Mak-
ing, Systems Analysis Modelling Simulation 3, No.2 (1986), 153-161.
Evans G.W., An Overview of Techniques for Solving Multiobjective Mathematical Programs,
Management Science 30, No. 11 (1984), 1268-1282.
Federer H., Geometric Measure Theory, Springer-Verlag, Berlin, Heidelberg, 1969.
Ferreira P.A.V., Geromel J.C., An Interactive Projection Method for Multicriteria Optimiza-
tion Problems, IEEE Transactions on Systems, Man, and Cybernetics 20, No.3 (1990),
596-605.
Ferreira P.A.V., Machado M.E.S., Solving Multiple-Objective Problems in the Objective
Space, Journal of Optimization Theory and Applications 89, No.3 (1996), 659-680.
Fishburn P.C., Lexicographic Orders, Utilities and Decision Rules: A Survey, Management
Science 20, No. 11 (1974), 1442-1471.
Flavell R.B., A New Goal Programming Formulation, OMEGA, The International Journal
of Management Science 4, No.6 (1976), 731-732.
Frank M., Wolfe P., An Algorithm for Quadratic Programming, Naval Research Logistics
Quarterly 3, No. 1-2 (1956), 95-110.
French S., Interactive Multi-Objective Programming: Its Aims, Applications and Demands,
Journal of the Operational Research Society 35, No.9 (1984), 827-834.
Friedman L., Mehrez A., Towards a Theory of Determining Size in Multi-Objective Service
Systems, European Journal of Operational Research 63, No.3 (1992), 398-408.
Friesz T.L., Multiobjective Optimization in Transportation: The Case of Equilibrium Network
Design, Organizations: Multiple Agents with Multiple Criteria, Edited by N.N. Morse,
Lecture Notes in Economics and Mathematical Systems 190, Springer-Verlag, Berlin,
Heidelberg, 1981, pp. 116-127.
Gal T., On Efficient Sets in Vector Maximum Problems - A Brief Survey, European Journal
of Operational Research 24, No.2 (1986),253-264.
Gal T., Postoptimal Analyses, Parametric Programming, and Related Topics: Degeneracy,
Multicriteria Decision Making, Redundancy, Second Edition, Walter de Gruyter & Co.,
Berlin, 1995.
Gal T., Hanne T., On the Development and Future Aspects of Vector Optimization and
MCDM, Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidel-
berg, 1997, pp. 130-145.
Gal T., Leberling H., Redundant Objective Functions in Linear Vector Maximum Problems
and Their Determination, European Journal of Operational Research 1, No.3 (1977),
176-184.
Gal T., Wolf K., Stability in Vector Maximization - A Survey, European Journal of Opera-
tional Research 25, No.2 (1986), 169-182.
Gardiner L.R., Steuer R.E., Unified Interactive Multiple Objective Programming, European
Journal of Operational Research 74, No.3 (1994a), 391-406.
Gardiner L.R., Steuer R.E., Unified Interactive Multiple Programming: An Open Architecture
for Accommodating New Procedures, Journal of the Operational Research Society 45, No.
12 (1994b), 1456-1466.
Gardiner L.R., Vanderpooten D., Interactive Multiple Criteria Procedures: Some Reflections,
Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997,
pp. 290-301.
Gass S., Saaty T., The Computational Algorithm for the Parametric Objective Function,
Naval Research Logistics Quarterly 2 (1955), 39-45.
Geoffrion A.M., Proper Efficiency and the Theory of Vector Maximization, Journal of Math-
ematical Analysis and Applications 22, No.3 (1968), 618-630.
Geoffrion A.M., Dyer J.S., Feinberg A., An Interactive Approach for Multi-Criterion Opti-
mization, with an Application to the Operation of an Academic Department, Manage-
ment Science 19, No.4 (1972)' 357-368.
266 References

Geromel J.C., Ferreira P.A.V., An Upper Bound on Properly Efficient Solutions on Multi-
objective Optimization, Operations Research Letters 10, No.2 (1991), 83-86.
Gershon M., DuckBtein L., An Algorithm lor Choosing 01 a Multiobjective Technique, Essays
and Surveys on Multiple Criteria Decision Making, Edited by P. Hansen, Lecture Notes
in Economics and Mathematical Systems 209, Springer-Verlag, Berlin, Heidelberg, 1983,
pp.53-62.
Ghosh D., Pal B.B., Basu M., Implementation 01 Goal Programming in Long-Range Resource
Planning in University Management, Optimization 24, No. 3-4 (1992), 373-383.
Giannikos t, El-Darzi E., Lees P., An Integer Goal Programming Model to Allocate Offices
to Staff in an Academic Institution, Journal of the Operational Research Society 46, No.
6 (1995), 713-720.
Gibson M., Bernardo J.J., Chung C., Badinelli R., A Comparison 01 Interactive Mu/tiple-
Objective Decision Making Procedures, Computers & Operations Research 14, No. 2
(1987), 97-105.
Giokas D., Vassiloglou M., A Goal Programming Model lor Bank Assets and Liabilities
Management, European Journal of Operational Research 50, No.1 (1991), 48-60.
Goh C.J., Yang X.Q., Analytic Efficient Solution Set lor Multi-Criteria Quadratic Programs,
European Journal of Operational Research 92, No.1 (1996), 166-181.
Giipfert A., Multicriterial Duality, Examples and Advances, Large-Scale Modelling and In-
teractive Decision Analysis, Edited by G. Fandel, M. Grauer, A. Kurzhanski, A.P.
Wierzbicki, Lecture Notes in Economics and Mathematical Systems 273, Springer-Verlag,
1986, pp. 52-58.
Gouljashki V.G., Kirilov L.M., Narula S.C., Vassilev V.S., A Reference Direction Inter-
active Algorithm 01 the Multiple Objective Nonlinear Integer Programming, MUltiple
Criteria Decision Making: Proceedings of the Twelfth International Conference, Hagen
(Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical
Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 308-317.
Granat J., Kreglewski K., Paczynski J., Stachurski A., IAC-DIDASN++ Modular Modeling
and Optimization System Theoretical Foundations, Report of the Institute of Automatic
Control, Warsaw University of Technology (1994a).
Granat J., Kreglewski K., Paczynski J., Stachurski A., IAC-DIDASN++ Modular Modeling
and Optimization System Users Guide, Report of the Institute of Automatic Control,
Warsaw University of Technology (1994b).
Grauer M., A Dynamic Interactive Decision Analysis and Support System (DIDASS), User's
Guide, Working Paper WP-83-60, IIASA, Laxenburg (1983a).
Grauer M., Reference Point Optimization - The Nonlinear Case, Essays and Surveys on
Multiple Criteria Decision Making, Edited by P. Hansen, Lecture Notes in Economics
and Mathematical Systems 209, Springer-Verlag, Berlin, Heidelberg, 1983b, pp. 126-135.
Grauer M., Merten U., Analysis of Multimedia Concepts for Decision Support, Arbeits-
berichte Nr. 20, Universitat - Gesamthochschule - Siegen, 1995.
Grauer M., Lewandowski A., Wierzbicki A., DIDASS - Theory, Implementation and Expe-
riences, Interactive Decision Analysis, Edited by M. Grauer, A.P. Wierzbicki, Lecture
Notes in Economics and Mathematical Systems 229, Springer-Verlag, 1984, pp. 22-30.
Graves S.B., Ringuest J.L., Bard J.F., Recent Developments in Screening Methods for Non-
dominated Solutions in Mu/tiobjectiue Optimization, Computers & Operations Research
19, No.7 (1992), 683-694.
Guddat J., Guerra Vasquez F., Tammer K., Wendler K., Multiobjective and Stochastic Op-
timization Based on Parametric Optimization, Akademie-Verlag, Berlin, 1985.
Gulati T.R., Islam M.A., Proper Efficiency in a Linear Fractional Vector Maximum Problem
with Genemlized Convex Constraints, European Journal of Operational Research 36, No.
3 (1988), 339-345.
Gulati T.R., Islam M.A., Efficiency and Proper Efficiency in Nonlinear Vector Maximum
Problems, European Journal of Operational Research 44, No.3 (1990), 373-382.
 References 267

Haimes Y.Y., The Surrogate Worth Trade-off (SWT) Method and Its Extensions, Multiple
Criteria Decision Making Theory and Application, Edited by G. Fandel, T. Gal, Lecture
Notes in Economics and Mathematical Systems 177, Springer-Verlag, Berlin, Heidelberg,
1980, pp. 85-108.
Haimes Y.Y., Multiple-Criteria Decisionmaking: A Retrospective Analysis, IEEE Transac-
tions on Systems, Man, and Cybernetics 15, No.3 (1985), 313-315.
Haimes Y.Y., Chankong V., Kuhn-Thcker Multipliers as Trade-Offs in Multiobjective Deci-
sion-Making Analysis, Automatica 15, No.1 (1979), 59-72.
Haimes Y.Y., Hall W.A., Multiobjectives in Water Resource Systems Analysis: The Surrogate
Worth Trade Off Method, Water Resources Research 10, No.4 (1974), 615-624.
Haimes Y.Y., Li D., Hierarchical Multiobjective Analysis for Large-Scale Systems: Review
and Current Status, Automatica 24, No.1 (1988), 53-69.
Haimes Y.Y., Lasdon L.S., Wismer D.A., On a Bicriterion Formulation of the Problems
of Integrated System Identification and System Optimization, IEEE Transactions on
Systems, Man, and Cybernetics 1 (1971),296-297.
Haimes Y.Y., Hall W.A., Freedman H.T., Multiobjective Optimization in Water Resources
Systems, Elsevier Scientific Publishing Company, Amsterdam, 1975.
Haimes Y.Y., Tarvainen K, Shima T., Thadathil J., Hierarchical Multiobjective Analysis of
Large-Scale Systems, Hemisphere Publishing Corporation, New York, 1990.
Hall W.A., Haimes Y.Y., The Surrogate Worth Trade-Off Method with Multiple Decision-
Makers, Multiple Criteria Decision Making Kyoto 1975, Edited by M. Zeleny, Lecture
Notes in Economics and Mathematical Systems 123, Springer-Verlag, Berlin, Heidelberg,
1976, pp. 207-233.
Hallefjord A., Jiirnsten K, An Entropy Target-Point Approach to Multiobjective Program-
ming, International Journal of Systems Science 17, No.4 (1986), 639-653.
Halme M., Korhonen P., Nondominated Tradeoffs and Termination in Interactive Multiple
Objective Linear Programming, Improving Decision Making in Organisations, Edited
by A.G. Lockett, G. Islei, Lecture Notes in Economics and Mathematical Systems 335,
Springer-Verlag, Berlin, Heidelberg, 1989, pp. 410-423.
Harrison T.P., Rosenthal R.E., Optimization of Utility and Value Functions, Naval Research
Logistics 35, No.3 (1988), 411-418.
Haslinger J., Neittaanmaki P., Finite Element Approximation for Optimal Shape Design:
Theory and Applications, John Wiley & Sons, Inc., Chichester, 1988.
Hazen G.B., Differential Characterizations of Nonconical Dominance in Multiple Objective
Decision Making, Mathematics of Operations Research 13, No.1 (1988), 174-189.
Helbig S., On the Connectedness of the Set of Weakly Efficient Points of a Vector Optimiza-
tion Problem in Locally Convex Spaces, Journal of Optimization Theory and Applications
65, No.2 (1990), 257-270.
Helbig S., Approximation of the Efficient Point Set by Perturbation of the Ordering Cone,
ZOR - Methods and Models of Operations Research 35, No.3 (1991), 197-220.
Hemaida R.S., Kwak N.K, A Linear Goal Programming Model for Trans-Shipment Prob-
lems with Flexible Supply and Demand Constraints, Journal of the Operational Research
Society 45, No.2 (1994), 215-224.
Hemming T., Multiobjective Decision Making under Certainty, Doctoral Thesis, The Eco-
nomic Research Institute at the Stockholm School of Economics, Stockholm, 1978.
Hemming T., Some Modifications of a Large Step Gradient Method for Interactive Multi-
criterion Optimization, Organizations: Multiple Agents with Multiple Criteria, Edited
by N.N. Morse, Lecture Notes in Economics and Mathematical Systems 190, Springer-
Verlag, Berlin, Heidelberg, 1981, pp. 128-139.
Henig M.I., Existence and Characterization of Efficient Decisions with Respect to Cones,
Mathematical Programming 23, No.1 (1982a), 111-116.
Henig M.l., Proper Efficiency with Respect to Cones, Journal of Optimization Theory and
Applications 36, No.3 (1982b), 387-407.
268 References

Henig M.L, The Domination Property in Multicriteria Optimization, Journal of Mathemat-

ical Analysis and Applications 114, No.1 (1986), 7-16.
Henig M.L, Value Functions, Domination Cones and Proper Efficiency in Multicriteria
Optimization, Mathematical Programming 46, No.2 (1990), 205-217.
Henig M.L, Buchanan J.T., Generalized Tradeoff Directions in Multiobjective Optimization
Problems, Multiple Criteria Decision Making - Proceedings of the Tenth International
Conference: Expand and Enrich the Domains of Thinking and Application, Edited by
G.H. Tzeng, H.F. Wand, U.P. Wen, P.L. Yu, Springer-Verlag, New York, 1994, pp. 47-56.
Henig M.L, Buchanan J.T., Tradeoff Directions in Multiobjective Optimization Problems,
Mathematical Programming 78, No.3 (1997), 357-374.
Hobbs B.F., What Can We Learn From Experiments in Multiobjective Decision Analysis?,
IEEE Transactions on Systems, Man, and Cybernetics 16, No.3 (1986), 384-394.
Hobbs B.F., Chankong V., Hamadeh W., Stakhiv E.Z., Does Choice of Multicriteria Method
Matter? An Experiment in Water Resources Planning, Water Resources Research 28,
No.7 (1992),1767-1779.
Horst R., Thoai N.V., Utility Function Programs and Optimization over the Efficient Set in
Multiple-Objective Decision Making, Journal of Optimization Theory and Applications
92, No.3 (1997), 605-631.
Hu Y., Theory and Application of Multiobjective Optimization in China - A Survey, Annals
of Operations Research 24, No. 1-4 (1990), 45-51.
Hussein M.L., EI-Ghaffar F.S.A., An Interactive Approach for Vector Optimization Problems,
European Journal of Operational Research 89, No.1 (1996), 185-192.
Hwang C.-L., Lin M.-J., Group Decision Making under Multiple Criteria: Methods and Ap-
plications, Lecture Notes in Economics and Mathematical Systems 281, Springer-Verlag,
New York, 1987.
Hwang C.-L., Masud A.S.M., Multiple Objective Decision Making - Methods and Applica-
tions: A State-of-the-Art Survey, Lecture Notes in Economics and Mathematical Systems
164, Springer-Verlag, Berlin, Heidelberg, 1979.
Hwang C.L., Yoon K., Multiple Attribute Decision Making Methods and Applications: A
State-of-the-Art Survey, Lecture Notes in Economics and Mathematical Systems 186,
Springer-Verlag, Berlin, Heidelberg, 1981.
Hwang C.L., Paidy S.R., Yo on K., Masud A.S.M., Mathematical Programming with Multiple
Objectives: A Tutorial, Computers & OperationS Research 7 (1980), 5-31.
Ignizio J.P., Goal Programming and Extensions, Lexington Books, D.C. Heath and Company,
1976.
Ignizio J.P., Generalized Goal Programming: An Overview, Computers & Operations Re-
search 10, No.4 (1983a), 277-289.
Ignizio J.P., An Approach to the Modeling and Analysis of Multiobjective Generalized Net-
works, European Journal of Operational Research 12, No.4 (1983b), 357-361.
Ignizio J.P., Introduction to Linear Goal Programming, Sage Publications, Sage University
Press, Beverly Hills, 1985.
Inuiguchi M., Kume Y., Goal Programming Problems with Interval Coefficients and Target
Intervals, European Journal of Operational Research 52, No.3 (1991), 345-360.
Isermann H., Proper Efficiency and the Linear Vector Maximization Problem, Operations
Research 22, No.1 (1974), 189-191.
Isermann H., Steuer R.E., Computational Ezperience Concerning Payoff Tables and Mini-
mum Criterion Values over the Efficient Set, European Journal of Operational Research
33, No.1 (1988), 91-97.
Ishizuka Y., Shimizu K., Necessary and Sufficient Conditions for the Efficient Solutions of
Nondifferentiable Multiobjective Problems, IEEE Transactions on Systems, Man, and
Cybernetics 14, No.4 (1984), 625-629.
Jahn J., Duality in Vector Optimization, Mathematical Programming 25, No.3 (1983),
343-353.
 References 269

Jahn J., Scalarization in Vector Optimization, Mathematical Programming 29, No.2 (1984),
203-218.
Jahn J., Scalarization in Multi Objective Optimization, Mathematics of Multi Objective
Optimization, Edited by P. Serafini, Springer-Verlag, Wien, New York, 1985, pp. 45-88.
Jahn J., Mathematical Vector Optimization in Partially Ordered Linear Spaces, Methoden
und Verfahren der mathematischen Physik, Band 31, Verlag Peter Lang GmbH, Frankfurt
am Main, 1986a..
Jahn J., Existence Theorems in Vector Optimization, Journal of Optimization Theory and
Applications 50, No.3 (1986b), 397-406.
Jahn J., Parametric Approximation Problems Arising in Vector Optimization, Journal of
Optimization Theory and Applications 54, No.3 (1987), 503-516.
Janssen R., Multiobjective Decision Support for Environmental Management, Kluwer Aca-
demic Publishers, Dordrecht, 1992.
Jarvenpaa S.L., The Effect of Task Demands and Graphical Format on Information Pro-
cessing Strategies, Management Science 35, No.3 (1989), 285-30:-1.
Jaszkiewicz A., Slowinski R., The Light Beam Search over a Non-Dominated Surface of a
Multiple-Objective Programming Problem, Multiple Criteria Decision Making - Proceed-
ings of the Tenth International Conference: Expand and Enrich the Domains of Thinking
and Application, Edited by G.H. Tzeng, H.F. Wand, V.P. Wen, P.L. Yu, Springer-Verlag,
New York, 1994, pp. 87-99.
Jaszkiewicz A., Slowinski R., The Light Beam Search - Outranking Based Interactive Proce-
dure for Multiple-Objective Mathematical Programming, Advances in Multicriteria Anal-
ysis, Edited by P.M. Pardalos, Y. Siskos, C. Zopounidis, Kluwer Academic Publishers,
Dordrecht, 1995, pp. 129-146.
Jedrzejowicz P., Rosicka L., Multicriterial Reliability Optimization Problem, Foundations of
Control Engineering 8, No. 3-4 (1983), 165-173.
Jelassi M.T., Ozernoy V.M., A Framework for Building an Expert System for MCDM Models
Selection, Improving Decision Making in Organisations, Edited by A.G. Lockett, G. Islei,
Lecture Notes in Economics and Mathematical Systems 335, Springer-Verlag, Berlin,
Heidelberg, 1989, pp. 553-562.
Jelassi M.T., Jarke M., Stohr E.A., Designing a Generalized Multiple Criteria Decision
Support System, Decision Making with Multiple Objectives, Edited by Y.Y. Haimes, V.
Chankong, Lecture Notes in Economics and Mathematical Systems 242, Springer-Verlag,
1985, pp. 214-235.
Jendo S., Multicriteria Optimization of Single-Layer Cable System, Archives of Mechanics
38, No. :-I (1986), 219-234.
Jeyakumar V., Yang X.Q., Convex Composite Multi-Objective Nonsmooth Programming,
Mathematical Programming 59, No. :-I (1993),325-:-143.
Jurkiewicz E., Stability of Compromise Solution in Multicriteria Decision-Making Problems,
Journal of Optimization Theory and Applications 40, No.1 (1983), 77-83.
Kacprzyk J., Orlovski S.A. (Eds.), Optimization Models Using Fuzzy Sets and Possibility
Theory, Theory and Decision Library, Series B: Mathematical and Statistical Methods,
D. Reidel Publishing Company, 1987.
Kaliszewski I., A Characterization of Properly Efficient Solutions by an Augmented Tcheby-
chef] Norm, Bulletin of the Polish Academy of Sciences - Technical Sciences 33, No. 7-8
(1985), 415-420.
Kaliszewski I., Norm Scalarization and Proper Efficiency in Vector Optimization, Founda-
tions of Control Engineering 11, No.3 (1986), 117-131.
Kaliszewski I., A Modified Weighted Tchebychef] Metric for Multiple Objective Programming,
Computers & Opera.tions Research 14, No.4 (1987), 315-323.
Kaliszewski I., Quantitative Pareto Analysis by Cone Separation Technique, Kluwer Aca-
demic Publishers, 1994.
Kaliszewski I., A Theorem on Nonconvex Functions and Its Application to Vector Optimiza-
tion, European Journal of Operational Research 80, No.2 (1995), 439-445.
270 References

Kaliszewski I., Michalowski W., Generation of Outcomes with Selectively Bounded Trade-
Offs, Foundations of Computing and Decision Sciences 20, No.2 (1995), 113-122.
Kaliszewski I., Michalowski W., Efficient Solutions and Bounds on Tradeoffs, Journal of
Optimization Theory and Applications 94, No.2 (1997), 381-394.
Kaliszewski I., Michalowski W., Kersten G., A Hybrid Interactive Technique for the MCDM
Pmblems, Essays in Decision Making: A Volume in Honour of Stanley Zionts, Edited
by M.H. Karwan, J. Spronk, J. Wallenius, Springer-Verlag, Berlin, Heidelberg, 1997,
pp.48-59.
Kasanen E., Ostermark R., Zeleny M., Gestalt System of Holistic Graphics: New Manage-
ment Support View 01 MCDM, Computers & Operations Research 18, No.2 (1991),
233-239.
Kaul R.N., Lyall V., Kaur S., Semilocal Pseudolinearity and Efficiency, European Journal
of Operational Research 36, No.3 (1988), 402-409.
Keeney R.L., Raiffa H., Decisions with Multiple Objectives: Preferences and Value Tradeoffs,
John Wiley & Sons, Inc., 1976.
Kim S.H., Gal T., A New Interactive Algorithm for Multi-Objective Linear Programming Us-
ing Maximally Changeable Dominance Cone, European Journal of Operational Research
64, No.1 (1993), 126-137.
Kim S.-H., Ahn B.-S., Choi S.-H., An Efficient Force Planning System Using Multi-Objective
Linear Goal Programming, Computers & Operations Research 24, No.6 (1997), 569-580.
Kirilov L.M., Vassilev V.S., A Method for Solving Multiple Objective Linear Programming
Problems, Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidel-
berg, 1997, pp. 302-309.
Kitagawa H., Watanabe N., Nishimura Y., Matsubara M., Some Pathological Configura-
tions 01 Noninlerior Set Appearing in Multicriteria Optimization Problems 01 Chemical
Processes, Journal of Optimization Theory and Applications 38, No.4 (1982), 541-:-563.
Kiwiel KC., An Aggregate Subgradient Descent Method for Solving Large Convex Nonsmooth
Multiobjective Minimization Problems, Large Scale Systems: Theory and Applications
1983, Edited by A. Straszak, Pergamon Press, 1984, pp. 283-288.
Kiwiel KC., A Descent Method lor Nonsmooth Convex Multiobjective Minimization, Large
Scale Systems 8, No.2 (1985a), 119-129.
Kiwiel KC., An Algorithm for Linearly Constrained Nonsmooth Convex Multiobjective Min-
imization, Systems Analysis and Simulation 1985 Part I: Theory and Foundations, Edited
by A. Sydow, S.M. Thoma, R. Vichnevetsky, Akademie-Verlag, Pergamon Press, Berlin,
1985b, pp. 236-238.
Kiwiel KC., Methods of Descent for Nondifferentiable Optimization, Lecture Notes in Math-
ematics 1133, Springer-Verlag, Berlin, Heidelberg, 1985c.
Kiwiel KC., A Method for Solving Certain Quadratic Programming Problems Arising in
Nonsmooth Optimization, IMA Journal of Numerical Analysis 6, No.2 (1986), 137-152.
Kiwiel KC., Proximity Control in Bundle Methods lor Convex Nondifferentiable Minimiza-
tion, Mathematical Programming 46, No.1 (1990), 105-122.
Kiwiel KC., Stachurski A., Issues of Effectiveness Arising in the Design 01 a System 01
Nondifferentiable Optimization Algorithms, Aspiration Based Decision Support Systems:
Theory, Software and Applications, Edited by A. Lewandowski, A. Wierzbicki, Lecture
Notes in Economics and Mathematical Systems 331, Springer-Verlag, 1989, pp. 180-192.
Klimberg R., GRADS: A New Graphical Display System lor Visualizing Multiple Criteria
Solutions, Computers & Operations Research 19, No.7 (1992), 707-711.
Klinger A., Improper Solutions 01 the Vector Maximum Problem, Operations Research 15,
No.3 (1967), 570-572.
Kok M., Scalarization and the Interface with Decision Makers in Interactive Multi Objec-
tive Linear Programming, Mathematics of Multi Objective Optimization, Edited by P.
Serafini, Springer-Verlag, Wien, New York, 1985, pp. 433-438.
 References 271

Kok M., The Interface with Decision Makers and Some Experimental Results in Interactive
Multiple Objective Programming Method, European Journal of Operational Research 26,
No.1 (1986), 96-107.
Kok M., Lootsma F.A., Pairwise-Comparison Methods in Multiple Objective Programming,
with Applications in a Long-Term Energy-Planning Model, European Journal of Opera-
tional Research 22, No.1 (1985), 44-55.
Kiiksalan M.M., Moskowitz H., Solving the Multiobjective Decision Making Problem Using
a Distance Function, Multiple Criteria Decision Making - Proceedings of the Tenth
International Conference: Expand and Enrich the Domains of Thinking and Application,
Edited by G.H. Tzeng, H.F. Wand, U.P. Wen, P.L. Yu, Springer-Verlag, New York, 1994,
pp. 101-107.
Koopmans T.C., Analysis and Production as an Efficient Combination of Activities, Activity
Analysis of Production and Allocation, Edited by T.C. Koopmans, Yale University Press,
New Haven, London, 1971 (originally published in 1951), pp. 33-97.
Korhonen P.J., Solving Discrete Multiple Criteria Problems by Using Visual Interaction,
Large-Scale Modelling and Interactive Decision Analysis, Edited by G. Fandel, M. Grauer,
A. Kurzhanski, A.P. Wierzbicki, Lecture Notes in Economics and Mathematical Systems
273, Springer-Verlag, 1986, pp. 176-185.
Korhonen P.J., VIG (A Visual Interactive Approach to Goal Programming), User's Guide,
NumPlan (1987).
Korhonen P.J., The Multiobjective Linear Programming Decision Support System VIG and
Its Applications, Readings in Multiple Criteria Decision Aid, Edited by C.A. Bana e
Costa, Springer-Verlag, Berlin, Heidelberg, 1990, pp. 471-491.
Korhonen P.J., Two Decision Support Systems for Continuous and Discrete Multiple Criteria
Decision Making: VIG and VIMDA, Methodology, Implementation and Applications
of Decision Support Systems, Edited by A. Lewandowski, P. Serafini, M.G. Speranza,
International Centre for Mechanical Sciences, Courses and Lectures No. 320, Springer-
Verlag, Wien, New York, 1991a, pp. 85-103.
Korhonen P.J., Using Harmonious Houses for Visual Pairwise Comparison of Multiple Cri-
teria Alternatives, Decision Support Systems 7, No.1 (1991b), 47-54.
Korhonen P.J., Reference Direction Approach to Multiple Objective Linear Programming:
Historical Overview, Essays in Decision Making: A Volume in Honour of Stanley Zionts,
Edited by M.H. Karwan, J. Spronk, J. Wallenius, Springer-Verlag, Berlin, Heidelberg,
1997, pp. 74-92.
Korhonen P., Halme M., Using Lexicographic Parametric Programming for Searching a Non-
dominated Set in Multiple Objective Linear Programming, Journal of Multi-Criteria De-
cision Analysis 5, No.4 (1996), 291-300.
Korhonen P., Laakso J., A Visual Interactive Method for Solving the Multiple-Criteria Prob-
lem, Interactive Decision Analysis, Edited by M. Grauer, A.P. Wierzbicki, Lecture Notes
in Economics and Mathematical Systems 229, Springer-Verlag, 1984, pp. 146-153.
Korhonen P., Laakso J., On Developing a Visual Interactive Multiple Criteria Method -
An Out/ine, Decision Making with Multiple Objectives, Edited by Y.Y. Haimes, V.
Chankong, Lecture Notes in Economics and Mathematical Systems 242, Springer-Verlag,
Berlin, Heidelberg, 1985, pp. 272-281.
Korhonen P., Laakso J., A Visual Interactive Method for Solving the Multiple Criteria Prob-
lem, European Journal of Operational Research 24, No.2 (1986a), 277-287.
Korhonen P., Laakso J., Solving Generalized Goal Programming Problems Using a Visual
Interactive Approach, European Journal of Operational Research 26, No.3 (1986b),
355-363.
Korhonen P., Narula S.C., An Evolutionary Approach to Support Decision Making with
Linear Decision Models, Journal of Multi-Criteria Decision Analysis 2, No.2 (1993),
111-119.
Korhonen P., Wallenius J., A Pareto Race, Naval Research Logistics 35 (1988),615-623.
272 References

Korhonen P., Wallenius J., A Careful Look at Efficiency and Utility in Multiple Criteria
Decision Making: A Tutorial, Asia-Pacific Journal of Operational Research 6, No. 1
(1989a), 46-62.
Korhonen P., Wallenius J., Observations Regarding Choice Behaviour in Interactive Multiple
Criteria Decision-Making Environments: An Experimental Investigation, Methodology
and Software for Interactive Decision Support, Edited by A. Lewandowski, I. Stanchev,
Lecture Notes in Economics and Mathematical Systems 337, Springer-Verlag, 1989b,
pp.163-170.
Korhonen P., WalleniuB J., VIG - A Visual and Dynamic Decision Support System for
Multiple Objective Linear Programming, Multiple Criteria Decision Making and Risk
Analysis Using Microcomputers, Edited by B. Karpak, S. Zionts, Springer-Verlag, Berlin,
Heidelberg, 1989c, pp. 251-281.
Korhonen P., Wallenius J., A Multiple Objective Linear Programming Decision Support Sys-
tem, Decision Support Systems 6, No.3 (1990), 243-251.
Korhonen P., Wallen ius J., Behavioural Issues in MCDM: Neglected Research Questions,
Journal of Multi-Criteria Decision Analysis 5, No.3 (1996), 178-182.
Korhonen P., Wallenius J., Behavioral Issues in MCDM: Neglected Research Questions,
Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997,
pp. 412-422.
Korhonen P., Wallen ius J., Zionts S., Solving the Discrete Multiple Criteria Problem Using
Convex Cones, Management Science 30, No. 11 (1984), 1336-1345.
Korhonen P., Moskowitz H., Wallenius J., Choice Behavior in Interactive Multiple-Criteria
Decision Making, Annals of Operations Research 23, No. 1-4 (1990), 161-179.
Korhonen P., Moskowitz H., Wallenius J., Multiple Criteria Decision Support - A Review,
European Journal of Operational Research 63, No.3 (1992a), 361-375.
Korhonen P., Wallen ius J., Zionts S., A Computer Graphics-Based Decision Support System
for Multiple Objective Linear Programming, European Journal of Operational Research
60, No.3 (1992b), 280-286.
Korhonen P., Salo S., Steuer R.E., A Heuristic for Estimating Nadir Criterion Values in
Multiple Objective Linear Programming, Operations Research 45, No.5 (1997), 751-
757.
Kornbluth J.S.H., A Survey of Goal Programming, OMEGA 1, No.2 (1973), 193-205.
Koski J., Silvennoinen R., Norm Methods and Partial Weighting in Multicriterion Optimiza-
tion of Structures, International Journal for Numerical Methods in Engineering 24, No.
6 (1987), 1101-1121.
Kostreva M.M., Ordoyne T.J., Wiecek M., Multiple-Objective Programming with Polynomial
Objectives and Constraints, European Journal of Operational Research 57, No.3 (1992),
381-394.
Kostreva M.M., Zheng Q., Zhuang D., A Method for Approximating Solutions of Multicri-
terial Nonlinear Optimization Problems, Optimization Methods and Software 5, No.3
(1995), 209-226.
Kreglewski T., Nonlinear Optimization Techniques in Decision Support Systems, Aspira-
tion Based Decision Support Systems: Theory, Software and Applications, Edited by A.
Lewandowski, A. Wierzbicki, Lecture Notes in Economics and Mathematical Systems
331, Springer-Verlag, 1989, pp. 158-171.
Kreglewski T., Paczynski J., Wierzbicki A.P., IAC-DIDAS-N A Dynamic Interactive De-
cision Analysis and Support System for Multicriteria Analysis of Nonlinear Models on
Professional Microcomputers, Theory, Software and Testing Examples for Decision Sup-
port Systems, Edited by A. Lewandowski, A. Wierzbicki, Working Paper WP-87-26,
IIASA, Laxenburg, Austria, 1987, pp. 177-192.
Kreglewski T., Granat J., Wierzbicki A.P., IAS-DIDAS-N: A Dynamic Interactive Decision
Analysis and Support System for Multicriteria Analysis of Nonlinear Models, v. 4.0,
Collaborative Paper CP-91-010, IIASA, Laxenburg (1991).
 References 273

Kuhn H.W., Tucker A.W., Nonlinear Programming, Proceedings of the Second Berkeley
Symposium on Mathematical Statistics and Probability, Edited by J. Neyman, University
of California Press, Berkeley, Los Angeles, 1951, pp. 481-492.
Kuk H., Tanino T., Tanaka M., Sensitivity Analysis in Parametrized Convex Vector Opti-
mization, Journal of Mathematical Analysis and Applications 202, No.2 (1996), 511-522.
Kumar P., Singh N., Tewari N.K., A Nonlinear Goal Programming Model for Multistage,
Multiobjective Decision Problems with Application to Grouping and Loading Problem in
a Flexible Manufacturing System, European Journal of Operational Research 53, No.2
(1991), 166-171.
Larichev OJ., Polyakov O.A., Nikiforov A.O., Multicriterion Linear Programming Problems,
Journal of Economic Psychology 8 (1987), 389-407.
Lazimy R., Solving Multiple Criteria Problems by Interactive Decomposition, Mathematical
Programming 35, No.3 (1986a), 334-361.
Lazimy R., Interactive Relaxation Method for a Broad Class of Integer and Continuous Non-
linear Multiple Criteria Problems, Journal of Mathematical Analysis and Applications
116, No.2 (1986b), 553-573.
Lee G.M., On Efficiency in Nonlinear Fractional Vector Maximization Problem, Optimiza-
tion 25, No.1 (1992), 47-52.
Levary R.R., Optimal Control Problems with Multiple Goal Objectives, Optimal Control
Applications & Methods 1, No.2 (1986), 201-207.
Lewandowski A., Man-Machine Dialogue Interfaces in Decision Support Systems, Large-
Scale Modelling and Interactive Decision Analysis, Edited by G. Fandel, M. Grauer, A.
Kurzhanski, A.P. Wierzbicki, Lecture Notes in Economics and Mathematical Systems
273, Springer-Verlag, 1986, pp. 161-175.
Lewandowski A., Granat J., Dynamic BIPLOT as the Interaction Interface for Aspiration
Based Decision Support Systems, Multiple Criteria Decision Support, Edited by P. Kor-
honen, A. Lewandowski, J. Wallenius, Lecture Notes in Economics and Mathematical
Systems 356, Springer-Verlag, 1991, pp. 229-241.
Lewandowski A., Grauer M., The Reference Point Optimization Approach - Methods of
Efficient Implementation, Working Paper WP-82-26, IIASA, Laxenburg (1982).
Lewandowski A., Rogowski T., Kreglewski T., A Trajectory-Oriented Extension of DIDASS
and Its Applications, Plural Rationality and Interactive Decision Processes, Edited by M.
Grauer, M. Thompson, A.P. Wierzbicki, Lecture Notes in Economics and Mathematical
Systems 248, Springer- Verlag, Berlin, Heidelberg, 1985a, pp. 261-268.
Lewandowski A., Rogowski T., Kreglewski T., Application of DIDASS Methodology to a
Flood Control Problem - Numerical Experiments, Plural Rationality and Interactive
Decision Processes, Edited by M. Grauer, M. Thompson, A.P. Wierzbicki, Lecture Notes
in Economics and Mathematical Systems 248, Springer-Verlag, Berlin, Heidelberg, 1985b,
pp. 325-328.
Lewandowski A., Kreglewski T., Rogowski T., Wierzbicki A.P., Decision Support Systems of
DIDAS Family (Dynamic Interactive Decision Analysis & Support), Theory, Software
and Testing Examples for Decision Support Systems, Edited by A. Lewandowski, A.
Wierzbicki, Working Paper WP-87-26, IIASA, Laxenburg, 1987, pp. 4-26.
Li D., Convexification of a Noninferior Frontier, Journal of Optimization Theory and Ap-
plications 88, No.1 (1996), 177-196.
Li D., Haimes Y.Y., The Envelope Approach for Multiobjective Optimization Problems, IEEE
Transactions on Systems, Man, and Cybernetics 11, No.6 (1987),1026-1038; Correction,
IEEE Transactions on Systems, Man, and Cybernetics 18, No.2 (1988),332.
Lieberman E.R., Soviet Multi-Objective Programming Methods: An Overview, Multiobjec-
tive Problems of Mathematical Programming, Edited by A. Lewandowski, V. Volkovich,
Lecture Notes in Economics and Mathematical Systems 351, Springer-Verlag, 1991a,
pp.21-31.
Lieberman E.R., Soviet Multi-Objective Mathematical Programming Methods: An Overview,
Management Science 37, No.9 (l991b), 1147-1165.
274 References

Lin J.G., Multiple-Objective Problems: Pareto-Optimal Solutions by Methods of Proper Equal-

ity Constraints, IEEE Transactions on Automatic Control 21, No.5 (1976a), 641-650.
Lin J.G., Three Methods for Determining Pareto-Optimal Solutions of Multiple-Objective
Problems, Directions in Large-Scale Systems, Many-Person Optimization and Decen-
tralized Control, Edited by Y.C. Ho, S.K. Mitter, Plenum Press, New York, 1976b,
pp. 117-138.
Liu J .C., (-Pareto Optimality for Nondifferentiable Multiobjective Programming via Penalty
Function, Journal of Mathematical Analysis and Applications 198, No.1 (1996), 248-
261.
Liu L., The Second-Order Conditions of Nondominated Solutions for C 1 ,1 Generalized Mul-
tiobjective Mathematical Programming, Systems Science and Mathematical Sciences 4,
No.2 (1991), 128-138.
Liu L., Krizek M., The Second Order Optimality Conditions for Nonlinear Mathematical
Programming with C1,1 Data, Applications of Mathematics 42, No.4 (1997), 311-320.
Loganathan G.V., Sherali H.D., A Convergent Interactive Cutting-Plane Algorithm for Mul-
tiobjective Optimization, Operation Research 35, No.3 (1987), 365-377.
Lotti V., Teich J.E., Multiple Criteria Decision Making Using Personal Computers, Working
Paper No. 727, State University of New York at Buffalo, Buffalo, New York (1989).
Lotov A.V., An Estimate of Solution Set Perturbations for a System of Linear Inequalities,
Optimization Methods and Software 6, No.1 (1995), 1-24.
Lotov A.V., Comment on D.J. White "A Characterization of the Feasible Set 0/ Objective
Function Vectors in Linear Multiple Objective Problems", European Journal of Opera-
tional Research 89, No.1 (1996), 215-220.
Lotov A., Chernykh 0., Hellman 0., Multiple Objective Analysis of Long-Term Development
Strategies for a National Economy, European Journal of Operational Research 56, No.
2 (1992), 210-218.
Lotov A.V., Bushenkov V.A., Chernykh O.L., Multicriteria DSS for River Water-Quality
Planning, Microcomputers in Civil Engineering 12 (1997), 57-67.
Lowe T.J., Thisse J.-F., Ward J.E., Wendell R.E., On Efficient Solutions to Multiple Objec-
tive Mathematical Programs, Management Science 30, No. 11 (1984), 1346-1349.
Luc D.T., On the Domination Property in Vector Optimization, Journal of Optimization
Theory and Applications 43, No.2 (1984a), 327-330.
Luc D.T., On Duality Theory in Multiobjective Programming, Journal of Optimization The-
ory and Applications 43, No.4 (1984b), 557-582.
Luc D.T., On Scalarizing Method in Vector Optimization, Large-Scale Modelling and In-
teractive Decision Analysis, Edited by G. Fandel, M. Grauer, A. Kurzhanski, A.P.
Wierzbicki, Lecture Notes in Economics and Mathematical Systems 273, Springer-Verlag,
1986, pp. 46-51.
Luc D.T., About Duality and Alternative in Multiobjective Optimization, Journal of Opti-
mization Theory and Applications 53, No.2 (1987), 303-307.
Luc D.T., Theory of Vector Optimization, Lecture Notes in Economics and Mathematical
Systems 319, Springer-Verlag, Berlin, Heidelberg, 1989.
Luc D.T., Recession Cones and the Domination Property in Vector Optimization, Mathe-
matical Programming 49, No.1 (1990), 113-122.
Luc D.T., On the Properly Efficient Points of Nonconvex Sets, European Journal of Opera-
tional Research 86, No.2 (1995), 262-269.
Luc D.T., Jahn J., Axiomatic Approach to Duality in Optimization, Numerical Functional
Analysis and Optimization 13, No. 3-4 (1992), 305-326.
Luc D.T., Schaible S., Efficiency and Generalized Concavity, Journal of Optimization Theory
and Applications 94, No.1 (1997), 147-153.
Lucchetti R., Some Existence Results and Stability in Multi Objective Optimization, Math-
ematics of Multi Objective Optimization, Edited by P. Serafini, Springer-Verlag, Wien,
New York, 1985, pp. 179-187.
 References 275

Luenberger D.G., Linear and Nonlinear Programming, Second Edition, Addison-Wesley Pub-
lishing Company, Inc., 1984.
MacCrimmon K.R., An Overview of Multiple Objective Decision Making, Multiple Criteria
Decision Making, Edited by J .L. Cochrane, M. Zeleny, University of South Carolina
Press, Columbia, South Carolina, 1973, pp. 18-44.
Majumdar A.A.K., Optimality Conditions in Differentiable Multiobjective Programming,
Journal of Optimization Theory and Applications 92, No.2 (1997), 419-427.
Makela M.M., Issues of Implementing a Fortran Subroutine Package NSOLIB for Nonsmooth
Optimization, Report 5/1993, University of Jyviiskyla, Department. of Mathematics, Lab-
oratory of Scientific Computing, Jyviiskyla (1993).
Makela M.M., Neittaanmaki P., Nonsmooth Optimization: Analysis and Algorithms with
Applications to Optimal Control, World Scientific Publishing Co., Singapore, 1992.
Manas M., Graphical Methods of Multicriterial Optimization, Zeitschrift flir Angewandte
Mathematik und Mechanik 62, No.5 (1982), 375-377.
Mangasarian O.L., Nonlinear Programming, McGraw-Hill, Inc., 1969.
Mareschal B., Brans J.-P., Geometrical Representation for MCDA, European Journal of
Operational Research 34, No.1 (1988), 69-77.
Martein 1., Some Results on Regularity in Vector Optimization, Optimization 20, No.6
(1989), 787-798.
Martel J.-M., Aouni B., Incorporating the Decision-Maker's Preferences in the Goal-Pro-
gramming Model, Journal of the Operational Research Society 41, No. 12 (1990), 1121-
1132.
Martel J.-M., Aouni B., Diverse Imprecise Goal Programming Model Formulations, Journal
of Global Optimization 12, No.2 (1998), 127-138.
Martinez-Legaz J.-E., Lexicographical Order and Duality in Multiobjective Programming,
European Journal of Operational Research 33, No.3 (1988), 342-348.
Martfnez-Legaz J .-E., Singer 1., Surrogate Duality for Vector Optimization, Numerical Func-
tional Analysis and Optimization 9, No. 5-6 (1987), 547-568.
Marusciac I., On Fritz John Type Optimality Criterion in Multi-Objective Optimization,
Revue d'Analyse Numerique et de Theorie de l'Approximation 11, No. 1-2 (1982), 109-
114.
Masud A.S.M., Hwang C.L., Interactive Sequential Goal Programming, Journal of the Oper-
ational Research Society 32 (1981), 391-400.
Masud A.S.M., Zheng X., An Algorithm for Multiple-Objective Non-Linear Programming,
Journal of the Operational Research Society 40, No. 10 (1989), 895-906.
Matos M.A., Borges P., A Flexible Interface for Decision-Aid in Multicriteria Decision Prob-
lems, Multicriteria Analysis, Edited by J. Climaco, Springer-Verlag, Berlin, Heidelberg,
1997, pp. 390-400.
Meisel W.S., 7radeoff Decision in Multiple Criteria Decision Making, MUltiple Criteria De-
cision Making, Edited by J.L. Cochrane, M. Zeleny, University of South Carolina Press,
Columbia, South Carolina, 1973, pp. 461-476.
Michalowski W., Evaluation of a Multiple Criteria Interactive Programming Approach: An
Experiment, INFOR 25, No.2 (1987),165-173.
Michalowski W., MCDM at the Crossroads, Multicriteria Analysis, Edited by J. Climaco,
Springer-Verlag, Berlin, Heidelberg, 1997, pp. 579-584.
Michalowski W., Szapiro T., A Bi-Reference Procedure for Interactive Multiple Criteria
Programming, Operations Research 40, No.2 (1992), 247-258.
Miettinen K., On the Methodology of Multiobjective Optimization with Applications, Doctoral
Thesis, Report 60, University of Jyviiskyla, Department of Mathematics, Jyviiskyla,
1994.
Miettinen K., Makela M.M., Nonsmooth Multicriteria Optimization Applied to Optimal Con-
trol, Reports on Applied Mathematics and Computing, No.6, University of Jyviiskyla,
Jyviiskyla (1991).
276 References

Miettinen K., Makela M.M., An Interactive Method for Nonsmooth Multiobjective Opti-
mization with an Application to Optimal Control, Optimization Methods and Software
2 (1993), 31-44.
Miettinen K., Makela M.M., A Nondifferentiable Multiple Criteria Optimization Method
Applied to Continuous Casting Process, Proceedings of the Seventh European Confer-
ence on Mathematics in Industry, Edited by A. Fasano, M. Primicerio, B.G. Teubner,
Stuttgart, 1994, pp. 255-262.
Miettinen K, Makela M.M., Interactive Bundle-Based Method for Nondifferentiable Multi-
objective Optimization: NIMBUS, Optimization 34, No.3 (1995), 231-246.
Miettinen K., MakeUi. M.M., NIMBUS - Interactive Method for Nondifferentiable Multiobjec-
tive Optimization Problems, Multi-Objective Programming and Goal Programming: The-
ories and Applications, Edited by M. Tamiz, Lecture Notes in Economics and Mathe-
matical Systems 432, Springer-Verlag, Berlin, Heidelberg, 1996a, pp. 50-57.
Miettinen K., Makela M.M., Comparing Two Versions 0/ NIMBUS Optimization System,
Report 23/1996, University of Jyvii.skyla, Department of Mathematics, Laboratory of
Scientific Computing, Jyvii.skyla, 1996b.
Miettinen K., Makela M.M., Interactive Method NIMBUS for Nondifferentiable Multiob-
jective Optimization Problems, Multicriteria Analysis, Edited by J. Climaco, Springer-
Verlag, Berlin, Heidelberg, 1997, pp. 310-319.
Miettinen K, Miikela M.M., Theoretical and Computational Comparison 0/ Multiobjective
Optimization Methods NIMBUS and RD, Report 5/1998, University of Jyviiskyla, De-
partment of Mathematics, Laboratory of Scientific Computing, Jyviiskyla, 1998a.
Miettinen K., Makela M.M., Interactive MCDM Support System in the Internet, Trends
in Multicriteria Decision Making: Proceedings of the 13th International Conference on
Multiple Criteria Decision Making, Edited by T. Stewart, R. van den Honert, Springer-
Verlag, 1998b, pp. 419-428.
Miettinen K., Makela M.M., Makinen R.A.E., Interactive Multiobjective Optimization Sys-
tem NIMBUS Applied to Nonsmooth Structural Design Problems, System Modelling and
Optimization, Proceedings of the 17th IFIP Conference on System Modelling and Op-
timization, Prague, Czech Republic, Edited by J. Dolezal, J. Fidler, Chapman & Hall,
London, 1996a, pp. 379-385.
Miettinen K., Makela M.M., Mannikko T., Nondifferentiable Multiobjective Optimizer NIM-
BUS Applied to an Optimal Control Problem of Continuous Casting, Report 22/1996,
University of Jyvii.skyla, Department of Mathematics, Laboratory of Scientific Comput-
ing, Jyviiskyla, 1996b.
Miller G.A., The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity
/or Processing In/ormation, Psychological Review 63, No.2 (1956), 81-87.
Minami M., Weak Pareto Optimality of Multiobjective Problem in a Banach Space, Bulletin
of Mathematical Statistics 19, No. 3-4 (1980-81), 19-23.
Minami M., Weak Pareto Optimality of Multiobjective Problems in a Locally Convex Linear
Topological Space, Journal of Optimization Theory and Applications 34, No.4 (1981),
469-484.
Minami M., Weak Pareto-Optimal Necessary Conditions in a Nondifferentiable Multiobjec-
tive Program on a Banach Space, Journal of Optimization Theory and Applications 41,
No.3 (1983), 451-461.
Mishra S.K., Lagrange Multipliers Saddle Point and Scalarizations in Composite Multiob-
jective Nonsmooth Programming, Optimization 38, No.2 (1996), 93-105.
Mishra S.K., Mukherjee R.N., Generalized Convex Composite Multi-Objective Nonsmooth
Programming and Conditional Proper EJficiency, Optimization 34, No.1 (1995), 53-66.
Mitani K., Nakayama H., A Multiobjective Diet Planning Support System Using the Satis-
ficing Trade-Off Method, Journal of Multi-Criteria Decision Analysis 6, No.3 (1997),
131-139.
Mitra A., Patankar J.G., A Multi-Objective Model for Warranty Estimation, European Jour-
nal of Operational Research 45, No. 2-3 (1990), 347-355.
 References 277

Miyaji I., Ohno K., Mine H., Solution Method for Partitioning Students into Groups, Euro-
pean Journal of Operational Research 33, No.1 (1988),82-90.
Mocci U., Primicerio L., Ring Network Design: An MCDM Approach, Multiple Criteria De-
cision Making: Proceedings of the Twelfth International Conference, Hagen (Germany),
Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical Systems
448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 491-500.
Moldavskiy M.A., Singling Out a Set of Undominated Solutions in Continuous Vector Op-
timization Problems, Soviet Automatic Control 14, No.5 (1981), 47-53.
Monarchi D.E., Kisiel C.C., Duckstein L., Interactive Multiobjective Programming in Water
Resources: A Case Study, Water Resources Research 9, No.4 (1973), 837-850.
Monarchi D.E., Weber J.E., Duckstein L., An Interactive Multiple Objective Decision-Making
Aid Using Nonlinear Goal Programming, Multiple Criteria Decision Making Kyoto 1975,
Edited by M. Zeleny, Lecture Notes in Economics and Mathematical Systems 123,
Springer-Verlag, Berlin, Heidelberg, 1976, pp. 235-253.
Mor6n M.A., Romero C., Ruiz del Portal F.R., Generating Well-Behaved Utility Functions
for Compromise Programming, Journal of Optimization Theory and Applications 91,
No.3 (1996), 643-649.
Mote J., Olson D.L., Venkataramanan M.A., A Comparative Multiobjective Programming
Study, Mathematical and Computer Modelling 10, No. 10 (1988), 719-729.
M'silti A., Tolla P., An Interactive Multiobjective Nonlinear Programming Procedure, Euro-
pean Journal of Operational Research 64, No.1 (1993), 115-125.
Mukai H., Algorithms for Multicriterion Optimization, IEEE Transactions on Automatic
Control 25, No.2 (1980),177-186.
Musselman K., Talavage J., A Tradeoff Cut Approach to Multiple Objective Optimization,
Operations Research 28, No.6 (1980), 1424-1435.
Nakayama H., Geometric Consideration of Duality in Vector Optimization, Journal of Op-
timization Theory and Applications 44, No.4 (1984), 625-655.
Nakayama H., On the Components in Interactive Multiobjective Programming Methods, Plu-
ral Rationality and Interactive Decision Processes, Edited by M. Grauer, M. Thompson,
A.P. Wierzbicki, Lecture Notes in Economics and Mathematical Systems 248, Springer-
Verlag, Berlin, Heidelberg, 1985a, pp. 234-247.
Nakayama H., Lagrange Duality and Its Geometric Interpretation, Mathematics of Multi
Objective Optimization, Edited by P. Serafini, Springer-Verlag, Wien, New York, 1985b,
pp. 105-127.
Nakayama H., Duality Theory in Vector Optimization: An Overview, Decision Making with
Multiple Objectives, Edited by Y.Y. Haimes, V. Chankong, Lecture Notes in Economics
and Mathematical Systems 242, Springer-Verlag, 1985c, pp. 109-125.
Nakayama H., Sensitivity and Trade-Off Analysis in Multiobjective Programming, Method-
ology and Software for Interactive Decision Support, Edited by A. Lewandowski, I.
Stanchev, Lecture Notes in Economics and Mathematical Systems 337, Springer-Ver-
lag, 1989, pp. 86-93.
Nakayama H., Satisficing Trade-Off Method for Problems with Multiple Linear Fractional
Objectives and Its Applications, Multiobjective Problems of Mathematical Programming,
Edited by A. Lewandowski, V. Volkovich, Lecture Notes in Economics and Mathematical
Systems 351, Springer-Verlag, 1991a, pp. 42-50.
Nakayama H., Trade-off Analysis Based upon Parametric Optimization, Multiple Criteria
Decision Support, Edited by P. Korhonen, A. Lewandowski, J. Wallenius, Lecture Notes
in Economics and Mathematical Systems 356, Springer-Verlag, 1991b, pp. 42-52.
Nakayama II., Trade-Off Analysis Using Parametric Optimization Techniques, European
Journal of Operational Research 60, No.1 (1992a), 87-98.
Nakayama H., Theoretical Remarks on Dynamic Trade-Off, Multiple Criteria Decision Mak-
ing: Proceedings of the Ninth International Conference: Theory and Applications in
Business, Industry, and Government, Edited by A. Goicoechea, L. Duckstein, S. Zionts,
Springer-Verlag, New York, 1992b, pp. 297-309.
278 References

Nakayama H., Engineering Applications of Multi-Objective Programming: Recent Results,

MUltiple Criteria Decision Making - Proceedings of the Tenth International Confer-
ence: Expand and Enrich the Domains of Thinking and Application, Edited by G.H.
Tzeng, H.F. Wand, U.P. Wen, P.L. Yu, Springer-Verlag, New York, 1994, pp. 369-378.
Nakayama H., Aspiration Level Approach to Interactive Multi-Objective Programming and
Its Applications, Advances in Multicriteria Analysis, Edited by P.M. Pardalos, Y. Siskos,
C. Zopounidis, Kluwer Academic Publishers, Dordrecht, 1995, pp. 147-174.
Nakayama H., Some Remarks on Dualization in Vector Optimization, Journal of Multi-
Criteria Decision Analysis 5, No.3 (1996), 218-225.
Nakayama H., Furukawa K., Satisficing Trade-Off Method with an Application to Multiob-
jective Structural Design, Large Scale Systems 8 (1985), 47-57.
Nakayama H., Sawaragi Y., Satisficing Trade-Off Method for Multiobjective Programming,
Interactive Decision Analysis, Edited by M. Grauer, A.P. Wierzbicki, Lecture Notes in
Economics and Mathematical Systems 229, Springer-Verlag, 1984, pp. 113-122.
Nakayama H., Tanino T., Sawaragi Y., An Interactive Optimization Method in Multicriteria
Decisionmaking, IEEE Transactions on Systems, Man, and Cybernetics 10, No.3 (1980),
163-169.
Nakayama H., Nomura J., Sawada K., Nakajima R., An Application of Satisficing Trade-
Off Method to a Blending Problem of Industrial Materials, Large-Scale Modelling and
Interactive Decision Analysis, Edited by G. Fandel, M. Grauer, A. Kurzhanski, A.P.
Wierzbicki, Lecture Notes in Economics and Mathematical Systems 273, Springer-Verlag,
1986, pp. 303-313.
Nakayama H., Kaneshige K., Takemoto S., Watada Y., An Application of a Multi-Objective
Programming Technique to Construction Accuracy Control of Cable-Stayed Bridges, Eu-
ropean Journal of Operational Research 87, No.3 (1995), 731-738.
Nakayama H., Kaneshige K., Takemoto S., Watada Y., A Construction Accuracy Control
System of Cable Stayed Bridge Using a Multi-Objective Programming Technique, Mul-
tiple Criteria Decision Making: Proceedings of the Twelfth International Conference,
Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Math-
ematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 501-509.
Narula S.C., Weistroffer H.R., Algorithms for Multi-Objective Nonlinear Programming Prob-
lems: An Overview, Improving Decision Making in Organisations, Edited by A.G. Lock-
ett, G. Islei, Lecture Notes in Economics and Mathematical Systems 335, Springer-Verlag,
Berlin, Heidelberg, 1989a, pp. 434-443.
Narula S.C., Weistroffer H.R., A Flexible Method for Nonlinear Multicriteria Decisionmaking
Problems, IEEE Transactions on Systems, Man, and Cybernetics 19, No.4 (1989b), 883-
887.
Narula S.C., Kirilov L., Vassilev V., An Interactive Algorithm for Solving Multiple Objective
Nonlinear Programming Problems, Multiple Criteria Decision Making - Proceedings of
the Tenth International Conference: Expand and Enrich the Domains of Thinking and
Application, Edited by G.H. Tzeng, H.F. Wand, U.P. Wen, P.L. Yu, Springer-Verlag,
New York, 1994a, pp. 119-127.
Narula S.C., Kirilov L., Vassilev V., Reference Direction Approach for Solving Multiple Ob-
jective Nonlinear Programming Problems, IEEE Transactions on Systems, Man, and
Cybernetics 24, No.5 (1994b), 804-806.
Ng K.Y., A Multicriteria Optimization Approach to Aircraft Loading, Operations Research
40, No.6 (1992), 1200-1206.
Nijkamp P., Spronk J., Interactive Multiple Goal Programming: An Evaluation and Some
Results, Multiple Criteria Decision Making Theory and Applications, Edited by G. Fan-
del, T. Gal, Lecture Notes in Economics and Mathematical Systems 177, Springer-Verlag,
Berlin, Heidelberg, 1980, pp. 278-293.
Nijkamp P., Rietveld P., Spronk J., Open Problems in the Operationalization of Multiple
Criteria Decision Making, Systems Analysis Modelling Simulation 5, No.4 (1988), 311-
322.
 References 279

Ogryczak W., Preemptive Reference Point Method, Multicriteria Analysis, Edited by J. Clf-
maco, Springer-Verlag, Berlin, Heidelberg, 1997a, pp. 156-167.
Ogryczak W., Reference Distribution - An Interactive Approach to Multiple Homogeneous
and Anonymous Criteria, Multiple Criteria Decision Making: Proceedings of the Twelfth
International Conference, Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes
in Economics and Mathematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997b,
pp. 156-165.
Ohkubo S., Dissanayake P.B.R., Taniwaki K., An Approach to Multicriteria Fuzzy Optimiza-
tion of a Prestressed Concrete Bridge System Considering Cost and Aesthetic Feeling,
Structural Optimization 15, No.2 (1998), 132-140.
Ohta H., T. Yamaguchi, Linear Fractional Goal Programming in Consideration of Fuzzy
Solution, European Journal of Operational Research 92, No.1 (1996), 157-165.
Olbrisch M., The Interactive Reference Point Approach as a Solution Concept for Econo-
metric Decision Models, X. Symposium on Operations Research, Part 1, Sections 1-5,
Edited by M.J. Beckmann, K.-W. Gaede, K. Ritter, H. Schneeweiss, Verlag Anton Hain
Meisenheim GmbH, Knigstein/Ts., 1986, pp. 611-619.
Olkucu A., Conflicting Objectives in Software System Design, Multiple Criteria Decision
Making and Risk Analysis Using Microcomputers, Edited by B. Karpak, S. Zionts,
Springer-Verlag, Berlin, Heidelberg, 1989, pp. 357-394.
Olson D.L., Review of Empirical Studies in Multiobjective Mathematical Programming: Sub-
ject Reflection of Nonlinear Utility and Learning, Decision Sciences 23, No. 1 (1992),
1-20.
Olson D.L., Tchebycheff Norms in Multi-Objective Linear Programming, Mathematical and
Computer Modelling 17, No.1 (1993),113-124.
Oppenheimer K.R., A Proxy Approach to Multi-Attribute Decision Making, Management
Science 24, No.6 (1978), 675-689.
Osman M.S.A., Ragab A.M., A Unified Approach for Characterizing the Noninferior Solu-
tions of Multiobjective Nonlinear Programming Problems, X. Symposium on Operations
Research, Part 1, Sections 1-5, Edited by M.J. Beckmann, K.-W. Gaede, K. Ritter, H.
Schneeweiss, Verlag Anton Hain Meisenheim GmbH, Knigstein/Ts., 1986a, pp. 133-141.
Osman M.S.A., Ragab A.M., An Algorithm for Solving Multiobjective Nonlinear Program-
ming Problems, X. Symposium on Operations Research, Part 1, Sections 1-5, Edited by
M.J. Beckmann, K.-W. Gaede, K. Ritter, H. Schneeweiss, Verlag Anton Hain Meisenheim
GmbH, Knigstein/Ts., 1986b, pp. 143-147.
Osyczka A., Multicriterion Optimization in Engineering with FORTRAN Programs, Ellis
Horwood Limited, 1984.
Osyczka A., Multicriterion Decision Making with Min-Max Approach, Optimization Methods
in Structural Design, Edited by H. Eschenauer, N. Olhoff, Euromech-Colloquium 164,
Wissenschaftsverlag, 1989a.
Osyczka A., Computer Aided Multicriterion Optimization System (CAMOS), Discretization
Methods and Structural Optimization - Procedures and Applications, Edited by H.A.
Eschenauer, G. Thierauf, Lecture Notes in Engineering, Springer-Verlag, Berlin, Heidel-
berg, 1989b, pp. 263-270.
Osyczka A., Computer Aided Multicriterion Optimization System (CAMOS): Software Pack-
age in FORTRAN, International Software Publishers, 1992.
Osyczka A., Koski J., Selected Works Related to Multicriterion Optimization Methods for En-
gineering Design, Optimization Methods in Structural Design, Edited by H. Eschenauer,
N. Olhoff, Euromech-Colloquium 164, Wissenschaftsverlag, 1989.
Osyczka A., Kundu S., A New Method to Solve Generalized Multicriteria Optimization Prob-
lems Using the Simple Genetic Algorithm, Structural Optimization 10, No. 2 (1995),
94-99.
280 References

Osyczka A., Zajac J., Multicriteria Optimization 0/ Computationally Expensive Functions

and Its Application to Robot Spring Balancing Mechanism Design, Multicriteria De-
sign Optimization Procedures and Applications, Edited by H. Eschenauer, J. Koski, A.
Osyczka, Springer-Verlag, Berlin, Heidelberg, 1990, pp. 168-183.
Ozernoy V.M., Choosing the "Best" Multiple Criteria Decision-Making Method, INFOR 30,
No.2 (1992a), 159-171.
Ozernoy V.M., Developing an Expert System Prototype for Intelligent Decision Support, Mul-
tiple Criteria Decision Making: Proceedings of the Ninth International Conference: The-
ory and Applications in Business, Industry, and Government, Edited by A. Goicoechea,
L. Duckstein, S. Zionts, Springer-Verlag, New York, 1992b, pp. 325-333.
Pal B.B., Basu I., A Goal Programming Method lor Solving Fractional Programming Prob-
lems via Dynamic Programming, Optimization 35, No.2 (1995), 145-157.
Papageorgiou N.S., Pareto Efficiency in Locally Convex Spaces II: Stability Results, Numer-
ical Functional Analysis and Optimization 8, No.1 (1985), 117-136.
Pareto V., Cours d'Economie Politique, Libraire Droz, Geni!Ve, 1964 (the first edition in
1896).
Pareto V., Manual 0/ Political Economy, The MacMillan Press Ltd, 1971 (the original edition
in French in 1927).
Pascoletti A., Serafini P., Scalarizing Vector Optimization Problems, Journal of Optimization
Theory and Applications 42, No.4 (1984), 499-524.
Passy U., Levanon Y., Analysis 0/ Multiobjective Decision Problems by the Indifference Band
Approach, Journal of Optimization Theory and Applications 43, No.2 (1984), 205-235.
Peterson E.S., Applications 0/ Multiobjective Decision Making in the Mining Industry, Pa-
pers on Nonlinear Analysis and Mathematical Systems Theory, Edited by J. Szep, P.
Tallos, Department of Mathematics, Karl Marx University of Economics, Budapest, 1984,
pp.37-65.
Plastria F., Carrizosa E., Geometrical Characterization of Weakly Efficient Points, Journal
of Optimization Theory and Applications 90, No.1 (1996), 217-223.
Polak E., Payne A.N., On Multicriteria Optimization, Directions in Large-Scale Systems,
Many-Person Optimization and Decentralized Control, Edited by Y.C. Ho, S.K. Mitter,
Plenum Press, New York, 1976, pp. 77-94.
Powell J .G., Premachandra I.M., Accommodating Diverse Institutional Investment Objectives
and Constraints Using Non-Linear Goal Programming, European Journal of Operational
Research 105, No.3 (1998), 447-456.
Preda V., On Efficiency and Duality for Multiobjective Programs, Journal of Mathematical
Analysis and Applications 166, No.2 (1992), 365-377.
Preda V., Optimality Conditions and Duality in Multiple Objective Programming Involving
Semilocally Convex and Related Functions, Optimization 36, No.3 (1996), 219-230.
Preda V., Stancu-Minasian I.M., Generalized Mond- Weir Duality lor Multiobjective Non-
smooth Programming, Multiple Criteria Decision Making: Proceedings of the Twelfth
International Conference, Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes
in Economics and Mathematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997,
pp.55-63.
Rail L.B., Automatic Differentiation: Techniques and Applications, Lecture Notes in Com-
puter Science 120, Springer-Verlag, Berlin, Heidelberg, 1981.
Ramesh R., Karwan M.H., Zionts S., Theory 0/ Convex Cones in Multicriteria Decision
Making, Annals of Operations Research 16, No. 1-4 (1988), 131-147.
Ramesh R., Karwan M.H., Zionts S., Interactive Multicriteria Linear Programming: An
Extension 0/ the Method 0/ Zionts and Wallenius, Naval Research Logistics 36, No.3
(1989a), 321-335.
Ramesh R., Karwan M.H., Zionts S., Preference Structure Representation Using Convex
Cones in Multicriteria Integer Programming, Management Science 35, No.9 (1989b),
1092-1105.
 References 281

Roo S.S., Optimization Theory and Applications, Second Edition, Wiley Eastern Limited,
1984.
Roo S.S., Game Theory Approach for Multiobjective Structural Optimization, Computers
and Structures 25, No.1 (1987), 119-127.
Rarig H.M., Haimes Y.Y., Risk/Dispersion Index Method, IEEE Transactions on Systems,
Man, and Cybernetics 13, No.3 (1983), 317-328.
Reeves G.R., Gonzalez J.J., A Comparison of Two Interactive MCDM Procedures, European
Journal of Operational Research 41, No.2 (1989), 203-209.
Reeves G.R., Reid R.C., Minimum Values over the Efficient Set in Multiple Objective Deci-
sion Making, European Journal of Operational Research 36, No.3 (1988), 334-338.
ReVelle C., Equalizing Superpower Force Disparities with Optimized Arms Control Choices:
A Multi-Objective Approach, European Journal of Operational Research 33, No.1 (1988),
46-53.
Rietveld P., Multiple Objective Decision Methods and Regional Planning, North-Holland
Publishing Company, 1980.
Ringuest J .L., Multiohjective Optimization: Behavioral and Computational Considerations,
Kluwer Academic Publisher, 1992.
Rogowski T., Sobczyk J., Wierzbicki A., IAC-DIDAS-L: A Dynamic Interactive Decision
Analysis and Support System for Multicriteria Analysis of Linear and Dynamic Lin-
ear Models of Professional Microcomputers, Theory, Software and Testing Examples for
Decision Support Systems, Edited by A. Lewandowski, A. Wierzbicki, Working Paper
WP-87-26, IIASA, Laxenburg, 1987, pp. 106-124.
Romero C., Handbook of Critical Issues in Goal Programming, Pergamon Press, 1991.
Romero C., Goal Programming and Multiple Criteria Decision Making: Some Reflections,
Multiple Criteria Decision Making: Proceedings of the Twelfth International Conference,
Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Math-
ematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 192-198.
Rosenthal R.E., Goal Programming - A Critique, New Zealand Operational Research 11,
No.1 (1983), 1-7.
Rosenthal R.E., Principles of Multiobjective Optimization, Decision Sciences 16, No. 2
(1985), 133-152.
Rosinger E.E., Interactive Algorithm for Multiobjective Optimization, Journal of Optimiza-
tion Theory and Applications 35, No.3 (1981), 339-365; Errata Comge, Journal of
Optimization Theory and Applications 38, No.1 (1982), 147-148.
Rosinger E.E., Aids for Decision Making with Conflicting Objectives, Mathematics of Multi
Objective Optimization, Edited by P. Serafini, Springer-Verlag, Wien, New York, 1985,
pp. 275-315.
Rothermel M.A., Schilling D.A., Conjoint Measurement in Multiple Objective Decision Mak-
ing: A New Approach, European Journal of Operational Research 23, No.3 (1986),
31(}--319.
Roy A., Mackin P., Multicriteria Optimization (Linear and Nonlinear) Using Proxy Value
Functions, Multiple Criteria Decision Support, Edited by P. Korhonen, A. Lewandowski,
J. Wallenius, Lecture Notes in Economics and Mathematical Systems 356, Springer-
Verlag, 1991, pp. 128-134.
Roy A., Wallenius J., Nonlinear Multiobjective Optimization: An Algorithm and Some The-
ory, Mathematical Programming 55, No.2 (1992), 235-249.
Roy B., The Outranking Approach and the Foundations of ELECTRE Methods, Readings in
Multiple Criteria Decision Aid, Edited by C.A. Bana e Costa, Springer-Verlag, Berlin,
Heidelberg, 1990, pp. 155-183.
Roy B., Mousseau V., A Theoretical Framework for Analysing the Notion of Relative Impor-
tance 0/ Criteria, Journal of Multi-Criteria Decision Analysis 5, No.2 (1996), 145-159.
Rufz-Canales P., Rufian-Lizana A., A Characterization 0/ Weakly Efficient Points, Mathe-
matical Programming 68, No.2 (1995), 205-212.
282 References

Saber H.M., Ravindran A., A Partitioning Gradient Based (PGB) Algorithm for Solving
Nonlinear Goal' Programming Problems, Computers & Operations Research 23, No.2
(1996), 141-152.
Sadagopan S., Ravindran A., Interactive Algorithms for Multiple Criteria Nonlinear Pro-
gramming Problems, European Journal of Operational Research 25, No.2 (1986), 247-
257.
Sadek I.S., Bruch J.C. Jr., Sloss J.M., Adali S., Structural Control of a Variable Cross-
Section Beam by Distributed Forces, Mechanics of Structures and Machines 16, No.3
(1988-89), 313-333.
Sainfort F.C., Gustafson D.H., Bosworth K., Hawkins R.P., Decision Support Systems Ef-
fectiveness: Conceptual Framework and Empirical Evaluation, Organizational Behavior
and Human Decision Processes 45 (1990), 232-252.
Sakawa M., Interactive Multiobjective Decision Making by the Sequential Proxy Optimization
Technique: SPOT, European Journal of Operational Research 9, No.4 (1982), 386-396.
Sakawa M., Mori N., Interactive Multiobjective Decisionmaking for Nonconvex Problems
Based on the Weighted Tchebycheff Norm, Large Scale Systems 5, No.1 (1983), 69-82.
Sakawa M., Mori N., Interactive Multiobjective Decisionmaking for Nonconvex Problems
Based on the Penalty Scalarizing Functions, European Journal of Operational Research
17, No.3 (1984), 320-330.
Sakawa M., Seo F., Interactive Multiobjective Decisionmaking for Large-Scale Systems and
Its Application to Environmental Systems, IEEE Transactions on Systems, Man, and
Cybernetics 10, No. 12 (1980), 796-806.
Sakawa M., Seo F., Interactive Multiobjective Decision Making by the Sequential Proxy Op-
timization Technique (SPOT) and Its Application to Environmental Systems, Control
Science and Technology for the Progress of Society, Edited by H. Akashi, IFAC, vol. 2,
Pergamon Press, 1982a, pp. 1527-1532.
Sakawa M., Seo F., Interactive Multiobjective Decision Making in Environmental Systems
Using Sequential Proxy Optimization Techniques (SPOT), Automatica 18, No.2 (1982b),
155-165.
Sakawa M., Yano H., Interactive Multiobjective Reliability Design of a Standby System by
the Fuzzy Sequential Proxy Optimization Technique (FSPOT), International Journal of
Systems Science 16, No.2 (1985), 177-195.
Sakawa M., Yano H., Trade-Off Rates in the Hyperplane Method for Multiobjective Optimiza-
tion Problems, European Journal of Operational Research 44, No.1 (1990), 105-118.
Sakawa M., Yano H., Generalized Hyperplane Methods for Characterizing A-Extreme Points
and Trade-Off Rates for Multiobjective Optimization Problems, European Journal of
Operational Research 57, No.3 (1992), 368-380.
Sankaran S., Multiple Objective Decision Making Approach to Cell Formation: A Goal Pro-
gramming Model, Mathematical and Computer Modelling 13, No.9 (1990), 71-81.
Sarma G.V., Merouani H.F., Some Difficulties in Applying Interactive Compromise Pro-
gramming Illustrated by a Recent Method and a Case Study, Journal of the Operational
Research Society 46, No.1 (1995), 9-19.
Sawaragi Y., Nakayama H., Tanino T., Theory of Multiobjective Optimization, Academic
Press, Inc., Orlando, Florida, 1985.
Schilling D.A., ReVelle C., Cohon J., An Approach to the Display and Analysis of Multiob-
jective Problems, Socio-Economic Planning Sciences 17, No.2 (1983), 57-63.
Schniederjans M.J., Goal Programming: Methodology and Applications, Kluwer Academic
Publishers, Boston, Dordrecht, London, 1995a.
Schniederjans M.J., The Life Cycle of Goal Programming Research as Recorded in Journal
Articles, Operations Research 43, No.3 (1995b), 551-557.
Schniederjans M.J., Hoffman J., Multinational Acquisition Analysis: A Zero-One Goal Pro-
gramming Model, European Journal of Operational Research 62, No.2 (1992), 175-185.
 References 283

Sherali H.D., Equivalent Weights for Lexicographic Multi-Objective Programs: Characteriza-

tions and Computations, European Journal of Operational Research 11, No.4 (1982),
367-379.
Sherali H.D., Soyster A.L., Preemptive and Nonpreemptive Multi-Objective Programming:
Relationships and Counterexamples, Journal of Optimization Theory and Applications
39, No.2 (1983), 173-186.
Shimizu K, Ishizuka Y., Bard J.F., Nondifferentiable and Two-Level Mathematical Program-
ming, Kluwer Academic Publishers, 1997.
Shin W.S., Ravindran A., Interactive Multiple Objective Optimization: Survey I - Continu-
ous Case, Computers & Operations Research 18, No.1 (1991), 97-114.
Shin W.S., Ravindran A., A Comparative Study of Interactive Tradeoff Cutting Plane Meth-
ods/or MOMP, European Journal of Operational Research 56, No.3 (1992), 380-393.
Shor N.Z., Minimization Methods for Non-Differentiable Functions, Springer-Verlag, Berlin,
1985.
Silverman J., Steuer R.E., Whisman A.W., Computer Graphics at the Multicriterion Com-
puter/User Inter/ace, Decision Making with Multiple Objectives, Edited by Y.Y. Haimes,
V. Chankong, Lecture Notes in Economics and Mathematical Systems 242, Springer-Ver-
lag, Berlin, Heidelberg, 1985, pp. 201-213.
Silverman J., Steuer R.E., Whisman A.W., A Multi-Period, Multiple Criteria Optimization
System for Manpower Planning, European Journal of Operational Research 34, No.2
(1988), 160--170.
Simon C.P., Scalar and Vector Maximization: Calculus Techniques with Economics Applica-
tions, Studies in Mathematical Economics, Edited by S. Reiter, MAA Studies in Math-
ematics 25, The Mathematical Association of America, 1986, pp. 62-159.
Singh C., Bhatia D., Rueda N., Duality in Nonlinear Multiobjective Programming Using
Augmented Lagrangian Functions, Journal of Optimization Theory and Applications
88, No.3 (1996), 659-670.
Sinha S.B., Sastry S.V.C., Misra R.P., Management 0/ a Community Storage System: A
Goal Programming Model, European Journal of Operational Research 34, No.1 (1988),
92-98.
Skulimowski A.M.J., Classification and Properties of Dominating Points in Vector Optimiza-
tion, Methods of Operations Research 58, Edited by P. Kleinschmidt, F.J. Radermacher,
W. Schweitzer, H. Wildermann, Athenum Verlag, Frankfurt am Main, 1989, pp. 99-112.
Skulimowski A.M.J., Applicability of Ideal Points in Multicriteria Decision Making, Multiple
Criteria Decision Making: Proceedings of the Ninth International Conference: Theory
and Applications in Business, Industry, and Government, Edited by A. Goicoechea, L.
Duckstein, S. Zionts, Springer-Verlag, New York, 1992, pp. 387-400.
Skulimowski A.M.J., Decision Support Systems Based on Reference Sets, Wydawnictwa
AGH, Krak6w, 1996.
Slowinski R., Interactive Multiobjective Optimization Based on Ordinal Regression, Mul-
tiobjective Problems of Mathematical Programming, Edited by A. Lewandowski, V.
Volkovich, Lecture Notes in Economics and Mathematical Systems 351, Springer-Ver-
lag, 1991, pp. 93-100.
Slowinski R., Teghem J. (Eds.), Stochastic versus Fuzzy Approaches to Multiobjective Mathe-
matical Programming under Uncertainty, Kluwer Academic Publishers, Dordrecht, 1990.
Sobol I.M., An Efficient Approach to Multicriteria Optimum Design Problems, Surveys on
Mathematics for Industry 1, No.4 (1992), 259-281.
Sobol' I.M., Levitan Y.L., Error Estimates for the Crude Approximation of the Trade-off
Curve, Multiple Criteria Decision Making: Proceedings of the Twelfth International Con-
ference, Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes in Economics
and Mathematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 83-92.
Soland R.M., Multicriteria Optimization: A General Characterization of Efficient Solutions,
Decision Sciences 10, No.1 (1979), 26-38.
284 References

Soloveychik D., The Multiple Criteria Decision-Making Process Using the Methods of Mul-
tivariate Statistical Analysis, Fifth International Conference on Collective Phenomena,
Edited by J.L. Lebowitz, The New York Academy of Sciences, New York, 1983, pp. 205-
212.
Song A., Cheng W.-M., A Method for Multihuman and Multi-Criteria Decision Making,
Systems Analysis and Simulation 1988 I: Theory and Foundations, Edited by A. Sydow,
S.G. Tzafestas, R. Vichnevetsky, Akademie-Verlag, Berlin, 1988, pp. 213-216.
Sounderpandian J., Value Functions When Decision Criteria Are Not Totally Substitutable,
Operations Research 39, No.4 (1991), 592-600.
Soyibo A., Goal Programming Methods and Applications: A Survey, Journal of Information
and Optimization Sciences 6, No.3 (1985), 247-264.
Spronk J., Interactive Multifactorial Planning: State of the Art, Readings in Multiple Criteria
Decision Aid, Edited by C.A. Bana e Costa, Springer-Verlag, Berlin, Heidelberg, 1990,
pp. 512-534.
Stadler W., A Survey of Multicriteria Optimization or the Vector Maximum Problem, Part
I: 1776-1960, Journal of Optimization Theory and Applications 29, No.1 (1979), 1-52.
Stadler W. (Ed.), Multicriteria Optimization in Engineering and in the Sciences, Plenum
Press, New York, 1988a.
Stadler W., Fundamentals of Multicriteria Optimization, Multicriteria Optimization in En-
gineering and in the Sciences, Edited by W. Stadler, Plenum Press, New York, 1988b,
pp. 1-25.
Staib T., On Necessary and Sufficient Optimality Conditions for Multicriterial Optimization
Problems, ZOR - Methods and Models of Operations Research 35, No.3 (1991), 231-248.
Starn A., Lee Y.-R., Yu P.L., Value Functions and Preference Structures, Mathematics of
Multiobjective Optimization, Edited by P. Serafini, CISM Courses and Lectures 289,
1985, pp. 1-22.
Starn A., Kuula M., Cesar H., Transboundary Air Pollution in Europe: An Interactive Multi-
criteria Tradeoff Analysis, European Journal of Operational Research 56, No.2 (1992),
263-277.
Stancu-Minasian I.M., Stochastic Programming with MUltiple Objective Functions, Math-
ematics and Its Applications (East European Series), D. Reidel Publishing Company;
Editura Academiei, 1984.
Statnikov R.B., Matusov J., Use of Pr-Nets for the Approximation of the Edgeworth-Pareto
Set in Multicriteria Optimization, Journal of Optimization Theory and Applications 91,
No.3 (1996), 543-560.
Sterna-Karwat A., Continuous Dependence of Solutions on a Parameter in a Scalarization
Method, Journal of Optimization Theory and Applications 55, No.3 (1987), 417-434.
Steuer R.E., Multiple Criteria Optimization: Theory, Computation, and Applications, John
Wiley & Sons, Inc., 1986.
Steuer R.E., The Tchebycheff Procedure 01 Interactive Multiple Objective Programming, Mul-
tiple Criteria Decision Making and Risk Analysis Using Microcomputers, Edited by B.
Karpak, S. Zionts, Springer-Verlag, Berlin, Heidelberg, 1989a, pp. 235-249.
Steuer R.E., Trends in Interactive Multiple Objective Programming, Methodology and Soft-
ware for Interactive Decision Support, Edited by A. Lewandowski, I. Stanchev, Lecture
Notes in Economics and Mathematical Systems 337, Springer-Verlag, 1989b, pp. 107-119.
Steuer R.E., Implementin9 the Tchebycheff Method in a Spreadsheet, Essays in Decision
Making: A Volume in Honour of Stanley Zionts, Edited by M.H. Karwan, J. Spronk,
J. Wallenius, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 93-103.
Steuer R.E., Choo E.-U., An Interactive Weighted Tchebycheff Procedure lor Multiple Ob-
jective Programming, Mathematical Programming 26, No.3 (1983), 326-344.
Steuer R.E., Gardiner L.R., Interactive Multiple Objective Programming: Concepts, Current
Status, and Future Directions, Readings in Multiple Criteria Decision Aid, Edited by
C.A. Bana e Costa, Springer-Verlag, Berlin, Heidelberg, 1990, pp. 413-444.
 References 285

Steuer R.E., Gardiner L.R., On the Computational Testing of Procedures for Interactive
Multiple Objective Linear Programming, Operations Research, Edited by G. Fandel, H.
Gehring, Springer-Verlag, Berlin, Heidelberg, 1991, pp. 121-131.
Steuer R.E., Sun M., The Parameter Space Investigation Method of Multiple Objective Non-
linear Programming: A Computational Investigation, Operations Research 43, No. 4
(1995), 641-648.
Steuer R.E., Whisman A.W., Toward the Consolidation of Interactive Multiple Objective
Programming Procedures, Large-Scale Modelling and Interactive Decision Analysis, Ed-
ited by G. Fandel, M. Grauer, A. Kurzhanski, A.P. Wierzbicki, Lecture Notes in Eco-
nomics and Mathematical Systems 273, Springer-Verlag, 1986, pp. 232-241.
Steuer R.E., Silverman J., Whisman A.W., A Combined Tchebycheff/Aspiration Criterion
Vector Interactive Multiobjective Programming Procedure, Management Science 39, No.
10 (1993), 1255-1260.
Steuer R.E., Gardiner L.R., Gray J., A Bibliographic Survey of the Activities and Interna-
tional Nature of Multiple Criteria Decision Making, Journal of Multi-Criteria Decision
Analysis 5, No.3 (1996), 195-217.
Stewart T.J., A Critical Survey on the Status of Multiple Criteria Decision Making Theory
and Practice, OMEGA 20, No. 5-6 (1992), 569-586.
Stewart T.J., Convergence and Validation of Interactive Methods in MCDM: Simulation
Studies, Essays in Decision Making: A Volume in Honour of Stanley Zionts, Edited by
M.H. Karwan, J. Spronk, J. Wallenius, Springer-Verlag, Berlin, H.lidelberg, 1997, pp. 7-
18.
Sultan A.M., Templeman A.B., Generation of Pareto Solutions by Entropy-Based Meth-
ods, Multi-Objective Programming and Goal Programming: Theories and Applications,
Edited by M. Tamiz, Lecture Notes in Economics and Mathematical Systems 432, Spring-
er-Verlag, Berlin, Heidelberg, 1996, pp. 164-195.
Sun M., Starn A., Steuer R.E., Solving Multiple Objective Programming Problems Using Feed-
Forward Artificial Neural Networks: The Interactive FFANN Pmcedure, Management
Science 42, No.6 (1996), 835-849.
Sunaga T., Mazeed M.A., Kondo E., A Penalty Function Formulation for Interactive Mul-
tiobjective Programming Problems, System Modelling and Optimization, Edited by M.
Iri, K. Yajima, Lecture Notes in Control and Information Sciences 113, Springer-Verlag,
1988, pp. 221-230.
Szidarovszky F., Szenteleki K., A Multiobjective Optimization Model for Wine Production,
Applied Mathematics and Computation 22, No. 2-3, (1987), 255-275.
Tabucanon M.T., Multiple Criteria Decision Making in Industry, Elsevier Science Publishers
B.V., Amsterdam, 1988.
Tamiz M., Jones D.F., Algorithmic Improvements to the Method of Martel and Aouni, Jour-
nal of the Operational Research Society 46, No.2 (1995), 254-257.
Tamiz M., Jones D.F., An Overview of Current Solution Methods and Modelling Practices
in Goal Programming, Multi-Objective Programming and Goal Programming: Theories
and Applications, Edited by M. Tamiz, Lecture Notes in Economics and Mathematical
Systems 432, Springer-Verlag, Berlin, Heidelberg, 1996, pp. 198-211.
Tamiz M., Jones D.F., Interactive Frameworks for Investigation of Goal Programming Mod-
els: Theory and Practice, Journal of Multi-Criteria Decision Analysis 6, No.1 (1997a),
52-60.
Tamiz M., Jones D.F., A General Interactive Goal Programming Algorithm, Multiple Crite-
ria Decision Making: Proceedings of the Twelfth International Conference, Hagen (Ger-
many), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical
Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997b, pp. 433-444.
Tamiz M., Hasham R., Jones D.F., Hesni B., Fargher E.K., A Two Staged Goal Program-
ming Model for Portfolio Selection, Multi-Objective Programming and Goal Program-
ming: Theories and Applications, Edited by M. Tamiz, Lecture Notes in Economics and
Mathematical Systems 432, Springer-Verlag, Berlin, Heidelberg, 1996, pp. 286-299.
286 References

Tamura K., Arai S., On Proper and Improper Efficient Solutions of Optimal Problems with
Multicriteria, Journal of Optimization Theory and Applications 38, No.2 (1982), 191-
205.
Tan Y.S., Fraser N.M., The Modified Star Graph and the Petal Diagram: Two New Visual
Aids for Discrete Alternative Multicriteria Decision Making, Journal of Multi-Criteria
Decision Analysis 7, No.1 (1998), 20-33.
Tanino T., Sensitivity Analysis in Multiobjective Optimization, Journal of Optimization The-
ory and Applications 56, No.3 (1988a), 479-499.
Tanino T., Stability and Sensitivity Analysis in Convex Vector Optimization, SIAM Journal
on Control and Optimization 26, No.3 (1988b), 521-536.
Tanino T., Stability and Sensitivity Analysis in Multiobjective Nonlinear Programming, An-
nals of Operations Research 27, No. 1-4 (1990), 97-114.
Tanino T., Sawaragi Y., Stability of Nondominated Solutions in Multicriteria Decision-
Making, Journal of Optimization Theory and Applications 30, No.2 (1980), 229-253.
Tapia C.G., Murtagh B.A., The Use of Preference Criteria in Interactive Multiobjective
Mathematical Programming, Asia-Pacific Journal of Operational Research 6, No.2 (1989),
131-147.
Tapia C.G., Murtagh B.A., A Markovian Process in Interactive Multiobjective Decision-
Making, European Journal of Operational Research 57, No.3 (1992), 421-428.
Tarvainen K., On the Implementation of the Interactive Surrogate Worth Trade-Off (ISWT)
Method, Interactive Decision Analysis, Edited by M. Grauer, A.P. Wierzbicki, Lecture
Notes in Economics and Mathematical Systems 229, Springer-Verlag, Berlin, Heidelberg,
1984, pp. 154-161.
Tarvainen K, Duality Theory for Preferences in Multiobjective Decisionmaking, Journal of
Optimization Theory and Applications 88, No.1 (1996), 237-245.
Tecle A., Duckstein L., A Procedure for Selecting MCDM Techniques for Forest Resources
Management, Multiple Criteria Decision Making: Proceedings of the Ninth International
Conference: Theory and Applications in Business, Industry, and Government, Edited by
A. Goicoechea, L. Duckstein, S. Zionts, Springer-Verlag, New York, 1992, pp. 19-32.
Teghem J. Jr., Delhaye C., Kunsch P.L., An Interactive Decision Support System (IDSS)
for Multicriteria Decision Aid, Mathematical and Computer Modelling 12, No. 10-11
(1989), 1311-1320.
Tell R., Wallenius J., A Survey of Multiple-Criteria Decision Methods and Applications: Util-
ity Theory and Mathematical Programming, The Finnish Journal of Business Economics
I, No.1 (1979), 3-22.
Tenhuisen M.L., Wiecek M.M., On the Structure of the Non-Dominated Set for Bicriteria
Programmes, Journal of Multi-Criteria Decision Analysis 5, No.3 (1996), 232-243.
Thach P.T., Konno H., Yokota D., Dual Approach to Minimization on the Set of Pareto-
Optimal Solutions, Journal of Optimization Theory and Applications 88, No.3 (1996),
689-707.
Thore S., Nagurney A., Pan J., Generalized Goal Programming and Variational Inequalities,
Operations Research Letters 12, No.4 (1992), 217-226.
Torn A., A Sampling-Search-Clustering Approach for Solving Scalar (Local, Global) and
Vector Optimization Problems, Theory and Practice of Multiple Criteria Decision Mak-
ing, Edited by C. Carlsson, Y. Kochetkova, North-Holland Publishing Company, 1983,
pp. 119-141.
Udink ten Cate A., On the Determination of the Optimal Temperature for the Growth of an
Early Cucumber Crop in a Greenhouse, Plural Rationality and Interactive Decision Pro-
cesses, Edited by M. Grauer, M. Thompson, A.P. Wierzbicki, Lecture Notes in Economics
and Mathematical Systems 248, Springer-Verlag, Berlin, Heidelberg, 1985, pp. 311-318.
Vanderpooten D., The Use of Preference Information in Multiple Criteria Interactive Pro-
cedures, Improving Decision Making in Organisations, Edited by A.G. Lockett, G. Islei,
Lecture Notes in Economics and Mathematical Systems 335, Springer-Verlag, Berlin,
Heidelberg, 1989a, pp. 390-399.
 References 287

Vanderpooten D., The Interactive Approach in MCDA: A Technical Framework and Some
Basic Conceptions, Mathematical and Computer Modelling 12, No. 10-11 (1989b), 1213-
1220.
Vanderpooten D., Multiobjective Programming: Basic Concepts and Approaches, Stochastic
versus Fuzzy Approaches to Multiobjective Mathematical Programming under Uncer-
tainty, Edited by R. Slowinski, J. Teghem, Kluwer Academic Publishers, Dordrecht,
1990, pp. 7-22.
Vanderpooten D., Three Basic Conceptions Underlying Multiple Criteria Interactive Proce-
dures, Multiple Criteria Decision Making: Proceedings of the Ninth International Con-
ference: Theory and Applications in Business, Industry, and Government, Edited by A.
Goicoechea, L. Duckstein, S. Zionts, Springer-Verlag, New York, 1992, pp. 441-448.
Vanderpooten D., Vincke P., Description and Analysis of Some Representative Interactive
Multicriteria Procedures, Mathematical and Computer Modelling 12, No. 10-11 (1989),
1221-1238.
Vassilev V., Sgurev V., Atanassov A., Deianov A., Kichovich M., KiriIov L., Djambov V.,
Vachkov G., Software Product for Multiobjective Nonlinear Programming: MONP-16,
Version 1.1, General Description; User's Guide, Software Products and Systems Corpo-
ration, Institute of Industrial Cybernetics and Robotics, Sofia, 1990.
Verkama M., Heiskanen P., Comment on a Decision Support Approach for Negotiation: Soft-
ware 'lis. Methodology, European Journal of Operational Research 96, No.1 (1996), 202-
204.
Vessey I., Cognitive Fit: A Theory-Based Analysis of the Graphs vel'sus Tables Literature,
Decision Sciences 22 (1991), 219-240.
Vetschera R., Feedback-Oriented Group Decision Support in a Reference Point Framework,
Multiobjective Problems of Mathematical Programming, Edited by A. Lewandowski, V.
Volkovich, Lecture Notes in Economics and Mathematical Systems 351, Springer-Verlag,
1991a, pp. 309-314.
Vetschera R., A Note on Scalarizing Functions under Changing Sets of Criteria, European
Journal of Operational Research 52, No.1 (1991b), 113-118.
Vetschera R., A Preference-Preserving Projection Technique for MCDM, European Journal
of Operational Research 61, No. 1-2 (1992), 195-203.
Vincke P., Multicriteria Decision-Aid, John Wiley & Sons, Inc., Chichester, 1992.
Wallenius J., Comparative Evaluation of Some Interactive Approaches to Multicriterion Op-
timization, Management Science 21, No. 12 (1975), 1387-1396.
Wallenius J., Zionts S., A Research Project on Multicriteria Decision Making, Conflicting
Objectives in Decision, Edited by D.E. Bell, R. Keeney, H. Raiffa, John Wiley & Sons,
Inc., New York, 1977.
Wan Y.-H., On Local Pareto Optima, Journal of Mathematical Economics 2, No.1 (1975),
35-42.
Wang S., Lagrange Conditions in Nonsmooth and Multiobjective Mathematical Programming,
Mathematics in Economics 1 (1984), 183-193.
Wang S., Algorithms for Multiobjective and Nonsmooth Optimization, Methods of Oper-
ations Research 58, Edited by P. Kleinschmidt, F.J. Radermacher, W. Schweitzer, H.
Wildermann, Athenum Verlag, Frankfurt am Main, 1989, pp. 131--142.
Wang S., Second-Order Necessary and Sufficient Conditions in Multiobjective Programming,
Numerical Functional Analysis and Application 12, No. 1-2 (1991), 237-252.
Wang S., An Interactive Method for Multicriteria Decision Making, Optimization: Tech-
niques and Applications, vol. 1, Proceedings of the International Conference (ICOTA),
Edited by K.H. Phu3, C.M. Wang, W.Y. Yeong, T.Y. Leong, H.T. Loh, K.C. Tan, F.S.
Chou, World Scientific Publishing Co., Inc., 1992, pp. 307-316.
Wang S., Li Z., Scalarization and Lagrange Duality in Multiobjective Optimization, Opti-
mization 26, No. 3-4 (1992), 315-324.
288 References

Wang S., Zhou X., Multiobjective Optimization in Spillway Profile, Approximation, Opti-
mization and Computing: Theory and Applications, Edited by A.G. Law, C.L. Wang,
Elsevier Science Publishers B.V., 1990, pp. 309-312.
Warburton A., Quasiconcave Vector Maximization: Connectedness of the Sets of Pareto-
Optimal and Weak Pareto-Optimal Alternatives, Journal of Optimization Theory and
Applications 40, No.4 (1983), 537-557.
Weber M., Decision Making with Incomplete Information, European Journal of Operational
Research 28, No.1 (1987), 44-57.
Week M., Fortsch F., Application of Multicriteria Optimization to Structural Systems, Sys-
tem Modelling and Optimization, Edited by M. Iri, K. Yajima, Lecture Notes in Control
and Information Sciences 113, Springer-Verlag, 1988, pp. 471-483.
Weidner P., On the Characterization of Efficient Points by Means of Monotone Functionals,
Optimization 19, No.1 (1988), 53-69.
Weidner P., Complete Efficiency and Interdependencies Between Objective Functions in Vec-
tor Optimization, ZOR - Methods and Models of Operations Research 34, No.2 (1990),
91-115.
Weir T., Proper Efficiency and Duality for Vector Valued Optimization Problems, Journal
of the Australian Mathematical Society, Series A 43, No.1 (1987), 21-34.
Weir T., On Efficiency, Proper Efficiency and Duality in Multiobjective Programming, Asia-
Pacific Journal of Operational Research 7, No.1 (1990), 46-54.
Weistroffer H.R., Multiple Criteria Decision Making with Interactive Over-Achievement Pro-
gramming, Operations Research Letters I, No.6 (1982), 241-245.
Weistroffer H.R., An Interactive Goal-Programming Method for Non-Linear Mu.ltiple-Cri-
teria Decision-Making Problems, Computers & Operations Research 10, No.4 (1983),
311-320.
Weistroffer H.R., A Combined Over- and Under-Achievement Programming Approach to
Multiple Objective Decision-Making, Large Scale Systems 7, No.1 (1984), 47-58.
Weistroffer H.R., Careful Usage of Pessimistic Values is Needed in Multiple Objective Opti-
mization, Operations Research Letters 4, No.1 (1985), 23-25.
Weistroffer H.R., A Flexible Model for Multi-Objective Optimization, Recent Advances and
Historical Developments of Vector Optimization, Edited by J. Jahn, W. Krabs, Lecture
Notes in Economics and Mathematical Systems 294, Springer-Verlag, Berlin, Heidelberg,
1987, pp. 311-316.
Weistroffer H.R., Narula S., The Cu.rrent State of Nonlinear MUltiple Criteria Decision
Making, Operations Research, Edited by G. Fandel, H. Gehring, Springer-Verlag, Berlin,
Heidelberg, 1991, pp. 109-119.
Weistroffer H.R., Narula S.C., The State of Multiple Criteria Decision Support Software,
Annals of Operations Research 72 (1997), 299-313.
Wendell R.E., Lee D.N., Efficiency in Multiple Objective Optimization Problems, Mathemat-
ical Programming 12, No.3 (1977), 406-414.
White D.J., Concepts of Proper Efficiency, European Journal of Operational Research 13,
No.2 (1983a), 180-188.
White D.J., A Selection of Multi-Objective interactive Programming Methods, Multi-Ob-
jective Decision Making, Edited by S. French, R. Hartley, L.C. Thomas, D.J. White,
Academic Press, 1983b, pp. 99-126.
White D.J., A Bibliography on the Applications of Mathematical Programming Multiple-
Objective Methods, Journal of the Operational Research Society 41, No.8 (1990), 669-
691.
Wierzbicki A.P., Basic Properties of Scalarization Functionals for Multiobjective Optimiza-
tion, Mathematische Operationsforschung und Statistik, Ser. Optimization 8, No. 1
(1977), 55-60.
 References 289

Wierzbicki A.P., A Methodological Guide to Multiobjective Optimization, Optimization Tech-

niques, Part 1, Edited by K. Iracki, K. Malanowski, S. Walukiewicz, Lecture Notes in
Control and Information Sciences 22, Springer-Verlag, Berlin, Heidelberg, 1980a, pp. 99-
123.
Wierzbicki A.P., The Use of Reference Objectives in Multiobjective Optimization, Multiple
Criteria Decision Making Theory and Applications, Edited by G. Fandel, T. Gal, Lecture
Notes in Economics and Mathematical Systems 177, Springer-Verlag, Berlin, Heidelberg,
1980b, pp. 468-486.
Wierzbicki A.P., A Mathematical Basis for Satisficing Decision Making, Organizations: Mul-
tiple Agents with Multiple Criteria, Edited by N.N. Morse, Lecture Notes in Economics
and Mathematical Systems 190, Springer-Verlag, Berlin, Heidelberg, 1981, pp. 465-485.
Wierzbicki A.P., A Mathematical Basis for Satisficing Decision Making, Mathematical Mod-
elling 3, No. 25 (1982), 391-405.
Wierzbicki A.P., On the Completeness and Constructiveness of Parametric Characteriza-
tions to Vector Optimization Problems, OR Spektrum 8, No.2 (1986a), 73-87.
Wierzbicki A.P., A Methodological Approach to Comparing Parametric Characterizations of
Efficient Solutions, Large-Scale Modelling and Interactive Decision Analysis, Edited by
G. Fandel, M. Grauer, A. Kurzhanski, A.P. Wierzbicki, Lecture Notes in Economics and
Mathematical Systems 273, Springer-Verlag, 1986b, pp. 27-45.
Wierzbicki A.P., On the Role of Intuition in Decision Making and Some Ways of Multi-
criteria Aid of Intuition, Journal of Multi-Criteria Decision Analysis 6, No.2 (1997a),
65-76.
Wierzbicki A.P., Conllergence of Interactille Procedures of Multiobjective Optmization and
Decision Support, Essays in Decision Making: A Volume in Honour of Stanley Zionts,
Edited by M.H. Karwan, J. Spronk, J. Wallenius, Springer-Verlag, Berlin, Heidelberg,
1997b, pp. 19-47.
Wierzbicki A.P., Granat J., Multi-Objective Modeling for Engineering Application in Deci-
sion Support, Multiple Criteria Decision Making: Proceedings of the Twelfth Interna-
tional Conference, Hagen (Germany), Edited by G. Fandel, T. Gal, Lecture Notes in
Economics and Mathematical Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997,
pp. 529-540.
Winkels H.-M., Meika M., An Integration of Efficiency Projections into the Geoffrion Ap-
proach for Multiobjective Linear Programming, European Journal of Operational Re-
search 16, No.1 (1984), 113-127.
Wood E., Greis N.P., Steuer R.E., Linear and Nonlinear Applications of the Tchebycheff
Metric to the Multi Criteria Water Allocation Problem, Environmental Systems Anal-
ysis and Management, Edited by S. Rinaldi, North-Holland Publishing Company, 1982,
pp. 363-376.
Yang J.-B., Sen P., Preference Modelling by Estimating Local Utility Functions for Mul-
tiobjective Optimization, European Journal of Operational Research 95, No.1 (1996),
115-138.
Yang J.-B., Chen C., Zhang Z.-J., The Interactive Step Trade-Off Method (ISTM) for Mul-
tiobjective Optimization, IEEE Transactions on Systems, Man, and Cybernetics 20, No.
3 (1990), 688-695.
Yano H., Sakawa M., Trade-Off Rates in the Weighted Tchebycheff Norm Method, Large
Scale Systems 13, No.2 (1987), 167-177.
Yano H., Sakawa M., A Unified Approach for Characterizing Pareto Optimal Solutions of
Multiobjectille Optimization Problems: The Hyperplane Method, European Journal of
Operational Research 39, No 1 (1989), 61-70.
Yoon K.P., Hwang C.-L., Multiple Attribute Decision Making: An Introduction, Sage Publi-
cations, Inc., Thousand Oaks, 1995.
Yoshikawa K., Haruna M., Kobayashi K., Multi-Objective Multi-Level Systems Approach to
Planning of Public Works, Control Science and Technology for the Progress of Society,
Edited by H. Akashi, IFAC, vol. 2, Pergamon Press, 1982, pp. 1545-1551.
290 References

Youness E.A., A Direct Approach for Finding All Efficient Solutions for Multiobjective Pro-
gramming Problems, European Journal of Operational Research 81, No.2 (1995), 440-
443.
Yu P.L., A Class of Solutions for Group Decision Problems, Management Science 19, No.8
(1973), 936-946.
Yu P.L., Cone Convexity, Cone Extreme Points, and Nondominated Solutions in Decision
Problems with Multiobjectives, Journal of Optimization Theory and Applications 14, No.
3 (1974), 319-377.
Yu P.L., Multiple-Criteria Decision Making Concepts, Techniques, and Extensions, Plenum
Press, New York, 1985.
Yu P.L., Habitual Domains, Operations Research 39, No.6 (1991), 869-876.
Yu P.L., Toward Expanding and Enriching Domains of Thinking and Application, Multiple
Criteria Decision Making - Proceedings of the Tenth International Conference: Expand
and Enrich the Domains of Thinking and Application, Edited by G.H. Tzeng, H.F. Wand,
U.P. Wen, P.L. Yu, Springer-Verlag, New York, 1994, pp. 1-9.
Yu P.-L., Habitual Domains: Freeing Yourself from the Limits on Your Life, Highwater
Editions, Shawnee Mission, Kansas, 1995.
Yu P.-L., Liu L., A Foundation of Principles for Expanding Habitual Domains, Essays in De-
cision Making: A Volume in Honour of Stanley Zionts, Edited by M.H. Karwan, J. Spronk,
J. Wallenius, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 185-200.
Zadeh L., Optimality and Non-Scalar- Valued Performance Criteria, IEEE Transactions on
Automatic Control 8 (1963), 59-60.
Zangwill W.I., Nonlinear Programming: A Unified Approach, Prentice-Han, Inc., 1969.
Zeleny M., Compromise Programming, Multiple Criteria Decision Making, Edited by J.L.
Cochrane, M. Zeleny, University of South Carolina Press, Columbia, South Carolina,
1973, pp. 262-301.
Zeleny M., Linear Multiobjective Programming, Lecture Notes in Economics and Mathemat-
ical Systems 95, Springer-Verlag, Berlin, Heidelberg, 1974.
Zeleny M., The Theory of Displaced Ideal, Multiple Criteria Decision Making Kyoto 1975,
Edited by M. Zeleny, Lecture Notes in Economics and Mathematical Systems 123,
Springer-Verlag, Berlin, Heidelberg, 1976, pp. 153-206.
Zeleny M., Multiple Criteria Decision Making, McGraw-Hill, Inc., 1982.
Zeleny M., Stable Patterns from Decision-Producing Networks: New Interfaces of DSS and
MCDM, MCDM WorldScan 3, No. 2-3 (1989), 6-7.
Zeleny M., Towards the TradeoJJs-Free Optimality in MCDM, Multicriteria Analysis, Edited
by J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997, pp. 596-601.
Zhou X., Mokhtarian F.S., Zlobec S., A Simple Constraint Qualification in Convex Program-
ming, Mathematical Programming 61, No.3 (1993), 385-397.
Zionts S., Methods for Solving Management Problems Involving Multiple Objectives, Multiple
Criteria Decision Making Theory and Application, Edited by G. Fandel, T. Gal, Lecture
Notes in Economics and Mathematical Systems 177, Springer-Verlag, Berlin, Heidelberg,
1980, pp. 540-558.
Zionts S., Multiple Criteria Mathematical Programming: An Updated Overview and Several
Approaches, Multiple Criteria Decision Making and Risk Analysis Using Microcomputers,
Edited by B. Karpak, S. Zionts, Springer-Verlag, Berlin, Heidelberg, 1989, pp. 7-60.
Zionts S., The State of Multiple Criteria Decision Making: Past, Present, and Future, Multi-
ple Criteria Decision Making: Proceedings of the Ninth International Conference: Theory
and Applications in Business, Industry, and Government, Edited by A. Goicoechea, L.
Duckstein, S. Zionts, Springer-Verlag, New York, 1992, pp. 33-43.
Zionts S., Some Thoughts on MCDM: Myths and Ideas, Multicriteria Analysis, Edited by
J. Climaco, Springer-Verlag, Berlin, Heidelberg, 1997a, pp. 602-607.
 References 291

Zionts S., Decision Making: Some Experiences, Myths and Observations, Multiple Criteria
Decision Making: Proceedings of the Twelfth International Conference, Hagen (Ger-
many), Edited by G. Fandel, T. Gal, Lecture Notes in Economics and Mathematical
Systems 448, Springer-Verlag, Berlin, Heidelberg, 1997b, pp. 233--241.
Zionts S., Wallenius J., An Interactive Programming Method lor Solving the Multiple Criteria
Problem, Management Science 22, No.6 (1976), 652-663.
Zionts S., WaJlenius J., An Interactive Multiple Objective Linear Programming Method lor a
Class 01 Underlying Nonlinear Utility Functions, Management Science 29, No.5 (1983),
519-529.
Zlobec S., Two Characterizations 01 Pareto Minima in Convex Multicriteria Optimization,
ApJikace Matematiky 29, No.5 (1984), 342-349.
Zubiri J., Scalarization 01 Vector Optimization Problems via Generalized Chebyshev Norm,
Mathematical Analysis and Systems Theory V., Edited by J. Szep, P. Ta1l6s, Department
of Mathematics, Karl Marx University of Economics, Budapest, 1988, pp. 39-42.
 INDEX

A posteriori methods 63, 77 Conditional proper Pareto optimality

A priori methods 64, 115 30
Achievement function 108, 164, 168, Cone 6
180,184,186 Connectedness of (weakly) Pareto
order-approximating 108 optimal set 20
order-representing 108 Constraint function 5
Achievement problem 109 Constraint qualification 38
Achievement scalarizing function see Cottle 48, 49
Achievement function Kuhn-Tucker 38,39
Achievement scalarizing function regularity 42,90-93
approach 107-112 second-order 42
Active constraint 6 Slater 51,72,75
Ad hoc method 64 Continuously differentiable function 7
ADBASE 236 Convergence of method 134
Analyst 14 Convex combination 6
Anchoring 135, 224 Convex function 7
Archimedian goal programming see Convex hull 6
Weighted goal programming Convex multiobjective optimization
Aspiration function 196 problem 7
Aspiration level 14,62,108,121,164, Convex set 7
172,174,184,196 Cottle constraint qualification 48, 49
Augmented weighted Tchebycheff metric Criterion value see Objective value
101 Criterion vector see Objective vector
Augmented weighted Tchebycheff
problem 101, 160 Decision maker 14
Automatic trade-off 176 Decision map 240
Decision support system (DSS) 233
Bar chart 242 Decision variable space 5
Basic methods 77 Decision variable vector see Decision
Bi-reference procedure 211 vector
BIPLOT 246 Decision vector 5
Bundle methods 71,198 Decreasing function 8
DlDAS 167,235
CI,I-function 52 Differentiable function 7
CAM OS 235, 236 Direction-searching method 209
Characteristic neighbour 183 Dominating point 111
Combined goal programming 124 Domination 23
Combined procedure 169 Domination property 36
Complete efficiency 36 Duality gap 34, 84, 95
Composite programming 106
Compromise programming 67, 97 Edgeworth-Pareto optimality see
Concave function 7 Pareto optimality
294 Index

Efficiency 23, 24 GAIA 247

proper in the sense of Henig 32 GDF method 141-149,221,222,226
super 33 Generalized goal programming 121,
weak 25 188
Envelope approach 112 Generalized reachable sets method 113
E:-constraint method 85-95 Generalized second-order directional
E:-constraint problem 85 derivative 53
stable 90 Geoffrion-Dyer-Feinberg method see
€-Pareto optimality 31 GDF method
E:-proper Pareto optimality 30 Global Pareto optimality 12
E:-strongly increasing function 9 Global trade-off 27,81,104-106
Euclidean distance function 6 Goal 121
Euclidean norm 6 Goal programming 121-129
Exact trade-off 177 generalized 121,188
Existence of Pareto optimal solutions lexicographic 124
33-35 min-max 125
Exterior branching algorithm 163 weighted 122
Gradient 7
Feasible direction 6 GRADS 246
Feasible objective region 5 GUESS method 170-174,222,224
Feasible region 5
Final solution 15,22,28,29 Habitual domain 15
Flexible goal 121,188 Harmonious house 246
Fritz John necessary condition for Hessian matrix 9
optimality 46 Hierarchical optimization 120
Fritz John necessary condition for Pareto Hybrid interactive decision making
optimality 37,47 technique 210
Fritz John necessary condition for weak Hybrid method 96
Pareto optimality 38, 48 Hybrid problem 96
Function Hyperplane method 112
concave 7
continuously differentiable 7 Ideal objective vector 16,67,97,161,
convex 7 198
decreasing 8 Improvement function 72,197
differentiable 7 Increasing function 8
E:-strongly increasing 9 Indifference band 27
increasing 8 Indifference curve 27
locally Lipschitzian 9 Indifference threshold 181
monotonic 8 Indifference trade-off see Marginal rate
nondifferentiable 9 of substitution
pseudo concave 8 Indifferent objective vectors 21,181
pseudo convex 7 Inflexible goal 121,188
quasiconcave 7 Interactive goal programming method
quasiconvex 7 146
strictly convex 8 Interactive integrated approach 147
strictly increasing 8 Interactive methods 64,131,219,230
strictly quasiconvex 8 Interactive multiple goal programming
strongly decreasing 8 (IMGP) method 208
strongly increasing 8 Interactive sequential goal programming
twice continuously differentiable 9 (ISGP) 208
twice-differentiable 9 Interactive step trade-off method 212
upper semidifferentiable 10 Interactive surrogate worth trade-off
FW method 142 method see ISWT method
 Index 295

Interactive weighted Tchebycheff Lp-metric 67

procedure see Tchebycheff weighted 97
method L,,-problem 67
Interchanging roles of functions 178, weighted 97
188,205
Interior of set 6 Marginal rate of substitution 28, 83,
ISWT method 136-141 127, 141, 143, 149
Iteration counter 136 approximation of 144
Mathematical convergence 134
jth Lagrangian method 89 MATLAB 237
Just noticeable difference 144,180 MCDM see Multiple criteria decision
making
Karush-Kuhn-Tucker multiplier 38,91 Meta algorithm 252
Karush-K uhn-Tucker necessary Method of Steuer 211
condition for e-constraint problem Method of the displaced ideal 209
90 Method of the global criterion 67-71,
Karush-Kuhn-Tucker necessary 97
condition for Pareto optimality Method of weighted metrics 97-107,
39,49,51 124
Karush-K uhn-Tucker necessary Methods for generating Pareto optimal
condition for weak Pareto solutions see A posteriori methods
optimality 40,50,51 Min-max goal programming 125
Karush-Kuhn-Tucker sufficient condition Min-max goal programming problem
for optimality 40 125
Karush-Kuhn-Tucker sufficient condition Modified weighted Tchebycheff metric
for Pareto optimality 40,50 101
Karush-Kuhn-Tucker sufficient condition Modified weighted Tchebycheff problem
for weak Pareto optimality 41,51
101
K uhn-Tucker constraint qualification
MOLP problem see Linear multiobjec-
38,39
tive optimization problem
Kuhn-Tucker necessary condition for
proper Pareto optimality 44 Monotonic function 8
Kuhn-Tucker sufficient condition for MONP-16 178,235
proper Pareto optimality 44 Motzkin's theorem 39
MPB method 71-76,197,236
Lagrange function 90, 100 MPBNGC 76, 236
Lagrange multiplier 90 Multiattribute decision analysis xiii
Learning-oriented method 132 Multiobjective optimization Xlll
Less is preferred to more 22,23, 136 Multiobjective optimization problem 5
Level set 8 convex 7
Lexicographic goal programming 124 linear 6
Lexicographic ordering 118-121,124, nondifferentiable 10
149,186 nonlinear 6
Lexicographic problem 119 Multiobjective proximal bundle method
Lexicographic weighted Tchebycheff see MPB method
problem 155 Multiple criteria decision making xiii
Light beam search 179-184, 235, 252 Multiple criteria optimization see
Linear independence 6 Multiple criteria decision making
Linear multiobjective optimization
problem 6 Naive method see GUESS method
Lao-metric see Tchebycheff metric Nadir objective vector 16,162,171
Loa-problem see Tchebycheff problem Necessary condition for e-constraint
Local Pareto optimality 12 problem 90
Locally Lipschitzian function 9 Necessary condition for optimality 46
296 Index

Necessary condition for Pareto Pareto optimal set 12

optimality Pareto optimality 11
Fritz John 37,47 €- 31
Karush-Kuhn-Thcker 39,49,51 global 12
second-order 42, 53 local 12
Necessary condition for proper Pareto proper £- 30
optimality 44 proper in the sense of Geoffrion 29
Necessary condition for weak Pareto proper in the sense of Kuhn and
optimality Thcker 31
Fritz John 38,48 weak 19
Karush-Kuhn-Thcker 40,50,51 Pareto optimality tests 33-35
Negative deviation see Underachieve- Pareto race 188
ment Partial derivative 7
NIMBUS 235 Partial trade-off 26
NIMBUS method 195-208 Partial trade-off rate 26,27,93,137,
No-preference methods 63,67 150
NOA 237 Payoff table 16
Non ad hoc method 64 Penalty scalarizing function 111
Nondifferentiable function see Locally Perturbation function 89
Lipschitzian function Petal diagram 244
Nondifferentiable interactive multiob- Pointed cone 6, 23
jective bundle-based optimization Positive definite matrix 9
system see NIMBUS method Positive deviation see Overachieve-
Nondifferentiable multiobjective ment
optimization problem 10 Preemptive goal programming see
Nondominance see Pareto optimality Lexicographic goal programming
Noninferior set estimation (NISE) Preemptive reference point method
method 112 169
Noninferiority see Pareto optimality Preference threshold 181
Nonlinear multiobjective optimization Proper efficiency in the sense of Henig
problem 6 32
Normal boundary intersection (NBI) Proper equality method 85,140
method 113 Proper Pareto optimality
Normalizing objective functions 18 conditional 30
£- 30
Objective function 5 Geoffrion 29
Objective function value see Objective Kuhn and Thcker 31
value Proper Pareto optimality in the sense of
Objective function vector see Objective Geoffrion 29
vector Proper Pareto optimality in the sense of
Objective space 5 Kuhn and Thcker 31
Objective value 5 Proxy approach 147,151
Objective vector 5 Pseudo concave function 8
ideal 16,67,97,161,198 Pseudo convex function 7
nadir 16,162, 171
utopian 16,98,101,154,175 Quasiconcave function 7
Open ball 6 Quasiconvex function 7
Ordering cone 23
Outranking relation 180,181 Radar chart see Spider-web chart
Outranking trials method 167 Range equalization factor 18
Overachievement 122 RD method 190-193
RD problem 191
Parameter space investigation (PSI) Redundancy 36
method 213 Reference ball 167
 Index 297

Reference direction 184, 190 Strictly convex function 8

Reference direction approach 184-189 Strictly increasing function 8
Reference direction method see RD Strictly quasiconvex function 8
method Strongly decreasing function 8
Reference function 29 Strongly decreasing value function see
Reference point 14,164,171,180, 184 Less is preferred to more
Reference point method 164-170,235 Strongly increasing function 8
Regular point 42, 90-93 Subdifferential 10
Reservation level 168 Subgradient 10
Subgradient GDF method 148
Satisficing decision making 22,165, Substationary point 48, 75
174,180 Sufficient condition for g-constraint
Satisficing solution 22, 62, 175 problem 92
Satisficing trade-off method see STOM Sufficient condition for optimality 40
Scalar product 6 Sufficient condition for Pareto optimality
Scalar subproblem 198 40,50
Scalar version (of NIMBUS) 197 second-order 43, 53
Scalarization 3, 62 Sufficient condition for proper Pareto
Scalarizing function 62 optimality 44
Scatterplot matrix 245 Sufficient condition for weak Pareto
Searching-oriented method 132 optimality 41,51
Second-order constraint qualification Super efficiency 33
42 Surrogate worth 139
Second-order directional derivative Surrogate worth trade-off (SWT)
generalized 53 method 136
Second-order necessary condition for
Pareto optimality 42, 53 Tchebycheff method 154-161,222,225
Second-order partial derivative 9 Tchebycheff metric 68
Second-order sufficient condition for augmented weighted 101
g-constraint problem 92 modified weighted 101
Second-order sufficient condition for weighted 97
Pareto optimality 43, 53 Tchebycheff problem 68
Sensitivity 56 augmented weighted 101,160
Sensitivity analysis 56 lexicographic weighted 155
Sequential information generator modified weighted 101
for multiple objective problems weighted 97,155, 161, 162
(SIGMOP) 209 Threshold
Sequential multiobjective problem indifference 181
solving (SEMOPS) technique 208 preference 181
Sequential proxy optimization technique veto 181
see SPOT TOMMIX 252
SIMOLP method 223 Total trade-off 26
Slater constraint qualification 51,72, Total trade-off rate 26
75 Trade-off 26
Solution process 6 global 27,81,104--106
SOMMIX 252 partial 26
Spider-web chart 243 total 26
SPOT 149-154 Trade-off rate
Stability 56 partial 26,27,93, 137, 150
Stable c-constraint problem 90 total 26
Star coordinate system 243 Trial-and-error procedure 221,222
STEM 161-164,222 'lUcker's theorem 43
Step method see STEM Twice continuously differentiable
STOM 174-179,235 function 9
298 Index

Twice-differentiable function 9 Weak efficiency 25

Weak Pareto optimality 19
Underachievement 122 Weighted goal programming 122
Unified algorithm 252 Weighted goal programming problem
Upper semidifferentiable function 10 122
User 65 Weighted Lp-metric 97
Utility function 21 Weighted Lp-problem 97
Utopian objective vector 16,98,101, Weighted max-min problem 171
154,175 Weighted Tchebycheff metric 97,155
augmented 101
Value function 21,27,64,115,132,136,
modified 101
141,149,186,220,225
Value function method 115-118 Weighted Tchebycheff problem 97,155,
Value function problem 115 161, 162
Value path 240 augmented 101,160
Vector optimization 61 lexicographic 155
Vector subproblem 197 modified 101
Vector version (of NIMBUS) 197 Weighting method 78-85
Veto threshold 181 Weighting problem 78
VIC 188,236 WWW-NIMBUS 206,235
Visual interactive approach see
Reference direction approach Zionts-Wallenius (ZW) method 212