(Applied Optimization 44) Evangelos Triantaphyllou (Auth.) - Multi-Criteria Decision Making Methods - A Comparative Study-Springer US (2000)

Multi-Criteria Decision Making Methods:
A Comparative Study
Applied Optimization
Volume 44
Series Editors:
Panos M. Pardalos
University of Florida, US.A.
Donald Hearn
University of Florida, US.A.
The titles published in this series are listed at the end of this volume.
Multi-Criteria Decision
Making Methods:
A Comparative Study
by
Evangelos Triantaphyllou
Department of Industrial and Manufacturing Systems Engineering,
College of Engineering,
Louisiana State University,
Baton Rouge, Louisiana, US.A.
SPRINGER-SCIENCE+BUSINESS MEDIA B.V.

A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-1-4419-4838-0 ISBN 978-1-4757-3157-6 (eBook)

DOI 10.1007/978-1-4757-3157-6
Printed an acid-free paper
AU Rights Reserved
© 2000 Springer Science+Business Media Dordrecht
OriginaIly published by Kluwer Academic Publishers in 2000
Softcover reprint of the hardcover 1st edition 2000
No part ot the matenal protecteo by thlS copynght not1ce may be reproouceO or
utilized in any fonn or by any means, electronic or mechanical,
including photocopying, recording or by any infonnation storage and
retrieval system, without written pennission from the copyright owner
This book is gratefully dedicated to all my students;
of the past, the present, and the future.
TABLE OF CONTENTS
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. xiii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxix
1 Introduction to Multi-Criteria Decision Making . . . . . . . •• 1

1.1 Multi-Criteria Decision Making:
A General Overview . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Classification of MCDM Methods . . . . . . . . . . . . " 3
2 Multi-Criteria Decision Making Methods .••.......... 5

2.1 Background Information . . . . . . . . . . . . . . . . . . . 5
2.2 Description of Some MCDM Methods . . . . . . . . . . . 5
2.2.1 The WSM Method . . . . . . . . . . . . . . . . . . 6
2.2.2 The WPM Method . . . . . . . . . . . . . . . . . . 8
2.2.3 The AHP Method . . . . . . . . . . . . . . . . . .. 9
2.2.4 The Revised AHP Method . . . . . . . . . . . . . . 11
2.2.5 The ELECTRE Method . . . . . . . . . . . . . . . 13
2.2.6 The TOPSIS Method . . . . . . . . . . . . . . . . . 18
3 Quantification of Qualitative Data for

MCDM Problems . • . . . . . . . . . . . . . . . . . . • . . . . . . . 23
3. 1 Background Information . . . . . . . . . . . . . . . . . . . . 23
3.2 Scales for Quantifying Pairwise Comparisons . . . . . . . 25
3.2.1 Scales Defined on the Interval [9, 1/9] ...... 26
3.2.2 Exponential Scales . . . . . . . . . . . . . . . . . . 28
3.2.3 Some Examples of the Use of
Exponential Scales . . . . . . . . . . . . . . . . . . 29
3.3 Evaluating Different Scales . . . . . . . . . . . . . . . . . . 32
3.3.1 The Concepts of the RCP and CDP Matrices .. 32
3.3.2 On The Consistency of CDP Matrices ...... 35
3.3.3 Two Evaluative Criteria . . . . . . . . . . . . . . . 43
3.4 A Simulation Evaluation of Different Scales . . . . . . . . 44
3.5 Analysis of the Computational Results . . . . . . . . . . . 50
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
viii MCDM Methods: A Comparative Study, by E. Triantaphyllou
4 Deriving Relative Weights from Ratio Comparisons ...... 57

4.1 Background Information . . . . . . . . . . . . . . . . . . . . 57
4.2 The Eigenvalue Approach . . . . . . . . . . . . . . . . . . . 58
4.3 Some Optimization Approaches . . . . . . . . . . . . . . . 60
4.4 Considering The Human Rationality Factor . . . . . . . . 61
4.5 First Extensive Numerical Example . . . . . . . . . . . . . 65
4.6 Second Extensive Numerical Example . . . . . . . . . . . 66
4.7 Average Error per Comparison for Sets
of Different Size . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5 Deriving Relative Weights from Difference Comparisons . .. 73

5.2 Pairwise Comparisons of Relative Similarity ....... 76
5.2.1 Quantifying Pairwise Comparisons
of Relative Similarity . . . . . . . . . . . . . . . . . 76
5.2.2 Processing Pairwise Comparisons
of Relative Similarity . . . . . . . . . . . . . . . . . 77
5.2.3 An Extensive Numerical Example . . . . . . . . . 79
5.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6 A Decomposition Approach for Evaluating Relative

Weights Derived from Comparisons . . . . . . . . . . . . . . . . 87
6.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . 88
6.3 Two Solution Approaches . . . . . . . . . . . . . . . . . . . 91
6.3.1 A Simple Approach . . . . . . . . . . . . . . . . . . 91
6.3.2 A Linear Programming Approach . . . . . . . . . 92
6.4 An Extensive Numerical Example . . . . . . . . . . . . . . 95
6.5 Some Computational Experiments . . . . . . . . . . . . . . 97
6.6 Analysis of the Computational Results . . . . . . . . . . 100
6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .. 112
7 Reduction of Pairwise Comparisons Via a

Duality Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.1 Background Information . . . . . . . . . . . . . . . . . .. 115
7.2 A Duality Approach for Eliciting Comparisons . . . . . 116
7.3 An Extensive Numerical Example . . . . . . . . . . . . . 120
7.3.1 Applying the Primal Approach . . . . . . . . . . 121
Table of Contents ix
7.3.2 Applying the Dual Approach . . . . . . . . . . . 122

7.4 Some Numerical Results for Problems of
Different Sizes . . . . . . . . . . . . . . . . . . . . . . . .. 124
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .. 128
8 A Sensitivity Analysis Approach for MCDM Methods •••• 131

8.2 Description of the Two Major Sensitivity
Analysis Problems . . . . . . . . . . . . . . . . . . . . . . . 133
8.3. Problem 1: Determining the Most Critical
Criterion . . . . . . . . . . . . . . . . . . . . . . . . 135
8.3.1 Definitions and Terminology . . . . . . . . . . . 135
8.3.2 Some Theoretical Results in Determining
the Most Critical Criterion . . . . . . . . . . . . 137
8.3.2.1 Case (i): Using the WSM or the
AHP Method . . . . . . . . . . . . . . . . 137
8.3.2.2 An Extensive Numerical Example
for the WSM Case . . . . . . . . . . . . 138
8.3.2.3 Case (ii): Using the WPM Method . . 142
for the WPM Case . . . . . . . . . . . . 143
8.3.3 Some Computational Experiments . . . . . . . . 145
8.4 Problem 2: Determining the Most Critical aij
Measure of Performance . . . . . . . . . . . . . . . . . . . 155
8.4.1 Definitions and Terminology . . . . . . . . . . . 155
8.4.2 Determining the Threshold
Values <,j,k . . . . . . . . . . . . . . . . . . . . .. 157
8.4.2.1 Case (i): When Using the WSM
or the AHP Method . . . . . . . . . . . . 157
When the WSM or the
AHP Method is Used . . . . . . . . . .. 158
8.4.2.3 Case (ii): When Using the WPM
Method . . . . . . . . . . . . . . . . . . . 161
When the WPM Method is Used . . .. 161
8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .. 165
x MCDM Methods: A Comparative Study, by E. Triantaphyllou
Appendix to Chapter 8 . . . . . . . • . . . . . . . . . . . . . . .. 167

8.6 Calculation of the 01 12 Quantity When
the AHP or the WSM Method is Used ..... 167
8.7 Calculation of the 01•1•2 Quantity When
the WPM Method is Used . . . . . . . . . . . . . 169
8.8 Calculation of the 7 3.4.5 Quantity When
the WSM Method is Used . . . . . . . . . . . .. 170
8.9 Calculation of the 7 3•4,5 Quantity When
the AHP Method is Used . . . . . . . . . . . . . 171
8.10 Calculation of the 7 3,4,5 Quantity When
the WPM Method is Used . . . . . . . . . . . .. 174
9 Evaluation of Methods for Processing a

Decision Matrix and Some Cases
of Ranking Abnormalities . . . . . . . . . . . . . . . . . . . . . 177
9.1 Background Information . . . . . . . . . . . . . . . . . .. 177
9.2 Two Evaluative Criteria . . . . . . . . . . . . . . . . . .. 177
9.3 Testing the Methods by Using the First
Evaluative Criterion . . . . . . . . . . . . . . . . . . . . .. 179
9.4 Testing the Methods by Using the Second
Evaluative Criterion. . . . . . . . . . . . . . . . . . . . .. 186
9.5 Analysis of the Computational Results . . . . . . . . . . 192
9.6 Evaluating the TOPSIS Method . . . . . . . . . . . . . . 194
9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .. 197
10 A Computational Evaluation of the Original

and the Revised AHP . . . . . . . . . . . . . . . . . . . . . . . . . 201
10.2 An Extensive Numerical Example . . . . . . . . . . . . . 202
10.3 Some Computational Experiments . . . . . . . . . . . . . 206
10.4 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . .. 212
11 More Cases of Ranking Abnormalities When Some

MCDM Methods Are Used . . . . . . . . . . . . . . . . . . . . . 213
11. 1 Background Information . . . . . . . . . . . . . . . . . . . 213
11.2 Ranking Irregularities When Alternatives Are
Compared Two at a Time . . . . . . . . . . . . . . . . . . 215
11.3 Ranking Irregularities When Alternatives Are
Compared Two at a Time and Also as a Group . . . . . 220
Table of Contents xi
11.4 Some Computational Results . . . . . . . . . . . . . . . . 223

11.5 A Multiplicative Version of the AHP . . . . . . . . . . . 228
11.6 Results from Two Real Life Case Studies . . . . . . . . 230
11.6.1 Comparative Ranking Analysis of
the "Bridge Evaluation" Problem . . . . . . . . 230
11.6.2 Comparative Ranking Analysis of
the "Site Selection" Problem . . . . . . . . . . . 232
11.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
12 Fuzzy Sets and Their Operations ..•.......••••... 235

12.2 Fuzzy Operations . . . . . . . . . . . . . . . . . . . . . . . 236
12.3 Ranking of Fuzzy Numbers . . . . . . . . . . . . . . . .. 238
13 Fuzzy Multi-Criteria Decision Making • • . . . . . . . . • . . . 241

13.2 The Fuzzy WSM Method . . . . . . . . . . . . . . . . . . 242
13.3 The Fuzzy WPM Method . . . . . . . . . . . . . . . . . . 244
13.4 The Fuzzy AHP Method . . . . . . . . . . . . . . . . . . . 245
13.5 The Fuzzy Revised AHP Method . . . . . . . . . . . . . 247
13.6 The Fuzzy TOPSIS Method . . . . . . . . . . . . . . . . . 248
13.7 Two Fuzzy Evaluative Criteria for
Fuzzy MCDM Methods . . . . . . . . . . . . . . . . . . . 250
13.7.1 Testing the Methods by Using the First
Fuzzy Evaluative Criterion . . . . . . . . . . . . 251
13.7.2 Testing the Methods by Using the Second
Fuzzy Evaluative Criterion . . . . . . . . . . . . 255
13.8 Computational Experiments . . . . . . . . . . . . . . . . . 257
13.8.1 Description of the Computational
Results . . . . . . . . . . . . . . . . . . . . . . . .. 258
13.8.2 Analysis of the Computational
Results . . . . . . . . . . . . . . . . . . . . . . . .. 261
13.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
14 Conclusions and Discussion for Future Research .....•• 263

14.1 The Study of MCDM Methods:
Future Trends . . . . . . . . . . . . . . . . . . . . . . . . . 263
14.2 Lessons Learned . . . . . . . . . . . . . . . . . . . . . . .. 263
xii MCDM Methods: A Comparative Study, by E. Triantaphyllou
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 275
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289

LIST OF FIGURES
1 Introduction to Multi-Criteria Decision Making • • .. 1

Figure 1-1: A Typical Decision Matrix . . . . . . . . . . . . . . . . . . 3
Figure 1-2: A Taxonomy of MCDM methods (according to
Chen and Hwang [1991]) . . . . . . . . . . . . . . . . . . . 4
2 Multi-Criteria Decision Making Methods ••••••.• 5

MCDM Problems ••.•••••...••...•.•••.. 23
Figure 3-1: Actual Comparison Values . . . . . . . . . . . . . . . . . . 37
Figure 3-2: Maximum, Average, and Minimum CI Values of
Random CDP Matrices When the Original
Saaty Scale is used . . . . . . . . . . . . . . . . . . . . . . . 42
Figure 3-3: Inversion Rates for Different Scales and Size
of Set (Class 1 Scales) . . . . . . . . . . . . . . . . . . . . . 46
Figure 3-4: Indiscrimination Rates for Different Scales
and Size of Set (Class 1 Scales) . . . . . . . . . . . . . . . 47
Figure 3-5: Inversion Rates for Different Scales and Size
of Set (Class 2 Scales) . . . . . . . . . . . . . . . . . . . . . 48
Figure 3-6: Indiscrimination Rates for Different Scales
and Size of Set (Class 2 Scales) . . . . . . . . . . . . . . . 49
Figure 3-7: The Best Scales . . . . . . . . . . . . . . . . . . . . . . . . . 51
Figure 3-8: The Worst Scales . . . . . . . . . . . . . . . . . . . . . . . . 52
4 Deriving Relative Weights from Ratio Comparisons . 57

Figure 4-1: Average Residual and CI versus Order of Set
When the Human Rationality Assumption is Used
(the Results Correspond to 100 Random Observations) . 70
Figure 4-2: Average Residual and CI versus Order of Set
When the Eigenvalue Method is Used
(the Results Correspond to 100 Random Observations) . 71
5 Deriving Relative Weights from Difference

Comparisons . . . . • • • . • . . • • • • • • . • . . . • . • . . 73
xiv MCDM Methods: A Comparative Study, by E. Triantaphyllou

Weights Derived from Comparisons ••••••••••• 87
Figure 6-1: Partitioning of the n(n-1)/2 Pairwise
Comparisons ...... . . . . . . . . . . . ... . . . . . . . . 90
Figure 6-2: Error Rates Under the LP Approach for Sets
of Different Size as a Function of the
Available Comparisons. . . . . . . . . . . . . . . . . . .. 106
Figure 6-3: Error Rates Under the Non-LP Approach for Sets
Available Comparisons. . . . . . . . . . . . . . . . . . .. 107
Figure 6-4: Error Rates Under the LP Approach for Sets
Common Comparisons . . . . . . . . . . . . . . . . . . ., 108
Figure 6-5: Error Rates Under the Non-LP Approach for Sets
Common Comparisons . . . . . . . . . . . . . . . . . . .. 109
Figure 6-6: Error Rates for the two Approaches as a
Function of the Available Comparisons. . . . . . . . .. 110
Figure 6-7: Error Rates for the two Approaches as a
Function of the Common Comparisons . . . . . . . . .. 111

Duality Approach •....•................ 115
Figure 7-1: Total Number of Comparisons and Reduction
Achieved When the Dual Approach is Used.
The Number of Criteria n = 5 . . . . . . . . . . . . . . 125
Figure 7-4: Total number of Comparisons and Reduction
Figure 7-5: Net Reduction on the Number of
Comparisons When the Dual Approach is used.
Results for Problems of Various Sizes . . . . . . . . . . 127
Figure 7-6: Percent (%) Reduction on the Number of
Comparisons When the Dual Approach is used.
Results for Problems of Various Sizes . . . . . . . . . . 127
List of Figures xv
8 A Sensitivity Analysis Approach

for MCDM Methods . . . . . . . . . . . . . . . . . . . . 131
Figure 8-1: Frequency of the time that the PT Critical
Criterion is the Criterion with
the Highest Weight . . . . . . . . . . . . . . . . . . . . .. 149
Figure 8-2: Frequency of the time that the PT Critical
the Lowest Weight . . . . . . . . . . . . . . . . . . . . . . 149
Figure 8-3: Frequency of the time that the PA Critical
Figure 8-4: Frequency of the time that the PA Critical
Figure 8-5: Frequency of the time that the AT Critical
Figure 8-6: Frequency of the time that the AT Critical
Figure 8-7: Frequency of the time that the AA Critical
Figure 8-8: Frequency of the time that the AA Critical
the Lowest Weight . . . . . . . . . . . . . . . . . 152
Figure 8-9: Frequency of the time that the AT and PT
Definitions point to the Same Criterion. . . . . . . . .. 153
Figure 8-10: Frequency of the time that the AA and PA
Definitions point to the Same Criterion. . . . . . . . .. 153
Figure 8-11: Frequency of the time that the AT, PT, AA, and PA
Definitions point to the Same Criterion
Under the WSM Method . . . . . . . . . . . . . . . . . . 154
Figure 8-12: Rate that the AT Criterion is the one
with the Lowest Weight for Different Size
Problems Under the WPM Method . . . . . . . . . . . . 154

of Ranking Abnormalities . . . . . . . . . . . . . . . . . 177
Figure 9-1: Contradiction Rate (%) Between the
XVI MCDM Methods: A Comparative Study, by E. Triantaphyllou
WSM and the AHP . . . . . . . . . . . . . . . . . . . . .. 184

WSM and the Revised AHP . . . . . . . . . . . . . . . .. 185
WSM and the WPM . . . . . . . . . . . . . . . . . . . . . 185
Figure 9-4: Rate of Change (%) of the Indication of the
Optimum Alternative When a Non-Optimum
Alternative is Replaced by a Worse one.
The AHP Case. . . . . . . . . . . . . . . . . . . . . . . .. 191
Figure 9-5: Rate of Change (%) of the indication of the
The Revised AHP Case . . . . . . . . . . . . . . . . . . . 191
Figure 9-6: Contradiction Rate (%) Between the WSM
and TOPSIS Method . . . . . . . . . . . . . . . . . . . . . 196
Figure 9-7: Rate of Change (%) of the Indication of the
Optimum Alternative When aNon-Optimum
The TOPSIS Case . . . . . . . . . . . . . . . . . . . . . .. 196
Figure 9-8: Indication of the Best MCDM Method According
to Different MCDM Methods. . . . . . . . . . . . . . .. 198

and the Revised AHP . . . . . . . . . . . . . . . . . . .. 201
Figure 10-1: The Failure Rates are Based on 1,000 Randomly
Generated Problems. The AHP Case . . . . . . . . . . . 210
Figure 10-2: The Failure Rates are Based on 1,000 Randomly
Generated Problems. The Revised AHP Case . . . . . 211
11 More Cases of Ranking Abnormalities When Some

MCDM Methods Are Used • . . . . . . . . . . . . . . . 213
Figure 11-1: Contradiction Rates on the Indication of the
Best Alternative When Alternatives are
Considered Together and in Pairs.
The Original AHP Case . . . . . . . . . . . . . . . . . . . 225
Figure 11-2: Contradiction Rates on the Indication of the
Best Alternative When Alternatives are
The Ideal Mode (Revised) AHP Case . . . . . . . . . . . 225
Figure 11-3: Contradiction Rates on the Indication of
List of Figures xvii
Any Alternative When Alternatives are

Figure 11-4: Contradiction Rates on the Indication of
The Ideal Mode (Revised) AHP Case . . . . . . . . . . . 226
Figure 11-5: Contradiction Rates on the indication of
Considered in Pairs.
Figure 11-6: Contradiction Rates on the indication of
Considered in Pairs.
The Ideal Mode AHP Case . . . . . . . . . . . . . . . . . 227
12 Fuzzy Sets and Their Operations ............ 235

Figure 12-1: Membership Functions for the Two Fuzzy
Alternatives AJ and A2 .................... 239
13 Fuzzy Multi-Criteria Decision Making . . . . . . . . . 241

Figure 13-1: Membership Functions of the Fuzzy Alternatives
AI' A2, and A3 of Example 13-1 According
to the Fuzzy WSM Method . . . . . . . . . . . . . . . . . 243
Figure 13-2: Membership Functions of the Fuzzy Alternatives
AJ , A27 and A3 of Example 13-2 According
to the Fuzzy WPM Method . . . . . . . . . . . . . . . . . 244
Figure 13-3: Contradiction Rate R11 When the Number of
Fuzzy Alternatives is Equal to 3 . . . . . . . . . . . . . . 259
Fuzzy Alternatives is Equal to 21 . . . . . . . . . . . . . 259
Fuzzy Alternatives is Equal to 21 . . . . . . . . . . . . . 260
14 Conclusions and Discussion for Future Research .. 263

LIST OF TABLES
1 Introduction to Multi-Criteria Decision Making . . .. 1
2 Multi-Criteria Decision Making Methods .•.••••. 5

MCDM Problems ••••.•.•.•.•••••.••.... 23
Table 3-1: Scale of Relative Importances
(according to Saaty[1980]) . . . . . . . . . . . . . . . . . . . 27
Table 3-2: Scale of Relative Importances
(According to Lootsma[1988]) . . . . . . . . . . . . . . . . 28
Table 3-3: Two Exponential Scales . . . . . . . . . . . . . . . . . . . . 29
4 Deriving Relative Weights from Ratio

Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Table 4-1: RCI Values of Sets of Different Order n . . . . . . . . . . 59
Table 4-2: Data for the Second Extensive Numerical Example ... 66
Table 4-3: Comparison of the Weight Values for
the Data in Table 4-2 . . . . . . . . . . . . . . . . . . . . . . 67
Table 4-4: Average Residual and CI Versus Order of Set and
CR When the Human Rationality Assumption (HR)
and the Eigenvalue Method (EM) is used.
Results Correspond to 100 Random Observations .... 69
5 Deriving Relative Weights from Difference

Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Table 5-1: Proposed Similarity Scale . . . . . . . . . . . . . . . . . . . 77

Weights Derived from Comparisons ..••....... 87
Table 6-1a: Computational Results, Part A . . . . . . . . . . . . . .. 101
Table 6-1b: Computational Results, Part B . . . . . . . . . . . . . .. 102
Table 6-1c: Computational Results, Part C . . . . . . . . . . . . . .. 103
Table 6-1d: Computational Results, Part D . . . . . . . . . . . . . .. 104
xx MCDM Methods: A Comparative Study, by Eo Triantaphyllou

Duality Approach . . . . . . . . . . . . . . . • • . . . . . 115
8 A Sensitivity Analysis Approach

for MCDM Methods . . . . . . . . . . . . . . . . . . . . 131
Table 8-1: Decision Matrix for the Numerical Example
on the WSM 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 139
Table 8-2: Current Final Preferences 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 139
Table 8-3: All Possible 0k.i,j Values (Absolute Change
in Criteria Weights) 0 0 0 0 0 0 • 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 140
Table 8-4: All Possible Olk.i,j Values (Percent Change
in Criteria Weights) 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 141
Table 8-5: Decision Matrix for the Numerical Example
on the WPM 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 143
Table 8-6: Current Ranking 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 144
Table 8-7: All Possible K Values for the WPM Example 0 0 0 0 0 0 145
Table 8-8: Decision Matrix and Initial Preferences for
the Example 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 158
Table 8-9: Threshold Values 7 /i•j •k (%) in Relative
Terms for the WSM/AHP Example 0 0 0 0 0 0 0 0 0 0 0 0 159
Table 8-10: Criticality Degrees j),,iij (%) for each aij
Performance Measure 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 160
Table 8-11: Sensitivity Coefficients sens(aij) for each aij
Performance Measure 0 0 0 • • • • • • • • • • • • • • • •• 160
Table 8-12: Decision Matrix for Numerical Example . . . . . . . .. 162
Table 8-13: Initial Ranking . 0 0 • • • • • • • 0 0 • 0 0 • • • 0 • • • • •• 162
Table 8-14: Threshold Values 7/ioj.k (%) in Relative
Terms for the WPM Example o. o. 0 0 0 •• 0 •• 0 •• 163
Table 8-15: Criticality Degrees b.. lij (in %) for each aij
Measure of Performance . . . . 0 • • • 0 • • • 0 0 • • • •• 164
Table 8-16: Sensitivity Coefficients sens(aij) for each aij
Measure of Performance . . . . . . . . . . . . . . o. 0 • • 164

of Ranking Abnormalities . . . . . . . . . . . . . . . . . 177
Table 9-1: Contradiction Rate (%) Between the
WSM and the AHP . 0 • • • • • • • • 181 0 • • • • 0 •• 0 0 • 0

WSM and the Revised AHP . . . . . . . . . . . . . 182 0 • • 0
List of Tables xxi

WSM and the WPM ..................... 183
Table 9-4: Rate of Change (%) of the Indication of the
Alternative is Replaced by a Worse One.
The AHP Case . . . . . . . . . . . . . . . . . . . . . . . . . 188
The Case of the Revised AHP ............... 188
Table 9-6: Summary of the Computational Results . . . . . . . . . . 190
Table 9-7: Contradiction Rate (%) Between the WSM and
the TOPSIS Method . . . . . . . . . . . . . . . . . . . . . . 194
The TOPSIS Case . . . . . . . . . . . . . . . . . . . . . . . 195

and the Revised AHP . . . . . . . . . . . . . . . . . . . . 201
Table 10-1: The Failure Rates are Based on 1,000 Randomly
Generated Problems. The AHP Case . . . . . . . . . . . 208
Table 10-2: The Failure Rates are Based on 1,000 Randomly
Generated Problems. The Revised AHP Case ..... 209
11 More Ranking Abnormalities When Some

MCDM Methods Are Used . . . . . . . . . . . . . . . . 213
Table 11-1: Priorities and Rankings of the Alternatives in the
"Bridge Evaluation" Case Study [Saaty, 1994] ..... 231
12 Fuzzy Sets and Their Operations . . . • . . . . . . . . 235
13 Fuzzy Multi-Criteria Decision Making . . . . . . . . . 241
14 Conclusions and Discussion for Future Research .. 263

FOREWORD
Multi-Criteria Decision Making (MCDM) has been one of the fastest

growing problem areas during at least the last two decades. In business,
decision making has changed over the last decades. From a single person (the
Boss!) and a single criterion (profit), decision environments have developed
increasingly to become multi-person and multi-criteria situations. The
awareness of this development is growing in practice. In theory many methods
have been proposed and developed since the sixties to solve this problem in
numerous ways.
Two main theoretical streams can be distinguished. First, multi-
objective decision making models which assume continuous solution spaces
(and therefore are based on continuous mathematics), try to determine optimal
compromise solutions and generally assume, that the problem to be solved can
be modeled as a mathematical programming model. This is primarily the
realm of theoreticians since continuous mathematics is very elegant and
powerful and readily allows for many modifications of a basic model or
method. Unfortunately mathematical programming does not solve the majority
of MCDM-problems in practice, and so these nice and powerful methods are
only of limited value for the practitioner. The second stream focuses on
problems with discrete decision spaces, i.e. with countable few decision
alternatives and basically uses approaches from discrete mathematics, which
are mathematically not as elegant as the former. This stream is often called
"Multi-Attribute Decision Making". In this book the more general term
MCDM is used. These models do not try to compute an optimal solution, but
they try to determine via various ranking procedures either a ranking of the
relevant actions (decision alternatives) that is "optimal" with respect to several
criteria, or they try to find the "optimal" actions amongst the existing
solutions (decision alternatives). Even though this type of problem is much
more relevant and frequent in practice, there are many fewer methods
available and their quality is much harder to determine than in the continuous
case. Therefore, the question "Which is the best method for a given
problem?" has become one of the most important but also most difficult to
answer.
This is exactly where the book of Dr. Triantaphyllou has its focus and
why it is that important. Rather than suggesting another MCDM method
without any convincing justification, he concentrates on the best known and
most frequently used methods. He extensively compares them and makes the
reader aware of quite a number of "abnormalities" of some of the methods
of which users are often not conscious. He also considers very critically the
touchiest points in solving real MCDM problems, namely, quantification of
xxiv MCDM Methods: A Comparative Study, by E. Triantaphyllou
qualitative data, deriving weights from ratio and difference comparisons, and
especially sensitivity analysis of MCDM methods. This to me seems as
valuable or even more so than suggesting a new method which may solve
another variant of the MCDM problem. At the end of the book Fuzzy MCDM
methods are described and evaluated.
What makes this book so valuable and different from other MCDM
books is, that even though the analyses are very rigorous, the results are
described very clearly and are understandable even to the non-specialist. Also,
very extensive numerical studies and comparisons are presented, which are
hard to find in any other text that I know. This book, in fact, provides a
unique perspective into the core of MCDM methods and practice. The
presented theoretical and empirical analyses are complementary to each other,
thus allowing the reader to gain a deep theoretical and practical insight into
the topics covered in this book. In addition to this, the author offers at the end
of each chapter and at the end of the book suggestions for further research
and I can only hope, that his suggestions will be accepted by many scientists.
Dr. Triantaphyllou has been involved in MCDM for almost two
decades. He has become internationally known as one of the leading experts
in the field and he is, therefore, qualified as hardly anybody else to write this
book. I can only congratulate him on his achievement and hope that many
practitioners will benefit from this excellent book and that scientists will
accept his suggestions for further research as fascinating challenges.
Aachen, Germany, April 2000
Hans-Jtirgen Zimmermann
PREFACE
Probably the most perpetual intellectual challenge in science and

engineering is how to make the optimal decision in a given situation. This is
a problem as old as mankind. In some ancient civilizations people attempted
to solve complex and risky decision problems by seeking advice from priests
or the few knowledgeable individuals. In ancient Egypt it was believed that
only the kings and the upper clergy could find what is the best solution to a
given problem. In classical Greece oracles served a similar purpose.
Many centuries passed since then. Today mankind has replaced the
old methods with modern science and technology. The development of
scientific disciplines such as operations research, management science,
computer science, and statistics, in combination with the use of modern
computers, are nothing but aids in assisting people in making the best decision
for a given situation. Theories such as linear programming, dynamic
programming, hypothesis testing, inventory control, optimization of queuing
systems, and multi-criteria decision making have as a common element the
search for an optimal decision (solution).
Among the previous methods, there is one class of methods which
probably has captured the attention of most of the people for most of the time.
This is multi-criteria decision making (MCDM). That is, given a set of
alternatives and a set of decision criteria, then what is the best alternative?
This problem may come in many different forms. For instance, the
alternatives or the criteria may not be well defined, or even more commonly,
the -related data may not be well defined. In many real life cases it may even
be impossible to accurately and objectively quantify the pertinent data. Often
a decision problem can be structured as a multi-level hierarchy. Also, it is
not unusual to have a case in which all or part of the data are stochastic or
even fuzzy.
The central decision problem examined in this book is how to evaluate
and rank the performance of a finite set of alternatives in terms of a number
of decision criteria. It is assumed that the decision maker is capable of
expressing his/her opinion of the performance of each individual alternative
in terms of each one of the decision criteria. The problem then is how to
rank the alternatives when all the decision criteria are considered
simultaneously.
In the main treatment the data are assumed to be deterministic. In the
latter part of this book we also consider the case in which the data are fuzzy.
That is, this book does not consider stochastic or probabilistic data. Although
this may sound restricted, nevertheless it captures many real life situations,
for stochastic data are difficult to be obtained or individual decision makers
xxvi MCDM Methods: A Comparative Study, by E. Triantaphyllou
feel uncomfortable dealing with them.

The author of this book became actively involved with research in this
area of decision making when he was a graduate student at Penn State
University, more than seventeen years ago. What has captured his attention
since the early days was the plethora of alternative methods for solving the
same MCDM problem. In most cases the authors and supporters of these
methods have identified some weaknesses of the previous methods and then
they propose a new method claiming to be the best method. As a result,
today a decision maker has an array of methods which all claim that they can
correctly solve a given MCDM problem. The subjectivity and the tremendous
conceptual complexity involved in many MCDM problems make the problem
of comparing MCDM methods a challenging and urgent one.
This book presents the research experiences of the author gathered
during a long search in finding which is the best MCDM method. Although
the final goal of determining the best method seems to be unattainable and
utopian, some useful lessons have been learned in the process and are
presented here in a comprehensive and systematic manner.
A methodology has been developed for evaluating MCDM methods.
This methodology examines methods for estimating the pertinent data and
methods for processing these data. A number of evaluative criteria and
testing procedures have been developed for this purpose. What became clear
very soon is that there is no single method which outperforms all the other
methods in all aspects. Therefore, the need which rises is how one can
conclude which one is the best method. However, for one to answer the
problem of which is the best MCDM method, he/she will first need to use the
best MCDM method! Thus, a decision paradox is reached.
This is the main reason why a comparative approach is needed in
dealing with MCDM methods. By simply stating various MCDM theories
and methods one fails to capture the very real and practical essence of
MCDM. The present book attempts to bridge exactly this gap. Although not
every MCDM method has been considered in this book, the procedures
followed here can be easily expanded to deal with any MCDM method which
examines the problem of evaluating a discrete set of alternatives in terms of
a set of decision criteria.
This book provides a unique perspective into the core of MCDM
methods and practice. It provides many theoretical foundations for the
behavior and capabilities of various MCDM methods. This is done by
describing a number of lemmas, theorems, corollaries, and by using a
rigorous and consistent notation and terminology. It also presents a rich
collection of examples, some of which are extensive. A truly unique
characteristic of this book is that almost all theoretical developments are
accompanied by an extensive empirical analysis which often involved the
Preface xxvii
solution of hundreds of thousands or millions of simulated test MCDM

problems. The results of these empirical analyses are tabulated, graphically
depicted, and analyzed in depth. In this way, the theoretical and empirical
analyses presented in this book are complementary to each other, so the
reader can gain both a deep theoretical and practical insight of the covered
subjects. Another unique characteristic of this monograph is that at the end
of almost each chapter there is description of some possible research problems
for future research. It also presents an extensive and updated bibliography
and references of all the subjects covered. These are very valuable
characteristics for people who wish to get involved with new research in
MCDM theory and applications. Some of the findings of these comparative
analyses are so startling and counter intuitive, that are presented as decision
making paradoxes.
Therefore, this book can provide a useful insight for people who are
interested in obtaining a deep understanding of some of the most frequently
used MCDM methods. It can be used as a textbook for senior undergraduate
or graduate courses in decision making in engineering and business schools.
It can also provide a panoramic and systematic exposure to the related
methods and problems to researchers in the MCDM area. Finally, it can
become a valuable guidance for practitioners who wish to take a more
effective and critical approach to problem solving of real life multi-criteria
decision making problems.
The arrangement of the chapters follows a natural exposition of the
main subjects in MCDM theory and practice. Thus, the first two chapters
provide an outline and background information of the most popular MCDM
methods used today. These are the weighted sum model (WSM) , the
weighted product model (WPM), the analytic hierarchy process (AHP) with
some of its variants, and the ELECTRE and TOPSIS methods.
The third chapter provides an exposition of some ways for quantifying
qualitative data in MCDM problems. This includes discussions on the
elicitation of pairwise comparisons and the use of different scales for
quantifying them. Chapters four to seven describe some different approaches
for extracting relative priorities from pairwise comparisons and also of ways
for reducing the number of the required judgments.
Chapter eight is the longest one and it deals with a unified sensitivity
analysis approach for MCDM methods. Since no real life decision problem
can be considered completely analyzed without a sensitivity analysis, this is
a critical subject. As with most of the chapters, this chapter provides an in
depth theoretical and empirical analysis of some key sensitivity analysis
problems.
Chapters nine to eleven deal with the comparison of different MCDM
methods and procedures. Chapter nine presents a comparison of different
xxviii MCDM Methods: A Comparative Study. by E. Triantaphyllou
ways for processing a decision matrix. Chapter ten presents a computational

study of the AHP and the Revised AHP. Chapter eleven presents some new
cases of ranking irregularities when the AHP and some of its additive variants
are used. One can claim that these new cases of ranking irregularities are
strongly counter intuitive. They have been analyzed both theoretically and
empirically.
Chapters twelve and thirteen present some fundamental concepts of
fuzzy decision making. As always, the treatments here are accompanied with
extensive comparative empirical analyses. Finally, some conclusions and
possible directions for future research are discussed in the last chapter.
ACKNOWLEDGMENTS
The research and the writing of this book would never had been
accomplished without the decisive help and inspiration from a number of
people to which the author is deeply indebted. His most special thanks go to
his first M.S. Advisor and Mentor, Professor Stuart H. Mann, currently the
Dean of the W.F. Harrah College of Hotel Administration at the University
of Nevada. The author would also like to thank his other M.S. Advisor
Professor Panos M. Pardalos currently at the University of Florida and his
Ph.D. Advisor Professor Allen L. Soyster, currently the Dean of Engineering
at the Northeastern University. He would also like to thank Professors F.A.
Lootsma, S. Konz, J. Elzinga, T.L. Saaty, L.G. Vargas, E.H. Forman, Dr.
P. Murphy; the CEO of InfoHarvest, Inc., and the late Drs. I.H. LaValle,
Dr. H.K. Eldin, and C.L. Hwang. The author is also highly appreciative for
the excellent comments made by Professor H.-J. Zimmermann.
Many thanks go to his colleagues at LSU; especially to Dr. E.
McLaughlin; Dean Emeritus of the College of Engineering at LSU and to Dr.
L.W. Jelinski; Vice-Chancellor for Graduate Studies and Research at LSU for
her support and inspiration. The author is also very indebted to his
distinguished colleague Dr. Tryfon T. Charalampopoulos at LSU. Many
special thanks are also given to the Editor Mr. John Martindale at Kluwer
Academic Publishers for his encouragement and incredible patience and to his
graduate student Vetle I. Torvik for his always thoughtful comments. The
author would also like to recognize here the significant contributions to his
published papers made by the numerous and anonymous referees and editors
of the journals in which his papers have been published.
The author would also like to acknowledge his most sincere gratitude
to his graduate and undergraduate students, which have always provided him
with unlimited inspiration, motivation and joy.
Evangelos (Vangelis) Triantaphyllou
April 2000
Department of Industrial and Manufacturing Systems Engineering

3128 CEBA Building
Louisiana State University
Baton Rouge, LA 70803-6409, U.S.A.
E-mail: trianta@fsu.edu
Personal Web page: http://www.imse. fsu.edu/vangeiis/
Chapter 1
INTRODUCTION TO MULTI-CRITERIA
DECISION MAKING
1.1 MULTI-CRITERIA DECISION MAKING:

A GENERAL OVERVIEW
The analysis of the way people make decisions (prescriptive theories)

or the way people ought to make decisions (normative theories) is perhaps as
old as the recorded history of mankind. Of course, not all these analyses
were characterized by the rigorous scientific approaches we see in the
literature today. Therefore, it is not surprising that the literature in decision
making is humongous and continuously increasing. At the same time,
however, the development of the perfect decision making method for rational
real life decision making still remains an elusive goal. This contradiction
between the extensiveness of the study on this subject and the elusiveness of
the final goal of the real life applicability of the findings, constitutes in a way
the ultimate decision making paradox.
Multi-criteria decision making (MCDM) is one of the most well
known branches of decision making. According to many authors (see, for
instance, [Zimmermann, 1996]) MCDM is divided into multi-objective
decision making (MODM) and multi-attribute decision making (MADM).
However, very often the terms MADM and MCDM are used to mean the
same class of models (Le., MCDM).
MODM studies decision problems in which the decision space is
continuous. A typical example is mathematical programming problems with
multiple objective functions. The first reference to this problem, also known
as the "vector-maximum" problem, is attributed to Kuhn and Tucker [1951].
On the other hand, MCDM/MADM concentrates on problems with discrete
decision spaces. In these problems the set of decision alternatives has been
predetermined.
Although MCDM methods may be widely diverse, many of them have
certain aspects in common [Chen and Hwang, 1991]. These are the notions
of alternatives and attributes (also often called goals or decision criteria) as
described next.
Alternatives:
Usually alternatives represent the different choices of action available to the
decision maker. In this book the set of alternatives is assumed to be finite,
E. Triantaphyllou, Multi-criteria Decision Making Methods: A Comparative Study

© Springer Science+Business Media Dordrecht 2000
2 MCDM Methods: A Comparative Study, by E. Triantaphyllou
ranging from several to hundreds. They are supposed to be screened,

prioritized, and eventually ranked.
Multiple Attributes:
Each MCDM problem is associated with multiple attributes. Attributes are
also referred to as "goals" or .,decision criteria". Attributes represent the
different dimensions from which the alternatives can be viewed.
In cases in which the number of criteria is large (e. g., more than a
dozen), criteria may be arranged in a hierarchical manner. That is, some
criteria may be major ones. Each major criterion may be associated with
several sub-criteria. Similarly, each sub-criterion may be associated with
several sub-sub-criteria and so on. Although some MCDM methods may
explicitly consider a hierarchical structure in the criteria of a decision
problem, most of them assume a single level of criteria (e.g., no hierarchies).
Conflict among Criteria:

Since different criteria represent different dimensions of the alternatives, they
may conflict with each other. For instance, cost may conflict with profit etc.
In this book no such conflict is assumed unless it is explicitly stated
otherwise.
Incommensurable Units:
Different criteria may be associated with different units of measure. For
instance, in the case of buying a used car, the criteria "cost" and "mileage"
may be measure~ in terms of dollars and thousands of miles, respectively.
It is this nature of having to consider different units which makes MCDM
problems intrinsically hard to solve.
Decision Weights:
Most of the MCDM methods require that the criteria be assigned weights of
importance. Usually, these weights are normalized to add up to one. How
these weights can be determined is described in Chapters 4 and 5.
Decision Matrix:
An MCDM problem can be easily expressed in a matrix format. A decision
matrix A is an (m X n) matrix in which element au indicates the performance
of alternative Ai when it is evaluated in terms of decision criterion C; (for i =
1,2,3, ... , m, andj = 1,2,3, ... , n). It is also assumed that the decision
maker has determined the weights of relative performance of the decision
criteria (denoted as wj ' for j = 1, 2, 3, ... , n). This information is best
summarized in Definition 1-1 (according to [Zimmermann, 1996]) and Figure
Chapter 1: Introduction to MCDM 3
1-1. Please note that in this definition the term "goals" is used instead of the
usual term "criteria".
Definition 1-1:
Let A = {Ai' for i = 1, 2, 3, ... , n} be a (finite) set of decision alternatives
and G = {gj' for j = 1, 2, 3, ... , m} a (finite) set of goals according to
which the desirability of an action is judged. Determine the optimal
alternative A· with the highest degree of desirability with respect to all
relevant goals gj'
Criteria
CI C2 C3 Cn
Alts. ( WI w2 W3 wn )
Al all a l2 aJ3 a ln
A2 a21 a22 a23 a2n
Am amI am2 amJ amn
Figure 1-1: A Typical Decision Matrix.
1.2 CLASSIFICATION OF MCDM METHODS

As it was stated in the preface, there are many MCDM methods
available in the literature. Each method has its own characteristics. There
are many ways one can classify MCDM methods. One way is to classify
them according to the type of data they use. That is, we may have
deterministic, stochastic, or fuzzy MCDM methods (for an overview of
fuzzy MCDM methods see [Chen and Hwang, 1991]). However, there may
be situations which involve combinations of all the above (such as stochastic
and fuzzy) data types.
Another way of classifying MCDM methods is according to the
number of decision makers involved in the decision process. Hence, we have
single decision maker MCDM methods and group decision makers MCDM
(for more information on the later class, the reader may want to check the
journal of Group Decision Making). In this book we concentrate our attention
to single decision maker deterministic MCDM methods.
In [Chen and Hwang, 1991] deterministic -- single decision maker --
MCDM methods were also classified according to the type of information and
the salient features of the information. A taxonomy of a number of MCDM

methods is given in Figure 1-2 (as adopted from [Chen and Hwang, 1991]).
For a brief description of the methods mentioned in Figure 1-2, the interested
reader may want to consult with [Hwang, 1987] or [Chen and Hwang, 1991].
The weighted sum model (WSM), the analytic hierarchy process (AHP), the
revised AHP, the weighted product model (WPM), the ELECTRE, and the
TOPSIS method are described in later sections and are the ones which are the
most widely used. Finally, it should be stated here that there are many other
alternative ways for classifying MCDM methods [Chen and Hwang, 1991].
However, the previous ones are the most common schemes.
Type of Information Salient Feature Major Classes of Methods

From the Decision of Information
Maker
Dominance
No Information r---------t~ Maximin
7'-------.J Maximax
Multi-Attribute Conjunctive Method

Decision (Satistying Method)
Making
Disjunctive Method
Elimination by Aspect
Information Lexicographic Semi order
on the l...-_---'r---+i Lexicographic Method
Attributes
Weighted Sum Model

Weighted Product Model
\---+1 Analytic Hierarchy Process
l...-_---' ELECTRE
TOPSIS
Figure ]-2: A Taxonomy of MCDM Methods (according to Chen and Hwang [1991]).
Chapter 2
MULTI-CRITERIA DECISION MAKING

METHODS
2.1 BACKGROUND INFORMATION

With the continuing proliferation of decision methods and their
variants, it is important to have an understanding of their comparative value.
Each of the methods uses numeric techniques to help decision makers choose
among a discrete set of alternative decisions. This is achieved on the basis
of the impact of the alternatives on certain criteria and thereby on the overall
utility of the decision maker(s). The difficulty that always occurs when trying
to compare decision methods and choose the best one is that a paradox is
reached, i.e., What decision-making method should be used to choose the
best decision-making method? This problem is examined in Chapter 9.
Despite the criticism that multi-dimensional methods have received,
some of them are widely used. The weighted sum model (WSM) is the
earliest and probably the most widely used method. The weighted product
model (WPM) can be considered as a modification of the WSM, and has been
proposed in order to overcome some of its weaknesses. The analytic
hierarchy process (AHP), as proposed by Saaty [Saaty, 1980 and 1994], is a
later development and it has recently become increasingly popular. Professors
Belton and Gear [1983] suggested a modification to the AHP (which we will
call the revised AHP) that appears (as it is demonstrated in later chapters) to
be more consistent than the original approach. Some other widely used
methods are the ELECTRE and the TOPSIS methods.
In the section that follows these methods are presented in detail. In
Chapter 9 the methods are tested in terms of two evaluative criteria. The same
chapter also uses the test findings and examines the implication of these
findings on the effectiveness of the various decision making approaches.
2.2 DESCRIPTION OF SOME MCDM METHODS

There are three steps in utilizing any decision-making technique
involving numerical analysis of alternatives:
1) Determine the relevant criteria and alternatives.

2) Attach numerical measures to the relative importance
of the criteria and to the impacts of the alternatives

on these criteria.
3) Process the numerical values to determine a ranking
of each alternative.
This section is only concerned with the way the WSM, WPM, AHP,
revised AHP, ELECTRE, and TOPSIS methods process the numerical data
in step 3. The central decision problem examined in this book is described
as follows. Given is a set of m alternatives denoted as AI' A2, A3, ••• , Am and
a set of n decision criteria denoted as Cl , C2 , C3, ... , Cn. It is assumed that
the decision maker has determined (the absolute or relative) performance
value aij (for i = 1, 2, 3, ... , m andj = 1, 2, 3, ... , n) of each alternative
in terms of each criterion. That is, he/she has determined the matrix A with
the aij values, along with the criteria weights Wj (i.e., the decision matrix as
it was defined in Section 1.2). In the next chapter a number of procedures for
determining these data is discussed.
Given the aij and Wj values, then the problem examined in this chapter
is how one can rank the alternatives when all the decision criteria are
considered simultaneously. In the following sections, the criteria are assumed
to represent some kind of profit. That is, the higher the value, the better it
is. Next, a number of MCDM models for solving the above problem (i.e.,
step 3 above) is presented.
2.2.1 The WSM Method
The weighted sum model (WSM) is probably the most commonly used
approach, especially in single dimensional problems. If there are m
alternatives and n criteria then, the best alternative is the one that satisfies (in
the maximization case) the following expression [Fishburn, 1967]:
n
A~M-scor" = ~ L aijwj , for i =1, 2, 3, ... , m. (2-1)
I j = 1
where: A*WSM-score is the WSM score of the best alternative, n is the number
of decision criteria, (lij is the actual value of the i-th alternative in terms of the
j-th criterion, and Wj is the weight of importance of the j-th criterion.
The assumption that governs this model is the additive utility
assumption. That is, the total value of each alternative is equal to the sum
of the products given as (2-1). In single-dimensional cases, where all the
units are the same (e.g., dollars, feet, seconds), the WSM can be used
without difficulty. Difficulty with this method emerges when it is applied to
Chapter 2: MCDM Methods 7
multi- dimensional MCDM problems. Then, in combining different

dimensions, and consequently different units, the additive utility assumption
is violated and the result is equivalent to "adding apples and oranges".
Example 2-1:
Suppose that an MCDM problem involves four criteria, which are
expressed in exactly the same unit, and three alternatives. The relative
weights of the four criteria were determined to be: wJ = 0.20, w2 = 0.15,
W 3 = 0.40, and W 4 = 0.25. Also, the performance values of the three
alternatives in terms of the four decision criteria are assumed to be as follows:
25 20 15 30
A = 10 30 20 30
30 10 30 10
Therefore, the data for this MCDM problem are summarized in the following
decision matrix:
Criteria
C1 C2 C3 C4
Alts. ( 0.20 0.15 0.40 0.25)
A1 25 20 15 30
A2 10 30 20 30
A3 30 10 30 10
When formula (2-1) is applied on the previous data the scores of the three
alternatives are:
A J• WSM-score = 25 xO.20 + 20xO.15 + 15 x 0.40 + 30xO.25 =
= 21.50.
Similarly, we get:
A 2• WSM-score = 22.00,
and A 3. WSM-score = 20.00.
Therefore, the best alternative (in the maximization case) is alternative
A2 (because it has the highest WSM score; 22.00). Moreover, the following
ranking is derived: A2 > A J > A3 (where the symbol" > " stands for "better
than") . •
2.2.2 The WPM Method

The weighted product model (WPM) is very similar to the WSM.
The main difference is that instead of addition in the model there is
multiplication. Each alternative is compared with the others by multiplying
a number of ratios, one for each criterion. Each ratio is raised to the power
equivalent to the relative weight of the corresponding criterion. In general,
in order to compare two alternatives AK and AL, the following product
(Bridgman [1922] and Miller and Starr [1969]) has to be calculated:
11
R(AxIAL) ;: II (aKjlaLir J, (2-2)

i = 1
where n is the number of criteria, aij is the actual value of the i-th alternative
in terms of the j-th criterion, and Wj is the weight of importance of the j-th
criterion.
If the term R(AK / AJ is greater than or equal to one, then it indicates
that alternative AK is more desirable than alternative AL (in the maximization
case). The best alternative is the one that is better than or at least equal to all
other alternatives.
The WPM is sometimes called dimensionless analysis because its
structure eliminates any units of measure. Thus, the WPM can be used in
single- and multi-dimensional MCDM. An advantage of the method is that
instead of the actual values it can use relative ones. This is true because:
11
aKjI E
i = 1
aKj
;:
I
aKj
(2-3)
11
aIJI E
i = 1
ali
A relative value d Kj is calculated using the formula:
where the aKj's are the actual values.
Example 2-2:
Consider the problem presented in the previous Example 2-1.
However, now the restriction to express all criteria in terms of the same unit
is not needed. When the WPM is applied, then the following values are
derived:
R(A/A:zJ = (25/10)°·20 x (20/30)°·15 X (15/20)°·40 X (30/30)°·25 =
= 1.007 > 1.
Similarly, we also get:
R(A/A~ = 1.067 > 1,
and R(A/A3J = 1.059 > 1.

Therefore, the best alternative is AI' since it is superior to all the other
alternatives. Moreover, the ranking of these alternatives is as follows: Al >
A z > A3 · •
An alternative approach with the WPM method is for the decision

maker to use only products without ratios. That is, to use the following
variant of formula (2-2);
n
P(A K } = II (aKjr J, (2-4)
j = 1
In the previous expression the term P(A k ) denotes the performance value (not
a relative one) of alternative Ak when all the criteria are considered under the
WPM model. Then, when the previous data are used, exactly the same
ranking is derived. Some interesting properties of this method are discussed
in Chapter 11, Sections 11.6 and 11.7.
2.2.3 The AHP Method

The analytic hierarchy process (AHP) ([Saaty, 1980 and 1994])
decomposes a complex MCDM problem into a system of hierarchies (more
on these hierarchies can be found in [Saaty, 1980]). The final step in the AHP
deals with the structure of an mXn matrix (where m is the number of
alternatives and n is the number of criteria). The matrix is constructed by
using the relative importances of the alternatives in terms of each criterion.
The vector (ail' aiZ' ai3 , ... , ain ) for each i is the principal eigenvector of an
n X n reciprocal matrix which is determined by pairwise comparisons of the
impact of the m alternatives on the i-th criterion (more on this, and other
related techniques, is presented in Chapter 3).
The importance of the AHP, its variants, and the use of pairwise
comparisons in decision making is best illustrated in the more than 1,000
references cited in [Saaty, 1994]. A number of special issues in refereed
journals have been devoted to the AHP and the use of pairwise comparisons
in decision making. These issues are: Socio-Economic Planning Sciences
[Vol. 10, No.6, 1986]; Mathematical Modelling [Vol. 9, No. 3-5, 1987];
European Journal of Operational Research [Vol. 48, No.1, 1990]; and
Mathematical and Computer Modelling [Vol. 17, No. 4/5, 1993]. Also, four
international symposia (called ISAHP) have been dedicated on the same topic
so far and one such event is now scheduled every two years.
Some evidence is presented in [Saaty, 1980]) (see also Chapters 3 and
4 in this book) which supports the technique of pairwise comparisons for

eliciting numerical evaluations of qualitative phenomena from experts and
decision makers. However, here we are not concerned with the possible
advantages and disadvantages of the pairwise comparison and eigenvector
methods for determining the aij values. Instead, we examine the method used
in the AHP to process the aij values after they have been determined. The
entry aij' in the m x n matrix, represents the relative value of alternative Ai
n
when it is considered in terms of criterion C;. In the original AHP the sum L
aij
is equal to one. i = 1
According to the AHP the best alternative (in the maximization case)
is indicated by the following relationship (2-5):
n
A;HP-score = ~x L aijwj' for i =1, 2, 3, ,." m. (2-5)
I j = 1
The similarity between the WSM and the AHP is clear. The AHP
uses relative values instead of actual ones. Thus, it can be used in single- or
multi-dimensional decision making problems.
Example 2-3:
As before, we consider the data used in the previous two examples
(note that as in the WPM case the restriction to express all criteria in terms
of the same unit is not needed). The AHP uses a series of pairwise
comparisons (more on this can be found in Chapter 3) to determine the
relative performance of each alternative in terms of each of the decision
criteria. In other words, instead of the absolute data, the AHP would use the
following relative data:
Criteria
CI C2 C3 C4
Alts. ( 0.20 0.15 OAO 0.25)
Al 25/65 20/55 15/65 30/65

A2 10/65 30/55 20/65 30/65
A3 30/65 5/55 30/65 5/65
That is, the columns in the decision matrix have been normalized to add up
to one. When formula (2-5) is applied on the previous data, the following
scores are derived:
A 1, AHP.score = (25/65)xO.20 + (20/55)xO.15 +
+ (15/65)xOAO + (30/65)xO.25 =
0.34.
Similarly, we get:
A 2• AHP-score = 0.35,
and A 3. AHP.score 0.31.
Therefore, the best alternative (in the maximization case) is alternative A2
(because it has the highest AHP score; 0.35). Moreover, the following
ranking is derived: A2 > Al > A3 • •
2.2.4 The Revised AHP Method

Belton and Gear [1983] proposed a revised version of the original
AHP model. They demonstrated that a ranking inconsistency can occur when
the AHP is used. A numerical example was presented that consists of three
criteria and three alternatives. In that example (which is also shown next) the
indication of the best alternative changes when an identical alternative to one
of the nonoptimal alternatives is introduced now creating four alternatives.
According to Belton and Gear the root for that inconsistency is the fact that
the relative values for each criterion sum up to one. Instead of having the
relative values of the alternatives AI> A21 A31 "'1 Am sum up to one, they
proposed to divide each relative value by the maximum value of the relative
values. In particular, they elaborated on the following example.
Example 2-4: (from [Belton and Gear, 1983], p. 228)

Suppose that the actual data of a MCDM problem with three
alternatives and three criteria are as follows:
Criteria
C1 C2 C3
Alts. (113 113 113)
1 9 8
9 1 9
1 1 1
In real life problems the decision maker may never know the previous
real data. Instead, he/she can use the method of pairwise comparisons (as
described in Chapters 3 and 4) to derive the relative data. When the AHP is
applied on the previous data, the following decision matrix with the relative
data is derived:
Criteria
CI C2 C3
Alts. (113 113 113)
Al 1111 9/11 8/18

A2 9/11 1111 9/18
A3 1111 1111 1118
Therefore, it can be easily shown that the vector with the final AHP scores
is: (0.45, 0.47, 0.08). That is, the three alternatives are ranked as follows:
A2 > Al > A3 •
Next, we introduce a new alternative, say A 4 , which is an identical
Q!ID.'. of the existing alternative A2 (i.e., A2 == A4)' Furthermore, it is also
assumed that the relative weights of importance of the three criteria remain
the same (i.e., 113, 113, 113). When the new alternative A4 is considered,
it can be easily verified that the new decision matrix is as follows:
Criteria
CI Cz C3
Alts. ( 113 1/3 113)
Al 1/20 9/12 8/27

Az 9/20 1112 9/27
A3 1120 1112 1/27
A4 9/20 1112 9/27
Similarly as above, it can be verified that the vector with the final AHP scores
is: (0.37, 0.29, 0.06, 0.29). That is, the four alternatives are ranked as
follows: Al > A2 == A4 > A 3 • Belton and Gear claimed that this result is in
logical contradiction with the previous result (in which A2 > Aj).
When the revised AHP is applied on the last data (that is, when the
data are normalized by dividing by the largest entry in each column), the
following decision matrix is derived:
Criteria
CI C2 C3
Alts. ( 113 113 113)
Al 119 1 8/9
Az 1 119 1
A3 119 119 119
A4 1 119 1
The vector with the final scores is: (2/3, 19/27, 1/9, 19/27). That is, the
four alternatives are ranked as follows: A2 == A4 > Al > A3 • The last
ranking is, obviously, the desired one. •
The revised AHP was sharply criticized by Saaty in [1990]. He

claimed that identical alternatives should not be considered in the decision
process. However, even earlier Triantaphyllou and Mann in [1989] had
demonstrated that similar logical contradictions are possible with the original
AHP, as well as with the revised AHP, even when non-identical alternatives
are introduced (see also Section 9.4 in Chapter 9). Some discussion on this
controversy is also presented in Chapter 11.
Two other MCDM methods are presented next. These methods are
of limited acceptance by the scientific and practitioners communities. These
are the ELECTRE and TOPSIS methods.
2.2.5 The ELECTRE Method
The ELECTRE (for Elimination and Choice Translating Reality;

English translation from the French original) method was first introduced in
[Benayoun, et al., 1966]. The basic concept of the ELECTRE method is to
deal with "outranking relations" by using pairwise comparisons among
alternatives under each one of the criteria separately. The outranking
relationship of the two alternatives Ai and Aj , denoted as A[~Aj' describes that
even when the i-th alternative does not dominate the J-th alternative
quantitatively, then the decision maker may still take the risk of regarding Ai
as almost surely better than Aj [Roy, 1973]. Alternatives are said to be
dominated, if there is another alternative which excels them in one or more
criteria and equals in the remaining criteria.
The ELECTRE method begins with pairwise comparisons of
alternatives under each criterion. Using physical or monetary values, denoted
as gJAj} and gdAJ of the alternatives Aj and Ak respectively, and by
introducing threshold levels for the difference gJAj } - gdAk }, the decision
maker may declare that he/she is indifferent between the alternatives under
consideration, that he/she has a weak or a strict preference for one of the two,
or that he/she is unable to express any of these preference relations.
Therefore, a set of binary relations of alternatives, the so-called outranking
relations, may be complete or incomplete. Next, the decision maker is
requested to assign weights or importance factors to the criteria in order to
express their relative importance.
Through the consecutive assessments of the outranking relations of the
alternatives, the ELECTRE method elicits the so-called concordance index,

defined as the amount of evidence to support the conclusion that alternative
Aj outranks, or dominates, alternative Ak , as well as the discordance index,
the counter-part of the concordance index.
Finally, the ELECTRE method yields a system of binary outranking
relations between the alternatives. Because this system is not necessarily
complete, the ELECTRE method is sometimes unable to identify the most
preferred alternative. It only produces a core of leading alternatives. This
method has a clearer view of alternatives by eliminating less favorable ones.
This method is especially convenient when there are decision problems that
involve a few criteria with a large number of alternatives [Lootsma, 1990].
There are many variants of the ELECTRE method. The organization
of the original version of the ELECTRE method is illustrated in the following
steps [Benayoun, et at., 1966]:
Step 1: Nonnalizing the Decision Matrix

This procedure transforms the entries of the decision matrix into
dimensionless comparable entries by using the following equation:
(2-6)
~i>~
k=l
Therefore, the normalized matrix X is defined as follows:
where m is the number of alternatives, n is the number of criteria, and xi} is

the normalized preference measure of the i-th alternative in terms of the j-th
criterion.
Step 2: Weighting the Nonnalized Decision Matrix

Next, each one of the columns of the previous X matrix is multiplied
by the associated weight of importance of the corresponding decision
criterion. These weights, denoted as (WI> W 2 , w3 , ... , wn), were determined
by the decision maker. Therefore, the weighted matrix, denoted as Y, is:
Y = X W, or:
Yl1 Y12 Y13 YIn
Y21 Y22 Y23 Y2n
Yml Y m2 Ym3 .•• Y mn
WIX m1 W 2 X m2 W3 X m3 ... wnxmn
where:
WI 0 0 0
0 w2 0 0
n
W and LW i l.
i= I
0 0 0 ... Wn
SteR 3: Determine the Concordance and Discordance Sets

The concordance set Ckl of two alternatives Ak and AI' where m ;?;
k, I ;?; 1, is defined as the set of all the criteria for which Ak is preferred to
AI. That is, the following is true:
Ckl = {j, Ykj ;?; Yij}, for j = 1,2,3, ... , n.
The complementary subset is called the discordance set and it is described as
follows:
Dkl = {j, Ykj < Ylj}' for j = 1, 2, 3, ... , n.
Step 4: Construct the Concordance and Discordance Matrices

The relative value of the elements in the concordance matrix C is calculated
by means of the concordance index. The concordance index ckt is the sum
of the weights associated with the criteria contained in the concordance set.
That is, the following is true:
ekl = L Wj' for j = 1, 2, 3, ... , n.

jECk 1
The concordance index indicates the relative importance of alternative

Ak with respect to alternative At. Apparently, 0 ~ ckt ~ 1. The
concordance matrix C is defined as follows:
where the entries of matrix C are not defined when k = I.

The discordance matrix D expresses the degree that a certain
alternative Ak is worse than a competing alternative At. The elements dkt of
the discordance matrix are defined as follows:
max
JEDkl
IYkj - Ylj I
dkl = (2-7)
max
j
IYkj - Ylj I
The discordance matrix is defined as follows:
d m1 d m2 dm3 ...
as with the C matrix, the entries of matrix D are not defined when k = I.
Finally, it should be noted here that the previous two m x m matrices are
not symmetric.
Step 5: Detennine the Concordance and Discordance Dominance

Matrices
The concordance dominance matrix is constructed by means of a
threshold value for the concordance index. For example, alternative Ak will
only have a chance to dominate alternative Al if its corresponding concordance
index Ckl exceeds at least a certain threshold value f. That is, this happens if
the following condition is true:
cld ;?! f·
The threshold value f can be determined as the average concordance
index. That is, the following relation could be true:
1 m m
" = m(m - 1)
E E
k=l 1= 1
C/d'
(2-8)
tIIt4 k., tIIt4 I.k
Based on the threshold value, the elements of the concordance dominance

matrix F are next determined as follows:
hi = 1, if CId ;?! f,
hi = 0, if cld < f·
Similarly, the discordance dominance matrix G is defined by using
a threshold value d, where d could be defined as follows:
m m
1
d =
m(m - 1) E E
k=l 1= 1
dId' (2-9)
tIIt4 k., tIIt4 I.k
and gkl = 1, if did d,

;?!
gkl = 0, if did < d·

Step 6: Determine the A22l"e2ate Dominance Matrix
The elements of the ag2l"egate dominance matrix E are next defined
as follows:
(2-10)
Step 7: Eliminate the Less Favorable Alternatives

From the aggregate dominance matrix one can derive a partial
preference ordering of the alternatives. If eld = 1, then this means that
alternative Ak is preferred to alternative Al by using both the concordance and
discordance criteria.
If any column of the aggregate dominance matrix has at least one
element equal to 1, then this column is "ELECTREally" dominated by the
corresponding row. Therefore, one can simply eliminate any column(s) which
have an element equal to one. Then, the best alternative is the one which
dominates all other alternatives in this manner.
2.2.6 The TOPSIS Method
TOPSIS (for the Technique for Order Preference by Similarity to

Ideal Solution) was developed by Yo on and Hwang [1980] as an alternative
to the ELECTRE method and can be considered as one of its most widely
accepted variants. The basic concept of this method is that the selected
alternative should have the shortest distance from the ideal solution and the
farthest distance from the negative-ideal solution in some geometrical sense.
The TOPSIS method assumes that each criterion has a tendency of
monotonically increasing or decreasing utility. Therefore, it is easy to define
the ideal and negative-ideal solutions. The Euclidean distance approach was
proposed to evaluate the relative closeness of the alternatives to the ideal
solution. Thus, the preference order of the alternatives can be derived by a
series of comparisons of these relative distances.
The TOPSIS method evaluates the following decision matrix which
refers to m alternatives which are evaluated in terms of n criteria:
where xij denotes the performance measure of the i-th alternative in terms
of the J-th criterion. Next, the steps of the TOPSIS method are presented.
Step 1: Construct the Nonnalized Decision Matrix

The TOPSIS method first converts the various criteria dimensions into
non-dimensional criteria as was the case with the ELECTRE method (i.e.,
relation (2-6». An element rij of the normalized decision matrix R is thus
calculated as follows:
Chapter 2: MCDM ¥ethods 19
(2-11)
~i;x~
&=1
As a remark, it should be stated here that in the ELECTRE and

TOPSIS methods the Euclidean distances defined in expressions (2-6) and (2-
11), respectively, represent some plausible assumptions. Other alternative
distance measures could be used as well, in which case it is possible for one
to get different answers for the same problem.
Step 2: Construct the Weighted Normalized Decision Matrix

A set of weights W = (wI> w2 , W 3 , ... , wn ), (where: Ew; = 1) defined
by the decision maker is next used with the decision matrix to generate the
weighted normalized matrix V as follows:
Step 3: Determine the Ideal and the Negative-Ideal Solutions

The ideal, denoted as A*, and the negative-ideal, denoted as A-,
alternatives (solutions) are defined as follows:
A" = { (max vii I jEJ), (min vii I j E JI), i = 1,2,3, ... ,m}
= { VI*, V 2*, ... , vn* }. (2-12)

A- = { (min vii I jEJ), (max vii I j E JI), i = 1,2,3, ... ,m}
= { VI_, v2_, ... , vn_ }. (2-13)

where: J = {j = 1, 2, 3, ... , n and j is associated with benefit criteria},
J I = {j = 1, 2, 3, ... , nand j is associated with cost/loss criteria}.
The previous two alternatives are fictitious. However, it is reasonable

to assume here that for the benefit criteria, the decision maker wants to have
a maximum value among the alternatives. For the cost criteria, the decision
maker wants to have a minimum value among the alternatives. From the
previous definitions it follows that alternative A * indicates the most preferable
alternative or the ideal solution. Similarly, alternative A- indicates the least
preferable alternative or the negative-ideal solution.
Step 4: Calculate the Separation Measure

The n-dimensional Euclidean distance method is next applied to
measure the separation distances of each alternative from the ideal solution
and negative-ideal solution. Thus, for the distances from the ideal solution
we have:
n
L (vij - Vj J2, for i = 1,2,3, ... , m, (2-14)
j = 1
where Si* is the distance (in the Euclidean sense) of each alternative from the
ideal solution.
Similarly, for the distances from the negative-ideal solution we have:
n
S.1 - L (vij - Vj _)2, for 1, 2, 3, ... , m, (2-15)
j = 1
where Si_ is the distance (in the Euclidean sense) of each alternative from the
negative-ideal solution.
Step 5: Calculate the Relative Closeness to the Ideal Solution

The relative closeness of an alternative Ai with respect to the ideal
solution A* is defined as follows:
c. =
S.'
1*
Si. + 1-
and i = 1, 2, 3, ... , m. (2-16)

Apparently, Ci* = 1, if Ai = A*, and Ci_ = 0, if Ai = A-.
Step 6: Rank the Preference Order

The best (optimal) alternative can now be decided according to the
preference rank order of C,"*. Therefore, the best alternative is the one that
has the shortest distance to the ideal solution. The previous definition can
also be used to demonstrate that any alternative which has the shortest
distance from the ideal solution is also guaranteed to have the longest distance
from the negative-ideal solution.
Chapter 3
QUANTIFICATION OF QUALITATIVE DATA FOR

MCDM PROBLEMS
The first step in any MCDM problem is to define the set of

alternatives and the set of decision criteria that the alternatives need to be
evaluated with. Although this is an enormously critical step, its formulation
cannot easily be captured with a standard modeling procedure. This task
appeals more to the art aspect of MCDM than to the science one. It is this
realization that makes most experts in this area to preach that the single most
important step in solving any MCDM (and for that reason any decision
making) problem is to first correctly define the problem. The interested
reader may want to refer to the classic book with the characteristic fables by
Russell L. Ackoff [Ackoff, 1978] on the art and science of decision making.
Another very critical step in dealing with any MCDM (and also with
any decision making) problem is to accurately estimate the pertinent data.
Very often, in MCDM problems the data cannot be known in terms of
absolute values. For instance, what is the worth of the i-th car (i.e., an
alternative) in terms of its aesthetic value (i.e., a decision criterion)?
Although information about questions like the previous one may be vital in
making the correct decision, it is very difficult, if not impossible, to quantify
it correctly. Therefore, many decision making methods attempt to determine
the relative importance, or weight, of the alternatives in terms of each
criterion involved in a given MCDM problem. This is examined in this
chapter and the discussions are based on the results published in
[Triantaphyllou, Lootsma, et al., 1994].
Next we consider the case of having a single decision criterion and a
set of m alternatives, denoted as Ai (for i = 1, 2, 3, ... , m). The decision
maker wishes to determine the relative performance of these alternatives in
terms of the single criterion. An approach based on pairwise comparisons,
which was proposed by Saaty [1977; 1980; and 1994] as part of the AHP
method, has long attracted the interest of many researchers and practitioners
alike. This occurred because of both of its easy applicability and interesting
mathematical properties. Pairwise comparisons can be used to determine the
relative importance of each alternative in terms of each criterion.
In that approach the decision maker has to express his/her opinion
about the value of a single pairwise comparison at a time. Usually, the
24 MCDM Method: A Comparative Study. by E. Triantaphyllou
decision maker has to choose his/her answer among 10-17 discrete choices.
Each choice is a linguistic phrase. Some examples of such linguistic phrases
are: "A is more important than B, " or "A is of the same importance as B, "
or "A is a little more important than B, " and so on. The main focus of this
chapter is not the wording of these linguistic statements, but, instead, the
numerical values which should be associated with such statements.
The main challenge with the use of pairwise comparisons is how to
quantify the linguistic choices selected by the decision maker during the
evaluation of the pairwise comparisons. All the methods which use the
pairwise comparisons approach eventually express the qualitative answers
(i.e., the linguistic statements) of a decision maker into some numbers. The
present chapter examines the issue of quantifying pairwise comparisons. Since
pairwise comparisons are the keystone in solving many MCDM problems,
correctly quantifying them is a very crucial step in MCDM methods which
use qualitative or fuzzy data.
Pairwise comparisons are quantified by using a scale. Such a scale
is nothing but an one-to-one mapping between the set of discrete linguistic
choices available to the decision maker and a discrete set of numbers which
represent the importance, or weight, of the previous linguistic choices. There
are two major approaches in developing such scales. The first approach is
based on the linear scale proposed by Saaty [1980] as part of the analytic
hierarchy process (AHP). The second approach was proposed by Lootsma
[1988], [1990], and [1991] and determines exponential scales. Both
approaches depart from some psychological theories and develop the numbers
to be used based on these psychological theories.
This chapter is organized as follows. The second section illustrates
the principals of the two classes of scales. The second section also presents
some ways for generating even more scales based on Saaty's linear scale and
on the exponential scales proposed by Lootsma. The third section discusses
some ways for evaluating the performance of various scales. This is achieved
in terms of two evaluative criteria. The next section describes the problem
of selecting the appropriate scale (or scales) as an MCDM problem. Some
computational results presented in the fifth section reveal that under different
conditions some scales are more efficient than others. These findings are
presented in depth in the final section which is the conclusion section of this
chapter.
Chapter 3: Quantification of Qualitative Data in MCDM Problems 25
3.2 SCALES FOR QUANTIFYING PAIRWISE

COMPARISONS
As it was mentioned in the previous section, two classes of scales are
considered in this chapter. The first class of scales is defined on the interval
[9, 1/9] and is based on the original Saaty scale. The second class of scales
is based on the exponential scales introduced by Lootsma [1988], [1990], and
[1991]. Besides these two types of scales, there are other types of scales as
was first mentioned in [Churchman, Arnoff, Ackoff, 1957]. However, the
ones mentioned here are the ones which have been proposed to be used with
pairwise comparisons.
Once the pairwise comparisons are determined by using a scale, they
are processed in order to derive the final values. These values are estimates
of the relative magnitudes (or weights) of importance of the entities which are
compared in terms of a common characteristic they all share. Usually, one
compares a number of alternatives in terms of a single decision criterion.
Also, a decision maker may want to compare a set of decision criteria in
order to derive their relative weights of importance. Some alternative ways
of performing pairwise comparisons are discussed in Chapter 7, where a type
of dual comparisons is presented.
Saaty proposed in [Saaty, 1980; and 1994] an approach to process
pairwise comparisons which is based on eigenvalues. Another method, which
is based on a logarithmic regression model, is proposed by Lootsma [1988]
and [1991]. A critical discussion of these, as well some other approaches, is
presented in Chapter 4.
Also, an approach which uses differences instead of ratios is
presented in Chapter 5. That approach describes how similarity relations
among a group of entities (i.e., the alternatives or the criteria) can be
estimated by using an efficient quadratic programming formulation. All these
approaches are capable of estimating relative magnitudes. Therefore, the
actual values could only be derived, if at least one of them were known
apriori. However, this is not possible in real applications, thus only relative
magnitudes (or weights of importance) are derivable by using pairwise
comparisons.
It should be stated here that when pairwise comparisons are used the
entire process may become impractical when the number of the entities (i.e.,
alternatives or criteria) to be compared becomes large. If n is the number of
such entities, then the number of all possible comparisons is equal to n(n-1)/2.
For instance, for n = 100 the decision maker would have to make 4,950
pairwise comparisons!
26 MCDM Method: A Comparative Study, by E. Triantaphyllou
3.2.1 Scales Defined on the Interval [9, 1/9]

In 1846 Weber (as reported in [Saaty, 1980]) stated his law regarding
a stimulus of measurable magnitude. According to his psychological theory
a change in sensation is noticed if the stimulus is increased by a constant
percentage of the stimulus itself. That is, people are unable to make choices
from an infinite set. For example, people cannot distinguish between two
very close values of importance, say 3.00 and 3.02. Psychological
experiments have also shown the intriguing fact that most individuals cannot
simultaneously compare more than seven objects (plus or minus two) [Miller,
1956]. This is the main reasoning used by Saaty to establish 9 as the upper
limit of his scale, 1 as the lower limit and a unit difference between
successive scale values.
The values of the pairwise comparisons are determined according to
the instructions presented in Table 3-1 [Saaty, 1980]. According to this scale
(which we will call in this chapter Scale1) , the available values for the
pairwise comparisons are members of the set: {9, 8, 7, 6, 5,4, 3, 2, 1, 1/2,
1/3, 1/4, 1/5, 1/6, 117, 1/8, 1/9}. The above numbers illustrate that the
values for the pairwise comparisons can be grouped into the two intervals [9,
1] and [1, 1/9]. As it is implied from the above discussion, the values in the
interval [9, 1] are evenly distributed, while the values in the interval [1, 1/9]
are skewed to the right end of this interval.
However, one may reasonably claim at this point that there is no good
reason why for a scale defined on the interval [9, 1/9] the values on the sub-
interval [9, 1] should be evenly distributed. An alternative scale could have
the values evenly distributed in the interval [1, 119], while the values in the
interval [9, 1] could simply be the reciprocals of the values in the interval [1,
119]. This consideration leads to the scale (which we will call in this chapter
Scale2) with the following values: {9, 9/2, 9/3, 9/4, 9/5, 9/6, 917, 9/8, 1,
8/9, 7/9, 6/9, 5/9, 4/9, 3/9, 2/9, 119}. This scale was originally presented
in [Ma and Zheng, 1991]. In the second scale each successive value on the
interval [1, 1/9] is (1 - 1/9)/8 = 1/9 units apart. In this way, the values in
the interval [1, 1/9] are evenly distributed, while the values in [9, 1] are
simply the reciprocals of the values in [1, 1/9].
It should be stated here that the notion of having within a scale a
group of values evenly distributed, is followed in order to be in agreement
with the same characteristic present in the original Saaty scale (see also Table
3-1). As it will be seen in the next section, however, other scales can be
defined without having evenly distributed values.
Table 3-1: Scale of Relative Importances (according to Saaty [1980]).
Intensity of
Importance Definition Explanation
I Equal importance Two activities contribute

equally to the objective.
3 Weak importance of one Experience and judgment

over another slightly favor one activity
over another.
5 Essential or strong Experience and judgment

importance strongly favor one activity
over another.
7 Demonstrated importance An activity is strongly

favored and its dominance
demonstrated in practice.
9 Absolute importance The evidence favoring one

activity over another is of
the highest possible order of
affirmation.
2,4,6,8 Intermediate values between When compromise is needed

the two adjacent judgments
If activity i has one of the

Reciprocals of above above nonzero numbers
nonzero assigned to it when
compared with activity j,
then j has the reciprocal
value when compared with i.
Besides the second scale, many other scales can be generated. One
way to generate new scales is to consider weighted versions between the
previous two scales. That is, within the interval [1, 1/9] the values can be
calculated using the formula:
NewValue = Value(Scalel) +
+ [Value(Scale2) - Value(Scale1)] x (a/IOO),
where a can range from a to 100. Then, the values in the interval [9, 1] are
the reciprocals of the above values. Observe that for a = a Scale1 is
derived, while for a = 100 Scale2 is derived.
3.2.2 Exponential Scales

As was mentioned earlier in this chapter, a class of exponential scales
has been introduced by Lootsma [Lootsma, 1988; 1990; and 1991]. The
development of these scales is based on a different observation in psychology
about stimulus perception (denoted as e;). According to that observation,
which is due to Roberts [1979], the difference en+1 - en must be greater than
or equal to the smallest perceptible difference, which is proportional to en'
The permissible choices by the decision maker are summarized in Table 3-2.
As a result of Robert's theory the numerical equivalents of these linguistics
choices need to satisfy the following relations:
en+! - en = e en' (where e > 0) or:
en+ 1 = (1 + e) en = (1 + e)2 en_I =
= (1 + e)n+1 eo, (where eo = 1) or:
In the previous expressions the parameter 'Y is unknown (or,

equivalently, e is unknown), since 'Y = In(1 + e), and e is the basis of the
natural logarithms (please notice that in contrast e; is just the notation of a
variable). Table 3-3 presents the values of two exponential scales that
correspond to two different values of the 'Y parameter. Apparently, different
exponential scales can be generated by assigning different values to the 'Y
parameter.
Table 3-2: Scale of Relative Importances (according to Lootsma [1988]).
Intensity of
Importance Defmition
Indifference between Ai and Aj
Indifference threshold towards Ai
Weak preference for Ai
Commitment threshold towards Ai
Strong preference for Ai
Dominance threshold towards Ai
Very strong preference for Ai
Reciprocals of If member i has one of the above nonzero

above nonzero numbers assigned to it when compared with
member j, thenj has the reciprocal value when
compared with i.
Table 3-3: Two Exponential Scales.
NORMAL (y=1I2) STRETCHED (y=1) Dermition
e"lxo = 1.00 = 1.00 eo

e"y'Xl = 1.65 = 2.72 e,
e'Yx2 = 2.72 = 7.39 e2
e"'lX3 = 4.48 = 20.09 e,
e"fX4 = 7.39 = 54.60 e.
e"s = 12.18 = 148.41 es
e,X6 = 20.09 = 403.43 e6
Another difference between exponential scales and the Saaty scale

is on the number of categories allowed by the exponential scales. Now
there are only four major linguistically distinct categories, plus three
so-called threshold categories between them. The threshold categories can
be used if the decision maker hesitates between the main categories. In the
following section we present some evidence that human beings follow
exponential scales when they categorize an interval. More on these
examples can be found in Lootsma [1990] and [1991]. Very appropriate
here is also the new book by Lootsma [1999].
3.2.3 Some Examples of the Use of Exponential Scales

It is surprising to see how consistently humans categorize certain
intervals of interest in totally unrelated areas. In this section we present
some examples to show, for instance, how humans partition certain ranges
on the time, sound, and light intensities.
a) Historical periods. The written history of Europe, from 3000 Be until

today, is subdivided into a small number of major periods. Looking
backwards from 1989, the year when the Berlin Wall was opened, one can
distinguish the following turning points marking off the start of a
characteristic development:
1947: 42 years before 1989 beginning of cold war and

decolonization,
1815: 170 years before 1989 beginning of industrial and
colonial dominance,
1500: 500 years before 1989 beginning of world-wide trade

and modem science,
450: 1,550 years before 1989 beginning of middle ages,
-3000: 5,000 years before 1989 beginning of ancient history.
These major echelons, measured by the number of years before 1989,

constitute a geometric sequence with the progression factor equal to 3.3.
We can obtain a more refined subdivision when we introduce the years:
1914: 75 years before 1989 beginning of world wars,

1700: 300 years before 1989 modem science established,
1100: 900 years before 1989 beginning of high middle ages,
-800: 2,800 years before 1989 beginning of Greek/Roman
history.
With these turning points interpolated between the major ones, we find a
geometric sequence of echelons with the progression factor equal to 1.8.
b) Planning horizons. In industrial planning activities, we usually observe

a hierarchy of planning cycles where decisions under higher degrees of
uncertainty and with more important consequences for the company are
prepared at increasingly higher management levels. The planning horizons
constitute a geometric sequence, as it shown in the following list:
1 week: weekly production scheduling,

1 month: 4 weeks monthly production scheduling,
4 months: 16 weeks ABC planning of tools and labor,
1 year: 52 weeks capacity adjustment,
4 years: 200 weeks production planning,
10 years: 500 weeks strategic planning of company structure.
The progression factor of these major horizons is equal to 3.5. In practice

there are no planning horizons between these major ones.
c) Size of nations. The above categorization is not only found on the time
axis, but also in spatial dimensions when we categorize the nations on the
basis of the size of their population. Omitting the very small nations with
less than one million inhabitants, we have:
small nations: 4 million DK, N,

medium-size nations: 15 million NL, DDR,
large nations: 60 million D, F, GB, I,
very large nations: 200 million USA, Russia,
giant nations: 1,000 million China, India.
We find again a geometric sequence, with the progression factor equal to

4.0. Furthermore, it seems reasonable to interpolate the following threshold
echelons:
small/medium size: 8 million A, B, GR,
medium size/large: 30 million E, PL,
large/very large: 110 million Japan.
Because the respective nations fall typically between the major echelons, the
refined sequence of echelons has the progression factor equal to 2.0.
d) Loudness of sounds. The range of audible sounds (measured in dB

units) can roughly be categorized as follows:
40 dB: very quiet; whispering,
60 dB: quiet; conversation,
80 dB: moderately loud; electric mower and food blenders,
100 dB: very loud; farm tractors and motorcycles,
120 dB: uncomfortable loud; jets during take-off.
Although the precision should be taken with a grain of salt because

we have a mixture of sound frequencies at each of these major echelons, we
can obviously find here a geometric sequence of subjective sound intensities
with the progression factor equal to 4.0.
e) Brightness of light. Physically, the perception of light and sound

proceed in different ways, but these sensory systems follow a similar
pattern. The range of visible light intensities (measured in Lum units) can
roughly be categorized as follows:
30 Lum: star light,

50 Lum: full moon,
70 Lum: street lighting,
90 Lum: office space lighting,
110 Lum: sunlight in summer.
Under the precaution that the precision should not be taken too
seriously because we have at each of these major echelons a mixture of
wave lengths, we can observe that the subjective light intensities also
constitute a geometric sequence with the progression factor equal to 4.0.
In the previous paragraphs we have used five examples to
demonstrate that exponential scales are common in human comparative
judgment when dealing with historical periods, planning horizons, size of
nations, sound intensities, and perception of light. Therefore, these
examples make exponential scales only plausible. Lootsma [1991] has
studied the scale sensitivity of the resulting scores when exponential scales
are used. He observed that the rank order of the scores is not affected by
variations of the scale parameter. That is, the numerical values of the
calculated scores are weakly dependent on that parameter.
For a more detailed documentation on psychophysics we refer the
interested reader to [Marks, 1974], [Michon, et al., 1976), [Roberts,
1979], [Zwicker, 1982], and [Stevens and Hallowell Davis, 1983]. The
reader will find that the sensory systems for the perception of tastes, smells,
and touches follow the power law with exponents in the vicinity of one.
3.3 EVALUATING DIFFERENT SCALES
In order to evaluate different scales two evaluative criteria were

developed in [Triantaphyllou, Lootsma, et al., 1994]. Furthermore, a
special class of pairwise matrices was also developed. These special
matrices are then used in conjunction with the two evaluative criteria in
order to investigate some stability properties of different scales. These
developments are described in more detail in the following sections.
3.3.1 The Concepts of the RCP and CDP Matrices

As it was mentioned earlier, reciprocal matrices with pairwise
comparisons (also known as judgment matrices) were introduced by Saaty
[1980] as a tool for extracting all the pertinent information from a decision
maker. The same author also proposed a scale which results in matrices
with entries from the set e, where e is the set of integers 1, 2, 3, ... , 9 and
their reciprocals (see also Table 3-1). If a different scale is to be used, then
e will be the finite set of discrete numerical values which represent that
scale. Usually, each entry in these matrices numerically represents the value
of a pairwise comparison between two alternatives with respect to a single
decision criterion. These matrices are constructed as to be an effective way
for capturing the necessary information regarding the relative performance
of the alternatives or the criteria [Saaty, 1980].
Saaty's judgment matrices have received wide acceptance as being
an effective way for extracting qualitative information for real world
MCDM problems (see, for example, [Chu, et al., 1979], [Federov, et al.,
1982], [Hihn and Johnson, 1988], [Khurgin and Polyakov, 1986], [Lootsma,
et al., 1990], and [Vargas, 1982], and also in this book, especially in
Chapter 4).
A related study in [Triantaphyllou, et al., 1990b] was based on the

assumption that in the real world the actual performance values of the
alternatives take on continuous values. Let WI' W 2, W 3, ... , Wn be the real
(and thus unknown) such values of a set of n alternatives (or decision
criteria) when they are considered in terms of a single common
characteristic. Each of the Wi values is assumed to be in the interval [1, 0].
If the decision maker knew the above real values then, he/she would be able
to have constructed a matrix with the real pairwise comparisons. In that
matrix, say matrix A, the entry Oli} is equal to w/Wj.
That is, the entry Oli} represents the real (and thus unknown) value
of the comparison when the i-th entity (i.e., an alternative or a criterion) is
compared with the j-th entity. We call this matrix the Real Continuous
Pairwise matrix, or the Rep matrix. Since in the real world the w/s are
unknown, so are the entries Oli} of the previous matrix. However, we will
assume here that the decision maker, instead of an unknown entry Oli} is able
to determine the closest value taken from the set e of the numerical values
provided by a scale. In other words, instead of the real (and thus
unknown) value Oli} we will assume that one is able to determine the value
ai} such that:
the difference I Oli} - ai} I is minimum and ai} E e.

Therefore, judgments about the values of the pairwise comparison
of the i-th entity when it is compared with the j-th entity are assumed to be
so accurate that they are closest (in absolute value terms) to the true or real
values one is supposed to estimate when a scale with the discrete values e
is used.
It should be stated at this point that other norms, alternative to the
previous one, are also possible to be assumed as the way a decision maker
best approximates real (and thus unknown) pairwise comparisons. For
instance, such an alternative norm could be as follows:
the difference I 0l1(1 + Oli}) - al(1 + ai}) is minimum

and ai} E e.
However, any norm which attempts to approximate the real (and thus
unknown) ratios with ratios taken from a finite and discrete set of values,
will always allow for the possibility that some real ratios (which are close
enough to each other) will be mapped to the same discrete value from the
current scale. The last statement indicates that Theorem 3-1 (stated later in
Section 3.3.2) will still be valid if alternative norms are considered
(however, its present proof assumes that the first norm is used).
The matrix with the aij entries that we assume the decision maker is
able to construct as described above has entries from the discrete and finite
set e. We call this matrix the Closest Discrete Pairwise matrix or the CDP
matrix. The CDP matrix may not be perfectly consistent. That is, the
consistency index (CI) values (see the next section for an exact definition of
the CI index) of CDP matrices are not necessarily equal to zero. More on
this inconsistency issue will be discussed in the following section. It is
important to observe here that the CDP matrices are the reciprocal matrices
with pairwise comparisons that a decision maker will construct if we assume
that each of his/her pairwise comparisons is the closest possible to its actual
real value.
The decision maker is limited by the discrete numerical values (i.e.,
the values from the set e provided to him/her by a scale). He/She may
never know the actual values of his/her pairwise comparisons. He/She
simply attempts to approximate them. In other words, we assume here that
these approximations are the closest possible. Clearly, this is a highly
favorable assumption when one attempts to investigate the effectiveness of
various scales. The following example illustrates further the concepts of the
RCP and CDP matrices.
Example 3-1:
Suppose that the real (and hence unknown) relative performance
values (or weights of importance), after normalization, of a set with three
entities (i.e., three alternatives or three criteria) are WI = 0.77348, W 2 =
0.23804, and W3 = 0.23848. Then, the RCP matrix with the real values of
the pairwise comparisons is as follows:
1 3.24938 3.24342
Rep = 0.30775 1 0.99817
0.30832 1.00183 1
This is true because, for instance, entry (2, 1) is equal to 0.30775

(WI/W2) = (0.23804/0.77348), and so on. If, for instance, the original Saaty
scale is to be used (as it is defined in Table 3-1) then, it can be easily
verified with a simple exhaustive enumeration that the corresponding CDP
matrix is as follows:
1 3 3
-
1 1 1
CDP = 3
1
1 1
3
To see this consider the (1, 2) entry of the previous RCP matrix.
For this entry we have a l2 = 3.24938. Therefore, when the values in Table
3-1 are to be used in order to quantify the (1, 2) pairwise comparison then,
the a l2 entry is approximated by the value 3. The value 3 is the closest one
to the value 3.24938 when the values in Table 3-1 are used. Clearly, this
is an assumption which is made here in order to study different scales. A
similar explanation holds for the rest of the entries in the previous CDP
matrix . •
As a final technical note it is stated here that it is possible for the

approximation of an aij element not to be the reciprocal of the
approximation of the corresponding 1Iaij element. For instance, if aij =
3.49 and the set e = {9, 8, 7, ... , 1, 112, 113, ... , 119} is used for the
discrete value approximations, then the approximation of aij is equal to 3,
while the value of the approximation of 1Iaij is equal to 114, and not equal
to 113. This happens because the value of the ratio 113.49 is equal to
0.2865329, which is closest to the element 114 than to the element 113 when
the previous e set is considered. However, in order to preserve the
reciprocity property of the derived CDP matrices (and thus the methods of
Chapter 4 for extracting the relative weights to be applicable), the reciprocal
elements were determined from the relationship aji = 1Iaij (for j > i and i
= 1,2,3, ... , n-l) and not by performing independent approximations with
the elements from the e set.
3.3.2 On The Consistency of CDP Matrices

If all the pairwise comparisons are perfectly consistent with each
other then, the following relation should always be true among any three
comparisons aik , akj , and aij [Saaty, 1980]:
ai k x akj = aij , for any 1 ~ i, j, k ~ n.
Saaty expresses the inconsistency of a pairwise comparison matrix in terms
of the consistency index (CI). The CI index is defined as follows:
A - n
CI = _max_-::--
n - 1 '
where Amax is the maximum eigenvalue of the matrix with the pairwise
comparisons and n is the order of that matrix.
In the following paragraphs we will show that CDP matrices can be
inconsistent regardless of the scale used to quantify the actual pairwise
comparisons. This is stated in terms of the following theorem:
Theorem 3-1:
Regardless of the scale that is used to quantify the pairwise comparisons of
n (where n ~ 3) entities, the corresponding CDP matrices may be
inconsistent.
Proof:
Without loss of generality, suppose that AI> A 2 , and A3 are three items
(alternatives or criteria) of a collection of n (where n ~ 3) items that we
need to compare in terms of some common characteristic. Let the current
scale be defined on the following (2k + 1) discrete values (where k ~ 1):
[Vk' 1/Vk_l , Vk_2 , ••• , V2 , VI> 1, 1/VI , 1/V2 , ... , 1/Vk_2 , 1/Vk_l , 1/Vk],
where V; > 0 for any i = 1, 2, 3, ... , k, and VI > 1.
In this proof it will be shown that, when the previous scale is used,
then it is possible for the three comparisons a 12 , a 13 , a32 made by the
decision maker not to satisfy the consistency requirement:
a l2 = a13 X Cl:32'
Suppose that the actual values of the pairwise comparisons that
correspond to the previous three items AI' A 2 , and A3 are as follows:
Al 3Vl + 1
= (X 12 =
Az 4
Al VI + 3
= (X 13 =
A3 4
A3 (A l/A 2) (X 12 3Vl + 1
and = (X 32
A2 (Al/A3) (Xu VI + 3
Using the above relations it can be easily verified (since VI > 1) that the
following conditions (i) are true (see also Figure 3-1):
V1 + 1
V1 > cx 12 > 2
V1 + 1
2 ~ cx 13 >1,
V1 + 1
2 > CX 32 > 1.
From Figure 3-1, or conditions (z), it follows that in the

corresponding CDP matrix the decision maker will assign the following
three values au, a 13 , ~ (taken from the current scale) to the previous three
pairwise comparisons:
au = VI'
a l3 = 1,
~ = 1.
Clearly, the consistency requirement does not hold for these three values
because:
au ¢ a l3 x tl:32.
(Since a l2 = VI ¢ a l3 x tl:32 = 1 x 1 = 1, and Yt > 1.)
In other words, the entire CDP matrix is inconsistent. •
1/V1 1.00
! ! MI
I
Note: MI = (VI + 1)/2; is the middle point of the interval [VI, 1].
Figure 3-1: Actual Comparison Values.

The previous theorem states that under the most favorable

assumption that the decision maker is capable of determining only the
closest values of all the pairwise comparisons, the resulting CDP matrices
may be inconsistent. The following paragraphs of this section discuss the
issue of the maximum consistency, denoted as CImax , of CDP matrices. The
following lemma provides an interesting result regarding the maximum
error omax associated with the pairwise comparisons of a CDP matrix. The
maximum error omax is defined as follows:
omax = max( eij - 1),
where: eij = ay(wj / w;), for any i, j = 1,2, 3, ... , n.
The au's are the entries of a pairwise matrix and Wi' Wj are the real weights
of the entities i and j, respectively.
Lemma 3-1:
Let a scale for quantifying pairwise comparisons be defined on the following
(2k+ 1) discrete values:
fVk' Vk _I, Vk _Z' ••• , Vz, VI' 1, 1lVI' IIVz, ... , IlVk _z, IIVk _I, IIVJ,
where V; > 0, for any i = 1, 2, 3, ... , k, and VI > 1.
Then, the maximum error, 0max' of the pairwise comparisons in a CDP
matrix is given by the formula:
V-V.
} }- I }
max { for j=1,2,3, ... ,k, and Vo 1.
°max
lJ + lJ-I '
Proof:
Suppose that a pairwise comparison has actual (and hence unknown) value
equal to ex, where: 1Il-j;;::: ex ;;::: 1Il-j_1' for some k ;;::: j ;;::: 1. Let M be the
middle point of the interval [1Il-j, lI"1-a. That is:
M =
lIlJ_I - lIV
--'---:::------..:..} +
1 V} + V}- I
2 lJ 2 V} V'} - I
Then, the largest 0 value for this particular pairwise comparison occurs
when the value of ex coincides with the middle point M. This is true
because in this case the closest value from the values permitted by the
current scale has the largest distance from ex. That is, under the assumption
that the decision maker will choose the closest value (i.e., he/she will be as
much accurate as possible), the value of this pairwise comparison will
become equal either to 1Il-j or to 11"1-1' In the first case the corresponding
ovalue, which we will call 0], becomes:
° = __
1
lIV.
M
J -1
VJ
V.J + V.J- I -1
VJ-I - VJ
VJ-I + VJ
2 V.J V·J- I
Similarly, in the second case the value of 0, which we will call ~, becomes:
v. - V. I
02 = J J-, that is: 1011 = 1021.
V.J- I + V.J
Since, in general, it is assumed that: lIVk ::;; O! ::;; Vk , it follows that the
maximum value of 0, denoted as 0max, is given by the following formula:
°mu
= max {
V.-V.
J J-
V.J + V.J- I
I }
'
for j = 1, 2, 3, ... , k, and Vo = 1.

Finally, it is worth stating here that both the expressions 1 and 02 remain
the same if the values ~ and ~.I are replaced by their reciprocals. •
°
In the previous considerations, and throughout this chapter, it is
assumed that the real values of the pairwise comparisons are within the
interval [Vk' 1/VJ. If, instead, the real ratios were allowed to be from the
range zero to infinity, then the associated errors could be infinitely large.
In other words, the real ratios are assumed to take on values according to
the scale under consideration.
Although this may appear to be restrictive, it eliminates the
possibility of having infinitely large errors when the decision maker attempts
to approximate pairwise comparisons by using a discrete and finite scale.
Furthermore, this is a plausible assumption since, most of the time, the
entities to be compared in the pairwise manner are assumed to be somehow
closely associated (i.e., similar) with each other and do not allow for
extreme cases. Therefore, it makes sense not to permit to have infinitely
large errors in the estimation process. Next, Lemma 3-1 is used to prove
Theorem 3-2 which deals with the value of the maximum CI index, denoted
as ClmilX , of random CDP matrices.
Theorem 3-2:
Let a scale for quantifying pairwise comparisons be defined on the following
(2k+ 1) discrete values:
[Vk' Vk _]' Vk-2' ••• , V2, V]' 1, 1IV], IIV2, •• ·, IlVk_2, IlVk_], IIVJ,
where V; > 0 for any i = 1, 2, 3, ... , k.
Then an upper bound of the maximum consistency index, denoted as Clmox,
of the resulting CDP matrices is given by the following relation:
fl
CI ~~,
max 2
V.-V. I }
where' 5 = max { j j-
'max V V'
j + j-I
for } = 1, 2, 3, ... , k, and Vo = 1.
Proof:
The proof of this theorem is based on Theorem 7-16, stated in [Saaty,
1980]. According to that theorem the following relation is always true:
~x - n ~ (n - 1)/2 02max, (3-1)
where 5max is defined as:
Omax = max {eij - I}, and eij = ay(w/w i ),
for any i,} = 1, 2, 3, ... , n.
The aij's are the entries of the pairwise (judgment) matrix and Wi' Wj are the
real weights of entities i and}, respectively. From relation (3-1), above, we
get:
Amax -n 02max
~ or:
n- 1 T'
(3-2)
o~x
C1max ~
T'
For the case of CDP matrices the value of the maximum 0, denoted as 5max ,
can be determined as follows (see also Lemma 3-1):
5 = max { V.-V.
j }-
I }
(3-3)
max V . + V. I '
} }-
for} = 1, 2, 3, ... , k and Vo = 1.

Therefore, the maximum consistency index, CI.nax, of CDP matrices satisfies
the relation:
O~ax
C1rnax ::::; 2'
where omax is given by the previous relation (3-3). •
In the original Saaty scale a pairwise comparison takes on values

from the discrete set: e = {9, 8, 7, ... , 3, 2, 1, 112, 113, ... , 117, 118,
1I9}. Therefore, it can be easily verified that omax is equal to 113 and thus
the following Corollary 3-1 is true when the original Saaty scale is used.
Corollary 3-1:
When the original Saaty scale is used, an upper bound of the maximum
consistency index, Clmax , of the corresponding CDP matrices is:
(113)2 1
C1max ::::; 2 18·
Figure 3-2 depicts the maximum, average, and minimum consistency

indexes of randomly generated CDP matrices which were based on the
original Saaty scale. That is, first a RCP matrix was randomly generated.
Next, the corresponding CDP matrix was derived and its CI value was
calculated and recorded (see also [Triantaphyllou, et al., 1990]). This
experiment was performed 1,000 times for each value of n equal to 3,4,5,
... , 100. It is interesting to observe that the curves which correspond to the
maximum and minimum CI values of samples of 1,000 randomly generated
CDP matrices, are rather irregular. This was anticipated since it is very
likely to find one extreme case from a sample of 1,000 CI values of random
CD P matrices.
One the other hand, however, the middle curve, which depicts the
average CI values of random CDP matrices, is very regular. This was also
anticipated because the impact of a few extreme CI values diminishes when
a large sample (such as the ones we used of size 1,000) of random CDP
matrices is considered. Moreover, the same results indicate that the average
CI value approaches the number 0.0145 when the value of n is greater than
20. More on the CI values of random Saaty matrices (i.e., not necessarily
CDP matrices) can be found in [Donegan and Dodd, 1991].
0.030
0.028
0.026
0.024
0.022
0.020
0.018
-
rI.I 0.016
Average
=
~
;;;. 0.014
CIS
1-1
U 0.012 Minimum
0.010
0.008
0.006
0.004
0.002
O.OOO~.~-~~~~~~~~~r-~r-~r-~r-~r-r-o
o 10 20 30 40 50 60 70 80 90 100
Order of Set
Figure 3-2: Maximum, Average, and Minimum CI Values of Random CDP Matrices
When the Original Saaty Scale is Used.
The results in this section reveal that CDP matrices (which are
assumed to be the result of a highly effective elicitation of the pertinent
pairwise comparisons) are very unlikely to be perfectly consistent. That
is, some small inconsistency may be better than no inconsistency at all!
(since no CDP matrix with CI is equal to zero was found when sets with
more than five elements were considered). This counter-intuitive situation
forms another paradoxical phenomenon which, however, can be explained
theoretically by the lemmas and theorems in this section.
3.3.3 Two Evaluative Criteria

In [Triantaphyllou and Mann, 1990], the evaluation of the
effectiveness of Saaty's eigenvalue method was based on a continuity
assumption (see also Chapter 10). Under that continuity assumption it turns
out that the eigenvalue approach may cause worse alternatives to appear to
be better than alternatives that are truly better in reality.
In that study two kinds of ranking inconsistency were examined.
The first kind was "ranking reversal". For example, if the real ranking of
a set of three entities was equal to (1, 3, 2) and a method yielded (1, 2, 3),
then a case of a ranking reversal was said to had been occurred. The
second kind was "ranking indiscrimination". For example, if the real
ranking of a set of three entities was equal to (1,3,2) and a method yielded
(1, 2, 2), that is, a tie between two or more entities, then a case of ranking
indiscrimination had occurred. In order to examine the effectiveness of
various scales the concept of the CDP matrices can be used. That is, the
ranking implied by a CDP matrix (which, as mentioned in the previous
section, represents the best decisions that a decision maker can make) has
to be identical with the actual ranking indicated by the corresponding RCP
matrix. Therefore, the following two evaluative criteria can be introduced
to investigate the effectiveness of any scale which quantifies pairwise
comparisons:
First Evaluative Criterion:

Consider the ranking derived when an RCP matrix with the actual values of
the pairwise comparisons of n entities is considered. Next consider the
ranking derived when the corresponding CDP, a scale, and some methodfor
extracting the relative weights, are considered. Then, these two rankings
should do not demonstrate any ranking inversions.
Second Evaluative Criterion:

Consider the ranking derived when an RCP matrix with the actual values of
the pairwise comparisons of n entities is considered. Next consider the
ranking derived when the corresponding CDP, a scale, and some methodfor
extracting the relative weights, are considered. Then, these two rankings
should do not demonstrate any ranking indiscriminations.
Since the previous two ranking anomalies are independent of the

scale under consideration or the method used to process matrices with
pairwise comparisons, the previous two evaluative criteria can be used to
evaluate any scale and method.
3.4 A SIMULATION EVALUATION OF DIFFERENT

SCALES
In [Triantaphyllou. Lootsma. et al .• 1994] a number of different

scales were evaluated by generating test problems and then recording the
inversion and indiscrimination rates as described in the two evaluative
criteria presented in the previous section. Suppose that a scale defined on
the interval [9. 119] (as described in Section 3.2.1) or an exponential scale
(as described in Section 3.2.2) is defined on the interval [X, 1IX]. That is,
the numerical value that is assigned to a pairwise comparison that was
evaluated as: "A is absolutely more important than B" (Le .• the highest
value) is equal to X. For instance. in the original Saaty scale (as well as in
all the other scales in Section 3.2.1) X equals to 9. Under the assumption
that a scale on the interval [X. 1/X] is used, the pairwise comparisons also
take numerical values from the interval [X. 1IX]. In this case the entries of
RCP matrices (as defined in Section 3.2.1) are any numbers from the
interval [X. 1IX]. However, in CDP matrices the entries take on values
only from the discrete and finite set that is defined on the interval [X. 1IX].
As before. we will denote this as set e. For example. in the case of the
original Saaty scale the entries of CDP matrices are members of the set e
= {9. 8.7 ..... 2. 1. 112 ..... 117, 118. 1I9}.
For the above reasons test problems for the case of the first and
second evaluative criterion were generated as follows. First, n weights of
n entities were randomly generated from the interval [1. 0]. These weight
values were such that no ratio of any pair of them would be larger than X
or less than 1IX. After the random weight values were generated. the
corresponding RCP matrix was constructed. Next. from the RCP matrix
and the discrete and finite set e the corresponding CDP matrix was
determined. Then. the eigenvalue approach was applied on this CDP matrix
and the new ranking of the n elements was derived.
The eigenvalue method was used because it is rather simple to apply
and is the method used widely in the literature when only one decision
maker is considered. The recommended ranking of the n elements was
compared with the actual ranking which is determined from the real weight
values that were generated randomly at the beginning of this process. If a
ranking inversion or ranking indiscrimination was observed. it was recorded
so. This is exactly the testing procedure followed in the investigation of the
original Saaty scale as it was reported in [Triantaphyllou and Mann, 1990].
A FORTRAN program was written which generated for each test
problem the n random weight values. the RCP and CDP matrices, and
compared the two rankings as described above. Sets with n = 3, 4. 5 .....
30 elements were considered. For each such set 21 scales defined on the
interval [9, 1/9] (which correspond to ex values equal to 0, 5, 10, 15, ... ,
90, 95, 100) and 57 exponential scales which correspond to 'Y values equal
to 0.02, 0.04, 0.06, ... , 1.10, 1.12, 1.14 were generated. The previous 78
(Le., 21 + 57) scales will also be indexed as scale #1, scale #2, scale #3,
... , scale #78.
In Figures 3-3 and 3-4 the results of the evaluations of scales #1,
#2, #3, ... , #21 (also called Class 1 scales) in terms of the first and second
evaluative criterion, respectively, are presented. Similarly, in Figures 3-5
and 3-6 the results of the evaluations of scales #22, #23, #24, ... , #78 (also
called Class 2 scales) in terms of the first and second evaluative criterion,
respectively, are presented. It should be noted here that only 57 exponential
scales were generated because in this way values of 'Y from zero to around
to 1.00 can be considered. In the original Lootsma scales the value of 'Y
was 0.50 and 1.00. In this investigation all the scales with 'Y = 0.02, 0.04,
0.06, ... , 0.50, ... , 1.00, ... , 1.14 are considered. For each case of a value
of n and one of the 78 scales, 1,000 random test problems were generated
and tested according to the procedure described in the previous paragraphs.
The computational results of this investigation are depicted in Figures 3-5
and 3-6.
At this point it should be emphasized that the present simulation
results are contingent on how the random weight values were generated.
Other possibilities, such as assigning weight values from a nonuniform
distribution (such as the normal distribution), would probably favor other
scales. However, the uniform distribution from the interval [1, 0] was
chosen in this study (despite the inherited restrictions of this choice) because
it is the simplest and most widely statistical distribution used in similar
simulation investigations.
46 MCDM Method: A Comparative Study. by E. TriantaphylLou
Inversion
100 Rate (%) Scale 1
~-s~~::: Scale 2
90
80
70
Scale 20
60
Scale 21
50
40
30
20
10
Size of Set
Figure 3-3: fuversion Rates for Different Scales and Size of Set
(Class 1 Scales).
Indiscrimination
100 Rate (%)
90
80
70
60
50
40
30
20
10
Size of Set
Figure 3-4: Indiscrimination Rates for Different Scales and Size of Set
(Class 1 Scales).
Inversion
Rate (%)
:l~~I~~~~~~ Scale 7778
100
"~
90
80
70
60
50
40
30
20
Scale 23
10 Scale 22
5 10 15 20 25 30
Size of Set
Figure 3-5: Inversion Rates for Different Scales and Size of Set
(Class 2 Scales).
Indiscrimination
100 Rate (%)
90
Scale 22
80
Scale 23
70
60
50
40
30
20
10
10 15 20 25 30
Size of Set
Figure 3-6: Indiscrimination Rates for Different Scales and Size of Set
(Class 2 Scales).
3.5 ANALYSIS OF THE COMPUTATIONAL RESULTS
Figures 3-3, 3-4, 3-5, and 3-6 depict how the previous 78 different
scales perform in terms of the two evaluative criteria given in Section 3.3.3.
Figures 3-3 and 3-4 depict the inversion and indiscrimination rates (as
derived after applying the two evaluative criteria) for Class 1 scales. That
is, for the scales defined in the interval [9, 1/9]. Similarly, Figures 3-5 and
3-6 depict the inversion and indiscrimination rates for the exponential scales
(or Class 2 scales). It is also interesting to observe here that when scales
from both classes of scales are evaluated in terms of the second evaluative
criterion (see also the indiscrimination rates in Figures 3-4 and 3-6), then
they perform worse when the size of the set with the entities to be compared
is between 8 and 12.
Clearly, there is no single scale which outperforms all the other
scales for any size of set. Therefore, there is no single scale, or a group of
scales, which is better than the rest of the scales in terms of both evaluative
criteria. However, the main problem is to determine which scale or scales
are more efficient.
Since there are 78 different scales for which there are relative
performance data in terms of two evaluative criteria, it can be concluded
that this is a classical multi-criteria decision making problem. That is, the
78 scales can be treated as the alternatives in this decision making problem.
The only difficulty in this consideration is how to assess the weights of
importance of the two evaluative criteria. Which evaluative criterion is the
most important one? Which is the least important? Apparently, these type
of questions cannot be answered in a universal manner.
The weights for these criteria depend on the specific application
under consideration. For instance, if ranking indiscrimination of the entities
to be compared is not of main concern to the decision maker, then the
weight of the ranking reversals should assume its maximum value (i.e.,
becomes equal to 1). However, one may argue that, in general, ranking
indiscrimination is less severe than ranking reversal. Depending on how
more critical ranking reversals are, one may want to assign a higher weight
to the ranking reversal criterion. If both ranking reversal and ranking
indiscrimination are equally severe, then the weights of the two criteria are
equal (i.e., they are set equal to 0.50).
For the above reasons, the previous decision making problem was
solved for all possible weights of the two evaluative criteria. Criterion 1
was assigned weight WI while the second evaluative criterion (or Criterion
2) was assigned weight w2 = 1 - WI (where 1 ;:::: WI ;:::: 0). In this way, a
total of 100 different combinations of weights were examined.
Scale No Description Scale No Description
~
19 Class 1
with a = 90 II 69 Class 2
with y = 0 .96
21 Class 1 77 Class 2
Weight Wi
~ with a = 100 with y = 1.12
22 Class 2
with y = 0.02 II 78 Class 2
with y = 1.14
24 Class 2
with y = 0.06 D other scales
Size of Set
Figure 3-7: The Best Scales.

Scale No Description
II 78 Class 2
with y = 1.14 Scale No Description
Weight WI
77 Class 2
with y '" 1.12 lID 22 Class 2
with y = 0.02
23 Class 2
~
75 Class 2
with y = 1.08 with y = 0.04
I@I 24 Class 2
with y = 0.06 D other scales
1.00
0.80
0.60
0.40
o.zo
0.00
o 10 15 20 25 30
Size of Set
Figure 3-8: The Worst Scales.

For each one of these combinations of the weights of importance of

the two evaluative criteria, the decision making problem was solved by
using the revised AHP (as introduced by Belton and Gear [1983], see also
Section 2.2.4). In [Triantaphyllou and Mann, 1989] (or in Chapter 9) the
revised AHP was found to perform better when it was compared with other
MCDM methods. For each of the above decision making problems the best
and the worst alternative (i.e., a scale) was recorded.
The results regarding the best scales are depicted in Figure 3-7.
Similarly, the results regarding the worst scales are depicted in Figure 3-8.
In both cases the best or worst scales are given for different values of the
weight of the first evaluative criterion (or equivalently, the second evaluative
criterion) and the size of the set.
The computational results demonstrate that only very few scales can
be classified either as the best or the worst scales. It is possible that the
same scale (for instance, scale #78) to be classified as one of the best scales
for some values of the weight WI and also as the worst scale for other values
of the weight WI' Probably, the most important observation is that the
results illustrate very clearly that there is no single scale which is the best
scale for all cases. Similarly, the same results also illustrate that there is
no single scale which is the worst scale for all cases.
However, according to these computational results, the best scale
can be determined only if n (i.e., the size of the set) is known and the
relative importance of the weights of the two evaluative criteria has been
assessed. It is also interesting to observe from Figure 3-7 that sometimes
under similar weights of the two evaluative criteria, the same scale might be
classified as the best. The same is also true for the worst scales depicted in
Figure 3-8. This phenomenon suggests that sometimes an approximated
assessment of the relative weights is adequate to successfully determine
either the best or worst scale.
3.6 CONCLUSIONS
Both the theoretical and empirical analyses described in this chapter
revealed that the scale issue is a complex one. The various theoretical and
empirical results demonstrated that there is no single scale which can always
be classified as the best scale or as the worst scale for all cases. The
present investigation is based on the assumption that there exists a
real-valued rating of the comparison between two entities, that ideally
represents the individual preference. However, the decision maker cannot
express it, hence he/she has to use a scale with finite and discrete options.
In order to study the effectiveness of various scales, we furthermore
assumed the scenario in which the decision maker is able to express his/her
judgments as accurately as possible. Under this scenario, it is assumed that
the decision maker is able to construct CDP matrices with pairwise
comparisons instead of the unknown RCP matrices. Based on this setting,
a number of computational experiments was performed to study how the
ranking derived by using CDP matrices differs from the real (and hence
unknown) ranking implied by the RCP matrices. The computational results
revealed that there is no single scale which is best in all cases. It should be
emphasized here that given an RCP matrix (and a scale with numerical
values), then there is one and only one CDP matrix which best
approximates it. Moreover, this CDP matrix mayor may not yield a
ranking different than the ranking implied by the RCP matrix.
An alternative assumption to the current one, which accepts that
there exists a real-valued rating of the comparison between entities, is to
consider the premise that maybe the real entity is the CDP matrix as given
by the decision maker. In this case the RCP matrix is maybe just an
illusion. In the later case the preference reversal leads to a very different
conclusion: if the CDP is the only "real thing", then it means that the
individual should point at the interval [1/~, l/Vi_l ] or [~-l' ~] rather than
to the values V;. That is, the preference reversal effects indicate that two
entities will be indifferent (since their ranking changes in the interval).
To determine the appropriate scale in a given situation certain
factors will have to be analyzed. First the number n, of the entities to be
compared, has to be known. Secondly, the relative importance of the two
evaluative criteria has to be assessed. These evaluative criteria deal with
possible ranking inversions and ranking indiscriminations that may result
when a scale is used. When these factors have been assessed, Figure 3-7
depicts the best scale for each case. Similarly, Figure 3-8 depicts the worst
scale for each case.
For instance, suppose that one has to evaluate the weight values of
a set with 15 members. Furthermore, suppose that ranking reversal is
considered, in a particular application, far more severe than ranking
indiscrimination. In other words, the weight of the first evaluative criterion
is considered to be higher than the weight of the second criterion. Using
this information, we can see that Figure 3-7 suggests to use scale #22 from
Class 2 (i.e., an exponential scale with parameter 'Y = 0.02). Moreover,
Figure 3-8 suggests that the worst scale for this case is scale #77 from Class
2 (i.e., an exponential scale with parameter 'Y = 1.12).
The same figures also indicate that the choice of the best or worst
scale is not clear under certain conditions. For instance, when the number
of members is greater than 15 and the two evaluative criteria are of almost
equal importance. In cases like this, it is recommended to experiment with
different scales in order to increase the insight into the problem, before
deciding on what is the best scale for a given application.
The computational experiments in this chapter indicate (as shown in
Figure 3-7) that exponential scales are numerically more efficient (i.e., more
stable) than the original Saaty scale (i.e., Scale 1). Only two Saaty-based
scales (i.e., scales #19 and #21) are present in Figure 3-7. In a matter of
fact, for sets with up to 10 entities scale #21 was best over a wide range of
weights. It is also worth noting that all the worst scales in Figure 3-8 came
from the exponential class (i.e., Class 2).
However, as the various examples in Section 3.3 suggest, human
beings seem to use exponential scales in many diverse situations. Therefore,
exponential scales appear to be the most reasonable way for quantifying
pairwise comparisons. The computational results in this chapter provide a
guide for selecting the most appropriate exponential scale for quantifying a
given set of pairwise comparisons.
Finally, it needs to be emphasized here that the scale problem is a
very crucial issue when the data are qualitative and thus they need to be
determined by using pairwise comparisons. Such qualitative data can
capture the pertinent information for many real life decision making
problems. An alternative point of view of the study described in this
chapter would be to perform in the future a similar investigation with
methods which do not use pairwise comparisons and thus are counterparts
of the pairwise comparison methodologies. However, since pairwise
comparisons provide a flexible and also realistic way for estimating these
type of qualitative data, it follows that an in depth understanding of all the
aspects of the scale problem is required for a successful solution to an
MCDM problem.
Chapter 4
DERIVING RELATIVE WEIGHTS FROM RATIO

COMPARISONS
As it was mentioned in Chapter 3, an important issue in MCDM

methods is to be able to determine the relative weights of importance of a
collection of entities (such as the alternatives to be studied in terms of a single
decision criterion). This task is similar and closely related to the problem of
determining the degree of membership of the elements of a fuzzy set.
Usually, such values are between 0 and 1 and they add up to 1. Such weights
of degrees of membership are supposed to be a good model of the way people
perceive categories [Dubois and Prade, 1980]. Often, the most representative
members in the set are assigned to the value of 1 and non-members to the
value of O. Then, the main problem is to determine the degree of
membership (i.e., a number between 0 and 1) of the between members.
Psychologists [Lakoff, 1973] have found that people can easily identify
representative members in a fuzzy set, while they have difficulties in
identifying the other members. The importance of evaluating the membership
degrees in applications of fuzzy set theory in engineering and scientific fields
is best illustrated in the more than 1,800 references given in [Gupta, et al.,
1979] (see also Chapter 12 for discussions on some related problems).
As it was described in the previous chapter, qualitative information
for MCDM problems can be captured by means of pairwise comparisons.
Next, one needs to process these pairwise comparisons and extract the implied
relative weights of importance of the compared items. Saaty [1977; 1980;
and 1994] has suggested a possible solution to the problem of finding the
relative weights of importance from pairwise comparisons. His method is
based on eigenvalue theory. In both [Chu, et al., 1979] and [Federov, et ai.,
1982] Saaty's method has been viewed as a modified least squares problem.
In this chapter a new least squares approach is used on the data derived from
the pairwise comparisons as proposed by Saaty. That method is based on the
approach presented in [Triantaphyllou, Pardalos, and Mann, 1994]. This
approach uses an error minimization function that attempts to reflect the real
intentions of the decision maker while eliciting the pairwise comparisons.

4.2 THE EIGENVALUE APPROACH

Let AI' A 2, ... , An be n entities (concepts, alternatives, or criteria) to
be compared. We are interested in evaluating the relative weights of the
above entities when they are compared with each other in terms of a single
common characteristic. As described in Chapter 3, Saaty [1977; 1980; and
1994] proposed to use a matrix A of rational numbers taken from the set {9,
8, 7, ... , 2, 1, 112, ... , 117, 118, 1I9}. Each entry of the above matrix A
represents a pairwise comparison (judgment). Specifically, the entry aij
denotes the number that estimates the relative importance of element Ai when
it is compared with element Aj • Obviously, aij = lIaji and au = 1. That is,
this matrix is a reciprocal one.
Let us first examine the case in which it is possible to have perfect aij
values. In this case it is aij = w/Wj (where wk denotes the actual weight of
importance of element Ak ) and the previous reciprocal matrix A is consistent.
That is:
aij = ail< x akj' for i, j, k = 1, 2, 3, ... , n, (4-1)
where n is the number of elements to be compared. It can be proved [Saaty,
1980] that A has rank 1 and A. = n is its nonzero principal right eigenvalue.
Thus, we have:
A x = n x, (4-2)
where x is a principal right eigenvector. From the fact that aij = w/Wj' the
following are obtained:
for i = 1, 2, 3, ... , n, (4-3)

or: A W = n W. (4-4)
Equation (4-4) states that n is a right eigenvalue of A with W a
corresponding principal right eigenvector. The same equation also states that
in the perfectly consistent case (i.e., when aij = ail< x akj)' the vector W, with
the relative weights of the elements 1, 2, 3, ... , n, is a principal right
eigenvector of matrix A.
In the non-consistent case (which is more common in real life
applications) the pairwise comparisons are not perfect, that is, the entry aij
might deviate from the real ratio w/Wj (i.e., from the ratio of the real weight
values Wi and wj ). In this case, the previous expression (4-1) does not hold
for all possible combinations. Now the new matrix A can be considered as
a perturbation of the previous consistent case. When the entries aij change
slightly, then the eigenvalues change in a similar fashion [Saaty, 1980].
Moreover, the maximum eigenvalue is close to n (greater than n) while the
Chapter 4: Deriving Relative Weights from Ratio Comparisons 59
remaining eigenvalues are close to zero. Thus, in order to find the weights
in the non-consistent case, one should find an eigenvector that corresponds to
the maximum right eigenvalue Amax. That is, to find the principal right
eigenvector W that satisfies:
A W = Amax W, where Amax ~ n.
Saaty proposed to estimate the reciprocal right eigenvector W by
multiplying the entries in each row of matrix A together and taking the n-th
root (n being the number of the elements to be compared). Since we desire
to have values that add up to 1, one needs to normalize the previously found
vector by the sum of the above values. If one wants to have the element with
the highest value to have a weight value equal to 1, then he/she needs to
divide the previously found vector by the highest value.
Under the assumption of total consistency, if the judgments are
gamma distributed (something that Saaty claims to be the case), the principal
right eigenvector of the resultant reciprocal matrix A is Dirichlet distributed.
If the assumption of total consistency is relaxed, then Vargas [1982] proved
that the hypothesis that the principal right eigenvector follows a Dirichlet
distribution is accepted if the consistency ratio is equal to or less than 10%.
The consistency ratio (CR) is obtained by first estimating Amax. Saaty
proposed to estimate the value of >--max by adding the columns of matrix A and
then multiplying the resulting vector with the vector W. Then, he uses what
he calls the consistency index (Cn of the matrix A. He defined CI as
follows:
CI = (>--max - n)/(n -1).
Then, the consistency ratio CR is obtained by dividing the CI by the
Random Consistency Index (RCI) as given in Table 4-1. Each RCI is an
average random consistency index derived from a sample of size 500 of
randomly generated reciprocal matrices with entries from the set {9, 8, 7, ... ,
2, 1, 1/2, ... , 117, 1/8, 1/9} to see if its CI is 10% or less. The number of
500 replications was used due to the limitations of the early computers when
the above concepts were established by Saaty. If the previous approach yields
a CR greater than 10 %, then a re-examination of the pairwise judgments is
recommended until a CR less than or equal to 10% is achieved.
Table 4-1: RCI Values of Sets of Different Order n.
n 1 2 3 4 5 6 7 8 9
ReI 0 0 0.58 0.90 1.12 1.24 1.32 1.41 1.45

60 MCDM Methods: A Comparative Study. by E. Triantaphyllou
4.3 SOME OPTIMIZATION APPROACHES

Chu, et at. [1979] claimed that given the aij data, the values Wi to be
estimated are desired to have the property:
aij .. Wi/Wi" (4-5)

This is reasonable since aij is meant to be the estimation of the ratio w/Wj.
Then, in order to get the estimates for the Wi given the aij data, they proposed
the following constrained optimization problem:
n n
min S = L L (aijwj -
i=j j=j
wji,
n
S.t. LWj = 1,
1=1
and Wi > 0, for i = 1, 2, 3, ... , n. (4-6)

The same authors also provided an alternative expression SI that is more
difficult to solve numerically. That expression was as follows:
n n
SI = LL(aij - wiwY· (4-7)
;=1 j=1
In [Federov, et at., 1982] a variation of the above least squares

formulation was proposed. For the case of only one decision maker they
recommended the following models:
and aij = wdWj + 1j12(Wj , wj)e ij' (4-9)

In the previous two expressions Wi and Wj are the true (and hence unknown)
weight values; vl(X, Z) andv2(X, Z) are given positive functions (where X,
Z> 0). The random errors eij are assumed to be independent with zero mean
and unit variance. Using these two assumptions they were able to calculate
the variance of each individual estimated weight value. However, they fail to
give a way of selecting the appropriate positive functions. In the second
example, presented later in Section 4.6, a sample problem which originates
in [Saaty, 1977] and later in [Federov, et al., 1982] is solved for different VI
and '112 functions by using the Federov, et al., method.
4.4 CONSIDERING THE HUMAN RATIONALITY

FACTOR
According to the Human Rationality assumption [Triantaphyllou,
Pardalos, and Mann, 1990] a decision maker is a rational person. Rational
persons are defined here as individuals who try to minimize their regret
[Simon, 1961], to minimize losses, or to maximize profit [Write and Tate,
1973]. In the relative weight evaluation problem, minimization of regret,
losses, or maximization of profit could be interpreted as the effort of the
decision maker to minimize the errors involved in the pairwise comparisons.
As it was stated in the previous paragraphs, in the inconsistent case
the entry aij of the matrix A is an estimation of the real ratio w;lwj' Since it
is an estimation, the following is true:
aij = (wtfw}dij' for i, j = 1, 2, 3, ... , n. (4-10)

In the above relation dij denotes the deviation of aij from being an accurate
judgment. Obviously, if dij = 1, then the aij was perfectly estimated. From
the previous formulation we conclude that the errors involved in these
pairwise comparisons are given by:
t ij = dij - 1,
or after using expression (4-10), above:
tij = aij(wiwj) - 1. (4-11)

As it was described in Chapter 3, when the set of the entities to be
compared contains n elements, then Saaty's method requires the estimation of
the following n(n - 1)/2 pairwise comparisons:
W2 /W1' W3/ Wl' W4/Wl' ... , w,a!wl'

W3/ W 2' W 4 /W2' ... , wll /w2 '
W 4 /W3' ... , wll /w3 ,
(4-12)
wll _ 1/wll •
Therefore, the corresponding n(n - 1)/2 errors are (after using
relations (4-11) and (4-12»:
(4-13)
for i,} = 1,2,3, ... , n, and} > i.

Since the w;'s are relative weights of importance that (usually) add up to 1,
the following relation (4-14) should also be satisfied:
n
(4-14)
Apparently, since the Wi weights of importance are positive numbers,

we should also have:
Wi > 0, for i:: 1, 2, 3, ... , n.
Relations (4-13) and (4-14), when the data are consistent (i. e., all the errors
are equal to zero), can be written as follows:
B W = b. (4-15)
The vector b has zero entries everywhere except the last one that is equal to
1, and the matrix B (on the next page) has the following form (please notice
that blank entries represent zeros. Also, the first row and last column
indicate the cell numbers of the rows and columns, respectively, and are not
part of the matrix):
1 2 3 4 5 6 7 n-1 n
-1 a 1,2 1
-1 a 1,3 2
-1 a 1,4 3
-1
1
B 2
3
a2,n-l
-1 a 2 ,n n-2
-1 an-1,n 1
1 1 1 1 1 1 1 . . . 1 1
The error minimization issue is interpreted in many cases (e. g. ,

regression analysis, linear least squares problem) as the minimization of the
sum of squares of the residual vector: r = b - B W (see, for example,
[Stewart, 1973]). In terms of formulation (4-15) this means that in a real life
situation (i.e., when errors are not zero any more) the real intention of the
decision maker is to minimize the expression:
(4-16)
which, apparently, expresses a typical linear least squares problem.
If we use the notation described previously, then the quantity S in
expression (4-6) which is minimized in [Chu, et al., 1979] becomes:
n n n n
S ::; L L (aijwj - w i )2 LL (eijw;)2.

i=1 j=1 i=1 j=1
and the alternative expression (4-7) becomes:
n n n n
S. ::; L L (aij - Wj/wi

;=1 j=1
Clearly, both expressions are too complicated to reflect in an intuitive way the
intentions of the decision maker.
The optimization models proposed by Federov, et al. [1982] are
closer to the one developed under the human rationality assumption. The only
difference is that instead of the relations:
and aij ::; wJwj + W2(Wi , w)e ij •

the following simpler expression is used:
or:
(4-17)
However, as the second example illustrates, the performance of the Federov,
et al., method is greatly dependent on the selection of the WI or w2 functions.
Now, however, these functions are further modified by (4-17).
4.5 FIRST EXTENSIVE NUMERICAL EXAMPLE

Let us assume that the following is the matrix with the pairwise
comparisons for a set of four elements:
1 2/1 1/5 1/9

1/2 1 1/8 1/9
A
5/1 8/1 1 1/4
9/1 9/1 4/1 1
By using the numerical procedures presented in the previous sections we can

see that:
Amax 4.226,
Cl = (4.226-4)/(4-1) = 0.053,
CR Cl/0.90 = 0.0837 < 0.10.
The formulation (4-15) that corresponds to this example is as follows:
-1 2/1 0.0 0
-1 0.0 1/5 0 0
Vl 0
-1 0.0 0 1/9
V2 0
0.0 -1 1/8 0 x = 0
V3 0
0.0 -1 0 1/9 V4 0
0.0 0.0 -1 1/4 1
1 1 1 1
The vector V that solves the above least squares problem can be shown to be
as follows:
V = (0.065841, 0.039398, 0.186926, 0.704808).
Hence, the sum of squares of the residual vector components is 0.003030.
The average squared residual for this problem is 0.003030/«4 x(4 - 1)/2)+ 1)
= 0.000433; that is, the average residual is V(0.000433) = 0.020806.

4.6 SECOND EXTENSIVE NUMERICAL EXAMPLE
The second extensive numerical example uses the same data used
originally in the paper by Saaty [1977], and later in the two papers by Chu,
et al. [1979] and Federov, et al. [1982]. These data are next presented in
Table 4-2.
Table 4-3 presents a summary of the results (as found in the
corresponding references) when the methods described in Sections 4.2, 4.3,
and 4.4 are used. The power method for deriving the eigenvector was applied
as presented by Kalaba, et al. [1979]. In the last row of Table 4-3 are the
results obtained by using the least squares method under the human rationality
assumption (HR).
As it is shown in the last column of Table 4-3, the performance of
each method is very different as far the mean residual is concerned. The
same results also illustrate how critical is the role of the functions it,(X, Z)
and itiX, Z) in Federov's, et al. [1982] method. The mean residual obtained
by using the least squares method under the human rationality assumption is,
as expected, the smallest one (actually by 16%).
Table 4-2: Data for the Second Extensive Numerical Example.
itemlitem (1) (2) (3) (4) (5) (6) (7)
(1) 1 4 9 6 6 5 5
(2) 114 1 7 5 5 3 5
(3) 119 117 1 115 115 117 115
(4) 116 115 5 1 1 1/3 1/3
(5) 116 115 5 1 1 1/3 113

(6) 115 113 7 3 3 1 2
(7) 115 1/4 5 3 3 112 1

Table 4-3: Comparison of the Weight Values for the Data in Table 4-2.
Element's (Item) ID No. in Set

Average
(1) (2) (3) (4) (5) (6) (7) residual
Method used
Saaty's eigenvector 0.429 0.231 0.021 0.053 0.053 0.119 0.095 0.134
method
Power method's 0.427 0.230 0.021 0.052 0.052 0.123 0.094 0.135
eigenvector
Chu's, et aI., method 0.487 0.175 0.030 0.059 0.059 0.104 0.085 0.097
Federov's, et al., Model 0.422 0.232 0.021 0.052 0.052 0.127 0.094 0.138
1 with 'lr, = 1
2 with 'lr2 = 1
2 with 'lr2 = I w;- Wj I
2 with 'lr2 = (w; / w)
Least squares method 0.408 0.147 0.037 0.054 0.054 0.080 0.066 0.082
under the HR assumption
4.7 AVERAGE ERROR PER COMPARISON FOR SETS

OF DIFFERENT ORDER
In this part of this study we generated random reciprocal matrices

with pairwise comparisons of different order and Consistency Index (CI)
values. For each test problem the least squares problem, as derived under the
human rationality assumption (as described in Section 4.4), was solved and
the average residual was recorded. The same problem was also solved by
using the eigenvalue method. The average residual was considered here as
an indicator of the effectiveness in estimating successfully the relative weights
from a set of pairwise comparisons. The simulation program was written in
FORTRAN and the resulting least squares problems were solved using the
appropriate IMSL subroutines. The results are presented in Table 4-4 and are
also depicted in Figures 4-1 and 4-2.
For each case the number of the generated random matrices was
varying but large enough to ensure representative results. It was observed
that for large matrices with small CI values the sample size could be small
(less than 100) and still achieve statistically significant results. However, the
opposite is true for small matrices with large CI values. In Figures 4-1 and
4-2 the vertical axis represents the average residual while the horizontal one,
the consistency index (CI). The CI index was selected such that the CR index
would be equal to 0.02, 0.03, 0.04, ... , 0.10. Due to the difficulty in
obtaining large random matrices with the CI (or the corresponding CR) value
being very small, the Supercomputer facilities at the Cornell University were
used (those experiments took place in 1988). Figure 4-1 depicts average
residuals when the human rationality assumption is used, while Figure 4-2
depicts residuals when the eigenvector method is used.
As the plots in Figures 4-1 and 4-2 illustrate, the average residual is
a function of both the consistency index (CI) of the data as well as the order
of the input matrix (Le., the number of elements to be compared in the
pairwise manner). Regression analyses suggested that for sets of a given
order the average residual is linearly related to the consistency index (CI).
Also, as the CI index of the data decreases so does the average residual. This
is expected because as the CI reaches zero the input data tend to be perfectly
consistent. From these results it can be seen that the average residuals are
significantly smaller when the method that is based on the human rationality
assumption is used.
The same plots also illustrate the importance of the number of the
elements to be compared. For small sets the CI index has to be small in
order to keep the average error at a low level. However, as the size of the
sets increases the CI index can be relatively larger and still achieve small
average errors per comparison. That is, the input data in the case of large
sets can be more inconsistent than in the case of smaller sets, as far the error
per comparison issue is concerned.
Table 4-4: Average Residual and CI Versus Order of Set and CR When the Human
Rationality (HR) Assumption and the Eigenvalue Method (EM) are Used.
The Results Correspond to 100 Random Observations.
Corresponding CR Index
Order
of
Method 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10
Set
Used
CI 0.0264 0.0396 0.0528 0.0660 0.0792 0.0924 0.1056 0.1188 0.1320
7 EM (.) (.) (.) (.) 0.6446 0.6674 0.8497 0.8930 0.9414
HR (.) (.) (*) (*) 0.5094 0.5279 0.5560 0.6024 0.6112
CI 0.0248 0.0372 0.0496 0.0620 0.0744 0.0868 0.0992 0.1116 0.1240
6 EM 0.4049 0.5546 0.6336 0.7100 0.7798 0.8369 0.9347 0.9681 1.0554
HR 0.2982 0.4035 0.4467 0.4615 0.5240 0.5560 0.5956 0.6127 0.6503
CI 0.0224 0.0336 0.0448 0.0560 0.0672 0.0784 0.0896 0.1008 0.1120
5 EM 0.4957 0.6192 0.7598 0.7927 0.8682 0.9769 1.0303 1.0984 1.2097
HR 0.3382 0.4276 0.4912 0.4934 0.5575 0.6030 0.6334 0.6742 0.6913
CI 0.0180 0.0270 0.0360 0.0450 0.0540 0.0630 0.0720 0.0810 0.0900
4 EM 0.5420 0.6629 0.7787 0.8534 0.9779 1.0696 1.1388 1.2451 1.2759
HR 0.3509 0.4193 0.4883 0.5284 0.5777 0.6323 0.6807 0.7252 0.7340
CI 0.0116 0.0174 0.0232 0.0290 0.0348 0.0406 0.0464 0.0522 0.0580
3 EM 0.5005 0.6690 0.7847 0.8196 0.9979 1.0486 1.1012 1.2225 1.3060
HR 0.2869 0.3554 0.3836 0.4785 0.4998 0.5535 0.6190 0.6215 0.6449
(*) Values were not found for these cases due to the excessive requirement for CPU time.
0.8
n=4
0.7
0.6
0.5
Average 0.4
Residual
0.3
0.2
0.1
0.00 0.03 0.06 0.09 0.12 0.15
CI Values
Figure 4-1: Average Residual and CI Versus Order of Set When the Human Rationality
Assumption is Used (the Results Correspond to 100 Random Observations).
1.4
1.2
1.0
n=7
0.8
Average
Residual
0.6
0.4
0.2
0.00 0.03 0.06 0.09 0.12 0.15
CI Values
Figure 4-2: Average Residual and CI Versus Order of Set When the Eigenvalue
Method is Used (the Results Correspond to 100 Random Observations).
4.8 CONCLUSIONS
The eigenvalue method, as proposed by Saaty, for evaluating the
relative weights of entities that share some mutual characteristic of interest has
captured the interest of many researchers. However, the findings of the
present chapter demonstrate that there is a new dimension to the weight
estimation problem. Although the CR value can be kept less than 10% (and
hence ensure satisfaction of the Dirichlet distribution criterion as described
previously) the mean residual can vary significantly.
The results in this chapter demonstrate that even with data that yield
CR values less than 10%, the CR (or the corresponding CI) value has to be
kept at low levels for small comparison sets and at somewhat higher levels for
large sets, even when the method which is based on the human rationality
assumption is used. However, it should also be stated here that Saaty's
method is the easiest one from the computational point of view.
As it was demonstrated in other chapters of this book, small changes
in the weight values can mean the difference between selecting one alternative
instead of another in many MCDM problems. Accurately estimating relative
weights of key decision components (alternatives and criteria) is crucial in
many real life MCDM problems. Thus, a deeper understanding of this
problem is very critical in MCDM theory and practice.
Chapter 5
DERIVING RELATIVE WEIGHTS FROM

DIFFERENCE COMPARISONS

The previous two chapters analyzed the case of eliciting information
from the decision maker(s) by means of a number of pairwise comparisons.
However, in other related domains such information may be elicited in terms
of difference pairwise comparisons. That is, instead of asking questions of
the type "How many times is item A more important than item B under
criterion X?", now such questions look like "How much is item A more
important than item B under criterion X?".
Therefore, there is a compelling reason to study pairwise comparisons
which are based on the difference rather than on the ratio of the importance
of n items (such as n alternatives or n criteria) considered two at a time.
Difference comparisons have been used in the past to express similarity and
dissimilarity relations among the members of a set that share a common
characteristic. In doing so, various authors have proposed to express such
differences as a distance in a Euclidean sense. This problem was studied in
[Triantaphyllou, 1993]. Thus, the present chapter presents the mathematical
foundation for using such difference comparisons and also an optimization
approach for extracting the final relative priorities from a set (not necessarily
of size n(n-l)/2) of difference pairwise comparisons.
Next, we consider a nonempty set of possible worlds U which is
introduced to represent different states of the system being modeled by a set
of sentences. Such a world here is the set of the entities to be compared in
order to extract their relative weights. Then a similarity function maps pairs
of possible worlds into a number in the interval [1, 0]. In other words, a
similarity function, denoted as:
s: U X U -+ [0, 1],
assigns to each pair of two entities (A, A ') a unique degree of similarity
between 1 and 0 [Ruspini, 1991].
In this chapter we assume that the value 1 corresponds to the
maximum similarity between two entities, while the value 0 to the maximum
dissimilarity. It should be stated here that this assignment of the values 1 and
oto the maximum similarity and dissimilarity, respectively, is quite arbitrary.
However, here we use 1 for the maximum similarity in order to capture the
intuitive feeling that the degree of similarity between any world and itself

should be as high as possible.

In general, a function s: U x U - R+ is a similarity measure if it
satisfies the following properties between any two entities Ai' Aj E U:
(i) S(Ai' Aj) = s(Aj' Ai)' for all Ai' Aj E U,
(ii) s(A i , Ai) ~ s(Aj , Ai)' for all Ai' Aj E U.
(Please notice that the left-hand-side part of (ii) refers to (Ai' Ai) and not to
(Aj , Ai ).)
On the other hand, a dissimilarity relation d: U x U - R+ satisfies the
following properties:
(i) d(A i , A) d(Aj , A;), for all Ai' Aj E U,
(ii) d(Ai' A) ~ 0, for all Ai' Aj E U,
(iii) d(Ai' A;) 0, for all Ai E U.
Examples of similarity and dissimilarity measures can be found in any

book on data analysis (e.g., [Anderberg, 1973] and [Sneath and Sakal,
1973]). A dissimilarity function which has also the following properties:
(iv) if d(W;, llj) = 0,
then W; = llj, (definiteness property)
(v) d(W;, llj) :::; d(W;, Wk) + d(Wk' llj),
for all W;, llj, Wk E U. (triangular inequality property)
is called a distance. This is the case presented in this chapter. Through out
this chapter the similarities among the entities are measured by a dissimilarity
(not similarity) relation, particularly by the distance named in the literature
as the "city-block-metric" [Mardia, 1979].
From the previous expressions it follows that the function d is a
symmetric one. Furthermore, in this chapter it is assumed that entities can
be compared because their similarities can be measured by means of
difference of degrees of a common characteristic or feature. We denote by
WA and WA , the degree that this common feature is present in the two entities
A and A I, respectively. The values WA and WA , can also be viewed as
membership values of the members of a fuzzy set defined in terms of a
common feature. More on fuzzy sets can be found in Chapters 4 and 12 of
this book or in [Dubois and Prade, 1980] and [Kaufmann, 1975]. Usually,
these membership values take on values in the interval [1, 0]. A value of 1
indicates that the feature is fully present, while a value of 0 indicates that the
feature is completely absent.
Using the definition of the two values ~ and WA , it follows that the
closer the two values ~ and WA , are with each other, the more similar the
two entities A and A I should be. Therefore, a dissimilarity function can also
be defined in terms of the degrees two entities share a common feature as
Chapter 5: Deriving Relative Weights from Difference Comparisons 75
follows:
d(A, A 1 = I WA - WA , I'
for any entities A, A lEU.
The main problem that this chapter examines is how to estimate the
similarity relation among any pair of n entities, concepts, or objects. It is
assumed that these n entities share some common feature and that these
similarity relations are viewed in terms of that common feature.
As input data for determining the previous similarity relations we will
use a set of pairwise comparisons. Since there are n(n - 1)/2 possible pairs
of entities, the maximum number of pairwise comparisons is also equal to n(n
- 1)12.
For each comparison the decision maker is asked to do his/her best
in estimating the similarity only between two entities A and A I at a time.
The answer of the decision maker is (as was the case earlier in Chapter 3) a
phrase from a finite set of linguistic phrases. Each such linguistic phrase is
pre-assigned to a numerical value which attempts to capture the numerical
value of the difference I WA - WA , I (where WA and WA , are as defined
above). This is done by using a scale (see also Chapter 3).
In the previous chapter pairwise comparisons have been used to
estimate the relative importance of n entities when they are examined in
terms of a common feature. In that context pairwise comparisons express
ratios of relative importance. Usually, the values of these comparisons are
not in the interval [1, 0]. In the present treatment, however, pairwise
comparisons express the relative similarity of a pair of entities. Furthermore,
they express differences instead of ratios and take values in the interval [1,
0]. The issue which is raised in the light of these pairwise comparisons of
relative similarity is how to combine all of them together and estimate the
actual similarity relations among any pair of the n entities.
The present chapter is organized as follows. The next section
introduces the concept of pairwise comparisons of relative similarity.
Moreover, the same section illustrates how a scale of discrete choices can be
used to quantify this type of comparisons. The third section describes the
main contribution of this chapter. It presents the methodology for processing
pairwise comparisons of relative similarity. The comparisons lead to the
formulation of a quadratic programming problem which minimizes the
errors associated with these comparisons. In this way, the actual similarity
relations among a number of entities can be estimated effectively. Finally,
the last section presents a summary of the contributions described in this
chapter and some possible extensions for more research in this area.
5.2 PAIRWISE COMPARISONS OF RELATIVE

SIMILARITY
In the case of using pairwise comparisons of relative importance, the
comparative judgments express ratios of relative importance among pairs of
entities. For this reason the matrices which contain these pairwise
comparisons are reciprocal (see also Chapter 4). That is, the following
relation always holds:
aji = lIaij and a ii = 1, (for n ~ i, j ~ 1).
However, in the case of pairwise comparisons of relative similarity
the comparative judgments express the difference of the degree a certain
feature is present in pairs of entities. Therefore, the matrices which are
formed with this type of pairwise comparisons are symmetric. That is, the
following expression is true now:
aij = aji' (for n ~ i,j ~ 1).
Every pairwise comparison of relative similarity, aij (for n ~ i, j ~
1), represents the assessment of the decision maker of the absolute difference
I ~ - ~ I of the degrees a certain feature is present in the i-th and j-th
entity, respectively. In the case of this type of comparisons the decision
maker focuses directly on the similarity relations among pairs of entities. The
interest here is not how to estimate the ~ values, but instead, how to estimate
the previous differences I ~ - ~ I .
Since the decision maker is restricted to use a similarity scale with
discrete choices, two methodological problems arise. The first problem is
how to quantify the similarity comparisons. The second problem is how to
combine all n(n - 1)/2 possible pairwise comparisons and estimate the actual
similarity relations among the n entities. It should be stated here that no
matter what are the numerical values associated with the choices given by a
discrete similarity scale, there is always a need to combine the similarity
pairwise comparisons and estimate the actual similarity relations. Apparently,
these two problems are analogous to problems described in previous chapters.
5.2.1 Quantifying Pairwise Comparisons of Relative Similarity

As it was in the case with pairwise comparisons of relative
importance, a scale is needed to quantify pairwise comparisons of relative
similarity. A decision maker cannot directly assign numerical values to
his/her judgments. Instead, he/she can use some linguistic phrases to assess
his/her comparative judgments efficiently and effectively. The observations
made by Weber in 1846 and later by Miller, which were used in developing
scales for quantifying pairwise comparisons of relative importance (see also
Chapter 3), are also applicable when dealing with similarity based
comparisons.
Although the main goal of this chapter is not the development of a
similarity based scale, such a scale is presented in Table 5-1. This scale uses
as linguistic choices an extension of the symbolic structures highlighted in
[Ruspini, 1991]. A close examination of the scales depicted in Tables 3-1,
3-2, and 3-3 with this scale reveals that the proposed scale focuses explicitly
on similarity relations as opposed on the relative importance of the entities.
Table 5-1: Proposed New Similarity Scale [Triantaphyllou, 1993].
Intensity of Similarity Definition
0 The two entities are identical
0.10 Almost identical
0.20 Very similar
0.30 Almost very similar
0.40 Almost similar
0.50 Similar
0.60 Almost dissimilar
0.70 Almost very dissimilar
0.80 Very dissimilar
0.90 Almost completely dissimilar
1 Completely dissimilar
5.2.2 Processing Pairwise Comparisons of Relative Similarity
Suppose that the real (and hence unknown to the decision maker)
value of the (i, j) pairwise comparison is equal to (Xi} (where (Xi} ~ 0). This
value (Xi} is equal to the absolute value of the difference (W; - llj), where Wi
and llj is the degree (similarity value) a certain feature is present in the i-th
andj-th entities, respectively. That is, the following is true:
(Xi} = (Xji = I W; -llj I . (5-1)
Since the decision maker (in his/her assessment of the value of the
(i, j) pairwise comparison) has to use a similarity scale with discrete

numerical values, most likely he/she will select a linguistic choice (such as
"very similar" or "almost similar" etc.) which is associated with a numerical
value (denoted as aij)' Hopefully, this value will be very close to the actual
value (Xij' Therefore, there is an error factor Xij introduced with each
comparison. Thus, the following relation is true:
Xij aij = ~i a ji = (Xij = (Xji = I W; - "') I . (5-2)
From the previous relation it follows that the error factor Xij is equal
to 1, if and only if the value aij (given by the decision maker) and the actual
value (Xij are identical and aij > O. Otherwise, the more far away Xij is from
being 1, the more different the two values aij and (Xij are.
At this point, without loss of generality, suppose that the following
ordering exists among the values WI' W 2 , W 3, ... , Wn:
WI ~ W 2 ~ W3 ~ '" ~ W n · (5-3)
This ordering is always possible because if the W; (n ~ i ~ 1) values of the
n entities are not as in (5-3), above, then a rearrangement of their indices can
achieve the ordering expressed as (5-3). .
Next, let us consider all possible pairwise comparisons among any
three entities Ai' A j , and Ak with similarity values equal to W;, "'}, and W k ,
respectively, (where: n ~ i > j > k ~ 1). Then, by combining the
previous expressions (5-2) and (5-3), the following expressions are derived:
X ik a ik = I W; - W k I W; - W k,
Xkj a kj = I W k - "'} I - (Wk - "'}) "'} - W k ,

and ~i a ji I "'} - W; I - ("') - W) W; - "'}-
By adding up the previous three expressions, the following expression (5-4)

is derived:
X ik a ik + Xkj akj + ~i a ji = 2 (W; - W k ), or:
X ik a ik + Xkj a kj + ~i a ji = 2 X ik a ik , or:
Xkj akj + ~i a ji = X ik a ik ,
for any n ~ i >j > k ~ 1. (5-4)
Given n entities of interest, there are C3n = n(n - l)(n - 2) / 6

possible expressions like the previous one. These expressions involve n(n -
1)/2 variables (note that Xij = ~i and aij = a ji , for any n ~ i, j ~ 1). An
obvious solution to system (5-4) is to set: Xij = 0, for any n ~ i, j ~ 1.
However, it makes sense here to seek to determine the Xij values which are
as close to 1 as possible. That is, to find those Xij values which minimize the
following sum of squares :
n-l n
E E (1 - Xi)2, or equivalently:
i = 1 j = i+l
(5-5)
n-l n n-l n
L L
i = 1 j = i+ 1
Xi~ - 2x L L
i = 1 j = i+l
Xij'
subject to the constraints expressed as (5-4).

The concept of minimizing a sum of squared errors is very common
in many error estimation problems in science and engineering (see also
Sections 4.3 and 4.4). Expression (5-5) reaches an optimal value of 0 if and
only if all the Xi} variables are equal to 1. From the previous discussion it
follows that the n(n - l)(n - 2)/6 expressions of type (5-4), above, comprise
the body of constraints, while expression (5-5) is the objective function of a
quadratic problem with linear constraints which needs to be minimized
subject to the previous constraints. This quadratic programming problem can
be easily transformed into an equivalent system of linear equations. This
system of linear equations has a special structure which can be exploited and
thus it can be solved very efficiently. The previous considerations are
explained further via the next extensive numerical example.
5.2.3 An Extensive Numerical Example

Suppose that a decision maker has to estimate the similarity relations
among the four (i.e., n = 4) entities of interest. denoted as AI' A 2 • A 3 , and
A 4 • Furthermore, let us assume that the real (and hence unknown to the
decision maker) values WI' W2 , W3 • and W4 are equal to 0.92.0.74.0.53, and
0.28, respectively. In other words, the real (and hence unknown) similarity
pairwise comparisons form matrix A as follows:
o 0.18 0.39 0.64

0.18 0 0.21 0.46
A
0.39 0.21 0 0.25
0.64 0.46 0.25 0
In the above matrix the entry (1,2) is equal to 0.18 because 0.92 - 0.74 =
0.18. A similar explanation holds for the remaining entries in matrix A.
The decision maker cannot determine the exact values of the previous
comparisons. However, he/she can use the scale depicted in Table 5-1 to
quantify his/her judgments. If we asswne that the decision maker is always
able to make that selection from the scale which has a numerical value closest
to the corresponding actual value in matrix A then, the following matrix B
represents the pairwise comparisons which we assume that the decision maker
can derive for this example:
0 0.20 0.40 0.60

0.20 0 0.20 0.50
B =
0.40 0.20 0 0.30
0.60 0.50 0.30 0
In this matrix the entry (1, 2) is equal to 0.20 because this value is the closest
value to 0.18 when the scale in Table 5-1 is used. A similar explanation
holds for the remaining entries in matrix B.
From the above considerations it follows that matrix A corresponds
to the notion of the Real Continuous Pairwise (RCP) matrix, while matrix
B corresponds to the notion of the Closest Discrete Pairwise (CDP) matrix
described in detail in Section 3.3.1. These two classes of matrices were
originally introduced in order to study certain phenomena in decision making
problems which use pairwise comparisons as the input data.
Furthermore, suppose that the decision maker has determined that the
ranking of the W; similarity values (recall that the W; values are unknown to
the decision maker) is as follows:
WI ~ W2 ~ W3 ~ W4 •
The decision maker can reach the above conclusion by first asking which of
the n entities has the highest degree of the similarity feature, then which one
has the second highest, etc. At this point, the decision maker does not have
to estimate the values of Wit W2 , W3, ... , Wn • He/She simply needs to
determine only their relative ranking.
The corresponding quadratic programming problem with the 4(4 - 1)(4
- 2) / 6 = 4 linear constraints has the following general form:
minimize f(XJ2> X 13• XI 4> X23• X 24> X34 ) =

= X 122 + X132 + X l42 + X2/ + Xu2 + x3i -
- 2 X12 - 2 X13 - 2 X l4 - 2 X 23 - 2 Xu - 2 X34
subject to:
gl(X12' X 13• X I 4> •••• X34) = X32 a32 + X21 a21 - X13 a l3 = 0
g2(X12' X 13• X I 4> •••• X34) = X42 a42 + X21 a21 - X l4 a l4 = 0
g3(X12' X 13• X I 4> •••• X34) = X43 a43 + X31 a31 - X l4 a l4 = 0
g4(X12' X 13• X l4• •..• X34) X43 a43 + X32 a32 - Xu au = 0
and all the Xii variables assume real positive values.
In general, this quadratic programming problem takes the following form:

n-l n n-l n
minimize f Li L
;=1 = ;+1
X~ - 2 xL i LXii
i=1 = ;+1
subject to:
Xkj akj + ~i aji - Xile aile = 0 (i)
(for any n ~ i > j > k ~ 1)
and all the Xg's ~ O.
As it was stated in the previous section this quadratic programming

problem has C3n = n(n - 1)(n - 2) / 6 linear constraints. Furthermore, its
objective function is always convex. In order to find an optimal solution to
this problem we first need to associate a Lagrangian multiplier Ai with the
i-th constraint and form the Lagrangian function (see, for instance, [Winston,
1991]). For the current example the Lagrangian function is:
L(X12' X 13, XJ4> .00' X34, AI' A2, A3• A.J =
= f(X12' X 13• Xl4 • 000' X3.J -
4
- L A;xgj(XI2' X13 , Xl4' ... , X34 )·
;=1
- - - - ----
Then, any point (X12' X 13• X I4> .00' X34• AI' A2, A3• A4) which satisfies the
following relations:
6L 6L 6L 6L (5-6)
= - - = - - = .•. = - - = 0,
6 X34 6 Al 6 A4
is also an optimal solution to the previous quadratic programming problem.
For the present extensive numerical example expressions (5-6) indicate
that the quadratic programming system (i), above, is equivalent to the
following system of linear equations (given in matrix form):
1 0 0 0 0 0 -a 12 -a 12 0 0 X 12 1
0 1 0 0 0 0 a 13 0 -a 13 0 X13 1
0 0 1 0 0 0 0 a 14 a 14 0 X 14 1
0 0 0 1 0 0 -a23 0 0 -a23 X23 1
0 0 0 0 1 0 0 -a24 0 a 24 X24 1
x
0 0 0 0 0 1 0 0 -a34 -a34 X34 1
a 12 -a13 0 a 23 0 0 0 0 0 0 1..1 0
a 12 0 -a14 0 a 24 0 0 0 0 0 1..2 0
0 a 13 -a14 0 0 a 34 0 0 0 0 1..3 0
0 0 0 a 23 -a24 a 34 0 0 0 0 1.. 4 0
This system of linear equations has the following general form:
[ 1 , -A T
m
A, 0"
xl X I
A
= [! l.
0
(ii)
In the above formulation 1m is the identity matrix of order m (where m = n(n

- 1)/2), A is an Nxm matrix with the coefficients of expressions (5-4) (where
N = C 3n = n(n - l)(n - 2)/6), AT is the transpose of A, ON is a square
matrix of order N and with all entries equal to 0, X is a vector of size m with
the Xij variables (n ~ j > i > ~ 1), ,,_ is a vector of size N with the
Lagrangian coefficients A; (N ~ i ~ 1), 1 is a vector of size m with all
entries equal to 1, and 5 is a vector of size N with all entries equal to 0 .
The previous system (ii) leads to the following derivations:
=
=
~}
_
0
or:
(iii)
The matrix - A AT is always symmetric and of order N (please note

of the difference between the n and N parameters as they were defined
earlier). For the current example this matrix takes the following form:
222 2 2 2
-(a I2 + a l3 + a23 ) -a l2 +a13 -az3
-(a I2 + a l4 + a2J
2 222 2 2
-a 12 -a 14 +~
2 2 222 2
+a13 -a 14 -(a I3 + a l4 + a3J -a34
2 2 222
-az3 +a24 -a34 -(~+~+a3J
The above matrix is of rank N-1. This is true because any column (or
row) is linearly dependent on the remaining columns (or rows) and the
remaining columns (rows) are linearly independent. Therefore, anyone of
the \ variables can be set to an arbitrary value and then solve for the
remaining N-1 variables. Then the Xu variables can be determined from the
first relation in system (iii).
For instance, suppose that in the current example we set A4 = O.
When the numerical data (that is, matrix B) of this example are use..Q., the
linear syste!!J. defined by the ~econd relation in (iii) yields the solution Al = -
0.087015, A2 = -0.066934.z..A3 = -0.147256 (and A4 = 0).
From the previous Ai values and the first relation in (iii) the following
optimal solution to the original quadratic programming problem is derived:
-
X 12
0.96921
X13 0.97590
-
X 14 1.12851
=
X23 0.98260
x24 0.96653
X34
0.95582
It should be_emphasized here that the previous optimal solution is

independent of the Ai values. To see this, suppose that A' and A" (where t.!
;t. A") are two solutions to the equations denoted by the second relation in
system (iii). Then the following derivations are true:
AATA' = AATA" = -A I, or: AAT(A'-A") = O. (5-7)
From the structure of the AAT matrix (as it was depicted earlier) it follows
that expression (5-7) is true if and only if the difference (AI - A") is equal to
the following vector:
+1
-1
+1 xt,
-1
where t is any real number.
Given the previous observation on the difference (AI - A"), the
structure of the matrix A\ and relations (iii), it follows that the following
relations are also true:
Xl - X" = AT (AI - A") = 0, or: Xl = X",
t
where Xl = i + AT AI, and X" = + AT A". In other words, altho.!!gh
system (iii) may have infinitely many Ai solutions, the optimal solution X is
unique.
As it was stated earlier, this is also the optimal solution to the
original quadratic programming problem. In general, if there are n entities
to be compared, then the resulting system of linear equations has C3n - 1 =
n(n-l)(n-2)/6 - 1 real valued variables and the same number of equatiops.

By using expression (5-2) and the previous optimal solution X, the
decision maker can determine aij' the estimated similarity relations among the
n entities, as follows:
-
X12 a 12
-
X a13
:::: 0.193842
13
:::: 0.390360
-
X a 14
14 0.677106
:::: -
X a23 0.196520
23
:::: -
X24 a24
0.483265
0.286746
-
X a34
::::
34
It is interesting to observe here that for this particular illustrative

example these estimates are closer to the actual values in matrix A than the
original input data (i.e., the pairwise comparisons of relative similarity)
presented in matrix B.
A critical issue which arises here is whether the proposed pairwise
comparison approach always works. There is only one situation in which the
pairwise comparisons may yield the wrong results. This is the case when the
triangular property ai j :::;; ai k + ak j does not hold for all possible
combinations of the i, k, and j indices. From the definition of the pairwise
comparisons of relative similarity, however, it was assumed that the triangular
inequality should always be satisfied. This is introduced in order to capture
the intuitive feeling by many decision makers that there is a close relation
between the notion of similarity and that of distance (see, for example,
[Ruspini, 1991]). Therefore, if the decision maker reaches a situation in
which the triangular property does not hold for all possible combinations, then
some or all of his/her comparative judgments need to be revised until the
triangular property holds true for all combinations of the indices.
Another interesting issue here is to observe that the proposed
approach always reaches a feasible solution. This is indeed the case
because from the transformation of problem (i) into its equivalent form (iii)
it follows that the variables A.; (for i = 1, 2,_3, ... , N) can always be
calculated. Furthermore, the solution vector X cannot have· a negative
element (and hence be infeasible).
To see this cOI.!.sider the two relations in system (iii). If some element
in the solution vector X is negative, then the system (iii), and the fact that the
matrix A is formed from the coefficients of the constraints expressed as (5-4),
implies that the expression (that is, an element of the vector Ai): alrj + aij -
aile is negative for some n 2:: i > j > k 2:: 1.
However, the above situation can never occur because from the
triangular property aile 2:: aij + ajk (for any n 2:: i, j, k 2:: 1) and the fact that
aij = aji (for any n 2:: i, j 2:: 1) it follows that the expression akj + aij - aile
can never be negative. Therefore, the proposed approach always reaches a
feasible (and consecutively an optimal) solution.
5.3 CONCLUSIONS
This chapter presented an approach for estimating similarity relations
among n entities. Pairwise comparisons have been used intensively as the
means for extracting the pertinent data for many decision making problems
(as discussed in the previous chapters). In this way, imprecise judgments of
an expert can be processed and accurate estimates of the unknown parameters
of a problem can be derived.
In the past, pairwise comparisons have been used in estimating the
relative importance among the members of a set (of alternatives or criteria).
In that context pairwise comparisons estimate the ratio of the relative
importance of two entities when they are considered in terms of a property
which is present in both entities.

On the other hand, the pairwise comparisons used in this chapter refer
to the relative similarity among pairs of entities. For each comparison
between two entities the decision maker is asked to estimate the difference of
the degrees that a given feature is present in these two entities. In this way,
the resulting similarity relations exhibit the triangular inequality. Furthermore,
this type of pairwise comparisons focuses directly on the similarity relations
among the entities.
In order for the proposed type of pairwise comparisons to be
quantified, a discrete scale (given in Table 5-1) was also defined in this
chapter. Finally, a quadratic programming formulation was proposed as the
means for estimating the desired similarity relations among a number of
entities. This formulation minimizes a sum of squared errors. The proposed
method is very efficient because the quadratic programming problem can be
reduced into the problem of solving a system of linear equations.
An interesting question at this point is to examine why one may need
all the n(n-1)/2 pairwise comparisons. It is easy to verify that only n-1
independent comparisons are enough to estimate the rest of the n(n-1)/2
comparisons. The reason for seeking to evaluate all possible comparisons is
that in this way the proposed approach can use information from many more
sources (e.g., comparative judgments) in order to effectively estimate the
similarity relations among n entities. Only n-1 comparisons would be enough
if the decision maker were perfectly accurate in all his/her judgments.
However, if the decision maker is heavily inaccurate in a few comparisons,
then the negative impact is hopefully diminished if all the independent
n(n-1 )/2 comparisons are incorporated into the estimation process. This
justification is identical to the case of using all possible ratio pairwise
comparisons (as described in the previous chapters).
A number of extensions are possible at this point. An interesting
issue is to devise an approach for estimating similarity relations among entities
even if some of the pairwise comparisons are missing. Problems with
incomplete data are common in many real life applications. A related issue
is to examine how the order of deriving pairwise comparisons affects the final
results. In other words, if the decision maker can only estimate, say 80%,
of the total number of comparisons, then, which should be these comparisons?
This chapter provides some of the necessary foundations for investigating the
previous two critical problems further.
Chapter 6
A DECOMPOSITION APPROACH FOR

EVALUATING RELATIVE WEIGHTS DERIVED
FROM COMPARISONS
As it was described in the previous chapters, pairwise comparisons

play an important role in MCDM problems. They often provide an effective
and efficient manner for eliciting qualitative information from the decision
maker(s). However, a severe drawback of their application is the often large
number of them. If there are n objects (also called entities, elements, or
concepts) to be analyzed, then a complete set of pairwise comparisons is of
size n(n-l)l2. This chapter describes an approach for reducing this number
without severely affecting the benefits of having redundancy in the data
elicitation process.
The main problem examined in this chapter is how to estimate the
relative weights of n elements when the pairwise comparisons of two subsets
with n] and n2 (where n] + n2 > n) elements are known. Successfully
addressing this problem is of critical importance for a number of reasons.
When the number of entities to be processed is large, the number of all
possible comparisons can be too large. For instance, for a collection of 20
entities one needs to perform 190 (i.e., 20 X 19/2) pairwise comparisons.
Therefore, finding a way for reducing the total number of comparisons is of
great practical importance. One may want to partition a large collection of
entities into a number of smaller groups. The elements in each group may be
clustered together by placing very similar ones in the same group. In this
way, the decision maker can evaluate entities which are more homogenous.
This strategy has the potential of deriving more accurate comparisons than
when one has to compare very different elements.
Another application comes from the area of performing the union
operation on the membership values of two fuzzy sets. One can view these
relative weights as the membership values of belonging to a given set (after
normalization the value 1 may denote perfect membership while a value of 0
no membership). If the pairwise comparisons used to derive these
membership values are available, then one may want to utilize them in order
to determine the membership values of the union of the two fuzzy sets. This
operation can be applicable when performing the union operation in fuzzy
databases.

A similar problem of trying to reduce the number of comparisons was

examined by Harker [1987]. In that approach the decision maker starts with
a minimum set of n comparisons (where n is the number of elements). Then
an expert system-like approach is developed for determining what should be
the next comparisons to be made. In this way the decision .maker determines
the comparisons in a guided manner. The main difference of the problem
examined in this chapter, is that here the pairwise comparisons are assumed
to be clustered into two groups of nt and n2 elements each (where n t + ~ >
n). Therefore, in the present investigation we assume that we have available
two complete collections of pairwise comparisons. Thus, the first collection
has nt(n t - 1)/2 and the second has ~(~ - 1)/2 comparisons.
In this chapter two methods are developed for solving the problem of
estimating the relative weights from two complete collections of comparisons.
The first approach is a simple and straightforward approach, while the
second one is more sophisticated and uses a linear programming (LP)
formulation. The LP approach estimates the missing comparisons of the
reciprocal matrix by attempting to minimize the CI (consistency index) value
of the matrix defined on all the n elements. The two methods are also
evaluated by means of a forward error analysis. The computational results
reveal that when the number of common pairwise comparisons is high enough
(Le., the sum n t + n2 is significantly greater than n), then the LP approach
is the most reliable one.
6.2 PROBLEM DESCRIPfION

The main problem examined in this chapter is best described in terms
of an illustrative example. Suppose that there are five entities, say AI' A 2 , A3 ,
A4 , and As, for which a decision maker wishes to find their relative weights
by using pairwise comparisons. These entities may be the alternatives of an
MCDM problem and the decision maker wishes to find their relative weights
(or priorities) in terms of a single decision criterion. Furthermore, suppose
that the decision maker has available the pairwise comparisons when these
five entities are grouped into two sub-sets of four and three members as
follows: The first sub-set is: {At, A2 , A3, A4 }, while the second sub-set is:
{A3' A4, As}·
Let the following two matrices Ml and Ml be the reciprocal matrices
with the pairwise comparisons for the previous two sub-sets, respectively:
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 89
1.000 2.000 1.000 0.333

0.500 1.000 0.500 0.200
1.000 2.000 1.000 0.333
3.000 5.000 3.000 1.000
1.000 0.333 0.500]

and Ml = [ 3.000 1.000 2.000 .
2.000 0.500 1.000
In this chapter it is always assumed that the pairwise comparisons for the
same pair of entities is the same regardless of the matrix. This is the reason
why the comparison between the elements A3 and A4 is the same (i.e., equal
to 0.333) in both the Ml and Ml matrices.
It can be verified that both of the previous matrices are satisfactorily
consistent (i.e., their CI values are less than 0.10) and thus they can be used
to derive the relative priorities of the two groups of elements. In [Saaty,
1980], or in Section 4.2, an effective way for estimating the relative weights
of importance from reciprocal matrices is described. According to that
procedure, one has first to calculate the geometric means of each row and
then normalize these means so they add up to 1. When this procedure is
applied to the previous two Ml and M2 matrices, then the following two
vectors PI and P2 with the relative priorities are derived, respectively.
0.1855
0.9710 0.1634 ]
, and [ 0.5396 .
0.1855
0.2970
0.5318
It can be observed from the first vector that the ratio of the relative
priorities of the elements A3 and A4 is equal to 0.3488 (= 0.1855/0.5318),
while from the second vector the same ratio is equal to 0.3028 (= 0.1634/
0.5396). Therefore, the question we seek to answer in this example is what
are the relative priorities when all the five elements are considered together?
One can view the previous two matrices as parts of a larger matrix
which is defined on the entire five elements. When matrices Ml and Ml are
combined, then the following 5 x 5 matrix M is derived (where "*" indicates
an undetermined comparison):
1.000 2.000 1.000 0.333 *

0.500 1.000 0.500 0.200 *
M 1.000 2.000 1.000 0.333 0.500
3.000 5.000 3.000 1.000 2.000
* * 2.000 0.500 1.000
In other words, only the comparisons for the two pairs {At> As} and {A2' As}
are missing.
In general, suppose that there are n entities of interest which have
been divided into two sub-sets of n1 and nz elements each (where: n1 + nz >
n). Without loss of generality, suppose that the k (where: k = (n 1 + nz) -
n) last elements in the first sub-set correspond, with the same order, to the
first k elements in the second sub-set (as was the case in the previous
numerical example with k = 2 = (4 + 3) - 5).
Next we consider the matrix with all possible comparisons for the n
elements. Then from the above considerations it follows that the decision
maker has available the following comparisons:
aij' for i, j = 1, 2, 3, ... , nt>
and aij , for i,j = liz, (nz+l), (lIz+2), ... , n.
while the comparisons aij , for i = (nl + 1), (nl +2), ... , n, andj = 1, 2, 3,
... , (nz-l) are undetermined (recall that aj i = 11aij). In other words, the n x n
matrix with the n(n - 1)/2 comparisons can be viewed to be partitioned into
the sub-matrices as depicted in Figure 6-1.
111
Missing comparisons
/'
1 JI'
1
1
..".
'
..
Reciprocals of
missing com parisons
~ ". .
"
.. •
..
II,
II
Figure 6-1: Partitioning of the n(n - 1)/2 Pairwise Comparisons.

In the next section two procedures are developed for estimating the
missing comparisons. The first procedure is a simple and straightforward
one, while the second procedure attempts an error minimization strategy and
it is based on a linear programming (LP) formulation.
6.3 TWO SOLUTION APPROACHES
6.3.1 A Simple Approach
As it was mentioned in Chapter 4, perfectly consistent reciprocal

matrices with pairwise comparisons satisfy the following relationship:
aij = aik x ajk , for i,j, k = 1,2,3, ... , n. (6-1)
From relationship (6-1) it follows that, in the perfectly consistent case, the
missing comparisons (denoted here as Xij) can be determined as follows (see
also Figure 6-1):
for i = 1, 2, 3, ... , n-~,
j = (n\+1), (n\+2), ... , n,
and k = (n-n 2 + 1), ... , n\. (6-2)
In non consistent cases, the previous relationship (6-2) is not always

true. However, the unknown terms Xij can be expected to be as close to the
products aik X ajk as possible. Therefore, it is reasonable to determine the
unknown Xij terms as the (arithmetic) averages of all possible products. In
other words, a simple way is to calculate the Xij terms as follows:
n1
L
k=n-"2+1 ajk
(n 1 + n2 - n)'
for i = 1,2,3, ... , n-~,
andj = (n\+1), (n\+2), (n\+3), ... , n. (6-3)
After these averages are calculated the missing entries of the entire
matrix have been estimated. Next, the eigenvector approach (or any other
pertinent approach like the ones discussed in Chapter 4) can be applied on the
complete matrix and thus the final weights of the n entities can be estimated.
Although the above averages can be calculated in a straightforward
manner, the above approach fails to capture the requirement for the following
relationship (6-4):
92 MCDM Methods: A Comparative Study. by E, Triantaphyllou
Xu """ (aii' X aj'j) X Xi'j"

for i, if = 1, 2, 3, .. " n-~,
and},/ = (n l +l), (n l +2), .. " n, (6-4)
It should be noticed that relationship (6-4) follows directly from the fact that
au = lIaji , for any i,} = 1,2,3, .. " n,
6.3.2 A Linear Programming Approach
The previous approach can be modified and transformed into a more

sophisticated procedure, Let us consider relationship (6-3), When one wishes
to incorporate relationship (6-4), then relationship (6-3) may not hold as an
equality but instead it will be satisfied as follows:
"1
L
k = 11-"2+1 ajk
(n 1 + n2 - n)'
for i = 1, 2, 3, .. " n-~,
and} = (nl + 1), (nl +2), .. " n, (6-5)
That is, now the left-hand-side is approximately equal to the

right-hand-side, If we wish to make relationship (6-5) to be an equality, then
an error term, denoted as eu' needs to be introduced as follows:
LNl
ail;
k=N-N2+1 ajk
(n1 + n2 - n)
for i = 1, 2, 3, .. " n-~,
and} = (n l +l), (n l +2), .. ,' n, (6-6)
Similarly, relationship (6-4) can be transformed into an equality by

introducing an error term, denoted as e/i', as follows:
Xu = (a j j • X aj'j) x X;'i' + e/i',

for i, if = 1, 2, 3, "" n-~,
and },j' = (n l +l), (nl+2), .. ,' n, (6-7)
Relationships (6-6) and (6-7) suggest that a reasonable treatment to the

problem of estimating the missing entries (and thus determining the final
relative weights) is to attempt to minimize the sum of all the previous error
terms, This is in accordance with the implicit assumption that the decision
maker attempts to be as consistent in his/her judgements as possible, Since
the errors may be positive or negative, we would like to minimize the sum of
their absolute values, Thus, this consideration leads to the following LP
formulation:
n-~ n
Minimize f = L L I eij I +
i=1 j=n 1 +1
n n
I eiji' j' I
n-~ n-~
+ L L L L
i=1 j=n t +l i'=1 j'=n t +l
subject to:
nt
L a ik
k=n-~+1 ajk
(n 1 + n2 - n)
for i = 1,2,3, .. " n-~,
and} = (n 1 +1), (n 1 +1), ,." n,
(a i i' x ajj') x X;, j' + e/ j',
for i, il = 1,2,3, .. " n-~,
and},/ = (n 1 +1), (n 1 +2), (n 1 +3), .. ,' n,
where all Xy's ~ 0, and all eij and e/j' variables are unrestricted,
The absolute values in the previous LP model can be eliminated by

introducing the following transformations in the body of constraints:
eij = Pij - nij'

for i = 1, 2, 3, .. " n-~,
and} = (n 1+1), (n 1+2), (n 1+3), .. " n,
e/j' = p/j' - N/j',
e
for i, = 1, 2, 3, .. " n-~,
and},/ = (n 1+1), (n 1+2), (n 1 +3), .. " n,
where all the new variables Pij' nij, P/ j', and N/ j' are ~ 0,
The above nij (unknown) variables need not to be confused with the sizes
(i.e., the cardinality values), denoted as n1 and ~, of the two previous input
sets.
Similarly to the above constraint transformations, the previous
objective function! can be modified as follows:
11-"2 II
! = L L [Pij + nij] +
i=1 j=1I1 +1
11-"2 II 11-"2 II
+ L L L i' L
i = 1 j =111 + 1 j' = 1 =111 + 1
This is true because if, say the actual term eij' has to be negative then
it can be easily seen (from the linear dependencies in the columns of the new
constraints) that the value of the variable Pij will be equal to zero (as a
nonbasic variable) while the variable nij will be greater than zero. In other
words:
Pij + nij,
and P iji'j' + Ni'j'
ij .
Therefore, the previous LP model takes the following form:
11-"2 II
Min ! = L L [Pij + nij] +
i= 1 j=1I1 + 1
11-"2 II 11-"2 II
+
+ LLLL
i'
i = 1 j = III + 1 j' = 1 = III + 1
[pf
y
j' N yi ' j']
subject to:
III
L a ik
k=II-"2+1 ajk
(n 1 + n2 - n)
for i = 1, 2, 3, .,., n-~,
andj = (n 1+1), (n 1+2), (n 1+3), ... , n.
Xij = (a ii , x ajj') x Xi'j' + P/j' - N/j',
for i, il = 1,2,3, ... , n-~,
where all the Xii' Pii' nii' and P/' j', N/, j' variables are ~ O.
From the above formulations and definitions of the variables it follows

that the proposed LP model uses (n-n,)(n-n 2)[3 + 2(n-n,)(n-n2)] continuous
variables and (n-n,)(n-n 2)[1 + (n-n,)(n-n 2)] constraints. It can also easily be
seen that if the input pairwise comparisons (i.e., the ones defined on the two
sub-groups of n, and n2 members) are perfectly consistent, then at optimality
the value of the objective function of the previous LP problem is equal to zero
(i.e., all errors vanish). Moreover, the optimal solution of the Xii variables
is given by the rather simple relationships (6-3) (Le., the arithmetic means
calculations). The previous concepts are further illustrated in the following
extensive example.
6.4 AN EXTENSIVE NUMERICAL EXAMPLE
In this example we use the same data as in the illustrative example

described earlier in Section 6.2. Therefore, the pertinent LP model will have
14 variables and 6 constraints (since n = 5, n, = 4 and nz = 3). Please note
that:
a 13 a 14 1.000 0.333
- + +
a53 a54 2.000 0.500
0.583.
2 2
Similarly, we get:
a23 a24
+
a53 a54
= 0.325.
2
and (all x a55 ) = 1.00, (a'2 X a55 ) = 2.00.
Therefore, the constraints of the LP formulation for this numerical example
are as follows:
X'5 0.583 + P'5 - n'5
X25 = 0.325 + P25 - n25
X'5 2.000 X25 + P'5 25 - N'5 25
X25 = 0.500 X'5 + P25 '5 - N25'5
X'5 1.000 X'5 + P 15 '5 - N'5'5
X25 = 1.000 X25 + P25 25 - N25 25
It can be observed at this point that the last two constraints are
redundant, since we can directly set P I5 15 = N I5 15 = P25 15 = N2525 = O. In
general, we can always set P;/ = N;/ = O. This observation suggests that the
LP formulation, defined in the previous section, can have fewer variables and
constraints than the ones used here.
From the previous discussions it follows that the LP formulation for
this example is as follows:
Min f = + n15 + P25 + n 25 +

PI5
+ p 1525 + N Il 5 + P2/ 5 + N 2/ 5
subject to:
X I5 P15 + n15 = 0.583
X 25 P25 + n 25 = 0.325
X15 2.000 P1525
X 25 - + N
15
25 = 0
X 25 0.500 X15 - P2515 + N25 15 = 0
and P 25 P 15 N 25 N 15 ~ O.
X 15, X 25• PI5' P25' n 15• n 25• 15' 25' 15' 25
An optimal solution to this LP problem has: X 15• = 0.5830, X25* = 0.2915,

n25• = 0.0335, and all other variables are equal to zero. The value of the
objective function at optimality is equal to 0.0335.
From this optimal solution it follows that the complete 5 x 5 matrix
(denoted as M '), with all the pairwise comparisons, is as follows:
1.000 2.000 1.000 0.333 0.583

0.500 1.000 0.500 0.200 0.292
M' = 1.000 2.000 1.000 0.333 0.500
3.000 5.000 3.000 1.000 2.000
1.715 3.431 2.000 0.500 1.000
When the eigenvector approximation (as described in Section 4.2 or 6.2) is

applied on the previous pairwise comparisons, the following relative weights
are derived for the entire set of the five elements AI' A 2 , A 3 , A 4 , and A 5 •
0.1392
0.0722
p' 0.1350
0.4137
0.2398
If the non-LP approach is used, then it can be easily verified that X/5
0.583, and X25 = 0.325. Therefore, M" and p" the matrix with the
pairwise comparisons and the vector with the relative priorities, respectively,
are as follows:
1.000 2.000 1.000 0.333 0.583

0.500 1.000 0.500 0.200 0.325
M" 1.000 2.000 1.000 0.333 0.500
3.000 5.000 3.000 1.000 2.000
1.715 3.077 2.000 0.500 1.000
0.1478
0.0783
and P" = 0.1433
0.4390
0.1916
6.5 SOME COMPUTATIONAL EXPERIMENTS

One challenging issue in designing computational experiments is how
to generate the pertinent data. In this study we follow a similar strategy, in
deriving the pertinent data, as in the rest of most of the computational
experiments reported in this book.
The following forward error analysis is based on'the assumption that
in the real world the actual relative weights (or priorities) of the members of
the collection of entities take on continuous values. This continuity
assumption is believed to be a reasonable one since it captures the majority
of real world cases. As it was mentioned earlier, these entities may form a
set of alternatives and the weights reflect the degrees that these alternatives
meet a single decision criterion.
Let WI' W3, W3, ... , Wn be the real (and thus unknown) relative weights
of a set with n members. If the decision maker knew the above real values,
then he/she would be able to have constructed a matrix with the real pairwise
comparisons. In this matrix, say matrix A, the entries are cxij = w/Wj' This
matrix is called the Real Continuous Pairwise (RCP) matrix. Since in the
real world the w;'s are unknown, so are the cxij entries of the previous RCP
matrix. However, we will assume here that the decision maker instead of an
unknown entry cx he/she is able to determine the closest values taken from the
set: {9, 8, 7, ... , 2, 1, 112, ... , 1/7, 118, 1I9} (if the original Saaty scale is
to be used). That is, instead of the real (and thus unknown) value ex one is
able to determine ali such that:
the difference I ex - ali I is minimum,
and ali E {9, 8, 7, ... , 2, 1, 112, ... , 117, 118, 1I9}.
In other words, it is assumed here that one's judgments about the value of the
pairwise comparison of the i-th element when it is compared with the j-th one,
is so accurate that in real life it is the closest (in absolute value terms) of the
values one is supposed to choose from.
The matrix with the ali entries that we assume the decision maker is
able to construct has entries from the discrete and finite set: {9, 8, ... , 2, 1,
1/2, ... , 1/8, 1I9}. This second matrix is called the Closest Discrete
Pairwise (CDP) matrix. More on some intriguing properties of RCP and
CDP matrices can be found in Section 3.3.
For illustrative purposes suppose that the actual relative weights of a
set of five elements, denoted as {AI' A 2, A 3, A 4 , and A5}, are as follows:
0.1328
0.0745
n 0.1542
0.3888
0.2498
By using the previous data, the corresponding RCP matrix can be found to
be as follows:
1.000 1.783 0.861 0.342 0.532

0.561 1.000 0.483 0.192 0.298
Rep 1.161 2.070 1.000 0.397 0.617
2.928 5.220 2.521 1.000 1.556
1.881 3.354 1.620 0.643 1.000
This is true because, say entry (1,2) is equal to 1.783 (= Wl/W2 = 0.1328 /
0.0745). A similar interpretation holds for the remaining entries in this RCP
matrix. Given the previous RCP matrix it can be easily verified that the
corresponding CDP matrix is as follows:
1.000 2.000 1.000 0.333 0.500

0.500 1.000 0.500 0.200 0.333
CDP 1.000 2.000 1.000 0.333 0.500
3.000 5.000 3.000 1.000 2.000
2.000 3.000 2.000 0.500 1.000
This is true because, say entry (1, 2), is equal to 2.000, since the value 2.000
(taken from the current scale on use) is the closest value to the corresponding
entry in the RCP matrix (i.e., 1.783). A similar interpretation holds for the
remaining entries in this CDP matrix. Given the previous CDP matrix the
relative weights (priorities) derived by using the eigenvector approximation
approach (see also Section 4.2) are:
0.1352
0.0743
P 0.1352
0.4143
0.2410
Next, we consider the case in which n1 = 4 and ~ = 3. This setting,

along with the previous CDP matrix, creates the data considered in the
numerical example discussed earlier in Section 6.4. In that example it was
found that the LP approach and the non-LP approach yield the following
relative weight (priorities) vectors PI and Pz, respectively:
0.1392 0.1478
0.0722 0.0783
0.1350 and P2 0.1433
0.4137 0.4390
0.2398 0.1916
From the previous two sets of relative weights we can observe that the
weights derived by using the LP approach are closer to the weights when the
CDP matrix is considered (i.e., with the ones implied by vector P). However,
the ranking of the five elements, when the LP and non-LP approaches are
used, is different than the ranking implied by the weights in vector P in this
illustrative example.
The computational experiments were conducted in a similar manner

to the analysis of this illustrative example. In particular, sets with n = 4, 5,
6, ... , 13 members were assumed. For each case all possible pairs of values
of n 1 and nz were considered. For instance, for the case with n = 6, the (n j ,
n2) pairs were: (3,3), (3,4), (3,5), (4,4), (4,5), and (5,5). In this manner, for
the previous values of n a total of 170 different combinations of n, nj , and n2
values were generated. These values, along with the derived computational
results, are depicted in Table 6-1 (parts A, B, C, and D) (note that the results
for the same nand n2 values are sorted in ascending order of the entries in the
fifth column).
For each such case, 100 random problems were generated and tested
to see if the ranking derived by the non-LP and the LP approaches were
identical to the ranking derived when the CDP matrix is processed with the
eigenvalue approximation approach. Note that sometimes the LP formulation
with n > 13 required excessive CPU time (when 100 random test problems
had to be solved), and thus values of n > 13 were dropped from further
consideration. For each such random problem the actual relative weights
were assumed as in the numerical example. However, because the Saaty
matrices use values from the set: {9, 8,7, ... , 2, 1, 112, ... , 117, 118, 1I9}
only the random problems which are associated with RCP matrices with
entries within the continuous interval [9, 119] were considered. These
computational results are described in more detail in the next section.

The computational results are presented in Table 6-1 (parts A, B, C,
and D). The first three columns in that table present the values of n, nj , and
nZ, respectively. The forth column gives the percent of common pairwise
comparisons (PCs) of the number of available comparisons. Furthermore, the
fifth column presents the percentage of available pairwise comparisons of the
total number of comparisons. Finally, the last two columns in this table
present the number of contradictions (Le., when the derived ranking is
different than the ranking derived by the original CDP matrix) under the LP
and the non-LP approaches.
Table 6-1a: Computational Results, Part A.
n1 nz n Common Available Error rate Error rate

PCs(%) PCs (%) (%), LP (%), non-LP
approach approach
3 3 4 20 83.33 2 2
3 3 5 0 60 15 5
3 3 6 0 40 82 100
3 4 5 12.5 80 7 5
3 4 6 0 60 21 11
3 4 7 0 42.86 88 100
3 5 6 8.33 80 8 10
3 5 7 0 61.9 27 21
3 5 8 0 46.43 94 100
3 6 7 5.88 80.95 5 8
3 6 8 0 64.29 38 27
3 6 9 0 50 92 100
3 7 8 4.35 82.14 21 19
3 7 9 0 66.67 40 25
3 7 10 0 53.33 98 100
3 8 9 3.33 83.33 15 16
3 8 10 0 68.89 48 41
3 8 11 0 56.36 75 45
3 9 10 2.63 84.44 23 18
3 9 11 0 70.91 57 32
3 9 12 0 59.09 99 100
3 10 11 2.13 85.45 34 27
3 10 12 0 72.73 53 42
3 10 13 0 61.54 99 100
3 11 12 1.75 86.36 33 30
3 11 13 0 74.36 74 48
3 12 13 1.47 87.18 33 35
4 4 5 33.33 90 0 0
4 4 6 9.09 73.33 12 9
4 4 7 0 57.14 21 9
4 4 8 0 42.86 93 100
4 5 6 23.08 86.67 0 2
4 5 7 6.67 71.43 14 15
4 5 8 0 57.14 33 18
4 5 9 0 44.44 94 100
4 6 7 16.67 85.71 3 10
4 6 8 5 71.43 28 18
4 6 9 0 58.33 47 33
4 6 10 0 46.67 100 100
4 7 8 12.5 85.71 7 5
4 7 9 3.85 72.22 32 21
4 7 10 0 60 58 40
4 7 11 0 49.09 98 100
4 8 9 9.68 86.11 13 16
4 8 10 3.03 73.33 23 25
Table 6-2b: Computational Results, Part B.
4 8 11 0 61.82 72 72
4 8 12 0 51.52 100 100
4 9 10 7.69 86.67 5 13
4 9 11 2.44 74.55 47 35
4 9 12 0 63.64 76 49
4 9 13 0 53.85 100 100
4 10 11 6.25 87.27 18 23
4 10 12 2 75.76 51 42
4 10 13 0 65.38 88 61
4 11 12 5.17 87.88 24 26
4 11 13 1.67 76.92 55 44
4 12 13 4.35 88.46 11 19
5 5 6 42.86 93.33 0 1
5 5 7 17.65 80.95 9 11
5 5 8 5.26 67.86 24 16
5 5 9 0 55.56 56 28
5 5 10 0 44.44 98 100
5 6 7 31.58 90.48 4 2
5 6 8 13.64 78.57 12 18
5 6 9 4.17 66.67 38 28
5 6 10 0 55.56 59 35
5 6 11 0 45.45 100 100
5 7 8 24 89.29 4 4
5 7 9 10.71 77.78 18 16
5 7 10 3.33 66.67 35 32
5 7 11 0 56.36 98 100
5 7 12 0 46.97 89 79
5 8 9 18.75 88.89 6 8
5 8 10 8.57 77.78 26 36
5 8 11 2.7 67.27 53 45
5 8 12 0 57.58 84 62
5 8 13 0 48.72 90 75
5 9 10 15 88.89 11 14
5 9 11 6.98 78.18 24 29
5 9 12 2.22 68.18 57 45
5 9 13 0 58.97 95 66
5 10 11 12.24 89.09 11 24
5 10 12 5.77 78.79 43 27
5 10 13 1.85 69.23 55 49
5 11 12 10.17 89.39 13 18
5 11 13 4.84 79.49 43 50
5 12 13 8.57 89.74 18 21
6 6 7 50 95.24 0 0
6 6 8 25 85.71 3 11
6 6 9 11.11 75 19 30
6 6 10 3.45 64.44 44 34
6 6 11 0 54.55 79 51
6 6 12 0 45.45 99 100
6 7 8 38.46 92.86 2 2
6 7 9 20 83.33 10 15
Table 6-2c: Computational Results, Part C.
6 7 10 9.09 73.33 39 31
6 7 11 2.86 63.64 57 42
6 7 12 0 54.55 83 52
6 7 13 0 46.15 98 100
6 8 9 30.3 91.67 1 6
6 8 10 16.22 82.22 20 21
6 8 11 7.5 72.73 32 38
6 8 12 2.38 63.64 65 50
6 8 13 0 55.13 97 66
6 9 10 24.39 91.11 5 10
6 9 11 13.33 81.82 20 30
6 9 12 6.25 72.73 45 43
6 9 13 2 64.1 76 51
6 9 11 20 90.91 10 13
6 10 12 11.11 81.82 20 20
6 10 13 5.26 73.08 47 54
6 11 12 16.67 90.91 6 8
6 11 13 9.38 82.05 29 35
6 12 13 14.08 91.03 14 24
7 7 8 55.56 96.43 0 1
7 7 9 31.25 88.89 6 8
7 7 10 16.67 80 16 22
7 7 11 7.69 70.91 45 36
7 7 12 2.44 62.12 68 56
7 7 13 0 53.85 93 66
7 8 9 44.12 94.44 2 3
7 8 10 25.64 86.67 14 20
7 8 11 13.95 78.18 26 23
7 8 12 6.52 69.7 46 40
7 8 13 2.08 61.54 77 55
7 9 10 35.71 93.33 3 7
7 9 11 21.28 85.45 14 20
7 9 12 11.76 77.27 30 32
7 9 13 5.56 69.23 66 52
7 10 11 29.41 92.73 8 12
7 10 12 17.86 84.85 16 23
7 10 13 10 76.92 41 48
7 11 12 24.59 92.42 7 19
7 11 13 15.15 84.62 20 24
7 12 13 20.83 92.31 9 9
8 8 9 60 97.22 0 0
8 8 10 36.59 91.11 4 6
8 8 11 21.74 83.64 13 22
8 8 12 12 75.76 38 40
8 8 13 5.66 67.95 51 47
8 9 10 48.84 95.56 4 4
8 9 11 30.61 89.09 12 14
8 9 12 18.52 81.82 19 25
8 9 13 10.34 74.36 46 38
8 10 11 40.38 94.55 3 7
Table 6-2d: Computational Results, Part D.
8 10 12 25.86 87.88 14 25
8 10 13 15.87 80.77 23 31
8 11 12 33.87 93.94 7 10
8 11 13 22.06 87.18 10 26
8 12 13 28.77 93.59 8 10
9 9 10 63.64 97.78 1 1
9 9 11 41.18 92.73 7 13
9 9 12 26.32 86.36 17 21
9 9 13 16.13 79.49 24 38
9 10 11 52.83 96.36 1 8
9 10 12 35 90.91 14 23
9 10 13 22.73 84.62 17 26
9 11 12 44.44 95.45 4 15
9 11 13 30 89.74 13 22
9 12 13 37.84 94.87 3 6
10 10 11 66.67 98.18 0 0
10 10 12 45.16 93.94 4 10
10 10 13 30.43 88.46 23 26
10 11 12 56.25 96.97 4 3
10 11 13 38.89 92.31 8 16
10 12 13 48 96.15 3 9
11 11 12 69.23 98.48 1 0
11 11 13 48.65 94.87 6 12
11 12 13 59.21 97.44 3 4
12 12 13 7l.43 98.72 1 0
For instance, consider the seventh row in Table 6-1 (part A). This
row has the numbers: [3, 5, 6, 8.33, 80, 8, 10]. In this case the size of the
original set is equal to 6. Also, the same data indicate that the first sub-set
is comprised of the first 3 elements, while the second one is comprised of the
last 5 elements. From the values of n, nj, and n2 we can verify that the
number of the available comparisons is equal to 12, the number of common
comparisons is 1, and the number of all possible comparisons is equal to 15.
Therefore, from the available comparisons, 8.33 % are common, and the
available comparisons represent 80.00% of all possible comparisons. When
the LP approach was applied on 100 randomly generated test problems with
the previous characteristics, in 8 cases (i.e., 8 % of the cases) the derived
ranking was different than the one implied by the corresponding CDP
matrices. For the case of using the non-LP approach, however, the same rate
was equal to 10% (or 10 cases out of the 100 random replications).
These results are also plotted in Figures 6-2 through 6-7. Figures 6-2
and 6-3 depict the error rates, for sets of different size, when the LP and non-
LP approaches are used, respectively. In these figures the error rates are
presented as a function of the size of the set (i.e., the n value) and the percent
of available comparisons (i.e., the fifth column in Table 6-2, all parts).
Figures 6-4 and 6-5 also depict the error rates for different sets, but now the
percent of common comparisons (i.e., the fourth column in Table 6-2) are
used instead. Figures 6-6 and 6-7 present the error rates of the two
approaches when the results are in terms of the averages of all set sizes. This
is the reason why there are only two curves in Figures 6-6 and 6-7; one
represents the performance of the LP approach, while the other one the
performance of the non-LP approach. Figure 6-6 refers to the available
comparisons, while Figure 6-7 refers to the common comparisons.
From the previous results a number of interesting conclusions can be
derived. First of all, as it is natural the percent of common comparisons
increases with the number of available comparisons. From Figures 6-2 to 6-5
we can observe that for larger set sizes the error rates are lower.
Furthermore, the LP approach yields, on the average, smaller error rates.
The effectiveness of the two approaches becomes more transparent in Figures
6-6 and 6-7. We can observe that when the percentages of available or
common comparisons are low, then both methods are unreliable, since they
result in too high error rates. For instance, if the percent of available
comparisons is less than 50 %, then both methods yield error rates higher than
50% (see also Figure 6-6).
106 MCDM Methods: A Comparative Study. by E. Trianraphyllou
o? 0 Q
o ~Q) r~r [' 0".1 IT 0 r 00

II
;;
il Ii
o
lifJ ::
::
r
rfJ
(J
q
II
.,0
fIJ
c,. 1J
~
~
,..
Li
y
!/J
fJ
It
~
10
. 9'
Figure 6-2: Error Rates Under the LP Approach for Sets of Different Size as
a Function of the Available Comparisons.
o
Q
i
?,.
Q? o
r,.!. :.,: : f
,rfl
Irtl
I?
i)
II II I
rjJ ~ i ::
~ ::
to ~j
~
c.- 1J
~
.-
~
a fJ
!/
I/J
II
fJ
10
o·
Figure 6-3: Error Rates Under the Non-LP Approach for Sets of Different
Size as a Function of the Available Comparisons.
108 MCDM Methods: A Comparative Study . by E. Triantaphyllou
\rfJ
rjJ
~
to
UJ
c... 1J
!l
...
J/J
t
6y 1} ~
~
fJ
0
10 0
Figure 6-4: Error Rates Under the LP Approach for Sets of Different Size as
a Function of the Common Comparisons.
9
9
9 o
\rfJ
gO
tJ
l'
ffJ
~
... IjJ ~
..
!1
IP ff
t ?,
6
!/ fJ I ?
fJ ?
10 0 0 o o
Figure 6-5: Error Rates Under the Non-LP Approach for Sets of Different
Size as a Function of the Common Comparisons.
100
90 o ,.''-''-''-.-''~·---"-:··" ·" -, l, -, -" · ·, ·" , , , ·, -l·!, , ·, " ' ' ' ' -' ' ' '-:r-' ' '- -'-' ' ' ' '
Non-LP apprdach
-""-""'-""''' ':'''f--:;~-'''' . -.. "-1":7-":'"' ·rp"~pp-roiiCh·--. -t. -"·, , · " ."""'"' ' r' ' ' "" "--.._,,
". : 0
" 0 i i i i
80
! I 0 ! l
-------t·-.. . .·. . . .---·-t·--,---· . -_ ..--·-i----·-------
0 !
70 .------.--. --t......-:.. - i - _.. -
! A , A . l 0 .1 : I
r-.
; ' .to I 0 I l !
~ 60
(/)
1!0
.... 50
....
0
....
....
w 40
000
30
20
J
!
I
t
o !
. !
0
-------·--··'t. -------------:--r---·--_. . ,· . -.. .'1'-:'·'··. . ··'o-'--: ..
! o~.. ! ..
: OJ
..
,
~···,t,-: ·
tl·
..
- -0-
'"
~.. -a'.
~ao· ! :
f
I
- 6··-..·-···--..--·
· •
10 . - . .- . "". . ."L"""""-,,--I--,,-.-.
! . I . """.,,.-I.!. ~.-~"".-""-.,,
" ,o""l.:~_".:. . ;-<>~":'F}o
~ :. _..~"""~ . . . .". 0
! i ! . o..·!: 0 0
0
40 50 60 70 80 90 100
Available PCs (as a % of all possible PCs)
Figure 6-6: Error Rates for the two Approaches as a Function of the
Available Comparisons.
100
90
80
70
.,.....
~
'-"
60
III
!!0
..... 50
.....
0
l.
.....
w 40
30
20
10
0
a 10 20 30 40 50 60 70 80
Common PCs (as a % of available PCs)
Figure 6-7: Error Rates for the Two Approaches as a Function of the
Common Comparisons.
However, for error rates at lower levels, say less than 30 %, the LP
approach is always the best method. This is the case when the percent of the
available comparisons is 80 % or higher. This is also the case when we
consider the percent of the common comparisons. Figure 6-7 suggests that
when the percent of common comparisons is equal to 10% or higher, then the
two methods have error rates less than 20 %. Moreover, the LP approach
consistently outperforms the non-LP approach. As it is anticipated, the
performance of the methods converges to being perfect (Le, the error rates
vanish) when the percentages of available or common comparisons approach
the upper limit of 100 %.
The fact that in these results the LP approach performs better than the
non-LP approach as the number of common comparisons increases, is in
direct agreement with the way the two approaches were designed. The main
difference of the two approaches is that the LP approach utilizes the presence
of common comparisons in minimizing the sum of the absolute errors. In the
non-LP approach, the issue of the common comparisons is totally ignored.
However, it is interesting to observe here that when the percent of common
comparisons is very high (e.g., more than 60%), then the non-LP approach
is as good as the LP approach. Moreover, the non-LP approach is extremely
simpler (one just needs to calculate the means in relations (6-3» compared to
the more CPU time consuming LP formulation.
6.7 CONCLUSIONS
This chapter examined the problem of decomposing a large set of
pairwise comparisons into two sub-sets. These two sub-sets may have some
common comparisons. These sub-sets may be defined by groups of elements
which are more similar within these two sub-sets, than when they are
considered all together. Two approaches were developed and tested. The first
is a simple approach, while the second one is based on an LP formulation.
Both approaches first attempt to estimate the missing comparisons and then
to derive the relative priorities and rankings by using all the pairwise
comparisons .
The LP formulation attempts to minimize the sum of the absolute
errors in the missing comparisons. The simulation results suggest that the LP
approach is consistently better than the non-LP approach. Furthermore, the
effectiveness of this approach improves with the amount of the common
comparisons. It should be also stated here that an interesting issue for future
research in this area might be how to decompose a large set of elements to be
compared into more than two sub-sets.
Finally, it is important to notice that the computational results in this
chapter are contingent on the way the random data were generated. If the
random data are generated from other distributions, then it is possible for the
computational results to be different. However, it can still be highly
anticipated that the LP approach will perform better than the non-LP
approach, and moreover, its performance to improve with the amount of the
input data (since it processes the input data in a more comprehensive manner).
Since deriving relative weights from pairwise comparisons is a popular
approach in MCDM problem solving, finding ways of reducing the amount
of the required data, is of critical importance.
Chapter 7
REDUCTION OF PAIRWISE COMPARISONS

VIA A DUALITY APPROACH
As it was discussed in previous chapters, an appealing approach for

eliciting qualitative data for an MCDM problem is to use pairwise
comparisons. Next suppose that a decision maker wishes to elicit the relative
priorities, or weights of importance, of n entities via a sequence of pairwise
comparisons. As before, these n entities could be the decision criteria, or the
alternatives to be examined in terms of a single decision criterion in some
MCDM problem. Then, as it was illustrated in Chapter 3, the decision maker
must elicit the value of n(n-l)/2 pairwise comparisons. Therefore, if an
MCDM problem involves m alternatives and n decision criteria (multiple
hierarchical levels are not considered at this point), then the total number of
the required pairwise comparisons is equal to n(n-l)/2 + n(m(m-l)/2).
This number can be quite large, even for moderate values of m and
n. In order to get the feeling of the amount of the required pairwise
comparisons required, let us consider the case of a simple MCDM problem
with 15 alternatives and 10 decision criteria. That is, now we have m = 15
and n = 10. Then, the total number of the required comparisons is equal to
1,095. From the above discussion it follows that this number increases
quadratically as the values of m and n increase. This chapter presents an
approach for reducing the total number of the required pairwise comparisons
when the number of alternatives is larger than the number of decision criteria
plus one (Le., when m > n + 1).
This is achieved through a duality approach in the way the pairwise
comparisons are elicited. These developments were first reported in
[Triantaphyllou, 1999]. The proposed duality approach is presented in the
next section. The third section of this chapter demonstrates this approach in
terms of an extensive numerical example. The forth section presents some
numerical results that highlight the benefits of this approach. Some
concluding remarks, and areas of possible extensions, are presented in the last
section.

7.2 A DUALITY APPROACH FOR ELICITING

COMPARISONS
As it was primarily mentioned in Chapter 3, pairwise comparisons

elicit information from the decision maker on the relative performance of two
alternatives at a time when they are examined in terms of a single decision
criterion. Moreover, such comparisons can also be used to extract the relative
weight of importance of the decision criteria or sub-criteria in a hierarchy.
In this way the decision maker can determine the relative performance of the
alternatives in terms of a given criterion, or the relative weights of a set of
criteria. Thus, the columns of the decision matrix are derived normalized.
This normalization usually occurs by dividing the entry of each column by the
sum of the entries in that column. In that case, their sum is equal to one. An
alternative way is to divide each entry by the largest entry of the column it
belongs to. This is the main modification of the original AHP as proposed
by Belton and Gear [1983] and we call it the revised AHP (see also Section
2.2.4).
When two alternatives are considered, the decision maker is asked to
choose the best linguistic statement from a fmite set of statements which best
describes the answer to the following question (described in generic terms):
"What is the relative imponance of alternative Aj when it is
compared with alternative Aj in terms of the decision criterion
Ct·?"
Next, the linguistic statement chosen by the decision maker is assigned to

some numerical value according to a predetermined scale (such as the ones
described in Chapter 3).
In the dual approach described in [Triantaphyllou, 1999] the previous
question takes on a different format. Instead of comparing two alternatives
at a time, now the relative performance of two decision criteria is examined
in the context of a given alternative. That is, now the typical question is of
the following format:
"What is the relative imponance of criterion Cj when it is
compared with criterion q in terms of alternative At?"
In reality the previous type of question is not new. In the traditional

way of eliciting the value of the pairwise comparisons the decision maker is
also asked to compare two criteria at a time. Now, the decision maker is
asked to do the same within the domain of a single alternative at a time.
One can reasonably argue at this point that rarely a decision maker
examines a set of criteria without having in mind the alternatives to be
Chapter 7: Reduction of Pairwise Comparisons Via a Duality Approach 117
evaluated and vice-versa. In other words, since criteria and alternatives are
the key entities in a given MCDM problem, one has to simultaneously keep
them in focus all the time during the decision making process.
For instance, if one considers the two criteria "cost" and
"functionality " in purchasing a product, then these two criteria have different
relative importance if the problem is to purchase a TV set or a new house.
In the first case a 20% price change may not be so critical, while in the case
of purchasing a new house a 20 % price change may be more detrimental.
Therefore, it is also natural to accept the premise that perfect independence
of the criteria and the alternatives seldomly exists. This is a highly debatable
issue in the decision analysis community and different authors have expressed
different positions (see, for instance, [Saaty, 1994, Chapter 8]). Thus, the
use of the proposed comparison question format is not totally new.
When the decision maker elicits all possible comparisons for a given
alternative, then the weight vector of this judgment matrix corresponds to a
normalized row (as opposed to a column with the traditional approach) of the
decision matrix. Therefore, by using the previous type of dual comparisons
the decision maker can determine all the rows of the decision matrix in a
normalized manner (i.e., the elements in each row sum up to one or are
divided by the maximum entry).
Let by (for i = 1, 2, 3, ... , m andj = 1, 2, 3, ... , n) denote the (i,
j) entry in the decision matrix normalized in terms of each row via a sequence
of m judgment matrices which are defined with dual comparisons. It can be
noticed that now it is required to form m judgment matrices of size n xn,
while in the traditional (primal) approach there are n such matrices of size
mXm each. Please recall that a typical MCDM problem is assumed to have
m alternatives and n decision criteria. Moreover, suppose that the decision
maker also forms the pairwise comparisons of a single judgment matrix in the
traditional fashion. That is, the decision maker forms a single mxm
judgment matrix. To help fix ideas, suppose that this is the judgment matrix
which examines all the m alternatives in terms of the first decision criterion.
Let ail (for i = 1, 2, 3, ... , m) be the elements of the corresponding weight
vector derived from the last judgment matrix.
Given the values of the m rows (e.g., the by values) normalized as
above and the values of the single normalized column (e.g., the ail values),
it is now straightforward to derive the elements of any column in the decision
matrix. It can be easily verified that this can be achieved by employing the
following formula (7-1):
(7-1)
It is important to note here that the sum of all the ayl elements in a given
column (except the first one) is not necessarily equal to one.
In general (if the k-th column is selected for normalization via the
extra judgment matrix), then the previous expression becomes:
(7-2)
Next the previously derived elements are normalized by dividing each one of
them by the sum of the entries of its column. Thus, the last expression
yields:
(7-3)
Therefore, it is possible for one to derive the decision matrix normalized in

any desirable way by using dual pairwise comparisons and in conjunction with
a single judgment matrix with primal (traditional) pairwise comparisons.
Obviously, when these dual comparisons are used, a different sequence of
judgment matrices is formed of dimensions different than those formed when
comparisons are elicited as in the traditional approach. The question which
is naturally raised at this point is under what condition(s) the number of
comparisons is smaller in the dual approach. This is the subject of the
following theorem and corollaries [Triantaphyllou, 1999]:
Theorem 7-1:
The percent (%) of change of the number of comparisons between the prime
and the dual problem is given by the following formula:
m-'(n'---_1~)..,.(m_-_n_-_l~)
_ (7-4)
x 100.
n(n - 1 + m(m - 1»
Proof:
As it was described in the previous paragraphs, a typical MCDM problem
with m alternatives and n decision criteria, under the traditional (prime)
approach requires the formation of a single n x n judgment matrix to derive
the criteria weights. It also requires the formation of n judgement matrices
of size m x m each to derive the relative weights of the m alternatives in terms
of each one of the n decision criteria. Thus, according to the traditional
(prime) approach the total number of the required pairwise comparisons is

equal to:
n(n - 1) m(m - 1) (7-5)
+ n .
2 2
Similarly, for the dual problem the decision maker must form one
matrix of size n x n for the weights of the decision criteria, plus m judgement
matrices of size n Xn (one for each of the m alternatives of the decision
matrix), plus a single matrix of size mXm (in order to normalize anyone of
the n columns of the decision matrix). Therefore, the total number of
pairwise comparisons under the dual approach is equal to:
n(n - 1) n(n - 1) m(m - 1) (7-6)

+ m + •
222
Therefore, the net decrease on the number of comparisons can be
found as the difference of expression (7-6) from expression (7-5), given as (7-
7), below (after some elementary algebraic simplifications take place):
1 (7-7)
-m(n - 1)(m - n - 1).
2
Thus, the percent (%) change of the number of comparisons between the
prime and the dual problem is given as expression (7-4). •
Corollary 7-1:
The dual problem requires less pairwise comparisons if the number of
alternatives in the problem is greater than the number of decision criteria plus
one.
Proof:
The above corollary follows directly from the fact that expression (7-7) must
be greater than zero. Thus, m - n - 1 > 0, or m > n + 1. This is true
because the value of n is always greater than 1 (and also m > 0). •
Therefore, if an MCDM problem has more alternatives than decision

criteria plus one, then the data elicitation process can explicitly benefit from
the smaller number of comparisons needed by the proposed dual approach.
It should be stated here that in many real life problems there are more
decision criteria than alternatives. However, in some real life problems the
number of alternatives may be dramatically high. For instance, in ranking a
number of employees for possible pay raises, the number of alternatives (i.e.,
the individual employees) is often very large, when compared to the decision
criteria (which describe their job performance). The rate of reduction on the
number of required comparisons is given by expression (7-4) of Theorem 7-1.
The next corollary describes a theoretical statement and it states that these
reduction rates converge to a fixed quantity when the number of criteria is
kept constant and the number of alternatives approaches to infinity
[Triantaphyllou, 1999].
Corollary 7-2:
The percent (%) of change of the number of comparisons between the prime
and the dual problem, for a given number of criteria N, approaches the value
(N - J)/N when the number of alternatives approaches infinity.
Proof:
This follows directly from expression (7-4) in Theorem 7-1, if one sets n =
N and then takes the limit when m approaches to infinity. Also, the function
in (7-4) is continuous and increases monotonically. Therefore, this limit can
also serve as an ugper bound on the reduction rate which can be achieved
by using the proposed duality approach. •
A related issue is to examine what happens if a problem is defined in

a multi-level hierarchy. That is, the previous considerations can easily be
extended to this general case. The proposed duality approach can directly be
applied on each individual level of the hierarchy. In particular, the duality
approach will be beneficial if the number of sub-criteria in one level, is
greater than the number of criteria in the previous level plus one.

The previous analyses are next demonstrated in terms of an extensive
illustrative example. Suppose that a single-level hierarchy MCDM problem
involves the five alternatives AI' A 2 , A j , A4 , and As, which have to be
evaluated in terms of the three decision criteria CI , C2, and Cj • These three
criteria are assumed to have weights of importance equal to: (5/8, 118, 2/8).
Let the actual values of these alternatives in terms of the three decision
criteria be as in the following decision matrix:
Criteria
CI C2 C3
Alts. (5/8 1/8 2/8)
Al 5 3 4
A2 2 5 3
A3 4 2 5
A4 2 4 1
As 3 3 2
The previous criteria weights can be viewed as the weight vector of the
following 3 x 3 judgement matrix:
1 5/1 5/2
1/5 1 1/2
2/5 2/1 1
7.3.1 Applying the Primal Approach

Under the traditional AHP approach, the columns of the previous
decision matrix must be normalized by dividing each entry by the sum of the
entries in that column. Therefore, when the above decision matrix is
normalized, it takes the following form:
Criteria
CI C2 C3
Alts. (5/8 118 2/8)
Al 5/16 3/17 4/15

A2 2/16 5/17 3/15
A3 4/16 2/17 5/15
A4 2/16 4/17 1/15
As 3/16 3/17 2/15
In standard AHP practice (see also Section 2.2.3), the decision maker
derives the normalized columns from a sequence of three judgment matrices
of size 5 x 5 each. For instance, the first column is derived from the
following primal judgment matrix (please note that all the comparisons are
perfectly consistent for easy demonstration of the main concepts in this
chapter):
1 5/2 5/4 5/2 5/3

2/5 1 2/4 2/2 2/3
4/5 4/2 1 4/2 4/3
2/5 2/2 2/4 1 2/3
3/5 3/2 3/4 3/2 1
From the above considerations it follows that for this illustrative

example the decision maker needs to form 4 (i.e., 1 + 3) judgment matrices
with a total of 33 (= 3(3-1)/2 + 3[5(5-1)/2]) pairwise comparisons.
7.3.2 Applying the Dual Approach

Under the dual approach the decision maker is required to normalize
the rows of the decision matrix, by dividing each entry by the sum of the
entries in that row. Therefore, when the above decision matrix is normalized
in this manner, it takes the following form:
Criteria
CI C2 C3
Alts. (5/8 118 2/8)
Al 5/12 3/12 4/12

A2 2/10 5/10 3/10
A3 4/11 2/11 5/11
A4 217 417 117
As 3/8 3/8 2/8
In the proposed dual approach, the decision maker derives normalized

rows from a sequence of five judgment matrices of size 3 x 3 each. For
instance, the first row is derived from the following 3 x 3 dual judgment
matrix:
1 5/3 5/4
3/5 1 3/4
4/5 4/3 1
Next, the decision maker is also required to derive a normalized

vector (besides the three normalized rows) of the decision matrix. In other
words, to form a single 5x5 judgment matrix. This extra matrix represents
the pairwise comparisons derived when the five alternatives are compared in
terms of anyone of the three decision criteria. For easy demonstration,
suppose that the decision maker chooses to compare the five alternatives in
terms of the first decision criterion (the case of using the second or the third
decision criterion can be developed in an identical manner). The
corresponding judgment matrix was provided in the previous sub-section.
Therefore, the normalized column is: (5/16, 2/16, 4/16, 2/16, 3/16?
The next step is to use the previous three normalized rows and the
normalized column to derive the normalized columns of the decision matrix.
It can be easily observed that only the last two columns of the decision matrix
need to be calculated in this example. When formula (7-3) is applied to
calculate the value of al2 , it turns out that its value is equal to 3/17. This is
true because from relation (7-3) the value of al2 should be equal to:
5/16 x3/12
5/12
( 5/16 X3/12) +(2/16 X5/10) +(4/16 X2/11) +( 2/16 X4f7) +( 3/16 X3/8)
5/12 2/10 4/11 2f7 3/8
3/16 = 3/17.
3/16 + 5/16 + 2/16 + 4/16 + 3/16
In a similar manner, it can be shown that the second column (after

normalization) is equal to:
(3/17,5/17,2/17,4/17,3/17)T.
Similarly, the third column (after normalization) is equal to:
(4/15, 3115, 5/15, 1115, 2/15?
Therefore, the normalized decision matrix is as follows:
Criteria
C1 C2 C3
Alts. (5/8 1/8 2/8)
A1 5/16 3/17 4/15

A2 2/16 5/17 3/15
A3 4/16 2/17 5/15
A4 2/16 4/17 1/15
As 3/16 3/17 2/15
Obviously, the above matrix is identical to the matrix derived with the
primal approach. However, in the dual approach the decision maker needed
to form 5 (i.e., 1 + 3 + 1) judgment matrices which required a total of 28
(= 3(3-1)/2 + 5[3(3-1)/2] + 5(5-1)/2) pairwise comparisons. This represents
a reduction of 15.14 % from the total number of pairwise comparisons
required under the primal approach. Although this reduction may not seem
to be very significant, when the number of alternatives is much higher than
the number of decision criteria, then the benefits of using the dual approach
increase dramatically. This is further investigated in the numerical results
presented in the next section.
7.4 SOME NUMERICAL RESULTS FOR PROBLEMS

OF DIFFERENT SIZES
Let us consider expressions (7-5), (7-6), and (7-7) which calculate the
total numbers of comparisons required under the primal and dual approaches,
as well as their net difference. Figures 7-1, 7-2, 7-3, and 7-4, depict these
values when the number of decision criteria n is equal to 10, 15, 20, and 25,
respectively. As it is also shown in these expressions, the values of these
functions increase quadratically with the value of m (i.e., the number of
alternatives) .
Moreover, when the condition of Corollary 7-1 is satisfied (i.e., when
m > n + 1), then the net decrease due to duality is positive. It is also
interesting to observe here that in all these four representative plots, the
number of comparisons for the dual problem increases almost linearly with
the number of alternatives. This, of course, is a nice characteristic of the
number of comparisons required by using the dual approach. The previous
observation holds true for the ranges of the m and n parameters considered in
this study.
3000
. . . 1 l .
2500 . .
I -
·······················i·········N~~b·;~· ~f···c~~;~; ·i·~~~~···f; ···p;·i ·~;· *~b·i·;~·····"T········
...,
1 : :
c: 2000 ······t ........................... ····················•·· ..····....Net"·"[jec;;:e~~e···d~e·-r····· ~.:~ . . . . . ..

o
.t:
o
E '500 ··········r········..····.. ········~··················· ...... i. ················::::.:;r':····..······..·····
8 ....
/ ..
: ; :
'000 ········ ..········,·",4,········ .. ,·············,~······•..•................~ .... .•..__ i_.... !'::.. _........
. .
.......;
./ :
.,../ :
-2~-
! ~~::1=I -s=~~=::--~-
500
o . Num,er of c01parisons ifor Dual problem

l i E 1
-SOD
o 5 10 '5 20 25 30 J5
Number of Alternatives in Problem
Figure 7-1: Total Number of Comparisons and Reduction Achieved When the
Dual Approach is Used. The Number of Criteria n = 5.
6000
5000
Numb~r of Comparisons for Prlme !Problem l. . .·. . ·. ·. . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .1"0 . . . . . . . . . . . . . . . . . . . . . . . , . . . . . . . . . . . . . . . . . . . . . . . .
: ;
on • • • • • ~. . . . . . . . . . . . . . .
.g 4000 · .. ········t..·.. 1 •.......•••••••••• ~ •••• •••••••••• ,••••••••••l.........

: .
..j
:: ·····i····
'"
.t:
o
JOOO
Q.
E
8
'0
~ 2000
.I:J
E
:;)
z
-1000
o 5 10 15 20 25 30 J5
8000
6000
: :
3'" of Cdmparison~ roblem i

·c'" ~ i
o 4000 .. '·····..·····I···············
i
·'··········' ~ ···· ··· ·· ·
"'-
E ~
8
'0
;
.&J
2000 ............. !....................... !. ..
E
:>
z
o ........................,..............
~-~ ...- .
-2000 ~~~~--~~~--~~~~~~--~----~--~~
o 5 10 15 20 25 30 35
12000
.............. -- .._._ .. :..... _- ........ ---_.:.__ .............. "' .... ............ ........ ......--
. . _ _ + N":b=IOfC:·FO":~! P".~.. : . . . . . -...

10000 ~"' " ;
8000
'"
c:
. ,'",'''. In...' h.'.S1:''.,. PfOh......

0
'"
·c
0
"'- 6000
E
0
(.)
'0
; 4000
.&J
E
:>
z 2000
- 2000
0 5 10 15 20 25 30 35
Dual Approach is Used. The number of Criteria n = 20.
4000 ~~~-'~~~~~~~-r~~~'-~~~,-~~~~~--~
3000 .,", ... ,"', .. ,,,.,, .• ········t·················,,······!··········

(I)
c:
o
.. ; ......... i .
'8.E
(I)
2000
8
'0 1000 """""""".. ,""."."."" ....
"-
OJ
.<:>
E
z'" o
-1000
o 5 10 15 20 25 30 35
Figure 7-5: Net Reduction on the Number of Comparisons When the Dual
Approach is Used. Results for Problems of Various Sizes.
100 ~~~~r-~~~~~~~~~--~~~~~~--~~~ __ ~~
Number bf Criteriia n =4
50 .. ~~~~~
'"
<::
o ........ , ........................ ...................... ,.
o .;..
.~
<>
0..
E
o -50 ............................................
<..>
'0
~ - 100
(.J
".,
=>
a:: -150
N
"",j" '''·''''''''·.. ".... ''t''·'' '·' '·. ·. ···,',·.·.-_i"",: .....,.. ,.. ,
-200 ......... i .......................... ;.. ........ " ....... ,._ .......... ,....... ................
-250
o 5 10 15 20 25 30 35
Num~r of Alternatives in Problem
Figure 7-6: Percent (%) Reduction on the Number of Comparisons When the
Dual Approach is Used. Results for Problems of Various Sizes.
From the previous figures and analyses it follows that if the number
of alternatives is more than 35, then the number of comparisons under the
dual approach will assume a more quadratic rate of increase. However, one
may reasonably argue that most of the real life problems involve less than 35
alternatives (which is the upper limit in the previous plots).
On the other hand, the number of comparisons required under the
traditional (prime) approach noticeably increases quadratically in these
figures. The previous observations are a compelling reason why the duality
approach can be significantly beneficial to most real life decision problems
which use pairwise comparisons to elicit qualitative data.
Figure 7-5 depicts the net reduction on the number of comparisons
(given as expression (7-4)) achieved when the dual approach is applied, for
different size problems. Figure 7-6 presents the percent (%) reductions
achieved when the dual approach is used on problems of various sizes. As
before, these increases follow quadratic patterns on the size of the decision
problems. It is noticeable that in Figure 7-6 the reduction rates seem to
converge to a constant value when the number of alternatives increases.
Obviously, this is in direct agreement with Corollary 7-2.
7.5 CONCLUSIONS
The previous analyses demonstrate that the number of comparisons

required to solve an MCDM problem which has m alternatives and n criteria
can significantly be reduced when the number of alternatives is significantly
greater than the number of criteria plus one. This is achieved in terms of a
duality approach. In this approach the decision maker compares how well the
criteria perform within a single alternative at a time. It should be noticed that
in the traditional way of implementing pairwise comparisons, the decision
maker is asked to compare the alternatives in terms of a single criterion at a
time or to compare a set of criteria (when their relative weights of importance
are extracted).
The achieved reductions on the required total number of pairwise
comparisons are given through some analytical formulas. The main findings
of this study are also depicted in a number of figures. It is remarkable to
emphasize here that these reductions become more dramatic when the size of
the problem increases. Thus, the proposed duality approach becomes more
practical for large size decision problems.
Although the concept of duality is an old one in decision sciences
(e.g., in linear programming), the proposed duality approach is a novel
development in MCDM problem solving. The proposed method can be
applied to the AHP, or its variants, as well as to any other method which uses
pairwise comparisons to elicit qualitative or fuzzy data from the decision

maker(s). An interesting question is how to expand this duality approach to
difference pairwise comparisons (as defined in Chapter 5). Another goal
might be to expand these results to problems with multiple hierarchies on the
way the decision criteria are organized. More research in this intriguing area
may reveal more benefits of using information extracted from the dual
formulation of a given MCDM problem.
Chapter 8
A SENSITIVITY ANALYSIS APPROACH

FOR MCDM METHODS
S.l BACKGROUND INFORMATION
There is considerable research on sensitivity analysis for some

operations research and management science models such as linear
programming, inventory models, and investment analysis (e.g., [Wendel,
1992] and [Triantaphyllou, 1992]). However, research on sensitivity analysis
for deterministic MCDM models is rather limited. A brief overview of the
related literature can be found in [Triantaphyllou and Sanchez, 1997]).
Simon French in [1986; and 1989] emphasized the role of sensitivity
analysis on decision making. He performed an analysis of the use of
interactive decision aids to overcome some of the difficulties in modeling
judgments. The models examined were mostly stochastic as opposed to
deterministic. Furthermore, he stressed out the importance of having better
and more general sensitivity analysis tools. Also, French and Rios Insua in
[1989] used a distance minimization approach to determine competitors to a
current optimal solution.
A methodology for sensitivity analysis in multi-objective decision
making is described by Rios Insua in his seminal book [1990]. That treatment
introduced a general framework for sensitivity analysis which expanded
results of the traditional Bayesian approach to decision making. Emphasis is
given to cases which use partial and/or doubtful data. Also, that work
contains an analysis of why the ''flat maxima principle" is not valid. The
Rios Insua [1990] book also includes the description of SENSATO; a prototype
library of a sensitivity analysis package for decision aids. However, the
present chapter assumes that the data are not stochastic and it focuses on the
issue of sensitivity analysis on the weights of the decision criteria and the
performance measures of the alternatives in a deterministic MCDM
environment.
A recent development in sensitivity analysis when the AHP is used is
due to Masuda [1990]. In that work Masuda studied the effect that changes
in entire vectors of the decision matrix may have on the ranking of the
alternatives. That author considered multiple levels of a hierarchy. However,
he did not offer a procedure for performing a sensitivity analysis on changes
in an individual piece of data of a given problem (i.e., changes in a single
criterion weight or the performance value of an alternative in terms of a given
criterion) .
The sensitivity analysis approach proposed in [Triantaphyllou and
Sanchez, 1997], which is also described in detail in this chapter, is
complementary to the one developed by Masuda. Thus, these two approaches
can be used together (since the proposed approach can focus on individual
judgments while the Masuda approach considers a single vector at a time).
Also, Armacost and Hosseini in [1994] presented a procedure for determining
the most critical criterion for a single level hierarchy AHP problem. Their
work does not explicitly determine what should be the smallest change in the
current weight of a criterion, such that the existing ranking of the alternatives
will change.
As a related comment it should also be stated here that the Expert
Choice (http://expertchoice.com/) software package which is based on
the AHP, performs a type of elementary sensitivity analysis. The user has the
option to graphically alter the weights of the decision criteria and see on the
screen how the rankings of the alternatives will change. However, the issue
of criteria sensitivity analysis is not studied systematically. Moreover, Expert
Choice does not offer any means for studying the effects of changes in the
measures of performance of the alternatives (which is part of the methodology
proposed in this chapter).
In decision making the weights assigned to the decision criteria
attempt to represent the genuine importance of the criteria. When criteria
cannot be expressed in quantitative terms (such as cost, weight, volume, etc.),
then it is difficult to represent accurately the importance of these criteria. In
a situation like the above one, the decision making process could be improved
considerably by identifying the critical criteria (the formal definition is given
later) and then re-evaluate more accurately the weights of these criteria. The
intuitive belief is that the criterion with the highest weight is the most critical
one [Winston, 1991; p. 754]. This may not always be true and in some
instances the criterion with the lowest weight may be the most critical one.
The decision maker can make better decisions if he/she can determine
how critical each criterion is. In other words, how sensitive the actual
ranking of the alternatives is to changes in the current weights of the decision
criteria. Thus, in this chapter we examine two closely related sensitivity
analysis problems. In the first problem we determine how critical each
criterion is, by performing a sensitivity analysis on the weights of the
criteria. This sensitivity analysis approach determines what is the smallest
change in the current weights of the criteria, which can alter the existing
ranking of the alternatives. In the second problem, we use a similar concept
to determine how critical the various performance measures of the alternatives
(in terms of a single decision criterion at a time) are in the ranking of the
alternatives.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 133
This chapter is organized as follows. The next section presents the

formal definitions of the two sensitivity problems analyzed in [Triantaphyllou
and Sanchez, 1997]. The related concepts and methods are further illustrated
in terms of some demonstration examples. Some computational experiments
were also performed in order to increase the insight of these sensitivity issues.
Finally, the last section presents the main conclusions of the proposed
methodology.
8.2 DESCRIPTION OF THE TWO MAJOR SENSITIVITY

ANALYSIS PROBLEMS
As before, the typical MCDM problem to be studied in this chapter
is defined on m alternatives and n decision criteria. The available data form
a decision matrix with the aij and Wj entries (for i = 1,2,3, ... , m, and) =
1, 2, 3, ... , m) as defined in Chapter 1. Then, the central MCDM problem
is how to determine the best alternative or rank the entire set of alternatives.
Let Pi (for i = 1, 2, 3, ... , m) represent the final preference of
alternative Ai when all the decision criteria are considered simultaneously. As
it was shown in Chapter 2, different decision methods apply different
procedures in calculating the preference values Pi' In this chapter, it will be
assumed without loss of generality (the desired ranking can be achieved by a
simple rearrangement of the indexes), that the m alternatives are arranged in
such a way that the following relation (ranking) is satisfied (that is, the first
alternative is always the best alternative and so on):
PI ~ P2 ~ P3 ••• ~ Pm' (8-1)

This chapter deals with two major problems. The first major problem
is how to determine the most critical criterion in the previous typical MCDM
problem. Intuitively, one may think that the most critical criterion is the
criterion which corresponds to the highest weight Wj' However, this notion
of criticality may be misleading. In this chapter, the most critical criterion
is defined in two alternative ways. In the first way the interest is on whether
the indication of the best (top) alternative changes or not. On the second
definition the interest is on changes in the ranking of any alternative. These
definitions are provided formally in the next section.
In the previous notion of criticality, the term smallest change can also
be defined in two different ways. The first way is to define smallest change
in absolute terms. The second way is to define smallest change in relative
terms.
For instance, suppose that the two criteria C] and C2 have weights w]
= 0.30 and w2 = 0.50, respectively. Furthermore, suppose that when the

first weight becomes w 11 = 0.35, then the existing ranking of the alternatives
changes. Similarly, suppose that when the second weight becomes w 12 =
0.57 (and w] remains equal to 0.30), then the existing ranking of the
alternatives changes. In absolute terms, the first criterion is the most critical
criterion. This is true since the change of the weights for C] is equal to: I w]
- W 11 I = 0.05, while for C2 the corresponding change is: I w2 - W 12 I =
0.07. That is, for the first criterion the critical change is smaller than that for
the second criterion when the changes are computed in absolute terms.
However, when one considers relative terms, then the previous
picture reverses itself. In relative terms, the change of the weights for C] is
equal to: I w] - W I] I x 100/w] = 16.67 %, while for C2 the corresponding
change is: I W 2 - W l2 I X 100/w2 = 14.00%. That is, for the second criterion
the relative change is smaller than that for the first criterion. Thus, when the
relative changes are considered, then the most critical criterion is C2 •
Furthermore, it can be observed that for changes in the ranking of the
alternatives one may view them from two different perspectives as follows.
First, one might be interested to see when a change in the current data causes
any two alternatives to reverse their existing ranking. However, it is also
possible for one to be interested only when the best (top) alternative changes.
For instance, if the problem is to fill a single vacancy, then one is interested
in identifying the best candidate (i.e., the top alternative) from a number of
competing candidates. On the other hand, if the interest is to distribute a
budget to a number of competing needs, then the precise values of the Pi
values are of interest.
Therefore, a total of four alternative definitions can be considered.
These are coded as Absolute Any (AA), Absolute Top (AT), Percent Any
(PA), and Percent Top (PT). This approach, however, might be misleading.
After all, a change, say by 0.03, does not mean much unless someone is also
given the original value. A change of 0.03 is very different if the original
value was equal to 0.08 or 0.80. That is, it is more meaningful to use
relative changes. Therefore, in this chapter the emphasis will be on relative
(percent) changes and thus all developments are based on relative changes.
However, the proposed methodology and the illustrative numerical examples
presented later, do present how one can also derive changes in absolute terms.
It can be noticed that in order for one to derive changes in relative terms,
changes in absolute terms need to be calculated first.
The above notion of critical change is used to determine both the most
critical criterion (which we will call problem 1) and the most critical aij
performance measure (which we will call problem 2). As an extension of
determining the most critical criterion (or the aij performance measure), the
notion of critical change is used to determine how critical each criterion
weight Wi and performance measure aij is. These two problems are examined
in more detail in the subsequent sections.
8.3 PROBLEM 1:
DETERMINING THE MOST CRITICAL CRITERION
8.3.1 Defmitions and Terminology

First we consider the case of defining the changes in the current
weights of the decision criteria (please notice that in the following treatments
some of the indexes contain commas"," in order to enhance reading clarity).
Definition 8.1:
Let Ok,jJ (for 1 ~ i < j ~ m and 1 ~ k ~ n) denote the minimum change
in the current weight W k of criterion Ck such that the ranking of alternatives
Aj and Aj will be reversed.
Next, we define Ofk,jJ as follows:

Ofk,~
.. = Ok",~ X lOO/Wk,
for any 1 ~ i <
j ~ m and 1 ~ k ~ n. (8-2)
That is, the parameter Ofk,jJ expresses changes in relative terms. As it will be
shown later (in Theorem 8-1), it is possible for a given pair of alternatives
and a decision criterion the value of the critical change to be infeasible.
The most critical criterion is next defined in two possible ways (please
recall that from relations (8-1) alternative A I is always assumed to be the best
alternative). The first of these two definitions (Le., Definition 8-2) applies
when one is interested only in changes in the best alternative, while the
second definition (Le., Definition 8-3) applies when one is interested in
changes in the ranking of any alternative. Recall that Is I stands for the
absolute value function (e.g., I-51 = +5).
Definition 8-2:
The Percent-Top (or PT) critical criterion is the criterion which corresponds
to the smallest Iffk'lJI (for 1 ~ j ~ m and 1 ~ k ~ n) value.
Definition 8-3:
The Percent-Any (or PAl critical criterion is the criterion which corresponds
to the smallest Iffk,j,jl (for 1 ~ i < j ~ m and 1 ~ k ~ n) value.
As it was stated in the previous section, in this chapter we adopt the

definitions which correspond to relative changes. Thus, the following two
definitions express how critical a given decision criterion is.
Definition 8-4:
The criticality degree of criterion Cle• denoted as D I", is the smallest percent
amount by which the current value of w" must change, such that the existing
ranking of the alternatives will change. That is, the following relation is true:
Di = min
1:0: i<j :o:m
{I aLJ I}, for all n ~ k ~ l.
Definition 8-5:
The sensitivitv coefficient of criterion Cle• denoted as sens(Ck ), is the
reciprocal of its criticality degree. That is, the following relation is true:
1
sens(Ck ) = -, for any n ~ k ~ 1.
Di
lfthe criticality degree is infeasible (i.e., impossible to change any alternative
rank with any weight change), then the sensitivity coefficient is set to be equal
to zero.
The previous two Definitions 8-4 and 8-5 are based on changes in the
ranking of any alternative. However, one may be interested only in changes
in the ranking only of the best (top) alternative. For instance, in a problem
involving the purchase of a house, the focus is on the best house and the
ranking of all alternative houses may be of secondary interest. In cases like
the above one, one may want to use modifications of the criticality degree and
sensitivity coefficient concepts in which changes are only defined on the
ranking of the best alternative. It can also be observed from the above
definitions that since D Ik is always less than or equal to one, then it follows
that the value of sens(Ck) is always greater than or equal to one (except when
it is set to be equal to zero as described above).
8.3.2 Some Theoretical Results in Determining the Most

Critical Criteria
8.3.2.1 Case (z): Using the WSM or the AHP Method
For this case it is assumed that a decision maker uses the WSM or the
AHP method and he/she wishes to alter the existing ranking of the two
alternatives A] and A2 by modifying only the current weight w] of criterion C].
Currently, the following relation is true (as it was assumed in (8-1»: Pi ;;:::
P2 • In the Appendix of this chapter it is shown that the minimum quantity
0U2' needed to reverse the current ranking of the two alternatives A] and A2 ,
should satisfy the following relation:
(P2 - PI)
al ,l,2 <
(a21 - au)
, or :
(8-3a)
(P2 - PI)
al ,l,2 > -~--=---,
(a21 - au)
Furthermore, the following condition should also be satisfied for the new
weight w*] = w] - OJ,J,2 to be feasible:
o or:
o or: (8-3b)
a1,1,2 ;s; WI'
In these developments it is not required to have w*i ::;; 1 because these weights
are re-normalized to add up to one.
From the last two expressions (8-3a) and (8-3b), it can be seen that
sometimes the quantity OJ,J,2 may not have a feasible value. In other words,
it may be impossible to reverse the existing ranking of the two alternative A]
and A2 by making changes in the current weight w] of criterion C]. This
situation occurs when the value of the ratio:
( P2 - PI )
(a 21 - au)
is greater than w].
The previous considerations can be easily generalized and thus lead
to the proof of the following theorem which covers the general case (recall
that currently the following relation is assumed to be true from (8-1): Pi;;:::
Pj , for all 1 ::;; i ::;; j ::;; m).
Theorem 8-1:
When the WSM, or the AHP method is used, the quantity O/k. ;. j (jor 1 :::; i
< j :::; m and 1 :::; k :::; m), by which the current weight W k of criterion Ck
needs to be modified (after normalization) so that the ranking of the
alternatives A; and Aj will be reversed, is given as follows:
(Pj - Pi) 100

-~--'-x--,
(ajk - aik) Wk
(8-4a)
I (Pj - Pi) 100
Ok"
.IJ > --"---- x - - ,
(ajk - aik) Wk
Furthermore, the following condition should also be satisfied for the value of
o/k,;,j to be feasible:
(Pj - Pi)
(8-4b)
(a jk - aik)
From the previous considerations it can be seen that if alternative A;

dominates alternative Aj (i.e., if a ik ~ ajk, for all k =1, 2, ... , n) then, it is
impossible to make alternative Aj more preferred than alternative A; by
changing the weights of the criteria. Also, a criterion Ck is a robust
criterion if all () /k,;,j (jor 1 :::; i < j :::; m and 1 :::; k :::; n) quantities
associated with it are infeasible. In other words, if condition (8-4b) is
violated for all i, j = 1,2,3, ... , m, for some criterion Ck then, any change
in the weight of that criterion does not affect the existing ranking of any of
the alternatives and thus this criterion is a robust one and consequently it can
be dropped from further consideration (since it does not provide any
discriminatory power in ranking the alternatives).
Therefore, if one is interested in determining the most critical
criterion, then all possible {)/k,;,j (for 1 :::; i < j :::; m and 1 :::; k :::; n) values
need to be calculated. It can be observed here that there are n X (m(m-l))/2
such possible () /k,;,j quantities. This issue is further illustrated in the following
extensive numerical example.
8.3.2.2 An Extensive Numerical Example for the WSM Case
Consider an MCDM problem defined on the four alternatives AI' A z,

A3 , and A4 and the four decision criteria C1, Cz, C3 , and C4. Suppose that the
following Table 8-1 is its corresponding decision matrix when the WSM (or
the AHP with one hierarchical level) method is used. Note that the data were
normalized to add up to one, although this is not required by WSM (however,

it is required by the AHP).
Table 8-1: Decision Matrix for the Numerical Example on the WSM.
Criteria
CI C2 Cj C4
Alts. (0.3277 0.3058 0.2876 0.0790)
Al 0.3088 0.2897 0.3867 0.1922

A2 0.2163 0.3458 0.1755 0.6288
Aj 0.4509 0.2473 0.1194 0.0575
A4 0.0240 0.1172 0.3184 0.1215
Suppose that we want to apply the WSM (the case of the AHP is
identical since the data are already normalized). Then, by using formula (2-
1), in Section 2.2.1, the final preferences and the corresponding ranking of
the four alternatives are as shown in Table 8-2.
Table 8-2: Current Final Preferences.
I Alternative
I Preference (P)
I Ranking
I
Al 0.3162 1*
A2 0.2768 2
Aj 0.2621 3
A4 0.1449 4
Note: * indicates the most preferred (best) alternative.
Therefore, the relation PI ~ P 2 ~ P3 ~ P4 holds and as a result the

most preferred alternative is AI' It can be observed now that according to the
weights of the four criteria, criterion C1 appears to be the most important
one. The minimum change 0/,/.3 needed to alter the current weight WI' so that
the current ranking of the two alternatives Al and A3 will be reversed, can be
found by using relation (8-4a) of Theorem 8-1 as follows:
(0.2621 - 0.3162)
l>113 < , or
,, (0.4509 - 0.3088)
l> 1,1,3 < - 0.3807.

This quantity -0.3807 satisfies condition (8-4b), as it is less than WI (=

0.3277). Thus, the modified weight W*I of the first criterion (before
normalization) for this case is equal to:
w*J = [0.3277 - (-0.3807)] = 0.7084.
Working as above for all possible combinations of criteria and pairs of
alternatives, Table 8-3 is derived. Next, Table 8-4 depicts the changes in
relative terms (that is, the flk,i,j values as are computed by using relation (8-
4a) of Theorem 8-1). It can be observed that negative changes in Table 8-3
indicate increases, while positive changes indicate decreases. Also it can be
noticed that the changes (either percentages or in absolute terms) are given
before normalization. The boldfaced numbers in either table indicate
minimum critical changes (as it is explained in the following paragraphs).
The Percent-Top (PT) critical criterion can be found by looking for
the smallest relative value of all rows which are related to alternative A, (i.e.,
the best alternative) in Table 8-4. The smallest such percentage (i.e.,
64.8818%) corresponds to criterion C3 when the pair of the alternatives Al
and A2 is considered. For criterion C3 a reduction of its current weight by
64.8818% will make A2 to be the most preferred alternative and Al will not
be the best alternative any more.
Table 8-3: All possible fJk.i,j Values (Absolute Change in Criteria Weights).
Pair of Criterion
Alternatives C1 C2 CJ C4
A 1 - A2 N/F -0.7023 0.1866 -0.0902
A 1 - AJ -0.3807 N/F 0.2024 NIF

A 1 - A4 NIF NIF N/F NIF
A2 - AJ -0.0627 0.1492 0.0262 0.0257
A2 - A4 NIF NIF -0.9230 N/F

A J - A4 0.2745 NIF -0.5890 -1.8313
Note: N/F stands for Non-Feasible. That is, the corresponding fJ value does not
satisfy condition (8-4b).
The Percent-Any (PA) critical criterion can be found by looking for

the smallest relative rlk,i,j value in the entire Table 8-4. Such smallest value
is {/3,2,3 = 9.1099 % and it (again) corresponds to criterion C3. Therefore,
the PA critical criterion is C3 • Finally, it should be noticed here that it is a
coincidence that both definitions of the most critical criterion indicate the
same criterion (i.e., criterion C3) in this numerical example.
At this point it should be stated that if a decision maker wishes to
define the most critical criterion in absolute changes, then the previous two
definitions of Percent-Top (PT) and Percent-Any (PA) critical criterion
correspond to the Absolute-Top (AT) and the Absolute-Any (AA) critical
criterion, respectively. From Table 8-3 it can be easily verified that the AT
criterion is C4 and also, by coincidence, the AA criterion is C4 (the
corresponding minimum changes are boldfaced). Later, in Section 8.3.3,
some computational results indicate how frequently various alternative
definitions of the most critical criterion may point out to the same criterion.
When Definition 8-4 is used, then from Table 8-4 it follows that the
criticality degrees of the four criteria are: VI = 1-19.13341 = 19.1334, V z
= 48.7901, V3 = 9.1099, and V 4 = 32.5317. Therefore, the sensitivity
coefficients of the four decision criteria (according to Definition 8-5) are:
sens(CI ) = 0.0523, sens(Cz) = 0.0205, sens(C3) = 0.1098, and sens(C4) =
0.0307. That is, the most sensitive decision criterion is C3 , followed by the
criteria C I , C4 , and Cz in that order.
Table 8-4: All possible eik "ij Values (Percent Change in Criteria Weights).
Pair of Criterion
Alternatives CI C2 C3 C4
Al - A2 N/F -229.7 64.8818 -114.1772
Al - A3 -116.1733 N/F 70.3755 N/F
Al - A4 N/F N/F N/F N/F
A2 - A3 -19.1334 48.7901 9.1099 32.5317
A2 - A4 N/F N/F -320.9 N/F
A3 - A4 83.7656 N/F -204.8 -2,318.10
Note: N/F stands for Non-Feasible. That is, the corresponding ()' value does
not satisfy condition (8-4b).
8.3.2.3 Case (iz): Using the WPM Method
Suppose that we are interested in determining the quantity (\1,2 when

the WPM method is used. Recall that according to relation (2-2), in Section
2.2.2, alternative Al is more preferred than alternative A2 when the following
ratio is greater than or equal to one:
R A2
IT
(A1] -_ i=l (aa1i ]Wl .
2i
(8-5)
Furthermore, according to relation (8-1), it is currently assumed that

PI ~ P2 • Let p/ and p/ denote the new preferences of these two
alternatives. Then, when the ranking of these two alternatives is reversed, the
relation on the preferences becomes: p/ < P/' In the Appendix of this
chapter it is shown that the quantity 01,1,2 must satisfy the following
condition:
log ( IT (a 1y ]W' 1
Y =1 ~Y
5 1,1,2 > (8-6)
Cl all]
loaf
a21
The right-hand-side of the last relationship gives the minimum change
value needed to modify the current weight WI of criterion CI such that
alternative A2 will become more preferred than alternative Al (in the
maximization case). Similarly as in the previous section, this quantity needs
to satisfy condition (8-4b). The previous considerations can be easily
generalized and thus lead to the proof of the following theorem:
Theorem 8-2:
When the WPM method is used, the critical quantity 0 'k,iJ (jor 1 :::;; i < j :::;;
m and 1 :::;; k :::;; n), by which the current weight W k of criterion Ck needs to
be modified (after normalization) so that the ranking of the alternatives Ai and
Aj will be reversed, is given as follows:
5~ij > K, if K ~ 0, or:

a~ij < K, otherwise.
where the parameter K is defined as follows:
log :iI
y = 1
a iy
a jy
w" 1 100
K = (8-7a)
loafail:) x Wl; •
"ajl:
Furthermore, the following constraint should also be satisfied:
a~.. ~ 100. (8-7b)

....IJ
Similarly to Theorem 8-1, in order to determine the most critical

criterion a total of n xm(m - 1) critical changes (Le., f/k,i,j values) need to be
calculated. All the previous theoretical considerations for the WPM model
are further illustrated in the following extensive numerical example.
8.3.2.4 An Extensive Numerical Example for the WPM Case
Consider a decision making problem with the four alternatives AI' A2,
A3 , and A4 and the four decision criteria el , e2, e3 , and e4 • Please note that
this numerical example is different than the first one (and the two examples
which follow), in order to provide a wider exposure of various numerical
scenarios. Also, the decision matrices in these extensive numerical examples
are square (Le., m = n) of mere coincidence. The proposed procedures can
be applied to any size of decision matrix without any modification at all.
Next, suppose that Table 8-5 depicts the decision matrix for this numerical
example.
Table 8-5: Decision Matrix for the Numerical Example on the WPM.
Criteria
CI C2 C3 C4
Alts. (0.4504 0.1231 0.0848 0.3417)
Al 0.9381 0.3501 0.8811 0.5646

A2 0.7691 0.4812 0.1679 0.9336
A3 0.9445 0.1138 0.2219 0.0135
A4 0.1768 0.0221 0.9462 0.1024
Then, by applying the WPM approach (that is, by using formula (2-2)
in Section 2.2.2) the ranking of the four alternatives is as shown in Table 8-6.
The product expressed by (2-2) for alternative A] is greater than one for all
possible combinations which include A], thus, the most preferred alternative
is A]. Also, according to the weights of the four criteria, criterion C]
appears to be the most important one, because this is the criterion with the
highest weight.
Next, we consider the minimum quantity needed to alter the current
weight w4 , so that the current ranking of the two alternatives A1 and A2 will
be reversed. This quantity, denoted as K (and expressed as %), can be found
by using relation (8-7a) of Theorem 8-2 as follows:
10J( 0.9381 )0.4504( 0.3501 )0.1231 ... (0.5646)0.3417)

K _~
o~ 0.7692 0.4812
_ _---'_ _' - -_ _ L . -_ _ 0.9336
--'~_---'~--Lx _100
__
loj 0.5646) 0.3417

6~ 0.9336
= -11.04.
Table 8-6: Current Ranking.
Pair of R(AJAj ) Ratio Ranking

Alternatives Aj - Aj
Al - A2 1.0192 Al -+ I-st *
AJ - Aj 4.6082 A2 -+ 2-nd
Al - A4 5.3062 Aj -+ 3-rd
A2 - Aj 4.5216 A4 -+ 4-th
A2 - A4 5.2065
A3 - A4 1.1515
Note: * indicates the most preferred alternative (in the maximization case).
Therefore, the value of 0/4,1,2 should be smaller than K = - 11.04.

Note that this is a feasible value since it can easily be verified that it satisfies
condition (8-7b). In a similar manner, all possible K values can be
determined. These values are depicted in Table 8-7 (the boldfaced number
corresponds to the minimum change).
Table 8-7: All Possible K Values for the WPM Example.
Pair of Criterion
Alternatives C1 C] Cj C4
Al - A] 21.21 -48.44 13.50 -11.04
Al - Aj N/F N/F N/F N/F

Al - A4 N/F N/F N/F N/F
A] - Aj N/F N/F N/F N/F
A] - A4 N/F N/F N/F N/F
AJ - A4 18.69 69.97 -114.72 -20.37
Note: N/F stands for non-feasible. That is, the corresponding K value does not satisfy
condition (8-7b).
It is interesting to observe here that the PT, and P A critical criteria

happened to point out to the same criterion (i.e., criterion C4). However, the
criterion with the highest weight is criterion C]. Clearly, this is a counter-
intuitive (i.e., another paradox) conclusion. Also, when Definition 8-4 is
used, then from Table 8-7 it follows that the criticality degrees of the four
criteria are: Ii] = 18.69, Ii2 = 1-48.441 = 48.44, Ii3 = 13.50, and Ii4
= 1-11.041 = 11.04. Therefore, the sensitivity coefficients of the four
decision criteria (according to Definition 8-5) are: sens(C]) = 0.0535,
sens(C2) = 0.0206, sens(C3 ) = 0.0741, and sens(C4 ) = 0.0906. That is, the
most sensitive decision criterion is C4 , followed by criteria C3, C j , and C2 in
that order.
8.3.3 Some Computational Experiments

A computational study was undertaken to study how often the PT and
PA critical criteria were the criteria with the highest or with the lowest
weight. For that reason, random MCDM problems were generated and the
PT and PA critical criteria were determined. For the AHP case (only) the
data for these problems were generated in a manner similar to the procedure
used in [Triantaphyllou, Pardalos and Mann, 1990a], [Triantaphyllou,
Lootsma, et al., 1994], [Triantaphyllou and Mann, 1990] and also in most of
the experiments described in this book (as described in the following
paragraphs). This experimental strategy ensures that the problems are
generated completely randomly. For the WPM and WSM cases the data were
generated randomly from the uniform distribution in the interval [9, 1].
According to the~~test problem generation approach described in
et
[Triantaphyllou, Lootsma, al., 1994], or in Chapter 3 of this book, the data
were generated as follows. First a random weight vector W was generated
such that the ratio of the largest to the smallest element was less than 9 (in
order to comply with the values in the Saaty scale which was used here).
From these weights the entries of the matrix with the actual pairwise
comparisons were determined by using the relationship aij = w/Wj. It is
assumed that the decision maker does not know these values. This matrix is
called in [Triantaphyllou, Lootsma, et al., 1994] and also in Chapter 3 the
Real and Continuous Pairwise (RCP) matrix. However, it was again
assumed here that the decision maker is capable of estimating the entries of
the RCP matrix by forming a matrix in which each entry in the RCP matrix
is replaced by a number which is as close as possible to the values allowed
in the traditional Saaty scale (i.e., the numbers from the set {9, ... , 2, 1,
1/2, ... , 1/9}). This is the Closest and Discrete Pairwise (CDP) matrix (see
also Chapter 3). Next, the eigenvector of the COP matrix is estimated as
before and the corresponding vector of the decisioll: matrix is formed.
For instance, if the real (and hence unknowri1 performance values of
three alternatives in terms of a single criterion are: (0.77348, 0.23804,
0.23848), then, the (1,3) entry of the corresponding RCP matrix is equal to
3.24342 (= 0.77348/0.23848). Thus, the corresponding COP element will
be equal to 3 (because this value is the closest one from the Saaty scale
values: {9, ... , 2, 1, 1/2, ... , 1/9}). More on this approach and some
interesting properties of the COP matrices can be found in [Triantaphyllou,
Lootsma, et al., 1994] and also in Chapter 3.
Two parameters were considered in these test problems. The first
parameter was the number of decision criteria. The second parameter was the
number of alternatives. The number of decision criteria was equal to 3, 5,
7, ... , 21. Similarly, the number of alternatives was equal to 3, 5, 7, ... ,
21. In this way we formed 100 (= 10 x 10) different combinations of
numbers of criteria and alternatives and 1,000 random test MCOM problems
were generated for each such combination. This simulation program was
written in Fortran, using the IMSL library of subroutines for generating
random numbers. These results are depicted in Figures 8-1 to 8-12.
For each test problem we examined whether the PA or the PT critical
criterion was the criterion with the highest or the criterion with the lowest
weight. The results of the computational experiments, when the relative
(percent) changes are considered, are depicted in Figures 8-1 to 8-4. Figures
8-5 to 8-8 illustrate the same concepts but when changes are expressed in
absolute terms. Finally, Figure 8-9 depicts some specific results when the
WPM is used. In the present study, we solved each problem using the WSM,
the WPM, the AHP and the ideal mode (revised) AHP method.
The four curves in each figure represent the results from each one of
the three different MCDM methods used plus one curve for the ideal mode
AHP. The most profound observation is that all MCDM methods generated
almost identical results. This is indicated by the fact that their curves in
Figures 8-1 to 8-8 are very close to each other.
Figures 8-1 to 8-8 indicate that it makes a significant difference
whether critical changes are expressed as percentages (Le., in relative terms)
or in absolute terms. When changes are expressed as percentages, then more
frequently the criterion with the highest weight is the most critical criterion.
This is true both when the concept of the critical criterion is defined in terms
of changes in the ranking of the top alternative or in terms of changes in the
ranking of any alternative. This is evident when one compares Figure 8-1
with Figure 8-2 and Figure 8-3 with Figure 8-4. The reverse situation occurs
when one defines change in absolute changes. That is, now more frequently
the most critical criterion is the criterion with the lowest weight. Figures 8-5
to 8-8 depict the corresponding results.
As it was anticipated, the sensitivity importance of any weight
(including the highest or the lowest) reduces gradually as the number of
decision criteria in a problem increases. In a matter of fact, when changes
are measured in relative terms (Le., as a percentage), then the lowest weight
is hardly ever sensitive in problems with more than 10 criteria (see also
Figures 8-2 and 8-4). On the other hand, the number of alternatives has only
a minor practical influence. This is indicated in Figures 8-11 and 8-12 in
which the bottom curve corresponds to problems with 3 alternatives and the
top curve to problems with 21 alternatives.
The question which is raised at this point is what kind of changes a
decision maker should consider: The ones defined as percentages or the
ones defined in absolute tenus? One may argue here that percentage
changes are the most meaningful ones. After all, a change, say of 0.03, does
not mean much unless one also considers what was the initial value (for
instance, was the initial value equal to 0.95 or to 0.05?).
Figure 8-9 depicts how frequently the AT and PT definitions pointed
out to the same criterion. Please recall that this situation also occurred in
some of the illustrative examples analyzed earlier.
Similarly, Figure 8-10 depicts how frequently the AA and PA
definitions pointed out to the same criterion. As expected, the frequency of
matching the top rank is always higher than matching all ranks. Moreover,
this distinction fades away as the number of criteria in a problem increases.
Finally, Figure 8-11 depicts how frequently all alternative definitions (i.e.,
AT, PT, AA, and PA) pointed out to the same criterion when the WSM
model was used (the other models yielded similar results).

These computational results indicate that the previous coincidence
rates are rather high (Le., around 70 to 50 or 40 %) when the number of
decision criteria in a problem is rather small (Le., less than 7 or 9).
Therefore, if the number of decision criteria is small, one may not have to be
concerned on which definition to use. Finally, notice that the number of
alternatives in a test problem did not seem to be important. This is indicated
by the closeness of the curves (which correspond to problems with different
numbers of alternatives) in Figure 8-12 (which shows a particular set of
results when the WPM is used, the other methods yielded similar patterns).
55
50
45
40 ~
~
>- 30
~
..
______ MCDM methods
~
o
c
~ 25
~
0-
m
20
~~
~ ~
15 ---- ~
10 --------------
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-1: Frequency of the time that the PT Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40
~35
"--'
~
15 ~
10 ~ L MCDM methods
~,
5
o ~-
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-2: Frequency of the time that the PT Critical Criterion is the Criterion with
the Lowest Weight.
55
50
0,
45
40 ~
~
>- 30
~ ~ MCDM methods
~
o
c
~ 25
0-
OJ
~ 20
~
15
---- ~
10
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-3: Frequency of the time that the PA Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40
"
/ MCDM methods
15
/
10
5
\,
o ~
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-4: Frequency of the time that the PA Critical Criterion is the Criterion with
the Lowest Weight.
55
50
45
40
>- 30
u ./ MCDM methods
c
~ 25
0-
Q) ~~
.: 20
15
~
10
~
~
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-5: Frequency of the time that the AT Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40 \\
\\
\\
\\ MCDM methods
,~ /
15 ~'Y
10
~
~~
5 =
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-6: Frequency of the time that the AT Critical Criterion is the Criterion with
the Lowest Weight.
55
50
45
40
~35
.........
>- 30
()
c
~ 25
(J
OJ
~ 20
/ MCDM methods
15
10
~
5
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-7: Frequency of the time that the AA Critical Criterion is the Criterion with
the Highest Weight.
55
I
50
45 \
40
\
\
\
~35
.......
>-30
()
c \. / MCDM methods
~ 25
(J
OJ
~ 20
,,/
15 ~
10
~
~
5
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-8: Frequency of the time that the AA Critical Criterion is the Criterion with
the Lowest Weight.
100
90
80
70
r--.
~
"-" 60 ~ / MCDM methods
>-
u
c 50 ~
-----
OJ ~
::J
cr
OJ
L.
40 --
L.-
30
20
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-9: Frequency of the time that the AT and PT Definitions point to the Same
Criterion.
100
90
80
70
r--.
MCDM methods
~ 60 /
>-
u
c
(l)
50 ~ ,/
-
::J
cr
-===;
40
::::----::::
Q)
....
LL
30
20
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-10: Frequency of the time that the AA and PA Defmitions point to the Same
Criterion.
55
50
45
40
~35
'--'
No of alternatives
>- 30
u
c
~ 25
0-
OJ
~ 20
15
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-11: Frequency of the time that the AT, PT, AA, and PA Definitions point to
the Same Criterion Under the WSM Method.
55
50
45
40
~35
'--'
No of alternatives
>- 30
u
c
~ 25
0-
OJ
~ 20
15
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-12: Rate the AT Critical Criterion is the one with the Lowest Weight for
Different Size of Problems Under the WPM Method.
8.4 PROBLEM 2:
DETERMINING THE MOST CRITICAL au
MEASURE OF PERFORMANCE
8.4.1 Definitions and Terminology

The second major problem examined in this chapter is how to
determine the most critical aij measure of performance when the WSM,
AHP, or the WPM method is used. The following definitions are pertinent
to this problem.
Definition 8-6:
Let Tij,k (jor 1 : : :; i < k : : :; m and 1 : : :; j : : :; n) denote the threshold value of
!!g, which is the minimum change which has to occur in the current value of
aij such that the current ranking between alternatives A; and Ak will change.
Since there are m alternatives, each aij performance measure is

associated with a total of (m-l) such threshold values. In a similar way as
earlier regarding the definition of the c/k,;,j values, one can also consider
threshold values expressed in relative terms. We denote these relative term
threshold values as <,j,k' That is:
7/"k
I,j, = 7""k
I.j, x IOO/a""Y'
for any 1 : : :; i, k :::::;; m, and 1 : : :; j : : :; n. (8-8)
For the reasons explained earlier, when we consider threshold values

we will mean the ones defined in relative terms (i.e., the 7/;,j,k values). Given
the previous notion of threshold value, we define as the most sensitive
alternative the one which is associated with the smallest threshold value.
Also as before, one may be interested in changes in the ranking of (only) the
best alternative, or in changes in the ranking of any alternative.
As it was mentioned above, there are (m-l) possible threshold values
7/;,j,k (where i;;t.k, 1 : : :; i, k : : :; m, and 1 : : :; j : : :; n) for any aij measure of
performance (how to calculate these threshold values is given later in

Theorems 8-3, 8-4, and 8-5). The following three definitions are related to
the previous notion of threshold values 7 /;,j,k' It can be noticed that
analogous definitions are possible if one is interested in changes only in the
top alternative (as was the case earlier with Definitions 8-4 and 8-5).
Definition 8-7:
The criticality degree of alternative Ai in terms of criterion~, denoted as
~ lij' is the smallest amount (%) by which the current value of aij must change,
such that the existing ranldng of alternative Ai will change. That is, the
following relation is true:
a~j '" min { I 1:L.k I },

k .. i (8-9)
for all m ~ i ~ 1, and n ~ j ~ 1.
Definition 8-8:
Alternative AL is the most critical alternative if it is associated with the
smallest criticality degree. That is, if and only if the following relation is
true:
a~k = min {min {a"'}}

m~i2:1 n>j2:1 I) ,
(8-10)
for some n ~ k ~ 1.
Definition 8-9:
The sensitivity coefficient of alternative Ai in terms of criterion C, denoted J
as sens(aij)' is the reciprocal of its criticality degree. That is, the following
condition is true:
-,'
1
dij (8-11)
for any m ~ i ~ 1, and n ~ j ~ 1.
If the criticality degree is infeasible, then the sensitivity coefficient is set to be
equal to zero.
From Definition 8-7 it follows that the smaller the criticality degree
~ lij
is, the easier is for the ranking of alternative Ai to change. Alternatively,
Definition 8-9 indicates that ranking changes are easier, as the sensitivity
coefficients sens(aij) are higher. Finally, when Definition 8-8 is combined
with Definitions 8-7 and 8-9, it indicates that the most sensitive alternative is
the one with the highest sensitivity coefficient. The next section shows how
to calculate these new important terms.
8.4.2 Determining the Threshold Value T /iJ,k
8.4.2.1 Case (I): When Using the WSM or the AHP Method
The Appendix of this chapter presents the highlights for a proof for
Theorem 8-3. This theorem provides the main formula used to calculate the
threshold 7 / iJ,k values and it is stated next.
Theorem 8-3:
When the WSM method is used, the threshold value 7 /iJ'k (in %) by which the
pelj'ormance measure of alternative Ai in terms of criterion ~, denoted as ai}'
needs to be modified so that the ranking of the alternatives Ai and Ak will be
reversed, is given as follows:
't.' .
1.1. k
< R, when i < k, or :
(8-lOa)
't .' .
1.1. k > R, when i > k.
where R is defined as follows:
100
R = x --.
Wj aij
Furthermore, the following condition should also be satisfied for the threshold
value to be feasible:
-e.t. :.,; 100. (8-lOb)

1.1. k
From the requirement to have: 0 ~ ail - 7 i,j,k and relation (8-8), the
new condition 0 ~ ail - ail X 7 /iJ,,JlOO is derived. Next, this leads to condition
(8-lOb). For the case of the AHP method it can be easily shown (see also the
Appendix of this chapter) that the corresponding theorem is as follows:
Theorem 8-4:
When the AHP method is used, the threshold value 7 /iJ,k (in %) by which the
measure of pelj'ormance of alternative Ai in terms of criterion ~ needs to be
modified so that the ranking of alternatives Ai and Ak will change, is given as
follows:
(Pi - P k ) 100
• .1. ----------------------x---. (8-12a)
'." k [Pi - P k + wAakj - aij + 1)] aij
Furthermore, the following condition should also be satisfied for the threshold
value to be feasible:
.~.
',J. k
:s; 100. (8-12b)
The sensitivity analysis of the aij values, when the WSM or the AHP method
is used, is next demonstrated in terms of an extensive numerical example.
8.4.2.2 An Extensive Numerical Example When the WSM or

the AHP Method is Used
Consider the decision matrix depicted in Table 8-8 (along with the
corresponding final preferences P;) of an application of the WSM model.
This illustrative problem has five alternatives and five decision criteria). The
AHP case can be developed in an analogous fashion.
Table 8-8: Decision Matrix and Initial Preferences for the Example.
Criteria
C, C, C3 C4 Cs Preference Values
Alts. (0.4146 0.0129 0.2958 0.0604 0.2164) Pi
A, 0.3576 0.2483 0.2899 0.2961 0.3202 0.3244

Al 0.3603 0.2836 0.0407 0.0939 0.0172 0.1745
A3 0.0255 0.1745 0.2895 0.2212 0.2641 0.1690
A4 0.1609 0.2008 0.2960 0.0716 0.0315 0.1680
As 0.0957 0.0928 0.0839 0.3172 0.3670 0.1643
When Theorem 8-4 is used on the above data, the corresponding 7 / i.j,k
threshold values are as in Table 8-9. The boldfaced entries in Table 8-9
correspond to the criticality degrees !l. lij (i.e., the smallest entry per column
in each row section, as given in Definition 8-7). The criticality degrees are
next summarized in Table 8-10.
To help interpret the entries in Table 8-10, consider anyone of them,
say entry (3,1) (i.e., 89.3). This entry indicates that 7 11,1,4 = 89.3%. That
is, the measure of performance a11 must be decreased by 89.3 % from its
current value (i.e., 0.3576) to (1 - 0.893)x0.3576, in order for aiternativeA4
(which is shown on the right column in this table) to become more preferred
than alternative Al (note that currently Al is more preferred than A4)' A
similar interpretation holds for the rest of the entries. Note that some of the
entries in the previous table are marked as NIP, because they correspond to
infeasible values (Le., condition (8-12b) in Theorem 8-4 is violated).
It can be noticed that the entries in Table 8-9 are greater than 100
only when the sign is negative. Recall that negative changes in reality mean
increases. If a rating becomes greater than 100, that is all right. In the case
of the criteria weights the numbers will be re-normalized to add up to one. In
the case of the au performance measures, the numbers may be re-normalized
(for instance, in the AHP model) or they may become greater than one (for
instance, in the WPM or WSM models). As before, the boldfaced numbers
indicate minimum values.
Table 8-9: Threshold values T 'jJ.k (%) in Relative Terms for the WSM/AHP Example.
Criterion 0
Alt.
Alt. (AJ C} Cz CJ C4 Cs At
A} 74.1 N/F N/F N/F N/F Az
A} N/F N/F N/F N/F N/F AJ
Al 89.3 N/F N/F N/F N/F A4
Al 96.1 N/F N/F N/F N/F As
------- --------------------------------------- --------
Az -157.9 N/F -1,677.6 N/F N/F Al
Az 5.4 N/F 35.9 79.4 N/F A3
Az 5.3 N/F 41.9 N/F N/F A4
Az 8.9 N/F 78.4 N/F N/F As
------- -------------------------------------- --------
AJ -1,535.7 N/F N/F N/F N/F Al
A3 -39.3 -355.2 -8.7 -52.5 -13.2 Az
AJ 8.0 38.9 1.1 8.3 2.2 A4
AJ 41.1 N/F 6.8 29.9 7.3 As
------- --------------------------------------- --------
A4 -286.1 N/F N/F N/F N/F Al
A4 -8.2 N/F -10.2 -163.0 -98.9 Az
A4 -1.7 -41.4 -1.1 -19.7 -11.5 A3
A4 5.9 N/F 5.3 65.9 40.4 As
------- ----------------------------------- --------
As -460.8 N/F -970.6 N/F N/F Al
As -20.7 N/F -44.4 -87.4 -21.2 Az
As -12.8 -544.7 -15.9 -29.7 -6.7 AJ
As 10.0 N/F 18.7 14.9 3.5 A4
Table 8-9 was next used to derive Table 8-10. From Table 8-10 it
follows that the most critical alternatives (according to Definition 8-8) are
alternatives A3 and A4 • This is true because these alternatives correspond to
the minimum criticality degrees (which happened to be identical and equal to
1.1) among all the values in Table 8-10. It can be noticed that the
corresponding alternatives from the right column in Table 8-9 are now within
the parentheses in the entries in Table 8-10. As before, boldfaced numbers
represent the corresponding minimum values. Finally, Table 8-11 presents
the various sensitivity coefficients (as given in Definition 8-9). It can be
noticed that if any entry in Table 8-10 was infeasible (Le., it was denoted by
the "_" symbol), then the corresponding sensitivity coefficient in Table 8-
11 is defined to be equal to zero.
Table 8-10: Criticality Degrees A lij (%) for each aij Performance Measure.
Criterion Cj
CI C1 C3 C4 Cs
AIt.(AJ
AI 74. 1(A 2 ) - - - -
Al 5.3(A.) - 35.9(A3) 79.4(A3)
-
A3 8.0(A4) 38.9(A4) 1. 1(A4) 8.3(A4) 2.2(A.)
A4 1.7(A3) 41.4(A 3) -1.1 (A 3) 19.7(A3) 11.5(A3)
As IO.O(A.) 544.8(A3) 15.9(A3) 14.9(A.) 3.5(A4)
Table 8-11: Sensitivity Coefficients sens(aij) for each aij Performance Measure.
Criterion Cj
CI C2 C3 C4 Cs
Alt.(AJ
AI O.014(A 2 ) 0 0 0 0
A2 O.189(A.) 0 O.028(A 3) O.013(A) 0
A3 O.125(A.) O.026(A 4) O.909(A.) O.121(A4) O.455(A.)
A4 O.588(Aj) O.024(Aj ) O.909(Aj) O.051(A3) O.087(A 3)
As O.l00(A.) O.OO5(Aj) O.063(A3) O.067(A.) O.286(A.)

8.4.2.3 Case (if): When Using the WPM Method
The Appendix of this chapter also presents the highlights for a proof
for Theorem 8-5. This theorem provides the main formula for calculating the
threshold values r/i,j,k when the WPM method is used and it is stated next.
Theorem 8-5:
When the WPM model is used, then the threshold value r/i,j,k (in %) by which
the performance measure of alternative Ai in terms of criterion ~, denoted as
aij' needs to be modified so that the ranking of the alternatives Ai and Ak will
be reversed, is given as follows:
I
t: ij,k > Q, when i > k, or:
(8-14a)
I
t: iJ,k < Q, when i < k.
where Q is defined as follows:
Furthermore, the following condition should also be satisfied for the value to
be feasible:
(8-14b)
The application of the previous theorem is illustrated in the following

extensive numerical example.
8.4.2.4 An Extensive Example When the WPM Method is Used
Consider an MCDM problem which involves the five alternatives Ai'

A z, A 3, A4 and A5 and the five decision criteria Ci , Cz, C3, C4, and C5.
Suppose that Table 8-12 presents its corresponding decision matrix and the
WPM model is to be used:
Table 8-12: Decision Matrix for Numerical Example.
Criteria
CI C2 C3 C4 Cs
Alts. (0.2363 0.1998 0.0491 0.2695 0.2453)
AI 0.8366 0.5001 0.8179 0.8104 0.6951

A2 0.4307 0.4782 0.9407 0.2062 0.9259
Aj 0.7755 0.5548 0.6380 0.3407 0.0514
A4 0.3727 0.7447 0.3214 0.3709 0.0550
As 0.4259 0.7126 0.2195 0.0470 0.0014
Please recall that in the WPM method normalization of the aij values is not
required. Then, by applying formula (2-2) (in Section 2.2.2), the current
ranking of the alternatives is as shown in Table 8-13.
Table 8-13: Initial Ranking.
Pair of
Alternatives Ratio
Ap - Aq R(ApIAq) Ranking
AI - A2 1.580 A J .... I-st
AI - Aj 2.415 A2 .... 2-nd
AJ - A4 2.692 Aj .... 3-rd
AI - As 6.152 A4 .... 4-th
A2 - A3 1.529 A5 .... 5-th
A2 - A4 1.704
A2 - As 3.893
Aj - A4 1.115
A3 - As 2.547
A4 - As 2.285
From Table 8-13 it follows that relation PI ~ Pz ~ P3 ~ P4 ~ P5

holds and as a result the most preferred alternative is AI' When Theorem 8-5
(i.e., formulas (8-14a) and (8-14b» is applied on the previous data, then
Table 8-14 with all possible threshold values 7/i,j,k is derived. The entries in
Table 8-14 have a similar interpretation as the ones in Table 8-9.
Some entries in Table 8-14 are represented in standard exponential
format. This happens because they correspond to very high (negative) values.
For instance, the entry (8,1), (Le., -2E +05) actually represents the value: -
2.0x 105 = -200,000. It can be argued here that very significant changes
(such as the ones represented in exponential format or those which measure
in terms of thousands of % change) are not realistic and thus practically they
can also be classified as "NIF" (e.g., Non-Feasible) cases. Finally, it can be
observed that the highlighted entries in Table 8-14 correspond to the criticality
degrees Ll lij (as given by Definition 8-7). The criticality degrees are best
summarized in Table 8-15.
Table 8-14: Threshold Values 7 /;jk (%) in Relative Terms for the WPM Example.
"
Criterion ~
Alt.
AIt(A,) CI C2 C3 C4 Cs At
Al 85.6 89.9 N/F 81.7 84.5 A2
Al 97.6 98.8 N/F 96.2 97.3 A3
Al 98.5 99.3 N/F 97.5 98.2 A4
Al N/F N/F N/F As N/F N/F
------- ----------------------------------- ------
A2 -593 -887 -IE+06 -446 -546 Al
A2 83.4 88.0 N/F 79.3 82.3 A3
A2 89.5 93.0 N/F 86.2 88.6 A4
A2 99.7 99.9 N/F 99.4 99.6 As
------- ----------------------------------- ------
A3 -4,072 -8156 -6E+09 -2,538 -3,540 Al
A3 -502 -736 -6E+05 -383 -464 A2
A3 36.8 41.9 89.0 33.1 35.7 A4
A3 98.1 99.1 N/F 96.9 97.8 As
------- ----------------------------------- ------
A4 -6,501 -14,105 -6E+1O -3,844 -5,562 Al
A4 -853 -1,339 -5E+06 -622 -777 A2
A4 -58 -72 -811 -50 -56 A3
A4 97.0 98.4 N/F 95.3 96.6 As
------- ----------------------------------- ------
As -2E+05 -9E+05 -IE+ 18 -8E+04 -2E+05 Al
As -3E+04 -9E+04 -IE+ 14 -2E+04 -3E+04 A2
As -5,124 -10,672 -2E+1O -3,113 -4,420 A3
As -3,202 -6,161 -2E+09 -2,049 -2,805 A4
From Table 8-15 it follows that the most critical alternative (according
to Definition 8-8) is alternative A 3 . This is true because this alternative
corresponds to the minimum criticality degree (e.g., 33.1) among all the
values in Table 8-15. Table 8-16 presents the various sensitivity coefficients
(as defined by Definition 8-9). Note that if there was an infeasible entry
(denoted by the "_" symbol) in Table 8-15, then the corresponding
sensitivity coefficient in Table 8-16 is defined to be equal to zero. A
comparison of Tables 8-10 and 8-15 (or Tables 8-11 and 8-16) indicates that
the current WPM example is much more robust than the WSMI AHP
example. This is true because the sensitivity coefficients in the WPM
example are much smaller. This is a consequence of the specific data used
in these two examples and the different nature of the AHP and WPM
procedures.
Table 8-15: Criticality Degrees d lij (in %) for each aij Measure of Performance.
Criterion 0
AIt.(A)
C1 C2 C3 C4 Cs
A1 85.6(A2) 89.9(A2)
- 81.7 (A 2) 84.5(A 2)
A2 83.4(A 3) 88.0(A 3) 79.3(A J) 82.3(A J)
-
A3 36.8(A4) 41.9(A4) 89.0(A 4) 33.1(A 4) 35.7(A4)
A4 58.0(A3) n.0(A3) - 50.0(AJ) 56.0(A 3)
As - - - - -
Table 8-16: Sensitivity Coefficients sens(aij) for each aij Measure of Performance.
Criterion 0
C] C2 CJ C4 Cs
AIt.(AJ
A1 0.012(A 2) O.oII(A 2) 0 0.012(A 2) 0.012(A2)
A2 0.012(AJ) O.oII(A 3) 0 0.0 13 (A J) 0.012(AJ)
A3 0.027(A 4) 0.024(A 4) 0.011 (A4) 0.030(A 4) 0.028(A 4)
A4 0.0 17(AJ) 0.014(A 3) 0 0.020(A 3) 0.018(A 3)
As 0 0 0 0 0
8.5 CONCLUSIONS
The contributions described in this chapter are both theoretical and

empirical in nature. This chapter was based in the developments that first
appeared in [Triantaphyllou and Sanchez, 1997] and it presented a unified
approach for a sensitivity analysis for three major (and a variant of one)
MCDM methods. These methods are: the weighted sum model (WSM), the
weighted product model (WPM), and the analytic hierarchy process (AHP)
(both in its original and in ideal mode). The proposed sensitivity analysis
examines the impact of changes in the weights of importance of the decision
criteria (Le., the Wj values) and the measures of performance of the
alternatives in terms of a single decision criterion at a time (Le., the aij
values) on the final ranking of the alternatives. The theoretical contributions
of this chapter are best summarized in the five theorems presented in the
previous sections.
The empirical contributions are related to the sensitivity analysis of
changes in the weights of the decision criteria. We did not cover changes in
the aij values with an empirical study because that would result in too many
sensitivity scenarios under consideration for a given problem and thus divert
the attention from the central ideas. Recall that for a problem with m
alternatives and n criteria there are m X n different aij values.
The two most important empirical conclusions of this study are: (i)
The choice of the MCDM method or number of alternatives has little
influence on the sensitivity results; and (ii) most frequently the most sensitive
decision criterion is the one with the highest weight, if weight changes are
measured in relative terms (Le., as percentages), and it is the one with the
lowest weight if changes are measured in absolute terms.
The main observation of the computational experiments is that the
MCDM methods studied here, perform in similar patterns. These patterns
refer to the frequency the criterion with the highest (or lowest) weight is also
the most critical criterion, when changes are measured in relative (or absolute)
terms. Moreover, the same results seem to indicate that the number of
decision criteria is more important than the number of alternatives in a test
problem.
The proposed methodology can be used to carry out a standard
sensitivity analysis when one of the previous MCDM methods is used. The
benefit of doing a sensitivity analysis is too paramount to be ignored in
applications of MCDM methods to real life problems. As Dantzig [1963, p.
32] stated it: "Sensitivity analysis is a fundamental concept in the effective
use and implementation of quantitative decision models, whose purpose is to
assess the stability of an optimal solution under changes in the parameters. "
By knowing which data are more critical, the decision maker can more
effectively focus his/her attention on the most critical parts of a given MCDM
problem.
Another area of application is during the phase of gathering the data
for a MCDM problem, given a limited budget. Often, in real life applications
of MCDM, data are changeable and cannot be precisely determined. In such
cases it makes more sense to determine with higher accuracy the weights of
the criteria (as well as the aij measures of performance) which are more
critical and with less accuracy the less critical weights. A sensitivity analysis,
contacted at an early stage, may reveal which Wj and aij values have a
tendency to be more critical to the final (ranking) decisions. Therefore, these
data can be determined with higher accuracy at a second stage. Next, a new
sensitivity analysis cycle can be initiated again. This process can be repeated,
in this stepwise manner, for a number of times until the entire budget is used
or the decision maker is satisfied with the robustness of the final results.
The three (and a variant of one) MCDM methods examined in this
chapter have been fuzzified by Triantaphyllou and Lin [1996] and are also
presented in Chapter 13. Thus, a natural extension of this research is to
develop a sensitivity analysis approach for cases in which the data are
described by fuzzy numbers. An additional area of possible extension is to
extend these results to AHP problems with multiple hierarchies.
APPENDIX TO CHAPTER 8
8.6 CALCULATION OF THE 01,1,2 QUANTITY WHEN

THE AHP OR THE WSM METHOD IS USED
To help fix ideas suppose that we wish to determine the minimum

change in the weight of importance W 1 of criterion C1 so that the ranking of
the two alternatives A1 and A2 will be reversed (recall that from (8-1)
currently it is: P1 ~ P2 ). As it was introduced in the main part of the
chapter, Ok,i,j (for 1 ::; i < j ::; m and 1 ::; k ::; n) is the minimum change
in the current weight of criterion Ck such that the ranking of alternatives Ai
and Aj will be reversed. In the current setting we have: k = 1, i = 1 and j = 2.
Therefore, the new (Le., the modified) weight, denoted as W*1' of the first
criterion is:
W; = WI - 01.1.2' (AI)
As it was stated in the definitions given in Section 1.1, usually the
criteria weights are normalized to add up to one. Therefore, the new
normalized weights, denoted as W Ii' will be as follows:
WI*
wII
WI* + w2 + .•. + wn
Wz
wI2
WI* + W2 + + Wn (A2)
WIn =
W; + W2 + ... + wn
Given the new weights W Ii (for i = 1, 2, ... , n) it is necessary to
express the conditions for the new ranking. Let p I1 and p I2 denote the new
final preference values for the two alternatives A1 and Ab respectively. Since
it is desired that the new ranking of the previous two alternatives to be
reversed, the following relation should be satisfied:
(A3)
Since the WSM (or AHP) model is used, the previous two preference
values are given by the formulas:
/I /I
pI = L w{a lj , and pi = L w{ a j" 2

i = I i=I
Thus, when the above two definitions of pi] and pl2 are applied on (A3), the
following relation is derived:
/I /I
P~ = L (w;a l) < P~ = L (w;a2) , -

~l ~l
/I /I
-
Lwja Ij Lwj a2j
WI* all j=2 WI* a21 j=2
+ < +
/I /I /I /I
WI* + LWj
•
WI + LWj WI* + LWj WI* + LWj
j=2 j=2 j=2 j=2
/I /I
W; all + L (wjalj) < W; a21 + L (Wj a2). (A4)

j=2 j=2
If relations (A4) and (AI) are combined, then the following is derived:
/I /I
-c\,I,2a ll + L (wjalj) < -c\,1,2 a21 + L (wj a2) , -

j=l j=l
-l)1,1,.2a 11 + PI < -ol.1~a21 + P2' =:$0
-
(P2 - PI)
() 1,1,2 < if (a21 > all)' or :
(a21 - all)'
(A5)
(P2 - PI)
()
1,1,2
> if (a21 < all)·
(a21 - all)'
The above derivations can easily be expanded and are generalized in Theorem
8-1.
8.7 CALCULATION OF THE 01,1,2 QUANTITY WHEN

THE WPM METHOD IS USED
As in the previous sub-section, suppose that we are interested in
determining the quantity 0J,J,2 when the WPM method is used. Recall that
from relation (2-2), in Section 2.2.2, alternative Al is more preferred than
alternative A2 when the following ratio is greater than or equal to one:
R(Al)
A2
= IT
j=l
(a 1j
a2j
)W1• (A6)
Furthermore, according to the ordering in relation (8-1), it is currently

assumed that PI ~ P2 . Let P/ and p/ denote the new preferences of the
previous two alternatives, respectively. Then, when the ranking of these two
alternatives is reversed, the relation on the preferences becomes: p/ < P/'
Also, it can be observed that now the ratio defined in (A6) should be strictly
less than one.
By substituting the new weights (i.e., the ones derived after the
weight of the first criterion has been modified) into (A6), then the following
relations are derived:
RI[~) -
RI[ ~:) where: (A7)
rl = (all),
a
r2 = (aa I2 ), ... , and rn
l2 22
In the previous relations the new normalized weights are as follows:
... ,
w1
where:
=
w··w" w2
= w;
=
+
w··
w wn .
w··
2 + ••• +
Therefore, relation (A7) yields:
-
.
WI W2
- Wft
w" w" w"

rl r2 ... rn < l. (A8)
next it can be noticed that: w; WI - °

1, 1,2' Therefore, relation (A8)
yields (since the right-hand-side is greater than or equal to zero):
1og(r WI r r
a > 1 Z
W2 ••• Wft)
n
1,1,2 logr1
When relation (A7) is used on the previous expression, the following relation
is obtained after some simple algebraic manipulations:
log ( II
n a )W1
--.!l 1
y=1 a Zy
(A8)
109( au]
a Zl
8-2.
8.8 CALCULATION OF THE T 3,4,S QUANTITY WHEN

THE WSM METHOD IS USED
The WSM method is an additive MCDM model in which the final
preference Pi of alternative Ai is obtained according to relation (2-1), in
Section 2.2.1. Let pi; denote the new preferences, that is, the ones after the
change in the aif measure of performance. As before, let us assume that we
are interested in reversing the actual ranking between alternatives A2 and A 3 ,
which currently is from relation (8-1): P2 ~ P3 • Furthermore, in order to
get that rank reversal one must find the threshold value T3,4,2 of the a34
measure of performance required to get the following relation:
p I 2 < p13' (A9)
When relation (2-1) is applied to (A9), then the following is derived:
aZ1 w 1 + •.• + a2n w n < a 31 w1 + ••• +
P2 < -T3,4,2 w4 + P3 , (because d 34 = a34 - T3,4,~.

I
a34 w4 + ••• + a3nw n • -
Therefore, in order to reverse the ranking of the two alternatives A2
and A3 by modifying the a34 measure of performance, the threshold value T3,4,2
'1: 3,4,2 <

w4
Furthermore, the following condition should also be satisfied for the new d 34
value to have a feasible meaning:
o
o
~
I
a34 ,
~ a34 - "t 3,4,2'

-
"t 3,4,2 ~ a34 •
From the previous relations, it can be seen that the 7 3,4,2 value must be less
than or equal to a34 in order to have a feasible meaning.
Next, let us assume that the measure of performance that we wish to
alter is again a34 , but now we want to reverse the ranking between alternatives
A3 and A5 (observe that now: P3 ~ P5 from (8-1». Then, by following an
approach similar to the previous one, one can derive that in order to reverse
the current ranking of the two alternatives A3 and A5 the threshold value 73,4,5
Moreover, the following condition should be satisfied for the new a /34 value
to have a feasible meaning:
a34 •
t 3,4,5 ~
8-3.
8.9 CALCULATION OF THE T'3,4,5 QUANTITY WHEN

THE AHP METHOD IS USED
Suppose that the interest is in reversing the ranking between

alternatives Ai and Ak, As it was stated in Definition 8-6, the threshold value
7 i,},k (for i ~k, 1 :::;; i, k :::;; m and 1 :::;; j :::;; n) is the minimum change in the
current value of the a y measure of performance, such that the ranking between
the two alternatives Ai and Ak will be reversed, Let a'i} denote the modified
a y measure of performance, That is:
a'y = ay - 7 i ,i,k' (AlO)
For an easy demonstration of the main ideas, suppose that we are
interested in reversing the current ranking between alternatives A2 and A 3,
(where P2 ~ P 3 from (8-1» by altering the a34 value (only), Therefore, it is
necessary to determine the threshold value 7 3,4,2 required to get the following
relation (where P /2 and P /3 denote the new final preference values of
alternatives A2 and A3 , respectively):
pl2 < p13' (All)

Recall that the AHP requires normalization of the aij values. Also it can be
noticed that a *34 (= a34 - 73,4,2) denotes the altered value of aij' Therefore, the
new al ij values are as follows:
a14
a14 =
a14 + ~ + *
~4 + ... + am4
a24
~4 =
a14 + a24 + ~4* + + am4
* (A12)
~4
a34 =
a14 + a24 + ~4* + ... + am4
am4 =
a14 + ~ + ~4
* + + am4
Furthermore, the denominator of relations (A12) can be simplified by using
relation (AlO) as follows:
a14 + ~ + a34 - 't 3,4,2 + a44 +

- 't 3,4,2 + a 14 + ... + am4 =
1 - '1: 3,4,2'
(A13)
Therefore, relations (A12) can be expressed as follows:
----,
1 - 't 3,4,2
I a34* a34 - 't 3,4,2

(A14)
~4 = (from (AIO»
1 - 't 3•4,2 1 - 't 3•4,2 ,
I am4
am4 = ----=-'"'---
1 - 't 3,4,2
Consequently, by applying relations (A14) and (8-1) on (All) we have the

following:
-
(A15)
When we substitute relations (A14) on (A15) we get:
~ w4 (a 34 - '1: 3•4 ,2)W4

- a24w4 + P2 < - a34 w4 + P 3 •
(1 - '1: 3•4,2) (1 - '1:3.4~
which can be further reduced to (by observing that the denominator on the
right-hand-side is always a positive number):
'1: 3•4,2 < (A16)

[P3 - P2 + w4 (a24 - a34 + 1)]
Furthermore, the following condition should also be satisfied for the new a l34
value to have a feasible meaning:
o :0:; a34 :0:; 1. or (from (AiO»:

't 3.4,2 :0:; a34 •
From the previous relations, it can be seen that the T3,4,2 quantity must be
within the range [a34 , a34 - 1] in order to have a feasible value.
Now assume that the measure of performance to alter is again a34 , but
now we want to reverse the ranking between alternatives A3 and Aj (note that
from (8-1) now we have: P3 ~ Pj ). Then, by following a similar approach
as previously one can get that in order to reverse the current ranking of
alternatives A3 and A j , by modifying the a34 measure of performance, the
minimum quantity T 3,4,5 should satisfy the following relation:
P3 - Ps
'1: 3•4,5 >
The following condition should also be satisfied for the new d 34 value to have
a feasible meaning:
't" 3,4,5 ~ a 34 ·
8-4.
8.10 CALCULATION OF THE T '3,4,5 QUANTITY WHEN

THE WPM METHOD IS USED
In this sub-section we follow a similar developmental approach as in

the previous sub-sections. In the WPM model the ratio R(A/A.) should be
greater than or equal to one (according to relation (2-2) in Section 2.2.1) in
order for alternative Ai to be more preferred than alternative Ak • Let us
denote as RI (A/A.) the new ratio after the Oi,j,k change has occurred on the aij
measure of performance. The new ratio RI (A/A.) should be strictly less than
one. We want to find the threshold value 73,4,2 in order to get a rank reversal
between alternatives A2 and A 3 . In this case the new ratio RI (A/A.J will be
as follows:
R' (~) = (~:rx x(~rx . . x(::r < I.

(A17)
If we use the relation: al 34 = a34 - 73,4,2 in (A17), we get:
or:
(
a34 )W4 xR (A2)
- < 1.
a 34 - 't" 3,4,2 A3
After some simple algebraic manipulations we get:
<'.4, > a"x 1 - ·~l· (A18)
Furthermore, the following condition should also be satisfied for the new
a 134 value to be feasible:
o
o
-
=>
"3,4,2 ~ a34 ,
Now, assume that the measure of performance to alter is again a34 ,
but now we want to reverse the ranking between the two alternatives A3 and
A5 (now: P3 2!: P5 from (8-1», Then, by following an approach similar to the
previous one, one can derive that in order to reverse the current ranking of
the two alternatives A3 and A5 , the threshold value 7345 should satisfy the
following relation:
<3~' < a,.x 1 _ W~R[~:) .

Similarly with above, the following condition should also be satisfied for the
new a /34 to have a feasible meaning:
't 3,4,5 ~ a34 '

8-5,
Chapter 9
EVALUATION OF METHODS FOR PROCESSING

A DECISION MATRIX AND SOME CASES
OF RANKING ABNORMALITIES
The MCDM methods presented in the second chapter are among the
most widely used ones. As it will be seen in the following sections, however,
the part of the methods that processes a decision matrix (i.e., step 3 in Section
2.2) may give different answers to the same problem. Because only the
WPM, the AHP, the revised AHP, and the TOPSIS method are applicable
both in single- and multi- dimensional decision making, these are the methods
that are examined in this chapter (i.e., the WSM will not be tested). The
ELECTRE method (as described in Chapter 2) was not tested because it is
already known that this method presents a different ranking philosophy and
it does not assume that a unique ranking always exists in practice. The main
part of this chapter is based on the research results presented in
[Triantaphyllou and Mann, 1989] plus some additional investigations.
Since the truly best alternative is the same regardless of the method
chosen, an estimation of the accuracy of each method is highly desirable. The
most difficult problem that arises here is how one can evaluate a multi-
dimensional decision making method when the true best alternative is not
known. Two evaluative criteria are introduced for the above purpose in the
following section.
9.2 TWO EVALUATIVE CRITERIA
In [Triantaphyllou and Mann, 1989] two evaluative criteria were

developed and used for evaluating some MCDM methods. These two
evaluative criteria are as follows:
First Evaluative Criterion:

An MCDM method that is accurate in multi-dimensional problems should also
be accurate in single-dimensional problems.
There is no reason for an accurate multi-dimensional method to fail

in giving accurate results in single-dimensional problems, since single-

dimensional problems are special cases of multi-dimensional ones. Because

the first method, the WSM, gives the most acceptable results for the majority
of single-dimensional problems, the result of the WSM is used as the
standard for evaluating the other four methods in this context.
Second Evaluative Criterion:

An effective MCDM method should not change the indication of the best
alternative when an alternative (not the best) is replaced by another worse
alternative (given that the relative importance of each decision criterion
remains unchanged).
A number of similar evaluative criteria could be introduced as well.

For example, the fact that an accurate model should not indicate a change in
the best alternative after the introduction of identical (or worse) non-optimal
alternatives, could be the basis for other evaluative criteria. However, the
second criterion is stricter than these criteria and it affects more than one
method.
It should be stated here that more evaluative criteria can be generated
depending on the nature of the MCDM problem. For instance, in cases such
as when distributing a limited budget among a number of competing projects
one is interested on what happens to anyone of the projects (e.g.,
alternatives) and not only to the top alternative. Thus, a failure could be
assumed to have occurred when the WSM and the other methods disagree on
the ranking of any alternative (as far as the first evaluative criterion is
concerned). A similar consideration is also possible in terms of the second
evaluative criterion. On the other hand, if the main interest is to find which
is the best alternative (for example, if we are interested in filing out a single
job vacancy), then we want to concentrate our attention only on which is the
top ranked alternative. In cases like this one, the analyses described in the
following sections are more suitable.
Finally, it is important to note here that the testing (using the above
two evaluative criteria) is sufficient to reveal that a method is ineffective in
indicating the best alternative or correctly ranking a set of alternatives. But,
if a method is in fact ineffective, then it is not also necessarily true that the
above two evaluative criteria will detect this fact. The following sections
provide some numerical examples and present the results obtained from some
computational tests of the methods using the previous two evaluative criteria.
Chapter 9: Evaluation of Methods for Processing a Decision Matrix 179
9.3 TESTING THE METHODS BY USING THE FIRST
EVALUATIVE CRITERION
This is best illustrated in the following example.
Example 9-1: Testing the method used by the Analytic Hierarchy Process by
using the first evaluative criterion
Suppose that the matrix below depicts the actual values, measured in
the same units.
Criteria
C1 · C2 C3
Alts. (8/13 2/13 3/13)
Al 1 9 9
A2 5 2 2
A3 1 5 9
This previous decision matrix represents a problem defined on the

three alternatives AI' A2 , and A3 , in terms of three criteria with the following
weights: WI = 8/13, W 2 = 2/13, and W3 = 3/13. When the WSM is applied,
it can be shown that the alternative Al is the best one (since Al WSM.score =
A*wSM-score = 53/13).
From the matrix with the actual values we can see that the three 3 x 3
matrices with the pairwise comparisons (please also see Chapters 3 and 4 for
the use of pairwise comparisons) that correspond to this problem are as
follows (perfect consistency in the pairwise comparisons is assumed in order
to block out any effects due to inconsistent comparisons):
Criterion C1 Criterion C2
1
5/1
1/5
1
1/1
5/1
1 9/2
9/5
2/9 1 2/5
1
1/1 1/5 1 5/9 5/2 1
Criterion C3
1 9/2 9/9
2/9 1 2/9
9/9 9/2 1
Therefore, the mXn (Le., 3x3) decision matrix with the relative
importances of the alternatives in terms of each criterion, along with the
criteria weights, that is used in the final step of the AHP is:
Criteria
C1 Cz C3
Alts. (8/13 2113 3/13)
Al 117 9/16 9/20

A2 517 2/16 2/20
A3 117 5/16 9/20
Applying the last step of the AHP it turns out that alternative A2 is the best
one (A 2• AHP-score = A*AHP-score = 0.48). Obviously, this is in contradiction with
the conclusion derived using the WSM. •
A computer program was written to generate random data and to

solve problems with all possible combinations of 3,5,7, ... ,21 alternatives
and 3,5,7, ... , 21 criteria. Thus, 100 (= lOx 10) different cases were
examined. For each case, 10,000 different decision matrices were randomly
generated. The problem was solved each time assuming as data random
integer numbers between 1 and 9 and then by using the WSM and AHP
approaches as in Example 9-1. That is, the actual values were assumed to be
random integers from 1 to 9. This range of values was used because in AHP
each pairwise comparison is taken from the set: {xly, where x,y = 1,2,3, ... ,
9} (due to the scale that is recommended by the AHP, as described in Chapter
3). If there were a contradiction between the two approaches, the program
noted that the AHP yielded contradictory results.
In a similar manner, the revised AHP, and the WPM were examined
(the case of the TOPSIS method is studied in Section 9.6). Because at the
beginning of this study the purpose was to get a general idea of the
contradiction rates, the sample sizes were not determined using the standard
deviations of the observations. However, the findings indicate that the
sample sizes were satisfactorily large. This is the case with all the numerical
results presented in this book. The results are presented in the following
tables and figures.
Table 9-1: Contradiction Rate (%) Between the WSM and the AHP.
Number of Number of Criteria

Alternatives 3 5 7 9 11
3 8.2 10.5 12.3 11.7 12.2

5 8.4 13.2 12.6 13.2 13.2
7 8.1 10.1 12.3 12.6 13.3
9 8.5 9.5 11.0 12.5 12.5
11 7.5 10.0 10.9 12.1 12.1
13 7.1 8.5 10.6 11.4 11.7
15 6.6 8.7 9.6 10.6 11.1
17 6.8 9.3 10.0 10.6 10.7
19 6.7 8.1 9.1 9.3 10.7
21 6.0 7.9 9.2 9.6 10.3

3 11.8 11.8 12.1 12.2 12.5

5 13.5 13.6 13.3 14.0 13.7
7 13.1 13.5 13.1 14.6 13.7
9 12.0 13.0 13.1 13.5 13.2
11 12.3 12.2 12.5 12.8 13.0
13 11.8 12.1 12.0 13.1 12.4
15 11.5 12.0 12.1 11.2 12.4
17 11.1 11.5 11.5 12.0 11.0
19 10.7 10.9 10.9 11.4 10.7
21 10.7 10.8 10.5 11.0 10.9
Table 9-2: Contradiction Rate (%) Between the WSM and the Revised AHP.

3 7.4 8.2 8.8 9.6 10.5

5 5.6 7.1 9.3 8.4 8.7
7 4.4 6.3 6.2 6.2 7.7
9 4.2 4.2 4.9 6.9 7.5
11 3.0 3.8 5.6 4.8 5.1
13 2.7 3.3 3.8 5.0 4.5
15 2.2 3.1 3.9 4.4 3.6
17 1.7 2.2 2.6 3.7 3.6
19 1.8 2.7 2.6 2.8 3.1
21 1.4 1.9 2.2 2.2 2.7

3 10.3 10.9 10.5 9.5 10.6

5 9.1 8.7 8.8 8.8 9.1
7 6.6 7.1 7.6 8.3 7.3
9 6.5 6.2 6.8 6.4 7.6
11 5.4 6.0 6.3 6.0 6.2
13 4.0 6.3 5.3 5.0 5.3
15 3.9 3.4 4.5 4.3 5.3
17 3.8 3.4 3.7 4.2 4.0
19 3.2 3.4 3.5 3.1 3.4
21 2.9 2.5 2.9 3.1 3.0
Table 9-3: Contradiction Rate (%) Between the WSM and the WPM.

3 14.7 12.7 12.1 13.9 15.3

5 12.0 14.8 15.7 17.0 17.5
7 12.2 17.1 18.7 18.0 22.0
9 11.2 16.5 18.9 19.7 20.9
11 11.8 17.5 20.9 21.3 21.8
13 11.7 17.2 21.1 20.8 22.8
15 11.3 19.4 19.4 21.8 24.9
17 11.2 17.1 21.2 23.3 23.9
19 11.3 18.9 21.3 22.7 23.6
21 11.3 18.2 21.1 24.0 25.4

3 14.3 13.8 12.6 13.6 14.5

5 19.6 18.9 17.1 18.7 18.1
7 19.5 19.9 20.2 21.8 20.3
9 21.5 23.0 23.4 21.6 23.3
11 23.6 22.9 22.0 22.2 23.0
13 23.4 24.9 25.0 25.7 25.1
15 24.8 23.3 25.0 26.9 24.7
17 26.4 22.5 25.2 26.3 26.2
19 25.7 25.2 28.6 26.9 27.7
21 26.6 25.9 27.5 26.4 25.7
184 MCDM Methods: A Comparative Study. by E. TriantaphyLlou
Figure 9-1: Contradiction Rate (%) Between the WSM and the AHP.
20
lS~
16
-
~
~
14 c:
C
12
ti
:a
10 ~
C
U
Iv,
IIlJJber of C .
ljterJa
Figure 9-2: Contradiction Rate (%) Between the WSM and the RevisedAHP.
Figure 9-3: Contradiction Rate (%) Between the WSM and the WPM .
9.4 TESTING THE METHODS BY USING THE SECOND

EVALUATIVE CRITERION
As with the first evaluative criterion, this testing is best illustrated in the
following example.
Example 9-2: Testing the method used by the Analytic Hierarchy Process by
using the second evaluative criterion
Suppose that the following is a matrix of eigenvectors produced by the
AHP process. That is, the matrix contains the relative values of the
importance of the alternatives instead of the actual values. Also, assume that
the criteria have weights WI = 217, W2 = 217, and W3 = 317.
Criteria
CI C2 C3
Alts. (217 217 317)
Al 9/19 2/12 217

A2 5/19 1112 417 = Matrix Ml
A3 5/19 9/12 117
It can be shown (by multiplying the matrix with the relative

importances by the vector with the weights of the three criteria followed by
normalization) that the priority vector of the alternatives (according to AHP)
is: (0.305, 0.344, 0.351).
Apparently, the best alternative is A 3 • If in the above problem the
alternative Al (which is not the best one and was defined by the relative values
(9/192/12217), is replaced by alternative All which is worse than the original
alternative Al then, the above matrix is modified as follows:
Criteria
CI Cz C3
Alts. (217 217 317)
All 8/18 1111 116

A2 5/18 1111 4/6 = Matrix M2
A3 5/18 9/11 116
Matrix Ml can be considered as the matrix with the relative values

obtained from the following three 3 x 3 matrices with pairwise comparisons
(perfect consistency in the pairwise comparisons is again assumed as before):
Criterion C1 Criterion C2
1 9/5 9/5 1 2/1 2/9

5/9 1 5/5 1/2 1 1/9
5/9 5/5 1 9/2 9/1 1
Criterion C3
1 2/4 2/1
4/2 1 4/1
1/2 1/4 1
Matrix M2 has been derived from matrix Ml by substituting the

alternativeA J with the lesser alternativeA'J = (8/18 1111 116) < (9/192112
217). That is, instead of 9 it is now 8, instead of 2 it is now 1, and instead
of 19 it is now 18.
Similarly, the priority vector for matrix M2 and the same criteria
weights is (0.224, 0.391, 0.385). It is clear that now the best alternative is
A 2 • The last statement is, obviously, in contradiction with the original result
namely, that the best alternative is A 3 . Thus, by introducing a new worse
alternative (different from the best one) it is possible to have a change in the
indication of the best alternative. •
A computer program was written to solve random problems (5,000

trials per case) in a manner similar to Example 9-2. Again, 100 different
cases were examined. Each random problem was solved using the AHP and
thereafter replacing one random alternative (not the best) with another worse
one in a way similar to Example 9-2 (i.e., by subtracting 1 from both the
numerator and denominator).
The cases of the revised AHP and WPM were examined in a similar
way. However, it can be easily shown that the WPM does not yield
contradictory results when it is tested in terms of the second evaluative
criterion because of its mathematical structure (see also Section 11.6). The
following figures and tables present the results of the above testing. Please
note here that the vertical scale has been changed in the following figures in
order to better illustrate the impact. The case of the TOPSIS method is
described in Section 9.6.
Table 9-4: Rate of Change (%) of the Indication of the Optimum Alternative When a
Non-Optimum Alternative is Replaced by a Worse One. The AHP Case.

3 0.16 0.26 0.26 0.32 0.30

5 0.34 0.36 0.40 0.54 0.46
7 0.30 0.42 0.42 0.40 0.36
9 0.16 0.36 0.22 0.20 0.22
11 0.14 0.18 0.12 0.24 0.06
13 0.06 0.22 0.22 0.28 0.18
15 0.04 0.10 0.18 0.12 0.22
17 0.08 0.10 0.14 0.24 0.06
19 0.08 0.04 0.10 0.10 0.10
21 0.06 0.10 0.10 0.04 0.08

3 0.26 0.54 0.28 0.40 0.26

5 0.32 0.32 0.40 0.38 0.42
7 0.40 0.34 0.42 0.16 0.26
9 0.26 0.26 0.26 0.18 0.26
11 0.22 0.48 0.22 0.22 0.20
13 0.24 0.26 0.22 0.08 0.28
15 0.14 0.12 0.10 0.20 0.14
17 0.06 0.10 0.18 0.04 0.08
19 0.12 0.12 0.04 0.08 0.08
21 0.10 0.10 0.10 0.06 0.02
Non-Optimum Alternative is Replaced by a Worse One. The Revised AHP
Case.

3 0.50 0.50 0.60 0.30 0.30

5 0.50 0.50 0.80 0.80 1.30
7 0.60 0.60 1.70 0.90 0.80
9 0.70 0.30 1.30 0.80 0.80
11 0.30 1.00 0.30 1.00 0.50
13 0.20 0.60 1.40 0.80 1.20
15 0.30 0.90 0.80 0.40 0.90
17 0.00 0.40 0.80 0.40 0.90
19 0.20 0.10 0.20 0.50 0.80
21 0.10 0.20 0.30 0.40 0.70

3 0.10 0.00 0.10 0.10 0.10

5 0.80 0.10 0.50 0.50 0.00
7 0.50 0.50 0.50 0.40 0.20
9 1.00 1.10 0.70 0.60 0.20
11 1.10 0.60 0.60 0.80 0.70
13 1.20 0.80 0.80 0.40 0.40
15 1.20 0.60 0.60 0.60 0.90
17 0.80 0.80 1.20 0.50 0.30
19 1.20 0.80 0.60 1.00 0.70
21 0.60 0.70 0.40 1.20 0.60
Table 9-6: Summary of the Computational Results.

Criterion
2
Cases with
Nwnber of alts. 3 3 3 21 3 3 3 21
Nwnber of criteria 3 5 7 21 3 3 7 21
Sub-criterion 2 3 100 101 102 103 200
Method:
WPM 14.7 12.7 12.1 ... 25.7 0 0 0 ... 0
AHP 8.2 10.5 12.3 ... 10.9 0.16 0.26 0.26 0.02
RevisedAHP 7.4 8.2 8.8 ... 3.0 0.50 0.50 0.50 ... 0.60
o. Z4
~
'-"
0.11 ~
O.lS g
O.ll
~
:a
0.12 f!
0.09
-=o
U
06
Figure 9-4: Rate of Change (%) of the Indication of the Optimum Alternative
When a Non-Optimum Alternative is Replaced by a Worse One.
The AHP Case.
The Revised AHP Case.

In Figure 9-1 the contradiction rate between the WSM and the AHP
methods decreases as the number of alternatives increases. This occurs
m
because the quantities (1/ E aij) which are used to multiply each entry
i = 1
aij approach the same value (law of large numbers) as the number m of
alternatives increases (the a/s are random integer numbers equal to 1,2,
... ,9). In the WSM each entry can be multiplied by the same quantity without
altering the results. Hence, the AHP tends to behave like the WSM as the
number of alternatives increases.
Also in the same figure, the contradiction rate increases with the
number of criteria. This is true because as n, the number of criteria,
m
increases, the probability that the multipliers (1/ E aij' for j = 1,2,3,
i = 1
... , n) will deviate significantly increases. Thus, the AHP can differ
substantially from the WSM.
In the revised AHP method, the multipliers are the quantities
(lImax{aij' i = 1,2, ... , m}). As m (the number of alternatives) increases,
all of the multipliers tend to the same number (i.e., 9). Since this
convergence is very fast, the AHP approaches in this sense the WSM very
quickly (see also Figure 9-2).
The number of criteria plays exactly the same role in the revised
AHP as in the original AHP. The only difference is that the contradiction
rates are smaller, because the multipliers are not as different as in the original
AHP.
Figure 9-3 shows the contradiction rate for the WPM case. In order
to get an understanding of the shape of the graph in this figure we consider
the simple case where all the criteria are of the same importance. For the
purposes of Figure 9-3, we can view the WSM not in matrix form but
instead, with only two alternatives at a time (as we do with the WPM). Now,
we can see that if one entry aij is very different from the rest, it causes more
impact in the WPM (since it multiplies or divides a sequence of terms) than
in the WSM (where it has only an additive effect). The more criteria are
involved, the more likely it is that this phenomenon will take place. This is
why the rates in Figure 9-2 increase with the number of criteria. This
reasoning is also true as the number of alternatives increases (see also Table
9-3). Hence, the contradiction rate increases with the number of alternatives.
The rate of change (%) of the indication of the optimum alternative
when a non-optimum alternative is replaced by a worse one (for the AHP
case), does not depend on the number of criteria as seen in Figure 9-4. This
occurs because, for every criterion, the same constant value (Le., -1) was
used to alter the nonoptimal alternative. It is expected (although not tested)
that if the aij were to be changed by a random number and not a constant
value (Le., -1), the contradiction rate would be dependent on the number of
criteria. Because the constant value (Le., -1) was used the bias, if any, would
be in favor of the original AHP over the revised AHP.
However, the role of the alternatives is a critical one. When m, the
number of alternatives, increases the role of the multipliers is more critical
than the impact of replacing a nonoptimal alternative by a worse one. Since
the number of alternatives overrides the role of the previous replacements the
rates in Figure 9-4 decrease as the number of alternatives increases.
Finally, Figure 9-5, for the revised AHP, illustrates that the rate of
change of the indication of the optimal alternative when a nonoptimal
alternative is replaced by a worse one decreases as the number of alternatives
increases. This occurs because as the number of alternatives increases the
multipliers (lImax{aij' for i = 1,2,3, ... , m}) becomes closer to 9 (Le., the
maximum) and thus the altered alternative still remains nonoptimal. Similarly
with the previous paragraph, as the number of criteria increases there is
greater chance that the multipliers will differ substantially and alternative
changes will have greater impact.
Because in the revised AHP the multipliers (Le., the quantities
(lImax{aij, for i = 1,2,3, ... , m}) are more stable than the corresponding
ones of the original AHP the revised AHP performs better in terms of the
first evaluative criterion. This is not the case when the second evaluative
criterion is used. However, now the difference in the rates does not seem to
be so dramatic.
9.6 EVALUATING THE TOPSIS METHOD

The TOPSIS method (as was described in Section 2.2.6) was also
evaluated in terms of the previous two evaluative criteria. This was
accomplished during a different investigation. Thus, these findings are
presented separately from the ones which were originally presented in
[Triantaphyllou and Mann 1989]. The corresponding results are presented in
Tables 9-7 and 9-8 and are also depicted in Figures 9-7 and 9-8, respectively.
As it can be seen from these results TOPSIS also exhibits certain failure rates.
However, these rates are higher than those of the methods examined in earlier
in this chapter. For this reason, this method was not compared with the
previous three methods (Le., the WSM, AHP, and revised AHP).
Table 9-7: Contradiction Rate Between the WSM and the TOPSIS Method.

3 17.3 24.2 29.9 31.4 31.8

5 25.9 32.4 37.6 38.7 38.2
7 27.4 33.7 41.1 39.0 41.3
9 27.1 35.2 41.5 44.2 42.2
11 27.2 38.3 40.2 45.4 43.8
13 28.8 39.0 42.4 46.1 48.3
15 31.7 41.3 44.3 45.5 49.2
17 31.8 43.4 48.0 46.7 46.3
19 31.6 40.3 47.1 51.2 48.0
21 30.8 40.6 47.9 46.6 53.4

3 32.0 31.9 29.1 30.9 33.7

5 37.6 37.6 38.9 40.7 40.1
7 42.8 42.4 40.5 42.2 45.7
9 43.2 45.1 43.0 43.5 46.7
11 45.1 46.4 44.6 44.4 48.6
13 49.7 48.6 48.3 46.1 48.2
15 48.8 51.2 50.1 51.5 53.6
17 46.8 52.0 51.3 53.5 50.1
19 51.8 50.0 51.6 51.9 52.0
21 50.5 52.0 49.0 51.2 52.1
Non-Optimum Alternative is Replaced by a Worse One. The TOPSIS
Case.

3 2.4 3.3 3.3 3.8 5.4

5 2.3 3.4 3.9 4.4 4.1
7 2.1 2.9 3.4 3.3 3.7
9 1.2 2.5 3.1 3.4 2.7
11 0.8 1.6 2.8 3.1 2.6
13 1.4 1.7 2.3 3.0 2.5
15 1.4 2.2 3.1 2.8 3.0
17 1.8 2.0 2.6 2.4 2.7
19 1.0 2.0 2.6 2.7 1.7
21 1.1 1.1 1.7 1.3 2.0

3 5.0 4.7 5.3 1.2 4.1

5 3.8 4.8 5.2 4.9 4.3
7 3.1 4.0 5.4 5.2 1.9
9 3.1 4.0 4.8 4.1 4.3
11 2.6 3.7 3.3 4.1 3.0
13 2.3 3.2 3.7 2.9 4.8
15 2.9 3.3 2.6 3.9 2.5
17 2.0 3.6 4.0 3.0 2.1
19 3.2 4.6 2.5 2.0 4.1
21 2.5 2.1 2.7 3.3 2.8
60
SO
()
g 40
~ cr ~
:30
~ ~
(l ZO
/0 Q".
~
~
'l,.."
"q,
,,"I. ~"i:>
,,~ :..~
,,~ -<;;:-r,,:
,," '<.,.<z--<'
q, ~
~ '" ::<:;:,<z--<' c3-.
~v~
Figure 9-6: Contradiction Rate (%) Between the WSM and the TOPSIS
Method.
The TOPSIS Case.
9.7 CONCLUSIONS
The previous analyses make it clear that none of the MCDM methods
reviewed is perfectly effective in terms of both evaluative criteria. The
results that were obtained by testing the methods using the two evaluative
criteria described in Section 9.2 indicate that the MCDM methods yield
different rates of contradiction.
The findings of the analyses are summarized in Table 9-8. This table
presents the structure of the typical decision-making problem with which the
typical MCDM method is concerned with and it also illustrates a decision
making paradox. In this table, MCDM methods are treated as alternatives
and the two evaluative criteria as the decision criteria of the problem. The
two evaluative criteria were considered in terms of cases with 3 alternatives
and 3 criteria, 3 alternatives and 5 criteria, 3 alternatives and 7 criteria, and
so forth. Thus, 100 (= 10 X 10) subcriteria were generated per criterion.
The numbers WI and w2 = 100 - WI> where WI = 1,2,3, ... , 100%, were
assumed to be the relative weights of the two evaluative criteria. The relative
weights of the first 100 sub criteria were assumed to be equal to WI / 100, and
for the last 100 subcriteria equal to W 2 / 100.
Since it is not clear which method is the best, the problem of selecting
the best decision making method was solved in [Triantaphyllou and Mann,
1989] by using successively the WSM, the AHP, and the revised AHP
method. The occurrence of zeros in Table 9-6 made the applicability of the
WPM infeasible (because of the divisions by zeros). Thus, the WPM was not
considered for solving the previous decision problem. Figure 9-8 presents the
results for different values for the weights WI and w2 • Since in Table 9-6 all
values are presented in the same unit (recall that these numbers represent
percentages), the WSM is the method that would yield the most reliable
answer to the problem of choosing the most effective decision making
method.
The first column in the chart of Figure 9-8 represents the evaluation
derived by using the WSM. That chart suggests that the revised AHP appears
to be the best method over a wide range of different combinations of weights
WI and w2 • Only for WI less than 10 per cent the revised AHP is not the best
method. For WI less than 10 and above 3 per cent the AHP appears to be
superior than the others, while for WI less than 3 per cent the WPM becomes
the best method.
From Figure 9-8 one can generally conclude that as WI increases, then
WPM is replaced by the AHP and later by the revised AHP as the best
method. This probably happens because both methods behave in the same
way for problems with 100 criteria and 3 alternatives (as is the case in Table
9-6).
100
80
60
KEY:
40
20
WPM AHP Revised

AHP
o
WSM AHP Revised
AHP
Method used to evaluate

the results
Figure 9-8: Indication of the Best MCDM Method According to Different

MCDM Methods.
However, both methods recommend the WPM to be the best and not
themselves! They also recommend the revised AHP to be the second best
approach, while the original AHP is the best one only for a very small portion
of combinations of weights WI and W 2 .
The previous study, with the findings summarized in Figure 9-8,
demonstrates that it is impossible to determine precisely the best MCDM
method, for to do so one needs to use the best MCDM method! This
problem of finding the best MCDM method always reaches a
decision-making paradox which makes any attempt in solving this problem
to be of limited success. However, the results in this chapter recommend that
for most of the cases of different weights of the two evaluative criteria the
revised AHP appears to be the best decision making method of the four
examined in terms of these two evaluative criteria, while the original AHP,
appears to be the most inaccurate one when the previous two evaluative
criteria are considered.
Chapter 10
A COMPUTATIONAL EVALUATION OF THE

ORIGINAL AND THE REVISED AHP
As it was mentioned in Chapter 1, the typical problem examined by

the AHP consists of a set of alternatives and a set of decision criteria. Since
this problem is very common in many engineering, science, and economic
applications, the AHP has been a very popular decision tool. Another reason
which contributed to the wide use of the AHP in such applications, is the
development of the Expert Choice software (http://www.expertchoice.
com/). Furthermore, many other computer packages have been developed
and are based on the principles of the AHP. A prime example is the
Criterium Decision Plus package by Info Harvest, Inc. (http://www.
infoharvest.com/). For an excellent description of this package the
interested reader may want to read the review in [Haerer, 2000].
Some of the engineering and science applications of the AHP and its
variants include its use in integrated manufacturing [Putrus, 1990], and
[Falkner and Benhajla, 1990], in the evaluation of technology investment
decisions (i.e., [Swann and O'Keefe, 1990], [Roper-Lowe and Sharp, 1990],
[Finnie, 1988], and [Boucher and McStravic, 1991]), in flexible
manufacturing systems [Wabalickis, 1988], layout design [Cambron and
Evans, 1991], and also in other decision problems (see, for instance, [Wang
and Raz, 1991], and [Arbel and Seidmann, 1984]). The revised AHP is a
variant of the original AHP and was proposed by Belton and Gear [1983]
(see also Section 2.2.4) and was also later adopted by Saaty as the "ideal
mode" AHP. Since these two methods are widely known today, these are the
two MCDM methods to be investigated and compared in this chapter.
The numerical investigations discussed in this chapter follow an
approach similar to the one presented in [Triantaphyllou and Mann, 1990].
Since an AHP decision making problem involves a sequence of reciprocal
(judgment) matrices, the testing strategy developed in [Triantaphyllou and
Mann, 1990] will be repeated a number of times. One time for each one of
the reciprocal matrices involved in a given decision problem. Recall that if
a problem is defined on m alternatives and n decision criteria, then there is
one such matrix of order n X nand n matrices of order m X m.
In [Triantaphyllou and Mann, 1990] it was found that when the
eigenvalue method is used to derive relative weights, then it is possible for
some entities (Le., decision criteria or alternatives) to have the same rank
while in reality they are distinct. However, the empirical analyses of this
chapter demonstrate that it is possible for alternatives which in reality are less
important than others to appear to be more important after the AHP or the
revised AHP are used. Furthermore, this problem of rank reversals is more
severe in the present investigation. Also, the present findings reveal that the
rate of the rank reversals increases dramatically as the number of
alternatives increases.
This chapter uses extensively the concepts of the RCP (Real and
Continuous Pairwise) and CDP (Closest Discrete Pairwise) matrices (which
are defined and analyzed in Section 3.3). The problem which is investigated
in this chapter is best described by the numerical example presented in the
next section. In that numerical example it is assumed that the actual relative
values in a decision making problem, which involves three alternatives and
four criteria, are known. The AHP and the revised AHP are applied to these
actual data and the relative weight and ranking of each alternative is derived
according to each one of these two MCDM methods.
Next, we will use the concepts of the RCP and CDP matrices to
derive the discrete data required in the test problems. These discrete data are
the data which a decision maker could have if one were able to make the
closest approximations to the real (and thus unknown) pairwise comparisons.
Again, this is a very favorable assumption since there is no certainty that in
reality one can achieve these closest approximations for all the pairwise
comparisons. The following numerical example reveals that it is possible for
the AHP and the revised AHP method to fail to preserve the ranking of the
alternatives when the continuity assumption is utilized and even if the decision
maker is as accurate as possible.
Suppose that an MCDM problem with three alternatives and four

criteria has in reality the following matrix as the decision matrix with the
actual relative weights:
Criteria
Ct C2 C3 C4
Alts. (0.1325 0.0890 0.5251 0.2533)
At 0.5008 0.4804 0.0850 0.2113

A2 0.3785 0.2133 0.4434 0.3799
A3 0.1207 0.3063 0.4716 0.4089
Chapter 10: A Computational Evaluation of the Original and Revised AHP 203
That is, the actual relative weights of the four criteria are: (0.1325
0.0890 0.5251 0.2533). The relative importances of the three alternatives
in terms of the first criterion are: (0.5008 0.3785 0.1207), and so on.
The above values can also be viewed as the principal right
eigenvectors of the following 1 + 4 = 5 perfectly consistent Rep matrices.
For the case of the four criteria we have:
1 1.4890 0.2524 0.5231

0.6716 1 0.1695 0.3513
3.9622 5.8995 1 2.0727
1.9116 2.8462 0.4825 1
For instance, in the previous matrix the (1, 2) entry is equal to 1.4890
because the corresponding pairwise comparison is: 0.1325 / 0.0890 (=
1.4890). Similar interpretations hold for the remaining entries.
For the three alternatives we have in terms of each one of the four
decision criteria:
1 1.3230 4.1502
0.7559 1 3.1371
0.2410 0.3188 1
1 2.2552 1.5682
0.4440 1 0.6963
0.6377 1.4362 1
1 0.1917 0.1803
5.2156 1 0.9403
5.5468 1.0635 1
1 0.5561 0.5167
1.7982 1 0.9291
1.9355 1.0763 1
By applying the AHP and the revised AHP (see Sections 2.2.3 and 2.2.4,
respectively, for the appropriate formulas) it can be shown that the following
results are obtained:
Derived AHP solution Derived Revised AHP solution

Alts. Weights Ranking Weights Ranking
Al 0.2073 3 0.2048 3
A2 0.3982 1 0.3980 1
A3 0.3945 2 0.3972 2
In reality, however, the decision maker does not know the previous
real values. Instead, one has to use matrices with entries from the set: {9, 8,
... , 2, 1, 1/2, ... , 118, 1/9} (that is, to use the scale presented in Table 3-1
in Section 3.2.1). The COP matrices which best approximate the previous
five RCP matrices are shown next.
For the case of the four decision criteria the corresponding COP
matrix and its corresponding relative weights (approximated eigenvector) is:
RCPo -+ CDPo =
1 1.5000 0.2500 0.5000 0.1304

0.6667 1 0.1667 0.3333 0.0870
= -+
4.0000 6.0000 1 2.0000 0.5217
2.0000 3.0000 0.5000 1 0.2609
For instance, the (1, 2) entry in the previous matrix is equal to 1.5000
because the corresponding entry in the RCP matrix is 1.4890 and the value
1.5000 is the closest value from the permissible values (when the scale
depicted in Table 3-1 in Section 3.2.1 is used). Similar interpretations hold
for the remaining entries.
Similarly with above, for the three alternatives in terms of each one
of the four decision criteria the COP matrices and their corresponding relative
weights are as follows:
1 1.3333 4.0000 0.5000
RCP l -;0.
CDP l 0.7500 1 3.0000 -;0. 0.3750
0.2500 0.3333 1 0.1250
1 2.2500 1.6000 0.4833
RCP2 -;0. CDP2 0.4444 1 0.7143 -;0.

0.2151
0.6250 1.4000 1 0.3016
1 0.2000 0.1667 0.0835
RCP3 -;0. CDP3 5.0000 1 0.8889 -;0. 0.4264
6.0000 1.1250 1 0.4901
1 0.5556 0.5000 0.2083
RCP4 -;0. CDP4 1.8000 1 0.8889 -;0. 0.3734
2.0000 1.1250 1 0.4183
That is, the data which are assumed to be available in reality to the
decision maker for the present problem are:
Criteria
C1 C2 C3 C4
Alts. (0.1304 0.0870 0.5217 0.2609)
Al 0.5000 0.4833 0.0835 0.2083

A2 0.3750 0.2151 0.4264 0.3734
A3 0.1250 0.3016 0.4901 0.4183
Similarly, as in the first part of this example, the AHP and the revised AHP
yield the following results:
Derived AHP solution Derived Revised AHP solution

Alts. Weights Ranking Weights Ranking
At 0.2051 3 0.2049 3
A2 0.3875 2 0.3867 2
A3 0.4074 1 0.4084 1
These findings clearly contradict the results found in the first part
of this analysis. In this numerical example the weights of the two alternatives
A z and A3 are almost identical when the AHP is applied (namely, 0.3982 and
0.3945, respectively) on the actual data. However, after the original Saaty
scale is used, and the assumption that the decision maker is as accurate as
possible is made, then the AHP yields the weights 0.3875 and 0.4074 for
alternatives A2 andA 3 , respectively. In other words, the AHP causes the two
alternatives to look very distinct, while in reality they are almost identical.
If for illustrative purposes one considers that the alternatives represent
projects to be funded, then the AHP would suggest that project A3 to be
funded 5.14% (i.e., (0.4074 - 0.3875) x 100 I 0.3875) more than project
A 2 , while in reality project A3 should be funded by almost 1.0% less than
project A z. That is, after the AHP or the revised AHP are applied, then it is
likely that alternatives which in reality are almost identical, to appear to be
rather distinct. The reverse situation is also possible to occur.
10.3 SOME COMPUTATIONAL EXPERIMENTS
In the previous numerical example both the AHP and the revised AHP
yield the same alternative ranking. However, in general this is not always the
case. It is possible for the AHP to yield the correct alternative ranking and
the revised AHP an incorrect one, or vice-versa. In order to study the
performance of the two methods in depth, a simulation approach was
followed. Random problems with different number of decision criteria and
alternatives were generated and then treated as in the previous extensive
numerical example.
For each such random problem actual relative weights were assumed
as in the previous numerical example. However, because the Saaty matrices
use values from the set: {9, 8, ... , 2, 1, 112, ... , 2/8, 2/9} only the random
problems which are associated with Rep matrices with entries within the
continuous interval [9, 119] were considered. Each test problem was treated
as in the previous example. If the AHP or the revised AHP yielded a ranking
different than the one derived when actual data were used, then the case was
recorded as a failure. Tables 10-1 and 10-2 present the results of this
simulation approach for random problems of different sizes. Table 10-1

deals with the AHP and Table 10-2 with the revised AHP. Figures 10-1 and
10-2 graphically depict the results presented in Tables 10-1 and 10-2,
respectively. The simulation program was written in FORTRAN and the
random numbers were generated by using the IMSL library of subroutines.
The findings reveal that the nmnber of alternatives in a decision
problem is very critical. As the number of alternatives increases, so does
the failure rate of the AHP and the revised AHP. This was expected since in
[Triantaphyllou and Mann, 1990] it was found that the number of entities to
be compared in the pairwise manner can dramatically influence the
performance of the eigenvalue approach in deriving accurate entity ratings.
However, in the present investigation the failure rates are much higher than
the ones reported in [Triantaphyllou and Mann, 1990]. It is worthy to note
here that in the tests reported in [Triantaphyllou and Mann, 1990] the revised
AHP performed significantly better than the original AHP. The same findings
also demonstrate that the number of decision criteria seems to be without any
particular influence on the final results.
It should be noted here that in these simulation experiments it was
assmned that the decision maker is as accurate as possible. This assumption
is imposed by the introduction of the concept of the CDP matrices (as
described in more detail in Section 3.3.1). Clearly, this is an assumption
which works in favor of the two methods. If one accepts the premise that the
decision maker cannot always be that effective in hislher judgments, then it
follows that in real life applications the failures of the two methods occur
even more frequently and intensively than what is reported in this chapter.
From the above observations it can be concluded that if one uses the
AHP or the revised AHP in order to select the best alternative, then he/she
may instead select an alternative which is not the best. Similarly, if one is
interested in determining the relative weights of the alternatives, then
distortions in the calculated weights are possible.
The answer to the critical question whether the numerical anomalies
reported here are severe or not, depends on the particular application under
consideration, and no generalizations can be made. If the impact of choosing
a non-optimal alternative is not that severe, then the decision maker should
not be highly concerned. However, if the application is very critical, then the
present investigation may serve as a warning for caution before reaching the
final decision. The present investigation provides some additional evidence
to the belief of many practitioners today, that MCD M methods should be used
only as decision support tools, and not to be taken literally.
Table 10-1: The Failure Rates are Based on 1,000 Randomly Generated Problems.
The AHP Case.
Nwnber of Alternatives
Nwnber of
Criteria 3 5 7 9 11
3 20.8 45.8 68.6 84.3 92.7
5 20.3 47.7 70.4 86.6 94.3
7 20.4 50.2 73.7 87.7 95.0
9 24.1 49.7 75.6 90.3 96.2
11 22.4 54.0 75.3 931 96.6
13 24.0 52.2 78.0 91.7 97.8
15 23.8 55.5 78.8 92.8 98.1
17 26.7 56.7 81.2 93.3 98.2
19 23.1 58.8 84.6 94.0 98.2
21 35.9 60.1 85.1 94.7 98.8
Nwnber of
Criteria 13 15 17 19 21
3 96.6 99.0 99.7 99.7 100.0
5 98.9 99.2 99.8 100.0 100.0
7 97.9 99.4 99.9 100.0 100.0
9 98.5 99.8 99.9 100.0 100.0
11 99.1 99.8 100.0 100.0 100.0
13 99.3 99.9 100.0 100.0 100.0
15 99.6 99.9 100.0 100.0 100.0
17 99.6 100.0 100.0 100.0 100.0
19 99.6 99.9 100.0 100.0 100.0
21 99.6 99.9 100.0 100.0 100.0

Table 10-2: The Failure Rates are Based on 1,000 Randomly Generated Problems.
Nwnber of
Criteria 3 5 7 9 11
3 22.8 48.1 69.3 86.0 93.4
5 20.4 46.8 72.6 87.5 93.5
7 21.1 51.4 75.3 88.5 95.7
9 24.1 52.4 75.0 89.7 97.4
11 23.6 56.1 76.6 92.8 96.0
13 25.7 55.2 79.3 92.6 98.0
15 26.9 55.9 82.1 94.3 97.7
17 27.7 57.0 82.7 93.9 98.7
19 25.5 56.6 84.0 94.0 98.7
21 28.2 59.8 83.1 93.6 98.5
Nwnber of
Criteria 13 15 17 19 21
3 97.5 99.3 99.6 99.8 100.0
5 98.4 99.5 99.8 99.9 100.0
7 99.1 99.5 100.0 100.0 100.0
9 99.0 100.0 99.8 100.0 100.0
11 99.9 99.7 100.0 100.0 100.0
13 99.2 100.0 100.0 100.0 100.0
15 99.5 100.0 99.9 100.0 100.0
17 99.3 100.0 100.0 100.0 100.0
19 99.7 99.9 100.0 100.0 100.0
21 99.7 100.0 100.0 100.0 100.0

Number of Alternatives
100 .....-' 21
90 ·~·=0·~ ·~·~·i·~··~· ~·~ ·~·~·~ +.:.;.~.~.~ .~.f·~·~;'·~·~·~·~·f ·~·~·~·~·~·~·f·~·=:·= ·~·~ .:. . 7
:·:~~=· =.t~~·~ :~~:¥:·::·:=~~-l·~~~·~·:1 "'·::::~l"~~·· ~

80
70
---~
'"
'"
"0
60
50
C···················· ,····························; · · · · · ··· ::·:.....==f=== 5
a::
'"
.2 40
'0 :3
LL
30
20
10
o
:3 6 9 12 15 18 21
Number of Decision Criteria
Figure 10-1,' The Failure Rates are Based on 1,000 Randomly Generated Problems.
The AHP Case.
Number of Alternatives
1~~.;.;'.~=.~.:r;.,. ;'.~.~.~. j.~~.~.=.:.=:.+~.

100 ~~~~=-~~~====~~~~~~~~~~~21
90 - - - - -< - - f -- ~ - ~ ~ ~ . =.;. ~ ~ ~.~ ~

.. . . .. ......~
....:.. ... . . ....-. . . . . . . ....:... . . . . ~.. . -... ..... .
. :. . . . . ...:.- . . . 7
80 ~ ........................... 1.........
.... ".. -', ...:.:.:.~:.; : . :. ~.:.~.:.:. :. ~.~.~.:. :. :. ~.i.;: :.:. ~, ~. : . ~.: . ~.:"' "
: \,;,
. " c; • • • • • • • • • '~:""""- <I"

• • • • .0- • • • • ; .
70 ····· __ ···········r······························f····.,.-,--.. ................. ..-...........-.. --.. -... -.. ................... '.......... .

~ ~
..-.
~
.,'"
60 ···················_······1······· ··················+········· · :·· =·· J·· =~··=···=···t··=···=···= ·):·· =~:::·· .:;;;. .-
- .+ . .
5
"0
a:: 50 ~. ::: ·~ ~ · =··. =r . ·~·· . · · . · · . . ·. ·I· .. ·..··· .. · ..···..·......·+··. ·. ··· . · . ··· . ·..··· . t·········· .. ········· ..·· .. ..
~
.2
'0
40 .............................~ .. , ..........-....- ........+. . . . . . . . . . . . . . . ~. . . . . . . . . . . . . .····t···. ·······..·············.. ·~·; ..·........·
:
.....
30 ......... ...................:.............................. ~ .... ·· .. ··· .. ··· .... ·..·.. +·························.. ··t...... ·..··· .... ·......······t···· .. ··· ..·.. ·· .... ~ 3
20 ~.. ........................ ,..... ·· ....·· .... ·.. t············ .. ·.. ··· .. ··· .. ··;· ..· ......................................................
10 ..............................:.. ... ---.....................; ...................... ,...·,··.;"···"···"··"··,····,····"·7"···'······.. ···.. ··.. ·······t .. ·,·· ..·
a L-~~~~~ __~-L__~~-J____~__~~__~-L~~~-J
:> 6 9 12 15 18 21
Number af Decision Criteria
Figure 10-2: The Failure Rates are Based on 1,000 Randomly Generated Problems .
10.4 CONCLUSIONS
In this chapter we investigated the reversals in the alternative ranking

which occur when Saaty pairwise comparison matrices are used as input to
MCDM problems. Saaty pairwise comparison matrices are widely considered
as successful means for gathering data in many real-world decision making
situations (see also Chapter 3). The AHP and some of its variants have found
a considerable appeal in approaching many real life engineering, science, and
economic applications. As it was mentioned in Section 10.1, many computer
packages have been developed (see, for instance, the Expert Choice or the
Criterium Decision Plus packages) to automate their application to real life
problems.
The forward error analysis described in this chapter reveals that when
the AHP or the revised AHP methods are used in combination with the
eigenvalue approach for processing the input data, then some severe
alternative ranking failures are possible. These failures are more dramatic as
the number of alternatives increases. However, even for decision problems
with a few alternatives the possibility of ranking alterations is significantly
high. The number of decision criteria or the MCDM method used (i.e., the
AHP or the revised AHP) seem to be without any particular influence in the
above rates. Therefore, these results can be used to properly guide the
decision maker when applying these methods to real life problems.
The present chapter provides an additional justification on why the
application of the AHP (or the revised AHP and perhaps all other MCDM
methods) should be treated with some skepticism and their results not to be
taken literally. This consideration comes in symphony with the concerns
expressed by many other researchers in the past (see, for instance, [Belton
and Gear, 1983], [Dyer, 1990a; and 1990b], [Dyer and Ravinder, 1983],
[Dyer and Wendell, 1985], [Winkler, 1990]). However, the assumption that
the decision maker is as accurate as possible in deriving his/her judgments
(i.e., the concept of the CDP matrices) and its application to evaluating
MCDM methods is a very new and powerful one.
It is worthy to emphasize here that in under less than perfect
conditions the real failure rates can be even more dramatic. It should also
be kept in mind, that the present investigation is based on the (rather
plausible) assumption that the actual data for evaluating alternatives in terms
of decision criteria, take on continuous values. Therefore, the previous
results are subject to this continuity assumption.
Finally, it should be stated here that the same investigation, besides
being a mechanism for testing other (present or future) MCDM models, it
also reveals that there is an open problem in this class of MCDM models and
more research in this topic is needed.
Chapter 11
MORE CASES OF RANKING ABNORMALITIES

WHEN SOME MCDM METHODS ARE USED
Chapter 9 presented an evaluation of the AHP, the revised AHP, the

WPM, and the TOPSIS methods in terms of two evaluative criteria. Another
evaluation was presented in Chapter 10 for the AHP and the revised AHP
methods. The issue of evaluating MCDM methods is a controversial one in
the decision analysis / decision making communities. This chapter is partially
based on some recent developments that are presented in [Triantaphyllou,
2000] and are related to the evaluation of MCDM methods. The present
chapter continues on the same subject and presents some new ranking
abnormalities when the AHP and some of its variants are used.
As it was also mentioned in Chapter 2 (Section 2.2.4) from the early
days of the AHP development it was observed that certain ranking
abnormalities may occur. Belton and Gear [1983] noticed that when identical
copies of existing alternatives are introduced in an MCDM problem, then it
is possible for the AHP to change the ranking of the alternatives. They
attributed this phenomenon to the fact that in the AHP the values of relative
performance of the alternatives in terms of each decision criterion in the
decision matrix are normalized so they add up to one. Thus, they proposed
that these values be normalized by dividing them by the largest entry of each
column of the decision matrix. When this modification is followed, then the
previous type of ranking reversal does not occur. This is the essence of the
revised AHP as was described in more detail in Section 2.2.4.
Saaty asserted that this kind of ranking reversal may be acceptable.
He used the example of a lady who wishes to buy a hat. She makes her first
selection, but then she discovers that her top choice is a very common one,
since there are many more copies of it in the store. As a result of this
discovery, she changes her mind and now she chooses another hat as the best
one. Although this illustration may capture a realistic situation, it cannot be
convincing in justifying the previous type of ranking abnormalities. In
Saaty's fable the lady does change her preferences and this results in the
alteration of her initial decision. In the example used by Belton and Gear,
however, no change on the numerical data is assumed. Nevertheless, a
ranking change does occur. Although Saaty was very hesitant to accept any
modifications to the AHP approach, he later accepted the previous variant of

the AHP and now it is also called the ideal mode AHP.
The issue of ranking reversals continues to be a hot and controversial
one. This subject has also been studied by Dyer and Ravinder [1983] and
Dyer and Wendell [1985]. Saaty in [1983] and [1987] responded to these
criticisms with some axioms and guidelines on how close a near copy can be
to an original alternative without causing a rank reversal. He asserted that the
decision maker has to eliminate alternatives from consideration that score
within 10% of another alternative. This recommendation was later sharply
criticized by Dyer [1990a].
Dyer in [1990a; and 1990b] claimed that the original AHP may result
in arbitrary rankings, even when no identical alternatives are considered. He
also asserted that in some cases the AHP may yield results which may be
highly correlated with the actual performances of a consistent decision maker,
and even provide an accurate ranking of alternatives in other special cases.
However, the same author claimed that, in general, this cannot be guaranteed
a priori.
Dyer attributed this problem to the principle of hierarchic
composition, which is one of the main assumptions in the AHP. This
principle assumes that the weights of the criteria do not depend on the
alternatives under consideration. As an alternative approach, Dyer suggested
that the AHP be combined with utility theory. That paper by Dyer stimulated
some critical responses by Saaty [1990] and Harker and Vargas [1990].
As Winkler in [1990] remarked in a related technical note, multi-
attribute utility theory (MAUT) is not a panacea. He cautions the reader to
be careful in accepting alternative theories. Clearly, these problems are still
controversial in decision theory. Some additional discussions beyond the ones
given in Chapters 9 and 10, on these truly important issues are due to Bell
and Farquhar [1986], Hogarth and Reder [1986], Weber and Camerer [1987],
Munier [1988], Fishburn and LaValle [1989], and Sarin [1989]. In contrast
with all previous analyses, the new ranking abnormalities described in this
chapter do not involve the introduction of any new alternatives.
The new type of ranking abnormalities are based on decomposing an
MCDM problem into a set of smaller problems. Each such problem is
defined on a pair of the alternatives. It is important to state here that no
change of the numerical data is assumed (except any normalizations which do
not alter the relative preferences). Then, the rankings derived from the
smaller problems are compared with the rankings derived when all the
alternatives are considered simultaneously. The above are best described in
terms of the following illustrative examples.
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 215
11.2 RANKING ABNORMALITIES WHEN

ALTERNATIVES ARE COMPARED
TWO AT A TIME
The main ideas in this section are best described in terms of two
illustrative examples. The first example deals with the original AHP and it
is presented next.
Example 11-1: The Case of the Original AHP
Suppose that the following is the decision matrix of a simple problem

with three criteria and the three alternatives AI' A 2, and A3:
Criteria
CI Cz Cj
Alts. (2/7 2/7 3/7) Combined
Priorities:
Al 9/19 2/12 2/7 0.305
Az 5/19 1112 4/7 0.344
Aj 5/19 9/12 117 0.351
Note that the above data are assumed to have been derived from
perfectly consistent judgment matrices with pairwise comparisons. This was
assumed in order to block out any effects due to inconsistent pairwise
comparisons among the alternatives in terms of each one of the decision
criteria. When the original AHP is applied on the above decision matrix, then
it can be easily seen that the three alternatives have the priority values shown
under the "Combined Priorities" vector. For instance, the combined priority
for alternative Al is 0.305, because: (9/19) X (2/7) + (2/12) X (2/7) +
(2/7)x(37) = 0.305. Therefore, in the maximization case, the above
combined priorities indicate that the three alternatives are ranked as follows:
Aj > A z > Al (where the symbol "> " as before stands for "is better than ").
In the following paragraphs the three alternatives are compared in a
pairwise fashion. That is, the original problem is broken into three smaller
problems, each one having two alternatives. The first such problem (which
considers the pair of alternatives A2 and A 3) is described by the following
decision matrix:
Criteria
CI C2 C3
Priorities: Derived Ranking:
5/10 1110 4/5 0.5143 A2 > AJ
5/10 9/10 1/5 0.4857
The columns in the above matrix have been normalized again, in order to be
consistent with the basic requirement of the original AHP (which requires the
vectors to add up to one).
Similarly, the second such problem (which considers the pair of
alternatives A] and A 2) is described by the following decision matrix:
Criteria
CI C2 C3
9/14 2/3 2/6 0.5170 A] > A2
5/14 1/3 4/6 0.4830
Finally, the third and last problem (which considers the pair of alternatives
A] and A J) is described by the following decision matrix:
Criteria
CI C2 C3
9/14 2111 2/3 0.5213 A] > A3
5/14 9/11 1/3 0.4787
The previous results, when are taken together, indicate that the
ranking of the three alternatives must be: Al > A2 > A 3. That is, this
ranking is different of the one derived when all the alternatives were
considered simultaneously (namely: A3 > A2 > AI). In matter of fact, the
second ranking is the reverse of the first one! Therefore, the questions which
naturally are raised at this point are: "Which one is the best ranking?" and
"What are the correct combined priorities of the three alternatives?" Before
we proceed to answering these questions, we present some further
elaborations on this illustrative example. Suppose that the combined priorities
of the three alternatives are as usually denoted as PI' P2 , and P3 ,
respectively. In the light of this notation, the previous three smaller problems
imply that:
P/Pj :::::: 0.5143/0.4857 = 1.058884,

P/P2 :::::: 0.517010.4830 = 1.070393,
and P/Pj :::::: 0.5213/0.4787 = 1.088991.
In the above relations the approximation symbol "::::::" is used instead of
equality. This is necessary because if these were equalities, then from the
obvious relation: (P/P2) x (P/P j) = P/Pj , then the following relation would
also had to be true: (1.070393) X (1.058884) = 1.088911. However, this m
not true because the left-hand-side product is equal to 1.133422.
However, the previous three pairwise comparisons lead to the
introduction of the following pairwise comparison matrix:
Alts. Al Az A3 Combined
Priorities
Al 1 1.070393 1.088991 0.3505205
A2 111.070393 1 1.058884 0.3318642
A3 111.088991 111.058884 1 0.3176152
The above combined priorities were derived from the previous pairwise
comparisons by applying the Saaty approximation of estimating the right
principal eigenvector of the previous reciprocal matrix. This approximation
is done by multiplying the elements in each row and then by taking the n-th
root (i.e., the third root in this example) [Saaty, 1980] (or Section 4.2). The
values obtained that way are then normalized and the "Combined Priorities"
vector is obtained.
It can be noticed now that in this matrix the alternatives are compared
in terms of all the three criteria combined. This is a novel step since
traditionally in the AHP context alternatives are compared in terms of a single
decision criterion at a time. At this point it becomes apparent that the ranking
of the three alternatives implied by the last combined priorities is: Al > A2
> A j • Obviously, this ranking is identical to the ranking obtained from the
combination of the solutions of the earlier three sub-problems. Again, this
ranking is different (actually reverse) than the ranking obtained when all the
alternatives are considered simultaneously . •
A natural thought at this point is whether one should always analyze

an AHP problem as in the previous example. The answer to this question is
negative. The reason is that sometimes the ranking obtained from the last
pairwise matrix may be contradictory to the ranking implied by the individual
pairwise comparisons. This situation is further demonstrated in the next
numerical example.
Example 11-2: The Case o/the Ideal Mode AHP
Similarly with the previous example, the case of the ideal mode AHP
is best described in terms of an example. Suppose that the following is the
decision matrix of a simple problem with three criteria and the three
alternatives AI' A 2, and A):
Criteria
CI C2 C)
Alts. (4122 9/22 9/22) Combined
Priorities:
Al 9/9 5/8 2/8 0.5398
A2 119 8/8 5/8 0.6850
A) 8/9 2/8 8/8 0.6730
As earlier, the above data are also assumed to have been derived from
perfectly consistent judgment matrices with pairwise comparisons. When the
ideal mode AHP is applied on the above decision matrix, then it can be easily
seen that the three alternatives have the priority values shown under the
"Combined Priorities" vector. Therefore, in the maximization case, the above
combined priorities indicate that the three alternatives are ranked as follows:
A2 > A3 > AI'
Similarly with the discussion in Example 11-1, next the three
alternatives are compared in a pairwise fashion. That is, the original problem
is broken into three smaller problems, each one having two alternatives. The
first such problem (which considers the pair of alternatives A2 and A) is
described by the following decision matrix:
Criteria
CI C2 C)
Alts. (4122 9/22 9122) Combined
118 8/8 5/8 0.687500 AJ > A2
8/8 2/8 8/8 0.693182
Now the columns in the above matrix have been normalized in the way
prescribed by the ideal mode AHP (that is, the elements in each vector are
divided by the largest value in that vector).
Similarly, the second such problem (which considers the pair of the
alternatives Al and A2) is described by the following decision matrix:
Criteria
CI C2 C3
Alts. (4/22 9/22 9/22) Combined
9/9 5/8 2/5 0.601136 A2 > Al
119 8/8 5/5 0.838384
Finally, the third and last problem (which considers the pair of
alternatives Al and A3) is described by the following decision matrix:
Criteria
CI C2 C3
Alts. (4122 9/22 9122) Combined
9/9 5/5 2/8 0.693181 A3 > Al
8/9 2/5 8/8 0.734343
The previous results, when are taken together, indicate that the ranking of the
three alternatives must be: A3 > A2 > AI' Similarly with the first example,
this ranking is different of the one derived when all the alternatives were
considered simultaneously (namely: A2 > A3 > AI)'
As with the previous illustrative example, it can be observed that the
following relations must be true:
PiP3
0.687500/0.693182 = 0.991803,
~
P/P2
0.601136/0.838384 = 0.717018,
~
and ~ 0.693182/0.734343 = 0.943948.
P/P3
These pairwise comparisons lead to the introduction of the following pairwise
comparison matrix and then to the attached combined priorities:
Alts. Al A2 AJ Combined
Priorities
Al 1 0.717018 0.943948 0.2912657
A2 1/0.717018 1 0.991803 0.3696254
AJ 110.943948 110.991803 1 0.3391087
These combined priority values (which now add up to one) indicate

that the ranking of the three alternatives must be: A2 > A3 > AI' However,
when the pairwise comparisons in the previous matrix are considered one at
a time, then the derived ranking is different. Namely, this ranking is: A3 >
A2 > AI. This ranking is (by definition) identical to the ranking derived from
solving the three smaller sub-problems and then combining the partial
solutions. That is, in this case the ranking derived from the combined
pairwise matrix is different than the ranking derived when the three smaller
problems are considered. Clearly, this was not the case in Example 11-1.
Besides the previous categories of ranking abnormalities, a different type of
ranking abnormalities may occur and it is discussed in the next section. •
11.3 RANKING ABNORMALITIES WHEN

ALTERNATIVES ARE COMPARED TWO
AT A TIME AND ALSO AS A GROUP
The previous discussions clearly raise the issue that when the original
or ideal mode AHP is used, then sometimes it may not be very obvious what
is the correct ranking. Should one accept the ranking derived when all the
alternatives and decision criteria are used? Or should one accept the ranking
derived from combining the solutions of the smaller problems which consider
two alternatives (and all the criteria together)? It may also be the case when
the matrix with the combined pairwise comparisons is considered, then one
may derive a third ranking which is different than the previous two rankings.
The above scenarios are possible when the original or ideal mode AHP is
used.
One may argue here that when the decision maker encounters a
problematic situation as in the previous scenarios, then he/she must be more
careful in accepting the final results. A re-evaluation of the characteristics of
the problem (especially if new decision criteria can be introduced) may result
in more robust data which may lead to less ambiguous conclusions.
However, as the two numerical examples discussed in the previous sections
demonstrate, it is possible to reach inconclusive results even when all the
pairwise comparisons (i.e., the ones which consider pairs of alternatives in
terms of a single criterion) are perfectly consistent. Recall that this was
actually the case with the data in the decision matrices in these examples.
Therefore, the first question to be answered is which is the right ranking of
the alternatives.
Deciding on which one is the right ranking can be an open-ended
question. One may argue that the correct ranking is the one derived when all
the alternatives and decision criteria are considered simultaneously. However,
one may also claim that the solutions of the smaller problems are more
reliable because these involve simpler (only two alternatives are considered
at a time) decision problems. If one accepts the premise that simpler

problems lead to more reliable solutions, then the next statement to be
accepted is that the ranldng obtained by combining the partial rankings of
the smaller problems is the most reliable one.
Besides the last reason, there is a second argument why the ranking
obtained from the smaller problems might be more reliable than the ranking
obtained when all the alternatives are considered simultaneously in terms of
all the criteria. As it was mentioned at the beginning of this chapter and also
in other chapters, both the original and ideal mode AHP have been sharply
criticized because their ranking of alternatives may change when new or
copies of existing alternatives are introduced in a decision problem. These
observations formed the basis of many controversial disputes in the scientific
and practitioners communities regarding the validity of the AHP method.
When the decision maker accepts the ranking derived by combing the
solutions of the smaller problems, then the previous abnormalities cannot
occur. This is obviously true because the ranking of the alternatives is based
on how they are ranked when they are considered two at the time. When new
alternatives are introduced, then the results of comparing existing alternatives
remain identical as before the introduction of the new alternatives. Therefore,
the relative ranking of the old alternatives will never change as result of
introducing additional alternatives.
However, it is possible the previous sequence of smaller problems
still to result in ranldng abnormalities. More specifically, when the original
or the ideal mode AHP are used for comparing two alternatives at the same
time, then it is possible to reach a new type of ranking contradiction. This
is best illustrated in the following example.
Example 11-3: The Case o/the Ideal Mode AHP
As in the previous illustrative examples, a problem with four criteria

and three alternatives is considered. The aij data are real (non-normalized)
numbers from the continuous interval [9, 1].
Criteria
C1 C2 C3 C4
Alts. (0.27 0.41 0.05 0.27)
Al 1.92 7.59 1.27 6.13

A2 3.12 4.31 8.57 7.11
A3 7.70 4.77 7.45 3.29
222 MCDM Methods: A Comparative Study. by E. TriantaphyLLou
Suppose that one uses the ideal mode AHP. When alternatives A I and
A z are compared as in the previous two examples, then it is derived that: Az
> AI. Similarly, when alternatives AI and AJ are compared, then it is derived
that: AI > AJ. That is, these relations suggest that the ranking of the three
alternatives must be as follows: A z > Al > A3. However, when the two
alternatives A3 and Az are compared, then the derived ranking is: A3 > A z.
That is, a logical contradiction (Le., a case of logical inconsistency) is
reached. A similar phenomenon can be observed with test problems which
use the original AHP method. •
Dyer in [1990a] claimed that any kind of ranking abnormalities will

be eliminated if the final stage of the AHP is modified. His suggestion was
to subtract the smallest value in each vector in the decision matrix and then
divide by the largest remaining element. However, as the following
numerical example illustrates, ranking abnormalities are still possible to
occur.
Example 11-4: The Case of an AHP Variant (as proposed by Dyer,

[1990aJ)
Consider the following problem with four criteria and three

alternatives. As before, the aij data are real (non-normalized) numbers from
the continuous interval [9, 1].
Criteria
CI Cz C3 C4
Alts. (0.43 0.12 0.06 0.39)
Al 6.78 7.19 3.28 5.10

Az 4.26 1.44 8.20 5.76
A3 5.52 3.99 7.26 8.14
When Dyer's suggestion is applied on the previous matrix, the following

decision matrix is derived:
Criteria
CI Cz C3 C4
Alts. (0.43 0.12 0.06 0.39)
Al 1.00 1.00 0 0
Az 0 0 1.00 0.22
A3 0.50 0.44 0.81 1.00
Then, the final priorities of the three alternatives are: PI = 0.550, P2 =

0.145, and P3 = 0.707. Therefore, the suggested ranking of these
alternatives is: A3 > Al > A2·
Next, suppose that the three alternatives are examined two at a time
(as was done in the previous examples). It can be easily verified that when
the pair Al and A2 is considered and Dyer's suggestion is applied on this
smaller problem (which now has two alternatives and four decision criteria),
the derived ranking is: Al > A2. Similarly, the derived ranking when the
pair A2 and A3 is considered the derived ranking is: A3 > A 2. Also, when the
pair Al and A3 is considered, the ranking becomes: Al > A 3. Thus, the
ranking derived when the alternatives are considered two a time is: Al > A3
> A2 • Obviously, this ranking is in contradiction with the one derived when
all the alternatives were considered simultaneously (i.e., A3 > Al > A2) . •
11.4 SOME COMPUTATIONAL RESULTS
The previous illustrative examples clearly demonstrate that certain

types of ranking abnormalities may occur when the AHP, or some of its
additive variants, are used. In order to gain a deeper understanding on how
frequently such abnormalities may occur on random test problems, a
computational study was undertaken. The data were random numbers from
the interval [9, 1] (in order to be compatible with the numerical properties of
the Saaty scale). In these test problems the number of alternatives was equal
to the following 10 different values: 3, 5, 7, ... , 21. Similarly, the number
of criteria was equal to 3, 5, 7, ... , 21. Thus, a total of 100 (= 10 x 10)
different cases were examined with 10,000 replications (in order to derive
statistically significant results) per each case. Each random problem was
solved using the original and the ideal mode AHP. The test problems were
treated as the previous illustrative examples (i.e., Examples 11-1 to 11-3).
Any ranking irregularity was recorded. Figures 11-1 to 11-6 summarize these
results.
Figures 11-1 and 11-2 depict how often the indication of the best
alternative was different when all the alternatives were considered
simultaneously and when they were considered in pairs (i.e., similar to the
analysis in Examples 11-1 and 11-2, respectively). Figure 11-1 refers to the
use of the original AHP, while Figure 11-2 to the use of the ideal mode AHP.
The different curves correspond to problems with different numbers of
alternatives. As it can be seen from these figures, problems with few
alternatives had smaller contradiction rates. The number of the decision
criteria in a problem seemed to play an insignificant role. Also, these figures
show similar contradiction rates for the two versions of the AHP.
On the other hand, Figures 11-3 and 11-4 depict contradictions in the
ranking of any alternative. Now the number of alternatives plays a decisive
role, while the number of decision criteria is not as important. Moreover, the
contradiction rates are significantly more dramatic. For instance, for
problems with five alternatives, the contradiction rates are almost 50%. As
before, there is no much difference between the results obtained when the
original or the ideal mode AHP was used. As it was expected, the
contradiction rates in Figures 11-3 and 114 are much higher than those in
Figures 11-1 and 11-2. This was expected because the cases of contradiction
in the first two figures are naturally included in the results of Figures 11-3
and 11-4.
Figures 11-5 and 11-6 present the contradiction rates when the
alternatives are compared two at a time and cases of logical inconsistencies
were found (as in Example 11-3). The roles of the number of alternatives and
number of decision criteria are similar as before. However, now the ideal
mode AHP performs significantly worse than the original AHP.
In all these results problems with less alternatives yielded fewer
ranking contradictions than problems with more alternatives. This was
expected because the number of alternative pairs to be considered for a given
case is directly related to the number of alternatives in the problem. In matter
of fact, if the number of alternatives is equal to m, then the number of pairs
to be considered is equal to m(m-l)/2. Thus, the chances of finding a logical
inconsistency increase accordingly. As it can be seen from the illustrative
examples, the number of criteria did not playa prime role. This is also
evident in the computational results by the almost horizontal curves.
100 ;
90 ---. ~- !i .
_______ .1 __ • ______L _ _ _ _ _ • !o------.~. .---.-.----.;.. . -.--
! ;
80 ----1-.. -------~-------~r-·----t·-----i--~---. ·-~ . -..···--·---·-·----·---r--
~! .:=t=+=+
.-...
~ 70
~
0::
0 60
c::: !! 1 ~
~
0
:;J 50
.~
'0
g 40
c:::
~+
0
u 30
20
10
. : '{..Probler6s w ti h 31 Alternot~es
0
4 6 8 10 12 14 16 18 20
Number of Criterio
Figure 11-1: Contradiction Rates on the Indication of the Best Alternative When
Alternatives are Considered Together and in Pairs. The Original AHP
Case.
100
I
~++=r++~=:::=:::::::::r
90
80
; I i i I f I
.-... 70
~
4>
'0 60
0::
c:::
0 50
~
'0
~ 40
C
0
u 30
i i i i I Prob l~ms with : 21 Alter?otives i
~~~;~r4i.:t3~~,~~::,:i~
20
10
' . ' i ' Problems with ! 3A1etre ~ otives
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 11-2: Contradiction Rates on the Indication of the Best Alternative When
Alternatives are Considered Together and in Pairs. The Ideal Mode
(Revised) AHP Case.
100 r- ....... . ... ..
90 ~i~:~ -=-t~:--=-.J.:.~: ~ [ ~ .:. ~1~ .: ~ ~! ~ ~ ~ ~1~:;~. ~.~.;. ~J .~~

.~ . . i i ! Prbblems with 7 Alterna tives j
tEL=3 ~F~
80
~
~
70
'"
'0
a:::
60
! ; . , I /Prpblems 1ith 5 Alt1rnotives !
c:
50
·r~tt31~
:e'6
0
40
~
c:
0
. --..+-/-.. +-..--···-l·······-..----~···-. ·-. ·. . ·. -..i·. .-

30
. . -.,. . . . -..+-.! . . -.-1--.i.-.. . -i:--..·i
u
!
Prbblems ~ith 3 Alt~ rnotives
20
10 ~ :-------.. -f-------.f----.----~----.-I
i j j t : i i i
--_ ..,!, ..... _ - ' - , .. .j,!,........ _ ...."".._-,!.: _._--
! ! ! ! !
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 11-3: Contradiction Rates on the Indication of Any Alternative When Alternatives
are Considered Together and in Pairs. The Original AHP Case.
100
90 ~J.~'~.~.;c.:.:.~)~-:: ~ ~ ~ .: ~i ~ .: ~ ~r ~ ·~J.;~~f~ : ~.l~

. i i ! !
:.
! Problem~ with 7 A tlemativ'es j
80
70
g
2 60
0
a:::
c: 50
:3
.~
'0 40
E!
C 30
0
u
I !
20 ~
i
. --I·---.. .---..i! . - . ·. ---t-------·j--7f-----t·----.
!
.t-.~JtemotiJes i
. .·---t--------i----·
Problemk with 3
1 0 _ ;., ·~-·-····'··- ..··l... +-..-----f---.----..t-- .1-----'-~.!' -

I ' . -~,!,! .- .1-_
o
I I i i
4 6 8 10 12 14 16 18 20
Number of Criteria
are Considered Together and in Pairs. The Ideal Mode (Revised) AHP Case.
100
90
80
g 70
"
OJ
60
cr
c:
.2 50
-0
v 40
~
C
a 30
U
20
10
0
4 6 8 10 12 14 16 18 20
Problems with 3 A tlernatives
Number of Criteria
are Considered in Pairs. The Original AHP Case.
Problems with 21 Alternatives
i _ .-+--* .- - ; - - - - ! - -*·-··-i-
100
·-·-l-··:::·..·~t""~·7;...- -j. .e.:::.-:.: ...-·-;-:.~:.:.:..::::::::f·:~:::~:·:::::.::[:·::~:::·:::~:::::::::·::~~~::~:~·;~:;~·i=~:·:·:

i
90
1"" ~/ ! i i i ' Problems with 15i Alternatives
80
g 70
OJ
"
cr
c:
.2
60
50
-0
'<;
e 40
ca ./1 .... i : i j l.Problems with 9: Alternatiites
u 30
20
10
0
4 6 8 10 12 14 16 18 20
Numbe r of Criteria
are Considered in Pairs. The Ideal Mode (Revised) AHP Case.
11.5 A MULTIPLICATIVE VERSION OF THE AHP
All the previous problematic situations are caused by the required

normalization (either by dividing by the sum of the elements or by the
maximum value in a vector) and the use of an additive function on the data
of the decision matrix for deriving the final preference values of the
alternatives. Absolute data cannot be used when the decision criteria are
defined on different units (such as dollars, minutes, pounds, etc.). Therefore,
one must normalize the pertinent data one way or another. Unfortunately,
this normalization step and the use of additive functions may lead to
inconclusive and erroneous results as was the case in the previous numerical
examples.
As it was stated in the previous paragraph, if the decision maker deals
with qualitative data, or with multiple units of measure, then he/she must use
some kind of relative data. That is, the aij data must be normalized by either
dividing by the sum of the entries in that vector, or by the maximum value,
or by some other way. However, the use of an additive function can be
easily avoided. This is the case, for instance, in the weighted product model
(WPM) which was described in Section 2.2.2.
The use of multiplicative formulas in deriving the relative priorities
in decision making is not new [Lootsma, 1991]. A critical development
appears to be to use multiplicative formulations when one aggregates the
performance values aij with the criteria weights wj • This is the core step in
the WPM method. Barzilai and Lootsma in [1994] and later Lootsma in
[1999] have proposed to use a multiplicative variant of the AHP in order to
model power relations in group decision making. Please recall from Section
2.2.2 that when two alternatives, say Al and A2 are compared, then alternative
Al is more preferred than alternative A2 (Le., Al > A2), if the following
relation is true:
iI (ali)w/
au1=1
> 1, or:
(11-1)
II (alit' II (a r'·
/I /I
> 2I
1= 1 i = 1
It can be noticed here that the last two products express the priority
values of the corresponding two alternatives under the WPM method. These
priority values' can be normalized and/or ranked and thus derive the ranking
of a set of alternatives. In [Triantaphyllou, 2000] it is proved that this
multiplicative variant of the AHP possesses certain interesting properties. In
particular, it is proven theoretically that the ranking abnormalities studied in
this chapter cannot occur under this method.
Example 11-5: Application of the Multiplicative AHP
We consider the data used in Example 11-3 (also presented below).
Criteria
CI C2 C3 C4
Alts. (0.27 0.41 0.05 0.27)
Al 1.92 7.59 1.27 6.13

A2 3.12 4.31 8.57 7.11
A3 7.70 4.77 7.45 3.29
When the ideal mode AHP was applied, the ranking derived when two
alternatives were considered at a time, had some internal contradictions.
Next, the multiplicative formula (11-1) is applied on these data. The first pair
to consider is alternatives Al and A2. The corresponding calculations are:
R(AlAJ = (1.92)0.27 X (7.59)0.41 x (1.27)0.05 X (6.13)0.27 = 0.966 < 1.

3.12 4.31 8.57 7.11
Therefore, these alternatives are ranked as follows: A2 > AI' It can
be observed here that the data within the parentheses can be normalized in any
possible way, but the values of the ratios will remain unchanged. Next, the
pair A2 and A3 is considered:
J«NAJ -_ (3.12)0.27
- (4.31)0.41 x (8.57)0.05
x - -
(7.11)0.27
x - = 0.932 < 1.
7.70 4.TI 7.45 3.29
Similarly with above, these alternatives are ranked as follows: A3 >
A 2 • From the last two rankings the following ranking is derived for all three
alternatives: A3 > A2 > AI' Observe, that when the last pair Al and A3 is
considered, then the derived ranking is always in perfect agreement with the
previous global ranking of the three alternatives. The corresponding
calculations for the pair Al and A3 are presented next:
_ (1.92)0.27 (7.59)0.41 (1.27)0.05 (6.13)0.27

R(AlAJ - 7.70 x 4.TI x 7.45 x 3.29 = 0.900 < 1.
The above result could also had been obtained by observing that:
R(AtfA 3) =: R(A 1/ A 2 ) I R(AJA2) =: R(AtfA 2) x R(AiA~
0.966 x 0.932 = 0.900 < 1.

This was naturally expected because as it was proved in

[Triantaphyllou, 2000], all the ranking abnormalities studied in this chapter
are not possible when the recommended multiplicative version of the AHP is
used. Obviously, this was not the case in Example 11-3, where a ranking
irregularity had occurred (it is worth recalling here that the ranking suggested
at the beginning of Example 11-3 was: Az > Al > A3).
Next, suppose that the relative performance values of the previous
three alternatives in terms of the four decision criteria were available. That
is, the following decision matrix is assumed to be known:
Criteria
CI C2 C3 C4
Alts. (0.27 0.41 0.05 0.27)
AI 0.151 0.455 0.073 0.371

A2 0.245 0.259 0.496 0.430
A3 0.604 0.286 0.431 0.199
When formula (11-1) is applied on the previous decision matrix, the derived
priority values (before normalization) are: PI = 0.735, Pz = 0.742, and P3
= 0.754. Therefore, the ranking of the three alternatives is: A3 > Az > AI'
which is identical (as it should be) with the one found earlier. •
11.6 COMPARATlVERESULTSFROMTWOREALLlFE
CASE STUDIES
The previous empirical analyses revealed that ranking abnormalities

may occur frequently on simulated problems. The question which is raised
at this point is whether the same is true with real life problems. This section
briefly describes what happens in two problems with real life data. These two
problems were collected randomly from the literature.
11.6.1 Ranking Analysis of the "Bridge Evaluation" Problem

The "Bridge Evaluation" problem is described in [Saaty, 1994, page
90]. In this problem the alternatives are seven bridges in Pittsburgh, PA.
These bridges are evaluated in terms of seven decision criteria. The pertinent
data can be summarized as follows:
Criteria
C1 C2 CJ C4 Cs C6 C7
Alts. (0.1760 0.2360 0.0700 0.0790 0.2190 0.0300 0.1900)
Al 0.2114 0.2468 0.2286 0.2895 0.3516 0.1667 0.3613

A2 0.1314 0.1106 0.1714 0.1711 0.1553 0.1667 0.1257
AJ 0.1314 0.1277 0.1429 0.1711 0.1370 0.1667 0.1309
A4 0.1314 0.1489 0.1143 0.1053 0.0822 0.3333 0.1204
As 0.1371 0.1234 0.1286 0.1053 0.1050 0.1333 0.1047
A6 0.1371 0.1149 0.0857 0.0921 0.1050 0.1333 0.0890
A7 0.1200 0.1277 0.1286 .0.0658 0.0639 0.1000 0.0681
When the original AHP, ideal mode AHP, and the multiplicative AHP are
applied on the previous data, the priorities and rankings depicted in Table Il-
l are derived.
It can be observed from this table that the original and ideal mode
AHP derived identical rankings, while the ranking derived by the
multiplicative AHP differs on the third and second position (marked with "*"
on this table).
Table 11-1: Priorities and Rankings of the Alternatives in the "Bridge Evaluation" Case
Study [Saaty, 1994].
Original AHP Ideal Mode AHP Multiplicative AHP
Alt. Relative Ranking Relative Ranking Relative Ranking

Priority Priority Priority
AI 0.2841 1 0.2718 1 0.2821 1
A2 0.1368 2 0.1358 2 0.1381 3'
A3 0.1358 3 0.1355 3 0.1383 2'
A4 0.1238 4 0.1269 4 0.1220 4
As 0.1168 5 0.1194 5 0.1186 5
A6 0.1078 6 0.1103 6 0.1088 6
A7 0.0949 7 0.1004 7 0.0919 7
When all the alternatives are evaluated simultaneously, then the

derived ranking is: Al > A z > A3 > A4 > A5 > A6 > A 7 • However, when
the pair of alternatives Az and A3 is considered, then it can be easily verified
that under the AHP and the ideal mode AHP the derived ranking is: A2 < A 3 .
That is, a case of ranking reversal occurs. All the other pairs yield rankings
consistent with the original ranking. This is true when the original or the
ideal mode AHP method is used. It should also be stated here that no logical
inconsistencies as the ones found in Example 11-3 were detected with these
data.
11.6.2 Ranking Analysis of the "Site Selection" Problem
As a second illustration the "Site Selection" problem, as described by

Hegde and Tadikamalla in [1990], is examined next. The numerical data of
this case are depicted in the next decision matrix. Details on the
interpretation of these data and the meaning of the decision criteria and
alternatives can be found in [Hegde and Tadikamalla, 1990].
Criteria
CI Cz C3 C4 Cs C6 C7
Alts. (0.1760 0.2360 0.0700 0.0790 0.2190 0.0300 0.1900)
Al 0.2114 0.2468 0.2286 0.2895 0.3516 0.1667 0.3613

Az 0.1314 0.1106 0.1714 0.1711 0.1553 0.1667 0.1257
A3 0.1314 0.1277 0.1429 0.1711 0.1370 0.1667 0.1309
A4 0.1314 0.1489 0.1143 0.1053 0.0822 0.3333 0.1204
As 0.1371 0.1234 0.1286 0.1053 0.1050 0.1333 0.1047
A6 0.1371 0.1149 0.0857 0.0921 0.1050 0.1333 0.0890
A7 0.1200 0.1277 0.1286 0.0658 0.0639 0.1000 0.0681
It can be easily shown that when all the alternatives are considered
simultaneously and the AHP is applied, then the derived ranking is: A1 > A2
> A 3 > A4 > A5 > A6 > A 7• However, when the alternatives A 3 and A2 are
compared as a pair and the AHP method is used, then their ranking becomes:
A3 > A2 . All the other pairs of alternatives yield rankings consistent with the
original one. The same phenomenon also occurs when the ideal mode AHP
is applied on these data. Finally, it should be stated here that many more
such case studies examinations can be found in [Triantaphyllou, 2000].
11. 7 CONCLUSIONS
The analyses on ranking abnormalities presented in this chapter, along
with other studies on ranking abnormalities, re-enforce a growing belief
among many decision analysts that ranking abnormalities are unavoidable
when the AHP or some of its additive variants are used. This phenomenon
seems to be an inherited difficulty when one deals with criteria which are
defined on different units. This chapter has demonstrated, via some
numerical examples, that ranking abnormalities are possible even when the
data are perfectly known (and thus perfectly consistent). Therefore, it can
very well be the case that the decision maker may never know the exact
ranking of the alternatives in a given MCDM problem when the original
AHP, or its current additive variants, are used. The computational study,
along with some additional analyses reported in [Triantaphyllou, 2000]
indicate that such ranking abnormalities may occur dramatically too often.
A multiplicative variant of the AHP, as described in [Barzilai and
Lootsma, 1994] and in [Lootsma, 1999] is also advocated in this chapter.
This multiplicative variant of the AHP is a direct adaptation of the well
known weighted product model (WPM) [Miller and Starr, 1969] and [Chen
and Hwang, 1992] in the last step of the AHP, in which the decision matrix
is processed and the final ranking of the alternatives is determined (see also
Section 2.2.2). In this way, no more ranking abnormalities of the type
described in this chapter are possible.
It should be emphasized at this point that the fact that the proposed
multiplicative AHP does not possess any of the ranking abnormalities studied
in this chapter does not necessarily guarantee that this method is perfect (i.e.,
the rankings derived are always the true ones). However, it is obvious that
the reverse argument should always be true. That is, a perfect MCDM
method should never possess any of the ranking abnormalities studied in this
chapter.
It is also important to state here that a fuzzy version of the WPM
performed poorly in the empirical study described in Chapter 13 (see
especially Section 13.7.1). Please also recall the contradiction rates between
the WSM and the WPM methods (in crisp environments) as studied in
Chapter 9. Ranking abnormalities in MCDM methods is a very important,
fascinating, and also very controversial issue in decision analysis, and more
research is required.
Chapter 12
FUZZY SETS AND THEIR OPERATIONS
12.1 BACKGROUND
- "
INFORMATION
For a long time it has been recognized that an exact description of
many real life physical situations may be virtually impossible. This is due to
the high degree of imprecision involved in real world situations. Zadeh, in
his seminal papers [Zadeh, 1965; and 1968], proposed fuzzy set theory as the
means for quantifying the inherent fuzziness that is present in ill-posed
problems (which by many accounts are the majority of the real life problems
in decision making). Fuzziness is a type of imprecision which may be
associated with sets in which there is no sharp transition from membership to
nonmembership [Bellman and Zadeh, 1970]. Examples of fuzzy sets are
classes of objects (entities) characterized by such adjectives as large, small,
serious, simple, approximate, etc. [Bellman and Zadeh, 1970].
A comprehensive description of the importance of fuzzy set theory in
engineering and scientific problems is best illustrated in the more than 1,800
references given in [Chang, 1971], [Dubois and Prade, 1980], [Gupta, Ragade
and Yager, 1979], [Xie and Berdosian, 1983], [Zadeh, Fu, Tanaka, and
Shimura, 1975], and [Zadeh, 1976; 1978; and 1979].
Currently, an increasingly large number of researchers has been faced
with the problem that either their data or their background knowledge is
fuzzy. This is particularly critical to people which build expert systems and
decision support systems, for the knowledge they are dealing with is almost
always riddled with vague concepts and judgmental rules (e.g., [Lee, 1971],
[Lee, Grize, and Dehnad, 1987], [Prade and Negoita, 1986], [Ramsay, 1988],
and [Zadeh, 1983]. The most critical step in any application of fuzzy set
theory is to effectively estimate the pertinent data (i.e., the membership
values). Although this is a fundamental problem, there is not a unique way
of determining membership values in a fuzzy set. This is mainly due to the
way different researchers perceive this problem.
The above problems and challenges are particularly pervasive in many
MCDM problems. Very often in MCDM problems data are imprecise and
fuzzy. For instance, what is the value of the j-th alternative in terms of an
environmental impact criterion? A decision maker may encounter difficulty
in quantifying such linguistic statements so they can be used in deterministic
decision making. Related are the developments presented in Chapter 3.
Therefore, it is also desirable to develop and compare decision making
methods which use fuzzy data in making a decision. It is also crucial to
evaluate the performance of these fuzzy MCDM methods.

Lootsma in [1989; and 1997] presented a fuzzy version of Saaty's
AHP method. In his approach triangular fuzzy numbers were used with
pairwise comparisons in order to compute the weights of importance of the
decision criteria. Similarly, the fuzzy performance values of the alternatives
in terms of each decision criterion were computed by also using triangular
fuzzy numbers. The fuzzy operations which were used by Lootsma are also
applied in the next chapter to fuzzify four more decision-making methods.
These four methods were presented extensively in Chapter 2 and they are: the
weighted sum model (WSM), the weighted product model (WPM), the revised
analytic hierarchy process (RAHP) (as proposed by Belton and Gear [1983]),
and the TOPSIS method [Hwang and Yoon, 1981]. These methods are
fuzzified in Chapter 13 according to the fuzzy operations described in the
following section.
12.2 FUZZY OPERATIONS

Most of the decision making in the physical world takes place in a
situation in which the pertinent data and the sequences of possible actions are
not precisely known. Therefore, it is very important to adopt fuzzy data to
express such situations in decision making problems. In order to fuzzify the
previous four crisp decision making methods, it is important to know how
fuzzy operations are used on fuzzy numbers. Fuzzy operations were first
introduced by [Dubois and Prade, 1979; and 1980]. Other authors, such as
Laarhoven and Pedrycz [1983], Buckley [1985] and Boender, et al., [1989],
treated a fuzzy version of the AHP by using the fuzzy operations introduced
by Dubois and Prade.
When the decision maker considers the problem of ranking the m
alternatives AI' A2 , ••• , Am with respect to the n criteria C.. C2 , .•• , Cn' then,
the decision maker may feel great difficulty in assigning numbers, or ratios
of numbers, to alternatives in terms of these criteria. Thus the crisp MCDM
methods described and analyzed earlier may not be directly applicable in a
fuzzy environment. From now on fuzzy alternatives will be denoted as Ai
in order to distinguish them from their crisp version counterparts which are
denoted as Ai. Fuzzy alternatives are discrete alternatives for which the
pertinent information can only be expressed in terms of fuzzy numbers.
Similar statements also hold true between fuzzy decision criteria, denoted as
Ci , and their crisp counterparts denoted as Ci
The merit of using a fuzzy approach is to express the relative
importances of the alternatives and the criteria with fuzzy numbers instead of
using crisp numbers. For fuzzy numbers we will use triangular fuzzy
Chapter 12: Fuzzy Sets and their Operations 237
numbers (that is, fuzzy numbers with lower, modal, and upper values, see
also next definition) because they are simpler when compared to more
complex trapezoid fuzzy numbers. A fuzzy triangular number is formally
defined as follows:
DefInition 12-1 [Dubois and Prade, 1980]:

A fuzzy number M on R E (- 00, + 00 » is defined to be a fuzzy triangular
number if its membership function I-'m'" R -+ [0,1] is equal to."
1 I
--x - --, if x E [I, m]
m - I m - I
Ilm(X) = 1 I (12-1)
x - , if x E [m, u]
m-u m-u
0, otherwise.
In the above relation (12-1) I :::;; m :::;; u, and I and u stand for the
lower and upper value of the support of fuzzy number M, respectively, and
m for the modal value. A fuzzy triangular number, as expressed by equation
(12-1), will be denoted as (I, m, u).
Laarhoven and Pedrycz [1983], Buckley [1985] and Boender, et al.,
[1989] introduced fuzzy number operations in Saaty's AHP method by
substituting crisp numbers with triangular fuzzy numbers. The distinction of
Buckley's method from Boender's is that the fuzzy solution of a decision
making problem does not need to be approximated by fuzzy triangular
numbers. However, the triangular approximation of fuzzy operations is
plausible to fuzzify the implicit solution of a decision making problem and
provides fuzzy solutions with much smaller spread than Buckley's method
[Boender, et al., 1989]. Also, Boender, et al., proposed the use of a
geometric ratio scale as opposed to the original Saaty equi-distant scale in
quantifying the gradations of a human's comparative judgements. The basic
operations of fuzzy triangular numbers which were developed and used in
[Laarhoven and Pedrycz, 1983] are defined as follows:
nl EB n2 = (n ll +n2t, nlm +n2m, nlu +n2u ) for addition

nl ® n2 = (nU xn 2t' nlm xn2"" nlU xn2u ) for multiplication
e nl = (-n lu , -nlm' -nn) for negation
lInl - (lIn lu , lIn lm, linn) for division
Rn(nl) - (Rn(n u), Rn(nlJ, Rn(nlu » for natural logarithm
exp(nl) - (exp(nu),exp(nlm),exp(nlu» for exponential
where "=: " denotes approximation, and fl.] = (nu, n]m> n]u) and fl.2 = (n2f ,
n2m> n2u) represent two fuzzy triangular numbers with lower, modal, and
upper values. For the special case of raising a fuzzy triangular number to the
power of another fuzzy triangular number, the following approximation was
used:
- i!2 _ I n2f n2m n2u )
n] = \nlf ' n]m ,n]u
Please note that this formula was used only in the development of the fuzzy
WPM (as explained in the next chapter).
12.3 RANKING OF FUZZY NUMBERS
The problem of ranking fuzzy numbers appears very often in the

literature. For instance, a comparison and evaluation of different ranking
approaches is described in [McCahon and Lee, 1988] and [Zhu and Lee,
1991]. As each method of ranking fuzzy numbers has its advantages over the
others in certain situations, it is very difficult to determine which method is
the best one. Some important factors in deciding which ranking method is the
most appropriate one for a given situation, include the complexity of the
algorithm, its flexibility, accuracy, ease of interpretation and the shape of the
fuzzy numbers which are used.
A widely accepted method for comparing fuzzy numbers was first
introduced in [Baas and Kwakernaak, 1977]. In [Tong and Bonissone, 1981]
the concept of a dominance measure was introduced and it was proved to be
equivalent to Baas and Kwakernaak's ranking measure. This method was also
later adopted by Buckley [1985]. Later Tong and Bonnisone [1984] stated an
interesting observation. Namely, that most of the MCDM methods use the
degree of discrimination as a measure of quality and they wonder whether this
is a valid measure if one wishes to model human decision makers.
According to Zhu and Lee [1991] this ranking method is less complex
but still effective. It allows a decision-maker to implement it without
difficulty and with ease of interpretation. Therefore, in this book we use this
method for ranking fuzzy triangular numbers. However, if a given problem
requires for some reason a different method, then another method should be
used.
The above procedure for ranking triangular fuzzy numbers is used as
follows. Let I'.-;(x) denote the membership function for the fuzzy number fl. i .
Next, we define:
eij = max {min(Il/X), Il/Y))},

x~y (12-2)
for all i,j = 1,2,3, ... , m.
Chapter 12: Fuzzy Sets and their Operations 239
Then, it; dominates (or outranks) itj , written as it; > itj , if and only if eij = 1
and ej ; < Q, where Q is some fixed positive fraction less than 1. Values such
as 0.70, 0.80, or 0.90 might be appropriate for Q and the value of Q should
be set by the analyst and possibly be varied for a sensitivity analysis. In the
computational experiments reported in the next chapter the value of Q was set
equal to 0.90. The above concepts are best explained in the following
Example 12-1:
Suppose that the importances of two fuzzy alternatives A] and A2 are
represented by the two fuzzy triangular numbers it] = (0.20, 0.40, 0.60) and
it2 = (0.40,0.70,0.90), respectively. Next, it can be observed from Figure
12-1 that e2] = 1 and el2 = 0.40 < Q = 0.90. Therefore, A] > A2 and thus
the best (fuzzy) alternative is A]. •
Applications of using fuzzy triangular numbers in Saaty's AHP

method can be found in [Laarhoven, et ai., 1983] and [Lootsma, 1989; and
1997]. One of the goals here is to fuzzify the crisp and deterministic MCDM
methods which were described in the second chapter in a manner similar to
the one used in [Lootsma, 1989]. Thus, the previous concepts and definitions
form the foundation for the analyses presented in the next chapter.
Figure 12-1: Membership Functions for the Two Fuzzy Alternatives Al and A2•
Chapter 13
FUZZY MULTI-CRITERIA DECISION MAKING
In this chapter four deterministic MCDM methods of the ones

presented in the second chapter are fuzzified. These are the WSM, the
WPM, the AHP (original and ideal mode), and the TOPSIS method. The
ELECTRE is not examined (since the TOPSIS method seems to be superior
to it). The multiplicative AHP (as described in Section 11.4) is not studied
either since it has the same numerical properties as the WPM model. Some
crisp MCDM methods have been fuzzified in [Buckley, 1985], [Laarhoven,
et al., 1983], and [Lootsma, 1989; and 1997]. The methodology which we
are going to explore is different from the methods introduced in the previous
three references. In this chapter we will fuzzify the crisp MCDM methods
under the assumption that only one decision maker is involved. The
developments presented in this chapter are based on the work reported in
[Triantaphyllou and Lin, 1996].
In this chapter two fuzzy evaluative criteria are used to examine the
performance of the fuzzy MCDM methods in an attempt to find the best such
method. These two fuzzy evaluative criteria are analogous to the ones
presented in Section 9.2 for the crisp MCDM methods. Please recall that the
first evaluative criterion for the crisp case was to see if a method, which is
supposed to be accurate in multi-dimensional problems, is also accurate in
single dimensional problems. Hence, the first fuzzy evaluative criterion uses
the results of the fuzzy WSM as the standard in evaluating the rest of the
fuzzy MCDM methods.
Similarly, the second fuzzy evaluative criterion considers the premise
that an accurate fuzzy MCDM method should not change the indication of the
best alternative when a nonoptimal alternative is replaced by a worse
alternative (given that the importance of each decision criterion remains
unchanged). As with the crisp case, this is a reasonable assumption because
intuitively the truly best alternative should remain the same no matter which
nonoptimal alternative is replaced by a worse alternative. More on these two
fuzzy evaluative criteria is described in Section 13.6.
In this chapter two main research objectives are pursued. The first
objective is to fuzzify the previous four existing crisp MCDM methods in a
manner similar to one used by Lootsma in fuzzifing the AHP method in
[Lootsma, 1989; and 1997]. These fuzzified MCDM methods are then
evaluated in terms of the previous two evaluative criteria. This is done by

using the Saaty scale, and the Lootsma scale in order to study the impact of
different scales in the final results. Determining the best fuzzy MCDM
method is the second major objective of this chapter.
In the context of this chapter the values which the decision maker
assigns to the alternatives in terms of the decision criteria in a fuzzy MCDM
problem are assumed to be fuzzy triangular numbers instead of crisp
numbers. ThuS,-'these fuzzy triangular-numbers will be denoted as fl = (I, m,
u), where I !5; m !5; u, and I, m, U are again the lower, modal and upper
value of fl, respectively. The parameter m (modal value) should not be
confused here with the number of alternatives which was also denoted as m
earlier in this book.
In the following sections the procedures applied by Laarhoven, et al. ,
[1983] and Boender, et al., [1989] will be used on the crisp MCDM methods
which have been introduced in the second chapter. As usually, some
numerical examples will also be analyzed in order to achieve a better
illustration of the key concepts in this chapter.
13.2 THE FUZZY WSM METHOD

As it was stated in the second chapter according to the crisp WSM
method the best alternative is the one which satisfies equation (2-1). Now,
the performance value of the i-th alternative in terms of the j-th criterion is
a fuzzy triangular number denoted as: flij = (aijl' aijm, aiju). Analogously, it
is assumed that the decision maker will use fuzzy triangular numbers in order
to express the weights of importance of the criteria. Thus, these weights will
be denoted as: Wj = (wjl , Wjm , wju ). Also, in order to be consistent with the
basic requirement that the weights should add up to one (in a crisp
environment), now it is required that the sum of the wjm values (i.e., the
modal values of the fuzzy triangular numbers which represent the criteria
weights) should be equal to 1. Similarly with the discussion in Section 9.2
and the above considerations, it follows that under the fuzzy WSM (denoted
as F-WSM) the best alternative is the one which satisfies the following
relation (where m is the number of alternatives):
for i = 1, 2, 3, ..., m. (13-1)
Example 13-1:
Let a decision problem be defined on the four fuzzy decision criteria
CI, C2, C3, C4, and the three fuzzy alternatives AI' A2, and A3 . Next, let us
Chapter 13: Fuzzy Multi-Criteria Decision Making 243
assume that the data for this problem are given in the following fuzzy
decision matrix.
Criteria
C1 C1 C3 C 4
AIts. (0.13,0.20,0.31) (0.08,0.15,0.25) (0.29,0.40,0.56) (0.17,0.25,0.38)
Al (3.00,4.00,5.00) (5.00,6.00,7.00) (5.00,6.00,7.00) (2.00,3.00,4.00)

A2 (6.00,7.00,8.00) (5.00,6.00,7.00) (0.50, 1.00,2.00) (4.00,5.00,6.00)
AJ (4.00,5.00,6.00) (3.00,4.00,5.00) (7.00,8.00,9.00) (6.00,7.00,8.00)
Therefore, when the fuzzy WSM (i.e., the FWSM) approach is used, then the
fInal priority scores (denoted as PI P2 • and P3 • respectively) of the
three alternatives are: •
PI = (0.13, 0.20, 0.31) x (3.00,4.00, 5.00) +
(0.08,0.15,0.25) x (5.00,6.00, 7.00) +
(0.29, 0.40, 0.56) x (5.00, 6.00, 7.00) +
(0.17,0.25,0.38) x (2.00,3.00,4.00) =
=
(2.583, 4.850, 8.750).
In a similar manner, the following results are also derived:
P2 = (1.979, 3.950, 7.625),
and P3 = (3.792,6.550, 11.188).
Q • 8.98
x
Figure 13-1: Membership Preference Functions of the Fuzzy Alternatives AI' A2• and AJ
of Example 13-1 According to the Fuzzy WSM Method.
Figure 13-1 displays the membership functions of these final results.

They could be interpreted as a measure of the ability of each alternative to
meet the decision criteria. From this figure and the definition of the eij
coefficient as given by relation (12-2) (in Section 12.3) it is clear that: eJI =
eJ2 = e]2 = 1, and eIJ • e2J , and e2I are less than Q (= 0.90). Thus, according
to the ranking procedure which was discussed in Section 12.3, fuzzy
alternative AJ is the most preferred one. •
13.3 THE FUZZY WPM METHOD
The best alternative with the crisp version of this model is the one
which satisfies equation (2-2). For the fuzzy version of this model the
corresponding formula becomes:
R[~Kl
AL
= iI (~Kj)-WJ.
J=l alJ
(13-2)
where o'Kj' o'lj and Wj are Afuzzy triangular numbers. Si~ilarly to the crisp
setting, fuzzy alternative Ax dominates fuzzy alternative AL if and only if the
value of the numerator in equation (13-2) is greater than its denominator.
The application of formula (13-2) is also illustrated in the following example:
x
Figure 13-2: Membership Functions of the Fuzzy Alternatives A/. A2 • and AJ of Example
13-2 According to the Fuzzy WPM Method.
Example 13-2:
The data used in Example 13-1 are also used here. Therefore, when
relation (13-2) is applied, then the following ratios are obtained:
[ (3.00,4.00,5.00)(·13,.20,.31) X (5.00,6.00,7.00)(·08,.15,.25) X
X (5.00,6.00,7.00)(·29,.40,.56) X (2.00,3.00,4.00)(·17,.25,.38) ] /
[ (6.00,7.00,8.00)(·13,.20,.31) x (5.00,6.00,7.00)(·08,.15,.25) X
X (0.50,1.00,2.00)(·29,.40,.56) X (4.00,5.00,6.00)(·17,.25,.38)] _
(2.355, 4.652, 13.516) / (1.473, 2.887, 9.008).
In the previous calculations the approximation symbol " == " was used,
because the corresponding formula in Section 12.2 uses an approximation as
well. Next, in a similar manner, we get:
R(A/AJ - (2.355, 4.652, 13.516) / (3.099, 6.348, 19.649),

and R(A/AJ - (1.473, 2.887, 9.008) / (3.099, 6.348, 19.649).
Furthermore, according to formula (12-2) the pertinent eij coefficients

are as follows: e3l = e32 = en = 1, and e13, e23, e2l are less than Q (= 0.90).
Obviously, fuzzy alternative A3 dominates all the other alternatives. The
priority values (expressed as fuzzy numbers) are shown in Figure 13-3. It can
be noticed that the best alternative according to the fuzzy WSM and the fuzzy
WPM approaches is identical (in this numerical example) although the fuzzy
WPM requires more complicated operations to reach the final results. •
13.4 THE FUZZY AHP METHOD

A fuzzy variant of the AHP was presented in [Laarhoven and
Pedrycz, 1983] and in [Boender, et al., 1989]. That approach is next
illustrated in the following numerical example by using fuzzy triangular
numbers.
Example 13-3:
Consider a decision problem with four fuzzy decision criteria and
three fuzzy alternatives. Now pairwise comparisons are represented as ratios
of fuzzy triangular numbers. Next suppose that when the decision maker is
asked to compare the three fuzzy alternatives in terms of the first fuzzy
criterion by using fuzzy pairwise comparisons, then the following reciprocal
judgment matrix was derived:
Comparisons Under the First Fuzzy Criterion: (;]

..41 ..42 ..43
..41 (1, 1, 1) (116, 112, 2) (1110, 114, 2/3)
..42 (112, 2, 6) (1, 1, 1) (118, 113, 1)
..43 (3/2, 4, 10) (1, 3, 8) (1, 1, 1)
It should be noticed here that when an alternative is compared with

itself, then the fuzzy triangular number (1, 1, 1) is used instead of the crisp
number 1. Next, the fuzzy eigenvector of the above matrix is estimated.
Recall that given a crisp reciprocal matrix, then according to Saaty the right
principal eigenvector of the matrix expresses the importances of the
alternatives. When we combine Saaty's eigenvector approximation method
(see also Section 4.2) with the fuzzy arithmetic operations described in the
previous chapter, we get the following importances of the three alternatives
(expressed as fuzzy triangular numbers):
= [(1, 1, 1) x (116, 112, 2) x (1110, 114, 2/3)]1/3 _
(0.25, 0.50, 1.10),
= [(112, 2, 6) x (1, 1, 1) x (118, 113, 1)]1/3 _

(0.40, 0.87, 1.82),
and = [(3/2, 4, 10) x (1, 3, 8) x (1, 1, 1)]1/3 _

_ (1.14, 2.29, 4.31).
As before, in the previous calculations the approximation symbol " == " was
used, because the corresponding formula in Section 12.2 uses an
approximation as well.
Next, the previous vector is normalized according to the similar
requirement in the original crisp AHP. The normalized vector is derived by
dividing each entry by the sum of the entries in the vector. It can be easily
verified that the normalized priority vector is as follows:
(0.02, 0.14, 0.99)
(0.06, 0.24, 1.02)
(0.16, 0.62, 2.41)
At this point let us assume that the approximation of the fuzzy

eigenvectors of the pairwise comparisons when the three alternatives are
compared in terms of each one in terms of the remaining decision criteria are
also derived in a similar manner. The same is also done for the weights of
importance of the four fuzzy criteria. These data now are used to form the
vectors in the following fuzzy decision matrix:
Criteria
C1 C2 C 3 C4
Alts. (0.08,0.18,0.46) (0.08,0.16,0.39) (0.17,0.40,0.86) (0.11,0.26,0.61)
~
Al (0.02,0.14,0.99) (0.18, 0.44,0.95) (0.22, 0.37, 0.64) (0.12, 0.23, 0.55)

~2 (0.06, 0.24, 1.02) (0.14, 0.35,0.83) (0.07, 0.10, 0.15) (0.13,0.30,0.69)
A3 (0.16,0.62,2.41) (0.11,0.21,0.53) (0.30,0.53, 0.91) (0.19,0.47, 1.00)
Then, according to equation (13-1) the final fuzzy priority scores (denoted as
PI' P2 • and P3 ) of the three fuzzy alternatives are as follows:
(0.02, 0.14, 0.99) x (0.08, 0.18, 0.46) +
(0.18, 0.44, 0.95) x (0.08, 0.16, 0.39) +
(0.22,0.37,0.64) x (0.17,0.40,0.86) +
(0.12, 0.23, 0.55) x (0.11, 0.26, 0.61)
= (0.068, 0.474, 1.887).
Similarly, the following two fuzzy preference scores are derived:
= (0.772, 0.208, 1.908),
and (0.897, 0.480, 2.696).
When the procedure described in Section 12.3 for ranking triangular fuzzy
numbers is applied to PI' P2 • and P3 • it can be easily shown that fuzzy
alternative A3 is the best one. •
13.5 THE FUZZY REVISED ANALYTIC HIERARCHY

PROCESS
As mentioned in Section 2.4., the revised version of the AHP (Le.,

the "ideal mode" AHP [Saaty, 1994]) as proposed by Belton and Gear [1983]
is to normalize the relative performance measures of the alternatives in terms
of each criterion by dividing the values with the largest one. This is the only
difference from the original AHP method. The fuzzy version of the revised
AHP is best illustrated in the following example which uses the same
numerical data as the last example.
Example 13-4:
In this example the vectors in the fuzzy decision matrix of Example
13-1 are divided by the largest entry in that vector. In this way the following
fuzzy decision matrix is derived:
Criteria
C1 C2 C3 C4
A1ts. (0.08,0.18,0.46) (0.08,0.16,0.39) (0.17,0.40,0.86) (0.11,0.26,0.61)
Al (0.01,0.21,9.90) (0.44, 1.00,2.29) (0.41,0.69, 1.26) (0.26,0.50, 1.26)

A2 (0.01,0.38, 1.14) (0.35,0.79,2.00) (0.13,0.19,0.30) (0.30,0.63, 1.59)
A3 (0.07,1.00,15.1) (0.26,0.50,1.26) (0.56,1.00,1.78) (0.44,1.00,2.29)
In a manner similar to the one used in the original AHP, the final fuzzy
preference scores of the three fuzzy alternatives are calculated as follows:
= (0.08, 0.18, 0.46) x (0.01, 0.21, 9.90) +
(0.08, 0.16, 0.39) x (0.44, 1.00, 2.29) +
(0.17,0.40,0.86) x (0.41,0.69, 1.26) +
(0.11, 0.26, 0.61) x (0.26, 0.50, 1.26) =
= (0.130, 0.605, 2.923).
Similarly, the following two fuzzy preference scores are derived:

= (0.081, 0.422, 2.525),
and = (0.167, 0.932, 11.45).
When the ranking procedure described in Section 12.3 is applied to

the Pl' P2 • and Pl' values, it can be easily shown that fuzzy alternative
13 is the best one. •
13.6 THE FUZZY TOPSIS METHOD

The fuzzy version of the TOPSIS method is best illustrated in the
following numerical example.
Example 13-5:
In this example we follow the fuzzy versions of the steps for the crisp
TOPSIS method as described in Section 2.2.6.
Step 1. Construction of the Fuzzy Nonnalized Decision Matrix

Suppose that when a decision problem with four criteria and three alternatives
is considered, then the following fuzzy decision matrix was derived:
Criteria
C1 C2 C3 C4
AIts. (0.13,0.20,0.31) (0.08,0.15,0.25) (0.29, 0.40, 0.56) (0.17,0.25,0.38)
"
Al (0.08, 0.25, 0.94) (0.25, 0.93, 2.96) (0.34, 0.70, 1.71) (0.12, 0.24, 0.92)
..4.2 (0.23, 1.00, 3.10) (0.13,0.60,2.24) (0.03, 0.05, 0.09) (0.12, 0.40, 1.48)
..4.3 (0.15, 0.40, 1.48) (0.13, 0.20, 0.88) (0.62, 1.48, 3.41) (0.24, 1.00,3.03)
Step 2. Construction of the Fuzzy Weighted Normalized Decision

Matrix
Given the previous fuzzy normalized decision matrix, the corresponding fuzzy
weighted normalized matrix is:
Criteria
Alts. C1 C2 C3 C4
"
Al (0.01,0.05,0.29) (0.02,0.14,0.74) (0.10,0.28,0.96) (0.02 ,0.06 ,0.35)
..4.2 (0.03, 0.20, 0.96) (0.01,0.09,0.56) (0.01,0.02,0.05) (0.02, 0.10, 0.55)
..4.3 (0.02, 0.08, 0.46) (0.01,0,03,0.22) (0.18, 0.59, 1.91) (0.04, 0.25, 1.15)
Step 3. Determine the Fuzzy Ideal and Fuzzy Negative-Ideal Solutions

According to the fuzzy version of equations (2-12) and (2-13) in Section
2.2.6, the fuzzy ideal solutionj* and the fuzzy negative-ideal solutionj- are
as follows:
j* = {(0.03,O.20,O.96), (0.02,0.14,0.74), (0.18,0.59,1.91), (0.04,0.25, 1.15)}.
and j- = {(0.01,O.05,O.29), (0.01,0.02,0.22), (0.01,0.02,0.05), (0.02,0.06,O.35)}.
Step 4. Calculation of the Fuzzy Separation Measure

When the fuzzy version of equations (2-14) and (2-15) (in Section 2.2.6) are
used, the following fuzzy separation distances between each fuzzy alternative
and the fuzzy ideal/negative-ideal solutions are derived:
8/* - (0.09, 0.39, 1.41), 8/_ - (0.09, 0.28, 1.04),
82* - (0.17,0.59, 1.95), 82_ - (0.02,0.16,0.76),
83* - (0.02, 0.16, 0.71), 83_ - (0.17, 0.60, 2.03).
For instance, 8/* = { [(0.01,0.05,0.29) - (0.03,0.20,0.96)]2 +
[(0.02,0.14,0.74) - (0.02,0.14,0.74)]2 +
[(0.02,0.06,0.35) - (0.04,0.25,1.15)]2 +
[(0.10,0.28,0.96) - (0.18,0.59,1.91)]2}'A ==
(0.09, 0.39, 1.41).
As before, the. approximation symbol " == " was used instead of equality.
Step 5. Calculation of the Relative Closeness to the Fuzzy Ideal Solution

The relative closeness to ideal solution of the three alternatives is
defined by the fuzzy version of equation (2-16) as follows:
eJ * = Si_ / (SJ* + SJ_) ==

- (0.09, 0.39, 1.41) / [ (0.09, 0.39, 1.41) + (0.09, 0.28, 1.04)] =
= (0.04, 0.42, 5.83).
Similarly, the following values are derived:

e 2* - (0.01, 0.21, 3.99),
and e 3* == (0.06, 0.79, 10.42).
When the ranking procedure described in Section 12.3 is applied here, it can
be easily shown that the previous closeness measures are ranked as follows:
e 3* > e J* > e 2*. Therefore, the preference order of the three fuzzy
alternatives is: 13 > 1J > 1 2 , That is, the best alternative is 1 3, •
13.7 TWO FUZZY EVALUATIVE CRITERIA FOR FUZZY

MCDM METHODS
The previous four fuzzy MCDM methods can be used in fuzzy
MCDM problems which involve a single decision maker. However, these
methods may derive different answers for the same problem. Since the best
alternative should be the same no matter which method is used, an
examination of the accuracy and consistency of these methods is highly
desirable. Thus, similarly to the evaluative procedures described in Chapter
9, two fuzzy evaluative criteria are introduced in this section to examine the
performance of these fuzzy MCDM methods.
As it was stated in Chapter 9, in a crisp and single-dimensional
environment the WSM yields the most reasonable results. Therefore, in a
crisp single-dimensional problem one may want to compare the results of the
WPM, AHP, and TOPSIS with the results derived by using the WSM. In a
fuzzy setting, however, one may want to apply the above evaluative criterion
by comparing the result of the fuzzy WSM with those obtained by applying
the fuzzy versions of the WPM, the AHP, and the TOPSIS methods. This is
the first fuzzy evaluative criterion. In comparing the ranking derived by
using the fuzzy WSM and anyone of the other methods, two contradiction
rates can be determined. The first one is to record the rate the best
alternative is the same by both methods. The second contradiction rate is to
record the times the two rankings are different in terms of any (i.e., not just
the best) alternative.

The second evaluative criterion in Chapter 9 examined the stability of
the results derived by a method when a nonoptimal alternative is replaced by
a worse one. A perfectly accurate method should rank some alternative as the
best, even after a nonoptimal alternative is replaced by a worse alternative
(and assuming that the rest of the data remain the same). Thus, the second
fuzzy evaluative criterion considers the premise that a fuzzy MCDM method
should not change its indication of the best alternative when a nonoptimal
alternative is replaced by a worse one.
Furthermore, the concepts of the RCP (Real and Continuous
Pairwise) and the CDP (Closest and Discrete Pairwise) matrices will be
fuzzified here accordingly in order to carry out an empirical study of these
fuzzy methods. These matrices are used in this book to emulate the pairwise
comparison of a decision maker under the assumption that he/she is as much
accurate as possible. The notion of these matrices, along with their fuzzy
extensions, are discussed in the following examples which demonstrate the
application of the previous two fuzzy evaluative criteria.
13.7.1. Testing the Methods by Using The First Fuzzy

Evaluative Criterion
The following example illustrates the procedure of testing the fuzzy
decision-making methods by using the first fuzzy evaluative criterion.
Example 13-6:
Let a decision problem involve four decision criteria and three
alternatives. Let all the data be expressed in terms of the same unit of
measurement (e.g. dollars, or time, or weight, etc.). Suppose that the
following decision matrix depicts the actual (and thus unknown to the
decision maker) crisp data for this problem.
Criteria
CI C2 C3 C4
Alts. (6.015 5.526 8.349) 5.721
Al 4.454 3.253 3.987 5.816

A2 3.647 8.450 1.447 2.189
A3 6.232 7.273 2.496 4.756
The decision maker is assumed not to know these data. Next, he/she
is asked to use triangular fuzzy numbers in which the lower, modal, and
upper values are members of the set: {9, 8, 7, ... 2,1,112, ... 117, 118, 1I9}.
These are the values recommended by the Saaty scale [1994] when one wishes
to quantify pairwise comparisons (see also Chapter 3).
It is also assumed here that the decision maker is as accurate as
possible. For instance, when the decision maker attempts to estimate the
performance of the first alternative in terms of the first criterion, the fuzzy
number (3, 4, 5) is used. It can be observed now that the modal value of this
number is the closest number from the above set of numbers to the actual
value of 4.454. Also, the lower and upper values of that fuzzy number are
assumed to be one unit apart. In a similar manner the rest of the entries are
estimated and thus the following fuzzy decision matrix is assumed to had been
obtained by the decision maker.
Criteria
(;1 (;2 (;3 (;4
Alts. (5, 6, 7) (5, 6, 7) (7, 8, 9) (5, 6, 7)
'" (3, 4, 5)
Al (2, 3,4) (3, 4, 5) (5, 6, 7)
12 (3, 4, 5) (7, 8, 9)
(6, 7, 8)
(0.5, 1, 2)
(1, 2, 3)
(1, 2, 3)
(4, 5, 6)
A3 (5, 6, 7)
Therefore, the final fuzzy priority scores of the fuzzy alternatives are
calculated as follows:
= (5,6,7) x (3,4,5) + (5,6,7) x (2,3,4) +
(7,8,9) x (3,4,5) + (5,6,7) x (5,6,7)
= (71, 105, 157).
Similarly, the following two fuzzy priority values are derived:

(58.5, 92, 135),
and = (82, 124, 174).

When the ranking procedure which was discussed in Section 12.4 is used
here, then it can be easily determined that the best fuzzy alternative is A3 and
the three alternatives are ranked as follows: A3 > A} > A2 •
In the following paragraphs it is assumed that the decision maker is
using the fuzzy AHP method. If the decision maker knew the actual data
shown in the original crisp decision matrix, then the matrix of the actual
pairwise comparisons (say matrix ex) when the four decision criteria are
compared with each other would be as follows (Le., this is the corresponding
RCP matrix):
6.015 1 1.09 0.72 1.05

5.526 0.91 1 0.66 0.97
8.349
= 1.39 1.52 1 1.46
a.
5.721 0.95 1.03 0.68 1
For instance, the entry a 12 = 1.09 = 6.015/5.526. However, in

reality the decision maker will never know the actual values of the pairwise
comparisons. At this point, it is assumed that the decision maker is as
accurate as possible and, when he/she is asked to use fuzzy triangular
numbers, then he/she is able to determine the following matrix (say matrix (3)
with the closest fuzzy pairwise comparisons (this is the corresponding fuzzy
CDP matrix):
(1, 1, 1) (0.50, 1, 2) (0.30, 0.50, 1) (0.50, 1, 2)

(0.50, 1, 2) (1, 1, 1) (0.30, 0.50, 1) (0.50, 1, 2)
~
(0.50, 1, 2) (1, 2, 3) (1, 1, 1) (0.50, 1, 2)
(0.50, 1, 2) (0.50, 1, 2) (0.30, 0.50, 1) (1, 1, 1)
For instance, the value {332 = (1, 2, 3) is derived from a32 (= 1.52)
in which 1.52 has the absolute minimal difference with the value 2 from
Saaty's original scale. Therefore, the modal value of {332 is set to be equal to
2 and the upper and lower values of {332 are equal to 1 and 3, respectively,
which are one unit apart from the modal value. Similarly, the fuzzy pairwise
comparisons (i.e., the corresponding fuzzy CDP matrices) of the three
alternatives in terms of each decision criterion are derived in accordance with
the Saaty scale and are as follows:
- -
4.454 1.22 0.71 (1,1,1) (0.50,1,2) (0.30,0.50,1)
3.647 0.82 1 0.59 (0.50,1,2) (1,1,1) (0.30,0.50,1)
6.232 1.41 1.69 1 (1,2,3) (2,3,4) (1,1,1)
- -
3.253 0.38 0.45 (1,1,1) (0.25,0.30,0.50) (0.30,0.50, 1)
8.450 2.63 1 1.16 (2,3,4) (1,1,1) (0.50,1,2)
7.273 2.22 0.86 1 (1,2,3) (0.50,1,2) (1,1,1)
- -
3.987 1 2.76 (1,1,1) (2,3,4) (1,2,3)
1.447 0.36 1
!.OO] (0.25,0.30,0.50) (1,1,1) (0.30,0.50,1)
0.:8
2.496 0.63 1.72 (0.30,0.50,1) (1,2,3) (1,1,1)
5.816 1 2.66 1.23 (1,1,1) (2,3,4) (0.50,1,2)

2.189 - 0.38 1 0.46 - (0.25,0.30,0.50) (1,1,1) (0.30,0.50,1)
4.736 0.81 2.17 1 (0.50,1,2) (1,2,3) (1,1,1)
Next, the approximation eigenvector approach and normalization

procedure are applied as before on the above fuzzy reciprocal matrices in
order to derive the relative preference values of the alternatives in terms of
the criteria along with the weights of importance of the four decision criteria.
Therefore, the derived fuzzy decision matrix (which we assume the decision
maker has estimated) is as follows:
Criteria
~ ~ ~ ~
A1ts. (0.09,0.22,0.69) (0.08,0.23,0.61) (0.11,0.32,0.73) (0.08,0.23,0.61)
Al (0.13,0.29,0.67) (0.10,0.17,0.35) (0.29,0.55,0.98) (0.23,0.46,0.91)

..42 (0.10,0.22,0.52) (0.24,0.47,0.90) (0.09,0.16,0.34) (0.09,0.16,0.36)
..43 (0.24,0.49,0.84) (0.16,0.36,0.66) (0.14,0.29,0.60) (0.16,0.38,0.81)
The final fuzzy priority scores PI , p2' and p3' of the three fuzzy
alternatives are derived in a similar manner as stated earlier and are as
follows:
= (0.09,0.26,0.72) x (0.12,0.28,0.66) +
(0.08, 0.22, 0.60) x (0.09, 0.16, 0.34) +
(0.11, 0.30, 0.72) x (0.28, 0.54, 0.97) +
(0.08, 0.22, 0.60) x (0.22, 0.45, 0.90) =
= (0.070, 0.384, 1.946).
Similarly, the following two fuzzy priorities are derived:
= (0.045, 0.245, 1.376),
and = (0.063, 0.371, 1.914).
When these priority scores are ranked as before, then the three fuzzy
alternatives are ranked as follows: "C
> A3 > A2. That is, fuzzy alternative
A1 turns out now to be the best one. Obviously, this is in contradiction with
the results derived when the fuzzy WSM was applied at the beginning of this
At the same time, it can also be observed that the entire ranking of
the fuzzy alternatives as derived by the fuzzy WSM and the fuzzy AHP has
also changed (that is, from: A3 > A1 > A2 it has changed to: A1 > A3 >
A2). Therefore, a contradiction occurs between the fuzzy WSM and the fuzzy
AHP when one compares the entire ranking orders of the three alternatives
for this illustrative example. •
It is possible that the best alternative derived from the fuzzy WSM
and the other fuzzy methods be identical but the remaining alternatives change
their orders. The revised (i.e., ideal mode) AHP, the WPM and the TOPSIS
methods can be examined as above and it can be similarly demonstrated that
they also yield contradictions when a single-dimensional environment is
assumed and the fuzzy WSM is used as the norm.
13.7.2 Testing the Methods by Using the Second Fuzzy

Evaluative Criterion
Similarly to the developments described in the previous section, the
use of the second fuzzy evaluative criterion is best illustrated with an
Example 13-7:
As with the previous numerical example suppose that a decision
problem with four criteria and three alternatives has the following decision
matrix. Again, it is assumed that these values are unknown to the decision
maker.
Criteria
C1 C2 C3 C4
Alts. (2.885 3.987 2.434 4.894)
Al 8.283 7.851 1.064 4.554

A2 5.284 8.626 4.161 3.750
A3 1.358 2.905 8.501 4.888
When one generates the fuzzy triangular numbers as before and

proceeds as in the previous example, it can be easily shown that the
corresponding fuzzy decision matrix (which we assume the decision maker
can derive for this problem) is as follows:
Criteria
C1 C2 C3 C4
Alts. (O.OS, O.lS, 0.47) (0.11,0.30,0.74) (O.OS, O.lS, 0.47) (0.13,O.35,O.S2)
41
~
(0.35,0.59,0.91) (0.22, 0.43, 0.S3) (0.06, O.OS, 0.11) (0.13,O.33,O.S4)

A2 (0.21,0.32,0.57) (0.22, 0.43, 0.S3) (0.20,0.31,0.54) (0.13,O.33,O.S4)
13 (0.06,0.09,0.13) (0.09, 0.14, 0.26) (0.3S, 0.62, 0.94) (0.13,0.33,0.84)
Also, the final priority scores PI' P2 • and P3 • of the three alternatives
can be shown to be as follows:
PI = (0.073, 0.361, 1.889),
P2 = (0.073, 0.355, 1.824),
and P3 = (0.061, 0.2~4, 1.388).
Apparently, fuzzy alternative A 1 is the best one.
Next, the crisp alternative A3 (which is not the best one) in the
original matrix is replaced by A/ which is worse than the original A3 • The
performance values of A/ are the same as in the original alternative A3 except
that the third value in terms of criterion C3 of 8.501 is substituted by the least
one 1.064. Thus, the original matrix of crisp numbers is modified as follows:
Criteria
CJ C2 C3 C4
Alts. (2.885 3.987 2.434 4.894)
AJ 8.283 7.851 1.064 4.554

A2 5.284 8.626 4.161 3.750
A/ 1.358 2.905 1.064 4.888
In order to test the stability of the fuzzy AHP method, the same
procedure as above is repeated in order to determine the best alternative. The
new fuzzy decision matrix can be shown to be as follows:
Criteria
C J C2 C3 C4
Alts. (0.08, 0.18, 0.47) (0.11,0.30,0.74) (0.08,0.18,0.47) (0.13, 0.35, 0.82)
Al (0.35, 0.59, 0.91) (0.22, 0.43, 0.83) (0.10, 0.17, 0.29) (0.13, 0.33, 0.84)
A2
A /
(0.21,0.32,0.57) (0.22, 0.43, 0.83) (0.45,0.67,0.97) (0.13, 0.33, 0.84)
A3 (0.06,0.09,0.13) (0.09,0.14,0.26) (0.10,0.17,0.29) (0.13, 0.33, 0.84)
Working as before, it can be seen that the final fuzzy priority scores now
become:
PI = (0.077, 0.377, 1.872),
P2 = (0.092, 0.418, 2.030),
and P!.3 = (0.040, 0.205, 1.082).
From the above fuzzy scores it is obvious that now the best fuzzy
alternative is A2 . This result is in contradiction with the earlier result, namely
that the best fuzzy alternative is AI. This analysis indicates that a
contradiction may occur when the fuzzy AHP (original version) is used and
a nonoptimal alternative is replaced by a worse one. In a similar manner it
can be shown that the ideal mode (revised) AHP, WPM and TOPSIS methods
may also fail when they are tested in a similar manner. •
13.8 COMPUTATIONAL EXPERIMENTS
The previous fuzzy MCDM methods were evaluated by generating test

problems which were treated as the problems in Examples 13-6 and 13-7 and
then recording the contradiction rates. These experiments were conducted in
an attempt to examine the performance of the previously mentioned fuzzy
MCDM methods in terms of the two fuzzy evaluative criteria [Triantaphyllou
and Lin, 1996]. Therefore, the three contradiction rates which were
considered in these computational experiments are summarized as follows:
Rll: Is the rate that the fuzzy WSM and another fuzzy method
disagree in the indication of the best alternative.
R12: Is the rate that the fuzzy WSM and another fuzzy method
disagree on the entire ranking of the alternatives.
R21: Is the rate that a method changes the indication of the best
alternative when a nonoptimal alternative is replaced by a
worse alternative.
The computer program for this empirical study was written in

FORTRAN and run on an IBM 3090 mainframe computer. A total number
of 400 (i.e., 10 alternatives x 10 criteria x 4 methods) cases were examined
with number of 3,5,7 ... , 19,21 alternatives and number of 3,5,7, ... , 19,
21 criteria. Three kinds of contradiction rates were recorded for each case
by running each case (according to the previous three definitions) with 500
random replications per case.
As it was stated in the previous sections, in this empirical study it was
assumed that the decision maker did not know the actual values of the
alternatives in terms of the decision criteria or the weights of importance of
the decision criteria. For the purpose of these simulations, the original
importance measures of the alternatives in terms of the decision criteria were
generated randomly within the interval [9, 1] (which is the interval of the
values according to the original Saaty scale). The fuzzy AHP (original and
revised), the WPM, and the TOPSIS methods were then examined in terms
of the two fuzzy evaluative criteria in a manner similar to the procedures
described in Examples 13-6 and 13-7. The computational results are depicted
in Figures 13-4 to 13-8 and are also discussed in the following two sub-
sections.
13.8.1 Description of the Computational Results

In Figures 13-4 to 13-8 the fuzzy AHP, the fuzzy revised AHP, the
fuzzy WPM, and the fuzzy TOPSIS methods are denoted as F-AHP,
F-RAHP, F-WPM, and F-TOPSIS, respectively. The computational results
suggest that the contradiction rate Rll increases as the number of alternatives
increases. For a more transparent illustration, Figures 13-4, and 13-5 depict
the results only for the cases when the number of alternatives is equal to 3
and 21, respectively. These results reveal that the rate the indication of the
best alternative changes increases as the number of alternatives increases, no
matter which method is used. However, the number of criteria does not seem
to be important here.
For the contradiction rate R21, which indicates changes of the best
alternative when a nonoptimal alternative is replaced by a worse one, it can
be seen from the plots in Figures 13-5 and 13-6 that the fuzzy revised (i.e.,
ideal mode) AHP is slightly better than the other methods.
Similarly, the contradiction rate R12, which indicates changes in the
entire ranking of alternatives between the fuzzy WSM and other methods,
shows a similar behavior. In this case only the results for number of
alternatives equal to 3 are shown (in Figure 13-7). This is because the results
for rate R12 when the number of fuzzy alternatives is 21 were almost always
equal to 100 %.
Despite the increasing inaccuracy of these fuzzy methods when the
decision-making problems become more complex, it can be seen from these
graphs that the fuzzy revised AHP is better than any other fuzzy MCDM
methods in most cases. This result turns out to be in agreement with the
results originally reported in [Triantaphyllou and Mann, 1989] (which
examined the same problems in a crisp setting) and are also described in
Chapter 9. However, now the fuzzy WPM becomes the worst method in
terms of the contradiction rate R21 (however, the WPM was the second
best method in the crisp setting).
.35
.3
.25
20
0::
c .2
0
~
'0
.1 5 ,
+---i---r-+-b.:-;j2-: t-+-r-
~c ~-r=':~~-~~---;
0
<->
.1
.05 ··.··+.······.···--····i·····---·····---··i··------·--t-····..--··--···---·-····-···-··t·······---·---T----··-----r--·--·---·. t-..--·
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 13-3: Contradiction Rate R11 When the Number of Fuzzy Alternatives
is Equal to 3 .
.55
: :
.5
.45
.4
20 .35 ~- - - .. _..... .
0::
c
0 .3
:.-
0
'0 .25
~c
~4~g
0 .2
<->
.15
.1
.05
0
4 6 8 10 12 14 16 18 20
Number of Criteria
is Equal to 21.
260 MCDM Methods: A Comparative Study. by E. Triantaphyl/Ou
.1
r ,
i F-IjpM
.08 -.•••••• ---~, ir'_~M ____ M.'mr···--······-·······t,·-··--····--·'··1---"-"----'j"---""
\" , ,
.06 ···"..:··,~,,-·····---······--·1····-···-·--·----··!--·
G>
'0 \L ~
c::
c
.2
t) .04
"6
~
c:0
u .02
o
4 6 8 10 12 14 16 18 20
Number of Criteria
is Equal to 3 .
.2
.15
',I
......... 1" __ .,_.~ __.._.,i:,"'" ~

..-___ ._. ___ i ...... _.._._. __ J._____.;. ._. .. !
1
! I :
-----.r-----------.
i ........
! !
i.!,
I • 1
20 l i ! I !
i, ! i, ! j
c::
c:
!"" !
i i I
l
!
i
:8u
"6
E
.1
iii:
i .--...-.. ---.--..i-.-.---..-.'.'.--~~ ..'-.....--.-....1·-----·---··-1..---.................L..............--•.•.. -·l--··-··-·····-~----------;-
! l .
t1t
c:
0
(.)
.05
o ~-L~__~~____~~__J-~__-L__~~~__~~~
...•~._
.. -_ •.~
.-~ ...._...~~~
4 6 8 10 12 14 16 18 20
Number of Criteria
is Equal to 21.
.55 r -" --""T,--""Tl--""Tl---',---',---',-~---r-

i - -'
, --,
.5
.45
.4
a
II>
.35
cr
.2 .3
t>
'g... .25
C
8 .2
.15
.05 ·.·---·.i---·-----.-----i·---··-·---··--t··-·------·.·-l--------·-f-·----·-·---·-r--···--·-·-···t---··~-··--·---·t-----·····--·--·····[·---····
! ! . 1 j ! ~
4 6 8 10 12 14 16 18 20
Number af Criteria
is Equal to 3.
13.8.2 Analysis of the Computational Results

The results derived from the computational experiments lead to some
interesting observations. First, none of the fuzzy MCDM methods which
were examined in this study is perfectly effective in terms of both fuzzy
evaluative criteria. The results indicate that each method yields different rates
of contradiction.
Secondly, the results reveal that the contradiction rates increase when
the number of alternatives increases. That is , the methods are less accurate
when the MCDM problems become more complex.
Finally, it appears that the fuzzy revised AHP is the best fuzzy
MCDM method in most cases, although the difference in performance may
be small in certain cases. From these results, the fuzzy revised AHP has the
smallest contradiction rates in terms of both fuzzy evaluative criteria. As it
was stated in the previous section, this should not come as a surprise.
At this point it should also be stated that in [Triantaphyllou and Lin,
1996] more scales were used for quantifying the pairwise comparisons in
similar experiments. In particular, we considered two geometric scales
introduced by Lootsma [1988] (see also Chapter 3) in which the gamma (y)
262 MCDM Methods: A Comparative Study, by E. TriantaphyLLou
parameter was equal to 0.50 and 1.00. However, the derived contradiction
rates were significantly higher than the ones derived when the Saaty scale was
used, and thus these results were not plotted. In summary, the findings of
this study reveal that some fuzzy MCDM methods are better than others in
some cases even though none of the methods is perfectly accurate.
13.9 CONCLUSIONS
The analyses presented in this chapter reveal that none of the four
fuzzy MCDM methods is completely perfect in terms of both fuzzy evaluative
criteria. Different contradiction rates are yielded when these fuzzy MCDM
methods are tested according to the two evaluative criteria. The fuzzy WSM
could be the simplest method to solve single-dimensional decision making
problems. However, the other more systematic approaches; the fuzzy AHP,
the fuzzy revised RAHP, the fuzzy WPM, and the fuzzy TOPSIS, are more
capable of capturing a human's appraisal of ambiguity when complex MCDM
problems are considered. This is true because pairwise comparisons provide
a flexible and realistic way to accommodate real life data. The experimental
results reveal that the fuzzy revised (ideal mode) AHP is better than the other
methods in terms of the previous two fuzzy evaluative criteria.
It needs to be emphasized here that these fuzzy MCDM methods are
best to be used as decision tools. Individual decision makers may reach their
own solution after applying anyone of them. This study provided only a
general view of different methods under certain situations. A broader
understanding of the characteristics of the methods, and fuzzy evaluative
criteria is required for a successful accomplishment in solving real life fuzzy
MCDM problems.
Chapter 14
CONCLUSIONS AND DISCUSSION FOR

FUTURE RESEARCH
14.1 THE STUDY OF MCDM METHODS:

FUTURE TRENDS
It is widely accepted today that people do not always behave the way
the well studied normative theories say they ought to behave (see, for
instance, [Allais and Hagen, 1979], [Bell, et al., 1988], [Ellsberg, 1961], and
Raiffa [1984]. Many decision theories (especially game theories) assume that
the decision makers are always perfectly rational. Too often, however,
important decisions are based on non-scientific clues (see, for instance,
[Kadane and Larkey, 1982a; 1982b; and 1983], [McMillan, 1992], [Neale
and Bazerman, 1991], [Raiffa, 1982], and [Sebenius, 1992].
Even well defined classical statistic inference procedures can be of
limited benefit because the input probabilities are hard to be estimated or the
underlying assumptions are not valid at a given situation or they cannot be
validated properly (see, for instance, [French, 1986], [Pratt, et al., 1994],
[Raiffa, 1968], and [Raiffa and Schlaifer, 1968]). The above issues constitute
the three major concerns on the future of research and practice as was
expressed recently by Raiffa [1994].
As an antidote to the above problems, Raiffa in [1994] proposed a
new orientation to the study of decision making which he calls the prescriptive
orientation. According to this orientation, people should not be considered
as super rational individuals. Instead new methods should seek to establish
a systematic and wise approach in making decisions. Raiffa also proposed
that the study of decision making should combine normative theories with a
deep understanding of the cognitive and behavioral aspects involved in real
life decision making.
14.2 LESSONS LEARNED

The MCDM methods studied in this book are mostly deterministic and
assume that the decision maker has a near perfect knowledge of the
information involved in a given MCDM problem. That is, he/she knows all
the pertinent alternatives and decision criteria that define the main components
of the MCDM problem on hand. Furthermore, the decision maker knows,

or can satisfactorily assess, the numerical data involved in solving such a

decision problem.
The previous chapters presented some of the most widely used
methods for addressing all the steps in solving an MCDM problem. From the
quantification of qualitative data to the processing of the final numerical data.
However, even under the most favorable assumption that the decision maker
is a rational person and the information involved is perfectly known, we
discovered that still different methods may result in different conclusions for
the same problem. This situation keeps stimulating the introduction of even
more such methods. This in tum, may compound the problems caused to
practitioners by the already plethora of such competing methods.
The main conclusion of the comparative study presented in this book
is that for certain problems one may never know what the best decision is,
even if perfect knowledge in the input data and the structure of the MCDM
problem are assumed. Moreover, there may not be a single MCDM method
that can always ensure the best decision. All the methods studied failed in
terms of some of the evaluative criteria considered in the investigations
reported in this book. However, some methods failed more dramatically than
others. Finding the "best" MCDM method is a very elusive goal
[Zimmermann and Gutsche, 1991] and this quest may never be answered.
On the other hand, for some problems the ranking of the alternatives
may be the correct one. This depends on how dominant the alternatives are
in terms of the most discriminatory decision criteria. Ironically, for such
problems most methods may point out to the same final conclusion any way.
The situation becomes blurred when the alternatives are difficult to distinguish
and rank. This suggests that the decision maker should always be vigilant
before accepting the results of an MCDM method. If the results strongly
suggest a particular alternative to be the best one and the method used to
derive this conclusion was determined to be resilient in terms of the
evaluations described here, then this is probably the right conclusion.
At this point one may reasonably claim that there may never be a
single MCDM method that can ensure that the derived ranking of the
alternatives is the correct one. The problem of determining the best
alternative is an ill defined one by its own nature. It is hard to know the best
solution even under perfect knowledge of the input data. What can be done,
however, is to analyze the behavior of the numerical methods under certain
evaluative criteria that are motivated by widely accepted notions of logical
stability. This was done in the treatments presented in this book.
For the successful research and application of MCDM methods, one
needs both a deep understanding of the pure numerical properties of the
methods involved, as well as the cognitive and behavioral aspects of the
decision making process. Multi-criteria decision making often involves the
Chapter 14: Conclusions and Discussion for Future Research 265
processing of hard to quantify aspects of a problem. The present study

examined the numerical aspects and some of the cognitive and behavioral
issues involved in the research and application of some of the most widely
used MCDM methods. More research and application of such methods will
certainly follow in the future. This study hopefully provides a systematic
methodology for evaluating and better understanding existing and even future
MCDM methods.
REFERENCES
Ackoff, R.L., (1978), "The Art of Problem Solving, " John Wiley & Sons, Inc., New York,
NY, USA.
Allais, M., and O. Hagen (Eds.), (1979), "Expected Utility Hypothesis and the Allais
Paradox, " D. Reidel, Dordrecht, The Netherlands.
Anderberg, M.R., (1973};, "Probability and Mathematical Statistics, " Cluster Analysis for
Applications. Academic Press, New York, NY, USA.
Arbel, A., and A. Seidmann, (1984), "Selecting a Microcomputer for Process Control and
Data Acquisition, " IEEE Transactions, Vol. 16, No. I, pp. 16-25.
Armacost, R.L., and J.C. Hosseini, (1994), "Identification of Determinant Attributes Using
the Analytic Hierarchy Process, "Journal of the Academy of Marketing Science, Vol.
22, No.4, pp. 383-392.
Baas S.J., and H. Kwakernaak, (1977), "Rating and Ranking of MUlti-Aspect Alternatives
Using Fuzzy Sets, "Automatica, Vol. 13, pp. 47-58.
Barzilai, J., and F.A. Lootsma, (1994). "Power Relations and Group Aggregation in the
Multiplicative AHP and SMART, " Proceedings of the Third International Symposium
on the AHP, George Washington University, Washington, DC, USA, 157-168.
Bell, R.E., and P.H. Farquhar, (1986), "Perspectives on Utility Theory," Operations
Research, Vol. 24, pp. 179-183.
Bell, D.E., H. Raiffa, and A. Tversky, (1988), Decision Making: Descriptive, Normative,
and Prescriptive Interactions, Cambridge University Press, Cambridge, England.
Bellman, R.E., and L.A. Zadeh, (1970), "Decision-Making in a Fuzzy Environment,"
Management Science, Vol. 17B, No.4, pp. 141-164.
Belton, V., and T. Gear, (1983), "On a Short-Coming of Saaty's Method of Analytic
Hierarchies, " Omega, pp. 228-230.
Ben-Arieh, D., and E. Triantaphyllou, (1992), "Quantifying Datafor Group Technology with
Weighted Fuzzy Features, " International Journal of Production Research, Vol. 30,
No.6, pp. 1285-1299.
Benayoun, R., B. Roy, and N. Sussman, (1966), "Manual de Reference du Programme
Electre, " Note de Synthese et Formaton, No. 2S, Direction Scientifque SEMA, Paris.
France.
Boender, C.G.E., J.G. de Graan, and F.A. Lootsma, (1989), "Multi-Criteria Decision
Analysis with Fuzzy Pairwise Comparisons, " Fuzzy Sets and Systems, Vol. 29, pp.
133-143.
Boucher, T.O., and E.L. McStravic, (1991), "Multi-Attribute Evaluation within a Present
Value Framework and Its Relation to the Analytic Hierarchy Process," The
Engineering Economist, Vol. 37, No. I, pp. 55-71.
Bridgman, P.W., (1922), "Dimensional Analysis," Yale University Press, New Haven, CT,
USA.
Buckley, J.J., (1985), "Ranking Alternatives Using Fuzzy Numbers," Fuzzy Sets and Systems,
Vol. 15, pp. 21-31.
Buckley, J.J., (1985), "Fuzzy Hierarchical Analysis," Fuzzy Sets and Systems. Vol. 17, pp.
233-247.
Cambron, K.E., and G.W. Evans, (1991), "Layout Design Using the Analytic Hierarchy
Process, " Computers and Industrial Engineering, Vol. 20, No.2, pp. 221-229.
Chang, S.K., (1971), "Proceedings of the Brooklyn Poly technical Institute Symposium on
Computers and Automata," Fuzzy Programs, Brooklyn, New York, NY, USA. Vol.
21, pp. 124-135.

Chen, S.J., and C.L. Hwang, (1991), "Fuzzy Multiple Attribute Decision Making: Methods
and Applications," Lecture Notes in Economics and Mathematical Systems, No.
375, Sringer-Verlag, Berlin, Germany.
Chu, A.T.W., R.E. KaIaba, and K. Spingarn, (1979), '~ Comparison of Two Methodsfor
Determining the Weights ofBelonging to Fuzzy Sets, " Journal of Optimization Theory
and Applications, Vol. 27, pp. 531-538.
Churchman, C.W., E.L. Arnoff, R.L. Ackoff, (1957), "Introduction to Operations
Research," Wiley, New York, NY, USA.
Dantzig, G.B., (1963). "Linear Programming and Extensions," Princeton University Press,
Princeton, NJ, USA.
Decision Support Software Inc., (1986), "Expert Choice Software Package, "McLean, VA,
USA.
Donegan, H.A., and F.J. Dodd, (1991), "A Note on Saaty 's Random Indexes, " Mathematical
and Computer Modeling, Vol. 15, No. 10, pp. 135-137.
Dubois, D., and H. Prade, (1980), "Fuzzy Sets and Systems: Theory and Applications, "
Academic Press, New York, NY, USA.
Dyer, J.S., and H.V. Ravinder, (1983), "Irrelevant Alternatives and the Analytic Hierarchy
Process," Technical Report, Department of Management, The University of Texas
at Austin, Austin, TX, USA.
Dyer, J.S., and W.E. Wendell; (1985), "A Critique of the Analytic Hierarchy Process,"
Technical Report 84/85-4-24, Department of Management, The University of Texas
at Austin, Austin, TX, USA.
Dyer, J.S., (1990), "Remarks on the Analytic Hierarchy Process, " Management Science, Vol.
36, No.3, pp. 249-258.
Dyer, J.S., (1990), ~ Clarification of "Remarks on the Analytic Hierarchy Process",'
Management Science, Vol. 36, No.3, pp. 274-275.
Dubois, D., and B. Prade, (1980), "Fuzzy Sets and Systems, " Academic Press, New York,
NY, USA.
Dubois, D., and H. Prade, (1979), "Fuzzy Real Algebra: Some Results, " Fuzzy Sets and
Systems, Vol. 2, pp. 327-348.
EUsberg, D., (1961), "Risk, Ambiguity, and the Savage Axioms," Quarterly Journal of
Economics. No. 75, pp. 643-649.
Gupta, M.M., R.K. Ragade, and R.Y. Yager, (Editors), (1979), "Fuzzy Set Theory and
Applications," North-Holland, New York, NY, USA.
Falkner, C., and S. BenhajIa, (1990), "Multi-Attribute Decision Models in the Justification
of CIM Systems, " The Engineering Economist, Vol. 35, pp. 91-114.
Federov, V.V., V.B. Kuz'min, V.B., and A.I. Vereskov, (1982), "Membership Degrees
Determinationfrom Saaty Matrix Totalities," Institute for System Studies, Moscow,
former USSR. Paper appeared in: Approximate Reasoning in Decision Analysis,
M.M. Gupta, and E. Sanchez (Editors), North-Holland Publishing Company, pp.
23-30.
Finnie, J., (1988), "The Role of Financial Appraisal in Decisions to Acquire Advanced
Manufacturing Technology," Accounting and Business Research, Vol. 18, No.7, pp.
133-139.
Fishburn, P. C., (1967), "Additive Utilities with Incomplete Product Set: Applications to
Priorities and Assignments," Operations Research Society of America (ORSA),
Baltimore, MD, USA.
Fishburn, P. C., and I.H. LaValle, (Editors), (1989), "Annals of Operations Research.
References 269
Volume i9: Choice Under Uncertainty," J.C. Baltzer, Basel, Switzerland.
French, S., (1986), Decision Theory, Ellis Horwood Limited, London, England.
French, S. (1989), Readings in Decision Analysis, Chapman and Hall, London, England.
French, S., and R.D. Insua, (1989), "Partial information and Sensitivity Analysis in Multi-
Objective Decision Making". In P. Lockett (Ed.), Improving Decision Making in
Organizations. pp. 23-41, LNEMS, Springer-Verlag, London, England.
de Graan, J.G., (1980), "Extension of the Multiple Criteria Analysis Methods ofT.L. Saaty, "
National Institute for Water Supply, Voorburg, The Netherlands.
Haerer, W., (2000), "Criterium Decision Plus 3.0: Versatile Multi-Criteria Tool Excels in
Ability to Support Decision-Making," ORiMS Today, the flagship magazine
published by INFORMS (ORIMS Society of America), Vol. 27, No.1, February,
pp.40-43.
Harker, P. T., (1987), "incomplete Pairwise Comparisons in the Analytic Hierarchy Process, "
Mathematical Modeling, Vol. 9, pp. 837-848.
Harker, P.T., and L.G. Vargas, (1987), "The Theory of Ratio Scale Estimation: Saaty's
Analytic Hierarchy Process," Management Science, Vol. 33, No. 11, November,
pp. 1383-1403.
Harker, P.T., and L.G. Vargas, (1990), "Reply to 'Remarks on the Analytic Hierarchy
Process'," Management Science, Vol. 36, No.3, pp. 269-273.
Hedge, G.G., and P.R. TadikamaUa, (1990), "Site Selection for a Sure Service Terminal, "
European Journal of Operational Research, 48, pp. 77-80.
Hogarth, R.M., and M.W. Reder, (Editors), (1986), "Rational Choice: The Contrast
Between Economics and Psychology, " University of Chicago Press, Chicago, IL,
USA.
Hibn, J.M., and C.R. Johnson, (1988), "Evaluation Techniquesfor Paired Ratio-Comparison
Matrices in a Hierarchical Decision Model, " Measurement in Economics, Psysics-
Verlag, pp. 269-288.
Hwang C.L., and K. Yoon, (1981), "Multiple Attribute Decision Making: Methods and
Applications, " Springer-Verlag, New York, NY, USA.
Hwang, F., (1987), "An Expert Decision Making Support System For Multiple Attribute
Decision Making," Ph.D. Thesis, Dept. of Industrial Engineering, Kansas State
University, Manhattan, KS, USA.
Insua, R.D., (1990), "Sensitivity Analysis in Multi-Objective Decision Making," Lecture
Notes in Economics and Mathematical Systems Series, Springer-Verlag, Berlin,
Germany.
Kadane, J.B., and P.D. Larkey, (1983), "The Confusion of is and Ought in Game Theoretic
Contexts, " Management Science, No. 29, pp. 1365-1379.
Kadane, J.B., and P.D. Larkey, (1982a), ·Subjective Probability and the Theory of Games, "
Management Science. No. 28, pp. 113-120.
Kadane, J.B., and P.D. Larkey, (1982b), "Reply to Professor Harsanyi," Management
Science, No. 28, pp. 124.
Kalaba, R., and K. Spingarn, (1979), "Numerical Approaches to the Eigenvalues of Saaty 's
Matrices for Fuzzy Sets, " Computers and Mathematics with Applications, Vol 4, pp.
369-375.
Kaufmann A., (1975), "introduction to the Theory of Fuzzy Sets, " Academic Press, New
York, NY, USA.
Khurgin, J.I., and V.V. Polyakov, (1986), "Fuzzy Analysis of the Group Concordance of
Expert Preferences, defined by Saaty Matrices," Fuzzy Sets Applications.
Methodological Approaches and Results, Akademie-Verlag, Berlin, Germany, pp.
270 MCDM Methods: A Comparative Study, by E. TrianraphyUou
111-115.
Khurgin, J. I., and V.V. Polyakov, (1986), "Fuzzy Approach to the Analysis of Expert
Data," Fuzzy Sets Applications. Methodological Approaches and Results,
Akademie-Verlag, Berlin, Germany, pp. 116-124.
Kuhn, H.W., and A.W. Tucker, (1951), "Nonlinear Programming," in Proceedings of the
Second Berkeley Symposium on Mathematical Statistics and Probability. J. Neyman
(Ed.), University of California at Berkeley Press, Berkeley, Calif., USA, pp. 481-
492.
Laarhoven, P.J.M., and W. Pedrycz, (1983), "A Fu¥y Extension of Saaty's Priority
Theory, " Fuzzy Sets and Systems. Vol. 11, pp. 229-241.
Lakoff, G., (1973), "Hedges: A Study in Meaning Criteria and the Logic of Fuzzy Concepts, "
J. Philos. Logic. pp. 12-22.
Lee, C.T.R., (1971), "Fuzzy Logic and the Resolution Principle," Second International Joint
Conference on Artificial Intelligence. London, England, pp. 560-567.
Lee, S.N., Y.L. Grlze, and K. Dehnad, (1987), "Quantitative Modelsfor Reasoning Under
Uncertainty in Knowledge Based Expert Systems", International Journal of Intelligent
Systems. Vol. 2, pp. 15-38.
Legrady, K., F.A. Lootsma, J. Meisner and F. ScheUemans, (1984), "Multi-Criteria
Decision Analysis to Aid Budget Allocation,· in M. Grauer en A.P. Wierzbicki,
(Editors), Interactive Decision Analysis, Springer-Verlag, Berlin-New York, NY,
USA, pp. 164-174.
Lootsma, F .A., (1999), "Multi-Criteria Decision Analysis via Ratio and Difference Judgment. "
Kluwer Academic Publishers. Applied Optimization Series, Vol. 29, Dordrecht, The
Netherlands.
Lootsma, F .A., (1997), "Fuzzy Logic for Planning and Decision Making." Kluwer Academic
Publishers. Applied Optimization series, Vol. 8, Dordrecht, The Netherlands.
Lootsma, F.A., (1990), "The French and The American School in Multi-Criteria Decision
Analysis,· Recherche Operationnele / Operations Research, Vol. 24, No.3, pp.
263-285.
Lootsma, F .A., (1980), "Saaty's Priority Theory and the Nomination of a Senior Professor in
Operations Research, • European Journal of Operational Research, Vol. 4, pp. 380-
388.
Lootsma, F.A., T.C.A. Mensch, and F.A. Vos, (1990), "Multi-Criteria Analysis and Budget
Reallocation in Long-Term Research Planning, " European Journal of Operational
Research, Vol. 47, pp. 293-305.
Lootsma, F.A., (1991), "Scale Sensitivity and Rank Preservation in a Multiplicative Variant
of the AHP and SMART," Technical Report 91-67, Faculty of Technical
Mathematics and Informatics, Delft University of Technology, Delft, The
Netherlands.
Lootsma, F.A., (1988), "Numerical Scaling of Human Judgment In Pairwise-Comparison
Methods For Fuzzy Multi-Criteria Decision Analysis," Mathematical Models for
Decision Support, NATO ASI Series F, Computer and System Sciences, Springer-
Verlag, Berlin, Germany, Vol. 48, pp. 57-88.
Ma, D., and X. Zheng, (1979), "9/9-9/1 Scale Method of the AHP," Proceedings of the 2nd
International Symposium on the AHP, Vol. I, Pittsburgh, PA, USA, pp. 197-202.
Mardia, K.V., J.T. Kent, and J.M. Bibby, (1979), "Multivariate Analysis." Academic
Press, New York, NY, USA.
Marks, L.E., (1974), "Sensory Processes. The New Psychophysics. " Academic Press, New
York, NY, USA.
References 271
Masuda, T., (1990), "Hierarchical Sensitivity Analysis of the Priorities Used in the Analytic
Hierarchy Process," Inter'l Journal of Systems Science, Vol. 21, No.2, pp.
415-427.
McCahon, C.S., and E.S. Lee, (1988), "Comparing Fuzzy Numbers: The Proportion of the
Optimum Method," Inter'l Journal of Approx. Reasoning, Vol. 4, pp. 159-163.
McMillan, J., (1992), Games, Strategies, and Managers, Oxford University Press, Oxford,
England.
Michon, J.A., Eijkman, E.G.J., and L.F.W. de Klerk, (Editors), (1976), "Handboek der
Psychonomie, " (in Dutch), Van Loghum Slaterus, Deventer, The Netherlands.
Miller, D.W., and M.K. Starr, (1969), "Executive Decisions and Operations Research, "
Prentice-Hall, Inc., Englewood Cliffs, NJ, USA.
Miller, C.A., (1956), "The Magic Number Seven Plus or Minus Two: Some Limits on Our
Capacity for Processing Information, " Psychological Review, Vol. 13, pp. 81-97.
Munier, B.R., (Editor), (1988), "Risk, Decision and Rationality," D. Reidel, Dordrecht,
The Netherlands.
Neale, M.S., and M.H. Bazerman, (1991), Cognition and Rationality in Negotiation, The
Free Press, Amsterdam, The Netherlands.
Pratt, J.W., H. Raiffa, and R.O. Schlaifer, (1994), Introduction to Statistical Decision
Theory, M.I.T. Press, Boston, MA, USA.
Prade, H., and C.V. Negoita, (Editors), (1986), "Fuzzy Logic in Knowledge Engineering,"
Verlag TUV Rheinland, Berlin, Germany.
Putrus, R., (1990), "Accounting for Intangibles in Integrated ManUfacturing (nonfinancial
justification based on the Analytical Hierarchy Process, "Information Strategy, Vol.
6, pp. 25-30.
Raiffa, H., (1994), "The Prescriptive Orientation of Decision Making: A Synthesis of
Decision Analysis, Behavioral Decision Making, and Game Theory, " in Decision
Theory and Decision Analysis: Trends and Challenges, Sixto Rios, (Ed.), Kluwer
Academic Publishers, Boston, Mass., USA.
Raiffa, H., (1984), Back from Prospect Theory to Utility Theory, Plural Rationality and
Interactive Decision Processes, LNEMS, Springer-Verlag, New York, NY, USA,
pp. 100-113.
Raiffa, H., (1982), The Art and Science of Negotiations, Harvard University Press. Boston,
Mass., USA.
Raiffa, H., (1968), Decision Analysis, Addison Wesley, Boston, Mass., USA.
Raiffa, H., and P.J.H. Schlaifer, (1968), Applied Statistical Decision Theory, M.I.T. Press,
Boston, Mass., USA.
Ramsay, A., (1988), "Formal Methods in Artificial Intelligence, " Cambridge University
Press, Cambridge, England.
Ray, T., and E. Triantaphyllou, (1999), "Procedures For the Evaluation of Conflicts in
Rankings of Alternatives, " Computers and Industrial Engineering, Vol. 36, No.1,
pp. 35-44.
Ray, T., and E. TriantaphyUou, (1998), "Evaluation of Rankings With Regard to the
Possible Number of Agreements and Conflicts, " European Journal of Operational
Research, Vol. 106, No.1, pp. 129-136.
Roberts, F.S., (1979), "Measurement Theory, " Addison-Wesley, Reading, Mass., USA.
Roper-Lowe, G.C., and J.A. Sharp, (1990), "The Analytic Hierarchy Process and its
Application to an Information Technology Decision, " Operational Research Society
Ltd., Vol. 41, No.1, pp. 49-59.
Ruspini, E,H., (1991), "On the Semantics of Fuzzy Logic," Inter'l Journal of Approximate
272 MCDM Methods: A Comparative Study, by E. TriantaphyUou
Reasoning. Vol. 5, No.1, pp. 45-88.

Saaty, T.L., (1994), Fundamentals of Decision Making and Priority Theory with the AHP.
RWS Publications, Pittsburgh, PA, USA.
Saaty, T.L., (1990), 'An Exposition of the AHP in Reply to the Paper "Remarks on the
Analytic Hierarchy Process",' Management Science, Vol. 36, No.3, pp. 259-268.
Saaty, T.L., (1987), "Rank Generation, Preservation, and Reversal in the Analytic Hierarchy
Process," Decision Sciences, Vol. 18, pp. 157-177.
Saaty, T.L., (1983a), "Priority Setting in Complex Problem," IEEE Transactions on
Engineering Management, Vol. EM-30, No.3, pp. 57-68, 1983.
Saaty, T.L., (1983b), "Axiomatic Foundations of the Analytic Hierarchy Process,"
Management Science, Vol. 32, No.7, pp. 841-855.
Saaty, T.L., (1980), "The Analytic Hierarchy Process, "McGraw-Hill, New York, NY, USA.
Saaty, T.L., (1978), "Exploring the Interface Between Hierarchies, Multiple Objects and
Fuzzy Sets, " Fuzzy Sets and Systems, Vol. 1, pp. 57-68.
Saaty, T.L., (1977), "A Scaling Methodfor Priorities in Hierarchical Structures," Journal
of Mathematical Psychology, Vol. 15, No.3, pp. 57-68.
Saaty, T.L., (1974), "Measuring the Fuzziness of Sets, "Cybernetics, Vol. 4, No.4, pp. 53-
61.
Sarin, R. K., (1989), "Analytical Issues in Decision Methodology," in I. Horowitz (Editor),
Organization and Decision Theory, Kluwer-Nijhoff, Dordrecht, The Netherlands, pp.
156-172.
Sebenius, J.K., (1992), "Negotiation Analysis: A Characterization and Review, "Management
Science. No. 38, pp. 18-38.
Shrestha, G., and S. Rahman, (1991), "A Statistical Representation of Imprecision in Expert
Judgments, " Inter'l Journal of Approximate Reasoning, Vol. 5, No.1, pp. 1-25.
Simon, H.A., (1961), "Models of Man." John Wiley & Sons, Inc., (Second Edition), New
York, NY, USA.
Sneath, P.H.A., and R.R. Sokal, (1973), "Numerical Taxonomy. " W.H. Freeman and
Company, San Francisco, Calif., USA.
Stewart, S.M., (1973), "Introduction to Matrix Computations," Academic Press, New York,
NY, USA.
Stevens, S.S., and M.D. Hallowell Davis, (1983), "Hearing. its Psychology and Physiology",
American Institute of Physics, New York, NY, USA.
Swann, K., and W.O. O'Keefe, (1990), "Advanced Manufacturing Technology: Investment
Decision Process," Management Decisions, Vol. 28, No.1, pp. 20-31.
Tong, R.M., and P.P. Bonissone, (1980), "A Linguistic Approach to Decision Making with
Fuzzy Sets," IEEE Transactions on Systems. Man and Cybernetics. Vol. SMC-10,
No. 11, pp. 716-723.
Tong, R.M., and P.P. Bonissone, (1984), "Linguistic Solutions to Fuzzy Decision
Problems," Fuzzy Sets and .Decision Analysis. TIMS Studies in Management Sci. ,
Vol. 20, pp. 323-334.
Triantaphyllou, E., (2000), "Some New Cases of Ranking Irregularities When the AHP and
Some of Its Variants Are Used," Working Paper, Dept. of Industrial and
Manufacturing Systems Engineering, Louisiana State University, Baton Rouge, LA,
USA.
Triantaphyllou, E., (1999), "Reduction of Pairwise Comparisons in Decision Making via a
Duality Approach, "Multi-Criteria Decision Analysis. Vol. 8, No.6, pp. 299-310.
Triantaphyllou, E., and A. Sanchez, (1997), "A Sensitivity Analysis Approach for Some
Deterministic Multi-Criteria Decision-Making Methods, " Decision Sciences. Vol. 28,
References 273
No.1, pp. 151-194.
TriantaphyUou, E., B. Kovalerchuk, L. Mann, Jr., and G. Knapp, (1997), "Determining
the Most Important Criteria in Maintenance Decision MaJdng," Quality in
Maintenance Engineering. Vol. 3, No.1, pp. 16-26.
TriantaphyUou, E., and C-.T. Lin, (1996), "Development and Evaluation of Five Fuzzy
Multi-Anribute Decision-MaJdng Methods," inter'l Journal of Approximate
Reasoning. Vol. 14, No.4, pp. 281-310.
TriantaphyUou, E., and S.H. Mann, (1995), "Using the Analytic Hierarchy Process For
Decision MaJdng in Engineering Applications: Some Challenges, "inter'l Journal of
Industrial Engineering: Applications and Practice. Vol. 2, No.1, pp. 35-44.
TriantaphyUou, E., (1995), "A Linear Programming Based Decomposition Approach in
Evaluating Relalive Priorities from Pairwise Comparisons and Error Analysis, "
Journal of Optimization Theory and Applications, Vol. 84, No.1, pp. 207-234.
TriantaphyUou, E., F.A. Lootsma, P.M. PardaIos, and S.H. Mann, (1994), "On the
Evaluation And Application of Different Scales For Quantifying Pairwise
Comparisons in Fuzzy Sets," Journal of Multi-Criteria Decision Analysis, Vol. 3,
No.3, pp. 133-155.
TriantaphyUou, E., (1994), "Decision MaJdng in Engineering and Business: A Cost-Effective
Approach," Proceedings of Northcon '94. Seattle, WA, USA, October 11-13, pp.
232-237.
TriantaphyUou, E., and S.H. Mann, (1994), "Some Critical Issues in MaJdng Decisions with
Pairwise Comparisons, " Proceedings of the Third International Symposium on the
Analytic Hierarchy Process. George Washington University, Washington, DC, USA,
July 11-13, pp. 225-236.
TriantaphyUou, E., and S.H. Mann, (1994), "A Computational Evaluation of the AHP and
the Revised AHP when the Eigenvalue Method is Used under a Continuity
Assumption, " Computers and industrial Engineering, Vol. 26, No.3, pp. 609-618.
TriantaphyUou, E., (1993), "A Quadratic Programming Approach In Estimating Similarity
Relations, " IEEE Transactions on Fuzzy Systems, Vol. 1, No.2, pp. 138-145.
TriantaphyUou, E., (1992), "A Sensitivity Analysis of a (l'i' SJ Inventory Policy with
Increasing Demand, " Operations Research Letters, Vol. 11, No.3, pp. 167-172.
TriantaphyUou, E., P.M. Pardalos, and S.H. Mann, (1990), "The Problem of Determining
Membership Values in Fuzzy Sets in Real World Situations," Operations Research
and Artificial Intelligence: The Integration of Problem Solving Strategies. (D.E.
Brown and C.C. White III, Editors), Kluwer Academic Publishers, New York, NY,
USA, pp. 197-214.
TriantaphyUou, E., and S.H. Mann, (1990), "An Evaluation of the Eigenvalue Approachfor
Determining the Membership Values in Fuzzy Sets," Fuzzy Sets and Systems, Vol.
35, No.3, pp. 295-301.
TriantaphyUou, E., P.M. Pardalos, and S.H. Mann, (1990), "A Minimization Approach to
Membership Evaluation in Fuzzy Sets and Error Analysis," Journal of Optimization
Theory and Applications, Vol. 66, No.2, pp. 275-287.
TriantaphyUou, E., and S.H. Mann, (1989), "An Examination of the Effectiveness of Multi-
Dimensional Decision-Making Methods: A Decision-Making Paradox," inter'l
Journal of Decision Support Systems, No.5, pp. 303-312.
Vargas, L. G., (1982), "Reciprocal Matrices with Random Coefficients," Mathematical
Modeling, Vol. 3, pp. 69-81.
Xie, L.A., and S.D. Berdosian, (1983), "The Information in a Fuzzy Set and the Relalion
Between Shannon and Fuzzy Information," Proceedings of the 7th Annual
274 MCDM Methods: A Comparative Study. by E. TriantaphyUou
Conference on Information Sciences and Systems. Johns Hopkins University.

Baltimore. MD. USA. pp. 212-233.
Wabalickis, R.N •• (1988). wJustiftcation of FMS with the Analytic Hierarchy Process, "
Manufacturing Systems. Vol. 17. No.3. pp. 175-182.
Wang. L., and T. Raz. (1991). "Analytic Hierarchy Process Based on Data Flow Problem, "
Computers and Industrgl Engineering. Vol. 20. No. 3.pp. 355-365.
Weber, M., and C. Camerer. (1987). WRecent Developments in Modeling Preferences Under
Risk, W OR. Spectrum. Vol. 9. pp. 129-151.
Wmkler, R.L •• (1990). "Decision Modeling and Rational Choice: AHP and Utility Theory, "
Management Science. Vol. 36. No.3. pp. 247-248.
Winston, W.L., (1991). "Operations Research, Applications and Algorithms," Second
Edition. PWS-Kent Publishing Co .• New York. NY. USA.
Write, C., and M.D. Tate, (1973). "Economics and Systems Analysis: Introduction for Public
Managers," Addison-Wesley. Reading. MA. USA.
Zadeh, L.A•• (1965). "Fuzzy Sets, W Information and Control. Vol. 8. pp. 338-353.
Zadeh, L.A•• (1968). wFuzzy Algorithms, W Information and Control. Vol. 12. pp. 94-102.
Zadeh, L.A., K.S. Fu, K. Tanaka, K., and M. Shimura, (Editors). (1975). "Fuzzy Sets
and their Applications to Cognitive and Decision Processes. " Academic Press. New
York. NY. USA.
Zadeh, L.A.• (1978). "Fuzzy Sets as a Basis for a Theory of Possibility," Fuzzy Sets and
Systems. Vol. 1. pp. 3-28.
Zadeh, L.A.• (1979). wA Theory of Approximate Reasoning," Machine Intelligence. Edited
by J. Hayes. D. Michie. and L.I. Mikulich. John Wiley & Sons. Inc .• New York.
NY. USA. Vol. 9. pp. 149-194.
Zadeh, L.A.. "The Concept of a Linguistic Variable and Its Applications to Approximate
Reasoning," Parts 1-3. Information Sciences. Vol. 8. pp. 199-249. 1975; Vol. 8.
pp. 301-357. 1975; Vol. 9. pp. 43-80. 1976.
Zadeh, L.A.• (1983). "ean Expert Systems be Designed Without Using Fuzzy Logic?",
Proceedings of the 17th Annual Conference on Information Sciences and Systems;
Johns Hopkins University. Baltimore, MD. USA. pp. 23-31.
Zimmermann, H.4.• (1985). "Fuzzy Set Theory - and Its Applications. Kluwer Academic
Publishers. Amsterdam. The Netherlands.
Zimmermann, H.4.• (1996), Fuzzy Set Theory and Its Applications. Kluwer Academic
Publishers. Third Revised Edition. Boston. MA. USA.
Zimmermann, H.4., and L. Gutsche. (1991). Multi-Criteria Analyze-Einfiihrung in die
Theorie der Entscheidngenbei Mehrfachzielsetzungen. (in German) Springer-Verlag.
Berlin. Heidelberg. Germany.
Zhu, Q, and E.S. Lee, (1991). "Comparison and Ranking of Fuzzy Numbers," Fuzzy
Regression Analysis. Omnitech Press. pp. 23-32. Warsaw. Poland.
Zwicker, E •• (1982). Psychoakustik. Springer. Berlin, Germany.
SUBJECT INDEX
A
absolute-any (AA) critical criterion 134, 141, 152
absolute terms 133
absolute-top (AT) critical criterion 134, 141, 151
actual relative weights 98
additive function 228
additive utility assumption 6-7
AHP axioms 215
AHP variant by Dyer 222-223
all possible comparisons 105-112
alternatives 1
analytic hierarchy process (AHP) 9-11, 116, 137-141, 157-160,
165-166, 179-182, 186-188, 197-199,
201-212, 215-217, 225-227, 241
attributes 2
available comparisons 105-112
average consistency index of CDP matrices 41-42
average squared residual 65
B
best scales 51,53
best MCDM method 199
best fuzzy MCDM method 261-262
bound of maximum consistency 40
"Bridge Evaluation" problem 230-231
C
case studies 230-232
common characteristic 74
common comparisons 105-112
complete pairwise comparisons 88
conflict among criteria 1
concordance index 14, see also TOPSIS (fuzzy and crisp)
concordance matrix 16, see also TOPSIS (fuzzy and crisp)
concordance set 16,
see also under TOPSIS (fuzzy and crisp)
consistency see perfect consistency
consistency index (CI) 59, 67-71

consistency of CDP matrices 35-42
consistency ratio (CR) 59
consistent data 91
"city-block-metric" 74
class 1 scales 45-47
class 2 scales 45,48-49
Closest Discrete Pairwise (CDP) matrix 32-43,
80, 98, 145, 202, 251
closest value 34
criteria 2
critical alternatives (most) 156-175
criticality degree (of a decision criterion) 136
criticality degree (of an alternative) 156
criticality degree (of measures of performance) 160, 164
critical measure of performance (most) 155-175
Criterium Decision Plus (computer software) 201, 212
D
decision criteria 1
decision making paradox see paradox
decision matrix 2-3
decision space 1
decision support tool 207
decision weights 2
decomposition of judgment matrices 88-113
definiteness property 74
descriptive theories 265
deterministic MCDM 2
difference comparisons 74-86
difference judgments 76-86
dimensionless analysis 8
discordance index 14, see also TOPSIS (fuzzy and crisp)
discordance matrix 16-17, see also TOPSIS (fuzzy and crisp)
discordance set 16, see also TOPSIS (fuzzy and crisp)
discrete decision spaces 1
dissimilarity relation 74
distance (of similarity) 74
dominance matrices 17-18, see also TOPSIS (fuzzy and crisp)
duality approach 115-129
dual problem see duality approach
Subject Index 277
E
eigenvalue approach 44, 58-59, 217
eigenvector see eigenvalue approach
eigenvector approximation 58-59
see eigenvalue approach
eigenvector method see eigenvalue approach
ELECTRE method 13-18, 241
error terms 93
evaluative criteria (crisp) 43, 177-199
evaluative criteria (fuzzy) 250-262
Euclidean distance 19-20
examples of exponential scales 29-32
Expert Choice (computer software) 132, 201, 212
exponential scales 24,28-32
H
hierarchic composition 214
hierarchies (multiple) 131
hierarchical structure 1
human rationality assumption 61-71
F
feasible solution 85
feasible value 137, 163
''flat maxima principle" 131
fuzzy AHP 245-247, 262
fuzzy alternatives 236, 239, 242-262
fuzzy decision criteria 236, 242-262
fuzzy CDP matrix 251-257
fuzzy data 235
fuzzy databases 87
fuzzy decision matrix 243, 254-256
fuzzy evaluative criteria see evaluative criteria
fuzzy operations 236-329
fuzzy MCDM 2, 241-262
fuzzy numbers 166, 236-237, 239, 241
fuzzy RCP matrix 251-257
fuzzy reciprocal (judgment) matrix 253-260
fuzzy revised AHP 247-248, 258, 262
fuzzy sets 57,87

fuzzy TOPSIS 248-250, 262
fuzzy trapezoid numbers 237
fuzzy triangular numbers see fuzzy numbers
fuzzy WSM 242-244, 262
fuzzy WPM 244-245, 262
G
goals 1-2
group decision making 2
guided (pairwise comparisons) 88
Q
quadratic problem see quadratic programming
quadratic programming 75,79-85
I
ideal mode AHP see, revised AHP
ideal solution 20
identical alternatives 12, 214
incommensurable units 2
inconsistent CDP matrix 37
InfoHarvest, Inc. 201
L
Langrangian multipliers 81
Langrangian (the) 81
large size decision problems 128
law of stimulus of measurable magnitude 26
law of stimulus perception 28
least squares 63-67
linguistic choices 24,28
linear equation, system of 79,82
linear programming 92-97, 112
linear scale 24, see also Saaty scale
Lootsma scales see exponential scales
logical contradiction 222
Subject Index 279
M
matrix partitioning see decomposition
matrix transpose 82
maximum eigenvalue see eigenvalue approach
maximum similarity 73-74
maximum dissimilarity 73-74
maximum consistency of CDM matrices 38-42
maximum consistency index 40-42
membership value 57, also see fuzzy numbers
modal value 237
most critical criterion 133-135, 138-139, 141
most important criterion 144
most sensitive alternative 155-156
missing pairwise comparisons 86,91
Multi-Attribute Decision Making (MADM) 1
multi-attribute utility theory (MAUT) 214
Multi-Criteria Decision Making (MCDM) (definitions) 1-22
multi-dimensional MCDM 8
Multi-Objective Decision Making (MODM) 1
multiple attributes 1
multiple hierarchies 131-132
multiple objective functions 1
multiplicative AHP 228-233, see also WPM
N
negative-ideal solution 20
normalized decision matrix 14-16, 19
normalized columns 117-118
normalized rows 117-118
normative theories 265
number of alternatives (role of) 207
o
optimization approaches 60-67
optimal solution 83-85
outranking relations 13-14
p
pairwise comparisons 23, 25-32
paradox 2, 42, 145, 197-199, 265
partial ranking 221
partitioning of pairwise comparisons 90
see also decomposition of judgment matrices
percent-any (PA) critical criterion 134-135, 140-141, 145, 150
percent-top (PT) critical criterion 134-135, 140-141, 145, 149
perfect consistency 95, 215, 218
performance values see decision matrix
power law 32
power method 67
prescriptive theories 265
prime approach see duality approach
prime problem see duality approach
primal approach see duality approach
R
random consistency index (RCI) 59, see also consistency index
ranking abnormalities 213-233
ranking of fuzzy numbers 238-239
ranking reversal 43
ranking indiscrimination 43
ratio comparisons 57-72
Real Continuous Pairwise (RCP) matrix 32-34, 80, 97, 146,
202-205, 251
real life case studies 230-232
reciprocal comparisons see pairwise comparisons
reciprocal matrices 88-89, see also pairwise comparisons
redundant constraints 96
relative closeness 20
relative importance 75,85
relative priorities 89, see also relative importance
relative magnitudes 25
relative ranking 80
relative similarity 75-85, also, see similarity
relative terms 132
relative weights 57-72, 73-86, 87-113
revised AHP 11-13, 116, 189, 191-193, 197-199,
201-212, 214, 218-222, 225-227
Subject Index 281
robust criterion 138

robust decision problem 164
S
Saaty scale 26-27
scale 24-55,75
scale generation 27
scale evaluation 32-55
separation measure 20
SENSA ITO library 131
sensitivity analysis 131-175
sensitivity coefficient (of a criterion) 136
sensitivity coefficient (of an alternative) 156
sensitivity coefficient (of measures of performance) 160, 164
similarity function 73
similarity scale 77
similarity measure 74
single-dimensional MCDM 8
"Site Selection" problem 232
stepwise approach 166
stochastic MCDM 2
stimulus of measurable magnitude 26
stimulus perception 28
symmetric comparisons see difference judgments
symmetric comparisons 76
system of linear equations see quadratic programming
T
taxonomy of MCDM methods 4
threshold value(s) 155, 158-159, 161, 163, 175
TOPSIS method 18-21, 194-196, 241
triangular property 74,85
V
"vector-maximum" problem 1
U
union operation 87
units of measure 8
upper bound of reduction rate 120
unrestricted variables 93
utility theory 214
W
weighted product model (WPM) 8-9, 142-145, 154, 161-164, 165,
169-170, 174-175, 180, 183, 185,
192, 197-199, 228-233, 241
weighted sum model (WSM) 6-7, 137-141, 157-160, 165, 167-168,
170-171, 179-185, 241, 197-199
worst scales 52-54
AUTHOR INDEX
A
Ackoff, R. L. 23,25
Allais, M. 265
Anderberg, M.R. 74
Arbel, A. 201
Armacost, R. 313
Amoff 25
B
Baas, S.J. 238
Barzilai, J. 228,233
Bazerman, M.H. 265
Bell, D.E. 265
Bellman, R.E. 235
Belton, V. 5, 11, 12,53, 116201,212,
213, 236, 247
Benayoun, R. 13, 14
Benhajla, S. 201
Berdosian, S.D. 235
Boender, e.G.E. 236, 237, 242, 245
Bonissone, P.P. 238
Boucher, T.O. 201
Bridgman, P. W. 8
Buckley, J.J. 236, 237, 238, 241
C
Cambron, K.E. 201
Camerer, C. 214
Chang, S.K. 235
Chen, S.J. 1, 3, 4, 233
Chu, A.T.W. 32,57,60,64
Churchman 25
D
Dantzig, G. 165
Dehnad, K. 235
Dodd, F.J. 41
Donegan, H.A. 41
Dubois, D. 57, 74, 235, 236, 237
Dyer, J.S. 212, 214, 222
E
Ellsberg, D. 265
Evans, G.W. 201
F
Falkner, C. 201
Federov, V.V. 32, 57, 60, 64, 66
Finnie, J. 201
Fishburn, P. C. 6,214
French, S. 131, 265
Fu, K.S. 235
G
Gear, T. 5, 11, 12, 53, 116, 201,
212, 213, 247
Grize, Y.L. 235
Gupta, M.M. 57,235
Gutsche 264
H
Haerer, W. 201
Hagen, O. 265
Hallowell Davis, M.D. 32
Harker, P.T. 88,214
Hedge, G.G. 232
Hihn, J.M. 32
Hogarth, R.M. 214
Hosseini, J.C. 132
Hwang, F. 3
Hwang, C.-L. 1, 3, 4, 18, 233, 236
Author Index 285
J
Johnson, C.R 32
Insua, RD. 131
K
Kadane, J.B. 265
Kalaba, R.E. 66
Kaufmann, A. 74
Khurgin, J.1. 32
Kuhn H.W. 1
Kwakernaak, H. 238
L
Laarhoven, PJ.M. 236, 237, 239, 241, 242, 245
Lakoff, G. 57
Larkey, P.D. 265
LaValle, I.H. 214
Lee, C.T.R. 235
Lee, E.S. 238
Lee, S.N. 235
Lootsma, F.A. 14, 23, 24, 25, 28, 29,
32, 45, 146, 228, 233, 236, 239, 241, 261
Lin, C.-T. 166,241,267,261
M
Ma, D. 26
Mann, S.H. 13, 43, 44, 53, 57, 60, 145,
177, 194,201,207,258
Mardia, K. V. 74
Masuda, T. 131
Marks, L.E. 32
McCahon, C. S. 238
McMillan, J. 265
McStravic, E.L. 201
Michon, J .A. 32
Miller, C.A. 26,76
Miller, D.W. 8,233
Munier, B.R 214
N
Neale, M.A. 265
Negoita C.V. 235
o
O'Keefe, W.D. 201
p
Pardalos, P.M. 57, 61, 145
Pedrycz, W. 236, 237, 245
Polyakov, V. V. 32
Prade, H. 57, 74, 235, 236, 237
Pratt, J.W. 265
Putrus, R 201
R
Raiffa, H. 265
Ramsay, A. 235
Ragade, RK. 235
Ravinder, H. V. 212, 214
Raz, T. 201
Reder, M.W. 214
Roberts, F.S. 28,32
Roper-Lowe, G.C. 201
Roy, B. 13
Ruspini, E.H. 73, 77, 85
S
Saaty, T.L. 5, 9, 10, 23, 24, 25, 26, 27, 32,
40, 57, 58, 66, 89, 117, 201, 214, 217, 231
Sanchez, A. 131, 132, 133, 165
Sarin, RK. 214
Schlaifer, R. O. 265
Sebenius, J.K. 265
Seidmann, A. 201
Sharp, J.A. 201
Shimura, M. 235
Author Index 287
Sokal, RR. 74
Sneath, P.R.A. 74
Starr, M.K. 8,233
Stevens, S.S. 32
Stewart, S.M. 63
Swann, K. 201
T
Tadikamalla, P.R. 232
Tate, M.D. 61
Tong, RM. 238
Triantaphyllou, E. 13, 23, 32, 33, 41, 44, 53,
57,61,73,77, 115, 116, 118, 120, 131,
132, 133, 145, 146, 165, 166, 177, 184,201,
207, 213, 228, 232, 233, 241, 257, 258, 261
Tucker, A.W. 1
V
Vargas, L.G. 32,59,214
X
Xie, L.A. 235
y
Yager, R.Y. 235
Yoon, K. 18, 236
W
Wabalickis, RN. 201
Wang. L. 201
Weber, A. 26,76
Weber, M. 214
Wendell, W.E. 131, 212, 214
Winkler, RL. 212, 214
Winston, W. L. 81, 131
Write, C. 61
Z
Zadeh, L.A. 235
Zheng, X. 26
Zhu, Q. 238
Zimmermann, H.-J. Foreword, 1,2,264
Zwicker, E. 32
ABOUT THE AUTHOR
Dr. Triantaphyllou did his graduate studies at Penn State University

from 1984 to 1990. While at Penn State, he earned a Dual M.S. degree in
Environment and Operations Research, an M.S. degree in Computer Science
and a Dual Ph.D. degree in Industrial Engineering and Operations Research.
Currently, he is an Associate Professor in the Industrial and Manufacturing
Systems Engineering Department at the Louisiana State University in Baton
Rouge, LA, USA.
His research is focused in decision making theory and applications,
data mining and knowledge discovery, and the interface of operations research
and computer science. Since the years he was a graduate student, he has
developed new methods for data mining and knowledge discovery and also
has explored some of the most fundamental and intriguing subjects in decision
making. In 1999 he has received the prestigious lIE (Institute of Industrial
Engineers), OR Division, Research Award for his research contributions in
the above fields. Some of his graduate students have also received national
awards and distinctions.
He always enjoys sharing the results of his research with his students
and is also getting them actively involved in his research activities. He has
received teaching awards and distinctions. His research has been funded by
federal and state agencies, and the private sector. He has extensively
published in some of the top refereed journals and made numerous
presentations in national and international conferences.
Dr. Triantaphyllou has a strong inter-disciplinary background. He has
always enjoyed organizing multi-disciplinary teams of researchers and
practitioners with complementary expertise. These groups try to
comprehensively attack some of the most urgent problems in the sciences and
engineering. He is a strong believer of the premise that the next round of
major scientific and engineering discoveries will come from the work of such
inter-disciplinary groups. More details of his work can be found in his web
site (http://www.imse.isu.edu/vangeiis/).

(Applied Optimization 44) Evangelos Triantaphyllou (Auth.) - Multi-Criteria Decision Making Methods - A Comparative Study-Springer US (2000)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Applied Optimization 44) Evangelos Triantaphyllou (Auth.) - Multi-Criteria Decision Making Methods - A Comparative Study-Springer US (2000)

Uploaded by

Copyright:

Available Formats

Multi-Criteria Decision Making Methods:

SPRINGER-SCIENCE+BUSINESS MEDIA B.V.

ISBN 978-1-4419-4838-0 ISBN 978-1-4757-3157-6 (eBook)

Printed an acid-free paper

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. xiii

1 Introduction to Multi-Criteria Decision Making . . . . . . . •• 1

2 Multi-Criteria Decision Making Methods .••.......... 5

3 Quantification of Qualitative Data for

4 Deriving Relative Weights from Ratio Comparisons ...... 57

5 Deriving Relative Weights from Difference Comparisons . .. 73

6 A Decomposition Approach for Evaluating Relative

7 Reduction of Pairwise Comparisons Via a

7.3.2 Applying the Dual Approach . . . . . . . . . . . 122

8 A Sensitivity Analysis Approach for MCDM Methods •••• 131

Appendix to Chapter 8 . . . . . . . • . . . . . . . . . . . . . . .. 167

9 Evaluation of Methods for Processing a

10 A Computational Evaluation of the Original

11 More Cases of Ranking Abnormalities When Some

11.4 Some Computational Results . . . . . . . . . . . . . . . . 223

12 Fuzzy Sets and Their Operations ..•.......••••... 235

13 Fuzzy Multi-Criteria Decision Making • • . . . . . . . . • . . . 241

14 Conclusions and Discussion for Future Research .....•• 263

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 275

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

About the Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289

1 Introduction to Multi-Criteria Decision Making • • .. 1

2 Multi-Criteria Decision Making Methods ••••••.• 5

3 Quantification of Qualitative Data for

4 Deriving Relative Weights from Ratio Comparisons . 57

5 Deriving Relative Weights from Difference

6 A Decomposition Approach for Evaluating Relative

7 Reduction of Pairwise Comparisons Via a

8 A Sensitivity Analysis Approach

9 Evaluation of Methods for Processing a

WSM and the AHP . . . . . . . . . . . . . . . . . . . . .. 184

10 A Computational Evaluation of the Original

11 More Cases of Ranking Abnormalities When Some

Any Alternative When Alternatives are

12 Fuzzy Sets and Their Operations ............ 235

13 Fuzzy Multi-Criteria Decision Making . . . . . . . . . 241

14 Conclusions and Discussion for Future Research .. 263

1 Introduction to Multi-Criteria Decision Making . . .. 1

2 Multi-Criteria Decision Making Methods .•.••••. 5

3 Quantification of Qualitative Data for

4 Deriving Relative Weights from Ratio

5 Deriving Relative Weights from Difference

6 A Decomposition Approach for Evaluating Relative

7 Reduction of Pairwise Comparisons Via a

8 A Sensitivity Analysis Approach

9 Evaluation of Methods for Processing a

Table 9-2: Contradiction Rate (%) Between the

Table 9-3: Contradiction Rate (%) Between the

10 A Computational Evaluation of the Original

11 More Ranking Abnormalities When Some

12 Fuzzy Sets and Their Operations . . . • . . . . . . . . 235

13 Fuzzy Multi-Criteria Decision Making . . . . . . . . . 241

14 Conclusions and Discussion for Future Research .. 263

Multi-Criteria Decision Making (MCDM) has been one of the fastest

Aachen, Germany, April 2000

Probably the most perpetual intellectual challenge in science and

feel uncomfortable dealing with them.

solution of hundreds of thousands or millions of simulated test MCDM

ways for processing a decision matrix. Chapter ten presents a computational

= { VI, V 2, ... , vn* }. (2-12)