Professional Documents
Culture Documents
Readings in Multiple Criteria Decision Aid
Readings in Multiple Criteria Decision Aid
Readings in Multiple Criteria Decision Aid
Readings in
Multiple Criteria
Decision Aid
With 124 Figures
Springer-Verlag
Berlin Heidelberg N ew York
London Paris Tokyo
Hong Kong Barcelona
Carlos A. Bana e Costa
Technical University of Lisbon (1ST)
IST-CESUR
Av. Rovisco Pais
1000 Lisbon, Portugal
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically
the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in other
ways, and storage in data banks. Duplication of this publication or parts thereof is only permitted under the provisions of the
German Copyright Law ofSeptember9, 1965, in its version of June 24,1985, and a copyright fee must always be paid. Violati-
ons fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin· Heidelberg 1990
Softcover reprint of the hardcover 1st edition 1990
The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement,
that such names are exempt from the relevant protective laws and regulations and therefore free for general use.
214217130-543210
FOREWORD
ACKNOWLEDGEMENTS
INTRODUCTORY CHAPTER
CHAPTER
MODELLING DECISION SITUATIONS
Bernard Roy
DECISION-AID AND DECISION-MAKING 17
Anna Ostanello
ACTION EVALUATION AND ACTION STRUCTURING: DIFFERENT
DECISION AID SITUATIONS REVIEWED THROUGH TWO ACTUAL CASES 36
Denis BOuyssou
BUILDING CRITERIA: A PREREQUISITE FOR MCDA 58
Jean-Claude Vansnick
MEASUREMENT THEORY AND DECISION AID 81
Philippe Vincke
BASIC CONCEPTS OF PREFERENCE MODELLING 101
Hans-Jurgen Zimmermann
DECISION MAKING IN ILL-STRUCTURED ENVIRONMENTS AND WITH
MULTIPLE CRITERIA 119
x
CHAPTER II
THE OUTRANKING APPROACH
Bernard Roy
THE OUTRANKING APPROACH AND THE FOUNDATIONS OF ELECTRE
METHODS 155
Daniel Vanderpooten
THE CONSTRUCTION OF PRESCRIPTIONS IN OUTRANKING METHODS 184
Benedetto Matarazzo
A PAIRWISE CRITERION COMPARISON APPROACH: THE MAPPAC AND
PRAGMA METHODS 253
CHAPTER III
VALUE AND UTILITY THEORY APPROACH
Rakesh K. Sarin
CONJOINT MEASUREMENT: THEORY AND METHODS 277
Ernest H. Forman
MULTI CRITERIA DECISION MAKING AND THE ANALYTIC
HIERARCHY PROCESS 295
Eric Jacquet-Lagreze
INTERACTIVE ASSESSMENT OF PREFERENCES USING HOLISTIC
JUDGMENTS: THE PREFCALC SYSTEM 335
CHAPTER IV
INTERACTIVE MULTIPLE OBJECTIVE PROGRAMMING
Pekka Korhonen
THE MULTIOBJECTIVE LINEAR PROGRAMMING DECISION SUPPORT
SYSTEM VIG AND ITS APPLICATIONS 471
Jaap Spronk
INTERACTIVE MULTIFACTORIAL PLANNING: STATE OF THE ART 512
XII
CHAPTER V
GROUP DECISION AND NEGOTIATION
Gunter Fandel
GROUP DECISION MAKING: METHODOLOGY AND APPLICATIONS 569
CHAPTER VI
THE SCHOOL CASE-STUDY
Marc Pirlot
A REPORT ON THE STUDY OF THE PORTUGUESE NAVY CASE 648
INTRODUCTORY CHAPTER
MULTIPLE CRITERIA DECISION AID:
AN OVERVIEW
development and other kind acti vi ties and thus to control the
modelling process evolution within the decision aid process".
Carlos Bana e Costa remarks also that "those familiar with the
practice of decision aid shall recognise that ( ... ) the question-
ing procedures appropriate to specify the weights of the criteria
are intrinsically complex and hard to tackle. This fact by itself
can create uncertainty and imprecision and/or may cause the
interruption of the process close to initial stages". So, as we
have been seeing, the phase of modelling inter-attribute prefer-
ences is crucial for the success of MCDA. This topic is also
discussed in the paper by Bana e Costa. He remarks that "si tua-
tions involving poor weighting information are very common in the
practice of decision aid. Nevertheless, there exists a signifi-
cant lack of operational MCDA approaches explicitly devoted to
support decision making under those circumstances. As a matter of
fact, traditional multicriteria methods only operate with stable
and single weights". Also in the context of MAUT, the Outweigh
Approach proposed by Bana e Costa is a contribution towards the
development of procedures able of directly facing those si tua-
tions of poor preferential information of three main types: only
some conditions relating the weighting constants are known, a
rank-order of the weights is known and lower and upper preference
bounds for value tradeoffs, or for substitution rates, or for
weights are known.
Group decison making and negotiation support are more and more
studied in the literature and they constitute a field having
strong links with multicriteria decision aid. The paper by Tawfik
Jelassi, Gregory Kersten and Stanley Zionts provides "an over-
view of formal models for group decision making and negotiation
wi th a special focus on those which can be used in developing
computer-based support systems." First they distinguish four types
of multi-person decision-making situations: individual decision-
making in a group setting; hierarchical or bureaucratic decision-
making (centralized and decentralized); group decision-making or
one-party decision-making; and multi-party decision-making.
14
Bernard Roy
ABSTRACT
o. INTRODUCTION
Within the analytic form of A, gk(a) = gk(xl, ... ,>Gn) and then:
U ( a) = U ( xl , ... , >Gn) .
Table 1
n
(1) General additive Uta) '" E wj[gj(a)]
form j=l
n
(2) Weighted sum Uta) E k j . gj (a) with k j 2: O.
j=l
(3) General form of uta) '" E u(Yl"" ,Yn) .8a (Ylr··· ,Yn )
expected utility Yl""'Y n
value
where u(Ylr' .. ,Yn) is a multi-attribute
utility function.
E Uj (Yj) . 83 (Yj )
Yj
n
(5) Utility IT [1+k.k j .gj (a) ]-1
mu I tip I i cat i v e j=1
form Uta) kj2:0, k#O, k>-l
k
n
with E k j #l, gj (a) E Uj (Yj ) . 83 (Yj )
j=l Yj
For this purpose, each criterion must take into account one or
more precise attributes (or consequences) allowing us to clearly
apprehend a well-identified point of view (see 2.2 a). The con-
sensus deals with the way of defining, and consequently comput-
ing, the performances gk(a). On the other hand, it does not deal
with the importance which should be given to each of these crite-
ria in order to carry out the comparisons, as the value systems
on which different actors base their opinions on the topic may
not be the same.
- the family F,
- additional inter-criteria information (weights, veto thresh-
olds, substitution rates, ... ),
Here, a major problem comes from the fact that each criterion
has to playa specific role according to its own importance. It
is essential to bear in mind that this notion of importance has
no absolute meaning. To be meaningful, the importance of a crite-
rion must refer to a given aggregation model.
4. CONCLUSION
REFERENCES
Anna Ostanello
1. INTRODUCTION
For what concerns the presented cases, they have been both
conducted by the author and her collaborators (Simoni and Ver-
noni, in 1978; Norese, in 1982-1983); for each of them, the story
and most of the activities had been registered during the proc-
ess, memorized and verified.
2. FRAME MODEL
The frame model does not assume a priori phases in the proc-
ess. It assumes five different Routines - empirically derived
as basic steps, leaving a phase recognition as a result of the
process reconstruction (cf also: Witte, 1972; Mintzberg et al.,
1976). Some combinations of the Routines, named Development
Procedures, - also empirically identified - help in giving a more
structured representation of the DA process.
2.1. Routines
(..)
to
40
2.2. D-procedures
INFORMATION · It! It ip Ie sources It! It ip les sources, • One or few IMultiples sources,
SOURCE both i nterna 1 and mainly internal coperat ing sources !mainly internal,
.Unicity.: one or One or more sources external showing divergences, normally cooperating,
more cooperat ing genera lly cooperat i ng, · Not necessari ly redundances, conf 1icts bringing different
sources with possible cooperat ing :elements that can vary
cofl icting points • Conflicting points from structured data
of view of view to partial analyses,
i nf onna 1 representat.,
verba 1 i nd icat ions, •••
STARTING Deve loped, stab Ie • Partially developed • Developed • Partially developed • Totally undeve loped • Tota lly undeve loped
STATE OF • Unstable • Stable • Unstable
THE MODEL
.SHAPE- BY
THE ROUTI NES I R3 / R4 II ~ R4/R2 - > R3/R5 ~ I~R2<~R5~- II R4 <-> R3 i I III R4 <-> R3 !- I R4 <-> R3 !
I
.j:>.
-"
42
3. DA PROCESS PROFILES
STEPS
Identifie. I III
Struetur.
Developm. (no C's involvement) (with organization
Control involvement)
Comunic.
i==A and C
C/I II IV
cis
C/Co (C's involvement) (with both C's and
C/Co organization
C/D involvement)
CLIENT'S CLIENT'S
ORGANIZATION POSITION
ORGANIZATION
PROCESS
.•=- --~
:ksr~~¥~}~,: ,
::;:::'''1 ~:;~' j
rASE 1 Regional Head Relocation Decision/ • RegiOnal1
L
Administrat. Service of firms at
of a regiona 1
Regiona 1 level relocation of firm
Study area owners
Centre Strategica 1 with a
strong
D
leader
r- ---- - -- _..
-~-=~ i=--="~~-l
--=--~~
Case 2)
Dissatisfaction with the present decision routine,
essentially based on inter-personal relationships and heavily
conditioned by political criteria. Intention of elaborating some
action to change this situation.
fJ<_>_3
r(i) m(e) i(ma) e(d)
R2 III
R4
R3
R1
R2 IV
R4
R3 6< 5<
47
CENTRAL STEPS -
a) In cases of «participating Client», they are
«Development» activities, mainly conducted at collec-
tive level; complexity arises in the case of organiza-
tion involvement (see figure 3);
b) In cases of «non-participating Client», «Structur-
ing» at individual level is heavily based on modelling
activities (R4), as i(ss), i(ma), e(ma), iterating
with «Communication» for «Structuring and Control» in
the organization environment.
m(e) i(ma) R2 R3 R5
4 _ r-I-----=-~-2--""-3---9
,Structur.
[1 161
1- --1t>..J
--;0 15 L.<--I
Dcvelopm.
Control l :] L1 IL=:20 9
4.4.1. Case 1
1) Structuring stage:
r"1--->'
2 3 Consequence analysis:
L..< I
4) Criteria modelling:
5 -> 6) S i mu I a t ion:
4.4.2. Case 2
PROBLEMS --->--"----:>---..I.
. COGN ITI VE
. ORGANIZATIONAL ~
AND POll TI CAL
PROBLEMS - - - - - - - -
I OFFICIAl SOURCES,
0 • Indication for Theoret i ca I Models of:
E the elaboration I-> model of the -> · Demand (D)
N of a fOnD for activity • Supply (S)
T technical spaces • D/S
I requ irements
F • Indication for
I the elaboration Census fOnD Data contro I
C of a fOnD for I-> for public -> and
A the management and private integration
T of facilities facil ities with
I · Data on the competent
0 cons i stency sources
N of facil ities f-> Inquiry on ->
• Data on actua I users
users behaviour
OFFICIAl SOURCES,
S • Not homogeneous Elaboration Elaboration
T data, without -> of a final f-> of a shared
R a comnon code fOnD to model:
U • Private and/or co llect data · Data control
C inaccessible on and
T data cons i stency integration
U · Non usable and managem. · Data
R indications of public trans format .
I (uncomp leted, and private • Control on
N diverting) facilities partial
G representat.
• modelling of
Data criteria
gathering >
CONCLUSIVE CONSIDERATIONS
REFERENCES
Denis BOuyssou
1. INTRODUCTION
2_ WHAT IS A CRITERION?
1. Though these two terms are not synonymous (an axis is the
operational counterpart of a given point of view (see Roy, 1985))
we will use them interchangeably in this paper.
60
ii- Once a point of view has been defined and accepted, the
method allowing to arrive at the evaluation on the criterion for
each alternative, should also be understood and accepted by all
the actors of the decision process. The search for a simple and
transparent model of evaluation should therefore be an important
preoccupation of the analyst. Furthermore this method should be
as free as possible from elements deeply linked to a particular
value system. In fact, the presence of such elements may well
lead some actors to question or even reject the validity of the
comparisons made on the basis of the criterion.
for all a, bE A:
3. AN EXAMPLE
n
g(s) l: Wihj{ s) ,
i=l
n
g(s) l: Wihj{ s) (1+a.;)d,
i=l
n
g(s) l: Wih;(s) (1+a.;(k) )d] P(Ekl
k=l i=l
where there are n zones surrounding each site wi being the weight
affected to the inhabitants of the ith zone, hi(s) being the
number of people living in the ith zone of site s in 1989 (possi-
bly corrected in order to take into account "privileged" inhabi-
tants), a.i being the annual increase rate of that population for
the next d years, a.i(k) being the same rate in scenario Ek having
a probability P(E k), IDs scenarii being considered for the site s.
Let us note that one may want to include in these formulas var-
ious scenarii concerning the orientation of the runaways or the
possible date of installation of the airport.
g(a) = c(a),
c(a) being the evaluation of alternative a on the consequence.
g(a) c(a).
g(a) = c(a),
p(g(a» [a+a' + (f3 + f3')g(a») / (1 W),
q (g (a » = Min [a + f3g ( a ); ( a' + f3' g ( a » / (1 - W»).
g(a) (1 )
n
g(a) l: fa(k)/(l+i)k,
k=l
the expected profit of the latter (1 000 000) is far greater than
that of the former. In order to build a criterion taking into
account such a preference, it suffices to choose u such that:
2. This is all the more true that there are many controversies
concerning the normative, descriptive and prescriptive virtues of
Expected Utility. For a recent overview of the debates, see,
e.g., Munier (1988).
74
5. CONCLUSION
1. Sometimes the analyst is only hired for these two phases, see,
e.g. Grassin (1986).
REFERENCES
Jean-Claude Vansnick
University of Mons-Hainaut
Place Warocque, 17
B - 7000 Mons BELGIUM
1. INTRODUCTION
For those who wish to get further into the field of measure-
ment we highly recommend the book of F.S. Roberts (1979).
XI 1-----'--.
The graphical representation of P is given in figure 1.
X2
X3 .~--------__~---------. X4
Figure 1
which is impossible.
1) P is transitive
2) the relation I defined by [Vx t y E X: x I y iff x F y and
y F x] is transitive.
Remarks
Let us point out that the way of using the numbers ~ (x) in
order to represent the information considered would normally be
included in the definition of the notion of scale of measurement.
86
Figure 2
10 if x xl
8 if X x2
0 if X x3
8 if X x4
This is the reason why we will say that the above statement is
not meaningful; it is meaningless.
x ~(1:ol1)(x) 1:(l1(x)).
X + information on X
R
11
(represent i ng the
information
given on Xl
Figure 3
89
Example:
1:: Jl(X) -f R
We have
23 8 if x xl
21 2 if x = x2
1:0Jl X -f ~ X -f(1:oJl)(x)
2° 1 if x x3
21 2 if X x4
1: : Jl(x) -f II
Example:
Proof:
Jl(X3) + Jl(x4)
(3 + 4) / 2
2 7/21
and 7/2 > 6/2
~(xll + ~(xs)
(1 + 5) / 2 6/2
2
1:(Il(X3)) + 1:(Il(x4))
2 12 )
and 12 < 17.
1:(Il(xIl) + 1:(Il(xs))
17
2
a.(3 + 4) + 2.13
= 7/2.a + 13
2
a.(l + 6) + 2.13
= 6/2.a + 13
2
and 7/2.a + 13 > 6/2.a + 13 since a > O.
Example:
is not uniquely determined (we can take for ~(xl) any real
number greater than 1).
94
Rema rks :
Now, we say a few words about the notion of ratio scale. Such
a scale is characterized by admissible transformations of the
form ~(r) = a.r (a > 0). For each a > 0, such a transformation
has one fixed point: r = O. This shows that it is only conceiva-
ble to obtain a ratio scale when the measurement concerns a no-
tion for which there is a natural zero. In physics, mass defines
a ratio scale as null mass is a natural zero and the only degree
of freedom is the choice of the unit (gram, kilogramme, ton, ... ).
In decision aid, the notion of difference of preference can lead
to a ratio scale on {(x, y) E X x X I x P y) because speaking of
null difference of preference makes sense. It is more difficult
to conceive the existence of a ratio scale on X (in the framework
of decision aid) but we think that it is possible to obtain such
a scale from the idea of algebraic «attractiveness» of the
elements of X (negative attractiveness repulsiveness).
A A A
,..,
IfI
....a.
.... a.:I
C Q.CI
a. c·...
"
E a."
a. ....
" a. ....
:I CootB
IfI a.
a. "a.
tB
IW:
E ....
Coo
Coo QCoo
Q Q
:JJ
IfI .... a.
...tBa. "Q
a.C
!:loa.
Q Q'CI
CI)
"C
!:loa.
.... !:Io
a.
'CI
C
....
'II'
f;I;l
a:
...=
CI
~
98
GJD x w v
11 ( . )
~D 2 1
U
The scale
11'(·) =
REFERENCES
Philippe Vincke
1. INTRODUCTION
2. PREFERENCE STRUCTURE
We denote
a P b i f a is preferred to b,
b P a i f b is preferred to a,
a I b i f a and b are indifferent,
a ? b i f a and b are incomparable.
102
The pre fer en c e ( P), i n d iff ere n c e ( I) and inc omp a r a b iii t y (?)
relations are the sets of couples (a, b) such that, respectively,
a P b, a I b and a ? b. These relations are used in most of the
works devoted to preference modelling.
a P b => b F a P is asymmetric,
a I a I is reflexive,
a I b => b I a I is symmetric,
a '1 a ? is irreflexive,
a ? b => b ? a ? is symmetric.
a S b iff a P b or a I b (S P U I).
We have then
a P b iff a S b and b S a,
a I b iff a S b and b S a,
a ? b iff a S b and b S a.
a. .. .b a. .b a. .b
a P b a I b a ? b
Va, b, c E A:
'1 b no incomparability,
\: P b and b P c => a P c
I b and b I c => a I c
P is transitive,
I is transitive.
Remarks
a l b : no incomparability,
a P b, b I c,c P d => a P d,
a P b, b P c, a I d => d P c.
Remarks
4) The
DOD
semiorder structure has a very particular matrix
representation, leading to the interesting concept of
«minimal representation» (Pirlot, 1990) which generalizes
the notion of rank usually associated to the complete
order.
ja
a P b iff g(a) > g(b) + q(g(b»,
I b iff g(a) ~ g(b) + q(g(b», «variable threshold
g(b) ~ g(a) + q(g(a». model"
(i)
If the function q satisfies the following «coherence condi-
tion»: "la, bEA:
(ii)
If the function q does not verify the coherence condition
then the preference structure implied by the variable threshold
model has to satisfy the following conditions: "I a, b, c, d E A:
a ? b : no incomparability,
{
P b, b I c, c P d => a P d.
Remarks
1) P is still transitive.
and
4) The
DO
interval order structure also has a particular matrix
representation (see Roubens and Vincke, 1985).
109
(i)
If the preference only arises when the intervals are disjoint,
i.e. if
a P b iff xa > y b ,
{
I b iff [xa' Ya] n [xb' Yb] '" <I>
then, putting, V a E A:
xa = g (a) ,
{
Ya = g(a) + q(g(a)),
(ii )
In the case where
l
a P b iff xa > Yb'
a Q b iff Ya > Yb > xa > xb'
a I b iff one interval is included in the other.
S is reflexive,
:a : : and b S a => a = b S is antisymmetric,
\
S b, b S C => a S c S is transitive;
a P b, b P c => a P c P is transitive,
a I b, b I c => a I c I is transitive,
a P b, b I C => a P c,
a I b, b P c => a P c.
where U(x) and U(y) are the «values» of x and y. This means that
the comparison of probability distributions is reduced to the
comparison of numbers: this attitude sometimes leads to criticiz-
able conclusions (see Allais, 1953). Other concepts have been
introduced in this field (see Fishburn, 1982, Jaffray et al.
1982, Bouyssou, 1983).
16. REFERENCES
Hans-Jurgen Zimmermann
ABSTRACT
the final judgments are also represented by fuzzy sets which have
to be ordered to determine the optimal alternative. Then the
second stage is, of course, by far not trivial.
(1)
The rij as well as the Wj' however, will in many cases more
appropriately be modelled by fuzzy numbers. This has the follow-
ing consequences:
A. Hierarchical Aggregation
1. Hierarchical aggregation using crisp weights (Yager, 1978)
m
i . e. 115 (xi) min Il\'ij(xi), i=l, ... ,n,
j=l
Yager now allows for different importance of the goals and ex-
presses this by exponential weighting of the membership functions
of the goals. If Wj are the weights of the goals the weighted
membership functions, Il\'i, are
Other models along the same line have been suggested by Laar-
hoven and Pedrycz (1983). They concentrate on step 1 and suggest
Saaty's AHP modified for triangular fuzzy numbers.
125
Step 2: Ranking
For the final ranking of the xi Baas and Kwakernaak start
from the observation that if the xi had received crisp
rating, ri = R i , then a reasonable procedure would be to
select the xi that have received the highest rating,
i.e. to determine the set of preferred alternatives as
{xiEI ri>rj' \;tjEI}, I = {l, ... ,m}.
Since here the final ratings are fuzzy the problem is
somewhat more complicated. The authors suggest in their
model two different fuzzy sets in addition to Ri which
supply different kinds of information about the prefera-
bility of an alternative.
127
(7 )
Summarizing:
1- IlRi(r) is the fuzzy rating of Xi
2. Ili (x;) is the degree to which Xi is best alternative, and
3. IlPi(P) is the degree of preferability of Xi over all other
alternatives.
128
r:
with
n
IlM (r) (11 )
I 1
and
rmax sup r.
rES
with
f1 M
, , "
1
,, 1
,
,
,,
/
Figure 1
M
-i
'1
,, I
/
I
/
I
I
/ ,,'
,
/,,"
I"
/"
".I
,
,,"I
I
, ," /
,,
, "/
Figure 2
130
The authors are only concerned with step 2 of MADM, i.e. the
ranking of (normal convex) fuzzy sets, ri' which represent rat-
ings of decision alternatives, xi' and include already the aggre-
gation of the merits of an alternative with respect to different
criteria (goals).
The "ranking" of the final ratings, ri, i=l, ... ,n, of the
alternatives, xi' i=1, ... ,n is achieved in 2 steps:
for rj ::; r*
( 12 )
for rj ~ r*,
/L
1. 0+-----..,.. r-- .
/
.;" " \ • -r3
\
.5 \
\
x* x
Figure 3
n
l: v(xd .ri
i=l
irk
r
k
- (13 )
n
l: v(xd
i=l
ifk
n
maxf min lli'i(Xi) (14 )
rl.··· .rn i=l
1. Possibility of dominance
PD(r) Poss (ri ~ rj)
PD(ril sup min[J..lfj{rd, llfj(rj)] (15 )
ri,rj
ri~rj
3. Necessity of dominance
ND(rj) Nec (r; > rj)
inf (17)
r;
~ /
"" /'
//
.... / / //
I ;'
/'
/'
/'
_-
....
- - - - - -.ll.;1;:':-:':-:':-:":::~=-:-=-=-:=--vv7"i--=======.;lIIv;-;e:ft-;;otthhrreeStsh 01 d
~s.~ Maximum non-significant:
I gi (a)
I threshold
.
g. ( b)
I
tl-oo-----'---evaluation scalt' G j for ----_+101
I I
t i-th criterion 1
Figure 4
135
di (a,b) di (a,b)
t
!
Fuzzy Concordance
Criteria
r-+ Relation
weights
(a,b)
d(a,b)
Choic@ of a* E Adefined by
FIGURE 5
136
2.3.1. Comparisons
--r,
---- r 2
r r
Situation 1 Situation 2
: \
\
\
\
\
\
/ \
\
• I \
:/
r r
Situation 3 Situation 4
137
\
\
\
\
\
\
\
Situation 5 r Situation 6 r
Figure 6
a) b)
.'
.'
;:1
c) d)
I
-r I
31
I
I
1
I
1
,...-
l:
e) 1)
Figure 7
Case a b c d e f
Criterion
Generality
Discrimination
Sophistication ~
Information
. ----
Fuzzification
Requirement
-----------
Baas & Kwaker- fuzzy fuzzy sets max - min preference crisp
naak sets
Laarhiven &
BI fuzzy
I hierarchical
I fuzzy
I
Pedrycz triangular crisp aggregation ----- ratings
number
D I I I
Jain II fuzzy maximizing crisp
set
Max z = cx
s.th. Ax ~ b (20 )
x ~ 0
Bx ~ d, (22 )
x ~ o.
m+1
~ii(x) min ~j{x) (23)
i=l
m+1
max min IIi (x) (24)
x;<O i=l
1 i f (Bx) i ~ di
lid x) E [0,1] if di ~ (Bx) i ~ di+Pi, i=l, ... ,m+1 ( 25)
0 i f (Bx) i > di+Pi
1 i f (Bx)i ~ di
(Bx)i-di
IIi (x) 1 i f d.
1
~ (Bx) i ~ di+P i , i=l, ... ,m+1 (26 )
Pi
0 i f (Bx) i > di+Pi
Maximize >-
such that >-Pi+(Bx)i ~ di+Pi, i=l, ... ,m+1, (28 )
o ~ >- ~ 1,
x ~ O.
146
Maximize ~
Assuming that the decision maker has upper and lower bounds ci
and ~i for his aspirations, such that he does not accept solu-
tions for which cix is smaller than ~i and that he is fully
satisfied whenever cix is equal or larger than ci' we can express
his objectives by fuzzy sets whose membership functions increase
monotonically from 0 at ~i to 1 at ci' Denoting these fuzzy sets
as fuzzy constraints C(x). (21) can be written as
Find x
such that Cx ~ ~
Ax ~ b (30 )
x ;?: O.
147
Now we can apply directly model (29). The only problem that
remains is how to establish the ~i and ci' We can either assume
that the decision maker establishes apiration levels for himself,
or we can compute those as a function of the solution space. For
the latter approach it was suggested to use the individual optima
as upper bounds and "least justifiable" solutions as lower
bounds.
"Maximize" z = Cx
such that Ax $ b (31)
Dx $ b'
x ~ O.
k
:; min f' (32)
1
1 '=1
149
or
f1 (x) ~ min fj (x;,) f'
a
I' =1
Then f01 (x) min ( f1 ' f 1, ) . The crisp model equivalent to (30)
is then
Maximize >-
such that >- (f 1'-f 0') - Cx ~ -f'0
P + Ax ~ b + P (33 )
Dx ~ b'
>- , x ~ O.
REFERENCES
Bernard Roy
These two possibilities underline the fact that maps are not
territories: the vector of performances g(a) = [gl(a), ... ,gn(a)]
is like a map of that territory which is the action a. We want to
compare actions, i. e . territories, not maps. These comparisons
are made by means of maps, and we have to avoid working as if
maps do not differ from territories.
1.For more details on this concept, see Roy (1985, ch. 7) or Roy
and Vincke (1987).
Let us point out that «at least as good as» must be considered
as «not worse than».
In this case:
F gl g2 g3 g4 gs
A V'jEF: %=5
a 50 50 50 50 50 and
b 55 46 48 54 55
c 90 90 90 45 42
d 90 90 90 10 10
a'. _ _~,~__ .a
From this last property combined with the previous one, we can
deduce (for more details, see Roy and Bouyssou (1987a)):
and shares many points in common with it, the binary relations
they refer to are not, strictly speaking, outranking relations
(for more details on this point, see Roy and BOuyssou (1989».
a) Concordant criterion
b) Discordant criterion
1. For more details, see Roy (1985) or Roy and Vincke (1987).
166
c) Consequences
C(a'Sa) UC(aPa') oF F.
(2.1 )
By definition:
1
l: k. with k l: k., (2.2)
k jEC(a'Sa) J jEF J
1
l: rp .• k. with rp. (2.3)
k jEC(aQa') J J J
o ~ c(a', a) ~ 1,
c(a', a) 0 if C(a P a') F,
c(a', a) 1 i f C(a' Sa) F.
We can also admit that this veto effect can occur for a
difference gj(a) - gj(a') smaller than Vj when c(a', a) < 1 - kj/k.
This leads to reinforce the veto effect all the more as c (a', a)
decreases.
d,(a', a) ~ v, (2.4)
J J
3.1. ELECTRE IS
1
c(a', a) 2: s, 1/2 < s s s* with s* 1 - min k J., (3.1 )
k jEF
(3.2)
(more rigorously, in the above formula, qj should be re-
placed by min {qj' Vj - Pj})
1 - c(a', a) - kj/k
wi th w( s, c)'
1 - s - k/k
The condition (3.2) simply expresses the fact that the veto
effect does not occur for any criterion. The coefficient w(s, c)
is used to modelize the reinforcement of the veto effect intro-
duced in 2.3 c) above.
172
1 - d j (a', a)
a(a',a) c(a',a). '1'1" (3.3)
jEDc(a' ,a) 1 - c(a', a)
Let us remember that ELECTRE I (see Roy (1968» was the first
decision-aid method using the concept of outranking relation.
173
(1) All outranking relations are based on concordance and discordance concepts; except for
boxes containing «fuzzy)), the figures refer to non-fuzzy relations.
175
is not too important, nonetheless bring the veto into play for a
small difference in performance). Here again, if it is difficult
to give a precise value to a Vj threshold, we can assign it a
variation interval.
5. CONCLUSION
REFERENCES
Daniel Vanderpooten
1. INTRODUCTION
2. INTRODUCTORY CONCEPTS
.3
.4
189
We first deal with the case where the preference model only
consists of one crisp outranking relation S. Other cases are
introduced afterwards. Before presenting any concept or procedure,
we must state requirement (r. 6.1) in a more precise and opera-
tional way. Many operational restatements are possible resulting
in various concepts and procedures. We shall restrict our atten-
tion to a restatement consistent with the outranking approach in
the sense that it emphasizes a prudent attitude. This means that,
when characterizing AO, we must be able, at least, to justify the
elimination of any other alternative. We may then restate (r.6.1)
as follows:
H:
al~~
The reader should notice that (7.1), (7.2) and (7.3) are
mutually exclusive because of the above minimal requirements.
Thus, (7.1-3) could be used for sorting the candidate actions.
However, the resulting procedure would not be exhaustive (except
for n=l) in the sense that some candidate actions may verify none
of these conditions. Moreover, because of the imprecision inher-
ent in the definition of the criteria and in the specification of
the reference profiles, these conditions should be modulated. For
instance, a may be assigned to A+ even if some of the inequali-
ties in (7.1) are slightly violated. The use of outranking rela-
tions allows to handle both of these difficulties.
where:
B' {b E B / a(a, b) ~ t}, B" {bEB \ B' / a(b, a) ~ t'},
C' {CEC / a(c, a) ~ v}, C" {CEC \ C' / a(a, c) ~ v'}.
bO
, , bi
I
b h-I
~-T------~---+---r------4\---+i'~
bh b h+1
I
b k-I
I \ I
I
\
\ I \ I \ JI ~
\
Al
I \
Ah Ah + I
I \ ' :
We first deal with the case where the preference model only
consists of one crisp outranking relation S. The case with sever-
al crisp relations or one fuzzy relation is introduced after-
wards.
t
Zl: a1 ~:
t
a2 ~
~ t
a4
- w~ ( a ) =:E a ( b , a),
bEB
q~ ( a ) p~ ( a ) w~ ( a) .
9. CONCLUSION
REFERENCES
CONTENTS:
3. APPLICATIONS
A Location Problem.
(1.1 )
fl ( . ) f2 ( . ) ... fj(. ) . .. fk ( . )
a) f1 (a)) f 2(aj} .. . fj(aj} ... fk (a))
a~ fj{ a 2) f2 (a2 ) · .. fj (a2) · .. fk (a2)
.. .. . .. . · .. . .. · .. . ..
... · ..
I ai f1 (a;) f 2( ai) fj (ai) fk (ai)
.. . .. . .. · .. ... .. . . ..
an fI (an) f2 (an) · .. f j (an) · .. fk (an)
f1 f2 f1 f2 f1 f2 f1 f2 . f1 f2
3. 100 100 100 20 100 99 100 99 100 100
EXAMPLE IV: Again a and b are efficient. But they are not
incomparable like in Example II. In this case they could be
considered as indifferent.
219
o ~ p(a,b) ~ 1 (1. 4)
P( a, b)
d = f (a) - f(b)
P(a,b) d;':O,
H(d) 1 (1. 6)
P(b,a) dS;O.
H( d)
H (d) {~
d-O H[
Idl>O
No parameter to fix.
0 d
l [
IdlSq
H(d) {~ Idl>q HI'
q to fix.
0 Q d
~
{ I~I/P IdlSp
H (d)
Idl >p
p to fix.
0 P d
~ ~
IdlSq
H (d)
{1!2 q<ldlSp
Idl>p
HI'
q and p to fix.
q
r
0 P d
~
H (d) { 0 IdlSq
(Jdl-q)/(p-q) q<ldlSp HI'
r Idl>p
q and p to fix.
0 q P d
\l!
TYRe VI : Gaussian criteril!n.
25
H (d) = 1 - exp ( - d /2 2 )
5 fix.
to
0 • d
It has been observed that the GAUSSIAN criterion has been the
most selected by users for practical applications. After the
GAUSSIAN criterion, the V SHAPE CRITERION WITH INDIFFERENCE AREA
also has been often selected.
k
rr(a,b) :E Wj P j (a,b) 1) (1. 9)
j=l
1 k
n:(a,b) l: P. (a,b) (1.10)
k j=l J
n:(a,a) = 0 (1.11)
O$n:(a,b)$l \ia,bEA
n:(a,b) 0 implies a weak global preference of a over b,
~
n: (a, b)
~ 1 implies a strong global preference of a over b I
~f)~(l
complete fuzzy outranking graph
on A. This graph is a consider-
able enrichment of the domi-
nance graph because appropriate
a •
d levels of intensities of pref-
erence were taken into account
thanks to the associated gen-
eralised criteria.
¢+ ( a) = l; n: ( a , x ) (1.12)
xEA
¢- ( a) = l; n: ( x , a ) (1.13)
xEA
00
~~
The positive outranking flow expresses how each alternative is
outranking all the others. The higher ¢+(a), the better the
alternative. ¢+(a) represents the power of a, it gives its
outranking character.
{a S- b iff ¢- ( a ) <¢- ( b )
(1. 15)
a 1- b iff ¢- ( a ) =¢- ( b )
228
r
The PROMETHEE I partial relation is the intersection of these
two preorders
and a S- b
S>b
a pI b if a S+ b and a 1- b
a r b and a S- b
a II b if a r b and a 1- b (1.16)
aR b otherwise
(1.17 )
It is the balance of flow. The higher the net flow the better
the alternative.
or alternatively:
k
<I>(a) = ~ <l>j(a) Wj (2.3)
j=l
where
(2.4)
<I>(a)
r(a) n-1
=
k
~
j=l
rj(a}w j (2.5)
where
1
r.J (a) n-1 xEA
~ [P.(a,x) - P.(x,a)]
J J
(2.6)
231
(2.7)
(2.8)
71 ( . ) 72 ( . ) . .. 7j ( • ) ... 7k ( . )
4>1
• •
..
In R'< In
GAIA
Actil:ms at ai
5::ritt=:r;ia ej Yj
k k
MAX {u'Cu+v'Cv} :E c jj I ¥j 112 + 2:E :E Cjs ( ~j' ~s) ( 2 .9)
u,v j=l j=l sfj
where:
The (u,v) plane or the GAIA plane is the one on which as less
information as possible get lost by projection. A measure of the
information having been preserved is given by
S = (2.10)
Let us consider the projections ~j' j=1,2, ... ,k, on the GAIA
plane. These projections have different lengths and different
orientations. Thanks to (2.9) it is possible to obtain a very
clear interpretation of these lengths and orientations.
2. Similar criteria
3. Independent criteria
will be nearly zero. As, in this case, Cjs ( ~j 1 ~s) cannot have a
significant contribution to (2.9), the GAIA plane will be chosen
so that the scalar product will be nearly zero. SOl independent
cri teria will have ~ vectors nearly orthogonal.
4. Conflicting criteria
v
GAIA plane
Y2
u
235
k
(o,i'w) = l; wj7j(aj} (2.12)
j:l
The weight vector w looks like a stick located above the GAIA
plane, a stick that the decision maker can move according to his
preferences in favour of particular criteria. When the stick w is
moved, the PROMETHEE decision axis also moves in the GAIA plane
where the consequences for decision making can be clearly
observed.
GAIA plane v
3. APPLICATIONS
The basic data are given in the left hand side of the follow-
ing table. For each criterion i t is mentioned i f i t is to be
maximised or to be minimised.
CRI ARAMETERS
f3( . ) min 600 200 400 1000 600 700 V q=50 ,p=500
f4 ( . ) min 54 97 72 75 20 36 IV q=10,p=60
f5 ( . ) min 8 1 4 7 3 5 I
f6 ( . ) max 5 1 7 10 8 VI s=5
al a2 a3 a4 as Cifi ~+ ( • ) ~- ( .) ~( .)
al 0 0.296 0.250 0.268 0.100 0.185 1.099 1.827 -0.728
PROMETHEE I PROMETHEE II
241
v
GAIA plane
• a 6 (France)
a 3
(Germany) Y3 (Const. C sts)
.a 1 (Italy)
a2 (Bel\lium)
__ • Y2 (Power)
Y4 (Maintenance)
Y6 (Security)
a5
(Portu\lal )
(Manpower) Y1
France and Italy are powerful plants. Belgium and Germany are
not but they are cheap. Portugal is selected by PROMETHEE II, it
is located in the direction of the PROMETHEE decision axis.
4.1. Introduction
For instance fixed assets, equity, long term debt, net current
assets, inventories, accounts receivable, other assets, accounts
payable, other liabilities, requirements, liquid funds, short
term borrowings, cash, total assets, ... , from the balance sheet;
and sales, gross profit, value added, operations profit, finan-
cial profit, other profit, net profit, depreciation, . cash-flow,
generated cash-flow, ... , from the income statement.
f1 f2 .. . fj ... fk
CLIENT 1 fdcll f2 ( cll ... fj( cd . .. fdcd
CLIENT 2 fd c 2) f2 (c2) .. . f j ( c2 ) ... fd c2)
.. . .. . . ..
CLIENT i fdcil f2 (cil .. . fj (ci ) ... fk (ci)
.. . .. . . ..
CLIENT n f1 ( c n ) f2 ( en) . .. f j ( c n) ... fdcn)
- RFB is stable
- The RFB is the same standard for each client. After some
use, its features are well known by the decision-maker.
- The evaluation of a client on the RFB requests a limited
computation time. Real time evaluations become possible.
- Different RFB may be considered: representative clients,
only good clients, target clients, ... , , ..
CLIENT a
CLIENT b
CLIENT c C L B
01
2
REF.CL. u
REF.CL. v
3 REF.CL. w
~I CLIENT i
R F B
REF.CL. Y
CLIENT n-1
REF.CL. z
CLIENT n
The clients are ranked from the best to the worst one. The
ranking of client i provides its positioning with regard to the
standard. This ranking is obviously a global evaluati.on of the
client. This global evaluation is displayed on a strip graduated
from 1 to 50 in the BANKADVISER display.
I :l~ II +
I -1
U I:
S 1 EQUITY RATJO 3U x +l1Yo L 4 RECEIU, TURNOVER 49 d +7x
S 2 C-F/LG TERM DEBT 196.0 x +2Si. }L 5 PAYABLE TURNOVER
!I
48 d -5X(
Hex
I
iLl ACID TEST 199,2 Yo +12Yo P1 HET PROFIT/SALES 1.1 X
I L2 CASH INCREASE -59,4 Yo P 2 OPER, PROF/SALES 1.7 x +25x
Ii L3 INVENT, TURNOVER 14 a I P 3 CASH-FLOW/EQUITY 26,8 x -9X
r,
\11
W-e-i9-ht-:--4-,-4-!.----------~ I
~----------------
Statistics : Average 126 d KinilllJll : 5d
( 51 clients) St, dev, 88 d Ka.xi~ : 437 d ,
4.7. Conclusion
1. Treat an example
2. Define a problem
3. View of the problem
4. Use the spreadsheet
5. Solve the problem
6. Modify the data
7. Save the problem on a file
8. Print all results of the problem
9. Quit PROMCALC (or GAIA)
REFERENCES
Benedetto Matarazzo
Istituto di Matematica
Facolta di Economia e Commercio
Universita di Catania, ITALY
1. INTRODUCTION
The aim of the two methods briefly outlined here is to aid the
decision maker with disc~ete problems in the presence of more
than one cardinal criterion. Methodologically the two have much
in common with the original Pairwise Criterion Comparison
Approach (PCCA) whose specific quality consists in the
possibility to compare feasible actions with respect to all the
possible unranked pairs of distinct criteria considered. The
partial results thus obtained are then suitably aggregated and
used in order to aid the decision maker in a variety of problems.
From this point of view, therefore, the approach proposed may be
considered as an attempt to make explicit the limited capacity of
the human mind to make comparisons between numerous and often
conflicting evaluations simultaneously; it offers, instead, a
series of comparisons easy to execute one at a time.
i f aj < bj
gj{ ah) > gj{ak) and gj (ah) > gj (ak) <=> chi > cki and Chj > Ckj
gj{ ah) gj{ak) and gj(~) > gj (ak) <=> chi cki and ~j > Ckj
gj{ ah) > gd ak) and gj (ah) gj(ad <=> chi > cki and Chj Ckj
(1)
in particular:
TI ij ( ah , ak ) TIij ( ak , ah )
or
TIij(ak,ah) TIij (ah ,ak)
Kj Hi = Ai chi
Hj = AjChj
Ki = AicKi
H' K j =_AjCKj
IJ = 1
T
x
Observe that Rij (ah' ak) = 0.5 may occur both if ah and ~ are
indifferent with respect to the criteria gi and gj (that is,
g;(ah) = g;(ak) and gj(ah) = gj(ak»' and i f ah and ak are equiva-
lent, that is if:
'1T
_lJ
.. 1,2, ... ,n; k 1,2, ... ,n).
259
=l.: i .j
~
_lJ
.. (3 )
(i<j) m- 1
In particular we obtain:
a(1)(a) = n - 1 or a(1)(a) 0
+ h + h '
This iteration will make it possible to form class C2, and so on.
and placing in the last class Cs the action(s) which present the
highest value for this index. We then proceed with the calcula-
tion of the indices 0(2) related to the subset A - Cs and so on.
Ai Aj
Fig.2
i
Rj
,
'- Tj partial broken line - 1
Sjr--_ _ '_B,_ _ _ _D...-"'- Sj
~ Rj BDTj
..... "C....
partial broken line - 2
Tj. .. =Sj BCDSj
Ti
f(1) (as)
I)
3
0,3 ;
(2)
f ij (as)= 0.7 ; ,(3) (a s )=
I)
° ;
Ai B C D
n n
:E fij(k)(ah) 1, 'V ah E A and:E f d k) (ah) 1, k = 1,2, ... ,n.
k = 1 h = 1
n n
:E 1, 'V ah E A and :E 1, k 1,2, ... , n.
k =1 h =1
k
F(l) (ah) f(1)(ah) and F(k)(ah) = 1: f(i)(ah)' k 2,3, ... ,n.
i =1
268
q
~ ~ F(k) (ah) , ah E Ai 1 ~ O-t ~ ~ ~ ••• ~ O'.q > o. (5)
k =1
6. FINAL CONSIDERATIONS
REFERENCES
Economica, Paris.
Roy, B. ( 1990): "The outranking approach and the foundations of
the ELECTRE methods", in this volume.
Scharlig, A. (1985): D~cider sur plusieurs crit~res, Presses
Poly techniques Romandes, Lausanne.
Vansnick, J.-C. (1990): "Measurement theory and decision aid", in
this volume.
Vincke, Ph. (1985): "Multiattribute utility theory as a basic
approach" in G. Fandel, J. Spronk and B. Matarazzo (eds.),
Multiple Criteria Decision Methods and Applications,
Springer-Verlag, Berlin, 27-40.
Vincke, Ph. (1990): "Basic concepts of preference modelling", in
this volume.
CHAPTER I I I
Rakesh K. Sarin
1. INTRODUCTION
or
2. WHAT IS MEASUREMENT?
Note, and note well, that cIJ 1 ' c1J 2 ' c1J 3 are not constructed inde-
pendently and then cIJ obtained by simply summing these three
scales. Thus, you cannot interpret cIJ l by itself and nor can you
construct it independent of other components. I hammer on this
point because I have seen many mis-applications of conjoint meas-
urement where c1J.' s are constructed using independent processes
1
such as rating scales and cIJ is obtained by simply summing these
component scales.
4. PRELIMINARIES
Now, think of our car example and we can order all the cars in
set X using above expression.
~ (x)
This axiom requires that the d.m. is able to compare any two
objects XI y E X (indicate X i y or y i x) and these preferences
are transitive (if x i YI Y i ZI then x i Z for all XI YI Z EX).
(x~ I x 21 x 3) i (X{' I
XU
2 I
x3 ) I
(x~ , x 21 Y 3 ) i (x{' , XU
2 '
Y3 ) •
Xl XU
i component i i
285
For any xi' xi' E Xi and x 1+I ' xi~l E X i +l , suppose it is true that
then
then
Note that i f cfJ (x) = F(cfJ (Xl)' cfJ 2 (x2 ), ... , cfJn(xn )), where F is
strictly increasing, then weak independence must be satisfied.
When we specialize F further so thatcfJ(x) =cfJI(xI ) +cfJ 2 (x2 ) + '" +
cfJ (x) then Axiom 2 becomes necessary.
n n
Pick a component say Xl' Now suppose that xlm and xlm-l are equi-
distant in the sense that for each m = 2, 3, , xlm xlm-l and ...
m
(Xl' X2 ' X3 ' ...
, Xn) ( m-l ,
xl Y2' Y3'
~
, Yn) , for two fixed sets of ...
points x j ' Yj E Xj , for j > 1.
rl £ cfJ.(y.)
j=2 J J
- ~
j=2
cfJ.(x.)j > 0
J J
level of . (x",
1
x2 )
Preference
Level
level of (xl' X 2)
level of (x;' f ~2)
x'2
288
n n
x l y if and only if:E <p.(x.) ~:E <p.(y.), for all x, y EX.
i=l 1 1 i=l 1 1
Further, <P ~ =
1
0: <p.1 + /3.,
1
0: > 0 will also provide additive con-
joint representation. Thus, <p. 's are unique up to positive
1
linear transformation with change in unit 0: the same for all
components.
The above Theorem applies only when there are three or more
essential components. If only two components are essential then
we must add a condition called Thomsen condition to the list of
axioms.
289
X3 Y3
xl YI Zl Xl YI Zl
x2 12 14 18 x2 11 13 17
Y2 4 10 16 Y2 3 9 15
z2 2 6 8 z2 1 5 7
We start with an origin (~l' ~2) and set its value to be zero
(for simplicity). Thus <I>(~l' ~2) == 0, <l>l(~l) == 0, <l>2(~2) == 0. Now
we select a unit (xl' ~2) and set its value 1 (again for simplic-
i ty). Thus <I> (Xl' ~2) == 1, <I> I (Xl) == 1.
X"
2
X2
x'2
X2
~2
~l Xl x'I x"
I
Xl
( xl' X2 ) ~ ( Xi ' ~2 )
(Xi, x 2 ) ~ (x{', ~2)' and so on.
3, ... .
REFERENCE
Ernest H. Forman
1. INTRODUCTION
0.509
COST
0.243
VISIBLE
0.155
CUST .FIT - - - - - - - -
0.094
COMPET' N - - - - -
INCONSISTENCY RATIO = 0.098
n~
Now let us consider the case where we do not know~, and where
we have only estimates of the aijs in the matrix ~ and the strong
consistency property most likely does not hold. (This allows for
small errors and inconsistencies in judgments). It has been shown
that for any matrix, small perturbations in the entries imply
similar perturbations in the eigenvalues, thus the eigenvalue
problem for the inconsistent case is:
matrix equation 1.
Expert Choice since the larger the value, the more inconsistent
the judgments.
Main Street
The Mall
The first step in using AHP and the Expert Choice software is
to develop a hierarchy by breaking the problem down into its
components. The three major levels of the hierarchy shown in
Figures 3 and 4 are the goal, criteria, and alternatives.
306
c:;;:J
I
~ ~ iCUST.FITIICOMPET'NI
I -SUBoCTRol-SUBoCTRol-SUBoCTRol-SUBoCTRo
-THE MALL -THE MALL -THE MALL -THE MALL
-MAIN ST
0 -MAIN ST
0 -MAIN ST0 -MAIN ST0
P:'AIN ST.-
COMPET' N1r~HE MALL-
SUB.CTR.-
P:'AIN ST.-
CUST .FIT1r~HE MALL-
SUB.CTR.-
GOAL-
P:'AIN ST.-
VISIBLE-,,~HE MALL-
SUB.CTR.-
P:'AIN ST.-
COST--,,~HE MALL-
SUBoCTR.-
THE MALL-
COMPET'N-I:MAIN ST.-
SUB.CTR.-
THE MALL-
CUST. FIT-I:MAIN ST.-
SUB.CTR.-
THE MALL-
RENT--tMAIN ST.-
SUB.CTR.-
COST- {
THE MALL-
AINTNCE-I:MAIN ST.-
SUB.CTR.-
THE MALL-
COMPET'N-tMA1N ST.-
SUB .CTR.-
THE MALL-
CUST .FIT-tMAIN ST.-
SUB.CTR.-
STAT .QU
THE MALL-
VISIBLE-[MAIN ST.-
SUB.CTR.-
THE MALL-
COST--tMA1N ST.-
SUB.CTR.-
THE MALL-
COMPET' N-tMAIN ST.-
SUB .CTR.-
THE MALL-
CUST .FIT-tMAIN ST.-
SUB.CTR.-
GOAL BOOM
THE MALL-
VISIBLE-[MAIN ST.-
SUB.CTR.-
THE MALL-
COST--tMA1N ST.-
SUB.CTR.-
THE MALL-
COMPET ' N-tMA I N ST.-
SUB.CTR.-
THE MALL-
CUST. FIT-tMAIN ST.-
SUB.CTR.-
GLOOMY
THE MALL-
VISIBLE-[MAIN ST.-
SUB.CTR.-
THE MALL-
COST--tMAIN ST.-
SUB.CTR.-
Figure 6 - EC Model with Envirorvnental Scenarios
309
THE MALL-
COMPETIN-tMA1N ST.-
SUB.CTR.-
THE MALL-
CUST .FIT-tMA1N ST.-
SUB.CTR.-
CONSULT.
THE MALL-
V1S1BLE-tMA1N ST.-
SUB.CTR.-
THE MAU-
COST---tMA1N ST.-
SUB.CTR.-
THE MAU-
OMPETIN-tMA1N ST.-
SUB.CTR.-
THE MALL-
CUST.FIT-tMA1N ST.-
SUB.CTR.-
GOAL MKTG.D1R
THE MAU-
V1S1BLE-tMA1N ST.-
SUB.CTR.-
THE MALL-
COST---tMA1N ST.-
SUB.CTR.-
THE MALL-
COMPETIN-tMA1N ST.-
SUB.CTR.-
THE MAU-
CUST • FIT-tMA1N ST.-
SUB.CTR.-
lCEPRES
THE MAU-
V1S1BLE-tMA1N ST.-
SUB.CTR.-
THE MAU-
COST---tMA1N ST.-
SUB.CTR.-
L1TTLE-
COMPET ' NiMODERATE-
STRONG-
INTENSE-
POOR--
CUST.FlTiFAIR--
GOOD--
EXCELLNT-
GOAL-
HlDDEN-
VISIBLEiLOII--
O.K.--
GREAT--
CHEAP--
COST-[AVERAGE-
EXPENSIV-
VERY EXP-
1 SUBURBAN CNTR#l
2
3
SUBURBAN CNTR#2
OLD TOliN AREA
CHEAP
CHEAP
VERY EXP
O.K.
LOll
GREAT
GOOD
GOOD
EXCELLNT
STRONG
MODERATE
INTENSE
8:jll
0.270
4 THE MALL VERY EXP GREAT EXCELLNT STRONG 0.266
5 MAIN ST/HI RISE EXPENSIV O.K. GODD STRONG 0.232
6 NEAR APT .CLUSTER AVERAGE LOll GOOD MODERATE 0.229
7 OFF INTERSTATE EXPENSIV O.K. EXCELLNT MODERATE 0.221
8 SUBURBAN CNTR#3 AVERAGE HlDDEN FAIR MODERATE 0.194
311
0.515
SUB.CTR. •
0.196
THE MALL - - - - - - - - -
0.290
MAIN ST. - - - - - - - - -. . . . . . . . . .
INCONSISTENCY RATIO = 0.000
Figure 9 - Preference for Alternatives with Respect to Cost
3.3. Synthesis
Once judgments have been entered for each part of the model,
the information is synthesized to achieve an overall preference.
The synthesis produces a report which ranks the alternatives in
relation to the overall goal. This report includes a detailed
ranking showing how each alternative was evaluated with respect
to each criterion. Figure 10 shows the details followed by a
ranking of the alternatives.
314
COST =0.509
SUB.CTR. =0.262
MAIN ST. =0.147
THE MALL =0.099
VISIBLE =0.243
THE MALL =0.158
SUB.CTR. =0.068
• MAIN ST. =0.017
CUST.FIT =0.155
SUB. CTR. =0.070
THE MALL =0.070
• MAIN ST. =0.014
COMPET 'N =0.094
SUB.CTR. =0.053
MAIN ST. =0.031
THE MALL =0.009
SUB.CTR. 0.453 - - - - - - - - - - - - - - - - - - - - - - -
3.4. Sensitivity
,~~r-----------;------------,
19~
,30
17~
,6~
I _____ UB, CTR.
_-+------
15~
-----=----------
,40 - - - -
I 3~
____
-------r---__
12~
.1~
L___--..-----..-"----~---=::::::.~;::::::::::=IAHEINSf.
MALL
• ~~'-----L..---'--..L.---l.----1--...L.---.lL......---L..---L----l
.1 .2 .3 .4 .5 .6 .7 ,8 .9
COST
4. OTHER ASPECTS
Other aspects of AHP and Expert Choice may interest the read-
er, but are beyond the scope of this paper. These include the use
of subsidiary models to evaluate the importance of players or the
likelihoods of scenarios, geometric averages for group judgments,
structural adjustment of priorities based on the number of
descendants, product adjustment for priorities based on multipli-
cative factors and other sensitivity analyses modes.
5. SUMMARY
BIBLIOGRAPHY
Valerie Belton
University of Strathclyde
and
Stephen Vickers
Heriot-Watt University
1. INTRODUCTION
/ Recommendation
Sensitivity 1m plementation
\
analysis
(
Synthesis Define
\
Ilnatives
~score
alternatives
value value
100 100
km m2
distance from shops size of garden
They say how much of one criterion you would be prepared to give
up in return for an improvement on another criterion. The actual
values of the weights are related to the measurement scales used
for the scores. Thus the weight of a criterion captures both the
psychological concept of importance and the discriminatory power
of the scale on which the criterion is measured. It is possible
for a criterion which is very important to the decision maker to
have a low weight because it does not discriminate between the
alternatives under consideration, i.e, the scores of a and 100
represent something very similar. This is particularly so when
using local scales. In order to distinguish between the psycho-
logical concept of importance and the concept captured by crite-
ria weights, we will refer to the former as importance and the
latter as significance. A suggested procedure for weighting is as
follows:
Step 8: Implementation
3. V'I'S'A
VISUAL INTERACTIVE SENSITIVITY ANALYSIS FOR MCDM
The top bar chart represents the overall scores of the alter-
natives being evaluated, given the current status of the model.
The bottom bar chart represents the distribution of criteria
weights within the selected family. By using the cursor control
keys the user can select a particular criterion and change the
weight allocated to it. All other weights change to preserve the
normalisation, their relative values being retained. The effect
of the change on the overall scores is seen immediately as the
bars grow or shrink. The user can display the numeric values on
either or both of the bar charts, obtain a printout of the visual
display, obtain a printout of the current status of the model in
tabular form, opt to retain the current set of values or return
to the initial set of values.
The simple weighted value function model has often been criti-
cised as a decision aid. In this section we explore some of those
criticisms and discuss the strengths of the approach.
331
Criterion 1
Criterion 2
Figure F
REFERENCES
Eric Jacquet-Lagreze
Euro-Decision and Lamsade
Euro-Decision, 64 rue Royale
78000 VERSAILLES FRANCE
Evaluation
... Preference Model
Model Structure
......
Parameters
(weig hts, trade-off)
Matrix
Global Preference
u(j), uO') or
j s r
The main question here is WHY does the DM prefers j to j'? The
answer being choose a mathematical model and ask for holistic
judgments or preferences of the DM and estimate indirectly the
parameters of the model.
Evaluation
Matrix
.. Preference Model
Model Structure -
.... Hollistic Preference
(rank order, .... )
r
Parameters of the
model
weights, ....
- Additivity
- Monotonicity
Let us introduce
criterion . . the limit of the most desired level on
xi can be a theoretical value like value 100 on a
l.
- Cardinality
u (x,)
u (x') 1
u(x) = l:wivj{xil
i
1 for all i
341
The user can then modify the values uik to draw a marginal
utility function representing in a better way his preferences.
Remark 1
This model remains simple (additivity). Users not familiar
with MCDM concepts understand possibilities offered by the model
to express judgments such as "after such value xik' the impor-
tance of the criterion decreases (i. e. the slope of the curve
decreases)". They appreciate the extreme possibility to draw a
flat curve beyond a particular value xik'
342
Remark 2
The initial quadratic function is often kept because it is
concave and represents the situation for which a criterion is
important when the value is low on this criterion and once we are
good on this criterion it is much less important.
Remark 3
For an ordinal qualitative scale, it is recommended to choose
only one linear piece because in a certain way the initial values
are already defined as a preference scale. On the other hand it
is most important to choose more than one linear piece for
numerical scales (PREFCALC is limited to 5 linear pieces).
Max S - M :E Zj
j
Uik - uik+l ~ 0
since the variables introduced in PREFCALC are the differences:
1 + 10 + 3 x 10 = 41 variables
9 + 1 = 10 constraints.
Remark 1
The error function is not very satisfactory in case a) (no
solution). A better error function implies to introduce two error
variables per alternative, in order to minimize the sum of the
deviations. Such a modification has been introduced in the exten-
sion MINORA of UTA proposed by Siskos and Yannacopoulos (1983).
Remark 2
A point such as J is never computed by PREFCALC. This could be
misleading in some situations. The fuzziness could actually be
worse than the one computed by the program.
Remark 3
Only a few criteria (up to 10 in PREFCALC) can be taken into
account otherwise the disaggregation would yield too imprecise
estimation or the user would have to rank-order too many alterna-
tives.
Concluding Remark
a1
a3r-----________~
a3
o~--------------~
n
Min M:E Zj + :E ti
jEJ i=l
a and Wi = wi for i = 1 to n
:E WiXij - ui for all j E J.
REFERENCES
1. By «decision acton> we refer to those actors whose preference systems rrust be taken into account (such as
the decision-makers themselves or other agents directly involved in the decision making process). From now on
we wi 11 simp ly refer to them as • actors » •
2. Rules for the development of this dialogue have to be established. accepted and well understood by all the
participants. In particular. the procedure supporting the dialogue must be consistent with the type of
aggregation model previously accepted.
4. We refer here to the Outranking Approaches based in the concepts of concordance and discordance. which
strictly speaking are the .true. outranking concepts (see Roy. 1990).
352
1. «Weights» Wj (j=1, ... ,n) are used to refer either to «scaling factors. or to «coefficients of importance"
assuming that it is clear in the reader's mind that their meaning and characteristics depend upon on the type
of operational approach being used (see Vansnick, 1986). As Roy (1990) states, the weights have an intrinsic
character in an ELECTRE method, but they are non-intrinsic in the AHP and MAUT methods because they depend on
the nature of the scale chosen for evaluation (which further complicates the weighting process). To ignore
this fact can give rise to meaningless practices, like it is to combine (apparently simpler) weighting tech-
niques with aggregation mode Is with which they are not theoretically compatible. As an example, between an
(unfortunately) large collection of cases, one can refer to applications of Saaty's Method to specify the
weights and in sequence to make use of those substitution rates as coefficients of importance within an ELECTRE
Method.
2. The absence of a single and/or well identified decision maker is not so rare as one could imagine, mainly in
appl ications in the publ ic sector.
353
iii) X = {X1, ... ,Xj, ... ,Xn } is a set of quantitative and/or quali-
tative attributes, an attribute Xj being a set of at least two
elements Xj expressing different levels of some underlying dimen-
sion. An alternative can be typified by a tuple (xl""'~""~)
1. I.e .• being Xj and Yj any two evaluation levels on the attribute Xj. if the decision-maker prefers Xj to Yj
than the real numbers Vj(Xj) and Vj(Yj) must be so that Vj(Xj) be greater than the value Vj(Yj)'
355
Let (xu ... ,Xj, ••• ,x n ) and ( Y P " " Y j " " ' Yn ) represent the
evaluation profiles of two alternatives a and b of A. On each
attribute Xj E X one, and only one, of the two following prefer-
ence situations can hold:
2. This paper does not detail the development of the phase of modelling intra-attribute preferences. Neverthe-
less, we want to remark that as a rule the construction of the value functions Vj is not a easy task. Basic
references on this subject are Fishburn (1967), Keeney and Raiffa (1976). Keeney (1977) and Dyer and Sarin
(1979) .
356
n
v(a) L WjVj (a) .
j=1
1. "Weight is simply a transfonnation of a scale, and that scale is meaningful only if its end points are
clearly defined" (Edwards et a1., 1988). This is why it is more appropriate to call them scaling constants. See
illustration in Keeney and Raiffa (1976, section 5.9).
3. See Chankong and Haimes (1983) for a theoretical review of the additive model conditions and Roy and
Bouyssou (1987) for a critical discussion. See also the discussion in Bell et a1. (1977, chapter 19) and French
(1988) .
4. "Simplicity helps in many ways. It helps a nontechnician (or even a technician) to understand what is going
on. It makes analysis briefer and easier to perform. It makes jud9fents easier to make. Above all it helps the
decision maker to use and communicate about the tools that have been developed and their results, and this
promotes adopt ion and use" (Edwards et a 1., 1988).
5. As Edwards (1977) remarks, "probably the most widely used, and certainly the simplest, aggregation rule
( ... ) cons ists of simply tak ing a weighted 1 inear average".
358
n
vi) L: w· 1 and Wj > 0 for all j.
j=l J
n
Strict preference ( »: a > b iff ~ Wj. [Vj(Xj)-Vj(Yj)] > 0 or
J=l
or
n
Indifference (-): a - b iff ~ Wj. [Vj(Xj)-Vj(Yj)] O.
J=l
iii) Lower and upper preference bounds for value tradeoffs, for
substitution rates or for weights 4.
(j l, ... ,n and j # r)
1. Although developed within undoubtedly different operational MCDA approaches, both the questioning procedures
proposed by Keeney and Raiffa (1976, section 3.7) and by Roy et al. (1986) for the assessment of weights
(scaling constants or coefficients of importance, respectively) are based upon on comparisons between subsets
or coalitions of criteria. If for some reason the dialogue is broken at initial stages, their output will be
nothing better than a few inequalities relating weights, for example, "one weight in bigger (or smaller) and
the other we ights are not rank-ordered".
2. Here included not only the case of a complete rank-order of the weights, but also some variations of
pract ica 1 interest, 1ike "the constants are rank-ordered with two constants equa 1" or "subsets of constants are
rank-ordered" (these two cases are both considered by Kirkwood and Sarin (1985)).
3. Two MAUT applications both dealing with a complete rank-order of the scaling constants are described in
(Kirkwood and Sarin, 1985) and (Bana e Costa and Neves, 1989).
4. Intervals (bands) of no preference, i. e., no value between the bounds is more representative in terms of
inter-criteria preferences than any other, such that narrow bands correspond to richer information (less
uncerta i nty, imprec i s i on and/or conf 1i ct) •
5. See Keeney and Raiffa (1976), Keeney and Nair (1977) and Keeney (1980) for description and practical
appl ications of these procedures.
6. "Often a decision maker is unable or unwi 11 ing to provide precise multiattribute tradeoffs" (Hazen, 1986).
360
(forallj~r).
1. Hazen (1986) illustrates this type of poor information with an hypothetical example constructed from the
real-world decision analysis case of selecting nuclear power plant sites described by Keeney and Nair (1977)
and Keeney (1980).
2. This is the case of the choice problem of a suburban highway design in the Paris forest environment
described by Bertier and de Montgolfier (1974). A simplified version of this problem is used as an example by
Bana e Costa (1989 b) to illustrate the applicability of the Outweigh Approach for supporting decision-making
in situations characterised by the fact that an interval for each tradeoff rate is independently defined by
each of the actors involved.
3. It is important to remark that (practical) simplicity and (theoretic) adequacy are often incompatible aims,
and, unfortunately, many applications of simple weighting techniques ignore this fact.
361
1. A very illustrative example of this type of situation can be found in Edwards et a1. (1988). They describe
the design of a system for evaluation and prioritisation of research projects, to be systematically used from
year to year, thus being easily understandable that the weights needed to be available before the value meas-
ures were.
2. See Bana e Costa (1989 b) for an appl ication developed in the space of the tradeoff rates.
362
i) ii)
iiLl) iiL2)
1. Nevertheless, we do not assume it is possible to totally overcome incomparability. As the outranking ap-
proaches, the outweigh approach admits as realist ic that «indecision» situations can not be totally el iminated
in many (most of) decision aid processes.
2. See (Bana e Costa, 1989 a) for a more detailed list of references, both in the framework of the so-called
Outranking Family and in bayesian probabilistic or expected-utility contexts with incomplete knowledge of
probabil ities in MAUl.
364
a ~P b if f vP ( a ) = vP (b)
r(a,b) O}.
It can happen that r(a,b) does not intersect the set of feasi-
ble weights W (W n r(a,b) = ¢). In this case, one of the two
alternatives will be always globally preferred to the other in W,
and we will say that a situation of «restricted dominance» 0. r )
occurs, even if neither a dominates b nor b dominates a (i. e. ,
Q n r( a , b ) " ¢) - see figure 2- i) 1:
2. Following Hazen (1986). we can distinguish between .additive (restricted) dominance. for the case of an
additive multiattribute aggregation function and «multiplicative dominance' if the multiplicative model is
used. Let us also point out that the «restricted dominance relation. is equivalent to the outranking relation
built in the framework of the Method of Progressive Reduction of Incomparability. So, taking the concept of
outranking relation in a non strict sense, we could use «additive outranking. to refer to the same concept of
.addit ive dominance».
366
..,
.,
"
ii)
n
W(a,b) {P E W: l: ~. [Vj(Xj)-Vj(Yj)] > O}
j=l
and
n
W(b,a) {P E W: l: ~. [Vj(Xj)-Vj(Yj)] < O}.
j=l
"
.,
Figure 3
1. This way of proceeding will (very probably) enrich the dominance analysis, provided that it can be possible
to go back to previous phases. As a matter of fact, many times the evolution of the decision aid process
will not permit to contract the initial feasible set loll, as for instances, when it is not possible to estab-
I ish new consensus between the actors.
368
n
1) Solve the linear program: Min :E Wj' [Vj (Xj ) -Vj (Yj ) ] z* ,
W j=l
THEOREM - If wI > ••. > Wj > .•• > wn ' then a is guaranteed to be
preferred to b if and only if
m
:E 1, ... ,n
j=l
m
:E 1, ... ,no
j=l
1. That is. if W = {(wI •...• Wj ••..• wn) EO: aj < w/wr < Ilj (j=l •...• n and jk)} - case iii.l) in section 2.4.
370
..
Figure 4
RANK-ORDERING ALGORITHM
#W(a,b)
d(a,b) =
#W
As W is by definition continuous in the space lfI- l , by analogy
the measure of the credibility degree d (a, b) between any two
alternatives a and b of A will be given by:
d(a,b)
J W(a, b) oWl' ••• , OWj' ••• ,own_l
V[W(a,b)]
d(a,b)
V[W]
1. More rigorously, we have to include also the points for which a and b are indifferent.
2. Within the context of the REGIME Analysis Hinloopen et al. (1983) also proposed to associate the value of
the credibil ity degree d(a,b) with the relative size of w(a,b).
374
- d: A xA --. [0,1], that is, 0 :<:; d(a,b) :<:; 1, for all (a,b);
01 al a2
0
a3
0
a4
.133
as
0
al
a2 .991
a3 .009
a4 .867 0 0 .678
as 0 0 .322
375
1. Being the fuzzy outweigh relation max-min transitive, the derived crisp relations are obviously transitive.
2. The nature of the conditions which must be satisfied to validate the assertion «a outweighs bll (in the sense
of «a is at least as good as b.) is obviously influenced by .the strength of the arguments required: the
strongest we can imagine is a dominates b but the outranking concept is an interesting one only because weak
arguments can be sufficient. This is way S is usually a notably richer relation than dominance.> (Roy, 1990).
376
.
TABLE 4 - PAIRWISE COMPARISONS (5 = 0.85)
II CO 1
;\ rota 1 4 0 1 2 2
jiI
ll..:...-:.-:-= .. -...,.-::-:-=-==-:::..:....~-::=~_=_""-.....-:--:--_~":".----=-:
Figure 5
377
6. FINAL COMMENTS
1. See the VISA approach. presented in this volume by Belton and Vickers (1990).
378
7. REFERENCES
1. INTRODUCTION
These methods share the property that the weights are assumed
.0 be non-negative and add up to 1. Thus, if criteria have been
'anked in decreasing order of importance, the set of weights S
~ich is consistent with the information on the ranking reads as
ollows:
386
S (1 )
B (0,0, \)
B (0,0,1)
C~ _ _ ~A
(0,1,0) E (1.1.0) 11,0,0)
B (O,O,ll
c~~
(0,1,0)
A
(1,0,0)
o elsewhere
E ( >-1) 1/J2
E ( >-2) 1 1J2 + 1 1 [ J ( J -1 ) ]
(3)
EPJ_Il = 1/J2 + 1/[J(J-1)] + ... + 1/[J(J-2)]
E ( >- J ) = 1 1J2 + 1 1 [ J ( J -1 )] + ... + 1 1 [ J ( J - 2 )] + 1 1 [ J ( J -1 ) ] .
E (4 )
rii' (6 )
(8)
1
( 10 )
I-I
However, the crucial problem here is to assess Pii' and Pi. This
implies that an assumption needs to be made about the probability
distribution function both of the >-j and of the rii'j. In view of
the ordinal nature of the >-j, it is plausible to assume for the
whole relevant area a uniform density function for the >-j (see
also section 2). The motive is that if the ordinal weights vec-
tor, >-, is interpreted as originating from a stochastic weight
vector, >-*, there is, without any prior information, no reason to
assume that a certain numerical value of >-* has a higher proba-
bility than any other value. In other words, the weights vector,
~*, can adopt with equal probability each value that is in agree-
4.1. Introduction
- Problem definition.
- Problem presentation.
- Problem evaluation.
- Sensitivity analysis.
Population 51 49 16 27 30 43 100 19 21
Evacuation 1 2 1 1 1 2 1 1 1
Agricult at risk 2 2 2 2 2 1 3 2 2
Industry at risk 1 4 3 2 1 5 3 1 1
Fr water at risk 1 1 1 2 2 1 2 2 2
Cool-water quant 2 1 1 1 1 1 3 1 1
Cool-water qual 2 1 2 3 3 1 2 3 3
Air pollution 2 2 2 2 2 1 2 2 2
Thermal poll. 3 2 2 2 3 1 2 3 3
Indirect landuse 2 3 3 2 1 2 4 1 1
Landscape 3 1 1 1 3 1 2 3 3
Nat environment 3 1 3 1 2 1 1 1 1
National grid 2 2 3 1 1 2 2 2 3
Infrastructure 2 1 1 2 2 1 1 2 2
Coal-location 3 6 4 3 2 7 5 1 1
~opul~~i
Eu;cu;t i
A~ricul~k
lndustry~
Fr w;terk
Coo I-u;tt
Coo I'w;~
Air poll
IherH; I
Ind i recte
L~ndsnp
H~t enui
H;t ion; I
Infr;str
(o~l-Ioc
B B E F K H H H ~
A D E L E A 0 0 I
r R H E r A E P E
H S S U E 5 R D R
S 0 L U 0 L 1
E L 1 D
L
E
A
K
J
K
E
R "
G
The results of the Expected value and the Evamix method are
shown in table 4. Since these methods require quantitative
information on weights, the expected value method was used to
transform the ordinal priority information into quantitative
weights.
Uoawlakt.
2 Eem.
3 NO Paid ...
~
.D
E
~
5
~a Wia"ingtrm ....
0::
+ _ - - - - _ 4 - - - - -..... l,loordijk
Table 6 shows that Eems ranks higher then NO Polder for any
score assigned to the criterion Industry at risk. Rank 3 is
assigned to Coal water quality for alternative NO Polder. Table 6
shows that with a rank of 1 for this criterion the ranking of
Maasclak and NO Polder is reversed. Rank 2 for this criterion
results in a tie of both alternatives.
5. CONCLUDING REMARKS
REFERENCES
APPENDIX 2
o elsewhere
where c = (J-1)!J!
For "2' ... ' "J-l the following transformation has to be used:
Ralph E. Steuer
Department of Management Science & Information Technology
Brooks Hall, University of Georgia
Athens, Georgia 30602, USA
Lorraine R. Gardiner
Department of Management
Auburn University
Auburn, Alabama 36849, USA
min {cost}
min {imported crude}
min {high sulfur crude}
min {deviations from demand state}
414
Production Planning
Forest Management
Water Resources
concepts
Current Status
Future Directions
25. Switching
26. Preserving Convergence Achievements
27. Trajectory Applications
28. Polyscreen Workstations
29. Cognitive Equilibrium
30. Network Optimization Applications
31. Russian Interactive Research
32. Mainframes Versus Microcomputers
417
Z3
Z2
3. PAYOFF TABLES
zl z2 zk
1
z Z*
z12 zlk
1
2
z zZl z·
2 zZk
k
z zkl zk2 Z*
k
419
Nhere the rows are the criterion vectors resulting from individu-
~lly maximizing the objectives. If one is worried about al-
ternative optima, a way to assure that all row criterion vectors
~re nondominated is to lexicographically maximize the objectives.
z**
; z*; + E;
420
where the z~
1
are the maximal criterion values obtained from the
main diagonal of the payoff table, and the Ei are moderately
small positive values. An E.1 value that raises z~*
1
to the small-
est integer greater than Zj is normally sufficient.
where R.1 is the difference between z~1 and the minimum value in
the i-th column of the payoff table. Ideally, we would want to
use R'
i
of R.' s, but
s insteadl l
the R. values are typically not
available.
421
8. WEIGHTED-SUMS PROGRAM
k
max {c x = Zk}
s.t. XES {x E Rn I Ax = b, x ~ 0, bERm}.
max {),7 Z I cx
i
XES}
422
v-max{Cx z I XES}
called upon to select 10 ~-vectors that are as far apart from one
another as it is practicable to compute.
X E S
ex E R unrestricted, Z E Rk unrestricted
f;(x) = z; i E K
XES
ex E R unrestricted, Z E Rk unrestricted
425
a. E R unrestricted, Z E Rk unrestricted
max { f j ( x ) zj }
s.t. fi(x) ~ ei i h
XES
is formed. From the solution of this program and the new current
solution, another line of search is defined. Then the most pre-
ferred solution along this line of search provides the current
solution for the next iteration, and so forth.
z, ~ e, i E H
1 1
Z, + d~ ~ t, i E I
1 1 1
Z, - d~ ~ u, i E J
1 1 1
fi (x) Z, i E K
1
X E S
d- ,d+ ~ 0
a. E R unrestricted, Z E Rk unrestricted
where:
(a) L = 1
(b) all) = 0, ~l) = 1 and "(1) 0
(C) G I = J = <I>
(d) H =K - {j}
min {-z.}
J
s.t. z. ~ e. iE(K-{j})
1 1
fj(X) Zj i E K
XES
Z E Rk unrestricted
(a) L = 1
(b) all) = 1, ~l) = 1, and "(1) o
(c) G K, H = I = J = <I>
min {a. - f ~. ZJ
1
s.t. a. ~ \ (Zj ** - Zj) i E K
f. (X) = z. i E K
1 1
XES
a. E R unrestricted, Z E Rk unrestricted
25. SWITCHING
max { c T X I XES}
s.t. S {x E Rn I Ax = b, x ~ 0, b E RID}
If, apart from upper and lower bounds on the variables, A has (i)
no more than two nonzero elements per column and (ii) the nonzero
elements are either l's or -l's, the LP is a network. Because of
the special structure of A, networks can be solved using high-
speed solution procedures.
REFERENCES
SUMMARY
menus for choosing the available actions and dialogue boxes for
exchanging information with the user in order to facilitate the
decision maker (DM) tasks.
The first one is the triangle filled with the regions correspond-
ing to each of the already known Pareto optimal vertices. Eventu-
al constraints on the variation of ~ will also be presented. The
second one presents a projection on the plan f1f2 of the Pareto
optimal solutions already computed. The area of each region of
the triangle and the Tchebycheff distance (L.) to the "ideal
solution" corresponding to each Pareto optimal extreme point
already computed are also presented. The so-called "ideal solu-
tion" is the one that would optimise all objective functions. Of
course it is not feasible in general!
min a
K, a
s .t . (f ~ - Qj K) b j ~ a, j = I, • • . ,p
x E Fk
a ~ 0
The process ends when the weighting space has been reduced to
a small enough region for a final solution to be identified, thus
converging to the optimum of the implicit DM's utility function.
The responses of the DM in the dialogue phases must be coherent
with this implicit utility function. The authors admit as possi-
ble implicit utility functions, not only weighted sums of the
linear objective functions but also more general pseudo-concave
functions. In this case the method only guarantees the conver-
gence to the greatest utility Pareto optimal extreme point.
2.1.1. Cost
2.1.3. Reliability
2.
f3=1. 9792 * 105
\. 1
".
3.8986 f3 =2.1454* 105 3. f3= 1.2454* 105
;..{ 2.87 a9-f-_ _ _--r-"1_ _ _..,---..--. f 1
3.5954 6.6866 8.0 (* 107)
OPTIMUM %
DO
Fl F2 F3
Figure 4
The first run of this model using the weights ),1 = ),2 1/3
ed to the optimal solution of f2 (solution A) (fig. 5).
460
Fl
min
------OpU6
F2
max Opt.96
~:o:r-------
I~A
F3
min Opt.96
........
~
.....
.....
...........
,/'
Figure 5
There are two adjacent vertices, but only the one that makes
us to move to the lower right part of the triangle is distinct
from the present one in the sense indicated in (Steuer, 1986,
chapter 13): two adjacent vertices are distinct (meaning they are
different enough for a pairwise comparison) when at least for one
of the objective functions the corresponding values differ more
than 10%.
Figure 6
462
f1 1.0000
f2 0.0012
f3 0.0010
).,.2
Figure 7
Note that the DM mayor may not accept the trade off that
leads to the elimination of the constraint of opposite direction
previously introduced and its replacement by the new constraint.
Let us suppose that the response is positive. Then the method
selects a new weighting system which satisfies the constraints
signalled in fig. 7, and calculates the corresponding non-domi-
nated vertex (fig. 8).
463
OpUS
-----
Opt.9IS
~:r------
A B
Opt.9IS
-----
A B
Figure 8
Nuclear 1 Coal 3
3673.3 441. 3
Fl
min
-----OPu.s
c
F2
max Opt.96
F3
min Opt.96
A C
Figure 9
f1 = 37700264.5
f2= 5368.7
f3 = 154082.0
>..~
Fl
min Opt.96
~---
F2
max
A C D
F3
min
--- Opt.96
3 A C D
4
Figure 10
Note that despite these two vertices being distinct (C and D),
the corresponding objective function values are not very differ-
ent. Let us suppose that the DM prefers the new vertex to C. The
constraint 3 (fig. 10) is generated and a new set of weights is
computed which satisfies the constraints 1, 2 and 3 leading to a
new candidate extreme point whose objective function values are:
f1 = 37045318.1
f2 = 4712.0
f3 = 184587.0
Coal 1 Coal 2
1370.0 1826.3
This procedure, although being less rigid than the one pro-
posed in the Zionts-Wallenius method, contradicts the general
philosophy of TRIMAP. Here we seek to widen the flexibility of
the dialogue with the DM. However, the introduction of this type
of procedure (or the use of the Zionts-Wallenius method) may be
justified to explore some specific regions of the weighting
space. It should be noticed that in many real cases (such as the
468
(c) The experiments carried out with the STEM method confirm
its main advantages: simplicity not only computational but also
from the point of view of the OM's comprehension. However, it is
not always easy for the OM to state which objectives are to be
relaxed and by how much.
ACKNOWLEDGEMENTS
REFERENCES
Pekka Korhonen
ABSTRACT
1. INTRODUCTION
y y(x) Ax,
min E
subject to: (2.1)
Cx + wE ;::: b,
x ;::: 0,
E unrestricted,
wi
{ : 0, i f i refers to a constraint,
0, i f i refers to an objective.
they are tired the next day. Therefore, if possible, I would like
to avoid it. Finally, product 3 is very important to a major
customer, and I cannot totally exclude it from the production
plan" .
V I G
+++++++
Menu
4+++
The model is entered in three phases. Each phase has its own
spreadsheet format. Using arrow keys, we can move from one field
to another. The use of any other key activates the field, and
makes it ready for our input.
FIO:Exit to Menu
Ranges for flexible goals are needed for defining the lengths
of the bars on the screen. To obtain a maximum benefit from the
visual feature of PARETO RACE, we recommend careful consideration
of the ranges (even though they are automatically updated by the
system). Sometimes it is advisable to respecify the ranges during
the race. In the use of the program, experience is a good teacher.
4. PARETO RACE
F5: "Brakes"
Reduce speed.
Num: "Turn"
Change the direction of motion by pressing the number
key corresponding to the goal's ordinal number once or
several times.
FlO: "EXIT"
Exit to the main menu.
484
Pareto Race
Pareto Race
5. APPLICATIONS
columns), but there clearly exists a need (and the potential) for
solving large-scale problems using VIG.
We also cite several studies that have been made, that have
not yet resulted in concrete applications.
The waters of the New Ycrk Bight are among the most intensive-
ly utilized in the world. Concern for water quality in this
489
6. CONCLUSION
REFERENCES
Nr. 1r 3-15.
Korhonen, P. (1987b), VIG (A Visual Interactive Approach to Goal
programming) - user's Guide.
Korhonen r P' r Narula r S., and wallenius, J. (1988)r "An Evolu-
tionary Approach to Decision-making r with an Application to
media Selection"r (unpublished manuscript).
Korhonen r P.and Soismaa r M. (1988), "A Multiple Criteria Model
for Pricing Alcoholic Beverages", European Journal of
Operational Research, 37, 165-175.
Korhonen r P. and Wallenius, J. (1988), "A Pareto race", Naval
Research Logistics, 35, 615-623.
Korhonen, P. and Wallenius, J. (1989a): "On Using Qualitative
Data in Multiple Objective Linear programming", (Forthcom-
ing in European Journal of Operat ional Research).
Pukkala, T., Saramaki, J., and Mubita, O. (1989), "Management
Planning System for Tree Plantations. A Case Study for
Pinus kesiya in Zambia", Working paper, The Finnish Forest
Research Institute, Joensuu, Finland.
Sakawa, M. (1983), "Interactive Computer Programs for Fuzzy
Linear Programming with Multiple Objectives", International
Journal of Man - Machine Studies, 18, 489-503.
Sprague, R.H. and Carlson, E.C. (1982), Building Effective Deci-
sion Support Systems, Prentice-Hall.
Steuer, R. (1986), Multiple Criteria Optimization: Theory, Compu-
tation, and Application, John Wiley & Sons, New York.
Wallenius, H., Leschine, T.M. and Verdini, W. (1987), "Multiple
Criteria Decision Methods in Formulating Marine Pollution
Policy: A Comparative Investigation", unpublished manu-
script.
Zimmermann, H.-J. (1985), Fuzzy Set Theory and Its Applications,
Kluwer-Nijhoff Publishing, Boston.
A PERSONAL COMPUTER VERSION OF THE MCDA APPROACH STRANGE
1. INTRODUCTION
R
:E qr 1.
r=l
3. AN ILLUSTRATIVE EXAMPLE
First objective
Second objective
Constraints
r = 1 (ql = O. 25 ) :
r = 2 (q2 = 0.50):
r = 3 (q3 = 0.25) :
",
u
II
~ ••
'n
'D
(ell
\
\ 'y"
y' "
\
,,/ \
\
\
.'
.'
",
4. PC-VERSION OF STRANGE
"KIN" Ohjective ti 2
Scenario # 2
Pl'ohahility : 0.500000
Scenal'i 0 D2
Constraint ~ 3
Fig. 2 and 3 show of the screens used for inputting the illus-
trative example.
B) Calculation sequence
The user has of course the option to directly aim for the
first compromise y(m) for m = 1. The illustration is in fig 5.
499
Menu _mt'!t1
Figure 4: The pay-off table
" 1"1' I 1 IKE FIRSI mmnlSE : I .- 1.'1 " '" i ' i ,n.': I
HI! v, I,!: (,i tr,t om i; t,I!1 m
f, 1 : un
xz : Ui ~
UI : (1.900 U1 : 1,m
U!: HOI) U~ : U'lO N? : (i j : U!: : U~(I
Menu miMi!'
1RSI CDnPAD"m I I· I .. . i I. . · . . I I • •
f fl....·\........·,....................................~·············································
COMproMi se nO 1
CR ITER !ON 31
Absolute Value: 3.696 .
Ua~iability Range: [ 1.000,11.809 1
lsi t OK ? No
4 Graphics
Uisualisation
, ,
'n!H
" "",,: I .1 A8S0Lun UALUES :aF, IKEICiI1ERIA I, I.. ','
1 II 2 1Z : 11
"i
~ 1~
:' ,
MM'·
. :1
XO .2!O Z\.in II .m li .m !.(i!(' ?. ~S€
o.~V(I zO .m Il.m IUft ~ . ~~1 ,Z: ~
0.3!') 10 .0ti IU~ I!.m UIZ .m
o. (II) IU('(' 11 .77(' I~ .t~~ i.7H ! .W
10 . !~ 1!.101 I(' .m I$ .m I . Oi:~ em
~o.s: lI .m l .m It .m UI t.m
IO .H1
{1 . tO~
11 .1\1
17 .m
i .m
i .m
IUil
I UO ~ ,,.'
I.m
. J .."
Lm
I .:.t
om 11 .m t.m Ii .• € U ~ . I)n
, 11, 1'·' .1 11 I! I ,II' . '.: ,., 'UftR188UI II. 111 H I ' , I" " ,., ',. . I' II '
" 1 :~ : ~ 1 .0 I i ~.
~ ~ ;, i 7 U i U) i 1(' 11 i 1:
n .H(1 ~ H~ l ll: (t . I)O~ L!l& 1.m 0.00(' o.m 0.000 U) (0.00(' 0.000 O.MO 0,1)0 (t .(uj~
(UO (, 1.m : .(·!I 0.000 I.m Ul1 uoo 0.000 0.000 1.:11 U,)(· 0.000 u~o I) .000 1),0')(0
~J ~ (I 1. ~1 : .11. 0,('00 I.m Ut!. 0.000 1).00(' (, .000 U!:~ 0.0(1) 0.000 0.000 I) .0)00 '),0')0
I) .\0
)(1 -:'1) ?m 0.1,0') ~ J(I' U~. 0.000 0). 000 UO(I :.H1 0),00(' OUOO 0,(000 i) .'NO 0.00(·
JO . ;: ~ I(o! ~ . ~i7 f'.OOI) { (IOO 1.513 HO(o uoo 9.Of!!0 l .m 0.000 1),(000 0.00(0 (I 000 (0.1)('0
I IL :~'1 <.m L1Si 'uoo Ull !.O(oO UM '),000 OJH ~ .00(, 0.(000 (0.000 0.('00 0.')00 1) . (I{1 ~
l(o.m 1m ~ . ~]1 O.HO U~i $.II, 0.000 0j,900 o.m 1.112 0.000 0.000 (1 .('00 a.HZ (I . ~(II)
U03 ; .: l~ ::.\1(, 0.000 Ull I.m 0. 101 0).000 o.m I .m o.I!O(o 1) .000 I) .000 O. n 0.00)('
o.m t .€S2 2.m ol.OOO U!9 I.m (l, i03 i!.OO~ (iJlj I .m (t,Ij9~ 0.000 0.000 o.m <) .( 0
) (1
o.m I " ,. : .118 oJ.OOo) l .m !.O1: o.m U(lO 1,~s~ ~ ,1)75 (•. (")(0 0.0(1,) 0.000 LV> (1 .0(1(.
O.i ~: !.nI 3m \' .0(10) ]!10 3.m om '). 0)«(' IS!! ! .!~~ (0,000 UG') (0.('00 1.m (•. ['00
0.$01 un ~" SO ,) .0 01, , .m !.m Hot 'LOM 1)1 !.7H O.l!OO (1 ,1)(11) (1.0(") Uli
1) .1)1) (1
u s; ! .il( u.(I( (1 .00') U~l 10.m 0.110 0.000 i.(ln ~ .037 0.0,00 ').000 (1 .00(' :.o~ (' .0')['
UO) ,. ~ ).u
·tl i .m OGOO U!I IO .m o. m 0.000 2.H ! (.m (1 .000 9.0( 0) '),0('0 . }t '
... <- 0.000
11.0('0 ! .m ?72t o.m U!I 10.m 0.90l 0.000 Z.?H UI! I) .Olj(1 0).00(' 0.0)0 Zji; oU,)')
i i " : I I •• ' : ,
, ' , I: .. , m !HE (WIRIII; I, I I I ' , , ' 1',0 '. I·: ,', ..r.• .
1.00
l) j~
(I. ~ 1
(0.1 (
tl,Si
!) .7i
(. ):.
o) .i f
o) ,i
(I . ~~
(I .! ~
(I . S(~ ., :'1 •••••.
I) . ~~. #
. --,_04
(I . Ovt-"T"....,.-.,.....,......,.-'T"'"...,.;...,-"'I""'-,.--,.....;,......,.-~..,.."""'I'-'!"'
0.00 O.Oi 0.11 0.11011 US O.ll U! 0.15 0.50 O.Si O.iI U! (' .n U* HI U! U S 1.00
The user can select any value for the level of relaxation and
call for a vizualisation of the corresponding solution. The
previous compromise is displayed for comparison purposes in a
separate window on the same screen (fig. 14).
5. CONCLUSIONS
5 iP";; ·,~···.A.
---------------~~~-
0,1, ---------------~~~~~~~~~~~-------------------
Z 11 Z 12 Z 21 Z 22 231
ulDllll I
the Y' lues 01 the m i.t, Ie! m
~I: Lm
Xl: l,m
UI : 0.000 U!: t,m Ul: UB U~: (I,m
U~ : (1.000 ut: o,m U7: Hil us : (1.00(1
U9 : HOO UIO : 0,')00 UI1 : O,H) UI2 : (1.')(10
Menu 1IfiIilIMIrI
REFERENCES
APPENDIX
R
zK+1,sK+1 (Y) = '=1 qr(Wr )
s .t. Y EDdY)
(r)
min zk (Yk ),
rE{l ••••• R} sk sk
'tt"k\ with
K+l mk s· I I ck S II
k k
:E
k=l
and ~ S
k
minimize 8
Sk
S • t. :E Pk sk (ck S Y - ~ S ) 11k s :-: ; 8 k = 1, . .. I K+ 1 (2 )
sk=l k k k
of variations [ ] ;
minimize 8
\
s.t. :£ P (c k Y - ~ ) n: k ~ 8 k=ll'" k+ 1
k\
I
\=1 \ \ \
C(k G .)* •
"1<
Y = M
k~
)
* + >- (m( kSk ) * - M ) )
-l<sk *
>- ~
a
>- ~ A (3)
where >-
a
and A are given respectively by:
M + >- (m M) - z(m)
k~ )* 0 (ksk)* - -l<sk)* - k~ )*
and n:(k~)* set equal to 0 in the course of the procedure (see (1)).
511
In this way, the value accepted by the decision maker for the
relaxed criterion (k \)* will not be increased during the next
interactive phases. This limits the number of interactive phases.
Remark
n:(k~)* is set equal to zero during the next iteration, but set
Jaap Spronk
Erasmus University
P.O. Box 1738,
3000 DR Rotterdam
THE NETHERLANDS
1. INTRODUCTION
b = f (x) for ( 1)
x E R, R = {x I h(x) $ h},
516
A.x = b (2 )
x E R, R {x I B.x ~ h},
(3 )
under A.x y+ + y- b,
where y+, y- 2': 0 and y~ .y~
1 1
o for i 1, 2, ... , m
and x E R.
under A.x - y+ + Y = bl
where y+, Y- ~ 0 and Y~ 'Y~ o for i I, ... , m
and X E R
under A.x - y+ + Y- = b,
where y+, Y o for i 1, ••• , m
and X E R
the n: ... ,
under A. x - y+ + Y = b,
where y+, Y o for i 1, ••. , m
and X E R
and under f. (y+, Y-) 1, ... , p-l.
1
m
min Q { ~ (5)
i=l
+
Yi = {Ia 1.. x - b.1
1
+ (a 1.. x - b.)}/2
1
for i I, ••• I m (7a)
Yi { la 1.. x - b·1
1
- (a 1.. x - b.1 ) } / 2 for i 1, '" .. , m ( 7b)
519
(8)
min Q (9)
(10)
(11 )
Obviously both (9) and (10) or its weighted forms, may serve
as the distance function f(y+, y-) in (3). Moreover, subtracting
(7b) from (7a) gives:
I U
r. :E e .. x. +:E (14 )
J i=l lJ 1 u=l
where e lJ.. denotes central activity i's pro unit use of resource j
and r~ represents the amount of resource j distributed to unit u.
J
Obviously, e lJ.. may be negative in the case that central activity
i adds to the amount of type j resources (for example, borrowing
increases the amount of funds available). In addition, it is
possible to define for each central resource j, an activity
variable measuring the unused amount of this resource. This type
of variables can be actively manipulated in order to guarantee
the required flexibility in the plans to be developed. Of course,
the central decision unit is not completely free in choosing the
values of the instrumental variables. We assume the choice of
instrumental variables restricted to the feasible set X. Thus
x E X, with x the vector of instrumental variables.
u
gm (X 1(u) ... xi(u) ... xI(u)), for m = 1, ... ,M; and (15)
gm(u) (xl(u) xi(u) ... xI(u)), for m(u) = 1, ... ,M(u).
The
choice of the local instrumental variables is restricted
by local constraints, i.e. Xu E Xu with Xu the vector of unit u's
instrumental variables and X the convex feasible set described
u
by local constraints. In addition, the choice is restricted by
the amounts of resources r~, j
J
= I, ... , J, available to unit u.
This leads to the constraints
I(u)
L S i(u)j xi(u)' j 1, ... , J; (16 )
i(u)=l
so, whether it adopts x U2 and so on. For each project xU'k ' lower
level management has calculated the minimum contribution to each
of the central goal variables. Thus, assuming that central man-
agement receives a series of plans from each of the lower level
decision units, we can reformulate formula (13) for the m'th
central goal variable as
u k
gm
c
gm + L L U
uk
gm x uk for m 1, ... , M (17 )
u~l k~l
I u k
uk
r. L 6 .. + L LU r. (18 )
J 1J J
i~l u~l k~l
REFERENCES
TawIik Jelassi
Technology Management Area
INSEAD
Boulevard de Constance
77305 Fontainebleau
FRANCE
Gregory Kersten
Decision Analysis Laboratory
School of Business
Carleton University
Ottawa, Ontario KIS 5B6
CANADA
and
Stanley Zionts
Department of Management Science and Systems
School of Management
State University of New York
Buffalo, New York 14260
U.S.A.
ABSTRACT
CONTENTS
1. Introduction
6. Concluding Remarks
7. References
539
1. INTRODUCTION
by the group. More about the issues discussed here can be found
in Keeney and Raiffa (1976) and Kersten et al. (1988).
Hence, one can incorporate the aspirations which may cause the
objective functions or the utility functions to be defined for
the aspiration levels intervals. The reservation prices (the
lowest/highest acceptable values), shifts in negotiation tactics,
and context dependent concessions can also be introduced. The
paradigm is, however, that a Pareto optimal solution is desira-
ble, and that the decision problem can be expressed as a problem
of identifying acceptable Pareto optimal solutions.
identified some sixty GDSS and NSS. In the GDSS area, significant
experiments took place in North American university laboratories,
in particular at the University of Arizona where the PLEXSYS
system was developed and at the University of Minnesota, home of
the SAMM system.
[OJ
While Board Wall Mounted Projection Screen While Board
Facilitator Console
and Network D
DO 00
File Server Control
Room
D= DO 00
DO D 00
Break-Oul ~
Workstations
Room
/
DO 00
Barco
Projector
DO 00
Break-Out
Room
D
=
DO 00
D D D D
~
D CJ CJ CJ
Break-Oul
Room
D
= Slorage Break Area
/
Figure 1. The University of Arizona Small GDSS Facility
Conference Table
:I
.
Pri ....ate
Terminal
6. CONCLUDING REMARKS
REFERENCES
Anson, R.G. and M.T. Jelassi (1990), "A Development Framework for
Computer-Supported Conflict Resolution", European Journal
of Operational Research, Special Issue on Group Decision
and Negotiation Support Systems, forthcoming.
Axelrod, R. (1984), The Evolution of Cooperation, New York, N.Y.:
Basic Books.
Belton, V. and S. Vickers (1990), "Use of a Simple Multi-Attribute
Value Function Incorporating Visual Interactive Sensitivity
Analysis for Multiple Criteria Decision Making", in this
Volume.
Bui, T.X. (1985), "N.A.I.: A Consensus Seeking Algorithm for
Group Decision Support Systems", IEEE Conference on Cyber-
netics and Society, Atlanta, GA., 380-384.
563
Gunter Fandel
Fernuniversitat Hagen
D - 5800 Hagen I - WEST GERMANY
further suppose that every decision maker n controls one and only
one decision component an of a vector a, where an E AN and a E
A = Al X . . . x AN; An designates the set of decision alternatives
of the n-th decision maker.
(AI)
(A3 )
(A4)
N N
{u* E U I 'IT max 'IT
n=l a E A n=l
S N-S - - S N-S
max min I: un(a,a )forallScN,a=(a,a )EA;(l)
as Ji-S n E S
(S-l)! (N-S) !
~n(v) = I: [v( S) - v(S - {n} ) ) ,
~N N!
(2)
nE [N), S II S II and N II N II.
LQ = {u* E U I u* ~(v)} .
N
U {u I E un S; v(N)} (3 )
n=l
according to (2).
(7)
a'
Allowing (8), the supergame strategy G = (a', a', ... ) repre-
sents a non-cooperative equilibrium if it fulfils the condition
00
E (9 )
t=l
regrouping, is equivalent to
a
n
[un (a' ) -un ( c)] > u n (a',
n b n) -u n(a'), n E [N] , ( 12)
1 - a
n
L={u*EU
Q
u*=u(a*) EP(~') and vn(a*)=vn, (a*) for all n, n' E [N]}
u (a*, b) - u (a*)
_-"n_.;.;.n_-"n_ _---"n___ , n E [N]. (13 )
579
o (14.1)
C
n
f (UN) =
n
of/aun uN n E N: (14.2)
S
un N-S (14.3)
tN
n'
) for all n, n' EN: (14.5)
1:: 1::
n ES n' E S'
(14.6)
S S' S S'
= max min c u (a , a
1:: ) - 1:: c n' un' (a , a )],
nES n n n' E 5'
S' S'
a EA
t~), n,kdl:
L=
Q
{U*EU I u* ~ z(t) ~ y}.
n E [N] . (16 )
(17)
( 18 )
(19 )
( 20)
585
and
o otherwise, (15' )
n,n' E {1,2} and nln'
with
2 2
'Tr (Ur-Ur ) > 'Tr [U~(t+l)-Ur]'
r=l r=l
(21 )
for the connection between the actual contract and the naive
solution where the perturbation variable Et is to be normally
distributed with the expected value zero, one obtains the regres-
sion line
wt = 0,96 wt (21')
U
Initial claim Initial offer Actual Naive
of IG Metall made by contract solution
Gesamtmetall achieved
wt W* W d~(w +w )
~ t t t -t
The close connection between the naive solution and the actual
agreement which has been described in the last section calls for
an enquiry of the question to what extent the agreement points of
the wage negotiations are interpretable in the sense of the
cooperative Nash solution, since in case of linearly transferable
utility this game-theoretic concept divides the bargaining cake
available between two parties into equal shares. For an analytic
reconstruction of this possibility of interpretation, let us
assume in the following that the cake to be divided between IG
Metall and Gesamtmetall is each time defined by the difference of
the changes of the wage sums resulting from the initial claim of
the union and the initial offer made by the employers, thus reads
(22 )
Metall and Gesamt-metall are assigned the indices n=l and n=2
respectively, the utility functions remaining constant over the
years with respect to the shares can be written as follows:
wt - ~
a. (23)
-
w t - ~
and
w- - W
t t
b. (24)
-w -
t ~
w -w
-
w -w
t -t b._t _ t_
max u t = (u lt - 0) (u 2t - 0) u lt • u 2t = a.-_-- (25)
w -w
-
w -w
t -t t -t
a.b
(wt -w
-t
) (wt -wt ) C(Wt -!'4) (Wt-Wt )·
dU t
c( -2wt +wt+w
-t
) o (26 )
dW t
w =w =5 W =12· w =5
t -t t '-t
t=1969
Ut=c(Wt-~t) (Wt-W t )
3
/
u= (0.0)/
2 3 4 5
L
While any points above L are not feasible since for them the
union's claims are always higher than the offers made by the
employers, which means that no agrements can be reached there,
the points below L represent a waste of the cake to be divided.
From the condition of optima~ity w~ - ~ = t - w~ according tow
(26) follows the Nash solution w~ = 8.5, with 8.5 - 5 = 12 - 8,5
= 3.5, which in Figure 1 is near point C; the actual contract
2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
1 2 1 2 1 2 1 2 1
'\ 6
7 12 15 11 18 11 8 9.5 8
10 8 9
7.9 14 7 8 9.5 8 6
10 7 9 6.5 11 15
5.67 5.5
1) 1)
hese values are not considered in the verification according to the theorem by zeuthen
c.n
~
594
Let Rt = {rt I r t =1,2, ... ,rt} denote the set of round indices in
year t, where r t indicates the number of bargaining rounds in
this year required for a (possible) agreement. Let w~(rt) charac-
terize the proposal made by partner n, n E {1, 2} in the round
rtER t of year t; in this connection it should be pointed out that
according to the arrangement described in the last paragraph
index n=1 (IG Metall) can appear only in the case of odd and
595
(28)
must hold.
1, if
{ (30)
o otherwise
Table 3
Result of the verification with respect to Zeuthen rationality
of the concessions made by IG Metall and Gesamtmetall during the
wage negotiations 1961-62 to 1979
~i!r~r~~~~~f~~f:~~c~~
~~t~~;i!±±i~9]if±£:il~~~I;~;j
C!:JI 1969
I 1970
I 1971
I 1974
f (l"t l 1 1 2 1 2 2 1 2 1 2 1 2 1 2 1 2 1 2
!(f'tl 1 0 1 1 1 1 1 1 1 1 o 1 0 1 1 1 0 1
LJ 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
C!:JI 1975
I~ 1977
I 1978
I 1979
I
f't 3 4 5 6 4 5 6 4 5 6 7 8 4 5 6 7 8 4 5 6
f (l"t l 1 2 1 2 2 1 2 2 1 2 1 2 1 1 2 1 2 2 1 2
YCf't) 1 1 0 1 1 1 1 1 1 1 0 1 1 1 1 0 1 1 1 1
47
f ( y= 1 ) 0.82 (31 )
57
holds.
597
Table 4
Derivation of the results in table 3 for the year t=1969
H: f ~ f = 0.70 ( 32)
o 0
since the test variable f exceeds the upper limit c of the zero
o
hypothesis (Wetzel, 1973, p. 195 ff.), i.e.
2) do the wage claim curves of the union and/or the wage offer
curves of the employers take decreasing convex and/or
increasing concave courses?
b1
1 1 t
wt (r t ) at .rt + Et (35)
and/or
b2
2 2 t
wt (r t ) at .r t + Et ( 36 )
enough, the regression has shown here all wage ofter functions of
the years under review increase and follow concave courses on
account of a~ > 0 and 0 < b~ < 1, thus the wage ofter behaviour
of Gesamtmetall can be approximated fairly well by a function
whose curves follow these courses. While consequently the empir-
ically founded proposition concerning the existence of decreasing
convex concession functions is supported by the regression
results for Gesamtmetall, the same cannot be said with respect to
the case of IG Metall, sinc~ expression (35) implies convex
courses of the curves with wage claims decreasing, but the
preliminary reflections mentioned did not lead to such unique
results in this sense.
Table 5
Regression results for the annual wage claim and wage after
functions of IG Metall and Gesamtmetall.
t IG Metall Gesamtmetall
I 2
at b it r at b2 r
t
1961-62 10.89 -0.22 -0.77 0.76 0.97 0.99
1963 8.18 -0.19 -0.97 2.12 0.43 0.96
1965-66 9.39 -0.24 -0.74 1. 65 0.72 0.93
1968 7.23 -0.l3 -0.93 3.54 0.17 0.95
1969 12.34 -0.15 -0.81 3.46 0.40 0.88
1970 15.66 -0.16 -0.74 5.31 0.42 0.98
1971 10.76 -0.23 -0.98 3.69 0.32 0.98
1974 18.04 -0.24 -0.99 6.80 0.34 0.98
1975 10.73 -0.32 -0.96 5.55 0.11 0.99
1976 8.45 -0.21 -0.74 3.28 0.26 0.96
1977 10.00 -0.18 -0.84 3.80 0.28 0.96
1978 8.52 -0.21 -0.85 2.34 0.42 0.97
1979 6.28 -0.17 -0.74 2.41 0.34 0.98
1
at /3 2 • Wt + Et and/or at2 /3 3 • Wt + Et (37), (38)
Al
at 1. 03 wt and/or a 2
t
-
0.78 w t (37'), (38')
using such wage proposal curves. Yet, due to the results of the
regression analysis it can reasonably be assumed that certain
rules do exist. This should encourage further investigations with
a view to clarifying the rationality of wage bargaining process-
es. For, even irrespective of the methodical reflections made
here, the naive solution and the data compiled in Table 2 allow
the assumption that the decision making of the two bargaining
partners shows certain features of a heuristic procedure: On an
average the partners expect about four double rounds of wage
bargaining until an agreement is reached, and knowing the initial
proposals they adapt their concessions accordingly, so that
eventually the commonly defined bargaining cake is divided almost
into halves.
t=1971
12
~!= 10.76 r~0.23
11 e/
10
3 3.69 r~·32
2 3 4 5 6 7 8
Figure 2. Wage claim and wage offer curves of IG Metall and
Gesamtmetall in 1971
605
REFERENCES
ABSTRACT
1. INTRODUCTION
2. MODELLING FRAMEWORK
To be rich you have to have money and there are three ways to
get mo n e y : ear n Q..L S tea I Q..L b 0 r row mo n e y. To ear n mo n e y you
have to find a job and keep it, and to find a job you have to
prepare your CV and look respectable and visit an employment
agency. To keep the job you have to work hard and be coopera-
tive. If you steal money you have to find an appropriate place
and you must not be caught. If you decide to borrow money you
can ask you r f r i end, Q..L you can see a ban k rna nag e r and fin d
security and look respectable.
AND VISIT_AGENCY.
KEEP_JOB <= WORK_HARD AND BE_COOPERATIVE. (1)
STEAL_MONEY <= FIND_PLACE AND NOT GET_CAUGHT.
BORROW_MONEY <= ASKJRIEND OR ( VISIT_BANK AND
~
A
EARN_MONEY STEAL_MONEY BORROW MONEY
FIND_JOB KEAEP
peaUGHT
_JOB FIND_PLACE
~
¢NO_NAME -
FRIEND
WORK_HARD
VISIT BANK
BE_COOPERATIVE - FIND_SECURITY
LOOK_RESPECTABLE LOOK_RESPECTABLE
(3 )
cates PREPARE_CV and WORK_HARD to true and then solve the problem
(1). The rule used in this conditional bounding is as follows:
GET_RIO_OF_BAD_HABITS] .
hand side in model (1). Thus, restructure rule (8) means that if
someone is caught then the statement "In order to steal money you
find place and you are not caught" is replaced with new statement
"In order to get low sentence you get a lawyer and you look
respectable". Note that one restructure rule may introduce
several new rules as in (9), and it also may introduce new predi-
cates.
(10)
GET_RIO_OF_BAD_HABITS.
GEUA:f!\ LOOK_RESPECTABLE
LONG_TERM_GOAL (police).
SHORT_TERM_GOAL (police) <= NOT LOSE_LIVES (POLICE).
ERADICATE_TERRORISM (police)] OR
INTRODUCE_TOUGHER_LAWS (police).
LOSE_IMAGE (police) <= CASUALTIES (police) OR
YIELD_WITHOUTJIGHT(police, terroris~.
ERADICATE_TERRORISM (police) <= NOT RELEASE_PRISONERS (terroris~ OR
MAKE_RECRUITING_DIFFICULT (police).
MAKE_RECRUITING_DIFFICULT (police) <= KEEP_LOW_PROFILE (police) OR
RECOGNIZE_MOVEMENT (pOlice) ] ..
KEEP _LOW_PROFILE (police) <= RANSOM_DEMAND (terrorist) AND
LOSE_LIVES (police)] OR
LOSE_LIVES (police) ].
GET_INFO (terrorist) <= STAND_TRIAL (terroris~ OR
RELEASE_WOMEN(terrorist, crimina~ ::= true AND KILL_MEN (terrorist, crimina~ ::= false
¢:=
OBTAIN_RANSOM (police) ::= true AND STAND_TRIAL (police) ::= false AND
FOOD_SUPPLY (police)) ::= any.
RELEASE_MEN (terrorist, crimina~ ::= true
¢:=
RELEASE_WOMEN (terrorist, crimina~ ::= true AND STAND_TRIAL (police) ::= false.
RELEASE_WOMEN (terrorist, psycho) ::= true
¢:=
->
[ PSYCHO]RESSURE (police) ¢= NOT TV_COVERAGE (terrorist, police) AND
STAND_ TRIAL(terrorist) ].
->
[ PSYCHO_PRESSURE(police) ¢= SHOW_OFJORCE(police) AND
POSTPONE_NEGO(police) ].
The editors allow the user to enter and edit working environ-
ment (knowledge base, meta-knowledge base). They work in both a
terse and a normal mode, accepting either complete it~ms (fact,
rule, restructure rule) to be manipulated, or allowing for the
construction of an item with expanded description in a natural
language. This description is later saved in a lexicon. The other
function of the editors is related to the possibility of setting
the elements of the position under consideration. If the user
wants a predicate to have a particular value, then he may signal
it by using the set option of the editor.
628
release_women(terrorist)
short_term_goals(police) lose_lives(police)
release_men(terrorist)
loseJmage(police)
eradicate_terrorism(police)
introduce_tougherJaws( police)
Figure 4.
NEGOPLAN
..
editor ...... .......... .. ....... gst ........................... ..... ..
nego graph .......... ...... ( CONTINUE ) preset fact ......................
Figure 5.
630
inflict_casualties(deception,police
inflict_casualties(attack,police)
release_women(terrorist)
lose_lives(policel
release_men(terroristl
Figure 6.
Appendix
O'.i {= r i ( [ O'.ij] j E Ji ), (1 )
s ([ r(O'.j) ] jEJ) = [ r(0'.1) = v' and / 0 r / not , •.• ,a n d / 0 r / not r(~) = v"]. (4 )
(5 )
s( r( [O'.d i E Itl
~ (7 )
p ( [O'.i {:: r j{ [O'.ij] j EJil ] i E Ig) -> p ( [0'.1 <= r 1 ( [01< h E 1* ) ] 1 EL) ,
634
References
1. SHORT DESCRIPTION
2. SCALE OF MEASUREMENT
- Insufficient (5)
- Level I
- Level II
- Level III
- level IV.
SUMMARY
ANNEX A
CANDIDATE C1
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 31 83 13 0 0 38 46 16
I I
A2 0 0 6 24 12 0 0 14 57 29
A3 0 0 3 44 9 0 0 5 79 16
A4 0 0 1 20 7 0 0 4 71 25
As 0 0 4 27 9 0 0 10 68 22
At; 0 0 1 8 1 0 0 10 80 10
A7 0 0 0 28 0 0 0 0 100 0
TOTALS 0 0 46 189 41 ***
I I
CANDIDATE C2
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 9 57 26 0 0 10 62 28
A2 0 0 9 25 16 0 0 18 50 32
A3 0 0 4 37 20 0 0 6 61 33
A4 0 0 2 26 8 0 0 6 72 22
As 0 0 0 26 10 0 0 0 72 28
I At;
I 0 0 2 6 0 0 0 25 75 0
A7 0 0 0 8 10 0 0 0 44 66
TOTALS 0 0 26 182 90 ***
643
CANDIDATE C3
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Ai 0 2 26 74 6 0 2 24 68 6
A2 0 0 7 37 13 0 0 12 65 23
A3 0 0 '1 29 44 0 0 1 39 60
A4 0 0 0 23 13
I0 0 0 64 36
I
As 0 1 12 31 11
I0 2 22 56 20
I
~ 0 0 2 7 4 0 0 15 54 31
A7 0 0 2 27 4 0 0 6 82 12
TOTALS 0 3 50 228 95
I *** I
CANDIDATE C4
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Ai 0 1 33 58 28 0 1 28 48 23
A2 0 0 13 26 21 0 0 22 43 35
A3 0 0 10 63 7 0 0 12 79 9
A4 0 0 5 34 1 0 0 13 85 2
As 0 0 27 31 0 0 0 47 53 0
~ 0 0 2 27 11 0 0 5 68 27
A7 0 0 1 27 10
I0 0 3 71 26
I
TOTALS 0 2 91 266 78 ***
644
CANDIDATE C5
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 2 27 32 17 0 2 35 41 22
A2 0 2 16 26 10 0 4 30 48 18
A3 0 5 14 28 12 0 9 24 47 20
A4 0 0 3 17 6 0 0 12 65 23
0 4 13 31 5 0 8 25 58 9
~
I I
At; 0 0 6 19 3 0 0 21 68 11
A7 0 0 4 9 0 0 0 31 69 0
TOTALS 0 13 83 162 53 ***
CAND IDATE C6
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 2 29 45 15 0 2 32 49 17
A2 0 3 23 21 8 0 5 42 38 15
A3 0 2 15 29 22 0 3 22 43 32
A4 0 0 8 21 3 0 0 25 66 9
~ 0 0 17 22 9 0 8 35 46 19
At; 0 0 0 9 1 0 0 0 90 10
A7 0 0 3 15 4 0 0 14 68 18
TOTALS 0 7 95 162 62 ***
645
CANDIDATE C7
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 18 77 7 0 0 18 75 7
A2 0 0 11 35 5 0 0 21 69 10
A3 0 0 '9 38 20 0 0 13 57 30
A4 0 0 0 28 8
I0 0 0 78 22
I
As 0 0 4 26 10 0 0 10 65 25
At; 0 0 0 15 1 0 0 0 94 6
A7 0 0 7 17 2 0 0 27 65 8
TOTALS 0 0 49 226 53 ***
CANDIDATE C8
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 30 85 2 0 0 25 73 2
I I
A2 0 6 32 13 9 0 10 53 22 15
I I I
A3 0 0 25 50 4 0 0 32 63 5
I I I I
~ 0 0 3 36 1 0 0 8 90 2
As 0 6 32 19 1 0 10 55 33 2
At; 0 0 8 27 1 0 0 22 75 3
A7 0 0 10 23 2 0 0 28 66 6
I I
TOTALS 0 12 140 253 20 ***
I I
646
CANDIDATE Cg
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 2 75 43 0 0 2 62 36
A2 0 0 13 7 40 0 0 22 11 67
A3 0 0 21 47 11 0 0 27 59 14
A4 0 0 5 26 9 0 0 12 65 23
As 0 0 1 24 35 0 0 2 40 58
At; 0 0 7 6 23 0 0 19 17 64
A7 0 0 0 34 3 0 0 0 92 8
TOTALS 0 0 49 219 164 ***
...
CANDIDATE ClO
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
I Al
I 0 0 46 64 1 0 0 41 58 1
I
A2 0 0 13 41 1 0 0 24 74 2
I A3 0 0 16 46 13 0 0 21 62 17
I A4
I 0 0 2 26 10 0 0 5 69 26
As 0 0 7 38 3 0 0 15 79 6
At; 0 0 7 8 1 0 0 44 50 6
A7 0 0 12 25 0 0 0 32 68 0
TOTALS 0 0 103 248 29 ***
647
CANDIDATE Cn
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
Al 0 0 57 26 1 0 0 89 9 2
A2 0 2 23 19 4 0 4 48 40 8
A3 0 1 .18 40 1 0 2 30 66 2
A4 0 0 8 19 12 0 0 28 65 7
As 0 1 23 15 1 0 2 58 38 2
At; 0 0 10 6 0 0 0 63 37 0
A7 0 0 14 10 0 0 0 58 42 0
TOTALS 0 4 153 135 9 ***
CANDIDATE CI2
FREQUENCIES
ATTRIBUTES ABSOLUTE RELATIVE (%)
S 1 2 3 4 S 1 2 3 4
I Al
I 0 1 25 38 19 0 1 30 46 23
I A2
I 0 3 21 15 4 0 7 49 35 9
A3 1 2 28 18 7 2 4 50 32 12
A4 0 1 5 21 1 0 4 17 75 4
As 0 4 15 15 3 0 10 41 41 8
At; 0 1 3 2 0 0 17 50 33 0
A7 0 1 10 11 0 0 5 45 50 0
TOTALS 1 l3 107 120 34 ***
A REPORT ON THE STUDY OF THE PORTUGUESE NAVY CASE
Marc Pirlot
1. GENERAL ATMOSPHERE
The best way for the reader to get a feeling of the case is
probably to have a look at the histograms reproduced in Appendix
1 (for the original data, the reader is referred to the presenta-
tion of the case in this book). Drawing histograms was actually
the first approach of many participants. For each candidate, Cl
to C12 , an histogram is drawn representing the relative frequen-
cies of the levels S, I, II, III, IV, obtained by counting the
assessments in each class without distinguishing between the
attributes.
Were there no «S» and no «I» at all, the above criterion for
breaking ties would yield exactly the same ranking as with giving
equidistant valuations to II, III and IV (the same for all
attributes) and calculating frequency-weighted means: the only
difference with the former ranking is an inversion of the posi-
tions of C7 and C1 • The mild dispersion of the assessments can
certainly be considered as one of the main stability factors.
3. ATTITUDES
Another trend was to use the data for testing the methods
rather than solving the problem. Group 5 experimented with the
valuation of the ordinal scale (S, I, II, III, IV) on an interval
scale (see the paper by J.-C. Vansnick in this book). Using the
two degrees of freedom of the scale, they attributed value 0 to
«II» and value 10 to «III». Afterwards they asked elevep. Navy
Officers to locate the other «levels» w.r.t. to 0 and 10, for
each attribute. For attribute A7 e.g., this resulted in Figure 1
which shows the relative position of level «IV» for the different
officers. Similar experiments showed that the same person gives
very different valuations to the «same» level IV depending on the
attribute: e.g. «IV» is 22 on A2 and 40 on AI. From the grea t
variability in the valuations the group concluded in the necessi-
ty of using an ordinal method in this case.
LUE
1 2 5 6 7 6 10 11
30
IV
20 -- -
10 - IIf
0 - -/I
The group begun with asking his Navy Officer for intuitive
valuation of the levels {Sf I, II, III, IV} which resulted in the
first row of Table 1. Then they asked the same Navy Officer to
use the EXPERT CHOICE software to obtain the pairwise comparison
matrix on the basis of which EXPERT CHOICE operates. This yielded
the valuations (as a ratio scale) given in the second line of
Table 1.
S I II III IV
Intuitive -40 o 10 20 50
A2
Intuitive 25 40 25 25 50 40 50
4. CONCLUSIONS
Though the case was not «solved» during the School, it has
been a particularly important element in the success of the
School. First of all it has been highly profitable for the par-
ticipants to think about a real situation and become aware of all
kinds of difficulties that can arise in concrete cases. Second,
the case was a good exercise for familiarising the participants
with the methods taught during the School and this motivated them
and maintained their attention at a high level during the whole
course of this intensive series of conferences.
s /I 11/ IV /I III IV
1 0.2
s 1/ III IV s /I III IV
s /I /II IV 5
658
s /I 11/ IV /I /1/ IV
380 AS5ESSHENT5
")2 A55E5SHENT5
/I 11/ IV
1/ 11/ IV
380 AS5ESSHENT5
275 ASSESSMENTS
Q3
/I /I //I IV
659
REFERENCES