Professional Documents
Culture Documents
Edited by
DENIS BOUYSSOU
ERIC JACQUET-LAGREZE
PATRICE PERNY
ROMAN SLOWINSKI
DANIEL VANDERPOOTEN
PHILIPPE VINCKE
....
~" Electronic Services <hup://www.wkap.nl>
Preface ix
The main contributions of Bernard Roy are focused on two broad themes:
We think that the influence of Bernard Roy is well reflected by the qual-
ity and the variety of the contributions that are gathered in this volume.
In order to make a volume of reasonable size, the editors chose not to
solicit contributions from the Graph Theory community. Had it not
been the case, two volumes would probably have been necessary. We
were really impressed by the willingness of everyone who was contacted
to participate in the project. This reflects the real impact of Bernard
Roy on the scientific community of his time - in our opinion much better
than a long list of his various distinctions.
Part I contains two papers related to the early career of Bernard Roy
when, working at SEMA, he developed many new techniques and con-
cepts in Graph Theory in order to cope with complex real-world prob-
lems. Jacques Lesourne, former director of SEMA, recalls the role of
Bernard Roy in popularizing Operational Research techniques in France
as well as his role in the development of SEMA. Dominique de Werra
and Pierre Hansen reflect on the influence of Bernard Roy's contribu-
tion in Graph Theory. More than 30 years after the publication of his
well-known books, this influence is still present.
The rest of the book consists of contributions related to the second part
of the career of Bernard Roy - to "Multi-Criteria Decision-Aiding".
convergence and complexity. This part and the whole book ends with
the paper by Pekka Korhonen on free searching over the efficient fron-
tier in Data Envelopment Analysis; the search is useful when preference
information is desired to incorporate into efficiency analysis.
The editors wish to extend their warmest thanks to all the contributing
authors. This book is a fruit of friendly co-operation between editors
and authors, motivated by a joint will of celebrating Bernard Roy. The
editors had the privilege of working closely with Bernard Roy during
many years. The authors invited to contribute a paper are also close to
him for various reasons.
Denis Bouyssou
Eric Jacquet-Lagreze
Patrice Perny
Roman Slowinski
Daniel Vanderpooten
Philippe Vincke
1. Books
Bernard Roy, Denis Bouyssou, Aide multicritere a la decision : Methodes
et cas, Paris, Economica, mai 1993,695 pages.
Bernard Roy, Multicriteria Methodology for Decision Analysis, Kluwer
Academic Publishers, 1996
(original version in French: Methodologie multicritere d'aide a la
decision, Paris, Economica, 1985, 423 pages; Polish translation :
Wielokryterialne wspomaganie decyzji, Wydawnictwa Naukowo-
Techniczne, Warszawa, 1990,281 pages).
Bernard Roy, Algebre moderne et theorie des graphes orientees vers les
sciences economiques et sociales :
- Tome 1: Notions et resultats fondamentaux, Paris, Dunod, 1969,
518 pages.
- Tome 2: Applications et problemes specifiques, Paris, Dunod, 1970,
784 pages.
Bernard Roy (in collaboration), Les problemes d'ordonnancement -
Applications et methodes, Paris, Dunod, Monographie de Recherche
Operationnelle, 1964 (German translation : Ablaufplanung -
Anwendungen und Methoden, Oldenburg Verlag, 1968).
2. Edited Volumes
Alberto Colorni, Massimo Paruccini, Bernard Roy (eds.), A-MCD-A - 25th
year, EURO Working group, Multiple Criteria Decision Aiding, EUR
Report, The European Commission, Ispra 2001.
Bernard Roy, Combinatorial Programming: Methods and Applications
Dordrecht, Holland, D. Reidel Publishing Company, 1975.
Bernard Roy, La decision : ses disciplines, ses acteurs, Presses
Universitaires de Lyon, Monographie de l'AFCET, 1983.
2 AIDING DECISIONS WITH MULTIPLE CRITERIA
Bernard Roy, Raphael Benayoun, Jean Tergny, From SEP procedure to the
mixed OPHELIE program, in Jean Abadie (ed.), Integer and Non Linear
Programing, North-Holland Publishing Company and John Wiley and
Sons, 1970,419-436
(see also Revue METRA, Vol. IX, nO 1, 1970, 141-156, in French).
Bernard Roy, Optimisation et analyse multicritere, in Claude Jessua,
Christian Labrousse, Daniel Vitry, Damien Gaumont (sous la direction
de), Dictionnaire des Sciences Economiques, Presses Universitaires de
France, 2001,640-643.
Bernard Roy, L'aide a la decision aujourd'hui : que devrait-on en attendre ?,
in Albert David, Armand Hatchuel, Romain Laufer (eds.), Les nouvelles
fondations des sciences de gestion - Elements d'epistemologie de la
recherche en management, Editions Vuibert, Collection FNEGE, 2001,
141-174.
Bernard Roy, Refiexion sur Ie theme quete de l'optimum et aide a la
decision, in Decision, Prospective, Auto-Organisation - Melanges en
I 'honneur de Jacques Lesourne, Textes reunis par J. Thepot, M. Godet,
F. Roubelat, A.E. Saad, Paris, Dunod, 2000, 61-83.
Bernard Roy, Denis Bouyssou, Aide a la decision, in J.P. Helfer, J. Orsoni
(coordinateurs), Encyclopedie du Management, TGme 1, Vuibert, janvier
1992,447-457
(see also AFCETIINTERFACESN° 65, mars 1988,4-13).
Bernard Roy, Denis Bouyssou, Comparaison, sur un cas precis, de deux
modeles concurrents d'aide a la decision, in Marches, Capital et
Incertitude, Essais en I 'Honneur de Maurice Allais, sous la direction de
Marcel Boiteux, Thierry de Montbrial, Bertrand Munier, Economica,
1986, 155-177.
Bernard Roy, Formatage et singularites du projet «Reseau 2000», in Jacques
Le Goff, Louis Guieysse (eds.), Crise de l'Urbain, Futur de la Ville,
Economica, 1985,45-50.
Jean-Christophe Hugonnard, Bernard Roy, Le plan d'extension du metro en
banlieue parisienne, in Methode de decision multicritere, Textes
rassembles par Eric Jacquet-Lagreze et Jean Siskos, Monographies de
I' AFCET, Division Gestion-Informatisation-Decision, Editions Hommes
et Techniques, 1983,39-65
(see also Les Cahiers Scientifiques de la Revue Transports nO 6, 1er
trimestre 1982, 77-108).
12 AIDING DECISIONS WITH MULTIPLE CRITERIA
5. Papers in Proceedings
Bernard Roy, Daniel Vanderpooten, The European School of MCDA: A
historical review, in Roman Slowinski (ed.), Proceedings of the 14th
European Conference on Operational Research «OR; Towards
Intelligent Decision Support», Jerusalem, Israel, July 3-6, 1995, 39-65.
Jean-Charles Pomerol, Bernard Roy, Camille Rosenthal-Sabroux,
Developing an «intelligent» DSS for the multicriteria evaluation or
railway timetables: Problems and issues, in The International Society for
Decision Support Systems, Third International Conference, Conference
Proceedings, Volume 1, lOSS '95, June 22-23, 1995, 161-172
(see also Revue des Systemes de Decision, Volume 5, n° 3-4, 1996,249-
267, and Foundations of Computing and Decision Sciences, Vol. 20, No.
3, 1995,219-238).
Bernard Roy, Daniel Vanderpooten, The European School of MCDA: A
historical review, in European Conference on Operational Research OR:
Towards Intelligent Decision Support, Jerusalem, Israel, July 3-6, 1995,
39-65.
Bernard Roy, Eric Jacquet-Lagreze, Concepts and methods used in
multicriteria decision models: Their application to transportation
problems, in H. Strobel, R. Genser, M.M.Etschmaier (eds.), Optimization
Applied to Transport Systems, IIASA, Laxenburg, Austria, 1977,9-26.
Selected Publications of Bernard ROY 13
Bernard Roy, Why multicriteria decision aid may not vit with the assessment
of a unique criterion, in Milan Zeleny (ed.), Multiple Criteria Decision
Making, Springer-Verlag, 1976,283-286.
Bernard Roy, From optimization to multicriteria decision-aid: Three main
operational attitudes, in Proceedings of a Conference, Jouy-en-Josas,
France, May 21-23, 1975, Herve Thiriez, Stanley Zionts (eds.), Multiple
Criteria DecisionMaking, Springer-Verlag, 1976, 1-34.
Bernard Roy, Hubert Le Boulanger, Traffic assignment - The ATCODE
model, in Vehicular Traffic Science, Proceedings of the Third
International Symposium on the Theory of Traffic Flow, American
Elsevier Publishing Co., New York, 1967.
Tullio Joseph Tanzi, Bernard Roy, Michel Flages, D. Voncken, Indicateurs
de dangerosite appliques aux transports collectifs, in Actes du 12e
Colloque National de Surete de Fonctionnement, Montpellier, 28-30
mars 2000, 703-708.
Bernard Roy, Recherche operationnelle et aide a la decision, in Claude
Sayettat (ed.), L'Intelligence Artificielle - Une Discipline et un
Carrefour Interdisciplinaire, Compiegne, 10-12 decembre 1992, 139-
145.
Albert David, Bernard Roy, Existe-t-il une approche systemique du
changement organisationnel ? Discussion a partir de I' exemple de la
modernisation de la RATP engagee par Christian Blanc, Actes du rr
Congres biennal de I 'association franr;aise des sciences et technologies
de I'information et des systemes «Systemique et Cognition», Versailles,
8-10 juin 1993, 293-308.
Guy Casteignau, Bernard Roy, L'analyse multicritere interactive comme
outil d'aide ala decision pour la gestion des risques environnementaux et
industriels, in Actes du Congres International Innovation, Progres
Industriel et Environnement - Preparer Ie XXIieme Siecle, Strasbourg, 4-
6 juin 1991,93-102.
Bernard Roy, Frederic Letellier, Vne approche multicritere pour piloter la
gestion des approvisionnements dans une structure de stockage a deux
niveaux, Actes du Colloque AFCET sur Ie Developpement des Sciences
et Pratiques de I 'Organisation et 4e Journees Francophones sur la
Logistique et les Transports, Theme 1989 : Logistique, Production,
Distribution, Transports, Paris, 13-15 decembre 1989,63-70.
14 AIDING DECISIONS WITH MULTIPLE CRITERIA
Jacques Lesourne*
Paris, France
Bernard Roy, is almost blind, reads with the help of a thick lens and writes
with big letters; the latter, Patrice Bertier is handicapped by a severe
poliomyelitis which constrains him to a rolling chair, but does not prevent
him from working. The blind and the paralysed as in some tales of the old
days.
The first act begins a few weeks later. Brand-new offices with white
walls and big ash tables. The first contracts are signed and the team
develops. The first topic on which, as far as I can remember, Bernard Roy
operates concerns the production and inventory policy for tubes used in oil
drilling. Bernard elaborates, out of numerical data, a statistical law of the
life-lengths of the tubes. Quickly, my opinion is made: this young guy open
and joyful has a good brain and the faculty to pass easily from observation to
modelling and vice-versa.
But the real test occurred a little later when Electricite de France asked
us to solve the scheduling problem of the nuclear power plant at Chinon.
Bernard's contribution to the modelling was decisive. He proposed to
introduce a graph, nodes of which representing the tasks and arrows the
anteriority constraints. Without knowing it, in parallel with an American
team working on the Polaris submarines (but without any contact with
them), we had discovered the modem treatment of scheduling problems. So
appeared the METRN potential method (MPM).
In the initial modelling, the availability constraints were not introduced,
but new developments made possible to apply it to public works when only
one crane is available on the site or to the construction of the liner France for
which it was necessary to take into account the available staff of different
professions.
Bernard Roy pursued the development of graph theory which enabled
him to obtain brightly his thesis and to publish at Dunod his first book
"Theorie des graphes et applications" (1969).
Along this road, he made incursions into integer linear programming
which in these days were of interest to us.
Act II opened in 1960 when SMA was transformed into SEMA (Societe
d'Economie et de Mathematiques Appliquees) and transferred -due to its
rapid growth- into larger offices, in La Boetie street. The size of the team
implied now the definition of a structure. It seemed to me that, in addition to
the various operational units, it was necessary to create a group which would
ensure the relations with the scientific world, would conceive new tools and
1 Name given to the group composed of Serna and its European subsidiaries.
Bernard ROY, Forty Years ofEsteem and Friendship 19
would assist the teams in the modelling of some of the problems they had to
deal with.
A name was obvious for the appointment of the head of this Direction
Scientifique: Bernard Roy. All the members of the staff supported this
choice. But the task was far from easy. The presence of this department
increased the central costs and was only acceptable if the operating staff
considered that the department was offering them a real service which it
could not carry out itself. The Directeur Scientifique had therefore to prove
daily that he was managing an efficient team, helping equitably all the
operational units with diplomacy and humility. In these responsibilities,
never was Bernard Roy criticized, which is a remarkable performance.
With the European development of the Serna group, Bernard had to
put the Direction Scientifique at the disposal of foreign companies, generally
smaller, in great need of technical transfers, but wanting to keep jealously
their identity. In this field also, Bernard Roy was recognised, thanks to the
Metra review, a quarterly journal freely distributed but with the quality of a
scientific periodical, conceived to promote, out of real examples, the OR
techniques in the broadest meaning of the concept and not to make publicity
for the group. Courage was indeed necessary for the man in charge of it
since I had imposed the constraint -which seems now crazy to me- for the
authors to present their papers in their own language, French, English,
German, Italian, Spanish so that, except for the summaries, only a few
people could read the whole set. Nevertheless, one may still read nowadays
the issues of Metra without being ashamed.
This period was the golden age of the Direction Scientifique. Among
the topics on which Bernard Roy has personally worked during these years,
I shall point out two fields:
the models for traffic analysis and forecasting, the first of which, called
"the model of preferential equilibrium" enabled to estimate the volume of
trips from home to work, from a zone i to a zone j when computed the
generalized cost of transportation between the zones and known the
geographical distribution of workers and jobs. This model was successfully
used for predicting traffic on the lines of the new underground system, the
Reseau Express Regional (RER).
the models of multicriteria choices in which Bernard Roy was a pioneer,
which contributed to establish through the years his international reputation.
The initial issue was to distribute optimally an advertising budget between
different medias, or for a given media, the press for example, between the
various papers. Bernard Roy realized that it would have no meaning to look
for a function to maximize. Too many and heterogeneous were the criteria:
20 AIDING DECISIONS WITH MULTIPLE CRITERIA
the cost of insertion in relation with date and size, the characteristics of the
audience, the impact of repeated insertions, ... He understood that it was not
satisfactory to give to each program a mark for each of the criteria, then to
add up these marks weighted by coefficients expressing the relative
importance of the various criteria. Pragmatisms which could lead to different
solutions from one problem to the next had to be based on a rigorous set of
axioms and on solid theoretical foundations. Electre -since it was the name
given to the first member of the family- was born in these years. It had a
numerous offspring conceived in Serna and later in Lamsade.
Difficult to situate precisely the beginning of act III, but it started under
the influence of two evolutions: the explosion of informatics (the name was
invented at SEMA by Philippe Dreyfus) and the absorption by SEMA of
OTAD a sister company operating in the fields of organization, training and
selection. Hence it became more difficult for the Direction Scientifique to
meet the needs of a very diversified international group covering many fields
from scientific informatics to urban planning including automatic data
processing, business economic studies, marketing research and surveys.
Nevertheless, I wanted the presence of a unit asserting to all that the group
wanted to go on invented methods coming from the fertile ground of
practical cases and from the development of sciences, mathematics,
economics and social sciences. In the dull climate generated by the group's
financial difficulties, Bernard Roy had to diversify his team, facilitate the
insertion of scientific advisers, broaden the scope of his research themes.
However, his primary task was to lead the elaboration of a yearly research
programme, each of the projects selected being financed by the centre and
executed by an operation unit in cooperation or not with the Direction
Scientifique. The proposals could be sent by any unit of the group, the
results of the projects being freely diffused throughout the group. Bernard
became then the head of the application of multicriteria methods to establish,
out of the bundle of offers, the group's research program. A serious task
which required openness, diplomacy and an aptitude to be accepted by the
representatives of all the sectors. Bernard perfectly assured this
responsibility. Some of the projects were achieved magnificently, as for
instance one on the internal staff relations in a hospital, but the practical
results were uneven, the efforts made to reduce the group's expenditures
modifying constantly the organization and diminishing the central costs. Hit
by this storm, the Direction Scientifique disappeared and Bernard Roy was
elected as a professor at University Paris IX Dauphine, where, for thirty
years, he has been developing a second career.
Bernard ROY, Forty Years ofEsteem and Friendship 21
As for the epilogue, it is not obviously to an end. It took the shape for
me of a many years' presidency of the Lamsade Scientific Committee, which
meant for me a double privilege, to receive the research documents of the
team and to have the opportunity every year to spend a stimulating day with
Bernard and his colleagues. Of course, my own research had developed in
other directions but I had not lost my interest for the immense field related to
decision and more generally action. On his side, Bernard has procured for
me the great pleasure of discovering in the book of M(?ianges offered to me
at the beginning of 2000, a deeply-thought paper on the meaning of
optimality in a decision context.
This long historical cohabitation is not however sufficient to explain
the friendship and esteem I feel for Bernard Roy. If I try to analyse it, several
dimensions come up to my mind: his attitude with respect to his handicap, a
handicap mastered, neither hidden nor borne as a medal, which facilitates
greatly relations with others; his authentic good temper which likes humour
and expresses an internal joy, certainly supported by the equilibrium of his
family; a great continuity of mood which helps him in facing and dominating
rationally difficulties; a constant curiosity and an openness which makes him
a precious partner for the specialists for all fields concerned with decision;
an ability to pass easily from reality to abstraction and vice-versa, which is
absolutely necessary in applied sciences; a constant thinking capacity which
enables him to plough and pursue the furrows he has chosen; at last, his
faithfulness in friendship to which a forty year period gives a precious
thickness.
CONNECTIVITY. TRANSITIVITY AND
CHROMATICITY: THE PIONEERING WORK
OF BERNARD ROY IN GRAPH THEORY
Pierre Hansen
GERAD and Ecole des HEC, Montreal, Canada
pierreh@crt.urnontreal.ca
Dominique de Werra
Ecole Poly technique Pedemle de Lausanne, Switzerland
dewerra@drna.epfl.ch
Abstract We review the work of B. Roy in graph theory and its posterity.
Introduction
Before exploring in depth and breadth Decision Aid Methods, Bernard
Roy devoted a few years to the study of graph theory. This led to a series
of contributions including a large book (Roy 1969/70). Generated in a
context of intensive research in graph theory, pioneering ideas in several
of his papers induced long streams of results by many authors up to the
present time.
In this chapter, we review the work of B. Roy on graph theory about
forty years after its publication. We also outline its posterity by pre-
senting a sample of the extensions of his seminal results and we describe
the general context in which his research was carried out. We do not
aim at exhaustivity, but rather, in a tutorial spirit, try to present to a
large audience the main themes of this research.
We assume the reader is familiar with the basic concepts of graph
theory and refer to the book of C. Berge, "Graphs and Hypergraphs"
(Berge 1973) for definitions not given here.
24 AIDING DECISIONS WITH MULTIPLE CRITERIA
The graph G in Fig 1 satisfies the conditions (i), (ii), (iii), as can be
verified.
In the graph G in Fig. 1 the subset A = {X2' X3} contains its image
f(A) = A, hence G is not strongly connected.
D. Gale (1959) observes that this last result appears in a long paper of
R. Rado (1943) on linear combinatorial topology and general measure.
P. Camion (1959) uses the proof technique of (Roy 1958) to derive a
well-known result on Hamiltonian circuits:
o
(a)
Theorem 6 (Roy 1959b): The matrix Tn· Tn-I ... Tl . A is equal to the
transitive closure A. of A.
Tn . Tn-I' .. Tl . L = t.
A further extension to the maximum capacity path can be made by
starting from a matrix C = (Chk) where Chk is the capacity of arc (Xh' Xk)
if it exists and 0 otherwise. The capacity of a path is defined as the
smallest capacity of its arcs. 7i is defined by Ti . C = (Chk) with
the Min operations for maximum capacity path and the Max operation
and the usual product for reliability. This suggests there should be a
general algebraic structure subsuming all these cases. It is indeed so. In
a general setting it is a dioid (Gondran and Minoux 1984) (also called
semi-ring by other authors), defined as a set S with two operations
(i) the operation EB ("add") gives S a structure of commutative monoid
(closure, commutativity, associativity) with neutral element E.
(ii) the operation ® ("multiply") gives S a structure of monoid (clo-
sure, associativity) with neutral element e (unit). Moreover, E is
absorbing (a ® E = E for any a in S) and ® is right and left dis-
tributive with respect to EB. Moreover,
(iii) the preorder relation ~ (reflexivity, transitivity) induced by EB
(canonical preordering) and defined by a ~ b if and only if there is
a c in S such that a = b EB c is a partial order, i.e., it satisfies a ~ b
and b ~ a implies a = b (antisymmetry).
Many authors have studied this structure (e.g. (Shimbel1954), (Cun-
ninghame-Green 1960, 1962, 1979), (Yoeli 1961), (Robert and Ferland
1968), (Tomescu 1968), (Gondran 1975) and (Wongseelashote 1976)).
The reader is referred to (Gondran and Minoux 1984) for further refer-
ences.
Properties of classical linear algebra, and in particular algorithms
to solve systems of linear equations (Gauss, Gauss-Seidel, Jacobi, Jor-
dan, etc.) can be adapted to dioids. Algorithms of (Dantzig 1967)
and (Tabourier 1973) can be viewed in this light, as well of course and
that of (Roy 1959b) which corresponds to a generalized Jordan method.
So the work of B. Roy on transitive closure is part of a vast stream
which dates back to the middle fifties. This research program has
been expanded further recently by M. Gondran (1996a, 1996b) and
M. Gondran and M. Minoux (1997) who have shown how the use of
dioids can extend nonlinear analysis.
Among possible applications of the transitive closure algorithm, we
mention finding the transitive reduction of an oriented graph G with-
out circuits, i.e., removing a maximal subset of arcs without changing
the transitive closure (Gries, Martin, van de Snepscheut and Udding
1989). Another application is the efficient evaluation of single-rule Dat-
alog programs with a slight generalization of the Floyd-Roy-Warshall
algorithm (Papadimitriou and Sideri 1999).
P.L. Hammer and S. Nguyen (1977) consider the following logical
problem, which generalizes the question of computing the transitive clo-
sure of a graph: we are given a set of binary relations between boolean
30 AIDING DECISIONS WITH MULTIPLE CRITERIA
This result combined with the observation that the vertices of a graph
G can always be colored with .6.{G) + 1 colors (where .6.{G) denotes the
maximum degree of G, i.e., the maximum number of arcs adjacent to a
vertex) gives the following:
3 2
(a) (b)
(e)
As an illustration, one may verify that the 4-coloring of Fig. 4 (c) sat-
isfies the above properties. It is interesting to observe that such colorings
are similar in spirit to some solutions of sequencing problems which are
named "squeezed to the left" (Fr: "cale it gauche") which were also stud-
ied by B. Roy (1962). Here we may consider the graph G as associated
to a chromatic scheduling problem as follows:
The nodes correspond to jobs with equal processing times, say 1. An
arc (x, y) means that job x must precede job y. There is a one-to-one
correspondence between k-colorings of G and feasible schedules in k time
units where each job starts at some integer time.
The k-colorings satisfying (a)-(d) are precisely schedules squeezed to
the left: according to (b), if a job x is scheduled at period c(x) = h, it
is because all jobs in Sl, S2,'" ,Sh-1 need h - 1 periods and x cannot
be scheduled in a period i ~ h - 1. It is thus a schedule where each job
starts at its earliest date.
34 AIDING DECISIONS WITH MULTIPLE CRITERIA
At the end of his paper B. Roy mentions briefly the special case of
perfect graphs G; these are the graphs in which each induced subgraph
H of G satisfies X(H) = w(H) where w(H) is the maximum cardinality
of a clique of H.
In such graphs, the colorings satisfying (a)-(d) have the following
characteristics:
For each vertex xESh there exists a clique K h(x) 3 x with K h(x) n Si i=
o for i = h, h - 1, ... , 1.
Such colorings have been called canonical in (Preissmann and de
Werra 1985) where strongly canonical colorings have also been defined
for strongly perfect graphs (Berge 1984) (a graph G is strongly perfect if
in every subgraph H of G there exists a stable set S such that S n K i= 0
for every inclusion-wise maximal clique K of H).
A coloring is strongly canonical if for any clique K of C there is a
clique C :J K such that en Si i= 0 for i = 1,2, ... ,min{ flS1 n K}.
As observed in (Preissmann and de Werra 1985), a graph G is strongly
perfect if and only if every induced subgraph of G has a strongly perfect
coloring. The graph G in Fig. 4 (a) is perfect (one may check that the
4- coloring in Fig. 4 (c) is canonical); it is also strongly perfect, as can
be verified.
Coming back to the general case handled by B. Roy, one observes
that the orientation constructed is such that there is a path P meeting
(consecutively) S1, S2, ... ,Sk; in fact, the coloring satisfies the following:
every vertex z with c(z) = h is on a path which meets (consecutively)
S1, S2,"" Sh·
Starting from this observation, Berge has obtained the following re-
sult:
Theorem 10 (Berge 1982): Let k be the maximum number of vertices
in a path of G. Then for every path P with k vertices, there exists
a k-coloring (S1,"" Sk) such that ISh n PI = 1 for h = 1,2, ... , k.
Connectivity, Transitivity and Chromaticity 35
SI ={a,c}
S2= {b,e}
S3= {d}
A set of non adjacent vertices is a stable set and the maximum cardi-
nality of a stable set in G is denoted by a(G). Theorem 12 amounts to
saying that min IMI :::; a(G). The graph G in Fig. 5 has a path partition
M = ({a,b,c},{d,e}) and a(G) = 2.
C. Berge has defined in an analogous way a-diperfect graphs which
are those in which every subgraph H has the following property:
given any maximum stable set 8, there exists a path-partition M =
(PI, P2, ... ,Pk ) with k = a(H) and 18 n Pil = 1 for i = 1, ... ,k.
Again perfect graphs and symmetric graphs are shown to be a-diperfect
(Berge 1982). The graph G in Fig. 5 is not a-diperfect:
for 8 = {a, c}, the only path partition (PI = {a, b, c}, P2 = {d, e}) in
a(G) = 2 paths is such that P2 n 8 = 0.
In general, for an arbitrary graph, the known proofs of Theorem 12
do not imply the existence of a maximum stable set 8 and of a path-
partition (PI,'" ,Plsl) with 18 n ~I = 1 for all i.
We would now like to recall the conjecture of C. Berge which would
unify Theorems 7 and 12.
A path partition M(PI,'" ,Pq ) is called k-optimal if it minimizes the
quantity
q
Bk(M) =L min {k, IPi !}.
i=1
Connectivity, Transitivity and Chromaticity 37
PI = {a,b,e}
P2 ={d,e}
SI ={b,d}
S2= {e,e}
k-coloring (SI, S2,'" ,Sk) such that for every k-optimal path partition
M = (PI"'" Pq) every path Pi of M meets min {k, JPi!} color classes.
Conclusion
We have presented in a condensed form the contributions of B. Roy
to graph theory. Our discussion has shown that several of his results as
well as some of the questions raised in his papers have undoubtedly had
an impact on the work of numerous researchers.
Many generalizations and variations have followed; what is now called
the "theorem of Roy-Gallai" has in particular been a source of inspiration
for a number of researchers and we conjecture that it will continue for
many years to come.
Acknowledgments
This research was carried out during a visit of the second author
to GERAD in Montreal. Support of both authors by Grant NSERC
#GP0105574 is gratefully acknowledged.
REFERENCES 39
References
Aharoni, R., Ben-Arroyo Hartman, I., and Hoffman, A.J. (1985). Path partitions and
packs of acyclic digraphs. Pacific J. of Math. 118:249-259.
Bang-Jensen, J., and Gutin, G. (1988). Generalization of tournaments: a survey.
J. Graph Theory 28: 171-202.
Berge, C. (1973). Graphs and Hypergraphs. North Holland, Amsterdam.
Berge, C. (1982). Diperfect graphs. Combinatorica 2:213-222.
Berge, C. (1984). A property of k-optimal path-partitions. In: Progress in Graph
Theory (Waterloo, Onto 1982). Academic Press, Toronto, 105-108.
Berge, C., and Duchet, P. (1984). Strongly perfect graphs. Ann. of Discrete Mathe-
matics 21:57-6l.
Berge, C., and Ghouila-Houri, A. (1965). Programming, games and transportation
networks. Methuen, London.
Bermond, J.C., and Thomassen, C. (1981). Cycles in digraphs - A survey. J. of Graph
Theory 5:1-43.
Bilde, 0., and Krarup, J. (1969). A modified cascade algorithm for shortest paths.
METRA VIII 231-24l.
Bondy, J. A. (1976). Disconnected orientations and a conjecture of Las Vergnas. J. Lon-
don Math. Soc. 2:277-282.
40 AIDING DECISIONS WITH MULTIPLE CRITERIA
Albert David
Evry-Val d'Essonne Universit!, France
albert.david@cgs.ensmp.fr
Abstract: There has been a wealth of very varied literature on decisions, decision aid and
decision aiding tools since the initial work carried out by Barnard [1938] and
Simon [1947], which marked the sudden emergence of decision-related issues
in organizational theory. Whether the decisions concern strategy, finance,
marketing or « operations », decision aiding tools have signalled the waves of
rationalisation that have taken place in the history of management sciences and
organizational theory. Our aim here is not to go into the typology of the
different tools, but to explore two particular questions. Do decision aiding
tools have specific structural properties? Which concepts can be used to
analyse and understand the dynamics of their introduction into organizations
and the resulting learning processes? We will begin by asking what a tool is,
what aid it can offer, for which decisions. We will see that knowledge cannot
be produced without tools, that decision-making is a complex process and that
decision aiding is a prescriptive relationship. We will confront decision aiding
tools with two « functional» and two « critical» decision models. We will
then examine the structure and dynamics of managerial innovations in general.
Finally, we will analyse to what extent decision aiding tools are specific
managerial innovations, by their technical substratum, their management
philosophy and their simplified view of organizational relations. We will
conclude by putting into perspective conforming and exploring approaches,
methods for managing change and the nature of learning during the tool
contextualisation process.
I A first version of this paper was published in French with the title "L'aide ala decision entre
outil et organisation", Entreprise et Histoire, n° 13, 1996.
2 The research activity of Albert David is also with the Centre de Gestion Scientifique (Ecole
des Mines de Paris) and the LAMSADE (Paris-Dauphine University).
46 AIDING DECISIONS WITH MULTIPLE CRITERIA
tools were used in a closed, conforming perspective, and all the weaker
when the tools were used in an open, exploratory perspective6 .
The crisis in operational research that took place during the second half
of the 1970s, when a certain number of tool designers and managers using
them began to have misgivings about the performance of models, can be
analysed as a crisis in the prescription process [Hatchuel, 1996]. Similarly to
a number of other management instrument approaches, one of the major
changes in the field of decision-making, begun at the end of the 1960s [Roy,
1968], was the gradual abandonment of strong recommendations based on
universal rationality. The science of decisions gave way to a weaker form of
prescriptive action, embodied in the idea of a science of decision aiding
[Roy, 1992].
According to Roy [1985], «decision aiding is the activity of the person
who, through the use of clearly explicit but not necessarily completely
formalised models, helps obtain elements of responses to the questions posed
by a stakeholder of a decision process. These elements work towards
clarifying the decision and usually towards recommending, or simply
favouring, a behaviour that will increase the consistency between the
evolution of the process and this stakeholder's objectives and value
system ».
This definition appears to give decision aiding a relatively modest role.
This conception of aiding is, in reality, quite sophisticated. A decision aiding
tool is considered as a model which is «clearly explicit but not necessarily
completely formalised », which helps the person intervening to act in a more
coherent manner with respect to his own objectives and value system.
Naturally, the neutrality is only on the surface: we will see below that this
definition of decision aiding tends to direct the decision-making process in a
particular manner.
In this context, « deciding, or in a wider scope, intervening in a decision-
making process only very rarely implies finding a solution to a problem.
Often, it is a question of finding a compromise or making people accept
arbitration in a conflict situation» [Roy, 1993, p.21]. It can be seen that the
definition of « deciding» is very wide in this case. Nevertheless, it should be
noted that if too wide a definition is accepted, anything that serves to
improve efficiency in decision-making must be considered a decision aiding
tool. In this way, quality circles, a new structure, assessment interviews,
7 We have taken these four models as examples, to confront functional and critical
approaches. For a more in-depth analysis of the types of rationality at work in decision
models, see, for example, Munier [1994].
Decision-Aid Between Tools And Organisations 51
real decision
virtual decision
representation
understanding of the organisation
8 Although they are in the confined, restricted space defined by the garbage can, meaning that
it is not total anarchy.
54 AIDING DECISIONS WITH MULTIPLE CRITERIA
relations /knOWledge
organisation
framework
relations knowledge
knowl edge
relati onal oriented
procedure procedure
detailed
procedure
Figure 3. The four possible starting points for the process of introducing managerial
innovations
10 This notion of contextualisation is therefore stronger than the more traditional notion of
adoption and less ambiguous than that of codification, which refers to both the
formalisation of the innovation in manuals and directions for use and to the concept of
encoding used in the cognitive sciences in particular.
58 AIDING DECISIONS WITH MULTIPLE CRITERIA
organisation, between the innovation and the organisation. The greater the
distance between the innovation and the organisation adopting it, the smaller
the degree of internal contextualisation. On the contrary, the nearer the
innovation is to the organisation adopting it, the higher the degree of internal
contextualisation. In qualitative terms, the « distance» not only corresponds
to the gap between the way things operate at present and the way they are
imagined in the future, but also to the time it will take and the difficulties
that will be encountered before the innovation works effectively in the
organisation.
If, at the start of the process, we consider the organisation (relations and
knowledge) on the one hand, and the managerial innovation on the other,
and if we acknowledge that each of the two includes an incomplete vision of
the other (initially, the players have an incomplete vision of the innovation;
the innovations convey a simplified vision of the organisation), then the
contextualisation of a managerial innovation in an organisation can be seen
as a process of cross exploration. At the start, the technical substratum is
controlled or controllable at varying degrees by the players, the managerial
philosophy is understood and accepted at varying degrees by the players, and
the simplified vision of relations or knowledge is to a greater or lesser extent
schematic and removed from current relations and knowledge. All these
factors illustrate a certain « distance» between the innovation and the
organisation. If all goes well, the process will converge towards a full
integration of the innovation and the organisation, at the price of more or
less significant transformations on either side. At that point, the innovation is
fully contextualised, meaning that the technical substratum is working, the
management philosophy is well-adapted and the simplified vision of relation
and/or knowledge has become explicit and complete: it can then be said that
the distance between the innovation and the organisation is equal to zero.
framework
starting
point degree of
contextualisation
minimum maximum
relations knowledge
detail
11 See Goody [1977] on the way in which lists, classifications and tables are a « graphical
reasoning ».
62 AIDING DECISIONS WITH MULTIPLE CRITERIA
12 On the other hand, constructing the indicators to be included in a performance indicator can
often be very difficult.
13 For full explanations, see Roy [1985] and Roy, Bouyssou [1993]
Decision-Aid Between Tools And Organisations 63
issues, (2) separating robust conclusions from fragile ones, (3) dispelling
certain forms of misunderstandings in communication, (4) avoiding the trap
of false reasoning and (5) highlighting uncontrovertible results once
understood [Roy, 1993]. It is quite clear that this is not just a question of
techniques designed to enable reasoning to be based on several criteria
instead of a single one. Admittedly, the management philosophy of the
multicriteria approach borrowed from operational research the idea of
working on a model to help select solutions - as reflected in the four issues
mentioned above - but the origin of these methods should first be sought in
social choice theory. As several authors have noted [for example, Pomerol
and Barba-Romero, 1993, Munier, 1993], there is conceptual identity
between the issue of the aggregation of the opinions of judges on actions and
that of the aggregation of the evaluations of actions on multiple criteria.
Social choice theories (Borda, Condorcet, Arrow, etc.) were drawn up by
researchers working in the field of political sciences on, to simplify, issues
of democracy such as drafting procedures for the aggregation of individual
votes that best reflected collective demands, and hence the general interests.
This concern can be found explicitly in the multicriteria approach, where it is
a question of introducing as much reason as possible in a decision-making
process, whilst respecting the players free play. The democratic, honest ideal
conveyed by the multicriteria approach can also be compared with the
principle of isonomy - lack of prejudice - proposed by Hatchuel [1994] to
guarantee the scientific nature of a researcher's intervention in a company
and with the principle of « low normativeness» which, according to
Lautmann [1994], qualifies the Crozerian approach to organisations in which
everything takes place as though the player who aids in decision-making is
« the ally of the ideal reformers of the system, [... ] whether such reformers
exist or not» 14.
The simplified vision of the organisation may appear not to exist, due to
the apparently very general way in which the problem is formulated. No one
would disagree that deciding means choosing, sorting, ranking or simply
describing potential actions evaluated using a series of criteria. In reality,
similarly to performance indicators or scores, the simplified vision of the
organisation concerns relations between players. In multicriteria analysis, it
focuses on two elements:
14 Lautmann, 1994, p.187 : « A low fonn of nonnativeness is inserted, which is to address the
issue to the sincere, free thinking decision-maker who is not the sociologists double but
who makes a couple with him». There is an obvious parallel with Roy's «decision-
maker/researcher» couple.
64 AIDING DECISIONS WITH MULTIPLE CRITERIA
framework
_.... --;....-....-.":--;;'
degree of
mmimlm contextualisation JI'8JI,.
"
I"
/'
,
",,,
I
",
"
,I
I
I minimum
JI
I I ,
I , ,
re",la",tio""n",-s+'___ +'+-__
, ,
----4'_--+.::knowledge
\ \ \
~~\
,
A more or less important
...
....... ".... \
\
degree of
transl~m~~uionof . .... ... ___: - _ ...... \I
orgamsatlOnal relations + fonnalisation
is necessary
detail
framework
Degree of
contextualisation
minimu.m maximum
minimum
relations knowledge
maximllm
Degree of
formalisation
detail
15 For a detailed explanation of these models, together with the political model and the
conquest model that are not mentioned here, see David [1996, 1998a].
66 AIDING DECISIONS WITH MULTIPLE CRITERIA
The difference as far as those steering the process are concerned, is that
in the management model, the tool's implicit relational dimension is
explicitly taken into account in the organisation of the change. For example,
corporate management may ask for performance indicators to be designed in
each entity of the organisation. This is a framework that the players
concerned must then try to draft themselves, in detail. If the management
supports the process and agrees to assess the results only at a later stage, this
guarantees that the formalisation and the contextualisation take place at the
same time. This result will also be obtained if the outside players
(researchers-players, for example) play this role of mediation between the
tool and the organisation.
16 The terms « conforming» and « exploring» are borrowed from Moisdon [1997]. The terms
« constraining» and « enabling» can be found in particular in Landry [1993].
Decision-Aid Between Tools And Organisations 67
involved or, on the contrary, feel that there is a threat to their autonomy.
Although the tool exists formally, it will fail either to control the business or
to explore the conditions under which actions could be effective.
It is therefore clear that the nature of learning due to the introduction of a
tool, in particular a decision aiding tool, varies depending on whether the
tool is designed with a conforming or an exploring approach, depending on
the model for steering the contextualisation process and depending on the
time at which the tool goes from the design stage to a more autonomous
stage of current use. The three aspects are related: the more the approach is
one of conformation and the greater the chances that the contextualisation is
steered in a technocratic, centralised manner, the more the tool is likely to be
delivered to the users without the co-operation of designers-users that would
enable a sufficiently high level of crossed learning to ensure that the tool
retained part of its exploratory capacities. There is then a risk that the tool
becomes completely autonomous, to such an extent that the players will be
unable to challenge their representations of reality if the context changes. On
the contrary, the more the tool is designed with an exploring angle and the
greater the chances that the contextualisation process is management-
oriented and decentralised, the more the tool will involve a process of
« simultaneous engineering» and result in far richer crossed learning. But if
the players attain this degree of clear-sightedness, up to what point will the
tool still be necessary?
There is a long continuum of tools/use couples ranging from normative
tools to disposable tools. It is no longer a question of comparing real
organisations and organisations that are implicitly driven by tools, but of
drawing up an appropriate theory for steering change, that it an intervention
theory. The question remains open as to the « right distance» from the
organisation at which the decision aiding tools should be designed and
inserted and how to steer contextualisation. If it is too near or too far, or if
the model for steering the change is out of place, it can fail to create tension
that generates learning.
Finally, we have shown [David, 1996] that methods for the design and
implementation of tools can be considered as tools in their own right, with a
technical substratum, a management philosophy and a simplified vision of
the organisation. The very notion of crossed learning, when viewed in the
context of management philosophy for multicriteria analysis as explained
above, refers back, at a higher level, to the question of the design,
deliberation and steering of projects in a democratic world. In the case of the
business world, shared design of managerial innovations can go as far as
challenging the hierarchy of employment, competence and remuneration. In
the case of a public multi-institutional world, shared design can, implicitly or
68 AIDING DECISIONS WITH MULTIPLE CRITERIA
References
Allouche, l , Schmidt, G. (1995), Les outils de la decision strategique, tomes 1 et 2, La
decouverte.
Argyris C. et Schon D. (1978), Organizational Learning: A Theory of Action Perspective,
Addison-Wesley.
Barnard, C. (1938), The fonctions ofthe executive, Cambridge, Harvard University Press.
Boulaire, C. , Landry, M. et Martel, lM. , « L'outil quantitatif d'aide a la decision comme jeu
et enjeu », Revue INFOR, 1996.
Bouquin, H. (1996), article « Controle », Encyc/opedie de gestion, Econornica.
Courbon J.C. (1982), « Processus de decision et aide a la decision» , Economies et Societes,
series Sciences de Gestion nO 3, tome XVI nO 12, decembre, p. 1455-1476.
David, A. (1988), Negociation et cooperation pour Ie developpement des produits nouveaux
chez un grand constructeur automobile - Analyse critique et role des outils d'aide a la
decision, PhD dissertation, Universite Paris-Dauphine, septembre.
David A. et Giordano lL. (1990), « Representer c'est s'organiser », Gerer et Comprendre,
Annales des Mines, juin.
David, A. (1995), RATP : la metamorphose - Realites et theorie du pilotage du changement,
InterEditions.
David, A. (1996), « Structure et dynamique des innovations manageriales », Centre de
Gestion Scientifique, cahier de recherche n° 12, juin.
David, A. (1998a), « Model implementation: a state of the art », EURO Conference,
Brussels. To be published in European Journal of Operational Research, 2001.
David, A. (1998b), « Outils de gestion et pilotage du changement », Revue Fran{:aise de
Gestion, septembre-octobre.
Erschler, l et Thurio, C. (1992), « Approche par contrainte pour I'aide aux decisions
d'ordonnancement », in Les nouvelles rationalisations de la production, de Terrsac, G. et
Dubois, P. (Editeurs), Editions Cepadues, Toulouse, pp. 249-266.
Girin, l (1981) « Les machines de gestion », Ecole Polytechnique.
Goody, J. (1977) La raison graphique, Editions de Minuit.
Hatchuel A. et Weil B. (1992), L'expert et Ie systeme, Econornica. English translation
published in 1995, Experts in Organizations, W. de Gruyter.
Hatchuei, A. (1994) « Apprentissages collectifs et activites de conception », Revue Fran{:aise
de Gestion, juin-juillet-aout.
Hatchue1, A. (1994), « Les savoirs de I'intervention en entreprise », Entreprise et histoire nO
7, pp 59-75.
Hatchuel, A. (1996), « Cooperation et conception collective: variete et crises des rapports de
prescription », working paper, Ecole des Mines de Paris.
Hatchuel, A, Molet, H. (1986) « Rational modelling in understanding human decision
making: about two case studies », European Journal of Operational Research, nO 24, p.
178-186.
Jacquet-Lagreze, E. (1995), «Optimisation sous contraintes et programmation lineaire»,
a
dossier Les techniques de I'aide la decision, cahiers de I'ANVIE, novembre 1995.
Landry, M. , Banville, C. , Oral, M. (1996) « Modellegitimisation in operational research »,
European Journal of Operational Research 92, 443-457.
Decision-Aid Between Tools And Organisations 69
Valerie Belton
University of Strathclyde, Glasgow, UnitedKingdom
val@mansci.strath.ac.uk.
Jacques Pictet
Bureau AD, Lausanne, Switzerland
Jpictet@fastnet.ch
I: "Yes, definitely ... but perhaps not on the same points! What interested
you most?"
The situation
The real world is a mess
c: "There were different aspects, but the first was when she spoke about
dealing with messy situations. It brings to mind the situation we are facing at
work right now."
I: "What is that, if I may ask?"
c: "Sure ... I am a senior manager in a public administration. The current
organisation was formed some years ago, in a merger of three independent
bodies, all operating in the area of food standards. Each of these had its own
culture, sphere of influence, working processes, etc., which they were keen
to protect. To gain approval for the merger, these bodies were given
guarantees about the future, mainly in terms of dedicated resources.
However, changes in priorities and financial cutbacks have led to the
questioning of these guarantees and the search for a more effective use of
limited resources. To achieve this, the vice-director has come up with a new
management structure, which he refers to as "management by project". He
intends to employ consultants to advise him on how to implement his
scheme. The idea seems to be good, but I wonder whether he has a clear
understanding of the possible consequences of implementing his ideas. He
used to work in a private company and thinks that the public sector should
follow the same path. Possibly he is right, but I am not sure that he has
thought through the broader repercussions for the impact on the organisation,
in particular on working relationships."
I: "In what sense?"
Aspects of complexity
C: "Well, firstly, it makes people feel insecure about the future. Like
many public organisations we have been through various changes such as
this one over the years and almost inevitably undesirable things happen:
people have to apply for their own jobs, they get displaced or even fired.
You never know where you will be at the end of it."
I: "Yes, I've heard of such situations. How does your boss expect his new
policy to work?"
C: "The general idea is to make resources more 'flexible'. This means
that part of the money and peoples' time will be taken away from their
'department' or 'sector' and allocated to a central resource pool. The
Talking About the Practice ofMCDA 73
management will decide how best to use these resources by comparing the
projects proposed by the various sectors. Potentially, people will have to
move from one sector to the other, or work on projects put forward by a
sector other than their own. I've already heard some people muttering that
they are experts in area X and it would not be making best use of their
expertise to have to become part of a project team in area Y. Others have
commented on the dangers of fragmenting teams that have been built up
carefully and operated effectively over a number of years. Not to mention
the nightmare of managing individuals' time effectively in such a matrix
structure."
I: "What you are saying is that it is difficult, if not impossible, to
differentiate the problem from the people connected with itl. I guess it is the
reason why situations are messy."
C: "Not only is it impossible differentiate the problem and the people
involved, everyone involved sees the problem differently - through their
own 'frame'2. However, that is only one important reason. Based on my
experience, I can suggest a couple of additional aspects. A second point
relates to the interdependence between the various people involved, their
objectives, and other related issues. You cannot expect these elements to
form nice little piles waiting for someone to take care of them. Usually, they
are scattered in a complete chaos, moreover they are moving all the time and
it is difficult to know where to draw the boundary3. This leads on to a third
factor contributing to the 'mess' - time itself. Situations are continually
evolving as both the external and the internals contexts change. Furthermore,
people are usually quite poor at keeping track of all these changes and tend
to reconstruct the past according to their beliefs, when asked for
explanations. "
There have been many terms used to capture the complexity of real life
situations. Rittel and Webber [1973] referred to 'wicked problems', Ackoff
[1981] to 'messes' and Schon [1987] differentiated between the 'swamp'
and the 'high ground'. Real life situations are messy for many reasons,
including the following:
- They are dynamic
- Issues cannot be considered in isolation
- Different people perceive the same issue differently
Different people have different values and objectives
I Melese [1987].
2 Russo and Shoemaker [1990].
3 Roy [1985] considers this activity more as an art than a science.
74 AIDING DECISIONS WITH MULTIPLE CRITERIA
Making sense
I: "That's a lot of food for thought! It sounds as if your boss is trying to
achieve a 'mission impossible'. It makes you wonder how anyone ever
manages to make sense of any situation, never mind decide on and
implement change. But of course, the way someone makes sense of the
situation will have a strong influence on how they might try to change it."
C: "Yes. Do you remember when the speaker mentioned various images
- or metaphors - of the organisation? She talked about visions originating
from different academic disciplines: the organisation as a body, a brain or a
machine 5 • Moreover, she suggested that a person's mental representation of
an organisation has strong implications for the way they conceptualise
interventions in the organisation. I can easily imagine that consultants tend
to use the models of their adopted metaphor: the surgeon uses a scalpel on a
body, the psychologist therapy on a mentally ill patient, the engineer a
screwdriver on a machine, etc.
I am quite sure that my boss thinks of the administration as a big
computer, which is dedicated to processing information. To a certain extent,
it is so, but I would feel more comfortable with a model that is less
restrictive. To me, each of these visions is describing only one dimension of
a whole."
I: "I don't know a lot about this; but I would tend to describe an
organisation as a "system" - it seems less restrictive than the metaphors you
mention. However, if you use this term there is a danger of people
interpreting it in the sense of cybernetics - that is, as a controllable system.
A more meaningful description is that of a 'negotiated system'."
C: "Yes, it seems a reasonable expression ... and it corresponds well to
my perception of my own institution. I am quite convinced that you cannot
oblige someone to do something they don't want to do. They might not
disobey openly, but they may deliberately attempt to sabotage the process or
through inaction allow the situation to rot'."
4 Many authors write about ways of understanding, interpreting and making sense of
situations. Weick [1995] brings together much of this discussion in an ongoing
conversation on sensemaking. Weick's view of sensemaking is encapsulated in the well-
known quotation "how can I know what I think until I see what I say". See also Dery
[\983] and Eden [1987].
5 Morgan [1989].
6 Bemoux [1985], following Crozier and Friedberg [1977] specify that in an organisation,
unlike technical systems, the people are not obliged to co-operate. Thus, there is a need for
some kind of agreement among them. See also Eden [1989].
Talking About the Practice ofMCDA 75
There are many different ways to view organisations, the way they work
and how decisions are taken within them. The following models are quite
influential:
- Simon [1976] changed the view of rationality within organisations,
shifting from substantive to procedural rationality and introducing the
notion of 'satisficing' (see also Bourgine and Le Moigne [1992]).
- Crozier [1977] founded the school of thought mown as the sociologie
des organisations. He analysed the way actors behave in order to increase
their power and influence on the system, highlighting the importance of
negotiated decisions (see also Bemoux [1985]).
- Mintzberg's model [1982] allows for different structures of organisation,
explained by the relative dominance of the different components (X, Y or
Z).
- Morgan [1989] proposes a number of metaphors for organisations and
suggests that one's understanding of how an organisation works and how
to change it is influenced by the image adopted.
Inter-organisational working
I: "I don't mow, perhaps you are right. .. I wonder what the implications
are for activities involving multiple organisations. In my own work, inter-
organisational structures are as important nowadays as intra-organisational
7 Eden [1992].
76 AIDING DECISIONS WITH MULTIPLE CRITERIA
Interventions
What's in a name?
I: "I have to admit that the name tends to put off clients who feel it
sounds terribly academic ... and, in any case, the academics themselves can't
decide what to call it!"
C: "Yes, that's a problem. Names are very important, as any marketing
expert will tell you. It's quite difficult to know what you are letting yourself
in for the first time you employ a new bunch of consultants. I don't know
how some people expect to get work when they can't convey clearly what
type of consultancy they are offering. Why can't the academics decide?"
I: "It is a long story. Historically, the first methods emerged under the
name of Multiple Criteria Decision Making (MCDM) - since they are
concerned with helping people make better decisions through taking account
of multiple factors, that makes some sense. However, one author argued that
this name was potentially misleading, as it engendered confusion between
the real-world, ongoing process of decision-making, which has to take into
account many factors and the contribution of these methods. So he proposed
instead Multiple Criteria Decision Aid (MCDA) - with the emphasis on
aiding or supporting the broader process of decision-making lO •
For the broader field, the same author proposed, using a similar
argument, that if there were to be a science, it could only a science about
how to help the people decide, and not a science of the decision itselfll."
C: "That's very interesting. But is it just that there are different views on
the name, or are there more fundamental differences between MCDM and
MCDA?"
I: "Hmm. That question would take a long time to answer fully. In fact,
there are many more than two approaches to MCDMlMCDA. These differ,
for example, theoretically in the assumptions they make about the way in
which preferences can be measured, and in the mathematics they employl2.
However, perhaps of even greater significance is the extent to which beliefs
about the nature of models and social intervention influences the way
consultants seek to intervene in a problem situation - this cuts across the
MCDMlMCDA division. For example, one of the most successful MCDA
practitioners bases his work on a theory drawn from the MCDM camp, but
applies this in a manner consistent with the philosophical basis of MCDA.
It's interesting to note - referring back to what we were just saying about
marketing - that he refers to his work as "Decision Conferencing" rather
than MCDN3. On a slightly more frivolous level, I am convinced that the
divide originates partly from the use of different languages: exponents of one
side use the French language as a way to resist!"
10 Roy [1985).
II Roy [1992,1994).
12 Roy and Vanderpooten [1996). Pictet and Belton [2000] try to reduce this gap.
13 Phillips [1990).
78 AIDING DECISIONS WITH MULTIPLE CRITERIA
difficult to convince potential clients of the benefits they can expect from
MCDA, particularly as these are often intangible, such as shared
understanding. But that takes us back to the nature of consultancy. You
made a comment that you don't know how people can expect to get work
when they are not able to describe what kind of consultancy they are offering
- it sounded as if you were speaking from bitter experience - have you had
someone try to sell you a magic potion which didn't work?"
Models of consultation
C: "No, not really. I was really thinking back to the comments made by
the speaker about different models of consultancyI4 and wondering which
applied in my current situation. One was the client as a purchaser of
expertise from the consultant; I thought that in this case the client would
need to know what they wanted before they could engage a consultant. This
seems to best describe the mode our vice-director is operating in, but as I
said earlier, I'm not sure he has a good enough understanding of our
organisation to know what is needed. In my view, the second model, the
notion of a partnership between the client and consultant, would be better
suited to our needs at the moment. In that model, the consultant helps the
client find out what their problem is and to work towards a solution. I can't
remember the third model ... "
I: "It was the doctor-patient analogy, whereby a consultant is employed
to diagnose what is wrong with the organisation and prescribe a solution. I
remembered that one because I was wondering whether it was an old-
fashioned relationship in which the doctor knew what was best for the
patient or a more modem one in which the patient is allowed to exercise
their own choice in selecting an appropriate treatment. I was trying to match
my own practice to one of the models."
C: "And did you succeed? Anyway, tell me more about yourself and
what you do ... we've been focusing on my problems. You said you were a
MCDA specialist?"
I: "Yes, but a rather junior one. I work with a very small company
"Better Solutions pIc" - just four of us - and I spend most of my time selling
MCDA to potential clients."
C: "Do you mean to imply that you spend all your time selling MCDA
rather than doing MCDA - and I'm still wanting to know what it is!"
I: "I do manage to spend sometime doing, but as I said earlier the
"selling" is a real challenge. At the moment I'm working on a project which
14 Schein [1988].
Talking About the Practice ofMCDA 79
15 About government procurement, see Roy and Bouyssou [1993]; Bana e Costa et al. [to be
published]; Pictet and Bollinger [2000).
80 AIDING DECISIONS WITH MULTIPLE CRITERIA
There are different ways in which someone can seek to help through
intervening in a problem situation.
"Selling" MCDA
I: "Well, that's a sensitive issue and it relates to your second question ...
there are many organisations facing issues of the type I described ... but there
is a problem finding enough who are willing to bring us in at the point at
which our expertise becomes relevant. We have good contacts, who we refer
to as our product champions, in a couple of organisations. They know what
we do and how we can help. These people are purchasing our expertise in
the use of MCDA to support decision making. But it is difficult getting into
new organisations. It comes back again to the question of what we're trying
to sell. Is it ourselves? Is it a process? Is it a method? You hit the nail on the
head earlier."
C: "Suppose I asked you to help me to make sense of the problem I
described earlier ... could you do that?"
which solution would best match your objectives. However, as I'm learning
from my experience and as your problem clearly demonstrates, there are
many different perspectives on the issue - and it sounds as if you don't yet
know what the objectives are."
C: "Does that mean that your models would be useless?"
I: "No ... I think they are still potentially useful. But you have to look at
them in a different light. They don't reflect a tangible 'reality' - it's not as if
there is something physical out there, like a manufacturing process, that we
are trying to capture in a simplified form. Our models are more about
intangibles - values, preferences, priorities - but it's not as if the model is
making explicit, or trying to simplify something which already exists in
someone's mind I6 ••• it's more about helping them (or more often helping a
group of people ... so there's more than one mind to worry about!) to
discover what is important to them - to learn about their values (and about
each other's values) - essentially to construct their preferences and to
facilitate their thinking."
C: "OK - that's an important aspect of our problem - but not actually
knowing what we want is only part of the whole issue."
The use of models is central to MCDA. However, the nature and meaning
of models can be perceived differently.
16 Roy [1989].
82 AIDING DECISIONS WITH MULTIPLE CRITERIA
be appropriate'. What ifit doesn't? What ifit emerges that something else is
needed?"
I: "That just adds further to the dilemma, doesn't it? I don't think it's
possible for a single person, or even a small team, to cover all the areas of
expertise that might be required. The ethical solution could be to build up a
network of contacts with companies having complementary expertise.
However, I fear that a more common response is to try to fit the problem to
your solution method ... you know the saying 'The danger, if all you have is
a hammer, is to see nails everywhere'19".
C: "Yes - and it relates back to what I was saying earlier about
consultants not being sensitive to the nature of the organisation. Can I pick
you up on something else you touched on earlier, working with groups of
people, how do you do that?"
19 B. Roy in Colasse and Pave [1997]. To a certain extent, it is very similar to the one of a
drunkard looking for his keys below a streetlight to take advantage of the light, even
though he knows he lost them somewhere in the dark [Roy, 1985).
20 Belton and Pictet [1997).
21 Sims [1979).
84 AIDING DECISIONS WITH MULTIPLE CRITERIA
22 Arrow [1951].
23 For a discussion, see Eden [1992]. For the related concept of 'procedural justice', see Eden
and Sims [1979].
24 Maystre and Bollinger [1999].
25 Edwards and Peppard [1994].
26 Huxham [1990].
Talking About the Practice ofMCDA 85
27 A distinction is often necessary between the demandeur (client) and the decideur
(decision-maker) [Roy, 1985].
86 AIDING DECISIONS WITH MULTIPLE CRITERIA
Afterword
This conversation, although completely fictional, summarises some of the
major issues a practitioner faces in her or his activity. Our aim in writing the
paper was to bring them to the attention of academics, particularly those who
are more often preoccupied with theory. As we have written eIsewhere28 , it is
our view that MCDA is a practical subject, which is worthless unless it is
applied, and so research and theoretical developments must be grounded in
practice. The development of theory, its implementation and evaluation in
practice should form a continuous loop, as proposed by Kolb29 • It may not be
the case that the theory is developed and the practice effected by the same
people. However, it is essential that practitioners and theoreticians
collaborate and communicate, in order that each is aware of the others
preoccupations, and through synergy to achieve the full potential for MCDA
as a management tool.
References
Ackermann F., Belton V., 1994, "Managing corporate knowledge experience with SODA and
VoloSoA", British Journal ofManagement 5, pp. 163-176.
Ackermann F., Belton V., 1999, Mixing methods: Balancing equivocality with precision,
Management Science Theory, Method and Practice Series 99/4, University of Strathclyde,
Glasgow.
Ackoff, R. L., 1981, "The art and science of mess management", Interfaces II, 20-26
Arrow K. J., 1951, Social choice and individual values, Wiley, New York.
Bana e Costa C. A., Antunes Ferreira J. A., Correa E. C., 1999, "A multicriteria methodology
supporting bid evaluation in public call for tenders" (to be published).
Bana e Costa C. A., Ensslin L., Correa E. C., Vansnick I-C., 1999,"Decision support systems
in action: integrated application in a multicriteria decision aid process", European Journal
of Operational Research 112 (2), pp. 315-335.
Belton V., Ackermann F., Shepherd I., 1997, "Integrative support from problem structuring
through to alternative evaluation using COPE and VoIoSoA", Journal of Multicriteria
Decision Analysis 6, pp. 115-130.
Belton V., Pictet I, 1997, "A framework for group decision using a MCDA model: Sharing,
aggregating or comparing individual information?", Journal of decision systems 6 (3),
pp. 283-303.
Bernoux P., 1985, La sociologie des organisations: Initiation, Seuil, Paris.
Bourgine P., Le Moigne J.-L., 1992, "Les 'bonnes decisions' sont-elles optimales ou
adequates ?", in Bourcier D., Costa I-P. (Eds), L 'administration et les nouveaux outits
d'aide ala deciSion, Editions STH, Paris.
Colasse B., Pave F., 1997, "Entretien avec Bernard Roy: La recherche operationnelle entre
acteurs et realites", Gerer et comprendre / Annales des Mines 47, pp. 16-27.
Crozier M., Friedberg E., 1977, L 'acteur et Ie systeme, Seuil, Paris.
Dery D., 1983, "Decision-making, problem-solving and organizational learning", Omega 11,
pp.321-328.
Eden C., 1987, "Problem-solving or problem-finishing?", in Jackson M., Keys P. (Eds), New
directions in management science, Gower, Hants.
Eden C., 1988, "Cognitive mapping", European Journal of Operational Research 36, pp. 1-
13.
Eden c., 1989, "Operational research as negotiation", in Jackson M., Keys P., Cropper S.
(Eds), Operational research and the social sciences, Plenum, New York.
Eden C., 1992, "A framework to think about group decision support systems", Group
Decision and Negotiation 1, pp. 199-218.
Eden C., Ackermann F., 1998, Making strategy: The journey of strategic management, Sage,
London.
Eden C., Huxham c., Vangen S., 1996, The dynamics of purpose in multi-organisational
collaborative groups: Achieving collaborative advantage for social development,
Management Science Theory, Method and Practice Series 96/3, University of StrathcJyde,
Glasgow.
Eden C., Sims D., 1979, "On the nature of problems in consulting practice", Omega 7 (2),
pp.119-127.
Edwards C., Peppard I W., 1994, "Business process redesign: hype, hope or hypocrisy?",
Journal of Information Technology 9, pp. 251-266.
Huxham c., 1990, "On trivialities of process", in Eden C. and Radford I (Eds), Tackling
Strategic Problems, Sage, London ..
Kolb D. A., Rubin I. M., McIntyre I M., 1984, Organisation psychology: An experimental
approach to organisational behaviour, Prentice-Hall, Englewood Cliffs.
Landry M., 1995, "A note on the concept of problem", Organization Studies 16 (2), pp. 315-
343.
Landry M., Banville C., Oral M., [1996], "Model legitirnisation in operational research",
European Journal of Operational Research 92, pp. 443-457.
Maystre L. Y., Bollinger D., 1999, Aide a la negociation multicritere, Presses polytechniques
et universitaires romandes, Lausanne.
Melese I, 1987, "Interventions systemiques dans les organisations", Revue internationale de
systemique 1 (4), pp. 457-470.
88 AIDING DECISIONS WITH MULTIPLE CRITERIA
Mintzberg H., 1982, Structure et dynamique des organisations, Editions d'organisation, Paris.
Morgan G., 1989, Images de I'organisation, Presses de I'Universite Laval et Editions Eska,
Quebec.
Phillips L. D., 1984, "A theory of requisite decision models", Acta Psychologica 56, pp. 29-
48.
Phillips L. D., 1990, "Decision analysis for group decision support", in Eden C. and Radford 1.
(Eds), Tackling Strategic Problems, Sage, London, pp. 142-150.
Pictet J., 1996, Depasser !'evaluation environnementale, Presses polytechniques et
universitaires roman des, Lausanne.
Pictet 1., Belton V., 1997, "MCDA: What Message?", Newsletter of the European Working
Group on Multicriteria Aid for Decisions II, pp. 1-3.
Pictet 1., Belton V., 2000, "ACIDE : Analyse de la compensation et de I'incomparabilite dans
la decision. Vers une prise en compte pratique dans MAVT", in AMCD - Aide Multicritere
ala Decision (Multiple Criteria Decision Aiding), Colomi A., Paruccini M., Roy B. (Eds),
Joint Research Centre, EUR Report, The European Commission.
Pictet J., Bollinger D., 2000, "Aide multicritere Ii la decision. Aspects mathematiques du droit
des marches publics", Baurecht I Droit de la construction 2/00, pp. 63-65.
Rittel H. W. 1., Webber M. M., 1973, "Dilemmas in a general theory of planning", Policy Sci.
4, pp. 155-169.
Roy B., 1985, Methodologie multicritere d 'aide a la decision, Economica, Paris. (English
version: Roy B., 1996, Multicriteria methodology for decision aiding, Kluwer, Dordrecht.)
Roy B., 1989, "Main sources of inaccurate determination, uncertainty and imprecision in
decision models", Mathematical Computer Modelling 12 (10111), pp. 1245-1254.
Roy B., 1992, "Science de la decision ou science de I'aide Ii la decision ?", Revue
internationale de systemique 6 (5), pp. 497-529. (English version: Roy B., 1993, "Decision
science or decision-aid science?", European Journal of Operational Research 66, pp. 184-
203.)
Roy B., 1994, "On operational research and decision aid", EURO Gold Medal Speech,
European Journal of Operational Research 73, pp. 23-26.
Roy B., Bouyssou D., 1993, Aide multicritere a la decision: Methodes et cas, Economica,
Paris.
Roy B., Vanderpooten D., 1996, "The European school of MCDA: Emergence, basic features
and current works", Journal ofMulti-Criteria Decision Analysis 5, pp. 22-38.
Schein E. H., 1988, Process consultation (Volume 1): its role in organization development,
Addison Wesley, USA
Russo J. E., Shoemaker P. 1. H., 1990, "Decision traps: Ten barriers to brilliant decision-
making and how to overcome them", Fireside, New York.
Schon D. A., 1987, Educating the reflective practitioner: towards a new design for teaching
and learning in the profeSSions, Jossey-Bass, San Francisco.
Simon H., 1976, "From substantive to procedural rationality", in Models of bounded
rationality (volume 2), 1982, The MIT Press, Cambridge Mass., pp. 424-443.
Sims D., 1979, "A framework for understanding the definition and formulation of problems in
teams", Human Relations 32, 909-921
Weick K. E., 1995, Sensemaking in Organisations, Sage, Thousand Oaks.
MULTI-CRITERIA DECISION-AID
IN A PHILOSOPHICAL PERSPECTIVE *
Jean-Louis Genard
ULB, FUSL, ISA "La Cambre", Belgium
jgenard@ulb.ac.be
Marc Pirlot
Faculte Poly technique de Mons, Belgium
marc.pirlot@fpms.ac.be
1. Introduction
According to B. Roy (Roy 1985, English version, p.1O), decision aiding
is the activity of the person who, through the use of explicit but not nec-
essarily completely formalized models, helps obtain elements of responses
to the questions posed by a stakeholder of a decision process. In his arti-
cle "Decision science or decision-aid science", Roy distinguishes in par-
ticular between the realist and constructivist approaches to decision-aid
(Roy 1992, English version: Roy 1993). According to the first approach,
• An initial version of this text was presented on 12 October 1995 at the 42th European
Multi-Criteria Decision-Aid Working Group Days in Namur, Belgium.
90 AIDING DECISIONS WITH MULTIPLE CRITERIA
Society). We are all too aware of the failure of the "solutions" that
Operational Research (OR) technicians seek to impose in the name of
"optimality". In a concrete problem, a deep understanding and dialogue
are essential between analyst(s) and decision-makers(s) - many essentials
can escape an outside consultant. In particular, who are the intervening
parties and the decision-makers, and what are the objective(s)? The raw
data, when these exist, require interpretation. Here the lack of precision,
the indeterminate state or the inaccessibility of certain data can lead to
the use of one model rather than another. English OR in particular has
been very sensitive to these questions and Rosenhead (Rosenhead 1989)
goes as far as to speak of revolution (in Kuhn's meaning of the term),
calling into question the dominant OR paradigm, i.e. the scientistic
and objectivist conception of decision-making problems (what Roy (Roy
1992), calls the "path of realism"). We will not enter into this debate,
but we will quote (Rosenhead 1989), p. 6, who expresses clearly the
extent to which the decision-making situation is not a given fact:
The clarity of the well-structured problem is simply unavailable, and an
OR approach which asserts otherwise does violence to the nature of the
situation.
"facts" are known through the judgements of experts) and, on the other
hand, the preferences of the decision-maker or the value judgements of
the intervening parties (to which the category of "true" does not apply).
This does not, however, mean that the validity criteria of decision-aid
are identical to those of decision-making. We shall attempt to clarify
this point later.
• and differentiated, that is that the spheres of the '!rue, the Good
and the Beautiful (the targets of Kant's three critiques) will in
future obey forms of validation and argument having differentiated
underlying logics.
It is this final point that Habermas was to pick up and amplify, but
from a somewhat different angle from that of M. Weber.
Whereas Weber interpreted his hypothesis of the polytheism of values
within an irrationalist-or more exactly decisionist perspective as regards
the political-ethical sphere (the choice of values being ultimately unde-
cidable, with reason applying solely to the field of science), Habermas
seeks, on the contrary, to rehabilitate a rationalist perspective, by ad-
mitting not one, but many forms of rationality and validation.
Here is what he writes (Habermas 1970, p. 285):
I will defend the idea that there exist at least four equally original preten-
sions to validity, and that these four pretensions, that is intelligibility,
truth, justness and sincerity form a whole that we can call rationality.
ity, none of which, however, appears to be able to account for the process
as a whole.
3.4. Habermas'positions
Before going further and addressing the more specific question of the
relationship between these theories of Habermas and the questions posed
by decision-aid, we would like to draw the reader's attention to three
points that seem to us to be important.
1 Through his proposals, Habermas seeks to distance himself from
two positions that were very widespread during the 20th century:
• a positivist or scientistic rationalism which tends to reduce
normative questions to factual, technical or similar questions,
that can then be decided according to the forms of rationality
specific to the sphere ofthe true (a trend which was very cer-
tainly present in decision-sciences from the outset and which
is still very likely present today) j
• decisionism according which ultimate choices of values are in
any event irrational and boil down therefore to subjective and
non-argument able preferences (a position which would refuse
any ambition to produce more just norms). This was the
position of M. Weber himself (Weber 1919).
2 Very generally, Habermas adheres to what he calls a consensus
theory of truth. For him, the truth of a proposition is intrinsically
linked to the ability to justify it in a discussion. This means that
he rejects theories which define truth as conformity with an ob-
ject considered as given. This consensual theory of truth applies,
for him, very certainly both to the sphere of truth itself and to
that of normative justness. However, given the nature of the types
of rationality specific to each of these spheres, he insists on the
fact that if the results of practical discussions can lead to chang-
ing social reality, theoretical discussions cannot be directed against
reality (nature) itself, but only against false affirmations about re-
ality (Habermas 1970, p. 296). In this way he gives an intrinsi-
cally constraining dimension to the objective world that gives it
its specificity, but without for as much adhering to a realist vision
of scientific theories.
3 With regard to the validation of regulatory statements (sphere of
the "good"), his reflections were to lead him to a rehabilitation
of discursive practice, argumentation, and public debate. For him,
Philosophical Perspective 99
Nor has it changed the concept of time in its everyday use. Another
aspect, illustrated by the particle and wave interpretations of quantum
mechanics, is that two descriptions of the same reality can exist side
by side (the wave-particle duality) and both prove pertinent in different
contexts.
the value (in the sphere of the good) of the decision that will be taken.
In developing the model, particular attention can be paid to supporting
the decision-maker's position as well as possible, to putting together the
"file" that will be the most convincing for all the parties concerned. This
is a form of "external" validity for the model; it is here too that qualities
such as the transparency of the model can reveal themselves, as this can
contribute to establishing the climate of trust which is necessary for a
true public discussion.
precise and technical form, have no place here as they would generally
appear as being technocratic, unless the decision-maker himself is able
to perceive their precise scope. In any event, they cannot prevail against
the decision-maker's perceptions of the decision-making situation and of
his own preferences. In particular, where the decision-maker formulates
preferences which, given what is known (for example, evaluations of al-
ternatives), appear to the analyst to be incompatible with one form of
rationality (often translated into axioms, such as monotonicity, transi-
tivity of preferences, etc.), and where the decision-maker sticks to his
viewpoint after the analyst has drawn his attention to what he considers
to be incoherent, it is the analyst's task to question the data, the model
and/or the axioms. In any event it would be totally aberrant to force
the decision-maker to adopt a way of thinking which seems foreign to
him-this would inevitably lead to a breakdown of dialogue and of the
whole decision-aid process.
In general therefore, whilst remaining transparent for the decision-
maker, the theoretical results may, during an aid process, lead the an-
alyst to pose questions (to himself) as to the pertinence of a formal
procedure. One typical case would be for example the use of a proce-
dure making full use of the cardinality of a numerical representation,
whereas it would seem that the decision-maker considers this informa-
tion as ordinal. What we have here are demands for formal consistency
and "faithfulness of translation" that can be borne only by the analyst.
It is in these demands that the analyst places the essential of what,
for him, makes up the validity of the model (the "logical validity" us-
ing the terminology of Oral and Kettani 1993). Where the analyst is
careful to avoid a normative and dogmatic attitude to rationality or a
choice of model, the theoretical results by developing the consequences
of the axioms to their logical conclusion, suggest means of directing the
decision-maker's attention to the implications of his affirmations. For
example, in cases of decision-making under uncertainty, one can invoke
"money-pump" type arguments. Of course, we repeat, it is the affir-
mations of the decision-maker which, in the final instance, carry the
day, possibly forcing the analyst to change the bases on which he has
constructed the decision-aid model.
Apart from suggesting to the analyst ways of testing the pertinence
of the use of a particular procedure in a particular decision-making pro-
cess, axiomatic characterisations can fulfil other roles. More positively,
they can be advanced in order to demonstrate to the decision-maker the
internal consistency, or the limitations, of a particular approach. For ex-
ample, it seems to us that Arrow's theorem (interpreted in the context of
decision-making in the presence of multiple criteria: see Bouyssou 1992
112 AIDING DECISIONS WITH MULTIPLE CRITERIA
and Perny 1992) enables us to understand and accept the following fact:
one cannot expect that the comparison of two alternatives, based mainly
on ordinal considerations, not take into account the other available alter-
natives (independence of irrelevant alternatives) and at the same time,
that the global preference be systematically transitive.
This having been said, and the reader will already have read this
between the lines above, there is nothing to oppose a constructivist
approach aimed at constructing a global utility or a global evaluation
function. The result remains pertinent: it can be interpreted as a pos-
sibility result of an approach consisting of constructing a global evalu-
ation function in an additive fashion. It therefore suggests a particular
construction strategy consisting of collecting information ("preference
fragments") from which it is then possible to construct a partial utility
function for each criterion and to assess the trade-offs, and then sum
these functions, weighted by the trade-offs. Here the theoretical result
serves both to guarantee the consistency of the approach and to define
which "fragments" are needed (see, for example, Fishburn 1967 for a
description of 24 methods of constructing an additive utility function;
these methods apply in different contexts and are based on the gath-
ering of preference fragments. One may even, in this context, consider
abandoning the hypothesis that the global preference is a ranking by
interpreting (exploiting) the global evaluation function more leanly, for
example by using a threshold: given the imprecise nature of the matter
we are dealing with, one can consider that one alternative is globally
preferable to another if the difference between the evaluations of them
exceeds a threshold.
7. Conclusions
It seems to us that Habermas's theory of orders of validity supplies
a framework which allows us to situate in a light-shedding manner the
different decision-aid approaches from an epistemological viewpoint in
the broad sense of the term. In particular, comparing the realist and
constructivist approaches, it seems to us that one of the strengths of con-
structivism is to make it more possible to maintain a true dialogue with
the decision-maker(s). This way of conceiving the validity of a model,
based on genuine dialogue and critical discussion, far from being specific
to the decision-aid process and opposed to the model validation meth-
ods of the sciences, represents a continuation of the validation models
that are valid in all the spheres defined by Habermas, as also in day-to-
day discursive practice. The particular context of decision-aid simply
reduces the "public" discussion to its simplest expression, a form of di-
alogue, and means that the resulting model is generally inapplicable to
third parties. On the other hand, the types of "realities" reflected by the
model are fairly particular, consisting as they do of the decision-maker's
own evaluations and preference judgements. The model constructed in
this way therefore remains hypothetical.
114 AIDING DECISIONS WITH MULTIPLE CRITERIA
In order to reassure (or further disturb) the reader who is afraid that,
with these concepts, the practice of decision-aid moves too far away from
scientific rigour, we shall end by mentioning two positions that show that
similar problems are echoed in other disciplines.
Acknowledgments
The authors would like to thank Bernard Roy for the interest he
showed for this work and for his pertinent comments on the original
version. They are indebted to Denis Bouyssou for his many comments
on the second version of this work, which have helped clarify certain
points that were poorly or insufficiently developed. His remarks together
with those of two anonymous referees helped us to improve the overall
readability of the paper. The authors thank Michael Lomax for the
quality of his translation into English of the original text.
References
R.L. Ackoff (1979) "The future of operational research is past," J. OpJ. Res. Soc. 30
(2): 93-104.
R.L. Ackoff (1979) "Resurrecting the future of operational research," J. OpJ. Res.
Soc. 30 (3): 189-199.
D. Bouyssou (1992) "On some properties of outranking relations based on a concordance-
discordance principle," in A. Goicoechea, L. Duckstein and S. Zionts (eds.), Mul-
tiple criteria decision making. Berlin: Springer 93-106.
116 AIDING DECISIONS WITH MULTIPLE CRITERIA
D. Bouyssou, P. Perny, M. Piriot, A. Tsoukias, P. Vincke (1993) "A Manifesto for the
new MCDM era," Journal of Multicriteria Decision Analysis, 2: 125-127.
D. Bauyssou, T. Marchant, P. Perny, M. Pirlot, A. Tsoukias, P. Vincke (2000) Eval-
uation and decision models: a critical perspective. Dordrecht: Kluwer Acad. Pub!.
R. Dery, M. Landry, C. Banville (1993) "Revisiting the issue of model validation in
OR: an epistemological view," European Journal of Operational Research, 66 (2):
168-183.
P.C. Fishburn (1967) "Methods of estimating additive utilities," Management Science,
13 (7): 435-453.
H.G. Gadamer (1960) Wahrheit und Methode. Tiibingen: J.C.B. Mohr. French trans!.:
Verite et methode (1976). Paris: Seui!.
J. Habermas (1970) Zur Logik der Sozialwissenschaften. French trans!.: Logique des
sciences sociales et autres essais (1987). Paris: PUF. English trans!.: On the logic
of social sciences (1990). Boston: MIT Press.
I. Kant (1797) Die Metaphysik der Sitten. French trans!.: Fondements de La metaphysique
des moeurs (1974). Paris: Delagrave.
G.B. Kleindorfer, L. O'Neill, R. Ganesham (1998) "Validation in simulation: various
positions in the philosophy of science," Management Science, 44 (8):1087-1099.
T.S. Kuhn (1962) The structun of scientific revolutions. Chicago: University of Chicago
Press.
T. S. Kuhn (1977) The essential tension. Chicago: University of Chicago Press.
M. Landry 1998 "L'aide a la decision comme support a la construction du sens dans
I'organisation," Systemes d'Information et Management, 3: 5-39.
M. Landry, C. Banville, M. Oral (1983) "Model validation in Operations Research",
European Journal of Operational Research, 14: 207-220.
M. Landry, C. Banville, M. Oral (1996) "Model legitimation in Operations Research,"
European Journal of Operational Research, 92: 443-457.
J.-L. Le Moigne (1980) "Les sciences de la decision: sciences d'analyse ou sciences de
genie? Interpretations epistemologiques," in R. Nadeau et M. Landry (eds.), L'aide
d La decision-Nature, instruments et perspectives d'avenir. Quebec: Les Presses de
I'Universite Lava!.
Th. Marchant, M. Piriot (1999) "Modern decisive wives don't wear corsets," Journal
of Multi-Criteria Decision Analysis, 8: 237-238.
P. McCullagh, J.A. Neider (1983) Generalised linear models. London: Chapman and
Hal!.
M. Oral, O. Kettani (1993) "The facets of the modeling and validation process in
operations research," European Journal of Operational Research, 66 (2): 216-234.
P. Perny (1992) "Sur Ie non-respect de l'axiome d'independance dans les methodes
de type ELECTRE," Bruxelles: Cahiers du CERO 34: 211-232.
R. Rorty (1982) Consequences of pragmatism. Minnesota: University of Minnesota
Press. French trans!.: Consequences du pragmatisme (1993). Paris: Seui!.
R. Rorty (1991) "Objectivity, relativism and truth", Philosophical papers 1. Cam-
bridge: University Press. French trans!.: Objectivisme, relativisme et verite (1994).
Paris: Presses Universitaires de France.
J. Rosenhead (ed.) (1989) Rational analysis for a problematic world. Chichester: Wi-
ley.
REFERENCES 117
Denis Bouyssou
CNRS - LAMSADE, Universite de Paris Dauphine, France
bouyssou@lamsade.dauphine.fr
Marc Pirlot *
Faculte Poly technique de Mons, Belgium
marc.pirlot@fpms.ac.be
Abstract Based on a general framework for conjoint measurement that allows for
intransitive preferences, this paper proposes a characterization of "strict
concordance relations". This characterization shows that the originality
of such relations lies in their very crude way to distinguish various levels
of "preference differences" on each attribute.
1. Introduction
A basic problem in the field of Multiple Criteria Decision Making
(MCDM) is to build a preference relation on a set of alternatives evalu-
ated on several attributes on the basis of preferences expressed on each
attribute and inter-attribute information such as weights or trade-offs.
B. Roy proposed several outranking methods (see Roy, 1968; Roy,
1996b; Roy and Bouyssou, 1993; Vincke, 1992; Vincke, 1999; Bouys-
sou, 2001) as alternatives to the dominant value function approach (see
Fishburn, 1970; Keeney and Raiffa, 1976; Wakker, 1989). In outranking
methods, the construction of a preference relation is based on pairwise
• Corresponding author: Marc Pirlot, Faculte Polytechnique de Mons, rue de Houdain 9, 7000
Mons, Belgium
122 AIDING DECISIONS WITH MULTIPLE CRITERIA
of Pi implies that one and only one of the following propositions is true:
XiPiYi or YiPiXi or XiIiYi (i.e. N ot[XiPiYi] and N ot[YiPiXi])
When comparing x to y, the following subsets of attributes play a
vital part in TACTIC:
xPy {:}
L':iEP(X,y) Wi > P L':jEP(y,X) Wj
and (2)
N ot[Yj VjXj] for all j E P(y, x)
124 AIDING DECISIONS WITH MULTIPLE CRITERIA
where Pi and Vi are strict semiorders such that Vi ~ Pi, Wi > 0 and
p ~ 1. We refer to Vansnick, 1986 for a thorough analysis of this method
including possible assessment techniques for Pi, Vi, Wi and p.
Simple examples show that, in general, a relation P built using (1)
or (2) may not be transitive and may even contain circuits. The use
of such a relation P for decision-aid purposes therefore calls for the
application of specific techniques, see Roy, 1991; Roy and Bouyssou,
1993; Vanderpooten, 1990.
Proof of proposition 1
i. The asymmetry of ~ follows from the skew symmetry of all Pi and
the oddness of F. Since Pi(Xi, Xi) = 0, the independence of ~ follows.
ii. Observe that Xi ~i Yi is equivalent to F(Pi(Xi, yd, 0) > (using
obvious notations). Since F(O) = 0, the nondecreasingness of F leads to
°
Pi(Xi, Vi) > 0. The desired property easily follows using the nondecreas-
ingness of F. 0
for all Xi, Yi, Zi, Wi E Xi and all a-i, L i , C-i, d_ i E X-i. We say that r-
satisfies ARC1 (resp. ARC2) if it satisfies ARC1i (resp. ARC2i) for all
i E N.
Condition ARC1i (Asymmetric inteR-attribute Cancellation) suggests
that r- induces on Xl a relation that compares "preference differences"
in a well-behaved way: if (Xi, Yi) is a larger preference difference than
(Zi,Wi) and (zi,c-d r- (wi,d_i) then we should have (xi,c-d r- (Yi,d_d
and vice versa. The idea that the comparison of preference differences is
central to the analysis of conjoint measurement models was powerfully
stressed by Wakker, 1988; Wakker, 1989.
Condition ARC2i suggests that the preference difference (Xi, Yi) is
linked to the "opposite" preference difference (Yi, Xi). It says that if the
preference difference between Zi and Wi is not larger than the preference
difference between Xi and Yi then the preference difference between Wi
and Zi should be larger than the preference difference between Yi and
Xi· Taking Xi = Yi, Zi = Wi, a-i = Ci and b-i = d-i shows that ARC2i
implies that r- is independent for N \ {i} and, hence, independent.
The following lemma shows that these two conditions are independent
and necessary for model (M).
Lemma 1
z. Model (M) implies ARC1 and ARC2,
ii. In the class of asymmetric relations, ARC1 and ARC2 are inde-
pendent conditions.
Proof of lemma 1
i. Suppose that (xi,a-i) r- (Yi,b_d and (zi,c-d r- (wi,d_i). Using
model (M) we have:
and
F(Pi(Zi, Wi), (Pj(Cj, dj))j:;I=i) > 0
abusing notations in an obvious way.
If Pi(Xi, Yi) ~ Pi(Zi, Wi) then using the nondecreasingness of F, we
have F(Pi(Xi, Yi), (Pj(Cj, dj))j:;I=i) > 0 so that (Xi, c-d r- (Yi, d-i). If
Pi(Zi, Wi) > Pi(Xi, Yi) we have F(Pi(Zi, Wi), (pj(aj, bj))j:;I=i) > 0 so that
(Zi' a-i) r- (Wi, b-i). Hence ARC1 holds.
Similarly, suppose that (Xi, a-i) r- (Yi, b-i) and (Yi, cd r- (Xi, d-d·
We thus have:
128 AIDING DECISIONS WITH MULTIPLE CRITERIA
and
F(Pi(Yi, Xi), (Pj(Cj, dj))ji:i) > O.
If Pi(Xi,Yi) ~ Pi(Zi,wd, the skew symmetry of Pi implies Pi(Wi,Zi) >
Pi(Yi,xd. Using the nondecreasingness of F we have F(Pi(Wi,Zi), (Pj(Cj,
dj))ji:i) > 0 so that (Wi, C-i) >- (Zi' d-i). Similarly, ifpi(zi, Wi) > Pi(Xi, Yi)
we have, using the nondecreasingness of F, F(Pi(Zi, Wi)' (pj(aj, bj))ji:i) >
o so that (Zi' a-i) >- (Wi, b_ i ). Hence ARC2 holds.
ii. It is easy to build asymmetric relations violating ARC1 and ARC2.
Using theorem 1 below, it is clear that there are asymmetric relations
satisfying both ARC1 and ARC2. We provide here the remaining two
examples.
2 Let X = {a,b} x {x,y} and >- on X be empty except that (a,x) >-
(a, y). It is clear that >- is asymmetric but not independent, so
that ARC2 is violated. Condition ARC1 is trivially satisfied. 0
and
Lemma 2
i ARC1i {:} [ti is complete],
ii ARC2i {:}
flor all Xi,Yi,Zi,Wi E Xi,Not[(Xi,Yi) ti (Zi,Wi)] ~ (Yi,xd ti
(Wi, zdJ,
iii [ARC1i and ARC2i] {:} [ti* is complete].
For the sake of easy reference, we note a few useful connections be-
tween ti, ti* and >- in the following lemma.
:::i,
let Pi(Xi, Yi) = qi(Xi, Yi) - qi(Yi, Xi). It is obvious that Pi is skew
symmetric and represents :::i*.
Define F as follows:
where 9 is any function from IR.n to IR. increasing in all its arguments and
odd (e.g. E) and f is any increasing function from IR. into (0, +00) (e.g.
exp(·) or arctan(·) + ~).
The well-definedness of F follows from part iv. of lemma 3 and the
definition of the pi'S. It is odd by construction.
To show that F is nondecreasing, suppose that Pi(Zi, Wi) > Pi(Xi, Yi),
i.e. that (Zi, Wi) ~i* (Xi, Yi). If X ~ Y, we know from part i. of lemma 3
that (Zi, X-i) ~ (Wi, Y-i) and the conclusion follows from the definition
of F. If x'" Y, we know from part iii. of lemma 3 that Not[(Wi, Y-i) ~
(Zi, X-i)] and the conclusion follows from the definition of F. If Y ~ X we
have either (Wi,Y-i) ~ (Zi,X-i) or (Zi,X-i) ::: (Wi,Y-i). In either case,
the conclusion follows from the definition of F. 0
(4)
I
Hence IPi(Xl> is an upper bound for the number of equivalence classes
of :::i*.
Strict Concordance Relations 131
4. A characterization of
strict concordance relations
Our main result in this section says that all strict concordance rela-
tions (definition 1) can be represented in model (M) with relations ti*
having at most three equivalence classes and vice versa.
Theorem 2 The following are equivalent:
i >- has a representation in model (M) with all relations ti* having
at most three distinct equivalence classes,
ii >- is a strict concordance relation.
Proof of theorem 2
ii => i. Given equation (4), the claim will be proven if we build a
representation of >- in model (M) with functions Pi taking only three
distinct values. Define Pi as:
I if XiPiYi,
Pi(Xi, yd = { 0 if xihYi,
-1 ifYiPixi.
Since ~ is asymmetric, the function Pi is well-defined and skew-symmetric.
Define F letting:
I if x >- y,
F(PI (Xl, Yl),P2{X2, Y2), ... ,Pn(xn , Yn)) = { -1 if Y >- x,
o otherwise.
Since, by hypothesis, [P{x, y) = P(z, w) and P(y, x) = P(w, z)] => [x >-
Y ¢:} z >- w], it is easy to see that F is well-defined. It is clearly odd. The
monotonicity of [> implies that F is nondecreasing in all its arguments.
i => ii. Define Pi letting, for all Xi, Yi E Xi, Xi~Yi ¢:} (Xi, Yi) >-i*
(Yi,Yi).
Suppose that XiPiYi and YiPiXi so that (Xi, Yi) >- i* (Yi, Yi) and (Yi, Xi)
>-i* (Xi, Xi). Since >- is independent, we have (Yi, Yi) "'i* (Xi, xd so that
(Yi, Xi) >-i* (Yi, Yi). The reversibility of ti* leads to (Yi, Yi) >-i* (Xi, Yi),
a contradiction. Hence, Pi is asymmetric.
Two cases arise:
• If attribute i E N is degenerate then >-i= 0. Hence ti* has only
equivalence class and Pi is empty. We clearly have [Xi Ii Yi and
ziIiwi] => (Xi,Yi) rvi* (zi,wd·
• If attribute i E N is influent, we claim that Pi is non empty
and that ti* has exactly three equivalence class. Indeed, ti
132 AIDING DECISIONS WITH MULTIPLE CRITERIA
being complete, there are Zi, Wi,Xi, Yi E Xi such that (Xi, Yi) >-i
(Zi' Wi). Since ti* is complete, this implies (Xi, Yi) >-i* (Zi' Wi). If
(Xi,Yi) >-i* (Yi,Yi) then XiPiYi. If not, then (Yi,Yi) ti* (Xi,Yi) and
ti* being a weak order, we obtain (Yi,Yi) >-i* (Zi,Wi). Using the
definition of ti*, this clearly implies (Wi, Zi) >- i* (Yi, yd. Since>- is
independent, we have (Yi,Yi) "'i* (zi,zd· Thus (wi,zd >-i* (Zi,Zi)
so that WiPiZi. Therefore Pi is not empty.
Since ti* has at most three distinct equivalence classes and XiPiYi
{:::} (Xi, Yi) >-i* (Yi, Yi) {:::} (Yi, Yi) >-i* (Yi, Xi), we conclude that ti*
has exactly three distinct equivalence classes. Therefore, XiPiYi
implies that (Xi, Yi) belongs to the first equivalence class of ti*·
This implies [XiPiYi and ZiPiWi] => (Xi, Yi) "'i* (Zi' Wi). Similarly,
it is easy to prove that [xi1iYi and zi1iwi] => (Xi, Yi) "'i* (Zi' Wi).
Therefore, [P(x,y) = P(z,w) and P(y,x) = P(w,z)] implies [(Zi,Wi)
"'i* (Xi, Yi), for all i EN]. From part iv. of lemma 3 we obtain:
[P(X, y) = P(z, w) and P(y, x) = P(w, z)] => [x >- Y {:::} Z >- w]. (5)
[P(X, y) ~ P(z, w) and P(y, x) :2 P(w, z)] => [x >- Y => Z >- w]. (6)
A c> B {:::}
[x >- y, for some X,Y E X such that P(x,y) = A and P(y,x) = B]
Equations (5) and (6) show that c> is asymmetric and monotonic. In
view of (5), it is clear that:
cannot be beaten and its "opposite" cannot beat any preference differ-
ence. Similarly, if a preference difference is smaller than a null preference
difference, then it cannot beat any preference difference and its "oppo-
site" cannot be beaten. It is not difficult to find relations >- satisfying
Ci but not Cj for j =1= i. We say that >- is coarse (C) if it is coarse on
alliEN.
Proposition 2 We have:
C, ARC1 and ARC2 are independent conditions,
zz if>- satisfies ARC1 and ARC2 then [C holds] ¢:} [~i* has at most
three equivalence classes, for all i EN].
Proof of proposition 2
i. Using a nontrivial additive utility model, it is easy to build ex-
amples of relations satisfying ARC1 and ARC2 and violating C. The
two examples used in the proof of part ii .. of lemma 1 show that there
are asymmetric relations >- satisfying C and ARC1 (resp. ARC2) but
violating ARC2 (resp. ARC1).
ii. Suppose that ARC1 and ARC2 hold. Let us show that [~i* has
at most three equivalence classes, for all i E N] ::} C. Suppose that C
is violated with (Xi, Yi) >-i (Yi, Yi). We have either (Zi' wd >-i (Xi, Yi)
or (Yi, Xi) >-i (Wi, zd, for some Zi, Wi E Xi. Since >-i ~ >-i* and ~i*
is a reversible weak order, it is easy to see that either case implies that
~i* has at least five equivalence classes. The case (Yi, yd >-i (Yi, xd is
similar.
Let us now show that C ::} [~i* has at most three equivalence classes,
for all i EN]. Suppose that (Xi,Yi) >-i* (Yi,Yi) so that either (xi,yd >-i
(Yi, Yi) or (Yi, yd >-i (Yi, Xi)' In either case, C implies, for all Zi, Wi E Xi,
(Xi, yd ~i (Zi' wd and (Wi, Zi) ~i (Yi, Xi) so that (Xi, Yi) ~i* (Zi' Wi).
Therefore if (Xi, Yi) >-i* (Yi, Yi) then (Xi, Yi) ~i* (Zi' Wi) for all Zi, Wi E
Xi. Similarly, it is easy to prove that (Yi, Yi) >-i* (Xi, Yi) implies (Zi' Wi)
~i* (Xi, yd for all Zi, Wi E Xi. This implies that ~i* has at most three
equivalence classes. 0
}~
(Xi, X-i) ~ (Yi,Y-i) and Not[(Yi,X-d ~ (Yi,Y-i)]
or
N ot[(Yi, x-d ~ (Xi, Y-i)] and (Yi, x-d ~ (Yi, Y-i)
iii Xj >-j Yj for some j E N and Xi "'i Yi for all i E N \ {j} :::} x >- y,
iv all influent attributes are essential.
Proof of proposition 4
i. Since "'i is reflexive by construction, the definition of noncompensa-
tion implies that >- is independent for N\ {i}. Hence, >- is independent.
ii. Suppose that Xi "'i Yi for all i EN and x >- y. Since >- is noncom-
pensatory and "'i is reflexive, this would lead to x >- x, contradicting
the asymmetry of >-.
iii. By definition, x >-i Y ¢:} [(Xi, Z-i) >- (Yi, z-d for all Z-i E X-i].
Since "'i is reflexive, the desired conclusion follows from the definition
of noncompensation.
iv. Attribute i E N being influent, there are Xi, Yi,Zi, Wi E Xi and
X-i, Y-i E X-i such that (Xi, X-i) >- (Yi, Y-i) and N ot[(Zi, x-d >- (Wi, Y-i)].
In view of NC, it is impossible that Xi "'i Yi and Zi "'i Wi. Hence at-
tribute i is essential. 0
Definition 3
The binary relation >- is said to be monotonically noncompensatory (in
the asymmetric sense) if:
Proof of proposition 5
i. ::::} ii. Since each attribute is essential, it is easy to see that {i} [> 0
so that Pi = h. The conclusion therefore follows.
ii. ::::} i. Letting Pi = >-i and defining [> by:
[A [> B) {:}
[x >- y, for some X,y E X such that >- (x,y) = A and >- (y,x) = B).
easily leads to the desired conclusion. o
Strict Concordance Relations 137
ii does not allow to point out the specific features of strict concor-
dance relations within a general framework of conjoint measure-
ment (conditions NC and MNC are indeed quite different from the
classical cancellation conditions used in most conjoint measure-
ment models, and most importantly, the additive utility model
(see Krantz et al., 1971; Debreu, 1960; Fishburn, 1970; Wakker,
1989)),
iii amounts to using very strong conditions (see the simple proof of
proposition 5).
Proof of proposition 6
We say that a nonempty set J ~ N is:
• decisive if, for all x, Y E X, [XiPiYi for all i E J] => x ~ y,
• semi-decisive if, for all x, Y EX, [XiPiYi for all i E J] => N ot[y ~
x],
Hence, an oligarchy 0 is a decisive set such that all {i} ~ a are semi-
decisive.
Since >- is a strict concordance relation, it is easy to prove that:
[P(X,y) = J,P(y,x) = N\ J and x ~ y, for some X,Y E X]
=> J is decisive,
and
[P(x,y} = J,P(y,x) = N \ J and Not[y ~ x], for some X,y E X]
Strict Concordance Relations 139
::::} J is semi-decisive.
Since )- is nonempty, we have, for all x, Y E X:
so that N is decisive.
Since N is finite, there exists (at least) one decisive set of minimal
cardinality. Let J be one of them. We have [XiPiYi for all i E J] ::::}
x )- y. If IJI = 1, then the conclusion follows. If not, consider i E J
and use the elements ai, bi , Ci E Xi such that aiPibi, biPiCi and aiPici to
build the following alternatives in X:
{i} J \ {i} N\J
a Ci aj bt
b ai bj Ct
c bi Cj at
5.5. Discordance
An immediate generalization of definition 1 is the following:
Definition 4 (Strict concordance-discordance relations)
A binary relation P on X is said to be a strict concordance-discordance
relation if there are:
• an asymmetric binary relation!> between disjoint subsets of N that
is monotonic and,
• asymmetric binary relations Pi and Vi such that Vi ~ Pi on each
Xi (i = 1,2, ... ,n),
such that, for all x, y E Y:
xPy ¢:} [P(x,y)!> P(y,x) and (Not[YjVjXj]' for all j E P(y,x))],(8)
Proposition 7
If >- is a strict concordance-discordance relation then >- satisfies
model (M) with all relations ti* having at most 5 distinct equiva-
lence classes.
ii There are relations t satisfying model (M) with all relations ti*
having at most 5 equivalence classes which are not strict concordan-
ce-discordance relations.
Proof of proposition 7
i. Given the properties of model (M), the claim will be proven if we
build a representation of >- in model (M) with functions Pi taking only
142 AIDING DECISIONS WITH MULTIPLE CRITERIA
2 if Xi ViYi,
1 if XiPiYi and N ot[Xi ViYi],
o if xi1iYi,
-1 if YiPiXi and N ot[Yi ViXi],
-2 if Yi ViXi.
1 if x ~ y,
-1 if Y ~ x,
o otherwise.
Using the definition of a strict concordance-discordance relation, it is
routine to show that F is well-defined, odd and nondecreasing.
ii. Using an additive utility model, it is easy to build examples of
relations having a representation in model (M) with all relations ti*
having at most 5 equivalence classes which are not strict concordance-
discordance relations. 0
5.6. Discussion
The main contribution of this paper was to propose a characterization
of strict concordance relations within the framework of a general model
for nontransitive conjoint measurement. This characterization allows
to show the common features between various conjoint measurement
models and to isolate the specific feature of strict concordance relations,
i.e. the option not to distinguish a rich preference difference relation on
each attribute. It was shown to be more general than previous ones
based on NC or MNC.
Although we restricted our attention to asymmetric relations, it is not
difficult to extend our analysis, using the results in Bouyssou and Pirlot,
2000, to cover the reflexive case studied in Fargier and Perny, ming in
which:
xSy {::> [S(x, y) ~ S(y, x)]
where S is a reflexive binary relation on X, Si is a complete binary
relation on Xi, ~ is a reflexive binary relation on 2N and S(x, y) = {i E
N: xiSiyd.
REFERENCES 143
Acknowledgments
We wish to thank Patrice Perny for his helpful comments on an earlier
draft of this text. The usual caveat applies.
References
Arrow, K.J. (1963). Social choice and individual values. Wiley, New York, 2nd edition.
Bouyssou, D. (1986). Some remarks on the notion of compensation in MCDM. European
Journal of Operational Research, 26:150-160.
Bouyssou, D. (1992). On some properties of outranking relations based on a concordance-
discordance principle. In Duckstein, L., Goicoechea, A., and Zionts, S., editors,
Multiple criteria decision making, pages 93-106. Springer-Verlag, Berlin.
Bouyssou, D. (1996). Outranking relations: Do they have special properties? Journal
of Multi-Criteria Decision Analysis, 5:99-11l.
Bouyssou, D. (2001). Outranking methods. In Floudas, C. and Pardalos, P., editors,
Encyclopedia of optimization. Kluwer.
Bouyssou, D., Perny, P., and Pirlot, M. (2000). Nontransitive decomposable conjoint
measurement as a general framework for MCDM and decision under uncertainty.
Communication to EURO XVII, Budapest, Hungary, 16-19 July.
Bouyssou, D., Perny, P., Pirlot, M., Tsoukias, A., and Vincke, Ph. (1993). A manifesto
for the new MCDM era. Journal of Multi-Criteria Decision Analysis, 2:125-127.
Bouyssou, D. and Pirlot, M. (2000). Non transitive decomposable conjoint measure-
ment: General representation of non transitive preferences on product sets. Work-
ing Paper.
Bouyssou, D. and Pirlot, M. (2001). 'Additive difference' models without additivity
and subtractivity. Working Paper.
Bouyssou, D. and Vansnick, J.-C. (1986). Noncompensatory and generalized noncom-
pensatory preference structures. Theory and Decision, 21:251-266.
144 AIDING DECISIONS WITH MULTIPLE CRITERIA
Alexis Tsoukias
LAMSADE - CNRS, Universite Paris Dauphine, France
tsoukias@lamsade.dauphine.fr
Patrice Perny
LIP6, Universite Paris 6, France
Patrice.Perny@lip6.fr
Philippe Vincke
SMG - ISRO, Universite Libre de Bruxelles, Belgium
pvincke@ulb.ac.be
1. Introduction
Consider a Parliament. The government has the support of the ma-
jority of seats, although not a very strong one. Suppose now that a law
on a very sensitive issue (such as education, religion, national defence,
minority rights etc.) is introduced for discussion by the government.
Several political, social and ethical issues are involved. Suppose finally
that the opposition strongly mobilises, considering that the law is a ma-
jor attack against "something". Massive demonstrations are organised,
an aggressive media campaign is pursued etc.. It is quite reasonable
that the government will try to find a compromise on some aspects of
the law in order to improve its "acceptability". Note, however, that
such a compromise concerns aspects argued by the minority and not the
majority.
Which decision rule is the government using to choose an appropriate
law proposal in such a situation? A law proposal x is considered "better"
than proposal y iff it meets the majority will and does not mobilise the
minority aversion. It should be observed that the minority is considered
here as an independent decision power source. Such a "decision rule"
is a regular practice in all mature democracies. Although the minority
does not have the power to impose its political will, it has the possibility
of expressing a "veto", at least occasionally. Such a "negative power"
may not necessarily be codified somewhere, but is accepted. Actually, it
is also a guarantee of the democratic game. When the present majority
becomes a minority it will be able to use the same "negative power".
Consider now the Security Council of the United Nations. Here, a
number of nations are officially endowed with a veto power such that
resolutions taken with a majority of votes (even the highest ones) can
be withdrawn if such a veto is used. We observe that in this case the
decision rule "x is better than y if it is the case for the majority and no
veto is used against x" is officially adopted. Again we observe that the
countries having a veto power do not have a "positive power" (impose a
decision), but only a "negative" one.
Finally, consider the very common situation where the faculty has to
deliberate on the admission of candidates to a course (let's say a manage-
ment course). Then consider two candidates: the first, x, having quite
good grades, systematically better than the second, y, but with a very
bad grade in management science; then candidate y, who is systemati-
cally worse than x, but has an excellent grade in management science.
Several faculty members will claim that, although candidate y is not
better than candidate x, it is also difficult to consider x better than y
due to their inverse quality concerning the key class of the course, man-
Modelling Positive and Negative Reasons in Decision Aiding 149
agement science. The same faculty members will also claim that the two
candidates cannot be considered indifferent because they are completely
different. These members are intuitively adopting the same decision rule
as in the previous two examples: candidate x is better than candidate y
iff (s}he has a majority of grades in (her}his favour and is not worse in
a number of key classes. For an extensive discussion on the question of
grades in decision support see Bouyssou et al., 2000.
If we consider a class grade of a candidate as (her)his value on a
criterion, the reader will observe that in the above decision rule there
exist criteria having a "negative power". Such a "negative power" is not
compensated by the "positive power" of the majority of criteria. It acts
independently and only in a negative sense.
We could continue with several other real life examples going from
vendor rating to bid selection and loan allowance. In all such cases it is
frequent to find the intuitive decision rule: alternative x is better than
alternative y iff there is a majority of "reasons" supporting x wrt to y
and there is no strong opposition to x wrt to y.
In order to be more formal we will use a large preference relation of
the type "x is at least as good as y" (denoted S (x, y), also known as
"outranking" relation) such that:
S(x, y) {::::::::? C(x, y) /\ -,D(x, y) (1)
where:
C(x, y) means there is a majority of reasons supporting x wrt to y;
D(x, y) means there is a strong opposition to x wrt to y;
/\ and -, being the conjunction and negation operators respectively.
C(x,y) (2)
D(x, y) (3)
where:
iff the score difference gj(Y) - gj(x) does not exceed a (indifference)
threshold qj. However, fixing a precise value for qj is not easy and
the concordance test (2) can be artificially sensitive to modifications of
criterion values, especially when the scale of criterion gj is continuous
(see Perny and Roy, 1992 and Perny, 1998 for a precise discussion on
this topic). A useful solution to overcome this difficulty was proposed
by B. Roy a long time ago (see Roy, 1978). The idea was to define
a concordance index Cj(x, y), valued in the unit interval, and defined
from the quantities gj(x) and gj(Y) for each criterion gj and each pair
(x,y). By convention, Cj(x,y) = 1 means that the criterion gj is fully
concordant with the assertion S (x, y) whereas Cj (x, y) = 0 means that
criterion gj is definitely not concordant with this assertion. There is,
of course, the possibility of considering intermediate values between 1
and 0 which makes the construction more expressive. It leaves room for
a continuum of intermediary situations between concordance and non-
concordance. As an example, we recall the definition of concordance
indices proposed by Roy in the Electre III method (Roy, 1978)
where:
o~--~----~------~~----------~------~
gj(Y)
(5)
The same ideas apply to the discordance test whose role is to check
whether some criteria are strongly conflicting with the proposition S(x, y).
The classical test, with the veto threshold, is not always convenient. This
is particularly true when the criterion scale is continuous; it does not
seem appropriate to declare that a given criterion gj should have a right
of veto over S(x, y) when gj{Y) - gj(x) > Vj{gj{x)) but should entirely
lose this right as soon as the inequality no longer holds. A continuous
transition seems preferable. For this reason, discordance indices mea-
suring the extent to which criterion j is strongly opposed to a statement
S(x, y) were introduced in (Roy, 1978).
Modelling Positive and Negative Reasons in Decision Aiding 155
where J~'Y = {j E {1, ... ,n},Dj(x,y) > 'Y} and 'Y,o E (0,1) are
the overall concordance and discordance thresholds respectively.
Note that the test is defined in such a way that the presence of at
least one fully discordant criterion 9j (such that Dj(x, y) = 1) is
sufficient to make the discordant test positive;
2 interpreting the discordance test in a multi-valued logic. This is
the option implicitly used in Electre III (Roy, 1978). This amounts
to defining the level to which the discordance test is fulfilled. Con-
sistent with the previous proposition the truth value d(x, y) E [0,1]
returned by the discordance test can be defined, for example, by:
When the concordance test is (5) and the discordance test is (7) the
construction of the outranking relation S is obviously defined by equa-
tion (1). When the concordance test is (6) and the discordance test is
(8) the equation (1) must be interpreted in a multivalued logic. This
leads to defining the overall outranking index s(x,y) E [0,1] for any pair
(x, y) of alternatives as a non-decreasing function of c(x, y) and a non-
increasing function of d(x, y). As an example, B. Roy uses the following
equality in Electre III:
The reader is referred to Perny and Roy, 1992 and Perny, 1998 for a
more general and systematic construction of outranking relations in the
framework of fuzzy set theory.
2.3. Problems
The use of the so-called outranking methods in MCDA is now largely
acknowledged and several empirical validations may be found in the
literature (see Roy and Bouyssou, 1993; Vincke, 1992b; Bouyssou et al.,
2000). It is nevertheless possible to note a number of significant open
questions.
but only two valuations are possible (either the outranking relation
holds or it does not).
• The definition of the overall outranking relation, at least as it
usually appears in outranking methods, implicitly imposes that
the criteria to be aggregated should at least be weak orders.
If the preference models of the criteria to be aggregated are Pseudo-
orders (preference structures allowing a numerical representation
using thresholds), there is no way to use such a specific informa-
tion in the establishment of the outranking relation. Only in the
case where the outranking relation is a fuzzy binary relation is it
possible to use the specific information included in pseudo orders
when these are represented as fuzzy relations themselves (see Roy,
1991). If the preference models of the criteria under aggregation
are partial orders then it is possible that the absence of preference
or indifference at the single criterion level could lead to a "non out-
ranking", not due to conflicting preferences, but due to ignorance.
There is however no way to distinguish such situations.
• As already mentioned by Vincke, 1982, each preference aggregation
step leads to a result which is (from a relational point of view)
poorer than the original information. This is obvious, since the
aggregation procedure eliminates some information. Moreover, as
already reported by Bouyssou, 1996, an outranking relation is not
necessarily a complete relation (not even a partial order). From
this point of view, there is a problem if such an approach has to be
used in presence of a hierarchy of criteria. If at each layer we keep
the result of the aggregation as it is and then we aggregate at the
next layer, we will very soon obtain an (almost) empty relation.
On the other hand, if at each layer, after aggregation, we transform
the outranking relation into a weak order (so that we can correctly
apply the aggregation procedure again), we introduce a bias in each
aggregation step the consequences of which are unknown. While
in usual situations of decision support the use of an exploitation
procedure can be discussed with the client, this is not possible
in a hierarchical aggregation problem and the above problem can
become severe.
The DDT Logic. The DDT logic, which is a four-valued first order
language, is based on a net distinction between the "negation" (which
Modelling Positive and Negative Reasons in Decision Aiding 159
and therefore :
u k
f
Figure 3. The continuous hi-lattice
x
y
• P(x, y)
x
y
• Q(x,y)
y
x
x
y
• I(x,y)
We are able to show that using the PC preference structure and the
positive / negative reasons approach we can model such situations of
hesitation due to the presence of an interval representation in a positive
way.
Theorem 2 (see Ngo The and Tsoukids, 2001) Definitions 1 and 2 are
equivalent.
-I=TIUTKUTK- l
{=::} (TS /\ TS- l ) V (TS /\ US-I) V (US /\ TS- l )
- i} I = Io = {(x,x),x E A};
- ii} \Ix, y, TS(x, y) V TS(y, x);
- iii} \Ix, y, z, FS(x, y) 1\ --,TS(y, z) => FS(x, z);
- ivY \Ix, y, z, US (x, y) 1\ (KS V FS)(y, z) => --,TS(x, z);
- v} \Ix, y, z, US(x, y) 1\ US(y, z) => US(x, z);
- vi} \lx,y,z, KS(x,y) 1\ FS(y,z) => FS(x,z);
- vii} \lx,y,z, KS(x,y) 1\ KS(y,z) => (KS V FS)(x,z);
- viii} \Ix, y, z, KS(x, y) 1\ US(y, z) => (KS V US)(x, z);
- ix} \Ix, y, z, US(x, y) 1\ US(z, y) => --,FS(x, z) 1\ --,FS(z, x);
where:
- J;;S is the set of criteria gj E G for which b.Sj(x, y) holds;
- J;;-'S is the set of criteria gj E G for which b.,Sj(x, y) holds;
- 11-1 and 11-2 represent two functions (11-1,11-2 : 2G I-t R "measuring" the
"strength" or "importance" of criteria coalitions;
- 'Y and 0 being thresholds representing a "security level" for the decision
maker.
For the same reasons as those mentioned in the preference modelling
section, it is worth considering a fuzzy extension of the aggregation
method proposed above. A natural proposition for this derives directly
from fuzzy concordance and discordance indices introduced in section
2.2. The most natural is:
b(S(x, y))
b(,S(x, y))
1 In the previous examples, for a given (x, y), b.S (b.,S) is evaluated
wrt the criteria for which b.Sj (b.,Sj) holds. It could be the
case that the criteria for which ,b.,Sj (,b.Sj) holds could also
be considered. It is necessary to study what the consequences of
such choices are, but a priori it is up to the will and intuition of
170 AIDING DECISIONS WITH MULTIPLE CRITERIA
What can we obtain as a final result? The two relations l::,.S and l::,.-,S
are just two binary relations for which nothing, except reflexivity, can
be imposed. What happens if a final result in the form of an ordering
relation is expected? As an example, in the following we will briefly
present three procedures (suggested in Tsoukias, 1997; the first also
studied and axiomatised in Greco et al., 1998).
1 Given the two binary relations l::,.S and l::,. -,S compute the following
score for each alternative:
4. Conclusions
Following the example of concordance and discordance concepts in-
troduced by B. Roy in multiple criteria aggregation methods, we have
presented some elements of a new and non-conventional approach to
preference modelling. The main originality of this approach is to con-
sider positive and negative arguments independently with respect to a
given assertion about preferences. Hence, the various possible combina-
tions of these two independent views on preferences provide a richer set
of situations than usual. This potential richness is represented by a lat-
tice of truth values from which a new multi-valued logic is constructed.
In the context of preference modelling, the increased expressive power
of the resulting language should be useful. Among others, it enables the
description of the possible belief states of a decision-maker facing a com-
plex situation due to imprecise evaluations or conflicting criteria. Due
172 AIDING DECISIONS WITH MULTIPLE CRITERIA
Acknowledgements
We would like to thank the referees for several useful comments. Ob-
viously we want to express our deep gratitude to Bernard Roy who really
inspired us with his pioneering work in decision aiding.
References
Belnap N.D., (1977), "A useful four-valued logic", in G. Epstein, J. Dumm, (eds.),
Modern uses of multiple valued logics, D. Reidel, Dordrecht, 8-37.
Bouyssou D., (1992a), "On some properties of outranking relations based on a concor-
dance - discordance principle", in A. Goicoechea, L. Duckstein and S. Zionts, (eds.),
Multiple criteria decision making, Springer-Verlag, Berlin, 93-106.
Bouyssou D., (1992b), "Ranking methods based on valued preference relations: A
characterization of the net flow network", European Journal of Operational Re-
search, vol. 60, 61-67.
Bouyssou D., (1996), "Outranking relations: do they have special properties?", Jour-
nal of Multi-Criteria Decision Analysis, vol. 5, 99-111.
Bouyssou D., Perny, P., (1992), "Ranking method for valued preference relations:
A characterization of a method based on leaving and entering flows", European
Journal of Operational Research, vol. 61, 186-194.
Bouyssou D., Pirlot M., (1997), "Choosing and ranking on the basis offuzzy preference
relations with the 'Min in Favor", in G. Fandel, T. Gal, (eds.), Multiple criteria
decision making, Springer Verlag, Berlin, 115-127.
Bouyssou D., Pirlot M., (1999), "Conjoint measurement without additivity and transi-
tivity", in N. Meskens, M. Roubens, (eds.), Advances in Decision Analysis, Kluwer
Academic, Dordrecht, 13-29.
Bouyssou D., Vincke Ph., (1997), "Ranking alternatives on the basis of preference
relations: a progress report with special emphasis on outranking relations" , Journal
of Multiple Criteria Decision Analysis, vol. 6, 77-85.
Bouyssou D., Pirlot M., Vincke Ph., (1997), "A general model of preference aggrega-
tion", in M.H. Karwan, J. Spronk, J. Wallenius, (eds.), Essays in decision making,
Springer Verlag, Berlin, 120-134.
Bouyssou D., Marchant Th., Perny P., Pirlot M., Tsoukias A., Vincke Ph., (2000),
Evaluation and Decision Models: a critical perspective, Kluwer Academic, Dor-
drecht.
Brans J.P., Vincke Ph., (1985), "A preference ranking organization method", Man-
agement Science, vol. 31, 647-656.
Dubois D., Prade H., (1988), Possibility Theory, Plenum Press, New York.
REFERENCES 173
Fargier H., Pemy P., (2001), "Modelisation des preferences par une regIe de con-
cordance generalisee", in A. Colomi, M. Parruccini, B. Roy (eds.), Selected Pa-
pers from 4gth and 5d h meetings of the EURO Working Group on MCDA, EUR-
Report, to appear.
Fitting M.C., (1991), "Bilattices and the semantics of Logic Programming", Journal
of Logic Programming, vol. 11,91-116.
Fortemps Ph., Slowinski R., (2001), "A graded quadrivalent logic for ordinal prefer-
ence modelling", Fuzzy Optimization and Decision Making, voU, no.l..
Grabisch M., Perny P., (2001), "Agregation Multicritere", in B. Bouchon-Meunier, C.
Marsala (eds.). Utilisation de la logique floue, Hermes, Paris, to appear.
Ginsberg M., (1988), "Multivalued logics: A uniform approach to reasoning in artificial
intelligence", Computational Intelligence, vol. 4, 265-316.
Greco S., Matarazzo B., Slowinski R., Tsoukias A., (1998), "Exploitation of a rough
approximation of the outranking relation in Multi-Criteria Choice and Ranking",
in T.J. Stewart and R.C. van der Honert (eds.), Trends in Multi-Criteria Decision
Making, Springer Verlag, LNEMS 465, Berlin, 45-60.
Greco S., Matarazzo B., Slowinski R., (2001), "Conjoint measurement and rough set
approach for multicriteria sorting problems in presence of ordinal criteria", in A.
Colorni, M. Paruccini, B. Roy, (eds.), AMCDA: Selected papers from the 4gth and
5d h meetings of the EURO MCDA Working Group, EUR-report, Ispra, 114 - 141.
Marchant Th., (1996), "Valued relations aggregation with the Borda method", Jour-
nal of Multiple Criteria Decision Analysis, vol. 5, 127-132.
Ngo The A., Tsoukias A., Vincke Ph., (2000), "A polynomial time algorithm to detect
PQI interval orders", International Transactions of Operations Research, vol. 7,
609-623.
Ngo The A., Tsoukias A., (2001), "A new axiomatisation of PQI interval orders",
submitted.
Ostanello A., (1985), "Outranking relations", in G. Fandel, J. Spronk, Multiple Cri-
teria Decision Methods and Applications, Springer Verlag, Berlin, 41-60.
Perny P., (1998), "Multicriteria filtering methods based on concordance and non-
discordance principles", Annals of Operations Research, vol. 80, 137-165.
Perny P., Roubens M., (1998), "Preference Modelling", in Handbook of Fuzzy Sets
and Possibility Theory, Operations Research and Statistics, R. Slowinski (ed.), D.
Dubois et H. Prade (series eds.), Kluwer Academic Publishers, 3-30.
Perny P., Roy B., (1992), "The use of fuzzy Outranking Relations in Preference Mod-
elling", in Fuzzy sets and Systems, vol. 49, 33-53.
Perny P., Tsoukias A., (1998), "On the continuous extension of a four valued logic
for preference modelling", in Proceedings of the IPMU 1998 conference, 302-309.
Pirlot M., (1995), "A characterization of 'min' as a procedure for exploiting valued
preference relations and related results", Journal of Multiple Criteria Decision
Analysis, vol. 4, 37-56.
Pirlot M., (1997), "A common framework for describing some outranking methods",
Journal of Multiple Criteria Decision Analysis, vol. 6, 86-92.
Roy B., (1968), "Classement et choix en presence de points de vue multiples (Le
methode ELECTRE)", Revue Fran~aise d'Informatique et de Recherche Operation-
nelle, vol. 8, 57-75.
174 AIDING DECISIONS WITH MULTIPLE CRITERIA
Roy B., (1978), ELECTRE III: Un algorithme de classement fonde sur une representation
Houe des preferences en presence de criteres multiples, Cahiers du CERO, vol. 20,
1,3-24.
Roy B., (1991), "The outranking approach and the foundations of ELECTRE meth-
ods", Theory and Decision, vol. 31, 49-73.
Roy B., Bouyssou D., (1993), Aide Multicritere d la Decision: Methodes et cas, Eco-
nomica, Paris.
Roy B., Mousseau V., (1996), "A theoretical framework for analysing the notion of
relative importance of criteria", Journal of Multi-criteria Decision Analysis, vol.
5,145-159.
Roy B., Skalka J.-M., (1984), "ELECTRE IS: Aspects methodologiques et guide
d'utilisation", Technical Report, Document du LAMSADE N°30, Universite Paris
Dauphine, Paris.
Roy B., Vincke Ph., (1984), "Relational systems of preferences with one or more
pseudo-criteria: some new concepts and results", Management Science, vol. 30,
1323-1335.
Roy B., Vincke Ph., (1987), "Pseudo-orders: definition, properties and numerical rep-
resentation", Mathematical Social Sciences, vol. 14, 263-274.
Tsoukias A., (1996), "A first-order, four valued, weakly paraconsistent logic", Cahier
du LAMSADE n. 139.
Tsoukias A., (1997), "Sur la generalisation des concepts de concordance et discor-
dance en aide multicritre a la decision", Memoire presente pour l'obtention de
l'habilitation a diriger des recherches, Universite Paris Dauphine, appeared also as
Document du LAMSADE. n. 117.
Tsoukias A., Vincke Ph., (1995), "A new axiomatic foundation of partial comparabil-
ity", Theory and Decision, vol. 39, 79-114.
Tsoukias A., Vincke Ph., (1997), "Extended preference structures in MCDA", in J.
CHmaco (ed.), Multi-criteria Analysis, Springer Verlag, Berlin, 37-50.
Tsoukias A., Vincke Ph., (1998), "Double Threshold Orders: a new axiomatisation",
Journal of Multiple Criteria Decision Analysis, vol. 7, 285-30l.
Tsoukias A., Vincke Ph., (2001), "A characterization of PQI interval orders", to ap-
pear in Discrete Applied Mathematics.
Vanderpooten, D., (1990), "The construction of prescriptions in outranking methods",
in Bana e Costa C.A., (ed.), Readings in Multiple Criteria Decision Aid, Springer-
Verlag, Berlin, 184-215.
Vincke Ph., (1982), "Aggregation of preferences: a review", European Journal of Op-
erational Research, vol. 9, 17-22.
Vincke Ph., (1988), "P,Q,I preference structures", in J. Kacprzyk, M. Roubens, eds.,
Non conventional preference relations in decision making, LNEMS 301, Springer
Verlag, Berlin, 72-8l.
Vincke Ph., (1992a), "Exploitation of a crisp relation in a ranking problem", Theory
and Decision, vol. 32, 221-240.
Vincke Ph., (1992b), Multicriteria Decision Aid, J. Wiley, New York.
Vincke Ph., (1999), "The Outranking Approach", in T. Gal, Th. Stewart, Th. Hanne,
(eds.), Multicriteria Decision Making,. Advances in MCDM Models, Algorithms,
Theory and Applications, Kluwer Academic, Dordrecht.
EXPLORING THE CONSEQUENCES
OF IMPRECISE INFORMATION IN
CHOICE PROBLEMS USING ELECTRE
Luis C. Dias
INESC and Faculty of Economics, University of Coimbra, Portugal
Idias@pombo.inescc.pt
J oao Climaco
INESC and Faculty of Economics, University of Coimbra, Portugal
jciimaco@pombo.inescc.pt
Abstract ELECTRE I, IS and its variants are well-known methods to help Deci-
sion Makers (DMs) choose one action (alternative) from a discrete set.
Here, we consider the case where the DMs have difficulties in fixing pre-
cise values for all the parameters required by ELECTRE, such as the
importance of the criteria, the veto thresholds or the performances of
the actions. We indicate how to obtain robust conclusions when the
DMs specify a set of multiple acceptable combinations of values for the
parameters. We argue that robust conclusions should be studied at the
outranking relation level, and then suggest some approaches to enrich
such conclusions (introducing a tolerance) and to exploit them.
1. Introduction
We owe Bernard Roy and his colleagues for the ELECTRE methodol-
ogy (see Roy, 1991; Roy and Bouyssou, 1993) and the concept ofrobust-
ness analysis in multicriteria decision aid (Roy, 1998; Roy and Bouys-
sou, 1993). This paper addresses robustness analysis in the context of
ELECTRE methods (ELECTRE I and its variants) for choosing one
action from a discrete set considering multiple evaluation criteria.
Using ELECTRE requires setting the value of its input parameters
(we use this word in a broad sense), such as the performance of the
176 AIDING DECISIONS WITH MULTIPLE CRITERIA
actions (alternatives), the importance and veto power of the criteria, etc.
Providing precise figures for all the parameters is often hard for Decision
Makers (DMs): some data may be missing, uncertain, contradictory or
imprecise, some modeling options are subject to arbitrariness, and some
parameters reflect the values of the DMs, which they may find difficult
to express and that may change over time (see also Roy, 1989; French,
1995).
To deal with these difficulties, we consider a setting where multi-
ple combination of values for the parameters are accepted, instead of a
precise one. This type of information is often named "imprecise" (e.g.
Miettinen and Salminen, 1999), "partial" (e.g. Charnetski and Soland,
1978) or "poor" (e.g. Bana e Costa and Vincke, 1995). The multiple
combinations of parameter values can be either enumerated or defined by
mathematical constraints (provided by the DMs or inferred from holistic
comparisons as in Mousseau, 1993).
Robustness analysis concerns the study of the results that are valid
for the multiple combinations of input values. We deem that this type
of analysis is useful from the very beginning of the decision process.
Firstly, it produces information (which conclusions are robust? which
results are more affected by the imprecision?) that may guide the DMs
in revising the information they provide, progressively narrowing the
range of acceptable values for the parameters. Secondly, it allows the
DMs to avoid the questions they consider difficult, or at least to postpone
these questions until they feel more familiar with the problem and more
confident about the answers. In group decisions, accepting mUltiple
combinations of parameter values may foster cooperation among the
DMs, as they may accept multiple opinions and may reach consensus
on the conclusions that are robust (see Dias and Climaco, 2000b, in the
context of sorting problems). This type of analysis may also be used
in the context of an interactive aggregation/ disaggregation approach to
decision aid (see Dias et al., 2000, in the context of sorting problems).
This paper discusses how to choose the most preferred action consid-
ering imprecise information, within the general context of ELECTRE
methods. Previous research concerning choice problems has focused on
the multiattribute value function model (for a recent review see Dias and
Climaco, 2000a), with few exceptions. For instance, Roy and Bouyssou
(1993) present a real-world study using ELECTRE IS where a robust-
ness analysis was conducted considering a discrete set of combinations
of input values; Miettinen and Salminen (1999) present an approach to
find the criteria weights (if any exist) that make each certain action the
best in terms of the min procedure (Pirlot, 1995).
Consequences of Imprecise Information 177
40 20 20 30 30
30 30 30 20 10
20 40 20 40 20
20 10 40 10 40
40 25 15 40 40
that the importance of the last two criteria has been fixed (k 4 = 0.1
and k5 = 0.2), hence there is imprecise information concerning only the
180 AIDING DECISIONS WITH MULTIPLE CRITERIA
f/! J/!
k=(.1,.4,.2,.1,.2) k=(.2,.4,. 1,. 1,.2) k=(.275,.275,.15,.1,.2)
N={tlj,a4} N={aJ,a4} N={a},a4}
k=(.4,.2,.1,.1,.2) k=(.4,.1,.2,.1,.2)
N={a},a4} N={a},aJ
Figure 1. Outranking relations (arrows denote outranking) and kernels for the
combinations in Ts
find that a2 always outranks a4, whereas a4 never outranks a2! This
paradox, resulting from the intransitivity of the outranking relation and
the rules for forming the kernel, forced us to ponder which kind of robust
conclusion should the analysis focus on. This example should not be dis-
missed on the grounds that it deals with a discrete set of combinations.
Indeed, a discrete T may correspond to a set of future scenarios or to a
group of DMs that decides to form T as the union of the combinations
of values that each of them deems worth considering. (NOTE: if we
tt
considered T instead of T s , " a4 E K " and " a2 K " would hold for
every combination exept (.15, .35, .2), which lays on the boundary of T,
whereas "a2Sa4" and "not a4Sa2" would always hold).
To our opinion, searching for binary robust conclusions concerning
the outranking relation fits well into the ELECTRE's philosophy of con-
fronting actions in pairs. Moreover, this has the advantage of being
more transparent to the DMs to the extent that the concept of outrank-
ing is easier to understand than the concept of kernel. It also avoids
the embarrassment of electing a winner (a4 in the previous example) in
a situation where there is another action (a2) that is preferable to it
(outranks it without being outranked) according to all the combinations
of values for the parameters.
Given an ordered pair of actions (ax, ay), from (1), considering r (ax, ay)
as a function of a combination of parameter values, we may test the ro-
bustness of conclusions concerning the outranking relation as follows:
These relations are more general and flexible than SR and N R : they
coincide with SR and N R when t = 0 and become richer (i.e. hold for a
larger number of pairs of actions) as t increases.
Table 3 presents the relative volumes of T where each outranking oc-
curs concerning Example 1. If the DMs regard as robust any conclusion
that holds for 95% of the acceptable combinations (setting t = 0.05),
then the following conclusions would be added to the list presented in
the previous section:
- a2 does not outrank al;
- a2 does not outrank a3.
The conclusion "al outranks a3" would also be accepted if the DMs
increased t up to 0.2. Of course, DMs should be given an idea of the
combinations in the region they are neglecting, which might contain the
values that they consider the most adequate.
The volume ofT may be analytically computed in the case of a polyhe-
dral T (e.g. Lasserre, 1983; Bueler et al., 1998). An alternative approach
is to compute approximate volumes using Monte-Carlo simulation (e.g.
Charnetski and Soland, 1978).
184 AIDING DECISIONS WITH MULTIPLE CRITERIA
The two procedures outlined above can also be used, since they do not
differentiate "unknown" from "contradictory" outrankings.
Notice that although we are interested in choice problems, all these
procedures provide a ranking of the actions.
6. Illustrative example
As an illustration, let us consider the choice of a machine to sort
packages, a problem faced by the French postal service (presentation
188 AIDING DECISIONS WITH MULTIPLE CRITERIA
Table 4. Performances (to be maximized) and thresholds for the example by Roy
and Bouyssou (the preference thresholds pj coincide with qj)
The criteria weights used in the original example are depicted in the
kj row of Table 4. These weights were chosen to satisfy the following
system, which reflects the opinion of the DMs:
(i) klO < k2 = k7 = ks < kl = k4 = k5 < k3 = k6 = kg = kn,
(ii) klO :::; k12 :::; kn,
(iii) kl = k2 + klO,
(iv) kl1 = kl + k2'
(v) kj ~ 0 (j = 1, ... , 12).
The original study set s = 0.7, although it admitted that s E [0.63, 0.73]
when performing a robustness analysis a posteriori. The outranking re-
lation from the original study is depicted in Figure 2. Based on this
relation, a5, a6, and as can obviously be excluded, while al justifies the
exclusion of a2, a3,and a4, which form an indifference class. Hence, the
kernel is K = {al,a7,ag}.
In our example, we will proceed by considering the imprecise infor-
mation defined by the constraints above and see what conclusions may
be drawn. We will use the same values of the original study for the
thresholds associated with the criteria (rows qj and Vj in Table 4). The
set T is defined by the constraints (i)-(v), the constraint L:/i=l
kj = 1
(which is not restrictive), and the bounds s E [0.63,0.73].
Let us now define r(.) to account for discordance as a ratio to the veto
thresholds, in order to attribute some meaning to inter-criteria compar-
isons of discordance:
Consequences of Imprecise Information 189
Figure 2. Outranking relation for the example by Roy and Bouyssou (a5 does not
appear since it is outranked by every other action).
as
- [-.08,.04] [-.12,.01] [.05,.18]. [.18,.31]
[-.28,.05] - [.13,.25] [-.03,.12] [.18,.31]
[-.16,.01] [.18,.31] - [.1,.25] [.08,.31]
[-.09,.09] [.01,.12] [-.03,.08] - [.27,.37]
[-4.34,-3.82] [-3.95,-3.57] [-4.29,-3.82] [-4.18,-3.71] -
[-1.21,-.8] [-1.10,-.75] [-1.14,-.83] [-.81,-.59] [.2,.33]
[-.09,.08] [-.24,.01] [-.48,-.17] [-.26,0] [.13,.25]
[-.23,-.07] [-.31,-.15] [-.38,-.19] [-.31,-.15] [-.09,.06]
[-.67,-.15] [-.31,-.14] [-.44,-.23] [-.38,-.19] [-.02,.13]
as
[-.05,.23] J-.09,.16J. 1~·23,-.14J 1~·38,-.26]
[-.12,.06] [-.69,-.45] [-.3,-.03] [-.11,.08]
[.01,.15] [-.60,-.23] [.01,.23] [-.17,.01]
[-.05,.16] [-.67,-.25] [-.33,-.22] [-.42,-.29]
[-2.81,-2.32] [-3.47,-2.98] [-3.90,-3.56] [-4.10,-3.79]
- [-.57,-.17] [-1.05,-.74] [-1.15,-.91]
[.18,.31] - [.01,.18] [-.48,-.24]
[-.2,-.1] [-.2,-.1] - [-.12,.06]
[-.54,-.40] [-1.20,-.56] [-.42,-.03] -
outranks any other (except a5) and is always outranked by a3 and a7.
Obviously, a5 and a6 are not contenders for the best action and can be
deleted.
Figure 3 represents the relation SR through thick arrows and the
complement of N R as segmented arrows. Hence, a thick arrow may
be read as "always outranks", a segmented arrow may be read as "may
outrank" , and the absence of an arrow indicates "never outranks". These
relations could be exploited by any of the techniques from the previous
section. As an example, the exploitation of the relation SR according to
the rules ofELECTRE IllS leads to the kernel K = {al,a7,ag}, which
is equal to the original study's.
To continue this example, let us suppose that the DMs were invited to
think about the doubtful outrankings and they would answer that they
were expecting that al would outrank a7. At this point, they could learn
that the combination yielding the minimum r( aI, a7) was kl = k4 = k5 =
0.08, k2 = k7 = kg = klO = 0.04, k3 = k6 = kg = kn = k12 = 0.12,
and s = 0.73. Analyzing this information, suppose that the DMs would
state that k2 should not be less than k12' which corresponds to a new
constraint on T (in this case, if they would state that al S a7, then
it could also be coded as a linear constraint). The ranges for the r(.)
functions would become smaller, as shown in Table 6.
By now it is clear that al S a7, although the DMs did not require
it directly. Let us also suppose that the DMs would accept a Type 2
relaxation with a tolerance to = 0.03. Figure 4 depicts the conclusions
corresponding to SZ (0.03) and N Z (0.03). The decision process could
then continue, either asking the DMs for more information (e.g. is al
preferred to a4 or indifferent to it?), or exploiting the relations obtained.
Action al would have the highest net-flow score and would appear at the
top of ELECTRE II's exploitation ranking. It would also belong to the
as ag
- [-.04,.04] [-.09,.01] [.05,.17] [.03,.16] [-.19,-.14] [-.34,-.26]
[-.28,.02] - [.13,.24] [-.03,.10] [-.69,-.47] [-.17,-.03] [-.06,.08]
[-.16,-.02] [.18,.30] - [.1,.24] [-.60,-.25] [.07,.23] [-.13,.01]
[-.09,.07] [.06,.12] [.01,.08] - [-.58,-.25] [-.29,-.22] [-.39,-.29]
[-.09,.06] [-.18,.01] [-.43,-.17] [-.26,-.02] - [.07,.18] [-.42,-.24]
[-.23,-.10] [-.31,-.17] [-.38,-.20] [-.31,-.17] [-.20,-.11] - [-.12,.04]
[-.67,-.20] [-.31,-.16] [-.44,-.25] [-.38,-.20] [-1.20,-.61] [-.42,-.06] -
Figure 4. Relations "always outranks" and "may outrank" after additional con-
straint and accounting for a tolerance of 0.03.
7. Concluding remarks
Instead of bulldozing the difficulties and hesitation of the DMs, through
a quest for the right combination of values for the parameters, we deem
that imprecision should be accepted from the very beginning of the de-
cision aid process. This allows to alleviate the DMs' cognitive burden at
the beginning, postponing the most difficult questions to a stage when
192 AIDING DECISIONS WITH MULTIPLE CRITERIA
they are more familiar with the problem at hand and the decision aid
method.
The analysis proposed here explores the consequences of the impre-
cise information that the DMs are able (or willing) to provide. We focus
on exploration rather than aggregation, avoiding the computation of
averages, median values, and other usual aggregation means. The ex-
ploration allows to discover which conclusions are robust and allows to
identify which conclusions are more affected by the imprecision. This is
particularly important in what regards the questions that can be posed
to the DMs when more information is needed.
This paper does not propose a precise method. Rather, we feel that
the array of tools to be used will depend on the problem at hand. A
decision support system implementing several of these tools would hence
be quite helpful. Most of the approaches proposed here have the pos-
sible drawback of demanding some computational effort (optimization
or volume computation). However, the DMs will not perceive this for
two reasons. On the one hand, today's low cost personal computers are
sufficiently powerful to solve these problems in acceptable time for the
problem dimensions usually found in practice. On the other hand, the
DMs need only understand the results, and not the algorithmic details
of their computation. The concept of a robust conclusion such as never
outranks or always outranks is, of course, easy to apprehend. We might
say that this is a case of using "hard" tools for a "softer" decision aid.
References
Bana e Costa. C.A. and Ph. Vincke (1995). Measuring credibility of compensatory
preference statements when trade-offs are interval determined, Theory and Deci-
sion 39, 127-155.
Bouyssou, D. (1992), Ranking methods based on valued preference relations: a char-
acterization of the net-flow method, European Journal of Operational Research
60,61-68.
Biieler, B., A. Enge and K. Fukuda (1998), Exact volume computation for polytopes:
a practical study, in Polytopes - Combinatorics and Computation, DMV-Seminars,
Birkhauser Verlag (to appear).
Charnetski J.R. and R.M. Soland (1978). Multiple-attribute decision making with
partial information: the comparative hypervolume criterion. Naval Research Lo-
gistics 25, 278-288.
Dias, L. C. and J. N. Clfmaco (1999), On computing ELECTRE's credibility indices
under partial information, Journal of Multi-Criteria Decision Analysis 8, 74-92.
Dias, L. C. and J. N. Clfmaco (2000a), Additive aggregation with variable interdepen-
dent parameters: the VIP Analysis software, Journal of the Operational Research
Society 51, 1070-1082.
Dias, L. C. and J. N. Clfmaco (2000b), ELECTRE TRI for Groups with Imprecise
Information on Parameter Values, Group Decision and Negotiation 9, 355-377.
REFERENCES 193
Dias, L., V. Mousseau, J. Figueira and J. CHmaco (2000), An Aggregation/Disaggregattion
approach to obtain robust conclusions with ELECTRE TRI, Cahier du LAM-
SADE, No. 174, Universite Paris-Dauphine. (To appear in EJOR).
French, S. (1995), Uncertainty and imprecision: modelling and analysis, Journal of
the Operational Research Society 46, 70-79.
Greco, S., B. Matarazzo, R. Slowinski amd A. Tsoukias (1997), Exploitation of a
rough approximation of the outranking relation, Cahier du LAMSADE, No. 152,
Universite Paris-Dauphine.
Lasserre, J.B. (1983), An analytical expression and an algorithm for the volume of
a convex polyhedron in RR, Journal of Optimization Theory and Applications 39,
363-377.
Miettinen, K. and P. Salminen (1999), Decision-aid for discrete multiple criteria de-
cision making problems with imprecise data, European Journal of Operational
Research 119, 50-60.
Mousseau, V. (1993), Problmes lies a l'evaluation de l'importance relative des criteres
en aide multicritere a la decision: refiexions tMoriques, experimentation et imple-
mentation informatique, PhD Thesis, Universite Paris-Dauphine.
Pirlot, M. (1995), A characterization of 'Min' as a procedure for exploiting valued
preference relations and related results, Journal of Multi-Criteria Decision Analysis
4,37-56.
Roy, B. (1968), Classement et choix en presence de points de vue multiples (la methode
ELECTRE), Revue Informatique et Recherche Oprationelle, 2e. Anne, No.8, 57-
75.
Roy, B. (1989), Main sources of inaccurate determination, uncertainty and imprecision
in decision models, Mathematical and Computer Modelling 12, 1245-1254.
Roy, B. (1991), The outranking approach and the foundations ofELECTRE methods,
Theory and Decision 31, 49-73.
Roy, B. (1998), A missing link in OR-DA: robustness analysis, Foundations of Com-
puting and Decision Sciences 23, 141-160.
Roy B. and Bouyssou, D. (1993), Aide multicritere a la decision: methodes et cas,
Economica, Paris.
Roy, B. e J. M. Skalka (1984), ELECTRE IS: aspects methodologiques et guide
d'utilisation, Document du LAMSADE, No. 30, Universite Paris-Dauphine.
Saltelli, A., S. Tarantola and K. Chan (1999), A role for sensitivity analysis in present-
ing the results from MCDA studies to decision makers, Journal of Multi-Criteria
Decision Analysis 8, 139-145.
Solymosi, T and J. Dombi (1986), A method for determining the weights of the
criteria: the centralized weights, European Journal of Operational Research 26,
35-41.
Tsoukias, A. and Ph. Vincke (1997), Extended preference structures in multicriteria
decision aid, in Climaco, J. (ed.), Multicriteria analysis, Springer, 37-50.
Vincke, Ph. (1999), Robust solutions and methods in decision aid, Journal of Multi-
Criteria Decision Analysis 8, 181-187.
MODELLING IN DECISION AIDING
Daniel Vanderpooten
LAMSADE - UniversiU Paris Dauphine, France
vdp@lamsade.dauphine.fr
Abstract This paper focusses on the central role of modelling in decision aid-
ing. All stages of the modelling process require choices which cannot
be totally rationalized. We believe, however, that adopting a certain
perspective (accepting ambiguity, favoring flexibility, ... ) in relation to
some specificities of the decision context may prove helpful to guide the
modelling process and to motivate some technical choices.
1. Introduction
The most remarkable characteristic of the contributions of Bernard
Roy is that they address all aspects of decision aiding, including:
• theoretical achievements,
• development of methods,
• epistemological reflection.
3. Definition of a model
The activity of decision aiding, when claiming to be founded on sci-
entific bases, requires an abstraction of the decision situation under the
form of models.
As defined by Roy, 1985 'a model is a schema which, for a certain
family of questions, is considered as a representation of a class of phe-
nomena that an observer has more or less carefully removed from their
environment to help in an investigation and to facilitate communication'.
A model is thus necessarily a partial representation of the situation to
be studied which emphasizes aspects that are judged to be relevant. As
outlined by Roy, 1999, 'a model is more a caricature ofreal-life situations
than an impoverished or approximate photograph of it'.
In the widest possible sense, we define a mathematical model as a
formal representation of a set of hypotheses which are judged to be
relevant in order to shed light on the situation under study.
The perception of the role of models in decision aiding, and correla-
tively the way of elaborating them, are closely related to the attitude (see
section 2). Therefore, depending on the underlying attitude, a model
may appear as:
1 a formal approximation of the reality under study (descriptive
model)
2 a formal translation of norms of rationality (normative model)
3 a formal framework for reasoning (constructive model)
Moreover, we shall introduce the following distinction between closed
and open models.
Modelling in Decision Aiding 199
This will then favor the selection of a sorting problem formulation which
focuses on the intrinsic value of alternatives. Closed models are also less
appropriate in an evolving context.
• Second level preference models are used when multiple criteria, de-
fined at the first level, must be aggregated in order to model overall
preferences, taking into account relative importance of criteria.
Optimization and simulation models include within the same entity
a model of alternatives and a first-level preference model. Unlike an
optimization model which is a closed model and involves a unique crite-
rion in its preference (sub}model, a simulation model is an open model
which may involve several criteria. Multicriteria models gather within
the same entity a model of alternatives, a first-level preference model (in-
cluding several criteria) and, possibly, a second-level preference model.
Indeed, while some multicriteria models explicitly include an aggregation
mechanism to represent overall preferences, other multicriteria models
replace this second·level preference model with an interactive process
which aims at exploring the model of alternatives using the first-level
preference model (see also section 5.2).
Observe that optimization models as well as multicriteria models in-
cluding first and second level preference (sub}models are closed models.
In the second case, however, explicitly expressing two levels of preference
modelling shows a lesser degree of closure and, above all, the possibil-
ity of reopening the model. We shall refer to this type of models as
semi-closed models.
When using a closed or semi-closed model, it is important to strengthen
the results through a sensitivity analysis or, more generally, a robustness
analysis (see section 5.1). As shown in section 5.2, interactivity provides
an efficient way to take advantage of an open model.
6. Conclusions
Building a model, as we have discussed in this article, refers essen-
tially to a constructivist attitude. Indeed, the modelling process includes
a series of modelling choices which cannot be fully rationalized and may
even appear sometimes as an act of faith. We believe, however, that
adopting a certain perspective (which accepts ambiguity, favors flexibil-
ity, ... ) in relation to some specificities of the decision context may prove
helpful to guide the modelling process and to consolidate some technical
choices.
In this paper we mainly focused on aspects related to the technical
validity of a model. However, a model also plays a key role in commu-
nication. Even if a model is technically valid, it must also be acceptable
REFERENCES 207
to the actors. This refers to what is called 'model legitimisation' by
Landryet al., 1996. These authors give some general guidelines to favor
the legitimacy of a model, such as working in close cooperation with
the strategic stakeholders, striking a balance between the level of model
sophistication and the competence level of the stakeholders,... We also
believe that specific modelling choices may influence the model legiti-
macy depending on the decision situation. For instance, when several
decision makers with different value systems are involved in the deci-
sion process, an acceptable representation must include the viewpoints
of each actor. A multicriteria model, where each actor sees his/her view-
point represented through at least one criterion, is already a good basis
for negotiation and discussion. Another example is in the not so un-
common situation where decision makers are willing to be supported by
a model but do not accept that the model dictates a prescription. It
is then extremely important to design an open model which leaves the
possibility to explore various solutions.
Acknowledgments
The author would like to thank John Buchanan and Lorraine Gardiner
for their helpful comments on an earlier version of this paper.
References
Azibi, R. and Vanderpooten, D. (2001). Elaboration de criteres agregeant des conse-
quences dispersees : deux exemples concrets. In Colorni, A., Paruccini, M., and
Roy, B., editors, AMCDA - Aide MultiCritre La Dcision Multi Criteria Decision
Aid, pages 13~30. EUR Report 19808.
Bana e Costa, C. (1996). Les problmatiques de l'aide la dcision : vers l'enrichissement
de la trilogie choix-tri-rangement. RAIRO - RO, 30(2):191~216.
Bell, D., Raiffa, H., and Tversky, A. (1988). Descriptive, normative and prescriptive
interactions in decision making. In Bell, D., Raiffa, H., and Tversky, A., editors,
Decision making: descriptive, normative and prescriptive interactions, pages 9~30.
Cambridge University Press.
Bonczek, R., Holsapple, C., and Whinston, A. (1981). Foundations of Decision Support
Systems. Academic Press, Orlando.
Bouyssou, D. (1989). Modelling inaccurate determination, uncertainty, imprecision
using multiple criteria. In Lockett, A. and Islei, G., editors, Improving Decision
Making in Organisations, LNEMS 335, pages 78~87. Springer-Verlag, Berlin.
Bouyssou, D. (1990). Building criteria: A prerequisite for MCDA. In Bana e Costa, C.,
editor, Readings in Multiple Criteria Decision Aid, pages 58~80. Springer-Verlag,
Berlin.
Bouyssou, D., Perny, P., Pirlot, M., Tsoukias, A., and Vincke, P. (1993). A Manifesto
for the new MCDA era. Journal of Multi Criteria Decision Analysis, 2:125~127.
Bouyssou, D. and Roy, B. (1987). La notion de seuils de discrimination en analyse
multicritere. INFOR, 25:302~313.
208 AIDING DECISIONS WITH MULTIPLE CRITERIA
Chalmers, A. (1982). What is this thing called Science? An assessment of the nature
and status of science and its method. University of Queensland Press, St Lucia.
traduction fran<;aise, Editions La Decouverte, Paris, 1987.
D'Avignon, G. (1992). Demarche d'aide la preparation d'un plan directeur: Le cas
des palais de justice. Documents du LAMSADE no. 73, Universite Paris-Dauphine,
France.
Dubois, D. and Prade, H. (1985). Theorie des possibilites - Applications d la represen-
tation des connaissances en informatique. Masson, Paris.
Fischer, M. (1985). Interactive optimization. Annals of Operations Research, 5:541-
556.
Fishburn, P. (1970). Utility theory for decision making. Wiley, New-York.
Fodor, J. and Roubens, M., editors (1994). Fuzzy preference modelling and multicri-
teria decision support. Kluwer, Dordrecht.
Gardiner, L. and Vanderpooten, D. (1997). Interactive multiple criteria procedures:
Some reflections. In Climaco, J., editor, Multicriteria Analysis, pages 290-301.
Proc. of the XIth International Conference on MCDM, Coimbra, Portugal, Springer
Verlag, Berlin.
Grassin, N. (1986). Constructing population criteria for the comparison of different
options for a high voltage line route. European Journal of Operational Research,
26:42-57.
Greco, S., Matarazzo, B., and Slowinski, R. (1999). The use of rough sets and fuzzy
sets in MCDM. In Advances in MCDM models, Algorithms, Theory, and Applica-
tions, pages 14.1-14.59. Kluwer Academic Publishers, Boston.
Grzymala-Busse, J. (1991). Managing uncertainty in expert systems. Kluwer Academic
Publishers, Boston.
Jacquet-Lagreze, E. (1973). Le probleme de l'agregation des preferences: une classe
de procedures a seuils. Mathmatiques et Sciences Humaines, 43:29-37.
Keeney, R (1988). Structuring objectives for problems of public interest. Operations
Research, 36:396-405.
Keeney, R (1992). Value-focused thinking: a path to creative decision making. Harvard
University Press.
Keeney, Rand Raiffa, H. (1976). Decisions with multiple objectives - preferences and
value trade-offs. Wiley and Sons.
Kouvelis, P. and Yu, G. (1997). Robust discrete optimization and its applications.
Kluwer.
Kruse, R, Schwecke, E., and Heinsohn, J., editors (1991). Uncertainty and vagueness
in knowledge based systems. Springer-Verlag, Berlin.
Landry, M., Banville, C., and Oral, M. (1996). Model legitimisation in operational
research. European Journal of Operational Research, 92(3}:443-457.
Levine, P. and Pomerol, J. (1989). Systemes interactifs d'aide d la decision et systemes
experts. Hermes, Paris.
Martel, J.-M., Guitouni, A., Abi-Zeid, I., Blanger, M., Chraygane, Z., and Zaras,
K. (1998). Construction d'une famille de critres dans Ie cadre du projet CESA.
Documents du LAMSADE no. 106, Universite Paris-Dauphine, France.
Norese, M. and Ostanello, A. (1989). Identification and development of alternatives:
introduction to the recognition of process typologies. In Lockett, A. and Islei, G.,
editors, Improving Decision Making in Organisations, LNEMS 335, pages 112-123.
Springer-Verlag, Berlin.
REFERENCES 209
Pawlak, Z. (1982). Rough sets. Int. J. Computer and Information Sci., 11:341-356.
Perny, P. and Roy, B. (1992). The use of fuzzy outranking relations in preference
modelling. Fuzzy Sets and Systems, 49:33-53.
Perny, P. and Vanderpooten, D. (1998). An interactive multiobjective procedure for
selecting medium-term countermeasures after nuclear accidents. Journal of Multi
Criteria Decision Analysis, 7(1):48-60.
Pollack, M. (1976). Interactive models in Operations Research - an introduction and
some future research directions. Computers and Operations Research, 3(4):305-312.
Roy, B. (1976). From optimization to multicriteria decision aid: Three main oper-
ational attitudes. In Thiriez, H. and Zionts, S., editors, Multiple Criteria Deci-
sion Making, Proceedings Jouy-en-Josas France 1975, pages 1-32. LNEMS 130,
Springer-Verlag, Berlin.
Roy, B. (1981). The optimisation problem formulation: criticism and overstepping.
Journal of the Operational Research Society, 32(6):427-436.
Roy, B. (1985). Methodologie Multicritere d'Aide d la Decision: Methodes et Cas.
Economica, Paris. english translation: Multicriteria Methodology for Decision Aid-
ing,Kluwer Academic Publishers, 1996.
Roy, B. (1987). Meaning and validity of interactive procedures as tools for decision
making. European Journal of Operational Research, 31(3):297-303.
Roy, B. (1989). Main sources of inaccurate determination, uncertainty and impreci-
sion. Mathematical and Computer Modelling, 12(10/11):1245-1254.
Roy, B. (1993). Decision science or decision-aid science? European Journal of Opera-
tional Research, 66(2):184-203.
Roy, B. (1998). A missing link in OR-DA: Robustness analysis. Foundations of Com-
puting and Decision Sciences, 23(3):141-160.
Roy, B. (1999). Decision-aiding today: what should we expect? In Gal, T., S. T. and
Hanne, T., editors, Multicriteria Decision Making, Advances in MCDM models,
Algorithms, Theory and Applications, pages 1.1-1.35. Kluwer Academic Publishers,
Boston.
Roy, B. and Bouyssou, D. (1993). Aide Multicritere d la Decision: Methodes et Cas.
Economica, Paris.
Roy, B. and Mousseau, V. (1996). A theoretical framework for analysing the notion
of relative importance of criteria. Journal of Multi Criteria Decision Analysis,
5(2):145-159.
Roy, B. and Slowinski, R. (1993). Criterion of distance between technical programming
and socio-economic priority. RAIRO - RO, 27:45-60.
Roy, B. and Vanderpooten, D. (1996). The European school of MCDA: Emergence,
basic features and current works. Journal of Multi Criteria Decision Analysis,
5(1):22-37.
Sims, R. (1997). Interactivity: a forgotten art? Computers in Human Behavior, 13(2):
157-180.
Slowinski, R. and Teghem, J., editors (1990). Stochastic versus Fuzzy approaches to
Multiobjective Mathematical Programming under uncertainty. Kluwer, Dordrecht.
Sprague Jr, R. and Carlson, E. (1982). Building effective Decision Support Systems.
Academic Press, Orlando.
Tsoukias, A. (1996). A first-order, four valued, weakly paraconsistent logic. Cahier
du LAMSADE no. 139, Universite de Paris Dauphine, France.
210 AIDING DECISIONS WITH MULTIPLE CRITERIA
Michael Doumpos
Technical University of Crete,Greece
dmichael@ergasya.tuc.gr
Constantin Zopounidis
Technical University of Crete, Greece
kostas@ergasya.tuc.gr
1. Introduction
The classification problem involves the assignment of a set of
alternatives (objects, alternatives) described over some attributes or criteria
into predefined homogeneous classes. Such problems are often encountered
in many fields, including finance, marketing, agricultural management,
human recourses management, environmental management, medicine,
pattern recognition, etc. This major practical interest of the classification
problem has motivated researchers in developing an arsenal of methods in
order to develop quantitative models for classification purposes. The linear
discriminant analysis (LDA) was the first method developed to address the
classification problem from a multidimensional perspective. LDA has been
used for decades as the main classification technique and it is still being used
at least as a reference point for comparing the performance of new
techniques that are developed. Other widely used parametric classification
techniques developed to overcome LDA's restrictive assumptions
(multivariate normality, equality of dispersion matrices between groups)
include quadratic discriminant analysis (QDA), logit and probit analysis and
the linear probability model (for two-group classification problems).
During the last two decades several alternative non-parametric
classification techniques have been developed, including mathematical
programming techniques (Freed and Glover, 1981), multicriteria decision aid
methods (Roy and Moscarola, 1977; Doumpos et aI., 2000), neural networks
(Patuwo et aI., 1993), machine learning approaches (Quinlan, 1986), and
rough sets (Pawlak, 1982; Pawlak and Slowinski, 1994). Among these
techniques, multicriteria decision aid (MCDA) has several distinguishing
and attractive features, involving, mainly, its decision support orientation.
Within the context of the MCDA theory, the preference disaggregation
approach (PDA) provides a sound methodological framework for developing
decision-making models both for ranking and classification (sorting
purposes). PDA (Jacquet-Lagreze and Siskos, 2001; Pardalos et aI., 1995)
refers to the analysis (disaggregation) of the global preferences (judgement
policy) of the decision maker in order to identify the criteria aggregation
model that underlies the preference result (ranking, classification). PDA uses
common utility decomposition forms to represent the decision maker's
preferences. The major distinguishing characteristic of PDA as opposed to
other MCDA approaches (multiattribute utility theory, outranking relations)
is that regression-based techniques are used for model development (indirect
estimation procedure) instead of direct interrogation techniques that are
employed in other MCDA approaches. While several case studies have
demonstrated the efficiency of methods based on PDA, in supporting
On the Use ofMulticriteria Classification Methods 213
where U (a) is the global utility of an alternative aEA, and uj[g.{a)] is the
marginal utility of an alternative aEA on the evaluation criterion gj.
To classify the alternatives into their original classes, it is necessary to
estimate utility thresholds UI, U2, ... , Uq_1 (threshold Uk distinguishes the
classes Ck and Ck+), "d k~q-l). Comparing the global utilities of an
alternative a with the utility thresholds, the classification of this alternative is
achieved in the following way:
U(a)~uI ~aECI
u2~U(a)<uI ~aEC2
(1)
U(a) < u q _ 1
The estimation of the global utility model (additive utility function) and
the utility thresholds is accomplished through linear programming
techniques, and more specifically, through the solution of the following
linear program:
MinimiseF= La+(a)+ ... + L[a+(a)+a-(a)]+ ... + La-(a) (2)
aeq aECk aeCq
s.t.
"daECk (4)
LuJg;(a)]-u k +a+(a)~O
;=1
(6)
On the Use ofMulticriteria Classification Methods 215
- aj is the number of subintervals [g( ,gt I ] into which the range of values
of criterion gi is divided,
- u\a) and (J-(a) are the classification errors (over-estimation and under-
estimation errors respectively). These errors indicate differences between
the classification of the alternatives performed by the developed additive
utility model and the classification specified by the decision maker. These
errors can be due to three reasons: (1) inability of the additive utility
model to fully represent the decision maker's preferences, (2) lack of data
or inappropriate data, (3) misjudgement by the decision maker during the
specification of the classification of the alternatives.
The objective function (2) of the above linear program minimises the
sum of all classification errors (J+ and (J- for all alternatives in the reference
set. This is achieved subject to the constraints (3)-(8). Constraints (3)-(5)
result from the classification rules (1) and they are used to define the
classification errors. In particular, constraint (3) implies that the global
utility of an alternative aECI should be greater or equal to the utility
threshold U\ which is the lower bound of class C\. If this is not possible, then
an amount of utility equal to (J+(a) should be added to the global utility of
this alternative, indicating that the alternative is classified to a lower class
than the one it actually belongs. Constraint set (4) is used for alternatives
which are classified by the decision maker in an intermediate class Ck • Since
Uk-\ is the lower bound of the class Ck-\ and Uk is the lower bound of class Ck,
the correct classification of an alternative a which belongs in class Ck can be
achieved only if the global utility of the alternative is strictly lower than the
utility threshold Uk-\ [by definition, according to the classification rule (1) if
U(a)=uk-\ then aECk] and greater or equal to the utility threshold Uk. In order
to ensure the strict inequality between U(a) and Uk-\ a user-provided small
positive constant £5 is used. If either of the above two conditions is not
satisfied then the corresponding amount of utility u+(a) or (J-(a) should be
216 AIDING DECISIONS WITH MULTIPLE CRITERIA
3. Experimental design
3.1 Methods
In order to perform a thorough examination of the classification
performance of the UTADIS method as opposed to traditional classification
procedures, an extensive experimental study is conducted using several
different data conditions.
Except for the UTADIS method, three other statistical and econometric
procedures are considered: linear discriminant analysis (LDA), quadratic
discriminant analysis (QDA) , and logit analysis (LA). These methods are
widely used in many scientific fields for the development of classification
models. Furthermore, they are also used as a benchmark in many
comparative studies investigating the classification performance of non-
parametric techniques such as mathematical programming, neural networks,
machine learning algorithms, etc. These are the two main reasons for
On the Use ofMulticriteria Classification Methods 217
alternatives using a given sample (reference set). On the other hand, using
the preference disaggregation framework employed in the UT ADIS method,
the aim is to analyse a given sample of decision instances (reference set) that
incorporate all the preferential information that characterizes the decision
maker's preferences and judgment. This enables the elicitation of the
necessary preferential information in an indirect way and the development of
classification models that are in accordance with the decision maker's
preferences. Such an approach supports the decision maker in understanding
the peculiarities of the considered alternatives, identifying and correcting the
possible inconsistencies in his/her judgments, thus improving the decision
making process. Of course, the use of the classification error as the
optimisation criterion in the linear programming formulation of the UTADIS
method is not necessarily in accordance with this objective. However, as
mentioned above, the use of linear programming enables the analyst to
incorporate specific constraints in the model development process in order to
elicit the decision maker's preferences as accurately as possible.
All the above functional, theoretical and practical (implementation)
differences between the methods and the associated advantages and
disadvantages of the methods are investigated in this experimental study in
terms of their impact to the classification performance (accuracy) of the
methods).
3.2 Factors
The comparison of the methods outlined in the previous subsection is
performed considering six factors regarding the properties of the data that
are used during model development and testing. A complete description of
the factors considered in this experimental design is presented in Table 1.
The first of the factors involving the properties of the data investigated in
the conducted experimental design involves their distributional form (F2).
While many studies conducting similar experiments have been concentrated
on univariate distributions to consider multivariate non-normality, in this
study multivariate distributions are considered. This specification enables
the investigation of additional factors in the experiment, such as the
correlation of the criteria and the homogeneity of the group dispersion
matrices. Actually, the use of a univariate distribution implies that the
criteria are independent, a case that is hardly the situation encountered in
real-world problems. The first two of the multivariate distributions that are
considered (normal and uniform) are symmetric, while the exponential (this
is actually a multivariate distribution that resembles the exponential
distribution in terms of its skewness and kurtosis) and log-normal
distributions are asymmetric, thus leading to a significant violation of
220 AIDING DECISIONS WITH MULTIPLE CRITERIA
The specified correlation coefficients for every pair of criteria define the
off-diagonal elements of the dispersion matrices of the groups, The elements
in the diagonal of the dispersion matrices, representing the variance of the
criteria are specified by the sixth factor, which is considered in two levels, In
the first level, the variances of the criteria are equal for all groups, whereas
in the second level the variances differ, Denoting the variance of criterion g;
for group y as O't, the realization of these two situations regarding the
homogeneity of the group dispersion matrices is performed as follows:
- For the multivariate normal, uniform and exponential distributions:
Levell: O';~ =0'/2 =0';; =1, V i=l, 2, ",,5,
Level 2: O'~ =1, O'j; =4, (jj; =16, V i=l, 2, .. ,,5,
- For the multivariate log-normal distribution, the variances are specified so
as to assure that the kurtosis of the data ranges within reasonable levels),
as follows:
a) In the case of two groups:
12' if m =36
Levell: O'j~ = O'j; ={14, if m =72 ,V i=l, 2, .. ,,5,
16, if m =108
12' if m = 36
Level 2: O';~ = 14, if m =72 , 0';22
{ =1.50';~, V i=l, 2, .. ,,5,
16,if m =108
b) In the case of three groups
4' if m =36
Levell: O';~ = 0';22 =0';; = 7, if m =72 ,V i=l, 2, .. ,,5,
{
10, if m =108
2'if m =36
Level 2: O'j~ = 4, if m =72 ,O'j;
{ =1.50'j~' O'j~ =1.50'j;, V i=1,2, .. ,,5,
6,if m =108
1 In the log-nonnal distribution the skewness and kurtosis are defined by the mean and the
variance of the criteria for each group. The procedures for generating multivariate non-
nonnal data can replicate satisfactory the prespecified values of the first three moments
(mean, standard deviation and skewness) ofa statistical distributions. However, the error is
higher for the fourth moment (kurtosis). Therefore, in order to reduce this error and
consequently to have better control of the generated data, both the mean and the variance of
the criteria for each group in the case of the multivariate log-normal distribution, are
specified so as the coefficient of kurtosis is lower than 40.
222 AIDING DECISIONS WITH MULTIPLE CRITERIA
The final factor defines the degree of group overlap. The higher the
overlapping is between each pair of groups, the higher is the complexity of
the classification problem due to the difficulty in discriminating between the
alternatives of each group. The degree of group overlap in this experimental
design is considered using the Hotelling r statistic. This is a multivariate
measure of difference between the means of two groups, assuming that the
criteria are multivariate normal and that the group dispersion matrices are
equal. Studies conducted on the first of these assumptions (multivariate
normality) have shown that actually the Hotelling r is quite robust to
departures from multivariate normality even for small samples (Mardia,
r
1975). Therefore, using the Hotelling in the non-multivariate distributions
considered in this experimental design does not pose a significant problem.
To overcome the second assumption regarding the homogeneity of the group
dispersion matrices, the modified version of the Hotelling r defined by
Anderson (1958) is employed in the case where the dispersion matrices are
not equal. The use of these measures of group overlap in the conducted
experimental design is performed as follows: Initially, the means of all five
criteria for the first group is fixed to a specific value (1 for the case of
multivariate normal, uniform and exponential distribution, and 8 in the case
of the log-normal distribution). Then, the means of the criteria for the second
r
group are specified so that the Hotelling (or its modified version) for the
differences in the means of groups 1 and 2 is significant at the 1% and the
10% significance level (low and high degree of group overlap). Similarly,
the means of the third group are specified so as the Hotelling r (or its
modified version) for the differences in the means of groups 2 and 3 is
significant at the 1% and the 10% significance level.
To ensure the consistency in the ordering of the classes (groups) from the
best (group l-CI ) to the worst one (group l-C3), all data are generated so
that the following condition is satisfied: g,{a»g,{b), VaECh bECk+ b i=l, 2,
... , 5, k=1, 2. This ensures that there is no alternative of group Ck that is
dominated by an alternative of group Ck+I.
For each combination of the factors F2-F7 (192 combinations) a training
sample and a validation sample are generated, having all the properties that
these factors specify. The size of the training sample is defined by the factor
F4 , while the size of the validation sample (holdout sample) is fixed at 216 in
all cases. For each factor combination 20 replications are performed.
Therefore, during this experiment the number of samples considered is 7,680
(192x20=3,840 training samples matched to 3,840 validation samples).
Overall, the conducted experiment involves a 4x4x2x3x2x2x2 full-level
factorial design consisting of768 treatments (factor combinations).
On the Use ofMulticriteria Classification Methods 223
4. Results
Table 2. Major explanatory effects regarding the classification performance of the methods
(seven-way ANOV A results)
The interaction effects are of major interest in analysing the results of the
experimental design with regard to the relative performance of the
considered methods. All three interaction effects that are found to explain a
high proportion of the total variance in the obtained results, involve the
interaction of the factor FI (classification procedures) with other factors, in
particular the homogeneity of the group dispersion matrices (F6), the
distributional form of the data (F2)' and the training sample size (F 4). Table 3
summarizes the results of all methods throughout all experiments, while
Tables 4-6 provide further details on the comparison of the methods in terms
of the aforementioned two and three-way interaction effects that are found
significant through the analysis of variance. Each of these tables reports the
average transformed error rate, the untransformed error rate (in parentheses)
224 AIDING DECISIONS WITH MULTIPLE CRITERIA
and the grouping obtained through the Tukey's test on the average
transformed error rates [cf. equation (9)].
Table 3. Tukey's test for significant differences between methods along all experiments
methods (factor: F 1)
of disEersion matrices
Homogeneit~
Egual Unegual
Tukey's Tukey's
Mean Mean
~ouEing ~ouEing
1.2391 1.1610
LDA C C
(33.98%) (30.42%)
1.3283 0.8059
QDA D B
(38.17%) (15.80%)
1.2183 1.1599
LOGIT B C
(33.01 %) (30.37%)
0.9932 0.7545
UTADIS A A
{24.20%) {15.01%}
Distribution
Nonnal Unifonn Exponential Log-nonnal
Tukey's Tukey's Tukey's Tukey's
Mean Mean Mean Mean
grouping grouping grouping grouping
1.1900 1.2312 1.1642 1.2147
LDA B C C C
(31.69%) (33.57%) (30.66%) (32.86%)
1.0917 1.0634 1.0392 1.0740
QDA A B B B
(27.79%) (27.02%) (26.97%) (27.16%)
1.1827 1.2200 1.1517 1.2020
LOGIT B C C C
(31.36%) (33.03%) (30.11%) (32.25%)
1.1061 0.9851 0.5126 0.8916
UTADIS A A A A
(28.\0%) (23.43%) (7.36%) (19.52%)
On the Use ofMulticriteria Classification Methods 225
Table 6. Tukey's test for significant differences between methods (factor: F4)
The results indicate that, overall, the UTADIS method outperforms all
statistical procedures, followed by QDA, while the performances of LDA
and LA are similar. The analysis, regarding the performance of the methods
when the three two-way significant interactions are considered (methods by
dispersion matrices, distribution and size; cf. Tables 4-6), further pronounces
the superiority of the UTADIS method. The results indicate that in all cases
the UTADIS method provides consistently lower error rates than the
considered statistical classification techniques. The differences that are
evident between the performances of the methods are found significant
through the Tukey's test at the 5% significance level. The only exception to
these remarks is the case of multivariate normal data, where QDA provides
slightly lower error rate than UTADIS (27.79% for QDA vs. 28.10% for
UTADIS). However, even in this case the difference between the two
methods is not found to be statistically significant at the 5% level.
These results indicate that theoretical advantages (cf. section 3.1) of the
proposed MCDA methodology over the statistical techniques is associated
with an improved classification performance. Furthermore, the higher fitting
ability that results from the increased number of degrees of freedom in the
additive utility models of the UTADIS method and the increased
computational effort required for model development are by higher
generalizing classification performance which is crucial in most real world
cases.
226 AIDING DECISIONS WITH MULTIPLE CRITERIA
5. Conclusions
References
Altman, E.I., Avery, R, Eisenbeis, Rand Stinkey, 1. (1981), Application of Classification
Techniques in Business, Banking and Finance, Contemporary Studies in Economic and
Financial Analysis, Vol. 3, JAI Press, Greenwich.
Anderson, T.W. (1958), An Introduction to Multivariate Statistical Analysis, Wiley, New
York.
Bajgier, S.M. and Hill, A.V. (1982), "An experimental comparison of statistical and linear
programming approaches to the discriminant problem", Decision Sciences, 13,604-618.
Dias, L., Mousseau, V., Figueira, 1. and Clirnaco, J. (2000), "An aggregation/disaggregation
approach to obtain robust conclusions with ELECTRE TRJ", Cahier du LAMSADE, No 174,
Universite de Paris-Dauphine.
Doumpos, M., Zopounidis, C. and Pardalos, P.M. (2000), "Multicriteria sorting methodology:
Application to financial decision problems", Parallel Algorithms and Applications, 15/1-2,
113-129.
Freed, N. and Glover, F. (1981), "Simple but powerful goal programming models for
discriminant problems", European Journal of Operational Research, 7, 44-60.
Jacquet-Lagreze, E. (1995), "An application of the UTA discriminant model for the
evaluation of R&D projects", in: P.M. Pardalos, Y. Siskos, C. Zopounidis (eds.),
Advances in Multicriteria AnalysiS, Kluwer Academic Publishers, Dordrecht, 203-211.
Jacquet-Lagreze, E. and Siskos, Y. (1982), "Assessing a set of additive utility functions for
multicriteria decision making: The UTA method", European Journal of Operational
Research, 10, 151-164.
Jacquet-Lagreze, E. and Siskos, J. (2001), "Preference disaggregation: Twenty years of
MCDA experience", European Journal of Operational Research, 130/2,233-245.
Joachimsthaler, E.A. and Starn, A. (1988), "Four approaches to the classification problem in
discriminant analysis: An experimental study", Decision Sciences, 19,322-333.
Keeney, R.L. and Raiffa, H. (1993), Decisions with Multiple Objectives: Preferences and
Value Trade-offs, Cambridge University Press, Cambridge.
Mardia, K.V. (1975), "Assessment of multinormality and the robustness of Hotelling's T2
test", Applied Statistics, 24, 163-171.
Mousseau, V. and Slowinski, R. (1998), "Inferring an ELECTRE-TRI model from assignment
examples", Journal of Global Optimization, 12/2,157-174.
Mousseau, V., Figueira, 1. and Naux, J.-Ph. (2001), "Using assignment examples to infer
weights for ELECTRE TRI method: Some experimental results", European Journal of
Operational Research, 130/2,263-275.
Pardalos, P.M., Siskos, Y. and Zopounidis, C. (1995), Advances in Multicriteria Analysis,
Kluwer Academic Publishers, Dordrecht.
Patuwo, E., Hu, M.Y. and Hung, M.S. (1993), "Two-group classification using neural
networks", Decision Sciences, 24, 825-845.
Pawlak, Z. (1982), "Rough sets", International Journal of Information and Computer
Sciences, 11,341-356.
Pawlak, Z. and Slowinski, R (1994), "Rough set approach to multi-attribute decision
analysis", European Journal of Operational Research, 72, 443-459.
Quinlan J.R. (1986), "Induction of decision trees", Machine Learning, I, 81-106.
Roy, B. and Moscarola, 1. (1977), "Procedure automatique d'examem de dossiers fondee sur
une segmentation trichotomique en presence de criteres multiples", RAIRO Recherche
Operationnele, 11/2, 145-173.
228 AIDING DECISIONS WITH MULTIPLE CRITERIA
Slowinski, R. and Zopounidis, C. (1995), "Application of the rough set approach to evaluation
of bankruptcy risk", International Journal of Intelligent Systems in Accounting, Finance
and Management, 4, 27-41.
Vale, D.C. and Maurelli, V.A. (1983), "Simulating multivariate nonnormal distributions",
Psychometrika, 48/3, 465-471.
Yu, W. (1992), "ELECTRE TRI: Aspects methodologiques et manuel d'utilisation",
Document du Lamsade No 74, Universite de Paris-Dauphine, 1992.
Zopounidis, C. and Doumpos, M. (1999), "A multicriteria decision aid methodology for
sorting decision problems: The case of financial distress", Computational Economies,
14/3,197-218.
Zopounidis, C. and Doumpos, M. (2000), "Building additive utilities for multi-group
hierarchical discrimination: The M.H.DIS method", Optimization Methods and Software,
14/3,219-240.
ORDINAL MULTIATTRIBUTE SORTING
AND ORDERING IN THE PRESENCE
OF INTERACTING POINTS OF VIEW
Marc Roubens
University of Liege, Institute of Mathematics, Belgium
M.Roubens@ulg.ac.be
Abstract In this paper, we use the Choquet integral as a general tool for dealing
with ordinal multiattribute sorting and ordering problems in the pres-
ence of interacting points of view. The technique that is used proceeds
in two steps : a pre-scoring phase determines for each point of view
and for each alternative a net score (number of times a given alterna-
tive beats all the other alternatives minus the number of times that this
alternative is beaten by the others) and is followed by an aggregation
phase that produces a global net score associated to each alternative.
The assessment of a capacity linked to the Choquet integral is ob-
tained by the solution of a constraint satisfaction problem deriving from
a reference set of alternatives (prototypes) that have been previously
ordered or sorted. We give examples of application comparing this ap-
proach with a rule based methodology.
1. Introduction
Let X = TIi=l Xi be a product space, where Xi is the ordered set
of possible evaluations of alternatives from the set A of cardinality m
for point of view i, belonging to the set N of cardinality n.
The performance scale is considered to be a totally ordered set Xi :
{gi --< ... --< g~J, i.e. a ni -point scale.
A profile that corresponds to alternative x E A is a vector x =
(Xl, ... ,Xn )
E X, where Xi = gi(x).(XiZ-d represents a profile that corresponds to z
except that its i-th component is equal to Xi.
230 AIDING DECISIONS WITH MULTIPLE CRITERIA
x t y iff Xi t Yi·
i
Ri (x, y) = 1 if x t y
i
0, otherwise.
Si (x) represents the number of times that x is better than any other
alternative minus the number of times that any other alternative is better
than x for point of view i.
The definition of the partial net score clearly indicates that this utility
function is measured on an interval scale. Positive linear transformations
are meaningful with respect to such a scale.
In order to obtain normalized measures, we consider the net scores
(this transformation is legitimate) :
sty ( ) = Si(X) + (m - 1)
z x 2(m _ 1)
to obtain
sf (x) 1 if x ~
l
y, for all y 1= x EA
= 0 if Y ~ x, for all y 1= x E A.
l
A ranking O:A) is a total preorder on the set A which does not con-
tradict the principle of coherence with respect to dominance (CD R) :
xDy ::::} x ~A y.
There exists a numerical representation of the ranking, i.e. a mapping
F : A -7 IR such that
v(0) 0
v(N) = 1
and A(i) is the coalition of points of view {(i), ... ,(n)}, <5i v(A(i)} is the
marginal contribution in capacity of (i) to those points of view which
are ranked before (i) by 1r :
2 there exists
As the net scores can serve as criteria, we can rewrite the last propo-
sition replacing
by
J[sf (x), ... , S~ (x)] ~ Zr {:} x E CE~
where Zr = minxEclr f[Sf (x), ... , S~ (x)].
Mv = L v(T}fT,
TeN
m(B) = L (-l)IB\Clv(C), VB c N.
GeB
v(B) = L m(C)
GeB
L m(T) ~ 0, VB C N, Vi E B
T:iETeB,
ITI~k
L m(i) + L m(i,j) =1
i {i,j}eN
m(0) = °
m(i) ~ 0, Vi E N
m(i) +L m(i,j) ~ 0, Vi E N, VB C N \ i
jEB
Cv(SN (Xl)) > Cv(SN (X2)) ~ ... ~ Cv(SN (X)) ~ ... ~ Cv(SN (Xi))
> min Cv(SN (x)) = Zr
xENdr
obtain a non-empty set of solution. For each step k, we solve the linear
programme
maxc: (2)
under constraints linked to a k-additive capacity and sorting constraints
of type (1) where the strict inequalities are replaced by
Cv(SN (x)) ~ Cv(SN (y)) + c:, for all x E Nd r , y E ND r- 1
with c: ~ o.
The objective fonction is introduced in order to maximize the separa-
tion between the classes of the sorting.
The stopping rule corresponds to the first k such that c: > o.
( 22 12) >
17 17). .
Cv ( 24 ' 24 Implies
Cv 24' 24
( 22
Cv 24' 24
7) > Cv ( 24'
17 12)
24 = Cv ((3), (2)) or
Cv ((4), (1)) > Cv ((3), (2)) which contradicts the DM's preferences.
5. An illustrative example of sorting
Let us consider a typical sorting problem presented in Greco and al.
[5]. Suppose that a school director wants to assign students to different
classes of merits on the basis of their scores in Mathematics (m), Physics
(p) and Literature (1). The ordinal scales of the evaluation in the three
courses as well as the global evaluation scale have been composed of
three grades : "bad" (B), "medium"(M), "good" (G).
Table 1 reproduces the evaluation of twenty seven students with re-
spect to three criteria and a sorting decision made by the director to-
gether with the net scores and the values of the Choquet integrals
(m = 27). In this case P = A.
Figure 1 presents the subgraphs G(Cfr, rr), r = 3,2,1 that corre-
sponds to the classes of "good" students (r = 3), "medium" students
(r = 2) and "bad" students (r = 1), as a Hasse diagram. The non-
dominating sets and non-dominated sets are :
NDa = ({GGG)}, Nda = ({GMM),{MGM)},
ND2 = ({MMG), (GGB)}, Nd2 = ({GMB), (MGB), (MMM)},
NDI = ({BGG), (GBG), (MMB)}, Nd 1 = ({BBB)}
240 AIDING DECISIONS WITH MULTIPLE CRITERIA
81 (BBB) B 4 4 4 4
82 (MBB) B 13 4 4 4 + 9v(m)
83 (GBB) B 22 4 4 4 + 18v(m)
84 (BMB) B 4 13 4 4 + 9v(p)
85 (MMB) B 13 13 4 4 + 9v(m,p)
86 (GMB) M 22 13 4 4 + 9v(m,p) + 9v(m)
87 (BGB) B 4 22 4 4 + 18v(p)
88 (MGB) M 13 22 4 4 + 9v(m,p) + 9v(P)
89 (GGB) M 22 22 4 4 + 18v(m,p)
810 (BBM) B 4 4 13 4 + 9v(£)
811 (MBM) B 13 4 13 4 + 9v(m, £)
812 (GBM) B 22 4 13 4 + 9v(£, m) + 9v(m)
813 (BMM) B 4 13 13 4 + 9v(p,£)
814 (MMM) M 13 13 13 13
815 (GMM) G 22 13 13 13 + 9v(m)
816 (BGM) B 4 22 13 4 + 9v(p, £) + 9v(p)
817 (MGM) G 13 22 13 13 + 9v(p)
818 (GGM) G 22 22 13 13 + 9v(m,p)
819 (BBG) B 4 4 22 4 + 18v(£)
820 (MBG) B 13 4 22 4 + 9v(m,£) + 9v(£)
821 (GBG) B 22 4 22 4 + 18v(m, £)
822 (BMG) B 4 13 22 4 + 9v(p, £) + 9v(£)
823 (MMG) M 13 13 22 13 + 9v(£)
824 (GMG) G 22 13 22 13 + 9v(m,£)
825 (BGG) B 4 22 22 4 + 18v(p, £)
826 (MGG) G 13 22 22 13 + 9v(p, £)
827 (GGG) G 22 22 22 22
Ordinal Multiattribute Sorting and Ordering 241
v(0) = 0, v(m,p, £) = 1
1 ~ v(m,p) ~ v(m) ~ 0, 1 ~ v(m,p) ~ v(p) ~ 0,
1 ~ v(m,£) ~ v(m) ~ 0, 1 ~ v(m,£) ~ v(£) ~ 0,
1 ~ v(p,£) ~ v(p), 1 ~ v(p,£) ~ v(£)
242 AIDING DECISIONS WITH MULTIPLE CRITERIA
E---~CI'3
Acknowledgment
The author is grateful to an anonymous referee whose remarks im-
proved the contents of the part related to the rule based methodology.
246 AIDING DECISIONS WITH MULTIPLE CRITERIA
References
[IJ Choquet, G., (1953), Theory of capacities, Annales de l'Institut Fourier, 5, 131-
295.
[2J Grabisch, M. (1997), k-order additive discrete fuzzy measures and their represen-
tation, Fuzzy Sets and Systems 92, 167-189.
[3J Grabisch, M., Roubens, M. (2000), Application of Choquet Integral in Multicrite-
ria Decision Making, in Grabisch, M., Murofushi, T. and Sugeno, M. (Eds) Fuzzy
Measures and Integrals, Theory and Applications, Physica Verlag, Heidelberg.
[4J Grabisch, M., Marichal, J.-L., Roubens, M. (2000), Equivalent representations of
set functions, Mathematics of Operations Research 25/2, 157-178.
[5J Greco, S., Matarazzo, B., Slowinski, R. (1998), A rough set approach to multicri-
teria and multiattribute classification, in Polkowski, L. and Skowron, A. (Eds),
Rough sets and current trends in computing, vol. 1424 of Lecture Notes in Arti-
ficial Intelligence, 60-67, Springer Verlag, Berlin.
[6J Greco, S., Matarazzo, B., Slowinski, R. (2000), Conjoint measurement and rough
set approach for multicriteria sorting problems in presence of ordinal criteria, in:
Colorni, A., Parruccini, M., Roy, B. (Eds), Selected papers from 49th and 50th
Meeting of the EURO working group on MCDA, EUR-Report, Ispra-Paris, in
print.
[7J Marichal, J.-L. (2000), Behaviorial Analysis of Aggregation in Multicriteria De-
cision Aid, in Fodor, J., De Baets, B., and Perny, P. (Eds), Preferences and
Decisions under Incomplete Knowledge, Physica Verlag, Heidelberg.
[8J Marichal, J.-L. (2000), An axiomatic approach of the discrete Choquet integral
as a tool to aggregate interacting criteria, IEEE Trans. Fuzzy Syst., to appear.
[9J Schmeidler, D. (1986), Integral representation without additivity, Roc. Amer.
Math. Soc. 97, 255-261.
IV
PREFERENCE MODELING
MULTIATTRIBUTE INTERVAL ORDERS
Peter C. Fishburn
AT&T Shannon Laboratory
fish@research.att.com
1. Introduction
In this introduction we present basic terminology and assumptions
along with two versions of our focal representation. Section 2 describes
implications of the representation and comments on some of its features
and specializations. Section 3 takes a closer look at the model's indepen-
dence or cancellation conditions and notes a scheme of such conditions
that is necessary and sufficient for the representation. The paper con-
cludes with a brief discussion.
We assume throughout that >-, interpreted as strict preference, is an
asymmetric binary relation on a nonempty finite set X = Xl X X 2 x··· X
X n . We denote the symmetric complement of>- by so X Y if neither
f"V, f"V
even though the reasons for x y might vary from pair to pair. For
f"V
sense, incomparable.
250 AIDING DECISIONS WITH MULTIPLE CRITERIA
2. Implications
This section identifies implications of Modell or Modell * (Lemmas 1
through 4), then notes special cases of the model. The inapplicability of
the Scott-Suppes representation for Modell * when attribute preference
relations are semiorders is verified by the proof of Theorem 1. The
section concludes with critical comments on the model. Prior to that,
Model 1 or Model 1* is assumed to hold with f, g, and I = 9 - f as
defined therein, and with n ~ 2.
252 AIDING DECISIONS WITH MULTIPLE CRITERIA
Proof. If Xi ""i Yi for all i, then f(xi) ::; g(Yd for all i, so not
(x >- y). Similarly, not(y >- x), so x "" y. If Xi h Yi for all i E I(x,y),
then f(xd > g(Yi) for all i E I(x, Y), so L:I(x,y)[J(Xi) - g(Yi) > 0 and
x >- y . •
Even when n is large, x >- Y is not assured when Xi ""i Yi and Xj'r---j-Yj
for all j i- i. For example, if Xl ""1 YI and Xj >-j Yj for j = 2, ... , n, we
have x"" Y iflf(xd-g(Ydl > L:j~2[J(Xj)-g(Yj)]. This may be unlikely,
but is plausible when one attribute is substantially more important than
the others.
We have already remarked on the importance of I(x, Y), whereby
attributes for which Xi = Yi have no bearing on X >- y. If I(x, y) were
omitted from Modell, with L:~=I replacing L:iEI(x,y) , then >- on X
would be an interval order. We note shortly that this need not be true
for ModelL First, an elementary observation about I.
Lemma 3 I(x,z) ~ I(x,y) UI(y,z) for all X,y,Z E X.
iEI(x,y) iEI(y,z)
> L l(Zi).
iEI(x,z)
and it is easily seen that all four inequalities can hold simultaneously. •
Our first special case of Model 1 is the basic additive utility repre-
sentation in which! = 9 or 1 == 0 with )- and every )-i a weak order.
Axioms for this case are included in Scott (1964), Krantz, Luce, Suppes
and Tversky (1971), and Fishburn (1970).
A second specialization arises when every )-i is a semiorder. The
threshold structure for this case is more uniform than for the interval
orders case, but a proof like that for the second part of Lemma 4 shows
that )- need not be a semi order or an interval order. In addition, the
Scott-Suppes semiorder representation for each h may be inapplicable
for Modell.
and
bl )-2 b3 )-2 b4 )-2 b5 , b2 )-2 b4, bl rv2 b2 ""2 b3 .
Then )-1 and )-2 are semiorders. Suppose Modell * holds with 1 == (St ~ 0
on Xl and 1 == 02 ~ 0 on X 2. Suppose further that (a2,bd )- (al,b3)
254 AIDING DECISIONS WITH MULTIPLE CRITERIA
and (a3, b5) >- (a5, b4), which are consistent with the original form of the
model. We have
Modell * holds here when l == 0 and, for each i < n for which f on Xi
is not constant,
n
min{J(xi) - f(Yi) : Xi h Yi} > L max{J(Xj) - f(Yj} : Xj >-j Yj}·
j=i+l
Multiattribute Interval Orders 255
and
n
X >- Y {::} I)!{Xi) - !(Yi)] > II{x, y)l·
i=l
3. Independence
We conclude our technical analysis of Model 1 by noting a scheme
of independence/cancellation conditions that is necessary and sufficient
for the model. We precede it with a simpler scheme that is necessary
but not sufficient. Both are based on a cancellation or balance relation
E between finite lists of members of X. Given m E {I, 2, ... } and
1 ... ,yffi EX ,wewne
x 1, ... ,xffi ,y, 't
(Xl, ... , Xffi)E(yl, ... , yffi) if yI, yr, ... , yrn if a permutation
of Xi1, Xi2, ... , Xiffi£or z'-1
- , ... , n.
where in (II) we have used yj '" xj. Because (xl, ... , xm)E(yl, . .. ,ym),
the left sides of the preceding inequalities are equal, and therefore
m k
L L l(xi) > L L l(vi)·
j=k+l iEI(xi ,yi) j=l iEI(xi ,yi)
However, the new ~ hypothesis of 82 implies
k m
L L l(vi) ~ L l(x{)
j=l iEI(xi ,yi) j=k+l iEI(xi ,yi)
(v) the sum of all {3j in (i)-(iii) multiplied by the same corresponding
integers equals O.
This gives 2N equations, two for each Cj.
Proceeding under the supposition that there is no (8, t) solution, let
k ~ 1 be the sum of the integers for (i), let m - k ~ 0 be the sum of
the absolute values of the integers for (ii), and let rj ~ 0 be the absolute
value of the integer for j in (iii). Then, with multiples of the x >- y
and x '" y pairs according to the corresponding integers, we have lists
xl >- yl, ... , xk >- yk and x k+ l '" yk+1, ... ,xm '" ym such that, for every
Cj E Xi and every i E {I, ... ,n}, (iv) and (v) yield
I{l :S 'Y :S k: xl = Cj and i E I(x'Y, y'Y)}1
-I{k < 'Y :S m : xl = Cj and i E I(x'Y, y'Y)}1 - rj =0
4. Discussion
Several approaches to multiattribute/multicriteria decision making
have risen to prominence during the past generation. The present paper
focuses on the approach in which a primitive strict preference relation
on a product set is represented numerically by an algebraic structure
based on the attributes of the product sets. Its particular emphasis is
an additive representation with attribute-preference interval orders and
strict-preference thresholds that accumulate additively over attributes
on which alteratives differ.
I have mentioned aspects of the representation that are liable to refu-
tation, including independence across attributes and its algebraic struc-
ture. A related concern is model validation, testing S2 or some of its
simpler implications as in Lemmas 1, 2 and 4. If the model is grossly
inadequate, this can usually be discovered quickly. A more interesting
situation occurs when no simple refutation is uncovered. If small-m cases
of S2 hold, then the representation is probably adequate for all practical
purposes. It is true for finite X that S2 holds in general if it holds up
to some value m* of m for given values of the lXii, but m* will usually
be large enough to discourage extensive testing of S2 for nearby values
ofm.
References
Abbas, M. and Ph. Vincke: Preference structures and threshold models, Journal of
Multi-Criteria Decision Analysis 2 (1993), 171-178.
Doignon, J.-P.: Threshold representations of preference relations, manu-script (1984).
Doignon, J.-P., B. Monjardet, M. Roubens and Ph. Vincke: Biorder families, valued
relations, and preference modelling, Journal of Mathematical Psychology 30 (1986),
435-480.
Fishburn, P. C.: Utility Theory for Decision Making. New York: Wiley, 1970.
Fishburn, P. C.: Mathematics of Decision Theory. Paris: Mouton, 1972.
Fishburn, P. C.: Axioms for lexicographic preferences, Review of Economic Studies
42 (1975), 415-419.
Fishburn, P. C.: Interval Orders and Interval Graphs: A Study of Partially Ordered
Sets. New York: Wiley, 1985.
Fishburn, P. C.: Nontransitive preferences in decision theory, Journal of Risk and
Uncertainty 4 (1991), 113-134.
Fishburn, P. C.: Stochastic utility, in Handbook of Utility Theory (ed. S. Barbera, P.
J. Hammond and C. Seidl), 273-319. Dordrecht: Kluwer, 1997.
Hokkanen, J. and P. Salminen: ELECTRE III and IV decision aids in an environmen-
tal problem, Journal of Multi-criteria Decision Analysis 6 (1997), 215-226.
Krantz, D. H., R. D. Luce, P. Suppes and A. Tversky: Foundations of Measurement:
Volume 1. New York: Academic Press, 1971.
Piriot, M. and Ph. Vincke: Semiorders: Properties, Representations, Applications.
Dordrecht: Kluwer, 1997.
REFERENCES 261
Roy, B.: Classement et choix en presence de points de vue mUltiples (Ia methode
Electre), Revue Franr;aise d'Informatique et de Recherche Operationnelle 8 (1968),
57-75.
Roy, B.: The outranking approach and the foundations of ELECTRE methods, Theory
and Decision 31 (1991), 49-73.
Scott, D.: Measurement structures and linear inequalities, Journal of Mathematical
Psychology 1 (1964), 233-247.
Scott, D. and P. Suppes: Foundational aspects of theories of measurement, Journal
of Symbolic Logic 23 (1958), 113-128.
Suppes, P., D. H. Krantz, R. D. Luce and A. Tversky: Foundations of Measurement:
Volume 2. New York: Academic Press, 1989.
Tsoukias, A. and Ph. Vincke: Double threshold orders: a new axiomatization, Journal
of Multi-Criteria Decision Analysis 7 (1998), 285-301.
Tversky, A.: Intransitivity of preferences, Psychological Review 76 (1969), 31-48.
Vincke, Ph.: {P, Q, I}-preference structures, in Non-Conventional Preference Rela-
tions in Decision Making (ed. J. Kacprzyk and M. Roubens), 72-81. Berlin: Springer,
1988.
PREFERENCE REPRESENTATION
BY MEANS OF CONJOINT MEASUREMENT
AND DECISION RULE MODEL
Salvatore Greco
Facolta di Economia, Universita di Catania, Italy
salgreco@mbox.unict.it
Benedetto Matarazzo
Facolta di Economia, Universita di Catania, Italy
matarazz@mbox.unict.it
Roman Slowinski
Institute o/Computing Science, Poznan University o/Technology, Poland
slowinsk@sol.put.poznan.pl
Key words: Preference modeling; Conjoint measurement; Decision rules; Ordinal criteria;
Inconsistency; Rough sets; Axiomatization
264 AIDING DECISIONS WITH MULTIPLE CRITERIA
1. Introduction
In economics, social choice theory and multicriteria decision making, a
lot of attention has been paid to conjoint measurement models representing
preferences by means of numerical functions. Researchers investigated a
wide variety of models going from simplest classical additive and transitive
models to most sophisticated non-additive and non-transitive models.
Recently, an alternative approach to representation of preferences has been
considered: preference models in terms of"ij. .. , then ... " decision rules.
The decision rules concern pairs (x, y) of actions and conclude either
presence or absence of a comprehensive preference relation between x and y;
conditions for the presence are expressed in "at least" comparison terms, and
for the absence in "at most" comparison terms, on particular criteria. For
example, in multicriteria choice and ranking problems, the two kinds of rules
are like: "if x is at least better than y on criterion i and x is at least weakly
preferred to y on criterionj, than x is at least as good as y", or "if x is at most
worse than y (i.e. worse or much worse) on criterion i and x is at most
indifferent (i.e. indifferent or worse) on criterion j, then x is not as good as
y".
Traditionally, preferences have been modeled using a value function u(-)
such that action x is at least as good as action y, denoted by xSy, iff
u(x);:::u(y). This implies that the relation S is complete (for each couple of
actions x,y, xSy or ySx) and transitive (for each triple of actions x,y,z, xSy
and ySz imply xSz). In a multicriteria context, each action x is generally
seen as a vector X=[Xt,X2,""Xo] of features of x with respect to criteria 1, ... ,no
It is often assumed that the value function is additive (see, e.g., Keeney and
Raiffa 1976, Krantz et al. 1978, and Wakker 1989, for an axiomatic
characterization), i.e.
o
u(x)= L uj(x j), (1)
j=l
where Uj is a marginal utility of action x with respect to criterion i (i=1, ... ,n).
The additive and transitive model represented by the additive value
function is inappropriate in many situations, because in reality:
• the indifference (the symmetric part of S) may not be transitive,
• S may not be complete, i.e. some pairs of actions may be incomparable,
• the compensation between evaluations of conflicting criteria and the
interaction between concordant criteria are far more complex than the
capacity of representation by the additive value function.
Conjoint Measurement and Decision Rule Model 265
3 if u;{x;}-u;{y;}=2
1 if U;{X;}-U;{y;} =1
o if U;{X;}-U;{y;} =0
-I if u;{x;}-u;{Y;}=-1
-2 if u;{x;}-u;{y;}=-2
0.75 ifuj{xj)-uj{yJ= 2
0.5 if uj{x.)-uj{Y·)=1
o if uj(XJ-Uj(yJ= 0
-I ~f uj(xj)-uj(yJ=-1
-2 If Uj(XJ-Uj(yJ=-2
and
G{'Pj[Uj(Xj), Uj(Yj)], 'Pj[Uj(Xj), Uj(Yj)]} = 'Pj[Uj(Xj), Uj(Yj)] + 'Pj[Uj(Xj), Uj(Yj)]·
The above equivalence is not by chance. The equivalence of preference
models represented by numerical function (2) and by a set of decision rules
is a fundamental result of this paper. The advantage of model (2) over
decision rules is that numerical representation of preferences simplifies the
calculation. The advantage of decision rules over numerical function models
relies on their intelligibility and capacity of explanation. Preference
representation in terms of decision rules seems closer to human reasoning in
everyday decision making because, as remarked by Slovic (1975), "people
make decisions by searching for rules which provide good justification of
their choices". In result of the established equivalence of the two classes of
preference models, all the multicriteria decision aiding methods based on
model (2) can be presented in terms of decision rules.
Another important aspect related to preference representation is
hesitation of DMs while expressing their preferences. Every preference
model is build upon some preferential information acquired from the DM.
The DMs have no doubt when expressing this information rather rarely. This
is due to complexity of multicriteria comparisons and unstable character of
DM's preferences. Hesitation may lead to inconsistencies that should be
handled by the modeling procedure and by the model itself. Model (2),
although very general, does not handle the inconsistencies.
Hesitations are manifested in different ways. In this paper we consider
two cases:
1) The case of inconsistent preferences: there are actions, say x,y,W,Z, such
that on all considered criteria the preference of x over y is at least as
Conjoint Measurement and Decision Rule Model 267
strong as the preference ofw over z but, unexpectedly, the DM considers
w comprehensively at least as good as z, and x comprehensively not at
least as good as y. The reason for such an inconsistency could be a
missing criterion, however, construction of this criterion may be either
very difficult, or impossible. Hesitations manifested in this way can be
represented by approximate decision rules: "if [conditions on particular
criteria], then there are no sufficient reasons to conclude that x is, or is
not, at least as good as y". For example: ''ifaction x is worse than action
y on criterion i and x is better than y on criterion j, then there are no
sufficient reasons to conclude that x is, or is not, at least as good as y".
In the paper we show that also a numerical function model, similar to
model (2), can handle these inconsistent preferences as follows:
G['Pi[ulxi), Ui(Yi)], H ..... n] ~ t2 iff x is at least as good as y without
doubt;
G['Pi[Ui(Xi), UlYi)], i=I .....n] ~ tl iff x is not at least as good as y without
doubt;
tl < G['Pi[ Ui(Xi), UlYi)], i=I .....n] < t2 iff there are no sufficient reasons to
conclude that x is, or is not, at least as good as y;
t\,t2ER such that t l< t2.
In this model the value of numerical function G['Pi[Ui(Xi), Ui(yi)].i=I ..... n]
can be interpreted as strength of the arguments in favor of the statement
"x is at least as good as y". Consequently, the value of tl is the upper
bound of the strength of arguments in favor of the conclusion "x is not
at least as good as y", while t2 is the lower bound for the conclusion "x
is at least as good as y". The values between tl and t2 represent an
intermediate strength of arguments corresponding to the situation of
hesitation.
2) The case of positive and negative reasons for preferences: in this case
the DM considers arguments for the statement: "x is at least as good as
y", and for the opposite statement: "x is not at least as good as y"
(Tsoukias and Vincke 1995, 1997). In this context, the hesitation occurs
when for the statement "x is at least as good as y":
a) there are both, reasons in favor and reasons against (contradiction),
b) there are neither reasons in favor nor reasons against (ignorance).
Decision rules can express naturally the reasons in favor and the reasons
against. As an example consider the following three decision rules:
268 AIDING DECISIONS WITH MULTIPLE CRITERIA
where G['I'i[Ui(Xi), Ui(Yi)], i=I .....n] and GC['I'i[Ui(Xi), Ui(Yi)], i=I .....n] are
functions non-decreasing in each argument.
In this model, function G['I'i[uiCxi), UiCYi)], i=I .....n] represents reasons in
favor of the statement "x is at least as good as y", while function
GC['I'i[Ui(Xi), Ui(Yi)], i=I .....n] represents reasons against. Four cases are
possible with respect to each pair of actions (x,y):
a) G['I'i[uiCxi), Ui(Yi)], i=I .....n] ~ 0 and GC['I'i[Ui(Xi), Ui(Yi)], i=I .....n] ~ 0: in
this case there are reasons in favor and there are no reasons against,
thus x is at least as good as y, without doubt;
[3) G['I'i[Ui(Xi), Ui(Yi)], i=I .....n] < 0 and GC['I'i[Ui(Xi), Ui(Yi)], i=I .....n] < 0: in
this case there are reasons against and there are no reasons in favor,
thus x is not at least as good as y, without doubt;
y) G['I'i[Ui(Xi), Ui(Yi)]>i=I .....n] ~ 0 and GTI'i[Ui(Xi), UiCYi)], i=I .....n] < 0: in
this case there are both, reasons in favor and reasons against, thus
there is a contradictory information about the statement "x is at least
as good as y";
0) G['I'i[Ui(Xi), Ui(Yi)], i=I .....n] < 0 and GC['I'i[Ui(Xi), Ui(Yi)], i=I .....n] ~ 0: in
this case there are neither reasons in favor nor reasons against, thus
there is no information about the statement "x is at least as good as y".
Conjoint Measurement and Decision Rule Model 269
Xi sf Yi iff for all a.iEX.i and ZEX, [(Yi. a.i) S Z ~ (Xi. a.v S z],
XiS~ Yi iff for all a.iEX.i and ZEX, [z S (Xi. a.i) ~ Z S (Yi. a.i)].
The relation (Xi. Yi) S~ (Wi. Zi) reads: "the strength of preference of Xi
over Yi is at least as great as the strength of preference of Wi over Zj". Due to
the implication in its definition, binary relation S~ is transitive.
Let us remark that the binary relation S~ is coherent with the marginal
outranking relation sf, i.e.
272 AIDING DECISIONS WITH MULTIPLE CRITERIA
i) x·s~y·
I I I and
Xi S~ Yi implies [(Wi, Xi) S~ (Zj, Ui) => (Wi, Yi) S~ (Zi, Ui)]. (4)
For each Xi,Yi,Wj,ZiEXi, i=l, ... ,n, and hi,kiEHi' hi~ki, XiPr i Yi and
Wi pfi Zi mean that Xi is preferred to Yi at least as much as Wi is preferred to
Zi, i.e. (Xi'Yi)S~(Wj,zJ For example, ifH i={1,2, ... ,5}, then for each Xi,yiEX
we could have:
~ Xi Pfhi Yi' which means that "Xi is preferred to Yi in degree at least h;",
if there exists kiEHi such that ki~hi and Xi pfi Yi.
X· S®
1
. y.
1 1
and u·1 ph.
. X.1
1
I
........
-.'" u.1 p>h.
;- y.l'
1
I
Condition Cl(i) ensures that, on the basis ofS, the relation S~ (i=l,oo.,n)
is strongly complete on XiXX i: in fact, S; would be incomplete if there were
Xi,yi,Zi,WiEXh a_i,b_i,c_i,d_iEX.i such that
i) (Xi, a_i) S (Yh b_i) and not (Wi, a-i) S (Zi, b_i): by definition of S; this
implies not (Wi, Zi) S; (Xi, Yi);
ii) (Wi, C_i) S (Zh d_ i) and not (Xi, C-i) S (Yh d_ i): by definition of S; this
implies not (Xi, Yi) S; (Wi, Zi).
However, for Cl(i) we cannot have i) and ii) at the same time. Therefore, if
condition Cl(i) holds, binary relation S; is a complete preorder on XiXXi
because it is transitive (due to implication in its definition) and strongly
complete (for Cl(i».
i) (Xi, a-i) S (Yi, b_ i) and not (Wi, a-i) S (yi, b_i): by definition of sf this
implies not Wi sf Xi;
ii) (Wi, C_i) s (Zi, d_ i) and not (Xi, C_i) S (Zi, d_ i): by definition of sf this
implies not Xi sf Wi·
However, for C2(i) we cannot have i) and ii) at the same time. Therefore, if
condition C2(i) holds, binary relation sf is a complete preorder on Xi
because it is transitive (due to implication in its definition) and strongly
complete (for C2(i)).
Analogously, condition C3(i) says that marginal outranking relation sf
(i=l, ... ,n) is a complete preorder on Xi.
Condition C4(i), together with C2(i) and C3(i), ensures that the orderings
obtained from sf and sf are compatible, i.e. there is no Xi,YiEXi such that
Xi sf Yi and not Yi sf Xi (i.e. Xi is preferred to Yi with respect to Sf), and not
Xi sf Yi and Yi S~ Xi (i.e. Yi is preferred to Xi with respect to S~), for i=l, ... ,n.
In fact, sf and S~ would not be compatible if there were XhYj,Zj,WjEXj,
a_i,b_i,c_hd_iEX_i such that
i) (Xi, a_i) S (Wi, b_i) and not (Yi, a-i) S (Wi, b_i): by definition of sf this
implies not Yi sf Xi, and for completeness of sf (due to C2(i)), also
Xi sf Yi;
ii) (Zi, C-i) s (Xi, d_ i) and not (zj, C_i) S (yi, d_ i): by definition of S~ this
implies not Xi s~ Yi, and for completeness of S~ (due to C3(i)),
also Yi S~ Xi.
However, for C4(i) we cannot have i) and ii) at the same time.
The information given by outranking relations sf and S~ can be
synthesized by the binary relation Si= sf n s~ . Analogously to sf and S~ ,
the relation XjSiYi reads: "Xi is at least as good as yi". Transitivity of sf and
S~ ensures transitivity of Si. Compatibility of the orderings of sf and S~
ensures completeness of Si. Since transitivity of sf and S~ is satisfied by
276 AIDING DECISIONS WITH MULTIPLE CRITERIA
if Xii p~hil Yil and ... Xid p~hid Yid and Xid+ ISid+ Irid+ 1 and ... XieSierie and
Sie+ISie+lYie+1 and ... SjfSifYif, then xSY,
where
- strength of preference can be defined for criteria gil, ... ,gid, while not
for criteria gid+h" .,gif,
- {hil, ... ,hid)eHilx ... xHid, (rid+I, ... ,rie)eXid+lx ... XXie, (Sie+l, ... ,sif)eXie+1
x ... XXif, {id+ 1, ... ,ie} and {ie+ 1, ... ,if} are not necessarily disjoint;
An example of a D>-decision rule: "if X is (at least) better than yon
criterion i and x is at least medium on criterion j and Y is at most
medium on criterion k, then x is at least as good as y";
2) D=:;;-decision rule, called "at most decision rule", being a statement of the
type:
if Xii pfthil Yil and ... XidP~hid Yid and rid+lSid+lXid+l and ... rieSiexie and
Yie+ISie+lSie+1 and ... YitSitSif, then not xSY,
where
- strength of preference can be defined for criteria gil, ... ,gid, while not for
criteria gid+h ... ,gif,
Conjoint Measurement and Decision Rule Model 277
- (hib ... , hid)EHiI X ... X Hid, (rid+b ... , rie)EXid+1 x X X ie, (Sie+b ... ,
Sif) EX ie+IX ... XXif, {id+ 1, ... ,ie} and {ie+ 1, ... ,if} are not necessarily
disjoint;
An example of a D:$-decision rule: "if x is (at most) worse than y on
criterion i and x is at most medium on criterion j and y is at least
medium on criterion k, then x is not at least as good as y";
3) D~:$-decision rule, called "at least-at most decision rule", being a
statement of the type:
if XiI PfIh il Yil and ... Xid p~hid Yid and Xid+1 P~~i+1 Yil and ... Xie p~hie Yie and
Xie+ISie+lrie+l and ... XirSitfif and rif+1Sif+lXif+l and ... rigSigXig and
Sig+ISig+IYig+1 and ... SioSioYio and Yio+lSio+ISio+l and ... YipSipsip,
then xSy or not xSy,
where
- strength of preference can be defined for criteria gib ... ,gie, while not
for criteria gie+b·· .,gip,
- (hqb ... , hqe)EHql x ... x Hqe, (rie+b ... , rif)EX ie+1 x ... x Xif, (rif+b ... ,
rig)EX if+1 x x Xig, (Sig+b ... , sip) EXig+1 X X Xio,
(Sio+b ... ,Sip)EXio+IX ... XXip, {il, ... ,id} and {id+l, ... ,ie}, {ie+l, ... ,if} and
{if+l, ... ,ig}, {ig+l, ... ,io} and {io+l, ... ,ip} are not necessarily
disjoint, respectively.
An example of a D~:$-decisionrule: "if X is (at most) worse than Y on
criterion i and at least better on criterion j, then there are no sufficient
reasons to conclude that x is, or is not, at least as good as y".
Wid+1Sid+lrid+l and ... and WieSierie and Sie+1Sie+1Zie+l and ... and SitSitZif, while
not w S z. For example, a pair of actions (w,z) such that
- w is better than z on criterion i,
- w is medium on criterionj,
- z is medium on criterion k
- w is not at least as good as z,
contradicts decision rule "i/x is (at least) better than y on criterion i and x is
at least medium on criterion j and y is at most medium on criterion k, then x
is at least as good as y" .
Set R"? of D"?-decision rules is non-contradictory iff there is no rule
contradicted by a pair (W,Z)EXXX.
We say that set R"? of D"?-decision rules represents the outranking
relation S on X iff it is complete and non-contradictory.
Let us remark that decision rules describe DM's preferences in terms of
"condition profiles" relative to pairs of actions. As the condition profiles
concern, in general, subsets of criteria, they are called partial profiles. They
involve two specific marginal relations: a marginal outranking relation Sj, for
all criteria, and a marginal relation S;, for comparison of "preference
strength" on criteria permitting definition of the strength of preference.
These marginal relations are used to define a specific dominance relation
with respect to pairs of actions, such that pair (x,y) dominates pair (w,z) if
action x is preferred to y at least as much as action w is preferred to z for
each considered criterion. Thus, the decision rule model R"? involves partial
profiles for pairs of actions and a specific dominance relation.
Analogous definitions hold for set R5, ofD5,-decision rules.
Theorem 2.1 (Greco, Matarazzo and Slowinski 2000b) The following four
propositions are equivalent:
1) the binary relation S on X satisfies properties C2, C3, C4 and Cl(i) for
i=I, ... ,k, 05,k5,n
2) there exist
• function G:Rk+2(n-k)-+R, non-decreasing in the first k arguments, non-
decreasing in each (k+s)-th argument with s odd and non-increasing in
each (k+s)-th argument with s even, where s=I, ... ,2(n-k),
• functions 'I'i:R2~ R, for all i= 1, ... ,k, non-decreasing in the first
argument and non-increasing in the second argument,
Conjoint Measurement and Decision Rule Model 279
• functions Ui:Xi~R,for all i=l, ... ,n,
such thatfor all x,yeX
G['Pi[Ui(Xi), Ui(Yi)], i=I, ... ,k> Ui(Xi), Ui(Yi), i=k+I, ... ,n] ~ 0 iff xSy;
3) there exist
• for all i=l, ... ,n, a marginal outranking relation Si being a complete
preorder on Xi>
• for all i=l, ... ,k, a binary relation S; being a complete preorder on
XiXXi>
• one set ofD~-decision rules representing S;
4) there exist
• for all i= I, ... ,n, a marginal outranking relation Si being a complete
preorder,
• for all i=l, ... ,k, a binary relation S; being a complete preorder,
• one set ofD<-decision rules representing S.
Proof. First, we prove that I) implies 2). Since C2, C3 and C4 hold, then Si
is a complete preorder for each i=l,oo .,n. Therefore, there exists a function
Uj:Xi~R such that XiSiYi iffui(xi)~ui(Yi)' Since CI(i) holds for criteria l,oo.,k,
then binary relation S;, for i=l, ... ,k, is a complete preorder. Therefore, there
exists a function <l>i:XiXXi~R such that (Xi,yi) S; (wj,Zj) iff
<l>i(Xj,Yi)~<l>i(Wj,Zi)' Taking into account the coherence of S; with Sj, we can
conclude that there exists a function 'Pj:R2~R, non-decreasing in the first
argument and non-increasing in the second argument, such that
<l>i(Xi,yi)='Pi[Uj(Xi), Ui(Yi)]. On the basis of binary relations S;, i=l,oo.,k, and
Sj, i=k+ 1,00' ,n, a dominance relation D can be defined on XxX as follows:
for each (x,y),(w,z)eXxX:
(x,y)D(w,z) iff (xj, yj) S; (wj, Zi) for each i=l,oo .,k, and Xj Sj Wi and Zi Si Yi
for each i=k+l,oo.,n.
The dominance relation D can be defined on XxX also In terms of
functions Uj and 'Pj, as follows: for each (x,y),(w,z)eXxX
(x,y)D(w,z) iff 'Pj[uiCxj), Ui(yj)]~'Pj[Uj(Wi)' Uj(Zj)], for each i=l, ... ,k, and
Uj(Xj)~Ui(Wi) and Uj(Zj)~Uj(Yi) for each i=k+l,oo.,n.
280 AIDING DECISIONS WITH MULTIPLE CRITERIA
On the basis of b), this means that there exists function V:Rk+2(n-k)~R,
such that
V['Pi[Ui(Wj), Ui(Zi)]oi=I, ... ,k. Ui(Wi), Ui(Zi), i=k+l, ... ,n] <
min V['Pi[Ui(Si), ulti)] , i=I, ... ,k, Ui(Si), Ui(ti), i=k+l, ... ,n]'
sSt
Given the function V ['Pi[UlXi) , Ui(Yi)], i=I, ... ,k, Ui(Xi), UlYi), i=k+l, ... ,n] we can
define function G:Rk+2(n-k)~R as follows
G['Pi[ Ui(Wi), Ui(Zi)], i=I, ... ,k. Ui(Wi), Ui(Zi), i=k+1, ... ,n] =
V ['Pi[UlWi) , Ui(Zi)], i=I, ... ,k, Ui(Wi), Ui(Zi), i=k+l, ... ,n] - Q
where Q = min V['Pi[ulsi), Ui(ti)]oi=I, ... ,k, Ui(Sj), Ui(ti), i=k+l, ... ,nl
sSt
if Xl Pfh l Yl and ... XkP~hk Yk and Xk+lSk+lrk+l and ... xnSnrn and Sk+lSk+lYk+l
and ... snSnYn, then xSy,
if Xl pfh l Yl and ... XkP~hk Yk and rk+lSk+lXk+l and ... rnSnxn and Yk+lSk+lSk+l
and ... YnSnsn, then not xSy,
Let us remark that the sets of D~-decision rules and D:s;-decision rules
considered above in the proof of 1) => 3) and 1) => 4), respectively, are not
unique. There is a decision rule for each pair (W,Z)EXXX, so these sets are
maximal in the sense that they contain all rules that can be defined with
complete profiles (all criteria are considered). Practically, much more
synthetic representations can be considered, involving less rules and partial
Conjoint Measurement and Decision Rule Model 283
profiles. The authors proved that the minimal (i.e. the most synthetic)
representation of outranking relation S by decision rules is unique (Greco,
Matarazzo and Slowinski 2000b).
Two remarks are necessary with respect to Theorem 2.1 :
1. As already pointed out in the introduction, proposition 2) of Theorem 2.1
is a bit more general than the corresponding proposition formulated by
Bouyssou and Pirlot (1997). Theorem 2.1 assumes that condition C1(i) is
satisfied for some criteria (i=1, ... ,k) and not satisfied for the other
(i=k+ 1, ... ,n), while Bouyssou and Pirlot consider that C 1(i) is satisfied,
for all criteria (i=1, ... ,n). This means that Theorem 2.1 permits to
distinguish between:
- criteria on which it is possible to measure preference strength by
function 'I'i (these are the criteria for which property C1(i) is satisfied)
- criteria on which it is not possible to measure preference strength by
function 'I'i (these are the criteria for which property Cl(i) is not
satisfied).
Consequently, the functional representation of S proposed in Theorem 2.1
IS
G['I'i[ Ui(Xi), Ui(Yi)], i=I, ... ,k, UlXi), UiCYi), i=k+I, ... ,n] ~ 0 iff xSy
while the functional representation of S proposed by Bouyssou and Pirlot
(1997) is
the last column of Table 2.1 (in the head of the column,
G(x,y)=G ['I' I(UI(XI),UI(YI»,U2(X2),U2(Y2)]). The outranking relation S is
represented by the function G as follows:
G['I'I(UI(XI),UI(YI»,u2Cx2),U2(Y2)] ~ 0 <=> xSy,
G['I'I(UI(XI),UI(YI»,U2(X2),U2(Y2)] < 0 <=> not xSy,
where
• 'I'1(UI(XI),UI(YI» represents the strength of preference with respect to
price of cars x and y, equal to Xl and Yb respectively (UI(XI) and UI(YI)
mean the utility of Xl and Yb respectively),
• U2(X2) and U2(Y2) represent the utility of X2 and Y2 equal to the speed of
cars X and y, respectively.
Finally, according to proposition 3) of Theorem 2.1, the outranking
relation from Table 2.1 can be represented by the following set of D~-
decision rules:
# 1) "if x is at least indifferent to Y on the price and the speed of x is at
least medium and the speed of Y is at most medium, then x is at least
as good as y";
#2) "if x is (at least) better than y on the price and the speed ofx is at least
medium, then x is at least as good as y";
#3) "if x is at least indifferent to y on the price and the speed of x is (at
least) high, then x is at least as good as y";
#4) for all uncovered pairs of cars (x,y), x is not at least as good as y.
Let us observe that, according to proposition 4) of Theorem 2.1, the
outranking relation from Table 2.1 can also be represented by a set of D::;-
decision rules:
#5) "if x is (at least) worse than y on the price, then x is not at least as
good as y",
#6) "if the speed ofx is (at most) low, then x is not at least as good as y",
#7) "if x is at most indifferent to y on the price and the speed of x is at
most medium and the speed ofy is (at least) high, then x is not at least
as good as y",
#8) for all uncovered pairs of cars (x,y), x is at least as good as y.
The identity numbers of rules matching the corresponding pair of cars are
indicated in the column "Outranking" of Table 2.1.
286 AIDING DECISIONS WITH MULTIPLE CRITERIA
Theorem 3.1 For any reflexive relation S on X, as well as for any set of
graded preference relations P~i , i= 1, ... ,k, O:S;k$;n, and for any outranking
relation Si being a complete preorder on Xj, i=k+ 1, ... ,n, there exist
• functions ui:Xr-~R, i=I, ... ,n, such that for each Xj,YiEXi, Ui(Xi) ~ Ui(Yi)
implies XiSiYj,
• functions <l>i:XiXXi ~R, i=I, ... ,k, such that for each Xj,Yi,Wj,ZiEXj,
<l>i(Xi,yi) ~ <l>i(Wj,Zi) implies Xi PP Yi and Wi pfi Zi with ri~sj,
Proof. Since Si is a complete preorder on Xj, i=k+ 1, ... ,n, there exist
functions Ui:Xi~R such that, for each Xj,YiEXj, Ui(Xi)~Ui(Yi) implies XiSYi.
Moreover, for each Xj,YiEXj, i=I, ... ,k, let us set <l>i(xj,Yi)=h if XiP~Yi'
Therefore, for each x,y,w,zEX, (x,y)D(w,z) can also be expressed by
<l>i(Xj,Yi)~<I>i(Wj,Zi)' i=I, ... ,k, and UlXi)~Ui(Wi) and Ui(Zi)~Ui(Yi)' i=k+l, ... ,n.
with at least one inequality strict. Thus, on the basis of a), G[<I>i(Xi,yi), i=I, ... ,k,
UlXi), Ui(Yi), i=k+I, ... ,n) is increasing in the first k arguments, increasing in each
(k+s)-th argument with s odd and decreasing in each (k+s)-th argument with
s even, where s=1, ... ,2(n-k). Therefore, function G satisfies the
monotonicity condition required by the theorem.
If (X,Y)E~ and (w,z)ENSuB, then due to condition 2) in the definition of
T, (x,y)T(w,z) and not (w,z)T(x,y), because otherwise, as already shown, we
should have <l>i(Xi,yi)=<I>i(Wi,Zi), for i=I, ... ,k, Ui(Xi)=Ui(Wi) and Ui(Zi)=Ui(Yi), for
i=k+ 1, ... ,n, which contradicts CCDA. This implies that for each x,y,w,zEX,
if(x,Y)E~ and (w,z)ENSuB, then h(x,y»h(w,z) and thus
One can prove analogously that for each (W,Z)EXXX, such that
(w,z)E~uB
G[<I>·(w-
1 h z·) '-1 , ...• k' u·(w·)
t ,1- 1 1, U·(7.) '-k+ 1,... ,n) >
I "'I , 1-
max G[<I>i(rj,ti), i=I, ... ,k, Ui(rj), Uj(ti), i=k+I, ... ,n).
(r,t)eNS
Therefore, if we set
tl = max G[<I>i(ri'~)' i=I, ... ,k, ui(ri), Ui(tj), j=k+I, ... ,n) and
(r,t)eNS
rules and a set R~'5, of D~'5,-decision rules. Set R= {R~, R'5" R~'5,} of decision
rules is complete iff
• each pair (X,Y)EXXX, such that (X,Y)ES, is covered by at least one D~
decision rule belonging to R~,
• each pair (X,y) EXXX, such that (x,Y)EB, is covered by at least one
D~'5,-decision rule belonging to R~'5,.
Theorem 3.2 For any reflexive relation S on X, for any set of graded
preference relations pri ,
i= 1, ... ,k, and for any outranking relation Si being
a complete preorder on Xi. i=k+l,,,.,n, there exists a set R={R~, R'5" R~'5,}
of decision rules representing the outranking relation S on X ..
Proof. For each pair (W,Z)ES one can build the following D~-decision rule:
if Xl p~hl Yl and ... XkP~hk Yk and Xk+1Sk+lrk+l and ... xnSnrn and
Sk+1Sk+1Yk+l and ... SnSnYn, then xSy,
where (WhZl)E prl '''., (Wk,Zk)E p~k , Wk+l=rk+lo Wn=rn, Zk+l=Sk+h Zn=Sn.
For each pair (w,z)ENS one can build the following D'5,-decision rule:
if Xl pfh l Yl and ... XkP~hk Yk and rk+1Sk+1Xk+l and ... rnSnxn and Yk+1Sk+1Sk+l
and ... YnSnsn, then not xSy,
292 AIDING DECISIONS WITH MULTIPLE CRITERIA
Let us remark that the set R={R~, R~, R~~}of decision rules considered
in the proof of Theorem 3.2 is not unique. Practically, much more synthetic
representations can be considered, involving less rules and partial profiles.
The authors proved that the minimal (i.e. the most synthetic) representation
of rough approximation £, NS and B by decision rules is unique (Greco,
Matarazzo and Slowinski 2000b).
- U2(X2) and Uz(Y2) represent the utility of X2 and Y2 equal to the speed of
cars x and y, respectively.
The thresholds tb t 2, are set on the values t l =-4, tz=4, and the
approximations of outranking relations Sand SC are represented as follows:
G[<fl I(XbYI),U2(X2),U2(Y2)] ~ 4 <:::> (X,Y)E~ (i.e. "x is at least as good as y
without hesitation"),
G[<flI(XbYI),Uz(X2),U2(Y2)] ~ -4 <:::> (x,Y)ENS (i.e. "x is not at least as good
as Y without hesitation"),
-4 < G[<fl I(XbYI),U2(X2),U2(Y2)] < 4 <:::> (x,Y)EB (i.e. "x is at least as good as
y, or x is not at least as good as y, with some hesitation between the two
possibilities").
The outranking relation from Table 3.1 can be represented by the
following set ofD~-, D~- and D~~-decision rules:
#1) "if x is (at least) better than y on the price, then x is at least as good as y",
#2) "if x is at least indifferent to y (i.e. indifferent or better) on the price
and the speed of y is at most medium, then x is at least as good as y ",
#3) "if x is (at least) worse than y on the price and the speed ofx is at most
medium, then x is not at least as good as y",
#4) "if x is at most indifferent to Y (i.e. indifferent or worse) on the price
and the speed of x is at most medium and the speed of y is (at least)
high, then x is not at least as good as y",
#5) "if x is (at least) worse than y on the price and the speed of x is (at
least) high, then x is at least as good as y or x is not at least as good as y",
#6) "if x is at most indifferent to y (i.e.
indifferent or worse) on the price
and the speed of x and Y is (at least) high, then x is at least as good as
y or x is not at least as good as y".
The identity numbers of rules matching the corresponding pair of cars are
indicated in the column "Outranking" of Table 3.1.
294 AIDING DECISIONS WITH MULTIPLE CRITERIA
Table 3.1 Inconsistent outranking on all 27 cases of possible evaluations of pairs of cars
Lowerappx.
Pairs
Price Speed ofx Speed ofy Outranking and G(x,y)
of cars
boundary
PI x is worse than y low high not xSy (#3,4) NS -II
P2 x is indifferent to y low high not xSy (#4) NS -5
P3 x is better than y low high xSy (#1) S 7
P4 x is worse than y medium high not xSy (#3,4) NS -7
P5 x is indifferent to y medium high not xSy (#4) NS -4
P6 x is better than y medium high xSy (#1) S 10
P7 x is worse than y high high xSy (#5,6) B -3
P8 x is indifferent to y high high not xSy (#6) B 0
P9 x is better than y high high xSy(#I) S 13
PIO x is worse than y low medium not xSy (#3) NS -7
PII x is indifferent to y low medium xSy (#2) S 4
PI2 x is better than y low medium xSy (#1,2) S 10
P13 x is worse than y medium medium not xSy (#3) NS -5
PI4 x is indifferent to y medium medium xSy (#2) S 8
PI5 x is better than y medium medium xSyi#I,2) S 12
PI6 x is worse than y high medium xSy(#51 B -I
PI7 x is indifferent to y high medium xSy (#21 S 12
PI8 x is better thany_ high medium xSy (#1,2) S 14
PI9 x is worse than y low low not xSy (#3J NS -5
P20 x is indifferent to J' low low xSy (#2) S 10
P21 x is better than y low low xSy(#1,2) S 13
P22 x is worse than y medium low not xSy(#3) NS -4
P23 x is indifferent to y medium low xSy (#2) S 12
P24 x is better than y medium low xSy (#1,2) S 14
P25 x is worse than y high low not xSy (#5) B 0
P26 x is indifferent to y high low xSy(#2) S 14
P27 x is better than y high low xSy(#1,2) S 15
• xSy means that there are some arguments in favor of the statement "x
is at least as good as y", i.e. there are some arguments in favor of the
outranking ofx over y,
• xScy means that there are some arguments against the statement "x is
at least as good as y", i.e. there are some arguments against the
outranking ofx over y.
While the only minimal requirement imposed on S is its reflexivity, the
only minimal requirement imposed on SC is its irreflexivity, i.e. for each
XEX, not xScx.
As the arguments for outranking and for the negative outranking may
come from different sources, the relations Sand SC are not
complementary. Thus, in general, not S 7:- Sc. In consequence, the four
following situations are possible:
• uSv and not uScv, that is true outranking (denotation uSTv),
• uScvand not uSv, that isfalse outranking (denotation uSFv),
• uSv and uScv, that is contradictory outranking (denotation uS~),
• not uSv and not uScv, that is unknown outranking (denotation uSuv).
The above four situations together constitute the so-called four-valued
outranking (see Tsoukias and Vincke 1995, 1997). It has been
introduced to underline the presence and the absence of positive and
negative reasons for the outranking. Moreover, it enables to distinguish
contradictory situations from unknown ones.
2. In Theorems 3.1 and 3.2 it is assumed that there exists a set of graded
preferences P~i with respect to each criterion i for which it is possible to
define a preference strength and one marginal outranking relation Sj with
respect to each criterion for which it is not possible to define a
preference strength. In this section, the graded preferences P~i and/or
the marginal outranking relation Sj, for i=I, ... ,n, do not pre-exist but
must be constructed starting from Sand Sc. To permit an appropriate
construction, Sand SC have to satisfy some specific conditions, in a
sense analogous to conditions Cl(i) - C4(i) above.
Considering comprehensive negative outranking binary relation SC, two
marginal outranking relations sjk and S~c, can be defined with respect to
each gi, i=l, ... ,n, as follows:
296 AIDING DECISIONS WITH MULTIPLE CRITERIA
Xi sfC Yi iff for all a_iEX_i and ZEX, [(Xi. a-i) SC Z => (Yi. a_i) SC z],
Xi S~C Yi iff for all a_iEXi and ZEX, [z SC (Yi, a-i) => Z SC (Xi. a_i)].
The relation (Xi. Yi) S~c (Wi. Zj) reads: "on the basis of negative
outranking SC, the strength of preference of Xi over Yi is at least as great as
the strength of preference of Wi over zi" .
Due to the implication in its definition, binary relation S~c is transitive.
While S~ is coherent with sf and S~ , S~c is coherent with S~ and S~c , i.e.
Xi sfC Yi implies [(Yi. Wi) S~C (Zi. Ui) => (Xj, Wi) S~C (Zj, Ui)],
Xi S~C Yi implies [(Wi. Xi) S~C (Zi. Ui) => (Wi. Yi) S~C (Zj, Ui)].
Representation of an outranking relation S and a negative outranking
relation SC proposed below is based on the following set of cancellation
properties:
la) for all i=l,oo .,k, for all Xi.Yi.Zi.WiEXi, a_i.b_i,c_i,d_iEX_i we have
Cl(i) [(xj, a-i)S(Yi, b_ i) and (Wi. C_i)S(Zj, d_ i)] =>
[(wj, a_i)S(zj, b_i) or (Xi. C-i)S(yi. d_ i)];
Cl'(i) [(xj, a_i)SC(yj, b_i) and (Wi,C_i)SC(Zj, d_ i)] =>
[(Wi. a_i)SC(Zj, b_i) or (xj, C_i)SC(yj, d_ i)];
Cl "(i) [(xj, a-i)S(Yi, b_i) and (Xi. C_i)SC(yi. d_ i)] =>
[(Wi. a_i)S(zj, b_i) or (Wi. C_i)SC(Zj, d_ i)];
Conjoint Measurement and Decision Rule Model 297
1') condition C1 '(i) ensures that, on the basis of Sc, the relation S~c
(i=1, ... ,k) is a complete preorder on XiXX i and it is fully meaningful to
speak about the strength of preference; otherwise the strength of
preference is not meaningful for criterion i,
2') condition C2'(i) says that marginal outranking relation src (i=1, ... ,n) is a
complete preorder on Xi,
3') analogously, condition C3'(i), says that marginal outranking relation S~c
(i=1, ... ,n) is a complete preorder on Xi>
298 AIDING DECISIONS WITH MULTIPLE CRITERIA
4') condition C4'(i), ensures for i=l, ... ,n, that the orderings obtained from
src and Sr c are compatible, i.e. there is no Xj,YiEXi such that XiSrc Yi
and not Yi src Xi (i.e. Xi is preferred to Yi on the basis of Src), and not
Xi Sr Yi and Yi Sr Xi (i.e. Yi is preferred to Xi on the basis of Sr
c c C ).
To show that properties based on Cl '(i), C2'(i), C3'(i) and C4'(i) with
respect to SC can be proved analogously as corresponding properties based
on, Cl(i), C2(i), C3(i) and C4(i) with respect to S, let us consider, as an
example, condition Cl '(i). It ensures that, on the basis of SC, the relation s~c
(i=l,oo .,k) is strongly complete: in fact, S~c would not be strongly complete
if there were xj,Yj,Zj,WiEXj, a_j,b_j,c_j,d_iEX_ i such that
ii) (Wj,C_i) sc (Zj,d_i) and not (Xi,C-i) Sc (Yj,d_ i): by definition of S~c this
implies not (Wj,Zi) S~c (xj,yJ
However for Cl '(i) we cannot have i) and ii) at the same time.
Therefore, if condition Cl '(i) holds, binary relation S~c, is a complete
preorder, because it is transitive (as pointed out before, by definition) and
strongly complete (for Cl '(i».
sr
Conditions Cl"(i), C2"(i) and C3"(i) ensure coherence between relations
S; , sf, from one side and relations S~c , src , Sr c from the other side.
More precisely,
1")Cl "(i) says that S~ and S~c are compatible, i.e. there is no Xi,yj,Wj,ZiEXi
such that (Xi,yi) S~ (Wj,Zi) and not (Wj,Zi) S~ (Xi,yi) (i.e. Xi is preferred to Yi
more than Wi is preferred to Zi on the basis of S~), and (Wj,Zi) S~c (Xi,Yi)
and not (Xi,yi) S~c (Wj,Zi) (i.e. Wi is preferred to Zj more than Xi is preferred
to Yi on the basis of S~c ),
2")C2" (i) says that sf and src are compatible, i.e. there is no Xj,YiEXi such
that Xi sf Yi and not Yi sf Xi (i.e. Xi is preferred to Yi on the basis of sf),
and Yi src Xi and not Xi sfC Yi (i.e. Yi is preferred to Xi on the basis of src ),
Conjoint Measurement and Decision Rule Model 299
3")C3"(i) says that S~ and S~c are compatible, i.e. there is no Xj,YiEXi such
that Xi S~ Yi and not Yi S~ Xi (i.e. Xi is preferred to Yi on the basis of S~),
and YiS~cXi and not XiS~cYi (i.e. Yi is preferred to Xi on the basis of
S~C).
5) C5(i) says that sf and S~c are compatible, i.e. there is no Xj,YiEXi such
that Xi sf Yi and not Yi sf Xi (i.e. Xi is preferred to Yi on the basis of sf),
and Yi S~c Xi and not Xi S~c Yi (i.e. Yi is preferred to Xi on the basis of
S~C)
1 ,
5') C5'(i) says that S~ and sfk are compatible, i.e. there is no Xj,YiEXi such
that Xi S~ Yi and not Yi S~ Xi (i.e. Xi is preferred to Yi on the basis of S~),
and Yi sfk Xi and not Xi sfk Yi (i.e. Yi is preferred to Xi on the basis of sfk ).
If the marginal outranking relations sf, s~, sfk and S~c are complete
preorders (i.e. if conditions C2(i), C3(i), C2'(i), C3'(i) hold) and are all
pairwise-compatible (i.e. if conditions C2"(i), C3"(i), C4(i), C4'(i), C5(i),
C5'(i) hold), then binary relation si=sf ns~ nSfk nS~c is a complete
pre order. In fact, Si is transitive because it is obtained as intersection of
transitive binary relations. Moreover, Si is strongly complete because
relations sf, s~ ,sfk ,s~c are strongly complete and pairwise-compatible.
Remark that Si would not be strongly complete in one of the two following
cases:
i) if there was KE {Sf ,s~ ,sfk ,S~C} and Xi,yiEXi such that not XiKYi and
not YiKxj, but this is impossible because sf, s~, sfk and S~c are all
strongly complete;
ii) ifthere were K),K 2 E {Sf ,s~ ,sfk ,s~c } and Xj,YiEXi such that not XiKIYi
and not YiK2Xi, but for the strong completeness of sf, s~ , sfk , S~c and,
300 AIDING DECISIONS WITH MULTIPLE CRITERIA
• each pair (X,y)EXXX, such that xSy, is covered by at least one D~
decision rule belonging to R~,
• each pair (X,y) EXXX, such that xSCy, is covered by at least one D~
decision rule belonging to R~.
Set R={R~, R~} of decision rules is non-contradictory iff
3) there exist
• two binary relations S~ and S~c on XiXX i for each i= 1, ... ,k, being
complete preorders, which are mutually compatible such that also
S~ ( l S~c is a complete preorder and from its equivalence classes a set
of graded preference relations pri on Xi can be defined,
• a set R={R~, R~} of decision rules representing the outranking
relation S and the negative outranking relation SC.
Proof. First we prove that 1) ~ 2). Since Si is a complete pre order for each
i=l, ... ,n, then there exists a function Ui:Xi~R such that, for each Xi,yiEXj,
XiSiYi if and only if UiCXi)~ulYi)' Moreover, since also binary relation
S~ ( l S~c , for i= 1, ... ,k, is a complete pre order coherent with binary relation
Sj, then there exists a function \}'i:R2 ~ R non-decreasing in the first
argument and non-increasing in the second argument such that
(Xj,Yi) S~ ( l S~c (Wj,Zi) if and only if \}'lUi(Xi),Ui(Yi»~\}'i(Ui(Xi),ulYi»' On the
302 AIDING DECISIONS WITH MULTIPLE CRITERIA
basis of binary relations S~ n S~c , i= 1, ... ,k, and Sj, i=k+ 1, ... ,n, a dominance
relation D can be defined on XxX as follows: for each (x,y),(w,Z) EXXX
(x,y)D(w,z) iff (Xi,yi)S~ nS~c (Wj,Zi) for each i=I, ... ,k, and Xi Si Wi and
ZiSiYi for each i=k+ 1, ... ,no
The definition of the binary relations S~ nS~c, i=I, ... ,k, and Si,
i=k+ 1, ... ,n, implies the following Coherence Conditions with respect to
Dominance for the outranking relation S (CCDS):
there is no x,y,w,zEX such that (x,y)D(w,z), and wSz and not xSy,
Analogously, there exists the following Coherence Conditions with
respect to Dominance for the negative outranking relation SC (CCDSC):
there is no x,y,w,zEX such that (x,y)D(w,z), and not wScz and xScy.
As far as representation of outranking relation S is concerned, on the
basis of dominance relation D and CCDS, function G:Rk+2(n-k)~R,
increasing in the first k arguments, increasing in each (k+s)-th argument with
s odd (s=I,3, ... ,2(n-k)-I) and decreasing in each (k+s)-th argument with s
even (s=2,4, ... ,2(n-k)), s=I, ... ,2(n-k), can be build in a way described in
Theorem 2.1, such that
G['Pi(Ui(Xi),Ui(Yi)), i=l, ... ,b Ui(Xi), Ui(Yi), i=k+l, ... ,n] ~ 0 iff xSy.
Let us concentrate, therefore, on representation of negative outranking
relation SC. To this end, let us consider the following binary relation T on
XxX, defined as follows: (x,y)T(w,z) if at least one of the two following
conditions is satisfied:
1) (x,y)D(w,z),
2) not xScy and wScz.
On the basis of dominance relation D and CCDSc, one can prove that
binary relation T is reflexive and transitive, i.e. it is a partial pre order.
Therefore, there is a function hC:XxX~R such that, for each
(x,y),(W,Z)EXXX, (x,y)T(w,z) implies hC(x,y)~hC(w,z).
Taking into account functions 'Pj, i=I, ... ,k, and Uj, i=k+1, ... ,n, and
CCDS c, one can construct a function VC:Rk+2(n-k)~R, increasing in the first k
arguments, increasing in each (k+s)-th argument with s odd and decreasing
in each (k+s)-th argument with s even, where s=I, ... ,2(n-k), such that
VC['Pi[Ui(Xi), Ui(Yi)], i=l, ... ,k, UiCXi)' Ui(Yi), i=k+l, ... ,n] = hC(x,y). On the basisi of
CCDS c, the following property of discriminating capacity of functions hC
Conjoint Measurement and Decision Rule Model 303
and yc holds: for each x,Y,W,ZEX, if not xScy and wScz, then hC(x,y»hC(w,z)
and thus
yC['Pi[Ui(Xi), UlYi)], i=I, ... ,b Ui(Xi), Ui(Yi), i=k+l, ... ,n] >
yC['Pi[Ui(Wi), Ui(Zi)], i=I, ... ,k, Ui(Wi), Ui(Zi), i=k+l, ... ,n],
or, in other terms, for each (W,Z)EXXX such that wScz
yC['Pi[Ui(Wi), Ui(Zi)], i=I, ... ,k, Uj(Wj), uiCzj), j=k+l, ... ,n] <
min yC['Pj[Uj(Sj), Uj(tj)], j=I, ... ,k, Uj(Sj), ultj), j=k+l, ... ,n].
not sSC t
Given the function yC['Pj[Uj(Xj), Uj(Yj)], j=I, ... ,k, Uj(Xj), Uj(Yj), j=k+l, ... ,n], one
can define function GC:Rk+2(n-k)~R, as follows
GC['Pj[Uj(Wj), Uj(Zj)], j=I, ... ,b Uj(Wj), Uj(Zj), j=k+l, ... ,n] =
yC['Pj[Uj(Wj), Uj(Zj)], j=I, ... ,k. Uj(Wj), Uj(Zj), j=k+l, ... ,n] - Q
where Q = min yC['Pj[Uj(Sj), Uj(tj)].i=l, ... ,k, Uj(Sj), Uj(tj), j=k+l, ... ,n].
not sSC t
Similarly to function yc, function GC is increasing in the first k
arguments, increasing in each (k+s)-th argument with s odd and decreasing
in each (k+s)-th argument with s even, where s=I, ... ,2(n-k). Consequently,
function GCsatisfies the monotonicity properties specified in proposition 2).
Moreover, by definition and due to discriminating capacity of function
yc, function GCsatisfies the following property: for all (X,Y)EXXX such that
not xScy we have
GC['Pj[Uj(Wj), Uj(Zj)], j=l .... ,k, Uj(Wj), UlZi), j=k+l, ... ,n] ~ 0
and for all (W,Z)EXXX such that wScz we have
GC['Pj[Uj(Wj), Uj(Zj)], j=I, ... ,k. Uj(Wj), Uj(Zi), j=k+l, ... ,n] < O.
Thus, we. proved that 1) =>2).
Now let's prove that 1) => 3). As explained before, in view of
cancellation properties assumed in proposition 1), Sj is a complete preorder
on Xj, while s7 n s7c is a complete preorder on XjxXj. Since s7 n s7 c , for
i=I, ... ,k, is a complete preorder, its equivalence classes define a set of
graded preference relations. For each pair (W,Z)EXXX, such that wSz, the
following D~-decision rule can be built:
if Xl p~hl Yl and ... XkP~hk Yk and Xk+1Sk+lrk+l and ... xnSnrn and Sk+1Sk+1Yk+l
and ... snSnYn, then xSy,
where (Wt.Zl)E p~l , ... , (Wk,Zk)E p~k , Wk+l=rk+t. wn=rn, Zk+l=Sk+t. Zn=Sn.
304 AIDING DECISIONS WITH MULTIPLE CRITERIA
For each pair (W,Z)EXXX, such that wScz, the following Ds;-decision rule
can be built:
ifXt pfh 1 Yt and ... Xk p~hk Yk and rk+tSk+tXk+t and '" rnSnxn and Yk+tSk+tSk+t
and ... YnSnsn, then xScy,
where (WhZt)E pfl , ... , (Wk,Zk)E p~k , Wk+t=rk+h wn=rn, Zk+t=Sk+h Zn=Sn.
The set R={R-e., R::;;} of decision rules corresponding to all pairs
(W,Z)EXXX is obviously complete. Moreover, it is non-contradictory
because otherwise CCDS for D~-decision rules and CCDSc for Ds;-decision
rules would not hold. Therefore, this set R={R-e., R:s;} of decision rules
represents the outranking relation S and the negative outranking relation SC.
This completes the proof of I) => 3).
The proof of2) => I) and 3) => I) is simple and left to the reader. 0
Let us remark that the set R={R-e., R:s;}of decision rules considered in the
proof of Theorem 3.3 is not unique. There is a decision rule for each pair
(W,Z)EXXX, so these sets are maximal in the sense that they contain all rules
that can be defined with complete profiles (all criteria are considered).
Practically, much more synthetic representations can be considered,
involving less rules and partial profiles. The authors proved that the minimal
(i.e. the most synthetic) representation of outranking and negative outranking
relations Sand SC by decision rules is unique (Greco, Matarazzo and
Slowinski 2000b).
Let us remark that in Theorem 3.3 conditions C2"(i), C3"(i), C4(i), C4'(i),
C5(i), C5'(i) are supposed to be satisfied for all i=I, ... ,n. If this was not the
case, then we would have the following consequences in the representation
ofS and Sc:
I) As to the functional representation, one should consider four functions
Ui:Xi~R, Vi:Xi~R, uf :Xi~R, and vf :Xi~R for i=I, ... n, and functions
G['I'i(uiCxi), Vi(Yi», i=t, ... ,k> Ui(Xi), ViCYi), i=k+t, ....n] -e. 0 iff xSY
G C[ 'l'ic (uf (Xi), Vf (Yi», i=t, ... ,k, uf (Xi), vf (Yi), i=k+t .... ,n] < 0 iff xScy.
This means that in order to represent the outranking relation S and the
negative outranking relation SC for two actions X and y, we need function
UiCXi) defining the "value" of Xi and function ViCYi) defining the "value" of
Conjoint Measurement and Decision Rule Model 305
. Thus, if C2"(i), C3"(i), C4(i), C4'(i), C5(i), C5'(i) are not satisfied, a D~-
decision rule has the following syntax:
if Xii Pil~h·
l II Yil an d ... XidP id
~h·,d Yid an d Xid+l S#id+1 rid+l an d ... Xie S#ierie an d
d) pairs of actions (x,y) not covered by any rule whose consequent is "xSy"
and not covered by any rule whose consequent is "xSCy" are in relation
of unknown outranking xSUy; these are the pairs: P3, P6, P8, P9, PII,
P12.
Table 3.2 Four-valued outranking on all 27 cases of possible evaluaions of pairs of cars
4-valued
Pairs of Speed Speed
Price S S' outran- G(x,y) G'(x,y)
cars ofx Ofy
king
PI x is worse thany low high SC (#3,4) SF -15 -15
P2 x is indifferent to y low high SC (#4) SF -10 -6
P3 x is better than y low high SU -6 3
P4 x is worse than y medium high SC (#3,4) SF -8 -II
P5 x is indifferent to y medium high SC (#4) SF -6 -5
P6 x is better than y medium high SU -4 6
P7 x is worse than y high high SC (#3) SF -5 -7
P8 x is indifferent to 'I high high SU -4 6
P9 x is better than y high high SU -3 9
PIO x is worse than y low medium SC (#3) SF -6 -10
Pll x is indifferent to 'I low medium SU -4 0
P12 x is better than y low medium SU -3 6
PI3 x is worse than y medium medium S (#1) SC (#3) SK 0 -8
P14 x is indifferent to 'I medium medium S (#1) ST 4 4
P15 x is better than y medium medium S (#1) ST 8 8
Pl6 x is worse than y high medium S (#1) SC (#3) SK 6 -6
PI7 x is indifferent to y high medium S (#1) ST 8 8
P18 x is better than y high medium S (#1) ST 10 10
P19 x is worse than y low low SC (#3) SF -3 -7
P20 x is indifferent to 'I low low S (#2) ST 6 6
P2l x is better than y low low S (#2) ST 9 9
P22 x is worse than y medium low S (#1) SC (#3) SK 6 -6
P23 x is indifferent to 'I medium low S (#1,2) ST 8 8
P24 x is better than y medium low S (#1,2) ST 10 10
P25 x is worse than y high low S (#1) SC (#3) SK 9 -5
P26 x is indifferent to )< high low S (#1,2) ST 10 10
P27 x is better than y high low S (#1,2) ST 11 11
if Xl pfO Yl and X2 P~o Y2 and ... and Xn-2 P~~2 Yn-2 and Xn-l P~~l Yn-!, then xSy,
1if Xl PI>0 Yl and X2 P>0
2 Y2 an d .. .an d Xn-l Pii-l
>0 Yn-l an d Xn Pil
>0 Yn, t h en XSy.
such that {il,i2, ... ,ip}c;;{1,2, ... ,n}, (Wil+ Wi2+ ... + wip)/IJ=1 w j ~ s.
Let us observe that, in comparison with the lexicographic aggregation,
not all the previous decision rules are necessary. Indeed, we can obtain a
representation of the binary relation S on X by majority aggregation
considering only those decision rules that do not involve any i=i1,i2, ... ,ip for
which (Wil+ Wi2+ ... + Wip-Wi)/IJ=IWj ~ s.
if XiI Pn~o Yil an d Xi2 P i2~o Yi2 and ... Xip Pip~o Yip and Xkl P kl~-I Ykl an d Xk2
l
such that {il,i2, ... ,ip }u{kl,k2, ... ,kq}={1,2, ... ,n}, (WiI+Wi2+ ... +Wip)/
IJ=I Wj ~ s.
Also in this case not all the previous decision rules are necessary, i.e. we
can obtain a representation of the binary relation S on X by ELECTRE I
considering only those decision rules that involve sets {i 1,i2, ... ,ip} including
no i=i1 ,i2, ... ,ip for which (WiI+ Wi2+ ... + Wip-Wi)/ IJ=1 W j ~ s.
lif XiI P il~I YiI an d ... XipP ip~I Yip and Xjl P~o
jl Yjl an d ... XjgP ~o
jg Yjg
and Xkl P~ll Ykl and ... XkqP~~1 Ykq, then xSy,
such that {il,i2, ... ,ip }u{jl, ... jg}u{kl,k2, ... ,kq}={1,2, ... ,n},
S(WiI+Wi2+ ... +Wip) ~ Wkl+Wk2+ ... +Wkq.
Conjoint Measurement and Decision Rule Model 311
5. Conclusions
This paper shows the equivalence of preference representation by
numerical functions and by "if. .. , then ... " decision rules in multicriteria
choice and ranking problems. The decision rule model representing a
comprehensive preference relation involves partial profiles defined for
subsets of criteria plus a dominance relation on these profiles and pairs of
actions. Moreover, we considered representation of hesitation in preference
modeling. Within this context, two approaches were considered: dominance-
based rough set approach, handling inconsistencies in expression of
preferences through examples, and four-valued logic, modeling the presence
of positive and negative reasons for preference. Equivalent representation by
numerical functions and by decision rules was proposed and a specific
axiomatic foundation was given for preference structure based on the
presence of positive and negative reasons. Finally, some well known
multicriteria aggregation procedures were represented in terms of the
decision rule model; these are: lexicographic aggregation, majority
aggregation, ELECTRE I and TACTIC. In these cases the decision rules
decompose the synthetic aggregation formula into several possible
"scenarios", involving partial profiles and a dominance relation, that lead to
comprehensive outranking. Such a decomposition may be very instructive
for the decision maker.
We want to conclude this paper, by stressing one advantage of decision
rule models: decision rule representation of preferences does not need an
intermediation of numbers assigned by a synthetic utility function. This
reminds an excellent citation of Bachelard opening a chapter devoted to
preference structures in the book of Roy (1985): "Si l'ordre apparait
quelque part dans la qualite, pourquoi chercherions-nous a passer par
l'intermMiaire du nombre? ". Indeed, the decision rules do not convert
ordinal information into numeric one but keep the ordinal character due to
the syntax proposed. We think that this could be decisive for adoption of
decision rule model in real world problems.
Acknowledgements
The authors are grateful to Patrice Perny for his remarks on the first draft of
this paper. The research of S. Greco and B. Matarazzo has been supported by
the Italian Ministry of University and Scientific Research (MURST). R.
Slowinski wishes to acknowledge financial support of the KBN research
grant no. 8T11F 006 19 from the State Committee for Scientific Research,
and of the subsidy no. 1112001 from the Foundation for Polish Science.
312 AIDING DECISIONS WITH MULTIPLE CRITERIA
References
Bouyssou, D., Pirlot, M.: A general framework for the aggregation of semi0 rders. Technical
Report, ESSEC, Cergy-Pontoise, 1996.
Bouyssou, D., Pirlot, M., Vincke, Ph.: "A General Model of Preference Aggregation", in
M.H. Karwan, l. Spronk, J. Wallenius (eds.), A volume in Honour of Stanley Zionts,
Springer-Verlag, Berlin, 1997, 120-134.
Cozzens, M., Roberts, F.: "Multiple semiorders and multiple indifference graphs", SIAM
Journal ofAlgebraic Discrete Methods 3 (1982) 566-583.
Doignon, J.P.: "Threshold representation of multiple semiorders", SIAM Journal ofAlgebraic
Discrete Methods 8 (1987) 77-84.
Doignon, J.P., Monjardet, 8., Roubens, M., Vincke, Ph.: "Biorder families, valued relations
and preference modelling", Journal of Mathematical Psychology 30 (1986) 435-480.
Fishburn, P.c.: "Lexicographic orders, utilities and decision rules: A survey", Management
Science 20 (1974) 1442-1471.
Fishburn, P.C.: "Axioms for lexicographic preferences", Review of Economic Studies 42
(1975) 415-419.
Fishburn, P.C.: "Nontransitive additive conjoint measurement". Journal of Mathematical
Psychology 35 (1991) 1-40.
Greco, S., Matarazzo, B., Slowinski, R.: "The use of rough sets and fuzzy sets in MCDM", in
T. Gal, T. Stewart and T. Hanne (eds.) Advances in Multiple Criteria Decision Making,
chapter 14, Kluwer Academic Publishers, Boston, 1999, 14.1-14.59.
Greco, S., Matarazzo, B., Slowinski, R.: "Extension of the rough set approach to multicriteria
decision support", INFOR 38 (2000a) 161-196.
Greco, S., Matarazzo, B., Slowinski, R.: Conjoint measurement using non-additive and non-
transitive preference model able to represent preference inconsistencies. Report RA
10/2000, Institute of Computing Science, Poznan University of Technology, Poznan,
2000b.
Greco, S., Matarazzo, 8., Slowinski, R.: "Rough sets theory for multicriteria decision
analysis", European J. of Operational Research 129 (2001) 1-47.
Keeney, R.L., Raiffa, H.: Decision with Multiple Objectives - Preferences and value
TradeojJs. Wiley, New York, 1976.
Krantz, D.M., Luce, R.D., Suppes, P., Tversky, A.: Foundations of Measurements l.
Academic Press, New York, 1978.
Moreno, l.A., Tsoukias, A.: "On nested interval orders and semiorders", submitted to Annals
of Operations Research, 1996.
Roberts, F.S.: "Homogeneous families of semiorders and the theory of probabilistic
consistency", J. Math. Psychology 8 (1971) 248-263.
Rochat, l.C.: Mathematiques pour la gestion de l'environnement, Birkaiiser, BlUe, 1980.
Roubens, M., Vincke, Ph.: Preference Modelling, Springer-Verlag, Berlin, 1985.
Roy, B.: "Classement et choix en presence de points de vue multiples (Ia methode Electre)",
Revue Franraise d'lnformatique et de Recherche Operationnelle 8 (1968) 57-75.
a
Roy, 8.: Methodologie Multicritere d 'Aide la Decision. Economica, Paris 1985.
Roy, B., Bouyssou, D.: Aide Multicritere a la Decision: Methodes et Cas. Economica, Paris
1993.
Slowinski, R., Stefanowski, J., Greco, S., Matarazzo, B.: "Rough sets based processing of
inconsistent information in decision analysis", Control and Cybernetics 29 (2000) 379-
404.
Conjoint Measurement and Decision Rule Model 313
Tsoukias, A., Vincke, Ph.: "A new axiomatic foundation of the partial comparability theory",
Theory and Decision 39 (1995) 79-114.
Tsoukias, A., Vincke, Ph.: "Extended preference structures in MCDA". In: l Climaco (ed.):
Multicriteria Analysis, Springer-Verlag, Berlin 1997, pp. 37-50.
Tversky, A.: "Intransitivity of preferences". Psychological Review 76 (1969) 31-48.
Vansnick, lC.: "On the problem of weights in multiple criteria decision making", European
Journal of Operational Research 24 (288-294) 1986.
Wakker, P.P.: Additive representions of preferences. A new foundation of decision analysis.
Kluwer Academic Publishers, Dordrecht, 1989.
TOWARDS A POSSIBILISTIC LOGIC
HANDLING OF PREFERENCES
Salem Benferhat
IRIT-CNRS, Universite Paul Sabatier, France
benferhat@irit.fr
Didier Dubois
IRIT-CNRS, Universite Paul Sabatier, France
dubois@irit.fr
Henri Prade
IRIT-CNRS, Universite Paul Sabatier, France
prade@irit.fr
Key words: Preference; Goal; Priority; Combination; Possibility theory; Possibilistic logic
1. Introduction
Decision Analysis and Artificial Intelligence have been developed almost
separately in the last half-century. Decision Analysis, with its different
315
316 AIDING DECISIONS WITH MULTIPLE CRITERIA
branches, Multicriteria Decision Analysis, Decision under Uncertainty,
Group Decision-Making, is concerned with aggregation schemes and has
often privileged quantitative approaches, while Artificial Intelligence deals
with reasoning and has an important logically-oriented tradition (Minker,
2000). Logic and decision have different concerns indeed; the first one
focuses on consistency and inference and is oriented towards symbolic
processing, while the other deals with trade-offs (and possibly with
uncertainty), and is more numerically inclined.
Central to decision is the modeling of agents' preferences. Approaches to
multicriteria decision can be roughly divided into two families respectively
based on : i) the construction of a global multi-attribute utility function used
for ranking the candidates: it is obtained by the aggregation of the
evaluations corresponding to the different criteria (Keeney and Raiffa,
1976); ii) the use of binary relations for comparing choices according to each
criterion, which are then combined in order to obtain a global ranking of the
candidates. This second approach, which has been especially developed by
Roy and his school (Roy and Bouyssou, 1993), is more faithful to the often
non-numerical expression of user's preferences, than multi-attribute utility
function methods.
Artificial Intelligence methods can contribute to a more implicit
specification of value functions, for instance in terms of constraints. This
will be in agreement with a more granular expression of preferences, which
is often the case in practice. This general line of research has been recently
illustrated in various ways by AI researchers (e.g., Boutilier, 1994; Tan and
Pearl, 1994; Lang 1996; Boutilier et al. 1999), and usually concentrate on
more qualitative evaluations which only require ordinal scales (rather than
numerical ones). The expected benefit of the logical handling of deeisioo
problems is not only to allow for a less abstract, and thus more human-like
expression of knowledge and preferences, but also to facilitate explanation
capabilities for the candidates selected among allowed choices by decision
support systems.
Non-classical logics, recently developed in Artificial Intelligence, are
often using ordering structures, and can thus handle some forms of
qualitative preferences. Among weighted logics, possibilistic logic based on
the conjoint use of classical logic and qualitative possibility theory (Dubois
and Prade, 1998; Zadeh, 1978) offers a framework at the meeting point of
the logical and decision traditions. Besides, possibilistic logic has been
already shown to be convenient for handling nonmonotonic reasoning
(Benferhat et al., 1997a). Its framework has also been used for modelling
preferences in terms of a set of prioritized goals (Lang, 1991a) and more
recently for decision under uncertainty (Dubois et al. , 1998).
Towards a Possibilistic Logic Handling ofPreferences 317
scale L = {aO = 0 < 0.1 < ... < an = I}. C is equivalently represented by the
set of crisp sets CCl"' called aj-cut of ~c' and defined by:
1
ignored.
This gives birth to a possibilistic knowledge base of the form K = {(ca"
1
1- aj-I), i = 1, ... , n} where ca' denotes the proposition whose set of models
1
is Ca" In the following, the notations ca. and Ca. are abridged into Cj and
1 1 1
Cj. Preferences are thus expressed in terms of sets of crisp (nested) goals
having different levels of priority. Clearly, 11K computed by (1) is equal to
~c. 7f'K(u) gets the highest value when u satisfies all the goals, and gets the
lowest value when u violates the highest priority goal. More generally, 7f'K(u)
is all the smaller as u violates g0als with higher priority. Decisions violating
goals with priority 1 have a level of acceptability equal to O.
This representation plays a basic role in the manipulations that we
perform on preferences in this approach.
Example 2:
Let us consider the following three criteria-based evaluation:
- if u satisfies A and B, u is completely satisfactory, and
- if A is not satisfied, solutions should at least satisfy C.
Such an evaluation function can be encountered in multiple criteria
problems for handling "special" cases (here situations '.vhere A is not
satisfied) which coexist with normal cases (here situations where both A and
B can be satisfied). It can be directly represented by the disjunctive form:
~D(u) = max(min(~A(u), ~B(u», min(~du), 1 - ~A(u), I-p» with p < 1.
The reading of this expression is easy. Either the candidate satisfies both
A and B, or ifit falsifies A, it satisfies C, which is less sattsfactory.
This expression of ~D obtained as the weighted union of the different
classes of more or less acceptable solutions can be transformed into an
equivalent conjunctive form like (4) or more generally (1); it can be checked
that this conjunctive form corresponds to the base K = {(avc, 1); C--'avb, 1);
(a, p); (b, p)}, where A, B, C are the sets of models of a, band c
respectively; this provides a logical, equivalent description of the evaluation
process in terms of prioritized requirements to be satisfied by acceptable
solutions. Note that in this knowledge base, the formula (b, p) can be
removed since it can be recovered from (--'avb, 1) and (a, p) using the
possibilistic resolution principle described in Section 2.2.
(8 ~ A) 1\ «C, p) * B)
326 AIDING DECISIONS WITH MULTIPLE CRITERIA
where 1\ stands for the min aggregation. In the example we use the
satisfaction scale {ao, aI, a2, a3, CX4, as} = {O, 0.2, 0.4, 0.6, 0.8, I} for A
and B and we take 0 = 0.8, P = 0.6. The problem can be then translated
under the form of a stratified possibilistic base. Namely, let Aj = Aaj +l'
e.g., Al = Aa2, (here the O.4-cut).
The fuzzy criterion A is encoded by
A = {(ao, 1), (aI, 0.8), (a2, 0.6), (a3, 0.4) (a4, 0.2)}.
Then (0 ~ A) corresponds to the following knowledge base:
KO~A = {(ao, 1), (aI, 0.8), (a2, 0.6), (a3, 0.4)}.
Now the base associated to (C, 0.6) * B is :
K (C, .6) * B = {(c, 0.3), (bO, 0.5), (bI, 0.4), (b2, 0.3), (b3, 0.2),
(b4, 0.1), (bo v c, 0.8), (bI v c, 0.7), (b2 v c, 0.6),
(b3 v c, 0.5),(b4 v c, 0.4)}.
Thus, the level cuts of the fuzzy set (0 ~ A) 1\ «C, p) * B) lead to the
possibilistic base:
{(ao ,1), (aI, 0.8), (bo v c, 0.8), (bI v c, 0.7), (a2, 0.6), (b2 v c, 0.6),
(bo ,0.5), (b3v c, 0.5), (a3 , 0.4), (bI, 0.4), (b4 v c, 0.4), (b2, 0.3),
(c, 0.3), (b3, 0.2),(b4, 0.1) }.
Viewing (0 ~ A) 1\ «C, p) * B) as a fuzzy constraint satisfaction
problem, this can be exploited for relaxing it into crisp problems
corresponding to the different level cuts of the above possibilistic logic base;
see (Moura-Pires and Prade, 1998). See also (Moura-Pires et aI., 1998).
another possibility distribution 'It if and only if for each interpretation u we have 7r{u)
328 AIDING DECISIONS WITH MULTIPLE CRITERIA
::;;1I'(u) and there exists at least one interpretation u' such that 1!'(u') < 1I'(u'). In other
words, 1C'is more informative, or more requiring than 11'.
2 i.e., {l = EJ u ... u Em, and for i ;9 we have Ei!l Ej = 0, moreover Ej ;t!0 for j ~.
Towards a Possibilistic Logic Handling of Preferences 329
a. m=O
b. While {) is not empty repeat c.l.-cA.:
~
b.l. m~m+ 1
b.2. Put in Em every model which is not in any R(Ci) of Cf),
b.3. If Em == 0 then f) is inconsistent.
b.4. Remove the elements of Em from {) ,
b.5. Remove from Cf) any pair (L(Cj), R(Ci)) containing elements of
Em-
Example 1 (continued)
In Section 3.3, the set of stratified goals Cl, C2, C3 was directly given.
However, such a stratification can be related to the possibility distribution
which can be selected from a set of constraints of the forms (5)-(6). For
instance if an agent expresses that he wants coffee, and if coffee is not
available he would like tea. This corresponds to the two following
constraints:
where Cl = coffee and C2 = tea. Let {) = {u1: cI'\c2' u2: Cll\-,cZ, u3: -,ClI\CZ'
u4: -,cll\-,c2} where u1: Cll\CZ reads u1 is the model which makes Cll\CZ
true, etc. We have:
Cf) == {(rub u2}, {u3' ud), ({u3}, {u4})}'
Applying the above algorithm leads to the following partition:
E1 ::::: {u1' u2} > E2::::: {u3} > E3::::: {u4}'
A numerical counterpart can be:
7r*(U 1) == 7r*(u2) = 1,
7r*(u3) == 2/3,
330 AIDING DECISIONS WITH MULTIPLE CRITERIA
'II*(u4) = 1/3.
When a set of constraints expressing weak comparative preferences is
inconsistent (step b.3), it reveals that for each constraint there does not exist
any "ideal situation", where the preference indeed takes place. To illustrate
this, let us take the following example:
i) "sea" is preferred to "not sea"
lI(s} > TI(-,s)
ii) "mountain" is preferred to "not mountain"
I1(m) > TI(-,m)
iii) "sea and mountain" is not preferred
TI(-,m v-,s) > TI(m 1\ s)
It can be checked that these three constraints are inconsistent. Indeed, for
the first rule there exist two pos8ible ideal worlds (expressed in the
language), namely "m 1\ s" and "-,m /\ s". "m 1\ s" cannot be considered due
to constraint iii). The choice of "-,m 1\ s" contradicts the second constraint.
However if we weaken both preferences i} and ii) by adding a specific
context to (i) and (ii) by casting the statement "sea is preferred to not sea" in
the context of flat land (-,m), and "mountain to not mountain" in the context
of being far from the sea (-,s); the set of constraints can now become
consistent as shown below
i)' TI(s 1\ -,m) > TI(-,s 1\ -,m)
ii}' TI(m 1\ -,s) > TI(-.In 1\ -,s)
iii)' max(TI(m 1\ -,s), TI(-,m 1\ s), TI(-,m 1\ -,s» > TI(m 1\ s).
An obvious solution is indeed TI(s 1\ -,m) = TI(-,s 1\ m) = 1, and II(s 1\ m)
= TI(-,s 1\ -,m} < 1.
It may also happen that, even by weakening the preferences by
contextualizing them with a situation expressible in the language, there is no
way of making the set of constraints consistent. This indicates that the
language is not expressive enough for specifying the "ideal" situation(s}.
Then new literals should be added. Thus, the extreme example of the two
constraints,
i) TI(p} > TI(-p}
ii} TI(jJ) > TI(p}
it can be mad~ consistent by adding a context to one of the constraints, e.g.,
i}' TI(p 1\ x} > TI(jJ 1\ x}
ii)' II( jJ) > II(P)
Towards a Possibilistic Logic Handling of Preferences 331
<=> max (I1(--p 1\ x), I1(--p 1\ -,x)) > max (I1(p 1\ x), I1(p 1\ -,x»
a solution to which is, for instance.
I1(--p 1\ -,x) ::: 1 ::: I1(p 1\ x) ::: 0.5, I1(p 1\ -,x) ::: 0 ::: I1(--p 1\ x).
These two examples point out that even if weak comparative preferences
constraint form a rather permissive manner of expressing preferences,
preferences which are not sufficiently specific can become inconsistent. This
inconsistency reveals a lack of details in the expression of the preferences.
However, they are weaker than constraints of types (8) and (5)
respectively. The latter one, weaker than II(q) > II(-'q), only expresses that
there exists a model of q which is preferred to a model of -'q; in other words,
it states that it is not the case that the (strong) constraint .:l(-'q) ~(q) holds,
which is indeed a weak piece of information. The constraint .:l(q) > .:l(~) is
weaker than .:l(q) > II(-'q) and means that all the models of q are preferred to
at least one model of -'q; in particular the least satisfactory candidates among
the models of q is more satisfactory than at least one candidate which is a
model of~.
3 This example has been recently used by Michel Grabisch and Marc Roubens (with a=IS, b=\6,
c=\5, d=14, e=12, f=1O) for illustrating the case where no weighted average aggregation
function can agree with both the proposed orderings between the candidates and the
respective importance of the criteria, while a Choquet integral (see, e.g., Grabisch et aI.,
1995) can represent the situation.
Towards a Possibilistic Logic Handling of Preferences 333
family (e. g., Choquet integral (Grabisch, 1996», as classical approaches do.
Let us emphasize that the interest of such an approach would be to obtain a
ranking of the situations directly from the specification of users' preferences,
without having to identify an aggregation function for the criteria. It also
enables us to check the consistency of the user requirements, and to restate
preferences as a set of stratified goals (which can be then checked, or
modified by the user). The development of such an approach also raises
computational issues which are not addressed here.
5. Concluding remarks
This paper has advocated and outlined the use of possibility theory and
possibilistic logic in decision analysis for the representation and
combination of preferences. It has proposed a discussion of the expression of
qualitative preferences by providing sets of prioritized goals, or sets of
solutions reaching some given level of satisfaction, by specifying
preferences through constraints pertaining to goals, to examples or to
importance assesments. This approach shares some motivations and ideas
with the rough set-based model recently proposed by (Greco et aI., 1998),
where the indiscemibility relations underlying ordinary rough sets are
changed into dominance relations, for approximating preference relations
and getting sets of decision rules playing the role of a comprehensive
preference model. Links and differences between the two approaches are still
to be clarified.
Towards a Possibilistic Logic Handling ofPreferences 335
In connection with the approach proposed in this paper, one may think of
other lines of research. First, the logical framework does not only provide a
convenient representation tool, but also provides a basis for generating
explanations of interest for the user. Another topic of interest for further
research, briefly considered in (Benferhat et aI., 1999), is the revision of
preferences expressed as a stratified set of goals by a new input asking for
the incorporation of further preferences; see also (Ryan and Williams, 1997).
Lastly, another worth investigating issue, where a layered logic
framework may be useful, is the analysis of conflict between preferences.
Suppose that different preference profiles, expressing different points of
view, are to be combined symmetrically. Taking these different preference
profiles together very often creates inconsistencies (see, e.g., Felix, 1992).
The problem is then to determine what goals can be relaxed or put at smaller
levels of priority, taking advantage of the stratification of the preferences.
Methods developed for reasoning from stratified inconsistent propositional
logic bases may be very useful for that purpose: these methods are based on
the selection of particular consistent subbases, or on the search for pro and
cons arguments (Benferhat, Dubois, and Prade, 1996), or on the exploitation
of minimally inconsistent subsets (Benferhat and Garcia, 1998).
References
Bellman R., Zadeh L.A. (1970) Decision-making in a fuzzy environment. Management
Sciences, 17, 141-164.
Benferhat S., Dubois D., Prade H. (1996) Reasoning in inconsistent stratified knowledge
bases. Proc. of the 26 Inter. Symp. on Multiple-Valued Logic (ISMVL'96), Santiago de
Compostela, Spain, 29-31 May, 184-189.
Benferhat S., Dubois D., Prade H. (1997a) Nonmonotonic reasoning, conditional objects and
possibility theory. Artificial Intelligence, 92, 259-276.
Benferhat S., Dubois D., Prade H. (1997b) From semantic to syntactic approaches to
information combination in possibilistic logic. In: Aggregation and Fusion of Imperfect
Information, Studies in Fuzziness and Soft Computing Series, (B.Bouchon-Meunier, ed.),
Physica. Verlag, 141-161.
Benferhat S., Dubois D., Prade H. (1998) Practical handling of exception-tainted rules and
independence information in possibilistic logic. Applied Intelligence, 9,101-127.
Benferhat S., Dubois D., Prade H. (1999) Towards a possibilistic logic handling of
preferences. Proc. 16th Int. Joint Conf. on Artificial Intelligence (IJCAI-99), Stockholm,
Sweden, 31 July- 6 august, 1999, 1370-1375.
Benferhat S., Garcia, L. (1998) Dealing with locally-prioritized inconsistent knowledge bases
and its application to default reasoning. In Applications of Uncertainty Formalisms (T.
Hunter, and S. Parsons eds.), LNAI 1455, Springer, Berlin, pages 323-353.
Boutilier C.(1994) Toward a logic for qualitative decision theory Proc. of the 4th Int. Conf. on
Principles of Knowledge Representation and Reasoning (KR-94), Bonn, (J.Doyle,
E.Sandewall, P.Torasso, eds.), Morgan Kaufmann, 75-86.
336 AIDING DECISIONS WITH MULTIPLE CRITERIA
Boutilier C., Brafman R. I., Hoos H.H., Poole D. (1999) Reasoning with conditional ceteris
paribus preference statements. Proc. of the 15th Conf. on Uncertainty in Artificial
Intelligence (UAI99), (K.B. Laskey, H. Prade, eds.), Morgan Kaufmann, 71-80.
Doyle 1., Wellman M.P. (1991) Preferential semantics for goals. In Proc. of the 9th National
Conf. on Artificial Intelligence (AAAI-90), Anaheim, 698-703.
Dubois D., Fargier H., Prade H. (1996) Refinements of the maximin approach to decision-
making in a fuzzy environment. Fuzzy Sets and Systems, 81, 103-122
Dubois D., Farinas L., Herzig A., Prade H., (1997) Qualitative relevance and independence: a
roadmap. In Proceedings of the fiftheen International Joint Conference on Artificial
Intelligence (IJCAI-97), 62-67.
Dubois D., Lang J., Prade H. (1994) Automated reasoning using possibilistic logic: semantics,
belief revision and variable certainty weights. IEEE Trans. on Data and Knowledge
Engineering, 6(1), 64-71.
Dubois D., Le Berre D., Prade H., Sabbadin R. (1998) Logical representation and
computation of optimal decisions in a qualitative setting. Proc. AAAI-98, 588-593.
Dubois D., Prade H. (1987) Necessity measures and the resolution principle, IEEE Trans.
Systems, Man and Cybernetics, Vol. 17, pp. 474-478.
Dubois D., Prade H. (1998) Possibility theory: qualitative and quantitative aspects. In
Handbook of defeasible reasoning and uncertainty management systems. Vol. 1, pp. 169-
226, Kluwer Academic Press.
Felix R. (1992) Towards a goal-oriented application of aggregation operators in fuzzy
decision-making. Proc. of the Int. Conf. on Information Processing and Management of
Uncertainty in Knowledge-Based Systems (IPMU-92), Mallorca, July 6-10, 585-588.
Grabisch M. (1996) The application of fuzzy integrals in multicriteria decision making.
Europ. 1. of Operational Research, 89, 445-456.
Grabisch M, H. T. Nguyen, and E. A. Walker (1995) Fundamentals of uncertainty calculi,
with applications to fuzzy inference. Kluwer Academic.
Greco S., Matarazzo B., Slowinski R. (1998) Rough set theory approach to decision analysis.
In Proc. 3rd Europ. Workshop on Fuzzy Decision Analysis and Neural Networks for
Management, Planning and Optimization (EFDAN'98), (R. Felix, ed.), Dortmund,
Germany, June 16-17, 1998, 1-28.
Keeney R. and Raiffa H. (1976). Decision with Multiple Objectives: Preferences and Value
Trade-offs, Wiley, New York.
Lacroix M., Lavency P. (1987) Preferences: Putting more knowledge into queries. Proc. of the
13rd Inter. Conf. on Very Large Data Bases, Brighton, UK, 215-225.
Lang 1. (1991 a) Possibilistic logic as a logical framework for min-max discrete optimisation
problems and prioritized constraints. In Fundamentals of Artificial Intelligence Research
(FAIR'91), (P. Jorrand, 1. Kelemen, eds.), L.N.C.S. n0535, Springer Verlag, 112-126.
Lang 1. ( 1991 b) Logique possibiliste: aspects formels, deduction automatique, et applications.
PhD Thesis, Universite P. Sabatier, Toulouse, France, January 1991.
Lang J. (1996) Conditional desires and utilities - an alternative logical framework for
qualitative decision theory. Proc. 12th European Conf. on Artif. Intellig. (ECAI-96),
Budapest, Wiley, U.K., 318-322.
Minker 1., ed. (2000) Logic-based Artifcial Intelligence. Kluwer Academics Publisher,
Boston.
Moura Pires 1., Prade H. (1998) Logical analysis of fuzzy constraint satisfaction problems.
Proc. of the 1998 IEEE Int. Conf. on Fuzzy Systems (FUZZ-IEEE'98), Anchorage, Alaska,
May 4-9, 1998, 857-862.
Towards a Possibilistic Logic Handling of Preferences 337
Moura-Pires 1., Dubois D., Prade H. (1998) Fuzzy constraint problems with general
aggregation operations under possibilistic logic form. Proc. 6th Europ. Congo on Intellig.
Techniques & Soft Comput., Aachen, Germany, Sept. 7-10,1998, pp. 535-539.
Roy B., Bouyssou D. (1993) Aide Multicritere a la Decision: Methodes et Cas. Ed.
Economica, Paris.
Ryan J., Williams, M.-A. (1997) Modelling changes in preference: an implementation. ISRR-
027-1997, Dept. of Management, Univ. of Newcastle, NSW, Australia.
Schiex T. (1992) Possibilistic constraint sarisfaction problems or "How to handle soft
constraints?" In Proc. of the 8th Conf. on Uncertainty in Artificial Intelligence (UAI92),
(D.Dubois, M.P.Wellman, B. D'Ambrosio, P. Smets, eds.), Morgan Kaufmann, 268-275.
Spohn W. (1988) Ordinal conditional functions: a dynamic theory of epistemic states. In:
Causation in Decicion, Belief Change and Statistics, Vol. 2 , (W.L.Harper and B.Skyrms,
eds.), Reidel, Dordrecht, 105-134.
Tan S.-W., Pearl J. (1994) Qualitative decision theory. In: Proc. of the 12th National Conf. on
Artificial Intelligence (AAAI-94), Seattle, Wa., July 31 - Aug. 4, 1994,928-933.
Van der Torre L., Weydert E. (2001) Parameters for utilitarian desires in qualitative decision
theory. Applied Intelligence, to appear.
Williams, M.-A. (1994) Transmutations of know ledges systems. Proc. KR-94, 619-629.
Yager R.R. (1 992) On the specificity of a possibility distribution. Fuzzy Sets and Systems, 50,
279-292.
Zadeh L.A. (1978) Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and
Systems, 1, 3-28.
EMPIRICAL COMPARISON OF LOTTERY- AND
RATING-BASED PREFERENCE ASSESSMENT
Oscar Franzese
Oak Ridge National Laboratory, Oak Ridge, TN, USA;
iranzese@ornl.gov
Mark R. McCord
The Ohio State University, Columbus, OH, USA
mccord.2@osu.edu
1. Introduction
In multi-criteria decision aiding (MCDA), different decision makers may
evaluate the criteria of a decision problem differently. Similarly, the same
decision maker may evaluate the criteria differently under different
circumstances. Such flexibility is an essential characteristic that
340 AIDING DECISIONS WITH MULTIPLE CRITERIA
2. Experimental design
2.1 Concepts
Individuals evaluated outcomes represented by the time and cost of
obligatory trips using the rating (R), probability equivalent (PE), and lottery
equivalent (LE) assessment techniques. The trips were to begin at 5 PM and
arrive at the destination at a specified time that ranged between 7 PM the
same evening and 3 AM the next morning. Therefore, the travel time
attribute had lower (To) and upper (T) bounds, respectively, of2 hours (i.e.,
depart at 5 PM, and arrive at 7 PM) and 10 hrs (i.e., depart at 5 PM, and
arrive at 3 AM). The trip cost attribute had lower bound Co of $20. We used
the technique presented in McCord and Franzese (1993) to fix the value of
the cost upper bound (Co). Specifically, for each individual i we elicited that
individual's value C/ such that:
*
(T = 10, Co = 20) ~ (To = 2, Cj• ), (1)
where "~,, represents indifference between the alternative on its left and that
on its right. In this way, C/ was the maximum dollar amount that subject i
was willing to pay to obtain a trip taking the minimum travel time
Comparison ofLottery- and Rating-Based Preferencel Assessment 343
* (1- Pa,;)],
(10, 20) ~ [(2,20), Pa,;; (10, C;), (4a)
and
(2, C;* ) ~ [(2, 20), Pb,;; (10, C;* ), (1- Pb,; )], (4b)
344 AIDING DECISIONS WITH MULTIPLE CRITERIA
where [A, P; A', (1 - p)] indicates a lottery offering A with probability p and
A' with probability (1 - p). The indifference probabilities pa.; and Pb.; can be
shown to be constrained to lie between 0 and 1.0, and no normalization was
required. Then, for the PE technique:
[(10,20),0.675; (10,C;), 0.325] - [(2, 20), qa,;; (10,C;), (1- qa,; )], (6a)
and
[(2, C;), 0.675; (10, C;), 0.325] - [(2, 20), qb,;; (10, C;), (1- qb,; )], (6b)
2.2 Protocol
Ninety-one Civil Engineering graduate and undergraduate students at The
Ohio State University participated in the experiment. The data were
collected as a required exercise during a course module on transportation
demand modeling. The students were told that although there were no right
or wrong answers, the computer software could track parameters that would
shed light on how seriously they considered their responses. Moreover, they
were informed that the data could also be used to gain insights into
preferences for the time and cost of intercity trips-a topic of interest in the
course module. We used the PE and R techniques in eliciting the VE's for 21
subjects. For 25 other subjects we used the LE and R assessment techniques.
For the remaining 45 subjects we performed the assessments using the PE
and LE techniques.
The experiments were conducted using an interactive computer program
(Franzese, 1993). Approximately one week before the actual interview, we
conducted "warm-up" sessions with the subjects. In these preliminary
sessions we explained the experiment, stressed that there were no right or
wrong answers, presented questions similar to the ones that would be used in
the actual interview, and demonstrated the software.
During the actual experiments, the subject was given a description of the
choice scenario. He or she was asked to consider returning from a job
interview in New York, NY, to Columbus, OH, a 600 mile trip, on the
following Friday. The trip would depart New York at 5 PM. The subject was
then told that he or she would have to choose between two alternatives to
make the trip to Columbus. The alternatives were to be considered identical
in every aspect except travel time and cost (price). Arrival time would range
from an earliest time of 7 PM (i.e., 2-hr travel time) to a latest time of 3 AM
(i.e., lO-hr travel time), and cost would range from a minimum of $20 to a
maximum of C/o The subject was asked to imagine that he or she would not
be reimbursed for the cost of the trip.
Following this introduction, the data collection phase began. For
questions in the lottery-based methods and when eliciting C/ of relation (1),
the subject was shown two alternatives and asked to express a preference for
one, or an indifference between the two, by pressing one of three keys on the
computer keyboard. If the subject preferred one of the alternatives, the
software used the bracketing method (McCord and de Neufville, 1985) to
change either cost in relation (1) or probabilities p or q (and the
complementary probability) in the lotteries on the right in relation (4) or (6).
This process was repeated until the subject indicated indifference or until the
difference between two consecutive levels of the changing variable was less
346 AIDING DECISIONS WITH MULTIPLE CRITERIA
then press an assigned key to actuate a random drawing from this set of 40
squares. This "random drawing game" could be repeated as long as desired
and was provided to allow the individual to gain a better feeling for the
probability of occurrence of the outcomes considered.
Press D to UTAS
draw a card
UTILITY ASSESSMENT
• • • 00000
• • • 00000
• • 000000 John Doe
•• 000000
Which one do you prefer? •• 000000
Time Cost
Trip starts at: 5 PM
[7:00 PM; $115)
First Attribute
Name: Arrival Time
Units: hh:mm
1·p.=70%
Second Attribute
Name: Trip Cost
This combination OR this lottery
Units: Dollars
3. Results
All 91 subjects completed the interview successfully. The average times
taken to answer questions with the R, PE, and LE techniques were 7.6
minutes, 7.3 minutes, and lOA minutes, respectively.
We assessed a subject's C· value at the beginning of the session. Equality
of the VE/s in Eq. (2) depends on Ci•• Therefore, at the end of the session we
elicited a second value of C· for subject i, which we denote Ce/. For further
analysis we considered only those subjects for whom C/ and Ce/ were
within 10% of each other. Specifically, we calculated the variable RELC· for
each subject i as:
348 AIDING DECISIONS WITH MULTIPLE CRITERIA
(8)
and considered only those subjects with RELC/ < 0.10. Of the 91 subjects
participating, 41 had such RELC/, with 28 providing the same C· values at
the beginning and end of the session (RELC/ = 0). (The values and other
data for all 91 subjects can be found in Franzese, 1993.) Of the 41 subjects
whose responses were retained for further analysis, 13 had provided
preference information using the Rand PE techniques, 12 had provided
preference information using the Rand LE techniques, and 16 had provided
preference information using the PE and LE techniques.
A systematic bias in the elicited value of C;* would affect the
comparisons of the R, PE, and LE techniques. If the elicitation procedure led
to a C;* that was less than (greater than) that which the individual would
wish to use to represent indifference, substituting the elicited value in
outcome (2, Cit) would make this outcome more desirable (less desirable)
than if the unknown value the individual wished to use for indifference was
used. Outcome (2, Cit) should then be preferred (less preferred) to outcome
(10,20), and we could no longer equate the VE;'s in Eq. (2).
To investigate the possibility that a biased elicitation of C;* invalidated
the assumed indifference between outcomes (10, 20) and (2, Cit), we
calculated the difference in the responses RDt for subject i when we used
technique m - m = R, PE, or LE - as:
where VEt(2, C/) and VEt (10, 20) are the VE's of subject i obtained from
Eqs. (3), (5), and (7) for the appropriate technique m. We used a paired t-test
of the null hypothesis that the mean difference of the assessed VE's-i.e., the
mean of the RD's-was o. The alternative hypothesis was that the mean was
different from 0 (a two-sided test). Rejecting the null hypothesis in favor of
the alternative hypothesis would indicate that for some reason the subjects
were evaluating either (2, C;) or (10, 20) as more preferred. In Table 1 we
present summary statistics of the response difference RD using the different
assessment techniques and the results of the hypothesis tests. The null
hypothesis of mean RD equal zero could not be rejected for any of the three
techniques at the 90% confidence levels, either when using data obtained at
the beginning or at the end of the session. Of course, not rejecting the null
Comparison ofLottery- and Rating-Based Preferencel Assessment 349
hypothesis does not mean that the hypothesis of no systematic bias can be
accepted. Similarly, since the investigation was conducted across the set of
individuals, it cannot rule out the possibility that the elicited C;*'s were "too
large" for some individuals and "too small" for others. Still, the results of the
investigation increased our confidence that our objective, if somewhat
arbitrary criterion (RELC;* < 0.10) led to a set of data for which Eq. (2)
could be used to compare the three assessment techniques.
Rating PE LE
Be~innin~ End Be~innin~ End Be~innin~ End
Mean 0.1080 0.0640 0.0220 -0.0470 -0.0271 -0.0738
SD 0.3290 0.2548 0.2909 0.2526 0.4196 0.2346
N 25 25 29 29 28 28
T 1.6411 1.2561 0.4076 -1.0029 -0.3412 -1.6636
Alpha/2 0.0562 0.1106 0.3433 0.1622 0.3678 0.0539
Conf. Lev. 0.8862 0.7788 0.3133 0.6755 0.2644 0.8922
where VEt(2, C;) and VEt (10, 20) were determined from Eq. (3), (5), or
(7) for m = R, PE, or LE, respectively, and the vertical bars indicate absolute
value. If the group of subjects was randomly generating responses, any VE
between 0 and 1 would be equally likely to appear, and the distribution of
ARD values would be that of the difference of two random numbers
uniformly distributed between 0 and 1. The mean and variance of such a
distribution are 1/3 and 1/18, respectively (e.g., Larson and Odoni, 1981).
We tested the null hypothesis that the mean of the ARD's taken across
subjects could have been generated by independent random variables with
mean 1/3 and variance 1/18. The alternative hypothesis was that the mean
was smaller than 1/3 (a one-sided test). In Table 2 we present summary
statistics of the ARD data and the results of our tests on the mean of the
ARD's. For data obtained at the beginning of the session, we were able to
reject the null hypothesis of random generation with more than 95%
350 AIDING DECISIONS WITH MULTIPLE CRITERIA
confidence for the R technique and with very close to 100% confidence for
the PE technique. On the other hand, we could not reject with 90%
confidence that the LE responses were being generated at random at the
beginning of the session. However, for the data obtained near the end of the
session, we could reject the null hypothesis with almost 100% confidence for
all three techniques. That is, after gaining familiarity with the techniques, the
group of subjects could not be considered to have provided responses at
random. Rather, the individuals seemed to be considering preferences for the
times and costs of the alternatives.
Comparing the sample means of the ARD distributions also allowed us to
compare the techniques at the aggregate level. The sample means in Table 2
show that the PE technique performed best on average (lowest mean ARD)
when using data collected either at the beginning or the end of the session.
Table 2. Summary Statistics of Absolute Response Differences ARD for the Three
Assessment Techniques Obtained with Data at the Beginning and End of Session
Rating PE LE
Be~innin~ End Be~innin~ End Be~innin~ End
Mean (P) 0.3333 0.3333 0.3333 0.3333 0.3333 0.3333
SD (P) 0.0471 0.0471 0.0438 0.0438 0.0445 0.0445
Mean (S) 0.2520 0.1920 0.1787 0.1338 0.2905 0.1596
N 25 25 29 29 28 28
Z -1.7253 -2.9981 -3.5330 -4.5582 -0.9605 -3.9003
Conf. Lev. 0.9578 0.9986 0.9998 1.0000 0.8316 1.0000
P: Assumed Population; S: Sample
The LE technique performed slightly worse than the R technique when using
data at the beginning of the session. However, when using data collected at
the end of the session, the LE technique outperformed the R technique, and
its average ARD was almost as low as that of the PE technique. The average
ARD was smaller when using data obtained at the end of the session than
when using data obtained at the beginning of the session for all three
techniques, suggesting that performance for all techniques improved after
the subjects gained familiarity with their use.
Since we assessed preference data from the same subject using two
techniques-either PE and R, LE and R, or PE and LE-we could also
investigate performance at the individual level. We computed the relative
response differences between techniques as:
RRD,!,k
I
= ARD'!' -
I
ARD~I ' (11)
Comparison ofLottery- and Rating-Based Preferencel Assessment 351
where ARDt and ARD/, respectively, are the absolute response differences
of Eq. (10) for subject i when the assessments were conducted using
techniques m and k-mk = PE,R; LE,R; or LE,PE.
In Table 3 we present summary statistics of the relative response
difference RRD using the different assessment techniques. (The sample sizes
were too small to expect to obtain traditional significance levels in formal
hypotheses tests.) Notice that the calculated mean RRD was always
negative. Therefore, the first technique m in the pair produced smaller
deviations on average than the second technique k in the pair. That is, when
considered at the individual level, the PE and LE techniques produced VE's
that were closer on average to the supposed equality than those produced
with the R technique, and the PE technique produced VE's that were closer
on average than those produced with the LE technique. These better "on
average" performances held for data obtained both at the beginning and at
the end of the session. We see from Table 3, however, that the superior "on
average" performance of the PE technique over the R technique was large
when using either the data obtained at the beginning or at the end of the
session. The superiority of the LE technique over the R technique was slight
when using either the data obtained at the beginning or at the end of the
session. The superiority of the PE technique over the LE technique was large
when using the data obtained at the beginning of the session but not when
using the data obtained at the end of the session.
outperformed the other for essentially the same number of subjects. Finally,
we see that the PE technique outperformed the LE technique more than
twice as often at the beginning of the session, but by the end of the session
each outperformed the other for essentially the same number of subjects.
4. Discussion
The data retained for analysis could not be considered to have been
randomly generated for any of the three techniques. Rather, it appears that,
on average, the subjects providing these data considered their preferences for
time and cost when evaluating the outcomes. Rejecting data as being
randomly generated may be a minimal requirement, but we have seen little
evidence demonstrating that preference data elicited for use in MCDA can
pass this test. Passing this test argues against claims that individuals cannot
respond meaningfully to hypothetical questions, in general, and to lottery-
based questions, in particular. In fact, we were surprised to see that the
assessments from the rating technique were also able to pass such a test.
Of course, the subjects could have been anchoring their responses on
numbers that did not necessarily reflect preferences, and the statistics of
Table 2 (ARD) on the differences between the supposedly indifferent
outcomes demonstrate that there is still a lot of "noise" in preference
responses. Moreover, the data from more than half the subjects were
eliminated because they did not provide beginning- and end-of-session
values of C· that differed by less than 10%. (Investigating the performance
of these data would make for an interesting subsequent study.) We do not
suggest, therefore, that preference assessment is a straightforward task. Still,
it is noteworthy that the noise, as measured by the ARD, decreased on
average after the subjects gained familiarity with any of the techniques, i.e.,
when comparing results based on data obtained at the end of the session to
results based on data obtained at the beginning of the session. If one thinks
of "constructing," rather than "collecting" preferences, then a process in
which the decision maker gains familiarity with preference assessment
Comparison ofLottery- and Rating-Based Preferencel Assessment 353
times for the "easier" R questions would likely have been much less. This
result would not be surprising, either, since individuals are much more
frequently asked to rate something "on a scale from X to Y" than to state an
indifference involving lotteries. However, MCDA is not intended to be easy.
On the contrary, the benefits from decision aiding are in great part derived
from using a methodology to arrive at a clearer understanding of the issues
of a decision problem and of the role of individual preferences in resolving
these issues. If this were an easy task, MCDA would likely not exist as a
discipline. Performance, rather than ease, should be the overriding criterion
when judging among different techniques.
We were somewhat surprised to see that the added complexity of the LE
technique did not lead to a marked decrease in performance as compared to
the PE technique. The statistics in Tables 2, 3, and 4 show that the PE
technique did, indeed, perform much better than the LE technique when
using the data collected at the beginning of the session. However, when
considering data collected after the individuals gained familiarity with the
two techniques, there was essentially little difference in their performance.
Again, more tests would be needed before one can say these techniques
perform equally well, even if performance is based only on the types of tests
considered here. Still, our results indicate that it may indeed be possible to
obtain the potential benefits offered by the more complex LE technique
when conducting MAEUT assessments and that there is no evidence that the
lottery-based techniques are any more difficult than direct assessment
techniques for MCDA preference assessment.
Acknowledgments
The authors gratefully acknowledge the valuable suggestions offered by
two anonymous reviewers. The second author acknowledges the value of
frequent discussions with Bernard Roy in helping him arrive at a better
understanding of MCDA and the context of the type of work reported here.
He also acknowledges the value of his several discussions with Martin
Rogers and his collaboration with Cathal Brugha and Michael Bruen in
helping form the context of this work. The views presented here are those of
the authors, however, and should not be taken to represent those of these
other researchers. This work was partially funded by National Science
Foundation grant #MSS-8657342.
Comparison ofLottery- and Rating-Based Preferencel Assessment 355
References
Allais, M. "The So-Called Allais Paradox and Rational Decisions Under Uncertainty,"
Expected Utility Hypotheses and the Allais Paradox. M. Allais and O. Hagen (eds.), D.
Reidel, Dordrecht, NL, 1979, pp. 437-682.
Franzese, O. Errors and Impacts of Preference Assessments in a Multiattribute Utility
Framework. Ph.D. dissertation, The Ohio State University, Columbus, OH, USA, 1993.
Fischhoff, B, N. Welch, and S. Frederick, "Construal Processes in Preference Assessment,"
Risk and Uncertainty, 19: 1-3, pp. 139-64, 1999.
French, S., Decision Theory: An Introduction to the Mathematics of Rationality. Ellis
Horwood, Chichester, UK, 1986.
Larson, R. and A. Odoni, Urban Operations Research. Prentice-Hall, 1981.
Law, A., D. Pathak, and M. McCord, "Health Status Utility Assessment by Standard Gamble:
A Comparison of the Probability Equivalence and the Lottery Equivalence Approaches,"
Pharmaceutical Research, 15(1), 1998, pp. 105-109.
McCord, M. R. and O. Franzese, "Empirical Evidence of Two-Attribute Utility on
Probability," Theory and DeciSion, 35, pp. 337-51, 1993.
McCord, M. R., O. Franzese, and X. D. Sun, "Multicriteria Analysis of Aeromedical Fleet
Expansion," Journal of Applied Mathematics and Computing, 54(2 & 3), pp. 101-29,
1993.
McCord, M. R. and A.Y.e. Leu, "Sensitivity of Optimal Hazmat Routes to Limited
Preference Specification," Information Systems and Operational Research, 33(2), pp. 68-
83, 1995.
McCord, M. R. and R. de Neufville, "Lottery Equivalents: Reduction of the Certainty Effect
Problem in Utility Assessment," Management Science, 32(1), pp. 56-60, 1986.
McCord, M. R. and R. de Neufville, "Assessment Response Surface: Investigating Utility
Dependence on Probability," Theory and Decision, 18, pp. 263-285, 1985.
McCord, M. R. and R. de Neufville, "Empirical Demonstration that Expected Utility Decision
Analysis is not Operational," Foundations of Utility and Risk Theory with Applications,
pp. 181-199, B. P. Stigum and F. Wenstop (eds.), D. Reidel, Dordrecht, NL, 1983.
Payne, J. W., 1. R. Bettman, and D. A. Schkade, "Measuring Constructed Preferences:
Towards a Building Code," Risk and Uncertainty, 19:1-3, pp. 243-70,1999.
Rogers, M., M. Bruen, and L.-Y. Maystre, ELECTRE and Decision Support: Methods and
Applications in Engineering and Infrastructure Investment. Kluwer Academic Publishers,
Boston USA, 2000, 208 pp.
Roy, B., Multicriteria Methodology for Decision Aiding. Kluwer Academic Publishers,
Dordrecht, NL, 1996, 292 pp.
von Winterfeldt, D. and W. Edwards, Decision Analysis and Behavioral AnalYSis. Cambridge
University Press, Cambridge, England, 1986.
RISK ATTITUDES APPRAISAL AND COGNITIVE
COORDINATION IN DECENTRALIZED
DECISION SYSTEMS
Bertrand Munier
GRID - CNRS, Ecole Normale Superieure de Cachan, France
munier@grid.ens-cachan.fr
1. Introduction
In group decision analysis, conflicts and fairness issues are not the only
types of problems to be dealt with. Indeed, one of the frequently encountered
problems in modem global corporations is the question of cognitive
coordination of individuals. Yet solving such coordination questions turns
out to be increasingly difficult because individuals have to work more and
more on common tasks while having different representations of the task,
indeed often different backgrounds on which they approach the task.
358 AIDING DECISIONS WITH MULTIPLE CRITERIA
The example on which this paper will rely all along refers to a subset of
the risk management system within modem corporations, namely to the
maintenance system in nuclear power plants, but a similar analysis could be
relevant to quality control and to many other types of multiple agents
systems as long as (i) risk considerations play an important role for the
system and (ii) the system's members display some heterogeneity in terms of
personal risk attitude and to some extent in terms of corporate culture, the
latter being often very important.
Typically, indeed, a risk management system entails specialists in
production processes, others in organizational reliability, still others in
workforce health safety, who all make risk assessments of their specific
problems and try to take 'self protection' steps to the best of their
knowledge. For example, in the production process, machine inspection is
reinforced, or maintenance made more focused on reliability. At the same
time, organizational routines undergo substantial changes as a result of more
stringent reliability requirements [Weick, 1987]. And, finally, at the 'end' of
the system, 'risk managers' deal with the question of how to finance the
residual risks (insurance, self-financing, market financing or combinations of
these possibilities) once all the different preventive measures have been
taken.
The problem raised by such situations to the higher management of the
firm is one of efficiency-driven coordination. Investing some amount of
resources in the reduction of a probability of default of a given equipment
should bring about "equivalent" results to the ones provided by such or
such other step taken at similar cost towards greater reliability of the
organization. A large part of executives agree with that view, but another
large part of the same executives admit their company does not perform the
necessary adjustments because they reputedly represent an 'infeasible' task.
How to compare so widely different actions, originating from so widely
different agents in the corporation ? Yet, the costs incurred in risk
management of modem technology are huge and must imperatively be
monitored.
Economists have a common practice of cost-benefit analysis to proceed
with such comparisons, based on the 'ability to pay' approach or one of its
numerous variants. But contingent valuation is cognitively too demanding to
be performed in cases similar to the ones just mentioned above. One is then
left with the common practice of using rules of thumb, emerging from past
experience and informal conversations between the risk manager and
engineers, to design risk management policies. Relying on government
regulations only looks like an alternative course. In fact, such regulations
raise in their turn the very same problem, for they usually fall on one or the
Risk Attitudes and Cognition 359
I Indeed, they are worse, for they let employees believe that they take care of safety and keep
them from analyzing the effective risks incurred in the diverse parts of their activity, as
underlined in the remarkable 2000 Lloyds' Register Lecture [Brinded, 2000].
2 One extreme case is represented by the probability to experience a core meltdown in a
nuclear plant within a year: such a probability is of the 10-5 order.
360 AIDING DECISIONS WITH MULTIPLE CRITERIA
formed, even though the actors are in the first place heterogeneous, as has
been stressed in the introduction to this paper.
Alternatives
K-------------->lK------------------:>I K------------;>I
Fig.I. Representation ofa scenario using an influence chart until impact on chosen attribute
Cbange in default
probabDity
Cbange In the probabDity
of melting down of the core
Safety In manipulation
Fineness of tuning
Change in avaDabiJity
of power: annual hours
of plant stalled
Change in effectiveness
Expenditures on
maintenance
Etc.
Maintenance Perceived consequences Dose-response Criteria or attnbutes
Fig 2. Mapping ofa space of variables meaningful to the engineers into a reduced space of
final common attributes. Example from the maintenance problem in a nuclear power plant.
;=1
(1 + k k ; )
h
were vl= 1
\.-I. 2
" ... ,n,kj=u (. -0) ,
Xj,Xj k ij=u ( . . -0) and -0
Xj,Xj,xij xij deSlgnates
.
levels of all attributes except Xi and 0 .
Theorem 3 (Dyckerhoff, 1994)
If the decision maker uses a rank-dependent evaluation of prospects as a
decision criterion, the set of X and Y attributes are additively independent if
and only if the following two conditions are met:
i) e = id[o,lj and ii) U is decomposable under an additive form.
One can write in this special case:
;=n i=n
U(x) = LU(xpX;°) = L kjUj (Xi)
i=1 i=1
5 The ai'S functions are respective probability transformation functions of the attributes i.
Note that, if they were aU the same function, one could interpret them as perception
functions of the probabilities. In fact, they are not, which shows that they represent in fact
the way the individual is prepared to deal with risk in view of a given attribute. This is in
line with what sociologists like Slovic have shown: risks are not aU of interchangeably
dealt with by individuals, depending on which attribute they bear among other things.
Risk Attitudes and Cognition 367
Ci
n n
V = Iki fOi[GUj (.)]d. + I Ikij fOi[Gu/.)]d. fOj[Gu/.)]d. +
i=1 i=1 j>i
6 SERUM was first developed by GRID (CNRSIENS de Cachan) and Electricite de France, It
stands for 'Systeme d'Estimation dans Ie Risque des Utilites Multi-attributs',
Risk Attitudes and Cognition 369
supposed to have exactly the same system of preferences as his own. Twin A
benefits from a risk reduction which turns the lottery (X,p;O) into the less
risky lottery (X,q;O). Twin B faces a risk reduction which turns the lottery
(X,q;O) into the even less risky lottery (X,S;O). The actor is then asked
which of these twins is more satisfied by the risk reduction he experiences.
The outcome X is fixed, and the analyst varies the probabilities p, q, r, s until
the decision maker reveals indifference between two risk reductions. To start
with, p is set to 1 and s to O. If the decision maker acts according to rank
dependent utility, the indifference between the two risk reductions for
q = q~ implies the equality
(x;,p;xn - (x; ,q;xn with 0 :;t P :;t 1 and 0 :;t q:;t 1. According to RDM, one
can write
u.(x') = (}j(q)
I I (}j(p)
As in this phase of the analysis the probability transformation functions
0.0 have been already elicited we easily obtain a first point on the partial
utility function u,{Xj).
In fact, the results obtained depend on the way the indifferences such as
(x;,p;xn - (x; ,q;xn are obtained. Direct elicitation of an "equivalent
lottery" to a given reference lottery leads to inconsistent results. In fact, it is
a common finding that this specific valuation task is extremely difficult for
most subjects to perform and entails therefore many errors. The "Closing
in" method [Abdellaoui and Munier, 1994a] does not use direct indifference
elicitation, but rather choice tasks. These are much simpler to fulfill and lead
to less inconsistencies which could be found to be errors in a large part. It
has been therefore incorporated into SERUM.
the steam generators when they are started and when the reactor is stopped in
the middle of normal operations in order to push the system into a state of
cold stop.
Three strategies in preventive maintenance had been singled out:
A gammagraphic check every 5 years
A gammagraphic check every 10 years
No gamma graphic control at all.
Engineers have demonstrated for long periods competence in dealing
with this question, but all expressed doubts as to the efficiency of the choices
they made, for they had no way in weighing costs and benefits in this
activity. For that prototype study, organization specialists were temporarily
ignored. The idea was rather to design a workable system and test it on a
limited problem.
The OMF version of ReM used in the operator of power plants in
question stresses four general objectives:
1) Maintain safety of the production system at its best possible
level
2) Obtain the best possible annual availability of production
3) Decrease radiation doses to the lowest achievable level
4) Keep maintenance costs under control and decrease
intervention costs.
As ASG is outside the confinement enclosure building, general objective
Nr. 3 is here irrelevant. Besides, step 1 of the study turned out to be
relatively trivial in this limited case, and quickly converged on three axes of
meaning which were more or less equivalent appraisals of the three other
general goals. So, the study consisted mainly in designing the three
corresponding criteria and - more importantly - the decision analysis view
and its foundations, as reported in sections 2 to 4 above, and in encoding the
probability transformation function as well as the partial utility function for
each of the three criteria and for each individual of a sample considered
representative among the people in charge of maintenance" on the spot" in
several power plants. 20 people were thus investigated under an anonimity
guarantee. Each interactive SERUM session lasted for a little less than 100
minutes on average, which is some kind of maximum in experimental
research.
Availability was represented through the number of hours lost per year in
production. It varies between 0 and 600 hours.
Safety was represented through the factor by which core meltdown
probability is increased, due to degrees of default in the ASG system. That
factor was considered as varying between 0% and 1200%.
372 AIDING DECISIONS WITH MULTIPLE CRITERIA
SERUM: An Example
O~PlI 50'10
l5% 0,(PlI -50'10 ORJI -_
ZI'Io _ _
u,(X1) '~[J
o.~
0.5
up:2)
o,~
o.s
U,llCl)
'~B
o.~
o.s
0;5 O;!!I U2I
0 0
0.00 18.00 36.00 5<,00 72,00 0.00 112,50 125,<Xl 181,!i) lOO.lD 00,00 J31,SO 625,00 912,50
X, X, x,
Scaling constans
k. I kz I kJ I kl2 i klJ I k13 I km
0,1425 I 0,9702 I 0,1425 I -0,9711 I -0,1425 I -0,9831 I 1,8416
Fig3. Results obtained from a SERUM session with a nuclear powerplant maintenance
engineer (see comments in the text).
with the third and last attribute, namely maintenance cost (utility is defined
in terms of 103 F).
Fairly general findings of the study appear already on this example:
Costs are dealt with by a large majority of subjects nearly like
expected utility, i.e. with a probability transformation curve
following the first diagonal of the unit square, whereas it is not
the case for the two other attributes.
Convex probability transformation functions (meaning a very
"cautious" or "pessimistic" treatment of probabilities by the
individual) appear more often for the safety attribute (core
meltdown) than for the availability issue and they never appear
for the cost issue. In other words, the shape of the probability
transformation curve expresses at least part of the individual's
psychological representation of the issue at stake: core
meltdown is subsumed under the " dreadful" and
" uncontrollable" risks, as social psychologists put it.
Of course, the shape of the probability transformation curve
does not have any specific correlation with the shape of the
utility function. Each of these curves expresses independently its
part of the attitude towards risk (whereas Neumannian theory
considers only the part linked to the utility function).
Once obtained for a given individual, these different functions allow to
rank the alternatives on a scale ranging between 0 and 1 (table 1). This
ranking represents what some specific engineer would have done if
confronted to the choices already mentioned and recalled in column one of
table 1.
These results are obtained with a set of scaling constants emerging from
the investigation of one subject-engineer as : k\ = 0,4699, k2 = 0,9944, k3 =
0,3296, k12 = -1,1382, k\3 = -0,3296, k23 = -0,8566, k123 = 1,5306.
374 AIDING DECISIONS WITH MULTIPLE CRITERIA
7 Of course, freedom should still be left to plant managers to deviate from the reference
support system recommendations, to the extent that plant managers can have locally
specific informations not contained in the system.
Risk Attitudes and Cognition 375
References
Abdellaoui, M., 2000 « Parameter-free Elicitation of Utilities and Probability Weighting
Functions », Management Science. 46 (11),1497-1512.
Abdellaoui, M., Munier, B., 1994a « The 'Closing In' Method: An Experimental Tool to
Investigate Individual Choice Pattems Under Risk », in: B. Munier and MJ. Machina,
Models and Experiments in Risk and Uncertainty. DordrechtIBoston, Kluwer Academic
Publishers. p. 141-155.
Abdellaoui, M., Munier, B., 1994b « On the Fundamental Risk-Structure Dependence of
Individual Preferences under Risk: an experimental investigation », Note de Recherche
GRID nO 94-07.
Abdellaoui, M., Munier B., Leblanc, G., 1996, « La transformation subjective de faibles
probabilites face au risque: Ie cas de I'exposition aux rayonnements ionisants », Note de
Recherche GRID nO 96-04.
Abdellaoui, M., Munier, B., 1998, « The risk-Structure Dependence Effect: Experimenting
with an Eye to Decision-Aiding », Annals of Operations Research, 80,237-252.
Allais, M., 1953, « Le comportement de I'homme rationnel devant Ie risque: critique des
postulats et axiomes de I'ecole americaine », Econometrica. 11, 503-546.
Allais, M., 1988, « The General Theory of Random Choices in Relation to the Invariant
Cardinal Utility Function and the Specific Probability Function, the (U,e) Model: A
General Overview», in: B. Munier, ed., Risk. Decision and Rationality.
DordrechtIBoston, Reidel, 231-289.
Beaudouin, F., Munier B., Serquin Y., 1999, « Multi-Attribute Decision making and
Generalized Expected Utility in Nuclear Power Plant Maintenance », in : M. 1 Machina
and B. Munier, Beliefs. Interactions and Preferences in Decision Making.
BostonIDordrecht, Kluwer academic Publishers. p. 341-357.
Bouyssou, D., Vansnick, lC., 1990, « Utilite cardinale dans Ie certain et choix dans Ie
risque », Revue Economique. 41, 979-1000.
Brinded. M. A., 2000, Perception versus Analysis: How to Handle Risk. The 2000 Lloyd's
Register Lecture. London, The Royal Academy of Engineering,Spring.
Currim, S.H., Sarin, R.K., 1989, « Prospect versus Utility », Management Science. 35 (1),
22-40.
Denneberg, R., 1994, Non Additive Measures and Integrals. DordrechtIBoston, Kluwer
Academic Publishers.
Dyckerhoff, R., 1994, « Decomposition of Multivariate Utility Functions in Non Additive
Expected Utility Theory», Journal ofMulti Criteria Decision Analysis. 3, 41-58.
Fishburn, P.C., 1984a, « Dominance in SSB Utility Theory», Journal of Economic Theory.
34,130-148.
Fishburn, P.C., 1984b, « Multiattribute Nonlinear Utility Theory », Management Science. 30,
1301-1310.
Howard, R.A., 1990, « From Influence to Relevance to Knowledge », in: R.M. Oliver and
lQ. Smith, eds., Influence Diagrams. Belief Nets and Decision Analysis. New York,
Wiley. P. 3-23.
Keeney, R.L., 1996, Value-Focused Thinking:A Path to Creative Decision Making. Boston,
Harvard University Press.
Keeney, R.L., Raiffa, H., 1976, Decisions with Multiple Objectives: Preferences and Value
Tradeo./fs. New York, Wiley.
Risk Attitudes and Cognition 377
Raymond Bisdorff
Centre Universitaire de Luxembourg, Luxemburg
bisdhein@pt.lu
Foreword
Let me thank beforehand the editors for having invited me to con-
tribute to this book in honour of Bernard Roy. When I obtained in
summer 1975 a three years NATO fellowship in Operations Research, I
got the opportunity to join, apart from several universities in the USA,
two European OR laboratories. One was directed by H.-J. Zimmer-
mann in Aachen and the other by Bernard Roy in Paris. Having made
my under-graduate studies in Liege (Belgium), I knew well the nearby
German city of Aachen and I decided therefore to preferably go to Paris
and join Bernard Roy at the newly founded Universite Paris-Dauphine.
380 AIDING DECISIONS WITH MULTIPLE CRITERIA
It is only later that I realized how important this innocent choice would
be for my scientific career. Indeed, I joined the LAMSADE, Roy's OR
laboratory, at a moment of great scientific activities. We may remem-
ber that 1975 is the birth year of EURO, the Federation of European
OR Societies within IFORS and more specifically the birth year of the
EURO Working Group on Multicriteria Decision Aid coordinated by
Bernard Roy so that I became an active participant in the birth of the
European School in Operations Research. One can better understand
25 years later that joining the LAMSADE at that precise moment made
an ever lasting positive effect on me. May Bernard recognize in this con-
tribution, a bit of the scientific enthusiasm he has communicated to all
his collaborators. Indeed, I rarely met any other person of such arguing
clarity when trying to match formal logical constructions with pragmatic
operational problems which often, if not always, appear uncertain and
fuzzy in nature. It is my ambition in this chapter to continue with this
tradition.
R.BisdorfJ, June 2001
1. Introduction
"Du point de vue de la connaissance, nous sommes capables de connaitre
une chose au moyen de son espece et nous sommes incapables de la nom-
mer si nous ne la connaissons pas; par consequent, si nous emettons une
vox significativa, c'est que nous avons une chose a l'esprit." (Umberto
Eco, Kant et l'ornithorynque [11, p.437])
Here, the term 'family' refers to the fact that the considered set of
criterion functions supports exhaustively the pragmatic preferences of
the decision-maker. More generally, we notice in Definition 3 that the
universal assertion 'a 8 b' is truth or falseness warranted from multiple
points of view depending on the decomposition of its cloud of conse-
quences into separated preference dimensions.
1 Neglecting in this definition a possible indifference threshold, Roy uses normally a single
implication. But we prefer to work with a double implication as it allows to capture at the
same time the semantics of the negated assertion.
Logical Foundation of Multicriteria Preference Aggregation 385
Definition 5. Let T = (A, F) be a given performance tableau and let
kg E Q+, measure numerically the significance, criterion 9 E F takes
in the eye of the decision maker with resect to the truth assessment of
the universal outranking situation. Let kF denote the universal closure
of the significance weights over the whole family of criteria, i.e. kF =
E9EF (kg). We denote F+, the subset of criteria that clearly support the
truthfulness of a given outranking assertion and we define the credibility
r( a S b) of the universal outranking assertion 'a S b' as follows:
r(aSb) = L k
(kg)·
gEF+ F
o
Proposition 1 evidently relies again upon the three properties of the
coherent family of criteria and it has an interesting logical corollary.
Corollary 1. Let A be a set of decision actions evaluated on a given
coherent family of criteria F and let r(a S b), Va, b E A denote the de-
gree of credibility of a pairwise outranking situation computed following
Definition 5.
• if r( as b) 2:: ~, then 'a S b' is considered to be more or less true,
3The universality of the semiotical reference over all possible pairs of decision actions, as
assumed in Definition 8, is a highly problematic assumption from a cognitive point of view,
but due to space limitations, we do not discuss this issue within this contribution.
Logical Foundation of Multicriteria Preference Aggregation 391
4Generally speaking, semiotics only apply to social interpretations of signs. But in accordance
with a Peircian approach, we may very well specialize semiotical considerations to local
cognitive contexts, here the social working context of the decision maker. In this way, we
explicitly introduce a social dimension into the decision aid model.
392 AIDING DECISIONS WITH MULTIPLE CRITERIA
r"(
PIRe
) = {I
0
if Re certainly confirms assertion PIll"
otherwise
true if r"(PIRJ = 1,
false if r"(Plll,,) = O.
Indeed, restricted to elementary preference dimensions, the construc-
tive methodology allows us to assume that the relative measure of signif-
icance of the argument restricted to an elementary reference, is 1, in the
sense that it is precisely the operational purpose of a criterion-function
to most clearly signify, under the principle "toute chose pareille par
ailleurs" , what of the two possible truth values (true or false) is actually
the casewhen looking at a given outranking situation.
Based upon these elementary references, we may now recursively de-
fine the degree of credibility of the universal assertion.
Definition 14. Let P represent an affirmative outranking assertion as-
sociated with a reference family 'RF supporting a weight distribution
k and let {Ri E RF / i = 1 ... n} denote a complete semiotical refer-
ence system. The degree r" of credibility of assertion P is given by the
following recursive definition:
n
r"(p) = L (k(~) x r"(PIRJ).
i=l
student la ca st
a 12 12 19
b 16 16 15
c 19 19 12
r(xSy) a b c
a 1.0 0,42 0,42
b 0,58 1,0 0,42
c 0,58 0,58 1,0
Besides these inconsistencies above, one may naturally question the pre-
cise numerical measurability of the relative significance of each criteria.
This is a well known weak point of the logical approach for multicrite-
ria reference aggregation and many work-arounds have been proposed
(see Mousseau [15]). We do not have the space here to discuss this
issue, therefore we postpone this topic to a future publication6 and con-
centrate now our attention first, on a situation where we may observe
partly redundant criteria.
6We have presented a purely ordinal version of the concordance principle in a communication
at the 22nd Linz Seminar on Fuzzy Set Theory on Valued Relations and Capacities in Deci-
sion Theory, organised by E.P. Klement and M. Roubens, February 2001 (see Bisdorff [7]).
Logical Foundation of Multicriteria Preference Aggregation 397
Wi - O=~:;t'i=l W{~k))
kii = w{R2)
(5)
W{Rij)
kij (6)
w{R2)
Logical Foundation of Multicriteria Preference Aggregation 399
In Table 3 we show the corresponding decomposition for the three
subjects underlying the evaluation of the statistics students under the
hypothesis that the calculus reference presents a 50% overlap with re-
spect to statistics reference. The marginal distributions ki. and k.j shown
topics la ca st U
ki 0,29 0,29 0,42 1,00
kij la ea st kt.
la 0,37 0 0 0,37
ea 0,10 0,265 0,365
st 0,265 0,265
k·.) 0,37 0,10 0,53 1.00
r(PIRJ + r(PIRj)
r(PIRij) = 2
If both elementary references give unanimous results, either zero or
one, the resulting credibility will be the same as the credibilities of the
underlying elementary references. If they disagree, the degree of cred-
ibility of their shared reference part will be put to ~, i.e. the logically
undetermined value.
400 AIDING DECISIONS WITH MULTIPLE CRITERIA
r(xSy) a b c
a 1,0 0,4 0,4
b 0,6 1,0 0,4
c 0,6 0,6 1,0
5. Conclusion
In this chapter, we have investigated the logical and semiotical foun-
dation of the concordance principle, i.e. the logical approach to multi-
criteria preference aggregation promoted since 1970 by Bernard Roy.
We first showed the denotational isomorphism which exists between
this concordance principle and the proceeding of balancing reasons as
promoted by C.S. Peirce. This result illustrates the split truth versus
falseness denotation installed by the concordance principle.
Taking furthermore support on the Peircian distinction between credi-
bility and state of belief concerning a preferential assertion, we propose a
semiotical foundation for the numerical determination of the significance,
i.e. the truth assessment knowledge carried by the family of criteria.
This approach makes apparent the semiotical requirements guarantee-
ing the coherence of a family of criteria.
Finally, based on these semiotical requirements, we propose an exten-
sion of the concordance principle in order to support pairwise overlapping
criteria and/or incomplete performance tableaux.
References
[1] R. Bisdorff and M. Roubens (1996), On defining fuzzy kernels from C-valued
simple graphs, in: Proceedings Information Processing and Management of Un-
certainty, IPMU'96, Granada, 593-599.
[2] R. Bisdorff and M. Roubens (1996), On defining and computing fuzzy kernels
from C-valued simple graphs, in: Da Ruan and al., (eds.), Intelligent Systems
and Soft Computing for Nuclear Science and Industry, FLINS'96 workshop, pp
113-123, World Scientific Publishers, Singapoure.
[3] R. Bisdorff (1997), On computing kernels from C-valued simple graphs. In Pro-
ceedings 5th European Congress on Intelligent Techniques and Soft Computing
EUFIT'97, vol. 197-103, Aachen.
[4] R. Bisdorff (1998), Bi-pole ranking from pairwise comparisons by using initial
and terminal L-valued kernels. In Proceedings of the conference IPMU'98, pp
318-323, Editions E.D.K., Paris.
[5] R. Bisdorff (1999), Bi-pole ranking from pairwise fuzzy outranking. Belgian Jour-
nal of Operations Research, Statistics and Computer Science 37(4) 53-70.
[6] R. Bisdorff (2000), Logical foundation of fuzzy preferential systems with applica-
tion to the Electre decision aid methods. Computers (3 Operations Research 27
pp 673-687.
[7] R. Bisdorff (2001), Semiotical Foundation of Muticriteria Preference Aggregation.
In E.P. Klement and M. Roubens (eds), Abstracts of the 22nd Linz Seminar on
Fuzzy Set Theory, pp 61-63, Universitatsdirektion, Johannes Kepler Universitat,
Linz, Austria.
[8] R. Bisdorff (2001), Electre like clustering from a pairwise fuzzy proximity index,
European Journal of Operational Research, (to appear).
REFERENCES 403
[9] D. Bouyssou and M. Pirlot (1999), Non-transitive decomposable conjoint mea-
surement. In N. Meskens and M. Roubens, Advances in Decision Analysis,
Kluwer, Dordrecht
[10] D. Dubois, H. Fargier, P. Perny and H. Prade (2001), On Concordance Rules
Based on Non-Additive measures: An Axiomatic Approach. In E.P. Klement and
M. Roubens (eds), Abstracts of the 22nd Linz Seminar on Fuzzy Set Theory, pp
44 - 47, Universitiitsdirektion, Johannes Kepler Universitiit, Linz, Austria.
[11] U. Eco (1999), Kant et l'ornithorynque, Grasset, Paris.
[12] H. Fargier and P. Perny (2001), Modelisation des preferences par une regIe de
concordance generalisee. In M. Parruccini, A. Colorni and B. Roy (eds), Selected
Papers from 49th and 50th meetings of the BURO Working Group on MeDA,
European Union, forthcoming.
[13] J. Fodor and M. Roubens (1994), Fuzzy preference modelling and multi-criteria
decision support, Kluwer Academic Publishers, Dordrecht.
[14] J.-L. Marichal (1999), Aggregation operators for multicriteria decision aid, PhD
Thesis, University of Liege, Belgium.
[15] V. Mousseau (1995), Eliciting information concerning the relative importance
of criteria. In P.M. Pardalos, Y. Syskos and C. Zopounidis (eds.), Advances in
Multicriteria Decsion Aid. Kluwer, Nonconvex Optimization and its Application,
vol. 5 pp 17-43.
[16] C. S. Peirce (1878), The probability of induction, Popular Science Monthly, in
J. Buchler (ed.), Philisophical writings of Peirce, Dover, New York, 1955.
[17] B. Roy (1985), Methodologie Multicritere d'Aide d la Decision, Economica, Paris.
[18] B. Roy and D. Bouyssou (1993), Aide Multicritere d la Decision, Economica,
Paris.
[19] B. Roy and V. Mousseau (1996), A theoretical framework for analysing the no-
tion of relative importance of criteria. Journal of Multicriteria Decision Analysis,
Wiley, vol. 5 pp 145-159.
v
APPLICATIONS
OF MULTI-CRITERIA DECISION-AIDING
A STUDY OF THE INTERACTIONS
BETWEEN THE ENERGY SYSTEM
AND THE ECONOMY USING TRIMAP
Carla Oliveira
INESC, Coimbm, Portugal
coliv@pombo.inescc.pt
J oao Climaco
INESC and University of Coimbm, Portugal
jclimaco@pombo.inescc.pt
Abstract In this paper a study is presented which is aimed at modelling the inter-
actions of the energy system with the economy on a national level. For
this purpose a multiple objective linear programming (MOLP) model
is developed, which is based on an input-output (10) table where the
energy sectors as well as the relevant economic activity sectors are dis-
aggregated. The results provided by the MOLP model with a set of
data representative of the portuguese situation are analyzed using the
TRIMAP interactive environment.
1. Introduction
The problem of energy dependence is a crucial issue, in the case of
Portugal, since, except for some renewable resources (such as hydro,
wind, solar and biomass) and some industrial by-products, all types of
primary energy are imported. In fact fossil fuel imports account for more
408 AIDING DECISIONS WITH MULTIPLE CRITERIA
est and, in a second phase, to search in more detail the regions which
the initial screening revealed as potentially interesting. In this way the
interactive environment contributes to stimulate the DM to structure
and stabilize his/her own preferences by bringing him/her arguments
able to strengthen or weaken his/her own convictions (Roy, 1987, 1990).
TRIMAP is designed for problems with three objective functions, which
allows for the use of graphical tools that are suited for an open communi-
cation scheme, capitalizing on the DM's main strengths, namely pattern
recognition through visual inspection. A block diagram of TRIMAP is
depicted in figure 1.
TRIMAP combines three main procedures: weight space decompo-
sition, introduction of constraints on the objective function space, and
introduction of constraints on the weight space. Moreover, the introduc-
tion of constraints on the objective function values is translated into the
weight space.
The dialogue with the DM is made mainly in terms of the objective
function values, which is the type of information that does not place
an excessive burden on the DM. In general, the weight space is used in
TRIMAP as a valuable means for collecting and presenting the infor-
mation, and not as a way for eliciting the preferences of the DM. The
weight space is filled with the indifference regions corresponding to the
(basic) nondominated solutions already computed. Direct limitations
introduced on the variation of the weights or inferior bounds on the
objective values translated into the weight space are also displayed. A
2-dimensional projection of the objective function space, shows all non-
dominated solutions already computed and enables also to identify the
nondominated edges and faces. By comparing the two graphs (weight
space and projection of the objective function space) it is possible, by
simple visual inspection, to avoid the search of sub-regions of the weight
space which are of no interest.
In each step of the Human-computer dialogue (when a new phase of
computation is being prepared) the DM needs only to express some indi-
cations about the options to be considered in the progressive search for
nondominated solutions (region of the weight space to search, bounds
on the objective function values, etc.). The interactive process con-
tinues until the DM has gathered "sufficient knowledge" about the set
of nondominated solutions, not requiring irrevocable intermediate deci-
sions, until a final satisfying solution (or a set of solutions for further
screening) is made explicit in face of the DM's evolutionary preferences
(Climaco and Antunes, 1987 and 1989).
Interactions Between Energy System and Economy Using TRIMAP 411
.B
.c
NumeIical infonnation:
fk, x i, LI, L2, L-, AIea(~)
~ DM consideIll the;~a
--
decision can be made based
on the infonnation ob1ained
END
~
R P I cc' --;;;;-
E "Ie cclf'
'" O"P
- m;,
E
A",
(wAu)
Am
(wAu)
~
(wAu)
An
(wAu)
"Ie.
(w-)
0 0
"".
(w)
'''P.
(w)
ARE 0 0 ARJ 0 0 0 0 0 ~
0
R
(wAu) (wAu)
p
APE 0 0 API ~er 0 0 ocr e"Pr II
(wAu)
AlE Am 0
(w-)
An
(w-)
ocl
(w)
e"Pl
- m.;
I
(\noAu) (wnAu) (wn-) (wnAm) (wn) (wn) (wn) (wn)
-
(wn)
Primary iDpuu
~
aa
(wnAu)
a..
0
0
0
0
au
(uo"")
aol
II ~
y ==
8+ =[]+8
~(pc + cc*+ gcff* + SIC + exp
(wnAu) (wn-) SIC'"" sc+. sc·
o"E 0 0 0,,1
maxJI = a~mpx
in which a~mp is the vector of the number of employees (per unit output
of each sector).
1 - k2 [ ka ( )( ,
pc= 1-kl 1- k5 1-tf aw x + k4a,oX ) -kasfaw, x 1
in which kl is a constant (non-residents' consumption in the country/total
private consumption), k2 is a constant (residents' consumption out of the
country /total residents' consumption), ka is the average propensity to
consume (household consumption/gross household disposable income),
t f is the average household income tax rate and sf is the social contri-
bution rate.
The objective function is
maxh =pc
min fa = i/amE
in which ia is a vector of ones with convenient dimensions and mE is the
energy imports vector.
bp 2: lbbp
in which lbbp is the lower bound for the balance of payments.
in which ps, by analogy with pm and pe, is the vector of average prices
for stock changes of hydrocarbons (per toe) and i1and i2 are vectors of
ones, both with convenient dimensions, and
in which atm is the vector of average tax rates for imports, and atg is
the average gross added value tax rate.
The intrinsic coherence of the model does not guarantee, however,
that the two definitions lead to the same results. Notice that the first
definition has two exogenous components, and that imports, exports
and stock changes are not connected to the model coefficients. In this
context, both have been explicitly considered.
lb ::; x ::; ub
in which lb and ub are the vectors of lower bounds and upper bounds
for sector outputs, respectively.
exp::; ub exp
in which ub exp and ubmc are the vectors of upper bounds for total
exports and imports, respectively.
Ct ~ ubct
in which ubct is the upper bound for CO 2 emissions.
sc +
P - sc-P ~ ss* - e O*
in which e O* is the vector of initial stocks, sc* is the vector of storage
capacity and ss* is the vector of security stocks (the last two vectors are
both exogenous).
The solution search process began with the computation of the non-
dominated solutions which optimize individually each objective function
(table 1). This information is useful to have a first overview of the range
of variation of the objective values within the nondominated region as
well as the main characteristics of (in principle) well dispersed solutions.
Figure 3 shows the weight space filled with indifference regions cor-
responding to 20 nondominated (basic) solutions. These solutions have
been computed in a selective and progressive way. The DM indicates the
region of the weight space where a new solution shall be computed. This
action results from the analysis of the information regarding the (basic)
nondominated solutions already known, whose indifference regions are
displayed.
For instance, the analysis of the objective function values of solutions
4, 9 and 10 may certainly lead the DM to opt for not computing new
solutions in the not yet searched region located between the indifference
regions corresponding to those solutions. In fact, the DM now knows
that for nondominated solution therein computed, the corresponding it
values would be better than in solution 9 but worst than in solution 4.
This selectivity of the search, which is enhanced by the interactive visual
information displayed on the weight space, contributes to minimize the
computational effort and the number of irrelevant solutions generated
during the exploitation of the problem. This type of assistance has
been particularly helpful in avoiding the search in the region between
solutions 2 and 5, in which a huge number of (basic) nondominated
solutions exists. Let us suppose that as a results of the knowledge on the
nondominated solution set acquired throughout the interactive process,
the DM wants to impose further limitations (reservation levels) on the
objective functions. The additional constraints it ~ 5 500 000 and
is ~ 19 950 000 restrict the search to regions above solutions 3 and 13,
and below solutions 4 and 5, respectively. The result of this feature of
TRIMAP, which is particularly useful values in order to reduce the scope
of the search (in a revocable manner) according to the indications of the
DM, is displayed in figure 4 (see Climaco and Antunes, 1987 and 1989
for technical details). New nondominated solutions computed by using
sets of weights within this region, which the initial screening revealed
as potentially interesting, are also displayed in figure 3. Let us consider
that after the computation of these 20 nondominated solutions the DM
finds them sufficiently representative of the universe of different policies.
Although several important economic indicators (GAV, GDP, etc.)
are computed in the framework of this model, allowing a detailed anal-
ysis of the characteristics of the solutions, the illustrative results herein
422 AIDING DECISIONS WITH MULTIPLE CRITERIA
reaches the lowest level, denoting that the country is highly dependent
on the energy obtained externally. Although total energy imports are
at the lowest level, the structure of this solution possesses some char-
acteristics worth noting: the level of electricity imports is the highest
(476 512 toes) representing a difference of about 35% from the nearest
value obtained with the optimization of private consumption. This is
explained by the fact that electricity production and self-production of
electricity are highly hydrocarbon intensive and the level of electricity
imports compensates the imports of hydrocarbons that would be neces-
sary to produce the same amount of electricity. A significant difference
exists in the values of iI and h regarding solutions in this lower left
part of the weight space and solutions immediately above (14 and 6; for
which more attention is paid to employment and private consumption
objective functions).
Solution It h h
(employment) (private consumption) (energy imports)
1 5795435 13444893 20885591
2 5788730 13453733 20891 825
3 5 285 947 12378740 19356817
4 5770 183 13390334 20 002 774
5 5764580 13394417 19 999 096
6 5668910 13136696 19658520
7 5672 009 13147573 19664468
8 5748408 13343734 19 835 499
9 5754016 13358331 19863325
10 5751091 13347213 19841 769
11 5742511 13314886 19808328
12 5745124 13318433 19814536
13 5403214 12444103 19 383 808
14 5663068 13129939 19 653 483
15 5666981 13141724 19660118
16 5752297 13378746 19912 142
17 5681 873 13154143 19675466
18 5700706 13249027 19739093
19 5718186 13 219 655 19740774
20 5751662 13376254 19903953
private consumption (13 444 893) and energy imports (20 885 591) reach
high values. Metallurgy, construction, water/steam and electricity out-
puts reach their highest level in this solution. The level of production
of hydrocarbons is slightly below the one obtained with the optimiza-
tion of private consumption (as the result of lower outputs reached for
diesel-oil and naphtha). The corresponding level of self-production of
electricity is below the one obtained in solution 2. The lower level at-
tained for energy imports, when compared to solution 2, is due to the
imports of the following types of energy: electricity, crude oil, shale oil,
paraffin, solvents and petroleum coke. The structure of production of
by-products has more representative values for pitch, coke oven gas and
coke gas which are inputs of steam production. In solution 2, which max-
imizes private consumption, the level of employment is a reasonable one
(5 788 730) and energy imports reach the worst value (20 891 825). The
non-energetic outputs which have particularly high values (when com-
pared to solutions 1 and 3) are metal-electrical-mechanics, food, services
and city gas (notice that services are highly intensive in city gas). The
production of hydrocarbons reaches the top level (due to diesel-oil and
naphtha). The highest level of energy imports obtained in this solution is
the result of the level of imports of crude oil, shale oil, paraffin, solvents
and petroleum coke. The level of electricity imports is not the highest
obtained, but is nevertheless higher than the one reached with the opti-
mization of employment. The output of self-production of electricity is
the highest (385 572 toes). This fact is also consistent with the highest
outputs of black liquors, other by-products, biogas and hydrogen, which
are important inputs of self-production of electricity.
The other region which can be identified on the weight space, regard-
ing the main characteristics of the corresponding solutions therein, is
between solutions 14 and 6 and solutions 4 and 5. Solutions within this
region present slight changes between them for the values of the three
objective functions. Solution 10 is representative of the main character-
istics of the solutions comprised in this region, and it is a good candidate
for a compromise solution. In this solution the level of employment is
high (5 751 091), and private consumption (13 347 213) and energy im-
ports (19 841 769) achieve reasonable values. Hydrocarbon production
achieves levels between those attained in solutions 1 and 2 and solution
3. LPG, fueloil and naphtha reach output values below those of solutions
1 and 2 and above the one obtained in solution 3. A similar behaviour is
verified in the analysis of the non-energetic output. Nevertheless some
exceptions can be pointed out, namely for metal-electrical-mechanics
which reaches a level above solutions 1 and 2. The output of electricity
has the same value reached in solution 3 {minimization of energy im-
Interactions Between Energy System and Economy Using TRIMAP 425
f3~19950000
f1~5500000
ports) and the value obtained for city gas is between those of solutions 1
and 2 and solution 3. Self-production of electricity is only 2% above the
value achieved in solution 3, and less 11% and 19% the value obtained
in solutions 1 and 2, respectively. The output of by-products is always
above the ones in solution 3 and below solutions 1 and 2. Energy im-
ports, concerning crude oil, shale oil, paraffin and petroleum coke follow
the same pattern. For the same reasons pointed out in solution 3, elec-
tricity imports reach the highest value (when compared to solution 1, 2
and 3), although in solution 10 it must be considered, in addition, that
the level of global output is higher than in solution 3.
Although the analysis has been made in terms of basic (vertex) solu-
tions only, these solutions convey sufficient information to broadly char-
acterize the policies corresponding to distinct regions in the weight space.
Nondominated edges and faces can also be identified by visual inspection
of the weight space decomposition: a nondominated face corresponding
to a point shared by several indifference regions, and a nondominated
edge corresponding to a common frontier between two indifference re-
gions. Moreover, nondominated solutions located on edges and faces can
be obtained by making a convex combination of basic (vertex) nondom-
426 AIDING DECISIONS WITH MULTIPLE CRITERIA
5. Conclusions
The energy sector is of great importance to the analysis of an economy
on a national level, because of its direct impact on final demand and the
production system, with consequences on the employment level, internal
provisions, international trade relationships and, in general, on social
well-being. A multiple objective model based on input-output analysis
has been presented, which is suited for the study of the energy sector in
the context of the economic system.
This study is aimed at outlining how the multiple objective input-
output model and the TRIMAP interactive environment can be used
to provide decision support to DMs and planners for evaluating poli-
cies taking into account macro-economic indicators, energy flows and
environmental impacts. The model is designed to accomodate for new
fuels, forms of electricity production and air pollutants. A description of
the main results achieved with some experiments using actual data has
been presented, namely paying attention to the characterization of the
distinct policies. The availability of statistical information has been the
main dificulty found in the experiments which have been carried out.
The TRIMAP interactive environment has been used to perform an
overview of the main characteristics of nondominated solutions. A pro-
gressive and selective search based on the weight space enabled to iden-
tify regions where solutions sharing the same features are located. This
approach overcomes the difficulties associated with generating methods
(by avoiding an exhaustive computation of nondominated solutions) and
some interactive methods (not requiring dichotomic responses before a
global knowledge of the nondominated solution set). The interactive
process is therefore associated with "creation" rather than "discovery"
of an optimal solution to an underlying preference structure (Roy, 1987,
1990).
References
Banco de Portugal, New Presentation of the Balance of Payments' Statistics, 1995.
(in Portuguese)
Bouyssou , D. Decision multicritere au aide multicritere?, Newletter of the European
Working Group Multicriteria Aid for Decisions, no. 2, 1993. http://www.inescc.pt
/ -ewgmcda/Bouyssou.html (in French)
Climaco, J., and Antunes C. H., TRIMAP - an interactive tricriteria linear program-
ming package, Foundations of Control Engineering, vo1.12, 101-119, 1987.
REFERENCES 427
Climaco, J., and Antunes C. H., Implementation of a user-friendly software package -
a guided tour of TRIMAP, Mathematical and Computer Modelling, vo1.12, 1299-
1309, 1989.
Dorfman, R., Samuelson, P., and Solow, R., Linear Programming and Economic Anal-
ysis, Dover Publications, New York, 1987.
INE, National Accounts, 1995. (in Portuguese)
IPCC, Revised 1996 IPCC Guidelines for National Greenhouse Gas Inventories Ref-
erence Manual. http://www.iea.org/ipcc/inv6.htm. 1996.
Leontieff, W., The Structure of the American Economy - 1919-1939, Oxford University
Press, New York, 1951.
Oliveira, C., A Multiple Objective Model for Energy Planning Based on Input-Output
Analysis, Dissertation of M.Sc. on Information Management in Organizations, Fac-
ulty of Economics, University of Coimbra, 2000. (in Portuguese)
Roy, B. Meaning and validity of interactive procedures as tools for decision making,
European Journal of Operational Research, Vol. 31(3), 297-303, 1987.
Roy, B. Decision-aid and decision-making, In C. Bana e Costa (Ed.), Readings in
Multiple Criteria Decision Aid, Springer Verlag, 17-35, 1990.
Steuer, R., Multiple Criteria Optimization: Theory, Computation and Application.
Wiley, 1986.
MULTICRITERIA APPROACH
FOR STRATEGIC TOWN PLANNING
The Case of Barcelos
Manuel L. da Costa-Lobo
Instituto Superior Tecnico, Technical University of Lisbon, Portugal
cesur@civil.ist.utl.pt
Isabel A. Ramos
University of Evora, Portugal
iar@uevora.pt
Abstract: Barcelos was one of the medium sized Portuguese towns selected to be
included in a governmental program aiming at restructuring the country urban
network. Each town had to prepare a "City Strategic Plan" as a previous
condition to get financial support for the implementation of its strategy. This
paper describes how a multicriteria methodology, enhanced with problem
structuring techniques, supported the construction of a strategy for Barcelos
in direct interaction with planners and the local elected politicians in a
decision-conferencing framework. This was a socio-technicalleaming process
that successfully implemented, in a strategic town-planning context, what
Bernard Roy defined as "decision aiding".
Key words: Decision aiding; Strategic planning; Decision conferencing; Mixing method
429
430 AIDING DECISIONS WITH MULTIPLE CRITERIA
Figure 1. Location of Barcelos Municipality in Portugal and the regional context of the town
Mapping (Eden and Ackermann, 1998; see also Bana e Costa et al., 1999b)
and AIDA - the Analysis of Interconnected Decision Areas included in the
Strategic Choice approach (Friend and Hickling, 1987). Eventually, MASP
is a type of "multimethodology" (Mingers and Gill, 1997) named
"methodology enhancement" by Mingers and Brocklesby (1997). Actually,
we enriched the multicriteria methodology with contributions borrowed
from other decision support methodologies, under a single paradigm - the
learning paradigm - that guaranteed the theoretical and practical cohesion
and consistency of the process at all stages of development (Midgley, 1997).
(See brief introductions to Cognitive Mapping, Strategic Choice and
Decision Conferencing in (Eden and Radford, 1990, appendix).)
2. Structuring
2.1 Identifying and structuring objectives
Structuring started in the middle of November 1994 with a one-month
search of development priorities and their framing in the planning process.
Objectives and proposals in previous urban plans and policy documents
were collected and systematised in collaboration with E&R. The main
references were the 1977 Oporto Region Plan, the Master Plan of Barcelos,
the ongoing planning studies for the Urban Centre of Barcelos, and also the
political measures and programs settled by the Municipal Council.
We tried to put all the ideas together and check its coherence to see if
they would be able to promote Barcelos as a meaningful regional pole, as it
was required. This first observation allowed to detect some lack of basic
conditions (such as in technical infrastructures) essential to get the results
aimed at and to avoid unbalanced developments from the social viewpoint.
As said above, E&R had already discussed with stakeholders in Barcelos
their main concerns. We also conducted several individual interviews with
politicians. To make objectives and their relationships emerge, we made use
of a "question-answer" procedure - what do you think it is important to
achieve? Why and for what is this objective important? How can it be
achieved? (Keeney, 1994) - while hand-drawing a cognitive map.
The (individual) maps and the notes taken by the planners of E&R
during their interviews were analysed together, in a workshop that took
place on 21 December 1994. It revealed that the local politicians were
mainly concerned with local strategic objectives and were neglecting that
Barcelos could be an important developing pole for the Oporto region. The
list of strategic objectives so far identified was therefore far from being
exhaustive. To take also in consideration the regional scale of strategic
436 AIDING DECISIONS WITH MULTIPLE CRITERIA
planning (cl Filiatre, 1993), the planners added to the local development
objectives of the local politicians some other objectives having in view to
strengthen the regional structure. The politicians accepted those additional
strategic objectives as they developed a shared understanding of the
importance of a correct integration of Barcelos within the regional urban
network. Using features offered by the Decision Explorer DSS (Banxia,
1997), redundancies were eliminated and all objectives assembled in
clusters in the synthesis map shown in Fig. 3. The map has a "means-ends
network-like structure" (Belton et al., 1997) with arrows indicating that one
objective influences, leads or has implications for another.
In its subsequent discussion with the participation of the Mayor the map
was several times redesigned, until four coherent, comprehensive and logic
key Barcelos' strategic development objectives were identified:
1. To improve the quality of human settlements, to create conditions to
stabilise the population.
2. To develop human resources, as a way to face future challenges.
3. Strengthening Barcelos ' importance in the regional context.
4. To improve access to and circulation capacity within Barcelos, to
allow flexible land planning.
The strategic development objectives 1 and 4 were decomposed into
more operational objectives (most of them already identified in the maps).
Multicriteria Approach for Strategic Town Planning 437
'----114. To improve acce ss to and circulation capacity within Barcelo s. 10 allow lIexibility land planning
f---_ 4 I IReg,onol
' - - -_ 4 12 Lo I
1 - -__ 4 2 IReg,onol
' - - -_ 422 Local
Medium
ononIy
N_ .
WfUt Supply (flI;gh prMsure)
Impl~I.;:I'Itc:it' 01 SGRSU
2.3 Impact-assessment
An expert-based procedure was followed to appraise the packages'
effects. "Expert-panels" (Wenstap et al., 1997; Wenstap and Carlsen, 1998)
of planners and municipal technicians in each intervention area reflected on
how each package would improve the achievement of each (bottom)
objective and selected from Table 3 the reference impact-level that in their
opinion best appraises the perceived effect.
From this process resulted the profile (a line of the Impact Table 4) of
qualitative impacts of each package in the full family of objectives. Note in
Table 4 that in general a package contributes to improve the status quo only
in a few objectives.
Figure 7. Building a value scale with MACBETH visual construction of the value scale
Multicriteria Approach for Strategic Town Planning 445
Table 5. Packages' values (Sector 5) and objectives' weights (see section 3.3)
Objective.
Sector Packag..
11 III III 1131 1131 1133 1134 1135 114 115 13 1 3 411 411 421 421 43
S5 p5.1 0 0 10 0 0 0 0 2 .5 0 0 0 0 0 0
p5.2 0 0 0 0 85 10 0 0 0 0 2 .5 0 0 0 0 0 0
p5.3 0 2.5 0 0 65 85 10 65 0 0 0 2.5 0 0 0 0 0
p5.4 0 10 0 0 65 85 10 85 0 0 0 2.5 0 0 0 0 0 0
p5.5 10 0 0 85 85 10 85 0 0 0 2.5 0 0 0 0 0
WeigbtsW·) 3.7 1.9 1.9 1.0 1.0 0.3 0.4 0 .4 6 .1 5 .6 14.7 11.0 24.0 4.9 4.4 11 .0 4.4 3 .3
Politicians
F"',-=,;;;;:;..".".,.:;:~r=-=-......,...,~C"'l"'1l Facilitator
[J
(1]
::H(i ...
::~::: :::n:~Uwn
.-.. ::;::::-:--.--
.-
Municipal
technicians Planners
The meeting started with a short description of the work done so far and
the validation of value-Table 5. Next, the MCDA-facilitator made a brief
methodological explanation about value trade-offs based on simple
examples and the group was invited to discuss the acceptability of the
principle of compensation behind an additive aggregation of value-scores. A
few comparisons of adequate combinations of impact-levels by pairs of
objectives served to validate the working hypothesis of additivity. Once
446 AIDING DECISIONS WITH MULTIPLE CRITERIA
Exhibit 1 Exhibit 2
Direct, necessary, sufficient and complementary Direct, necessary, sufficient and complementary
improvement of the status quo on improvement of the status quo on
• • • •
Health Leisure Health Leisure
or or
2nd 1st
aj bj
30 100
• • • • • • • • • •
Education Leisure Culture Social S. Health Regional Local Regional local protection
or or or or or or or or
1st 1st 3rd 3rd 5th 2nd 3rd 1sl 3rd 5th
100 100 45 45 30 45 40 100 40 30
~~ (~ @~ €~ V ~il ~w @9 0 D
Status quo Status quo
Exhibit 5 Exhibit 6
Direct, necessary, sufficient and complementary Direct, necessary, sufficient and complementary
improvement of the status quo on improvement of the status quo on
• • • • • • • • • • •
1.2.1 1.2.2 1.2.3.1 1.2.4 1.2.5 1.1 1.2.4 1.3 2 3 4.2.1
or or or or or or or or or
4th 4th 3rd 1st 2nd 6th 5th 2nd 3rd 1st 3rd
30 30 50 100 90 15 25 60 45 100 45
weights in exhibit 4 (40x 100 = 100x40). Similar cross checks were made
within each of the sub-sets of objectives in exhibits 3 and 4.
The swing weighting process was then moved a level up on the value
tree, and questions similar to the above ones were asked for the objectives
on exhibit 5 (Fig. 9). The swing on objective (1.2.3.1) Improvement of
leisure infrastructures, one of the most attractive (bottom) swings on social-
infrastructures, was included in exhibit 5, instead of a complex swing on the
objective (1.2.3) Improvement of (all) social infrastructures, to ensure an
easy judgemental transition between the two levels. Finally, moving again
the process a level up, the swing weights in exhibit 6 (Fig. 9) were obtained.
The relative weights of all objectives (shown on the last row in Table V)
were computed by re-scaling the swing weights and were validated by the
politicians. They accepted, in particular, the weight for strategic
development objectives 1 and 2 together (37% + 11 %) to be (almost) equal
to the weight for objectives 3 and 4 together (24% + 28%). Substantively,
this means that, a strategy strongly improving simultaneously the (at that
time) quality of the human settlements and human resources would
contribute to Barcelos' strategic development the same as a strategy strongly
strengthening the (at that time) Barcelos' regional role and accessibility.
resource allocation over sectors; the model included all the action-packages
(see Fig. 11), each one characterised by its cost and its benefit.
S
o
~' .:P% .J,1%
po2 1 p2. 2
~t
•
p2.J
...
p2."
S D.03%
o p3' p3 2' 1'3.) p3"
nCTOA .f SECTOR ~
.o .':1.5··"::U'·ili·"!B!tl~id.._
11'10.1 11'10.2 p10.3 p10."
Figure 10. Benefit of each action-package as a percentage of the overall benefit of all actions
~
flle
Sector 1 pl .0 Pl.l pl .2 01 .3
St.5tut quo E:w:ecu:ion
Sector 2 .,2.0. 021 p2.2 .,2.3 oH
S'~Ut~O E:-:ecution F..
Sector 3 pJ.o. DJ.2 pJ.•
.
1)3. 1 1)3.3
S~us~1? ,_J~01 F..
Sector. ~ . o. I ~. 1 ~. 2 ~3 ~
SI4'tu$ quo Execution _I- Ful
Seclor 5 0)5.0. <>5.1 05.2 p5,3 OS.•
SI.&hJ"ClUO EIo:e-etlion
Seclor 8 p5.0. 06.1 06·2 p5.3 06.·
SIbtu$ quo Elf.ee~ion Ful
Sector T 07.0. p71 07.2 p7.3 pH
Slatus quo El<eCUl:ion
Seclor 8 08.0. 06.1 06.2 08.3 06.·
SI4tus <flO E..w<ion Ful
Sector t 03.0. D5.1 09.2 D5.3
Sl.otus qvo Ful
Sector 10. 01 0.0 010..1 01D.2
Slbl.us~ Execution
Figure 11. Snapshot of the main screen of EQUITY DSS showing the structure of the
intervention sectors and the respective action-packages
450 AIDING DECISIONS WITH MULTIPLE CRITERIA
Sector COSTS
Packages Total Tota l
C P B
1 1 2 2 of5 2734 49
2 2 1 of5 5507 17 BENEFITS
3 2 1 of5 60 0
4 1 2 1 of5 3043 5
1000
........ ." .......
5 1 2 2 of6 106 2
6 1 2 2 of5 1719 49
7 1 2 1 of6 100 0
B 2 2 2 of5 250 3B
9 3 1 4 of 4 0 0
10 1 2 5 of5 700 16
455
100000
B· package 11309
COSTS (0 to 148923)
C· package 2100 145
scenario" requires a total investment of 10.1 * 109 PTE and achieves 81.7 %
of benefit.
Secl or COSTS
Packages Total Tolal
C P B
1 2 4 5 of5 15464 97
BENEFITS Weighted Preference Volu
2 1 3 2 of5 12685 31
3 3 4 4 0(5 11585 49 1000 fEr " ..••.
4
5
4 3
3 4 3
4 0(5
0(6
7003
7016 10
7
© ,o®
;.4",t P
6 4 4 5 of5 19643 136
..
7 4 6 5 of6 17900 64
500
8 5 5 5 of5 6500 107
9 4 4 4 0( 4 2900 146 :
10 5 5 5 of5 3600 171
4 5 6
51 pl .O pl .l pl .2 pl .3
53 pl.l p3.4
54 p4 .3 p4.4
56
57
1000
.... @:T .... pS.3
pl .3
pG.4
p7.4
......--
58 p83 pS.4
500
59 : p9.3
:
pl O.3
pl0.4
0
0 100000
COSTS (0 to 148923)
5. Conclusion
The Multicriteria Approach for Strategic Planning (MASP) followed in
Barcelos' strategic planning differs from the traditional planning models in
that it allowed us to deal both with the dynamic nature of the process and its
uncertainties, and with qualitative and subjective aspects related with actors'
value systems. Careful structuring was essential to identify strategic
objectives and actions. The requisite evaluation model built in direct
interaction with politicians allowed the appraisal of actions' attractiveness
in terms of the degree to which the objectives are achieved. Above all, the
process enabled the final choice to emerge naturally from the flow of the
interaction, and offered the politicians a well founded and justified strategy
for the future.
454 AIDING DECISIONS WITH MULTIPLE CRITERIA
The success of the case rests mainly on the constructive nature ofMASP:
it allowed planners to work together with politicians and to develop a shared
understanding of the issues, a sense of common purpose and a mutual
commitment to action.
Designing and implementing MASP successfully was, ultimately,
nothing but putting into practice, in a context of strategic town planning,
what Bernard Roy has been for so long wisely teaching us to be the decision
aid activity:
Decision aiding is the activity of the person who, through the use of
explicit but not necessarily completely formalized models, helps obtain
elements of responses to the questions posed by a stakeholder of a decision
process. These elements work towards clarifoing the decision and usually
towards recommending, or simply favouring, a behaviour that will increase
the consistency between the evolution of the process and this stakeholder's
objective and value system. (Roy, 1996, p. 10.)
References
Bana e Costa, e.A. (ed.) (1990), Readings in Multiple Criteria Decision Aid, Springer-
Verlag, Berlin.
Bana e Costa e.A. (1992), Structuration, Construction et Exploitation d 'un Modele
Multicritere d' Aide it la Decision, PhD thesis, Technical University of Lisbon.
Bana e Costa, C.A. (forthcoming), "The use of multicriteria decision analysis to support the
search for less conflicting policy options in a multi-actor context: case-study", Journal of
Multi-Criteria Decision Analysis - Special Issue: Multi-Criteria Decision Analysis and
Environmental Management.
Bana e Costa, e.A., Costa-Lobo, M.L. (1999), "How to help 'jumping into the darkness'?",
Journal of Multi-Criteria Decision Analysis, 8, I (12-14).
Bana e Costa, C.A., Ensslin, L., Correa, E.C., Vansnick, J.C. (1999a), "Decision support
systems in action: Integrated application in a multicriteria decision aid process", European
Journal of Operational Research, 113,2 (315-335).
Bana e Costa, C.A., Ensslin, L., Correa, E.C., Vansnick, J.C. (l999b), "Mapping critical
factors for the survival of firms: A case-study in the Brazilian textile industry", in G.
Kersten, Z. Mikolajuk, A. Yeh (eds.), Decision Support Systems for Sustainable
Development, Kluwer Academic Publishers, Dordrecht (197-213).
Bana e Costa, e.A., Vansnick, J.C. (1997), "Applications of the MACBETH approach in the
framework of an additive aggregation model", Journal of Multi-Criteria Decision Analysis,
6,2(107-114).
Bana e Costa e.A., Vansnick J.C. (1999), "The MACBETH approach: Basic ideas, software
and an application" in N. Meskens, M. Roubens (eds.), Advances in Decision Analysis,
Kluwer Academic Publishers, Dordrecht (131-157).
Bana e Costa, C.A., Vansnick, J.e. (2000), "Cardinal value measurement with MACBETH",
in S.H .. Zanakis, G. Doukidis, C. Zopounidis (eds.), Decision Making: Recent
Developments and World-wide Applications, Kluwer Academic Publishers, Dordrecht
(317-329).
Multicriteria Approach for Strategic Town Planning 455
Phillips, L.D. (1990), "Decision analysis for group decision support", in e. Eden, J. Radford
(eds.), Tracking Strategic Problems, SAGE Publications, London (142-150).
Phillips, L.D., Phillips, M.e. (1993), "Facilitated work groups: Theory and practice", Journal
of the Operational Research Society, 44, 6 (533-549).
Rosenhead,1. (ed.) (1989), Rational Analysis for a Problematic World: Problem Structuring
Methods for Complexity, Uncertainty and Conflict, John Wiley & Sons, Chichester.
Roy, B. (1996), Multicriteria Methodology for Decision Aiding, Kluwer Academic
Publishers, Dordrecht.
Roy, B. (1999), "Decision-aiding today: What should we expect?", in T. Gal, T. Stewart, T.
Hanne (eds.), Multicriteria Decision Making, Advances in MCDM - Models, Algorithms,
Theory, and Applications, Kluwer Academic Publishers, Dordrecht (1-1-1-35).
Schein, E.H. (1999), Process Consultation Revisited - Building the Helping Relationship,
Addison-Wesley, Reading, MA.
Thompsom Jr., A.A., Strickland III, A.J. (1999), Strategic Management: Concepts and Case
(Eleventh Edition), IrwinlMcGraw-Hill, Boston.
Vansnick, J.C. (1990), "Measurement theory and decision aid", in e.A. Bana e Costa (ed.),
Readings in Multiple Criteria Decision Aid, Springer-Verlag, Berlin (81-100).
Watson, S.R., Buede, D.M. (1987), Decision Synthesis: The Principles and Practice of
Decision Analysis, Cambridge University Press, Cambridge.
Wenstl'Jp, F., Carlsen A.1. (1998), "Using decision panels to evaluate hydropower
development projects" in E. Beinat, P. Nijkamp (eds.), Multicriteria Analysis for Land-
Use Management, Kluwer Academic Publishers, Dordrecht (179-195).
Wenstl'Jp, F., Carlsen, AJ., Berglang, 0., Magnus, P. (1997), "Valuation of environmental
goods with expert panels", in J. Climaco (ed.), Multicriteria Analysis, Springer-Verlag,
Berlin (539-548).
von Winterfeldt, D., Edwards, W. (1986), Decision Analysis and Behavioral Research,
Cambridge University Press, Cambridge.
MEASURING CUSTOMER SATISFACTION
FOR VARIOUS SERVICES USING
MULTICRITERIA ANALYSIS
Yannis Siskos
Technical University of Crete. Greece
siskos@hercules.ergasya.tuc.gr
Evangelos Grigoroudis
Technical University of Crete. Greece
dsslab@dias.ergasya.tuc.gr
1. Introduction
Customer satisfaction is one of the most important issues concerning
business organisations of all types, which is justified by the customer-
orientation philosophy and the main principles of continuous improvement
of modem enterprises. For this reason, customer satisfaction should be
measured and translated into a number of measurable parameters. Customer
458 AIDING DECISIONS WITH MULTIPLE CRITERIA
(1)
where the value functions y" and Xi" are normalised in the interval
[0,100], n is the number of criteria, and bi is a positive weight of the i-th
criterion.
In several cases, as presented in Sections 4-5, it is useful to assume a
value or treelike structure of criteria, also mentioned as "value tree" or
"value hierarchy" (Keeney and Raiffa, 1976; Keeney, 1996; Kirkwood,
1997).
functions X j' . The main objective of the method is to achieve the maximum
consistency between the value function '" and the customers' judgements
" . Based on the modelling approach presented in the previous section and
introducing a double-error variable (see Figure 2), the ordinal regression
equation becomes as follows:
where Y' is the estimation of the global value function y', and u+ and u-
are the overestimation and the underestimation errors, respectively.
(J+
J
y'm
(J.-
J
y']
y
o L---~--_+--~~--+_--~---4--~
j=1
subject to
(4)
m=1
i=1 k=1
Zm ~ 0, W ik ~ 0 'r;j m,i,k
+>0 , (Jj->0
(Jj _ _ for j=I,2, ... ,M
where M is the size of the customers sample, n is the number of criteria, and
xl, yj are the j-th level on which variables X; and Yare estimated.
The preference disaggregation methodology includes also a post
optimality analysis stage in order to overcome the problem of model
stability. The final solution is obtained by exploring the polyhedron of
multiple or near optimal solutions, which is generated by the constraints of
the previous linear program. This solution is calculated by n linear programs
(equal to the number of criteria) of the following form:
LW
cxl-i
1 ~
IS m ·m
=-~py
100 m=!
(6)
= -Ip; x;
1 k 'k •
= 1,2, ... ,n
CX;
S; for 1
100 k=!
where Sand Si are the average global and partial satisfaction indices, and pm
and p/ are the frequencies of customers belonging to the ym and x/
satisfaction levels, respectively.
It can be easily observed in equation (6) that the average satisfaction
indices are basically the mean value of the global and partial satisfaction
functions. So, these indices can give the average level of satisfaction value
globally and per criterion.
for a> 2
(7)
where a and a; are the number of satisfaction levels in global and partial
satisfaction functions, respectively.
It should be mentioned that these indices are normalised in the interval
[-1, 1] while the following possible cases hold:
Measuring Customer Satisfaction 463
y'a
IOO(m-I) 'm
-y a
IOO(m-I) a~./
a-I
y'm
y
y'
where band S are the mean values of the criteria weights and the average
satisfaction indices, respectively.
This way, the cut-off level for axes is recalculated as the centroid of all
points in the diagram. This type of diagram is very useful, if points are
concentrated in a small area because of the low-variation that appears for the
average satisfaction indices (e.g. case of a high competitive market)
These diagrams are also mentioned as decision, strategic, perceptual, and
performance-importance maps (Dutka, 1995; Customers Satisfaction
Council, 1995; Naumann and Giel, 1995), or gap analysis (Hill, 1996;
Woodruff and Gardial, 1996; Vavra, 1997), and they are similar to SWOT
analysis.
Detailed presentation of the mathematical development of the MUSA
method may be found in Grigoroudis and Siskos (2001), and Siskos et al.
(1998).
Measuring Customer Satisfaction 465
Personnel
• Access
•
Product/Service •
Credibility
•
Low High
RELATIVE IMPORTANCE
100~----------------------------------~---,
75
50
25
0.0
O+-------,-------,--------r-------r------~
All these results can be explained by the way the course of studies is
implemented:
a) During the lSI year, students are basically taught elementary subjects
(mathematics, sociology, etc).
Measuring Customer Satisfaction 471
b) At the beginning of the 3rd year, students have to choose the sector of
studies they will follow. This will affect in great extent all of their next
choices.
Segmentation satisfaction analysis is performed through the
implementation of the MUSA method in each student's cluster separately.
For this reason, the fitting and the stability level of the results may vary
causing a problem of "inconsistency" when trying to compare global with
segmentation analysis results. In this particular application, the problem
mainly concerns the average satisfaction indices due to the high error level
in the global satisfaction analysis (the global set is less homogenous than the
segments of students).
Service in store
•
•
Customer loyalty services
Product/Service
Image
•
Stores •
GSM network •
•
Pricing
Low High
RELATIVE IMPORTANCE
The highest satisfaction indices appear for the criteria of "Tidiness" and
"Comfort/Service quality", 94.2% and 90.1% respectively. Also,
customers seem to give higher importance to these criteria.
The rest of the criteria have a low level of importance for the customers
(4.4-6.6%), while the performance of the company is rather modest
(average satisfaction indices 58-76%).
Regarding the improvement efforts for the airline company, an inspection
of the action diagram (Figure 9) reveals that there is no particularly critical
satisfaction dimension calling for an immediate improvement. Nevertheless,
the improvement priorities should be focused on the criteria with the lowest
satisfaction indices. Assuming that the company has small improvement
margins for the "Price" criterion due to the competition, its efforts should be
focused on the credibility and the provided service.
Tidiness
•
•
Comfort/Service quality
Credibility •
Service •
Pricing •
Low High
RELATIVE IMPORTANCE
selected customer sample and it took place in three restaurants of the fast
food company located in different towns.
The value hierarchy of customers satisfaction dimensions presented in
Figure 10 indicates the set of criteria and subcriteria used in the analysis.
The main satisfaction criteria include:
1. Personnel: this criterion includes all the characteristics concerning
personnel (skills and knowledge, responsiveness, friendliness,
communication and collaboration with customers, etc).
2. Product: this criterion refers mainly to the offered products (quality and
quantity of food, variety of dishes, and prices).
3. Service: this criterion refers to the service offered to the customers; it
includes the appearance and the cleanliness of the stores, the waiting time
during busy and non-busy hours, and the service time.
4. Access: the location and number of stores, as well as parking availability
are included in this criterion.
86% and 92%. Given the high competitive conditions of the market, this
performance is not considered relatively high.
The detailed results of Table 12 reveal the following:
The most important criterion, with a significant importance level of
45.2%, is "Product". Customers do not consider important the rest of the
criteria.
The low weight for the "Access" criterion can be explained by the fact
that the main competitors have no better performance in this particular
criterion.
•
Personnel
Service-
-
Access
Product
-
Low High
RELATIVE IMPORTANCE
6. Conclusions
The original applications presented in this paper illustrate the
implementation of the preference disaggregation MUSA method in several
business organisations from the public and the private sector. The most
important results include:
- the determination of the weak and the strong points of the business
organisation,
- the performance evaluation of the company (globally and per
criteria/subcriteria), and
- the identification of the distinctive critical groups of customers.
The applications show that the MUSA method can measure and analyse
customer satisfaction in a very concrete way, and thus it may be integrated in
any business organisation's total quality approach. Several applications of
the method in original customer satisfaction surveys can be found in
Grigoroudis et al. (1999a, 1999b), Mihelis et al. (2001), and Siskos et al.
(2001). Also, the MUSA method may be used in a similar way to measure
and analyse employee satisfaction (Grigoroudis, 1999). Furthermore,
analysing clients' preferences and expectations is the basic step to evaluate
customer loyalty.
The installation of a permanent customer satisfaction barometer is
considered necessary, given that it allows the establishment of a
benchmarking system (Edosomwan, 1993). Thus, the implementation of the
MUSA method through a period of time can serve the concept of continuous
improvement.
Grigoroudis and Siskos (2001) propose several extensions and future
research regarding the MUSA method. Among others, the comparative
analysis between the results of the MUSA method and the financial indices
(market share, profit, etc) of a business organisation can help the
development of business strategies and the evaluation of the cost of quality.
It should be mentioned that, although customer satisfaction is a necessary
but not a sufficient condition for the financial viability, several researches
have shown that there is a significant correlation among satisfaction level,
customer loyalty, and profitability (Dutka, 1995; Naumann and Giel, 1995).
Acknowledgements
The authors are thankful to all the students of the National Technical
University of Athens and the University of Cyprus who helped customer
satisfaction survey conduct and analysis. Especially, the authors would like
to thank: P. Apseros, N. Eteokleous, T. Georgiou, L. Giannoudiou,
Measuring Customer Satisfaction 481
References
Customers Satisfaction Council (1995). Customer Satisfaction Assessment Guide, Motorola
University Press.
Dutka A. (1995). AMA Handbook of customer satisfaction: A complete guide to research,
planning and implementation, NTC Business Books, Illinois
Edosomwan J. A. (1993). Customer and market-driven quality management, ASQC Quality
Press, Milwaukee.
Gerson R. F. (1993). Measuring customer satisfaction: A guide to managing quality service,
Crisp Publications, Menlo Park.
Grigoroudis E. (1999). Measuring and analysing satisfaction methodology: A multicriteria
aggregation-disaggregation approach, Ph.D. Thesis, Technical University of Crete,
Department of Production Engineering and Management, Chania (in greek).
Grigoroudis E. and Y. Siskos (2001). Preference disaggregation for measuring and analysing
customer satisfaction: The MUS A method, European Journal of Operational Research (to
appear).
Grigoroudis E., A. Samaras, N. F. Matsatsinis and Y. Siskos (1999a). Preference and
customer satisfaction analysis: An integrated multicriteria decision aid approach,
Proceedings of the 5th Decision Sciences Institute's International Conference on
Integrating Technology & Human Decisions: Global Bridges into the 21st Century,
Athens, Greece, (2), 1350-1352.
Grigoroudis E., 1. Malandrakis, 1. Politis and Y. Siskos (1999b). Customer satisfaction
measurement: An application to the Greek shipping sector, Proceedings of the 5th
Decision Sciences Institute's International Conference on Integrating Technology &
Human Decisions: Global Bridges into the 21st Century, Athens, Greece, (2), 1363-1365.
Hill, N. (1996). Handbook of customer satisfaction measurement, Gower Publishing,
Hampshire.
Jacquet-Lagreze E. and 1. Siskos (1982). Assessing a set of additive utility functions for
multicriteria decision-making: The UTA method, European Journal of Operational
Research, (10), 2, 151-164.
Keeney R. L. (1996). Value-focused thinking: A path to creative decision making, Harvard
University Press.
Keeney R. L. and H. Raiffa (1976). Decisions with multiple objectives: Preferences and value
tradeoffs, John Wiley and Sons, New York.
Kirkwood G. W. (1997). Strategic decision making, Duxbury Press, Belmont.
Mihelis G., E. Grigoroudis, Y. Siskos, Y. Politis and Y. Malandrakis (2001). Customer
satisfaction measurement in the private bank sector, European Journal of Operational
Research, (130), 2, 347-360.
Naumann E. and K. Giel (1995). Customer satisfaction measurement and management: Using
the voice ofthe customer, Thomson Executive Press, Cincinnati.
Siskos 1. (1985). Analyse de regression et prograrnmation lineaire, Revue de Statistique
Appliquee, 23 (2),41-55.
482 AIDING DECISIONS WITH MULTIPLE CRITERIA
Jean-Pierre Brans
Vrije Universiteit Brussel - Center for Statistics and OR, Belgium
jpbrans@vub.ac.be
Pierre L. Kunsch
Vrije Universiteit Brussel - Center for Statistics and OR, Belgium
p.kunsch@nirond.be
Bertrand Mareschal
Universite Libre de Bruxelles - Mathematiques de la Gestion, Belgium
bmaresc@ulb.ac.be
Socio-economic process
(Real-World)
Micro
Quantitative Approaches
Virtual space
Decision maker(s)
- Responsible of the R.W.
J Analyst(s)
- Modelling
objectives - Describe
- Decision stick - Understand
Research (OR) on the other hand. However up to now both fields are
not properly bridged. An area of interest and useful research lies in front
of us, perhaps for several decades. It is essential for the management
process as a whole. It is also the purpose of this paper.
I
I
I
I
Variable I
I
I
I
I
I
: ---------
I _--
+ I1 _____ - _ - - - - - - -
~------------------------------~----------------~Time
This technique, though simple and often useful, is of course not very
reliable as it depends on the chosen extrapolating process.
Future
Past
W 1-
-- --------
------
Neuron
Inputs Wi
- Outputs
--
(Black bOx)
: k
-
W
Figure 5. Action of a neuron. Wi'S are weights to be calibrated during the learning
process
Future
- Moving structures
- Chaotic evolution
- Catastrophic evolution
Present
Phase I:
Step 1: Mental Model.
Information System
Figure 7. Overview of the ll-step management procedure starting with the set-up
of the Information System (IS)
- Auxiliary variables.
-""0
I I
~-~~
W _ "',
'\' 0
'.\ \ ~
I
I
Suppose the system includes R decision variables (Dr, r = 1,2, ... , R).
A strategy S is a numerical positioning of these decision variables:
As the R decision variables (Dr, r = 1,2, ... ,R) may vary continu-
ously and each strategy is a numerical positioning of the decision vari-
ables, the set of possible strategies is usually a continuous set in a R
dimensional space. In most of the cases, it will not be appropriate to
consider all the possible strategies but only a suitable discrete subset of
some of them. Each strategy is a different managerial option, a scenario
of virtual reality.
Let us suppose a set of N possible strategies is being considered by the
decision-makers, the analysts, the consultants or the experts in function
of realistic dynamic values given to the decision variables. In order
to avoid unnecessary complexity, it is recommended to keep N rather
small, say for instance N < 20. The most appropriate strategy for the
management of the system must then be selected, in t = 0, among the
considered strategies (Si,; i = 1, ... , N):
G
8
Figure 9. The geometric representation in the Gaia-plane (A = representation of
actions; C = representation of criteria)
Several particularly user-friendly sensitivity tools allow to modify the
weights and to appreciate how the best choice could be modified.
5. Conclusion
In this paper we describe a dynamic Decision-Aid procedure to pi-
lot and restructure socio-economic systems. SYSTEM DYNAMICS is
500 AIDING DECISIONS WITH MULTIPLE CRITERIA
used to model the complex structures of the systems and the MCDA
PROMETHEE-GAIA procedure to select appropriate strategies. Long-
term forecasts are used to obtain a view on the evolution of the system
and to define strategic objectives. In many cases profound re-engineering
becomes indispensable. Extrinsic control feedback loops are then created
to achieve control by structure. Short-term evolutions are monitored for
the short-term on-line control of the system.
The selected strategies are never definitively fixed as they can be
adapted or changed according to the observed evolution. A watchdog
alarm procedure imposes to reconsider the management policy in emer-
gency cases in presence of uncontrolled chaotic events or of catastrophic
changes in the environment. Note that in case the adopted strategies
cannot be implemented for real, the control by structure can still be
used to investigate potential scenarios. It then provides a particularly
useful prospective tool in turbulent and uncertain environments Kunsch
et al.(1999), .
In our statistics and OR Lab of the VUB in Brussels we are applying
this management procedure in various fields:
I...............................................................................................................1+-------.
Phase I: Informatin System - Data
,
! I I !
t ...................................................................! . . . . . . . . . . . . . . . . . . .J
1. Mental Model
No
TI =0
Yes
rPh::;:·ii:·········..···········----···············f··············.................
i System 2. Influence diagram
i Dynamics
: Modelling ..
3.
4.
No
5.
Yes
rph;~~·iiI;···················
i Strategies 6.
7.
8.
No
,...................................................................
i Phase IV: r----.......:~---.....,
1Monitoring 9.
!and Control
lO.
11.
Figure 10. Flow-chart of the 4-stage methodology for the management of complex
socio-economic systems.
502 AIDING DECISIONS WITH MULTIPLE CRITERIA
References
[Brans (1996)] Brans J.P. (1996), "The space of freedom of the decision-maker mod-
elling the human brain", European Journal of Operational Research 92(1996),
593-602.
[Brans et al., 1998] Brans J.P., Macharis C., Kunsch P.L., Chevalier A., and
Schwaninger M., "Combining multicriteria decision aid and system dynamics for
the control of socio-economic processes. An iterative real-time procedure" , Euro-
pean Journal of Operational Research 109(1998), 428-441.
[Brans et al.(1994)] Brans J.P. and Mareschal B., "The PROMCALC and GAIA De-
cision Support System", Decision Support Systems 12(1994), 297-310.
[Brans et al.(1985)] Brans J.P. and Vincke P., "A preference ranking organiza-
tion method. The PROMETHEE method for MCDM", Management Science
31(1985), 647-656.
[Checkland (1981)] Checkland P. (1981), System Thinking, Systems Practice, Chich-
ester, Wiley.
[Forrester(1968)] Forrester J.W. (1968), Principles of systems, Wright Allen Press.
[Forrester (1992)] Forrester J.W., "Policies, Decision and Information Sources for
Modelling", European Journal of Operational Research 59(1992), 42-63.
[Kunsch et al. (1998)] Kunsch P.L. and Brans J.P. (1998), "Une methodologie multi-
critere et dynamique d'elaboration, d'evaluation et de contrOle de politiques et de
projets du developpement durable. Application un cas de reduction des emissions
de C02", Proceedings of the 1st international symposium of the APREMA, Uni-
versit de Corse, Corte
[Kunsch et al.(2000)] Kunsch P.L., Chevalier A., and Brans J.P., (2000), "Comparing
the "Adaptive Control Methodology" (ACM) to a case study in financial plan-
ning". VUB working paper STOOTW /294 May 2000, accepted for publication
by the European Journal of Operational Research.
[Kunsch et al.(1999)] Kunsch P.L., Springael J. and Brans J.P., "An adaptive mul-
ticriteria control methodology in sustainable development. Case study: a C02
ecotax" to appear in Colorni, Parrucini, and Roy (Eds), Euro Courses Environ-
mental Management Volume, Kluwer Academic Publishers for the Commission
of the European Communities.
[Mareschal et al.(1988)] Mareschal B. and Brans J.P., "GAIA. Geometrical represen-
tation for MCDM ", European Journal of Operational Research 34(1988), 69-77.
[Richardson et al.(1981)] Richardson G. P. and Pugh III A.L. (1981), Introduction to
System Dynamics Modelling, Productivity Press, Portland, OR.
[Roy (1985)] Roy B., (1985), Methodologie multicritere d' aide Ii la decision, Econom-
ica, Paris 423 p.
[Simon (1990)] Simon H.A., "Prediction and prescription in systems modelling", Op-
erations Research 38(1990), 7-14.
[Springael et al.(2000)] Springael J., Kunsch P.L. and Brans J.P., (2000), "Traffic
crowding in cities: Control with urban tolls and flexible working hours. A group
multicriteria aid and system dynamics approach" Presented at the 14th ORBEL
conference at FUCAM, Mons, January 2000. Submitted for publication to JOR-
BEL, Brussels.
VI
MULTI-OBJECTIVE
MATHEMATICAL PROGRAMMING
METHODOLOGIES FOR SOLVING
MULTIOBJECTIVE COMBINATORIAL
OPTIMIZATION PROBLEMS
Jacques Teghem
Faculte Poly technique de Mons, Belgium
teghem(l)mathro.fpms.ac.be
1. Introduction
It is well known that, on the one hand, combinatorial optimization
(CO) provides a powerful tool to formulate and model many optimiza-
tion problems, on the other hand, a multi-objective (MO) approach is
often a realistic and efficient way to treat many real world applications.
Nevertheless, until recently, Multi-Objective Combinatorial Optimiza-
tion (MOCO) did not receive much attention in spite of its potential
applications. One of the reasons is probably due to specific difficulties
of MOCO models. We can distinguish three main difficulties.
The first two are the same as those existing for Multi-Objective Integer
Linear Programming (MOILP) problem [Teghem86a; Teghem86b] i.e.
• the non convex character of the feasible set requires to devise spe-
cific techniques to generate the so called "non supported" efficient
solutions {LoukilOO; Teghem86a].
A particular single CO problem is characterized by some specificities of
the problem, generally a special form of the constraints; the existing
methods for such problem use these specificities to define efficient ways
to obtain an optimal solution. For MOCO problems, it appears interest-
ing to do the same to obtain the set of efficient solutions. Consequently,
and contrary to what is often done in MOLP and MOILP methods, a
third difficulty is to elaborate methods avoiding to introduce additional
constraints so that we preserve during all the procedure the particular
form of the constraints.
K
for the value of vector A E A = {A I >'k > 0 and L >'k = I}.
k=l
SE{P) is generally a proper subset of the set E(P) of all the efficient
solutions. This is still true even if the combinatorial optimization prob-
lem satisfies the so called "total unimodularity" property. It is the case
of the assignment problem (see ['fuyttensOO]).
2. Direct methods
The first idea is to use intensively classical methods for single objective
problem P existing in the literature to determine E(P). Of course, each
time a feasible solution is ob~ed the k values Zk{X) are calculated
and compared with the list E{P) containing all the feasible solutions
already obtained ~d non dominated by another generated feasible so-
lution. Clearly E{P) - called the set of potential efficient solutions -
plays the role of the so-called "incumbent solution" in single objective
methods. ~ each step, E{P) is updated and at the end of the procedure
E(P) = E{P). Such an extension of single objective method is specially
designed for enumerative procedures based on a Branch and Bound (BB)
approach. Unfortunately, in a MO framework, a node of the BB tree is
less often fathomed than in the single objective case, so that logically
such MO procedures are less efficient.
We describe below an example of such direct method - extending the well
known Martello and Toth procedure - in the context of multi-objective
knapsack problem as developed in [Ulungu97]. Problem (P) is thus
n
"max "Zk(X) = LC)k)Xj k=l, ... ,K
j=1
n
LWjXj::; W
j=l
Xj = (0,1).
The following typical definitions are used (k = 1, ... , K):
ck
• Ok: variables order according to decreasing values of u}.
J
BI \{t}
{Bo n {I, ... , t - I}) U {t}
{t+l, ... ,n}.
When the node is non fathomed, a new node of BB tree is build up for
next iteration in the following way.
Multi-Objective Combinatorial Optimization 509
3. Two-phase method
Such an approach is particularly well suited for bi-objective MOCO
problems.
This first phase is continued until all pairs (xr, X S ) of S have been
examined without extension of S.
Finally, we obtain SE(P) = SuS' as illustrated in Figure 1.
C(I) -
-
[~2 8
~ ~4 ~41' C(2) -
-
[~5 2
~ 2 31'
: ;
357 1 423 5
17 -
13 -
11- •
10
7 - •
6 9 12 16 19 22
Figure 2. The feasible points in the (Zl, Z2)-space for the didactic example
512 AIDING DECISIONS WITH MULTIPLE CRITERIA
s
,
where indices ir and jr (is and js) are such that in solution xr (Xs)
we have
Xirj = Xijr = 1 (Xi.j = Xij. = 1).
Figure 3. Test 1
solution is obtained and put in a list N Brs; at this time, the second step
is applied.
Second step. When non dominated points Zl, ... , Zm E N Brs are found
inside the triangle 6Zr Z s , the test 1 can be improved. Effectively (see
figure 4), in this test the value
(f) == . max
~=o, ... ,m
(>'1Z1,i+1 + A2Z2,i)
where Zo == Zr and Zm+1 == Zs
Figure 4. Test 2
At this moment the list NSrs contains the non supported solutions cor-
responding to the triangle 6Zr Z s .
When each triangle have been examined NSE(P) = UrsNSrs .
This two phase method has been applied for bi-objective assignment
problems of dimension varying from n = 5 till n = 50 (see [TuyttensOO]).
It appears:
We note that this second fact is different for the bi-objective knapsack
problem (see [Visee98]) in which the number of non supported solutions
increases faster compared with the number of supported efficient solu-
tions. For instance :
n E(P) SE(P) NSE(P)
10 4 2 1
100 134 18 116
200 410 36 374
300 834 55 778
400 1198 69 1129
500 1778 86 1689
4. Heuristics
As pointed out in [TuyttensOOj Ulungu94j Ulungu97j Visee98] it is
unrealistic to extend the exact methods describe above to MOCO prob-
lems with more than two criteria or more than a few hundred variablesj
the reason is that these methods are too much time consuming. Because
meta-heuristic - Simulating Annealing (SA), Tabu Search (TS), Genetic
Algorithms (GA), etc - provide, for the single objective problem, excel-
lent solutions in a reasonable time, it appeared logical to try to adapt
these metaheuristics to a multi-objective framework.
The seminal work in this direction is the Ph. D. thesis of Ulungu E.L.
in 1993 giving rise to the so called MOSA method to approximate E(P)
Multi-Objective Combinatorial Optimization 515
4.1.1 Preliminaries.
• A wide diversified set of weights is considered: different weight
vectors >..(/), l E L are generated where >..(1) = (>..~), k = 1, ... , K)
with >..~) ~ 0 Vk and fS >..~) = 1 V l E L.
k=l
This set of weights is
uniformly generated:
1 2 ... ,--,1
>"k(I) E { 0,-,-, r - 1 }
.
r r r
(a) Initialization
• A greedy step is considered to produce an initial solution Xo'
This step is problem-dependent.
• Evaluate Zk(Xo ) 'V k.
• PE(A(l» = {Xo }; Ncount = n = O.
(b) Iteration n
• Draw at random a solution Y E V(Xn).
• Evaluate Zk(Y) and determine AZk = Zk(Y) - Zk(Xn ) 'V k.
Multi-Objective Combinatorial Optimization 517
V{Xn) is a neighborhood of X n.
Solution Y;; is "better" than solution "Yj if its modification vector 6(Y;;)
is smaller than the modification vector 6 ("Yj) based on the infinite norm.
A(Y)
A* min A(X)
XEPE(>")
a) Determination of P(E(>..) ••
The following procedure is applied to generate a set of potentially effi-
cient solutions PE(>").
• Initialization:
If 0 = 1 and A{Yi)
o < min {A*, A{Xn +1)) : X n +1 ~
Yi.
Doi=i+l
b) Generation of E(P) ..
The procedure is similar to the one described in section 4.1.3
• A wide diversified set of weights is considered : different weight
vector >..(l), l E L are generated where >..(1) = {A~), k = 1, ... , K}
with >..~) E A 'if l E L. This set of weights is uniformly generated
as in MOSA method. For each of them, the procedure described
in paragraph 4.2.2a) is applied to obtain ILl lists PE{>..(I)).
ILl
• The set U P E ( >.. (I)) is filtered by pairwise comparisons in order
1=1
to remove the dominated solutions :
ILl
E(P) = /\ PE (>..(1))
1=1
5.1. Initializations
• The SA parameters: Po, O!, Nstep' Tstep and Nstop.
• Application of SA on each objective Zk individually.
The best solution found for objective Zk is noted 13k.
• Defining the interval variation [Mk' mk] of each objective Zk
Mk = Zk{13k) and mk = max Zk{13I). Mk and mk are approxi-
1=1, ... ,K
Multi-Objective Combinatorial Optimization 521
mations, respectively, of the coordinates of the ideal point and the
nadir point. The definitions of these values correspond to the case
of a minimization problem.
• modifies the satisfaction levels gk, taking into account the infor-
mation furnished by the preferred solution(s). Based on the goals
expressed by the DM and when the relation 3 k E {1, ... , K} such
that gk =1= mk holds, a new set AILI+1 of weights is defined as follow:
• Based on the goals expressed by the DM, new bounds are defined:
K
L Wk
k=l
K
<
-
<1
'V
1-
K
We have the relation 0::; L Wk ::; K.
k=l
K
When L Wk = 0 or 9k = mk V k (the DM defines no aspiration
k=l
levels), no new set of weights is defined because the complete set of
weights W is used and in this case the DM explores all the efficient
frontier.
K
The more is increase in L Wk > 0, the more is the reduction in
k=l
the explored part of the efficient frontier. The restriction set w(m)
allows to eliminate some weights becoming useless with respect to
the aspiration levels defined by the DM. In order not to become too
restrictive in the elimination of the sets of weights and to be sure
to cover the reduced part of the efficient frontier the DM wants to
explore, a not too small value is to be assigned to the parameter 'Y.
Fixing'Y to its lower value is equivalent to restricting the set w(m)
to the unique set of weights A(!L!+l). We used the value 'Y = 0.9 in
our numerical experiments.
6. Conclusions
MOCO appears as a general model with many potential applications,
and in this field there is still a large place for intensive research.
References
Ben Abdelaziz F., Chaouachi J. and Krichen S. (1999). A Hybrid Heuristic for Mul-
tiobjective Knapsack Problems, in Voss S., Martello S., Osman I. and Roucairol
C., (Eds) Metaheuristics, Advances and Trends in Local Search Paradigms for
optimization. Kluwer Academic Plublishers, 205-212.
Czyzak P. and Jaszkiewicz A. (1998). Pareto Simulated Annealing - A Metaheuristic
technique for Multiple Objective Combinatorial Optimization, Journal of Multi-
Criteria Decision Analysis, Vol. 7, 34-47.
Gandibleux X., Mezdaoui N. and Freville A. (1997). A Tabu Search procedure to
solve multiobjective combinatorial optimisation problems. In proceedings Volume
of MOPGP'96. R. Caballero and R. Steuer (Eds), Springer Verlag 291-300.
Hansen M.P. (2000). Tabu Search for Multiobjective Combinatorial Optimization: TA-
MaCa, Control and Cybernetics, vol. 29, (3), 799-818.
Loukil T., Teghem J. and Fortemps Ph. (2000). Solving multi-objective production
scheduling problems with tabu-search. Control and Cybernetics, vol. 29, (3), 819-
828.
Teghem J. and Kunsch P. (1986). A survey of techniques for finding efficient solutions
to multi-objective integer linear programming. Asia-Pacific Journal of Operations
Research, Vol. 3, 1195-1206.
Teghem J. and Kunsch P. (1986). Interactive methods for multi-objective integer linear
programming. In G. Fandel et al. (Eds), Large scale modelling and interactive
decision analysis, Springer, pp. 75-87.
Teghem J., Thyttens D. and Ulungu E.L. (2000). An interactive heuristic method for
multi-objective combinatorial optimization. Computers and Operations Research,
Vol. 27, 621-634.
Thyttens D., Teghem J., Fortemps Ph. and Van Nieuwenhuyse K. (2000). Performance
of the MOSA method for the bicriteria assignment problem. Journal of Heuristics,
6,295-310.
Ulungu E.L. and Teghem J. (1994). Multi-objective combinatorial optimization prob-
lems: a survey. Journal of Multi-Criteria Decision Analysis, Vol. 3, 83-104.
Ulungu E.L. and Teghem J. (1997). Solving multi-objective knapsack problem by a
branch and bound procedure, in Multicriteria Analysis, J. Climaco (Ed.), Springer-
Verlag, 269-278.
Ulungu E.L., Teghem J., Ph. Fortemps and Thyttens D. (1999). MOSA method: a
tool for solving MaCa problems. Journal of Multi-Criteria Decision Analysis, Vol.
8, 221-236.
Ulungu E.L., Teghem J. & Ost Ch. (1998). Efficiency of interactive multi-objective
simulated annealing through a case study. Journal of Operational Research Society,
Vol. 49, 1044-1050.
Viennet R. and Fontex M. (1996). Multi-objective combinatorial optimization using a
genetic algorithm for determining a Pareto set. International Journal of Systems
Science, Vol. 27, (2), 255-260.
REFERENCES 525
Visee M., Teghem J., Pirlot M. and Ulungu E.L. (1998). Two-phases method and
Branch and Bound procedures to solve the bi-objective Knapsack problem. Journal
of Global Optimization, Vol. 12, 139-155.
OUTCOME - BASED NEIGHBORHOOD
SEARCH (ONS)
A Concept for Discrete Vector Optimization Problems
Walter Habenicht
University of Hohenheim. Stuttgart. Germany
habenich@uni-hohenheim.de
Key words: Discrete vector optimization; Local search; Quad tree; Neighborhood search
1. Basic concepts
This paper deals with the discrete vector optimization problem (DVOP),
defined as follows: "minimize" y = f(x), x E X. With a countable decision
set X, and a vector valued outcome function f: X~ Y c mm. Restricting
attention to the outcome space, one gets the equivalent formulation of
(DVOP): "minimize" y E Y. Where Y is the countable set of feasible
outcomes. In solving DVOP one is interested in the nondominated set
N:={YEYI y'~y => Y'(;!: V}, and the efficient set eff(X):={XEXI f(X)EN},
respectively. Throughout this paper we assume, that the efficient set still has
a huge cardinality. In fact, this is the case in many combinatorial problems,
like complex routing problems or in scheduling.
The main approach in vector optimization is an interactive one.
Transforming the vector optimization problem to a classical optimization
problem by means of some scalarizing function, one realizes an interactive
search in the efficient set. The most prominent approach for discrete
528 AIDING DECISIONS WITH MULTIPLE CRITERIA
. ....... .
......... (partial) Enumeration
(param.) Optimization
Metaheuristics
Structure of tbe
Emdent set:
• Linear Ust
• Efficient Quadtree
Optimizationprocedures:
• Weighting
• Goalprogramming
• Referencepoint Approach
Neighborhood Searching:
• Tree-Neighborhood
• ~-Neighborhood
• p-Neighborhood
• cone-Neighborhood
The rationale for this approach lies in the idea, that discrete optimization
problems are often NP-hard. Hence, in the one-step procedure, we would
Outcome - Based Neighborhood Search (ONS) 529
Let x, y E mm, x:;t:y. x is called a k-successor ofy, iff k=Li: Xj>Yj i-I.
Now we can introduce the quad tree structure by adding an algorithm for
the insertion of a vector into a quad tree.
Figure 2 shows a quad tree in two dimensional space, storing the set Y =
{(5,6),(7, 7),(2,4),(6,3),(8,4),(3,7),(7, I)}. Obviously, the structure is not
unique, for a given set. It depends on the order the elements are inserted. The
numbers at the edges are the "k's" of the k-son-relation. In brackets we have
noticed them as dual numbers. This is of some interest, since the dual
representation is related to the relative position between the k-successor and
his predecessor in the following way. In the i-th position of the dual
representation of k, counted from right to left, we find aI, if and only if the
i-th component of the k-successor is greater then its predecessor. In the
Outcome - Based Neighborhood Search (ONS) 531
lower part of figure 2 we show the separation of the space induced by the
quad tree structure.
There is a strong relation between k-successorship and dominance, as
stated in the following theorem:
Theorem 1 (Habenicht(1984),p.43):
I
1 •
:H .•..
.7 H..... I:
··1:
1 ~
:
•
1
:- - . - - - - - - - - - - - -
1
-6 -"'---,..-:----~ :
.............I····r········.. ···~······················ ...
1 : 1
1 : 1
1 : 1
:~-.------------
:
1
1 1
-------,---~
: 1
: 1
1 2
With So(k):= {i I bj=O, i :;;m} with bi E {O,l}, such that k=Lj bji- I, i.e. iE
SoCk) ¢::> Xj :;; Yj. Accordingly, we define Sl(k):= {i I bj=l, i :;; m}, i.e. iE
SI(k) ¢::> Xi > Yi. Then, we can formulate another theorem, concerning
dominance relations among certain triples of vectors. This, in fact, is the
central result for the evaluation of dominance relations in quad trees.
532 AIDING DECISIONS WITH MULTIPLE CRITERIA
Theorem 2 (Habenicht(1984),p.44):
From theorem 1 it follows, that in an efficient quad tree, i.e. a quad tree
without any dominance relations among nodes, there exist no O-successors
and no (2m-I)-successors. Hence an efficient quad tree is a (2m-2)-ary rooted
tree. The implication of theorem 2 is demonstrated in figure 3. If, for
example, we want to evaluate the dominance relations of some vector z E
9t4 , which is a 5-successor of some node y, then, of course, it can be
dominated by the 5-successors of y, but additionally only by the 1- and 4-
successors ofy. On the other hand, it may dominate some 5-successors ofy,
and additionally only some 7- and 13-successors. Assuming, that all
successors of y are contained in the tree, everyone being the root of a
subtree, then, from number theoretic arguments, it can easily be shown, that
we have to view at most half of those subtrees (see Habenicht(1984),p.52).
~____~____L-~II~____- L____________~~____________~
I I
may dominate z may be dominated by z
Identify eff(Y)
(Create an efficient quad tree).
Choose a starting point y*
The main difference to other well known local search algorithms (see
Czyzak, Jaszkiewicz(1998), Glover, Laguna(1997), Osman, Kelly(1997),
Vogt, Kull, Habenicht(1996)) like simulated annealing or taboo search, lies
in the fact that in this approach the searching process is based on
neighborhoods defined in outcome space, whereas the classical approaches
use neighborhoods in decision space. Since neighborhood definitions in
outcome space, in general, do not rely on special properties of the underlying
decision problem, the applicability of this approach is quite general. The
most prominent existing approach using neighborhoods in outcome space,
too, is the Light Beam Searching(LBS) approach, introduced by
Jaszkiewicz, Slowinski(1997). In this approach special attention is given to
the derivation of a neighborhood corresponding to a certain preference
information derived from an outranking approach. The neighborhood used in
this approach is a polyhedral one. The authors do not deal with the problem
Outcome - Based Neighborhood Search (ONS) 535
Identify eff(Y)
(Create an efficient Quadtree).
Choose as the starting point y*
the root of the quadtree
rl
Detennine as neighborhood N(y*)
the (1'" -l-i)-sons of the
i-th nodes ofN(y*) Lety* :=y'
Figure 6. Tree-search
536 AIDING DECISIONS WITH MULTIPLE CRITERIA
• y. I. & 2. Iteration
o
• N(y*) I. Iteration
Chosen I. ~on
• (y*) 2. Iteration
•••
• : chosen 2. Iteration
•••
before. The search starts with the root. The neighbors are the six sons of the
root (#2 to #7). In this case the DM prefers the root. Therefore, in the 2nd
iteration the neighborhood is formed by the (2m-I-i) -sons of the i-sons of
the root (#3, #39, #16 ,#20, #12, #35). Among these nodes #12 is chosen.
Now the process stops, because #12 has no successors. In this case the
search stops in a local optimum. Obviously, #11 would have been the best
solution.
Under the aspect of complexity, tree-search is obviously optimal, since
the maximal number of steps to be performed corresponds to the longest
path in the quad tree. But there is a great risk of running into a local
optimum. Nevertheless, tree search may be a good means to reach in few
iterations at a fairly good starting solution.
3.2. P- Neighborhood
The 13 - Neighborhood N~(y*) is a distance based neighborhood. It
contains all points, whose Tschebytscheff -<iistance from y* is not greater
then 13.
Y dominates 0 }
y, ~(y.) = { y e Y: y is dominated by u
o
OJ - Yi· + P
u,- Yi· W ~
o
y,
Figure 8. (j - Neighborhood
538 AIDING DECISIONS WITH MULTIPLE CRITERIA
Infigure 9, we can see that quad trees support the identification of the 13-
Neighborhood. In this example, only the shaded nodes have to be examined,
if we use the results of theorems 1 and 2.
The disadvantage of this type of neighborhood lies in the fact, that the
cardinality of Np(y*) depends on the choice of 13. If we choose it too large,
we may get too much neighbors. On the other hand, if we choose it too
small, the neighborhood may be empty.
3.3. P - Neighborhood
We can avoid this disadvantage of the I3-Neighborhood by using a p-
Neighborhood NP(y*), which is defined as follows:
..
Figure 10. Local Optimum, using p - neighborhood (p = 5)
540 AIDING DECISIONS WITH MULTIPLE CRITERIA
3.4. Cone-Neighborhood
Cone-neighborhood is based on the consideration of the 2m_2 cones
originating at y*, that can be built by the separation of the set {1 ,... ,m} into
two non-empty subsets Sand R, as defined in figure J J. The neighborhood
consists the points nearest to y* in each cone, based on the Tschebytscheff-
distance. Hence, we have at most 2m_2 neighbors. Again, we have in general
a higher complexity for the identification of the neighborhood, because we
have to identify N 1(y*) for each of the 2m_2 cones.
kilt neighborhood-cone:
o
o
let (S,R) be a seperation of {1,2, .. ,m} and
k: = L ie R 2;-1, S,R ..0
4. Summary
In this paper, we introduced the concept of output-based neighborhood
search (ONS). As a core feature of this concept, we discussed a sequence of
neighborhood definitions with improving convergence properties, but
increasing complexity. Some experimental tests on complex routing
problems indicate, that using the following sequence of neighborhoods- tree-
neighborhood~p-neighborhood~cone-neighborhood- form a good compro-
mise between complexity and convergence of the ONS approach.
Outcome - Based Neighborhood Search (ONS) 541
References
Borges, P.C.(1999): CHESS - Changing Horizon Efficient Set Search. A simple principle for
multiobjective optimization. European Journal of Operations Research, 1999.
Czyzak,P., Jaszkiewicz,A.(l998): Pareto Simulated Annealing-a Metaheuristic Technique for
Multiple-Objective Combinatorial Optimization. In: 1. of Multi-Criteria Deciision
Analysis, vol. 7, 1998, pp. 34-47.
Finkel, R.A., Bentley, 1.L.(1974): Quad-Trees. A Data Structure for Retrieval on Composite
Keys. In: Acta Informatica, vol. 4, 1974, pp.I-9.
Glover, F, Laguna, M. (1997): Tabu Search. Boston et al. 1997.
Habenicht, W. (1983): Quad Trees, a Datastructure for Discrete Vector Optimization
Problems. In: Hansen, P (Hrsg.): Essays and Surveys on Multiple Criteria Decision
Making. (Lecture Notes in Economics and Mathematical Systems 209), Berlin et al. 1983,
pp. 136-145.
Habenicht, W. (1984): Interaktive Losungsverfahren flir diskrete Vektoroptimierungs-
probleme unter besonderer Beriicksichtigung von Wegeproblemen in Graphen.
(Mathematical Systems in Economics, Bd. 90), Konigstein 1984. (In German)
Habenicht, W. (1991): ENUQUAD - A DSS for discrete vector optimization problems.
Arbeitspapier 3, Lehrstuhl flir Industriebetriebslehre, Universitat Hohenheim 1991 (In
German)
Jaszkiewicz, A., Slowinski, R.(1997): Outranking-based interactive exploration of a set of
multicriteria alternatives. In: 1. of Multi-criteria Decision Analysis, vol. 6, 1997, pp. 93-
106.
Osman, I. H., Kelly, J. P.(Hrsg.) (1997): Meta-Heuristics: Theory & Applications, 2nd print,
Boston et al. 1997
a
Roy, B. (1985): Methodologie Multicritere d' Aide la Decision. Paris 1985 (In French)
Steuer, R.E. (1985): Multiple Criteria Optimization: Theory, Computation, and Application.
New York et al. 1985.
Sun, M., Steuer, R.E.(l996): InterQuad: An interactive quad tree based procedure for solving
the discrete alternative multiple criteria problem. In: European Journal of Operations
Research, vol. 89, 1996, pp.462-472.
Vogt, 1., Kull, H., Habenicht, W.(1996): Konzeption eines Tabu-Search-Algorithmus flir
multikriterielle Optimierungsprobleme. Arbeitspapier 21, Lehrstuhl flir Industri-
ebetriebslehre, Universitat Hohenheim 1996. (In German)
Wierzbicki, A. (I 986): On the completeness and constructiveness of parametric characteri-
zations to vector optimization problems. In: OR Spektrum vol. 8, 1986, pp. 73-87.
SEARCHING THE EFFICIENT FRONTIER
IN DATA ENVELOPMENT ANALYSIS
Pekka Korhonen
Helsinki School of Economics and Business Administration, Finland
korhonen@hkkk.fi
Abstract: In this paper, we deal with the problem of searching the efficient frontier in
Data Envelopment Analysis. The approach developed to make a free search on
the efficient frontier in multiple objective linear programming can also be used
in DEA. The search is useful, when preference information is desired to
incorporate into efficiency analysis. For instance, the DM may want to use
some flexibility in locating a reference unit for an inefficient unit or (s)he
would like to find the most preferred input- and output-values on the efficient
frontier. Using a numerical example, we will demonstrate how to make the
search and how to use the results.
1. Introduction
Data Envelopment Analysis (DEA), originally proposed by Chames,
Cooper and Rhodes [1978 and 1979], has become one of the most widely
used methods in operations research/management science. It is easy to agree
with Bouyssou [1999] that "DEA can safely be considered as one of the
recent "success stories" in OR. .. ". An obvious reason for this success is that
DEA is a problem-oriented approach and focuses on an important task: to
evaluate performance of Decision Making Units (DMU). The analysis is
based on the evaluation of relative efficiency of comparable decision making
units. Based on information about existing data on the performance of the
units and some preliminary assumptions, DEA forms an empirical efficient
surface (frontier). If a DMU lies on the surface, it is referred to as an
efficient unit, otherwise inefficient. DEA also provides efficiency scores and
reference units for inefficient DMUs. Reference units are hypothetical units
544 AIDING DECISIONS WITH MULTIPLE CRITERIA
on the efficient surface, which can be regarded as target units for inefficient
units. A reference unit is traditionally found in DEA by projecting an
inefficient DMU radially! to the efficient surface.
From a managerial point of view, there may be a need sometimes to have
some flexibility to choose a reference unit. The reference unit found for an
inefficient unit through radial projection dominates the inefficient unit, but it
is not the only one on the efficient frontier with that property. Dominance is
an important and often desirable property, but why always choose one
specific unit. Why not to take into account the preferences of the DM? There
are several possibilities to incorporate preference information into the
analysis, for instance:
I) Term "radial" means that an efficient frontier is tried to reach by increasing the values of
the current outputs and/or decreasing the values of the current inputs in the same
proportion.
Searching the Efficient Frontier in Data Envelopment Analysis 545
the frontier. A requisite technique has already been developed for Multiple
Objective Linear Programming (MOLP) models. Because DEA- and
MOLP-models can be presented structurally in the same formulation (see,
Joro, Korhonen and Wallenius [1998]), we may apply the technique to DEA
problems as well.
The approach recommended is based on a so-called reference direction
approach originally proposed by Korhonen and Laakso [1986]. Korhonen
and Wallenius [1988] developed from the original reference direction
approach a dynamic version called Pareto Race. Actually, Pareto Race is an
interface, which makes it possible to move along the efficient surface and
freely search various target units in DEA as well. We demonstrate with a
numerical example and discuss various situations in which the approach
might be useful in DEA.
The rest of this paper is organized as follows. In Section 2, we review the
reference point/direction approach to multiple objective programming, and
demonstrate its similarity with the approaches used in the data envelopment
analysis. In Section 3, using a numerical example we demonstrate how to
make a free search on the efficient frontier in DEA. Section 4 concludes the
paper.
2) Because the results concerning u and U are valid for (~) and G) as well, for simplicity,
we often refer to u and U, although we are factually interested in results concerning (~)
and G}
546 AIDING DECISIONS WITH MULTIPLE CRITERIA
where A c 9lft is a feasible set, matrix A E 9l k ..,. is of full row rank k and
vector b E 9l k • The set of feasible values of vector u c 9l m +p is called a
I
feasible region and denoted by T = {u U = VA, A E A}. In the model (2.1),
we try to maximize the linear combination of the outputs and minimize the
linear combination of the inputs.
The efficient solutions of the model (2.1) are defined in a usual way:
s.t. u E T,
0 (UNon-Archimedean") 0 (UNon-Archimedean")
!R:
&> &>
A = {l l EI and Al Sb}
Vector II consists of the aspiration levels for outputs and If the aspiration
levels for inputs. Vectors w > 0 and w'" > 0 (w= (;) ) are the weighting
vectors for outputs and inputs, respectively. Let's denote the optimal value
of the models Z* and W*.
Korhonen and Laakso [1986] further developed the reference point
approach by parameterizing the reference point using a reference direction.
In the approach, a direction r specified by the DM is projected onto the
efficient frontier (see, e.g. Korhonen and Laakso [1986]):
&>0
A = {A. IA. E 91: and AA. Sb},
LV;x;o
;=1
s.t.: (2.5)
p
L,ukYAj
.::...k:~1_ _ :$; 1, j = 1,2, ... , n
LV;Xij
1=1
model (2.5) seems to have nothing to do with the reference point model.
However, it turns out that the model (2.5) can be first transformed into a
linear model, then the dual of that model (called an envelopment model) is
structurally like a reference point model (see, e.g. Joro, Korhonen, and
Wallenius [1998]). The weighting vectors w" and ~ and reference points If
and It are determined in a certain specific way in DEA-models as shown in
Table 1.
Let's consider now (feasible) decision making units U E T from the
perspective of data envelopment analysis. Set T is called a Production
Possibility Set in the data envelopment analysis literature. Note that in this
paper, we assume T is convex. Other assumptions are possible as well. In
DEA, we are interested to recognize efficient DMUs, which are defined as a
subset of points of set T satisfying the efficiency condition defined below:
Definition 4. A point (y*, x*) E T is weakly efficient iff there does not exist
another (y, x) E T such thaty > y* and X <x*.
the region consisting of the points surely less or equally preferred to the most
preferred solution can be specified. This region is used in value efficiency
analysis.
Let's consider the following simple numerical example, in which the data
in Table 2 are extracted and modified from a real application. Four large
super-markets A, B, C, and D are evaluated on four criteria: two outputs
(Sales, Profit) and two inputs (Working Hours, Size). "Working Hours"
refers to labor force available within a certain period and "Size" is the total
area of the super-market. (The super-markets are located in Finland.)
A B e D
Sales (106 Fim) 225 79 66 99 max
Profit (106 Fim) 5.0 0.2 1.2 1.9 max
Working Hours (10 3 h) 127 50 48 69 min
Size (103 m2) 8.1 2.5 2.3 3.0 min
Table 3a: Efficiency Analysis of Unit C with the Output-Oriented CCR-Model (1)
Table 3b: Efficiency Analysis of Unit C with the Input-Oriented CCR-Model (2) 4)
Table 3c: Efficiency Analysis of Unit C with the Combined CCR-Model (5)
A B C D Max (a) C
A.-coefficients 0.0676 0.0881 0.0000 0.4745 0.0475 Ref.U"it
Sales 225 79 66 99 -66 66.0 69.13
Profit 5 0.2 1.2 1.9 -1.2 1.2 1.26
Work Hours 127 50 48 69 48 48.0 45.72
Size 8.1 2.5 2.3 3 2.3 2.3 2.19
Each model will find a different reference unit on the efficient frontier for
C. In the first model, the output values are projected radially onto the frontier
subject to the current levels of resources. In the second model the role of the
input- and output-variables have been changed. In the last model, the
projection is made by improving the values of the output- and input-
variables simultaneously.
The reference unit is like a target for the inefficient unit. Why to choose
this specific target? Perhaps, the DM would like to incorporate some
flexibility into the selection. (S)he has many possibilities to choose a target
unit. For instance, some other dominating unit for unit C might be more
desirable. To be able to make a choice, (s)he needs help in evaluating a
certain part of the efficient frontier. (S)he may take, for instance, the input
values given, and consider possible output values. The efficient frontier in
this problem is trivial - one line. The whole line can be obtained as a convex
combination of the output values in columns I and IV in Table 4. In column
ill, we have the solution on the line resulted by the output-oriented CCR-
model in Table 3a. The DM has thus several alternatives to choose the
reference unit (s)he prefers most. All solutions on the efficient line satisfy
the input constraints; they all consume the resources less or the same amount
I II III IV
Sales 73.61 72.74 72.58 72.40
Profit 0.55 1.20 1.32 1.45
Work Hours 48.00 48.000 48.00 48.00
Size 2.30 2.30 2.30 2.30
A 0.0599 0.0709 0.0826
B 0.6533 0.1799 0.0924
D 0.2222 0.4551 0.4981 0.5436
The above considerations are very easy to perform if we have only two
criteria. The efficient frontier is piecewise linear in two dimensions, and all
efficient solutions can always be displayed visually. (The efficient frontier of
our example was exceptionally simple.) The characterization of the efficient
frontier can also be carried out in a straightforward way e.g. by first
identifying all extreme points and then using those points for characterizing
all efficient edges.
Generally, the efficient frontier cannot be characterized by enumerating
all efficient facets by means of the efficient extreme points. Even in quite
small size problems, the number of efficient points is huge. Actually, it is not
necessary to approach the problem in this way. Even if we could characterize
all efficient facets, the OM would need help to evaluate solutions on
different facets. Therefore, we recommend a free search. Using Pareto Race
the OM can freely move on the efficient frontier by controlling the speed and
the direction of motion on the frontier as explained in Section 2.
Searching the Efficient Frontier in Data Envelopment Analysis 555
Pareto Race
the DM (see, Halme, Joro, Korhonen, Salo and Wallenius [1999]). In this
problem the value efficiency scores are exceptionally the same as technical
efficiency ones.
4. Conclusion
We have shown that the approaches developed to search the efficient
frontier in multiple objective linear programming (MOLP) are useful in
analyzing the efficiency in data envelopment analysis (DEA). To
characterize the efficient frontier of a MOLP-problem, a widely used
technique is to transform the problem into a single objective problem by
using an achievement scalarizing function as proposed by Wierzbicki
[1980]. This transformation leads to a so-called reference point model in
which the search on the efficient frontier is controlled by varying the
aspiration levels of the values of the objective functions. For each given
aspiration level point, the minimization of the achievement scalarizing
function produces a point on the efficient frontier. Because the reference
point model and models used in DEA are similar, the methods based on the
reference point approach can be used in DEA as well. One of those further
developments is Pareto Race, a dynamic and visual free search type of
interactive procedure for multiple objective linear programming proposed by
Korhonen and Wallenius [1988]. The theoretical foundations of Pareto Race
are based on the reference direction approach developed by Korhonen and
Laakso [1986]. The main idea in the reference direction approach was to
parameterize the achievement scalarizing function
The search of the efficient frontier in DEA-models is desirable for
instance, when the DM would like to have more flexibility in determining a
reference unit to an inefficient unit than a radial projection principle
provides. Sometimes a DM may be interested to make a search on the
efficient frontier just for finding the most preferred unit on the frontier. The
most preferred solution is needed for instance for value efficiency analysis
proposed by Halme, Joro, Korhonen, Salo and Wallenius [1999].
References
Arnold v., Bardhan, I., Cooper, w. W., Gallegos, A., "Primal and Dual Optimality in
Computer Codes Using Two-Stage Solution Procedures in DEA", (Forthcoming in
Aronson, J. and S. Zionts (Eds.): Operations Research: Models, Methods and
Applications, Kluwer, Norwell. (A Volume in honor ofG.L. Thompson)), 1997.
Banker, R.D., Chames, A. and Cooper, W.W., "Some Models for Estimating Technical and
Scale Inefficiencies in Data Envelopment Analysis", Management Science 30, 1078-1092,
1984.
Searching the Efficient Frontier in Data Envelopment Analysis 557
Bouyssou, D., "DEA as a Tool for MCDM: Some Remarks", Journal 0/ the Operational
Research Society 50,974-978,1999.
Chames, A., Cooper, W., Lewin, A.Y. and Seiford, L.M., Data Envelopment Analysis:
Theory, Methodology and Applications...Kluwer Academic Publishers, Norwell, 1994.
Chames, A., Cooper, W.W. and Rhodes, E., "Measuring Efficiency of Decision Making
Units", European Journal o/Operational Research 2, 429-444, 1978.
Chames, A., Cooper, W.W. and Rhodes, E. "Short Communication: Measuring Efficiency of
Decision Making Units", European Journal 0/ Operational Research 3, 339, 1979.
Chames, A., W.W. Cooper, Q.L. Wei, and Z.M. Huang, "Cone Ratio Data Envelopment
Analysis and Multi-Objective Programming", International Journal 0/ Systems Science
20,1099-1118,1989.
Chames, A., W.W. Cooper, Q.L. Wei, and Z.M. Huang, "Fundamental Theorems of
Nondominated Solutions Associated with Cones in Normed Linear Spaces", Journal 0/
Mathematical Analysis and Applications 150, 54-78, 1990.
Dyson, R.G. and E. Thanassoulis, "Reducing Weight Flexibility in Data Envelopment
Analysis", Journal o/Operational Research Society 6,563-576, 1988.
Geoffrion, A., "Proper Efficiency and the Theory of Vector Maximisation", Journal 0/
Mathematical Analysis and Applications 22, pp. 618-630, 1968.
Golany, B., "An Interactive MOLP Procedure for the Extension of DEA to Effectiveness
Analysis", Journal o/Operational Research Society 39, 725-734, 1988.
Halme, M., Joro, T., Korhonen, P., Salo, S., and Wallenius, J., "A Value Efficiency Approach
to Incorporating Preference Information in Data Envelopment Analysis", Management
Science 45, 103-115, 1999.
Joro, T., Korhonen, P. and Wallenius, l, "Structural Comparison of Data Envelopment
Analysis and Multiple Objective Linear Programming", Management Science 44,962-970,
1998.
Korhonen, P., "VIG - A Visual Interactive Support System for Multiple Criteria Decision
Making", Belgian Journal 0/ Operations Research, Statistics and Computer Science 27,
3-15, 1987.
Korhonen, P., and Laakso, J., "A Visual Interactive Method for Solving the Multiple Criteria
Problem", European Journal o/Operational Research 24, 277-287, 1986.
Korhonen, P. and Wallenius, J., "A Pareto Race", Naval Research Logistics 35, 615-623,
1988.
Steuer, R.E., Multiple Criteria Optimization: Theory, Computation, and Application, Wiley, New
York,1986.
Thanassoulis E. and Dyson R.G., "Estimating Preferred Target Input-Output Levels Using
Data Envelopment Analysis", European Journal o/Operational Research 56, 80-97, 1992.
Thompson R.G., Langemeier, L.M., Lee, C-T., Lee, E., and Thrall, R.M., "The Role of
Multiplier Bounds in Efficiency Analysis with Application to Kansas Farming", Journal 0/
Econometrics 46, 93- \08, 1990.
Thompson, R.G., Singleton, Jr., F.R., Thrall, R.M., and Smith, B.A., "Comparative Site
Evaluation for Locating a High-Energy Physics Lab in Texas", Interfaces 16,35-49, 1986.
Wierzbicki, A., "The Use of Reference Objectives in Multiobjective Optimization", in G.
Fandel and T. Gal (Eds.), Multiple Objective Decision Making, Theory and Application,
Springer-Verlag, New York, 1980.
Wierzbicki, A., "On the Completeness and Constructiveness of Parametric Characterizations
to Vector Optimization Problems", OR Spektrum 8, 73-87, 1986.
558 AIDING DECISIONS WITH MULTIPLE CRITERIA
Wong Y-H.B. and J.E. Beasley, "Restricting Weight Flexibility in Data Envelopment
Analysis", Journal o/Operational Research Society 41,829-835, 1990.
Zhu, J., "Data Envelopment Analysis with Preference Structure", Journal o/the Operational
Research Society 47, 136-150, 1996.