UTA methods

Abstract

UTA methods refer to the philosophy of assessing a set of value or utility functions, assuming the axiomatic basis of MAUT
and adopting the preference disaggregation principle. UTA methodology uses linear programming techniques in order to optimally
infer additive value/utility functions, so that these functions are as consistent as possible with the global decision-maker’s
preferences (inference principle). The main objective of this chapter is to analytically present the UTA method and its variants
and to summarize the progress made in this field. The historical background and the philosophy of the aggregation-disaggregation
approach are firstly given. The detailed presentation of the basic UTA algorithm is presented, including discussion on the
stability and sensitivity analyses. Several variants of the UTA method, which incorporate differentforms of optimality criteria,
are also discussed. The implementation of the UTA methods is illustrated by a general overview of UTA-based DSSs, as well
as real-world decision-making applications. Finally, several potential future research developments are discussed.

Full text

UTA METHODS
Yannis Siskos
University of Piraeus
Department of Informatics
80 Karaoli & Dimitriou Str.
18534 Piraeus – Greece
ysiskos@unipi.g
r
Evangelos Grigoroudis, Nikolaos F. Matsatsinis
Technical University of Crete
Decision Support Systems Laboratory
University Campus, Kounoupidiana
73100 Chania – Greece
{vangelis,nikos}@ergasya.tuc.gr
Abstract: UTA methods refer to the philosophy of assessing a set of value or utility
functions, assuming the axiomatic basis of MAUT and adopting the preference
disaggregation principle. UTA methodology uses linear programming
techniques in order to optimally infer additive value/utility functions, so that
these functions are as consistent as possible with the global decision-maker’s
preferences (inference principle). The main objective of this chapter is to
analytically present the UTA method and its variants and to summarize the
progress made in this field. The historical background and the philosophy of
the aggregation-disaggregation approach are firstly given in this chapter. The
detailed presentation of the basic UTA algorithm is presented, including the
discussion on the stability and sensitivity analyses. Several variants of the
UTA method, which incorporate different forms of optimality criteria used in
the LP formulation, are also discussed. The implementation of the UTA
methods is illustrated by a general overview of UTA-based DSSs, as well as
real-world decision-making applications. Finally, several potential future
research developments of the UTA methodologies within the context of
MCDA are discussed through this chapter.
Key words: UTA methods, Preference Disaggregation, Ordinal Regression, Additive
Utility, Multicriteria Analysis

2 Chapte
r
1. INTRODUCTION
1.1 General philosophy
In decision-making involving multiple criteria, the basic problem stated
by analysts and decision-makers concerns the way that the final decision
should be made. In many cases, however, this problem is posed in the
opposite way: assuming that the decision is given, how is it possible to find
the rational basis for the decision being made? Or equivalently, how is it
possible to assess the decision–maker’s preference model leading to exactly
the same decision as the actual one or at least the most “similar” decision?
The philosophy of preference disaggregation in multicriteria analysis is to
assess/infer preference models from given preferential structures and to
address decision-aiding activities through operational models within the
aforementioned framework.
Under the term “multicriteria analysis” two basic approaches have been
developed involving:
1. a set of methods or models enabling the aggregation of multiple
evaluation criteria to choose one or more actions from a set A and
2. an activity of decision-aid to a well-defined decision-maker (individual,
organization, etc.)
In both cases, the set A of potential actions (or objects, alternatives,
decisions) is analyzed in terms of multiple criteria in order to model all the
possible impacts, consequences or attributes related to the set A.
Roy (1985) outlines a general modeling methodology of decision-making
problems, which includes four modeling steps starting with the definition of
the set A and finishing with the activity of decision-aid, as follows:
– Level 1: Object of the decision, including the definition of the set of
potential actions A and the determination of a problematic on A.
– Level 2: Modeling of a consistent family of criteria assuming that these
criteria are non-decreasing value functions, exhaustive and non-
redundant.
– Level 3: Development of a global preference model, to aggregate the
marginal preferences on the criteria.
– Level 4: Decision-aid or decision support, based on the results of level 3
and the problematic of level 1.
In level 1, Roy (1985) distinguishes four reference problem statements,
each of which does not necessarily preclude the others. These problematics
can be employed separately, or in a complementary way, in all phases of the
decision-making process. The four problematics are the following:
– Problematic α: Choosing one action from A (choice).

. UTA Methods 3
– Problematic β: Sorting the actions into pre-defined and preference-
ordered categories (sorting).
– Problematic γ: Ranking the actions from the best one to the worst one
(ranking).
– Problematic δ: Describing the actions in terms of their performances on
the criteria (description).
In level 2, the modeling process must conclude on a consistent family of
criteria
{
}
12
,,,
n
g
ggK. Each criterion is a non-decreasing real valued
function defined on A, as follows:
*
*
:[,]/ ()
iii
gA gg a ga→⊂→∈ (7-1)
where [ *
*,
ii
g
g] is the criterion evaluation scale, *i
g
and *
i
g
are the worst
and the best level of the i-th criterion respectively, ()
i
gais the evaluation or
performance of action a on the i-th criterion and ( )ag is the vector of
performances of action a on the n criteria.
From the above definitions the following preferential situations can be
determined:
(
)
()
( ) ( ) is preferred to
( ) ( ) is indifferent to
ii
ii
g
agb aba b
g
agb aba b
>⇔

=⇔


f
 (7-2)
So, having a weak-order preference structure on a set of actions, the
problem is to adjust additive value or utility functions based on multiple
criteria, in such a way that the resulting structure would be as consistent as
possible with the initial structure. This principle underlies the
disaggregation-aggregation approach presented in the next section.
This chapter is devoted to UTA methods, which are regression based
approaches that have been developed as an alternative to multiattribute
utility theory (MAUT). UTA methods not only adopt the aggregation-
disaggregation principles, but they may also be considered as the main
initiative and the most representative example of preference disaggregation
theory. Another, more recent example of the preference disaggregation
theory is the dominance-based rough set approach (DRSA) leading to
decision rule preference model via inductive learning (see chapter 15 of this
book).
1.2 The disaggregation-aggregation paradigm
In the traditional aggregation paradigm, the criteria aggregation model is
known a priori, while the global preference is unknown. On the contrary, the
philosophy of disaggregation involves the inference of preference models
from given global preferences (Figure 7-1).

4 Chapte
r
CRITERIA GLOBAL
PREFERENCE
Aggregation model
Aggregation model ?
Figure -1. The aggregation and disaggregation paradigms in MCDA (Jacquet-Lagrèze and
Siskos, 2001)
The disaggregation-aggregation approach (Jacquet-Lagrèze and Siskos,
1982; 2001; Siskos, 1980; Siskos and Yannacopoulos, 1985; Siskos et al.,
1993) aims at analyzing the behavior and the cognitive style of the Decision
Maker (DM). Special iterative interactive procedures are used, where the
components of the problem and the DM’s global judgment policy are
analyzed and then they are aggregated into a value system (Figure 7-2). The
goal of this approach is to aid the DM to improve his/her knowledge about
the decision situation and his/her way of preferring that entails a consistent
decision to be achieved.
In order to use global preference given data, Jacquet-Lagrèze and Siskos
(2001) note that the clarification of the DM’s global preference necessitates
the use of a set of reference actions
R
A. Usually, this set could be:
1. a set of past decision alternatives (
R
A
: past actions),
2. a subset of decision actions, especially when A is large ( R
AA⊂),
3. a set of fictitious actions, consisting of performances on the criteria,
which can be easily judged by the decision-maker to perform global
comparisons (
R
A: fictitious actions).
In each of the above cases, the DM is asked to externalize and/or confirm
his/her global preferences on the set
R
A
taking into account the
performances of the reference actions on all criteria.

. UTA Methods 5
Consistency of the
preference model and
DM’s judgment policy
Preference model
construction
Decision data - DM’s
global judgment policy
Criteria modeling
Decision
Problem
DM’s preferences
Problem
Aggregation of
DM’s preferences
on the criteria
Value or utility
system Decision
(a) The value system approach
DM’s preferences
Problem
Aggregation of
DM’s preferences
and construction of
outranking relations
Exploitation of the
constructed
outranking relations
Aid the DM to make
a “good” decision
(b) The outranking relation approach
(c) The disaggregation-aggregation approach
Multiobjective
mathematical
programming
formulation
DM’s satisfaction
level and/or utility
model
Multiobjective
optimization
techniques
Decision
(d) The multiobjective optimization approach
Interactive procedure
Figure -2. The disaggregation-aggregation approach vs. other MCDA approaches (Siskos and
Spyridakos, 1999)
1.3 Historical background
The history of the disaggregation principle in multidimensional/
multicriteria analyses begins with the use of goal programming techniques, a
special form of linear programming structure, in assessing/inferring

6 Chapte
r
preference/aggregation models or in developing linear or non–linear
multidimensional regression analyses (Siskos, 1983).
Charnes et al. (1955) proposed a linear model of optimal estimation of
executive compensation by analyzing or disaggregating pairwise
comparisons and given measures (salaries); the model was estimated so that
it could be as consistent as possible with the data from the goal programming
point of view.
Karst (1958) minimized the sum of absolute deviations via goal
programming in linear regression with one variable, while Wagner (1959)
generalizes the Karst’s model in the multiple regression case. Later Kelley
(1958) proposed a similar model to minimize the Tchebycheff’s criterion in
linear regression.
Srinivasan and Shoker (1973) outlined the ORDREG ordinal regression
model to assess a linear value function by disaggregating pairwise
judgments. Freed and Glover (1981) proposed goal programming models to
infer the weights of linear value functions in the frame of discriminant
analysis (problematic β).
The research on handling ordinal criteria began with the studies of Young
et al. (1976), and Jacquet-Lagrèze and Siskos (1978). The latter research
refers to the presentation of the UTA method in the “Cahiers du
LAMSADE” series and indicates the actual initiation of the development of
disaggregation methods. Both research teams faced the same problem: to
infer additive value functions by disaggregating a ranking of reference
alternatives. Young et al. (1976) proposed alternating least squares
techniques, without ensuring, however, that the additive value function is
optimally consistent with the given ranking. In the case of the UTA method,
optimality is ensured through linear programming techniques.
2. THE UTA METHOD
2.1 Principles and notation
The UTA (UTilités Additives) method proposed by Jacquet-Lagrèze and
Siskos (1982) aims at inferring one or more additive value functions from a
given ranking on a reference set
R
A
. The method uses special linear
programming techniques to assess these functions so that the ranking(s)
obtained through these functions on
R
A
is (are) as consistent as possible
with the given one.
The criteria aggregation model in UTA is assumed to be an additive
value function of the following form (Jacquet-Lagrèze and Siskos, 1982):

. UTA Methods 7
1
() ( )
n
ii i
i
upug
=
=∑
g (7-3)
subject to normalization constraints:
1
*
*
1
( ) 0, ( ) 1 1,2, ,
n
i
i
ii ii
p
ug ug i n
=
=


==∀=…

∑ (7-4)
where i
u, 1, 2, ,in=K are non-decreasing real valued functions, named
marginal value or utility functions, which are normalized between 0 and 1,
and i
p
is the weight of i
u (Figure 7-3).
()
ii
ug
*i
g
*
i
g
Figure -3. The normalized marginal value function
Both the marginal and the global value functions have the monotonicity
property of the true criterion. For instance, in the case of the global value
function the following properties hold:
[ ( )] [ ( )] (preference)
[ ( )] [ ( )] (indifference)
ua ub ab
ua ub ab
>⇔

=⇔

gg
gg
f
 (7-5)
The UTA method infers an unweighted form of the additive value
function, equivalent to the form defined from relations (7-3) and (7-4), as
follows:
1
() ( )
n
ii
i
uug
=
=∑
g (7-6)
subject to normalization constraints:

8 Chapte
r
*
1
*
()1
()0 1,2,,
n
ii
i
ii
ug
ug i n
=
=


=∀=…

∑ (7-7)
Of course, the existence of such a preference model assumes the
preferential independence of the criteria for the DM (Keeney and Raiffa,
1976), while other conditions for additivity have been proposed by Fishburn
(1966, 1967). This assumption does not pose significant problems in a
posteriori analyses such as disaggregation analyses.
2.2 Development of the UTA method
On the basis of the additive model (7-6)-(7-7) and taking into account the
preference conditions (7-5), the value of each alternative
R
aA
∈
may be
written as:
1
[ ( )] [ ( )] ( )
n
ii R
i
ua uga a aA
σ
=
′=+∀∈
∑
g (7-8)
where ( )a
σ
is a potential error relative to [()]ua
′
g.
Moreover, in order to estimate the corresponding marginal value
functions in a piecewise linear form, Jacquet-Lagrèze and Siskos (1982)
propose the use of linear interpolation. For each criterion, the interval
*
*
[,]
ii
g
g is cut into ( 1)
i
α
− equal intervals, and thus the end points j
i
g
are
given by the formula:
()
*
**
1 1,2, ,
1
j
ii ii i
i
j
gg gg j
α
α
−
=+ − ∀=
−K (7-9)
The marginal value of an action a is approximated by a linear
interpolation, and thus, for 1
() [ , ]
jj
iii
ga g g
+
∈
1
1
()
[()] () ( ) ()
j
jjj
ii
ii ii ii ii
jj
ii
ga g
uga ug ug u g
gg
+
+
−
=+ −

− (7-10)
The set of reference actions 12
{, , }
R
m
Aaaa=K is also “rearranged” in
such a way that 1
a is the head of the ranking and m
a its tail. Since the
ranking has the form of a weak order
R
, for each pair of consecutive actions
1
(, )
kk
aa
+ it holds either 1kk
aa
+
f (preference) or 1kk
aa
+
 (indifference).
Thus, if
11
( , ) [( ( )] [( ( )]
kk k k
aa ua ua
++
′′
∆= −
gg (7-11)
then one of the following holds:
11
11
( , ) iff
( , ) 0 iff
kk k k
kk k k
aa a a
aa a a
δ
++
++
∆≥

∆=

f
 (7-12)
where
δ
is a small positive number so as to discriminate significantly two
successive equivalence classes of
R
.

. UTA Methods 9
Taking into account the hypothesis on monotonicity of preferences, the
marginal values ()
ii
ug must satisfy the set of the following constraints:
1
( ) ( ) 1, 2, , 1, 1, 2, ,
jj
ii ii i i
ug ug s j i n
α
+−≥∀= −=KK (7-13)
with 0
i
s≥ being indifference thresholds defined on each criterion i
g.
Jacquet-Lagrèze and Siskos (1982) urge that it is not necessary to use these
thresholds in the UTA model (0)
i
s=, but they can be useful in order to
avoid phenomena such as 1
()()
j
j
ii ii
ug ug
+= when 1
j
j
ii
g
g
+f.
The marginal value functions are finally estimated by means of the
following linear program (LP) with (7-6), (7-7), (7-12), (7-13) as constraints
and with an objective function depending on the ( )a
σ
and indicating the
amount of total deviation:
11
11
1
*
1
[min] ( )
subject to
( , ) if
( , ) 0 if
( ) ( ) 0 and
()1
R
aA
kk k k
kk k k
jj
ii ii
n
ii
i
Fa
aa a a
aa a a
ug ug i j
ug
u
σ
δ
∈
++
++
+
=
=
∆≥
∆=
−≥ ∀
=
∑
∑
f

*
()0, ()0, ()0 , and
j
ii ii R
g
ug a a A i j
σ













=≥≥∀∈∀

(7-14)
The stability analysis of the results provided by LP (7-14) is considered
as a post-optimality analysis problem. As Jacquet-Lagrèze and Siskos (1982)
note, if the optimum *0F
=
, the polyhedron of admissible solutions for
()
ii
ug is not empty and many value functions lead to a perfect
representation of the weak order
R
. Even when the optimal value *
F
is
strictly positive, other solutions, less good for
F
, can improve other
satisfactory criteria, like Kendall’s
τ
.
As shown in Figure 7-4, the post-optimal solutions space is defined by
the polyhedron:
**
()
all the constraints of LP (7-14)
FF kF
≤+

 (7-15)
where *
()kF is a positive threshold which is a small proportion of *
F
.
The algorithms which could be used to explore the polyhedron (7-15) are
branch and bound methods, like reverse simplex method (Van de Panne,
1975), or techniques dealing with the notion of the labyrinth in graph theory,
such as Tarry’s method (Charnes and Cooper, 1961), or the method of
Manas and Nedoma (1968). Jacquet-Lagrèze and Siskos (1982), in the

10 Chapte
r
original UTA method, propose the partial exploration of polyhedron (7-15)
by solving the following LPs:
**
[min] ( ) [max] ( )
in and in 1,2, ,
polyhedron (7-15) polyhedron (7-15)
ii ii
ug ug
in


∀= …



(7-16)
F=F*+k(F*)
F=F*
polyhedron of constraints (7-14)
x2
x1
Figure -4. Post-optimality analysis (Jacquet-Lagrèze and Siskos, 1982)
The average of the previous LPs may be considered as the final solution
of the problem. In case of instability, a large variation of the provided
solutions appears, and this average solution is less representative. In any
case, the solutions of the above LPs give the internal variation of the weight
of all criteria i
g, and consequently give an idea of the importance of these
criteria in the DM’s preference system.
2.3 The UTASTAR algorithm
The UTASTAR method proposed by Siskos and Yannacopoulos (1985)
is an improved version of the original UTA model presented in the previous
section. In the original version of UTA (Jacquet-Lagrèze and Siskos, 1982),
for each packed action
R
aA
∈
, a single error ( )a
σ
is introduced to be
minimized. This error function is not sufficient to minimize completely the
dispersion of points all around the monotone curve of Figure 7-5. The
problem is posed by points situated on the right of the curve, from which it
would be suitable to subtract an amount of value/utility and not increase the
values/utilities of the others.

. UTA Methods 11
overestimation
error σ
underestimation
error σ+
01
1
2
3
.
.
.
Ranking
Global value
Figure -5. Ordinal regression curve (ranking versus global value)
In UTASTAR method, Siskos and Yannacopoulos (1985) introduced a
double positive error function, so that formula (7-8) becomes:
1
[ ( )] [ ( )] ( ) ( )
n
ii R
i
ua uga a a aA
σσ
+−
=
′=−+∀∈
∑
g (7-17)
where
σ
+ and
σ
− are the overestimation and the underestimation error
respectively.
Moreover, another important modification concerns the monotonicity
constraints of the criteria, which are taken into account through the
transformations of the variables:
1
( ) ( ) 0 1, 2, , an d 1, 2, , 1
jj
ij ii ii i
wug ug i n j
α
+
=−≥∀= = −KK (7-18)
and thus, the monotonicity conditions (7-13) can be replaced by the non-
negative constraints for the variables ij
w (for 0
i
s
=
).
Consequently, the UTASTAR algorithm may be summarized in the
following steps:
Step 1:
Express the global value of reference actions [( )]
k
uag, 1, 2 , ,km
=
K,
first in terms of marginal values ( )
ii
ug , and then in terms of variables ij
w
according to the formula (7-17), by means of the following expressions:
1
1
1
( ) 0 1,2, ,
( ) 1,2, , and 2, 3, , 1
ii
j
j
ii it i
t
ug i n
ug w i n j
α
−
=
=∀=

=∀= = −

∑
K
KK
(7-19)
Step 2:

12 Chapte
r
Introduce two error functions
σ
+ and
σ
−
on
R
A by writing for each pair
of consecutive actions in the ranking the analytic expressions:
1
111
(, ) [()] () ()
[ ( )] ( ) ( )
kk k k k
kkk
aa u a a a
ua a a
σσ
σσ
+−
+
+−
+++
∆=−+
−+−
g
g (7-20)
Step 3:
Solve the linear program:
1
11
11
1
11
[min] [ () ()]
subject to
(, ) if
( , ) 0 if
1
0, ( ) 0, ( ) 0 , and
i
m
kk
k
kk k k
kk k k
n
ij
ij
ij k k
zaa
aa a a k
aa a a
w
wa aijk
α
σσ
δ
σσ
+−
=
++
++
−
==
+−

=+



∆≥ 
∀

∆= 


=

≥≥≥∀


∑
∑∑
f
 (7-21)
with
δ
being a small positive number.
Step 4:
Test the existence of multiple or near optimal solutions of the linear
program (7-21) (stability analysis); in case of non uniqueness, find the mean
additive value function of those (near) optimal solutions which maximize the
objective functions:
1
*
1
() 1,2,,
i
ii ij
j
ug w i n
α
−
=
=∀=
∑K (7-22)
on the polyhedron of the constraints of the LP (7-21) bounded by the new
constraint:
*
1
() ()
m
kk
k
aaz
σ
σε
+−
=

+≤+

∑ (7-23)
where *
z is the optimal value of the LP in step 3 and
ε
a very small positive
number.
A comparison analysis between UTA and UTASTAR algorithms is
presented by Siskos and Yannacopoulos (1985) through a variety of
experimental data. UTASTAR method has provided better results
concerning a number of comparison indicators, like:
1. The number of the necessary simplex iterations for arriving at the optimal
solution.
2. The Kendall’s
τ
between the initial weak order and the one produced by
the estimated model.
3. The minimized criterion z (sum of errors) taken as the indicator of
dispersion of the observations.

. UTA Methods 13
2.4 A numerical example
The implementation of the UTASTAR algorithm is illustrated by a
practical example presented by Siskos and Yannacopoulos (1985). The
problem concerns a DM who wishes to analyze the choice of transportation
means during the peak hours (home-work place). Suppose that the DM is
interested only in the following three criteria:
1. price (in monetary units),
2. time of journey (in minutes), and
3. comfort (possibility to have a seat).
The evaluation in terms of the previous criteria is presented in Table 7-1,
where it should be noted that the following qualitative scale has been used
for the comfort criterion: 0 (no chance of seating), + (little chance of seating)
++ (great chance of finding a seating place), and +++ (seat assured). Also,
the last column of Table 7-1 shows the DM’s ranking with respect to the five
alternative means of transportation.
Table -1. Criteria values and ranking of the DM
Means of
transportation
Price (mu) Time (min) Comfort Ranking of the
DM
RER 3 10 + 1
METRO (1) 4 20 ++ 2
METRO (2) 2 20 0 2
BUS 6 40 0 3
TAXI 30 30 +++ 4
The first step of UTASTAR, as presented in the previous section,
consists of making explicit the utilities of the five alternatives. For this
reason the following scales have been chosen:
[]
*
1* 1
,30,16,2gg

=

[]
*
2* 2
, 40, 30, 20,10gg

=

[]
*
3* 3
,0,,,gg

=++++++

Using linear interpolation for the criterion 1
g according to formula (7-
10), the value of each alternative may be written as:
1123
[ (RER)] 0.07 (16) 0.93 (2) (10) ( )uuuuu=++++g
1123
[ (METRO1)] 0.14 (16) 0.86 (2) (20) ( )uuuuu=+++++g
12 3
12
[ (METRO2)] (2) (20) (0)
(2) (20)
uuuu
uu
=+ +
=+
g
1123
11
[ (BUS)] 0.29 (16) 0.71 (2) (40) (0)
0.29 (16) 0.71 (2)
uuuuu
uu
=+++
=+
g

14 Chapte
r
123
23
[ (TAXI)] (30) (30) ( )
(30) ( )
uuuu
uu
=+++++
=++++
g
where the following normalization conditions for the marginal value
functions have been used: 123
(30) (40) (0) 0uu u===.
Also, according to formula (7-19), the global value of the alternatives
may be expressed in terms of the variables ij
w:
11 12 21 22 23 31
11 12 21 22 31 32
11 12 21 22
11 12
21 31 32 33
[(RER)] 0.93
[(METRO1)] 0.86
[(METRO2)]
[(BUS)] 0.71
[(TAXI)]
uwwwwww
uwwwwww
uwwww
uww
uwwww
=+ ++++
=+ ++++
=+++
=+
=+++
g
g
g
g
g
According to the second step of the UTASTAR algorithm, the following
expressions are written, for each pair of consecutive actions in the ranking:
12 23 32
RER RER METRO1 METRO1
12 31 32
METRO1 METRO1 METR O2 METRO2
12 21 22
METRO2 METRO2 BUS BUS
(RER , METRO1) 0.07
(METRO1, METRO2) 0.14
(METRO2, BUS) 0.29
(BUS,TAXI
www
www
www
σσσ σ
σσσσ
σσσσ
−
−
−
++−
++−
++−
∆=+−
−++ −
∆=−++
−++−
∆=++
−++−
∆11 12 21 31 32 33
BUS BUS TAXI TAXI
)0.71
wwwwww
σσσ σ
−
++−
=+ −−−−
−++ −
Based on the aforementioned expression, a linear program according to
(7-21) is formulated, with 0.05
δ
= (see Table 7-2). An optimal solution is:
11 0.5w=, 21 0.05w=, 23 0.05w
=
, 33 0.4w= with *
[min] 0zz
=
=. This
solution corresponds to the marginal value functions presented in Table 7-3
and produces a ranking which is consistent with the DM’s initial weak order.
Table -2. Initial linear programming formulation
w11 w12 w21 w22 w23 w31 w32 w33 variables σ+ and σ- RHS
0 0.07 0 0 1 0 -1 0 -111-1000000 ≥ 0.05
0 -0.14 0 0 0 1 1 0 0 0-111-10000= 0
0 0.29 1 1 0 0 0 0 0 0 0 0 -1 1 1 -1 0 0 ≥ 0.05
1 0.71 -1 0 0 -1 -1 -1 0 00000-111-1≥ 0.05
1 1 1 1 1 1 1 1 0 000000000= 1
1 1 1 1 1 1 1 1 0 000000000 z
It should be emphasized that this solution is not unique. Through post-
optimality analysis (step 4), the UTASTAR algorithm searches for multiple
optimal solutions, or more generally, for near optimal solutions

. UTA Methods 15
corresponding to error values between *
z and *
z
ε
+
. For this reason, the
error objective should be transformed to a constraint of the type (7-23).
Table -3. Marginal value functions (initial solution)
Price Time Comfort
u1(30) = 0.000 u2(40) = 0.000 u3(0) = 0.000
u1(16) = 0.500 u2(30) = 0.050 u3(+) = 0.000
u1(2) = 0.500 u2(20) = 0.050 u3(++) = 0.000
u2(10) = 0.100 u3(+++) = 0.400
In the presented numerical example, the initial linear program has
multiple optimal solutions, since *0z=. Thus, in the post-optimality
analysis step, the algorithm searches for more characteristic solutions, which
maximize the expressions (7-22), i.e. the weights of each criterion.
Furthermore, in this particular case we have:
*0()()0
kk
zaak
σσ
+−
=⇔ = = ∀
so the error variables may be excluded from the linear programs of the post-
optimality analysis. Table 7-4 presents the formulation of the linear
programs that have to be solved during this step.
Table -4. Linear programming formulation (post-optimality analysis)
w11 w12 w21 w22 w23 w31 w32 w33 RHS
0 0.07 0 0 1 0 -1 0 ≥ 0.05
0 -0.14 0 0 0 1 1 0 = 0
0 0.29 1 1 0 0 0 0 ≥ 0.05
1 0.71 -1 0 0 -1 -1 -1 ≥ 0.05
1 1 1 1 1 1 1 1 = 1
1 1 0 0 0 0 0 0 [max] u1(g1
*)
0 0 1 1 1 0 0 0 [max] u2(g2
*)
0 0 0 0 0 1 1 1 [max] u3(g3
*)
The solutions obtained during post-optimality analysis are presented in
Table 7-5. The average of these three solutions is also calculated in the last
row of Table 7-5. This centroid is taken as a unique utility function, provided
that it is considered as a more representative solution of this particular
problem.
Table -5. Post-optimality analysis and final solution
w11 w12 w21 w22 w23 w31 w32 w33
[max] u1(g1
*) 0.7625 0.175 0 0 0.0375 0.025 0 0
[max] u2(g2
*) 0.05 0 0 0.05 0.9 0 0 0
[max] u3(g3
*) 0.3562 0.175 0 0 0.0375 0.025 0 0.4063
Average 0.3896 0.1167 0 0.0167 0.3250 0.0167 0 0.1354

16 Chapte
r
This final solution corresponds to the marginal value functions presented
in Table 7-6. Also, the utilities for each alternative are calculated as follows:
[ (RER)] 0.856
[ (METRO1)] 0.523
[ (METRO2)] 0.523
[ (BUS)] 0.473
[ (TAXI)] 0.152
u
u
u
u
u
=
=
=
=
=
g
g
g
g
g
where it is obvious that these values are consistent with the DM’s weak
order.
Table -6. Marginal value functions (final solution)
Price Time Comfort
u1(30) = 0.000 u2(40) = 0.000 u3(0) = 0.000
u1(16) = 0.390 u2(30) = 0.000 u3(+) = 0.017
u1(2) = 0.506 u2(20) = 0.017 u3(++) = 0.017
u2(10) = 0.342 u3(+++) = 0.152
These marginal utilities may be normalized by dividing every value
()
j
ii
ug by *
()
ii
ug . In this case the additive utility can be written as:
11 2 2 33
( ) 0.506 ( ) 0.342 ( ) 0.152 ( )uugugug=+ +g
where the normalized marginal value functions are presented in Figure 7.6.
3. VARIANTS OF THE UTA METHOD
3.1 Alternative optimality criteria
Several variants of the UTA method have been developed, incorporating
different forms of global preference or different forms of optimality criteria
used in the linear programming formulation.
An extension of the UTA methods, where [ ( )]ua
g is inferred from
pairwise comparisons is proposed by Jacquet-Lagrèze and Siskos (1982).
This subjective preference obtained by pairwise judgments is most often not
transitive, and thus, the modified model may be written as in the following
LP:

. UTA Methods 17
Figure -6. Normalized marginal value functions
[][]
{}
[][]
{}
()
( , ): ( , ):
1
1
1
[min]
subject to
( ) ( ) 0 if
() () 0 if
( ) ( ) ,
(
ab ab ab ba
ab a b ab a b
n
ii ii ab
i
n
ii ii ab ba
i
jj
ii ii i
ii
Fz z
uga ugb z a b
uga ugb z z a b b a
ug ug s ij
ug
λλ
=
=
+
=+
−+≥
−+−= ⇒
−≥ ∀
∑∑
∑
∑
f
f

*
1
*
)1
( ) 0, ( ) 0, 0 , and ( , )
n
i
j
ii ii ab
ug ug z ij ab R
=













=


=≥≥∀ ∈


∑
(7-24)
ab
λ
being a non negative weight reflecting a degree of confidence in the
judgment between a and b.

18 Chapte
r
An alternative optimality criterion would be to minimize the number of
violated pairs of an order
R
provided by the DM in ranking
R
′
given by the
model, which is equivalent to maximize Kendall’s
τ
between the two
rankings. This extension is given by the mixed integer LP (7-25), where
0
ab
γ
= if [ ( )] [ ( )]ua ub
δ
−≥gg for a pair ( , )ab R
∈
and the judgment is
respected, otherwise 1
ab
γ
=
and the judgment is violated. Thus, the
objective function in this LP represents the number of violated pairs in the
overall preference aggregated by ( )ug.
[][]
{}
(,)
1
1
*
1
*
[min] [max] ( , )
subject to
() () (,)
( ) ( ) ,
()1
( ) 0, ( ) 0
ab
ab R
n
ii ii ab
i
jj
ii ii i
n
ii
i
j
ii ii
FRR
uga ugb M ab R
ug ug s ij
ug
ug ug
γτ
γδ
∈
=
+
=
′
=⇔
−+⋅≥∀∈
−≥ ∀
=
=≥
∑
∑
∑
,
0 or 1 ( , )
ab
ij
ab R
γ












∀

=∀∈


(7-25)
where
M
is a large number. Beuthe and Scanella (2001) propose to handle
separately the preference and indifference judgments, and modify the
previous LP using the constraints:
[][]
{}
[][]
{}
[][]
{}
1
1
1
( ) ( ) if
() () 0
if
() () 0
n
ii ii ab
i
n
ii ii ab
i
n
ii ii ba
i
uga ugb M a b
uga ugb M
ab
uga ugb M
γδ
γ
γ
=
=
=
−+⋅≥



−+⋅≥




−+⋅≥



∑
∑
∑
f

(7-26)
The assumption of monotonicity of preferences, in the context of
separable value functions, means that the marginal values are monotonic
functions of the criteria. This assumption, although widely used, is
sometimes not applicable to real-world situations. One way to deal with non-
monotonic preferences is to divide the range of the criteria into intervals, so
that the preferences are monotonic in each interval, and then treat each
interval separately (Keeney and Raiffa, 1976). In the same spirit, Despotis
and Zopounidis (1993) present a variation of the UTASTAR method for the
assessment of non-monotonic marginal value functions. In this model, the
range of each criterion is divided into two intervals (see also Figure 7-7):

. UTA Methods 19
0
()
ii
ug
1
*ii
g
g=
2
i
g
1
i
p
i
g
+
...
i
p
ii
dg=
...
*
ii
pq
ii
gg
+
=
i
g
Figure -7. A non-monotonic partial utility function (Despotis and Zopounidis, 1993)
{
}
{}
112
*
1
2*
,,,
,,,
i
ii ii
p
iiii ii
pp pq
iiii i i
Gggg gd
Gdgg g g
++
== =

== =


K
K
(7-27)
where i
d is the most desirable value of i
g, and the parameters i
p
and i
q
are determined according to the dispersion of the input data; of course it
holds that ii i
pq
α
+=. In this approach, the main modification concerns the
assessment of the decision variables ij
w of the LP (7-21). Hence, formula (7-
19) becomes:
1
1
11
1
if 1
()
if
i
i
j
it i
t
j
ii pj
it it i i
ttp
wjp
ug
wwpj
α
−
=
−−
==
<≤


=
−<≤


∑
∑∑
(7-28)
while the conditions 1
()0
ii
ug
=
remain.
Another extension of the UTA methods refers to the intensity of the
DM’s preferences, similar to the context proposed by Srinivasan and Socker
(1973). In this case, a series of constraints may be added during the LP
formulation. For example, if the preference of alternative a over alternative
b is stronger than the preference of b over c, then the following condition
may be written:
[][]
[()] [()] [()] [()]uaub ubuc
ϕ
′′ ′′
−−−≥
gg gg (7-29)
where 0
ϕ
> is a measure of preference intensity and ()u
′
g is given by
formula (7-8). Thus, using formula (7-11), the following constraint should be
added in LP (7-14):
(,) (,)ab bc
ϕ
∆−∆≥ (7-30)

20 Chapte
r
In general, if the DM wishes to expand these preferences to the whole set
of alternatives, a minimum number of 2m
−
constraints of type (7-30) is
required.
Despotis and Zopounidis (1993) consider the case where the DM ranks
the alternatives using an explicit overall index I. Thus, formula (7-12) may
be replaced by the following condition:
11
(, ) 1,2, , 1
kk k k
aa I I k m
++
∆=−∀= −K (7-31)
Besides the succession of the alternatives in the preference ranking, these
constraints state that the difference of global value of any successive
alternatives in the ranking should be consistent with the difference of their
evaluation on the ratio scale.
In the same context, Oral and Ketanni (1989) propose the optimization of
lexicographic criteria without discretisation of criteria scales i
G, where a
ratio scale is used in order to express intensity of preferences.
Other variants of the UTA method concerning different forms of global
preference are mainly focused on:
– additional properties of the assessed value functions, like concavity
(Despotis and Zopounidis, 1993);
– construction of fuzzy outranking relations based on multiple value
functions u provided by UTA’s post-optimality analysis (Siskos, 1982).
The dimensions of the aforementioned UTA models affect the
computational complexity of the formulated LPs. In most cases, as noted by
Jacquet-Lagrèze and Siskos (1982), it is preferable to solve the dual LP sue
to the structure of these LPs. Table 7-7 summarizes the size of all LPs
presented in the previous sections, where P and I denote the number of
preference and indifference relations respectively, considering all possible
pairwise comparisons in
R
. Also, it should be noted that LP (7-25) has
(1)2mm− binary variables.
Table -7. LP size of UTA models
Model and optimality
criterion
Constraints Variables
UTA - Min sum of errors
(LP 7-14) 1
(1)
n
i
i
m
α
=
+
−
∑
1
(1)
n
i
i
m
α
=
+
−
∑
UTASTAR - Min sum of
errors (LP 7-21) m
1
2(1)
n
i
i
m
α
=
+
−
∑
UTA - Min sum of errors
from pairwise judgments
(LP 7-24)
[]
1
1 ( 1) 2 ( 1)
n
i
i
mm
α
=
+
−+ −
∑
1
2(1)
n
i
i
PI
α
=
+
+−
∑
UTA – Max Kendall’s
τ
(LP 7-25)
[]
1
1 ( 1) 2 ( 1)
n
i
i
mm
α
=
+
−+ −
∑
[]
1
(1)2 ( 1)
n
i
i
mm
α
=
−
+−
∑

. UTA Methods 21
3.2 Meta-UTA techniques
Other techniques named meta-UTA, aimed at the improvement of the
value function with respect to near optimality analysis or to its exploitation
for decision support.
Despotis et al. (1990) propose to minimize the dispersion of errors
(Tchebycheff criterion) within the UTASTAR’s step 4 (see section 2.3). In
case of a strictly positive error *
(0)z>, the aim is to investigate the
existence of near optimal solutions of the LP (7-21) which give rankings
R
′
such that *
(,) ( ,)
R
RRR
ττ
′>, with *
R
being the ranking corresponding to
the optimal value functions. The experience with the model (cf. Despotis and
Yannacopoulos, 1990) confirms that apart from the total error *
z, it is also
the dispersion of the individual errors that is crucial for *
(,)
R
R
τ
. Therefore,
in the proposed post-optimality analysis, the difference between the
maximum max
()
σ
and the minimum error is minimized. As far as the
individual errors are non-negative, this requirement can be satisfied by
minimizing the maximum individual error (the L
∞
norm) according to the
following LP:
max
*
1
max
max
max
[min]
subject to
all the constraints of LP (7-21)
() ()
()0
( ) 0
0
m
kk
k
k
k
aaz
a
ak
σ
σσ ε
σσ
σσ
σ
+−
=
+
−







+≤+


−≥

−≥ ∀
≥


∑ (7-32)
With the incorporation of the model (7-29) in UTASTAR, the value
function assessment process becomes a lexicographic optimization process.
That is, the final solution is obtained by minimizing successively the 1
L and
the L∞ norms.
Another approach concerning meta-UTA techniques refers to the
UTAMP models. Beuthe and Scanella (1996, 2001) note that the values
given to parameters
s
and
δ
in the UTA and UTASTAR methods,
respectively, influence the results as well as the predictive quality of the
models. Hence, in the framework of the research by Srinivasan and Shocker
(1973), they look for optimal values of
s
and/or
δ
in the case of positive
errors *
(0)z>, as well as when UTA gives a sum of error equal to zero
*
(0)z=.
In the post-optimality analysis step of UTASTAR (see section 2.3),
UTAMP1 model maximizes
δ
, which is the minimum difference between

22 Chapte
r
the global value of two consecutive reference actions. The name of the
model denotes that, on the basis of UTA, maximizing
δ
leads to better
identification for the relations of preference between actions.
Beuthe and Scanella (1996) have also proposed to maximize the sum
()
s
δ
+ in order to stress not only the differences of utilities between actions,
but also the differences between values at successive bounds. This more
general approach was named UTAMP2. Note that
s
corresponds to the
minimum of marginal value step ij
w in the UTASTAR algorithm. Although
the simple addition of these parameters is legitimate since both of them are
defined in the same value units, Beuthe and Scanella (2001) note that a
weighted sum formula may also be considered.
3.3 Stochastic UTA method
Within the framework of multicriteria decision-aid under uncertainty,
Siskos (1983) developed a specific version of UTA (Stochastic UTA), in
which the aggregation model to infer from a reference ranking is an additive
utility function of the form:
11
() ( )()
i
n
aajj
iiii
ij
ugug
α
δ
==
=∑∑
δ (7-33)
subject to normalization constraints (7-7), where a
i
δ
is the distributional
evaluation of action a on the i-th criterion, ( )
aj
ii
g
δ
is the probability that
the performance of action a on the i-th criterion is
j
i
g
, ( )
j
ii
ug is the
marginal value of the performance
j
i
g
, a
δ is the vector of distributional
evaluations of action a and ( )
a
uδ is the global utility of action a (see also
Figure 7-8).
)(
j
ii
gu
a
i
δ
b
i
δ
j
i
g
)(
j
ii
g
δ
evaluation
scale Gi
Figure -8. Distributional evaluation and marginal value function

. UTA Methods 23
This global utility is of the von Neumann-Morgenstern form (cf. Keeney,
1980), in the case of discrete i
g, where:
1
()1
iaj
ii
j
g
α
δ
=
=
∑ (7-34)
Of course, the additive utility function (7-30) has the same properties as
the value function:
( ) ( ) (preference)
( ) ( ) (indifference)
ab
ab
uu ab
uu ab
>⇔

=⇔


δδ
δδ
f

(7-35)
Similarly to the cases of UTA and UTASTAR described in sections 2.2-
2.3, the stochastic UTA method disaggregates a ranking of reference actions
(Siskos and Assimakopoulos, 1989). The algorithmic procedure could be
expressed in the following way:
Step 1:
Express the global expected utilities of reference actions ( )
k
a
uδ,
1, 2, ,km=K, in terms of variables:
1
()()0
jj
ij ii ii
wug ug
+
=−≥
(7-36)
Step 2:
Introduce two error functions
σ
+ and
σ
−
by writing the following
expressions for each pair of consecutive actions in the ranking:
1
1
11
(, ) ( ) () ()
( ) ( ) ( )
k
k
a
kk k k
a
kk
aa u a a
uaa
σσ
σσ
+
+−
+
+−
++
∆=−+
−+ −
δ
δ
(7-37)
Step 3:
Solve the linear program (7-21) by using formulae (7-33) and (7-34).
Step 4:
Test the existence of multiple or near optimal solutions.
Of course, the ideas employed in all variants of the UTA method are also
applicable in the same way in the case of the stochastic UTA.
3.4 UTA-type sorting methods
The extension of the UTA method in the case of a discriminant analysis
model was firstly discussed by Jacquet-Lagrèze and Siskos (1982). The aim
is to infer u from assignment examples in the context of problematic β (cf.
Roy, 1985). In the presence of two classes, if the model is without errors, the
following inequalities must hold:
[]
[]
1
2
()
()
o
o
aA u a u
aA u a u
∈⇔ ≥

∈⇔ <


g
g (7-38)

24 Chapte
r
with o
u being the level of acceptance/rejection which must be found in order
to distinguish the set of accepted actions called 1
A
and the set of rejected
actions called 2
A.
Introducing the error variables ( )
a
σ
,
R
aA
∈
, the objective is to
minimize the sum of deviations from the threshold o
u for the ill classified
actions (see Figure 7-9). Hence, ( )ug can be estimated by means of the LP:
[]
[]
01
1
02
1
1
*
1
[min] ( )
subject to
( ) ( ) 0
( ) ( ) 0
( ) ( ) and
()1
R
aA
n
ii
i
n
ii
i
jj
ii ii i
ii
i
Fa
uga u a a A
uga u a a A
ug ug s i j
ug
σ
σ
σ
∈
=
=
+
=
=
−+ ≥ ∀∈
−− ≤ ∀∈
−≥ ∀
=
∑
∑
∑
*0
( ) 0, 0, ( ) 0, ( ) 0 , and
n
j
ii ii R
ug u ug a a A i j
σ
















=≥ ≥ ≥∀∈∀


∑
(7-39)
0
u
0 1
()ug
()a
σ
A1
A2
Figure -9. Distribution of the actions A1 and A2 on u(g) (Jacquet-Lagrèze and Siskos, 1982)
In the general case, the DM’s evaluation is expressed in terms of a
classification of the reference alternatives into homogenous ordinal groups
12 q
AA AffKf (i.e. group 1
A includes the most preferred alternatives,
whereas group q
A includes the least preferred ones). Within this context, the
assessed additive value model will be consistent with the DM’s global
judgment, if the following conditions are satisfied:

. UTA Methods 25
[]
[]
[]
11
1
1
( )
() ( 2,3, , 1)
( )
lll
qq
ua u aA
uua u aAl q
ua u aA
−
−
≥∀∈

≤<∀∈= −

<∀∈


g
g
g
K (7-40)
where 12 1q
uu u
−
>>>K are thresholds defined in the global value scale
[0,1] to discriminate the groups, and l
u is the lower bound of group l
A
.
This approach is named UTADIS method (UTilités Additives
DIScriminantes) and is presented by Devaud et al. (1980), Jacquet-Lagrèze
(1995), Zopounidis and Doumpos (1997), Zopounidis and Doumpos (2001),
Doumpos and Zopounidis (2002). Similarly, to the UTASTAR method, two
error variables are employed in the UTADIS method to measure the
differences between the model’s results and the predefined classification of
the reference alternatives. The additive value model is developed to
minimize these errors using a linear programming formulation of type (7-
36). In this case, the two types of errors are defined as follows:
1.
{
}
max 0, [ ( )]
klk
uua
σ
+=−g kl
aA∀∈ (1,2, 1)lq
=
−K represents the
error associated with the violation of the lower bound l
u of a group l
A
by an alternative kl
aA∈,
2.
{
}
1
max 0, [ ( )]
kkl
ua u
σ
−
−
=−g kl
aA∀∈ (2,3, )lq
=
K represents the error
associated with the violation of the upper bound 1l
u
−
of a group l
A by an
alternative kl
aA∈.
Recently, several new variants of the original UTADIS method have
been proposed (UTADIS I, II, III) to consider different optimality criteria
during the development of the additive value classification model
(Zopounidis and Doumpos, 1997; Zopounidis and Doumpos, 2001;
Doumpos and Zopounidis, 2002). The UTADIS I method considers both the
minimization of the classification errors, as well as the maximization of the
distances of the correctly classified alternatives from the value thresholds.
The objective in the UTADIS II method is to minimize the number of
misclassified alternatives, whereas UTADIS III combines the minimization
of the misclassified alternatives with the maximization of the distances of the
correctly classified alternatives from the value thresholds.
In the same context, Zopounidis and Doumpos (2000a) proposed the
MHDIS method (Multi-group Hierarchical DIScrimination) extending the
preference disaggregation analysis framework of the UTADIS method in
complex sorting/classification problems involving multiple-groups. MHDIS
addresses sorting problems through a hierarchical (sequential) procedure
starting by discriminating group 1
A from all the other groups
23
{,,, }
q
AA AK, and then proceeding to the discrimination between the
alternatives belonging to the other groups. At each stage of this
sequential/hierarchical process, two additive value functions are developed
for the classification of the alternatives. Assuming that the classification of

26 Chapte
r
the alternatives should be made into q ordered classes 12 q
AA AffKf,
2( 1)q− additive value functions are developed. These value functions have
the following additive form:
1
1
() ( )
() ( )
n
llii
i
n
llii
i
uug
uug
=
=
=



=


∑
∑
g
g

(7-41)
where l
u measures the value for the DM of a decision to assign an
alternative into group l
A
, whereas the l
u corresponds to the classification
into the set of groups 12
{, ,,}
lll q
AAA A
++
=
K and both functions are
normalized in the interval [0,1] .
The rules used to perform the classification of the alternatives have the
following form:
11 1
22 2
1(1) 1
if () () then
else if ( ) ( ) then
else if ( ) ( ) then
else
kkk
kkk
qk q k k q
kq
ua u a a A
ua u a a A
ua u a a A
aA
−− −
>∈
>∈



>∈

∈




KKKKKKKKKKKKK (7-42)
The development of all value functions in the MHDIS method is
performed through the solution of three mathematical programming
problems at each stage l of the discrimination process 1, 2, 1lq
=
−K.
Initially, a LP is solved to minimize the magnitude of the classification
errors (in distance terms similarly to the UTADIS approach). Then, a mixed-
integer LP is solved to minimize the total number of misclassifications
among the misclassifications that occur after the solution of the initial LP,
while retaining the correct classifications. Finally, a second LP is solved to
maximize the clarity of the classification obtained from the solutions of the
previous LPs.
3.5 Other variants and extensions
In all previous approaches, the value function was built in a one-step
process by formulating a LP that requires only the DM’s global preferences.
In some cases, however, it would be more appropriate to build such a
function from a two-step questioning process, by dissociating the
construction of the marginal value functions and the assessment of their
respective scaling constants.
In the first step, the various marginal value functions are built outside the
UTA algorithm. These functions may be facilitated, for instance, by

. UTA Methods 27
proposing specific parametrical marginal value functions to the DM and
asking him/her to choose the one that matches his/her preferences on that
specific criterion. Those functions should be normalized according to (7-4)
conditions. Generally, the approaches applied in this construction step are:
a) techniques based on MAUT theory and described by Keeney and Raiffa
(1976), and Klein et al. (1985),
b) the MACBETH method (Bana e Costa and Vansnick, 1994, 1997; Bana e
Costa et al., 2001),
c) the Quasi-UTA method by Beuthe et al. (2000), that uses “recursive
exponential” marginal value functions, and
d) the MIIDAS system (see section 4) that combines artificial intelligence
and visual procedures in order to extract the DM’s preferences (Siskos et
al., 1999).
In the second step, after the assessment of these value functions, the DM
is asked to give a global ranking of alternatives in a similar way as in the
basic UTA method. From this information, the problem may be formulated
via a LP, in order to assess only the weighting factors i
p
of the criteria
(scaling constants of criteria). Through this approach, initially named UTA
II model (Siskos, 1980), formula (7-11) becomes:
{}
1
(,) [ ()] [ ()]
( ) ( ) ( ) ( )
n
iii ii
i
ab p u g a u g b
aabb
σσσσ
=
+−+−
∆= −
−++−
∑ (7-43)
and the LP (7-14) is modified as follows
1
[min] ( ) ( )
subject to
( , ) if
( , ) 0 if
1
0, ( ) 0, ( ) 0 ,
R
aA
n
i
i
iR
Faa
ab a b
ab a b
p
p
aaaAi
σσ
δ
σσ
+−
∈
=
+−

=+




∆≥

∆=

=


≥≥≥∀∈∀

∑
∑
f
 (7-44)
The main principles of the UTA methods are also applicable in the
specific field of multiobjective optimization, mainly in the field of linear
programming with multiple objective functions. For instance, in the classical
methods of Geoffrion et al. (1972) and Zionts and Wallenius (1976), the
weights of the linear combinations of the objectives are inferred locally from
trade–offs or pairwise judgments given by the DM at each iteration of the
methods. Thus, these methods exploit in a direct way the DM’s value
functions and seek the best compromise solution through successive
maximization of these assessed value functions.

28 Chapte
r
Stewart (1987) proposed a procedure of pruning the decision alternatives
using the UTA method. In this approach a sequence of alternatives is
presented to the DM, who places each new presented alternative in rank
order relative to the earlier alternatives evaluated. This ranking of elements
in a subset of the decision space is used to eliminate other alternatives from
further consideration. In the same context, Jacquet-Lagrèze et al. (1987)
developed a disaggregation method, similar to UTA, to assess a whole value
function of multiple objectives for linear programming systems. This
methodology enables to find compromise solutions and is mainly based on
the following steps:
1. Generation of a limited subset of feasible efficient solutions as
representative as possible of the efficient set.
2. Assessment of an additive value function using PREFCALC system (see
section 4).
3. Optimization of the additive value function on the original set of feasible
alternatives.
Also, Siskos and Despotis (1989), in the context of UTA-based
approaches in multiobjective optimization problems, proposed the
ADELAIS method. This approach refers to an interactive method that uses
UTA iteratively, in order to optimize an additive value function within the
feasible region defined on the basis of the satisfaction levels and determined
in each iteration.
3.6 Other disaggregation methods
The main principles of the aggregation-disaggregation approach may be
combined with outranking relation methods. The most important efforts
concern the problem of determining the values of several parameters when
using these methods. The set of these parameters is used to construct a
preference model with which the DM accepts as a working hypothesis in the
decision aid study. Assuming that the DM is able to give explicitly the
values of each parameter is not realistic in several real-world applications.
In this framework, the ELECCALC system has been developed (Kiss et
al., 1994), which estimates indirectly the parameters of the ELECTRE II
method. The process is based on the DM’s responses to questions of the
system regarding his/her global preferences.
Furthermore, concerning problematic β, several approaches consist in
inferring the parameters of ELECTRE TRI through holistic information on
DM’s judgments. These approaches aim at substituting assignment examples
for direct elicitation of the model parameters. Usually, the values of these
parameters are inferred through a regression-type analysis on assignment
examples.

. UTA Methods 29
Mousseau and Slowinski (1998) propose an interactive aggregation-
disaggregation approach that infers ELECTRE TRI parameters
simultaneously starting from assignment examples. In this approach, the
determination of the parameters’ values (except the veto thresholds) that best
restore the assignment examples is formulated through a nonlinear
optimization program.
Several efforts have tried to overcome the limitations of the
aforementioned approach (computational difficulty, estimation of the veto
threshold):
a) Mousseau et al. (2000a; 2000b) consider the subproblem of the
determination of the weights only, assuming that the thresholds and
category limits have been fixed. This leads to formulate a linear program
(rather than nonlinear in the global inference model). Through
experimental analysis, they show that this approach is able to infer
weights that restore in a stable way the assignment examples and it is also
able to identify possible inconsistencies in these assignment examples.
b) Doumpos and Zopounidis (2002) use linear programming formulations in
order to estimate all the parameters of the outranking relation
classification model. However, in this approach, the parameters are
estimated sequentially rather than through a global inference process.
Thus, the proposed methodology does not specify the optimal parameters
of the outranking relation (i.e. the ones that lead to a global minimum of
the classification error). Therefore, the results of this approach
(“reasonable” specification of the parameters) serve rather as a basis for a
thorough decision-aid process.
The problem of robustness and sensitivity analysis, through the extension
of the previous research efforts is discussed by Dias et al. (2002). They
consider the case where the DM can not provide exact values for the
parameters of the ELECTRE TRI method, due to uncertain, imprecise or
inaccurately determined information, as well as from lack of consensus
among them. The proposed methodology combines the following
approaches:
1. The first approach infers the value of parameters from assignment
examples provided by the DM, as an elicitation aid.
2. The second approach considers a set of constraints on the parameter
values reflecting the imprecise information that the DM is able to
provide.
In the context of ordinal regression analysis, the MUSA method has been
developed in order to measure and analyze customer satisfaction (Siskos et
al., 1998; Grigoroudis and Siskos, 2002). The method is used for the
assessment of a set of marginal satisfaction functions in such a way that the
global satisfaction criterion becomes as consistent as possible with

30 Chapte
r
customer’s judgments. Thus, the main objective of the method is the
aggregation of individual judgments into a collective value function.
The MUSA method assesses global and partial satisfaction functions *
Y
and *
i
X
respectively, given customers’ ordinal judgments Y and i
X
(for the
i-th criterion). The ordinal regression analysis equation has the following
form:
**
1
n
ii
i
YbX
σ
σ
+−
=
=−+
∑
% (7-45)
where *
Y
% is the estimation of the global value function *
Y, n is the number
of criteria, i
b is a positive weight of the i-th criterion,
σ
+
and
σ
−
are the
overestimation and the underestimation errors, respectively, and the value
functions *
Y and *
i
X
are normalized in the interval [0,100].
Similarly to the UTASTAR algorithm, the following transformation
equations are used:
*1 *
*1 *
for 1 2 1
for 1 2 1 and 1 2
mm
m
kk
ik i i i i i
z y y m = , , ...,α
w b x b x k = , ,...,αi = , ,...,n
+
+
=− −

=− −


(7-46)
where *m
y is the value of the m
ysatisfaction level, *k
i
x
is the value of the
k
i
x
satisfaction level, and
α
and i
α
are the number global and partial
satisfaction levels.
According to the previous definitions and assumptions, the MUSA
estimation model can be written in a LP formulation, as follows:
[]
-
1
11
11 1
1
1
1
11
min
subject to
0 for 1 2
100
100
, , ,
jj
i
i
M
jj
j
xy
n
ik m j j
ik m
α
m
m
α
n
ik
ik
mik j j
F
wz j=,,,M
z
w
zw , m,i,jk
σσ
σσ
σσ
+
=
−−+−
== =
−
=
−
==
+−
=+




−−+=



=


=

∀


∑
∑∑ ∑
∑
∑∑
K
(7-47)
where
M
is the size of the customer sample, and
j
i
x
and
j
y are the j-th
level on which variables i
X
and Y are estimated (i.e. global and partial
satisfaction judgments of the j-th customer). The MUSA method includes
also a post-optimality analysis stage, similarly to step 4 of the UTASTAR
algorithm.
An analytical development of the method and the provided results is
given by Grigoroudis and Siskos (2002), while the presentation of the

. UTA Methods 31
MUSA DSS can be found in Grigoroudis et al. (2000) and Grigoroudis and
Siskos (2003).
The problem of building non-additive utility functions may also be
considered in the context of aggregation-disaggregation approach. A
characteristic case refers to positive interaction (synergy) or negative
interaction among criteria (redundancy). Two or more criteria are synergic
(redundant) when their joint weight is more (less) than the sum of the
weights given to the criteria considered singularly.
In order to represent interaction among criteria, some specific
formulations of the utility functions expressed in terms of fuzzy integrals
have been proposed (Murofushi and Sugeno, 1989; Grabisch, 1996;
Marichal and Roubens, 2000). In this context, Angilella et al. (2003) propose
a methodology that allows including additional information such as an
interaction among criteria. The method aims at searching a utility function
representing the DM’s preferences, while the resulting functional form is a
specific fuzzy integral (Choquet integral). As a result, the obtained weights
may be interpreted as the “importance” of coalitions of criteria, exploiting
the potential interaction between criteria. The method can also provide the
marginal utility functions relative to each one of the considered criteria,
evaluated on a common scale, as a consequence of the implemented
methodology.
The general scheme of the disaggregation philosophy is also employed in
other approaches, including rough sets (Pawlak, 1982; Slowinski, 1995;
Dimitras, et al., 1999; Zaras, 2000), machine learning (Quinlan, 1986) and
neural networks (Malakooti and Zhou, 1994; Stam et al., 1996). All these
approaches are used to infer some form of decision model (a set of decision
rules or a network) from given decision results involving assignment
examples, ordinal or measurable judgments.
4. APPLICATIONS AND UTA-BASED DSS
The methods presented in the previous sections adopt the aggregation-
disaggregation approach. This approach constitutes a basis for the interaction
between the analyst and the DM, which includes:
– the consistency between the assessed preference model and the a priori
preferences of the DM,
– the assessed values (values, weights, utilities, …), and
– the overall evaluation of potential actions (extrapolation output).
A general interaction scheme for this decision support process is given in
Figure 7-10.

32 Chapte
r
Start
Problem formulation
Criteria modelling
Expression of DM's global
preferences on a reference set
Assess the DM's preference model
Is there a satisfactory consistency
between the DM's preferences and
the ones resulted by the model?
Is there any intention of modifying
the preference model?
Is there any intention of modifying
the DM's preferences?
Is there any intention of modifying
the criteria modelling or the
problem formulation?
The preference model is judged
unacceptable
Extrapolation on the whole
set of alternatives
Is the preference model
accepted?
End
NO
YES
NO
NO
NO
End
YES
YES
YES
YES
NO
Figure -10. Simplified decision support process based on disaggregation approach (Jacquet-
Lagrèze and Siskos, 2001)
Several decision support systems (DSSs), based on the UTA model and
its variants, have been developed on the basis of disaggregatio