Вы находитесь на странице: 1из 21

Deriving

Parameters and
Preferential Model for a
Total Order in ELECTRE III
PA V E L A N S E L M O Á LVA R E Z C A R R I L L O , p a v e l . a l v a r e z @ u d o . m x
J U A N C A R L O S L E Y VA L Ó P E Z , j u a n . l e y v a @ u d o . m x
D I E G O A L O N S O G A S T E L U M C H A V I R A , d a g a s t e l u m c h @ C O N A C Y T. M X
U N I V E R S I T Y O F O C C I D E N T, M E X I C O
Content
Introduction
Multicriteria Decision Aids
Genetic Algorithm
Aggregation/disaggregation preference
Results
Conclusions
Limitations
Introduction
The multicriteria aids methods support Decision Maker (DM) to solve divers and real complex
problems.
It is very common that these methods require parameter definition to generate solutions for
decision-making.
It is considered preferential information, which is used to determine how the alternatives will be
evaluated.
Eliciting preferential information from the DMs and formalizing it through preferential
parameters is hence a crucial phase in a multiple criteria decision aid model (Dias et al., 2002)
Introduction
Parameters definition is considered a complex task and requires great effort from DM.
(Dias et al., 2002) identify some reasons problems concerning preference elicitation:
◦ Data regards decision problem can be imprecise or uncertain
◦ vague understanding what parameters represent their point of view
◦ the evolution during the elicitation process.

Such situations make DMs face difficulties at definition of precise values for preferential
parameters.
More preferred alternatives
1.X 2
2.X 1
Ranking 3.X 6
4.X 5
5.X 4
6.X 7
7.X 3 Less preferred alternatives

Multicriteria Alternatives Choice


X2

X1 X2 X3

Decision
X4 X5

X6 Description Characteristics of
X7 alternatives

Analysis
Group 1
X 1, X 2, X 6
Sorting

Group 2
X 3, X 4, X 5, X 7
Multicriteria Decision Analysis a) Problemas discretos b) P

Consumer’s preference m Alternatives


n Criteria g1

Product g1 g2 g3 g4 Criteria
g1 g2 ... gn
x1 g11 g12 ... g1n
A1 TS PS NS NS x2 g21 g22 ... g2n

Alternatives
x3 g31 g32 ... g3n
. . . ... .
A2 PS MS TS PS
. . . ... .
. . . ... .
A3 S PS S MS xm g11 g22 ... g1n

A4 S S NS MS
Multicriteria Decision Analysis
1.1 The E LECTRE III method 3

ELECTRE III A finite set Coherent family


of actions, A of pseudo-criteria, F

Thresholds Performance of the actions g j (a) Weights


v j (g j (a)) Thresholds q j (g j (a)) and p j (g j (a)) wj

Discordance indices, D j (a, b) Concordance indices, C j (a, b)

Concordance relation, C(a, b)

Fuzzy outranking relation, σ (a, b)

Distillation threshold Ranking algorithm


s(λ ) Two complete pre-orders

Final partial pre-order

Figure 1.1: General structure of E LECTRE III


Aggregation/disaggregation preference
The approach requires less effort from DM to express their preference
It requires holistic preferences as pair comparison of as a reference set (A*)

The aim of this research is to infer the intercriteria parameters (w, q, p, v) of ELECTRE III through
examples given by the decision maker (A*)
◦ Holistic judges
◦ Pairs comparison
Aggregation/disaggregation preference
For this problem the holistic judges will be disaggregated by examples given by DM
◦ The aim is to find the ELECTRE III model as compatible as possible with the example assignation given by
DM
opez, E. Fern!
J.C. Leyva-L! andez-Gonz!
alez / European Journal of Operational Research 148 (2003) 14–27 23

≻ Table 5 Table 8
ELECTRE IIICredibility matrix and final ranking associated to DM1: energy Net flow
Credibility matrix and final ranking associated to DM4: trade
production unions
w1, w2, …, wn A1 A2 A3 A4 A5 A6 A1 A2 A3 A4 A5 A6
A1 1 0.45 0.6 0.69 0.37 0.55 A1
q1, q2, ..., qn A2
1 0.5 0.43 0.7 0.5 0.62
0.86 1 0.87 0.71 0.71 0.69 A2 0.5 ≻
1 0.43 ≻ 0.75 0.94 0.7
≻ p1, p2, ..., pn A3
A4
0.94
0.91
0.67
0.47
1
0.47
0.77
1
0.45
0.45
0.71
0.47
A3 0.9 1 1 0.6 1 0.75
A4 0.5 0.5 0.5 1 0.5 0.5
v1, v2, ..., vn A5 0.86 0.69 0.73 0.71 1 0.87 A5 0.5 0.62 0.44 0.5 1 0.65
A6 0.71 0.57 0.49 0.71 0.71 1 A6 0.75 0.56 0.35 0.5 0.75 1
Credibility matrix: ðA2 " A5Þ > A3 > A6 > A4 > A1; final Credibility matrix: A3 > A2 > A6 > A1 > A4 > A5; final
ranking, k0 ¼ 0:689. ranking, k0 ¼ 0:699.

≻ Table 6
Credibility matrix and final ranking associated to DM2: envi- Table 9
Holistic preference ronment Fuzzy outranking relation Ranking of alternatives
Group performances. Classification of sites according to each
Intercriteria parameters
A1 A2(preferential model)
A3 A4 A5 A6 new criterion

A1 1 0.75 0.75 0.68 0.75 0.75 DM1:EP DM2:ENV DM3:FIN DM4:TU


A2 0.32 1 0.77 0.3 1 0.93 Italy 6 2 1 4
A3 0.50 0.50 1 0.43 0.50 0.75 Belgium 1 3 3 2
A4 1 0.75 0.75 1 0.75 0.75 Germany 3 5 5 1
A5 0.3 0.43 0.7 0.28 1 0.85 Sweden 5 1 4 5
A6 0.35 0.5 0.6 0.32 0.75 1 Austria 2 4 6 6
Credibility matrix: A4 > A1 > A2 > A5 > A3 > A6; final France 4 6 2 3
ranking, k0 ¼ 0:699.
working with the preference matrix of Section 4.1.
Table 7 Usually, the values of parameters k and b needed
Credibility matrix and final ranking associated to DM3: Fi-
for defining the nine zones of r should be deter-
disagreements between the d-th
rithm borrows fundamental elem
in further detail the fundamental
A solution is encoded as a set
rion defined by the DM correspo
disagreements between the d-th order and the collective temporary order. The algo- tion of each criterion in the m-ary
rithm borrows fundamental elements from NSGA II [20]. In the following we present algorithm borrows fundamental e
Genetic algorithm
in further detail the fundamental aspects of the algorithms.
A solution is encoded as a set of pseudo-criteria in an m-ary string. For each crite-
eters used for the model applicati
tions, 0.90 crossover index and 0
rion defined by the DM corresponds the w, q, p and v pseudo-criteria. The representa-
tion of each criterion in the m-ary string corresponds to 4 real values (see Fig 2.). The
borrows fundamental elements from NSGA II [20].  The  GA’s  value  param- x j , y j , z j
The parameters of ELECTRE III can be expressed in terms of decision variables
algorithm
eters used forzthe z j + y;j
model application:
j - y ;j
40 individuals for population size, 10000 genera-
q = p = v j = x j + 12 y j + 12 z j
tions, 0.90j crossover
2 index and
j
2 0.4 mutation index. Fig. 2. Individual representat

Objective functions.
vj = x j + 12 y j + 12 z j (6)
Fig. 2. Individual representation for the GA of inferring inter-criteria parameters
Basic definitions.
From this equation, together with (5), it can be easily seen that v j ≥ p j . Definition 1. Let A { a
G { g 1 , g 2 ,..., g n } be a set of n d
Objective functions.
Thus, given that x j , y j , z j ≥ 0 and z j > y j , the selected representation ensures that
ality, we can consider the first t
v j ≥ definitions.
Basic p j ≥ q j ≥ 0 ∀j = 1,2,...,n . the rest n-t criteria G s { g t 1 , g t
Definition 1. Let A { a 1 , a 2 ,..., a m } be a set of m alternatives and Definition 2. Let K d be the s
In addition, the above representation also defines a ….
G { g 1 , g 2 ,..., g n } be a set of n decision criteria defined on A. Without loss of gener- O and the collective temporary
d

ality, we canthe
To reduce consider
search the firstwe
space, t criteria
also impose
Go { g 1 , g 2 ,..., g t } as objective
the condition p j ≤ x*j − x j* ,criteria
where and
x*j = max{ xij } (ai , a j ) A
i∈A* d d G
the rest n-t criteria G s { g t 1 , g t 2 ,..., g n } as subjective criteria. K (O ,O )
and x j* = min{ *
xij }. This condition is ensured with the constraint: (O
d
(ai ) O
d

Definitioni∈A2. Let K d be the set of pairwise disagreements between the d-th order
O and the collective
The first complement of K
d
,
y j + z j ≤ 2( x*j − x j* ) temporary order O G defined as:
d
(7)
(ai , a j ) A A |i j , (O
d
(ai ) O
d
(a j ) O
G
(ai ) O
G
( a j )) first d-th order O
d
and the collec
d d G
K (O ,O )
The solution vector has
d n x, y,d z variables,
G n variables
G corresponding to the criteria weights, (ai , a j ) A
(O ( a i ) O (a j ) O (ai ) O ( a j )) d ,C d G
one variable corresponding to the cutting point λ and n binary variables t j ∈{0,1}introduced K first
(O ,O )
d
y j + z j ≤ 2( x*j − x j* ) (7)

The solution vector has n x, y, z variables, n variables corresponding to the criteria weights,
one variable corresponding to the cutting point λ and n binary variables t j ∈{0,1}introduced
to indicate whether a criterion j can pose veto ( t j = 1 ) or not ( t j = 0 ). Thus, overall the
solution vector has 5n+1decision variables.
We use a value-encoding scheme to represent a solution vector (a potential solution) p! . Let
p! = p1 p2 ... p5n+1 be the schematic representation of an individual’s chromosome. p! ∈ ∏5n+1
i=1 Ci

Genetic algorithm
, where Ci is the set of values that pi can takes. This set of values is dependent of the
problem.
! ! ! !!
The general Min
form ofK ( w,
the x, y, z, t , λproblem
optimization ) can be expressed as follows:
s.t.
!
*
y j + z j ≤ 2(xMin
j
− xK()! ! j! =! 1,2,...,n
j* w, x, y, z, t , λ )

w1 + w2 + ...+s.t.wn = 1
! ! y j + z j ≤ 2(x*j − x j* ) j = 1,2,...,n
z≥y
! ! ! ! ! w!1 + w2 +...+ wn = 1 (8)
w, x, y, z, t ≥!0 !
z≥y
t ∈ {0,1}, λ ∈ ! ! ! λ !> 0.5
! !(0,1),
w, x, y, z, t ≥ 0
t ∈ {0,1}, λ ∈ (0,1), λ > 0.5
! ! ! !!
Where w, x, y, z, t are the vectors of the decision variables w j , x j , y j , z j ,t j
j = 1,2,...,n , and
! ! ! ! !
K ( w, x, y, z, t , λ ) denotes the ranking accuracy of the model corresponding to a specific
! ! ! !!
solution ( w, x, y, z, t , λ ) .
Genetic algorithm
Genetic algorithm Product
Consumer’s preference

g1 g2 g3 g4

Input
A1 TS PS NS NS

A2 PS MS TS PS

◦ Performance matrix (Values of alternatives for each criterion)


A3 S PS S MS

A4 S S NS MS

◦ Holistic judgments of alternatives (pairs comparison of alternatives) ≻ ≻ ≻


, ,

◦ Holistic judgments of criteria (pairs comparison of criteria) C4 ≻ C2 ≻ C3 ≻ C1

Output w1, w2, …, wn


q1, q2, ..., qn
◦ Divers set of intercriteria parameters p1, p2, ..., pn
opez, E. Fern!
J.C. Leyva-L!
v1, v2, ..., vn
andez-Gonz!
alez / European Journal of Operational Research 148 (2003) 14–27 23

Table 5 Table 8
Credibility matrix and final ranking associated to DM1: energy Credibility matrix and final ranking associated to DM4: trade

◦ Fuzzy outranking relation


production unions
A1 A2 A3 A4 A5 A6 A1 A2 A3 A4 A5 A6
A1 1 0.45 0.6 0.69 0.37 0.55 A1 1 0.5 0.43 0.7 0.5 0.62
A2 0.86 1 0.87 0.71 0.71 0.69 A2 0.5 1 0.43 0.75 0.94 0.7
A3 0.94 0.67 1 0.77 0.45 0.71 A3 0.9 1 1 0.6 1 0.75
A4 0.91 0.47 0.47 1 0.45 0.47 A4 0.5 0.5 0.5 1 0.5 0.5
A5 0.86 0.69 0.73 0.71 1 0.87 A5 0.5 0.62 0.44 0.5 1 0.65
A6 0.71 0.57 0.49 0.71 0.71 1 A6 0.75 0.56 0.35 0.5 0.75 1
Credibility matrix: ðA2 " A5Þ > A3 > A6 > A4 > A1; final Credibility matrix: A3 > A2 > A6 > A1 > A4 > A5; final
ranking, k0 ¼ 0:689. ranking, k0 ¼ 0:699.

Table 6
Credibility matrix and final ranking associated to DM2: envi- Table 9
ronment Group performances. Classification of sites according to each
A1 A2 A3 A4 A5 A6 new criterion

A1 1 0.75 0.75 0.68 0.75 0.75 DM1:EP DM2:ENV DM3:FIN DM4:TU


A2 0.32 1 0.77 0.3 1 0.93 Italy 6 2 1 4
A3 0.50 0.50 1 0.43 0.50 0.75 Belgium 1 3 3 2
A4 1 0.75 0.75 1 0.75 0.75 Germany 3 5 5 1
A5 0.3 0.43 0.7 0.28 1 0.85 Sweden 5 1 4 5
A6 0.35 0.5 0.6 0.32 0.75 1 Austria 2 4 6 6
como una jerarquía de un conjunto de alte
intuición del DM es difícil de capturar por m
1.1, ilustra lo anterior en las diferentes estruc
completo ó un preorden completo o un preor
pueden plantear problemas para criterios de o
principal puede ser percibida fácilmente por lo
Genetic algorithm
Individual i in population Total order

w, q, p, v, 𝛌
Disaggregation procedure for
inferring intercriteria parameters
in a total order problem Contruct 𝛔 with ELCTRE
III

Exploit 𝛔 with Net Flow

Measure distance with


Kendall_distance (Ri, RDM)

Figura 1.1. Diferentes


Notas: C j representa una clase que agrupa altern
Fuente: Le
Results
Problem 1 – Energy production DM
DM1: holistic preferences
Order of criteria
C2 ≻ C5 ≻ C3 ≻ C7 ≻ C4 ≻ C1

Disaggregation procedure
Order of locations
Austria ≻ Be lgium ≻ Germany ≻ France ≻ Sweden ≻ Italy

6
2
3
5
1
4
Results
Problem 2 – Environment DM
DM2: holistic preferences
Order of criteria
C5 ≻ C6 ≻ C7 ≻ C8

Disaggregation procedure
Order of locations
Sweden ≻ Italy ≻ Be lgium ≻ Austria ≻ Germany ≻ France

2
3
5
1
4
6
Conclusions
Derives intercriteria parameters for the ELECTRE III method from holistic judgments
Outcome solutions (set of intercriteria parameters) for the ELECTRE III method generate the
same rankings that input order expressed by DM of energy production and DM of environment.
◦ It means that valued outranking relations constructed by the set of intercriteria parameters contain
DM’s preference.
opez, E. Fern!
J.C. Leyva-L! andez-Gonz!
alez / European Journal of Operational Research 148 (2003) 14–27 23

≻ Table 5 Table 8
ELECTRE IIICredibility matrix and final ranking associated to DM1: energy Net flow
Credibility matrix and final ranking associated to DM4: trade
production unions
w1, w2, …, wn A1 A2 A3 A4 A5 A6 A1 A2 A3 A4 A5 A6
A1 1 0.45 0.6 0.69 0.37 0.55 A1
q1, q2, ..., qn A2
1 0.5 0.43 0.7 0.5 0.62
0.86 1 0.87 0.71 0.71 0.69 A2 0.5 ≻11 0.43 ≻ 0.75 0.94 0.7
≻ p1, p2, ..., pn A3
A4
0.94
0.91
0.67
0.47
1
0.47
0.77
1
0.45
0.45
0.71
0.47
A3 0.9 1 0.6 1 0.75
A4 0.5 0.5 0.5 1 0.5 0.5
v1, v2, ..., vn A5 0.86 0.69 0.73 0.71 1 0.87 A5 0.5 0.62 0.44 0.5 1 0.65
A6 0.71 0.57 0.49 0.71 0.71 1 A6 0.75 0.56 0.35 0.5 0.75 1
Credibility matrix: ðA2 " A5Þ > A3 > A6 > A4 > A1; final Credibility matrix: A3 > A2 > A6 > A1 > A4 > A5; final
ranking, k0 ¼ 0:689. ranking, k0 ¼ 0:699.

≻ Table 6
Credibility matrix and final ranking associated to DM2: envi- Table 9
Holistic preference ronment Fuzzy outranking relation Ranking of alternatives
Group performances. Classification of sites according to each
Intercriteria parameters
A1 A2(preferential model)
A3 A4 A5 A6 new criterion

A1 1 0.75 0.75 0.68 0.75 0.75 DM1:EP DM2:ENV DM3:FIN DM4:TU


A2 0.32 1 0.77 0.3 1 0.93 Italy 6 2 1 4
A3 0.50 0.50 1 0.43 0.50 0.75 Belgium 1 3 3 2
A4 1 0.75 0.75 1 0.75 0.75 Germany 3 5 5 1
A5 0.3 0.43 0.7 0.28 1 0.85 Sweden 5 1 4 5
A6 0.35 0.5 0.6 0.32 0.75 1 Austria 2 4 6 6
Credibility matrix: A4 > A1 > A2 > A5 > A3 > A6; final France 4 6 2 3
ranking, k0 ¼ 0:699.
un orden definido sobre un conjunto de acciones
subconjunto conjunto
de dichooconjunto de de
alternativas, esacciones
decir, deocon
alternativas, alternativas,
dirigir es decir,
capaz dehacia
la investigación
el fin de ser dirigir la investigacióndehacia la determinación de
la determinación
de determinar las acciones un orden
que definido
podrían un subconjunto
sobre un orden
ser consideradas definido
como de sobre
dicho un subconjunto
conjunto
"suficientemente de dicho conjunto
de alternativas,
satisfactorias" con el finde
dealternativas,
ser capaz con el fin de ser capaz
de determinar
basado en un modelo de preferencias, mientras de que
las acciones
se tiene determinar
podrían
en cuenta las
que acciones
serel que
de podrían
consideradas
conjunto ser consideradas satisfactorias"
como "suficientemente
alternativas como "suficientemente satisfactorias"
basado
podría evolucionar. El ranking en un modelo
proporciona en síde basado enrecomendación
un mientras
preferencias,
mismo una modelo de
se preferencias,
tiene mientras
entomador
para el cuenta de el se
que tiene endecuenta
conjunto que el conjunto de alternativas
alternativas
podría
decisiones en forma de un evolucionar.
orden podría proporciona
parcial oElcompleto
ranking evolucionar.
formado por El sí
en ranking
mismo
clases proporciona
queunacontienenen sí mismo
recomendación para una recomendación
el tomador de para el tomador de
decisiones entre
alternativas consideradas equivalentes en forma dedecisiones
sí (Roy, un ordenenparcial
1996). formao decompleto
un orden parcial por
formado o completo
clases queformado por clases que contienen
contienen
alternativas consideradas alternativas
equivalentesconsideradas equivalentes
entre sí (Roy, 1996). entre sí (Roy, 1996).
El concepto de ordenamiento, es una generalización que los tomadores de decisiones perciben

Limitations El concepto
como una jerarquía de un conjunto
intuición del DM es difícil como una jerarquía
de capturar
El concepto
de ordenamiento,
de alternativas en es una
orden
como
de de
por medio
de ordenamiento,
generalización
decreciente
una objetivos
uncriterios
conjunto jerarquía
queeslos
unatomadores
de preferencias,generalización
de un conjunto
de alternativas
la
en orden
individuales.
que los tomadores
de decisiones
dedecreciente
perciben de decisiones perciben
alternativasdeenpreferencias,
La Figura orden decreciente
la de preferencias, la
Theilustra
1.1, procedure
lo anteriorisendesigned
intuición
las for
del
diferentes DM
estructuras However,
a es difícil
intuición
de in
dellas
DM
de capturar
rankings, some
poresmedio
cuales cases
difícil
pueden the
dedecriterios
capturar
ser: DM
por medio
un objetivos
orden Other
de case
criterios
individuales. La occurs
objetivos
Figura when the
individuales. DM
La Figura
complete
completo ó unorder,
preordenincompleto
this case,
1.1, ilustra we
o unlopreorden
could
anterior en1.1,
las ilustra identify
lo
diferentes
parcial. Mientras anterior en de
estructuras
cada una
that
las
de diferentes
rankings,
esas
two
estructuras
las cannot
de ser: define
rankings,
cuales pueden
estructuras unlas strict
cuales
orden preference
pueden ser: un orden
expect the DM isparawilling alternatives are indifferent
pueden plantear problemas completo ó unto
criterios preorden
de between
ordenamientos them
completoespecíficos,
óo un sharing
preorden
preorden the
completo
parcial.
la estructura o same
un preorden
Mientras
de
orde esas
parcial.
cada una
ranking
indifference
Mientras between
cada una de esas
estructuras estructuras
compare the total pairs position in thepara ranking alternatives, because of
pueden
principal puede ser percibida plantear
fácilmente pueden
por problemas paradeplantear
los tomadores criterios problemas et. al,criterios
de ordenamientos
decisiones (Leyva de ordenamientos
específicos,
2013a). la estructuraespecíficos,
de ranking la estructura de ranking
between alternatives and define (containing ties). inconsistent information.
principal puede
a strict preference for each pair. ser principal
percibida puede
fácilmente ser
por percibida
los fácilmente
tomadores de por los
decisiones tomadores
(Leyva de 2013a).
et. al, decisiones (Leyva et. al, 2013a).
◦ In this case, we are dealing with ◦ In this case, we are dealing with
Total order a total preorder problem. a partial preorder problem.
Total preorder Partial preorder

Figura 1.1. Diferentes estructuras de ordenamiento.


Notas: C j representa una clase que agrupa alternativas a km indiferentes;
Figura ak A
1.1. Diferentes {akFigura
.., akm
estructuras
1 de};ordenamiento.
j Diferentes
1.1. m estructuras de ordenamiento.
Fuente:
Notas: Leyva et.una
C j representa al (2013a).
Notas:
clase C j representa
que agrupa una aclase
alternativas ak A {aakm
que agrupa alternativas ..,indiferentes; ak A {ak1.., akm}; j m
km indiferentes; k1
akm}; j m
Fuente: Leyva et. al (2013a). Fuente: Leyva et. al (2013a).
3
3 3
Deriving Parameters and
Preferential Model for a
Total Order in ELECTRE III
PA V E L A N S E L M O Á LVA R E Z C A R R I L L O , p a v e l . a l v a r e z @ u d o . m x
J U A N C A R L O S L E Y VA L Ó P E Z , j u a n . l e y v a @ u d o . m x
D I E G O A L O N S O G A S T E L U M C H A V I R A , d a g a s t e l u m c h @ C O N A C Y T. M X
U N I V E R S I T Y O F O C C I D E N T, M E X I C O

Вам также может понравиться