Methodology to Select Solutions from the Pareto-Optimal

Set: A Comparative Study

J. C. Ferreira C. M. Fonseca A. Gaspar-Cunha
IPC- Institute for Polymer and Centre for Intelligent Systems IPC- Institute for Polymer and
Composites University of Algarve Composites
University of Minho Faro, Portugal University of Minho
Guimarães, Portugal +351289800045 Guimarães, Portugal
+351253510320 +351253510328
[email protected]
[email protected] [email protected]

ABSTRACT such problems typically admit multiple optimal solutions where any
The resolution of a Multi-Objective Optimization Problem (MOOP) improvement in one objective can only be obtained at the expense
does not end when the Pareto-optimal set is found. In real problems, of degradation in other objectives [1]. Thus, the final solution of a
a single solution must be selected. Ideally, this solution must belong multi-objective optimization problem results, not only from the
to the non-dominated solutions set and must take into account the optimization process (or search process), but also from the decision
preferences of a Decision Maker (DM). Therefore, the searching for process. The integration of the DM preferences on the optimization
a single solution (or solutions) in MOOP is done in two steps. First, process must allow the selection, from the set of optimal solutions,
a Pareto optimal set is found. Multi-Objective Evolutionary the single solution (or a region) that best satisfies these preferences.
Algorithms (MOEA), based on the principle of Pareto optimality, Three different classes of methods can be defined, taking into
are designed to produce the complete set of non-dominated account the way how the search and the decision processes are
solutions. Second, a methodology able to select a single solution integrated [2, 3]: i) A priori methods: The decision maker must
from the set of non-dominated solutions (or a region of the Pareto specify her or his preferences, expectations and options before the
frontier), and taking into account the preferences of a Decision optimization process takes place. These preferences are defined in
Maker (DM), can be applied. In this work, a method, based on a terms of an aggregating function which combines individual
weighted stress function, is proposed. It is able to integrate the objective values into a single utility value. ii) A posteriori methods:
user’s preferences in order to find the best region of the Pareto The Pareto optimal set after been generated, without taking into
frontier accordingly with these preferences. This method was tested consideration any DM preferences, is supplied to the DM, which
on some benchmark test problems, with two and three criteria. This selects the most preferred among the alternatives. iii) Interactive
methodology is able to select efficiently the best Pareto-frontier methods: Decision making and optimization occur simultaneously.
region for the specified relative importance of the criteria. At each step, partial preference information is supplied by the DM.
The optimization algorithm uses this information to generate better
alternatives accordingly with the information received from the DM.
Categories and Subject Descriptors
I.2.8 [Problem Solving, Control Methods, and Search]: Heuristic The difficulties arising from the process of decision making in a
methods. multi-objective environment makes the need, to define a
methodology able to combine the DM preferences with the search
procedure, necessary. A new methodology, based on a weight Stress
General Terms Function Approach, to select the best solution (or set of solutions)
Algorithms. from a Pareto frontier taking into account the preferences of the
DM, is proposed. This methodology is able to define the best region
Keywords of the Pareto frontier from a set of weights given by a weight vector.
Decision making, Optimization, Multi-objective. The attribution of a relative importance to each criterion helps the
optimization algorithm in converging to relevant regions of the
Pareto frontier. Thus, in this case the search will be made on a small
1. INTRODUCTION portion of the Pareto frontier, which facilitates the algorithm
Real-world optimization problems are characterized by the convergence for a set of optimal solutions that have particular
existence of multiple, often conflicting, objectives. In the absence of interest to the DM.
information about the relative importance of the various criteria,

Permission to make digital or hard copies of all or part of this work for 2. METHODOLOGY FOR DECISION
personal or classroom use is granted without fee provided that copies are MAKING
not made or distributed for profit or commercial advantage and that copies In this section three different methods for selecting solutions taking
bear this notice and the full citation on the first page. To copy otherwise, or
into account the preferences of a DM will be described, the aim
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee. being to compare the performance of these techniques using some
GECCO’07, July 7–11, 2007, London, England, United Kingdom. benchmark problems from the literature.
Copyright 2007 ACM 978-1-59593-697-4/07/0007...$5.00.

789
2.1 Reference Point based EMO Approach where, wi is the weight vector (taking into account that wi >0 for

∑
N
(RP-EMO) i=1,…, N and
i =1
wi = 1 ) and z i* is the ideal criteria vector
Deb and co-authors [4] proposed a reference point approach, were which represents the vector maximizing each one of the objective
the aim was to reach a Pareto frontier region located near a specific functions [2].
pre-defined reference point or points. This is apriori method that
can be applied using or not a weight vector. The main ideas of this The Weighted Tchebycheff metrics is obtained from equation 1
method are the following: making p=∞:
1. Solutions closer to the reference points, in the criteria
space, are to be more highlighted. Minimize
i =1,K, N
[
max wi | f i ( X ) − z i* | ] (3)
2. Solutions within a ε-neighborhood to a near-reference- Subject to x∈S
point solution are to be highlighted, in order to maintain a diverse
set of solutions near each reference point. The aim of the Weighted Metrics is to minimize the distance
between the solutions (on the criteria space) and the ideal criteria
The normalized Euclidian distance (dij) between each solution of the vector. If the ideal criteria vector belongs to the space of admissible
best non-dominance level and the reference point is calculated, solutions, then the best solution will be this ideal vector, since it is
allowing the identification of the solutions located near the reference the solution that minimizes each one of the individual objective
point: functions. However, this ideal criteria vector only can belongs to the
2 admissible region when the criteria are not conflicting.
N
⎛ f i ( x) − z i ⎞ (1)
d ij = ∑ w ⎜⎜ f
i
⎟⎟ The method of weighted Tchebycheff metric is often used to
⎝ i − fi
max min
i =1 ⎠ generate the Pareto optimal solutions. The main property of this
where, fimax and fimin are the maximum and the minimum value of the method is the following: for a known ideal objective vector, for each
optimal solution X, there is a weighted vector wi for which X is the
objective function for criterion i, respectively, and z i and wi are the
solution of the weighted Tchebycheff problem. Thus, all the optimal
i-th component of the reference point and weight vector, solutions of a MOOP can be determined through the resolution of
respectively. This equation takes into account simultaneously the the weighted Tchebycheff problem. However, the application of this
relative importance of the different criteria, quantified through the technique can originate weak Pareto optimal solutions, thus some
weight vector, and the distance between the solution and the auxiliary calculations will be needed to identify these weak ones [2].
reference point. The solutions with lower distance to the reference
point and, simultaneously, with high weight vector, will be selected 2.3 Weighted Stress Function Method
preferentially. (WSFM)
This procedure was been implemented through the modification of The method proposed here (WSFM) integrates the DM preferences
the NSGA-II algorithm [5,6]. The main change was made on the after the search process has been made. Therefore, this is a posterior
crowding operator, since in this case the aim is not to obtain a great method where the search and the decision process are sequential.
diversity of solutions along the Pareto frontier but a sub-set of these The main principle of the WSFM method consists on the idea that
solutions that minimize the distance to the reference point. The the best solution that will satisfy the preferences of the DM must
details of these changes have been described by Deb and co-authors take into account the two following aspects. First, the solution must
[4]. belong to the Pareto Frontier, i.e., to the set of non-dominated
solutions. Second, the selection procedure must take into
There are two difficulties when this methodology is applied to real consideration the ideal objective vector denoted as z*, i.e., the
problems: First, it is necessary to know beforehand the criteria vector maximizing each one of the objective functions. Since the
space, which is not an easy task mainly for problems with many aim is to maximize the criteria, the individual optimization of each
criteria and for problems were the Pareto front is not obtained from criterion corresponds to the maximum value for this criterion.
simple functions. Second, the definition of the reference points However, in multi-objective optimization, the relative importance
makes necessary that the DM needs to have some knowledge about attributed to each criterion will induce a “stress” for searching for
the location of the best solutions or regions of the Pareto front. solutions that maximize each one of the various criteria.
2.2 Weighted Metrics Method (WMM) The method will be explained with Figure 1, where a Pareto frontier
The Weighted Metrics [2] is a preference method where the of an optimization problem with two criteria, f1 and f2, to be
articulation of preference information is used a posteriori. Thus, the maximized is shown. For each solution belonging to the Pareto
solution is chosen from a group of results obtained by the frontier, two stresses, γw1 and γw2, are defined, each one associated to
optimization process. For that purpose the Weighted Metrics (Lp) is the corresponding criteria, f1 and f2 (w1 and w2 are the weight
applied: attributed to each criterion). These stresses are proportional to the
weighted distance between criteria i and the i-th component of ideal
1/ p objective vector.
⎛ N p
⎞
Minimize ⎜ ∑ wi f i ( X ) − z i* ⎟ (2)
⎜ i =1 ⎟
⎝ ⎠
Subject to x ∈ S , 1 ≤ p < ∞

790
 Figure 2 represents the stress function (γwi) and the geometric
interpretation of functions ϕ, ψ and ξ. For a specific location of wi, ξ
(wi) is equal to γwi (fi) for fi equal to wi (since fi and wi are
normalized, ranging in the interval [0,1]). Simultaneously, the
distance between asymptote A1 and wi is ψ(wi)/2 and the distance
between asymptote A2 and wi is ϕ(wi)/2. When wi and fi,
simultaneously, approaches 0 or 1 functions ψ(wi) and ϕ(wi),
f2 respectively, cannot be null in order to avoid that, in these cases
(i.e., for wi equal to 0 or 1), γ wi(fi) assumes an infinite value. To
enable that the weight range in the interval [0,1], a small value (δ1)
must be added to ψ(wi) and ϕ(wi) functions, without changing
considerably the shape of the γ wi (fi) function. Finally, function
ξ(wi) must be constant in all wi domain, except for wi values near the
interval limits (i.e., for values of wi near to 0 and 1). This is
f1 accomplished using the parameter δ2 in equation 7. Various values
of δ1 and δ2 were tested and the best values are δ1=0.002 and
Figure 1- Stresses associated with an optimal solution for δ2=0.008.
an optimization problem with two criteria.
γ w ( fi )
Since both criteria are conflicting the solution of the problem cannot A1 i
A2
be the ideal objective vector. In fact the stresses associated with the
best solution for the problem are not null. The best solution for the
problem, in an ideal situation, is the one that have null stresses, but
this situation is not possible do to the conflicting relation between
criteria. Thus, the best solution will be the one where the differences
between the stresses associated to each criterion have the minimum
possible value. This means that, for the best solution, the stress γw1
does not constitutes a sufficient stress for searching solutions with
ξ ( wi )
best values for f1, since an increase in this criterion implies a
decrease in f2, and consequently an increase on γw2. The increase of
0 wi 1
this stress redirects the search for solutions with best values of the fi
second criterion in detriment of losses in the first one. Taking into ψ ( wi ) ϕ ( wi )
account these considerations, the weighted stress function associated 2 2
to the weight wi is defined as:
Figure 2- Stress function for a specific weight.
⎧ wi ⎛ π ⎞
⎪ tan⎜⎜ − ( x − wi ) ⎟⎟ + ξ ( wi ) , x ≤ wi
(4) Thus, for a set of weights wi, the resolution of MOOP with N criteria
⎪ 2 ⎝ ψ ( w i ) ⎠
⎪⎪ consists in solving the following problem with a single objective:
γ wi ( x ) = ⎨
⎪− ξ ( wi ) ⎛ π ⎞
Minimize T ( X ) = ∑ γ wi ( fi ( X ) ) − γ w j ( f j ( X ) ) (9)
tan⎜ − ( x − wi ) ⎟⎟ + ξ ( wi ) , x > wi
⎪ ⎛ π ⎞ ⎜⎝ ϕ ( wi ) ⎠
⎪ tan⎜⎜ ( wi − 1) ⎟⎟ i< j
⎩⎪ ⎝ ϕ ( wi ) ⎠
Subject to X ∈S
where,
2.4 Methodology for DM
3 (5) In the majority of the optimization problems only small regions of
ϕ ( wi ) = (1 − wi ) 2 + 2(1 − wi ) + δ 1 the Pareto frontier have practical importance [7, 8]. These sub-sets
4
of the Pareto frontier are defined by the integration of the DM
preferences on the problem under study. The definition of an
ψ (wi ) = ϕ (wi ) + 4wi − 2 (6) “ability” function, combining the non-dominance concept and the
relative importance of each criterion, will allow the selection of the
solutions satisfying both features. For that purpose, a new fitness
1 ⎛ π ⎛ 1 ⎞⎞ (7) function (denoted as DF) will be defined, depending of both on the
ξ ( wi ) = − tan⎜ ⎜ wi − ⎟ ⎟⎟ + 1 rank value - Ranks(X) and on the T(X) value defined by equation 9.
⎛ π ⎞ ⎜⎝ 1 + δ 2 ⎝ 2 ⎠⎠
tan⎜⎜ − ⎟
⎝ 2 + 2δ 2 ⎟⎠ DF ( X ) = Rank s ( X ) +
T (X ) (10)
T (X ) +1
This definition considers that the criteria are normalized, i.e., fi∈[0; This function specifies with detail the fitness of each solution taking
1], consequently, the ideal vector has the following coordinates into consideration two objectives: the non-dominance level and the
(1,1). importance of the solution considering the preferences of the DM.

791
These two objectives are improved when DF decreases. This model the Pareto frontier that take into consideration the DM preferences.
for calculating the fitness can be used to define a new decision rank, Thus, in this work the comparison will be made visually.
since the first term of equation 10 corresponds to the standard rank
The Pareto frontiers were obtained with the NSGA-II algorithm
of the solution. Thus, the values of DF contain information
with real codification using the SBX operator with an index of 10
concerning the non-dominance level. The extension of the sub-set of
and polynomial mutation with and index of 20. The population size
the Pareto frontier obtained by the application of this methodology
was fixed at 100 for ZDT1 and ZDT2 problems and at 200 for
is controlled by a dispersion parameter (ε) that ranges in the interval
DTLZ2 problem, the number of search generations (N1) was 300 for
[0,1]. For small values of this parameter the methodology produces
ZDT1 and ZDT2 problems and 500 for DTLZ2 and the number of
a single solution and for high values the entire Pareto frontier will
decision generations (N2) was 30 generations for all problems. Due
be obtained. In order to implement the methodology proposed the
to the stochastic nature of the optimization algorithm, five different
Non-Dominated Sorting Genetic Algorithm (NSGA-II) proposed by
runs were carried out for each case using different seed values.
Deb et al. [5,6] was been modified. During N1 generations (search
generations) the NSGA-II was applied without modifications in
order to find a good approximation to the Pareto frontier. Then, 3.2 Study of the Influence of the Dispersion
during N2 generations (decision generations) the new methodology Parameter (ε)
(equation 10) was applied in order to the populations converges to This parameter controls the dispersion of the solutions along the
the region defined by the weights given by the DM. Pareto frontier when the proposed methodology is applied. The
ideal situation will be the one were only a single solution is found
3. RESULTS AND DISCUSSION and, simultaneously, the DM preferences are considered. However,
3.1 Case Studies a balance between the dispersion and the concentration of the
solutions must be done, mainly when the DM does not have a good
Three different test problems will be used to test the validity of the
knowledge of the process.
algorithm:
Figure 3 shows the influence of the dispersion parameter, for ZDT2
ZDT1 [5] (Minimize f1 and f2, L=30, x ∈ [0,1]):
test problem, on the performance of the decision methodology
f1 ( x) = x1 proposed. As expected, for low values of ε (0.005 and 0.2) the
(11)
[
f 2 ( x) = g ( x) 1 − x1 g ( x) ] solutions converge for small regions of the Pareto frontier. The
parameter allows to control the extension of the Pareto set obtained,
⎛ L ⎞ i.e., to control the balance between the preferences of DM and the
g ( x) = 1 + 9 ⎜ ∑ xi ⎟ ( L − 1)
⎝ i =2 ⎠ non-dominance relation. If the DM has a good knowledge of the
problem it can use low values for this parameter. In this case the
ZDT2 [5] (Minimize f1 and f2, L=30, x ∈ [0,1]): number of alternative solutions is reduced (ε=0.005). If the DM
wants to obtain more alternatives a reasonable option is to use ε
f1 ( x) = x1 equal to 0.2. For a dispersion parameter higher than 0.6 the entire
(12)
[
f 2 ( x) = g ( x) 1 − ( x1 g ( x) )
2
] Pareto frontier is obtained. The same type of conclusion can be
demonstrated from an equal study made for the remaining test
⎛ L ⎞ problems. Figure 4 shows these results for DTLZ2 test problem with
g ( x) = 1 + 9 ⎜ ∑ xi ⎟ ( L − 1)
⎝ i=2 ⎠ three criteria.

DTLZ2 [9] (Minimize f1, f2 and f3, L=12, x ∈ [0,1]): 3.3 Study of the Influence of the Weight Vector
Figure 5 shows the results obtained for ZDT1 test problem using
⎛ π⎞ ⎛ π⎞ different weight vectors and ε equal to 0.1. These results show that
f1 ( x) = (1 + g ( x) ). cos⎜ x1 ⎟. cos⎜ x2 ⎟ an increase of the importance of the first criterion (f1) produces
⎝ 2⎠ ⎝ 2⎠ solutions with lower values for this criterion. When a weight of 80%
⎛ π⎞ ⎛ π⎞ (13) is attributed to f1 (and, as a consequence, 20% to f2) the main aim is
f 2 ( x ) = (1 + g ( x ) ). cos⎜ x1 ⎟. sin ⎜ x2 ⎟
⎝ 2 ⎠ ⎝ 2⎠ to minimize f1 and not to minimize f2. The main task, in this case, is
the searching of solutions with lower value for f1, and the task of
⎛ π⎞
f3 ( x) = (1 + g ( x) ). sin ⎜ x1 ⎟ minimizing f2 is less important. The reverse situation occurs when
⎝ 2⎠ the importance of the second criterion is higher than the importance
L
attributed to the first one. In this case, the minimization of f2 will be
g ( x) = ∑ ( xi − 0.5) 2
i =3 the main task. The same type of behavior was observed for the
remaining test problems.
First, the influence on the algorithm performance, of the dispersion
parameter (ε) and of the weight vector, will be evaluated, and then
the methodology proposed (WSFM) will be compared with the RP-
EMO and WMM approaches.
There are metrics available able to compare entire Pareto frontiers,
but since the aim here is to evaluate the location of small portions of

792
 ε=0.005
ε=0.005

ε=0.2
ε=0.2

ε=0.6
ε=0.6

ε=0.99
ε=0.99

Figure 3- Results for ZDT2 problem for various values of the

dispersion parameter - weight vector: (0.8, 0.2).
Figure 4- Results for DTLZ2 problem for various values of
the dispersion parameter - weight vector: (0.5, 0.3, 0.2).

793
 consideration the equilibrium between criteria 1 and 3, the solutions
a) obtained preserve the importance of criterion 3.
Pareto frontier A global comparison between WSFM and RP-EMO approaches
seems to indicate that WSFM approach is able to take into
consideration in a more equilibrated way the DM preferences. The
b) WSFM approach is able to perform well for convex and concave
problems and for problems with more than two criteria.

c)
(A)

d) i)
e)

ii)
Figure 5- Results for ZDT1 problem for various weight vectors:
a) (0.98,0.02), b) (0.8,0.2), c) (0.5,0.5), d) (0.2,0.8) and e)
(0.02,0.98). iii)
3.4 Comparison with the RP-EMO Approach
Figure 6 shows the results obtained for the ZDT1using: A) the
WSFM and B) the RP-EMO approach, for different weights values
and a single reference point located at the origin (i.e., the ideal
objective vector). Differences between the two methods are attained
manly for the set of weights i) and iii). For the first case (i) the
results obtained by RP-EMO approach are located around the point (B)
(0.2, 0.55), while for the WSFM they are located around the point
(0.1, 0.7). Taking into account that the aim is, mainly, to minimize
criteria f1 (since the weight attributed to this criteria is higher), the
region obtained by the WSFM seems to satisfy better the DM i)
preferences. The same type of justification can be given for the set
of weights iii), but now the higher importance is attributed to criteria
2. ii)
Figure 7 presents the results obtained for test problem ZDT2 under
the same conditions. In this case the main differences are attained iii)
for the set of weights represented by i) and ii). The solutions
obtained for case i) seems to not attribute any importance to
criterion 2 (f1≈0). Simultaneously, the solutions obtained by WSFM
approach (A) have, also, low values of f1, but without “forgetting”
some of the importance given to criterion 2. In the case of weight
vector ii), equal importance is attributed to both criteria. However,
the RP-EMO approach produces solutions with higher values for f1
(i.e., solutions that confer more importance to criterion 1). Again,
the WSFM approach seems to produce solutions more balanced, Figure 6- Comparison between the WSFM (A) and the RP-
taking into account that the DM is searching for solutions were both EMO (B) approaches for ZDT1 problem- weight vectors: i)
criteria have the same importance. (0.8,0.2), ii) (0.5,0.5) and iii) (0.2,0.8).

Finally, Figure 8 presents identical results for the DTZL2 problem,

using WSFM (i) and RP-EMO (ii) approaches. The weight vector
used was (0.07, 0.8, 0.13) and the reference point was located at (0,
0, 0). Taking into account that the importance attributed to criteria 1
and 3 is identical, the WSFM approach produces solutions were this
equilibrium (between criteria 1 and 3) is achieved and,
simultaneously, the higher importance for criterion 2 is considered.
The results obtained by the RP-EMO approach does not take into

794
 i) 3.5 Comparison with the WMM
(A) Figures 9, 10 and 11 show the results obtained using the WMM
approach for test problems ZDT1, ZDT2 and DTLZ2, respectively.
The results for the test problems with two criteria (ZDT1 and
ZDT2) produced by WSFM and WMM approaches are very similar.
ii) However, the comparison between these approaches for the DTLZ2
test problem (Figure 11) evidences some differences. As the RP-
EMO approach, the WMM approach gave more importance to
criterion 3 than to criterion 1, which does not consider the
equilibrium between these two criteria as defined by the weight
preferences vector defined by the DM.
iii)

i)
(B) i)
ii)
ii)
iii)

iii)

Figure 9- Results for ZDT1 problem using the WMM approach-

weight vectors: i) (0.8,0.2), ii) (0.5,0.5) and iii) (0.2,0.8).
Figure 7- Comparison between the WSFM (A) and RP-EMO
(B) approaches for ZDT2 problem- weight vectors: i) (0.8,0.2),
ii) (0.5,0.5) and iii) (0.2,0.8).
i)

ii)

i)
iii)

ii)

Figure 10- Results for ZDT2 problem using the WMM

Figure 8- Comparison between the i) WSFM and ii) RP-EMO approach - weight vectors: i) (0.8,0.2), ii) (0.5,0.5) and iii)
approach for DTLZ2 problem - weight vector for both: (0.2,0.8).
(0.07,0.8,0.13).

795
 5. ACKNOWLEDGMENTS
This work was supported by the Portuguese Fundação para a
Ciência e Tecnologia under grant POCTI/EME/48448/2002.

6. REFERENCES
[1] Branke, J., and Deb, K. Integrating User Preferences into
Evolutionary Multi-Objective Optimization. Technical report
i) (http://www.iitk.ac.in/kangal/reports.shtml), 2004.
[2] Miettinen, K. M. Nonlinear Multiobjective Optimization.
ii) Kluwer, Boston, 1999.
[3] Kaliszewski, I. Soft Computing for Complex Multiple Criteria
Decision Making. Springer, NY, 2006.
[4] Deb, K., Sundar, J., Bhaskara, U., and Chaudhuri, S. Reference
Point Based Multi-Objective Optimization Using Evolutionary
Algorithms. ISSN International Journal of Computational
Intelligence Research, 2, 3, (2006), 273-286.
[5] Deb, K., Pratap, A., Agrawal, S., and Meyarivan, T. A fast and
elitist multi-objective genetic algorithm: NSGA-II. IEEE
Figure 11- Comparison between the i) WSFM and ii) WMM Trans. on Evol. Comp., 6, 2 (2002), 182-197.
approaches for DTLZ2 problem - weight vector for both: [6] Deb, K. Multi-Objective Optimization using Evolutionary
(0.07,0.8,0.13). Algorithms. Wiley, Chichester, 2001.
[7] Fonseca, C. M., and Fleming, P. J. Genetic algorithms for
4. CONCLUSIONS multi-objective optimization: Formulation, discussion and
The WSFM approach allows the selection of the best solutions generalization. In Proceedings of the Fifth International
taking into account the DM preferences. This methodology Conference on Genetic Algorithms (ICG93), Morgan
produces good results for problems with two and three criteria and Kaufmann, San Mateo, CA,1993, 141-153.
for convex and non-convex Pareto frontiers. The method proposed [8] Fonseca, C. M., and Fleming, P. J. Multiobjective optimization
was not suitable for discontinuous fronts. However, a modification and multiple constraint handling with evolutionary algorithms,
of this method was been tested with good results for this type of part I: A unified formulation. IEEE Transactions on Systems,
problems. Man and Cybernetics, 28, 1 (1998), 26-37.
The implementation of this method allows the possibility of the DM [9] Ded, K., Thiele, L., Laumanns, M., and Zitzler E. Scalable
to conduct the search process for specific regions of the Pareto Multi-Objective Optimization Test Problems. IEEE
frontier. Also, the dispersion parameter allows one to control the Transactions on Evolutionary Computation, 1, (2002), 825-
size of the final Pareto frontier region obtained. 830.
The assessment with other methods available shown that the WSFM
approach is a good alternative for selecting solutions taking into
account the DM preferences.

796