IEEE LATIN AMERICA TRANSACTIONS, VOL. 21, NO.

11, NOVEMBER 2023 1157

Learning Decision Variables in Many-Objective

Optimization Problems
Artur Leandro da Costa Oliveira , Renê Gusmão and André Britto

Abstract—Traditional Multi-Objective Evolutionary Algo- that is constructed based on the data obtained previously from
rithms (MOEAs) have shown poor scalability in solving Many- the real objective function, called surrogates.
Objective Optimization Problems (MaOPs). The use of machine Despite the existence of related work on surrogates applied
learning techniques to enhance optimization algorithms applied
to MaOPs has been drawing attention due to their ability to add to multi- and many-objective optimization, they only explore
domain knowledge during the search process. One method of this learning from decision variables to estimate the objective
kind is inverse modeling, which uses machine learning models functions. However, since in MaOPs we have objective vectors
to enhance MOEAs differently, mapping the objective function with higher dimensions, could it be possible to learn the
values to the decision variables. The Decision Variable Learning behavior from objective functions to estimate the decision
(DVL) algorithm uses the inverse model in its concept and has
shown good performance due to the ability to directly predict variables?; since it is possible to learn the behavior from
solutions closed to the Pareto-optimal front. The main goal of decision variables using surrogates, is it possible to learn
this work is to experimentally show the DVL as an optimization inverse functions and from that functions, to predict sub-
algorithm for MaOPs. Our results demonstrate that the DVL optimal solutions to the problem?
algorithm outperformed the NSGA-III, a well-known MOEA The work done in [3] answers part of this question by using
from the literature, in almost all scenarios with restriction on the
number of objective functions with a high number of objectives. inverse surrogate models to generate more non-dominated
solutions for the decision-maker (DM), based on the approxi-
Index Terms—Many-Objective Optimization, Machine Learn- mated Pareto set obtained by MOEAs at the end of evolution-
ing, Inverse Surrogate Models, Decision Variable Learning.
ary optimization, increasing the density of the Pareto set found.
Some related work has been exploring the use of surrogates in
I. I NTRODUCTION recent years. In [4] a many-objective hybrid optimizer, called
MOHO, that uses five constitutive algorithms and actively
Multi-Objective Problem (MOP) is an optimization prob- switches among them throughout the optimization process.
A lem having two or more objective functions to be
optimized. MOPs that have more than three objectives are
Also, SAMaOEA [5], a reference vector guided evolutionary
algorithm assisted by radial basis function (RBF) models.
called Many-objective Optimization problems (MaOPs), and Furthermore, a surrogate-based evolutionary algorithm was
the field that studies new solutions for these problems is called proposed in [6] for expensive multi-objective optimization
Many-Objective Optimization. It is known that MOEAs scale problems.
poorly in many-objective optimization problems [1]. To solve In [7] is proposed the Decision Variable Learning (DVL)
these complex problems, the combination of Optimization algorithm that explores an inverse model approach in many-
algorithms with machine learning techniques has been the objective optimization. The DVL algorithm aims to design a
subject of research in the last few years [2]. Machine learning model, using machine learning, to represent the relationship
methods can be used to extract knowledge and integrate into between objective functions and variables, using the objective
the optimization process. This knowledge can take different vectors as input to estimate a set of solutions in decision space.
forms and result in different ways of integrating the knowledge In this process, reference points - which are ideal points, not
into the metaheuristic. guaranteed if they are possible - will be passed as input to
There has been an effort in the area to develop techniques the model. Thus, the model will return a set of solutions, with
that provide approximate models to represent the original objective vectors as close as possible to those reference points.
problem in a way that facilitates solving it, called meta- These initial DVL results help us to answer part of our
models. There are several meta-modeling methods, but they research question, however, there are still open topics that
can be divided into two main categories: classic function should be further explored to understand if it is possible to
transformation methods and surrogate modeling methods. The estimate the decision variables. The first topic is if a machine
first category corresponds to the techniques that seek to learning model can be trained to understand the relationship
represent the set of functions of the problem in only one, and between objectives and decision variables of an optimization
the second seeks a new expression for each objective function problem, and what is the error rate of this learning. The second
topic is related to the understanding of the impact of training
Artur Leandro da Costa Oliveira is with Graduate Pro- a model in different scenarios during an optimization process
gram on Computer Science, Federal University of Sergipe e- in comparison to the NSGA-III.
mail:[email protected].
Renê Gusmão and André Britto are with Departament of Computing, Fed- In this paper, we present an analysis through empirical
eral University of Sergipe e-mail:[email protected], [email protected]. experiments to validate the use of inverse surrogate models
OLIVEIRA et al.:LEARNING DECISION VARIABLES IN MANY-OBJECTIVE OPTIMIZATION PROBLEMS 1158

in Many-Objective Optimization. Here, DVL algorithm is the reproduction operator who benefits from the inverse model.
inverse model used. The experiments are intended to demon- The purpose of their method is to use the inverse models to
strate if it is possible to learn decision variables from objective increase the density of solutions in the preferred regions at a
functions, and validate the DVL as an algorithm for MaOPs, low computational cost.
analyzing the impact of different machine learning models The main contributions of the literature on inverse models
have on the algorithm. The experiments will be performed are based on using the knowledge extracted from the inverse
using MaOPs benchmark problems, and the results obtained model at the end of the optimization process, or internally
will be compared with the results of the NSGA-III algorithm. creating offspring solutions. These strategies limit the use of
The main contributions of our work are: (i) validation of inverse models in generating good solutions at the beginning of
machine learning algorithms to be used for estimating the the optimization process, decreasing the evolutionary time to
Pareto optimal set of optimization problems; (ii) comparison reach an optimal set. In a different manner, the DVL algorithm
of different classical machine learning algorithms as inverse uses the inverse model to generate, at each iteration, an entirely
models; and, (iii) analysis of the performance of the DVL algo- new solution set. Using the knowledge from the previous
rithm in solving different scenarios of optimization problems. iteration, the DVL algorithm produces solutions as close as
The remaining sections of this paper are organized as possible to the Pareto optimal set.
follows: Section II describes the related works, Section III
presents the algorithm studied in this work, Section IV de- III. D ECISION VARIABLE L EARNING (DVL)
scribes how the experiments were executed with the results
found, and finally, Section V concludes the paper. The Decision Variable Learning algorithm [7] aims to use
machine learning models to generate an approximated Pareto
optimal set using information from samples of the objective
II. R ELATED W ORK
space and a set of reference points. Differently from the Pareto
According to our research, the first attempt of constructing Estimation (PE), the DVL is not a method used after the
an inverse modeling method is made in [3], called Pareto MOEA, instead, it is an algorithm that uses the inverse model
Estimation (PE). Where an inverse functional mapping from approach into the algorithm to solve multi-objective problems.
the Pareto Front (PF) to the Pareto Set (PS) was built based on Although in the elaboration of the DVL the authors were not
the non-dominated solutions set obtained by the execution of aware of the studies relative to the inverse model, the DVL has
an MOEA at the end of evolutionary optimization. This model the theoretical basis of the PE method. But, instead of using
is used to generate additional non-dominated solutions, thereby the inverse model to increase the number of non-dominated
enhancing the density of the solutions set. The motivation solutions, the DVL directly infers an approximation of the
is that in many MOPs, principally in MaOPs, the decision- Pareto optimal set.
maker is not satisfied with the PS obtained by the MOEA. DVL is based on three main steps, the first one is the
In this work, it should be noted that the inverse model is not creation of an initial sample dataset, that contains solutions
used during the optimization process differently from the DVL with their respective objective functions. This initial sample
algorithm. generation is provided by the LHS method, which generates
In [8] the PE method is extended to improve the density well-distributed samples over the search space. The second
of available non-dominated solutions in multi-objective multi- step is the training of the inverse method, a machine learn-
modal problems. This kind of problem has multiples decision ing model is used for this purpose, which is trained with
vectors that map to identical objective vectors on the Pareto the dataset previously generated. The idea that supports this
front. In this case, the authors propose to subdivide the objec- modeling is that, for continuous MOPs, it can be deduced
tive space, by using a clustering algorithm, into different Cm from the Karush-Kuhn-Tucker optimality conditions that the
clusterings. So an ANN is trained for each cluster of solutions Pareto optimal set is a piecewise continuous in the decision
to identify the map F̃P̃C : P̃ → Cm . In this work, it is noted variables [3]. Depending on the machine learning model, Mn
m
that the quality of the resulting inverse model depends heavily models are created, and each one estimates only one variable,
on the training samples used. In [9], it is proposed the MAEA- so each solution generated in this process will be composed
GD/RD, an optimization algorithm based on the MAEA-gD of variables estimated by different models. Another strategy
with sample redistribution, which reallocates the samples in is to use only one model M that can be used to estimate the
the current population in objective space for training the entire decision variable set. The last step is the prediction of
inverse model. Doing this, more samples are present in regions new solutions next to the Pareto optimal set. The objectives
where higher quality topological information is more likely to set used as input for this step is the same as the PE method,
be obtained and used to improve the identification of F̃P . which is based on the method in [11], that places points on a
In [10] is proposed the IM-MOEA algorithm, motivated by normalized hyper-plane equally inclined to all objective axes
the PE method and the EDAs, provides a different way of and has an intercept of one on each axis. The algorithm was
using inverse models. In the IM-MOEA, the inverse models projected to work iteratively, where, at each iteration, the
are inserted internally in the optimization algorithm and then training of the machine learning model and prediction of a
used to create offspring solutions by sampling the objective new population is executed. These steps are performed using
space. They propose a decomposition of the inverse function, a set of solutions that is incremented at each iteration. The
also reducing the number of inverse models, and a particular purpose of this iterative approach is to improve the quality of
1159 IEEE LATIN AMERICA TRANSACTIONS, Vol. 21, No. 11, NOVEMBER 2023

the dataset and thereby improve future estimations since the In both experiments, Friedman’s test was used to check
dataset will have solutions closer to the reference points. whether there are statistically significant differences consid-
One real-world application for DVL is the Traffic Lights ering all algorithms. Then, the Wilcoxon rank-sum test is
Signaling [12]. Traffic light signaling optimization is one way adopted to compare the differences between pairs of configura-
of improving traffic efficiency. It relies on optimizing the tions. It was adopted 0.05 of the significance level. As pointed
traffic-light cycle, which is the sequence of states of a traffic out in [14] assumptions such as independence, normality, and
light. A traffic simulator can be used to obtain data on traffic homoscedasticity are most probably violated when analyzing
flow through traffic lights synchronization, which provides all the performance of stochastic algorithms based on compu-
the data necessary to simulate a traffic flow. The problem can tational intelligence. Thus, normally non-parametric tests are
be modeled as a MaOPs, since the decision variable time of used. The results of the statistical test were used to compute
each traffic-light cycle, and the objective functions are metrics the number of wins, ties and losses of an algorithm. An
obtained through the simulation such as quality measures algorithm wins when it obtains the higher mean values of
such as travel and stopped time, the number of vehicles that the metric and its results are statistically different to any
arrive at their destination, the overall average speed, among other algorithms’ result. In these cases, a loss is computed
others. In this problem even for a single intersection there can for the other algorithms. It is a tie when there is no statistical
be no obvious optimal solution. Furthermore, the simulation difference between the algorithm with the higher mean and
is costly. This real optimization problem has an expensive some other algorithm.
objective function and constraints, providing a new scenario In the default implementation of the DVL algorithm, n
for studying the DVL inverse modeling. models are created to estimate each decision variable. Different
from the strategy used in [3] where a single model is used
IV. E XPERIMENTS AND R ESULTS to estimate the entire decision variable set. Because of the
In this section, we evaluate the DVL algorithm through dependence between the input and the outputs that can occur
empirical experiments that have two main motivations: first, in optimization problems, the two strategies are applied in our
it is necessary to validate the possibility to learn decision experiments.
variables from objective functions, for this, we propose the Re-
gression experiment. Secondly, in the optimization experiment,
A. Learning Models
we validate the impact that different learning models have on
the algorithm. Also, we validate the DVL as an algorithm for A set of different regression machine learning models, well
Many-objective Optimization in comparison with a state-of- known in the literature, are used in our experiments to evaluate
the-art MOEA, the NSGA-III. the impacts of these models in learning the decision variables.
In our experiments, we used the first seven test problems The Linear model, an artificial neural network, more precisely
from the DTLZ benchmark test suite [13]: DTLZ1 to DTLZ7, a Multilayer Perceptron (MLP), a Random Forest regression
which are multiobjective problems for optimization that have (RFR), and a Support Vector Regression (SVR). Each one has
different configurations and shapes of the Pareto-optimal front. unique characteristics that make them perform well in certain
The problems are parameterized from a variable k, where types of problems. The implementation and the parameters of
k = n − m + 1, being n the number of variables and m the each model were defined by Scikit-learn, a python module that
number of objectives of the problem. Two different instances integrates a wide range of state-of-the-art machine learning
were defined for each problem, one with 3 objectives and 12 algorithms. Different from the default implementation of each
variables and the other with 10 objectives and 12 variables. model in the Scikit-learn, were used as parameters: (11, 11,
Although a problem with 3 objective is not considered a 11) for the number of neurons in the ith hidden layer, not
MaOPS, we adopted this scenario to analyze the scalabil- counting the input and output layers of the MLP. 100 for
ity performance of the compared algorithms. By the DTLZ the number of trees in the forest of the RFR, and 0.1 for
definition, the complexity of 10 objectives is lower than the the regularization parameter of the SVR. These values of the
complexity of 3 objectives. However, the variation of the parameters described were the best found by empirical analysis
objective number allows the identification of the influence of of running an experiment with the chosen models in a smaller
the objectives in the learning model. In a previous work [7], the test case with different parameters empirically and randomly
variation in decision variables showed no impact on algorithm chosen. In this smaller test case, it was adopted the same
performance, so the number of variables is set to 12 in both methodology described in this section.
cases. All the algorithms in the experiments were subjected With few exceptions, Machine Learning algorithms do not
to the same conditions, and our experiments aim to make a perform well when the input contains numerical attributes with
comparative analysis between them. significantly different scales. It is the case of the objective
Two different test configurations are described and analyzed. values of the DTLZ benchmark test suit. Therefore, for each
The first one is called the regression experiment, and the model was used the standardization transformation which is
second is the optimization experiment. The parameter settings an algorithm of feature scaling. The standardization subtracts
and the performance indicators used in each test are described the mean value (so standardized values always have a zero
at the beginning of each test subsection. In all experiments, mean), and then it divides by the variance so that the resulting
the algorithms are executed independently 20 times for each distribution has unit variance. This transformation causes a
configuration. better performance in our experiments.
OLIVEIRA et al.:LEARNING DECISION VARIABLES IN MANY-OBJECTIVE OPTIMIZATION PROBLEMS 1160

To validate and analyze the strategy of using one model by these variables in terms of divergence and convergence
to predict the entire decision variable set or using one model produced. For that, we made a visual comparison with the
to each decision variable, in our experiments, we applied the results obtained in the experiment and the Pareto-optimal front
multioutput regression in the models to obtain knowledge from values of each problem.
the dependence between inputs and outputs. The exception is Fig. 1 shows the visual comparison of the regression experi-
the SVR model, which implementation in Scikit-learn does not ment. For each problem, the model that achieves the best value
have support to multioutput regression. The models with this of MSE is used in comparison with the Pareto-optimal front.
strategy are denoted with an "MO" before the model name in Different models were used for comparison which indicates
our results. that the use of a specific model is linked to the problem that
will be solved. Each pair of charts represent the comparison
for 3 and 10 objectives respectively. In each chart, the green
B. Regression Experiment
color lines represent the Pareto-optimal front, the blue lines the
The first experiment is motivated to answer if it is possible objectives from the predicted decision variables. Every line in
to learn decision variables from objective functions. The task the chart details an objective function vector, the x-axis is the
of understanding and predicting the variables of the DTLZ position in the vector, and the y-axis is the objective function
benchmark test problems can be classified as supervised value contained in the vector position.
learning since we have the input and output for the learning In all charts, the models used achieved a high convergence,
algorithm and can be classified as a regression task since i.e. the values found are very close to the ideal. Only in
we want to predict a numerical value. This experiment is the DTLZ6 problem, Fig. 1k and Fig. 1l, that the predicted
based on two phases: training and testing. First, the model is objectives are a little far away from the ideal. DTLZ6 has a
trained by the execution of the DVL algorithm. The generated degenerate Pareto-optimal front as like DTLZ5, besides that,
samples and modeling strategy follow the default flow of the the DTLZ6 has many locals Pareto fronts which makes it diffi-
algorithm. Secondly, in the testing step, a subset of the Pareto- cult to converge to the Pareto-optimal front. In Figs. 1a, and 1e,
optimal front of each DTLZ problem serves as input to the respectively, DTLZ1, and DTLZ3 problems, the models could
model for prediction. The DTLZ benchmark test suit allows not find all parts of the Pareto-optimal for 3 objectives. It
the generation of the Pareto-optimal front of each problem occurred mainly due to the fact of the high search space
analytically. The key idea of this experiment is to analyze presented in these problems. Also, they have many local
if the inverse modeling of the DVL algorithm can correctly Pareto-optimal fronts, which makes it difficult to obtain a
predict the optimal set of a problem. In a real optimization good divergence on the predicted values. In the rest of the
problem, the optimal front is unknown, but the purpose of problems the predicted values obtained have performed well
this experiment is to validate the inverse modeling strategy in divergence and convergence of values. In these problems,
trained with solutions far from the optimal front. the shape of the region formed by the predicted values is
Each model in this experiment is trained with an initial quite similar to the shape of the Pareto-optimal front. It is
number of solutions generated through the LHS algorithm. important to note that the models had a greater divergence of
The size of the initial sample solutions is defined as 10000. values in the problems with 10 objectives, demonstrating a
For each DTLZ problem, approximately 5000 solutions of the shape closer to the ideal. This is mainly due to the increased
Pareto-optimal front were generated for the experiment. The number of features used for training which makes the models
values of these Pareto-optimal front were obtained through the perform better than with 3 objectives. Concerning the strategy
execution of diverse MOEAs without restriction and a post- of using 1 or n models to estimate each decision variable,
processing routine to diversify these solutions. The objective the results show that both strategies can be used. In different
function values of these generated solutions are computed optimizations problems, each one gets the best results.
and passed as input for each model to predict the decision The results found by the regression experiment are quite
variables. The predicted values of each model are evaluated promising. We show that different models could learn the
to the objective space. The mean-squared error (MSE) metric decision variables by a set of objective function values. Even
is used to make empirical comparisons between the results in more restricted scenarios, the models achieved good results.
of each model. The MSE is the average squared difference For the scope of this work, the models and techniques used
between the predicted optimal front and the Pareto-optimal were enough to validate the learning of decision variables.
front values.
The results of the regression experiment for 3 and 10
objectives are shown in Table I. The wins or ties are high- C. Optimization Experiment
lighted in bold. The model with the lower MSE value was In this experiment, each algorithm runs until it reaches
more successful in predicting decision variables. The results a stop condition, in the experiments we use the number of
obtained in this experiment show that different models have evaluations of the objective function as a stop condition. A
better performance on different types of problems. Also, set of a maximum number of evaluations to our experiment is
the number of objectives has interference in the results of defined as (250, 500, 1000, 1500, and 10000), all algorithms
MSE on each model for each problem. Although the MSE run with the same number of calls to the objective function.
obtained has relatively low values, it is necessary to evaluate This validation set is used to simulate different scenarios of
the predicted decision variables and the objectives produced restrictions in the use of the objective function. In this set, we
1161 IEEE LATIN AMERICA TRANSACTIONS, Vol. 21, No. 11, NOVEMBER 2023

(a) DTLZ1 - 3 objectives (b) DTLZ1 - 10 objectives (c) DTLZ2 - 3 objectives

(d) DTLZ2 - 10 objectives (e) DTLZ3 - 3 objectives (f) DTLZ3 - 10 objectives

(g) DTLZ4 - 3 objectives (h) DTLZ4 - 10 objectives (i) DTLZ5 - 3 objectives

(j) DTLZ5 - 10 objectives (k) DTLZ6 - 3 objectives (l) DTLZ6 - 10 objectives

(m) DTLZ7 - 3 objectives (n) DTLZ7 - 10 objectives

Fig. 1. Visual Comparisons for the DTLZ benchmark. (a) SVR model, (b) RFR model, (c) SVR model, (d) RFR model, (e) SVR model, (f) RFR model, (g)
RFR model, (h) RFR model, (i) SVR model, (j) RFR model, (k) MO-MLP model, (l) MO-Linear model, (m) MO-Linear model, and (n) MO-Linear model.
OLIVEIRA et al.:LEARNING DECISION VARIABLES IN MANY-OBJECTIVE OPTIMIZATION PROBLEMS 1162

TABLE I
S TATISTICAL RESULTS ( MEAN AND STANDARD DEVIATION ) OF THE MSE VALUES FOR 3 AND 10 OBJECTIVES .

Obj. Prob. MO-Linear SVR MO-MLP MO-RFR Linear MLP RFR

3 DTLZ1 (1.42±0.01)e-2 (1.41±0.00)e-2 (1.84±0.26)e-2 (3.60±0.87)e-2 (1.42±0.01)e-2 (2.44±0.63)e-2 (3.62±0.21)e-2
3 DTLZ2 (4.49±0.02)e-3 (2.20±0.14)e-3 (3.97±0.99)e-3 (7.91±1.96)e-3 (4.50±0.02)e-3 (4.37±1.35)e-3 (3.69±0.37)e-3
3 DTLZ3 (1.43±0.00)e-2 (1.41±0.00)e-2 (1.98±0.39)e-2 (4.59±0.95)e-2 (1.43±0.01)e-2 (2.75±0.66)e-2 (3.80±0.30)e-2
3 DTLZ4 (1.20±0.04)e-2 (1.31±0.16)e-2 (3.67±1.27)e-2 (9.17±1.01)e-3 (1.19±0.05)e-2 (5.89±2.32)e-2 (9.39±1.26)e-3
3 DTLZ5 (8.58±0.02)e-3 (8.21±0.11)e-3 (9.98±1.67)e-3 (1.42±0.14)e-2 (8.59±0.02)e-3 (1.14±0.11)e-2 (1.03±0.03)e-2
3 DTLZ6 (1.11±0.03)e-1 (2.68±0.03)e-1 (9.41±3.83)e-2 (1.10±0.07)e-1 (1.11±0.04)e-1 (1.17±0.67)e-1 (1.20±0.08)e-1
3 DTLZ7 (2.91±0.75)e-4 (2.83±0.08)e-1 (5.08±1.97)e-2 (7.65±0.26)e-2 (3.00±0.75)e-4 (6.17±1.89)e-2 (6.77±0.72)e-2
10 DTLZ1 (6.39±0.00)e-2 (6.54±0.13)e-2 (6.45±0.07)e-2 (5.95±0.10)e-2 (6.39±0.00)e-2 (5.80±0.14)e-2 (1.76±0.33)e-2
10 DTLZ2 (3.53±0.01)e-2 (3.47±0.01)e-2 (3.07±0.21)e-2 (1.80±0.03)e-2 (3.52±0.01)e-2 (7.07±0.39)e-3 (1.32±0.04)e-3
10 DTLZ3 (6.39±0.00)e-2 (6.41±0.04)e-2 (6.44±0.04)e-2 (6.47±0.15)e-2 (6.39±0.00)e-2 (5.87±0.13)e-2 (1.90±0.19)e-2
10 DTLZ4 (5.41±0.01)e-2 (5.37±0.03)e-2 (6.87±0.16)e-2 (4.33±0.06)e-2 (5.41±0.02)e-2 (5.93±0.12)e-2 (3.91±0.04)e-2
10 DTLZ5 (5.69±0.00)e-2 (5.69±0.00)e-2 (5.75±0.03)e-2 (6.14±0.05)e-2 (5.69±0.00)e-2 (5.82±0.05)e-2 (6.17±0.04)e-2
10 DTLZ6 (6.81±0.07)e-2 (3.01±0.08)e-1 (1.45±0.29)e-1 (9.41±0.48)e-2 (6.83±0.08)e-2 (1.01±0.22)e-1 (8.25±0.18)e-2
10 DTLZ7 (4.99±0.59)e-5 (9.73±0.80)e-2 (3.10±1.35)e-2 (4.30±0.05)e-2 (4.74±0.64)e-5 (4.97±1.54)e-3 (1.98±0.30)e-3

classify as a higher restriction level a number of validations 34 of 35 scenarios. With the MLP model having 24 wins
lower or equal to 1000. The empirical comparisons between in comparison to the others. In the experimentation with 3
the results are made by the use of Hypervolume (HV), a perfor- objectives, the NSGA-III wins a total of 14 in 35, and the
mance indicator that assesses both convergence and divergence combination of all DVL different models wins 15.
of the solutions. Let y ∗ be a reference point dominated by For 3 objectives, only the SVR and the MO-Linear models
all the Pareto front solutions, and P the population resulting achieved good results of hypervolume in DTLZ1, DTLZ3, and
of a MOEA. The HV of P is the calculus of the volume DTLZ7, just as they obtained in the regression experiment.
of the region that dominates y ∗ and is dominated by P . For 10 objectives, no model that got the best result in the
The HV values presented Qm in this work are all normalized regression experiment achieved the best HV mean. What
to [0, 1] by dividing i=1 yi∗ . In this work, the following draws the most attention is the performance of the SVR in
reference points y ∗ are used for each problem: for DTLZ1, 3 objectives, and the two MLP versions, MLP and MO-MLP,
y ∗ = (300.0, 300.0, ..., 300.0); DTLZ2 and DTLZ4, y ∗ = which achieved good results in 3 and 10 objectives. Mainly
(2.0, 2.0, ..., 2.0); DTLZ3, y ∗ = (700.0, 700.0, ..., 700.0); because although the MLP does not obtain the lowest MSE
DTLZ5, y ∗ = (4.0, 6.0, ..., 2(m + 1)); DTLZ6 and DTLZ7, value, the generated solutions have a good diversity which
y ∗ = (10.0, 18.0, ..., 2(4m + 1)) for m number of objectives. implies a better HV. Besides that, the architecture of the DVL
The DVL algorithm needs four parameters to be predefined algorithm allows new training at each iteration with objectives
for execution. The value c of the number of closest solutions closer to the ideal. As the model is reused during iterations,
of the reference points used for training. A k value for the this task can be seen as transfer learning and many studies have
initial number of population used by the LHS. A value i demonstrated the superior performance of neural networks
of the maximum number of iterations and a value ϵ of the compared to statistical methods in transfer learning.
difference between the HV of the population generated. In this In comparison with the NSGA-III algorithm, for 3 objec-
experiment the c value is fixed in 200 and the ϵ is set to 0.001, tives, in the problems DTLZ1-4 that have simpler Pareto-
the i and the k have different values for each configuration of optimal fronts, all of them triangular with DTLZ1 being plan
number objectives vs. number of evaluations. For 3 objectives: and DTLZ2, DTLZ3 and DTLZ4 a quarter of a sphere, the
i = 1 and k = 159 for 250 number of evaluations; i = 3 and DVL algorithm was superior only when there is a higher
k = 227 for 500; i = 7 and k = 363 for 1000; i = 9 and restriction in the number of evaluations. In a less restrictive
k = 681 for 1500 and i = 14 and k = 8726 for 10000. For 10 scenario, where the number of evaluations is 1500 and 10000,
objectives: i = 1 and k = 50 for 250 number of evaluations; the NGSA-III got better results in these problems. For the
i = 2 and k = 112 for 500; i = 4 and k = 132 for 1000; problems DTLZ5-7 the scenario is quite different, with the
i = 5 and k = 464 for 1500 and i = 12 and k = 7388 for DVL being superior in most cases. These problems are char-
10000. acterized by having Pareto-optimal fronts with different shapes
The parameters used in NSGA-III were defined by Pymoo, a from the previous ones. The case of DTLZ5 and DTLZ6 both
Python framework for optimization. For the simulated binary have degenerated Pareto-optimal fronts, and DTLZ7 has a
crossover, the distribution index is set to ηc = 30, and the disconnected front. These kinds of problems make it difficult
crossover probability pc = 1.0. For the polynomial mutation, for MOEAs to find the ideal frontier.
the distribution index and the mutation probability are set to The results of the experiment with 10 objectives show
ηm = 20 and pm = 1/n, respectively, as recommended that the DVL algorithm was superior in the vast majority
in [15]. The population size used was 91 and 220 for the prob- of cases, except for the case of the DTLZ4 problem with
lem with 3 and 10 objectives, respectively, and the individual 10000 evaluations, in which the NSGA-III is superior to
has a length of 12 for both objective sizes. the DVL with any model. This shows that the DVL can
Table II summarizes the results of the statistical comparison also be used for optimization in problems with a higher
in terms of wins, draws, and losses. It is possible to note the number of objectives. We can deduce from the experiments
DVL superiority in problems with 10 objectives by winning that the DVL has a good performance in different kinds of
1163 IEEE LATIN AMERICA TRANSACTIONS, Vol. 21, No. 11, NOVEMBER 2023

problems with different Pareto-optimal shapes, search space [6] S. Liu, H. Wang, W. Yao, and W. Peng, “Surrogate-assisted environmen-
sizes, with different local and global optimal, and a different tal selection for fast hypervolume-based many-objective optimization,”
IEEE Transactions on Evolutionary Computation, pp. 1–1, 2023.
number of objectives. According to the experiments, this last [7] M. Santos, J. A. de Oliveira, and A. Britto, “Decision variable learning,”
characteristic, the number of objectives, is a key factor for the in 2019 8th Brazilian Conference on Intelligent Systems (BRACIS),
greater performance of the DVL in comparison to the NSGA- pp. 497–502, IEEE, 2019.
[8] R. Kudikala, I. Giagkiozis, and P. Fleming, “Increasing the density
III. The results also show that the strategy used for learning as of multi-objective multi-modal solutions using clustering and pareto
well as the model is problem-dependent. In different scenarios, estimation techniques,” in The 2013 World Congress in Computer
different strategies got better results, with the MO-MLP and Science Computer Engineering and Applied Computing, 2013.
[9] Y. Yan, I. Giagkiozis, and P. J. Fleming, “Improved sampling of
MLP having the best overall among these strategies. decision space for pareto estimation,” in Proceedings of the 2015 Annual
Conference on Genetic and Evolutionary Computation, pp. 767–774,
TABLE II 2015.
W INS /T IES /L OSSES OF THE S TATISTICAL C OMPARISON . [10] R. Cheng, Y. Jin, K. Narukawa, and B. Sendhoff, “A multiobjective
evolutionary algorithm using gaussian process-based inverse model-
ing,” IEEE Transactions on Evolutionary Computation, vol. 19, no. 6,
3 objectives 10 objectives pp. 838–856, 2015.
[11] I. Das and J. E. Dennis, “Normal-boundary intersection: A new method
MO-Linear 0 / 4 / 31 1 / 0 / 34 for generating the pareto surface in nonlinear multicriteria optimization
SVR 4 / 0 / 31 0 / 0 / 35 problems,” SIAM journal on optimization, vol. 8, no. 3, pp. 631–657,
MO-MLP 1 / 2 / 32 8 / 0 / 27 1998.
[12] J. J. Sánchez-Medina, M. J. Galán-Moreno, and E. Rubio-Royo, “Traffic
MO-RFR 0 / 0 / 35 0 / 0 / 35 signal optimization in" la almozara" district in saragossa under conges-
NSGA-III 14 / 2 / 19 1 / 1 / 33 tion conditions, using genetic algorithms, traffic microsimulation, and
Linear 1 / 3 / 31 0 / 1 / 34 cluster computing,” IEEE Transactions on Intelligent Transportation
Systems, vol. 11, no. 1, pp. 132–141, 2009.
MLP 5 / 1 / 29 24 / 0 / 11 [13] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, “Scalable test problems
RFR 4 / 0 / 31 0 / 0 / 35 for evolutionary multiobjective optimization,” in Evolutionary multiob-
jective optimization, pp. 105–145, Springer, 2005.
[14] J. Derrac, S. García, D. Molina, and F. Herrera, “A practical tutorial
on the use of nonparametric statistical tests as a methodology for
V. C ONCLUSION comparing evolutionary and swarm intelligence algorithms,” Swarm and
Evolutionary Computation, vol. 1, no. 1, pp. 3–18, 2011.
In this paper, we showed experimentally that the DVL algo- [15] K. Deb and H. Jain, “An evolutionary many-objective optimization
rithm and its strategy of inverse modeling can compete with algorithm using reference-point-based nondominated sorting approach,
well-known MOEAs of the literature, such as the NSGA-III in part i: Solving problems with box constraints,” IEEE Transactions on
Evolutionary Computation, vol. 18, no. 4, pp. 577–601, 2014.
the optimization of benchmark problems. Experimental results
demonstrate that the inverse model can estimate the Pareto-
optimal front of the DTLZ benchmark test suit problems, Artur Leandro da Costa Oliveira received the
and also, the DVL algorithm can outperform the NSGA-III B.S. and M.S. degrees in computer science from
the Federal University of Sergipe. São Cristóvão,
algorithm in restrictive scenarios, i.e. with restrictions of the Sergipe, Brazil, in 2011 and 2022 respectively. From
number of evaluations, and in problems with a higher number 2011 to 2022, he work as a System Analyst. His
of objectives. The overall result found is quite promising, the research interest includes optimization and machine
learning areas.
DVL produced a valuable performance against the NSGA-III
in many scenarios, even though the DVL algorithm is recent
and also it needs greater maturity for optimization tasks.
For future works, we suggest the improvement of the weak
points of the DVL algorithm, such as the deterioration of Renê Gusmão received the Ph.D. degree in com-
the search performance due to local optima and the shape puter science from Federal University of Pernam-
of the reference points and to test DVL algorithm needs to buco, Brazil. He is an adjunct professor of the
Federal University of Sergipe. His current research
be experimentally tested in real-world optimization expensive interests include cluster analysis, data mining and
problems. computational intelligence.

R EFERENCES
[1] O. Schütze, A. Lara, and C. A. C. Coello, “On the influence of the
number of objectives on the hardness of a multiobjective optimization
problem,” IEEE Trans. Evolutionary Computation, vol. 15, no. 4,
pp. 444–455, 2011. André Britto received the Ph.D. in computer sci-
[2] L. Calvet, J. de Armas, D. Masip, and A. A. Juan, “Learnheuristics: ence department from Federal University of Parana,
hybridizing metaheuristics with machine learning for optimization with Parana, Brazil. He is an adjunct professor at the
dynamic inputs,” Open Mathematics, vol. 15, no. 1, pp. 261–280, 2017. Federal University of Sergipe, Sergipe, Brazil. His
[3] I. Giagkiozis and P. J. Fleming, “Pareto front estimation for decision main interests are multi-objective optimization and
making,” Evolutionary computation, vol. 22, no. 4, pp. 651–678, 2014. machine learning.
[4] S. Reddy and G. S. Dulikravich, “A self-adapting algorithm for many-
objective optimization,” Applied Soft Computing, vol. 129, p. 109484,
2022.
[5] “Multi/many-objective evolutionary algorithm assisted by radial basis
function models for expensive optimization,” Applied Soft Computing,
vol. 122, p. 108798, 2022.