ISSN 1440-771X

Australia

Department of Econometrics and Business Statistics

http://www.buseco.monash.edu.au/depts/ebs/pubs/wpapers/

Efficient Identification of the Pareto Optimal Set

Ingrida Steponavičė Rob Hyndman, Kate Smith-Miles

and Laura Villanova

April 2014

Working Paper 12/14

Efficient Identification of the Pareto Optimal Set

Ingrida Steponavičė1 , Rob J. Hyndman2 , Kate Smith-Miles1

and Laura Villanova3
1
School of Mathematical Sciences, Monash University, Clayton, Australia
2
Department of Econometrics & Business Statistics, Monash University, Clayton,
Australia
3
Ceramic Fuel Cells Limited, Noble Park, Australia

Abstract. In this paper, we focus on expensive multiobjective opti-

mization problems and propose a method to predict an approximation
of the Pareto optimal set using classification of sampled decision vec-
tors as dominated or nondominated. The performance of our method,
called EPIC, is demonstrated on a set of benchmark problems used in
the multiobjective optimization literature and compared with state-of
the-art methods, ParEGO and PAL. The initial results are promising
and encourage further research in this direction.

Keywords: multiobjective optimization, classification, expensive black-box func-

tion

1 Introduction
Many real-world optimization applications in engineering involve problems where
analytical expression of the objective function is unavailable. Such problems usu-
ally require either an underlying numerical model or expensive experiments to
be conducted. In an optimization setting, where objective functions are evalu-
ated repeatedly, evaluation demands may result in unaffordably high cost for
obtaining solutions. Therefore, the number of function evaluations is limited by
available resources. Consequently, the solution of such global optimization prob-
lems is challenging because many global optimization methods require a large
number of function evaluations.
This task becomes even more difficult in the case of multiple conflicting
objectives, where there is no single optimal solution optimizing all objective
functions simultaneously. Rather there exists a set of solutions representing the
best possible trade-offs among the objectives — the so-called Pareto optimal
solutions forming a Pareto optimal set. Unreasonably high evaluation costs could
also prevent designers from comprehensively exploring the decision space and
learning about possible trade-offs. In these cases, it is essential to find reliable
and efficient methods for estimating the Pareto optimal set within a limited
number of objective function evaluations.
Recently, researchers have developed methods to solve expensive problems by
exploiting knowledge acquired during the solution process [1]. Knowledge of past
2

evaluations can be used to build an empirical model, called a surrogate model,

that approximates the objective function. This approximation can then be used
to predict promising new solutions at a smaller evaluation cost than that of the
original problem [2]. Commonly used surrogate models include Gaussian process
(or Kriging) models [3, 4], polynomial response surface models [5], radial basis
functions [6], support vector regression [7], and local polynomial [8].

One of the state-of-art methods for expensive multiobjective optimization

problems, named ParEGO, was developed by Knowles [9]. It is essentially a
multiobjective translation of the efficient global optimization (EGO) method [1],
where multiple objectives are converted to a single objective using a scalarization
function with different parameter values at each step.

The Pareto active learning (PAL) method for predicting the Pareto optimal
set at low cost has been proposed in [10]. Like ParEGO, it employs a Gaussian
process (GP) model to predict objective function values. PAL classifies all sam-
pled decision vectors as Pareto optimal or not based on the predicted objective
function values. The classification accuracy is controlled by a parameter defined
by the user which enables a trade-off between evaluation cost and predictive
accuracy.

The main disadvantage of employing a GP is that model construction can be

a very time-consuming process [11], where the time increases with the number of
evaluated vectors used to model the GP. To overcome this issue, ParEGO uses
a subset of the evaluated vectors to build the GP model, thus attempting to
balance model accuracy and computation time. Moreover, using a GP becomes
increasing problematic in high dimensional spaces [7], so these methods do not
scale well as the dimension of the problem increases.

In this paper, we propose an efficient Pareto iterative classification (EPIC)

method which approximates the Pareto optimal set with a limited number of
objective function evaluations. At the core of the proposed method is the clas-
sification of decision vectors to be either dominated or nondominated without
requiring any knowledge about the objective space. The method iteratively clas-
sifies sampled decision vectors to one of the two classes until a stopping criterion
is met. It returns the set of decision vectors predicted to be nondominated. A
major advantage of this method is that it does not use any statistical model of
the objective function, such as GP, and so it involves more modest computational
requirements, and scales easily to handle high dimensional spaces.

The paper is organized as follows. Section 2 introduces the main concepts

involved in multiobjective optimization. The EPIC method is described in Sec-
tion 3. In Section 4, we outline an experiment setup, demonstrate the perfor-
mance of EPIC on benchmark problems, and compare our experimental results
with that of the ParEGO and PAL methods. Section 5 draws some conclusions
and briefly discusses some future research directions.
 3

2 Background

A multiobjective optimization problem has the form [12]

T
min f (x) = f1 (x), . . . , fk (x) (1)
subject to x ∈ S,

where S ⊂ Rn is the feasible set and fi :→ R, i = 1, . . . , k (k ≥ 2), are objective

functions that are to be minimized simultaneously. All objective functions are
represented by the vector-valued function f : S → Rk . A vector x ∈ S is called
a decision vector and a vector z = f (x) ∈ Rk an objective vector.
In multiobjective optimization, the objective functions f1 , . . . , fk in (1) are
typically conflicting. In that case, there does not exist a decision vector x̄ ∈ S
such that x̄ minimizes fi in S for all i = 1, . . . , k, but there exists a number
(possibly infinite) of Pareto optimal solutions. In mathematical terms, a decision
vector x̄ ∈ S and its image z̄ = f (x̄) are said to be Pareto optimal or nondom-
inated if there does not exist a decision vector x ∈ S such that fi (x) ≤ fi (x̄)
for all i = 1, . . . , k and fj (x) < fj (x̄) for some j = 1, . . . , k. If such a decision
x ∈ S does exist, x̄ and z̄ are said to be dominated by and its image z = f (x),
respectively.
Sometimes in the literature, Pareto optimal and nondominated solutions are
regarded as synonyms. However, in this paper, nondominated solutions refer to
solutions which are not dominated by any other solution in the set of evaluated
solutions. If an objective vector z1 = f (x) does not dominate an objective vector
z2 = f (x̄), this does not imply that z2 dominates z1 (e.g., they can be both
nondominated in a given set). Moreover, this does not imply Pareto optimality
as nondominance is defined subject to the set of available objective vectors.

3 The EPIC method

The EPIC method is designed for expensive multiobjective optimization prob-

lems where the number of possible objective function evaluations is limited due
to time, financial, or other costs. The main idea behind the proposed method is
the classification of a set of sampled decision vectors into two classes: dominated
and nondominated. The method comprises five steps: (i) evaluation of an initial
set; (ii) training a classifier; (iii) predicting labels for unevaluated decision vec-
tors; (iv) checking stopping conditions; and (v) selection of a decision vector to
evaluate next. Steps (ii)–(v) are repeated until stopping conditions are satisfied.
The pseudocode of EPIC is given in Algorithm 1.
Given a set of decision vectors S which can be generated using any sampling
technique (e.g., Hammersley sequence or Latin hypercube sampling), the EPIC
method selects and evaluates expensive objective functions for a small number of
decision vectors called an initial set Xinit . Then the evaluated decision vectors,
which comprise the set of all evaluated decision vectors E, are checked for non-
dominance and divided into dominated and nondominated classes. The decision
4

Algorithm 1 EPIC pseudocode

Input: sampled decision space S, pnext , pnond
Output: predicted Pareto optimal set P
E = Xinit , XU = S\E, i = 0
while i ≤ n do
i = i + 1, P = ∅
obtain labels for all x ∈ E
train a classifier
calculate the probability p of each x ∈ XU
for all x ∈ XU do
if p ≥ pnond then P = P ∪ {x}
end if
end for
xe = arg minx∈XU |p − pnext |
E = E ∪ {xe }, XU = XU \{xe }
end while

vectors and their labels “nondominated” and “dominated” are used to train a
classifier, for example, support vector machines or a Naive Bayes classifier. After
a classifier is trained, unevaluated decision vectors comprising a set XU are given
as an input to the classifier to predict their labels. The classifier provides prob-
ability p for each decision vector x that it belongs to the nondominated class.
If the probability of nondominance p is not lower than a predefined probability
pnond , the decision vector is included in a predicted Pareto optimal set P . Then
a new decision vector xe is selected for evaluation. The implemented selection
strategy in this paper is simply to select a decision vector xe whose probability
of nondominance p is closest to a predefined value pnext . Next, the evaluated
set E and its labels are updated and used to retrain the classifier. The method
continues until the number of evaluations reaches a predefined limit n.
If the sampling of the design space is sufficiently dense, and the number of
evaluations n is sufficiently large, then EPIC should provide a good approx-
imation of the Pareto optimal set. The efficiency of EPIC is determined by
the classification quality and the strategy for determining which decision vector
should be evaluated next.
It is clear that the next vector to be evaluated should be selected based on the
maximal information gain or uncertainty reduction which would help to improve
classification. We assume that evaluating a decision vector near to the boundary
that separates the two classes is most likely to provide more information and
improve classifier performance. An obvious choice is to set pnext = 0.5. However,
further thought suggests that alternative values of pnext may be preferred. The
initial set used to train the classifier is a small subset of the sampled decision
space and the vectors labelled “nondominated” might be mis-classified since they
are assigned with respect to the evaluated vectors. Therefore it may be better to
first evaluate decision vectors that are most likely nondominated (i.e., pnext >
0.5) in order to get a better representation of the nondominated class. Later, it
 5

may be better to concentrate on uncertainty related to the boundary separating

the two classes. Alternatively, selecting several decision vectors based on different
pnext values and evaluating them simultaneously could improve classification
resulting in better overall method performance.
One more parameter that has to be chosen is the probability pnond , used to
determine which decision vectors will be included in the predicted Pareto optimal
set. The prediction of the Pareto optimal set is made at each iteration, with the
predicted set obtained in the previous iteration replaced by the newly predicted
set. A higher value of pnond will result in a smaller set of points classified to be
nondominated. Again, the value of this parameter needs to be selected carefully
in order to control the efficiency of the algorithm in determining the Pareto
optimal set.
We leave the optimal selection of pnext and pnond to future research. In this
paper, we have selected pnond = 0.6, and have considered several values of pnext =
{0.6, 0.7, 0.8, 0.9}.
The main advantages of the EPIC method are: i) simplicity to implement,
ii) computational speed; iii) multiple decision vectors can be selected at each
iteration; and iv) has no limitations on high dimensional problems.

4 Experimental Results
4.1 Experimental setup
To assess the performance of EPIC, we compare it to PAL and ParEGO. For
that we need to have some performance measures. One of the measures is based
on a hypervolume (HV) metric [13] that describes the spread of the solutions
over the Pareto optimal set in the objective space as well as the closeness of
the solutions to it. Moreover, HV is the only measure that reflects the Pareto
dominance [14]. That is, if one set entirely dominates another, the HV of the
former will be greater. As PAL and EPIC are based on classification, the quality
of prediction is also measured by the percentage of correctly classified decision
vectors. We calculated other metrics as well such as set coverage. However, they
are not so informative and we do not include them in this work.

4.2 Methods comparison

EPIC and PAL differ from ParEGO as they do not generate new decision vectors
but rather select them from the sampled decision space. We used Latin hypercube
sampling to sample the decision space for both EPIC and PAL. The sampling size
selected for all problems is 500 decision vectors. ParEGO and EPIC were allowed
to evaluate 200 or 250 decision vectors, while PAL was running until all the
decision vectors were assigned to one of the classes. As the method performance
depends on the initial set, we ran the methods with 100 different initial sets and
calculated the average values of the performance metrics. All the methods were
run with the same initial sets consisting of 11d − 1 decision vectors, where d is
the dimension of the decision space, as proposed in [1].
6

The performance of the methods was measured at every iteration to assess the
progress obtained after each decision vector evaluation. We calculated the HV
metric of evaluated decision vectors for all the methods. For ease of comparison,
we considered the ratio between the HV obtained by each method and the HV of
the true Pareto optimal set. It should be noted the HV value calculated using the
evaluated vectors does not decrease with an increasing number of evaluations.
For the average HV metric calculation, when for some runs the PAL method
terminated earlier and did not used the maximum number of iterations, we
used the same nondominated set evaluated at the last iteration for the rest of
iterations.
The methods were tested on the following set of standard benchmark prob-
lems in multiobjective optimization with different features:
OKA2 [15]. This problem has two objective functions and three decision vari-
ables. Its true Pareto optimal set in the objective space is a spiral shaped
curve and the density of the Pareto optimal solutions in the objective space
is low. (The reference point is R = (6.00, 6.0483).)
Kursawe [16]. This problem has two objective functions and a scalable number
of decision variables. In our experiment, three decision variables were used.
Its Pareto optimal set in the decision space is disconnected and symmetric,
and disconnected and concave in the objective space. (The reference point
is R = (-3.8623, 25.5735).)
ZDT3 [17]. This has two objective functions and three decision variables. The
Pareto optimal set in the objective space consists of several noncontiguous
convex parts. However, there is no discontinuity in the decision space. (The
reference point R = (2.0000, 2.6206).)
Viennet [18]. This consists of three objective functions and two decision vari-
ables. This problem has not been solved with the PAL algorithm as we used
an implementation provided by the authors which is suitable only for prob-
lems with two objective functions. (The reference point is R = (9.249, 62.68,
1.1964).)
DTLZ4 [19]. This problem is scalable and has M objective functions and k +
M −1 of decision variables, where k = 10 as recommended by the authors. We
solved this problem consisting of 5, 6 and 7 objectives and, respectively, 14, 15
and 16 decision variables. (The reference points for the problem with 5, 6 and
7 objectives is R = (3.9324, 3.2452, 3.4945, 3.4114, 3.3022), R = (3.9943,
3.2319, 3.3666, 3.1851, 3.3236, 3.2196) and R = (3.7703, 3.3593, 3.3192,
3.3825, 3.4326, 3.2446, 3.3209), respectively.)
To classify decision vectors as dominated and nondominated we applied a
support vector machine (SVM) [20]. The basic idea of SVM classifiers is to
choose the hyperplane that has the maximum distance between itself and the
nearest example of each class [20, 21]. SVMs are computationally deficient clas-
sifiers and can deal with both linear and nonlinear as well as separable and
nonseparable problems. We used SVM with a radial basis function kernel allow-
ing to capture nonlinear relation between class labels and features by mapping
data to a higher dimensional space. The drawback of using SVMs is that they
 7

do not directly produce probability estimates. However, these can be calculated

using different strategies, for example, Platt’s method [22] or isotonic regression
[23]. For EPIC implementation we used LIBSVM, a library for SVMs [24], which
provides probability estimates.
We experimented with different pnext values: 0.6, 0.7, 0.8, and 0.9. However,
we did not find any appreciable performance differences as well as no single pnext
value provided the best results for all considered problems.

4.3 Methods comparison

All of the experiments were repeated 100 times and the average metrics are
presented in Figures 1–7. Looking at the plots of correct classification, it can
be noted that at the very first iterations, EPIC classification is not as good as
that of PAL, but its classification improves fast and within 20 iterations EPIC
outperforms PAL for all three problems. This can be explained by the fact that
when the training set (consisting of evaluated vectors) increases, the classifier
learns about the boundary separating the two classes.
The HV metric plots show that the performance of the methods is problem
dependent. For example, we cannot distinguish any method to be best when
considering OKA2 and ZDT3 problems as the HV curves overlap. However,
when solving DTLZ4 problem with 5, 6 and 7 objectives, the HV measure clearly
indicates that EPIC outperforms ParEGO.

OKA2
100
Correct classification, %

99

98

97
0 20 40 60 80 100 120 140 160
Iterations
PAL EPIC ParEGO
80

75

70
HV, %

65

60

55
0 20 40 60 80 100 120 140 160
Iterations

Fig. 1. Comparative performance on the OKA2 problem

8

Kursawe
100

Correct classification, % 99.5

99

98.5

98

97.5
0 20 40 60 80 100 120 140 160
Iterations
PAL EPIC ParEGO
100

90
HV, %

80

70

60
0 20 40 60 80 100 120 140 160
Iterations

Fig. 2. Comparative performance on the Kursawe problem

ZDT3
100
Correct classification, %

95

90

85

80
0 20 40 60 80 100 120 140 160
Iterations
PAL EPIC ParEGO
100

98
HV, %

96

94

92
0 20 40 60 80 100 120 140 160
Iterations

Fig. 3. Comparative performance on the ZDT3 problem

 9

Viennet
100

Correct classification, % 95

90

85

80
0 20 40 60 80 100 120 140 160
Iterations
EPIC ParEGO
100

98

96
HV, %

94

92

90
0 20 40 60 80 100 120 140 160
Iterations

Fig. 4. Comparative performance on the Viennet problem

DTLZ4: obj = 5
82
Correct classification, %

81

80

79

78

77
0 10 20 30 40 50 60 70 80 90
Iterations
EPIC ParEGO
88

86
HV, %

84

82
0 10 20 30 40 50 60 70 80 90
Iterations

Fig. 5. Comparative performance on the DTLZ4 problem with 5 objectives

10

DTLZ4: obj = 6
65

Correct classification, %
60

55

50
0 10 20 30 40 50 60 70 80
Iterations
EPIC ParEGO
88

87

86
HV, %

85

84

83
0 10 20 30 40 50 60 70 80
Iterations

Fig. 6. Comparative performance on the DTLZ4 problem with 6 objectives

DTLZ4: obj = 7
52
Correct classification, %

51

50

49

48

47
0 10 20 30 40 50 60 70
Iterations
EPIC ParEGO
88

87
HV, %

86

85

84
0 10 20 30 40 50 60 70
Iterations

Fig. 7. Comparative performance on the DTLZ4 problem with 7 objectives. (Note:

here, HV is calculated using an approximation method not exactly.)
 11

5 Conclusions and Future Works

5.1 Conclusions

Multiobjective black-box optimization problems appear in many engineering ap-

plications whose function evaluations are very expensive with respect to time or
money and therefore are limited by available or affordable resources. To deal
with such problems, we propose a simple method called EPIC to approximate
the Pareto optimal set using a limited number of function evaluations. The key
idea is to use a classifier to classify decision vectors into two classes — nondom-
inated and dominated — and then predict the Pareto optimal set using only
information about the decision space and without evaluating the rest of sam-
pled decision vectors. The initial experimental results demonstrate that our new
method, EPIC, is competitive with existing methods PAL and ParEGO, and
even outperforms it on some problems. The results of all methods seem quite
dependent on the characteristics of the problem though, and this dependence is
worthy of further investigation to understand the strengths and weaknesses of
the different approaches.

5.2 Future works

Possible future research includes developing a strategy for selecting more than
one decision vector at each iteration for evaluation. It might consider some clus-
tering approach to choose more diverse vectors in the objective space in order to
ensure a better approximation of the Pareto optimal set in the sense of uniform
distribution of the vectors in the objective space. Also, we need to explore the
influence of the probability values pnext and pnond used to select and classify
decision vectors, and to investigate how these can be chosen in an automatic
way based on knowledge about the problem at hand.

References

1. D. R. Jones, M. Schonlau, W. J. Welch, Efficient global optimization of expensive

black-box functions, Journal of Global Optimization 13 (4) (1998) 455–492.
2. L. Santana-Quintero, A. Montao, C. C. Coello, A review of techniques for handling
expensive functions in evolutionary multi-objective optimization, in: Y. Tenne, C.-
K. Goh (Eds.), Computational Intelligence in Expensive Optimization Problems,
Vol. 2, Springer Berlin Heidelberg, 2010, pp. 29–59.
3. J. Sacks, W. J. Welch, T. J. Mitchell, H. P. Wynn, Design and analysis of computer
experiments, Statistical Science 4 (4) (1989) 409–423.
4. J. D. Martin, T. W. Simpson, Use of kriging models to approximate deterministic
computer models, AIAA journal 43 (4) (2005) 853–863.
5. G. E. Box, N. R. Draper, Empirical model-building and response surfaces, John
Wiley & Sons, 1987.
6. H. Fang, M. F. Horstemeyer, Global response approximation with radial basis
functions, Engineering Optimization 38 (4) (2006) 407–424.
12

7. A. I. Forrester, A. J. Keane, Recent advances in surrogate-based optimization,

Progress in Aerospace Sciences 45 (1–3) (2009) 50–79.
8. P. Lancaster, K. Salkauskas, Surfaces generated by moving least squares methods,
Mathematics of computation 37 (155) (1981) 141–158.
9. J. Knowles, Parego: A hybrid algorithm with on-line landscape approximation for
expensive multiobjective optimization problems, IEEE Transactions on Evolution-
ary Computation 10 (1) (2006) 50–66.
10. M. Zuluaga, A. Krause, G. Sergent, M. Püschel, Active learning for multi-objective
optimization, in: Proceedings of the 30th International Conference on Machine
Learning, 2013.
11. R. Jin, W. Chen, T. Simpson, Comparative studies of metamodelling techniques
under multiple modelling criteria, Structural and Multidisciplinary Optimization
23 (1) (2001) 1–13.
12. A. Chinchuluun, P. M. Pardalos, A. Migdalas, L. Pitsoulis, Pareto Optimality,
Game Theory and Equilibria, 2nd Edition, Springer, 2008.
13. E. Zitzler, L. Thiele, Multiobjective optimization using evolutionary algorithms –
a comparative case study, in: Parallel Problem Solving from Nature - PPSN-V,
Springer, 1998, pp. 292–301.
14. C. Azevedo, A. Araujo, Correlation between diversity and hypervolume in evolu-
tionary multiobjective optimization, in: IEEE Congress on Evolutionary Compu-
tation (CEC), 2011, pp. 2743–2750.
15. T. Okabe, Y. Jin, M. O. B. Sendhoff, On test functions for evolutionary multi-
objective optimization, in: X. Yao, E. Burke, J. Lozano, J. Smith, J. Merelo-Guervs,
J. Bullinaria, J. Rowe, P. Tio, A. Kabn, H.-P. Schwefel (Eds.), Parallel Problem
Solving from Nature – PPSN VIII, Vol. 3242, Springer Berlin Heidelberg, 2011,
pp. 792–802.
16. F. Kursawe, A variant of evolution strategies for vector optimization, in: H.-
P. Schwefel, R. Mnner (Eds.), Parallel Problem Solving from Nature, Vol. 496,
Springer Berlin Heidelberg, 1991, pp. 193–197.
17. E. Zitzler, K. Deb, L. Thiele, Comparison of multiobjective evolutionary algo-
rithms: Empirical results, Evolutionary Computation 8 (2) (2000) 173–195.
18. R. Viennet, C. Fonteix, I. Marc, New multicriteria optimization method based
on the use of a diploid genetic algorithm: Example of an industrial problem, in:
Selected Papers from the European conference on Artificial Evolution, Springer-
Verlag, London, UK,, 1996, pp. 120–127.
19. K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable multi-objective optimiza-
tion test problems, in: Congress on Evolutionary Computation (CEC 2002), IEEE
Press, 2002, pp. 825–830.
20. V. Vapnik, The Nature of Statistical Learning Theory, 2nd Edition, Springer, 1999.
21. K. P. Bennett, E. J. Bredensteiner, Duality and geometry in SVM classifiers, in:
Proceedings of 17th International Conference on Machine Learning, Morgan Kauf-
mann, 2000, pp. 57–64.
22. J. C. Platt, Probabilistic outputs for support vector machines and comparisons
to regularized likelihood methods, in: Advances in Large Margin Classifiers, MIT
Press, 1999, pp. 61–74.
23. B. Zadrozny, C. Elkan, Transforming classifier scores into accurate multiclass prob-
ability estimates, in: Proceedings of the International Conference on Knowledge
Discovery and Data Mining, 2002, pp. 694–699.
24. C.-C. Chang, C.-J. Lin, LIBSVM: A library for support vector machines, ACM
Transactions on Intelligent Systems and Technology 2 (2011) 27:1–27:27.