J Glob Optim manuscript No.

(will be inserted by the editor)

Dynamic Algorithm Selection for

Pareto Optimal Set Approximation

Ingrida Steponavičė · Rob J Hyndman ·

Kate Smith-Miles · Laura Villanova

Received: date / Accepted: date

Abstract This paper presents a meta-algorithm for approximating the Pareto optimal
set of costly black-box multiobjective optimization problems given a limited number
of objective function evaluations. The key idea is to switch among different algo-
rithms during the optimization search based on the predicted performance of each
algorithm at the time. Algorithm performance is modeled using a machine learning
technique based on the available information. The predicted best algorithm is then
selected to run for a limited number of evaluations. The proposed approach is tested
on several benchmark problems and the results are compared against those obtained
using any one of the candidate algorithms alone.
Keywords multiobjective optimization · expensive black-box function · machine
learning · classification · algorithm selection · hypervolume metric · features

1 Introduction

Many optimization problems involve multiple conflicting objectives, where there is

no single optimal solution optimizing all objective functions simultaneously, but a set

Ingrida Steponavičė
School of Mathematical Sciences, Monash University, Clayton, Australia
Tel.: +61 3 9905 8511
E-mail: [email protected]
Rob J Hyndman
Department of Econometrics & Business Statistics, Monash University, Clayton, Australia

Laura Villanova
School of Mathematical Sciences, Monash University, Clayton, Australia

Kate Smith-Miles
School of Mathematical Sciences, Monash University, Clayton, Australia
2 Ingrida Steponavičė et al.

of solutions representing the best possible trade-offs among the objectives. Therefore,
multiobjective optimization is a very important research area due to the multiobjec-
tive nature of most real-life problems, with many challenging issues to tackle.
The development of multiobjective optimization techniques has been an active
area of research for many years, resulting in a wide variety of approaches [4, 23, 24].
Besides the challenge caused by multiple objectives, practical problems arising in
engineering often require the solution of optimization problems where analytical ex-
pressions of the objective functions are unavailable and the evaluation of the objective
functions are very expensive. Such problems might involve computationally expen-
sive black-box simulation, or require costly experiments to be conducted in order to
obtain the objective function values. One simulation or experiment may take several
hours, days or even weeks. In addition to time restrictions, there can be other limita-
tions such as financial and physical constraints. Therefore, in order to keep the cost
affordable, it is important to find approximate solutions of the optimization problem
within a very restricted number of function evaluations (often only a few hundred
evaluations can be made).
Methods have been developed to solve expensive black-box optimization (BBO)
problems by building a surrogate model that approximates the objective function
and predicts promising new solutions at a smaller evaluation cost [14, 31]. One of
the state-of-art methods for expensive multiobjective optimization problems, named
ParEGO, was developed by Knowles [16]. It is essentially a multiobjective translation
of the efficient global optimization (EGO) method [14], where multiple objectives are
converted to a single objective using a scalarization function with different parameter
values at each step. The idea of modelling challenging functions by statistical mod-
els has a very long history, and was popularized for optimization problems in [25,
34]. Other EGO modifications to address costly multiobjective optimization prob-
lems are also available, including SMS-EGO [29], -EGO [37], MOEA/D-EGO [41],
and EGO-MO [7].
In addition to the EGO family of algorithms, we have previously proposed the
EPIC (Efficient Pareto Iterative Classification) algorithm [32]. In this approach, the
Pareto optimal set is identified by classifying regions of the decision space as likely to
be part of the Pareto set or not. A support-vector-machine (SVM) is applied in order
to capture nonlinear relationships between class labels and features (i.e., decision
variable values in this case). The advantage of this approach is that it does not depend
on the dimensionality of the objective space and so is suitable for high-dimensional
multiobjective problems.
Other approaches are also possible. It is an open question how to best select the
optimization algorithm for the particular problem of interest. Comparisons of various
methods for expensive multiobjective black-box optimization are prospering in the
literature, but they are usually limited and highlight the advantages of some proposed
modification to an existing method over its predecessor. There is a need for a deeper
analysis of the characteristics of the available algorithms and how well-suited they
are to specific problems.
To our knowledge, the methods developed so far each have some strengths and
weaknesses, and despite advances made in recent years, they are still far from being
able to solve a variety of real-life problems efficiently. Often, they are better suited to
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 3

some restricted problem classes. Moreover, for the same problem, some algorithms
can perform very well at the beginning and then lose their power, while other al-
gorithms can perform badly at the beginning but later demonstrate their superiority.
Such an example is demonstrated in Figure 1 which presents the percentage of times
when one of the considered algorithms (ParEGO, EPIC and a hybrid of EPIC and
Nelder-Mead) was outperforming others with respect to the hypervolume (HV) met-
ric (see Section 3.3) on a benchmark problem ZDT3 over 100 runs; the horizontal
axis represents the number of objective function evaluations (or iterations). This fig-
ure clearly shows that there is no single algorithm (at least among those considered)
performing better than the rest for all runs and at all points in time. For example, if
one can afford more than 70 evaluations, one should use ParEGO; in case of fewer
than 20 affordable evaluations, one should run EPIC-NM. As this figure suggests,
one might think that we can obtain good results by running EPIC-NM for the first 20
iterations, then EPIC for the next 50 iterations, and then ParEGO for the remaining
iterations. However, algorithm performance depends on the decision space already
explored. If we switch algorithms, then we would also have a different historical ex-
ploration of the decision space, so the performance may not match that presented in
Figure 1. Thus, there is no guarantee that the results obtained using this simple idea
will outperform the algorithms running separately.

ZDT3 problem

70

60
Percentage of times performing better, %

50

ParEGO
40 EPIC
EPIC−NM

30

20

10

0
0 20 40 60 80 100 120 140 160
Iterations

Fig. 1 Algorithms performance on ZDT3 problem

Therefore, we are interested in learning how to select the right algorithm at each
stage of the optimization process when very little is known about the multiobjective
optimization problem in advance. In particular, we focus on expensive black-box
problems where we wish to limit the number of function evaluations.
The paper has the following structure. Section 2 introduces the main concepts
involved in multiobjective optimization, and our proposed approach is described in
Section 3. In Section 4, we outline our experimental setup and the selected algo-
rithms, and present and analyse the results we have obtained on some test problems.
4 Ingrida Steponavičė et al.

Section 5 draws some conclusions and briefly discusses some future research direc-
tions.
This is an initial exploration of an approach to this problem, describing how it
can be implemented, and highlighting and discussing the results obtained on a small
number of optimization problems. Much larger computational experiments involving
many more optimization problems would be required in order to draw general con-
clusions, and validate the proposed approach. This would take a vast amount of time
and so is left for future research.

2 Expensive Multiobjective Optimization Problems

The problems considered here are formulated as follows:

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

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

black-box functions that are to be minimized simultaneously. All objective functions
are represented by the vector-valued function f : S → Rm . A vector x ∈ S is called
a decision vector and a vector z = f (x) ∈ Rm an objective vector.
In multiobjective optimization, the objective functions f1 , . . . , fm in (1) are typ-
ically conflicting. In that case, there does not exist a decision vector x̄ ∈ S such that
x̄ minimizes fi in S for all i = 1, . . . , m, 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 non-dominated if there does not
exist a decision vector x ∈ S such that fi (x) ≤ fi (x̄) for all i = 1, . . . , m and
fj (x) < fj (x̄) for some j = 1, . . . , m. If such a decision x ∈ S does exist, x̄ and z̄
are said to be dominated by x and its image z = f (x), respectively.

3 Dynamic Algorithm Selection

3.1 General Framework

There are many different algorithms that perform well on some problem classes and
struggle on others, and it is difficult to predict the accurate performance of an al-
gorithm on a particular problem. In practice, the number of function evaluations
required by the candidate algorithm to solve some particular problem can be vast.
Having a particular problem to be solved, especially in a limited number of func-
tion evaluations, one must select an algorithm without being sure of making the most
appropriate choice. Bad decisions may lead to an unacceptable number of function
evaluations and poor approximation of the true Pareto set. Algorithm selection is a
learning problem where we use a model to predict the expected performance of each
algorithm on a given problem; the model is trained on a set of performance data for a
number of problems [30]. For each new problem, the model is used to select the algo-
rithm that is expected to give the best results. In addition to static approaches where
the selection is performed before running the algorithms, there have been proposed
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 5

a number of dynamic algorithm selection approaches where the selection process is

adapted during the actual execution of the algorithms (e.g., see [2, 21, 15]).

Here, we suggest switching among different algorithms during the search, based
on the information collected in the objective and decision spaces. For this purpose, we
use a model that predicts which algorithm will perform the best in a given situation
according to a selected performance metric. In multiobjective optimization, algorithm
performance can be assessed taking into account different qualities of the estimated
Pareto optimal set such as spread, convergence, distribution, etc. The choice of met-
rics to use in evaluating algorithm performance is somewhat subjective.

The basic idea of dynamic algorithm selection tends to circumvent the following
challenges which are associated with expensive multiobjective black-box optimiza-
tion problems: (i) selecting the ‘right’ algorithm to solve the problem with very little
(or no) knowledge about it; and (ii) obtaining a high quality approximation of the
Pareto optimal set within a limited number of evaluations.

Algorithm performance prediction can be modelled as a classification problem

in which each new instance is assigned to the algorithm class that would lead to
the greatest improvement in the performance metric, based on models trained on
instances whose class membership is known. Therefore, the proposed approach can
be divided into two stages.

In Stage A, we collect the training data and use them to build the performance
prediction model(s). That is, a classification algorithm is used to learn the relation-
ship between the descriptive metrics of a current situation and subsequent algorithm
performance over the next few evaluations. This is based on a large dataset where all
considered algorithms have been applied to a large number of problems at various
points in time and their performance has been monitored.

Stage B employs the prediction model to approximate the Pareto optimal set and
can be decomposed into the following steps:

Step B1 Generate an initial set in the decision space and evaluate the objective func-
tions;
Step B2 Given some evaluated vectors, calculate descriptive metrics;
Step B3 Ask the prediction model to predict the best algorithm based on the calcu-
lated metrics;
Step B4 Run the suggested algorithm for a limited number of evaluations;
Step B5 Stop if the maximum number of function evaluations is reached. Otherwise,
go to B2.

Stage B is represented in Figure 2. The most important elements are the prediction
model and descriptive metrics which at the very beginning are calculated from the
initial set of evaluated solutions. After running a suggested algorithm for a small
number of iterations, the solution set is updated. Metrics are updated and the same
steps are repeated until the maximum number of evaluations is reached.
6 Ingrida Steponavičė et al.

Initial
Set Which algorithm
should I run?

Suggested
Descriptive Prediction
Metrics Model Algorithm

Add new solutions

If # sol < N, recalculate Solution

Set

If # sol ≥ N

Finish

Fig. 2 Stage B representation

3.2 Performance prediction model

The purpose of a performance prediction model is to help select an algorithm with

the best performance in a given situation. Therefore, it has to give an answer to the
question “Which algorithm will perform best in the current situation: X, Y or Z?”
In Stage A, a set of instances with features and the class to which each of the
instances belongs (i.e., the name of the algorithm which give the best performance
for each instance) is given to the machine learning algorithm (e.g., a random forest or
support vector machine) to learn the relationship between features and class labels.
The learning algorithm utilizes the label information as well as the data itself to learn
a mapping function g (or a classifier) from features to class labels: g(f eatures) →
labels. To build a performance prediction model we need to have metrics describing
different situations used as features associated with a class label, and the name of the
algorithm whose performance was the best under given circumstances.
In Stage B, instances are represented by the same features as used in training and
then the performance prediction model assigns labels; i.e., it predicts which of the
algorithms will perform best in each instance.
Building a reliable performance prediction model that is applicable to various
expensive multiobjective optimization problems requires as much data as possible
resulting in a time-consuming process of running algorithms on a large and diverse
set of test problems. However, collecting data is a very important step to ensure an
efficient performance of the proposed approach. It is also very important to select the
right features or descriptive metrics; we discuss these in the following subsection.

3.3 Descriptive metrics

To build a classifier, we need to have knowledge of what features make good predic-
tors of class membership for the algorithms we are considering. In real-world situa-
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 7

tions, we often have little knowledge about relevant features. Therefore, we have to
find a set of features that separates classes as cleanly as possible.
It is clear that the selected features must have some relationship with the perfor-
mance of the algorithms. Usually, in the published comparisons of different multi-
objective optimization algorithms, we can see only numerical experiments and dis-
cussion about the algorithm performance without deeper analysis about why some
algorithms work well on the problems, why some algorithms have some difficulties,
and what are the characteristics that make them succeed or fail.
Intuitively, the metrics characterizing the observed Pareto set and the progress
made in searching for non-dominated solutions should be important. There exist
many metrics used to assess the quality of the obtained solution set in multiobjective
optimization that can be categorized into cardinality (or capacity), convergence, di-
versity and hybrid measures [12]. Most of these metrics were developed to compare
an obtained solution set with the true Pareto optimal set, and include generational
distance [5], inverted generational distance [35], -indicator [45] and hypervolume
difference [38] among others. However, these are not suitable for our purpose as in
practice the true Pareto set is not known a priori. Hence, this significantly reduces our
choice.
We now describe the quality metrics we have considered.

Ratio of non-dominated solutions. This metric is a capacity measure quantifying

the ratio of non-dominated solutions in the obtained solution set:

|P |
RON(S, P ) = , (2)
|S|

where |S| is number of the solutions in the observed solution set and |P | is the
number of non-dominated solutions in the observed Pareto set P .
Generalized Spread metric [6]. This diversity metric indicates the distribution of
solutions in the observed Pareto set P :
Pm P ¯
∗ i=1 d(ei , P ) + X∈P |d(X, P ) − d|
∆ (P, T ) = Pm ¯ , (3)
i=1 d(ei , P ) + |P | ∗ d

where (e1 , . . . , em ) are m extreme solutions in T , the true Pareto optimal set, and

d(X, P ) = min ||f (X) − f (Y )||2 ,

Y ∈P,Y 6=X

1 X
d¯ = d(X, P ).
|P |
X∈P

Smaller values are preferable. This metric requires knowledge of extreme val-
ues of T . When solving real-word problems, this information is not available
beforehand. Therefore, one can use some estimates of the extreme values of the
objective space.
8 Ingrida Steponavičė et al.

Number of distinct choices [38]. This metric divides the objective space into a grid
of (1/µ)m m-dimensional hypercubes (µ ∈ [0, 1]) and calculates the number of
hypercubes containing solutions; i.e., it indicates the number of distinct solutions
that exists in an observed Pareto solution set P :
ν−1
X ν−1
X ν−1
X
NDCµ (P ) = ··· Nµ (q, P ) (4)
`m =0 `2 =0 `1 =0

where q = (q1 , q2 , . . . , qm ) with qi = `νi , ν = µ1 and Nµ (q, S) is equal to 1

provided there is at least one non-dominated solution falling into the hypercube
Tµ (q); otherwise, it is equal to 0. Solutions falling into the same hypercube are
considered similar to one another. The idea is to count only those generated non-
dominated points that are sufficiently distinct from each other. The higher value
is preferred over the smaller for diversity.
Dispersion. This metric was introduced in [38] where it was named ‘cluster’; it de-
scribes the average size of clusters composed of similar non-dominated solutions
in P , and is calculated as follows:

|P |
CLµ (P ) = , (5)
NDCµ (P )

In the ideal case where every non-dominated solution obtained is distinct, then
the value of metric CLµ (P ) is equal to 1. Also, the higher the value of metric
CLµ (P ), the more clustered the non-dominated solution set P , and hence the
less preferred. In our opinion, the term ‘cluster’ is misleading, therefore, we have
named it ‘dispersion’.
Correct classification. This is the percentage of correctly classified non-dominated
and dominated solutions obtained by a support vector machine (SVM). This met-
ric was selected because some of the considered algorithms use an SVM. Thus,
the quality of an SVM at a given point in the search is a useful descriptor of the
current situation and how an algorithm relying on SVM modelling is likely to
perform.
Correct classification of non-dominated class. This is the percentage of correctly
classified non-dominated solutions. As the non-dominated class is usually (sig-
nificantly) smaller than the dominated one, total correct classification may still be
good while all examples from non-dominated class can be misclassified.
Correct classification of dominated class. This is the percentage of dominated so-
lutions correctly classified by an SVM.
Hypervolume metric [44]. This metric has attracted a lot of interest in recent years
as it describes both the convergence towards the Pareto optimal set and the dis-
tribution along it. Basically it calculates the volume covered by non-dominated
solutions (see Figure 3). Mathematically, for each solution i ∈ P , a hypercube vi
is constructed with a reference point W and the solution i as the diagonal corners
of the hypercube. The reference point can simply be obtained by composing a
vector of the worst objective function values. Then, a union of all hypercubes is
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 9

found and its hypervolume is calculated:

|P |
[
HV(P, W ) = volume( vi ). (6)
i=1

One of the main reasons for the popularity of HV is that it not only reflects domi-
nance, but also promotes diverse sets. Moreover, it is the only indicator known to
be strictly monotonic with respect to Pareto dominance and thereby guaranteeing
that the Pareto optimal set achieves the maximum hypervolume possible, while
any worse set will be assigned a worse indicator value [1].

Fig. 3 HV metric calculation [12]

Despite the attractive features of HV, it has a few major issues. First, it is
computationally intensive, especially for high dimensional problems. Second, the
metric varies with the choice of the reference point [42]. Finally, if the scales of
the objective functions are very different, it can be biased in favour of objectives
with a larger scale. To eliminate the bias of different scales, it is suggested [4] to
calculate the HV metric using normalized objective function values.
We use the HV metric for two purposes: first, as one of the features to char-
acterize the current situation; and second, to assess algorithm performance or
superiority to derive class membership.
All the descriptive metrics that depend on the scale of the objective functions
should be calculated using normalized (scaled) objective function values in order to
eliminate any bias in the metric values. Therefore, we estimated the extreme values
of the objective space and used this information for normalization.

4 Experimental Analysis

4.1 Selected algorithms

To test the proposed approach, we switch among three algorithms: ParEGO, EPIC,
and Nelder-Mead (NM). The performance of our dynamic switching algorithm was
10 Ingrida Steponavičė et al.

compared with ParEGO, EPIC and EPIC-NM algorithms running separately. They
are shortly discussed below.

ParEGO This method employs a Gaussian process (GP) model to predict objective
function values. It converts the multiobjective optimization problem into a single ob-
jective problem using the augmented Tchebycheff function:

m
X


fλ (x) = max λj fj (x) ± ρ λj fj (x), (7)
j=1,...,m
j=1

where ρ > 0 is a small positive number and λ is a weight vector. At each itera-
tion of the algorithm, a different weight vector is drawn uniformly at random from
the set of evenly distributed vectors allowing the model to gradually build up an ap-
proximation to the true Pareto set. Before scalarization, the objective functions are
normalized with respect to the known (or estimated) limits of the objective space to
the range [0, 1]. At each iteration, the method uses a genetic algorithm to search for
the solutions that maximizes the expected improvement criterion with respect to a
surrogate model. After evaluation of the selected solution on the real expensive func-
tion, ParEGO updates the GP surrogate model of the landscape and repeats the same
steps.
The main disadvantage of employing a GP is that model construction can be a
very time-consuming process [13], where the time increases with the number of eval-
uated vectors used to model the GP. To overcome this issue, when the iteration num-
ber is greater or equal to 25, 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 increasingly problematic in high dimensional spaces
[8], so these methods do not scale well as the dimension of the problem increases.

EPIC The EPIC algorithm approximates the Pareto optimal set with a limited num-
ber of objective function evaluations. Its main idea is to learn about the evaluated
non-dominated and dominated vectors in the decision space and to predict which un-
evaluated vectors are likely to be non-dominated, thus gradually building an approx-
imation of the Pareto optimal set by evaluating the most promising decision vectors.
A discussion of how to select vectors for evaluation can be found in [32].
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. Moreover, it is
simple to implement, has no limitations on high dimensional problems, and multiple
decision vectors can be selected at each iteration [32]. However, its weakness is that
it does not generate new decision vectors but rather selects a decision vector from a
given set representing the decision space, whose quality has an impact on the method
performance. To overcome this issue, the EPIC method can be modified by introduc-
ing a mechanism for generating new decision vectors in the most promising areas of
the decision space.
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 11

Nelder-Mead algorithm for multiobjective optimization The NM method was devel-

oped for single objective optimization problems as a very efficient derivative-free
local search procedure [26]. A disadvantage of NM is that its convergence is sensi-
tive to the selected starting point and can fail because the search direction becomes
increasingly orthogonal to the direction of steepest descent [33]. Also, it can become
inefficient for large dimensional problems [11, 28]. Despite lacking a satisfactory con-
vergence theory, the NM method generally works well on small dimensional real-life
problems and remains one of the most popular direct search methods [20, 18]. Re-
cently, numerous improvements to the Nelder–Mead simplex algorithm have been
posed to address these issues [22, 39, 28, 9]. For multiobjective optimization prob-
lems, several NM modifications or hybridization have been proposed as well [17,
40].
To apply NM for multiobjective optimization, we scalarize a multiobjective prob-
lem to a single objective by using the augmented Tchebycheff function (7). An initial
simplex is constructed from the n + 1 vertices having the best scalarized objective
function values. Then the main simplex transformation operations (reflexion, expan-
sion, contraction and shrinking) are performed as in the original algorithm for single
objective problems. At every iteration a new starting simplex is selected using a dif-
ferent weighting vector to encourage exploration of the Pareto front.

EPIC-NM algorithm We have modified the EPIC algorithm to enable generation of

the decision vectors by hybridizing it with NM for a local search in the promising
areas. Here, NM is called when there is no significant improvement measured by the
HV metric using the EPIC algorithm and run for a limited number of evaluations
to look for improvement locally. Different weighting vectors are used to convert the
original problem to a scalarized one for the NM algorithm. The starting simplex is
composed of the best vectors evaluated so far in an appropriate scalarized objective
space. Thereafter, NM continues with regular simplex iterations for a few evaluations.
Then, the search is taken over by EPIC fed with an updated set of evaluated solutions.
For switching, EPIC-NM does not use model-based decisions but a rule determined
by the metric of the current state.

4.2 Experimental setup

All algorithms were implemented in Matlab and their parameters were set to de-
fault values. In particular, our ParEGO implementation was based on the C code by
Knowles, which can be downloaded from www.cs.bham.ac.uk/~jdk/parego/;
this implementation corresponds to the algorithm described in [16]. The default val-
ues for the ParEGO implementation were used, namely (i) population size equal to
20, (ii) number of restarts when optimising the likelihood function is equal to 30, and
(iii) crossover is equal to 0.2. The implementation of EPIC is described in [32]. In
both EPIC and EPIC-NM, we used SVM with a radial basis function kernel; SVM
kernel parameters were obtained through cross-validation performed at each iteration.
In EPIC-NM, an initial simplex was composed of the vertices having the best scalar-
ized problem values. A local search was called after EPIC could not make progress
12 Ingrida Steponavičė et al.

(i.e., no change occurred in HV metric values for the last four iterations), and run for
five evaluations. All algorithms started with the same initial set consisting of 11n − 1
decision vectors, where n is the dimension of the decision space, as suggested in
[14]. The Latin hypercube technique was used to sample the decision space. In addi-
tion, for EPIC and EPIC-NM, we sampled a design space representation consisting
of 500 vectors although the objective function values for these points were not eval-
uated unless selected by an algorithm. The maximum number of evaluations was
restricted to 200 including the initial sampling. The algorithms were run 100 times
with different initial sets (as their performance is influenced by the initial set), and the
average values of the HV metric were calculated. The performance of the algorithms
was measured at every iteration to assess the progress obtained after each objective
function evaluation.

4.3 Test problems

Our training set consisted of more than 12000 instances (snapshots in time) from
solving the following four benchmark problems: ZDT3 [43], OKA2 [27], Kursawe
[19] and Viennet [36]. These presented different challenges for approximating the
true Pareto optimal set.

ZDT3. This problem has two objective functions and three decision variables. The
Pareto optimal set comprises several discontinuous convex parts in the objective
space.
Kursawe. This problem has two objective functions and a scalable number of deci-
sion variables. In our experiment, three decision variables were used. Its Pareto
optimal set is disconnected and symmetric in the decision space, and disconnected
and concave in the objective space.
OKA2. This problem has two objective functions and three decision variables. Its
true Pareto optimal set is a spiral shaped curve in the objective space, and the
density of the Pareto optimal solutions in the objective space is low.
Viennet. This problem consists of three objective functions and two decision vari-
ables. Its true Pareto optimal set is convex in the objective space.

4.4 Building the performance prediction model

There are many classification methods available including linear classifiers, support
vector machines, decision trees and neural networks. To model our algorithm perfor-
mance we have used a random forest [3], which is an ensemble of randomly trained
decision trees. Algorithm performance was assessed with the HV metric calculated
using normalized objective function values. Each instance was associated with the
name of the algorithm that had the largest value of the HV metric. The distribution
of instances among the three classes was as follows: 4599 instances in the ParEGO
class, 5890 in the EPIC class and 2930 in the NM class. Hence, the largest class
consists of instances where EPIC was best, while NM’s was the smallest.
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 13

Table 1 Confusion matrix of algorithm performance model

predicted EPIC predicted NM predicted ParEGO

actual EPIC 5014 230 646
actual NM 374 1618 938
actual ParEGO 1108 715 2776

When building a model for a classification problem, it is always important to

look at the classification accuracy; i.e., the number of correct predictions from all
predictions made. Out-of-bag accuracy of the trained model was 70.11%. A clear
way to present the prediction results of a classifier is to use a confusion matrix, which
shows how the predictions are made by the model. The diagonal elements show the
number of correct classifications made for each class, and the off-diagonal elements
show the errors made. The confusion matrix of the prediction model is presented in
Table 1. It shows that the model tends to predict the EPIC algorithm will be best more
often and misclassifies more instances from the NM and ParEGO classes.
The feature space of the prediction model gives some insight about the values of
the metrics which are favorable for each algorithm. For example, Figure 4 shows that
the EPIC algorithm was best in more than 90% of instances where the HV values were
between 0.6 and 0.65. When the ratio of non-dominated solutions is small, ParEGO
or NM are preferred algorithms as seen in Figure 5. This can be simply explained by
the fact that the SVM used in the EPIC algorithm has limited information about the
non-dominated solution class when there are few non-dominated solutions, resulting
in poor EPIC performance. Figure 6 shows that EPIC prefers higher NDC values over
smaller ones; i.e., it needs more distinct evaluated solutions to work well.
The random forest model we used to select the preferred algorithm revealed that
the most important feature for classification prediction is the hypervolume metric fol-
lowed by generalized spread, dispersion, correct classification of dominated points,
overall classification accuracy and the ratio of nondominated solutions. On the other
hand, the least important features for algorithm prediction are the number of distinct
choices and the correct classification of nondominated points.
The accuracy of the prediction model can be affected by several factors. First, a
model may be based on a training set where classes were unbalanced. Therefore, the
model assigned more instances to the majority class and made more errors for smaller
classes. This can be overcome by collecting more examples for smaller classes, or
by using different accuracy measures such as precision and recall. Second, selected
descriptive metrics may not completely describe the situation at that time. Looking
for more and/or better descriptive metrics can be a solution.
A further consideration is where one or more algorithms may be equally good, or
approximately the same, in their performance. In that case, finding the best algorithm
is not as important as avoiding a poor algorithm. For that reason, misclassifications
will not necessarily lead to poor performance in the main task of estimating the Pareto
front. Table 2 shows that our prediction model suggested the best algorithm in 70.11%
of all cases, while the worst one in only 9.78% of all cases. Thus, our proposed
approach can minimize the risk of selecting the poorest algorithm.
14 Ingrida Steponavičė et al.

HV
100
ParEGO
90 EPIC
NM

80
Percent when an algorithm is the best

70

60

50

40

30

20

10

0
0.4 0.5 0.6 0.7 0.8 0.9 1
Metric value

Fig. 4 Percentage distribution of HV values when each algorithm was the best
Ratio of Nondominated
50
ParEGO
45 EPIC
NM

40
Percent when an algorithm is the best

35

30

25

20

15

10

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Metric value

Fig. 5 Percentage distribution of RoN values when each algorithm was the best
NDC
40
ParEGO
EPIC
35 NM

30
Percent when an algorithm is the best

25

20

15

10

0
0 2 4 6 8 10 12 14
Metric value

Fig. 6 Percentage distribution of NDC values when each algorithm was the best
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 15

Table 2 Distribution of suggested algorithms by the performance model

Best algorithm Second best algorithm Worst algorithm

70.11% 20.11% 9.78%

4.5 Numerical results

A number of computational experiments were carried out to test the ability of the pro-
posed approach to approximate the Pareto optimal set with a limited number of eval-
uations. The performance of our proposed dynamic switching algorithm is demon-
strated in Figure 7. Here, it dynamically switches among three algorithms at every
fifth evaluation using model predictions to decide which algorithm to use and these
moments are marked by dots. Also, its performance is compared with the perfor-
mance of three algorithms run separately.

ZDT3 run #13

0.67

0.66

0.65

0.64
HV

0.63

0.62
EPIC
EPIC−NM
0.61
ParEGO
Our approach
0.6
call ParEGO
call NM
0.59 call EPIC
0 20 40 60 80 100 120 140 160
Number of evaluations

Fig. 7 Performance on a single run of ZDT3 problem

The comparison, based on the average HV metric over 100 runs and calculated
using normalized and original values of ZDT3 and Kursawe, is presented in Fig-
ures 8–13. The figures depict the average HV measured after the initial sampling
(i.e., starting from the 11nth function evaluation). The initial sampling does not pro-
vide relevant information for algorithm comparisons because, for each of the 100
runs, all the algorithms have been evaluated on the same initial sample. Results simi-
lar to those reported in Figures 8–13 were obtained for OKA2 and Viennet problems.
They also show that the proposed approach is competitive. It can be noted that for the
ZDT3 problem, algorithm superiority depends on how the HV metric is calculated.
For example, Figure 10 demonstrates that the proposed approach is the most efficient
with respect to the HV metric calculated using the original scale while Figure 8 does
not provide a clear winner.
This raises the question of which metric should be used to judge the algorithm
performance. If the objectives have different scales and we aim to find a uniformly
16 Ingrida Steponavičė et al.

HV of ZDT3 problem
0.66

0.65

0.64
avg normalized HV

0.63

0.62

0.61

0.6
EPIC
0.59 EPIC−NM
ParEGO
Our approach
0.58
0 20 40 60 80 100 120 140 160
iterations

Fig. 8 ZDT3: average HV (normalized scale)

Variability of HV obtained by different
EPIC algorithms EPIC−NM
0.68 0.68

0.66 0.66
Avg. normalized HV

Avg. normalized HV

0.64 0.64

0.62 0.62

0.6 0.6

0.58 0.58

50 100 150 50 100 150

Evaluations Evaluations
ParEGO Our approach
0.68 0.68

0.66 0.66
Avg. normalized HV

Avg. normalized HV

0.64 0.64

0.62 0.62

0.6 0.6

0.58 0.58

50 100 150 50 100 150

Evaluations Evaluations

Fig. 9 ZDT3: average and standard deviations of HV (normalized scale)

Original HV of ZDT3 problem

5.95

5.9

5.85
avg original HV

5.8

5.75

5.7
EPIC
5.65 EPIC−NM
ParEGO
Our approach
5.6
0 20 40 60 80 100 120 140 160
iterations

Fig. 10 ZDT3: average HV (original scale)

Dynamic Algorithm Selection for Pareto Optimal Set Approximation 17

HV of Kursawe problem

0.9

0.85

0.8
avg normalized HV

0.75

0.7

0.65
EPIC
EPIC−NM
0.6
ParEGO
Our approach
0.55
0 20 40 60 80 100 120 140 160
iterations

Fig. 11 Kursawe: average HV (normalized scale)

Variability of HV obtained by different
EPIC algorithms EPIC−NM

0.9 0.9
Avg. normalized HV

Avg. normalized HV

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Evaluations Evaluations
ParEGO Our approach

0.9 0.9
Avg. normalized HV

Avg. normalized HV

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160
Evaluations Evaluations

Fig. 12 Kursawe: average and standard deviations of HV (normalized scale)

Original HV of Kursawe problem
540

520

500

480
avg original HV

460

440

420

400

380 EPIC
EPIC−NM
360 ParEGO
Our approach
340
0 20 40 60 80 100 120 140 160
iterations

Fig. 13 Kursawe: average HV (original scale)

18 Ingrida Steponavičė et al.

distributed representation of the Pareto optimal set, HV should be calculated using a

normalized objective function axis. However, if the priority is given to the objective
function with the larger scale, we might be willing to use the original scales in order
to have favorable bias to the larger objective.
Moreover, Figures 9 and 12 demonstrate the standard deviation of HV as a shaded
area. It can be seen that for the ZDT3 problem, the EPIC algorithm and the proposed
approach do not vary significantly which means that for all runs their HV metric
values converge to the average. The HV variance for the Kursawe problem is a little
larger indicating slower convergence.
In summary, taking into account that there is nothing or very little known about
the real-world optimization problem at hand, the proposed approach can be very
promising in generating the approximation of the true Pareto set.

5 Conclusions and Future Work

5.1 Conclusions

Summarizing, we have introduced an approach to approximating the Pareto optimal

set for expensive black-box problems by switching among different algorithms. In
the proposed approach, the algorithms are selected based on the prediction of their
performance using the information available at that time. A classification technique
is used to build a performance prediction model that is used to predict which of the
algorithms is likely to outperform the others in the current situation. The initial re-
sults of our proposed dynamic switching approach are very encouraging and deserve
further investigation.
Although the proposed approach is applicable to any costly multiobjective black-
box optimization problem, its performance depends on the prediction model which
is strongly affected by the training set data. Therefore, to get a reliable prediction
model, we have to perform extensive calculations where all the algorithms are pro-
vided with the same initial conditions. This is a very time consuming process, as each
algorithm has to be tested with many situations: problems with different characteris-
tics, an approximated optimal Pareto set at different moments in time, etc.
However, if one works in a specific application area with a well-defined set of
similar optimization problems, it can be beneficial to train our approach only for that
class of problems.

5.2 Discussion on future research directions

We plan to build a performance model with a higher classification accuracy by col-

lecting more data over a larger set of test problems. Future work includes searching
for new descriptive metrics that better characterize the situation at each instance. One
such metric is fitness distance correlation [10], a landscape metric for multiobjec-
tive optimization providing useful information concerning the relative difficulty of
moving “along” the Pareto optimal set in the objective space. Moreover, we plan to
Dynamic Algorithm Selection for Pareto Optimal Set Approximation 19

investigate the strengths and weaknesses of the different algorithms as optimization

results by different algorithms seem to depend on the characteristics of the problem.
Finally, we will augment the set of algorithms to include the most diverse set of state-
of-the-art algorithms devoted to multiobjective black-box optimization problems.

References

1. Bader, J., Zitzler, E.: Hype: An algorithm for fast hypervolume-based many-objective optimization.
Evolutionary computation 19(1), 45–76 (2011)
2. Borrett, J.E., Tsang, E.P.: Adaptive constraint satisfaction: the quickest first principle. In: Computa-
tional Intelligence, pp. 203–230. Springer (2009)
3. Breiman, L.: Random forests. Machine learning 45(1), 5–32 (2001)
4. Deb, K.: Multi-objective optimization using evolutionary algorithms, vol. 16. John Wiley & Sons
(2001)
5. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm:
Nsga-ii. Evolutionary Computation, IEEE Transactions on 6(2), 182–197 (2002)
6. Durillo, J.J., Nebro, A.J.: jmetal: A java framework for multi-objective optimization. Advances in
Engineering Software 42(10), 760–771 (2011)
7. Feng, Z., Zhang, Q., Zhang, Q., Tang, Q., Yang, T., Ma, Y.: A multiobjective optimization based
framework to balance the global exploration and local exploitation in expensive optimization. Journal
of Global Optimization pp. 1–18 (2014)
8. Forrester, A.I., Keane, A.J.: Recent advances in surrogate-based optimization. Progress in Aerospace
Sciences 45(1–3), 50–79 (2009)
9. Gao, F., Han, L.: Implementing the nelder-mead simplex algorithm with adaptive parameters. Com-
putational Optimization and Applications 51(1), 259–277 (2012)
10. Garrett, D., Dasgupta, D.: Multiobjective landscape analysis and the generalized assignment problem.
In: Learning and Intelligent Optimization, pp. 110–124. Springer (2008)
11. Han, L., Neumann, M.: Effect of dimensionality on the nelder–mead simplex method. Optimization
Methods and Software 21(1), 1–16 (2006)
12. Jiang, S., Ong, Y.S., Zhang, J., Feng, L.: Consistencies and contradictions of performance metrics in
multiobjective optimization. Cybernetics, IEEE Transactions on 44(12), 2391–2404 (2014)
13. Jin, R., Chen, W., Simpson, T.: Comparative studies of metamodelling techniques under multiple
modelling criteria. Structural and Multidisciplinary Optimization 23(1), 1–13 (2001)
14. Jones, D.R., Schonlau, M., Welch, W.J.: Efficient global optimization of expensive black-box func-
tions. Journal of Global Optimization 13(4), 455–492 (1998)
15. Kaelbling, L.P., Littman, M.L., Moore, A.W.: Reinforcement learning: A survey. Journal of artificial
intelligence research pp. 237–285 (1996)
16. Knowles, J.: Parego: A hybrid algorithm with on-line landscape approximation for expensive mul-
tiobjective optimization problems. IEEE Transactions on Evolutionary Computation 10(1), 50–66
(2006)
17. Koduru, P., Dong, Z., Das, S., Welch, S.M., Roe, J.L., Charbit, E.: A multiobjective evolutionary-
simplex hybrid approach for the optimization of differential equation models of gene networks. Evo-
lutionary Computation, IEEE Transactions on 12(5), 572–590 (2008)
18. Kolda, T.G., Lewis, R.M., Torczon, V.: Optimization by direct search: New perspectives on some
classical and modern methods. SIAM review 45(3), 385–482 (2003)
19. Kursawe, F.: A variant of evolution strategies for vector optimization. In: H.P. Schwefel, R. Mn-
ner (eds.) Parallel Problem Solving from Nature, vol. 496, pp. 193–197. Springer Berlin Heidelberg
(1991)
20. Lagarias, J.C., Reeds, J.A., Wright, M.H., Wright, P.E.: Convergence properties of the nelder–mead
simplex method in low dimensions. SIAM Journal on optimization 9(1), 112–147 (1998)
21. Lagoudakis, M.G., Littman, M.L.: Algorithm selection using reinforcement learning. In: ICML, pp.
511–518. Citeseer (2000)
22. Luersen, M.A., Le Riche, R.: Globalized nelder–mead method for engineering optimization. Com-
puters & structures 82(23), 2251–2260 (2004)
23. Marler, R.T., Arora, J.S.: Survey of multi-objective optimization methods for engineering. Structural
and multidisciplinary optimization 26(6), 369–395 (2004)
20 Ingrida Steponavičė et al.

24. Miettinen, K.: Nonlinear multiobjective optimization, vol. 12. Springer Science & Business Media
(1999)
25. Mockus, J.: Bayesian Approach to Global Optimization. Kluwer Academic Publishers, Dordrecht
(1989)
26. Nelder, J.A., Mead, R.: A simplex method for function minimization. The computer journal 7(4),
308–313 (1965)
27. Okabe, T., Jin, Y., Sendhoff, M.O.B.: 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, pp. 792–802.
Springer Berlin Heidelberg (2011)
28. Pham, N., Wilamowski, B.M.: Improved nelder mead’s simplex method and applications. Journal of
Computing 3(3), 55–63 (2011)
29. Ponweiser, W., Wagner, T., Biermann, D., Vincze, M.: Multiobjective optimization on a limited budget
of evaluations using model-assisted S-metric selection. In: G. Rudolph, T. Jansen, N. Beume, S. Lu-
cas, C. Poloni (eds.) Parallel Problem Solving from Nature – PPSN X, Lecture Notes in Computer
Science, vol. 5199, pp. 784–794. Springer Berlin Heidelberg (2008)
30. Rice, J.R.: The algorithm selection problem (1975)
31. Santana-Quintero, L., Montaño, A., Coello, C.C.: A review of techniques for handling expensive func-
tions in evolutionary multi-objective optimization. In: Y. Tenne, C.K. Goh (eds.) Computational In-
telligence in Expensive Optimization Problems, vol. 2, pp. 29–59. Springer Berlin Heidelberg (2010)
32. Steponavičė, I., Hyndman, R.J., Smith-Miles, K., Villanova, L.: Efficient identification of the pareto
optimal set. In: Learning and Intelligent Optimization, pp. 341–352. Springer International Publishing
(2014)
33. Torczon, V.J.: Multi-directional search: a direct search algorithm for parallel machines. Ph.D. thesis,
Citeseer (1989)
34. Törn, A., Žilinskas, A.: Global Optimization, Lecture Notes in Computer Science, vol. 350 (1989)
35. Van Veldhuizen, D.A., Lamont, G.B.: Multiobjective evolutionary algorithm test suites. In: Proceed-
ings of the 1999 ACM symposium on Applied computing, pp. 351–357. ACM (1999)
36. Viennet, R., Fonteix, C., Marc, I.: New multicriteria optimization method based on the use of a diploid
genetic algorithm: Example of an industrial problem. In: Selected Papers from the European confer-
ence on Artificial Evolution, pp. 120–127. Springer-Verlag, London, UK, (1996)
37. Wagner, T.: Planning and Multi-objective Optimization of Manufacturing Processes by Means of
Empirical Surrogate Models. Vulkan (2013)
38. Wu, J., Azarm, S.: Metrics for quality assessment of a multiobjective design optimization solution set.
Journal of Mechanical Design 123(1), 18–25 (2001)
39. Zahara, E., Kao, Y.T.: Hybrid nelder–mead simplex search and particle swarm optimization for con-
strained engineering design problems. Expert Systems with Applications 36(2), 3880–3886 (2009)
40. Zapotecas-Martínez, S., Coello, C.A.C.: Monss: A multi-objective nonlinear simplex search approach.
Engineering Optimization (ahead-of-print), 1–23 (2015)
41. Zhang, Q., Liu, W., Tsang, E., Virginas, B.: Expensive multiobjective optimization by moea/d with
gaussian process model. Evolutionary Computation, IEEE Transactions on 14(3), 456–474 (2010)
42. Zitzler, E., Brockhoff, D., Thiele, L.: The hypervolume indicator revisited: On the design of pareto-
compliant indicators via weighted integration. In: Evolutionary multi-criterion optimization, pp. 862–
876. Springer (2007)
43. Zitzler, E., Deb, K., Thiele, L.: Comparison of multiobjective evolutionary algorithms: Empirical
results. Evolutionary Computation 8(2), 173–195 (2000)
44. Zitzler, E., Thiele, L.: Multiobjective optimization using evolutionary algorithms – a comparative case
study. In: Parallel Problem Solving from Nature - PPSN-V, pp. 292–301. Springer (1998)
45. Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C.M., Da Fonseca, V.G.: Performance assessment of
multiobjective optimizers: an analysis and review. Evolutionary Computation, IEEE Transactions on
7(2), 117–132 (2003)