Comparing Parameter Tuning Methods for Evolutionary

Algorithms
S.K. Smit A.E. Eiben

Abstract— Tuning the parameters of an evolutionary al- showcase and some very interesting insights whose gener-
gorithm (EA) to a given problem at hand is essential for alization requires much more experimental research.
good algorithm performance. Optimizing parameter values is,
however, a non-trivial problem, beyond the limits of human II. PARAMETERS , T UNERS , AND U TILITY L ANDSCAPES
problem solving.In this light it is odd that no parameter tuning
algorithms are used widely in evolutionary computing. This Intuitively, there is a difference between choosing a good
paper is meant to be stepping stone towards a better practice crossover operator and choosing a good value for the related
by discussing the most important issues related to tuning EA crossover rate pc . This difference can be formalized if
parameters, describing a number of existing tuning methods, we distinguish parameters by their domains. The parameter
and presenting a modest experimental comparison among them.
The paper is concluded by suggestions for future research –
crossoveroperator has a finite domain with no sensible
hopefully inspiring fellow researchers for further work. distance metric, e.g., {onepoint, uniform, averaging},
Index Terms— evolutionary algorithms, parameter tuning whereas the domain of the parameter pc is a subset of IR
with the natural metric for real numbers.
I. BACKGROUND AND OBJECTIVES This difference is essential for searchability. For parame-
Evolutionary Algorithms (EA) form a rich class of stochas- ters with a domain that has a distance metric, one can use
tic search methods that share the basic principles of incre- heuristic search and optimization methods to find optimal
mentally improving the quality of a set of candidate solutions values. For the first type of parameters this is not possible
by means of variation and selection [7], [5]. Algorithms in because the domain has no exploitable structure. The only
this class are all based on the same generic framework whose option in this case is enumeration.
details need to be specified to obtain a particular EA. It is For a clear distinction between these cases we can use
customary to call these details EA parameters, and designing the terms symbolic parameter, e.g., crossoveroperator,
an EA for a given application amounts to selecting good and numeric parameter, e.g., crossover rate. For both types
values for these parameters. of parameters the elements of the parameter’s domain are
Setting EA parameters is commonly divided into two called parameter values and we instantiate a parameter by
cases, parameter tuning and parameter control [6]. In case of allocating a value to it.
parameter control the parameter values are changing during It is important to note that the number of parameters
an EA run. In this case one needs initial parameter values and of EAs is not specified in general. Depending on partic-
suitable control strategies, that in turn can be deterministic, ular design choices one might obtain different numbers
adaptive, or self-adaptive. Parameter tuning is easier in that of parameters. For instance, instantiating the symbolic pa-
the parameter values are not changing during a run, hence rameter parent-selection by tournament implies a
only a single value per parameter is required. Nevertheless, numeric parameter tournamentsize. However, choosing
even the problem of tuning an EA for a given application for roulette-wheel does not add any parameters. This
is hard because there is a large number of options, but only example also shows that there can be a hierarchy among
little knowledge about the effect of EA parameters on EA parameters. Namely, symbolic parameters may have numeric
performance. EA users mostly rely on conventions (mutation parameters under them. If an unambiguous treatment requires
rate should be low), ad hoc choices (why not use uniform we can call such parameters sub-parameters, always belong-
crossover), and experimental comparisons on a limited scale ing to a symbolic parameter.
(testing combinations of three different crossover rates and For positioning algorithms for parameter tuning it is help-
three different mutation rates). ful to distinguish three layers: application layer, algorithm
The main objective of this paper is to illustrate the fea- layer, and design layer, see Figure 1.
sibility of using tuning algorithms, thereby motivating their The lower part of this three-tier hierarchy consists of
usage. To this end, we describe three different approaches an EA on the algorithm layer trying to find an optimal
to algorithmic parameter tuning (meta-EA, meta-EDA, SPO) solution for the problem on the application layer, e.g., the
and show their (dis)advantages when tuning EA parameters traveling salesman problem. Simply put, the EA is iteratively
for solving the Rastrigin function. While the limited scale generating (candidate) solutions, e.g., permutations of city
(one single fitness landscape and one algorithm to be tuned) names, whose quality is determined by the given problem
prevents general conclusions, we do obtain a convincing on the application layer.
The upper part of the hierarchy contains a design method
Vrije Universiteit Amsterdam, The Netherlands, {sksmit, gusz}@cs.vu.nl that is trying to find optimal parameters for the EA on the

978-1-4244-2959-2/09/$25.00 
c 2009 IEEE 399

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
 objective function of the problem on the application layer and
differences between suitable fitness functions mostly concern
arithmetic details. The notion of utility, however, is based on
the performance of the EA that can be defined in essentially
different ways, for instance, based on solution quality or
algorithm speed. Furthermore, performance can be average
or peak performance over a number a EA runs. Consequently,
the definition of a good solution is more sensitive for user
(a) Control flow (b) Information flow preferences on the upper half (in the context of parameter
tuning) than on the lower half (in the context of an EA
Fig. 1. The 3-layered hierarchy of parameter tuning
application). Second, the performance of the EA depends
on the problem the EA is solving, that is, the definition of
algorithm layer. The design method can be, for instance, a utility depends on the definition of fitness.
heuristic procedure (algorithm) or an interactive session with III. A LGORITHMIC A PPROACHES TO PARAMETER
the user itself. We can formalize the problem to be solved T UNING
here by denoting the symbolic parameters and their domains
As mentioned in Section II, an EA has symbolic and
by q1 , . . . , qm and Q1 , . . . , Qm , likewise using the notation
numeric parameters. In general, the space of symbolic pa-
r1 , . . . , rn and R1 , . . . , Rn for the numeric parameters.1 The
rameters does not have a searchable structure and can only
problem of parameter tuning can then be seen as a search
be treated by enumeration or grid search methods. Therefore
problem in the parameter space.
we focus on the numeric parameters here and describe three
different approaches to optimizing them.
S = Q1 × Q2 · · · × Qm × R1 × R2 · · · × Rn (1)
Finding a good set of parameter values is a complex opti-
Solutions of the parameter tuning problem can then be mization task with a nonlinear objective function, interacting
defined as parameter vectors with maximum utility, where the variables, multiple local optima, noise (by the stochastic
utility of a given parameter vector p̄ ∈ S is the performance nature of the EA to be tuned), and a lack of analytic
of the EA using the values of p̄. Using this nomenclature we solvers. Ironically, it is exactly this type of problems where
can define the parameter-performance landscape, or utility EAs are very competitive heuristic solvers. It is therefore a
landscape, as an abstract landscape where the locations are natural idea to use an evolutionary approach to optimize the
the parameter vectors in S and the height of a p̄ ∈ S is its parameters of an evolutionary algorithm. Two of the three
utility. Intuitively it is quite obvious that fitness landscapes methods we describe in the following are based on this idea.
–commonly used in EC, i.e., within the context of the lower
A. Meta Evolutionary Algorithm
part of the hierarchy– have a lot in common with utility land-
scapes –as introduced here for the upper part. To be specific, Mercer and Sampson [13] were the first to introduce a
in both cases we have a search space (candidate solutions meta-EA, but due to the large computational costs, their
vs. parameter vectors), a quality measure (fitness vs. utility) research was very limited. Greffenstette [8] did conduct more
that is conceptualized as “height”, and a method to assess extensive experiments with his Meta-GA and showed its
the quality of a point in the search space (evaluation vs. effectiveness.
testing). Finally, we have a search method (an evolutionary The individuals used in such a meta-EA (on the design
algorithm vs. a tuning procedure) that is seeking for a point layer) are vectors of numerical values. Each of those values
with maximum height. Table I provides a quick overview of belong to one of the parameters of the baseline EA to be
the related vocabulary. tuned. To evaluate the utility of such a vector, the baseline EA
is ran several times using the given parameter values. Using
TABLE I this representation and utility as (meta) fitness, basically any
O NE - GLANCE OVERVIEW OF PARAMETER TUNING TERMINOLOGY evolutionary algorithm can be used as the meta-EA, if only
Lower part Upper part it can cope with real-valued vectors as individuals.
Method at work EA tuning procedure In this paper we use an Evolution Strategy (ES) with Co-
Search space solution vectors parameter vectors
Quality fitness utility
variance Matrix Adaptation (CMA) as proposed by Hansen
Assessment evaluation testing [9] as a meta-EA. This choice is motivated by the good rep-
utation of Evolutionary Strategies as numerical optimizers.
The CMA-ES is currently the state-of-the-art improvement
Despite the obvious analogies between the upper and the of the standard ES.
lower halfs, there are two differences we want to note here.
First, the notion of fitness is usually strongly related to the B. Meta Estimation of Distribution Algorithm
1 Observe
Nannen and Eiben have introduced a method for Relevance
that by the possible presence of sub-parameters the number
of numeric parameters n depends on the instantiations of q1 , . . . qm . This Estimation and Value Calibration of parameters (REVAC) in
makes the notation somewhat inaccurate, but use it for sake of simplicity. [16], [15]. Although the REVAC method was not designed

400 2009 IEEE Congress on Evolutionary Computation (CEC 2009)

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
with Estimation of Distribution Algorithms (EDA) in mind, that the number of tests performed to estimate the utility
it is based on the same general idea [10]. Like all EDAs, RE- of a parameter vector, n, is not used as a universal constant
VAC tries to find an optimal parameter vector by estimating throughout the search, but as a variable maximum. Using
the distribution of promising values over the domain of each racing we initially perform only a few tests for each vec-
parameter and creating specific vectors by drawing values tor, separate the ones that are clearly good, and iteratively
from these distributions. REVAC has a characteristic way increase the number of tests for those vectors only that are
of updating the distributions after having evaluated newly not significantly worse or better than the good ones. This
drawn vectors. In essence, REVAC is a population-based method can save a substantial number of tests compared to
stochastic search method, where the population consists of the simple each-vector-n-tests approach.
parameter vectors of the baseline EA and one individual (i.e., Yuan et al. [19] used this feature in their (1 + λ) ES for
one vector) is replaced in each cycle. After termination of the tuning an evolutionary algorithm. In their approach, at each
algorithm, the estimated distributions per parameter represent generation, a set of λ new vectors is created using a Gaussian
a model of the utility landscape. This model is rather simple distribution centered at the current best vector. Racing is
(separated by coordinates, resp. parameters, hence blind for then used to determine which vector has the highest utility.
parameter interactions), but it can be used to get insights into This approach can be easily extended to a (μ + λ) ES, by
the sensitivity and relevance of the different parameters and using racing to determine the μ best individuals instead of
the costs and benefits of tuning each parameter [17]. the single best.
In this paper we use REVAC with the settings from earlier
publications, not adjusted, let alone optimized, for the present B. Sharpening
case study. In this respect, it is different from the other two Sharpening has not been introduced before as a separate
methods, where we use variants that have been much studied technique for testing, although it has been used previously,
and improved since their ‘birth’. inside the SPOmethod by Bartz-Beielstein et al. [2]. Thus,
C. Sequential Parameter Optimization in this paper we do not invent it, but designate it as an
independent add-on, and give it the name sharpening. The
Sequential Parameter Optimization (SPO), as introduced
purpose of sharpening is to decrease the number of tests
by Bartz-Beielstein et al. [4], [1], is a search-method specif-
needed to estimate the quality of parameter vectors as
ically designed for parameter tuning and parameter analysis.
compared to the simple each-vector-n-tests approach. Like
The approach shows some similarities is similar to a meta-
racing, it is to reduce the total runtime of a tuner algorithm.
EDA in that it relies on a model of the utility landscape.
The main idea is to start the tuning algorithm with a small
SPO starts with a initial population of vectors. Those are
number of tests per vector, but when a certain threshold
tested several times to determine their utility. Based on the
is reached the amount of tests per vector is doubled. This
results, a (regression) model is fitted to represent the relation
means that the algorithm is able to explore the search space
between the vectors and the results. Then s new vectors are
very quickly. If a promising area is found, the method
generated and tested using this model. The most promising
focuses on improving the estimates by reducing the effect
points are then added to the population. Although in [3], [2]
of possible outliers on the utility. Therefore, at the moment
regression models are used to model the utilities, succeded by
of termination, the current best vector is tested very often.
stochastic models [4], it is in principle a general framework
This can lead to better results than algorithm that tests each
suited for a large range of modeling techniques.
vector only a couple of times.
In this paper we have chosen to use Kriging models
for approximating the utility landscape, because of their C. Combining Racing and Sharpening
excellent performance on tuning problems with numerical
parameters [4]. Observe that racing and sharpening are opposing forces.
Sharpening is increasing the number of tests, while racing
IV. A DD - ONS FOR PARAMETER T UNING A LGORITHMS is reducing them. Nevertheless, they can be combined very
A careful study of related work discloses that besides the easily. In a combined setup sharpening will increase the
principal parameter tuning algorithms, like meta-EA, REVAC maximum number of tests that can be used by racing to
or SPO, there are a number of useful ‘add-ons’, i.e., methods select the best parameter vectors. In the beginning of the
for increasing search efficiency, that are independent from tuning-run the effect of racing will be very small, due to
the main tuner and can be combined with different tuning the small ’budget’, but during the run, when more and
algorithms. It this section we highlight two of such promising more tests are required to sharpen the estimates, the role
add-ons. of racing will get more important. By using racing not
much effort is spent on vectors that are not very promising,
A. Racing even if sharpening already increased the number of tests.
Racing was introduced by Maron and Moore [12]. The In principle, we can get the best of both worlds using this
purpose of racing is to decrease the number of tests needed setup. By combining sharpening and racing much more effort
to estimate the quality of parameter vectors, and thereby is spent on promising vectors while the effort wasted on bad
the total runtime of a tuner algorithm. The main idea is vectors is reduced.

2009 IEEE Congress on Evolutionary Computation (CEC 2009) 401

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
 V. S YSTEM D ESCRIPTION AND E XPERIMENTAL S ETUP it can read and write SPO property files. This requires a bit
more work, but it is more flexible.
As described in Section II, the complete system consists of
three different layers with a control and information flow. On B. Information Flow
the application level we have chosen to use a 20 dimensional The information flow between the layers uses the same
Rastrigin function. The Rastrigin function [18] is a popular interfaces or property files as used in the control flow. The
non-linear, highly multimodal, scalable benchmark function information that is passed to the algorithm layer by the
with an optimal value2 of 0. In this system, we have chosen application layer is a single value representing the fitness
for the implementation from the ECJ [11] library which is of the current vector. The algorithm layer, however, does not
open source and freely available. send a single value to the design layer, but a list of utilities.
For the Algorithm layer we again choose for an imple- The design layer algorithm defines how many tests have to be
mentation from the ECJ library. This library is widely used executed in order to evaluate a single parameter vector. If the
by EC practitioners and serves as a framework for a whole design algorithm requests for s tests of a certain parameter
range of evolutionary algorithms. It is written in Java and vector, the corresponding information that is send contains
allows users to configure Evolutionary Algorithms using Java s values, which represent the best fitness values from each
code or parameter-files. We have chosen for a ’standard’ of the s test runs. In this setup the tuning is focused on
configuration of the middle-layer algorithm (Table II). improving the fitness.
This setup requires 6 parameters to be defined by the The same setup can be used to tune for speed. But in
design algorithm from the top layer. that case, the algorithm layer have to send the s number of
For the top layer, we have tested three different algorithms evaluations used to solution as information for the design
combined with the two additional add-ons namely: layer.
• CMA-ES
• CMA-ES with Racing
C. Measures
• CMA-ES with Sharpening Each of the tuning-algorithms is allowed to perform a total
• CMA-ES with Racing and Sharpening of 1000 tests and is repeated 10 times. To measure the quality
• SPO (uses Sharpening) of each tuning-algorithm the following criteria are used:
• REVAC • Average Performance
• REVAC with Racing • Maximum Performance
• REVAC with Sharpening • Variation of Performance
• REVAC with Racing and Sharpening • Effectiveness of the Tuning Algorithm

The three base algorithms, CMA-ES [9], SPO [2] and The average performance is measured by the Mean Best
Racing [14], are open source and freely available from the Utility over the 10 runs. To estimate the utility, the best
websites of the corresponding authors. The additional add- vector that is found in each run, is tested 50 times. The
ons are handcrafted changes to the three algorithms and utility is equal to the average fitness values over the 50 runs.
their exact implementations are algorithm specific. All of the Maximum performance is measured similarly, but instead of
tuning algorithms are ran with their default parameter values the average utility over the 10 runs, the maximum utility is
and setup (Table III, IV and V). used.
To measure the variation in performance, the difference
A. Control Flow in top and lower quantiles of the utility distribution of the
The control flow between the application and the algorithm 10 runs is used. Finally, to measure the effectiveness of the
is trivial and only requires the definition of two parameters. tuning algorithm we identify the ’target’ area, namely the
Both layers are defined in the ECJ library and fit into the area of the utility landscape with the highest performance.
same framework. However, implementing the the control For each parameter the minimum and the maximum value is
flow between the algorithm and the design level is slightly calculated over the 10 best vectors. The effectiveness of the
more difficult. The CMA-ES is implemented in Java, so algorithms is defined by its ability to reach the area enclosed
creating the link between both layers is quite straightforward. by these values.
By implementing a Java interface that sets the algorithm VI. R ESULTS
parameters and executes the ECJ algorithm, both layers
Tuning an algorithm requires a lot of computer power,
are connected. REVAC and SPO are both implemented in
while some people argue that this is a waste of time. General
Matlab, which is also able to communicate directly with
thumb rules as a population size of 100 and low mutation
Java libraries. For REVAC the user needs to write a Matlab
sizes are supposed to perform reasonably well. The question
function that communicates with a Java class. This Java class
rises how beneficial tuning, and more specific automated
can be the same as used with the CMA-ES implementation.
tuning, is even to experienced practitioners.
SPO can be linked to any function or executable as long as
For quick assessment of the added value of algorithmic
2 Because we have a preference for maximization problems, a negative tuning we tested an EA using parameter values defined by
transformation is applied ‘common sense’ (Table VI).

402 2009 IEEE Congress on Evolutionary Computation (CEC 2009)

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
S ETUP OF THE A LGORITHM L AYER
Value Parameters
Population Single population Population size (POPSIZE)
Parent Selection Tournament selection Tournament proportion (TOURPROP)
Elitism Yes Generation gap (GENGAP)
Crossover Uniform Crossover probability (CROSSPROB)
Mutation Gaussian Mutation probability (MUTPROB) and σ (MUTSIZE)
Termination 20.000 fitness evaluations

TABLE III TABLE IV TABLE V

PARAMETER VALUES CMA-ES PARAMETER VALUES SPO PARAMETER VALUES REVAC

Parameter Value Parameter Value Parameter Value

μ 4 # Initial design points 60 Population size 100
λ 9 # Samples per point 1 Selected Points 50
Racing # Candidates 10 Smoothing 5
Minimum # evaluations 2 Racing
Maximum # evaluations 5 Minimum # evaluations 2
Sharpening Maximum # evaluations 5
Factor 1.5 Sharpening
Threshold 18 Factor 1.5
Threshold 100

The average performance of this manually chosen param- components to the algorithm is therefore probably beneficial,
eter vector is -44.00, while the EAs using parameter vectors although it is difficult to predict the effect on performance.
optimized by the tuners easily reach utility levels around - Our experimental data can not only be used to learn about
0.05. tuning algorithms, but it could also tell about high quality
Table VII shows the minimum and maximum performance settings of the baseline EA. For this purpose we investigate
of the 10 runs. It also indicates the median performance, the 10 best parameter vectors found in all runs. Table VIII
and the four quantiles. Because algorithm tuning is a kind shows these 10 parameter vectors, together with the tuner that
of design problem [7] the maximum(peak) performance is found them and the corresponding utility, being the mean best
probably the most interesting value. Tuning an algorithm fitness of the EA using the given parameter vector, averaged
is not a repetitive task, in the sense that it is not required over 50 independent runs.
to deliver a good value each time it is ran. The average Interestingly, there are rather big differences between
performance is therefore less important than the maximum the top 10 parameter vectors, depending on the specific
performance that can be reached. parameters. For instance, the optimized population size varies
The best parameter settings are found by the CMA- between 11 and 448. Also for the generation gap we find
ES, racing and sharpening combination, followed by SPO. optimized values far from each other, e.g., 4% (in combi-
REVAC shows on average a significantly worse performance, nation with population size 23) and 84% (in combination
however when combined with both racing and sharpening, with population size 14). To obtain more information about
the best performance gets close to the performance of the the spreading of optimized parameter values we performed
other algorithms. The main cause of the bad performance of experiments with the best variant of each of the three basic
REVAC is the speed. From more detailed results, not shown methods: CMA-ES with racing and sharpening, REVAC
here, we observe a steady increase in performance during with racing and sharpening, and SPO. We executed 10
the run. However, the 1000 allowed tests, forces REVAC to independent runs with each of them, resulting in 3 times
terminate prematurely. 10 optimized parameter vectors. The outcomes are shown
It is clear that the combined effect of racing and sharpening in Figure 2, split by parameter, the dots showing the actual
increases the maximum performance in this setup, however, parameter values. These results show that even if we use the
the effect on the CMA-ES is quite different than the effect same tuning algorithm, we can get very different optimized
on REVAC. When combining racing or sharpening with values, although this picture varies per parameter and by
CMA-ES, the variance increases and sharpening alone even tuning algorithm. For instance, the CMA-ES and SPO are
decreases the overall performance. However, when combined consistent in their values for mutation probability, but this
with REVAC, all three combinations decrease variance and does not hold for REVAC. As for the population size, all
increase the best and the overall performance. Adding those three algorithms show a wide range of good values. For

2009 IEEE Congress on Evolutionary Computation (CEC 2009) 403

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
 TABLE VI
M ANUAL C HOICE OF PARAMETER VALUES
Value
Population size 100
Tournament proportion 3%
Generation gap 2%
Crossover probability 0.8
Mutation probability 1
σ 0.1

TABLE VII
T HE MINIMUM , MAXIMUM , MEDIAN AND QUANTILES OF THE ALGORITHM UTILITIES

Algorithm Max Q1 Q.75 Median Q.5 Q.25 Min

CMA-ES -0.0508 -0.0508 -0.0542 -0.0589 -0.0783 -0.0844 -0.1761
CMA-ES (Racing) -0.0545 -0.0545 -0.0608 -0.0667 -0.0747 -0.0847 -0.0847
CMA-ES (Sharpening) -0.0508 -0.0508 -0.0542 -0.0713 -0.0910 -0.1085 -0.2317
CMA-ES (Racing , Sharpening) -0.0388 -0.0388 -0.0490 -0.0751 -0.0966 -0.1381 -0.1381
SPO -0.0457 -0.0457 -0.0482 -0.0623 -0.0776 -0.0776 -0.0776
REVAC -0.1031 -0.1031 -0.1588 -0.3502 -83.6759 -40.9933 -83.6759
REVAC (Racing) -0.0678 -0.0678 -0.2934 -0.9423 -2.3175 -2.3175 -67.7445
REVAC (Sharpening) -0.0822 -0.0822 -0.0936 -0.2762 -5.1779 -5.1779 -357.4347
REVAC (Racing , Sharpening) -0.0573 -0.0573 -0.0784 -0.1642 -19.7132 -19.7132 -102.5977

Illustration of the quantiles. Mind the different scales

comparison with the results concerning the overall top 10 landscape and one EA to be tuned). While this prevents
vectors, we augmented Figure 2 with blocks exhibiting the general conclusions, we did obtain a convincing showcase
‘target areas’, where the upper/lower borders of the block and some very interesting insights that motivate further
show the maximum/minimum values from Table VIII, and research and development.
the middle line belongs to the mean. The overall picture that Perhaps the most important conclusion is that using algo-
arises is that, except the mutation parameters, it is hard for rithms for tuning parameters of EAs does pay off in terms
tuners to consistently reach the areas with the best EA setup. of EA performance. To be specific, the best guess (i.e.,
This shows that it is needed to run each tuning-algorithm the parameter vector with the highest utility) of all of the
several times in order to find a good parameter setup. algorithms we tested greatly outperforms the best guess of a
VII. C ONCLUSIONS AND O UTLOOK human user. Simply put, no matter what tuner algorithm you
The main objective of this paper is to illustrate the feasibil- use, you will likely get a much better EA than relying on
ity of using algorithms for tuning parameters of EAs. To this your intuition and the usual parameter setting conventions.
end, we performed experiments with ten tuning algorithms, Further to the EA performance benefits, tuner algorithms
based on three different approaches, meta-EA, REVAC, and are also useful for they go for the best parameter vectors
SPO. As mentioned before, due to computational and time without being hindered by those rules-of-thumb human users
limitations these tests have a limited scale (one single fitness rely on. This can lead to surprising parameter settings,

404 2009 IEEE Congress on Evolutionary Computation (CEC 2009)

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
 TABLE VIII
T HE 10 B EST P ERFORMING PARAMETER V ECTORS F OUND

Algorithm Utility POPSIZE MUTPROB MUTSIZE CROSSPROB TOURPROP GENGAP

1 CMA-ES (R, S) -0.0388 37 0.0510 0.5641 0.5225 0.8702 0.5535
2 CMA-ES (R, S) -0.0457 14 0.0502 0.5684 0.4537 0.8782 0.8443
3 SPO -0.0457 448 0.0540 0.6219 0.5000 0.3503 0.0215
4 SPO -0.0472 23 0.0686 0.6130 0.7789 0.8580 0.0408
5 SPO -0.0482 33 0.0472 0.6370 0.7536 0.6725 0.2173
6 CMA-ES (R, S) -0.0490 122 0.0631 0.5897 0.6525 0.5108 0.1582
7 CMA-ES (S) -0.0508 271 0.0557 0.5955 0.3886 0.9207 0.1078
8 CMA-ES -0.0508 271 0.0557 0.5955 0.3886 0.9207 0.1078
9 CMA-ES (S) -0.0514 11 0.0416 0.6358 0.5791 0.4946 0.6300
10 CMA-ES -0.0514 11 0.0416 0.6358 0.5791 0.4946 0.6300
Percentage of Total Range 54.63 % 2.70 % 7.29 % 39.02 % 57.04 % 82.28 %

Fig. 2. The parameter values found by the best variants of our three methods (dots) and the ’target’ area (blocks). See text for explanation.

thereby offering a critical look on such rules-of-thumb. For this scale we have the CMA-ES that is a highly specialized
instance, in most EA publications the tournament size is optimizer building no model of the utility landscape. REVAC,
typically in the range of 2 to 10. The optimal values, however, and meta-EDAs in general, does create a model, the marginal
seem to higher, in the range of tens (while population sizes distributions over the ranges of each parameter. The fact
are rather conventional). Strictly speaking, this only holds that these distributions only take one parameter into account
for the Rastrigin function and the EA we investigated here, means that the model is simple, it is blind to parameter
but we do believe that the conventional wisdom is wrong in interactions. On the other hand, REVAC is able to provide
many more cases and tuning algorithms can help to show information about the entropy associated with the parameters,
this. hence showing the amount of tuning each parameter requires.
SPO is situated on the high end of the insights scale, since
Our results also support preferences regarding the tuning
it is inherently based on a model of the utility landscape. In
algorithms to be used. For a careful advise, we need to
principle, this model is not restricted to a specific form or
distinguish two functionalities tuners can offer. First and
structure, offering the most flexibility and insights, includ-
foremost, they can optimize EA parameters, second they can
ing information on parameter interactions. Based on these
provide insights into the (combined) effects of parameters
considerations and the outcomes of our experiments our
on EA performance. Regarding the insights offered the three
preferred method is the CMA-ES if a very good parameter
methods we tested are quite different. On the low end of

2009 IEEE Congress on Evolutionary Computation (CEC 2009) 405

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.
vector is the most important objective, and SPO if one is also [17] V. Nannen, S.K. Smit, and A.E. Eiben. Costs and benefits of tuning
interested in detailed information over the EA parameters. parameters of evolutionary algorithms. In G. Rudolph, Th. Jansen,
S.M. Lucas, C. Poloni, and N. Beume, editors, Parallel Problem
Regarding future work we see a number of promising Solving from Nature – PPSN X, volume 5199 of Lecture Notes in
directions. The most straightforward track is to extend the Computer Science, pages 528–538. Springer Berlin / Heidelberg, 2008.
scope of the present study and perform much more experi- [18] Aimo Torn and Antanas Zilinskas. Global optimization. Springer-
Verlag New York, Inc., New York, NY, USA, 1989.
ments using more objective functions and different EAs. This [19] B. Yuan and M. Gallagher. Combining Meta-EAs and Racing for
is needed to refine and consolidate our present findings and Difficult EA Parameter Tuning Tasks. In F.G. Lobo, C.F. Lima,
will most likely disclose new facts. From the practical point and Z. Michalewicz, editors, Parameter Setting in Evolutionary Al-
gorithms, pages 121–142. Springer, 2007.
of view, the development of a toolbox is the most urgent task.
Such a toolbox should contain one or more parameter tuning
algorithms allowing their combinations with racing and/or
sharpening. Furthermore, such a toolbox should be easy to
use. That is, it should enable EC practitioners with limited
time and computer experience to plug in their EA and the
problem to be solved and produce good parameter settings.

R EFERENCES
[1] T. Bartz-Beielstein, C.W.G. Lasarczyk, and M. Preuss. Sequential pa-
rameter optimization. In IEEE Congress on Evolutionary Computation,
volume 1, pages 773–780 Vol.1. IEEE, Sept. 2005.
[2] T. Bartz-Beielstein, K.E. Parsopoulos, and M.N. Vrahatis. Analysis
of Particle Swarm Optimization Using Computational Statistics. In
Chalkis, editor, Proceedings of the International Conference of Nu-
merical Analysis and Applied Mathematics (ICNAAM 2004), pages
34–37, 2004.
[3] Thomas Bartz-Beielstein. Experimental Analysis of Evolution Strate-
gies: Overview and Comprehensive Introduction. Technical Report
Reihe CI 157/03, SFB 531, Universität Dortmund, Dortmund, Ger-
many, 2003.
[4] Thomas Bartz-Beielstein and Sandor Markon. Tuning search algo-
rithms for real-world applications: A regression tree based approach.
Technical Report of the Collaborative Research Centre 531 Computa-
tional Intelligence CI-172/04, University of Dortmund, March 2004.
[5] K.A. De Jong. Evolutionary Computation: A Unified Approach. The
MIT Press, 2006.
[6] A.E. Eiben, R. Hinterding, and Z. Michalewicz. Parameter Control
in Evolutionary Algorithms. IEEE Transactions on Evolutionary
Computation, 3(2):124–141, 1999.
[7] A.E. Eiben and J.E. Smith. Introduction to Evolutionary Computation.
Natural Computing Series. Springer, 2003.
[8] J.J Greffenstette. Optimisation of Control Parameters for Genetic
Algorithms. In IEEE Transactions on Systems, Man and Cybernetics,
volume 16, pages 122–128, 1986.
[9] N. Hansen. The CMA evolution strategy: a comparing review. In J.A.
Lozano, P. Larranaga, I. Inza, and E. Bengoetxea, editors, Towards a
new evolutionary computation. Advances on estimation of distribution
algorithms, pages 75–102. Springer, 2006.
[10] Pedro Larraanaga and Jose A. Lozano. Estimation of Distribution
Algorithms: A New Tool for Evolutionary Computation. Kluwer
Academic Publishers, Norwell, MA, USA, 2001.
[11] S. Luke et al. A java-based evolutionary computation research system.
http://www.cs.gmu.edu/∼eclab/projects/ecj/.
[12] O. Maron and A. Moore. The racing algorithm: Model selection for
lazy learners. In Artificial Intelligence Review, volume 11, pages 193–
225, April 1997.
[13] R.E. Mercer and J.R. Sampson. Adaptive search using a reproductive
metaplan. Kybernetes, 7:215–228, 1978.
[14] V. Nannen and A. E. Eiben. Efficient Relevance Estimation and Value
Calibration of Evolutionary Algorithm Parameters. In IEEE Congress
on Evolutionary Computation, pages 103–110. IEEE, 2007.
[15] V. Nannen and A. E. Eiben. Relevance Estimation and Value
Calibration of Evolutionary Algorithm Parameters. In Manuela M.
Veloso, editor, IJCAI, pages 1034–1039, 2007.
[16] V. Nannen and A.E. Eiben. A Method for Parameter Calibration and
Relevance Estimation in Evolutionary Algorithms. In M. Keijzer,
editor, Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO-2006), pages 183–190. Morgan Kaufmann, San
Francisco, 2006.

406 2009 IEEE Congress on Evolutionary Computation (CEC 2009)

Authorized licensed use limited to: Southern Federal University. Downloaded on February 24,2021 at 06:40:03 UTC from IEEE Xplore. Restrictions apply.