Disturbed Exploitation compact Differential Evolution for limited

memory optimization problems

Ferrante Neri

, Giovanni Iacca, Ernesto Mininno

Department of Mathematical Information Technology, P.O. Box 35 (Agora), 40014 University of Jyvskyl, Finland
a r t i c l e i n f o
Article history:
Received 15 August 2010
Received in revised form 2 February 2011
Accepted 4 February 2011
Available online 15 February 2011
Keywords:
Differential Evolution
Evolutionary algorithms
Compact algorithms
Memetic Computing
a b s t r a c t
This paper proposes a novel and unconventional Memetic Computing approach for solving
continuous optimization problems characterized by memory limitations. The proposed
algorithm, unlike employing an explorative evolutionary framework and a set of local
search algorithms, employs multiple exploitative search within the main framework and
performs a multiple step global search by means of a randomized perturbation of the vir-
tual population corresponding to a periodical randomization of the search for the exploit-
ative operators. The proposed Memetic Computing approach is based on a populationless
(compact) evolutionary framework which, instead of processing a population of solutions,
handles its statistical model. This evolutionary framework is based on a Differential Evolu-
tion which cooperatively employs two exploitative search operators: the rst is based on a
standard Differential Evolution mutation and exponential crossover, and the second is the
trigonometric mutation. These two search operators have an exploitative action on the
algorithmic framework and thus contribute to the rapid convergence of the virtual popu-
lation towards promising candidate solutions. The action of these search operators is coun-
terbalanced by a periodical stochastic perturbation of the virtual population, which has the
role of disturbing the excessively exploitative action of the framework and thus inhibits
its premature convergence. The proposed algorithm, namely Disturbed Exploitation com-
pact Differential Evolution, is a simple and memory-wise cheap structure that makes use
of the Memetic Computing paradigm in order to solve complex optimization problems.
The proposed approach has been tested on a set of various test problems and compared
with state-of-the-art compact algorithms and with some modern population based
meta-heuristics. Numerical results show that Disturbed Exploitation compact Differential
Evolution signicantly outperforms all the other compact algorithms present in literature
and reaches a competitive performance with respect to modern population algorithms,
including some memetic approaches and complex modern Differential Evolution based
algorithms. In order to show the potential of the proposed approach in real-world applica-
tions, Disturbed Exploitation compact Differential Evolution has been implemented for
performing the control of a space robot by simulating the implementation within the robot
micro-controller. Numerical results show the superiority of the proposed algorithm with
respect to other modern compact algorithms present in literature.
2011 Elsevier Inc. All rights reserved.
0020-0255/$ - see front matter 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.ins.2011.02.004

Corresponding author. Tel.: +358 14 260 1211; fax: +358 14 260 1021.
E-mail addresses: ferrante.neri@jyu. (F. Neri), giovanni.iacca@jyu. (G. Iacca), ernesto.mininno@jyu. (E. Mininno).
Information Sciences 181 (2011) 24692487
Contents lists available at ScienceDirect
Information Sciences
j our nal homepage: www. el sevi er . com/ l ocat e/ i ns
1. Introduction
Memetic Computing (MC) is a promising and broad research area in computer science which considers the meme as the
unit of information with the purpose of solving problems, see [56,57]. The concept of meme is borrowed from philosophy
and is intended as the unit of cultural transmission. In other words, complex ideas can be decomposed into memes which
propagate and mutate within a population. Culture, in this way, constantly undergoes evolution and tends towards progres-
sive improvements.
In computer science, the concept of meme can be identied as a search strategy, an agent, or an operator belonging to a
complex system, e.g. an optimization algorithm, which is evolving over time. The earliest example of the MC paradigm is the
Memetic Algorithm (MA) [47], i.e. a hybrid evolutionary algorithm which employs deterministic local search within the evo-
lutionary cycle in order to solve a given optimization problem.
The importance and need of MC in optimization must be put into relationship with the No Free Lunch Theorem (NFLT)
[77], which mathematically proves that the average performance of any pair of algorithms A and B across all possible prob-
lems is identical. Thus, if an algorithm performs well on a certain class of problems, then it necessarily pays for that with
degraded performance on the set of all remaining problems, as this is the only way that all algorithms can have the same
performance averaged over all functions. Strictly speaking, the proof of NFLT is made under the hypothesis that both the
algorithms A and B are non-revisiting, i.e. the algorithms do not perform the tness evaluation of the same candidate solu-
tion more often than once during the optimization run. Although this hypothesis is de facto not respected for most of the
computational intelligence optimization algorithms, the concept that there is no universal optimizer had a signicant impact
on the scientic community.
It must be observed that until the NFLT publication, Evolutionary Algorithms (EAs) were extremely popular among com-
puter scientists, and their related research was oriented towards the design of algorithms having a superior performance
with respect to all the other algorithms present in literature. This approach is visible in many famous texts published in
those years, e.g. [21]. After the NFLT diffusion, researchers in optimization had to dramatically change their view about
the subject. More specically, in light of increasing interest in general purpose optimization algorithms, it has become
important, to understand the relationship between the performance of an algorithm A and a given optimization problem
f. Thus, the problem f became the starting point for building up a suitable algorithm.
Since each optimization problem with its features, e.g. multimodality of the landscape, separability of the function, epis-
tasis, etc., must be considered as the basis for building up an optimization solver that addresses the specic features of the
problem, NFLT constitutes, in a certain sense, the Full Employment Theorem (FET) for optimization professionals. In com-
puter science and mathematics, the term FET is used to refer to a theorem that shows that no algorithm can optimally per-
form a particular task done by some class of professionals. In this sense, as no efcient general purpose solver exists, there is
always scope for improving algorithms for better performance on particular problems.
Since MC, as mentioned above, represents a broad class of algorithms that combine various algorithmic components
(memes), a suitable combination is necessary for a given problem. Since, during the last decade, computer scientists had
to observe the features of their optimization problem in order to propose an ad hoc optimization algorithm, the approach
of combining various search operators within the algorithmic design became a common practice. In this sense, the develop-
ment of NFLT implicitly encouraged the use and development of MC approaches, which became extremely popular and often
necessary, in computer science, at rst, and in engineering and applied science, more recently, thus constituting the FET for
MC.
1.1. Memetic and compact computing: a literature survey
A crucially important issue in MC is the selection and coordination of the memes within an algorithmic structure. By
updating the classication given in [58], MC approaches can be subdivided as: (1) adaptive hyper-heuristic, see e.g.
[9,32,5,34], where the coordination of the memes is performed by means of heuristic rules; (2) self-adaptive and co-evolu-
tionary, see e.g. [69,80,39], where the memes, either directly encoded within the candidate solutions or evolving in parallel
to them, take part in the evolution and undergo recombination and selection in order to select the most promising operators;
(3) meta-Lamarckian learning, see e.g. [55,35,53,41], where the success of the memes biases their activation probability, thus
performing an on-line algorithmic design which can exibly adapt to various optimization problems; (4) diversity-adaptive,
see e.g. [6,51,52,50,73,7,72], where a measure of the diversity is used to select and activate the most appropriate memes. In
addition, it is worthwhile commenting Baldwinian systems, i.e. those MC approaches which do not modify the solutions after
the employment of local search, see [82,22]. An early attempt of modeling a system capable to automatically generate MC
approaches for given problems is described in [43].
Although these adaptive systems allow the generation of robust algorithms which are able to handle diverse optimization
problems, some classes of problems followdifferent logics and are not easily solvable by the straightforward implementation
of the above-mentioned MC approaches. This paper addresses a class of problems falling within this category, i.e. the opti-
mization despite a limited computational device, especially memory hardware. This situation is typical in control engineer-
ing and robotics, most of all for commercial robotics. For example, robots for domestic purposes (e.g. a vacuumcleaner robot)
need to undergo a learning process which can result in a complex optimization problem. This problem must be solved
2470 F. Neri et al. / Information Sciences 181 (2011) 24692487
quickly and without counting on a full power computer, due to volume and cost constraints. Obviously, algorithms
employing an archive, computationally expensive learning processes, or a large population size are not affordable for the
hardware.
In order to address this class of problems compact Evolutionary Algorithms (cEAs) have been designed. A cEA is an EA
belonging to the class of Estimation of Distribution Algorithms (EDAs), see [40]. The algorithms belonging to this class do
not store and process an entire population and all its individuals therein but on the contrary make use of a statistical rep-
resentation of the population in order to perform the optimization process. In this way, a much smaller number of param-
eters must be stored in the memory. Thus, a run of these algorithms requires much less capacious memory devices compared
to their correspondent standard EAs.
The rst cEA was the compact Genetic Algorithm (cGA) introduced in [24]. The cGA simulates the behavior of a standard
binary encoded Genetic Algorithm (GA). In [24], it can be seen that cGA has a performance almost as good as that of GA. An
analysis of the convergence properties of cGA by using Markov chains is given in [63]. The possibility of designing parallel
versions of cGA is explored in [27], where a memetic approach is also proposed in [42] where a strategy for performing ne-
grained parallelization is presented. In [23] (see also [25]) the extended compact Genetic Algorithm (ecGA) has been pro-
posed. The ecGA is based on the idea that the choice of a good probability distribution is equivalent to linkage learning.
The measure of a good distribution is based on Minimum Description Length (MDL) models: simpler distributions are better
than the complex ones. The probability distribution used in ecGA is a class of probability models known as Marginal Product
Models (MPMs). A theoretical analysis of the ecGA behavior is presented in [66]. A hybrid version of ecGA integrating the
NelderMead algorithm is proposed in [68] and another hybrid version with Particle Swarm Optimization (PSO) logic is re-
cently proposed in [1]. A study on the scalability of ecGA is given in [67]. The cGA and its variants have been intensively used
to perform hardware implementation, see [3,18,31]. A cGA application to neural network training is given in [17]. Ahn and
Ramakrishna [2] analyze analogies and differences between cGAs and (1 + 1)-ES, and extends a mathematical model of ES
[65] to cGA obtaining useful information on the performance. Moreover, Ref. [2] introduces the concept of elitism, and pro-
poses two new variants, with strong and weak elitism respectively, that signicantly outperform both the original cGA and
(1 + 1)-ES. A real-encoded cGA (rcGA) has been introduced in [44]. Some examples of rcGA applications to control engineer-
ing are given in [10,11].
Recently, a compact algorithm based on Differential Evolution (DE) logic and namely compact Differential Evolution (cDE)
has been proposed in [45]. This algorithm encodes the population within a probabilistic distribution and employs the stan-
dard DE logic for generating new trial solutions and selecting the most promising search directions. The cDE algorithm
showed a performance superior to other compact algorithms for a large set of test problems. Finally, a simple compact
MC approach composed of a cDE framework and an ad hoc compact local search algorithm, in a traditional memetic fashion,
has been proposed in [48]. This MC approach, namely Memetic compact Differential Evolution (McDE), leads to some
improvement with respect to the evolutionary framework. These improvements, unlike for population-based algorithms ap-
pear not to be dramatic. In this sense, the memetic combination of evolutionary framework and local search algorithms is
likely not the most efcient choice for compact algorithms. This fact is, according to our interpretation, due to the exploit-
ative nature of the compact frameworks. Thus, the integration of an extra local search algorithm fails at offering a radically
different perspective in the tness landscape search, see [36,38].
1.2. Algorithmic proposal contained in this paper
This paper proposes a novel compact MC approach for continuous optimization, namely Disturbed Exploitation compact
Differential Evolution (DEcDE). The proposed memetic scheme, instead of delegating the global search to the evolutionary
framework and the local search to the local search algorithms (in a traditional memetic fashion as for McDE in [48]), employs
anintensively exploitative evolutionary framework basedona DE logic aided by a shallowdepthexploitative local searchalgo-
rithm. Thus, the exploration is delegated not explicitly to a search operator but to a perturbation mechanismwhich disturbs
the exploitation of the search operators and allows an on-line re-start of the exploitation of unexplored areas of the decision
space. The resulting algorithmis characterized by a simple structure which makes use of the MC paradigmin order to display a
high performance on a diverse set of problems despite its limitations in terms of complexity and memory employment.
The remainder of this paper is organized in the following way. The algorithmic description of DEcDE and its justication
and algorithmic philosophy is given in Section 2. Section 3 displays the numerical results on a set of test problems and sub-
divides them into two parts: a comparison of DEcDE against other modern compact algorithms, and a comparison with state-
of-the-art complex population based algorithms selected amongst modern MAs and DE based algorithms. The suitability of
cDE to solve challenging problems in environments with limited computational resources is assessed in Section 4 by an
application on a challenging real-world optimization problem in the eld of robotics. Finally Section 5 gives the conclusive
remarks of this work.
2. Disturbed Exploitation compact Differential Evolution
In order to clarify the notation used throughout this chapter, we refer to the minimization problem of an objective func-
tion f(x), where x is a vector of n design variables in a decision space D. Without loss of generality, let us assume that param-
F. Neri et al. / Information Sciences 181 (2011) 24692487 2471
eters are normalized so that each search interval is [1, 1]. In this study, the optimization should be performed by attempting
to employ the smallest possible memory capacity.
This paper proposes a Memetic Computing approach, namely Disturbed Exploitation compact Differential Evolution (DEc-
DE), in order to address the class of above-mentioned optimization problems. The proposed algorithm consists of the follow-
ing steps. At the beginning of the optimization process, a (2 n) Probability Vector (PV) is generated. More specically, PV is a
n 2 matrix:
PV
t
l
t
; r
t
; 1
where l and r are, respectively, vectors containing, for each design variable, mean and standard deviation values of a Gauss-
ian Probability Distribution Function (PDF) truncated within the interval [1, 1]. The height of the PDF has been normalized
in order to keep its area equal to 1. The apex t indicates the generation (number of performed comparison).
For each design variable i, l
1
[i] = 0 and r
1
[i] = k where k is a large positive constant (e.g. k = 10). This initialization of r
values is done in order to simulate a uniform distribution. Then, one individual is sampled from PV. This individual is indi-
cated as elite x
e
. Subsequently, the search mechanism is organized in the following way. Three individuals x
r
, x
s
and x
t
are
pseudo-randomly sampled from the PV.
More specically, the sampling mechanism of a design variable x
r
[i] associated to a generic candidate solution x
r
from PV
consists of the following steps. As mentioned above, for each design variable indexed by i, a truncated Gaussian PDF char-
acterized by a mean value l[i] and a standard deviation r[i] is associated. The formula of the PDF is:
PDF li; ri
e

xli
2
2ri
2

2
p

ri erf
li1

2
p
ri

erf
li1

2
p
ri
; 2
where erf is the error function, see [19].
From the PDF, the corresponding Cumulative Distribution Function (CDF) is constructed by means of Chebyshev polyno-
mials according to the procedure described in [8]. It must be observed that the codomain of CDF is [0, 1]. In order to sample
the design variable x
r
[i] from PV, a random number rand(0, 1) is sampled from a uniform distribution. The inverse function of
CDF, in correspondence of rand(0, 1), is then calculated. This latter value is x
r
[i].
A provisional offspring x
0
off
is then generated by mutation, according to a DE logic, as:
x
0
off
x
t
Fx
r
x
s
; 3
where usually F 2 [0, 2] is a scale factor which controls the length of the exploration vector (x
r
x
s
) and thus determines how
far from point x
t
the offspring should be generated. The mutation scheme shown in formula (3) is also known as DE/rand/1.
Although in this study we chose to employ the DE/rand/1 mutation scheme, it is important to consider that other variants
of the mutation rule have been proposed in literature for standard DE and thus can potentially be employed within the DEc-
DE, see [62]:
DE/best/1: x
0
off
x
e
Fx
r
x
s

DE/cur-to-best/1: x
0
off
x
k
Fx
e
x
k
Fx
r
x
s

DE/best/2: x
0
off
x
e
Fx
r
x
s
Fx
u
x
v

DE/rand/2: x
0
off
x
t
Fx
r
x
s
Fx
u
x
v

DE/rand-to-best/1: x
0
off
x
t
Fx
e
x
t
Fx
r
x
s

DE/rand-to-best/2: x
0
off
x
t
Fx
e
x
t
Fx
r
x
s
Fx
u
x
v

where x
e
is the elite, and x
u
and x
v
are two additional pseudo-randomly selected individuals. Schemes which include also
variable and randomized control parameters have been proposed in e.g. [13,4,75,49,14]. It is worthwhile to mention the rota-
tion invariant mutation shown in [59]:
DE/current-to-rand/1x
off
= x
k
+ K(x
t
x
k
) + F
0
(x
r
x
s
)
where K is the combination coefcient, which, as suggested in [59], should be chosen with a uniform random distribution
from [0, 1] and F
0
= K F. Since this mutation scheme already contains the crossover, the mutated solution does not undergo
the crossover operation described below.
Recently, in [60], a new mutation strategy has been dened. This strategy, namely DE/rand/1/either-or, consists of the
following:
x
0
off

x
t
Fx
r
x
s
if rand0; 1 < p
F
;
x
t
Kx
r
x
s
2x
t
otherwise;

4
where, for a given value of F, the parameter K is set equal to 0.5 (F + 1) and p
F
is a constant value. More advanced mutation
schemes and more generally search moves have been introduced within the DE community. It is important to mention the
2472 F. Neri et al. / Information Sciences 181 (2011) 24692487
neighborhood based mutation scheme proposed in [12], where the offspring generation is performed by means of the com-
position of two contributions, the rst one resulting from the entire population, the second from a subset of it.
The mutated individual x
0
off
and x
e
undergo exponential crossover. A design variable of the provisional offspring x
0
off
i is
randomly selected and copied into the ith design variable of the elite solution x
e
(its copy). This guarantees that elite and
offspring have different genotypes. Subsequently, a set of random numbers between 0 and 1 are generated. As long as
rand(0, 1) 6 Cr, where the crossover rate Cr is a predetermined parameter, the design variables from the provisional offspring
(mutant) are copied into the corresponding positions of the elite x
e
. The rst time that rand(0, 1) > Cr, the copy process is
interrupted. Thus, all the remaining design variables of the offspring are copied from the parent. For the sake of clarity,
the pseudo-code of the exponential crossover is shown in Fig. 1
In other words, this crossover operator generates offspring composed of the elite x
e
and contains, within it, a section of the
chromosome of the mutant vector x
0
off
. According to its original denition, see e.g. [60], for a xed Cr value the exploration
feature of the crossover operator is dependant on the dimensionality of the problem. For example, if Cr is prearranged in a
low dimensional problem, the offspring is composed mainly of the mutant vector, while for a high dimensional problem, the
offspring is mainly composed of the elite. In this study, we propose to slightly modify the denition of exponential crossover
by xing, instead of Cr, the approximate proportion of mutant genes within the offspring. Let us dene this proportion as
a
m
%
n
m
n
; 5
where n
m
is the number of mutant genes we expect to copy into the offspring. In order to achieve that on average n
m
are
copied into the offspring we need to impose that
Cr
nam
0:5: 6
It can easily be seen that, for a chosen a
m
, the crossover rate can be set on the basis of the dimensionality in the following
way:
Cr
1

2
nam
p : 7
The proposed DEcDE, with a prearranged probability M
t
, instead of applying DE/rand/1 mutation and exponential cross-
over, generates the individual by means of Trigonometric Mutation, see [16], consisting of the following:
x
off

x
r
x
s
x
t

3
p
s
p
r
x
r
x
s
p
t
p
s
x
s
x
t
p
r
p
t
x
t
x
r
; 8
where for k = r, s, t,
p
k

jf x
k
j
jf x
r
j jf x
s
j jf x
t
j
: 9
Thus, the trigonometric mutation is a greedy operator that for three given points generates an offspring by exploiting the
most promising search directions. The employment of this operator within DEcDE is supposed to offer an exploitative alter-
native to the standard exploration rule of DE. The trigonometric mutation thus has the role of promoting the generation of
the offspring along (locally) optimal directions. In this sense, this special mutation can be seen as a single step local search
(see [26]).
When the offspring is generated, its tness value is computed and compared with that of the elite individual. The com-
parison allows the denition of winner and loser solutions. The winner solution biases the virtual population by affecting the
PV values. The update rule for l values is given by:
l
t1
l
t

1
N
p
winner loser; 10
Fig. 1. Exponential crossover pseudo-code.
F. Neri et al. / Information Sciences 181 (2011) 24692487 2473
where N
p
is virtual population size. The update rule for r values is given by:
r
t1

2
r
t

2
l
t

2
l
t1

1
N
p
winner
2
loser
2
; 11
where N
p
is a parameter, namely virtual population size. Details for constructing formulas (10) and (11) are given in [44].
The formulas displayed in Eq. (10) and (11) rule the convergence of the virtual population. More specically, the mean
value of the PDF representing the population is moved towards the winner solution while the standard deviation tends to
progressively narrow around the most promising solution, thus resulting in a r value tending toward zero. The latter con-
dition is here indicated as convergence.
The proposed DEcDE introduces extra rules for modifying the PV values. More specically, with a probability M
p
the PV is
perturbed. The mean value l is perturbed according to the following formula:
l
t1
l
t1
2s rand0; 1 s; 12
where s is a weight representing the maximum amplitude of perturbation. Similar to typical DE schemes, a toroidal mech-
anism (see [60]) ensures that l is bounded by 0 and 1 (for example 1 + 0.1 = 0.1). The perturbation rule for the sigma is given
by:
r
t1

2
r
t1

2
s rand0; 1: 13
For the sake of clarity, the pseudo-code summarizing the working principles of DEcDE is shown in Fig. 2.
2.1. Disturbed Exploitation compact Differential Evolution: algorithmic philosophy
An important feature of the proposed MC approach is that, due to its simplicity and memory structure, it can easily be
implemented into hardware characterized by a limited memory. The DEcDE is designed for optimization directly within a
Fig. 2. DEcDE pseudo-code.
2474 F. Neri et al. / Information Sciences 181 (2011) 24692487
micro-controller, for example, a Microchip PIC18 or a Motorola 68HC11. These devices usually have 816 Kb of RAM and a
cost lower than 10 USD. Considering that a signicant portion of the memory and computational power is employed to per-
form the real-time control, the optimization of a population-based algorithm is likely to overow either the computational
power or the available memory resources. In the case of a standard MC approach, composed of an evolutionary framework
and a set of local search algorithms, where the computational overhead requirements are usually higher than a classical
meta-heuristic due to the simultaneous employment of global and local search, the algorithmic overhead is likely to be
prohibitive.
In this light, the proposed DEcDE can be considered as a powerful optimization tool that, despite the limited memory
requirements (only four memory slots, i.e. two for the PV, one for x
e
, and one for the offspring), makes full use of the poten-
tials of MC for solving complex and relatively multivariate optimization problems.
This goal is obtained by a careful and unconventional algorithmic design. More specically, traditional MAs are composed
of an evolutionary framework and one or more local search algorithms. According to a standard algorithmic philosophy, see
e.g. [30,55], the evolutionary framework is supposed to perform the global search while the local search algorithms are sup-
posed to exploit the most promising search directions and thus perform the local search. This way of designing algorithms
can be exported to compact algorithms and lead to some improvements, see [48]. However, the impact on the performance
of an additional local search algorithm in compact frameworks appears to be rather marginal with respect to the impact ob-
served on population-based algorithms. According to our interpretation, this phenomenon is due to the fact that a popula-
tion of solutions by itself contains a certain degree of exploration potential, see [6,51]. On the contrary, a statistical
representation of the population is, due to its nature, a fairly exploitative structure which tends to follow the improvements
of the elite and focus the search around it. In this sense a compact algorithm can be seen from a memetic perspective as a
stochastic local search. For this reason, a standard combination of a compact algorithm and a local search algorithm can re-
sult not an efcient solution for balancing global and local search, see [30,29,71], since the resulting algorithm would fail at
exploring the decision space from complementary perspectives [37].
Thus, the balance between global and local search in the proposed DEcDE algorithm is obtained on the basis of a different
idea. The search logic is based on a fairly explorative DE based mutation structure, DE/rand/1. On the other hand, the expo-
nential crossover with the proposed control on the portion of mutant genes, see a
m
above, makes the moving operator fairly
exploitative. In this study we x a
m
= 0.25 and thus most of the offspring genotype is inherited from the elite. Due to the
structure of the exponential crossover, the offspring generation is equivalent to make some steps in the direction of a set
of neighbor design variables. In this sense the variation operator can be considered fairly exploitative. In addition, the trig-
onometric mutation can be seen as a shallow depth local search algorithm which makes a gradient estimation and attempts
to detect a promising solution by following the gradient direction. The trigonometric mutation can also be considered as an
exploitative operator. The employment of multiple exploitative components locally following gradient estimations can be
punt into relationship with the study on DE functioning reported in [15] where it is mathematically shown that DE based
schemes perform a local search. This multiple shallow local search is counterbalanced by an unconventional global search.
This global search is performed indirectly by perturbing the PV. In other words, DEcDE does not contain an explicit global
search operator. The working principle of the DEcDE can be seen as a local search which is periodically disturbed. The mov-
ing operators of mutations and crossover are supposed to detect promising search directions and quickly exploit them. This
fact corresponds to the convergence of the virtual population towards the elite. This convergence is likely to be premature.
The perturbation mechanism then inhibits the algorithmic convergence and forces the algorithm to search elsewhere in the
decision space, possibly detecting new promising solutions. In other words, DEcDE can be seen as a multi-start local search
algorithm which performs a highly exploitative mini-search between each pair of PV perturbations. However, this mini-
search occurs while the memory of the previously achieved enhancements is kept.
The DEcDE perturbation can be analyzed with the aid of two different perspectives. If we consider a PV as the model of an
actual population, the perturbation is equivalent to a random re-sampling of some new solutions and then replacement of a
portion of the population. It can be observed that the replacement of some individuals of a population which is converging
towards a genotype, modies the mean value of the population and, most likely, produces an increase in the population
diversity (standard deviation in PV). For a complementary perspective, i.e. if we consider the compact algorithm as an opti-
mization search algorithm which samples solutions within a portion of the decision space, the perturbation mechanism cor-
responds to a displacement and broadening of that portion of the decision space where the solutions are sampled.
In a memetic view, DEcDE is the harmonic combination of a local search generated by a DE/rand/1/exp and trigonometric
mutation which aim at exploiting the promising search directions indicated by the elite and a global randomized search pro-
moted by the perturbation mechanism. As a metaphor, two operators perform an in depth learning of a portion of the tness
landscape while one operator has the role of periodically changing the learning perspective, thus offering new intuitions to
the learning operators.
It must be remarked that DEcDE is not a MA, in a classical sense, since it is not composed of an evolutionary framework
and one or more local search algorithms. However, DEcDE is a MC approach according to the denition given [57] since it
combines multiple agents (DE search, trigonometric mutation, and perturbation) which concur towards the solution of a glo-
bal goal.
Fig. 3 gives a graphical representation of the DEcDE functioning. More specically, the behavior of the algorithm before
and after the perturbation is represented. Before the perturbation, the search direction suggested by the elite x
e
with a nar-
F. Neri et al. / Information Sciences 181 (2011) 24692487 2475
rowly distributed virtual population, greedily leads towards a sub-optimal solution. After the perturbation the new center
and shape of the virtual population suggests new promising search directions.
3. Numerical results
The following test problems have been considered in this study.
f
1
Shifted sphere function: f
1
from [70] with n = 30.
f
2
Shifted Schwefels Problem 1.2: f
2
from [70] with n = 30.
f
3
Rosenbrocks function: f
3
from [61] with n = 30.
f
4
Shifted Ackleys function: f
5
from [61] with n = 30.
f
5
Shifted rotated Ackleys function: f
6
from [61] with n = 30.
f
6
Shifted Griewanks function: f
7
from [61] with n = 30.
f
7
Shifted rotated Griewanks function: f
8
from [61] with n = 30.
f
8
Shifted Rastrigins function: f
9
from [70] with n = 30.
f
9
Shifted rotated Rastrigins function: f
10
from [70] with n = 30.
f
10
Shifted non-continuous Rastrigins function: f
11
from [61] with n = 30.
f
11
Schwefels function: f
12
from [61] with n = 30.
f
12
Schwefel Problem 2.22: f
2
from [74] with n = 10.
f
13
Schwefel Problem 2.21: f
4
from [74] with n = 10.
f
14
Generalized penalized function 1: f
12
from [74] with n = 10.
f
15
Generalized penalized function 2: f
13
from [74] with n = 10.
f
16
Schwefels Problem 2.6 with Global Optimum on Bounds: f
5
from [70] with n = 30.
f
17
Shifted rotated Weierstrass function: f
11
from [70] with n = 30.
f
18
Schwefels Problem 2.13: f
12
from [70] with n = 30.
f
19
Kowaliks function: f
15
from [79] with n = 4.
f
20
Six-hump camel-back function: f
20
from [61] with n = 2.
f
21
Branin function: f
17
from [74] with n = 2.
f
22
Hartmans function 1: f
19
from [79] with n = 4.
f
23
Hartmans function 2: f
20
from [79] with n = 6.
f
24
f
26
Shekels family: f
21
f
24
from [79] with n = 4.
f
27
and f
29
Michalewicz function:

n
i1
sinxi sin
xi
2
p

20
in [0, p]
n
with n = 50 and n = 100, respectively.
f
28
and f
30
Schwefel function:

n
i1
ai sin

jaij

in [500, 500]
n
with n = 50 and n = 100, respectively.
All the algorithms in this paper have been run for above-mentioned test problems. Test problems f
19
f
26
are character-
ized by a unique dimensionality value (indicated in the list). Thus, 30 test problems in total contained are in this study. For
each algorithm, 30 independent runs have been performed. The budget of each single run has been xed equal to 5000 n
tness evaluations. The proposed DEcDE has been compared with modern compact and population-based algorithms.
Numerical results are reported in Sections 3.1 and 3.2, respectively. For all the experiments reported in this paper the DEcDE
has been run with N
p
= 300, F = 0.5, a
m
= 0.25, and s = 0.1. The probabilities of activating trigonometric mutation M
t
and per-
turbing the virtual population M
p
have been set equal to 0.003 and 0.001, respectively. It must be remarked that for compact
algorithms the so called generation is actually a single comparison/tness evaluation. Thus, the probabilities indicated
100 80 60 40 20 0 20 40 60 80 100
100
80
60
40
20
0
20
40
60
80
100
100 80 60 40 20 0 20 40 60 80 100
100
80
60
40
20
0
20
40
60
80
100
Xe
Xe
Fig. 3. Functioning of DEcDE, before the perturbation (left) and after the perturbation (right).
2476 F. Neri et al. / Information Sciences 181 (2011) 24692487
above refer to each tness evaluation. This explains why although the numbers appear to be low, the events occur fairly of-
ten during each run. For example, in the case of n = 10 a number of tness evaluations equal to 50,000 is performed. During a
single run the perturbation occurs on average 50 times, which are apparently enough to detect solutions with a high
performance.
3.1. Comparison with modern compact algorithms
The following compact algorithms have been considered for comparison against DEcDE.
compact Genetic Algorithm (cGA) with persistent elitism proposed in [2].
real compact Genetic Algorithm (rcGA) with persistent elitism proposed in [44].
compact Differential Evolution (cDE) proposed in [45] with persistent elitism, DE/rand/1 mutation and binomial cross-
over. The cDE has been run with F = 0.5 and Cr = 0.7.
Memetic compact Differential Evolution (McDE) proposed in [48] with persistent elitism, DE/rand/1 mutation and bino-
mial crossover. The McDE has been run with F = 0.7, Cr = 0.7, probability of local search activation p
ls
= 0.005, reduction
factor b = 0.8, and initial hyper-cube dimension d equal to 10% of the search space.
The virtual population size N
p
for cGA, rcGA, cDE, and McDE has been set equal to 300.
Table 1 shows the average of the nal results detected by each algorithm the corresponding standard deviation values
calculated over the performed 30 runs. The best results are highlighted in bold face. In order to strengthen the statistical
signicance of the results, the Wilcoxon Rank-Sum test has also been applied according to the description given in [76],
where the condence level has been xed to 0.95. Table 2 shows the results of the Wilcoxon test for each version of DEcDE
against the other algorithms considered in this study. A + indicates the case in which DEcDE statistically outperforms, for
the corresponding test problem, its corresponding population-based algorithm; a = indicates that a pairwise comparison
leads to success of the Wilcoxon Rank-Sum test, i.e. the two algorithms have the same performance; a indicates that
DEcDE is outperformed.
Numerical results show that the proposed DEcDE algorithm clearly outperforms cGA and rcGA and also outperforms on
average cDE and McDE on the set of test problems considered in this study. Most importantly, cDE and McDE have a similar
algorithmic structure and therefore tend to have a similar performance. On the contrary, DEcDE is based on a different logic
and, as shown in Section 3.1 and Table 2, appears to offer a good performance for those problems where cDE and McDE fail.
Table 1
Average nal tness values standard deviations for compact algorithms.
Test problem cGA rcGA cDE McDE DEcDE
f
1
1.446e+04 4.63e+03 1.906e+04 9.62e+03 4.520e28 1.74e27 6.526e25 8.34e25 1.681e+01 1.01e+01
f
2
1.628e+06 7.00e+05 2.677e+04 4.78e+03 9.865e+03 2.52e+03 2.322e+03 9.48e+02 1.799e+03 1.79e+03
f
3
2.432e+09 1.62e+09 1.803e+09 2.02e+09 9.898e+01 1.41e+02 1.479e+04 6.89e+04 8.473e+03 6.51e+03
f
4
1.681e+01 9.45e01 1.859e+01 4.15e01 1.074e+01 1.75e+00 1.887e+00 1.70e+00 1.673e+00 4.74e01
f
5
1.721e+01 1.41e+00 1.880e+01 4.54e01 1.028e+01 1.83e+00 3.505e+00 1.29e+00 1.540e+00 4.62e01
f
6
8.840e+02 3.08e+01 2.259e03 4.11e03 1.883e01 2.03e01 7.533e03 1.89e02 6.081e01 2.01e01
f
7
8.778e+02 3.34e+01 3.403e02 9.71e02 1.891e01 2.06e01 2.497e01 2.12e01 7.258e01 1.30e01
f
8
2.265e+02 4.06e+01 2.037e+02 2.74e+01 5.959e+01 1.33e+01 6.586e+01 1.36e+01 1.651e+01 3.31e+00
f
9
3.013e+02 4.72e+01 1.985e+02 3.06e+01 1.219e+02 2.58e+01 1.210e+02 2.61e+01 2.005e+02 2.09e+01
f
10
1.307e+05 4.96e+04 2.900e+03 3.07e+03 6.448e+03 2.75e+03 6.984e+03 3.39e+03 1.648e+03 3.71e+02
f
11
4.947e+03 7.03e+02 3.156e+03 7.54e+02 9.972e+02 3.25e+02 1.090e+03 3.02e+02 1.626e+02 4.32e+01
f
12
5.614e+00 2.91e+00 4.127e+00 4.90e+00 2.558e02 7.10e03 8.208e01 1.74e01 4.402e01 1.19e01
f
13
3.309e+01 1.13e+01 1.000e+02 5.06e09 1.000e+02 1.73e06 1.000e+02 1.21e05 9.962e+01 1.45e01
f
14
4.472e+04 1.37e+05 1.401e+00 1.91e+00 1.982e04 1.74e04 1.468e02 2.57e02 3.921e02 2.10e02
f
15
1.121e+06 3.66e+06 7.869e01 8.94e01 1.148e+00 1.67e03 1.047e+00 7.35e02 9.811e01 1.25e01
f
16
1.172e+04 2.57e+03 8.975e+03 2.38e+03 8.023e+03 3.42e+03 7.600e+03 2.53e+03 3.044e+03 1.91e+03
f
17
1.307e+02 2.03e+00 1.226e+02 3.21e+00 1.242e+02 3.21e+00 1.223e+02 4.11e+00 1.302e+02 1.46e+00
f
18
2.958e+05 1.05e+05 3.089e+05 1.38e+05 5.480e+04 3.21e+04 3.661e+03 6.25e+03 8.517e+04 1.40e+04
f
19
5.296e02 9.80e09 5.296e02 5.22e18 5.296e02 3.28e11 5.296e02 4.63e10 5.296e02 9.40e09
f
20
9.632e01 4.22e02 1.067e+00 4.09e16 1.067e+00 1.50e05 1.067e+00 2.82e05 1.067e+00 1.13e05
f
21
2.337e+01 1.14e+00 3.979e01 9.70e13 3.979e01 1.71e05 3.979e01 1.45e05 3.979e01 1.44e05
f
22
3.760e+00 1.95e02 3.863e+00 1.94e15 3.863e+00 1.21e06 3.863e+00 5.89e06 3.863e+00 6.59e06
f
23
4.819e01 4.27e02 3.238e+00 5.53e02 3.288e+00 5.54e02 3.288e+00 5.53e02 3.283e+00 4.61e02
f
24
2.773e+00 1.13e+00 6.458e+00 2.47e+00 5.451e+00 3.24e+00 4.927e+00 2.94e+00 9.955e+00 1.66e+00
f
25
2.628e+00 9.63e01 7.258e+00 3.01e+00 5.504e+00 3.33e+00 5.738e+00 3.18e+00 9.891e+00 1.96e+00
f
26
2.764e+00 9.00e01 6.940e+00 3.19e+00 6.239e+00 3.75e+00 4.912e+00 3.07e+00 9.916e+00 1.53e+00
f
27
1.905e+01 1.95e+00 1.553e+01 2.57e+00 3.296e+01 1.61e+00 3.224e+01 1.75e+00 3.648e+01 9.97e01
f
28
8.546e+03 9.24e+02 7.718e+03 1.36e+03 3.712e+03 7.37e+02 4.437e+03 7.25e+02 3.365e+02 9.17e+01
f
29
2.516e+01 2.39e+00 2.171e+01 3.00e+00 4.581e+01 3.93e+00 4.462e+01 2.97e+00 6.709e+01 1.81e+00
f
30
2.151e+04 1.37e+03 2.157e+04 1.86e+03 1.549e+04 1.49e+03 1.590e+04 1.33e+03 8.363e+02 1.24e+02
F. Neri et al. / Information Sciences 181 (2011) 24692487 2477
In the spirit of the NFLT, DEcDE can be an efcient alternative to optimize those problems which are tough for cDE and McDE.
A special attention should be paid to highly dimensional problems (n = 50 and n = 100). In these cases DEcDE appears to be
very promising with respect to other compact algorithms.
Fig. 4 shows average performance trends of the ve considered compact algorithms over a selection of the test problems
considered in this study.
3.2. Comparison with modern population-based algorithms
The following population-based algorithms have been considered for comparison against DEcDE.
Estimation of Distribution Algorithm with MultiVariate Gaussian model (EDA
mvg
) proposed in [81]. EDA
mvg
has been run
with learning rate a = 0.2, population size N
p
= 50, selection ratio s = 0.3, and maximum amplication value Q = 1.5.
Real Coded Memetic Algorithm (RCMA) proposed in [46]. RCMA has been run with population size N
p
= 50, crossover
parameter a = 0.5, mutation probability 0.125, maximum generation number T = 10,000, b = 5, maximum number of indi-
viduals taking part to the negative assortative mating N
nam
= 3, roulette wheel selection. The other xed parameters have
been set as suggested in [46].
Differential Evolution with adaptive hill-climb Simplex Crossover (DEahcSPX) proposed in [54]. DEahcSPX has been run
with N
p
= 50, scale factor F = 0.9 and crossover rate Cr = 0.9 as suggested in the paper, the number of points involved in the
hill-climb n
p
= 3, factor = 1.
Particle Swarm Optimizer (PSO) proposed in [33]. PSO has been run with 5 particles, an inertia equal to 0.9 and acceler-
ation constant equal to 2.
Harmony Search (HS) proposed in [20]. HS has been run with harmony memory size hms = 5, harmony memory consid-
ering rate hmcr = 0.9, pitch adjusting rate par = 0.2, and fret width fw = 0.4.
It must be remarked that EDA
mvg
has been chosen because it is an EDA and therefore belongs to the same category of
DEcDE; RCMA and DEahcSPX have been chosen because they are MC approaches, see [49,14].
Numerical results in terms of average nal values and Wilcoxon test are shown in Tables 3 and 4, respectively.
Numerical results show that the proposed DEcDE algorithm, albeit much less expensive than the population-based algo-
rithms considered in this study, displays a respectable performance. More specically, results in Table 4 show that DEcDE
performs better than EDA
mvg
and DEahcSPX for the set of problems taken into account. Regarding the comparison with RCMA
and HS, numerical results show that DEcDE performs comparably to it. Comparison with PSO shows that DEcDE is, on aver-
Table 2
Wilcoxon Rank-Sum test for compact algorithms.
Test problem cGA rcGA cDE McDE
f
1
+ +
f
2
+ + + +
f
3
+ + + +
f
4
+ + + +
f
5
+ + + +
f
6
+
f
7
+
f
8
+ + + +
f
9
+ =
f
10
+ + + +
f
11
+ + + +
f
12
+ + + +
f
13
+ = = =
f
14
+ + =
f
15
+ + =
f
16
+ + + +
f
17
= = = =
f
18
+ + =
f
19
= = = =
f
20
+ = = =
f
21
+ = = =
f
22
+ = = =
f
23
+ + = =
f
24
+ + + +
f
25
+ + + +
f
26
+ + + +
f
27
+ + + +
f
28
+ + + +
f
29
+ + + +
f
30
+ + + +
2478 F. Neri et al. / Information Sciences 181 (2011) 24692487
age, slightly outperformed but still displays a respectable performance. This result appears promising for those problems
characterized by a limited memory capacity where a reasonably good performance must be obtained despite the hardware
limitations. An important nding is that the proposed DEcDE is competitive with the population based algorithms also for
the relatively high-dimensional problems (with n = 50 and n = 100).
Fig. 5 shows average performance trends of the four considered algorithms over a selection of the test problems consid-
ered in this study.
4. Case of study: space robotic arm control
Space robots are playing a constantly growing important role in current space operations. The main reason for their suc-
cess is that they can prevent humans from having to perform extremely risky operations, e.g. extra-vehicular activities such
0 5 10 15
x 10
4
0
2000
4000
6000
8000
10000
12000
Fitness function call
F
i
t
n
e
s
s

v
a
l
u
e
cGA
rcGA
cDE
McDE
DEcDE
(a) f
11
0 5 10 15
x 10
4
0
1
2
3
4
5
6
7
x 10
4
Fitness function call
F
i
t
n
e
s
s

v
a
l
u
e
cGA
rcGA
cDE
McDE
DEcDE
(b) f
12
Fig. 4. Performance trends of the compact algorithms.
F. Neri et al. / Information Sciences 181 (2011) 24692487 2479
Table 3
Average nal tness values standard deviations for population-based algorithms.
Test problem EDA
mvg
RCMA DEahcSPX PSO HS DEcDE
f
1
6.96e+01 1.47e+02 3.41e16 6.85e16 9.69e+01 1.50e+01 2.05e14 4.15e14 6.67e01 3.64e01 1.68e+01 1.01e+01
f
2
3.14e+02 3.01e+02 1.25e13 2.95e13 1.80e+03 2.36e+02 7.60e+02 3.26e+02 1.35e+03 7.15e+02 1.80e+03 1.79e+03
f
3
2.32e+06 5.55e+06 2.83e+01 1.28e01 1.11e+05 3.11e+04 1.23e+02 2.35e+02 2.87e+02 1.40e+02 8.47e+03 6.51e+03
f
4
3.27e+00 4.71e01 7.01e09 1.04e08 2.57e+00 4.64e01 5.75e08 6.38e08 3.55e01 1.60e01 3.67e+00 4.74e01
f
5
3.41e+00 1.07e+00 8.12e09 8.45e09 2.51e+00 4.09e01 2.53e+00 6.29e01 4.73e01 2.20e01 3.54e+00 4.62e01
f
6
2.64e+02 3.90e+01 6.58e+00 3.85e+00 4.89e+00 3.60e01 1.51e+00 9.47e01 3.96e01 2.08e01 6.08e01 2.01e01
f
7
2.73e+02 4.50e+01 6.06e+00 3.50e+00 4.82e+00 3.50e01 1.90e+00 9.58e01 3.49e01 2.18e01 7.26e01 1.30e01
f
8
1.77e+02 1.35e+01 7.11e15 2.55e14 3.40e+01 3.48e+00 1.47e+01 5.06e+00 3.57e+00 2.61e+00 1.65e+01 3.31e+00
f
9
1.82e+02 1.26e+01 4.74e15 2.32e14 1.17e+02 2.56e+01 1.60e+02 4.69e+01 1.54e+02 3.74e+01 2.00e+02 2.09e+01
f
10
1.40e+04 5.35e+04 1.00e+04 1.78e+04 3.03e+03 3.75e+02 4.25e+04 2.06e+04 4.42e+02 1.02e+02 1.65e+03 3.71e+02
f
11
1.01e+04 4.23e+02 2.99e+03 5.98e+02 8.26e+02 8.20e+01 1.93e+03 3.45e+02 8.89e+03 1.04e+02 1.63e+02 4.32e+01
f
12
1.29e+00 6.90e01 6.33e10 2.63e09 1.33e01 3.16e02 7.64e24 1.49e23 4.25e04 2.77e04 4.40e01 1.19e01
f
13
9.48e+01 2.21e+01 6.84e+01 1.17e+01 8.56e+01 1.39e+00 1.00e+02 0.00e+00 3.05e+01 1.50e+02 9.96e+01 1.45e01
f
14
5.93e01 1.34e+00 1.03e06 1.29e06 1.16e02 8.74e03 4.71e32 5.59e48 3.18e+00 8.26e01 3.92e02 2.10e02
f
15
1.29e01 1.90e+00 1.15e+00 3.70e03 8.54e01 1.08e01 1.15e+00 2.04e02 3.03e01 3.77e01 8.51e01 1.25e01
f
16
1.01e+04 3.36e+03 9.32e+03 2.45e+03 9.31e+03 1.02e+03 5.80e+03 1.36e+03 1.67e+04 3.26e+03 3.04e+03 1.91e+03
f
17
1.30e+02 7.51e01 1.26e+02 3.65e+00 1.30e+02 1.15e+00 1.29e+02 1.50e+00 1.30e+02 1.21e+00 1.30e+02 1.46e+00
f
18
5.27e+05 2.93e+05 1.30e+05 4.67e+04 7.20e+04 1.03e+04 5.90e+05 1.89e+05 1.10e+05 3.68e+04 8.52e+04 1.40e+04
f
19
5.30e02 6.61e09 5.30e02 4.34e18 5.30e02 3.64e09 5.30e02 3.24e16 5.30e02 1.09e08 5.30e02 9.40e09
f
20
1.07e+00 4.54e16 1.07e+00 6.64e06 1.07e+00 2.16e05 1.07e+00 4.54e16 1.06e+00 1.89e02 1.07e+00 1.13e05
f
21
3.99e01 3.42e03 3.98e01 1.95e04 3.98e01 1.97e04 3.98e01 0.00e+00 3.98e01 3.03e04 3.98e01 1.44e05
f
22
3.85e+00 3.37e02 3.86e+00 5.91e06 3.86e+00 5.56e06 3.86e+00 2.16e15 3.86e+00 1.97e04 3.86e+00 6.59e06
f
23
3.13e+00 1.65e01 3.27e+00 6.07e02 3.32e+00 1.33e04 3.27e+00 6.00e02 3.24e+00 5.73e02 3.28e+00 4.61e02
f
24
5.63e+00 3.50e+00 5.30e+00 3.20e+00 1.00e+01 1.19e01 7.71e+00 3.30e+00 8.42e+00 2.43e+00 7.56e+00 1.66e+00
f
25
9.25e+00 2.42e+00 4.65e+00 3.68e+00 1.02e+01 1.06e01 1.01e+01 1.56e+00 8.58e+00 2.80e+00 7.49e+00 1.96e+00
f
26
8.45e+00 3.01e+00 5.48e+00 3.33e+00 1.04e+01 1.93e01 1.05e+01 3.51e15 8.51e+00 2.86e+00 7.62e+00 1.53e+00
f
27
1.065e+01 7.65e01 2.165e+01 2.22e+00 3.177e+01 5.93e01 4.512e+01 1.04e+00 3.559e+01 1.03e+00 3.648e+01 9.97e01
f
28
1.757e+04 4.49e+02 9.098e+03 8.02e+02 4.160e+03 2.50e+02 4.804e+03 6.66e+02 1.479e+04 6.18e+00 3.365e+02 9.17e+01
f
29
1.269e+01 4.31e01 2.659e+01 1.96e+00 4.405e+01 9.15e01 7.419e+01 1.18e+01 5.061e+01 1.74e+00 6.709e+01 1.81e+00
f
30
3.688e+04 7.90e+02 2.309e+04 1.45e+03 1.571e+04 3.10e+02 1.552e+04 2.02e+03 2.959e+04 4.02e+00 8.363e+02 1.24e+02
2
4
8
0
F
.
N
e
r
i
e
t
a
l
.
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
1
8
1
(
2
0
1
1
)
2
4
6
9

2
4
8
7
as reparations outside the spacecraft. Due to the enormous distances, the robot cannot be fully remotely controlled manually
from the Earth because the communication delay between the command and the execution of the robot operation can likely
be unacceptable in several cases. For this reason an efcient real-time control system is crucially important for this class of
applications.
The absence of gravity plays an important role in the dynamics of the robot and must be taken into account when the
control system is designed. In this case of study, a robotic arm connected to a base, e.g. a spacecraft or a satellite, is consid-
ered. In a nutshell, the control system aims to perform the robot movements in order to minimize the disturbances, i.e. iner-
tial movements, on the base. More specically, each new trajectory step is optimized online using a look-ahead optimized
algorithm for trajectory planning, see [64].
Space robots are highly nonlinear, coupled multi-body systems with nonlinear constraints. Moreover, the dynamic cou-
pling between the manipulator (robotic arm) and the base usually affects the performance of the manipulator. The dynamic
coupling is important to understand the relationship between the robot joint motion and the resultant base motion, and it is
useful in minimizing fuel consumption for base attitude control. The measure of dynamic coupling has been formulated in
[78].
Let us consider a manipulator composed of n
b
links (bodies) interconnected by joints and connected by means of an exter-
nal joint to the base. With reference to Fig. 6, let V
i
and X
i
be linear and angular velocities of the ith body of the manipulator
arm with respect to the absolute reference system R
I
, and let v
i
and x
i
be linear and angular velocities of the ith body of the
manipulator arm with respect to the base R
B
.
Thus, we can obtain that the velocities of the ith body are:
V
i
v
i
V
0
X
0
r
i
;
X
i
x
i
X
0
;
14
where the operator represents the outer product of R
3
vectors. V
0
and X
0
are, respectively, linear and angular velocities of
the centroid of the base with respect to R
I
. The variable r
i
represents the position vector related to the ith body pointing to-
wards the center of the base with reference to R
B
, see [78] for details. The velocities in the base of the coordinates of the
reference system R
B
can be calculated as:
v
i
w
i

J
i
q
_
q; 15
Table 4
Wilcoxon Rank-Sum test for population-based algorithms.
Test problem EDA
mvg
RCMA DEahcSPX PSO HS
f
1
= +
f
2
= =
f
3
+ +
f
4

f
5
= =
f
6
+ + + +
f
7
+ + + +
f
8
+ + =
f
9

f
10
= + + +
f
11
+ + + +
f
12
+
f
13
= + + = +
f
14
+ +
f
15
= = +
f
16
+ + + + +
f
17
= = =
f
18
+ + + +
f
19
= = = = =
f
20
= + =
f
21
= + + +
f
22
+ = = +
f
23
+ = = +
f
24
+ + = =
f
25
+ =
f
26
= + =
f
27
+ + + +
f
28
+ + + +
f
29
+ + + +
f
30
+ + + +
F. Neri et al. / Information Sciences 181 (2011) 24692487 2481
where q is the vector of the angular positions of each joint of the manipulator arm (see q
1
, q
2
, and q
3
in Fig. 6) and J
i
(q) is thus
the Jacobian of the ith body of manipulator arm. The Jacobian can be then decomposed into two sub-matrices related to its
translational and rotational movements, see e.g. [78]
J
i
q
J
Ti
q
J
Ri
q

: 16
The total linear (P) and angular (L) momenta of the entire robotic arm can be expressed as:
P m
0
V
0

n
b
i1
m
i
v
i
17
and
L I
0
X
0
m
0
R
B
V
0

n
b
i1
I
i
x
i
m
i
r
i
v
i
; 18
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10
4
100
80
60
40
20
0
20
40
60
80
Fitness function call
F
i
t
n
e
s
s

v
a
l
u
e
s
DEahcSPX
DEcDE
EDA
mvg
RCMA
PSO
HS
(a) f
13
0 5 10 15
x 10
4
0
0.5
1
1.5
2
2.5
3
3.5
4
x 10
4
Fitness function call
F
i
t
n
e
s
s

v
a
l
u
e
s
DEahcSPX
DEcDE
EDA
mvg
RCMA
PSO
HS
(b) f
16
Fig. 5. Performance trends of DEcDE and population-based algorithms.
2482 F. Neri et al. / Information Sciences 181 (2011) 24692487
Fig. 6. Space robotic arm scheme.
Table 5
Average nal tness values standard deviations and Wilcoxon test for the space robot application.
rcGA McDE DEcDE
18.861 0.5518 + 10.969 0.2887 + 4.971 0.1985
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
0
50
100
150
Fitness function call
F
i
t
n
e
s
s

v
a
l
u
e
rcGA
DECDE
McDE
Fig. 7. Performance trend of compact algorithms for space robot control.
F. Neri et al. / Information Sciences 181 (2011) 24692487 2483
where I
i
and m
i
are the inertia momentum and mass of each body composing the robot manipulator, and R
B
is a positioning
vector pointing towards the centroid of the base with reference to R
I
.
Eqs. (17) and (18) can then be combined:
P
L

H
B
V
0
X
0

H
m
_
q: 19
The details about the structures of the matrices H
B
and H
m
are given in [28]. In a free-oating situation (due to the fact that
both robot and base are in outer space), there are no external forces or momenta. If we consider the gravitational force to be
negligible, linear and angular momenta are conserved. We assume that the initial state of the systemis stationary, so that the
total linear and angular momenta are zero. Hence, from Eq. (19), the mapping relationship between the manipulator joint
motion and the base motion is thus given by:
V
0
X
0

H
1
m
H
B
_
q: 20
184.1 184.2 184.3 184.4 184.5 184.6 184.7 184.8
200
150
100
50
0
50
100
150
200
Time [s]
A
c
c
e
l
e
r
a
t
i
o
n

[
m
/
s
2
]
A
c
c
e
l
e
r
a
t
i
o
n

[
m
/
s
2
]
X
1
X
2
X
3
(a) without optimization (beginning of learning period)
4568.4 4568.5 4568.6 4568.7 4568.8 4568.9 4569 4569.1 4569.2
10
5
0
5
10
15
20
Time [s]
X
1
X
2
X
3
(b) with optimization (end of learning period)
Fig. 8. Acceleration at the base without and with control system.
2484 F. Neri et al. / Information Sciences 181 (2011) 24692487
For a given trajectory that must be followed by the robot, the optimization problem under investigation consists of
detecting the set angular positions q
i
, angular velocities
_
q
i
, and angular accelerations

q
i
of each joint i in each knot k iden-
tifying the trajectory such that the disturbance on the base is minimized. The tness to be minimized is, in our study, the
integral over time of the norm of the acceleration vector on the base. The acceleration values can be derived by Eq. (20).
Since the trajectory must be continuous, the function describing the position, over time, of the joints must also be con-
tinuous. In order to satisfy this condition, we modeled each function q
i
(t) (where t is the time) as a set of 5th polynomial
splines and imposed the continuity of q
i
t;
_
q
i
t, and

q
i
t. Each spline is a polynomial of the 5th order because six condi-
tions are imposed by the physics of the phenomenon (continuity of the function, its rst and second derivative in the knots).
With reference to Fig. 6, considering that the robot manipulator contains three joints, the trajectory is marked by two knots,
and for each joint it is necessary to control angular position, velocity and acceleration, our tness function depends on
3 2 3 = 18 variables.
In order to minimize the tness function, the DEcDE algorithm has been implemented and compared with rcGA and
McDE. The cGA has been discarded from this experiment since, for this class of applications, rcGA proved to be superior
to cGA, see [44]. The cDE has been discarded because it has the same structure as McDE except for the fact that McDE em-
ploys a local search algorithm. Thus cDE and McDE tend to display a similar performance but McDE appears to be, on a reg-
ular basis, more promising that cDE, see [48]. The algorithms have been run with the same parameter setting specied in
Section 3. Twenty-ve independent runs have been performed by each algorithm for 10,000 tness evaluations.
The nal results, in terms of average nal value and standard deviation are given in Table 5. In addition, Table 5 displays
the statistical signicance calculated by means of the Wilcoxon Rank-Sum test.
Fig. 7 displays the average performance trend of the algorithms involved in this real-world problem.
Numerical results clearly show the superiority of DEcDE with respect to the other modern compact algorithms.
In order to clearly show the physical meaning of this real-world application as well as the importance of performing an
efcient optimization, the three spatial components of acceleration in the base (during the movement of the robotic arm)
without optimization (i.e. at the beginning of the on-line learning) and with DEcDE optimization (at the end of the on-line
learning) are shown in Fig. 8. It can be seen how the control system optimized by the DEcDE algorithm leads to a relevant
reduction of the disturbances, and that the maximum amplitude of peaks is reduced by 30 times in the optimized system.
Finally, Fig. 9 shows the variation, during the movement along the trajectory, of the angles q
1
, q
2
, and q
3
in the three joints
of the optimized system. It is interesting to observe how, in order to minimize the disturbances on the base, the angles tend
to get arranged as to have a small instantaneous algebraic sum.
5. Conclusion
This paper proposes a novel memetic implementation for compact algorithms. The proposed DEcDE algorithm employs a
compact structure, moving operators composed of two exploitative versions of DE and a randomization of the virtual pop-
ulation. This randomization can be seen as an alternative search strategy that cooperates, during the evolution, with the
exploitative logic of the moving operators. In other words, DEcDE performs an highly exploitative search which is periodi-
cally disturbed by a randomized action which perturbs the virtual population. The proposed algorithm displayed a good
performance with respect to modern compact algorithms, recently proposed in literature, on a diverse set of test problems.
This fact shows that the proposed MC approach is very promising, for compact algorithms, with respect to traditional MAs
employing an evolutionary framework and a set of local search algorithms. Our proposal is based on the consideration that a
compact algorithm, by itself, shares important similarities with a local search algorithm, and thus an algorithmic design
4568.3 4568.4 4568.5 4568.6 4568.7 4568.8 4568.9 4569 4569.1 4569.2 4569.3
1
0.5
0
0.5
1
1.5
Time [s]
J
o
i
n
t

P
o
s
i
t
i
o
n

[
r
a
d
]
q
1
q
2
q
3
Fig. 9. Position of the joints after the optimization.
F. Neri et al. / Information Sciences 181 (2011) 24692487 2485
which considers the compact algorithm as a global search algorithm can be misleading. DEcDE has also been compared with
modern population-based algorithms, two of them being MC approaches. The proposed algorithm, despite its compact nat-
ure, and thus a much more modest memory requirement, displayed a respectable performance on the set of test problems
considered in this study, proving thus to be competitive with complex and memory-demanding algorithms. The proposed
approach has been tested on a real-world optimization problem: the on-line control of a robot manipulator for space oper-
ations in order to minimize the disturbance on the base, this base being a spacecraft or a satellite. Numerical results show
that the proposed algorithm has a better performance compared to other compact algorithms in terms of both nal detected
solutions and convergence speed.
Acknowledgments
This research is supported by the Academy of Finland, Akatemiatutkija 130600, Algorithmic Design Issues in Memetic
Computing and Tutkijatohtori 140487, Algorithmic Design and Software Implementation: a Novel Optimization Platform.
This research is also supported by Tekes the Finnish Funding Agency for Technology and Innovation, Grant 40214/08
(Dynergia).
References
[1] C.W. Ahn, J. An, J.-C. Yoo, Estimation of particle swarm distribution algorithms: combining the benets of PSO and EDAs, Information Sciences, in press.
[2] C.W. Ahn, R.S. Ramakrishna, Elitism based compact genetic algorithms, IEEE Transactions on Evolutionary Computation 7 (4) (2003) 367385.
[3] C. Aporntewan, P. Chongstitvatana, A hardware implementation of the compact genetic algorithm, in: Proceedings of the IEEE Congress on
Evolutionary Computation, vol. 1, 2001, pp. 624629.
[4] J. Brest, S. Greiner, B. Bokovic , M. Mernik, V. Z

umer, Self-adapting control parameters in differential evolution: a comparative study on numerical

benchmark problems, IEEE Transactions on Evolutionary Computation 10 (6) (2006) 646657.
[5] E.K. Burke, G. Kendall, E. Soubeiga, A tabu search hyperheuristic for timetabling and rostering, Journal of Heuristics 9 (6) (2003) 451470.
[6] A. Caponio, G.L. Cascella, F. Neri, N. Salvatore, M. Sumner, A fast adaptive memetic algorithm for on-line and off-line control design of pmsm drives,
IEEE Transactions on System Man and Cybernetics Part B, special issue on Memetic Algorithms 37 (1) (2007) 2841.
[7] A. Caponio, F. Neri, V. Tirronen, Super-t control adaptation in memetic differential evolution frameworks, Soft Computing A Fusion of Foundations,
Methodologies and Applications 13 (8) (2009) 811831.
[8] W.J. Cody, Rational Chebyshev approximations for the error function 23 (107) (1969) 631637.
[9] P. Cowling, G. Kendall, E. Soubeiga, A hyperheuristic approach to scheduling a sales summit, in: Proceedings of the Third International Conference on
Practice and Theory of Automated Timetabling, Lecture Notes in Computer Science, vol. 2079, Springer, 2000, pp. 176190.
[10] F. Cupertino, E. Mininno, D. Naso, Elitist compact genetic algorithms for induction motor self-tuning control, in: Proceedings of the IEEE Congress on
Evolutionary Computation, 2006.
[11] F. Cupertino, E. Mininno, D. Naso, Compact genetic algorithms for the optimization of induction motor cascaded control, in: Proceedings of the IEEE
International Conference on Electric Machines and Drives, vol. 1, 2007, pp. 8287.
[12] S. Das, A. Abraham, U.K. Chakraborty, A. Konar, Differential evolution with a neighborhood-based mutation operator, IEEE Transactions on
Evolutionary Computation 13 (2009) 526553.
[13] S. Das, A. Konar, U.K. Chakraborty, Two improved differential evolution schemes for faster global search, in: Proceedings of the 2005 Conference on
Genetic and Evolutionary Computation, ACM, 2005, pp. 991998.
[14] S. Das, P.N. Suganthan, Differential evolution a survey of the state-of-the-art, IEEE Transactions on Evolutionary Computation, in press.
[15] S. Dasgupta, S. Das, A. Biswas, A. Abraham, On stability and convergence of the population-dynamics in differential evolution, AI Communications
The European Journal on Articial Intelligence 22 (1) (2009) 120.
[16] H.-Y. Fan, J. Lampinen, A trigonometric mutation operation to differential evolution, Journal of Global Optimization 27 (1) (2003) 105129.
[17] J.C. Gallagher, S. Vigraham, A modied compact genetic algorithm for the intrinsic evolution of continuous time recurrent neural networks, in:
Proceedings of the Genetic and Evolutionary Computation Conference, 2002, pp. 163170.
[18] J.C. Gallagher, S. Vigraham, G. Kramer, A family of compact genetic algorithms for intrinsic evolvable hardware, IEEE Transactions Evolutionary
Computation 8 (2) (2004) 111126.
[19] W. Gautschi, Error function and fresnel integrals, in: M. Abramowitz, I.A. Stegun (Eds.), Handbook of Mathematical Functions with Formulas, Graphs,
and Mathematical Tables, Dover Publications, Inc., New York, 1972, pp. 297309. Chapter 7.
[20] Z.W. Geem, J.H. Kim, G.V. Loganathan, A new heuristic optimization algorithm: harmony search, Simulation 76 (2) (2001) 6068.
[21] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley Publishing Co., Reading, MA, USA, 1989.
[22] M. Gong, L. Jiao, L. Zhang, Baldwinian learning in clonal selection algorithm for optimization, Information Sciences 180 (8) (2010) 12181236.
[23] G. Harik, Linkage learning via probabilistic modeling in the ECGA, Tech. Rep. 99010, University of Illinois at Urbana-Champaign, Urbana, IL, 1999.
[24] G.R. Harik, F.G. Lobo, D.E. Goldberg, The compact genetic algorithm, IEEE Transactions on Evolutionary Computation 3 (4) (1999) 287297.
[25] G.R. Harik, F.G. Lobo, K. Sastry, Linkage learning via probabilistic modeling in the extended compact genetic algorithm (ECGA), in: M. Pelikan, K. Sastry,
E. Cant-Paz (Eds.), Scalable Optimization via Probabilistic Modeling, Studies in Computational intelligence, vol. 33, Springer, 2006, pp. 3961.
[26] W.E. Hart, N. Krasnogor, J.E. Smith, Memetic evolutionary algorithms, in: W.E. Hart, N. Krasnogor, J.E. Smith (Eds.), Recent Advances in Memetic
Algorithms, Springer, Berlin, Germany, 2004, pp. 327.
[27] J. Hidalgo, M. Prieto, J. Lanchares, R. Baraglia, F. Tirado, O. Garnica, Hybrid parallelization of a compact genetic algorithm, in: Proceeding of the Eleventh
Euromicro Conference on Parallel, Distributed and Network-Based Processing, 2003, pp. 449445.
[28] P. Huang, K. Chen, S. Xu, Optimal path planning for minimizing disturbance of space robot, in: Proceedings of the IEEE International Conference on
Control, Automation, Robotics, and Vision, 2006.
[29] H. Ishibuchi, Y. Hitotsuyanagi, Y. Nojima, An empirical study on the specication of the local search application probability in multiobjective memetic
algorithms, in: Proceedings of the IEEE Congress on Evolutionary Computation, September 2007, pp. 27882795.
[30] H. Ishibuchi, T. Yoshida, T. Murata, Balance between genetic search and local search in memetic algorithms for multiobjective permutation ow shop
scheduling, IEEE Transactions on Evolutionary Computation 7 (2003) 204223.
[31] Y. Jewajinda, P. Chongstitvatana, Cellular compact genetic algorithm for evolvable hardware, in: Proceedings of the International Conference on
Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology, vol. 1, 2008, pp. 14.
[32] G. Kendall, P. Cowling, E. Soubeiga, Choice function and random hyperheuristics, in: Proceedings of the Fourth AsiaPacic Conference on Simulated
Evolution and Learning, 2002, pp. 667671.
[33] J. Kennedy, R.C. Eberhart, Particle swarm optimization, in: Proceedings of the IEEE International Conference on Neural Networks, 1995, pp. 19421948.
[34] A.V. Kononova, D.B. Ingham, M. Pourkashanian, Simple scheduled memetic algorithm for inverse problems in higher dimensions: application to
chemical kinetics, in: Proceedings of the IEEE World Congress on Computational Intelligence, 2008, pp. 39063913.
2486 F. Neri et al. / Information Sciences 181 (2011) 24692487
[35] P. Koroec, J. ilc, B. Filipic, The differential ant-stigmergy algorithm. Information Sciences, in press.
[36] N. Krasnogor, Studies in the theory and design space of memetic algorithms, Ph.D. Thesis, University of West England, 2002.
[37] N. Krasnogor, Toward robust memetic algorithms, in: W.E. Hart, N. Krasnogor, J.E. Smith (Eds.), Recent Advances in Memetic Algorithms. Studies in
Fuzzines and Soft Computing, Springer, Berlin, Germany, 2004, pp. 185207.
[38] N. Krasnogor, B. Blackburne, E. Burke, J. Hirst, Multimeme algorithms for protein structure prediction, in: Proceeding of Parallel Problem Solving in
Nature VII, Lecture Notes in Computer Science, Springer-Verlag, 2002.
[39] N. Krasnogor, J. Smith, A tutorial for competent memetic algorithms: model, taxonomy, and design issues, IEEE Transactions on Evolutionary
Computation 9 (2005) 474488.
[40] P. Larraaga, J.A. Lozano, Estimation of Distribution Algorithms: A New Tool for Evolutionary Computation, Kluwer, 2001.
[41] M.N. Le, Y.S. Ong, Y. Jin, B. Sendhoff, Lamarckian memetic algorithms: local optimum and connectivity structure analysis, Memetic Computing Journal
1 (3) (2009) 175190.
[42] F.G. Lobo, C.F. Lima, H. Mrtires, An architecture for massive parallelization of the compact genetic algorithm, Lecture Notes in Computer Science, vol.
3103, Springer, 2004, pp. 412413.
[43] R. Meuth, M.H. Lim, Y.S. Ong, D.C. Wunsch II, A proposition on memes and meta-memes in computing for higher-order learning, Memetic Computing
Journal 1 (2) (2009) 85100.
[44] E. Mininno, F. Cupertino, D. Naso, Real-valued compact genetic algorithms for embedded microcontroller optimization, IEEE Transactions on
Evolutionary Computation 12 (2) (2008) 203219.
[45] E. Mininno, F. Neri, F. Cupertino, D. Naso, Compact differential evolution, IEEE Transactions on Evolutionary Computation, in press.
[46] D. Molina, F. Herrera, M. Lozano, Adaptive local search parameters for real-coded memetic algorithm, in: Proceedings of the IEEE Congress on
Evolutionary Computation, 2005, pp. 888895.
[47] P. Moscato, M. Norman, A competitive and cooperative approach to complex combinatorial search, Tech. Rep. 790, 1989.
[48] F. Neri, E. Mininno, Memetic compact differential evolution for Cartesian robot control, IEEE Computational Intelligence Magazine 5 (2) (2010) 5465.
[49] F. Neri, V. Tirronen, Recent advances in differential evolution: a review and experimental analysis, Articial Intelligence Review 33 (12) (2010) 61
106.
[50] F. Neri, V. Tirronen, T. Krkkinen, T. Rossi, Fitness diversity based adaptation in multimeme algorithms: a comparative study, in: Proceedings of the
IEEE Congress on Evolutionary Computation, 2007, pp. 23742381.
[51] F. Neri, J. Toivanen, G.L. Cascella, Y.S. Ong, An adaptive multimeme algorithm for designing HIV multidrug therapies, IEEE/ACM Transactions on
Computational Biology and Bioinformatics 4 (2) (2007) 264278.
[52] F. Neri, J. Toivanen, R.A.E. Mkinen, An adaptive evolutionary algorithm with intelligent mutation local searchers for designing multidrug therapies for
HIV, Applied Intelligence 27 (3) (2007) 219235.
[53] Q.C. Nguyen, Y.S. Ong, M.H. Lim, A probabilistic memetic framework, IEEE Transactions on Evolutionary Computation 13 (3) (2009) 604623.
[54] N. Noman, H. Iba, Accelerating differential evolution using an adaptive local search, IEEE Transactions on Evolutionary Computation 12 (1) (2008) 107
125.
[55] Y.S. Ong, A.J. Keane, Meta-Lamarkian learning in memetic algorithms, IEEE Transactions on Evolutionary Computation 8 (2) (2004) 99110.
[56] Y.-S. Ong, M.-H. Lim, X. Chen, Research frontier: towards memetic computing, Tech. Rep. C2i-1209, School of Computer Engineering, Nanyang
Technological University, Singapore, 2009.
[57] Y.-S. Ong, M.-H. Lim, X. Chen, Memetic computation past, present and future, IEEE Computational Intelligence Magazine 5 (2) (2010) 2431.
[58] Y.S. Ong, M.H. Lim, N. Zhu, K.W. Wong, Classication of adaptive memetic algorithms: a comparative study, IEEE Transactions on Systems, Man and
Cybernetics Part B 36 (1) (2006) 141152.
[59] K.V. Price, Mechanical engineering design optimization by differential evolution, in: D. Corne, M. Dorigo, F. Glover (Eds.), New Ideas in Optimization,
McGraw-Hill, 1999, pp. 293298.
[60] K.V. Price, R. Storn, J. Lampinen, Differential Evolution: A Practical Approach to Global Optimization, Springer, 2005.
[61] A.K. Qin, V.L. Huang, P.N. Suganthan, Differential evolution algorithm with strategy adaptation for global numerical optimization, IEEE Transactions on
Evolutionary Computation 13 (2009) 398417.
[62] A.K. Qin, P.N. Suganthan, Self-adaptive differential evolution algorithm for numerical optimization, in: Proceedings of the IEEE Congress on
Evolutionary Computation, vol. 2, pp. 17851791.
[63] R. Rastegar, A. Hariri, A step forward in studying the compact genetic algorithm, Evolutionary Computation 14 (3) (2006) 277289.
[64] K. Ren, J.Z. Fu, Z.C. Chen, A new linear interpolation method with lookahead for high speed machining, in: Technology and Innovation Conference,
2006, pp. 10561059.
[65] G. Rudolph, Self-adaptive mutations may lead to premature convergence, IEEE Transactions on Evolutionary Computation 5 (4) (2001) 410414.
[66] K. Sastry, D.E. Goldberg, On extended compact genetic algorithm, Tech. Rep. 2000026, University of Illinois at Urbana-Champaign, Urbana, IL, 2000.
[67] K. Sastry, D.E. Goldberg, D.D. Johnson, Scalability of a hybrid extended compact genetic algorithm for ground state optimization of clusters, Materials
and Manufacturing Processes 22 (5) (2007) 570576.
[68] K. Sastry, G. Xiao, Cluster optimization using extended compact genetic algorithm, Tech. Rep. 2001016, University of Illinois at Urbana-Champaign,
Urbana, IL, 2001.
[69] J.E. Smith, Coevolving memetic algorithms: a review and progress report, IEEE Transactions on Systems, Man, and Cybernetics Part B 37 (1) (2007) 6
17.
[70] P.N. Suganthan, N. Hansen, J.J. Liang, K. Deb, Y.-P. Chen, A. Auger, S. Tiwari, Problem denitions and evaluation criteria for the CEC 2005 special session
on real-parameter optimization, Tech. Rep. 2005005, Nanyang Technological University, KanGAL, Singapore and IIT Kanpur, India, 2005.
[71] K. Tan, S. Chiam, A. Mamun, C. Goh, Balancing exploration and exploitation with adaptive variation for evolutionary multi-objective optimization,
European Journal of Operational Research 197 (2009) 701713.
[72] J. Tang, M.H. Lim, Y.S. Ong, Diversity-adaptive parallel memetic algorithm for solving large scale combinatorial optimization problems, Soft Computing
A Fusion of Foundations, Methodologies and Applications 11 (9) (2007) 873888.
[73] V. Tirronen, F. Neri, T. Krkkinen, K. Majava, T. Rossi, An enhanced memetic differential evolution in lter design for defect detection in paper
production, Evolutionary Computation 16 (2008) 529555.
[74] J. Vesterstrm, R. Thomsen, A comparative study of differential evolution, particle swarm optimization and evolutionary algorithms on numerical
benchmark problems, in: Proceedings of the IEEE Congress on Evolutionary Computation, vol. 3, 2004, pp. 19801987.
[75] M. Weber, F. Neri, V. Tirronen, A study on scale factor in distributed differential evolution, Information Sciences 181 (12) (2011) 24882511.
[76] F. Wilcoxon, Individual comparisons by ranking methods, Biometrics Bulletin 1 (6) (1945) 8083.
[77] D. Wolpert, W. Macready, No free lunch theorems for optimization, IEEE Transactions on Evolutionary Computation 1 (1) (1997) 6782.
[78] Y. Xu, The measure of dynamic coupling of space robot system, in: Proceedings of the IEEE Conference on Robotics and Automation, 1993, pp. 615620.
[79] X. Yao, Y. Liu, G. Lin, Evolutionary programming made faster, IEEE Transactions on Evolutionary Computation 3 (2) (1999) 82102.
[80] E.L. Yu, P.N. Suganthan, Ensemble of niching algorithms, Information Sciences 180 (15) (2010) 28152833.
[81] B. Yuan, M. Gallagher, Experimental results for the special session on real-parameter optimization at CEC 2005: a simple, continuous EDA, in:
Proceedings of the IEEE Conference of Evolutionary Computation, 2005, pp. 17921799.
[82] Q. Yuan, F. Qian, W. Du, A hybrid genetic algorithm with the baldwin effect, Information Sciences 180 (5) (2010) 640652.
F. Neri et al. / Information Sciences 181 (2011) 24692487 2487