You are on page 1of 31

Adaptive Memetic Differential Evolution with Global and Local

neighborhood-based mutation operators


Adam P. Piotrowski

Institute of Geophysics, Polish Academy of Sciences, Ks. Janusza 64, 01-452 Warsaw, Poland
a r t i c l e i n f o
Article history:
Received 14 November 2012
Received in revised form 18 March 2013
Accepted 31 March 2013
Available online 10 April 2013
Keywords:
Differential Evolution
Evolutionary algorithm
Memetic algorithm
Memetic computing
Self-adaptation
Global numerical optimization
a b s t r a c t
Differential Evolution (DE) is one of the most popular optimization methods for real-valued
problems and a large number of its variants have been proposed so far. However, bringing
together different ideas that already led to successful DE versions is rare in the literature. In
the present paper a novel DE algorithm is proposed, in which three among the most ef-
cient concepts already applied separately within DE framework are gathered together,
namely: (1) the adaptation of algorithm control parameters and probabilities of using dif-
ferent mutation strategies; (2) the use of NelderMead algorithm as a local search method
hybridized with DE; and (3) the splitting mutation into Global and Local models, when
Local mutation model is based on the concept of neighborhood of individuals organized
on a ring topology. The performance of the novel algorithm, called Adaptive Memetic DE
with Global and Local neighborhood-based mutation operators is compared with 13 differ-
ent DE variants on a set of 25 popular problems which include rotated, shifted and hybrid
composition functions. It is found that, although none DE algorithm outperforms all the
others for the majority of problems, on average the proposed approach perform better than
all 13 DE algorithms selected for comparison.
The proposed algorithm is another heuristic approach developed to solve optimization
problems. The question may arise, whether proposing novel methods is useful as No Free
Lunch theorems for optimization state that the expected performance of all possible heu-
ristics on all possible problems is equal. In the last section of the paper the limitations and
implications of No Free Lunch theorems are discussed based on rich, but unfortunately fre-
quently neglected literature. A very simple continuous and differentiable minimization
problem is proposed, for which it is empirically veried that each among considered 14
DE algorithms perform poorer than random sampling. It is also empirically shown that
when all considered DE algorithms search for the maximum of the proposed problem, they
found lower minimum than DE algorithms searching for the minimum. Such result is not
unexpected according to No Free Lunch theorems and should be considered as a precaution
from generalization of good performance of heuristic optimization methods.
2013 Elsevier Inc. All rights reserved.
1. Introduction
Differential Evolution (DE) [123,124,113] is one of the most popular Evolutionary Algorithms (EAs) for solving continuous
single-objective optimization problems, due to its simplicity and reasonable performance. Without the loss of generality, the
minimization problems are considered in this study. The goal is to nd the global optimum vector x

such that
0020-0255/$ - see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.ins.2013.03.060

Tel.: +48 22 6915 858; fax: +48 22 6915 915.


E-mail address: adampp@igf.edu.pl
Information Sciences 241 (2013) 164194
Contents lists available at SciVerse ScienceDirect
Information Sciences
j our nal homepage: www. el sevi er . com/ l ocat e/ i ns
f (x
+
) = min
xX
f (x) (1)
where f(x) is the real-valued function, x is a D-dimensional vector, x = {x
1
, . . . , x
D
} with the domain X # R
D
. DE has been
applied to a number of engineering [23,121,84], environmental [111,89], nancial [158,56], medical [61,26,97] and various
other problems [113,3]. Similarly to other EAs, DE is a population-based method. In the original DE version, proposed in
[123,124], to create an offspring (child), rst a parent (target) vector is perturbed by weighted difference between two ran-
domly chosen individuals producing a donor vector (this operation is called mutation), and then a crossover between donor
and parent vectors is performed. Finally the tness function of the child is evaluated and the better individual within the
parent-offspring pair is promoted to the next generation (what is called greedy selection). Although the original DE
[123,124] is frequently claimed to be successful in solving a number of optimization problems, it also has a number of draw-
backs, which motivate the researchers to develop various novel DE variants. During last 15 years a large number of DE algo-
rithms have been proposed, its detailed survey may be found in [113,94,31].
Initial variant of DE relied on three user-dependent control parameters: scale factor (F), crossover (CR), and population
size (NP). Their proper selection for particular problem turned out to be a difcult task. It soon appeared that various com-
binations of control parameter values may be advantageous for different problems [48,113,31] and their universal setting
cannot be suggested. For example lower CR is better suited for separable problems, higher for rotated ones; larger F helps
in exploration of multi-modal problems, lower in exploitation of uni-modal ones. Large NP increases the number of possible
moves and helps in exploration, but signicantly slows down the convergence.
In the original DE only a single mutation strategy is used, what together with the xed values of control parameters limit
the number of possible moves and hence the search ability of the algorithm [94,144], frequently leading to stagnation
[76,46,31]. The development of a number of novel mutation strategies [113] has led to signicant improvement in optimi-
zation accuracy and speed, but it again turned out that its choice should depend on the problem. Some strategies are better
suited for exploitation, others help in exploration, but at additional computational cost. The user was hence faced with the
need to choose not only the correct set of DE control parameters, but also the proper mutation strategy; in other words the
need to optimize the optimizer rst emerged. Moreover, in the case of some problems the choices should be modied dur-
ing the search to balance the exploration and the exploitation capabilities when algorithm approaches the optimum. To alle-
viate the ancient problem who will guard the guards [122] the well known within other EA frameworks [39,40] adaptive
and self-adaptive approaches become widely used within DE framework, leading to a large number of novel algorithmic vari-
ants [80,114,100,13,14,156,115,42,52,55,135,50,16,157,121,11,105,85,67]. Among them probably two algorithms, jDE [13]
and SADE [115] gained largest attention so far.
Apart from the concept of self-adaptation, a number of other ways to improve the original DE algorithm have been pro-
posed. One of the most frequently used is the idea of Memetic Computing (MC) [74,103,22,63], recently very widely under-
stood [104,96] as a broad subject that studies the computing structures composed of interacting memes, or operators.
According to a modern denition, the term MC refers rather to the implementation philosophy that considers algorithms
as evolving structures composed of competitive, but also cooperative elements (memes) [104]. A number of classes of
EAs may be included within a broad MC concept. If different mutation strategies are understood as a memes that interact
to generate improved solutions, many self-adaptive DE algorithms may also belong to MC (see also discussion in [63]). How-
ever, within a concept of MC, a different class of methods called Memetic Algorithms (MA), that are classically understood as
a composition of global search heuristics and local search procedures, is more widely known. Memetic DE algorithms, a com-
bination of DE method that perform global search with a local search algorithm, form a sub-class of MA. Memetic DE algo-
rithms are widely proposed in the literature [98,132,92,18,78,93,69]. The aim of the local search procedure is to quicken
exploitation of local optima without loosing global exploration capabilities of the algorithm. As another class within MC
framework may be understood the concept of hybridization of DE with other EAs or Swarm Intelligence methods. DE has
been successfully hybridized with Particle Swarm Optimization [59,28,90,101,81,44,137] (for a review see [152]), Bacterial
Foraging Algorithm [106,136], Ant Colony Optimization [25,21], Harmony Search [20,140], Biogeography-Based Optimiza-
tion [53,12], and Covariance Matrix Approximation-Evolution Strategy (CMA-ES) [51]. The hybridization of different biolog-
ically-based concepts leads to algorithms that are less prone to stagnation or premature convergence. However, when
proposing hybrid methods the attention must be focused on possible hampering of the behavior of constituent algorithms
(see discussion in [44] and references given in [96], on the signicance of the proper cooperation of algorithms see also [86]).
If the constituent methods are not properly selected, the hybrid approach may perform poorer than each of its basic algo-
rithms. Another, much different concept of merging DE with other global search methods is to create a meta-algorithm
[138,107], which manage the time budget among constituent methods and allow migration of solutions between them,
but does not affect signicantly their behavior. It is important to underline that although meta-algorithms, self-adaptive
DE, Memetic DE and hybrid DE methods belong to broad philosophy of MC, such classes of DE methods differ signicantly
in algorithmic structure and operation.
The next widely used idea of DE improvement aims at distributed computing [127]. A number of Distributed DE algo-
rithms have been proposed to date [128,6,142,143,108,36,144,109,110,33], which mainly enhance DE exploration capabili-
ties and prevent premature convergence. The concepts of clustering [17,82], smart sampling [34] and orthogonal crossover
[141] have also been applied within DE framework. Other ideas include opposition-based DE algorithms [116], where oppo-
sition numbers are used to improve different stages of DE and the very successful DE with Global and Local neighborhood-
based mutation operators (DEGL) [29], in which the so called DE/target-to-best/ mutation strategy (see Section 2) that suffer
A.P. Piotrowski / Information Sciences 241 (2013) 164194 165
from premature convergence was improved by using two different mutation models to balance its exploration and exploi-
tation capabilities. Within DEGL the Global mutation model allows for the quick spread of information among individuals
and the Local mutation model utilize the concept of ring topology such that for each individual only information from its
closest neighborhood is available. This slows down the convergence and improves exploration capabilities of the algorithm.
DEGL quickly became one of the most popular DE variants [31]. Recently, a number of novel DE methods have also been pro-
posed for constrained [9,62,155,88,68] and multi-modal optimization problems (for a survey see [32]).
Unfortunately, different ways of improving DE algorithms are rarely gathered together into a single method. For example,
the vast majority of Memetic DE algorithms use only a single mutation strategy and do not utilize the concept of adaptation.
Most hybrid algorithms use only a simple version of each hybridized method (a number of notable exceptions have recently
been proposed in [44]). Adaptive approaches usually do not benet from the concepts of distributed computing, clustering,
opposition numbers or topological neighborhood. Although the idea of combining the most successful components of differ-
ent EAs is sometimes criticized as adding stuff (see [63], nonetheless a very inspiring paper based on the concept of Ock-
hams razor, which state that plurality must never be posited without necessity, and [19], where the discussion is
extended), alluring start from tabula rasa means shaking off the experience, what is rarely the best evolutionary choice.
Although it is easy to agree that one should not do with more things that which can be done with fewer, building on
the experience of the predecessors and merging already successful behaviors is often promoted by the evolution (for some
recent discussion whether evolutionary inspirations are important for function optimization, see [2]). Such approach turned
also successful in some recently proposed DE algorithmic structures. For example Wang et al. [139] proposed a DE algorithm
which randomly uses one of three pools of F and CR control parameters and one of three pools of mutation strategies, when
the pools are solely based on the successful ndings fromthe literature. Gong et al. [55] showed that merging self-adaptation
rules of updating F and CR control parameters, an external archive borrowed from JADE [156] and a concept of using a num-
ber of DE mutation strategies selected according to adaptive pursuit method [131] may help improving the performance of
DE methods.
In the present paper the novel algorithm is proposed which bring together the benets of different ideas already explored
separately within DE framework, namely: (1) the Global and Local neighborhood-based mutation operators proposed in
DEGL [29], (2) the adaptation of control parameters and mutation strategies in a way similar to the one proposed in one
of the most successful and popular adaptive DE variant (SADE, [115]) and (3) the idea of Memetic DE algorithms, realized
by incorporation of NelderMead algorithm [91,75], which is probably the most popular Direct Search method [72], within
DE framework following the Super-t Memetic DE (SFMDE) [18]. In the extensive experiments performed on 25 test prob-
lems proposed for IEEE Congress on Evolutionary Computation in 2005 (CEC 2005) [126] it is shown that the novel algorithm,
called Adaptive Memetic DEGL (AM-DEGL), outperform not only the classical SADE, DEGL and SFMDE approaches, from
which it inherits its constituent elements, but also a large number of recently proposed, conceptually much different DE
algorithms.
Although this seems encouraging, the question may arise whether a novel DE algorithms are worth attention since the
publication of No Free Lunch theorems for optimization (NFL) [149,150,120,73,65,66,151,147], which, although frequently
neglected in the literature [73], are of great importance to the understanding of limitations of applicability of heuristics.
Hence, the last section of the paper is devoted to this bitter pill that must be swallowed by the author of the present study,
sweetened by the fact that it is probably a part of a lunch of anyone who develop novel heuristic optimization methods.
Although it is now known [120,66] that the NFL does not hold for subsets of problems that are composed only of problems
that may be of practical interest to anyone, the precautions stated in [150] cannot simply be omitted and the lack of general
applicability of any heuristic methods should be emphasized. As an example in the last section of the paper a very simple
minimization problem is shown for which, what should not be a surprise according to [150], both random sampling and
algorithms that aim at maximization perform better than any DE algorithm used for minimization. It is also shown empir-
ically that at least for one among popular CEC 2005 problems some widely used DE methods perform poorer than random
sampling.
The rest of the paper is organized as follows. The classical DE algorithm is introduced in Section 2 and the related work is
discussed in Section 3. The novel AM-DEGL algorithm is dened in Section 4. Section 5 contains experimental comparison of
the performance of AM-DEGL and 13 DE algorithms from the literature, followed by inter-comparison of a number of AM-
DEGL variants. In Section 6 the implications of NFL are discussed and a simple problem is proposed for which all DE methods
perform poorer than random sampling; nally Section 7 concludes the paper.
2. Classical Differential Evolution algorithm
The original DE algorithm was proposed in [123,124] for single-objective unconstrained optimization problems. In the
successive generations g, DE evolves the population of NP candidate solutions, called also individuals
x
i;g
= x
1
i;g
; . . . ; x
D
i;g
_ _
; i = 1; NP, to nd the global optimum of f within the subset

D
j=1
[L
j
; U
j
[ in the search space R
D
. Although
various initialization methods may be used by population-based algorithms, in DE individuals are usually randomly initial-
ized at g = 0 within the subset of the search space from the uniform distribution
x
j
i;0
= L
j
rand
j
i
(0; 1) (U
j
L
j
); j = 1; . . . ; D; i = 1; . . . ; NP (2)
166 A.P. Piotrowski / Information Sciences 241 (2013) 164194
where rand
j
i
(0; 1) generates a random value within [0, 1] interval for jth element of ith individual.
In subsequent generations each individual (x
i,g
), considered a parent, creates an offspring (u
i,g
) in two steps: rstly a donor
vector (v
i,g
) is created by means of mutation, then crossover between a donor and a parent vectors produces an offspring. A
greedy selection between a parent and an offspring is performed and only the better one is promoted to the next generation.
In the basic DE the so called DE/rand/1 mutation scheme with a scaling control parameter F was proposed DE/rand/1
v
i;g
= x
r1;g
F (x
r2;g
x
r3;g
) (3)
Later a number of other mutation strategies have been developed, including DE/best/1
v
i;g
= x
best;g
F (x
r1;g
x
r2;g
) (4)
DE/current-to-best/1 (or DE/target-to-best/1)
v
i;g
= x
i;g
F (x
best;g
x
i;g
) F (x
r1;g
x
r2;g
) (5)
DE/best/2
v
i;g
= x
best;g
F (x
r1;g
x
r2;g
) F (x
r3;g
x
r4;g
) (6)
DE/rand-to-best/2 (or DE/target-to-best/2)
v
i;g
= x
i;g
F (x
best;g
x
i;g
) F (x
r1;g
x
r2;g
) F (x
r3;g
x
r4;g
) (7)
DE/rand/2
v
i;g
= x
r1;g
F (x
r2;g
x
r3;g
) F (x
r4;g
x
r5;g
) (8)
where r1, r2, r3, r4 and r5 are randomly selected integers from the range [1, NP], such that r1 r2 r3 r4 r5 i, and
x
best,g
represents the best individual in the current population at generation g. In the literature also trigonometric mutation
[45], random localization mutation [71], generalized differentials mutation [5], Gaussian parentcentric blend crossover
(PBX) mutation [36], proximity-based mutation [43], ranking-based mutation [54] and self-balancing mutation [41] schemes
have been proposed.
In the basic DE and most novel DE variants, after mutation a binomial crossover between a donor and a parent vector is
performed to produce an offspring (u
i,g
), what requires specication of the value of CR control parameter:
u
j
i;g
=
v
j
i;g
if rand
j
i
(0; 1) 6 CR or j = j
rand;i
x
j
i;g
otherwise
_
_
_
(9)
Selected CR should be within [0, 1] interval. In Eq. (9) j
rand,i
is a randomly selected integer from [1, D] range, used to guarantee
that an offspring inherits al least one element from a donor vector. In the literature also different crossover schemes have
been suggested, but are rarely used in practice for examples and the detailed discussion on crossover see
[113,4,153,79,141,159]. However, some mutation strategies have also been proposed to generate offspring directly, without
crossover, for example DE/current-to-rand/1
u
i;g
= x
i;g
K (x
r1;g
x
i;g
) F (x
r2;g
x
r3;g
) (10)
in which K is another control parameter, according to [112,115] randomly generated from [0, 1] range for each individual.
If u
i,g
is located outside the decision space, one of constraint handling methods must be applied [83]. Simple rebounding
approach is adopted in the present paper:
u
j
i;g
=
2 L
j
u
j
i;g
if u
j
i;g
< L
j
2 U
j
u
j
i;g
if u
j
i;g
> U
j
_
_
_
(11)
and is repeated until L
j
6 u
j
i;g
6 U
j
. Finally, in the basic DE and almost all other DE variants the greedy selection between off-
spring and parent is performed:
x
i;g1
=
u
i;g
if f (u
i;g
) 6 f (x
i;g
)
x
i;g
otherwise
_
(12)
The algorithm continues the search until the stopping criteria are met, that are usually dened by setting the maximum
number of allowed function calls.
3. Previous work on adaptive, self-adaptive, Memetic and topological neighborhood-based Differential Evolution
Since the Storn and Price papers [123,124] DE has gained a lot of attention among researchers and large number of its
improved variants have been proposed. Very brief review of DE methods was given in Section 1, for the detailed survey
A.P. Piotrowski / Information Sciences 241 (2013) 164194 167
the reader is referred to [113,94,31]. In [31] also the milestone state-of-the-art DE algorithms are reported. In the present
section only the selected concepts on which the constituent components of AM-DEGL are based are briey reviewed.
3.1. Adaptive and self-adaptive Differential Evolution
Initially the selection of DE control parameter values and the choice of proper mutation strategy for particular problem
were the main challenge for the researchers. However, soon it turned out that it is not possible to suggest the single best DE
parameters setting for variety of problems and it is advantageous to modify dynamically the choices during the search.
Hence in the recent years concepts of adaptation and self-adaptation, that are widely used by other EAs [39,40], has been
applied to DE. Adaptive approach uses feedback from the search process to modify the control parameters, when self-adap-
tive methods encode the control parameters in the DE genotype [39,55]. However, most of the proposed adaptive or self-
adaptive DE algorithms do not adapt all control parameters together with probabilities of using different mutation strategies.
For example, the population size NP, apart from a few proposals [130,15,129,133,16,160] is usually kept constant during the
search. The adaptation or self-adaptation of other DE control parameters may be performed in very different ways.
A number of adaptive methods of DE control parameters have been proposed. In [80] fuzzy rules are used to adapt F and
CR parameters. In [156,67] the CR and F parameters are generated from normal and Cauchy distributions with xed scale
parameter (not to be confused with F), but adapted location one. All successful values of F and CR have similar impact on
the adaptation of location parameter, without considering the rate of improvement that was achieved when they were used.
In SADE [115] a similar way of adaptation of CR and the probabilities of using four mutation strategies (dened in Eqs. (3),
(7), (8) and (10)) is suggested, but the scale parameter F is generated from a normal distribution with standard deviation and
mean xed to 0.3 and 0.5, respectively. According to the philosophy proposed by [115], the strategies that reached higher
ratio between the number of successful and the number of all generated offspring during the recent generations have higher
chance to be used in the future. Gong et al. [55] proposed Adaptive Strategy Selection-based JADE (AdapSS-JADE) algorithm,
in which CR, F and probability of using each of four mutation strategies are adopted. However, in AdapSS-JADE the proba-
bilities of using each mutation strategy depend on the relative tness improvements achieved by successful offspring gen-
erated recently by each strategy. Moreover, the best strategy is given much larger probability of being used in the next
generation than the others, according to best takes (almost) all rule. In [11] an adaptation scheme of control parameters
is introduced, that depends on both elapsed time and tness values of different individuals within the population. Other ver-
sions of adaptive approach have also been proposed in [52,50].
Self-adaptive approaches were introduced to DE in [1], where only the value of CR parameter was self-adaptively ad-
justed. Somehow contrary, in [100] only the F parameter was self-adapted. Brest et al. [13] used self-adapting approach
to adjust both F and CR parameters and further showed its superiority over older adaptive methods in [14]. Pan et al.
[105] proposed self-adaptation of F, CR and probability of using each of four mutation strategies. Mallipeddi et al. [85] pro-
posed a different self-adaptive approach called ensemble of mutation strategies and control parameters (EPSDE), in which a
pool of control parameters and mutation strategies was created and the elements from the pool which more frequently pro-
vide successful offspring increase own probability of being used in the next generations.
3.2. Memetic Differential Evolution algorithms
Memetic DE algorithms combine global search DE method with some local search procedures. To propose a novel Memet-
ic DE algorithm, a number of important questions must be addressed: which local search methods to choose, which individ-
uals within DE population should follow the local search, how frequently local search should be commenced and how many
tness calls may it use. Different answers have been suggested in the literature. Algorithm proposed in [98] used simple local
hill descending technique with simplex crossover operator to speed up convergence of only the best individual in the pop-
ulation, the rest search was done by classical DE. In [132] three local search methods, namely Hooke Jeeves Algorithm, Sto-
chastic Local Search and Simulated Annealing are adaptively coordinated with DE. In [78] a random walk with direction
exploitation local search method is applied to each DE individual with pre-dened xed probability; the number of function
calls used by each local search is also xed. In [69] it was proposed to combine DE with chaotic local search in such a way
that the scale of chaotic movements is shrinking when the number of function calls is increasing and the local search, com-
menced for the best individual only, may last no longer than pre-dened number of iterations. In [95,64,87] a novel type of
compact Memetic DE algorithms for memory-limited applications has been proposed, which uses statistical model of pop-
ulation instead of individuals. Caponio et al. [18] proposed to use two local search methods, NelderMead simplex (NMA)
[91,75] and Rosenbrock algorithm (RA) [117], both commenced with different probabilities that depend on tness of indi-
viduals within the population. RA may be applied only to the best individual in population, NMA to any individual. As
NMA requires D + 1 points in the search space, additional D-points are randomly chosen from the current population. Each
run of local searcher is limited to a xed value of 500 function calls.
3.3. Topological neighborhood-based Differential Evolution
The DEGL algorithm proposed in [29] improves DE/target-to-best/1 mutation scheme (see Eq. (5)) by introducing Global
and Local neighborhood-based mutation models. DEGL approach has been applied to a number of practical problems
168 A.P. Piotrowski / Information Sciences 241 (2013) 164194
[30,8,154,134,111] and was considered as one of the most important DE variants [31]. In DEGL initially all individuals are
randomly indexed and organized on a ring topology. Individuals with consecutive indices are called the closest neighbors,
i.e. 1st individual is the closest neighbor of the 2nd and the NPth one. The Global mutation model is exactly the same as
DE/current-to-best/1 strategy (Eq. (5)). However, in the Local mutation model instead of x
best,g
the best individual within
the pre-dened neighborhood of a parent vector is used (x
Lbest,i,g
) and both random indices required in DE/current-to-
best/1 strategy are also selected from the individuals within this neighborhood. Das et al. [29] empirically assessed the
neighborhood to 10% of the population, i.e. for NP = 100, only the individuals with indices {i + 5, . . . , i + 1, i 1, . . . , i 5} be-
long to the neighborhood of parent x
i
. Within the mutation, the weighted sum of Global and Local models is performed to
produce donor vectors, when both weights sum to the unity and none is lower than 0.05. The self-adapting method of
weights adjustment was proposed in [29].
In DEGL, F and CR parameters are not adapted and both Global and Local models are based on a single DE/current-to-best/
1 mutation strategy. Results obtained by SADE [115], SspDE [105] and AdapSS-JADE [55] conrm that it is advantageous to
use a number of mutation strategies and adjust the probabilities of their application together with the adaptation of DE con-
trol parameters. On the other hand, it was shown [98,132,18,69] that the proper hybridization of DE with local searcher may
signicantly improve the algorithm performance. Hence, bringing together the advantages of DEGL, adaptive and Memetic
approaches seems a tempting possibility, realized in a novel AM-DEGL algorithm, which is described in details in the next
section. The adaptation method based on SADE is chosen, due to its good performance, wide popularity and the fact that
SADE uses four mutation strategies that are very different from each other (what is not the case in e.g. AdapSS-JADE). Among
local search methods the NMA [91,75] has been chosen due to three reasons: (1) it is probably the most popular of Direct
Search methods and turned out efcient in huge number of applications [72]; (2) it was already successfully hybridized with
DE in [18]; and (3) although used as local searcher in this paper, it also has exploration capabilities and is able to leave the
basin of attraction.
4. Adaptive Memetic Differential Evolution with Global and Local neighborhoodbased mutation operators
In the following sub-sections all steps of AM-DEGL algorithm are described and motivated. The pseudocode of AM-DEGL
algorithm is shown in Fig. 1. For completeness, the pseudocode of NMA procedure, used by AM-DEGL as local search method,
is shown in Fig. 2.
4.1. Population size and initialization
The choice of population size (NP) is always a difcult task. Although there are a few methods that allow for modication
of DE population size during run [130,15,142,133,129,16,160], they are rarely applied in practice. In AM-DEGL the population
size is kept constant, as in overwhelming majority of DE algorithms. However, the proposed values of NP for different DE
methods vary from NP < D [142] to NP = 10D [124]. A number of values in between have also been suggested
[113,48,156]. The Authors of SADE left the choice of NP to the user [115]. In DEGL the large population size (NP = 10D)
was proposed [29], enabling high diversity of available steps that may be generated by Local mutation model and slow ex-
change of information between individuals. Contrary, in a number of Memetic DE algorithms low population size is sug-
gested [98,69], as large population diminish the importance of local search method, especially if it is applied to the best
individual only. To reconcile different needs of ancestral algorithms NP = 5D is suggested for AM-DEGL. Such value is also
within most of recommended NP limits found in the literature (e.g. [113,139]).
In AM-DEGL, like in classical DE, individuals x
i,0
, i = 1, . . . , NP are randomly initialized within the search bounds according
to Eq. (2).
4.2. Global and Local neighborhood-based mutation models
AM-DEGL benets from the concept of splitting the mutation into two parts, namely Global and Local mutation models,
which is the core idea of DEGL algorithm [29]. In the case of Global mutation models (GM
i,g
) the information from the whole
population may be available to the parent x
i
. In the case of Local mutation models (LM
i,g
) only the information from a pre-
dened number of neighboring individuals is available, when individuals are organized on a ring topology. Each individual
x
i,g
in the population is indexed with i = 1, . . . , NP and the indexes i are xed for the whole run of the algorithm. Individuals
with consecutive indexes are neighbors, for example four nearest neighbors of x
1,g
are x
NP1,g
, x
NP,g
, x
2,g
and x
3,g
.
In AM-DEGL, ve different pairs of Global and Local mutation models are proposed. Four of them are based on strategies
used by SADE (Eqs. (3), (7), (8) and (10)), the last one has been used in original DEGL method (Eq. (5)):
GM
i;g
= x
r1;g
F
i;g
(x
r2;g
x
r3;g
)
LM
i;g
= x
Lr1;i;g
F
i;g
(x
Lr2;i;g
x
Lr3;i;g
)
(13)
GM
i;g
= x
i;g
F
i;g
(x
best;g
x
i;g
) F
i;g
(x
r1;g
x
r2;g
) F
i;g
(x
r3;g
x
r4;g
)
LM
i;g
= x
i;g
F
i;g
(x
Lbest;i;g
x
i;g
) F
i;g
(x
Lr1;i;g
x
Lr2;i;g
) F
i;g
(x
Lr3;i;g
x
Lr4;i;g
)
(14)
A.P. Piotrowski / Information Sciences 241 (2013) 164194 169
GM
i;g
= x
r1;g
F
i;g
(x
r2;g
x
r3;g
) F
i;g
(x
r4;g
x
r5;g
)
LM
i;g
= x
Lr1;i;g
F
i;g
(x
Lr2;i;g
x
Lr3;i;g
) F
i;g
(x
Lr4;i;g
x
Lr5;i;g
)
(15)
GM
i;g
= x
i;g
K
i;g
(x
r1;g
x
i;g
) F
i;g
(x
r2;g
x
r3;g
)
LM
i;g
= x
i;g
K
i;g
(x
Lr1;i;g
x
i;g
) F
i;g
(x
Lr2;i;g
x
Lr3;i;g
)
(16)
GM
i;g
= x
i;g
F
i;g
(x
best;g
x
i;g
) F
i;g
(x
r1;g
x
r2;g
)
LM
i;g
= x
i;g
F
i;g
(x
Lbest;i;g
x
i;g
) F
i;g
(x
Lr1;i;g
x
Lr2;i;g
)
(17)
Fig. 1. Pseudocode of AM-DEGL algorithm.
170 A.P. Piotrowski / Information Sciences 241 (2013) 164194
where x
Lbest,i,g
is the best individual among the p% closest neighbors of the parent vector x
i,g
, and Lr1, Lr2, Lr3, Lr4 and Lr5 are
randomly selected mutually distinct indices of individuals which are among p% nearest neighbors of x
i,g
. In [29] the neigh-
borhood size p% = 10% was suggested. However, in AM-DEGL the larger neighborhood (20%) turned out more efcient (vari-
ants of AM-DEGL with various p% values and the variant of AM-DEGL where LM
i,g
is not used by strategies borrowed from
SADE are tested in Section 5.4). The scale factor F
i,g
and parameter K
i,g
are indexed with i and g to mark that they are gen-
erated separately for each individual at every generation. Following [112] K
i,g
is randomly generated from [0, 1] range, the
choice of F
i,g
is dened in Section 4.4.1.
Each individual in every generation randomly chooses one of ve available pairs of Global and Local mutation models,
probability of choosing each strategy (each pair) is adapted during the search. It is important to note that as DE/current-
to-rand/2 strategy (Eq. (10)) does not follow the crossover, when GM
i,g
and LM
i,g
are created by Eq. (16) the crossover
(see Section 4.3) is not performed.
Fig. 1. (continued)
A.P. Piotrowski / Information Sciences 241 (2013) 164194 171
The Global and Local mutation models are then coupled by weighted factor w
/
i;g
(which, following discussion in [29] is
self-adapted for each individual, see Section 4.4.3), creating either donor vector (v
i,g
) when one of strategies dened by
Eqs. (13)(15) or (17) is used
Fig. 2. Pseudocode of NMA algorithm.
172 A.P. Piotrowski / Information Sciences 241 (2013) 164194
v
i;g
= w
/
i;g
GM
i;g
1 w
/
i;g
_ _
LM
i;g
(18)
or, only if Eq. (16) is used, directly an offspring (u
i,g
) (without crossover)
u
i;g
= w
/
i;g
GM
i;g
(1 w
/
i;g
) LM
i;g
(19)
4.3. Crossover, bounds handling and selection
If for particular parent x
i,g
the donor vector v
i,g
was produced by Eq. (18), an offspring u
i,g
is created by means of the
widely used binomial crossover between donor and parent vectors
u
j
i;g
=
v
j
i;g
if rand
j
i
(0; 1) 6 CR
i;g
or j = j
rand;i
x
j
i;g
otherwise
_
_
_
(20)
As CR
i,g
is adapted and may vary for each individual and generation, it is indexed with i and g.
Regardless whether u
i,g
was created by means of Eq. (20) or Eq. (19), if an offspring is located outside the search space,
constraint handling method is applied (Eq. (11)).
According to a classic DE rule, the individual that is promoted to the next generation g + 1 is chosen according to the gree-
dy selection between an offspring u
i,g
and a parent x
i,g
following Eq. (12).
4.4. Adaptation of control parameters and probabilities of choosing mutation strategies
AM-DEGL uses the adaptation schemes proposed in SADE, but with few important modications.
4.4.1. Scale factor
In DEGL a scale factor used for both Global and Local mutation models is xed to 0.8. In SADE scale factor is also not
adapted, but for each individual at every generation it is randomly sampled from the normal distribution N (0.5, 0.3) (with
mean equal 0.5 and standard deviation equal 0.3). Like with population size, to successfully reconcile the need of larger scale
factor which improves the performance of DEGL and the variable scale factor required by SADE, in AM-DEGL the values of F
i,g
are sampled from N (0.8, 0.1). Other variants of choosing scale factors are tested in Section 5.4.
4.4.2. Crossover
In AM-DEGL the value of CR
i,g
is sampled for every individual at every generation from normal distribution N (CRm
k,g
, 0.1)
with mean CRm
k,g
that is adapted for each of four pairs (k = 1, . . . , 4) of Global and Local mutation models separately (the pair
dened by Eq. (16) do not follow crossover). Standard deviation is constant and equal to 0.1, as in SADE and the values of
CRm
k,g
are also adapted following the concept proposed in SADE [115]. CRMEM
k,g
is established to store all CR
i,g
values that
at generation g by means of kth strategy produced an offspring better than its parent. Starting from g = LP + 1, the CRm
k,g
is
computed as the median of all CR
i,g
values stored in the CRMEM
k,g
, . . . , CRMEM
k,gLP
. All CR
i,g
values stored in the memory
more than g-LP generations ago are discarded. In [115] it was shown that SADE is insensitive for LP between 20 and 60, hence
in AM-DEGL the LP period equal 40 generations is adopted. During rst LP generations, CRm
k,g
are kept 0.5.
4.4.3. Weight factor for combining Global and Local mutation models
Following DEGL [29], self-adaptation of weight factor is used to update the value of w
i,g
for each individual at every gen-
eration. The weight factor w
i,g
is incorporated into the vector x
i,g
. For every individual w
i,0
are initialized randomly within
[0.05, 0.95] interval. Than at every generation during mutation phase of x
i,g
the associated w
i,g
value is mutated in a similar
way to DE/current-to-best/1 strategy within [0.05, 0.95] interval:
w
/
i;g
= w
i;g
F
i;g
(w
best;g
w
i;g
) F
i;g
(w
r1;g
w
r2;g
)
if w
/
i;g
> 0:95 then w
/
i;g
= 0:95
if w
/
i;g
< 0:05 then w
/
i;g
= 0:05
(21)
where w
best,g
is the weight factor associated with the best individual in the population and w
r1,g
and w
r2,g
are two weight
factors associated with two rst individuals (x
r1,g
and x
r2,g
) that are to be used to mutate the parent (by GM
i,g
) in the current
generation. Irrespective of which mutation strategy is used by individual x
i,g
, w
i,g
is always mutated according to Eq. (21). As
each Global mutation model uses r1 and r2 indices, x
r1,g
and x
r2,g
affect both Global mutation model and mutation of w
i,g
,
however not always in a similar fashion. Using mutation schemes borrowed from SADE to mutate w
i,g
turned out inefcient
in preliminary tests and is not applied in this paper. Such created w
/
i;g
value is then used to combine Global and Local muta-
tion models according to Eq. (18) or (19). If the offspring u
i,g
created with use of w
/
i;g
is promoted to the next generation, then
associated weight factor is also promoted and w
i;g1
= w
/
i;g
, otherwise w
i,g+1
= w
i,g
.
A.P. Piotrowski / Information Sciences 241 (2013) 164194 173
4.4.4. Probability of using different mutation strategies
In AM-DEGL the adaptation of probabilities of using particular strategies is based on the method proposed in SADE [115],
but with a few modications. Because the individuals are initially randomly generated, during initial LP generations all 5
mutation strategies have equal probability of being used. Like in the case of crossover adaptation LP is set to 40 and during
this warm-up period the probability of each strategy being selected is constant (in Section 5.4 also an AM-DEGL variant with-
out warm-up period for adaptation of probabilities of using different strategies is tested). After LP-1 generations, the number
of successful (sf
k,g
) and unsuccessful (uf
k,g
) offspring generated by each kth strategy (k = 1, . . . , 5) in particular generation g is
counted. From the generation g = LP + 1 for each kth strategy the success rate in the last generation S
k,g
is computed accord-
ing to
S
k;g
= 0:01
sf
k;g1
sf
k;g1
uf
k;g1
if sf
k;g1
uf
k;g1
> 0
S
k;g
= 0:01 otherwise
(22)
Contrary to SADE, S
k
is based on the successful offspring from the most recent generation only. Although this formula may
lead to overcapitalize some strategies at particular generation, it also allows to quickly correct such mistakes and to
prompt response to changing requirements. The probability of using kth strategy in the next generation is estimated as
prob
k;g
= 0:01 0:95
S
k;g

5
k=1
S
k;g
(23)
hence contrary to SADE, in AM-DEGL the probability of using each strategy cannot be reduced to lower value than 0.01. In
fact it must be slightly higher as, following SADE, 0.01 constant is also used in Eq. (22). The larger minimum chance of being
used than used in SADE is advantageous to avoid a situation when a good strategy is unused for a number of generations due
to occasional poor performance, which may happen by chance as the adaptation is based on the results from the single gen-
eration only. The variant of AM-DEGL which uses equations proposed in SADE (Eq. (14) in [115] with LP = 1), instead of Eqs.
(22) and (23), is tested in Section 5.4.
4.5. Local search algorithm
Various local search methods are used by Memetic DE algorithms, from a very simple random walk [78] or chaotic search
[69] to heuristics like Adaptive hill descending with simplex crossover [98], Simulated Annealing, Hooke Jeeves Algorithm
[132], Rosenbrock method or NelderMead simplex [18]. Since the publication of [102] meta-Lamarckian learning (i.e.
hybridizing global search heuristic with more than one local search method) with adaptive probability of commencing each
local algorithm was also tested within DE framework [132,18]. However, in both mentioned papers only simple DE variants,
without adaptation of control parameters and with a single mutation strategy were used. Because in AM-DEGL a number of
Global and Local mutation strategies are applied, simple Lamarckian learning is chosen, namely a single NMA algorithm with
pre-dened probability to be commenced is adopted as local search method. Hybridization of DE with single local search
algorithm is also frequent in the literature [98,78,69]. NMA is chosen due to its ability to leave the basin of attraction, sig-
nicant, long-lasting popularity and a vast number of successful applications.
The key point is which individuals should be allowed to carry local search. Among Memetic DE algorithms various ap-
proaches were tested. In [98,69] and in the case of RA used in [18], only the best individuals have a chance to perform local
search. This saves the function calls, but allows speeding up the exploration of only the one local optimum that seems the
most promising at the moment. Contrary, in [78,132] and in the case of NMA used in [18], any individual may followthe local
search. Such approach may allow recognizing the presence of promising local optima in the vicinity of various individuals,
but frequently waste much more function calls.
In AM-DEGL an intermediate approach is proposed, which benet from the idea of Local mutation model only individ-
uals that are better than all other individuals within p% neighborhood have a chance to follow the local search (in AM-DEGL
p% is set to 20%, see discussion in Section 4.2). This way the number of individuals that have a chance to commence local
search may vary from 0 to almost 10% of the population, depending on relation in tness among neighboring individuals.
In practice it is observed that for 30D problems, when NP = 5D, usually 45 individuals in a generation are better than all
others within their neighborhood, and only they have a chance to follow local search. Because the information exchange be-
tween this few individuals is slow, the probability that they are located within the basins of attractions of different local min-
ima is higher than for neighboring individuals.
Another very important issue is the probability that an individual that is allowed to follow local search would in fact com-
mence it, and the length of local search. In the literature two options prevail (1) frequent but short or (2) rare but longer
local search. The rst option performs better for problems with simpler tness landscape when greedy local search meth-
ods are used, the second is advantageous if tness landscape is more complicated and more advanced heuristics that may in
fact leave the basin of attraction are used.
The rst option is applied in [98,78,69]. The local search may be commenced at every generation for the best individual
only [98] or even for every individual in the population [78]. In [78] maximum number of function calls available for each
local search is set to only 10. In [69] chaotic local search may last up to D/5 function calls, but terminates earlier when for the
174 A.P. Piotrowski / Information Sciences 241 (2013) 164194
rst time the better solution than the parents one is found. Interestingly, in [98] the contrary approach is used and the local
search is carried on as long as each offspring generated by simplex crossover is better than the previous one.
The second option is used in [18,132] and seems to be a good choice when NMA is used, because NMA is able to leave the
basin of attraction of local optimum. But this approach is more greedy in terms of function calls as NMA requires moving
D + 1 points throughout the search space.
Based on the above discussion, AM-DEGL follow the second option rare but longer local search. In AM-DEGL the NMA
may be employed for each individual which is better than all other individuals within p% neighborhood with the probability
LSp = 0.005, and last for 100D function calls.
NMA requires D + 1 points to be initialized. In AM-DEGL the rst point is represented by the individual for which the local
search is commenced which is the best one within own neighborhood (dened according to a ring topology). A concept
how to select additional D points is needed. The chosen points should show some similarities (if they span the whole search
space, the local search would probably be unsuccessful), but should also introduce some diversity to the simplex (to exploit
exploration capabilities of NMA and avoid problems with so called lost dimension). Hence in AM-DEGL the selection is split
into halves. The rst D/2 individuals are neighbors of x
i,g
with lower indexes i (i.e. x
i1,g
, . . . , x
iD/2,g
). These individuals are
within p% = 20% neighborhood of x
i,g
hence they exchange information among themselves and with x
i,g
frequently by means
of Local mutation model. This may be advantageous or not, depending on how similar closest neighbors are using too
similar individuals as starting points may diminish NMA search capabilities. Hence, as the another D/2 points the
consecutive individuals that have higher indexes than x
i,g
but are not within p% = 20% neighborhood of x
i,g
are used (i.e.
x
i+D/2+1,g
, . . . ,x
i+D,g
). They are localized close to x
i,g
according to a ring topology (hence may also show some similarities with
x
i,g
), however they exchange information by means of Local mutation model directly among themselves, but not with the
other points to be used within NMA (i.e. x
i,g
, . . . , x
iD/2,g
). One should remember that the NP individuals are organized on
a closed ring, hence the next index after NP is 1. If the problem is multimodal, one may expect that within such formed sim-
plex a number of local minima may exists, and the goal of NMA is to locate and exploit the best of them, what is time con-
suming. This explains why the proposed length of the local search is longer than in [18], where each local search lasted 500
function calls and mostly 10-dimensional benchmark problems were considered. A variant of AM-DEGL which uses the clos-
est individuals to x
i
in the search space according to Euclidean norm as additional D points is tested in Section 5.4.
In AM-DEGL the worst individual in DE population x
worst,g
is replaced by the best point found by the NMA. Hence the
local search helps improving the performance of the poorest part of the population. Note that x
worst,g
is always poorer
than the best solution found by NMA that start from x
i,g
. In the implementation used in the present paper (see pseudo-
code, Fig. 1) x
worst,g
is replaced immediately, but if parallel computing is used it may be replaced when the NP loop
terminates.
5. Results and discussion
5.1. Benchmark problems
The performance of AM-DEGL is tested on 10-, 30- and 50-dimensional versions of standard 25 benchmark problems pre-
pared for CEC 2005 [126]. Each algorithm is run 30-times on each version of every test problem. Since the publication of
[119] it is known that EAs should be tested on rotated problems, as many EAs lack rotational invariance properties. Also most
DE algorithms are not rotationally invariant due to the crossover operation, hence they performbetter for separable and non-
rotated problems. The CEC 2005 set of problems is chosen as it includes rotated, shifted and composition functions. Very
brief description of each problem is provided in Table 1, for more details the reader is referred to [126].
5.2. Algorithms for comparison
In the present paper the performance of the proposed AM-DEGL is compared with 13 different DE algorithms. The pop-
ulation size NP = 5D is set for AM-DEGL and most other DE methods, as this value lies in the middle of most frequently sug-
gested population sizes found in the literature [124,48,113,80,98,142,156]. However, due to their specic capabilities, some
Composite or Memetic DE algorithms require small population size, hence they are applied with NP = 1D; on the contrary,
DE-SG require NP = 10D (as its population is split into D groups with 10 individuals each [109]). The values of control param-
eters suggested in the source papers are used, with the exception of DE-SG, which was designed to solve problems when
large number of function calls is available in this case to quicken convergence two control parameters were modied,
see point 2 below. Some authors proposed more than one version of their algorithm in the source papers, in such case
the best performing variant is chosen and dened below. The algorithms for comparison include:
1. DE original version of Differential Evolution [123,124,113] with DE/rand/1/bin; NP = 5D; F = 0.8; CR = 0.5;
2. DE-SG DE with Separated Groups [109,110]; NP = 10D; PNI = 10; MigProb = 1/PNI;
3. DEGL DE with Global and Local neighborhood-based mutation operators [29]; NP = 5D; self-adaptive weight factor is
used;
4. SADE Self-Adaptive DE [115]; NP = 5D;
A.P. Piotrowski / Information Sciences 241 (2013) 164194 175
5. SspDE Self-Adaptive DE [105]; NP = 5D;
6. JADE [156]; NP = 5D; with an archive;
7. AdapSS-JADE Adaptive Strategy Selection-based JADE [55]; NP = 5D; with Averaged Normalized Reward method;
8. DECLS DE algorithm based on chaotic local search [69]; NP = 1D;
9. DEahcSPX DE with adaptive hill-climbing (in case of minimization descending) crossover-based local search [98];
NP = 1D;
10. CoDE Composite DE [139]; NP = 1D; note that in this approach three offspring are generated for each parent and t-
ness function is called for each of them, hence lower population size is required;
11. EPSDE Ensemble of mutation strategies and parameters values for DE [85]; NP = 5D;
12. SFMDE Super-t Memetic DE [18]; NP = 5D;
13. CDE Clustering-based DE [17]; NP = 5D.
5.3. Comparison
Tables 24 show the average best tness obtained by each algorithm for every 10-, 30- and 50-dimensional problem,
respectively, accompanied by the standard deviation. Three statistical procedures are employed to test the differences be-
tween the results obtained by different algorithms.
First, the version of Friedman test [47] extended by Iman and Davenport (see [35,49]) is applied to test the hypothesis
that there is no signicant difference between results obtained by means of all applied algorithms. Three F
F
statistics:
F
F
= 16.58, F
F
= 12.15 and F
F
= 20.74, that were obtained for 10-, 30- and 50-dimensional problems, respectively, show that
the hypothesis that the performance of all methods is similar may be rejected at very high level of signicance (three cor-
responding p-values are below 10
7
) and another tests may be applied to detect the best methods.
Than, a statistical signicance of difference between the results obtained by means of AM-DEGL and other algorithms is
tested by means of rank-based Holm variant [60] of one-tailed Bonferroni-Dunn test [35,49] that adjusts p-values when a
number of inter-comparisons are made. Holm procedure compares aggregated over all 25 problems performance of all
algorithms against the control method (AM-DEGL). The adjusted p-values (APv) and information whether AM-DEGL is sig-
nicantly better than each among 13 competitive algorithms at 5% signicance level are reported in the second lowest row in
Tables 24 (denoted as H/APv). Note that ++ means that AM-DEGL is signicantly better than the competitive algorithm, ==
means no signicant difference (if the APv value is higher than 0.5, what happens a few times for 10-dimensional problems,
the Holm procedure with competitive algorithm as a control method is applied to test if the competitive algorithm is signif-
icantly better than AM-DEGL in such cases tests always showed that the difference is not signicant).
Finally, for each algorithm (not only AM-DEGL) the statistical signicance of its performance against the performance of
13 other algorithms is assessed for each problem separately by very popular (although sometimes criticized when applied
for multi-comparisons, see [49]) Wilcoxon rank-sum test [148] at 5% signicance level (see +/=/ columns in Tables 24).
Table 1
Summary of CEC 2005 test problems.
Problem Type f (x

) Initialization range Bounds


F1 Unimodal 450 [100, 100]
D
[100, 100]
D
F2 Unimodal 450 [100, 100]
D
[100, 100]
D
F3 Unimodal 450 [100, 100]
D
[100, 100]
D
F4 Unimodal 450 [100, 100]
D
[100, 100]
D
F5 Unimodal 310 [100, 100]
D
[100, 100]
D
F6 Basic multimodal 390 [100, 100]
D
[100, 100]
D
F7 Basic multimodal 180 [0, 600]
D
None
F8 Basic multimodal 140 [32, 32]
D
[32, 32]
D
F9 Basic multimodal 330 [5, 5]
D
[5, 5]
D
F10 Basic multimodal 330 [5, 5]
D
[5, 5]
D
F11 Basic multimodal 90 [0.5, 0.5]
D
[0.5, 0.5]
D
F12 Basic multimodal 460 [p, p]
D
[p, p]
D
F13 Expanded multimodal 130 [3, 1]
D
[3, 1]
D
F14 Expanded multimodal 300 [100, 100]
D
[100, 100]
D
F15 Hybrid composition 120 [5, 5]
D
[5, 5]
D
F16 Hybrid composition 120 [5, 5]
D
[5, 5]
D
F17 Hybrid composition 120 [5, 5]
D
[5, 5]
D
F18 Hybrid composition 10 [5, 5]
D
[5, 5]
D
F19 Hybrid composition 10 [5, 5]
D
[5, 5]
D
F20 Hybrid composition 10 [5, 5]
D
[5, 5]
D
F21 Hybrid composition 360 [5, 5]
D
[5, 5]
D
F22 Hybrid composition 360 [5, 5]
D
[5, 5]
D
F23 Hybrid composition 360 [5, 5]
D
[5, 5]
D
F24 Hybrid composition 260 [5, 5]
D
[5, 5]
D
F25 Hybrid composition 260 [2, 5]
D
None
176 A.P. Piotrowski / Information Sciences 241 (2013) 164194
Table 2
The mean tness values (mean) and its standard deviations (std) for 10-dimensional problems. +/=/ shows than how many of 13 competing methods the particular algorithm is signicantly better/equal/
signicantly worse, according to Wilcoxon rank-sum test at 5% signicance level. W/Pav how many times, averaged over all problems, the particular algorithm is signicantly better/signicantly worse than other
algorithms according to Wilcoxon test. H/APv adjusted p-values according to Holm test procedure with AM-DEGL as a control method; ++ means that AM-DEGL is signicantly better than the competitive algorithm,
== means that there is no signicant difference between AM-DEGL and competitive method at 5% signicance level, according to Holms procedure.
AM-DEGL DE DE-SG DEGL SADE
mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F1 450.000 1.42E7 2/11/0 450.000 1.09E7 1/12/0 450.000 1.50E7 2/11/0 450.000 1.83E7 2/11/0 450.000 1.72E7 2/11/0
F2 450.000 1.98E7 1/12/0 449.999 8.40E4 0/13/0 450.000 1.69E7 1/12/0 450.000 1.82E7 1/12/0 450.000 1.68E7 1/12/0
F3 450.000 2.66E7 11/2/0 9.07E+5 2.86E+5 0/0/13 450.000 1.16E5 10/3/0 449.999 3.68E3 11/2/0 365.430 4.63E+2 9/1/3
F4 450.000 2.86E7 3/10/0 449.995 4.53E3 2/0/11 450.000 1.38E7 2/11/0 450.000 1.28E7 2/11/0 450.000 1.45E7 2/11/0
F5 310.000 8.21E8 1/12/0 310.000 7.67E8 1/12/0 310.000 4.12E6 1/12/0 310.000 1.09E7 1/12/0 310.000 4.95E8 1/12/0
F6 390.399 1.22E+0 5/2/6 390.361 3.65E1 7/1/5 390.010 1.73E2 12/0/1 391.063 1.79E+0 3/6/4 390.000 1.20E7 8/5/0
F7 179.972 2.38E2 8/5/0 179.517 6.94E2 2/0/11 179.926 3.61E2 5/1/7 179.944 3.48E2 7/0/6 179.977 1.92E2 8/5/0
F8 119.998 3.32E3 13/0/0 119.654 7.22E2 1/8/4 119.711 9.62E2 8/3/2 119.645 6.52E2 1/8/4 119.673 8.18E2 3/7/3
F9 330.000 1.28E7 4/9/0 330.000 1.82E7 4/9/0 330.000 1.76E7 4/9/0 325.316 2.99E+0 1/2/10 330.000 1.87E7 4/9/0
F10 324.548 1.99E+0 7/2/4 300.060 4.66E+0 0/0/13 323.693 1.99E+0 5/3/5 322.177 4.46E+0 5/2/6 325.679 2.01E+0 9/4/0
F11 93.162 1.50E+0 9/2/2 98.415 6.74E1 0/0/13 94.307 1.40E+0 2/6/5 91.290 1.26E+0 12/1/0 92.742 2.03E+0 9/2/2
F12 397.498 2.85E+2 8/2/3 461.170 3.74E+2 2/0/11 457.653 4.30E+0 9/4/0 664.703 2.48E+3 0/6/7 406.082 1.79E+2 5/7/1
F13 129.699 1.01E1 12/1/0 128.526 1.75E1 0/0/13 129.513 6.55E2 5/2/6 129.234 2.88E1 2/2/9 129.636 5.75E2 10/2/1
F14 297.213 4.03E1 6/4/3 296.399 1.69E1 0/1/12 296.913 1.88E1 4/1/8 297.727 4.99E1 13/0/0 297.228 3.82E1 6/5/2
F15 277.009 1.90E+2 5/7/1 153.538 7.76E+1 7/5/1 120.000 1.21E7 13/0/0 425.217 1.75E+2 0/3/10 244.360 1.71E+2 6/6/1
F16 220.187 5.42E+0 7/3/3 279.915 1.03E+1 0/0/13 219.936 8.21E+0 7/2/4 226.654 1.15E+1 6/1/6 217.079 8.13E+0 8/5/0
F17 230.356 7.31E+0 6/1/6 290.574 1.67E+1 0/0/13 235.790 7.64E+0 5/0/8 226.552 1.00E+1 7/5/1 223.405 9.65E+0 8/5/0
F18 576.667 2.54E+2 7/5/1 493.340 2.45E+2 6/6/1 343.333 1.27E+2 12/1/0 785.028 1.88E+2 1/5/7 665.760 2.38E+2 4/7/2
F19 593.333 2.52E+2 5/7/1 422.381 2.03E+2 8/5/0 393.333 1.90E+2 10/3/0 738.002 2.38E+2 1/6/6 649.834 2.38E+2 5/6/2
F20 726.667 1.90E+2 2/5/6 526.676 2.52E+2 4/8/1 393.333 1.90E+2 12/1/0 726.555 2.13E+2 1/6/6 561.538 2.56E+2 8/5/0
F21 853.333 1.80E+2 7/6/0 860.000 5.7E13 8/5/0 843.333 6.48E+1 8/5/0 1144.564 2.93E+2 0/3/10 896.667 1.65E+2 4/9/0
F22 1110.132 8.54E+0 6/5/2 1131.968 3.39E+0 1/3/9 1008.945 1.96E+2 7/6/0 1141.697 3.31E+1 0/6/7 1113.613 1.72E+1 6/5/2
F23 1033.581 1.41E+2 5/6/2 935.643 4.94E+1 12/1/0 919.468 3.3E13 12/1/0 1195.515 2.30E+2 1/4/8 1080.043 2.03E+2 4/7/2
F24 460.000 0.00E+0 7/6/0 460.000 0.00E+0 7/6/0 460.000 0.00E+0 7/6/0 592.008 2.45E+2 0/6/7 460.000 0.00E+0 7/6/0
F25 633.708 3.12E+0 9/4/0 646.711 3.41E+0 1/3/9 636.336 3.26E+0 6/5/2 783.013 2.65E+2 0/4/9 635.265 3.31E+0 6/7/0
H/APv 0.0037 (++) 1 (==) 0.0036 (++) 1 (==)
W/Pav 6.24/1.6 2.96/6.12 6.76/1.92 3.12/4.92 5.72/0.84
SspDE JADE AdapSS-JADE DECLS DEahcSPX
F1 450.000 1.41E7 2/11/0 450.000 1.47E7 2/11/0 450.000 1.85E7 1/12/0 449.506 2.54E+0 0/5/8 450.000 7.78E8 1/12/0
F2 450.000 9.22E8 1/12/0 450.000 1.82E7 1/12/0 450.000 1.30E7 1/12/0 448.455 4.82E+0 0/1/12 450.000 4.80E8 1/12/0
F3 1773.011 2.95E+3 4/4/5 5657.087 8.71E+3 3/5/5 7702.986 1.57E+4 4/4/5 5.93E+4 6.09E+4 1/0/12 1.06E+4 1.36E+4 3/1/9
F4 450.000 8.61E8 2/11/0 450.000 1.12E7 2/11/0 450.000 1.51E7 2/11/0 443.048 2.08E+1 1/0/12 450.000 1.05E7 2/11/0
F5 310.000 6.47E8 1/12/0 310.000 9.53E8 1/12/0 310.000 1.01E7 1/12/0 307.605 7.93E+0 1/1/11 310.000 4.99E8 1/12/0
F6 390.266 1.01E+0 7/3/3 390.399 1.22E+0 3/7/3 390.266 1.01E+0 5/7/1 435.243 5.52E+1 0/0/13 391.428 1.91E+0 2/1/10
F7 179.981 1.51E2 8/5/0 179.982 1.51E2 8/5/0 179.981 1.65E2 8/5/0 173.898 8.85E+0 0/0/13 179.697 2.75E1 2/1/10
F8 119.630 6.09E2 1/6/6 119.659 6.53E2 2/7/4 119.637 5.83E2 1/7/5 119.615 6.15E2 1/4/8 119.411 1.36E1 0/0/13
F9 330.000 1.74E7 4/9/0 330.000 1.47E7 4/9/0 330.000 1.33E7 4/9/0 325.755 4.37E+0 1/2/10 318.020 8.68E+0 0/0/13
F10 325.863 1.26E+0 9/4/0 326.199 1.08E+0 10/3/0 326.338 1.35E+0 10/3/0 317.717 7.56E+0 3/1/9 313.341 8.83E+0 1/1/11
F11 94.301 1.07E+0 4/4/5 94.205 7.68E1 4/5/4 93.498 1.39E+0 8/3/2 94.475 1.48E+0 2/6/5 95.548 2.04E+0 2/4/7
F12 430.945 1.29E+2 4/7/2 312.384 3.98E+2 3/8/2 149.507 5.31E+2 3/7/3 632.215 2.07E+3 1/0/12 419.974 2.18E+3 2/1/10
F13 129.660 4.56E2 10/2/1 129.545 8.59E2 5/4/4 129.551 9.19E2 6/3/4 129.667 1.26E1 10/3/0 128.777 5.87E1 1/0/12
F14 297.062 2.59E1 6/3/4 297.425 2.81E1 10/2/1 297.317 3.02E1 7/5/1 297.177 4.20E1 6/4/3 296.385 4.18E1 0/1/12
(continued on next page)
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
1
7
7
Table 2 (continued)
AM-DEGL DE DE-SG DEGL SADE
mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F15 215.076 1.59E+2 7/5/1 255.724 1.70E+2 5/5/3 244.643 1.73E+2 6/6/1 342.720 1.54E+2 1/3/9 466.038 1.43E+2 0/1/12
F16 216.704 4.74E+0 10/3/0 217.055 6.30E+0 10/3/0 216.856 5.12E+0 10/3/0 244.161 2.18E+1 1/3/9 246.752 2.00E+1 1/3/9
F17 224.472 6.65E+0 7/5/1 225.767 6.11E+0 7/5/1 221.583 5.14E+0 11/2/0 248.632 1.69E+1 2/2/9 248.799 2.05E+1 2/2/9
F18 610.000 2.49E+2 3/7/3 674.684 2.65E+2 2/7/4 739.543 2.22E+2 1/5/7 861.838 1.83E+2 0/0/13 818.845 1.68E+2 1/3/9
F19 586.033 2.52E+2 4/7/2 643.059 2.60E+2 4/6/3 768.109 2.13E+2 1/3/9 884.725 1.65E+2 0/1/12 774.892 2.43E+2 0/4/9
F20 685.352 2.22E+2 4/4/5 721.275 2.34E+2 2/6/5 704.132 2.45E+2 2/6/5 931.768 8.66E+1 0/0/13 790.109 1.86E+2 1/4/8
F21 943.333 1.91E+2 4/6/3 986.674 2.20E+2 4/5/4 982.101 2.20E+2 4/5/4 1234.758 2.62E+2 0/2/11 1232.715 2.78E+2 0/2/11
F22 1121.949 1.77E+1 5/2/6 1117.379 2.19E+1 6/6/1 1111.589 1.73E+1 6/5/2 1143.897 1.10E+2 0/5/8 1159.528 4.64E+1 0/2/11
F23 1080.218 1.92E+2 4/7/2 1128.118 1.75E+2 2/6/5 1200.437 2.05E+2 1/3/9 1296.338 2.73E+2 0/3/10 1322.804 2.35E+2 0/1/12
F24 470.000 5.48E+1 2/11/0 510.000 1.14E+2 1/7/5 500.000 1.04E+2 1/7/5 530.258 1.30E+2 0/1/12 523.481 1.59E+2 1/7/5
F25 635.583 3.40E+0 7/6/0 644.110 3.17E+1 5/5/3 644.447 3.16E+1 5/5/3 853.563 3.57E+2 0/1/12 683.080 1.44E+2 1/3/9
H/APv 0.3747 (==) 0.2527 (==) 0.1750 (==) 1.55E9 (++) 3.54E10 (++)
W/Pav 4.80/1.96 4.24/2.28 4.36/2.64 1.24/9.84 1.00/8.44
CoDE EPSDE SFMDE CDE
F1 450.000 1.39E7 1/12/0 450.000 1.37E7 1/12/0 449.761 1.31E+0 1/0/12 450.000 2.22E7 2/11/0
F2 450.000 1.17E7 1/12/0 450.000 1.57E7 1/12/0 450.000 4.53E4 1/12/0 450.000 1.68E7 1/12/0
F3 2869.529 6.44E+3 4/4/5 441.493 3.39E+1 10/0/3 2.12E+04 9.58E+4 2/3/8 6936.786 7.92E+3 3/3/7
F4 450.000 1.46E7 2/11/0 450.000 1.41E7 2/11/0 406.592 2.00E+2 0/9/4 450.000 1.62E7 3/10/0
F5 310.000 1.03E7 1/12/0 310.000 5.84E8 1/12/0 229.331 3.70E+2 0/11/2 310.000 8.72E8 1/12/0
F6 390.399 1.22E+0 3/6/4 390.133 7.28E1 9/2/2 390.930 1.71E+0 3/6/4 392.868 1.26E+0 1/0/12
F7 179.921 3.83E2 5/1/7 179.878 7.24E2 4/0/9 179.410 7.92E1 1/1/11 179.973 2.96E2 8/5/0
F8 119.684 1.05E1 4/7/2 119.651 5.08E2 2/7/4 119.962 4.52E2 12/0/1 119.767 1.71E1 9/2/2
F9 329.735 5.18E1 4/8/1 330.000 1.39E7 4/9/0 326.086 1.67E+0 1/2/10 330.000 2.05E7 5/8/0
F10 318.492 5.94E+0 3/2/8 316.864 3.06E+0 1/1/11 320.803 5.55E+0 4/3/6 324.826 2.82E+0 8/3/2
F11 94.319 2.10E+0 2/6/5 96.196 8.26E1 1/0/12 94.969 1.06E+0 2/4/7 90.936 1.09E+0 12/1/0
F12 372.249 1.98E+2 3/7/10 457.046 6.40E+0 11/2/0 236.118 5.08E+2 3/7/10 434.691 8.43E+1 5/6/2
F13 129.584 9.89E2 6/3/4 129.341 8.16E2 2/2/9 129.220 2.76E1 2/2/9 129.563 1.23E1 5/4/4
F14 296.921 3.46E1 4/1/8 296.738 1.96E1 2/1/10 296.681 2.47E1 2/1/10 297.383 3.48E1 9/3/1
F15 276.855 1.70E+2 4/4/5 266.709 1.84E+2 5/7/1 359.704 1.43E+2 1/2/10 333.379 1.71E+2 3/2/8
F16 233.090 8.69E+0 4/1/8 235.340 6.86E+0 1/4/8 246.523 2.73E+1 1/3/9 219.455 1.21E+1 6/4/3
F17 229.826 8.53E+0 6/4/3 256.999 1.01E+1 1/1/11 250.679 2.35E+1 1/3/9 224.373 1.08E+1 8/5/0
F18 634.968 2.53E+2 5/6/2 616.280 2.55E+2 5/6/2 671.417 2.32E+2 1/5/7 493.333 2.45E+2 10/3/0
F19 645.123 2.56E+2 4/7/2 512.282 2.52E+2 7/6/0 705.301 2.17E+2 1/3/9 516.627 2.53E+2 5/8/0
F20 606.876 2.66E+2 7/5/1 611.842 2.51E+2 7/5/1 698.316 2.19E+2 2/4/7 557.336 2.53E+2 7/5/1
F21 924.882 2.17E+2 4/9/0 866.667 1.66E+2 6/7/0 1099.838 2.49E+2 2/1/10 931.815 1.61E+2 4/7/1
F22 1107.818 8.79E+1 6/5/2 1091.634 1.17E+2 7/4/1 1131.778 2.26E+1 1/5/7 1081.334 1.16E+2 8/5/0
F23 981.053 1.17E+2 5/6/2 1020.324 1.55E+2 7/4/2 1055.451 2.34E+2 3/7/3 1077.018 1.65E+2 4/6/3
F24 510.000 1.14E+2 1/7/5 480.000 7.61E+1 2/11/0 460.005 2.98E2 1/7/5 460.000 0.00E+0 7/6/0
F25 637.963 3.84E+0 5/3/5 637.616 2.33E+1 8/3/2 647.718 9.88E+0 1/3/9 634.358 2.70E+0 9/4/0
H/APv 0.1273 (==) 0.1613 (==) 6.20E6 (++) 1 (==)
W/Pav 3.76/3.28 4.28/3.56 1.96/6.88 5.72/1.88
1
7
8
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
Table 3
The mean tness values (mean) and its standard deviations (std) for 30-dimensional problems. +/=/ shows then how many of 13 competing methods the particular algorithm is signicantly better/equal/
signicantly worse, according to Wilcoxon rank-sum test at 5% signicance level. W/Pav how many times, averaged over all problems, the particular algorithm is signicantly better/signicantly worse than other
algorithms according to Wilcoxon test. H/APv adjusted p-values according to Holm test procedure with AM-DEGL as a control method; ++ means that AM-DEGL is signicantly better than the competitive algorithm,
== means that there is no signicant difference between AM-DEGL and competitive method at 5% signicance level, according to Holms procedure.
AM-DEGL DE DE-SG DEGL SADE
Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F1 450.000 3.47E7 2/11/0 449.939 1.24E2 0/0/13 450.000 6.92E8 2/11/0 450.000 1.21E7 2/11/0 450.000 1.13E7 2/11/0
F2 450.000 2.37E7 3/10/0 1.72E+4 2.70E+3 0/0/13 449.962 3.07E2 2/0/11 450.000 5.41E8 3/10/0 450.000 3.35E8 3/10/0
F3 1780.313 4.43E+3 13/0/0 1.34E+8 2.68E+7 0/0/13 7.23E+4 4.64E+4 10/1/2 4.26E+4 2.39E+4 11/1/1 8.21E+4 9.98E+4 10/2/1
F4 450.000 6.16E4 11/2/0 2.48E+4 3.38E+3 0/0/13 448.872 5.23E1 4/0/9 450.000 5.06E8 11/2/0 449.997 9.62E3 9/1/3
F5 309.980 9.75E2 12/0/1 5456.407 3.90E+2 0/0/13 526.089 3.50E+2 7/0/6 309.999 1.56E3 13/0/0 68.838 5.61E+2 10/1/2
F6 391.595 5.44E+0 9/1/3 5797.662 1.56E+3 0/0/13 482.619 1.54E+1 1/1/11 391.063 1.79E+0 12/1/0 391.311 1.88E+0 10/1/2
F7 179.998 4.08E3 13/0/0 19.031 2.74E+1 0/0/13 179.255 2.45E1 2/0/11 179.987 1.16E2 4/6/3 179.988 7.94E3 5/4/4
F8 119.983 1.43E2 13/0/0 119.044 4.42E2 0/11/2 119.045 3.92E2 0/11/2 119.063 5.88E2 0/11/2 119.051 4.18E2 0/11/2
F9 326.658 2.45E+0 7/0/6 208.680 8.20E+0 0/0/13 324.188 2.09E+0 6/0/7 296.559 2.20E+1 2/0/11 330.000 7.94E8 10/3/0
F10 297.067 1.06E+1 11/2/0 108.776 1.09E+1 0/0/13 209.218 9.66E+0 3/0/10 274.584 4.64E+1 6/4/3 278.311 9.01E+0 8/0/5
F11 113.845 3.95E+0 10/2/1 129.637 1.15E+0 0/0/13 121.330 1.16E+0 5/1/7 101.430 8.38E+0 13/0/0 113.543 6.33E+0 9/3/1
F12 1915.835 4.45E+3 8/5/0 3.89E+5 3.73E+4 0/0/13 1.52E+4 1.58E+4 4/1/8 7351.982 8.08E+3 5/2/6 1372.484 2.53E+3 7/6/0
F13 128.027 4.02E1 11/1/1 110.999 1.13E+0 0/0/13 126.232 2.89E1 5/1/7 123.932 4.13E+0 2/3/8 127.151 2.16E1 8/2/3
F14 287.285 4.67E1 7/3/3 286.542 1.72E1 0/0/13 287.005 1.80E1 5/1/7 287.520 3.86E1 10/3/0 287.175 2.16E1 7/2/4
F15 520.000 9.10E+1 1/8/4 492.964 4.38E+1 1/11/1 486.429 1.03E+2 6/4/3 461.621 1.11E+2 5/8/0 514.642 8.41E+1 1/10/2
F16 197.896 6.86E+1 8/4/1 366.187 1.60E+1 0/0/13 240.114 1.08E+1 4/3/6 272.790 1.38E+2 3/6/4 177.164 1.41E+1 9/4/0
F17 220.789 4.43E+1 8/5/0 389.383 1.59E+1 0/1/12 288.726 1.10E+1 3/1/9 272.092 1.44E+2 4/9/0 221.070 7.26E+1 8/5/0
F18 914.442 5.06E1 4/4/5 917.051 1.78E1 0/3/10 874.285 5.34E+1 3/10/0 912.946 1.95E+1 8/5/0 914.999 1.44E+0 3/5/5
F19 900.701 3.62E+1 9/3/1 917.008 1.72E1 0/4/9 895.479 4.35E+1 9/4/0 909.618 2.71E+1 7/5/1 915.141 1.41E+0 4/4/5
F20 914.877 1.17E+0 4/4/5 917.027 1.77E1 0/1/12 884.611 4.97E+1 4/9/0 912.698 1.95E+1 7/5/1 914.946 1.24E+0 4/4/5
F21 860.000 1.1E13 4/9/0 860.013 2.36E3 3/0/10 860.000 6.19E9 4/9/0 906.667 1.22E+2 0/4/9 860.000 2.3E13 4/9/0
F22 1230.939 1.88E+1 12/1/0 1318.997 7.78E+0 0/1/12 1262.048 1.19E+1 2/2/9 1252.624 1.86E+1 4/3/6 1243.142 1.79E+1 6/6/1
F23 894.164 1.58E4 6/7/0 894.168 1.44E3 4/1/8 894.165 2.57E4 6/7/0 922.360 1.02E+2 0/3/10 894.164 5.81E5 6/7/0
F24 460.000 0.00E+0 1/12/0 460.034 6.47E3 0/0/13 460.000 7.84E8 1/12/0 460.000 0.00E+0 1/12/0 460.000 0.00E+0 1/12/0
F25 469.656 6.06E1 13/0/0 565.429 2.08E+1 0/0/13 470.340 6.51E1 9/3/1 490.420 7.33E+1 1/2/10 470.326 6.28E1 9/3/1
H/APv 3.33E16 (++) 2.98E4 (++) 0.0214 (++) 0.0855 (==)
W/Pav 8.00/1.24 0.32/11.36 4.28/5.04 5.36/3.00 6.12/1.84
SspDE JADE AdapSS-JADE DECLS DEahcSPX
F1 450.000 6.31E8 2/11/0 450.000 8.17E8 2/11/0 450.000 8.88E8 2/11/0 450.000 5.81E8 2/11/0 450.000 5.15E8 2/11/0
F2 450.000 4.06E5 3/10/0 450.000 3.79E8 3/10/0 450.000 3.41E5 3/10/0 450.000 2.33E6 3/10/0 450.000 6.21E5 3/10/0
F3 1.85E+05 8.95E+4 7/2/4 7.58E+05 3.00E+5 3/1/9 7.45E+05 3.57E+5 3/1/9 2.26E+05 1.20E+5 6/3/4 3.74E+05 2.09E+5 4/2/7
F4 422.121 6.18E+1 2/1/10 449.990 3.40E2 8/1/4 449.978 3.36E2 7/0/6 432.075 2.51E+1 2/1/10 449.120 1.40E+0 5/1/7
F5 2144.586 4.34E+2 2/0/11 1236.255 3.18E+2 4/1/8 974.882 3.62E+2 6/0/7 1491.183 6.75E+2 3/2/8 200.769 1.29E+2 10/1/2
F6 402.045 2.03E+1 4/2/7 396.426 9.05E+0 5/2/6 402.020 1.84E+1 4/2/7 394.317 1.24E+1 8/2/3 391.496 1.80E+0 9/1/3
F7 179.981 1.18E2 4/0/9 179.992 5.32E3 4/7/2 179.990 5.47E3 5/4/4 179.988 1.18E2 4/5/4 179.994 8.09E3 8/3/2
F8 119.046 4.82E2 0/11/2 119.047 4.24E2 0/11/2 119.052 5.31E2 0/11/2 119.056 5.05E2 0/11/2 119.046 6.07E2 0/11/2
F9 330.000 7.12E8 10/3/0 277.907 5.74E+0 1/0/12 329.012 2.86E+0 8/1/4 329.503 7.27E1 8/1/4 311.035 5.57E+0 5/0/8
F10 301.378 4.58E+0 11/2/0 288.959 6.89E+0 8/3/2 286.469 8.01E+0 8/2/3 291.466 9.46E+0 8/3/2 268.651 5.94E+1 6/4/3
F11 116.984 1.26E+0 9/1/3 118.628 1.65E+0 7/1/5 118.910 1.50E+0 7/1/5 115.171 3.04E+0 10/2/1 124.541 8.47E+0 2/0/11
F12 2204.386 3.30E+3 7/6/0 2099.096 3.05E+3 7/6/0 3935.267 5.42E+3 6/6/1 2579.561 4.53E+3 7/6/0 2140.800 2.91E+3 7/6/0
F13 127.986 1.85E1 11/1/1 124.816 4.32E1 4/0/9 125.189 5.75E1 4/1/8 128.782 1.47E1 13/0/0 127.283 6.08E1 8/2/3
F14 287.168 2.23E1 7/2/4 287.600 2.38E1 11/2/0 287.571 1.78E1 11/2/0 287.372 3.90E1 9/2/2 286.924 2.63E1 3/3/7
(continued on next page)
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
1
7
9
Table 3 (continued)
AM-DEGL DE DE-SG DEGL SADE
Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F15 513.400 7.41E+1 1/9/3 491.911 9.23E+1 1/11/1 476.667 9.35E+1 4/9/0 456.718 8.51E+1 10/3/0 507.154 8.60E+1 1/11/1
F16 183.381 6.74E+1 11/1/1 192.856 6.37E+1 7/5/1 213.502 1.01E+2 7/4/2 252.027 1.23E+2 4/1/8 204.940 8.18E+1 9/4/0
F17 212.292 2.27E+1 9/4/0 265.870 9.38E+1 5/4/4 260.366 9.59E+1 6/4/3 235.371 8.64E+1 8/5/0 279.492 8.45E+1 3/7/3
F18 909.721 2.72E+1 6/7/0 914.430 8.24E1 4/4/5 913.954 6.93E1 8/2/3 916.899 2.06E+0 0/4/9 914.084 9.62E1 8/2/3
F19 913.685 1.97E+1 7/4/2 914.388 6.94E1 5/3/5 913.990 5.00E1 7/2/4 916.640 2.39E+0 0/6/7 913.837 3.36E1 7/2/4
F20 910.564 2.74E+1 10/2/1 914.616 1.05E+0 4/4/5 914.042 6.56E1 7/2/4 916.729 2.26E+0 0/7/6 913.804 2.94E1 7/2/4
F21 860.000 4.4E13 3/10/0 860.000 2.2E13 2/11/0 860.000 2.3E13 3/10/0 860.000 3.2E13 2/11/0 860.000 1.4E13 4/9/0
F22 1247.129 1.83E+1 5/7/1 1239.633 1.74E+1 6/7/0 1245.344 1.23E+1 6/6/1 1242.046 1.80E+1 6/6/1 1247.500 1.62E+1 5/7/1
F23 894.164 9.23E5 5/8/0 907.592 7.35E+1 1/1/11 894.164 3.49E4 6/7/0 894.165 5.82E4 6/7/0 894.172 1.59E2 2/2/9
F24 460.000 5.7E13 1/12/0 460.000 0.00E+0 1/12/0 460.000 0.00E+0 1/12/0 460.000 0.00E+0 1/12/0 460.000 0.00E+0 1/12/0
F25 472.049 1.32E+0 3/4/6 471.705 6.39E1 3/4/6 471.758 5.64E1 3/4/6 472.893 2.75E+0 3/3/7 470.200 6.31E1 9/3/1
H/APv 0.0154 (++) 0.0159 (++) 0.0214 (++) 0.0083 (++) 0.0050 (++)
W/Pav 5.60/2.60 4.24/3.88 5.28/3.16 4.92/3.12 5.12/3.24
CoDE EPSDE SFMDE CDE
F1 450.000 7.43E8 2/11/0 450.000 9.46E8 2/11/0 449.974 1.43E1 1/0/12 450.000 1.76E7 2/11/0
F2 450.000 2.34E6 3/10/0 450.000 2.00E8 3/10/0 355.757 5.16E+2 1/0/12 450.000 6.14E8 3/10/0
F3 2.32E+05 1.12E+5 6/3/4 3.53E+05 1.82E+5 4/2/7 3.75E+06 9.85E+6 1/4/8 1.37E+06 5.36E+5 2/0/11
F4 449.418 6.45E1 5/1/7 449.998 3.79E3 8/2/3 1654.396 3.17E+3 1/0/12 450.000 1.56E5 11/2/0
F5 87.820 2.04E+2 8/1/4 258.714 6.01E+2 8/1/4 2865.587 1.09E+3 1/0/12 1520.590 4.55E+2 3/1/9
F6 403.509 1.12E+1 3/0/10 390.133 7.28E1 11/2/0 520.362 1.71E+2 1/1/11 397.526 1.16E+0 6/0/7
F7 179.972 5.57E2 3/4/6 179.997 5.09E3 12/0/1 175.777 1.31E+1 1/0/12 179.995 5.89E3 8/3/2
F8 119.049 4.58E2 0/11/2 119.048 5.24E2 0/11/2 119.293 1.64E1 12/0/1 119.060 4.78E2 0/11/2
F9 330.000 7.32E8 10/3/0 304.132 2.33E+0 3/1/9 306.477 1.02E+1 3/1/9 330.000 1.90E7 10/3/0
F10 154.278 1.30E+1 1/0/12 168.704 9.19E+0 2/0/11 244.334 4.94E+1 5/0/8 225.003 1.11E+1 4/0/9
F11 122.765 1.53E+0 3/1/9 125.424 1.43E+0 1/0/12 122.566 2.23E+0 3/1/9 121.327 1.40E+0 5/1/7
F12 8.71E+04 1.33E+4 1/1/11 6.49E+04 4.60E+4 1/1/11 2.42E+04 3.96E+4 4/2/7 5.75E+04 9.88E+3 3/0/10
F13 123.940 6.99E1 3/0/10 123.478 6.99E1 1/0/12 127.080 7.69E1 8/2/3 126.509 2.93E1 6/1/6
F14 286.790 1.95E1 1/3/9 286.743 2.10E1 1/2/10 286.844 2.62E1 1/4/8 286.856 1.85E1 2/3/8
F15 526.667 5.21E+1 0/8/5 510.000 1.09E+2 1/11/1 525.053 4.02E+1 1/1/11 503.333 6.48E+1 1/10/2
F16 312.983 2.06E+1 1/1/11 353.324 9.16E+1 1/1/11 254.364 5.58E+1 3/3/7 238.450 1.22E+1 4/3/6
F17 357.225 2.79E+1 1/1/11 399.793 9.39E+1 0/2/11 265.719 5.02E+1 4/4/5 273.235 2.68E+1 5/2/6
F18 916.093 2.41E1 2/4/7 914.921 1.28E+0 4/4/5 911.417 3.55E+1 6/6/1 915.487 6.10E1 3/4/6
F19 915.975 3.43E1 2/3/8 914.200 8.78E1 6/4/3 916.077 3.00E+1 1/2/10 915.328 6.18E1 3/2/8
F20 915.893 3.08E1 1/4/8 914.694 1.56E+0 4/6/3 910.145 2.74E+1 8/5/0 915.551 9.70E1 2/3/8
F21 860.000 1.1E13 4/9/0 870.000 5.48E+1 0/4/9 860.022 1.2E1 2/0/11 860.000 2.2E13 4/9/0
F22 1262.959 9.76E+0 2/2/9 1242.860 1.54E+1 6/6/1 1325.995 2.93E+1 0/2/12 1256.220 1.38E+1 2/3/8
F23 894.165 2.72E4 6/7/0 907.569 7.34E+1 2/1/10 894.251 4.72E1 2/2/9 894.164 7.53E5 5/8/0
F24 460.000 1.2E12 1/12/0 460.000 0.00E+0 1/12/0 460.000 4.2E10 1/12/0 460.000 0.00E+0 1/12/0
F25 470.800 8.25E1 7/2/3 470.416 7.01E1 8/4/1 475.174 4.47E+0 2/0/11 471.486 6.52E1 3/4/6
H/APv 4.21E7 (++) 5.13E5 (++) 3.65E8 (++) 0.0055 (++)
W/Pav 3.04/5.88 3.60/5.48 2.92/8.04 3.92/4.84
1
8
0
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
Table 4
The mean tness values (mean) and its standard deviations (std) for 50-dimensional problems. +/=/ shows than how many of 13 competing methods the particular algorithm is signicantly better/equal/
signicantly worse, according to Wilcoxon rank-sum test at 5% signicance level. W/Pav how many times, averaged over all problems, the particular algorithm is signicantly better/signicantly worse than other
algorithms according to Wilcoxon test. H/APv adjusted p-values according to Holm test procedure with AM-DEGL as a control method; ++ means that AM-DEGL is signicantly better than the competitive algorithm,
== means that there is no signicant difference between AM-DEGL and competitive method at 5% signicance level, according to Holms procedure.
AM-DEGL DE DE-SG DEGL SADE
Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F1 450.000 2.64E7 2/11/0 293.680 7.95E+1 0/0/13 449.902 4.20E2 1/0/12 450.000 8.94E8 2/11/0 450.000 5.43E8 2/11/0
F2 450.000 1.29E7 11/2/0 8.58E+04 6.94E+3 0/0/13 428.595 1.44E+1 3/1/9 450.000 2.47E8 11/2/0 450.000 1.39E5 11/2/0
F3 3.89E+04 5.35E+4 13/0/0 5.37E+08 6.83E+7 0/0/13 1.03E+06 4.40E+5 5/2/6 1.86E+05 8.38E+4 11/1/1 1.84E+05 7.00E+4 11/1/1
F4 418.021 1.05E+2 12/0/1 1.05E+05 9.37E+3 0/0/13 319.150 3.72E+2 5/2/6 449.998 3.67E3 13/0/0 237.338 2.76E+2 10/1/2
F5 305.269 5.32E+2 13/0/0 1.78E+04 8.63E+2 0/0/13 3463.695 6.42E+2 5/3/5 1231.147 6.45E+2 12/0/1 2120.398 5.70E+2 11/0/2
F6 399.550 9.89E+0 12/0/1 3.59E+07 5.40E+6 0/0/13 897.727 1.08E+2 1/0/12 390.930 1.71E+0 13/0/0 436.096 3.60E+1 6/4/3
F7 179.995 7.93E3 7/4/2 3999.297 2.59E+2 0/0/13 173.131 4.40E+0 2/0/11 179.994 9.41E3 6/7/0 179.994 1.15E2 5/8/0
F8 119.945 3.43E2 13/0/0 118.866 4.51E2 0/11/2 118.872 3.61E2 0/11/2 118.873 2.14E2 0/11/2 118.868 5.44E2 0/11/2
F9 296.287 7.61E+0 7/0/6 21.667 1.10E+1 0/0/13 265.928 5.89E+0 3/2/8 265.892 1.57E+1 3/2/8 329.998 2.47E3 12/0/1
F10 256.589 3.88E+1 13/0/0 125.001 1.49E+1 0/0/13 57.282 1.72E+1 4/1/8 126.071 1.28E+2 4/5/4 188.654 3.34E+1 6/1/6
F11 134.273 5.97E+0 12/1/0 162.863 1.39E+0 0/0/13 150.797 1.36E+0 5/1/7 131.208 2.42E+1 7/6/0 146.633 1.49E+0 9/2/2
F12 1.08E+04 8.73E+3 9/4/0 2.01E+06 1.38E+5 0/0/13 9.56E+04 4.94E+4 3/1/9 2.41E+04 2.28E+4 4/4/5 1.21E+04 8.90E+3 7/6/0
F13 125.660 9.39E1 12/0/1 69.971 3.99E+0 0/0/13 120.468 5.64E1 6/1/6 109.071 8.59E+0 1/0/12 122.097 4.04E1 7/2/4
F14 277.630 4.05E1 8/3/2 276.851 1.43E1 0/1/12 277.371 2.24E1 6/2/5 277.455 3.06E1 6/4/3 277.456 1.60E1 7/3/3
F15 440.206 9.63E+1 6/6/1 620.380 6.24E+0 0/0/13 520.016 8.31E3 1/0/12 441.202 8.49E+1 5/7/1 506.701 5.08E+1 2/2/9
F16 181.144 2.29E+1 12/0/1 450.983 8.81E+0 0/0/13 300.083 6.28E+0 4/1/8 297.440 1.34E+2 4/7/2 218.645 2.18E+1 7/4/2
F17 276.789 8.08E+1 11/2/0 486.354 1.21E+1 0/0/13 363.241 1.34E+1 5/0/8 388.075 8.66E+1 2/2/9 293.431 1.56E+1 7/2/4
F18 927.578 2.75E+0 7/0/6 967.641 2.60E+0 0/0/13 922.038 3.17E+1 7/3/3 913.364 1.03E+2 9/3/1 916.752 3.87E+1 7/5/1
F19 927.248 2.63E+0 8/0/5 968.580 2.64E+0 0/0/13 918.738 3.68E+1 10/3/0 928.048 2.54E+1 5/3/5 920.103 3.75E+1 9/4/0
F20 923.279 2.16E+1 11/2/0 967.125 2.94E+0 0/0/13 929.250 2.24E+1 4/4/5 931.005 1.40E+1 3/4/6 923.335 3.12E+1 5/6/2
F21 860.000 2.7E13 8/5/0 1392.258 8.55E1 0/0/13 860.009 3.96E3 5/2/6 948.999 1.87E+2 4/1/8 860.000 3.5E13 8/5/0
F22 1265.103 1.11E+1 11/2/0 1319.339 4.74E+0 0/1/12 1298.403 1.46E+1 2/2/9 1304.029 2.37E+1 0/5/8 1277.948 1.84E+1 7/3/3
F23 899.122 1.62E4 9/4/0 1384.574 9.87E1 0/0/13 899.126 4.39E3 8/1/4 990.950 1.62E+2 5/0/8 899.122 6.07E5 9/4/0
F24 460.000 0.00E+0 3/10/0 727.428 3.46E+1 1/0/12 481.296 1.08E+2 2/0/11 460.000 1.6E12 3/10/0 460.000 5.9E14 3/10/0
F25 474.191 8.22E1 13/0/0 924.502 3.51E+1 0/0/13 484.122 2.85E+0 3/0/10 477.374 1.70E+0 9/1/3 475.107 9.41E1 11/1/1
H/APv - 2.67E-21 (++) 7.23E-7 (++) 0.0047 (++) 0.0327 (++)
W/Pav 9.72/1.04 0.04/12.44 4.00/7.28 5.68/3.48 7.16/1.92
SspDE JADE AdapSS-JADE DECLS DEahcSPX
F1 450.000 3.05E8 2/11/0 450.000 6.28E8 2/11/0 450.000 6.72E8 2/11/0 450.000 4.18E8 2/11/0 450.000 3.86E8 2/11/0
F2 449.462 3.22E1 6/0/7 449.942 1.34E1 8/2/3 448.748 1.24E+0 5/0/8 449.762 4.89E1 7/0/6 164.376 2.83E+2 2/0/11
F3 6.45E+05 2.92E+5 8/1/4 1.15E+06 3.66E+5 4/3/6 1.14E+06 2.99E+5 4/3/7 7.31E+05 2.81E+5 8/1/4 4.56E+06 1.42E+6 2/0/11
F4 1361.086 9.65E+2 4/0/9 101.276 3.53E+2 8/0/5 374.127 4.33E+2 5/2/6 556.923 7.45E+2 5/2/6 4786.467 1.77E+3 2/1/10
F5 5361.632 8.37E+2 3/0/10 3632.487 5.43E+2 5/2/6 2985.857 3.70E+2 8/2/3 3641.678 5.92E+2 5/2/6 2747.168 5.82E+2 9/1/3
F6 457.461 3.19E+1 4/1/8 425.684 2.16E+1 6/3/4 426.804 2.16E+1 7/2/4 427.663 3.02E+1 6/2/5 509.278 1.15E+2 3/2/8
F7 179.995 9.33E3 5/6/2 179.998 6.08E3 8/5/0 179.998 5.95E3 8/5/4 179.991 1.37E2 5/4/4 179.984 1.68E2 4/1/8
F8 118.869 3.78E2 0/11/2 118.870 3.31E2 0/11/2 118.867 3.37E2 0/11/2 118.869 3.53E2 0/11/2 118.862 3.15E2 0/11/2
F9 330.000 3.99E8 12/1/0 174.367 9.92E+0 1/0/12 326.518 1.75E+0 10/0/3 329.702 6.99E1 11/1/1 288.378 1.43E+1 6/0/7
F10 254.397 6.61E+0 10/2/1 196.578 1.32E+1 7/2/4 195.907 1.32E+1 7/2/4 246.409 1.83E+1 10/2/1 17.705 3.01E+1 3/0/10
F11 143.870 1.84E+0 11/1/1 148.883 1.70E+0 7/2/4 149.070 1.44E+0 7/2/4 146.441 2.29E+0 9/2/2 161.354 1.87E+0 1/0/12
F12 2.25E+04 1.34E+4 4/4/5 9394.746 1.06E+4 9/4/0 1.15E+04 1.07E+4 9/4/0 1.28E+04 1.31E+4 7/6/0 1.98E+04 1.72E+4 5/5/3
F13 125.415 2.94E1 10/1/2 115.983 8.89E1 4/1/8 116.256 1.12E+0 4/1/8 127.391 2.01E1 13/0/0 119.892 5.65E+0 6/3/4
F14 277.404 2.12E1 6/3/4 278.021 2.38E1 12/1/0 277.908 1.99E1 12/1/0 277.683 3.11E1 10/1/2 277.031 2.44E1 1/3/9
(continued on next page)
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
1
8
1
Table 4 (continued)
AM-DEGL DE DE-SG DEGL SADE
Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/ Mean Std +/=/
F15 490.695 7.89E+1 3/3/7 446.667 9.80E+1 6/6/1 463.333 8.98E+1 5/6/2 407.348 9.28E+1 10/3/0 454.450 9.48E+1 5/3/5
F16 172.594 2.02E+1 13/0/0 219.504 1.78E+1 7/4/2 220.772 7.47E+0 6/4/3 219.763 6.52E+1 7/2/4 344.450 2.91E+1 3/0/10
F17 255.686 2.40E+1 11/2/0 283.548 1.10E+1 8/3/2 281.919 2.05E+1 8/3/2 295.089 3.33E+1 7/4/2 379.880 2.11E+1 3/1/9
F18 918.215 5.52E+1 10/0/3 930.791 8.44E+0 5/3/5 929.397 1.32E+0 5/1/7 932.074 3.49E+0 3/4/6 925.361 1.70E+0 8/1/4
F19 924.607 4.61E+1 11/0/2 929.138 3.81E+0 5/2/6 929.243 2.25E+0 5/2/6 932.998 3.07E+0 3/2/8 925.328 1.32E+0 9/1/3
F20 927.503 4.11E+1 9/0/4 928.498 4.39E+0 7/2/4 929.986 2.12E+0 6/2/5 932.430 2.73E+0 3/4/6 925.810 1.67E+0 10/2/1
F21 860.000 5.8E13 6/7/0 860.000 3.5E13 8/5/0 860.000 3.5E13 8/5/0 1113.801 2.58E+2 2/5/6 1280.999 1.92E+2 2/1/10
F22 1283.716 2.05E+1 5/5/3 1271.691 1.55E+1 10/3/0 1272.152 1.28E+1 7/5/1 1268.522 1.42E+1 10/3/0 1294.279 8.88E+0 2/2/9
F23 899.133 1.57E2 8/4/1 914.802 8.59E+1 7/1/5 899.122 3.43E5 9/4/0 1245.744 2.13E+2 3/1/9 1342.837 1.12E+2 2/0/11
F24 460.000 1.1E12 3/10/0 460.000 9.7E13 3/10/0 460.000 5.8E14 3/10/0 460.000 1.4E12 3/10/0 460.000 5.10E9 3/10/0
F25 479.342 2.17E+0 4/4/5 479.091 1.11E+0 4/4/5 478.766 6.27E1 4/4/5 478.081 2.07E+0 9/1/3 479.416 1.92E+0 4/4/5
H/APv 0.0052 (++) 0.0055 (++) 0.0055 (++) 0.0025 (++) 2.73E8 (++)
W/Pav 6.72/3.20 6.04/3.36 6.16/3.16 6.32/3.32 3.88/6.60
CoDE EPSDE SFMDE CDE
F1 450.000 4.89E8 2/11/0 450.000 6.93E8 2/11/0 450.000 3.80E6 2/11/0 450.000 1.36E7 2/11/0
F2 375.346 4.22E+1 3/0/10 449.974 2.00E2 9/1/3 870.534 7.13E+3 1/1/11 449.971 8.67E3 8/1/4
F3 2.86E+06 1.09E+6 3/0/10 1.38E+06 6.46E+5 4/2/7 4.68E+05 4.80E+5 10/0/3 2.79E+07 7.14E+6 1/0/12
F4 4932.250 1.89E+3 2/1/10 114.050 2.53E+2 9/0/4 1.59E+04 1.00E+4 1/0/12 367.831 3.02E+1 10/1/2
F5 1.12E+04 9.84E+2 1/0/12 3288.718 7.86E+2 7/2/4 6821.534 1.62E+3 2/0/11 4774.867 6.44E+2 4/0/9
F6 702.219 4.95E+1 3/0/10 409.735 1.79E+1 11/0/2 729.542 4.40E+2 2/1/10 423.880 5.55E1 9/1/3
F7 177.018 7.26E1 3/0/10 179.996 6.85E3 5/8/0 169.119 2.78E+1 1/8/4 179.932 1.32E2 4/0/9
F8 118.879 4.10E2 0/11/2 118.868 3.55E2 0/11/2 119.208 2.35E1 12/0/1 118.870 3.34E2 0/11/2
F9 299.747 2.19E+0 8/0/5 203.869 7.49E+0 2/0/11 254.396 4.73E+1 3/2/8 306.352 1.65E+0 9/0/4
F10 88.301 1.64E+1 1/0/12 9.460 1.19E+1 2/0/11 225.316 7.23E+1 10/2/1 111.091 1.73E+1 5/1/7
F11 157.221 1.46E+0 3/0/10 159.728 1.43E+0 2/0/11 155.391 2.80E+0 4/0/9 150.828 2.10E+0 5/1/7
F12 6.31E+05 6.41E+4 1/0/12 2.03E+05 2.31E+5 2/5/6 6.48E+04 2.27E+5 4/6/3 3.01E+05 3.14E+4 2/1/10
F13 109.704 1.63E+0 2/0/11 112.030 1.00E+0 3/0/10 125.188 1.01E+0 10/1/2 122.019 4.03E1 7/2/4
F14 276.938 1.42E1 1/3/9 276.924 1.74E1 0/4/9 276.985 2.36E1 1/3/9 277.232 1.89E1 5/0/8
F15 506.673 5.07E+1 3/1/9 450.000 9.52E+1 6/7/0 433.764 8.59E+1 3/8/2 423.896 1.00E+2 7/6/0
F16 414.369 1.71E+1 1/0/12 394.278 5.99E+1 2/0/11 236.067 7.13E+1 6/5/2 271.471 1.30E+1 5/1/7
F17 452.550 1.32E+1 1/0/12 405.291 4.62E+1 2/1/10 284.295 8.74E+1 7/6/0 325.528 1.05E+1 6/0/7
F18 931.994 9.00E1 3/3/7 876.721 1.57E+2 12/1/0 949.578 3.00E+1 1/0/12 936.087 1.98E+0 2/0/11
F19 931.634 9.67E1 4/1/8 925.393 3.46E+0 9/1/3 945.795 3.00E+1 1/0/12 935.784 2.22E+0 2/1/10
F20 931.579 8.76E1 3/3/7 885.407 1.56E+2 11/2/0 951.914 1.22E+1 1/0/12 936.414 1.85E+0 2/1/10
F21 1346.459 1.32E+2 1/0/12 1145.780 2.54E+2 3/5/5 924.787 1.12E+2 4/2/7 860.000 3.5E13 8/5/0
F22 1303.479 5.82E+0 1/1/11 1280.048 1.45E+1 6/2/5 1283.507 2.09E+1 6/4/3 1289.949 7.02E+0 4/2/7
F23 1377.713 9.03E1 1/0/12 1276.835 1.92E+2 3/1/9 929.046 7.76E+1 6/0/7 899.123 1.11E4 9/4/0
F24 460.000 1.06E6 3/10/0 460.000 5.8E14 3/10/0 768.498 1.20E+2 0/0/13 460.000 5.8E14 3/10/0
F25 479.724 2.33E+0 4/4/5 475.036 1.06E+0 11/1/1 733.659 1.61E+2 1/0/12 514.958 4.05E+0 2/0/11
H/APv 4.60E12 (++) 4.79E5 (++) 3.04E8 (++) 6.65E5 (++)
W/Pav 2.32/8.72 5.04/4.96 3.96/6.64 4.84/5.76
1
8
2
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
Table 5
The mean tness values of different variants of the proposed algorithm on 30-dimensional problems. W/Pav informs how many times, averaged over all problems, the particular algorithm is signicantly better/
signicantly worse than other algorithms according to Wilcoxon test.
AM-
DEGL
F-SADE F-xed RA-NMA RA SADE-
glob-only
LM-
greedy
Loc-Only-
best
Loc-
every
p-strat-
from-g2
p-strat-
SADE
p
6
p 10 p 40 NMA
-closest
Mean
F1 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00
F2 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00
F3 1780.31 7948.30 312.35 308.01 2382.32 120.69 10130.32 6.16 301.46 721.97 350.44 202.19 12.44 1319.68 29.59
F4 450.00 285.52 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 450.00 449.87 449.97 450.00 450.00
F5 309.98 1215.05 310.00 308.33 307.92 309.64 309.80 309.98 309.99 309.97 309.93 309.28 309.77 310.00 309.99
F6 391.59 391.77 392.70 399.28 397.60 401.90 390.80 392.67 392.26 390.63 392.01 394.48 392.68 392.98 391.04
F7 180.00 179.97 180.00 179.99 179.99 179.99 179.99 179.99 180.00 180.00 180.00 179.99 180.00 180.00 179.99
F8 119.98 119.98 119.98 119.87 119.67 119.99 119.98 120.00 119.98 119.98 119.98 120.00 119.99 119.96 119.99
F9 326.66 328.03 324.30 321.47 324.00 329.48 329.91 326.90 326.00 326.30 326.29 320.43 321.56 329.49 327.98
F10 297.07 274.12 295.64 285.94 271.34 291.99 296.07 272.96 297.89 294.50 299.39 294.68 293.92 290.26 287.91
F11 113.84 112.02 113.21 115.93 118.85 111.69 107.29 113.22 111.87 112.76 112.31 112.47 112.60 113.77 115.97
F12 1915.83 2274.20 2815.13 5539.57 8293.43 2699.04 4713.38 3901.60 2020.65 2515.67 3248.81 3052.01 2279.87 1904.84 1511.68
F13 128.03 127.99 128.03 127.87 127.59 128.09 128.03 127.98 128.20 128.12 127.99 128.02 128.07 127.94 128.19
F14 287.29 287.88 287.35 287.21 287.31 287.22 287.56 287.27 287.37 287.22 287.27 287.19 287.29 287.54 287.34
F15 520.00 524.18 520.00 500.07 510.08 523.52 470.00 506.67 506.72 540.07 516.67 490.07 507.69 513.49 490.07
F16 197.90 230.07 211.29 217.58 205.05 193.94 225.04 242.83 229.82 225.04 208.91 229.18 210.15 246.72 212.56
F17 220.79 241.13 239.01 258.37 340.39 219.28 222.95 269.48 235.35 216.26 231.61 226.02 228.86 221.74 272.38
F18 914.44 887.72 914.65 911.33 914.95 911.52 914.58 904.40 897.27 911.21 911.18 908.13 907.80 911.12 911.23
F19 900.70 887.75 907.70 911.48 914.70 908.06 914.33 911.30 910.83 907.43 907.72 904.87 900.51 911.48 908.03
F20 914.88 914.11 914.70 914.88 914.64 911.56 914.70 911.36 914.53 907.81 911.18 901.08 908.07 914.67 908.35
F21 860.00 911.91 870.00 860.00 860.00 860.00 860.00 860.00 860.00 880.00 860.00 860.00 880.00 870.00 860.00
F22 1230.94 1262.75 1225.23 1234.46 1233.34 1237.18 1235.19 1231.76 1235.04 1235.36 1228.29 1233.73 1235.62 1229.85 1229.81
F23 894.16 894.18 894.16 894.16 894.16 894.16 907.59 894.16 894.16 894.16 894.16 921.02 894.16 894.16 894.16
F24 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00 460.00
F25 469.66 474.93 469.39 469.92 469.65 469.53 471.02 469.65 469.67 469.78 469.65 470.18 469.92 469.51 469.82
W/Pav 2.68/1.04 3.20/5.64 2.88/1.36 1.28/4.56 1.24/5.72 2.68/2.04 3.44/2.52 2.16/1.96 2.60/1.32 3.08/1.20 2.40/1.12 2.32/ 3.00 2.48/2.08 2.68/2.08 2.64/2.12
A
.
P
.
P
i
o
t
r
o
w
s
k
i
/
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
s
2
4
1
(
2
0
1
3
)
1
6
4

1
9
4
1
8
3
The +/=/ values refer to: + the number of algorithms that perform signicantly worse than the considered algorithm;
= the number of algorithms with performance that does not differ signicantly from the performance of the considered
algorithm; the number of algorithms that perform signicantly better than the considered algorithm. Note that uni-
modal problems F1 and F2 are easily solved by most DE approaches. It was found that AM-DEGL is among the quickest meth-
od for such problems, hence it could be awarded by additional + just for quick exploitation. Hence, to not overestimate the
performance of the proposed method, it was ultimately decided that in this paper when Wilcoxon rank-sum test is used the
performance of two algorithms is considered equal (=) if their mean results differ by less than 10
3
.
F3
0E00
3.E+06
6.E+06
9.E+06
fitn
e
s
s
AM-DEGL DE DE-SG DEGL
SADE SspDE JADE AdapSS-JADE
DECLS DEahcSPX CoDE EPSDE
SFMDE CDE
F3
0.E+00
3.E+06
6.E+06
9.E+06
0 100000 200000 300000
n.f.c.
f
i
t
n
e
s
s
F10
-350
-200
-50
100
0 100000 200000 300000
n.f.c.
f
i
t
n
e
s
s
F12
0.E+00
1.E+05
2.E+05
3.E+05
0 100000 200000 300000
n.f.c.
f
i
t
n
e
s
s
F17
200
400
600
0 100000 200000 300000
n.f.c.
f
i
t
n
e
s
s
F19
880
930
980
n.f.c.
f
i
t
n
e
s
s
F22
1200
1300
1400
1500
0 100000 200000 300000
0 100000 200000 300000
n.f.c.
f
i
t
n
e
s
s
Fig. 3. Mean convergence characteristics of 14 compared algorithms for selected 30-dimensional problems.
184 A.P. Piotrowski / Information Sciences 241 (2013) 164194
In Tables 24 for each problem only the results obtained by the algorithm that is signicantly better according to Wilcoxon
test than all 13 competing methods are bolded, what is rare. Frequently at least two out of a large number of algorithms
perform similarly for a given problem and the difference in their performance is not statistically signicant the difculty
of nding an algorithm that perform better than all others for many problems is also seen in comparisons presented in
[101,94,43,36,44,54]. With respect to that in the last row in Tables 24 (denoted W/Pav) for every algorithm the averaged
over all problems number of DE variants that perform, according to Wilcoxon rank-sum test, worse (+) / better () then
the considered algorithm is shown.
According to rank-based Holm procedure in case of 10-dimensional problems AM-DEGL is signicantly better (at 5% sig-
nicance level) than 5 competitive methods. The proposed AM-DEGL is neither signicantly better nor signicantly worse
than other 8 algorithms. When 30-dimensional problems are compared, AM-DEGL is signicantly better than 12 methods,
SADE is an exception. However, at 10% signicance level (not shown in Table 3) AM-DEGL is signicantly better than all 13
other DE variants. In case of 50-dimensional problems AM-DEGL is signicantly better than all competitive algorithms at 5%
signicance level.
However, when Wilcoxon test is applied for inter-comparison between all methods for each problem separately, usually
the differences between best algorithms are not statistically signicant. Only for 3 out of 25 10-dimensional problems some
algorithms (AM-DEGL, DE-SG and DEGL, each only once) signicantly outperform all other methods. In case of 30-dimen-
sional problems AM-DEGL and DEGL signicantly outperform all other algorithms for 3 and 2 problems, respectively. For
50-dimensional problems AM-DEGL signicantly outperforms all other algorithms for 5 problems, DEGL and SspDE for one.
Comparing the averaged over all problems number of algorithms that perform worse/better than the considered algo-
rithm by means of Wilcoxon test, the best results for 10-dimensional problems are obtained by means of DE-SG (it is on aver-
age signicantly better than 6.76 algorithms) and SADE (it is on average signicantly worse than 0.84 algorithms). AM-DEGL
nish the second winning (6.24) and second the most rarely loosing (1.6) method. For 30-dimensional problems AM-DEGL is
the best method, on average better than 8.00 and worse than 1.24 algorithms. The second-best method (SADE) is on average
better than 6.12 and worse than 1.84 algorithms. For 50-dimensional problems AM-DEGL is also the best algorithm, on aver-
age better than 9.72 and worse than 1.04 competitors. The second best method (SADE) is on average better than 7.16 and
worse than 1.92 algorithms.
One may note that the higher dimensionality, the larger difference between AM-DEGL and other DE algorithms, both
according to rank-based Holm procedure and Wilcoxon test. Considering 30- and 50-dimensional variants of CEC 2005
problems, AM-DEGL performance is among the best methods for all test functions except F9, F18 and F19. For none of
considered problems AM-DEGL is signicantly worse than 6 out of 13 algorithms. Among 30- and 50-dimensional prob-
lems AM-DEGL performance is not signicantly worse than any among 13 algorithms for 14 and 15 test functions,
respectively.
It is not easy to answer the question for what kind of problems AM-DEGL performance is the best, or for which problems
is the poorest. For uni-modal problems (F1F5) AM-DEGL and DEGL are the two best methods, irrespective of problem
dimensionality. This indicates that AM-DEGL inherited the possibility of quick detection of the minimum from DEGL, one
of its ancestor methods, enhanced by using a local search algorithm. In case of basic multimodal functions AM-DEGL perfor-
mance is diversied for 50-dimensional problems F8 and F10 and 30-dimensional problems F7 and F8 it is signicantly
better than all other methods. Its performance is also among the best methods for F11 and F12, irrespective of dimension-
ality. However, comparing to other DE variants, AM-DEGL does not work well for 10-dimensional variant of F6 and especially
for problemF9 (shifted Rastrigins function), irrespective of dimensionality. One may note that for F9 problemboth DEGL and
SFMDE (which, like AM-DEGL, also uses NMA algorithm as local search method) also perform poor, while performance of
SADE is among the best. This indicate that for non-rotated shifted Rastrigins function the more exploitative features of
AM-DEGL may lead to the premature convergence. However, when rotated variant of shifted Rastrigins function is used
(problem F10), the performance of AM-DEGL is the best among tested methods, especially for higher-dimensional problems,
conrming good balance between time-consuming adaptive approaches that strength exploration, and exploitation capabil-
ities inherited from DEGL and SFMDE. For expanded multimodal (F13F14) and composition functions (F15F25) the perfor-
mance of AM-DEGL is among the best methods, with exceptions of 10- and 30-dimensional variants of problems F14, F18,
F19 and F20, and 50-dimensional variants of problems F18 and F19, for which AM-DEGL ranked middle positions among DE
algorithms. However, note that composition problems F18, F19 and F20 are composed from the same functions with differ-
ent parameterization, and for this three problems none of parent methods of AM-DEGL perform well. On the other hand,
for seven among ten 50-dimensional composition functions no algorithm performs better than AM-DEGL. It also seems that
AM-DEGL is not as signicantly affected by rotation as other DE variants according to Wilcoxon test, AM-DEGL outperform
other DE variants much more frequently for shifted rotated Rastrigins function (F10) than for shifted Rastrigins function
(F9) and for rotated composition function (F16) than its non-rotated variant (F15), especially when 30- and 50-dimensional
versions are considered. Such behavior is not observed for DEGL or SADE. As most DE methods are known for its weak ability
to cope with rotated problems due to the nature of crossover, the better performance for rotated problems by means of AM-
DEGL may be the reason of two factors the use of NMA as local search method (note that the performance of SFMDE, com-
paring to other DE variants, is also relatively better for problems F10 and F16, then for F9 and F15), or the quick adaptation of
probability of using different mutation strategies (of which one is not followed by crossover; although SADE also uses this
strategy, the adaptation of probability of using different strategies in SADE is much slower than in AM-DEGL).
A.P. Piotrowski / Information Sciences 241 (2013) 164194 185
Fig. 3 shows the convergence characteristics of all compared algorithms for 6 selected 30-dimensional test functions. It
may be seen that AM-DEGL is among the quickest methods for each problem. It is frequently slightly slower only to DEGL,
but in vast majority of cases it reaches signicantly better results than all its ancestral approaches.
The obtained results, especially for 30- and 50-dimensional problems, may suggest that AM-DEGL is more reliable meth-
od than its ancestral algorithms and other DE variants (but note the discussion in Section 6). In the future it is planned to test
AM-DEGL on high-dimensional problems. However, this may require a bit different settings of the population size, as in DE
algorithms used for large scale optimization usually NP D [145,157,16]. As a result some novel method will be needed to
determine D + 1 points to be used by NMA simplex, what may lead to an important modication of AM-DEGL.
5.4. Motivation of different parts of AM-DEGL
In this sub-section it is veried howthe performance of AM-DEGL depends on some of its components and control param-
eter values. The comparison of AM-DEGL with a number of its modied variants is presented for 30-dimensional CEC 2005
problems. The following modications of AM-DEGL algorithm dened in Section 4 are considered:
1. F-SADE the scale factor F
i,g
is generated from N(0.5, 0.3), following SADE;
2. F-xed the scale factor F
i,g
is xed to 0.8, following DEGL;
3. RA-NMA two local search methods, RA [117] and NMA may be used by each individual that is better than all its
neighbors, each with the same probability (LSp = 0.005). The maximum number of function calls by each local search
method is kept 100D as in AM-DEGL. In this variant the probability of commencing local search and the average num-
ber of function calls used by the local search methods is twice higher than in AM-DEGL;
4. RA only RA is used as local search method, instead of NMA. The probability of commencing local search and the max-
imum number of function calls used by local search algorithm are the same as in AM-DEGL;
5. SADE-glob-only the Local mutation model is not used by four strategies that were applied in SADE, it is used by
DEGL-based strategy only;
6. LM-greedy irrespective of Global mutation strategy, as Local mutation model the most greedy DE/best/1 strategy
(Eq. (4)) is applied, where instead of x
best,g
the x
Lbest,i,g
is used. This variant is considered to check the possible advan-
tage of highly exploitative local mutation model;
7. Loc-only-best the local search may be commenced by the best individual in the population only, following [98].
Because on average 45 individuals have a chance to commence local search by AM-DEGL for 30-dimensional prob-
lems, the probability of commencing local search in this variant is increased by a factor of 4 (LSp = 0.02), to keep the
number of local searches per run roughly similar to the original AM-DEGL;
8. Loc-every the local search may be performed by each individual in the population. Because on average 45 individ-
uals have a chance to commence local search by AM-DEGL for 30-dimensional problems, the probability of commenc-
ing local search by this variant is divided by 30 (LSp = 0.000167), to keep the number of local searches per run roughly
similar to the original AM-DEGL;
9. p-start-from-g2 probabilities of using different strategies are adapted from the second generation, without warm-
up period;
10. p-start-SADE probabilities of using different strategies are adapted following equations proposed in SADE (Eq. (14) in
[115] with LP = 1), instead of Eqs. (22) and (23);
11. p 6 the value of p% neighborhood is reduced to 6%;
12. p 10 the value of p% neighborhood proposed in DEGL is used (i.e. 10%);
13. p 40 the value of p% neighborhood is increased to 40%;
14. NMA-closest a different procedure of choosing the initial D + 1 points for NMA algorithm is used. In this variant the
x
Lbest,i,g
and its D closest individuals in the search space (according to Euclidean norm) are chosen. The motivation
beyond this scenario is that such chosen D + 1 points may facilitate quick exploitation of the local optimum. However,
the explorative features of NMA are not well exploited in this procedure.
The mean performance of AM-DEGL and its 14 variants are presented in Table 5. For each version the averaged over all
problems number of other variants that are, according to Wilcoxon test, statistically signicantly better/worse is presented
in W/Pav row. It is seen that on average the performances of AM-DEGL dened in Section 4 and most of its variants are com-
parable. As in each considered variant only one component of AM-DEGL has been modied, this conrms that the small
changes to AM-DEGL do not destroy its ability to nd the optima of various problems. It also conrms that the good perfor-
mance of AM-DEGL is not an effect of a single novelty proposed, but rather a successful combination of well known elements
of already developed DE variants, what was the intention of the present paper. However, a few AM-DEGL variants perform
poorer than the others:
1. The use of greedy local search RA algorithm instead of, or together with NMA (RA and RA-NMA) disturbs the search ability
of AM-DEGL. Although benecial in some cases (F3, F15, F18) it usually diminish the exploration capabilities of the
method.
186 A.P. Piotrowski / Information Sciences 241 (2013) 164194
2. The method of generating scale factor borrowed from SADE (F-SADE) turns out un-trustworthy when applied in AM-
DEGL. Although the number of variants from which F-SADE is signicantly better than other AM-DEGL variants is the
largest (also larger than AM-DEGL), the number of variants that are signicantly better than F-SADE is also much higher
than noted for any other variant. This suggests that the scale factor generation method used in SADE, if implemented
within AM-DEGL, is well suited for selected problems only. Note that in SADE the scale factor is generated from
N(0.5, 0.3), hence 95% of its values falls within a large interval [0.1, 1.1], when in two other ancestor algorithms of
AM-DEGL (DEGL and SFMDE) scale factors around 0.70.8 are suggested.
3. The use of greedy local mutation model (LM-greedy) leads to similar, although mitigated, effect as noted for F-SADE var-
iant and is not recommended.
AM-DEGL has been selected as the best among compared novel variants because the averaged over all problems number
of algorithms that perform signicantly better than AM-DEGL is the smallest, hence this method outperform other variants
in terms of reliability. Note that other variants of AM-DEGL may be obtained easily by modifying only one of AM-DEGL ele-
ments (see pseudocode in Fig. 1).
6. Empirical results and No Free Lunch theorems for optimization
In the present paper a novel DE algorithm has been proposed and its good performance was empirically veried for 25
benchmark problems. Although this is a very frequent approach, one may ask if it is justied to propose a novel heuristic
and show its usefulness based on small number of benchmark problems if according to NFL [149,150,73] the expected per-
formance of all possible algorithms averaged over all possible problems is equal. The fact that in many papers NFL is simply
ignored or just mentioned with a brief comment has been criticized in the literature [99,73]. Because understanding the lim-
itations of the novel methods is no less important than conrming their good performance for some problems, this section is
devoted to a short literature review of NFL and its interpretations, followed by simple experiment that empirically shows the
limitations of DE algorithms. For the more detailed discussion of NFL and its proofs the reader is referred to the classical lit-
erature [149151,37,38,120,73,65,66,7,147,118]. Below the notation proposed in [150,151] is used.
6.1. No Free Lunch theorem
To prove NFL Wolpert and Macready [149151] specied that:
1. NFL aims at combinatorial optimization, where both the search space X and the space of tness values Y are nite. As
pointed out in [150], when digital computers are used these restrictions are always met in practice, even if one claims
(like in this and many other papers) that the algorithms are developed for continuous optimization of real valued func-
tions. As a result the size of the space [X[ and [Y[, as well as the size of the space of all possible problems [F[ = [Y[
[X[
is nite,
although frequently very large.
2. Each heuristic algorithm samples the elements x from X and evaluates their tness y = f(x) (x X, f(x) Y) according to the
rule that is based only on information about previously visited xs and their objective function values. NFL is valid for both
deterministic and non-deterministic algorithms [150,38]. The important, usually not fullled assumption is that an algo-
rithm does not re-visit the sampled elements. Hence NFL, while comparing algorithms, refers only to the m distinct ele-
ments as a sample d
m
= d
x
m
(1); d
y
m
(1)
_ _
; . . . ; d
x
m
(m); d
y
m
(m)
_ _ _ _
, where d
x
m
(i) is the i-th sampled distinct element within
sample of size m and d
y
m
(i) is its associated tness value. d
m
does not contain the re-visited cases. This may affect the com-
parison if algorithmin fact re-visits some elements or the problems with noise in tness are considered (e.g. F4 and F17 in
[126]).
3. An algorithm performs mapping from a sample d
m
to d
x
m1
(m 1) that is dened by the probability distribution
P
m
d
x
m1
(m 1) = x[d
m
_ _
. Hence, the algorithm a is a specication of P
i
for different values of i [150,73,151].
With the above assumptions, NFL says that the averaged over all possible problems f (f F), specied with distribution
P(f) the expected performance (measured by a performance measure U(d
y
m
)) of an algorithm a, after visiting m distinct ele-
ments x, is E(U[m; a) =

f F
U d
y
m
_ _
P d
y
m
[f ; m; a
_ _
P(f ). When
1. P(f) = 1/[F[ (NFL version proved in [150]), or when
2. P(f) is uniform over any set of functions F
s
(F
s
# F) that is closed under permutations (c.u.p.) and P(f) equal 0 if f R F
s
(so-
called Sharpened NFL proved in [120], see also [38,66,151]),
then the expected performance is independent of a, in other words the expected performance do not depend on the algo-
rithm used.
When one is interested in U d
y
m
_ _
= min
i
d
y
m
(i)
_ _
, what is frequently the case of EAs, then m < [X[ is assumed. The explana-
tion of NFL is very simple [150,24,73,70]: if one considers all possible problems (by summing over all possible fs or all prob-
lems that are c.u.p. with equal P(f)), the already visited elements do not provide any information about the unvisited ones (all
A.P. Piotrowski / Information Sciences 241 (2013) 164194 187
combinations of tness values are equally possible for unvisited elements). It is worth noting that in [66,125] NFL has also
been proved for some specic non-uniform P(f)s.
6.2. Implications, limitations and interpretations of NFL
Implications of NFL for application of heuristic optimization methods to problems that may be of any practical interest to
anyone have been studied in a number of papers from different points of view. They led to the conclusions that NFL does not
dismiss the applicability of heuristic optimization algorithms to large number of practical problems, and hence do not rule
out the need to develop novel methods.
According to [73] NFL is sometimes considered by practitioners simply as a lack of universal applicability of any heuristic
method. Based on intuition and empirical experience it is easy to agree that some algorithms perform better for some sets of
problems than the others. Although from a practical point of view this seems correct when one aims at solving a chosen
problem, this attens the NFL ndings. According to NFL the expected performance of any heuristic algorithm is the same.
As pointed out in [150] if one aims at minimization, purely random sampling, hill descending or hill climbing perform
equally well as any other heuristic. However, few people search for minimum by random sampling or using hill climbing algo-
rithm (or any other method that aims at maximization). This issue will be addressed once more in Section 6.3.
In a few papers the question whether NFL can be generalized to continuous functions was asked. In [7] it is shown that
NFL does not hold for continuous functions, only the weaker form of NFL does hold. Streeter [125] argued that uniform dis-
tributions cannot be assumed when F is innite, and that NFL would hold only when specic restrictions on P(f) are set. How-
ever, Rowe et al. [118] reinterpreted NFL using set theory approach showing that NFL does hold even for innite domains.
The rst suggestion that the differences in performance between algorithms may be found despite NFL was given in [150]
and refers to the possible head-to-head minimax distinction between two different deterministic algorithms. Although DE is
considered as a stochastic method, one must note that in [150] it is argued that when pseudorandom numbers are generated
given a seed, the algorithms become in fact deterministic. Head-to-head minimax distinction is dened as follows: if for
some minimization problems f
k
F the difference between expected performance of algorithms a and b is not lower than
k (k > 0) but does not exist any problem f
kk
F for which the difference between expected performance of algorithms b
and a is not lower than k, then algorithm b is minimax superior to algorithm a. In [150] it is empirically shown that it is pos-
sible to nd minimax distinctions between algorithms for some specied small [X[, [Y[ and some specic m and U. This means
that, although the expected performance of two different algorithms is equal for all possible problems, one algorithm may
perform much better than the other for a few problems (the difference in their performance is at least equal k), when the
other algorithm may be only marginally better, but for large number of problems. As marginal differences between perfor-
mance of different optimization methods are rarely important in practice and are frequently ruled out (for example when the
statistical tests are used to compare the results obtained by different algorithms), the lack of NFL for minimax distinction
between algorithms may be important from practical point of view. Unfortunately, in the mentioned example for which
minimax distinction is shown in [150] the overlaps between elements x visited by algorithms a and b within d
m
samples
are noted and U d
y
m
_ _
is of limited interest. Such minimax distinction are a kind of a priori distinctions between algorithms
for discussion and another example of such distinction the reader is referred to [73].
Probably the most important issue regarding NFL is that most among all problems (F) are unlikely to be of interest to any-
one [77,66]. NFL holds when all possible problems have the same P(f) [151], or when all problems within a subset of prob-
lems that are c.u.p. (F
s
) have the same P(f) [120]. As shown in [65,66], the fraction of all possible problems that are c.u.p.
converge to zero very quickly when [X[ is increasing. According to [150] when real-valued problems are implemented on dig-
ital machines, [X[ and [Y[ are very large, hence the probability that a randomly chosen subset of problems is c.u.p. converges
to zero. Nonetheless, for large [X[ and [Y[ the total number of subsets of problems that are c.u.p. is very large. The question
arise, can some relevant classes of problems be c.u.p.? In [66] it is proven that if one requires that for all f F
ss
the elements x
a
and x
b
for which f(x
a
) is a global minimum and f(x
b
) is a global maximum cannot be neighbored, then F
ss
is not c.u.p. As prob-
lems for which the global minimum and the global maximum are neighbored are not important in practice, at least because
they would not allow for robust solutions (on robust optimization see [10]) they may be excluded from consideration. In-
stead of excluding only problems for which global minima and maxima are neighbored, one may set weaker constraints
[24], for example by specifying the maximum allowed variation of f(x
n
) (when x
n
is any neighbor of x) in the neighborhood
of each x. For discussion on other classes of problems for which NFL does not hold the reader is referred to [65,125,66].
According to above discussion NFL only holds for subsets of problems that include functions that cannot be of interest to
anyone. It was also shown that some algorithms (e.g. sub-median seeker and sub-threshold-seeker) outperform some others
(e.g. random search) for some classes of functions that are not c.u.p. [24,146,70].
On the other hand, a Focused No Free Lunch (FNFL) [147] should be considered as a caution of over-enthusiastic claims
that a subsets of problems that may be of interest to anyone are not c.u.p. FNFL is based on Sharpened NFL [120] and show
that when limited number of function evaluations (m [X[) is allowed (what is frequent in EAs) the performance of a number
of specic algorithms averaged over a set of problems F
foc
, such that F
foc
F
s
# F, is equal. All problems within F
foc
may be of
practical importance. Although, contrary to NFL, the FNFL is valid only for some specic algorithms compared on specic m
and F
foc
, in practical applications m [X[ and one compares the performance of limited number of algorithms that bear some
similarities on a small number of benchmark problems that are articially constructed from a simple functions. Hence, FNFL
is at least a warning that when algorithms are compared on a small number of problems that belong to F
foc
, another small
188 A.P. Piotrowski / Information Sciences 241 (2013) 164194
number of problems that also belong to F
foc
and are important from practical point of view may exist, for which the perfor-
mance of compared algorithms would be opposite.
6.3. An alluring continuous and differentiable function
Although NFL only holds for subsets of problems among which some cannot be of interest to anyone, some of its impli-
cations suggested in [150] may help understanding the limitations of the heuristic optimization methods. Consider as an
example a very simple version of a continuous, differentiable and scalable problem, which two-dimensional version is de-
picted in Fig. 4:
y =

D
i=1
1
x
10
i
x
i
_ _
; x [0:1; 1000[
D
x
+
i
= 0:1 f (x
+
) ~ D E10 (24)
In the present sub-section all 14 DE algorithms compared in Section 5 are used to nd the minimum of 2-dimensional ver-
sion of this problem (note that for better view the illustration on Fig. 4 is cut off for x
i
> 10). When D = 2, three local minima
exists: shallow local minimum at xs = [1000, 1000] (with f(xs) = 2000) and two local minima with f(x) ~ 1E10 at
x = [0.1, 1000] and x = [1000, 0.1]. The maximum is located at x

= [1.23285, 1.23285] (with f(x

) ~ 2.71226), much clo-


ser to x

than to xs. Maximumnumber of function calls is set to 10,000D, as for problems addressed in Section 5. A few minor
modications of control parameters of some algorithms are needed to apply them effectively for 2-dimensional problems,
because:
DEGL the neighborhood size must be large enough to include at least 2 nearest neighbors for each individual;
SFMDE population size of PSO algorithm (consult [18] for details) cannot be higher than SFMDE population size (here
NP = 5D = 10);
JADE and AdapSS-JADE the subset of p100% individuals cannot be empty;
DECLS, DEahcSPX and CoDE NP cannot be 1D as DE cannot work with just two individuals (it is increased to 5D);
In addition to DE methods, a purely random sampling is applied. The 30-run averaged results are reported in Table 6. The
mean performance of random sampling is signicantly better than the mean performance of all tested DE algorithms. Only
four DE methods were able to locate a local minimum with tness value ~ (10)
10
once (hence the averaged tness of
3.3(10)
8
). None DE method was able to converge to the global minimum. The reason of poor performance of DE methods
is simple if within initial few generations the method do not nd by chance any solution outside the basin of attraction of
xs, the whole population is quickly attracted to xs due to a concept to follow the better performing individuals, according
to which heuristics are constructed. Random sampling is not attracted anywhere, hence the possibility of nding a solution
Fig. 4. 2-Dimensional illustration of a simple differentiable problem for which the heuristic algorithms are allured to a local optimum. The problem dened
in the text is bounded by 0.1 6 x
1
, x
2
6 1000. On the left picture, on which only points with objective function value (OF) > 30. are shown, it is seen that
heuristic approaches would be allured to the local minimum at high x
1
and x
2
values. On the right picture the full span of objective function values is shown,
with global minimum at [0.1, 0.1]. For the sake of clarity both gures are cut to x
1
, x
2
< 10.
Table 6
Thirty-run averaged results for 2-dimensional variant of the proposed problem.
Random AM-DEGL DE DE-SG DEGL SADE SspDE JADE
Mean 2.8E9 2000 2000 3.3E8 2000 2000 3.3E8 2000
AdapSS-JADE DECLS DEahc-SPX CoDE EPSDE SF-MDE CDE CoDE
Mean 2000 2000 2000 2000 2000 3.3E8 3.3E8 2000
A.P. Piotrowski / Information Sciences 241 (2013) 164194 189
close to other optima by chance is the same throughout the whole run. Note that EAs like CMA-ES [57,58] would perform
even poorer than DE algorithms. This show empirically that for some problems the random sampling may indeed perform
better than heuristic optimization algorithms, as suggested in [150].
Note that although global maximum is much closer to the global minimum than the shallow minimum at xs, if the
problem dened by Eq. (24) is encoded on a computer, there is a very large number of elements between global maximum
and global minimum, which are still far frombeing neighbored (see discussion in Section 6.2). One may imagine a large num-
ber of similar problems (some very simple one-dimensional ones are given in [99]). It may seem that such problems are
rarely considered when different heuristics are compared, but this is not necessarily true. A slightly similar concept of deep
narrow global optimum on the bound was also exploited in problem F8 proposed in [126]. Can random sampling be effec-
tive for minimization of F8 problem? To address this question it was applied for 10-dimensional version of F8, which was
already used to compare DE methods in Section 5.3 (with maximum number of function calls set to 10.000D, the computa-
tions were repeated 30 times). The 30-run averaged performance of random sampling was found to be 119.63, hence it
performs poorer than most DE algorithms (see Table 2). However, random sampling performs almost equally well as SspDE
and better than DECLS and DEahcSPX. This conrms empirically that even some very popular heuristic methods used to solve
some widely known problems may perform poorer than random sampling.
Consider problem dened in Eq. (24) again. Note that even when the scope is to nd the minimum, all 14 DE algorithms
may be applied to search for the maximum of this problem. The question is what will be the lowest objective function value
found accidentally during run when DE variants are used for maximization of the problem dened by Eq. (24). This is an
analog of switching from hill descending (intuitively suggested when one aims at minimization) to hill climbing (intuitively
a wrong concept when minimization is required) as discussed in [150] this intuitive selection is not necessarily correct. The
30-run averaged lowest function values found by the way by DE algorithms during search for the maximum of 2-dimen-
sional problem dened by Eq. (24) are shown in Table 7. They are signicantly lower than values found during minimization
(reported in Table 6) for all DE variants. Interestingly, 5 DE methods used for maximization show even better minimization
performance than random sampling. Hence, for this very simple problem which minimum is to be found, heuristics that aim
at maximization outperform signicantly heuristics that aim at minimization what is not a surprising according to [150].
The results showed in this sub-section should be considered as an empirical illustration of limiting applicability of heu-
ristic methods and may serve as another precaution fromgeneralizing the good performance of any heuristic optimizer. They
may also help avoiding the over-enthusiastic opinion regarding a novel method, which may turn out pernicious for both
authors and readers (see discussion in [27]). The presented discussion also bridges the gap between different concepts in
algorithmic design which were mentioned in the Introduction, as both starting from tabula rasa following Ockhams razor
philosophy [63,19] or building on the experience following inspiration by the process of evolution may easily fail if the path
to the optimum cannot be deduced from the knowledge accumulated by the optimization algorithm during search. However,
it is hoped that AM-DEGL algorithm proposed in this paper can be successfully applicable to wide number of problems that
may be of interest, although surely not all of them.
7. Conclusions
In the present paper a novel Differential Evolution algorithm (AM-DEGL) is proposed. It is developed following an idea of
getting together the most important components of different already successful Differential Evolution methods. The novel
method puts together the slightly modied concept of adaptation of control parameters and mutation strategies proposed
in Self-Adaptive Differential Evolution [115], the concept of Global and Local mutation operators proposed in Differential
Evolution with Global and Local neighborhood-based mutation algorithm [29] and the idea of using NelderMead algorithm
as a local search method within Memetic Differential Evolution framework proposed in [18]. It is empirically shown that the
proposed method outperforms 13 various Differential Evolution algorithms on a popular set of 30- and 50-dimensional
benchmark problems proposed in [126]. The proposed method also shows reasonable performance for 10-dimensional prob-
lems. A number of variants of novel algorithm are tested and it is found that the performance of a novel approach is mod-
erately sensitive to small changes in its constituent components. For the majority of tested problems AM-DEGL is able both
to converge to good solutions quickly and to keep improving until the maximum number of function calls is reached, show-
ing a good balance between exploitation and exploration properties.
As a novel heuristic is proposed in this paper, the implications of No Free Lunch theorems for optimization are very briey
discussed. It is known that No Free Lunch hold when one considers all possible problems or a subset of problems that are
Table 7
The lowest found function values (averaged over 30-runs) of 2-dimensional variant of the proposed problem, when each DE algorithm searched for the
maximum of the function.
Random AM-DEGL DE DE-SG DEGL SADE SspDE JADE
Mean 2.8E9 1.9E9 2.1E9 4.2E9 3.1E9 2.0E9 2.4E9 3.8E9
AdapSS-JADE DECLS DEahc-SPX CoDE EPSDE SF-MDE CDE CoDE
Mean 2.9E9 1.9E9 4.0E9 2.4E9 2.6E9 2.7E9 2.5E9 2.3E9
190 A.P. Piotrowski / Information Sciences 241 (2013) 164194
closed under permutations, what is not the case for problems that may be of any practical interest. But this by no means
exclude the possibility that for many problems of interest the random sampling may perform better than heuristic algo-
rithms, what was postulated by No Free Lunch theorems [150]. A very simple continuous and differentiable problem is pro-
posed in this paper for which all considered Differential Evolution methods perform poorer than random sampling.
Moreover, all considered Differential Evolution algorithms aiming at maximization of the proposed problem found lower
objective function values than all Differential Evolution algorithms aiming at minimization. It is also found that at least
for one among 25 benchmark problems dened in [126], that are widely used to compare Evolutionary Algorithms, the ran-
dom sampling perform equally well as one, and better than two other well known Differential Evolution methods from the
literature. This should be a precaution from generalizing the good performance of any proposed heuristics.
Acknowledgments
This work has been supported by the Inner Grant of the Institute of Geophysics, Polish Academy of Sciences Nr. 1b/IGF
PAN/2012/M.
References
[1] H.A. Abbass, An evolutionary articial neural networks approach for breast cancer diagnosis, Articial Intelligence in Medicine 25 (3) (2002) 265281.
[2] J. Akhtar, B.B. Koshul, M.M. Awais, A framework for evolutionary algorithms based on Charles Sanders Peirces Evolutionary semiotics, Information
Sciences 236 (2013) 93108.
[3] R.M. Alguliev, R.M. Aliguliyev, C.A. Mehidyev, Sentence selection for generic document summarization using an adaptive differential evolution
algorithm, Swarm and Evolutionary Computation 1 (2011) 213222.
[4] M.M. Ali, Differential evolution with preferential crossover, European Journal of Operational Research 181 (3) (2007) 11371147.
[5] M.M. Ali, Differential evolution with generalized differentials, Journal of Computational and Applied Mathematics 235 (8) (2011) 22052216.
[6] J. Apolloni, G. Leguizamon, J. Garcia-Nieto, E. Alba, Island based distributed differential evolution: an experimental study on hybrid test beds, in: Proc.
of IEEE 8th International Conference on Hybrid Intelligent Systems, Barcelona, Spain, 2008, pp. 696701.
[7] A. Auger, O. Teytaud, Continuous lunches are free, in: Proc. Genet. Evolut. Comput. Conf., 2007, pp. 117129.
[8] K. Bandurski, W. Kwedlo, A lamarckian hybrid of differential evolution and conjugate gradients for neural network training, Neural Processing Letters
32 (1) (2010) 3144.
[9] R.L. Becerra, C.A. Coello, Cultured differential evolution for constrained optimization, Computer Methods in Applied Mechanics and Engineering 195
(33-36) (2006) 43034322.
[10] H.G. Beyer, B. Sendhoff, Robust optimization a comprehensive survey, Computer Methods in Applied Mechanics and Engineering 196 (3334)
(2007) 31903218.
[11] X.J. Bi, J. Xiao, Classication-based self-adaptive differential evolution with fast and reliable convergence performance, Soft Computing 15 (8) (2011)
15811599.
[12] B. Boussaid, A. Chatterjee, P. Siarry, M. Ahmed-Nacer, Two-stage update biogeography-based optimization using differential evolution algorithm
(DBBO), Computers & Operations Research 38 (2011) 11881198.
[13] J. Brest, S. Greiner, B. Boskovic, M. Mernik, V. Zumer, Self-adapting control parameters in differential evolution: a comparative study on numerical
benchmark problems, IEEE Transactions on Evolutionary Computation 10 (6) (2006) 646657.
[14] J. Brest, B. Boskovic, S. Greiner, V. Zumer, M.S. Maucec, Performance comparison of self-adaptive and adaptive differential evolution algorithms, Soft
Computing 11 (2007) 617629.
[15] J. Brest, M.S. Maucec, Population size reduction for the differential evolution algorithm, Applied Intelligence 29 (3) (2008) 228247.
[16] J. Brest, M.S. Maucec, Self-adaptive differential evolution algorithm using population size reduction and three strategies, Soft Computing 15 (11)
(2011) 21572174.
[17] Z.H. Cai, W.Y. Gong, C.X. Ling, H. Zhang, A clustering-based differential evolution for global optimization, Applied Soft Computing 11 (1) (2011) 1363
1379.
[18] A. Caponio, F. Neri, V. Tirronen, Super-t control adaptation in memetic differential evolution frameworks, Soft Computing 13 (89) (2009) 811831.
[19] F. Carafni, F. Neri, G. Iacca, A. Mol, Parallel memetic structures. Information Sciences 227 (2013) 6082.
[20] P. Chakraborty, G.G. Roy, S. Das, D. Jain, A. Abraham, An improved harmony search algorithm with differential mutation operator, Fundamenta
Informaticae 95 (4) (2009) 401426.
[21] L. Chang, C. Liao, W.B. Lin, L.L. Chen, X. Zheng, A hybrid method based on differential evolution and continuous ant colony optimization and its
application on wideband antenna design, Progress in Electromagnetics Research PIER 122 (2012) 105118.
[22] X.S. Chen, Y.S. Ong, M.H. Lim, K.C. Tan, A multi-facet survey on memetic computation, IEEE Transactions on Evolutionary Computation 15 (5) (2011)
591607.
[23] J.P. Chiou, C.F. Chang, C.T. Su, Ant direction hybrid differential evolution for solving large capacitor placement problems, IEEE Transactions on Power
Systems 19 (4) (2004) 17941800.
[24] S. Christensen, F. Oppacher, What can we learn from no free lunch? A rst attempt to characterize the concept of a searcheable function, in: Proc.
Genet. Evolut. Comput. Conf., 2001, pp. 12191226.
[25] L.D. Coelho, D.L.D. Bernert, A modied ant colony optimization algorithm based on differential evolution for chaotic synchronization, Expert Systems
with Applications 37 (6) (2010) 41984203.
[26] A. Cong, W.X. Cong, Y.J. Lu, P. Santago, A. Chattzioannou, G. Wang, Differential evolution approach for regularized bioluminescence tomography, IEEE
Transactions on Biomedical Engineering 57 (9) (2010) 22292238.
[27] M. Crepinsek, S.H. Liu, L. Mernik, A note on teaching-learning-based optimization algorithm, Information Sciences 212 (2012) 7993.
[28] S. Das, A. Konar, U.K. Chakraborty, Improving particle swarm optimization with differentially perturbed velocity, in: Proc. Genet. Evol. Comput. Conf.
(GECCO), 2005, pp. 177184.
[29] S. Das, A. Abraham, U.K. Chakraboty, A. Konar, Differential evolution using a neighborhood-based mutation operator, IEEE Transactions on
Evolutionary Computation 13 (3) (2009) 526553.
[30] S. Das, S. Sil, Kernel-induced fuzzy clustering of image pixels with an improved differential evolution algorithm, Information Sciences 180 (8) (2010)
12371256.
[31] S. Das, P.N. Suganthan, Differential evolution: a survey of the state-of-the-art, IEEE Transactions on Evolutionary Computation 15 (1) (2011) 2754.
[32] S. Das, S. Maity, B.Y. Qu, P.N. Suganthan, Real-parameter evolutionary multimodal optimization a survey of the state-of-the-art, Swarm and
Evolutionary Computation 1 (2011) 7188.
[33] I. De Falco, A. Della Cioppa, D. Maisto, U. Scafuri, E. Tarantino, Biological invasioninspired migration in distributed evolutionary algorithms,
Information Sciences 207 (2012) 5065.
A.P. Piotrowski / Information Sciences 241 (2013) 164194 191
[34] V.V. De Melo, A.C.B. Delbem, Investigating smart sampling as a population initialization method for differential evolution in continuous problems,
Information Sciences 193 (2012) 3653.
[35] J. Demsar, Statistical comparisons of classiers over multiple data sets, Journal of Machine Learning Research 7 (2006) 130.
[36] B. Dorronosoro, P. Bouvry, Improving classical and decentralized differential evolution with new mutation operator and population topologies, IEEE
Transactions on Evolutionary Computation 15 (1) (2011) 6798.
[37] S. Droste, T. Jansen, I. Wegener, Perheps not a Free Lunch but at Least a Free Appetizer, Tech. Rep. ISSN 1433-3325, Dep. Comp. Sci. University of
Dortmund, 1998.
[38] S. Droste, T. Jansen, I. Wegener, Optimization with randomized search heuristics the (A)NFL theorem, realistic scenarios, and difcult functions,
Theoretical Computer Science 287 (1) (2002) 131144.
[39] A.E. Eiben, R. Hinterding, Z. Michalewicz, Parameter control in evolutionary algorithms, IEEE Transactions on Evolutionary Computation 3 (2) (1999)
124141.
[40] A.E. Eiben, S.K. Smit, Parameter tuning for conguring and analyzing evolutionary algorithms, Swarm and Evolutionary Computation 1 (2011) 1931.
[41] M.G. Epitropakis, V.P. Plagianakos, M.N. Vrahatis, Balancing the exploration and exploitation capabilities of the differential evolution algorithm, in:
Proc. of the IEEE Congress on Evolutionary Computation, Hong Kong, China, 2008, pp. 26862693.
[42] M.G. Epitropakis, V.P. Plagianakos, M.N. Vrahatis, Evolutionary adaptation of the differential evolution control parameters, in: Proc. of the IEEE
Congress on Evolutionary Computation, Trondheim, Norway, 2009. pp. 13591366.
[43] M.G. Epitropakis, D.K. Tasoulis, N.G. Pavlidis, V.P. Plagianakos, M.N. Vrahatis, Enhancing differential evolution utilizing proximity-based mutation
operations, IEEE Transactions on Evolutionary Computation 15 (1) (2011) 99119.
[44] M.G. Epitropakis, V.P. Plagianakos, M.N. Vrahatis, Evolving cognitive and social experience in particle swarm optimization through differential
evolution: a hybrid approach, Information Sciences 216 (2012) 5092.
[45] H.Y. Fan, J. Lampinen, A Trigonometric mutation operation to Differential Evolution, Journal of Global Optimization 27 (1) (2003) 105129.
[46] V. Feoktisov, Differential evolution in search of solutions, Book Series: Springer Optimization and Its Applications, vol. 5, Springer, New York, 2006.
[47] M. Friedman, The use of ranks to avoid the assumption of normality implicit the in the analysis of variance, Journal of American Statistical Association
32 (1937) 674701.
[48] R. Gamperle, S.D. Muller, P. Koumoutsakos, A parameter study of differential evolution, in: Advances in Intelligent Systems, Fuzzy Systems,
Evolutionary Computation, WSEAS Press, Interlaken, Switzerland, 2002.
[49] S. Garcia, A. Fernandez, J. Luengo, F. Herdera, Advanced nonparametric tests for multiple comparison in the design of experiments in computational
intelligence and data mining: experimental analysis of power, Information Sciences 180 (10) (2010) 20442064.
[50] A. Ghosh, S. Das, A. Chowdhury, R. Gini, An improved differential evolution algorithm with tness-based adaptation of the control parameters,
Information Sciences 181 (18) (2011) 37493765.
[51] S. Ghosh, S. Das, S. Roy, S.K.M. Islam, P.N. Suganthan, A differential covariance matrix adaptation evolutionary algorithm for real parameter
optimization, Information Sciences 181 (1) (2012) 199219.
[52] W.Y. Gong, Z.H. Cai, C.X. Ling, H. Li, Enhanced differential evolution with adaptive strategies for numerical optimization, IEEE Transactions on
Systems, Man and Cybernetics, Part B Cybernetics 41 (2) (2011) 397413.
[53] W.Y. Gong, Z.H. Cai, C.Y. Ling, DE/BBO: a hybrid differential evolution with biogeography-based optimization for global numerical optimization, Soft
Computing 15 (4) (2011) 645665.
[54] W.Y. Gong, Z.H. Cai, Differential evolution with ranking-based mutation operators, IEEE Transactions on Systems, Man and Cybernetics, Part B
Cybernetics, http://dx.doi.org/10.1109/TCYB.2013.2239988.
[55] W.Y. Gong, A. Fialho, Z.H. Cai, H. Li, Adaptive strategy selection in differential evolution for numerical optimization: an empirical study, Information
Sciences 181 (24) (2011) 53645386.
[56] N. Hachicha, B. Jarboui, P. Siarry, A fuzzy logic control using a differential evolution algorithm aimed at modelling the nancial market dynamics,
Information Sciences 181 (1) (2011) 7991.
[57] N. Hansen, A. Ostermeier, Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation, in: Proc. IEEE
Int. Conf. Evol. Comput., Nagoya, Japan, 1996, pp. 312317.
[58] N. Hansen, A.S.P. Niederberger, L. Guzzella, P. Koumoutsakos, A method for handling uncertainty in evolutionary optimization with an application to
feedback control of combustion, IEEE Transactions on Evolutionary Computation 13 (1) (2009) 180197.
[59] T. Hendtlass, A combined swarm differential evolution algorithm for optimization problems, in: Lecture Notes in Computer Science, vol. 2070, 2001,
pp.1118.
[60] S. Holm, A simple sequentially rejective multiple test procedure, Scandinavian Journal of Statistics 6 (1979) 6570.
[61] X. Hu, T. Glenn, F. Scalzo, M. Bergsneider, C. Sarkiss, N. Martin, P. Vespa, Intracranial pressure pulse morphological features improved detection of
decreased cerebral blood ow, Physiological Measurement 31 (5) (2010) 679695.
[62] F.Z. Huang, L. Wang, Q. He, An effective co-evolutionary differential evolution for constrained optimization, Applied Mathematics and Computation
186 (1) (2007) 340356.
[63] G. Iacca, F. Neri, E. Mininno, Y.S. Ong, M.H. Lim, Ockhams razor in memetic computing: three stage optimal memetic exploration, Information
Sciences 188 (2012) 1743.
[64] G. Iacca, F. Neri, E. Mininno, Noise analysis compact differential evolution, International Journal of Systems Science 43 (7) (2012) 12481267.
[65] C. Igel, M. Toussaint, On classes of functions for which no free lunch results hold, Information Processing Letters 86 (6) (2003) 317321.
[66] C. Igel, M. Toussaint, A no-free lunch theorem for non-uniform distributions of target functions, Journal of Mathematical Modelling and Algorithms 3
(2004) 313322.
[67] S.M. Islam, S. Das, S. Ghosh, S. Roy, P.N. Suganthan, An adaptive differential evolution algorithm with novel mutation and crossover strategies for
global numerical optimization, IEEE Transactions on Systems, Man and Cybernetics, Part B Cybernetics 42 (2) (2012) 482500.
[68] G. Jia, Y. Wang, Z. Cai, Y. Jin, An improved (l k)-constrained differential evolution for constrained optimization, Information Sciences 222 (2013)
302322.
[69] D.L. Jia, G.X. Zheng, M.K. Khan, An effective memetic differential evolution algorithm based on chaotic local search, Information Sciences 181 (15)
(2011) 31753187.
[70] P. Jiang, Y.P. Chen, Free lunches on the discrete Lipschitz class, Theoretical Computer Science 412 (17) (2011) 16141628.
[71] P. Kaelo, M.M. Ali, A numerical study on some modied differential evolution algorithms, European Journal of Operational Research 169 (3) (2006)
11761184.
[72] T.G. Kolda, R.M. Lewis, V. Torczon, Optimization by direct search: new perspectives on some classical and modern methods, SIAM Review 45 (3)
(2003) 385482.
[73] M. Koppen, D.H. Wolpert, W.G. Macready, Remarks on a recent paper on the No free lunch theorems, IEEE Transactions on Evolutionary
Computation 5 (3) (2001) 295296.
[74] N. Krasnogor, J. Smith, A tutorial for competent memetic algorithms: model, taxonomy, and design issues, IEEE Transactions on Evolutionary
Computation 9 (5) (2005) 474488.
[75] J.C. Lagarias, J.A. Reeds, M.H. Wright, P.E. Wright, Convergence properties of the NelderMead simplex method in low dimensions, SIAM Journal of
Optimization 9 (1) (1998) 112147.
[76] J. Lampinen, I. Zelinka, On stagnation of the differential evolution algorithm, in: Proceedings of 6th International Mendel Conference on Soft
Computing, Brno, Czech Republic, 2000.
192 A.P. Piotrowski / Information Sciences 241 (2013) 164194
[77] W.B. Langdon, R. Poli, Evolving problems to learn about particle swarm optimizers and other search algorithms, IEEE Transactions on Evolutionary
Computation 11 (5) (2007) 561578.
[78] T.W. Liao, Two hybrid differential evolution algorithms for engineering design optimization, Applied Soft Computing 10 (4) (2010) 11881199.
[79] C. Lin, A.Y. Qing, Q.Y. Feng, A comparative study of crossover in differential evolution, Journal of Heuristics 17 (6) (2011) 675703.
[80] J. Liu, J. Lampinen, A fuzzy adaptive differential evolution algorithm, Soft Computing 9 (2005) 448462.
[81] H. Liu, Z. Cai, Y. Wang, Hybridizing particle swarm optimization with differential evolution for constrained numerical and engineering optimization,
Applied Soft Computing 10 (2) (2010) 629640.
[82] G. Liu, Y. Li, X. Nie, H. Zheng, A novel clustering-based differential evolution with 2 multi-parent crossovers for global optimization, Applied Soft
Computing 12 (2) (2012) 663681.
[83] R. Mallipeddi, P.N. Suganthan, Ensemble of constraint handling techniques, IEEE Transactions on Evolutionary Computation 14 (4) (2010) 561579.
[84] R. Mallipeddi, J.P. Lie, P.N. Suganthan, S.G. Razul, C.M.S. See, A differential evolution approach for robust adaptive beamforming based on joint
estimation of look direction and array geometry, Progress in Electromagnetics Research PIER 119 (2011) 381394.
[85] R. Mallipeddi, P.N. Suganthan, Q.K. Pan, M.F. Tasgetiren, Differential evolution algorithm with ensemble of parameters and mutation strategies,
Applied Soft Computing 11 (2) (2011) 16791696.
[86] A.D. Masegosa, D.A. Pelta, J.L. Verdegay, A centralised cooperative strategy for continuous optimisation: the inuence of cooperation in performance
and behaviour, Information Sciences 219 (2013) 7392.
[87] E. Mininno, F. Neri, F. Cupertino, D. Naso, Compact differential evolution, IEEE Transactions on Evolutionary Computation 15 (1) (2011) 3254.
[88] A.W. Mohamed, H.Z. Sabry, Constrained optimization based on modied differential evolution algorithm, Information Sciences 194 (2012) 171208.
[89] C.G. Moles, J.R. Banga, K. Keller, Solving nonconvex climate control problems: pitfalls and algorithm performances, Applied Soft Computing 5 (1)
(2004) 3544.
[90] P.W. Moore, G.K. Venayagamoorthy, Evolving digital circuits using hybrid particle swarm optimization and differential evolution, International
Journal of Neural Systems 16 (3) (2006) 163177.
[91] J.A. Nelder, R. Mead, A simplex-method for function minimization, Computer Journal 7 (4) (1965) 308313.
[92] F. Neri, V. Tirronen, On memetic differential evolution frameworks: a study of advantages and limitations in hybridization, in: Proc. of IEEE World
Congress on Computational Intelligence, Hong Kong, 2008, pp. 21352142.
[93] F. Neri, E. Mininno, Memetic compact differential evolution for Cartesian robot control, IEEE Computational Intelligence Magazine 5 (2) (2010) 5465.
[94] F. Neri, V. Tirronen, Recent advances in differential evolution: a survey and experimental analysis, Articial Intelligence Review 33 (12) (2010) 61
106.
[95] F. Neri, G. Iacca, E. Mininno, Disturbed exploitation compact differential evolution for limited memory optimization problems, Information Sciences
181 (12) (2011) 24692487.
[96] F. Neri, C. Cotta, Memetic algorithms and memetic computing optimization: a literature review, Swarm and Evolutionary Computation 2 (2012) 114.
[97] N. Noman, H. Iba, Inferring gene regulatory networks using differential evolution with local search heuristics, IEEE ACM Transactions on
Computational Biology and Bioinformatics 4 (4) (2007) 634647.
[98] N. Noman, H. Iba, Accelerating differential evolution using an adaptive local search, IEEE Transactions on Evolutionary Computation 12 (1) (2008)
107125.
[99] M. Oltean, Searching for a practical evidence for the no free lunch theorems, in: Bioinspired Approaches to Advanced Information Technology,
Springer, Lausanne, Switzerland, 2004.
[100] M.G.H. Omran, A. Salman, A.P. Engelbrecht, Self-adaptive differential evolution, in: Lecture Notes in Articial Intelligence, Springer-Verlag, Berlin,
Germany, 2005.
[101] M.G.H. Omran, A.P. Engelbrecht, A. Salman, Bare bones differential evolution, European Journal of Operational Research 196 (1) (2009) 128139.
[102] Y.S. Ong, A.J. Keane, Meta-lamarckian learning in memetic algorithms, IEEE Transactions on Evolutionary Computation 8 (2) (2004) 99110.
[103] 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 Cybernetics 36 (1) (2006) 141152.
[104] Y.S. Ong, M.H. Lim, X. Chen, Memetic computation past, present and future, IEEE Computational Intelligence Magazine 5 (2) (2010) 2431.
[105] Q.K. Pan, P.N. Suganthan, L. Wang, L. Gao, R. Mallipeddi, A differential evolution algorithm with self-adapting strategy and control parameters,
Computers & Operations Research 38 (2011) 394408.
[106] K.M. Passino, Biomimicry of bacterial foraging for distributed optimization and control, IEEE Control Systems Magazine 22 (3) (2002) 5267.
[107] F. Peng, K. Tang, G.L. Chen, X. Yao, Population-based algorithm portfolios for numerical optimization, IEEE Transactions on Evolutionary Computation
14 (5) (2010) 782800.
[108] A.P. Piotrowski, J.J. Napiorkowski, Grouping differential evolution algorithm for multi-dimensional optimization problems, Control and Cybernetics
39 (2) (2010) 527550.
[109] A.P. Piotrowski, J.J. Napiorkowski, A. Kiczko, Differential Evolution algorithm with separated groups for multi-dimensional optimization problems,
European Journal of Operational Research 216 (2012) 3346.
[110] A.P. Piotrowski, J.J. Napiorkowski, A. Kiczko, Corrigendum to: Differential evolution algorithm with separated groups for multi-dimensional
optimization problems [Eur. J. Oper. Res. 216 (2012) 3346], European Journal of Operational Research 219 (2012) 488.
[111] A.P. Piotrowski, P.M. Rowinski, J.J. Napiorkowski, Comparison of evolutionary computation techniques for noise injected neural network training to
estimate longitudinal dispersion coefcients in rivers, Expert Systems with Applications 39 (1) (2012) 13541361.
[112] K.V. Price, An introduction to differential evolution, in: New Ideas in Optimization, McGraw-Hill, London, UK, 1999.
[113] K.V. Price, R.M. Storn, J.A. Lampinen, Differential Evolution, A Practical Approach to Global Optimization, Springer-Verlag, Berlin, Heidelberg, 2005.
[114] A.K. Qin, P.N. Suganthan, Self-adaptive differential evolution algorithm for numerical optimization, in: Proc. IEEE Congress on Evolutionary
Computation, Edinburgh, UK, 2005.
[115] 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 (2) (2009) 398417.
[116] S. Rahnamayan, H.R. Tizhoosh, M.M.A. Salama, Opposition-based differential evolution, IEEE Transactions on Evolutionary Computation 12 (1) (2008)
6479.
[117] H.H. Rosenbrock, An automated method for nding the greatest or least value of a function, Computer Journal 3 (3) (1960) 175184.
[118] J.E. Rowe, M.D. Vose, A.H. Wright, Reinterpreting no free lunch, Evolutionary Computation 17 (1) (2009) 117129.
[119] R. Salomon, Re-evaluating genetic algorithm performance under coordinate rotation on benchmark functions. A survey of some theoretical and
practical aspects of genetic algorithms, BioSystems 39 (1996) 263278.
[120] C. Schumacher, M.D. Vose, L.D. Whitley, The no free lunch and problem description length, in: Proc. Genet. Evolut. Comput. Conf., 2001, pp. 565570.
[121] S. Sivananaithaperumal, J. Miruna Joe Amali, S. Baskar, P.N. Suganthsn, Constrained self adaptive differential evolution based design of robust optimal
xed structure controller, Engineering Applications of Articial Intelligence 24 (6) (2011) 10841093.
[122] J.D. Sosin, Ausonius Juvenal and the Winstedt fragment, Classical Philology 95 (2) (2000) 199206.
[123] R. Storn, K.V. Price, Differential Evolution A Simple and Efcient Adaptive Scheme for Global Optimization Over Continuous Spaces, Tech. Report TR-
95-012, International Computer Sciences Institute, Berkeley, California, USA, 1995.
[124] R. Storn, K.V. Price, Differential evolution a simple and efcient heuristic for global optimization over continuous spaces, Journal of Global
Optimization 11 (4) (1997) 341359.
[125] M.J. Streeter, Two broad classes of functions for which a no free lunch result does not hold, in: Proc. Genet. Evolut. Comput. Conf., 2003, pp. 1418
1430.
A.P. Piotrowski / Information Sciences 241 (2013) 164194 193
[126] 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, Nanyang Technol. Univ., Singapore, Tech. Rep. KanGAL #2005005, IIT Kanpur, India, 2005.
[127] R. Tanese, Distributed genetic algorithms, in: Proc. of 3rd Int. Conference on Genetic Algorithms, San Francisco, CA, USA, 1989, pp. 434439.
[128] D.K. Tasoulis, N.G. Pavlidis, V.P. Plagianakos, M.N. Vrahatis, Parallel differential evolution, in: Proceedings of the IEEE Congress on Evolutionary
Computation, Portland, Oregon, USA, 2004, pp. 20232029.
[129] N.S. Teng, J. Teo, M.H.A. Hijazi, Self-adaptive population sizing for a tune-free differential evolution, Soft Computing 13 (7) (2009) 709724.
[130] J. Teo, Exploring dynamic self-adaptive populations in differential evolution, Soft Computing 11 (8) (2006) 673686.
[131] D. Thierens, An adaptive pursuit strategy for allocating operator probabilities, in: Proc. Genet. Evol. Comput. Conf., 2005, pp. 15391546.
[132] V. Tirronen, F. Neri, T. Karkkainen, K. Majava, T. Rossi, An enhanced memetic differential evolution in lter design for defect detection in paper
production, Evolutionary Computation 16 (4) (2008) 529555.
[133] V. Tirronen, F. Neri, Differential evolution with tness diversity self-adaptation, Nature-Inspired Algorithms for Optimisation, SCI 193, Springer-
Verlag, Berlin, Heidelberg, 2009, pp. 199234.
[134] I. Triguero, S. Garcia, F. Herrera, Differential evolution for optimizing the positioning of prototypes in nearest neighbor classication, Pattern
Recognition 44 (4) (2011) 901916.
[135] O. Urfalioglu, O. Arikan, Self-adaptive randomized and rank-based differential evolution for multimodal problems, Journal of Global Optimization 51
(4) (2011) 607640.
[136] K. Vaisakh, P. Praveena, S. Rama Mohana Rao, Kala Maeh, Solving dynamic economic dispatch problem with security constraints using bacterial
foraging PSO-DE algorithm, Electrical Power and Energy Systems 39 (2012) 5667.
[137] C. Voglis, K.E. Parsopoulos, D.G. Papageorgiou, I.E. Lagaris, M.N. Vrahatis, MEMPSODE: a global optimization software based on hybridization of
population-based algorithms and local searches, Computer Physics Communications 183 (5) (2012) 11391154.
[138] J.A. Vrugt, B.A. Robinson, J.M. Hyman, Self-adaptive multimethod search for global optimization in real-parameter spaces, IEEE Transactions on
Evolutionary Computation 13 (2) (2009) 243259.
[139] Y. Wang, Z. Cai, Q. Zhang, Differential evolution with composite trial vector generation strategies and control parameters, IEEE Transactions on
Evolutionary Computation 15 (1) (2011) 5566.
[140] L. Wang, L.P. Li, A coevolutionary differential evolution with harmony search for reliabilityredundancy optimization, Expert Systems with
Applications 39 (5) (2012) 52715278.
[141] Y. Wang, Z. Cai, Q. Zhang, Enhancing the search ability of differential evolution through orthogonal crossover, Information Sciences 185 (1) (2012)
153177.
[142] M. Weber, F. Neri, V. Tirronen, Distributed differential evolution with explorativeexploitative population families, Genetic Programming and
Evolvable Machines 10 (4) (2009) 343371.
[143] M. Weber, V. Tirronen, F. Neri, Scale factor inheritance mechanism in distributed differential evolution, Soft Computing 14 (11) (2010) 11871207.
[144] M. Weber, F. Neri, V. Tirronen, A study on scale factor in distributed differential evolution, Information Sciences 181 (12) (2011) 24882511.
[145] M. Weber, F. Neri, V. Tirronen, Shufe or update parallel differential evolution for large-scale optimization, Soft Computing 15 (2011) 20892107.
[146] D. Whitley, J. Rowe, Subthreshold-seeking local search, Theoretical Computer Science 361 (1) (2006) 217.
[147] D. Whitley, J. Rowe, Focused no free lunch theorems, in: Proc. Genet. Evol. Comput. Conf., 2008, pp. 811818.
[148] F. Wilcoxon, Individual comparisons by ranking methods, Biometrics Bulletin 1 (6) (1945) 8083.
[149] D.H. Wolpert, W.G. Macready, No Free Lunch Theorems for Search, Tech. Rep. SFI-TR-05-010, Santa Fe, NM, USA, 1995.
[150] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE Transactions on Evolutionary Computation 1 (1) (1997) 6782.
[151] D.H. Wolpert, W.G. Macready, Coevolutionary free lunches, IEEE Transactions on Evolutionary Computation 9 (6) (2005) 721735.
[152] B. Xin, J. Chen, J. Zhang, H. Fang, Z.H. Peng, Hybridizing differential evolution and particle swarm optimization to design powerful optimizers: a
review and taxonomy, IEEE Transactions on Systems, Mac and Cybernetics, Part C Applications and Reviews 42 (5) (2012) 744767.
[153] D. Zaharie, Inuence of crossover on the behavior of differential evolution algorithms, Applied Soft Computing 9 (3) (2009) 11261138.
[154] T. Zhang, L. Chen, Z.H. Yan, B. Li, Design of dual offset shaped reector antenna based on DEGL algorithm, Journal of Electromagnetic Waves and
Applications 25 (56) (2011) 723732.
[155] M. Zhang, W. Luo, X.F. Wang, Differential evolution with dynamic stochastic selection for constrained optimization, Information Sciences 178 (15)
(2008) 30433074.
[156] J. Zhang, A.C. Sanderson, JADE: adaptive differential evolution with optional external archive, IEEE Transactions on Evolutionary Computation 13 (5)
(2009) 945958.
[157] S.Z. Zhao, P.N. Suganthan, S. Das, Self-adaptive differential evolution with multi-trajectory search for large-scale optimization, Soft Computing 15 (11)
(2011) 21752185.
[158] Z. Zhao, J.Z. Wang, J. Zhao, Z.Y. Su, Using a grey model optimized by differential evolution algorithm to forecast the per capita annual net income of
rural households in China, OMEGA International Journal of Management Science 40 (5) (2012) 525532.
[159] S.Z. Zhao, P.N. Suganthan, Empirical investigations into the exponential crossover of differential evolutions, Swarm and Evolutionary Computation 9
(2013) 2736.
[160] W. Zhu, Y. Tang, J.A. Fang, W. Zhang, Adaptive population tuning scheme for differential evolution, Information Sciences 223 (2012) 164191.
194 A.P. Piotrowski / Information Sciences 241 (2013) 164194

You might also like