Opposition-based learning in the shuffled differential evolution algorithm
Transcript of Opposition-based learning in the shuffled differential evolution algorithm
ORIGINAL PAPER
Opposition-based learning in the shuffled differential evolutionalgorithm
Morteza Alinia Ahandani • Hosein Alavi-Rad
Published online: 9 February 2012
� Springer-Verlag 2012
Abstract This paper proposes using the opposition-based
learning (OBL) strategy in the shuffled differential evolu-
tion (SDE). In the SDE, population is divided into several
memeplexes and each memeplex is improved by the dif-
ferential evolution (DE) algorithm. The OBL by comparing
the fitness of an individual to its opposite and retaining the
fitter one in the population accelerates search process. The
objective of this paper is to introduce new versions of the
DE which, on one hand, use the partitioning and shuffling
concepts of SDE to compensate for the limited amount of
search moves of the original DE and, on the other hand,
employ the OBL to accelerate the DE without making
premature convergence. Four versions of DE algorithm are
proposed based on the OBL and SDE strategies. All
algorithms similarly use the opposition-based population
initialization to achieve fitter initial individuals and their
difference is in applying opposition-based generation
jumping. Experiments on 25 benchmark functions designed
for the special session on real-parameter optimization of
CEC2005 and non-parametric analysis of obtained results
demonstrate that the performances of the proposed algo-
rithms are better than the SDE. The fourth version of
proposed algorithm has a significant difference compared
to the SDE in terms of all considered aspects. The
emphasis of comparison results is to obtain some suc-
cessful performances on unsolved functions for the first
time, which so far have not been reported any successful
runs on them. In a later part of the comparative
experiments, performance comparisons of the proposed
algorithm with some modern DE algorithms reported in the
literature confirm a significantly better performance of
our proposed algorithm, especially on high-dimensional
functions.
Keywords Opposition-based learning � Shuffled
differential evolution �Memeplex � Premature convergence
1 Introduction
The evolutionary algorithms (EAs) can be applied to search
the complex, discrete or continuous, linear or non-linear,
and convex or non-convex spaces. So those are recognized
as all-propose and direct search optimization methods.
Existence of a criterion to evaluate a candidate solution is
unique, necessary and with enough limitation to apply the
EAs to each problem. Differential evolution (DE) (Storn
and Price 1997) is a kind of EAs, which was proposed as a
modified version of genetic algorithms (GAs) (Holland
1975). Although the DE is not biologically inspired, its
name was derived from natural biological evolutionary
processes.
There are various ideas behind EAs for generation of
new trial points and candidate solutions. In the GAs,
inspired from the natural biological evolution and heredity
process, by employing three operators of biological sys-
tems, i.e., selection, crossover and mutation, poor chro-
mosomes and poor genes disappear from population and
newly generated members are replaced with them. This
algorithm says which better chromosomes containing bet-
ter genes and/or better combinations of genes tend to
produce better offspring. The particle swarm optimization
(PSO) (Kennedy and Eberhart 1995) simulates the
M. A. Ahandani (&) � H. Alavi-Rad
Department of Electrical Engineering, Langaroud Branch,
Islamic Azad University, Langaroud, Iran
e-mail: [email protected]
H. Alavi-Rad
e-mail: [email protected]
123
Soft Comput (2012) 16:1303–1337
DOI 10.1007/s00500-012-0813-9
movement of organisms and social behavior in a flock of
birds or school of fishes. This algorithm updates its mem-
bers using the particle’s previous best position and using
the previous best position of the particle’s neighbors or
whole swarm. The shuffled frog leaping (SFL) (Eusuff and
Lansey 2003), inspired from grouping search of frogs for
food resources in a swamp, combines the benefits of the
memetic algorithm (MA) (Moscato 1989), the social
behavior-based PSO algorithm and sharing information of
parallel local search algorithms. The DE on one hand, such
as GAs, employs three biological operators to generate new
offspring, and such as GAs, in this algorithm accepting or
rejecting a new offspring has a direct relation with its fit-
ness. On the other hand, the DE such as PSO and unlike
GAs, does not eliminate the poor members from the pop-
ulation and each member preserves its own position in the
population all over the search process and is simply pro-
pelled to different areas of search spaces without complete
generation of a new member. Liu et al. (2007) mentioned
some other attractive characteristics of DE compared to the
GAs and PSO. The DE uses a simple differential operator
to create new candidate solutions and one-to-one compe-
tition scheme to greedily select new candidate, which
works with real numbers in a natural manner and avoids
complicated generic searching operators in the GA. It has a
memory; so, knowledge of good solutions is retained in the
current population, whereas in the GA previous knowledge
of the problem is destroyed once the population changes,
and in the PSO a secondary archive is needed. It also has
constructive cooperation between individuals; individuals
in the population share information between them.
The remaining sections of this paper are organized as
follows. In the next section, a literature review of related
works is presented. The original DE algorithm is briefly
described in Sect. 3. In Sect. 4, the OBL is presented.
In Sect. 5, the proposed modified versions of DE are
explained. The simulation results are presented and ana-
lyzed in Sect. 6. Section 7 concludes the paper and some
concepts for future works are presented.
2 Related works
This section reviews the benefits and drawbacks of DE, as
well as some reported approaches to compensate for its
drawbacks. After that, we concentrate on researches that
deal with opposition-based strategy.
2.1 The DE: benefits and drawbacks
The DE has only three control parameters to be tuned, i.e.,
amplification factor of the difference vector, crossover rate
and population size. However, proper setting of control
parameters in the DE is not an easy task. In addition to
these attractive characteristics, simplicity and easy imple-
mentation are two main preferences of DE than other
EAs and a basic reason for widespread application of
DE on different optimization problems in recent years
(see Plagianakos et al. 2008). Feoktistov (2006), from an
algorithmic viewpoint, mentioned the reasons for the
success of DE: the success of DE is due to an implicit
self-adaptation contained within the algorithmic structure.
Besides the aforementioned preferences, some drawbacks
of the standard DE are as follows:
(I) Stagnation or premature convergence because of its
low or fast convergence speed. Neri and Tirronen (2010)
describes how the stagnation occurs in the DE: a DE
scheme is highly explorative at the beginning of the evo-
lution and subsequently becomes more exploitative during
the optimization. Although this mechanism seems, at first
glance, to be very efficient, it hides a limitation. If for some
reason the algorithm does not succeed in generating off-
spring solutions which outperform the corresponding par-
ent, the search is repeated again with similar step size
values and will likely fail by falling into an undesired
stagnation condition. Liu and Lampinen (2005) mentioned
three reasons for this drawback of standard DE: (1) control
parameters not being well chosen initially for a given task;
(2) parameters being kept fixed through the whole process
and having no response to the population’s information
even though the environment in which the DE operated
may be variable; (3) lack of knowledge from the search
space.
Some of the DE variants as a solution for this problem
employed self-adaptive settings for automatically and
dynamically adjusting evolutionary parameters. Liu and
Lampinen (2005) proposed a fuzzy logic control (FLC) for
controlling parameters of DE. The FLC was employed to
choose the initial control parameters freely and the control
parameters adjusted online to dynamically adapt to
changing situations. It was found that DE with a fuzzy
search parameter control could perform better than DE
using all fixed parameters. Brest et al. (2006) proposed a
new version of DE for obtaining self-adaptive setting of
two control parameters, i.e., amplification factor of the
difference vector and crossover rate. The results on
numerical benchmark problems showed that their proposed
DE with self-adaptive control parameter settings was better
than, or at least comparable to, the standard DE algorithm.
Teng et al. (2009), for solving the manual setting of
population size by user, carried out two new systems for
the self-adaptive population size to test two different
methodologies, absolute encoding and relative encoding.
The empirical testing results showed that DE with self-
adaptive population size using relative encoding performed
well in terms of the average performance as well as
1304 M. A. Ahandani, H. Alavi-Rad
123
stability compared to absolute encoding version, as well as
the original DE. Also, a self-adaptive DE (SaDE) algorithm
was proposed by Qin et al. (2009) in which both trial vector
generation strategies and their associated control parameter
values were gradually self-adapted by learning from their
previous experiences in generating promising solutions.
Consequently, a more suitable generation strategy along
with its parameter settings could be determined adaptively
to match different phases of the search process/evolution.
Those compared the performance of SaDE with the con-
ventional DE and three adaptive DE variants over a suite of
26 bound constrained numerical optimization problems and
concluded that the SaDE was more effective in obtaining
better quality solutions, which were more stable with the
relatively smaller standard deviation and had higher suc-
cess rates. An opposition-based DE (ODE) was proposed
by Rahnamayan et al. (2008) to accelerate the DE. The
ODE used opposite numbers during population initializa-
tion and also for generating new populations during the
evolutionary process. Experimental results confirmed that
the ODE outperformed the original DE and fuzzy adaptive
DE in terms of convergence speed and solution accuracy.
(II) Having a problem in accurately zooming to optimal
solution. The DE can efficiently find the neighborhood of
global optimal solution, but for many cases, it is not able to
zoom exactly to individual optimal point. Hybridizing the
DE with a local search method has been proposed as a
solution for this drawback. In the hybrid algorithms, the
DE has an exploration duty of whole search space to find
some promising areas, and then the local search is entered
to localize the found areas and for exploitation of optimum
point as accurately as possible.
Caponio et al. (2009) proposed a superfit memetic DE
(SFMDE) by employing a DE framework hybridized with
three meta-heuristics: the PSO as assister of the DE to
generate a super-fit individual, the Nelder–Mead algorithm
and the Rosenbrock algorithm as local searchers, which are
adaptively coordinated by means of an index measuring
quality of the super-fit individual with respect to the rest of
the population. Numerical results of the SFMDE on two
engineering problems demonstrated that the SFMDE had a
high performance standard in terms of both final solutions
detected and convergence speed.
Also, a scale factor local search DE based on MA was
proposed by Neri and Tirronen (2009). It employed, within
a self-adaptive scheme, two local search algorithms, i.e.,
golden section search and uni-dimensional hill climb. The
local search algorithms assisted in the global search and
generated offspring with high performance, which were
subsequently supposed to promote the generation of
enhanced solutions within the evolutionary framework.
Numerical results demonstrated that the efficiency of the
proposed algorithm seemed to be very high, especially for
large-scale problems and complex fitness landscapes. Perez-
Bellido et al. (2008) presented a memetic DE to solve the
spread spectrum radar poly phase code design problem. The
DE and other utilized global search algorithms hybridized
with a gradient-based local search procedure, which included
a dynamic step adaptation procedure to perform accurate and
efficient local search for better solutions. Simulations in
several numerical examples showed that their proposed
approaches improved the performance of previous approa-
ches existing in this problem.
(III) To limit the number and diversity of search moves
in the original DE. The original DE carries out an effort in
per iteration to improve each member. Limiting the number
and variety of search moves prevents exploration of whole
search space. Ahandani et al. (2010) proposed three mod-
ified versions of the DE to repair its defect in accurate
converging to individual optimal point and to compensate
the limited amount of search moves of original DE. These
algorithms carried out several efforts one by one to obtain a
fitter offspring so that each search move was different from
previously utilized moves. Their proposed DE algorithms
employed the bidirectional optimization and parallel search
concepts. A comparison of their proposed methods with
some modern DE algorithms and the other EAs reported in
the literature confirmed a better or at least comparable
performance of their proposed algorithms.
(IV) Utilizing greedy criterion in accepting or rejecting
a new generated offspring. In the greedy acceptance cri-
terion, only generating an offspring with a better fitness
compared to its corresponding parent is considered as an
admissive move and those of candidate solutions with a
worse fitness will be rejected. However, greedy criterion
ensures the fast convergence; it increases the probability of
getting stuck in local minimums. The selection operator of
DE follows a greedy criterion inspired by the hill-climbing
process. Other EAs by employing some criteria to accept
poor candidate solutions do not confine their exploration.
The PSO accepts all new position of particles and the SFL
after two stages and a random frog without evaluating its
fitness. In the simulated annealing (SA) (Kirkpatrick et al.
1983), a member with a worse fitness is accepted with a
probability and the TS accepts all moves that do not satisfy
some tabu restriction criteria. But to the best of our
knowledge, all proposed DE algorithms employ greedy
criterion and there is not any superseded operator.
(V) Poor performance of DE in noisy environment. A
noisy optimization problem can be seen as a problem
where the fitness landscape varies with time and, due to a
wide search space and the existence of many local opti-
mums, obtaining an optimal point is very hard. Thus, for a
noisy problem, a deterministic choice of the scale factor
can be inadequate and a standard DE can easily fail at
handling a noisy fitness function (Neri and Tirronen 2010).
OBL in the SDE algorithm 1305
123
Looking at the problem from a different perspective, DE
employs too much deterministic search logic for a noisy
environment and therefore tends to stagnate.
(VI) To require multiple runs for tuning parameters and
to be problem dependent of the best control parameter
settings. The adaptive or self-adaptive control of parame-
ters have been proposed as a solution to overcome these
disadvantages [for example, see Qin and Suganthan (2005);
Liu and Lampinen (2005); Brest et al. (2007); Teng et al.
(2009)].
This study proposes four modified versions of DE
algorithms. These proposed algorithms use a combination
of shuffled DE (SDE) proposed by Ahandani et al. (2010),
and ODE proposed by Rahnamayan et al. (2008). On one
hand, The SDE diversifies exploratory moves of DE and
increases the number of efforts carried out to independently
improve each member. On the other hand, the ODE
accelerates the DE and prevents its stagnation. So moti-
vated by the successful implementation of the mentioned
techniques to remove some drawbacks of the original DE,
this study integrates these features together to develop
some modified algorithms, which combine benefits of both
aforementioned algorithms.
2.2 Opposition-based strategy
Opposition-based strategy in the optimization algorithms
uses the concept of opposition-based learning (OBL),
which was introduced by Tizhoosh (2005). The OBL by
comparing the fitness of an individual to its opposite and
retaining the fitter one in the population accelerates the
EAs. The OBL was initially applied to accelerate rein-
forcement learning (Tizhoosh 2006; Shokri et al. 2006) and
back-propagation learning in neural networks (Ventresca
and Tizhoosh 2006). A mathematical proof was proposed
by Rahnamayan et al. (2006) to show that, in general,
opposite numbers are more likely to be closer to the opti-
mal solution than a purely random one.
The OBL was recently applied to different EAs.
Rahnamayan et al. (2008) for the first time utilized oppo-
site numbers to speed up the convergence rate of an opti-
mization algorithm. Their proposed opposition-based DE
(ODE) employed the OBL for population initialization and
also for generation jumping. Also, the performance of ODE
on large-scale problems was examined by Rahnamayan
and Wang (2008). Results confirmed that ODE performed
much better than DE when the dimensionality of the
problems was increased from 500D to 1000D.
Subudhi and Jena (2009) presented a new DE approach
based on OBL applied for training neural network used for
non-linear system identification. The obtained results for
identification of two non-linear system benchmark prob-
lems demonstrated that the opposition-based DE-neural
network method of non-linear system identification pro-
vided excellent identification performance in comparison
to both the DE and neuro-fuzzy approaches. Balamurugan
and Subramanian (2009) presented an opposition-based
self-adaptive DE for emission-constrained dynamic eco-
nomic dispatch problem with non-smooth fuel cost and
emission level functions. A multi-objective function was
formulated by assigning the relative weight to each of the
objective and then optimized by opposition-based self-
adaptive DE. The convergence rate of DE was improved by
employing an OBL scheme and a self-adaptive procedure
for control parameter settings. The simulation results on a
test system with five thermal generating units showed that
the proposed approach provided a higher quality solution
with better performance. Boskovis et al. (2011) presented a
DE-based approach to chess evaluation function tuning.
The DE with opposition-based optimization was employed
and upgraded with a history mechanism to improve the
evaluation of individuals and the tuning process. The
general idea was based on individual evaluations according
to played games through several generations and different
environments. They introduced a new history mechanism,
which used an auxiliary population containing good indi-
viduals. This new mechanism ensured that good individu-
als remained within the evolutionary process, even though
they died several generations back and later could be
brought back into the evolutionary process. In such a
manner, the evaluation of individuals was improved and
consequently the whole tuning process.
The PSO is another EA for which the OBL has been
recently used in its structure. Wang et al. (2007) applied an
OBL scheme to the PSO (OPSO), along with a Cauchy
mutation to keep the globally best particle moving and
avoid premature convergence of it. The main objective of
OPSO with Cauchy mutation was to help avoid premature
convergence on multi-modal functions. Using OBL, two
different positions, the particle’s own position and the
position opposite the center of the swarm, were evaluated
for each randomly selected particle. Experimental results
on benchmark optimization problems showed that the
OPSO could successfully deal with those difficult multi-
modal functions while maintaining fast search speed on
those simple unimodal functions in the function optimiza-
tion. Han and He (2007) used OBL to improve the per-
formance of PSO. They employed the OBL in the
initialization phase and also during each iteration. How-
ever, a constriction factor was used to enhance the con-
vergence speed. In both the above mentioned which used
the OBL in the PSO, several parameters were added to the
PSO that were difficult to tune. Omran (2009) used the
OBL to improve the performance of PSO and barebones
DE (BBDE) without adding any extra parameter. Two
opposition-based variants were proposed (namely, iPSO
1306 M. A. Ahandani, H. Alavi-Rad
123
and iBBDE). The iPSO and iBBDE algorithms replaced the
least-fit particle with its anti-particle. The results showed
that, in general, iPSO and iBBDE outperformed PSO and
BBDE, respectively. In addition, the results showed that
using the OBL enhanced the performance of PSO and
BBDE without requiring additional parameters. Also,
Rashid and Baig (2010) presented an improved opposition-
based PSO and applied this algorithm to feed-forward
neural network training. The improved opposition-based
PSO utilized opposition-based initialization, opposition-
based generation jumping and opposition-based velocity
calculation.
Also some researchers have applied the OBL in bioge-
ography-based optimization (BBO), which is an EA
developed for global optimization. Ergezer et al. (2009)
proposed a novel variation to the BBO. The new algorithm
employed the OBL alongside BBO’s migration rates to
create oppositional BBO (OBBO). Additionally, a new
opposition method named quasi-reflection was introduced.
Empirical results demonstrated that with the assistance of
quasi-reflection, the OBBO significantly outperformed the
BBO in terms of success rate and the number of fitness
function evaluations required for finding an optimal solu-
tion. Bhattacharya and Chattopadhyay (2010) presented a
quasi-reflection oppositional biogeography-based optimi-
zation to accelerate the convergence of BBO and to
improve solution quality for solving complex economic
load dispatch problems of thermal power plants. The pro-
posed method employed the OBL along with a BBO
algorithm. Instead of opposite numbers, they used quasi-
reflected numbers for population initialization and also for
generation jumping. The effectiveness of the proposed
algorithm was verified on four different test systems.
Compared with the other existing techniques, the proposed
algorithm was found to perform better in a number of
cases. Considering the quality of the solution and conver-
gence speed obtained, this method seemed to be a prom-
ising alternative approach for solving the economic load
dispatch problems.
Besides the aforementioned algorithms, there are some
other cases of application of this strategy to the EAs.
Malisia and Tizhoosh (2007) employed the opposite-
based concept to improve the quality of solutions and
convergence rate of an ant colony system (ACS) (Dorigo
and Gambardella 1997). The modifications focused on the
solution construction phase of the ACS. Results on the
application of these algorithms on traveling salesman
problem instances demonstrated that the concept of
opposition was not easily applied to the ACS. Only one of
the pheromone-based methods showed performance
improvements that were statistically significant. Also, an
improved vanilla version of the SA using opposite
neighbors called opposition-based simulated annealing
(OSA) was proposed by Ventresca and Tizhoosh (2007).
They provided a theoretical basis for the algorithm as well
as its practical implementation. Simulation results on
six common real optimization problems confirmed the
theoretical predictions as well as showed a significant
improvement in accuracy and convergence rate over tra-
ditional SA. They also provided experimental evidence
for the use of opposite neighbors over purely random
ones.
Also, this strategy has been employed for several opti-
mization problems such as training neural network
(Subudhi and Jena 2009; Rashid and Baig 2010) and eco-
nomic dispatch problem (Balamurugan and Subramanian
2009; Bhattacharya and Chattopadhyay 2010). Because
random generation does not necessarily provide a good
initial population, Kofjac and Kljajic (2008) used the
mirroring of initial population by inspiring from the
opposite-based population initialization to solve a job shop
scheduling problem.
3 The DE algorithm
The DE such as other EAs starts with an initial sampling of
individuals within the search space. Generally speaking,
the initialization is performed randomly with a uniform
distribution function. After initialization, three main stages
of DE, i.e., mutation, crossover and selection are carried
out.
In the mutation stage, at each generation and for each
individual a donor member using an operator is generated.
There are several operators for the mutation. In these
operators, a donor member is generated by adding a
weighted difference of two, four or six members to one
other member. Some of the mutation schemes proposed in
the literature are as follow:
DE=rand=1 : v ¼ xr1þ Fðxr3
� xr2Þ ð1Þ
DE=current=1 : v ¼ xi þ Fðxr2� xr1
Þ ð2ÞDE=best=1 : v ¼ xbest þ Fðxr2
� xr1Þ ð3Þ
DE=rand=2 : v ¼ xr1þ Fðxr3
� xr2Þ þ kðxr5
� xr4Þ ð4Þ
DE=current=2 : v ¼ xi þ Fðxr2� xr1
Þ þ kðxr4� xr3
Þ ð5ÞDE=best=2 : v ¼ xbest þ Fðxr2
� xr1Þ þ kðxr4
� xr3Þ ð6Þ
DE=current-to-rand=1 :
v ¼ xi þ kðxr1� xiÞ þ Fðxr3
� xr2Þ ð7Þ
DE=current-to-best=1 :
v ¼ xi þ kðxbest � xiÞ þ Fðxr2� xr1
Þ ð8Þ
DE=rand-to-best=2 :
v ¼ xr1þ kðxbest � xiÞ þ Fðxr2
� xr1Þ þ Kðxr4
� xr3Þ ð9Þ
OBL in the SDE algorithm 1307
123
where xi is the current point, xbest is the best point found so far
and xr 1, xr 2
, xr 3, xr 4
and xr 5are random points that are selected
from the population, where r1 = r2 = r3 = r4 = r5, and F,
K and k are three random numbers within a definite range.
After generation of a donor member, in the crossover
stage each member of the population is allowed to carry out
a crossover by mating with a donor member. There are a
few crossover strategies. One of the well-known crossover
strategies is named binary crossover. The binary crossover
operator is as follows:
uj ¼vj if randð0; 1Þ\CR _ j ¼ kxj otherwise
�ð10Þ
where uj, vj and xj are the jth gene of the offspring, donor
and current individuals, respectively. rand(0, 1) is a random
number in the range of [0,1] and CR is the user-supplied
crossover rate constant. k e {1, 2, …, N} is a randomly
chosen index, chosen once for each member to make sure
that in the u at least one parameter is always selected from v.
Also, N is the dimension of problem. In the binary cross-
over, a comparison of random variable, rand(0, 1) and
crossover rate, CR, determines which gene should be copied
from the current member or donor member.
There is another crossover operator called exponential
crossover, in which crossover rate regulates how many
consecutive genes of the donor individual v on average
should be copied to the offspring u. In this crossover, the
genes are taken from donor member until the random
variable is smaller or equal to CR for the first time. Then
the remaining genes are copied from the current member.
The exponential crossover is as follows:
uj ¼vj j ¼ hniN ; hnþ 1iN ; . . .; hnþ l� 1iNxj otherwise
�ð11Þ
where n and l are two integer numbers in the set of
{1, 2, …, N} that denote the starting point and number of
component, respectively. Also, angular brackets ‹ ›N denote
a modulo function with modulus N. According to Eq. (11),
to generate u, those of genes are between n - 1 and n ? l
must be copied from v and other genes will be copied from
x. The integer l is selected from {1, 2, …, N} according to
the following pseudo-code:
l ¼ 0;
While ðrandð0; 1Þ\CRÞ ^ ðl\NÞfl ¼ lþ 1;
gEnd While
After generation of an offspring, in the selection stage
and according to a one-to-one spawning strategy, the value
of the cost function in the point u is evaluated, and while
f(u) B f(x), x is replaced with u and otherwise no
replacement occurs:
x ¼ u f ðuÞ� f ðxÞx f ðuÞ[ f ðxÞ
�ð12Þ
This evolutionary process consisting of the mutation,
crossover and selection stages is repeated over several
generations until one of the stopping criteria is met.
4 The OBL
4.1 The concept behind OBL
In general, the EAs start with some initial solutions (initial
population) and try to improve them toward some optimal
solution(s). In the absence of a priori information about the
solution, starting with random guesses, generally with a
uniform distribution in whole search space, is a common
initialization. The computation time, among others, is
related to the distance of these initial guesses from the
optimal solution. We can improve our chance of starting
with a closer (fitter) solution by simultaneously checking
the opposite solution. By doing this, the fitter one (guess or
opposite guess) can be chosen as an initial solution. In fact,
according to probability theory, 50% of the time a guess is
further from the solution than its opposite guess. Therefore,
starting with the closer of the two guesses (as judged by its
fitness) has the potential to accelerate convergence. The
same approach can be applied not only to initial solutions,
but also continuously to each solution in the current pop-
ulation. However, before concentrating on OBL, we need
to define the concept of opposite numbers (Rahnamayan
et al. 2008).
An opposition-based number can be defined as follows.
Let x e [a, b] be a real number. The opposition number x̂ is
defined by
x̂ ¼ aþ b� x ð13Þ
Similarly, the opposite number in an N-dimensional
space can be defined as follows:
Let X = (x1, x2, …, xN) be a point in an N-dimensional
space where x1, x2, …, xN e R and xi e [ai, bi]Vi e
{1, 2, …, N}. The opposite point X̂ ¼ ðx̂1; x̂2; . . .; x̂NÞ is
completely defined by its coordinates
x̂i ¼ ai þ bi � xi ð14Þ
Now, by employing the definition of opposite point, the
opposition-based optimization can be defined as follows.
Let X = (x1, x2, …, xN) be a point in an N-dimensional
space (i.e., candidate solution). Assume f(�) is a fitness
function which is used to measure the candidate’s fitness.
1308 M. A. Ahandani, H. Alavi-Rad
123
According to the definition of the opposite point, X̂ ¼ðx̂1; x̂2; . . .; x̂NÞ is the opposite of X = (x1, x2, …, xN). Now
if f ðXÞ� f ðX̂Þ, i.e., X̂ has a better fitness than X, then point
X can be replaced with X̂; otherwise, we continue with X.
Hence, the point and its opposite point are evaluated
simultaneously to continue with the fitter one.
4.2 The OBL in DE
The OBL can be used in two stages of DE: firstly in the
initialization stage to achieve fitter starting candidate
solutions, while a priori knowledge about the initial
members does not exist; secondly, carrying out the DE to
force the current population to jump into some new can-
didate solutions, which ideally are fitter than the current
ones. Rahnamayan et al. (2008) called these two stages as
opposition-based population initialization and opposition-
based generation jumping, respectively.
4.2.1 Opposition-based population initialization
The following steps present the utilizing of the OBL for
population initialization (Rahnamayan et al. 2008).
1. Initialize the population pop with a size of Npop
randomly.
2. Calculate the opposite population by
Opopi;j ¼ aj þ bj � popi;j
i ¼ 1; 2; . . .;Npop; j ¼ 1; 2; . . .;Nð15Þ
where popi,j and Opopi,j denote the jth variable of the
ith member of the population and the opposite popu-
lation, respectively.
3. Select Npop the fittest individuals from {pop [ Opop}
as the initial population.
To demonstrate the efficiency of opposition-based strat-
egy to obtain members which are closer to optimum point
than simply random population, we use an example from a
minimization problem. Equation (16) and Fig. 1 show a
two-dimensional function and its surface plot, respectively.
f ðx; yÞ ¼ ðx2 þ y2Þ0:25sin 30 ðxþ 0:5Þ2 þ y2
h i0:1��
þ xj j þ yj jf ð0; 0Þ ¼ 0 � 10� x; y� 10 ð16Þ
Figure 2a shows a random population with Npop equal to
40 on surface plot of Eq. (16) and its corresponding
opposition members. Figure 2b shows the remaining
members after applying an opposition-based population
initialization technique. As can be seen, those of members
which are far from the minimum point are rejected and
only closer members to the minimum point are preserved.
4.2.2 Opposition-based generation jumping
In this stage, based on a jumping rate Jr (i.e., jumping
probability), after generating new population by DE oper-
ators, the opposite population is calculated and the Npop
fittest individuals are selected from the union of the current
population and the opposite population. Unlike opposition-
based initialization, generation jumping calculates the
opposite population dynamically. Instead of using vari-
ables’ predefined interval boundaries ([aj, bj]), generation
jumping calculates the opposite of each variable based on
minimum (MINjp) and maximum (MAXj
p) values of that
variable in the current population
Opopi;j ¼ MINpj þMAXp
j � popi;j
i ¼ 1; 2; . . .;Npop; j ¼ 1; 2; . . .;Nð17Þ
By staying within variables’ interval static boundaries,
we would jump outside of the already shrunken search
space and the knowledge of the current reduced space
(converged population) would be lost. Hence, we calculate
the opposite points by using variables’ current interval in
the population ([MINjp, MAXj
p]), which is, as the search
progresses, increasingly smaller than the corresponding
initial range [aj, bj] (Rahnamayan et al. 2008).
5 The modified versions of DE
In this section, we propose four modified versions of DE
which utilize the OBL techniques, i.e., opposition-based
population initialization as well as opposition-based gen-
eration jumping, in the SDE. Ahandani et al. (2010) by
inspiring from partitioning and shuffling processes
employed in the SFL algorithm gave the parallel search
ability to the DE algorithms and called them as the SDE.
The SDE such as SFL divides the population into several
Fig. 1 Surface plot of Eq. (16)
OBL in the SDE algorithm 1309
123
subsets referred to as memeplexes and each memeplex is
improved by the DE.
The SFL includes three exclusive stages: partitioning,
local search and shuffling. In this algorithm, after genera-
tion of the initial population, members of the population
are evaluated in the cost function. Then, all frogs are par-
titioned to several parallel subsets. In order to partition
frogs into memeplexes on the assumption of partitioning
m memeplexes, each containing n frogs, after sorting the
population in a decreasing order in terms of their function
evaluation value, frog ranking 1 goes to memeplex 1, frog
ranking 2 goes to memeplex 2,…, frog ranking m goes to
memeplex m; then the second member of each subset is
assigned as: frog ranking (m ? 1) goes to memeplex 1,
frog ranking (m ? 2) goes to memeplex 2,…, frog ranking
(m ? m) goes to memeplex m. This process continues to
assign all frogs into memeplexes. After partitioning, the
different subsets perform a local search independently
using an evolutionary process to evolve their quality. This
evolutionary process is iterated for a defined maximum
number of iterations. Then all subsets shuffle together and
the stopping criteria are checked, so that if these are not
met the algorithm will be continued.
Our proposed algorithms similar to SFL have all these
three stages, but, those such as the SDE, employ the DE as
evolutionary process to evolve the quality of each member
of memeplexes. Also, those employ the opposition-based
population initialization and opposition-based generation
jumping. All four proposed algorithms similarly use the
opposition-based population initialization to achieve fitter
initial individuals and their difference is in applying oppo-
sition-based generation jumping. Also these algorithms use
Eq. (8) (DE/current-to-best/1) and Eq. (11) (exponential
crossover) as mutation and crossover strategies, respectively.
In the exponential crossover, we randomly select n and
l from the set of {1, 2, …, N} and do not use the proposed
pseudo-code in Sect. 3 to decrease one of the control
parameters of DE.
The first modified version of SDE is called shuffled
opposition-based DE (SOBDE). The SOBDE employs the
opposition-based generation jumping after each iteration of
evolutionary process for each memeplex. Steps of the
SOBDE algorithm with m memeplexes, n members of each
memeplex and kmax defined iteration number of evolu-
tionary process are shown in Fig. 3. In this algorithm,
MINmemjp and MAXmemj
p denote minimum and maximum
values of jth variable in the current memeplex, respec-
tively. mem and Omem are members and corresponding
opposite members of the current memeplex, respectively.
Thus, the SOBDE considers each memeplex as an inde-
pendent set, for which each memeplex carries out abso-
lutely the DE stages and applies opposition-based
generation jumping on its members with a probability of Jr.
In other words, the SOBDE uses the ODE algorithm as
an evolutionary process to evolve members of each
memeplex.
The second modified version of SDE is called shuffled
extended opposition-based DE (SEOBDE). The SEOBDE
applies the opposition-based generation jumping as an
extra stage to improve the members of each memeplex.
Steps of the SEOBDE algorithm are shown in Fig. 4. In
this algorithm, to prevent fast convergence, Step 3.3.5 is
carried out based on Eq. (18).
v ¼ xi þ rem kðxg2� xiÞ þ Fðxr2
� xr1Þ;MaxSize
� �ð18Þ
where rem denotes the remainder of division kðxg2� xiÞ þ
Fðxr2� xr1
Þ on MaxSize. Also MaxSize is determined as
follows
X-axis
Y-a
xis
square:population triangle: opposition population
-10 -8 -6 -4 -2 0 2 4 6 8 10-10
-8
-6
-4
-2
0
2
4
6
8
10 square:population triangle: opposition population
X-axis
Y-a
xis
-10 -8 -6 -4 -2 0 2 4 6 8 10-10
-8
-6
-4
-2
0
2
4
6
8
10(a) (b)
Fig. 2 An example for opposition-based population strategy. a A set of 40 random members and their corresponding opposition members.
b Remaining members after applying opposition-based population technique
1310 M. A. Ahandani, H. Alavi-Rad
123
MaxSizej ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffijajj � jbjj
3� ðjajj þ jbjjÞ
sfor j ¼ 1; 2; . . .;N ð19Þ
where aj and bj are minimum and maximum variables’
interval boundaries, respectively. Thus based on the above
equation, the maximum allowable step size of donor
member for approaching the best member of all population,
i.e., xg2, is restricted. The MaxSize sets itself in terms of
minimum and maximum variables’ interval boundaries in
every problem.
The steps of the third modified version of SDE called
opposition-based SDE (OBSDE) are shown in Fig. 5. The
OBSDE, unlike two aforementioned algorithms which used
the opposition-based generation jumping inside evolution-
ary process to evolve members of a memeplex, employs it
after a complete iteration of SDE on the current population.
In this algorithm, MINjp and MAXj
p denote minimum and
maximum values of jth variable in the current population,
respectively.
The final modified version of SDE is called opposition-
based shuffled extended opposition-based DE (OB-SEO-
BDE). The OB-SEOBDE on one hand, such as the SEOBDE,
after generation of an offspring for current member of
memeplex, applies the opposition-based generation jumping
as an extra stage to improve the current offspring. On the
other hand, the OB-SEOBDE, such as OBSDE, employs
another opposition-based generation jumping stage after a
complete iteration of algorithm on all members of the current
population. Figure 6 shows the steps of the OB-SEOBDE
algorithm.
6 Computational results
In this section, different experiments were carried out to
assess the performance of the proposed algorithms. The
considered values to set parameters of different algorithms
are shown in Table 1. The values of parameters in this
table are results of our experiments carried out by each
algorithm on different functions. For example, we found
that the values 5 and 20 for m and n, respectively, led to a
better performance in the SOBDE, but value of 20 and 5
were suitable for three other algorithms.
Firstly, the focus of the experiments is to compare the
performance of the proposed algorithms with the original
SDE algorithm proposed by Ahandani et al. (2010). Then, a
Algorithm 1 (the SOBDE algorithm)
Begin SOBDE Step 1: generate and evaluate initial population of size popN based on "opposition-based population
initialization" strategy.Step 2: generate m memplexes with n members where popN m n= × .
Step 3: apply the OBDE (Step3. 0 to Step 3.5) to improve each memplex for maxk iterations:
Step 3.0: set counter 1k = . Step 3.1: While maxk k≤Improve each member of memplex ( ix ) as follow:
Step 3.2: set counter 1i = . Step 3.3: While i n≤Step 3.3.1: determine the best member of each memplex, g1x and the best member of population, g 2x .
Step 3.3.2: apply mutation and crossover operators for ix according to Eq. (8) and Eq. (11) with best g1x x= and
generate new member of u . Step 3.3.3: evaluate value of cost function in the point u and while ( ) ( )if u f x≤ , replace ix with u and set
1i i= + , else go to next step.Step 3.3.4: repeat steps 3.3.2 and 3.3.3 with best g2x x= .
End While Step 3.4: apply "opposition-based generation jumping" strategy based on a jumping rate of rJ :
0 1
1
1
,,
If ( , )
For :
For :
;
End For
End For
r
p pjiji j j
rand J
i n
j D
Omem MINmem MAXmem mem
<=
=
= + −
Select n fittest members from the set of { }mem Omem∪ as current memeplex.
Step 3.5: set 1k k= + . End While
Step 4: shuffle the population. Step 5: check the stopping criteria if are not met go to Step 2. End SOBDE
Fig. 3 The steps of SOBDE
OBL in the SDE algorithm 1311
123
comparison is carried out among the proposed algorithm
that obtained a better performance and some modern DE
algorithms proposed in the literature. The performance
comparison is made on 25 benchmark functions designed
for the special session on real-parameter optimization of
CEC2005 (Suganthan et al. 2005). Hansen (2005) split
these benchmark functions into three subsets:
• unimodal functions,
• solved multimodal functions (at least one algorithm
conducted at least one successful run),
• unsolved multimodal functions (no single run was
successful for any algorithm).
Unimodal functions are F1–F6, solved multimodal
functions are F7, F9–F12 and F15, and unsolved multimodal
functions include F8, F13 and F16–F25.
Twenty-five runs were performed for each benchmark
function. The results of 1st (min), 7th, 13th (median), 19th
and 25th (max) run are reported. All runs are averaged
(Avg) and standard deviation (SD) is also given. Also, a
non-parametric analysis over the obtained results using of
the Wilcoxon signed-ranks test with a = 0.05 (see Garcia
et al. 2009) in terms of the best run (1st run), average of all
runs (Avg) and success rate (SR) is provided. These studies
by means of pairwise comparisons compare the modified
versions of DE with the SDE and other DE algorithms
proposed in the literature. Experiments are carried out for
10, 30 and 50 variable numbers. The obtained results are
reported in Tables 2, 3, 4, 5, 6, 7, 8 and 9.
Table 2 shows a comparison among the SDE2 algorithm
proposed by Ahandani et al. (2010) and four versions of the
DE algorithm proposed in this research. These experiments
are carried out with ten variable numbers. In this table, we
intend to show how well the proposed algorithms perform
when compared with the conventional SDE.
On F1 to F4, all algorithms have a similar performance.
Those found optimal solutions on F1, F2 and F4 functions
with a success rate of 100%. On F5 to F7, success rate
obtained by the proposed algorithms are considerably
better than that obtained by the SDE2, and the SOBDE
obtains the best average results. It can be seen that, unlike
the SDE which has only a success rate of 44% on F7, our
proposed algorithms have at least a success rate of 84%.
On F8, the SEOBDE and OB-SEOBDE obtain a better
average and minimum results, respectively. On F9, all
algorithms have a successful performance. On F10, only the
SOBDE had a success rate of 4% and all other algorithms
did not converge to a minimum point at all. Also, this
algorithm has the best minimum and average results on
F11. On F12 and F13, the OB-SEOBDE obtains the best
Algorithm 2 (the SEOBDE algorithm)
Begin SEOBDE Step 1: generate and evaluate initial population of size popN based on "opposition-based population
initialization" strategy.Step 2: generate m memplexes with n members where popN m n= × .
Step 3: apply the extended OBDE (Step3. 0 to Step 3.5) to improve each memplex for maxk iterations:
Step 3.0: set counter 1k = . Step 3.1: While maxk k≤Improve each member of memplex ( ix ) as follow:
Step 3.2: set counter 1i = . Step 3.3: While i n≤Step 3.3.1: determine the best member of each memplex, g1x and the best member of population, g 2x .
Step 3.3.2: apply mutation and crossover operators for ix according to Eq. (8) and Eq. (11) with best g1x x= and
generate new member of u . Step 3.3.3: evaluate value of cost function in the point u and while ( ) ( )if u f x≤ , replace ix with u .
Step 3.3.4: apply "opposition-based generation jumping" strategy based on a jumping rate of rJ on u , and
generate Ou : 0 1
1
,,
If ( , )
For :
;
End For
r
p pjiji j j
rand J
j D
Ou MINmem MAXmem u
<=
= + −
While ( ) ( )if Ou f x≤ , replace ix with Ou and set 1i i= + , else go to next step.
Step 3.3.5: repeat steps 3.3.2 to 3.3.4 with best g2x x= .
End While Step 3.4: set 1k k= + . End While
Step 4: shuffle the population. Step 5: check the stopping criteria if are not met go to Step 2. End SEOBDE
Fig. 4 The steps of SEOBDE
1312 M. A. Ahandani, H. Alavi-Rad
123
success rate and average results. F13 is categorized in the
unsolved multimodal functions, but our proposed algo-
rithms solve this problem. A success rate of 36% is
obtained by the OB-SEOBDE on this function. To the best
of our knowledge, after Becker et al. (2005) obtained a
success rate of 4% with a minimum value of 9.8771e-3 on
F13, this is the first time that an algorithm has a success-
ful performance on this function. On F14, the SDE2
outperforms all other algorithms. On F15, the OBSDE and
OB-SEOBDE have a better success rate, but the OBSDE
has the best average results.
F16 to F25 are unsolved multimodal functions. Our
proposed algorithms for the first time obtain some prom-
ising results on these functions. On F16, the OB-SEOBDE
has a better success rate and average results. To the best of
our knowledge, this is for the first time that an algorithm
solves this function. On F17, also for the first time the
SEOBDE and OBSDE have a success rate of 4%, but the
OB-SEOBDE obtains a better average results with respect
to the other algorithms under investigation. On F18 and F19,
the OB-SEOBDE has the best performance. F18 is solved
for the first time by the OB-SEOBDE algorithm. On F20,
all algorithms have a similar performance in terms of
minimum obtained result, but the OBSDE has a better
average result. On F21 and F22, the SEOBDE and
OB-SEOBDE obtain the best results, respectively. On F23,
the SOBDE and OB-SEOBDE obtain a better minimum
and average results, respectively. On F24 and F25, the
SEOBDE and SOBDE obtain the best average results,
respectively.
Table 3 shows the pairwise comparisons of the obtained
results of Table 2. These analyses demonstrate that the
performance of OB-SEOBDE has a significant difference
with respect to the SDE2. The OB-SEOBDE considerably
outperforms the SDE2 in terms of all three considered
aspects. Also, the SEOBDE has a slightly better perfor-
mance in terms of the average obtained results and success
rate with respect to SDE2. The OBSDE has a better per-
formance in terms of the average obtained results com-
pared to the SDE2. Overall, all results of the Wilcoxon test
show a better performance of modified algorithms, but only
performance of the aforementioned algorithms is signifi-
cantly better compared to the SDE2. On the other side, a
pairwise comparison only among our algorithms proposed
in this research shows that any algorithm does not have a
significant difference from other algorithms.
Algorithm 3 (the OBSDE algorithm)
Begin OBSDE Step 1: generate and evaluate initial population of size popN based on "opposition-based population
initialization" strategy.Step 2: generate m memplexes with n members where popN m n= × .
Step 3: apply the DE (Step3. 0 to Step 3.4) to improve each memplex for maxk iterations:
Step 3.0: set counter 1k = . Step 3.1: While maxk k≤Improve each member of memplex ( ix ) as follow:
Step 3.2: set counter 1i = . Step 3.3: While i n≤Step 3.3.1: determine the best member of each memplex, g1x and the best member of population, g 2x .
Step 3.3.2: apply mutation and crossover operators for ix according to Eq. (8) and Eq. (11) with best g1x x= and
generate new member of u . Step 3.3.3: evaluate value of cost function in the point u and while ( ) ( )if u f x≤ , replace ix with u and set
1i i= + , else go to next step.Step 3.3.4: repeat steps 3.3.2 and 3.3.3 with best g2x x= .
End While Step 3.4: set 1k k= + . End While
Step 4: shuffle the population. Step 5: apply "opposition-based generation jumping" strategy based on a jumping rate of rJ :
0 1
1
1
,,
If ( , )
For :
For :
;
End For
End For
r
pop
p pjiji j j
rand J
i N
j D
Opop MIN MAX pop
<=
=
= + −
Select popN fittest members from the set of { }pop Opop∪ as current population.
Step 6: check the stopping criteria if are not met go to Step 2. End OBSDE
Fig. 5 The steps of OBSDE
OBL in the SDE algorithm 1313
123
Based on the results of Table 2 and clear statistical
evidence of Table 3, it can be concluded that combining
the OBL strategy in the SDE gives a promising and proper
approach to evolve the DE algorithm. This is clearly
reflected also by the Wilcoxon test. Our proposed algo-
rithms for the first time solve some of the functions that did
not have any successful reported results on them.
To evaluate the efficiency and quality of the proposed
DE algorithms, a comparison among our proposed
algorithms that obtained a better performance was per-
formed, i.e., the OB-SEOBDE, and some modern DE
algorithms. The following algorithms have been taken into
account for this goal: the Guided-DE and SaDE algorithms
were taken from (Bui et al. 2005) and (Qin and Suganthan
2005), respectively, and jDE and jDE-2 were taken from
(Brest et al. 2007). Table 4 shows the obtained results of
this comparison. Also Table 5 illustrates the results of
pairwise comparison for Table 4.
From Table 4, it is clearly observed that the OB-SEO-
BDE algorithm not only solves some unsolved functions
for the first time, but also obtains some promising and
comparable performances on the other algorithms. It has
the best success rate on F7 and also on F8, after the SaDE
obtains the best minimum result. On F12, only the
OB-SEOBDE and SaDE algorithms obtain a success rate of
100%. On F16 to F18, F22, F23 and F25, the OB-SEOBDE
obtains the best minimum result. Also on F16, F17, F22,
Algorithm 4 (the OB-SEOBDE algorithm) Begin SEOBDE Step 1: generate and evaluate initial population of size popN based on "opposition-based population
initialization" strategy.Step 2: generate m memplexes with n members where popN m n= × .
Step 3: apply the DE (Step3. 0 to Step 3.5) to improve each memplex for maxk iterations:
Step 3.0: set counter 1k = . Step 3.1: While maxk k≤Improve each member of memplex ( ix ) as follow:
Step 3.2: set counter 1i = . Step 3.3: While i n≤Step 3.3.1: determine the best member of each memplex, g1x and the best member of population, g 2x .
Step 3.3.2: apply mutation and crossover operators for ix according to Eq. (8) and Eq. (11) with best g1x x= and
generate new member of u . Step 3.3.3: evaluate value of cost function in the point u and while ( ) ( )if u f x≤ , replace ix with u .
Step 3.3.4: apply "opposition-based generation jumping" strategy based on a jumping rate of rJ on u , and
generate Ou : 0 1
1
,,
If ( , )
For :
;
End For
r
p pjiji j j
rand J
j D
Ou MINmem MAXmem u
<=
= + −
While ( ) ( )if Ou f x≤ , replace ix with Ou and set 1i i= + , else go to next step.
Step 3.3.5: repeat steps 3.3.2 to 3.3.4 with best g2x x= .
End While Step 3.5: set 1k k= + . End While
Step 4: shuffle the population. Step 5: apply "opposition-based generation jumping" strategy based on a jumping rate of rJ :
0 1
1
1
,,
If ( , )
For :
For :
;
End For
End For
r
pop
p pjiji j j
rand J
i N
j D
Opop MIN MAX pop
<=
=
= + −
Select popN fittest members from the set of { }pop Opop∪ as current population.
Step 6: check the stopping criteria if are not met go to Step 2. End OB-SEOBDE
Fig. 6 The steps of
OB-SEOBDE
Table 1 Parameter setting for algorithms
Algorithms Npop F k m n kmax Jr
SOBDE 100 [0,1] [0,1.5] 5 20 20 0.3
SEOBDE 100 [0,1] [0,1.5] 20 5 20 0.3
OBSDE 100 [0,1] [0,1.5] 20 5 20 0.3
OB-SEOBDE 100 [0,1] [0,1.5] 20 5 20 0.3
1314 M. A. Ahandani, H. Alavi-Rad
123
Table 2 A comparison among different modified versions of SDE algorithm with original SDE
F SDE2 SOBDE SEOBDE OBSDE OB-SEOBDE
1
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 0
25th 0 0 0 0 0
Avg 0 0 0 0 0
SD 0 0 0 0 0
SR 100% 100% 100% 100% 100%
2
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 0
25th 0 0 0 0 1.0097e-028
Avg 0 0 0 0 4.039e-030
SD 0 0 0 0 2.0195e-029
SR 100% 100% 100% 100% 100%
3
1st 1,482.1 8,798 26,665 81,040 48,558
7th 1.9042e?05 88,741 1.0264e?005 3.4653e?005 2.6209e?005
13th 6.8461e?05 242,730 3.127e?005 6.965e?005 5.2998e?005
19th 1.9124e?06 604,330 7.7417e?005 1.1847e?006 1.3261e?006
25th 6.8311e?06 991,490 4.1338e?006 5.2102e?006 4.7468e?006
Avg 1.4238e?06 3.6271e?005 6.0006e?005 1.1279e?006 1.0695e?006
SD 1.8216e?06 3.201e?005 8.5133e?005 1.3115e?006 1.2453e?006
SR 0% 0% 0% 0% 0%
4
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 0
25th 0 0 0 0 0
Avg 0 0 0 0 0
SD 0 0 0 0 0
SR 100% 100% 100% 100% 100%
5
1st 1.3529e-09 0 0 4.5475e-012 0
7th 2.2685e-08 0 1.819e-012 6.4574e-011 3.638e-012
13th 7.6028e-08 0 5.457e-012 1.7917e-010 1.819e-011
19th 1.9793e-07 1.819e-012 5.0932e-011 9.5679e-010 9.8225e-011
25th 0.00058102 2.7285e-012 2.2737e-009 3.5782e-007 7.2396e-010
Avg 2.3374e-05 7.8217e-013 1.4799e-010 1.4917e-008 1.1216e-010
SD 0.00011618 1.0128e-012 4.5788e-010 7.1443e-008 1.9874e-010
SR 96% 100% 100% 100% 100%
6
1st 0 0 0 0 0
7th 4.4366e-027 0 0 0 0
OBL in the SDE algorithm 1315
123
Table 2 continued
F SDE2 SOBDE SEOBDE OBSDE OB-SEOBDE
13th 5.3977e-026 0 0 0 0
19th 1.3765e-023 0 0 1.0747e-026 0
25th 3.9866 3.9866 3.9866 3.9866 3.9866
Avg 0.47839 0.15946 0.15946 0.31893 0.31893
SD 1.3222 0.79732 0.79732 1.1038 1.1038
SR 88% 96% 96% 92% 92%
7
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0.31471 0 0 0 0
19th 0.43267 0 0 0 0
25th 0.56817 0.2522 0.51323 0.52774 0.5001
Avg 0.23472 0.02132 0.055695 0.053074 0.071392
SD 0.22308 0.071751 0.15462 0.14981 0.16767
SR 44% 92% 84% 88% 84%
8
1st 20.185 20.182 20.175 20.207 20.161
7th 20.262 20.268 20.286 20.317 20.262
13th 20.314 20.327 20.333 20.345 20.330
19th 20.395 20.382 20.382 20.382 20.411
25th 20.436 20.459 20.449 20.423 20.457
Avg 20.352 20.357 20.322 20.338 20.337
SD 0.07194 0.070692 0.079901 0.055888 0.077742
SR 0% 0% 0% 0% 0%
9
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 0
25th 0 0 0 0 0
Avg 0 0 0 0 0
SD 0 0 0 0 0
SR 100% 100% 100% 100% 100%
10
1st 0.3246 0 2.5163 1.9930 3.9798
7th 3.9798 1.9899 5.6396 4.9748 4.9748
13th 6.1726 2.9849 6.9647 5.9737 5.9698
19th 9.2496 3.9798 8.2866 8.7600 7.8613
25th 11.941 4.9748 10.945 11.9740 11.939
Avg 6.115 2.9849 6.7233 6.5850 6.2845
SD 3.0802 1.3472 2.0618 2.4629 2.0102
SR 0% 4% 0% 0% 0%
11
1st 4.0696 0.20195 2.2851 0.2841 1.5927
7th 5.8809 0.34713 3.8167 4.0300 3.7245
13th 6.5069 1.1816 4.7275 4.3902 4.5041
19th 6.8359 3.977 5.3568 5.0413 5.1917
25th 7.1235 5.9717 6.4281 6.2281 5.9237
1316 M. A. Ahandani, H. Alavi-Rad
123
Table 2 continued
F SDE2 SOBDE SEOBDE OBSDE OB-SEOBDE
Avg 6.1784 2.4006 4.5245 4.0058 4.3741
SD 0.63647 1.918 1.0891 1.6514 1.0961
SR 0% 0% 0% 0% 0%
12
1st 8.3251e-021 0 0 2.6253e-027 0
7th 2.9908e-019 0 4.5438e-028 3.8562e-018 0
13th 5.0646e-016 1.1107e-027 1.7184e-024 9.1946e-010 2.2883e-026
19th 1.1069e-012 7.622e-020 3.2081e-022 1.086e-006 1.2591e-017
25th 0.33006 3.1991 10.26 30.909 0.00032304
Avg 0.01321 0.13665 0.41039 1.2716 1.2922e-005
SD 0.06601 0.63948 2.052 6.1767 6.4608e-005
SR 96% 92% 96% 88% 100%
13
1st 0 0 0 0 0
7th 0.53028 0.33998 0.0098704 0.1148 0
13th 0.61752 0.39505 0.25542 0.3631 0.23436
19th 0.65918 0.54487 0.45909 0.4519 0.40209
25th 0.75641 0.70635 0.66559 0.6088 0.58901
Avg 0.54499 0.40265 0.2658 0.2912 0.22196
SD 0.19019 0.19736 0.22287 0.1991 0.21824
SR 4% 8% 28% 20% 36%
14
1st 1.8054 2.1365 2.1794 2.4224 2.773
7th 2.4539 2.6075 2.9611 3.0985 3.1501
13th 2.9027 2.7998 3.1416 3.2188 3.2706
19th 3.1588 3.1274 3.2562 3.3144 3.3128
25th 3.6929 3.527 3.6604 3.5209 3.5885
Avg 2.8039 2.8313 3.0661 3.1697 3.2259
SD 0.42955 0.3566 0.3195 0.2435 0.20208
SR 0% 0% 0% 0% 0%
15
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 200.0000 200.0000 100.0000 0.0338 35.457
19th 400.0000 400.0000 200.0000 150.2511 200.0000
25th 404.5218 407.942 237.2780 300.0000 300.0000
Avg 171.4286 229.76 101.4132 74.0761 97.924
SD 179.9471 153.98 95.2440 94.3074 104.2
SR 36% 32% 36% 44% 44%
16
1st 15.0245 14.372 1.3797e-013 9.4760 0
7th 21.6688 24.841 18.6846 31.2192 18.057
13th 28.6613 30.908 27.6082 43.0641 37.004
19th 33.6794 43.463 46.2769 91.5970 46.277
25th 115.2318 107.414 134.8407 121.6007 96.174
Avg 43.5085 44.148 40.2680 57.4225 37.171
SD 56.7509 30.171 35.2638 36.0052 25.682
SR 0% 0% 4% 0% 8%
OBL in the SDE algorithm 1317
123
Table 2 continued
F SDE2 SOBDE SEOBDE OBSDE OB-SEOBDE
17
1st 30.5549 10.175 0 0 11.5331
7th 39.5409 26.252 19.8993 23.1548 20.3217
13th 56.9181 36.603 35.1286 36.4231 28.4788
19th 89.6497 55.552 43.4500 47.0098 41.5130
25th 150.1124 89.736 122.5260 150.5640 57.0901
Avg 68.7787 42.533 45.2879 43.9726 29.4416
SD 52.7338 22.427 34.7103 33.3869 12.3062
SR 0% 0% 4% 4% 0%
18
1st 300.0000 300.0000 300.0000 300.0000 3.2816e-013
7th 400.0000 300.0000 400.0000 400.0000 300.0000
13th 808.7119 408.5410 500.0000 500.0000 400.0000
19th 824.1082 814.5702 800.0000 808.7699 500.0000
25th 902.5931 910.6617 909.5770 866.3345 830.8944
Avg 747.2209 554.0830 587.4405 592.3354 462.8617
SD 204.0083 245.7911 231.8998 227.2923 218.2436
SR 0% 0% 0% 0% 4%
19
1st 300.0000 300.000 300.0000 300.0000 300.0000
7th 400.0000 400.000 400.0000 335.9206 336.21472
13th 811.0385 804.2170 407.1330 400.0000 400.0000
19th 822.2928 829.4831 802.8916 800.0000 500.0000
25th 885.8953 930.0147 971.3249 882.6131 849.1442
Avg 691.0391 646.4917 580.2510 529.0426 465.7913
SD 198.8886 230.5831 238.4771 225.8041 181.8147
SR 0% 0% 0% 0% 0%
20
1st 300.000 300.000 300.0000 300.0000 300.0000
7th 300.000 400.000 300.0000 400.0000 374.7275
13th 400.000 800.000 500.0000 400.0000 400.0000
19th 800.000 815.9904 800.0000 800.0000 800.0000
25th 848.9451 860.7317 903.3796 825.0095 852.1974
Avg 545.4350 615.1227 563.0376 528.6478 543.4026
SD 247.2186 219.95 233.6682 214.1895 224.0738
SR 0% 0% 0% 0% 0%
21
1st 500.0000 500.0000 410.4922 410.4945 500.0000
7th 531.9214 524.6717 500.0000 500.0000 500.0000
13th 614.3927 586.4908 557.6002 567.2293 584.2689
19th 673.65 687.9217 648.2883 616.4363 678.5980
25th 900.0000 828.7513 747.2052 780.8986 900.0000
Avg 640.0708 619.5562 572.3121 573.8298 605.6206
SD 42.6968 106.4992 98.4800 84.7315 104.2032
SR 0% 0% 0% 0% 0%
22
1st 500.0000 500.0000 200.0000 500.0000 200.0000
7th 500.0058 500.0024 500.0002 500.0012 500.0001
1318 M. A. Ahandani, H. Alavi-Rad
123
F23and F25, the OB-SEOBDE obtains the best average
results. Also on F13, F16 and F18, only the OB-SEOBDE
has a successful performance. Furthermore from Table 4,
we can observe that the OB-SEOBDE besides the Guided-
DE is the algorithm with the worst performance on F3.
Also from the results of non-parametric analysis of
Table 5, it is clearly shown that the OB-SEOBDE outper-
forms the Guided-DE. It has a considerably better perfor-
mance than the Guided-DE based on all three considered
aspects. Also, the OB-SEOBDE has a considerably better
performance than the jDE in terms of the best run. Also, the
paired pairwise comparison of the OB-SEOBDE and other
algorithms does not show a significant difference. As an
overall conclusion, the results of Table 5 show that
although the performance of OB-SEOBDE and some
algorithms does not have a significant difference, in all
cases the OB-SEOBDE is the fitter algorithm.
To examine the performance of the OB-SEOBDE
on high-dimensional functions, a comparison among the
OB-SEOBDE, the Guided-DE proposed in Bui et al. (2005)
and three modified versions of DE proposed in Ahandani
et al. (2010), i.e., the BDE2, SDE2 and SBDE2, is carried
out in Tables 6, 7, 8 and 9 for 30 and 50 variables.
Applying the OB-SEOBDE on high-dimensional func-
tions highlights its efficiency. Table 6 shows the compar-
ison results for 30 variables. The obtained results of
this table and the pairwise comparisons of those in Table 7
clearly demonstrate that the OB-SEOBDE has a
Table 2 continued
F SDE2 SOBDE SEOBDE OBSDE OB-SEOBDE
13th 500.2138 500.1100 500.0019 500.0121 500.0094
19th 518.6937 503.2314 500.0492 500.3966 500.0353
25th 889.2026 508.7924 802.3267 510.6857 576.6202
Avg 594.5276 502.2866 500.2884 500.6647 487.2779
SD 160.2612 2.5360 86.9411 2.1514 64.9543
SR 0% 0% 0% 0% 0%
23
1st 510.1981 508.0902 523.1287 507.0437 529.3801
7th 566.9675 548.0127 543.5150 526.9657 529.3801
13th 588.5153 613.0882 567.7894 554.4714 565.9457
19th 614.5676 809.5482 603.1247 635.4232 611.2536
25th 701.0727 949.4413 928.7413 827.3615 766.0608
Avg 594.8642 660.5117 589.5522 597.6233 575.1675
SD 76.0495 156.7053 84.3700 94.7825 77.5191
SR 0% 0% 0% 0% 0%
24
1st 200.0000 200.0000 200.0000 200.0000 200.0000
7th 200.0000 202.8504 202.5138 207.0925 200.0000
13th 203.4436 216.1271 207.3989 213.7113 205.9986
19th 216.5592 251.1467 220.4807 223.2402 248.8015
25th 500.0000 421.0412 400.0000 500.0000 500.0000
Avg 243.4304 254.1792 226.4891 237.7216 246.2228
SD 92.5408 78.5770 53.6241 79.8734 84.4419
SR 0% 0% 0% 0% 0%
25
1st 200.0001 200.0000 200.0000 200.0000 200.0000
7th 200.0528 205.1204 200.1249 200.9169 200.0000
13th 201.1032 229.5440 207.9837 207.1592 400.0000
19th 221.9227 272.5841 259.5602 313.7887 500.0000
25th 597.0077 816.2591 1122.2786 900.0000 500.0000
Avg 284.0225 290.1214 311.2877 302.0553 354.0000
SD 115.2128 142.7110 241.8858 181.2217 147.7973
SR 0% 0% 0% 0% 0%
OBL in the SDE algorithm 1319
123
significantly better performance than other algorithms in
terms of 1st run and Avg. It also outperforms the Guided-
DE in terms of success rate, but there is not a significant
difference among success rates of the OB-SEOBDE and
other algorithms.
Table 8 shows the comparison results for 50 variables.
The Guided-DE was not applied for 50 variables; thus, this
table is only a comparison of the OB-SEOBDE and three
DE algorithms proposed in Ahandani et al. (2010). Table 8
and the pairwise comparisons of its results in Table 9
clearly confirm the results of Tables 6 and 7. The OB-
SEOBDE considerably outperforms three other algorithms
in terms of 1st run and Avg. Also those do not have a
significant difference for success rate.
Table 3 Wilcoxon test applied over the all possible comparisons among algorithms obtained from Table 2 in terms of the best run (1st run),
average of all runs (Avg) and success rate (SR)
Comparison 1st run Avg SR
R? R- p value R? R- p value R? R- p value
SDE2–SOBDE 227.5 97.5 0.155 193 132 0.394 127.5 197.5 0.161
SDE2–SEOBDE 205 120 0.382 253 72 0.017 95 230 0.026
SDE2–OBSDE 200.5 144.5 0.695 239 86 0.042 108 217 0.105
SDE2–OB-SEOBDE 203 122 0.019 261 64 0.009 76.5 248.5 0.011
SOBDE–SEOBDE 164 161 0.878 180.5 144.5 0.502 130.5 194.5 0.380
SOBDE–OBSDE 141.5 183.5 0.859 161 164 0.876 169.5 155.5 0.725
SOBDE–OB-SEOBDE 144 181 0.169 217 108 0.115 136.5 188.5 0.203
SEOBDE–OBSDE 116.5 208.5 0.203 137 188 0.590 185 140 0.453
SEOBDE–OB-SEOBDE 136 189 0.508 196 129 0.372 127.5 197.5 0.160
OBSDE–OB-SEOBDE 172 153 0.221 192 133 0.414 137 188 0.168
Table 4 A comparison among our proposed OB-SEOBDE with some modern DE algorithms proposed in the literature
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
1
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 0
25th 0 0 0 0 0
Avg 0 0 0 0 0
SD 0 0 0 0 0
SR 100% 100% 100% 100% 100%
2
1st 0 0 0 0 0
7th 0 0 0 0 0
13th 0 0 0 0 0
19th 0 0 0 0 2.7755e-17
25th 1.0097e-028 0.958 2.5580e-12 0 4.4408e-16
Avg 4.039e-030 0.057 1.0459e-13 0 2.8865e-17
SD 2.0195e-029 0.193 5.1124e-13 0 9.0109e-17
SR 100% 80% 100% 100% 100%
3
1st 48,558 1.34e?04 0 0 1.0963e-14
7th 2.6209e?005 4.63e?04 0 0 6.1627e-12
13th 5.2998e?005 1.47e?05 0 0 1.3901e-10
19th 1.3261e?006 2.83e?05 9.9142e-06 2.7758e-17 9.5245e-09
1320 M. A. Ahandani, H. Alavi-Rad
123
Table 4 continued
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
25th 4.7468e?006 9.41e?05 1.0309e-04 1.3569e-11 6.2006e-06
Avg 1.0695e?006 2.09e?05 1.6720e-05 5.4316e-13 4.3401e-07
SD 1.2453e?006 2.04e?05 3.1196e-05 2.7138e-12 1.3459e-06
SR 0% 0% 64% 100% 88%
4
1st 0 0 0 0 0
7th 0 0 0 0 5.5511e-17
13th 0 0.065 0 0 1.9428e-16
19th 0 1.090 0 0 4.4408e-16
25th 0 2.942 3.5456e-04 0 4.0245e-15
Avg 0 0.619 1.4182e-05 0 4.4186e-16
SD 0 0.921 7.0912e-05 0 8.1713e-16
SR 100% 40% 96% 100% 100%
5
1st 0 37.703 1.1133e-06 0 1.7763e-15
7th 3.638e-012 66.064 0.0028 0 5.3290e-15
13th 1.819e-011 108.370 0.0073 0 1.0658e-14
19th 9.8225e-011 168.983 0.0168 0 1.5987e-14
25th 7.2396e-010 250.417 0.0626 0 5.5067e-14
Avg 1.1216e-010 121.796 0.0123 0 1.5312e-14
SD 1.9874e-010 63.770 0.0146 0 1.3994e-14
SR 100% 0% 0% 100% 100%
6
1st 0 0.082 0 0 7.8986e-13
7th 0 0.839 4.3190e-09 0 5.9842e-03
13th 0 1.722 5.1631e-09 0 0.0203
19th 0 4.096 9.1734e-09 2.7755e-17 0.0886
25th 3.9866 18.128 8.0479e-08 1.1379e-15 0.6044
Avg 0.31893 3.538 1.1987e-08 1.2767e-16 0.0760
SD 1.1038 4.068 1.9282e-08 3.0520e-16 0.1303
SR 92% 0% 100% 100% 32%
7
1st 0 64.788 4.6700e-10 0 1.8444e-13
7th 0 132.678 0.0148 7.3960e-03 0.0322
13th 0 153.585 0.0197 0.0123 0.0555
19th 0 186.077 0.0271 0.0270 0.0767
25th 0.5001 231.308 0.0369 0.0467 0.1124
Avg 0.071392 157.751 0.0199 0.0167 0.0531
SD 0.16767 39.583 0.0107 0.0140 0.0310
SR 84% 0% 24% 40% 12%
8
1st 20.161 20.267 20.0000 20.1999 20.1727
7th 20.262 20.390 20.0000 20.3017 20.2538
13th 20.330 20.449 20.0000 20.3444 20.3384
19th 20.411 20.497 20.0000 20.4060 20.3951
25th 20.457 20.572 20.0000 20.5016 20.4314
Avg 20.337 20.444 20.0000 20.3541 20.3206
SD 0.077742 0.077 5.3901e-08 0.0711 0.0750
OBL in the SDE algorithm 1321
123
Table 4 continued
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
SR 0% 0% 0% 0% 0%
9
1st 0 0 0 0 0
7th 0 1.990 0 0 0
13th 0 3.980 0 0 0
19th 0 4.975 0 0 0
25th 0 10.945 0 0 0
Avg 0 3.967 0 0 0
SD 0 2.911 0 0 0
SR 100% 12% 100% 100% 100%
10
1st 3.9798 3.980 1.9899 2.2576 4.4341
7th 4.9748 8.955 3.9798 5.1310 7.6225
13th 5.9698 12.935 4.9748 5.9697 9.6381
19th 7.8613 16.914 5.9698 7.4081 10.3258
25th 11.939 29.849 9.9496 9.9866 13.2220
Avg 6.2845 13.651 4.9685 6.1680 9.2627
SD 2.0102 6.414 1.6918 1.9945 1.9270
SR 0% 0% 0% 0% 0%
11
1st 1.5927 0.541 3.2352 8.9258e-10 0.0859
7th 3.7245 0.956 4.5129 2.0095e-05 4.7395
13th 4.5041 2.202 4.7649 3.2703e-03 5.1636
19th 5.1917 2.807 5.3823 0.1453 5.9770
25th 5.9237 4.759 5.9546 6.4392 6.8224
Avg 4.3741 2.196 4.8909 0.8058 5.0802
SD 1.0961 1.140 0.6619 1.9162 1.4210
SR 0% 0% 0% 56% 0%
12
1st 0 0 1.4120e-10 0 2.8310e-15
7th 0 1.668 1.7250e-08 0 1.0361e-13
13th 2.2883e-026 10.731 8.1600e-08 0 4.6544e-09
19th 1.2591e-017 21.505 3.8878e-07 10.0030 1.2524e-04
25th 0.00032304 1,556.660 3.3794e-06 712.2541 35.4601
Avg 1.2922e-005 262.015 4.5011e-07 67.8339 4.1255
SD 6.4608e-005 484.243 8.5062e-07 198.4822 8.7597
SR 100% 0% 100% 68% 76%
13
1st 0 0.272 0.1201 0.3701 0.3090
7th 0 0.444 0.1957 0.4394 0.3988
13th 0.23436 0.568 0.2170 0.5053 0.4573
19th 0.40209 0.702 0.2508 0.5700 0.5006
25th 0.58901 1.347 0.3117 0.6486 0.5458
Avg 0.22196 0.598 0.2202 0.5052 0.4504
SD 0.21824 0.238 0.0411 0.0813 0.0631
SR 36% 0% 0% 0% 0%
14
1st 2.773 2.558 2.5765 1.4669 2.8531
1322 M. A. Ahandani, H. Alavi-Rad
123
Table 4 continued
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
7th 3.1501 3.000 2.7576 2.6163 3.2235
13th 3.2706 3.233 2.8923 2.8582 3.2685
19th 3.3128 3.395 3.0258 2.9935 3.3636
25th 3.5885 3.742 3.3373 3.2399 3.5279
Avg 3.2259 3.208 2.9153 2.7747 3.2613
SD 0.20208 0.274 0.2063 0.3540 0.1668
SR 0% 0% 0% 0% 0%
15
1st 0 0 0 0 0
7th 0 77.738 0 0 0
13th 35.457 119.676 0 400.0000 67.8032
19th 200.0000 406.258 2.9559e-12 400.0000 100.5804
25th 300.0000 441.765 400.0000 400.0000 400.0000
Avg 97.924 183.862 32.0000 224.0000 95.1523
SD 104.2 153.572 110.7550 202.6491 122.2963
SR 44% 8% 92% 44% 28%
16
1st 0 101.171 86.3059 61.4505 95.3606
7th 18.057 115.255 98.5482 91.6652 107.4456
13th 37.004 120.640 101.4533 98.3280 112.8221
19th 46.277 126.891 104.9396 102.0676 115.7979
25th 96.174 159.818 111.9003 124.0614 121.4367
Avg 37.171 121.892 101.2093 96.7627 111.2191
SD 25.682 12.841 6.1686 12.0952 6.1818
SR 8% 0% 0% 0% 0%
17
1st 11.5331 87.447 99.0400 96.3383 108.1094
7th 20.3217 111.870 106.7286 101.0323 124.9650
13th 28.4788 116.812 113.6242 112.0854 129.6757
19th 41.5130 132.588 119.2813 116.7995 136.1688
25th 57.0901 148.346 135.5105 121.5371 153.8196
Avg 29.4416 120.604 114.0600 109.9146 130.4137
SD 12.3062 14.572 9.9679 8.5175 10.7680
SR 0% 0% 0% 0% 0%
18
1st 3.2816e-013 300.000 300.0000 300.0000 300.0000
7th 300.0000 300.0000 800.0000 800.0000 300.0000
13th 400.0000 724.071 800.0000 800.0000 300.0000
19th 500.0000 800.136 800.0000 800.0000 800.0000
25th 830.8944 842.140 900.8377 800.0000 800.0000
Avg 462.8617 586.495 719.3861 700.0000 440.0000
SD 218.2436 237.556 208.5161 204.1241 229.1288
SR 4% 0% 0% 0% 0%
19
1st 300.0000 300.000 300.0000 300.0000 300.0000
7th 336.21472 724.974 653.5664 300.0000 300.0000
13th 400.0000 835.172 800.0000 800.0000 300.0000
19th 500.0000 840.791 800.0000 800.0000 300.0000
OBL in the SDE algorithm 1323
123
Table 4 continued
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
25th 849.1442 972.051 930.7288 852.8697 800.0000
Avg 465.7913 725.042 704.9373 662.1148 400.0000
SD 181.8147 213.177 190.3959 230.7133 204.1241
SR 0% 0% 0% 0% 0%
20
1st 300.0000 300.000 300.0000 300.0000 300.0000
7th 374.7275 728.782 800.0000 300.0000 300.0000
13th 400.0000 832.504 800.0000 800.0000 300.0000
19th 800.0000 839.811 800.0000 800.0000 300.0000
25th 852.1974 841.154 907.0822 852.9271 800.0000
Avg 543.4026 723.011 713.0240 662.1171 400.0000
SD 224.0738 184.818 201.3396 230.7153 204.1241
SR 0% 0% 0% 0% 0%
21
1st 500.0000 862.119 300.0000 300.0000 500.0000
7th 500.0000 1,064.940 300.0000 500.0000 500.0000
13th 584.2689 1,080.770 500.0000 500.0000 500.0000
19th 678.5980 1,085.790 500.0000 500.0000 500.0000
25th 900.0000 1,092.470 800.0000 862.5199 800.0000
Avg 605.6206 1,068.658 464.0000 514.5008 512.0000
SD 104.2032 43.467 157.7973 173.9449 60.0000
SR 0% 0% 0% 0% 0%
22
1st 200.0000 749.223 300.0000 754.6395 749.4194
7th 500.0001 770.919 750.6537 757.3936 761.3741
13th 500.0094 781.991 752.4286 763.5070 763.6705
19th 500.0353 804.480 756.9808 766.7042 766.6464
25th 576.6202 884.563 800.0000 825.7207 768.1616
Avg 487.2779 796.623 734.9044 770.4906 763.2931
SD 64.9543 38.076 91.5229 22.1955 4.5850
SR 0% 0% 0% 0% 0%
23
1st 529.3801 559.468 559.4683 559.4683 559.4683
7th 529.3801 1,091.480 559.4683 559.4683 559.4683
13th 565.9457 1,102.440 559.4683 559.4683 559.4683
19th 611.2536 1,107.440 721.2327 721.2160 559.4683
25th 766.0608 1,125.880 970.5031 970.5031 970.5031
Avg 575.1675 1,032.814 664.0557 670.5235 605.2911
SD 77.5191 177.038 152.6608 151.4658 118.6263
SR 0% 0% 0% 0% 0%
24
1st 200.0000 399.931 200.0000 200.0000 200.0000
7th 200.0000 406.682 200.0000 200.0000 200.0000
13th 205.9986 408.827 200.0000 200.0000 200.0000
19th 248.8015 410.455 200.0000 200.0000 200.0000
25th 500.0000 412.857 200.0000 500.0000 200.0000
Avg 246.2228 408.229 200.0000 248.0000 200.0000
SD 84.4419 3.061 0 112.2497 0
1324 M. A. Ahandani, H. Alavi-Rad
123
Table 4 continued
F OB-SEOBDE Guided-DE SaDE jDE-2 jDE
SR 0% 0% 0% 0% 0%
25
1st 200.0000 200.002 370.9112 402.0977 401.7665
7th 200.0000 406.489 373.0349 402.3489 402.3340
13th 400.0000 408.979 375.4904 402.5183 402.5493
19th 500.0000 411.463 378.1761 402.6305 402.6452
25th 500.0000 415.074 381.5455 114.9815 402.9896
Avg 354.0000 368.090 375.8646 452.1622 402.4783
SD 147.7973 84.078 3.1453 166.3749 0.2858
SR 0% 0% 0% 0% 0%
Table 5 Wilcoxon test applied over the obtained results of Table 4 in terms of the best run (1st run), average of all runs (Avg) and success rate
(SR)
Comparison 1st run Avg SR
R? R- p value R? R- p value R? R- p value
Guided-DE–OB-SEOBDE 254 71 0.003 290.5 34.5 0.001 52.5 272.5 0.003
SaDE–OB-SEOBDE 208.5 116.5 0.278 185.5 139.5 0.503 133.5 191.5 0.553
jDE2–OB-SEOBDE 192 133 0.311 222 103 0.082 145 180 0.944
jDE–OB-SEOBDE 259.5 65.5 0.017 163.5 161.5 1.000 101.5 223.5 0.161
Table 6 A comparison among the OB-SEOBDE with Guided-DE algorithm and results reported in Ahandani et al. (2010) for 30 variable
numbers
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
1
1st 0 0 0 0 0.000
7th 0 0 0 0 0.000
13th 0 0 0 0 0.000
19th 0 0 0 0 0.137
25th 2.0195e-028 6.1781e-028 2.0195e-028 2.5244e-028 4.800
Avg 1.2117e-029 6.0182e-029 2.1457e-029 2.6506e-029 0.454
SD 4.4398e-029 9.9110e-029 6.3542e-029 7.9483e-029 1.078
SR 100% 100% 100% 100% –
2
1st 1.5146e-028 7.993e-029 0 5.1097e-029 0.119
7th 5.4905e-028 5.8376e-028 0 1.1967e-028 7.350
13th 7.2891e-028 8.141e-028 2.5244e-029 3.2585e-028 24.610
19th 9.9554e-028 1.0419e-027 4.4965e-029 6.9341e-028 64.701
25th 5.3011e-027 1.4414e-027 1.5146e-028 1.3373e-027 310.798
Avg 9.9009e-028 8.0404e-028 4.0469e-029 5.4863e-028 51.146
SD 1.0001e-027 3.6084e-028 4.9427e-029 3.2185e-028 67.704
SR 100% 100% 100% 100% 0%
3
1st 2.3512e?005 3.1476e?06 1.4467e?06 5.7461e?05 1.80e?06
7th 7.4967e?006 5.1924e?007 2.8219e?007 1.052e?007 2.57e?06
13th 2.4559e?007 2.4527e?008 6.5273e?007 2.3014e?007 3.38e?06
OBL in the SDE algorithm 1325
123
Table 6 continued
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
19th 8.110e?007 5.2713e?008 5.0448e?008 4.8214e?008 4.30e?06
25th 5.2278e?008 7.6017e?008 7.1908e?008 6.6334e?008 9.70e?06
Avg 6.0923e?007 5.1234e?008 3.1254e?008 1.5127e?008 3.81e?06
SD 8.0526e?007 3.7215e?008 5.7215e?008 6.2145e?008 1.94e?06
SR 0% 0% 0% 0% 0%
4
1st 5.418e-020 1.4361e-014 3.3937e-020 4.2468e-018 62.353
7th 3.7285e-016 1.4222e-010 1.3132e-018 6.6684e-017 113.070
13th 1.7778e-015 2.1623e-009 2.0363e-017 2.2164e-015 228.604
19th 9.5514e-015 7.7203e-009 8.1365e-016 4.5062e-014 662.856
25th 4.2682e-013 0.00010071 1.0174e-014 1.7014e-012 6,969.270
Avg 4.4738e-014 8.0174e-006 1.3049e-015 2.1602e-013 659.219
SD 1.0205e-013 3.1846e-005 3.1778e-015 5.3413e-013 1,347.355
SR 100% 96% 100% 100% 0%
5
1st 0.00020259 484.14 8,209.6 2.0049 942.059
7th 0.0051244 2,604.6 12,044 22.916 1,745.360
13th 0.01463 4,505.9 14,224 31.476 2,231.190
19th 0.17468 10,264 15,627 60.674 2,852.270
25th 13.175 13,722 17,497 259.48 4,235.990
Avg 0.67721 4,393.4 13,487 58.685 2,348.966
SD 2.6183 6,541.4 3,191.4 70.732 860.644
SR 0% 0% 0% 0% 0%
6
1st 9.1503e-026 1.3138e-026 0 0 0.840
7th 1.3898e-025 1.4176e-025 0 1.8923e-026 21.403
13th 2.9727e-025 2.2986e-025 0 2.1719e-025 31.882
19th 4.5041e-025 2.8887e-025 3.9781e-026 3.9866 73.881
25th 3.9866 3.9866 1.6749e-025 3.9866 163.846
Avg 0.79732 1.196 5.4807e-026 1.5946 56.180
SD 1.6275 1.9257 6.3257e-026 2.0587 45.049
SR 80% 76% 100% 68% 0%
7
1st 0 0.4988 12.51 1.9863e-008 1,346.130
7th 0 0.92976 21.587 1.5875e-006 1,518.540
13th 0 1.1029 41.93 1.6134e-005 1,626.350
19th 0 1.3849 54.576 0.0030908 1,703.580
25th 1.1797e-010 6.3554 175.42 0.88537 1,921.080
Avg 4.587e-010 1.6952 52.55 0.090865 1,629.623
SD 2.3127e-009 1.6891 56.487 0.090865 147.479
SR 100% 0% 0% 60% 0%
8
1st 20.6641 20.7102 20.6839 20.7236 20.791
7th 20.832 20.8330 20.827 20.894 20.980
13th 20.9007 20.9531 20.9138 20.9322 21.013
19th 20.937 20.9939 20.9698 20.9590 21.036
25th 20.967 21.1037 21.036 21.0214 21.130
Avg 20.8141 20.9360 20.829 20.9107 21.008
1326 M. A. Ahandani, H. Alavi-Rad
123
Table 6 continued
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
SD 0.067754 0.029065 0.10685 0.053022 0.063
SR 0% 0% 0% 0% 0%
9
1st 5.9698 27.849 6.1647 12.924 23.882
7th 9.9496 40.793 8.9546 21.889 44.795
13th 11.94 41.788 12.934 25.854 58.815
19th 16.914 49.748 16.914 29.824 64.692
25th 32.834 74.622 40.793 36.813 103.480
Avg 14.168 48.454 15.621 26.963 57.852
SD 6.6016 13.497 15.621 6.4897 17.005
SR 0% 0% 0% 0% 0%
10
1st 29.829 38.703 26.763 31.788 31.955
7th 50.743 83.526 46.108 61.687 62.688
13th 57.687 94.49 55.718 72.627 74.622
19th 64.657 102.45 65.167 86.536 83.577
25th 83.576 113.42 73.627 119.39 118.517
Avg 57.574 94.931 58.708 74.753 73.577
SD 14.257 18.823 11.835 22.981 21.372
SR 0% 0% 0% 0% 0%
11
1st 18.121 14.1608 20.756 11.692 14.364
7th 24.056 18.083 25.108 16.201 15.946
13th 26.121 20.325 27.547 19.591 21.013
19th 27.765 22.119 30.804 21.546 24.967
25th 31.050 26.725 32.211 23.813 35.773
Avg 25.903 20.927 27.852 18.986 21.126
SD 2.9576 6.49125 2.2754 3.0596 5.048
SR 0% 0% 0% 0% 0%
12
1st 6.4445e-016 14,845 5.7541e-019 1.3586e-007 52.735
7th 6.7361e-010 27,744 1.2288e-015 0.0042246 989.109
13th 2.0288e-006 35,429 5.4283e-015 0.1492 1,580.930
19th 0.0071141 87,639 1.4367e-014 29.868 2,348.470
25th 25.509 6.3404e?005 2.3969e-014 50,666 1.12e?04
Avg 1.3981 1.473e?005 9.1191e-015 5,695.6 2441.885
SD 8.9530 2.2534e?005 8.1169e-015 15,921 2,605.095
SR 80% 0% 100% 32% 0%
13
1st 0.71835 1.1346 1.2820 0.80427 2.098
7th 1.2231 1.9449 2.2275 1.035 3.622
13th 1.3054 2.5194 2.8052 1.5879 3.966
19th 1.5151 2.764 3.1576 1.8976 4.356
25th 2.7829 3.509 3.6905 2.6498 8.442
Avg 1.4279 2.6047 2.8583 1.5135 4.179
SD 0.42793 0.65802 0.61543 0.57842 1.342
SR 0% 0% 0% 0% 0%
OBL in the SDE algorithm 1327
123
Table 6 continued
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
14
1st 12.535 11.874 10.788 11.487 11.799
7th 12.769 12.174 11.926 12.126 12.895
13th 12.896 12.231 12.243 12.411 13.098
19th 13.061 12.317 12.581 12.686 13.390
25th 13.123 12.494 13.038 12.883 13.720
Avg 12.928 12.267 12.155 12.403 13.080
SD 0.1935 0.13036 0.543 0.29501 0.393
SR 0% 0% 0% 0% 0%
15
1st 202.68 202.03 205.03 200.00 204.393
7th 218.73 274.13 231.77 210.63 212.871
13th 230.000 307.63 306.74 218.5 230.674
19th 400.80 400 354.85 225.28 252.062
25th 500.000 401.17 400.00 401.62 401.536
Avg 315.31 321.93 320.11 227.14 258.444
SD 76.81 74.315 76.21 71.471 67.054
SR 0% 0% 0% 0% 0%
16
1st 31.3026 51.5240 35.4910 84.1800 73.644
7th 55.6228 77.7260 60.8648 131.108 103.422
13th 84.6663 92.4620 87.0665 151.7871 131.223
19th 142.2772 400.0000 168.5320 168.5100 215.766
25th 400.0000 400.7220 400.2981 186.4911 401.614
Avg 114.54 236.2554 122.6239 145.9877 191.250
SD 90.033 158.0505 117.8998 31.7808 121.847
SR 0% 0% 0% 0% 0%
17
1st 42.4480 65.573 58.9740 60.4920 79.762
7th 58.0234 91.21 88.0527 77.1940 102.751
13th 88.2147 117.092 114.5402 90.3100 127.307
19th 95.2247 152.173 128.2126 102.9162 159.434
25th 143.142 489.93 165.7200 193.0700 437.790
Avg 80.9177 169.58 109.9214 93.1725 144.814
SD 30.9687 117.19 35.4737 45.598 70.201
SR 0% 0% 0% 0% 0%
18
1st 300.0000 767.98 620.47 659.78 842.076
7th 571.0487 813.54 655.72 743.4 859.492
13th 721.6254 823.64 708.63 809.49 863.223
19th 815.4869 851.01 814.05 821.25 865.722
25th 869.6934 910.16 836.76 911.07 871.441
Avg 701.1381 845.91 747.58 809.61 862.275
SD 150.4512 48.489 82.136 94.38 6.058
SR 0% 0% 0% 0% 0%
19
1st 604.1339 735.1523 716.0900 758.7012 852.813
7th 772.4365 803.5175 804.5631 807.8211 858.968
1328 M. A. Ahandani, H. Alavi-Rad
123
Table 6 continued
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
13th 806.2962 815.6446 810.5315 811.6306 863.000
19th 811.9980 820.0809 814.4906 813.3052 864.898
25th 912.0979 875.1917 831.1900 817.0719 867.107
Avg 780.5621 809.6267 799.1730 810.6067 862.108
SD 65.5099 40.0791 36.0907 12.2955 4.009
SR 0% 0% 0% 0% 0%
20
1st 300.0000 756.3110 651.1036 742.6102 856.228
7th 574.1389 808.6283 781.7015 808.8116 860.962
13th 663.9974 814.1081 809.8928 811.5144 863.679
19th 814.1187 818.1978 813.5106 813.5534 865.482
25th 838.8724 827.7574 818.1264 819.1095 870.040
Avg 649.8701 812.970 773.4084 810.7644 863.248
SD 131.7891 24.3667 81.5913 31.8369 3.601
SR 0% 0% 0% 0% 0%
21
1st 559.38032 605.2532 589.1410 565.5196 863.243
7th 587.3724 642.5066 620.6930 602.9371 865.649
13th 628.3850 685.1409 635.1909 619.5048 867.489
19th 654.3850 708.7728 733.2443 642.1721 871.163
25th 702.4777 720.5542 773.4389 686.4944 883.513
Avg 619.5612 673.7939 672.7416 621.1317 868.876
SD 42.3437 41.1572 71.0482 40.8959 4.482
SR 0% 0% 0% 0% 0%
22
1st 500.0000 500.0000 500.0000 500.0000 550.801
7th 500.0000 500.0812 500.0000 500.0000 557.383
13th 500.0081 500.1611 500.0072 500.0094 560.569
19th 500.0912 501.2285 500.0728 500.0866 564.529
25th 500.2210 502.5235 500.8079 500.1660 568.612
Avg 500.08443 501.1073 500.2055 500.0658 560.973
SD 0.1108 1.0678 0.4016 0.1050 4.617
SR 0% 0% 0% 0% 0%
23
1st 553.5043 608.7612 539.3434 567.5573 867.525
7th 587.1616 621.5201 557.5559 618.8724 872.371
13th 613.3389 639.0388 589.5459 635.5267 875.118
19th 700.2539 689.5573 626.0184 665.7802 877.092
25th 742.4490 773.0118 886.8944 691.5946 883.600
Avg 630.9312 672.8516 621.5295 643.3877 874.996
SD 56.4911 61.8820 109.8517 44.1934 3.552
SR 0% 0% 0% 0% 0%
24
1st 200.1208 202.9346 200.0057 201.6099 242.567
7th 203.1023 212.5192 201.2007 208.0116 248.551
13th 207.8814 218.7320 202.2416 221.4537 250.417
19th 213.4412 224.6322 204.1864 234.6506 252.526
25th 400.0000 466.7289 206.8495 339.2744 256.223
OBL in the SDE algorithm 1329
123
Table 6 continued
F OB-SEOBDE BDE2 SDE2 SBDE2 Guided-DE
Avg 229.8914 279.4537 202.5502 247.3866 250.163
SD 65.0845 124.9821 2.2234 68.5517 3.222
SR 0% 0% 0% 0% 0%
25
1st 201.0048 205.0939 202.1478 201.0441 250.691
7th 205.3717 213.2079 208.296 205.2231 251.846
13th 212.0711 219.9711 217.1156 212.4859 252.521
19th 225.3719 250.8567 236.4612 230.9312 253.214
25th 327.9073 350.1024 365.7446 325.1075 256.565
Avg 222.1447 249.4860 239.55304 227.84348 252.770
SD 52.8059 57.8298 67.5495 54.1050 1.361
SR 0% 0% 0% 0% 0%
Table 7 Wilcoxon test applied over the obtained results of Table 6 in terms of the best run (1st run), average of all runs (Avg) and success rate
(SR) for 30 variables
Comparison 1st run Avg SR
R? R- p value R? R- p value R? R- p value
Guided-DE–OB-SEOBDE 311.5 13.5 0.000 283 42 0.001 95 230 0.026
BDE2–OB-SEOBDE 285.5 39.5 0.000 307 18 0.000 115.5 209.5 0.066
SDE2–OB-SEOBDE 252.5 72.5 0.008 265 60 0.004 173.5 151.5 1.000
SBDE2–OB-SEOBDE 173.5 51.5 0.003 278 47 0.002 126.5 198.5 0.109
Table 8 A comparison among the OB-SEOBDE with results reported in Ahandani et al. (2010) for 50 variable numbers
F OB-SEOBDE BDE2 SDE2 SBDE2
1
1st 0 2.5244e–028 0 1.2622e–029
7th 2.051e–028 4.6701e–028 0 5.0487e–029
13th 4.039e–028 7.6993e–028 5.0487e–029 2.5244e–028
19th 6.5633e–028 1.1612e–027 3.2622e–029 4.039e–028
25th 1.4641e–027 2.7894e–027 2.5244e–028 1.1107e–027
Avg 4.5868e–028 8.7582e–028 8.709e–029 3.1819e–028
Std 3.5398e–028 5.4428e–028 1.339e–028 3.1159e–028
SR 100% 100% 100% 100%
2
1st 3.3271e–027 4.6014e–027 3.1381e–028 3.1125e–027
7th 6.3109e–027 7.2189e–027 1.3251e–027 4.1232e–027
13th 9.8306e–027 1.0502e–026 1.7262e–027 5.8467e–027
19th 1.5614e–026 1.4453e–026 2.1481e–027 7.2694e–027
25th 2.0994e–025 3.7084e–026 4.4176e–027 1.1879e–026
Avg 2.1933e–026 1.3046e–026 1.8950e–027 6.0025e–027
Std 4.1779e–026 9.0369e–027 9.6816e–028 2.7723e–027
SR 100% 100% 100% 100%
3
1st 1.2498e?006 2.1424e?007 5.2104e?06 2.6318e?06
7th 3.3175e?007 7.3142e?007 3.05841e?007 2.1748e?007
1330 M. A. Ahandani, H. Alavi-Rad
123
Table 8 continued
F OB-SEOBDE BDE2 SDE2 SBDE2
13th 5.7749e?008 1.8254e?009 4.9214e?008 3.5228?009
19th 8.4187e?008 4.9215e?009 4.1827e?009 5.6117e?009
25th 3.1038e?009 8.6217e?009 8.3207e?009 8.319e?009
Avg 6.9917e?008 4.0253e?009 4.2825e?009 3.6128e?009
Std 5.5550e?008 4.8719e?009 4.9930e?009 5.934e?009
SR 0% 0% 0% 0%
4
1st 0.007192 57.175 1.146e-007 0.17815
7th 0.074442 97.81 0.00013438 2.3683
13th 0.40182 351.12 0.0010104 35.03
19th 1.813 558.49 0.0046718 366.78
25th 6.2445 1163.7 0.024412 508.72
Avg 1.151 435.08 0.0044463 3458.2
Std 1.6067 377.84 0.0073684 1080.6
SR 0% 0% 12% 0%
5
1st 154.4184 18100 16429 8795
7th 958.2955 21499 20194 11076
13th 1919.3028 23308 23010 13521
19th 4951.9688 25031 24783 14341
25th 15967.7129 32206 28329 19809
Avg 3492.3308 23469 23549 14653
Std 3845.7337 4683.1 3296.7 3062.1
SR 0% 0% 0% 0%
6
1st 2.2658e–025 1.9942e–025 1.639e–025 3.8133e–025
7th 1.2295e–024 6.4291e–025 3.1729e–025 6.52e–025
13th 2.873e–024 2.3965e–024 4.9086e–025 1.4118e–024
19th 8.4827e–024 3.9592e–023 7.9845e–025 3.9866
25th 3.9866 3.9866 3.9866 3.9866
Avg 0.79732 1.196 0.79732 2.1011
Std 1.6275 1.9257 1.6809 1.9933
SR 80% 76% 84% 60%
7
1st 3.7157e–005 91.461 269.417 0.018
7th 0.0321 121.017 365.661 4.960
13th 0.0817 176.926 5,655.723 6.523
19th 0.3455 182.782 629.715 9.229
25th 0.4806 227.048 939.808 16.173
Avg 0.1967 169.120 600.152 9.017
Std 0.1854 70.531 218.26 6.037
SR 20% 0% 0% 0%
8
1st 20.7014 21.009 20.829 20.912
7th 20.9103 21.106 21.014 21.088
13th 21.1160 21.121 21.109 21.119
19th 21.1254 21.172 21.123 21.168
25th 21.2014 21.221 21.189 21.207
OBL in the SDE algorithm 1331
123
Table 8 continued
F OB-SEOBDE BDE2 SDE2 SBDE2
Avg 21.0677 21.145 21.120 21.164
Std 0.1388 0.06077 0.0215 0.0314
SR 0% 0% 0% 0%
9
1st 34.8235 92.31 31.788 66.561
7th 52.7328 112.41 63.677 94.521
13th 58.7025 132.33 84.820 108.48
19th 69.6470 170.11 103.48 119.05
25th 95.5159 194.02 132.34 135.36
Avg 60.4934 152.22 79.099 107.06
Std 13.9501 31.811 31.936 17.748
SR 0% 0% 0% 0%
10
1st 82.5563 174.02 91.41 142.18
7th 111.4301 219.85 114.47 166.16
13th 134.3191 232.82 158.2 200.98
19th 152.2081 253.71 175.06 227.82
25th 201.9759 293.61 197 252.18
Avg 135.6279 236.2 154.81 198.79
Std 24.9865 39.934 42.352 39.128
SR 0% 0% 0% 0%
11
1st 34.9124 31.1350 47.6459 36.5600
7th 43.0279 41.6300 53.2561 42.6851
13th 50.0627 46.1078 58.8258 47.1627
19th 56.9028 49.6871 60.5207 52.7034
25th 60.2904 53.0820 64.2020 56.6207
Avg 50.6618 45.1217 59.5579 45.3162
Std 6.5476 9.7844 7.4571 9.4852
SR 0% 0% 0% 0%
12
1st 0.031176 1.5570e?005 1.1250e–006 40,171
7th 12.0872 4.0184e?005 0.0058 1.0525e?005
13th 328.1047 6.5205e?005 0.0375 2.1104e?005
19th 10,960 7.1822e?005 0.4749 2.7349e?005
25th 81,642 8.7402e?005 4.1647 3.8604e?005
Avg 20,311 5.8021e?005 0.49284 1.992e?005
Std 62,072 1.3941e?005 1.14172 99,800
SR 0% 0% 40% 0%
13
1st 2.1013 7.722 1.5037 4.3988
7th 3.4125 11.241 3.1887 5.6706
13th 4.3003 15.256 4.0646 6.5115
19th 4.7357 18.825 4.9627 7.2551
25th 6.7217 20.586 7.7219 8.129
Avg 4.1417 15.3472 4.7683 6.573
Std 1.0645 4.011 1.8919 1.2249
SR 0% 0% 0% 0%
1332 M. A. Ahandani, H. Alavi-Rad
123
Table 8 continued
F OB-SEOBDE BDE2 SDE2 SBDE2
14
1st 21.745 21.499 21.402 21.814
7th 22.267 22.157 21.637 22.139
13th 22.409 22.324 21.742 22.307
19th 22.611 22.521 21.824 22.546
25th 22.758 22.702 22.108 22.887
Avg 22.428 22.332 21.700 22.471
Std 0.2429 0.921 0.125 0.501
SR 0% 0% 0% 0%
15
1st 216.661 248.091 191.532 203.535
7th 265.227 314.992 236.091 251.199
13th 400.000 375.012 312.752 319.897
19th 400.000 404.494 375.416 371.191
25th 500.000 471.417 500.000 410.525
Avg 355.451 359.757 314.12 313.262
Std 73.702 93.039 97.888 74.860
SR 0% 0% 0% 0%
16
1st 47.6572 89.2525 85.7265 79.1045
7th 79.2944 120.1371 141.0857 109.5117
13th 123.5570 176.3244 213.1325 142.9316
19th 400.000 205.2484 264.5471 185.1913
25th 400.000 501.0105 500.0000 247.1839
Avg 168.2884 210.3252 275.0531 175.5309
Std 154.9750 183.2781 190.3011 160.4336
SR 0% 0% 0% 0%
17
1st 64.2291 112.2511 99.9632 85.0741
7th 79.8824 202.3107 181.1128 128.7871
13th 109.6147 235.8372 218.4971 155.4413
19th 236.6845 264.4430 251.4301 211.5576
25th 412.0214 501.5862 524.3217 420.3908
Avg 174.0152 253.2856 227.0086 186.8582
Std 134.7250 146.6668 161.8550 132.6859
SR 0% 0% 0% 0%
18
1st 775.3187 828.8015 809.5542 816.8407
7th 820.3135 840.0714 818.3673 826.1269
13th 830.0716 857.6108 828.1547 845.5234
19th 839.6455 862.8715 835.1807 852.1402
25th 885.0004 982.1534 848.1819 871.8631
Avg 825.9200 872.9411 830.0931 842.3385
Std 15.6037 75.1962 17.1256 27.5154
SR 0% 0% 0% 0%
19
1st 759.7364 824.8937 807.0158 815.5812
7th 815.849 839.1949 819.7194 824.5682
OBL in the SDE algorithm 1333
123
Table 8 continued
F OB-SEOBDE BDE2 SDE2 SBDE2
13th 827.3298 846.8178 826.4296 831.5799
19th 837.4564 858.7525 838.1690 838.01475
25th 923.4938 876.2723 849.5272 851.2634
Avg 827.8923 852.1862 829.7722 835.5764
Std 37.1305 19.9362 15.4446 13.6032
SR 0% 0% 0% 0%
20
1st 691.2085 822.5528 813.2102 814.1063
7th 818.4284 833.4069 824.5445 829.2518
13th 827.4300 842.01824 836.8216 842.1520
19th 833.2078 851.5638 849.5234 849.1725
25th 857.1437 866.6390 896.0999 1,055.0000
Avg 823.8865 842.8662 846.2849 859.4527
Std 30.9193 17.0701 26.4349 46.0739
SR 0% 0% 0% 0%
21
1st 534.7437 559.2351 572.9342 568.1000
7th 560.2876 568.2886 591.2531 589.1209
13th 582.3520 577.0805 608.3084 605.0933
19th 610.3835 594.1862 615.2314 624.6041
25th 651.8365 614.0174 622.4916 645.6796
Avg 596.0950 583.1976 600.2414 607.3692
Std 28.5720 23.3076 19.0997 34.7668
SR 0% 0% 0% 0%
22
1st 500.0000 500.0000 500.0000 500.0000
7th 500.0000 500.1107 500.0000 500.0000
13th 500.0140 500.4395 500.0372 500.0082
19th 500.0771 500.8737 500.1827 500.0511
25th 500.7239 501.2271 500.6119 500.1423
Avg 500.0878 500.5810 500.1418 500.0312
Std 0.2903 0.8370 0.3241 0.0585
SR 0% 0% 0% 0%
23
1st 554.1794 571.9187 565.0694 558.5649
7th 571.9907 604.2135 608.0819 582.3128
13th 593.5107 619.9710 624.2178 595.6842
19th 626.3521 625.5682 642.4327 603.3455
25th 657.0389 645.7019 662.6954 611.4208
Avg 601.5780 615.0344 623.0161 578.7269
Std 31.0642 42.2450 36.4146 19.3516
SR 0% 0% 0% 0%
24
1st 205.3700 211.5406 207.8617 210.1275
7th 215.0700 234.1469 223.1672 229.8107
13th 236.6101 248.4800 237.1524 256.49
19th 249.1028 310.4365 265.1287 292.8296
25th 268.2130 1,245.0724 1,158.1048 929.67542
1334 M. A. Ahandani, H. Alavi-Rad
123
7 Conclusions and future works
The DE is one simple and effective EA for global opti-
mization. It has only three control parameters to be tuned.
The DE employs simple differential operator to create new
candidate solutions and one-to-one competition scheme to
greedily select new candidates. Also, it has a simple and
compact structure that makes its implementation easy.
Besides the aforementioned benefits of the DE, stagnation
or premature convergence, not being able to accurately
zoom to optimal solution, the limited number and diversity
of search moves, greedy criterion as acceptance strategy,
poor performance in noisy environment, requiring multiple
runs for tuning parameters and the best control parameter
being problem dependent are some challenging drawbacks.
In an attempt on one hand to accelerate the classic DE and
on the other hand to compensate the limited amount of
number and diversity of search moves, we enhanced the
OBL strategy and SDE in this work.
In the SDE, the population is divided into several
memeplexes and each memeplex is improved by the DE.
The OBL by comparing the fitness of an individual to its
opposite and retaining the fitter one in the population
accelerates the search process. The emphasis of this paper
was to demonstrate how the OBL strategy can improve the
performance of SDE. Four versions of the DE algorithm
were proposed. All algorithms similarly used the opposi-
tion-based population initialization to achieve fitter initial
individuals and their difference was in applying opposi-
tion-based generation jumping.
Experiments were performed on 25 benchmark func-
tions designed for the special session on real-parameter
optimization of CEC2005. A non-parametric analysis over
the obtained results by using of the Wilcoxon signed-ranks
test showed that the proposed algorithms, despite their
simplicity, had a remarkable performance over a wide
and various set of test problems. The fourth version of
the proposed DE had a significant difference compared to
the SDE in terms of all three considered aspects. Also, the
proposed algorithms obtained some successful perfor-
mance on a set of functions named to unsolved functions
for the first time. In a later part of the comparative
experiments, performance comparisons of the proposed
algorithm with some modern DE algorithms reported in the
literature confirmed a significantly better performance of
our proposed algorithm, especially on high-dimensional
functions.
However, our proposed algorithms obtained a better or
at least comparable performance than other proposed
approaches, but those have many parameters for setting
Table 8 continued
F OB-SEOBDE BDE2 SDE2 SBDE2
Avg 235.7661 564.9503 454.38 490.0069
Std 18.8645 482.5603 499.2357 440.3765
SR 0% 0% 0% 0%
25
1st 204.9172 219.1167 210.2215 209.4638
7th 217.4204 237.0632 221.0362 235.1028
13th 234.6802 247.2539 244.6829 251.3718
19th 254.3583 344.1182 282.3716 275.1924
25th 1,304.4088 1,463.2813 1,342.1932 1,241.6122
Avg 343.9790 616.7214 515.2516 508.3351
Std 338.1631 562.3119 527.1362 501.4438
SR 0% 0% 0% 0%
Table 9 Wilcoxon test applied over the obtained results of Table 8 in terms of the best run (1st run), average of all runs (Avg) and success rate
(SR) for 50 variables
Comparison 1st run Avg SR
R? R- p value R? R- p value R? R- p value
BDE2–OB-SEOBDE 308.5 16.5 0.000 301 24 0.000 138 187 0.180
SDE2–OB-SEOBDE 261.5 63.5 0.005 263.5 61.5 0.007 185.5 139.5 0.465
SBDE2–OB-SEOBDE 307.5 17.5 0.000 278 47 0.002 138 187 0.157
OBL in the SDE algorithm 1335
123
which need several pre-runs to obtain the best composition.
These parameters are due to the original DE, partitioning
concept and opposition-based strategy. For the future work,
it might be interesting to employ adaptive or self-adaptive
tuning of parameters following the ideas presented in Liu
and Lampinen (2005), Brest et al. (2006) and Qin et al.
(2009). We considered a fixed value for MaxSize parame-
ter. It might also be interesting to limit the value of Max-
Size when the algorithm goes forward toward its final
iterations. In addition, the performance of the opposition-
based SDE variants may also be improved by employing
other mutation and crossover strategies and also to exper-
iment different population sizes.
References
Ahandani MA, Shirjoposht NP, Banimahd R (2010) Three modified
versions of differential evolution algorithm for continuous
optimization. Soft Comput 15:803–830
Rashid M, Baig, AR (2010) Improved opposition-based PSO for
feedforward neural network international conference on training.
Information Science and Applications (ICISA2010), Seoul,
pp 1–6
Balamurugan R, Subramanian S (2009) Emission-constrained
dynamic economic dispatch using opposition-based self-adaptive
differential evolution algorithm. Int Energy J 10:267–277
Becker W, Yu X, Tu J (2005) EvLib: a parameterless self-adaptive
real-valued optimisation library. In: The 2005 IEEE congress on
evolutionary computation CEC2005
Bhattacharya A, Chattopadhyay P (2010) Solution of economic power
dispatch problems using oppositional biogeography-based opti-
mization. Electr Power Compon Syst 38:1139–1160
Boskovis B, Brest J, Zamuda A, Greiner S, Zumer V (2011) History
mechanism supported differential evolution for chess evaluation
function tuning. Soft Comput 15:667–682
Brest J, Greiner S, Boskovic B, Mernik M, Zumer V (2006) Self-
adapting control parameters in differential evolution: a compar-
ative study on numerical benchmark problems. IEEE Trans Evol
Comput 10:646–657
Brest J, Boskovic B, Greiner S, Zumer V, Maucec MS (2007)
Performance comparison of self-adaptive and adaptive differen-
tial evolution algorithms. Soft Comput 11:617–629
Bui LT, Shan Y, Qi F, Abbass HA (2005) Comparing two versions of
differential evolution in real parameter optimization. In: The
2005 IEEE congress on evolutionary computation CEC2005
Caponio A, Neri F, Tirronen V (2009) Super-fit control adaptation in
memetic differential evolution frameworks. Soft Comput 13:811–
831
Dorigo JM, Gambardella LM (1997) Ant colony system: a cooper-
ative learning approach to the traveling salesman problem. IEEE
Trans Evol Comput 1(1):53–66
Ergezer M, Simon D, Du D (2009) Oppositional biogeography-based
optimization. In: IEEE Conference on Systems, Man, and
Cybernetics. San Antonio, Texas, pp 1035–1040
Eusuff MM, Lansey KE (2003) Optimization of water distribution
network design using the shuffled frog leaping algorithm.
J Water Res Plane Manag 129:210–225
Feoktistov V (2006) Differential evolution: in search of solutions. In:
Optimization and its applications, vol 5. Springer, New York
Garcia S, Molina D, Lozano M, Herrera F (2009) A study on the use
of non-parametric tests for analyzing the evolutionary algo-
rithms’ behaviour: a case study on the CEC’2005 special session
on real parameter optimization. J Heuristics 15:617–644
Han L, He X (2007) A novel opposition-based particle swarm
optimization for noisy problems. In: Proceedings of the third
international conference on natural computation, vol 3. IEEE
press, pp 624–629
Hansen N (2005) Compilation of results on the CEC benchmark
function set. http://www.ntu.edu.sg/home/epnsugan/index_files/
CEC05/compareresults.pdf
Holland J (1975) Adaptation in natural and artificial systems. The
University of Michigan Press, Ann Arbor
Kennedy J, Eberhart RC (1995) Particle swarm optimization. In:
Proceedings of the IEEE International Conference on Neural
Networks, pp 1942–1948
Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by
simulate annealing. Science 220:671–680
Kofjac D, Kljajic M (2008) Application of genetic algorithms and
visual simulation in a real-case production optimization.
WSEAS Trans Syst Control 3:992–1001
Liu J, Lampinen J (2005) A fuzzy adaptive differential evolution
algorithm. Soft Comput 9:448–469
Liu B, Wang L, Jin Y-H, Huang D-X, Tang F (2007) Control and
synchronization of chaotic systems by differential evolution
algorithm. Chaos Soliton Fract 34:412–419
Malisia AR, Tizhoosh HR (2007) Applying opposition-based ideas to
the ant colony system. In: Proceedings of the IEEE symposium
on foundations of computational intelligence (SIS 2007). Hono-
lulu, Hawaii, pp 182–189
Moscato P (1989) On evolution, search, optimization, genetic
algorithms and martial arts: toward memetic algorithm. Techni-
cal Report Caltech Concurrent Computation Program: Report 26,
California Institute of Technology
Neri F, Tirronen V (2009) Scale factor local search in differential
evolution. Memet Comp 1:153–171
Neri F, Tirronen V (2010) Recent advances in differential evolution: a
review and experimental analysis. Artif Intell Rev 33:61–106
Omran MGH (2009) Using opposition-based learning with particle
swarm optimization and barebones differential evolution. In:
Lazinica A (ed) Particle swarm optimization. In Tech, pp 373–
384
Perez-Bellido AM, Salcedo-Sanz S, Ortiz-Garcia EG, Portilla-
Figueras JA, Lopez-Ferreras F (2008) A comparison of memetic
algorithms for the spread spectrum radar polyphase codes design
problem. Eng Appl Artif Intel 21:1233–1238
Plagianakos VP, Tasoulis DK, Vrahatis MN (2008) A review of major
application areas of differential evolution. In: Chakraborty UK
(ed) Advances in differential evolution of studies in computa-
tional intelligence, vol 143. Springer, Berlin, pp 197–238
Qin AK, Suganthan PN (2005) Self-adaptive differential evolution
algorithm for numerical optimization. In: The 2005 IEEE
congress on evolutionary computation CEC2005, vol 13,
pp 1785–1791
Qin AK, Huang VL, Suganthan PN (2009) Differential evolution
algorithm with strategy adaptation for global numerical optimi-
zation. IEEE Trans Evol Comput 13:398–417
Rahnamayan S, Wang GG (2008) Investigating in scalability of
opposition-based differential evolution. WSEAS Trans Comput
7:1792–1804
Rahnamayan S, Tizhoosh HR, Salama MMA (2006) Opposition
versus randomness in soft computing techniques. Elsevier J Appl
Soft Comput 8(906):918
Rahnamayan S, Tizhoosh HR, Salama MMA (2008) Opposition-
based differential evolution. IEEE Trans Evol Comput 12:64–79
1336 M. A. Ahandani, H. Alavi-Rad
123
Shokri M, Tizhoosh HR, Kamel M (2006) Opposition-based Q(k)
algorithm. In: Proceedings of IEEE World Congress on Com-
putational Intelligence. Vancouver, BC, Canada, pp 646–653
Storn R, Price K (1997) Differential evolution—a simple and efficient
heuristic for global optimization over continuous spaces.
J Global Optim 11:341–359
Subudhi B, Jena D (2009) Nonlinear system identification using
opposition based Learning differential evolution and neural
network techniques. IEEE J Intell Cybern Syst 5:1–13
Suganthan PN, Hansen N, Liang JJ, Deb K, Chen Y-P, Auger A,
Tiwari S (2005) Problem definitions and evaluation criteria for
the CEC 2005 special session on real-parameter optimization.
Technical Report Report #2005005, Nanyang Technological
University, Singapore and IIT Kanpur, India. http://www.ntu.
edu.sg/home/EPNSugan/
Teng NS, Teo J, Hijazi MHA (2009) Self-adaptive population sizing
for a tune-free differential evolution. Soft Comput 13:709–724
Tizhoosh HR (2005) Opposition-based learning: a new scheme for
machine intelligence. In: Proc Int Conf Comput Intell Modeling
Control and Autom, vol 1. Vienna, Austria, pp 695–701
Tizhoosh HR (2006) Opposition-based reinforcement learning. J Adv
Comput Intell Intell Inf 10:578–585
Ventresca M, Tizhoosh HR (2006) Improving the convergence of
backpropagation by opposite transfer functions. In: Proceedings
of IEEE World Congress and Computational Intelligence.
Vancouver, BC, Canada, pp 9527–9534
Ventresca M, Tizhoosh HR (2007) Simulated annealing with opposite
neighbors. In: Proceedings of the IEEE Symposium on Founda-
tions of Computational Intelligence (SIS 2007). Honolulu,
Hawaii, pp 186–192
Wang H, Liu Y, Zeng S, Li C (2007) Opposition-based particle
swarm algorithm with Cauchy mutation. In: Proceedings of the
IEEE Congress on Evolutionary Computation, pp 4750–4756
OBL in the SDE algorithm 1337
123