IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO.
3, JUNE 2013
A Novel Artificial Bee Colony Algorithm Based on
Modified Search Equation and Orthogonal Learning
Wei-feng Gao, San-yang Liu, and Ling-ling Huang
Abstract—The artificial bee colony (ABC) algorithm is a rel-
atively new optimization technique which has been shown to be
competitive to other population-based algorithms. However, ABC
has an insufficiency regarding its solution search equation, which
is good at exploration but poor at exploitation. To address this
concerning issue, we first propose an improved ABC method
called as CABC where a modified search equation is applied to
generate a candidate solution to improve the search ability of ABC.
Furthermore, we use the orthogonal experimental design (OED)
to form an orthogonal learning (OL) strategy for variant ABCs
to discover more useful information from the search experiences.
Owing to OED’s good character of sampling a small number of
well representative combinations for testing, the OL strategy can
construct a more promising and efficient candidate solution. In
this paper, the OL strategy is applied to three versions of ABC, i.e.,
the standard ABC, global-best-guided ABC (GABC), and CABC,
which yields OABC, OGABC, and OCABC, respectively. The ex-
perimental results on a set of 22 benchmark functions demonstrate
the effectiveness and efficiency of the modified search equation and
the OL strategy. The comparisons with some other ABCs and sev-
eral state-of-the-art algorithms show that the proposed algorithms
significantly improve the performance of ABC. Moreover, OCABC
offers the highest solution quality, fastest global convergence, and
strongest robustness among all the contenders on almost all the
test functions.
Index Terms—Artificial bee colony (ABC) algorithm, orthogo-
nal experimental design (OED), orthogonal learning (OL), search
equation.
I. I NTRODUCTION
O PTIMIZATION techniques play a very important role
in engineering design, operational research, information
science, and related areas. They can be divided into two cate-
gories: derivative-based methods and derivative-free methods.
For a variety of reasons, there have always been many real-
world problems where derivatives are unavailable or unreliable.
Thus, as an important branch of derivative-free methods, evo-
lutionary algorithms (EAs) have shown considerable success
in solving optimization problems characterized as nonconvex,
Manuscript received December 29, 2011; revised April 28, 2012 and
July 19, 2012; accepted September 27, 2012. Date of publication October 18,
2012; date of current version May 10, 2013. This work was supported in part
by the National Nature Science Foundation of China under Grants 60974082
and 11126287 and in part by the Fundamental Research Funds for the Central
Universities under Grant K5051270002. This paper was recommended by
Associate Editor H. Takagi.
W. Gao and S. Liu are with Xidian University, Xi’an 710071, China (e-mail:
gaoweifeng2004@126.com; liusanyang@126.com).
L. Huang is with the China University of Petroleum, Qingdao 266580, China
(e-mail: huangzxp@yahoo.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSMCB.2012.2222373
2168-2267/$31.00 © 2012 IEEE
1011
discontinuous, nondifferentiable, and so on and attracted more
and more attention in recent years. The most prominent EAs
proposed in the literatures are genetic algorithm (GA) [1],
particle swarm optimization (PSO) [2], differential evolution
(DE) [3], ant colony optimization [4], biogeography-based
optimization [5], artificial bee colony (ABC) algorithm [6], and
so on.
In this paper, we concentrate on ABC, developed by
Karaboga [6] in 2005 based on simulating the foraging behavior
of the honeybee swarm. Numerical comparisons demonstrated
that the performance of ABC is competitive to that of other
population-based algorithms with an advantage of employing
fewer control parameters [7]–[10]. Due to its simplicity and
ease of implementation, ABC has captured much attention and
has been applied to solve many practical optimization problems
[11]–[16] since its invention.
However, similar to other EAs, ABC also faces up to the
poor convergence. The reasons are as follows. It is well known
that both exploration and exploitation are necessary for EAs.
In practice, the two aspects contradict each other. In order
to achieve good performances on problem optimizations, they
should be well balanced. While the solution search equation
of ABC, which is used to generate new candidate solutions
based on the information of previous solutions, is good at
exploration, it is poor at exploitation [17], which results in the
poor convergence.
To improve the performance of ABC, one active research
trend is to study its search equation. Until now, various search
equations have been suggested, such as [17]–[23] and so on.
The most representative one is that inspired by PSO. Zhu
and Kwong [17] proposed a global best (gbest)-guided ABC
(GABC) by incorporating the information of the gbest solution
into the solution search equation of ABC to improve the ex-
ploitation. However, like PSO, it may cause an “oscillation”
phenomenon because the guidance of the two terms in the
search equation may be in opposite directions. This is ineffi-
cient to the search ability of the algorithm and delays the con-
vergence speed. However, we note that a well-designed search
equation is usually beneficial to enhance the performance of
the algorithm. Hence, in this paper, we randomly choose a
solution instead of the current solution to make up a modified
search equation, which can bring more information and produce
a promising candidate solution to balance exploration and
exploitation. We use the modified search equation instead of
the old one in ABC and get a new algorithm named as CABC.
On the other hand, the search equation of ABC, which is
much like a blind mutation operator to search in a randomly
selected dimension of the solution vector, is short of an effective
1012
learning mechanism. This is because that one individual may
have good values on some dimensions of the solution vector
while the other individual may have good values on some other
dimensions. At the same time, we note that the orthogonal
experimental design (OED) offers an ability to discover the best
combination levels for different factors with a reasonably small
number of experimental samples [32]. Therefore, in this paper,
the OED is used to form an orthogonal learning (OL) strategy
for ABC to discover and preserve useful information from the
search experiences. Owing to the OED’s orthogonal test ability
and prediction ability, the OL strategy can bring better learning
efficiency to ABC, which results in faster convergence speed
and better global optimization performance. This learning strat-
egy is applicable to ABC, GABC, and CABC. Accordingly,
the OL strategy-based ABCs are termed as OABC, OGABC,
and OCABC, respectively. Moreover, the experiment results
on a set of benchmark functions demonstrate the effectiveness
of the modified search equation and the OL strategy in solv-
ing complex numerical optimization problems when compared
with other algorithms.
The rest of this paper is organized as follows. In Section II,
the standard ABC algorithm is presented, and the research
works on improved ABCs are reviewed. In Section III, the
OCABC algorithm is proposed based on the new solution
search equation and the OL strategy. Section IV presents and
discusses the experimental results. Finally, the conclusion is
drawn in Section V.
II. ABC F RAMEWORK
A. ABC Algorithm
The ABC algorithm is a recently introduced optimization
algorithm proposed by Karaboga [6], [11]–[13] which is in-
spired by the intelligent foraging behavior of the honeybee
swarm. In ABC, the colony of artificial bees is divided into
three groups: employed bees, onlookers, and scouts. Half of the
colony consists of the employed bees, and another half consists
of the onlookers. Employed bees are responsible for searching
available food sources and gathering required information.
They also pass their food information to onlooker bees. The
onlookers select good food sources from those found by the
employed bees to further search the foods. When the quality
of the food source is not improved through a predetermined
number of cycles, the food source is abandoned by its employed
bee, and then, the employed bee becomes a scout and starts to
search for a new food source in the vicinity of the hive.
In ABC, the position of a food source corresponds to a
possible solution to the optimization problem, and the nectar
amount of each food source represents the quality (fitness)
of the associated solution. The number of the employed bees
equals to the number of food sources. At the initialization step,
ABC generates a randomly distributed initial population P of
SN solutions (food source positions), where SN denotes the
size of employed bees or onlooker bees. Each initial solution
Xi = {xi,1 , xi,2 , . . . , xi,D } is produced randomly within the
range of the boundaries of the parameters as follows:
xi,j = xmin,j + rand(0, 1)(xmax,j − xmin,j ) (2.1)
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013
where i = 1, 2, . . . , SN , j = 1, 2, . . . , D, and D is the number
of optimization parameters; xmin,j and xmax,j are the lower
and upper bounds for the dimension j, respectively.
After the initialization, the population of the food sources
(solutions) is subjected to repeated cycles of the search pro-
cesses of the employed bees, onlookers, and scouts. Each
employed bee always remembers its previous best position and
produces a new position within its neighborhood in its memory.
If the new food source has equal or better quality than the old
source, the old source is replaced by the new source. Otherwise,
the old source is retained. After all employed bees finish their
search process, they will share the information about nectar
amounts and positions of food sources with onlookers. Each
onlooker chooses a food source depending on the probability
value associated with the food source and searches the area
within its neighborhood to generate a new candidate solution,
and then, as in the case of the employed bee, the greedy
selection is applied again. When the nectar of a food source is
abandoned, the corresponding employed bee becomes a scout.
The scout will randomly generate a new food source to replace
the abandoned position.
An onlooker bee chooses a food source depending on the
probability value pi associated with that food source, where
f iti
pi = (2.2)

SN
f itj
j=1
and f iti is the fitness value of solution i. In this way, the
employed bees exchange their information with the onlookers.
In order to produce a candidate food position Vi from the old
one Xi in memory, ABC uses the following expression:
vi,j = xi,j + φi,j (xi,j − xk,j ) (2.3)
where k ∈ {1, 2, . . . , SN } and j ∈ {1, 2, . . . , D} are randomly
chosen indices; k has to be different from i, and φij is a random
number in the range [−1, 1].
If a position cannot be improved further through a predeter-
mined number of cycles, then that food source is assumed to be
abandoned. The value of the predetermined number of cycles
is an important control parameter of ABC, which is called the
limit for abandonment. Assume that the abandoned source is
Xi . Then, the scout produces a food source randomly as in (2.1)
to replace Xi .
In particular, the frame of ABC is presented in Fig. 1.
B. Improved ABCs
Being fascinated by the prospect and potential of ABC,
many researchers are now working on its improvement, which
resulted in a number of variant ABCs. A brief overview of these
contemporary research efforts is presented in this section. These
improvements can be mainly classified as two categories.
1) Introduction of new solution search equations for gener-
ating new solutions [17]–[23]. For example, inspired by
PSO, Zhu and Kwong [17] proposed a GABC by incor-
porating the information of the gbest solution into the
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL
Fig. 1. Framework of ABC.
solution search equation to improve the exploitation. The
experimental results show that GABC can outperform
ABC in most experiments. Banharnsakun et al. [18]
presented a modified search equation for the onlooker
bees. In their method, the best feasible solutions found so
far are shared globally among the entire population. Thus,
the new candidate solutions are more likely to be close
to the current best solution. Li et al. [19] proposed an
improved ABC by introducing not only the best solutions
found so far but also the inertia weight and acceleration
coefficients to modify the search process. Tsai et al. [20]
introduced the concept of universal gravitation into the
consideration of the affection between the employed bees
and the onlooker bees and proposed the interactive ABC.
Coelho and Alotto [21] developed a novel alternative
search equation in which a parameter is responsible for
the balance between the Gaussian and the uniform dis-
tribution. Inspired by DE, Gao and Liu [22] proposed
two improved solution search equations. Furthermore,
in order to achieve a better balance between the ex-
ploration and exploitation of the two different solution
search equations, they used a selective probability P to
control the frequency of introducing the two different so-
lution search equations and got a new search mechanism.
Alam et al. [23] introduced an adaptation scheme of
mutation step size for the solution search equation of
ABC to find a suitable scaling factor.
2) Hybridization of ABC with other operators [24]–[31].
For example, Basturk and Karaboga [24] proposed a
1013
modified ABC by controlling the frequency of pertur-
bation and introducing the ratio of the variance opera-
tor. Kang et al. [25] developed a hybrid simplex ABC
which combines the Nelder–Mead simplex method with
ABC. Kang et al. [26] used the Rosenbrock’s rotational
direction method to implement the exploitation phase and
proposed the Rosenbrock ABC. Alatas [27] introduced
the chaotic maps into the initialization and chaotic search
into the scout bee phase and proposed the chaotic ABC.
Zhao et al. [28] presented a hybrid swarm intelligent
approach. The main idea of the approach is to obtain the
parallel computation merit of GA and self-improvement
merit of ABC by sharing information between the GA
population and bee colony. Duan et al. [29] adopted
ABC to increase the local search capacity as well as the
randomness of the populations. In this way, the improved
quantum EA can jump out of the premature convergence
and find the optimal value. Shi et al. [30] developed
a novel hybrid swarm intelligent algorithm. In their
method, two information exchanging processes are intro-
duced to share valuable information mutually between the
particle swarm and bee colony. Sumpavakup et al. [31]
proposed a cultural-based bee colony algorithm. The
cultural algorithm operates in two spaces, i.e., population
space and belief space. In this hybrid approach, ABC
works in the population space.
It is necessary to emphasize that our work falls in both of the
categories.
1014
Fig. 2. Oscillation phenomenon in GABC. (a) Moving toward Y . (b) Moving
toward Xk .
III. ABC A LGORITHM BASED ON M ODIFIED
S EARCH E QUATION AND OL
A. New Modified Search Equation
Zhu and Kwong [17] pointed out that the solution search
equation described by (2.3) is poor at exploitation. To address
this concerning issue, inspired by PSO, in order to improve
the exploitation and take advantage of the information of the
gbest solution to guide the search of candidate solutions, they
presented the following solution search equation in GABC:
vi,j = xi,j + φi,j (xi,j − xk,j ) + ψi,j (yj − xi,j )
where the third term in the right-hand side of (3.1) is a new
added term called the gbest term, yj is the jth element of the
gbest solution, and ψi,j is a uniform random number in [0,1.5].
However, since the guidance of the last two terms may be in
opposite directions, it may cause an “oscillation” phenomenon.
For the clearness and easiness of understanding, we first sim-
plify (3.1) as (3.1a) by removing the weight components and
the first term as follows:
vi,j = (xi,j − xk,j ) + (yj − xi,j ).
In (3.1a), we consider the following case for a minimization
problem where Xi is between Xk and Y , as shown in Fig. 2.
At first, the distance between Y and Xi may be farther than the
one between Xk and Xi , as in Fig. 2(a), and then, Xi will move
toward Y because of its larger pull. However, as moving toward
Y , the distance between Xk and Xi will increase, as shown
in Fig. 2(b). In this case, Xi will move toward Xk instead.
The oscillation would thus occur, and Xi will be puzzled
in deciding where to stay [32]. This oscillation phenomenon
causes inefficiency to the search ability of the algorithm and
delays convergence.
To overcome these issues, we present a new search equation
as follows:
vi,j = xr1,j + φi,j (xr1,j − xr2,j )
where the indices r1 and r2 are distinct integers uniformly
chosen from the range [1, SN ] and are also different from i and
φij is a random number in the range [−1, 1]. Since (3.2) looks
like the crossover operator of GA, we name the ABC with this
search equation as CABC in this paper.
In (3.2), the vectors for generating candidate solutions are
all selected from the population randomly. Consequently, it
has no bias to any search directions. The movement of a
bee in 2-D parametric space according to (3.1) and (3.2) has
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013
Fig. 3. Movement of a bee according to the search equation of GABC and
CABC in 2-D parametric space.
Fig. 4. Movement of a bee according to the search equation of IABC1 and
IABC2 in 2-D parametric space.
been diagrammatically shown in Fig. 3. Compared to (3.1),
(3.2) has two obvious advantages. First, since φi,j (Xi − Xk )
(3.1) and ψi,j (Y − Xi ) in (3.1) may be in opposite directions, it
may cause an “oscillation” phenomenon and inefficiency to
the search ability of the algorithm. However, (3.2) with the
guidance of the only one term φi,j (Xr1 − Xr2 ) can easily
avoid the oscillation phenomenon and maximize the search
ability of the algorithm. Second, what is more, in (3.2), the
new candidate solution is generated around a randomly selected
Xr1 , and in (3.1), the one is near Xi . Hence, (3.2) can bring
more information to the search equation and produce a more
promising candidate solution to improve the performance of the
(3.1a) algorithm when compared with (3.1).
Furthermore, in order to analyze the effectiveness of (3.2),
we compare (3.2) with the search equations of improved artifi-
cial bee colony algorithm (IABC) [22]. In IABC, two improved
solution search equations are proposed as follows:
vi,j = xbest,j + φi,j (xi,j − xr1,j ), (3.3)
vi,j = xr1,j + φi,j (xi,j − xr2,j ) (3.4)
where the indices r1 and r2 are mutually exclusive integers
randomly chosen from {1, 2, . . . , SN }, and both of them are
different from the base index i; Xbest is the best individual
vector with the best fitness in the current population, and j ∈
{1, 2, . . . , D} is a randomly chosen index. The movement of
a bee in 2-D parametric space according to the two equations
has been diagrammatically shown in Fig. 4. According to the
(3.2) aforementioned analysis, we can see that (3.3) and (3.4) can
also weaken the “oscillation” phenomenon. However, since
(3.3) generates a candidate solution around Xbest , it focuses
on the exploitation. As (3.4) is based on three vectors, Xi ,
Xr1 , and Xr2 , and the two latter ones are two different indi-
viduals randomly selected from the population, it emphasizes
the exploration. It is well known that both exploration and
exploitation are necessary for an EA. In practice, the explo-
ration and exploitation contradict each other. In order to achieve
good performances on problem optimizations, the two abilities
should be well balanced [17]. Therefore, (3.3) and (3.4) cannot
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL 1015

TABLE I
R ESULT C OMPARISONS OF IABC1, IABC2, AND CABC ON 30-D IMENSIONAL F UNCTIONS f1 − f20
AND 100-D IMENSIONAL F UNCTIONS f21 AND f22 W ITH 100 000 F UNCTION E VALUATIONS (FES)

TABLE II TABLE III

FACTORS AND L EVELS OF THE C HEMICAL E XPERIMENT E XAMPLE OA OF THE OED M ETHOD

maximize the search ability of the algorithm. For convenience,

we call the ABC with (3.3) and (3.4) as IABC1 and IABC2,
respectively. Here, the search equation of CABC is devoted
to seeking a balance between exploration and exploitation.
Obversely, CABC has a better exploration ability than IABC1.
At the same time, the search equation of CABC looks like case. Therefore, it is desirable to obtain or predict the best
the crossover operator of GA, and it has better guidance and combination by testing only few representative experimental
exploitation than that of IABC2. cases. The OED [37] using the orthogonal array (OA) and the
In order to make sure whether the performance of the pro- factor analysis (FA) is regarded as a very popular experimental
posed CABC is better than that of GABC, IABC1, and IABC2 design tool for this purpose.
or not, Table I provides the results of comparisons among the 1) OA: LM (QN ) represents an OA with N factors and Q
four algorithms. It can be clearly observed from Table I that levels per factor, where L denotes the OA and M is the number
CABC is superior to the other three algorithms on most test of combinations of test cases. The L9 (34 ) OA is shown in (3.5)
functions. As the only difference among the four algorithms [32]–[37]. In L9 (34 ), there are four factors, three levels per
is the search equation, we conclude that (3.2) has the best factor, and nine combinations of levels. In the first combination,
performance among the four contenders. the four factors have the respective levels of 1, 1, 1, and 1; in the
second combination, the four factors have the respective levels
of 1, 2, 2, and 2, etc.
B. OL Strategy ⎡ ⎤
1 1 1 1
1) OED: We use a simple example to illustrate how to ⎢1 2 2 2⎥
use the OED. For more details, see [32]–[36] and [39]. The ⎢ ⎥
⎢1 3 3 3⎥
conversion ratio of a chemical experiment depends on three ⎢ ⎥
⎢2 1 2 3⎥
variables: temperature (A), time (B), and alkali (C). These three 4 ⎢ ⎥
L9 (3 ) = ⎢ 2 2 3 1 ⎥ . (3.5)
quantities are called the factors of the experiment. Each factor ⎢ ⎥
⎢2 3 1 2⎥
has three possible values shown in Table II [32], and we say that ⎢ ⎥
⎢3 1 3 2⎥
each factor has three levels. ⎣ ⎦
3 2 1 3
To find the best combination of levels for a maximum conver- 3 3 2 1
sion ratio, we can do one experiment for every combination and
then select the best one. In the aforementioned example, there As any subcolumn of an OA is also an OA, for the example
are, in total, 33 = 27 combinations of experimental designs. shown in Table II, we apply only the first three columns
In general, when there are N factors and Q levels per factor, (or arbitrary three columns) of the L9 (34 ) OA to select nine
there are QN combinations. When N and Q are large, it is representative combinations to be tested, and these nine combi-
not possible or efficient to test all the combinations in the nations are shown in Table III [32].
1016
TABLE IV
FA OF THE OED M ETHOD
2) FA: The FA can evaluate the effects of individual factors
on the experimental result, rank the most effective factors, and
determine the best level for each factor. The FA is based on
the experimental results of all the M cases of the OA. The FA
results are shown in Table IV, and the process is described as
follows.
Let fm denote the experimental result of the mth (m =
1, . . . , M ) combination and Snq denote the effect of the qth
(q = 1, . . . , Q) level in the nth (n = A, B, C) factor
M
fm × zmnq
Snq = m=1 M (3.6)
m=1 zmnq
where zmnq is 1 if the level of the nth (n = A, B, C) factor
of the mth (m = 1, . . . , 9) combination is q (q = 1, . . . , 3);
otherwise, zmnq is 0. In this way, the effect of each level
on each factor can be calculated and compared, as shown in
Table IV. For example, when we calculate the effect of level 1
on factor A (A1), the experimental results of C1 , C2 , and C3
are summed up for (3.6) [32] because only these combinations
are involved in A1. Then, the sum divides the combination
number (3 in this case) to yield Snq (SA1 in this case). After
all the Snq ’s are calculated, the best level of each factor can be
determined by selecting the level of each factor that provides
the highest quality Snq . As the example shown in Table IV
is a maximization problem, the larger the Snq , the better the
qth level on factor n is. Otherwise, vice versa. Thus, the best
combination of (A3, B2, C2) can be obtained.
2) OL: Owing to the poor learning mechanism in ABC, it
does not carry out a systematic search. We use the OED to
form an OL strategy for ABC to construct an efficient and
promising candidate solution. At the same time, in order to add
more variation to the search in our framework, the following
formulation is used to generate a transmission vector Ti which
will take part in the OL:
Ti = Xk + rand(0, 1)(Xbest − Xk ) (3.7)
where rand(0, 1) denotes a uniformly distributed random num-
ber in (0, 1), k is an integer uniformly chosen from the range
[1, SN ] and is also different from i, and Xbest is the best
individual with the best fitness in the current population.
The OL is used to combine the information of Ti and Xi
to form a better candidate solution Vi . However, if the OL is
applied to every pair of Ti and Xi , the number of function eval-
uations is SN · (M + 1) at each generation when LM (QN )
is used, whereas the standard ABC only spends SN function
evaluations. It is unwise to apply the OL to every pair of Ti and
Xi . Hence, in order to save the computational cost and keep
the implementation simple, the OL is used only once at each
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013
generation. By doing this, the overhead of the OL in preparing
experiments at each generation is significantly reduced.
How to use the OL to generate a candidate solution Vi will
be described in the next section.
C. Algorithmic Framework
It is important to point out that, in this paper, we intend to
make use of the OL to improve the search ability of the standard
ABC, GABC, and CABC. As an example of this, an OL
strategy-based CABC (OCABC), which combines CABC with
OL, is presented in Fig. 5. The main steps of the algorithm are
summarized as follows. After the initialization, the algorithm
repeats the cycle of the search processes of the employed bees,
onlooker bees, and scout bees. In the employed bee phase, a
randomly selected individual from the population uses the OL
to construct a candidate solution, while other individuals em-
ploy (3.2) to generate a candidate solution. A greedy selection is
applied between Xi and Vi , and the better one is selected. After
all the employed bees complete their searches, the probability
values pi (i = 1, 2, . . . , SN ) are calculated by (2.2). Next,
the algorithm enters the onlooker phase. Each onlooker bee
chooses a food source site with a probability related to its nectar
amount and produces a candidate solution by (3.2). Moreover,
the greedy selection is also applied between Xi and Vi , and
the better one is selected. Finally, the scout process is executed.
Then, the algorithm turns to the employed bee phase and repeats
the aforementioned steps until a specified stop condition is met.
Accordingly, ABC and GABC combined with OL are called as
OABC and OGABC, respectively.
Here, the construction process of a candidate solution Vi by
the OL is described in the following six steps.
Step 1) Generate an LM (2N ) OA where M = 2log2 (D+1) ,
and a number of OAs can be found in http://www2.
research.att.com/~njas/oadir/.
Step 2) According to the OA, make up M tested solutions
Zj (1, . . . , M ) by selecting the corresponding value
from Ti or Xi .
Step 3) Evaluate each tested solution Zj (j = 1, . . . , M ),
and record the best (with best fitness) solution Zb .
Step 4) Calculate the effect of each level on each factor, and
determine the best level for each factor using (3.6).
Step 5) Derive a predictive solution Zp with the levels deter-
mined in Step 4), and evaluate Zp .
Step 6) Compare f (Zb ) with f (Zp ), and the better solution
is used as the vector Vi .
From the aforementioned analysis, it can been seen that a
better solution can be obtained with less number of function
evaluations by the OL strategy. In other words, it is impossible
to obtain such a better solution by so few computation cost
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL
Fig. 5. Framework of OCABC.
without the help of the OL strategy. Therefore, we conclude
that the OL strategy can improve the exploitation ability of the
algorithm.
IV. E XPERIMENTAL S TUDIES ON F UNCTION
O PTIMIZATION P ROBLEMS
A. Benchmark Functions and Parameter Settings
In this section, ABC-based algorithms are applied to mini-
mize a set of 20 scalable benchmark functions of dimensions
D = 15, 30, or 60 [9], [17], [36], [38] and a set of 2 func-
tions of dimensions D = 50, 100, or 200 [36], [38], as shown
in Table V. For convenience, according to different dimen-
sions, the functions are divided into the low-, middle-, and
high-dimensional functions. For example, the low-dimensional
functions contain 15-dimensional functions f1 − f20 and
50-dimensional functions f21 and f22 , etc.
1017
Summarized in Table V are the 22 scalable benchmark
functions. f1 − f6 and f8 are continuous unimodal functions.
f7 is a discontinuous step function, and f9 is a noisy quar-
tic function. f10 is the Rosenbrock function which is uni-
modal for D = 2 and 3 but may have multiple minima in
high dimension cases [40]. f11 − f22 are multimodal, and the
number of their local minima increases exponentially with
the problem dimension. In addition, f14 is the only bound-
constrained function investigated in this paper. Table V also
gives the global optimal value (column 3). Moreover, “Accept”
(column 4) is defined for each test function. If a solution found
by an algorithm falls between the acceptable value and the
actual global optimum (column 3), the run is judged to be
successful.
For a fair comparison among ABCs, they are tested using
the same settings of the parameters, i.e., the population size
SN = 100 [8] and limit = 200 [24]. Furthermore, we set
the maximum number of function evaluations to be 50 000,
1018
TABLE V
B ENCHMARK F UNCTIONS U SED IN E XPERIMENTS
100 000, and 200 000 in the case of D = low, middle, and high.
All the results reported in this section are obtained based on 50
independent runs. For clarity, the results of the best algorithms
are marked in boldface.
B. Comparison Among ABC Algorithms
The mean and the standard deviation of the results ob-
tained by each algorithm for f1 − f22 are summarized in
Tables VI–VIII. An interesting result is that the six ABC-based
algorithms have most reliably found the minimum of the Step
function f7 . It is a region rather than a point in f7 that is the op-
timum. Hence, this problem may relatively be easy to solve with
a 100% success rate. Important observations about the solution
accuracy of different algorithms can be made from the results
presented in Tables VI–VIII on the other 21 test functions.
First, the performance comparison among ABC, GABC, and
CABC is discussed. For the low-dimensional functions, CABC
is significantly better than ABC and GABC on 21 test functions.
For the middle- and high-dimensional functions, CABC always
outperforms ABC and GABC on 20 and 21 test functions, re-
spectively, except that ABC is better than GABC and CABC on
the Griewank function f13 . Second, the solutions obtained by
ABCs with OL are compared with the ones obtained by ABCs
without OL. For the low- and middle-dimensional functions,
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013
OABC, OGABC, and OCABC are always better than, or at
least comparable to, ABC, GABC, and CABC on 20, 19, and 20
test functions, respectively. For the high-dimensional functions,
OABC, OGABC, and OCABC are superior to ABC, GABC,
and CABC on 19, 19, and 17 test functions. Finally, for the
three OL-based ABCs, no method can surpass OCABC on any
test function except that OABC performs better than OCABC
on the Rosenbrock function (f10 ). Moreover, OCABC is better
than the five competitors on most test functions.
According to the aforementioned analyses, it can be con-
cluded that the modified search equation and the OL strategy
are beneficial to the ABC performance. What is more, the
ABC with the OL and modified search equation is much better
than the ABC with either or neither of them on most test
functions.
In order to compare the convergence rate and reliability of
different algorithms, more experimental results are given and
compared in Tables IX–XI. The results given there are the
average FES (AVEN) needed to reach the threshold expressed
as acceptable solutions specified in Table V. In addition, the
successful rates (SR%) of the 50 independent runs for each
function are also compared. Note that the AVEN are calculated
only for the runs that have been “successful.” For convenience
of illustration, we also plot the convergence graph for some test
functions in Fig. 6.
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL 1019

TABLE VI
R ESULT C OMPARISONS OF ABC S ON 15-D IMENSIONAL F UNCTIONS f1 − f20 AND 50-D IMENSIONAL F UNCTIONS f21 AND f22

TABLE VII
R ESULT C OMPARISONS OF ABC S ON 30-D IMENSIONAL F UNCTIONS f1 − f20 AND 100-D IMENSIONAL F UNCTIONS f21 AND f22

TABLE VIII
R ESULT C OMPARISONS OF ABC S ON 60-D IMENSIONAL F UNCTIONS f1 − f20 AND 200-D IMENSIONAL F UNCTIONS f21 AND f22
1020 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013

TABLE IX
C ONVERGENCE S PEED AND S UCCESSFUL R ATE C OMPARISONS OF ABC S ON 15-D IMENSIONAL
F UNCTIONS f1 − f20 AND 50-D IMENSIONAL F UNCTIONS f21 AND f22

TABLE X
C ONVERGENCE S PEED AND S UCCESSFUL R ATE C OMPARISONS OF ABC S ON 30-D IMENSIONAL
F UNCTIONS f1 − f20 AND 100-D IMENSIONAL F UNCTIONS f21 AND f22

It can be observed from Tables IX–XI and Fig. 6 that CABC
is faster than ABC and GABC and that OCABC is faster than
OABC and OGABC on all the test functions. This indeed shows
the advantages of the OL and the modified search equation in
constructing a promising candidate solution to make the search
faster. In particular, with a reasonable agreement to the fact
that GABC is always faster than ABC and slower than CABC,
OGABC is observed to be faster than OABC and slower than
OCABC, and OCABC is the fastest algorithm among the six
contenders.
Furthermore, the successful rates shown in Tables IX–XI also
indicate that the OL and modified search equation are very
promising in bringing a high reliability to ABC. In particu-
lar, OCABC results in higher algorithm reliability with 100%
successful rate on all the test functions except the Michalewicz
function (f22 ) with D = 200.
Experimental results and comparisons verify that the ABC
with the OL and modified search equation performs better than
the ABC with either or neither of them on most test functions.
Moreover, the OL and modified search equation work together
to improve the performance of ABC rather than contradict
each other. Thus, OCABC is the best algorithm among the six
contenders, in terms of solution accuracy, convergence speed,
and algorithm reliability.
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL 1021

TABLE XI
C ONVERGENCE S PEED AND S UCCESSFUL R ATE C OMPARISONS OF ABC S ON 60-D IMENSIONAL
F UNCTIONS f1 − f20 AND 200-D IMENSIONAL F UNCTIONS f21 AND f22

Fig. 6. Convergence performance of different ABCs on the 16 test functions.

with adaptive Levy mutation (ALEP) [41], fast evolutionary

C. Comparisons With Other State-of-the-Art Algorithms
programming (FEP) with Cauchy mutation [42], orthogonal
In Table XII, OCABC is compared with some variants GA with quantization (OGA/Q) [36], estimation of distribu-
of EAs. These algorithms include evolutionary programming tion algorithm with local search (EDA/L) [43], conventional
1022 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013

TABLE XII
C OMPARISONS B ETWEEN OCABC AND S EVERAL VARIANT EA S ON O PTIMIZING 30- OR 100-D IMENSIONAL F UNCTIONS

TABLE XIII
C OMPARISONS B ETWEEN OCABC AND DE S ON O PTIMIZING 30-D IMENSIONAL F UNCTIONS

TABLE XIV
C OMPARISONS B ETWEEN OCABC AND PSO S ON O PTIMIZING 30-D IMENSIONAL F UNCTIONS

evolutionary programming (CEP) [44], and EA based on level-
set evolution and latin squares (LEA) [38]. The results of
the compared algorithms are all derived directly from their
corresponding references. NA represents that the results are not
available in the corresponding reference. The results show that
OCABC yields good performance and does best on 7 of 9 test
functions with the smallest FES, while owing to the advantages
of the OED method, OGA/Q surpasses OCABC on functions
Sphere and Schwefel 2.22.
Further experimental results are listed in Tables XIII and
XIV. OCABC is compared with the standard DE [3], self-
adapting control parameters in DE (jDE) [45], DE with strategy
adaptation (SaDE) [46], and adaptive DE (JADE) with optional
external archive [47] in Table XIII. The results of these vari-
ant DEs are based on the reports in the literature [47], [48].
Additionally, OCABC is compared with the standard PSO, the
“fully informed” PSO (FIPS) [49], the self-organizing hierar-
chical PSO with time-varying acceleration coefficients (HPSO-
TVAC) [50], the comprehensive learning PSO (CLPSO) [51],
and orthogonal learning PSO (OLPSO-G) [32] in Table XIV.
The results of these variant PSOs are based on the reports in
the literature [32]. It is clear that OCABC works best in almost
all the cases and achieves overall better performance than other
competitive algorithms.
D. Discussions
OCABC, OLPSO-G, and OGA/Q are three population-based
algorithms which use the OED to enhance the performance
of the algorithms. However, the way of using the OED is
GAO et al.: NOVEL ABC ALGORITHM BASED ON MODIFIED SEARCH EQUATION AND OL
different. In OCABC, the OED is used to form an OL to
discover more useful information that lies in two solutions,
and the OL is only imposed to a randomly selected individual
from the population. In OLPSO-G, the OL is imposed to each
individual of the population, while in OGA/Q, the OED is used
to form an orthogonal initialization and orthogonal crossover
via a quantization technique.
It can be seen from Table XII that OCABC is superior to
OGA/Q on five functions and OGA/Q surpasses OCABC only
on functions Sphere and Schwefel 2.22. This may be due to the
fact that, in OCABC, the modified search equation is capable
of improving the search ability of the algorithm and the OL
employs the OED’s good character of sampling a small number
of well representative combinations for testing to construct a
more promising and efficient candidate solution for enhancing
the exploitation. Both the modified search equation and the OL
can remedy the poor exploitation of ABC. Thus, OCABC is
able to reach the balance between exploration and exploitation
and has faster convergence speed. At the same time, it can be
observed from Table XIV that OLPSO-G is worse than OCABC
on all the test functions except the function Step. This may be
due to the fact that, in OLPSO-G, global version PSO (GPSO)
uses the best historical experience of the entire swarm as its
neighborhood’s best historical experience [32] and the OL is
used to discover and preserve useful information to predict
the promising region for guiding the particle. Both GPSO and
the OL have strong ability of exploitation, which results in
providing a worse diversity and easily getting trapped in the
local optima when solving complex multimodal problems.
V. C ONCLUSION
To the problem of the poor convergence in ABC, we first
propose an improved ABC method called as the CABC. In this
method, the modified search equation results in improving the
exploitation ability of the standard ABC and easing the “os-
cillation” phenomenon in GABC [17]. Furthermore, we apply
an OL strategy to discover more useful information from the
search experiences. The OL strategy can help ABC construct a
more promising and efficient candidate solution. In this paper,
the OL strategy is applied to the three versions of ABC, i.e., the
standard ABC, GABC, and CABC, yielding OABC, OGABC,
and OCABC, respectively.
Comprehensive experimental tests have been conducted on
22 benchmarks. The experimental results show the high ef-
fectiveness and the high efficiency of the modified search
equation and the OL strategy. OCABC significantly surpasses
other compared algorithms on almost all the test functions,
contributing to higher solution accuracy, faster convergence
speed, and stronger algorithm reliability.
It is desirable to further apply OCABC to solve those
more complex real-world continuous optimization problems,
such as clustering, data mining, design, and optimization of
communication networks. The future work also includes the
studies on how to extend OCABC to handle those combinatorial
optimization problems, such as flow shop scheduling problem,
vehicle routing problem, and traveling salesman problem.
1023
R EFERENCES
[1] K. S. Tang, K. F. Man, S. Kwong, and Q. He, “Genetic algorithms and
their applications,” IEEE Signal Process. Mag., vol. 13, no. 6, pp. 22–37,
Nov. 1996.
[2] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc.
IEEE Int. Conf. Neural Netw., 1995, vol. 4, pp. 1942–1948.
[3] R. Storn and K. Price, “Differential evolution—A simple and efficient
heuristic for global optimization over continuous spaces,” J. Global
Optim., vol. 11, no. 4, pp. 341–359, Dec. 1997.
[4] M. Dorigo and T. Stutzle, Ant Colony Optimization. Cambridge, MA:
MIT Press, 2004.
[5] D. Simon, “Biogeography-based optimization,” IEEE Trans. Evol.
Comput., vol. 12, no. 6, pp. 702–713, Dec. 2008.
[6] D. Karaboga, “An Idea Based on Honey Bee Swarm For Numerical Opti-
mization,” Erciyes Univ., Kayseri, Turkey, Tech. Rep.-TR06, 2005.
[7] D. Karaboga and B. Basturk, “A powerful and efficient algorithm for
numerical function optimization: Artificial bee colony (ABC) algorithm,”
J. Global Optim., vol. 39, no. 3, pp. 459–471, Nov. 2007.
[8] D. Karaboga and B. Basturk, “On the performance of artificial bee
colony (ABC) algorithm,” Appl. Soft Comput., vol. 8, no. 1, pp. 687–697,
Jan. 2008.
[9] D. Karaboga and B. Basturk, “A comparative study of artificial bee colony
algorithm,” Appl. Math. Comput., vol. 214, no. 1, pp. 108–132, Aug. 2009.
[10] D. Karaboga, “Artificial bee colony algorithm,” Scholarpedia, vol. 5,
no. 3, p. 6915, 2010.
[11] D. Karaboga and C. Ozturk, “Neural networks training by artificial bee
colony algorithm on pattern classification,” Neural Netw. World, vol. 19,
no. 3, pp. 279–292, May 2009.
[12] D. Karaboga and C. Ozturk, “A novel clustering approach: Artificial bee
colony (ABC) algorithm,” Appl. Soft Comput., vol. 11, no. 1, pp. 652–657,
Jan. 2011.
[13] C. Ozturk, D. Karaboga, and B. Gorkemli, “Probabilistic dynamic de-
ployment of wireless sensor networks by artificial bee colony algorithm,”
Sensors, vol. 11, no. 6, pp. 6056–6065, Jun. 2011.
[14] M. F. Tasgetiren, Q. K. Pan, P. N. Suganthan, and A. L. Chen, “A dis-
crete artificial bee colony algorithm for the total flowtime minimization
in permutation flow shops,” Inf. Sci., vol. 181, no. 16, pp. 3459–3475,
Aug. 2011.
[15] Q. K. Pan, M. F. Tasgetiren, P. N. Suganthan, and T. J. Chua, “A discrete
artificial bee colony algorithm for the lot-streaming flow shop scheduling
problem,” Inf. Sci., vol. 181, no. 12, pp. 2455–2468, Jun. 2011.
[16] V. J. Manoj and E. Elias, “Artificial bee colony algorithm for the de-
sign of multiplier-less nonuniform filter bank transmultiplexer,” Inf. Sci.,
vol. 192, pp. 193–203, Jun. 2012.
[17] G. P. Zhu and S. Kwong, “Gbest-guided artificial bee colony algorithm for
numerical function optimization,” Appl. Math. Comput., vol. 217, no. 7,
pp. 3166–3173, Dec. 2010.
[18] A. Banharnsakun, T. Achalakul, and B. Sirinaovakul, “The best-so-far
selection in artificial bee colony algorithm,” Appl. Soft Comput., vol. 11,
no. 2, pp. 2888–2901, Mar. 2011.
[19] G. Q. Li, P. F. Niu, and X. J. Xiao, “Development and investigation of effi-
cient artificial bee colony algorithm for numerical function optimization,”
Appl. Soft Comput., vol. 12, no. 1, pp. 320–332, Jan. 2012.
[20] P. W. Tsai, J. S. Pan, B. Y. Liao, and S. C. Chu, “Enhanced artificial
bee colony optimization,” Int. J. Innovative Comput., Inf. Control, vol. 5,
no. 12, pp. 1–12, Dec. 2009.
[21] L. S. Coelho and P. Alotto, “Gaussian artificial bee colony algorithm
approach applied to Loneys solenoid benchmark problem,” IEEE Trans.
Magn., vol. 47, no. 5, pp. 1326–1329, May 2011.
[22] W. F. Gao and S. Y. Liu, “Improved artificial bee colony algorithm for
global optimization,” Inf. Process. Lett., vol. 111, no. 17, pp. 871–882,
Sep. 2011.
[23] M. S. Alam, M. W. Kabir, and M. M. Islam, “Self-adaptation of muta-
tion step size in artificial bee colony algorithm for continuous function
optimization,” in Proc. 13th Int. Conf. Comput. Inf. Technol., Dec. 2010,
vol. 3, pp. 69–74.
[24] B. Basturk and D. Karaboga, “A modified artificial bee colony algo-
rithm for real-parameter optimization,” Inf. Sci., vol. 192, pp. 120–142,
Jun. 2012.
[25] F. Kang, J. J. Li, and Q. Xu, “Structural inverse analysis by hybrid simplex
artificial bee colony algorithms,” Comput. Struct., vol. 87, no. 13/14,
pp. 861–870, Jul. 2009.
[26] F. Kang, J. J. Li, and Z. Y. Ma, “Rosenbrock artificial bee colony algo-
rithm for accurate global optimization of numerical functions,” Inf. Sci.,
vol. 181, no. 16, pp. 3508–3531, Aug. 2011.
[27] B. Alatas, “Chaotic bee colony algorithms for global numerical optimiza-
tion,” Expert Syst. Appl., vol. 37, no. 8, pp. 5682–5687, Aug. 2010.
1024
[28] H. Zhao, Z. Pei, J. Jiang, R. Guan, C. Wang, and X. Shi, “A hybrid swarm
intelligent method based on genetic algorithm and artificial bee colony,”
in Proc. ICSI, 2010, vol. 6145, pp. 558–565, Lect. Notes Comput. Sci.
[29] H. B. Duan, C. F. Xu, and Z. H. Xing, “A hybrid artificial bee colony opti-
mization and quantum evolutionary algorithm for continuous optimization
problems,” Int. J. Neural Syst., vol. 20, no. 1, pp. 39–50, Feb. 2010.
[30] X. Shi, Y. Li, H. Li, R. Guan, L. Wang, and Y. Liang, “An integrated
algorithm based on artificial bee colony and particle swarm optimization,”
in Proc. 6th Int. Conf. Nat. Comput., 2010, vol. 5, pp. 2586–2590.
[31] C. Sumpavakup, S. Chusanapiputt, and I. Srikun, “A hybrid cultural-based
bee colony algorithm for solving the optimal power flow,” in Proc. IEEE
Int. Midwest Symp. Circuits Syst., 2011, pp. 1–4.
[32] Z. H. Zhan, J. Zhang, Y. Li, and Y. H. Shi, “Orthogonal learning particle
swarm optimization,” IEEE Trans. Evol. Comput., vol. 15, no. 6, pp. 832–
847, Dec. 2011.
[33] S. Y. Ho, H. S. Lin, W. H. Liauh, and S. J. Ho, “OPSO: Orthogonal particle
swarm optimization and its application to task assignment problems,”
IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 38, no. 2, pp. 288–
298, Mar. 2008.
[34] Q. F. Zhang and Y. W. Leung, “An orthogonal genetic algorithm for
multimedia multicast routing,” IEEE Trans. Evol. Comput., vol. 3, no. 1,
pp. 53–62, Apr. 1999.
[35] Y. Wang, Z. X. Cai, and Q. F. Zhang, “Enhancing the search ability of
differential evolution through orthogonal crossover,” Inf. Sci., vol. 185,
no. 1, pp. 153–177, Feb. 2012.
[36] Y. W. Leung and Y. Wang, “An orthogonal genetic algorithm with quan-
tization for global numerical optimization,” IEEE Trans. Evol. Comput.,
vol. 5, no. 1, pp. 41–53, Feb. 2001.
[37] Math. Stat. Res. Group, Chinese Acad. Sci., Orthogonal Design
(in Chinese) Beijing, China: People Educ. Pub., 1975.
[38] Y. P. Wang and C. Y. Dang, “An evolutionary algorithm for global op-
timization based on level-set evolution and latin squares,” IEEE Trans.
Evol. Comput., vol. 11, no. 5, pp. 579–595, Oct. 2007.
[39] J. T. Tsai, T. K. Liu, and J. H. Chou, “Hybrid Taguchi-genetic algorithm
for global numerical optimization,” IEEE Trans. Evol. Comput., vol. 8,
no. 4, pp. 365–377, Aug. 2004.
[40] Y. W. Shang and Y. H. Qiu, “A note on the extended Rosenbrock func-
tion,” Evol. Comput., vol. 14, no. 1, pp. 119–126, Mar. 2006.
[41] C. Y. Lee and X. Yao, “Evolutionary programming using mutations based
on the Levy probability distribution,” IEEE Trans. Evol. Comput., vol. 8,
no. 1, pp. 1–13, Feb. 2004.
[42] X. Yao, Y. Liu, and G. M. Lin, “Evolutionary programming made faster,”
IEEE Trans. Evol. Comput., vol. 3, no. 2, pp. 82–102, Jul. 1999.
[43] Q. Zhang, J. Sun, E. Tsang, and J. Ford, “Hybrid estimation of distribution
algorithm for global optimization,” Eng. Comput., vol. 21, no. 1, pp. 91–
107, Jan. 2004.
[44] K. Chellapilla, “Combining mutation operators in evolutionary program-
ming,” IEEE Trans. Evol. Comput., vol. 2, no. 3, pp. 91–96, Sep. 1998.
[45] J. Brest, S. Greiner, B. Boskovic, M. Mernik, and V. Zumer, “Self-
adapting control parameters in differential evolution: A comparative study
on numerical benchmark problems,” IEEE Trans. Evol. Comput., vol. 10,
no. 6, pp. 646–657, Dec. 2006.
[46] A. K. Qin, V. L. Huang, and P. N. Suganthan, “Differential evolution
algorithm with strategy adaptation for global numerical optimization,”
IEEE Trans. Evol. Comput., vol. 13, no. 2, pp. 398–417, Apr. 2009.
[47] J. Zhang and A. C. Sanderson, “JADE: Adaptive differential evolution
with optional external archive,” IEEE Trans. Evol. Comput., vol. 13, no. 5,
pp. 945–958, Oct. 2009.
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 3, JUNE 2013
[48] W. Y. Gong, Z. H. Cai, C. X. Ling, and H. Li, “Enhanced differential evo-
lution with adaptive strategies for numerical optimization,” IEEE Trans.
Syst., Man, Cybern. B, Cybern., vol. 41, no. 2, pp. 397–413, Apr. 2011.
[49] R. Mendes, J. Kennedy, and J. Neves, “The fully informed particle swarm:
Simpler, maybe better,” IEEE Trans. Evol. Comput., vol. 8, no. 3, pp. 204–
210, Jun. 2004.
[50] A. Ratnaweera, S. Halgamuge, and H. Watson, “Self-organizing hierarchi-
cal particle swarm optimizer with time-varying acceleration coefficients,”
IEEE Trans. Evol. Comput., vol. 8, no. 3, pp. 240–255, Jun. 2004.
[51] J. J. Liang, A. K. Qin, P. N. Suganthan, and S. Baskar, “Comprehensive
learning particle swarm optimizer for global optimization of multimodal
functions,” IEEE Trans. Evol. Comput., vol. 10, no. 3, pp. 281–295,
Jun. 2006.
Wei-feng Gao received the B.S. degree in applied
mathematics from the School of Science, Xidian
University, Xi’an, China, in 2008, where he is cur-
rently working toward the Ph.D. degree.
His current research interests include evolutionary
algorithms and their applications in the real world.
San-yang Liu received the M.S. degree in applied
mathematics from Xidian University, Xi’an, China,
in 1984 and the Ph.D. degree in computational
mathematics from Xi’an Jiaotong University, Xi’an,
in 1989.
After finishing the Ph.D. degree, he spent one year
at the University Paul Sabatier, Toulouse, France,
as a postdoctoral fellow. He is currently a Profes-
sor and the Dean of the School of Science, Xidian
University. His research interests include nonlinear
optimization, combinatorial optimization, network
optimization, system reliability, and so on.
Ling-ling Huang received the Ph.D. degree in ap-
plied mathematics from Xidian University, Xi’an,
China, in 2012.
She is currently a Lecturer in the China University
of Petroleum, Qingdao, China. Her major research
interests concentrate on optimization theory, meth-
ods, and applications.