Applied Soft Computing 33 (2015) 170–182

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

A modified particle swarm optimization with multiple

subpopulations for multimodal function optimization problems
Wei-Der Chang ∗
Department of Computer and Communication, Shu-Te University, Kaohsiung 824, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a modified particle swarm optimization (PSO) algorithm is developed for solving multi-
Received 21 July 2014 modal function optimization problems. The difference between the proposed method and the general
Received in revised form PSO is to split up the original single population into several subpopulations according to the order of par-
15 December 2014
ticles. The best particle within each subpopulation is recorded and then applied into the velocity updating
Accepted 2 April 2015
formula to replace the original global best particle in the whole population. To update all particles in each
Available online 27 April 2015
subpopulation, the modified velocity formula is utilized. Based on the idea of multiple subpopulations,
for the multimodal function optimization the several optima including the global and local solutions may
Keywords:
Particle swarm optimization (PSO)
probably be found by these best particles separately. To show the efficiency of the proposed method, two
Multiple subpopulations kinds of function optimizations are provided, including a single modal function optimization and a com-
Multimodal optimization problem plex multimodal function optimization. Simulation results will demonstrate the convergence behavior
of particles by the number of iterations, and the global and local system solutions are solved by these
best particles of subpopulations.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction employed and developed in recent years [14–19]. For example,

a novel evolutionary algorithm which combines Niche technique
Particle swarm optimization (PSO) algorithm is a rather popular with the gravitational search algorithm (NGSA) was presented for
and useful tool for solving any kind of engineering optimizations solving multimodal function optimizations [14]. Three strategies
over the few years [1–13]. It was firstly proposed by Kennedy and consisting of a K-nearest neighbors strategy, an elitism strategy,
Eberhart in 1995 and has been proven to be a simple but power- and modification of active gravitational mass formulation are con-
ful optimal algorithm among a lot of evolutionary algorithms (EAs) sidered. In [15], for multimodal optimizations they developed a
[1]. The main features consist of the simplicity of updating formu- new variant of differential evolution (DE) algorithm in which new
las, easy for implementation on computer programs, real number individuals close to an isolated individual in a current population
manipulations, quicker convergence, etc. Due to these excellent are generated. This mechanism will allocate search resources for
advantages, a variety of engineering applications based on the PSO each optimum of the function. Another new algorithm named as
have successively been explored and developed such as job-shop the glowworm swarm optimization (GSO) was presented for simul-
scheduling problems [2,3], digital filter designs [4–6], controller taneous computation of multiple optima of multimodal functions
designs [7–9], power system designs [10,11], antenna array design [18]. This algorithm is to share some features with known swarm
[12] etc. This paper will focus on the issue of solving the multimodal intelligence algorithm like the ant colony optimization and particle
function optimization problem using the modified PSO algorithm. swarm optimization. In [19], they proposed a memetic algorithm
It is well known that most of function optimization problems with which hybridizes particle swarm optimization with a local search
several designed variables are of multimodal, i.e., including multi- technique for locating multiple global and local optimal solutions. A
ple global or/and local optimal solutions, not a single optimum, triggered re-initialization scheme is introduced into the proposed
over the search space. To solve this kind of optimization prob- algorithm to enhance the search performance.
lems, various methods based on evolutionary algorithms have been In the evolutionary algorithms, the concept of subpopulation
(or named as subswarm) is often discussed and utilized especially
for solving multiobjective optimization problems and multimodal
function optimization problems [14,17,20–24]. It is always to split
∗ Tel.: +886 7 6158000 ext. 4808; fax: +886 7 6158000 ext. 4899. up the original population into several small subpopulations with
E-mail address: [email protected]

http://dx.doi.org/10.1016/j.asoc.2015.04.002
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 171

a few particles for PSO, chromosomes for GA, or parameter vectors denoted by gbest is the best particle in the population. Every parti-
for DE algorithm for some special purposes. Then, certain evolving cle contained in the population is updated according to these two
mechanisms and operations are executed to achieve the system main information pbest and gbest, i.e., they guide the search direc-
design. In [20], a sub-population genetic algorithm (SPGA) was tion of all particles over the search space. How to evaluate the pbest
utilized and combined with a mining gene structure technique for each particle and the gbest for the whole population is deter-
for multiobjective flowshop scheduling problems. In the SPGA, mined by the corresponding cost function value. Generally, it is a
the original is divided into numerous subpopulations which are better particle whose cost function value is smaller.
assigned with different weights to search for optimal solutions in In the general PSO, let the vector X = [x1 , x2 , . . ., xn ] be a particle,
different directions. They claimed that this approach can consider where xi for i = 1, 2, . . ., n is the decision variable or the designed
both the expanding and converging natures of the system solutions variable of the optimized problem. The following two updating for-
[20]. In [21], the author presented a two-layer particle swarm opti- mulas are utilized to guide the moving of each particle, the velocity
mization (TLPSO) consisting of the top layer and the bottom layer to updating formula of Eq. (1) and the position updating formula of
increase the search diversity of the particles so that the drawback Eq. (2):
of trapping in some local optimum can be avoided. In the devel-
oped structure, it contains M swarms of particles in the bottom
vij (t + 1) = w ∗ vij (t) + c1 · r1 · (pbestij (t) − xij (t))
layer and one swarm of particles in the top layer. Each global best + c2 · r2 · (gbestj (t) − xij (t)), (1)
position in each swarm of the bottom layer is set to be the position
of the particle in the swarm of the top layer. A mutation mech-
anism is then introduced into the particles of each swarm in the
xij (t + 1) = xij (t) + vij (t), (2)
bottom layer so that the particles leap the local minimum to search
for the global one [21]. Furthermore, a competitive and cooperative where xij , pbestij , and gbestj represent the jth position components
co-evolutionary approach to multiobjective PSO algorithm design of the ith particle, the ith individual best particle, and the global
was presented, which the subswarms are employed to produce the best particle, respectively, vij is the jth velocity component of the
reasonable problem decompositions by exploiting any correlation ith particle, w is the inertia weight for balancing the global and
and interdependency between components of the problem [23]. It local search, c1 and c2 are two positive constants assigned by the
was validated by some comparisons with existing state-of-the-art designer, r1 and r2 are two uniformly random numbers chosen
algorithms. from the interval [0,1]. It is worthy of notice that only one global
This paper will propose a simple modified version of PSO algo- best particle of the population, gbest, is enrolled in the general PSO
rithm with multiple subpopulations for the multimodal function algorithm. In this manner, only one corresponding system solution
optimization. In the beginning, the population needs to be split up can be caught by the gbest for the optimized problem. For solving
into several subpopulations and in each subpopulation the best par- the multimodal function optimization, it is rather improper and
ticle is necessarily recorded, that is, each subpopulation has a best not applicable because the multimodal function consists of several
particle. With the information of the best particle instead of the system optima including global optimal solutions and local optimal
global best particle, all of particles contained in this subpopulation solutions.
are updated. In this manner, simultaneous finding multiple optima To tackle this problem, a modified version of the PSO algorithm is
is probably achieved for the multimodal function optimization. The developed, which the concept of subpopulations is utilized for solv-
remainder of this paper is organized as follows. In Section 2, the dif- ing multimodal functions. In the proposed algorithm, the original
ference between the general PSO algorithm and the proposed one single population is firstly requested to be divided into M sub-
is introduced in detail. A modified version of the velocity updating populations just according to the order of particles. For instance,
formula is described. In Section 3, a single modal function optimiza- assume that there are 40 particles in the original population, and it
tion problem is first simulated by the proposed method and it will will be split up into four subpopulations. In the case, the first sub-
reveal that all simulation results converge to the single optimum. population is composed of particles from number one to number
In Section 4, the proposed method is applied to solving the complex ten, the second subpopulation consists of particles from number
multimodal function optimization problem which consists of sev-
eral global and local minimum points. Simulation results reveal that
the proposed method can seek for most of optima. Finally, Section original population global best
5 gives a simple conclusion and future research work. with N particles

2. The proposed PSO with multiple subpopulations

It is well known that the PSO algorithm has been shown to be a

powerful search means to solving various engineering optimization
problems. It is an iterative and population-based algorithm which
is motivated by the social behavior of a flock of birds and a school
of fishes. In recent years, the PSO algorithm has received much 1st 2nd Mth
attention of many researches and also been applied to a variety subpopulation subpopulation subpopulation
of engineering fields as addressed in the previous section because
of its good features. In the general PSO algorithm, it begins with
generating a random initial population that consists of a number
of individuals (or called particles). For an optimization problem,
every particle represents a candidate solution of the optimized sys-
best particle best particle best particle
tem. It is necessary to enroll some information including both the
individual best for every particle and the global best for the whole M subpopulations
population. The individual best, pbest, denotes the best position of
each particle which it has found so far, and the global best which is Fig. 1. The diagram from a single population to M subpopulations.
172 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

Table 1
Design flow chart of the modified PSO algorithm with multiple subpopulations for
multimodal function optimization.

Begin 120

100

80
Create a population consisting of N particles which

f(x,y)
60
are generated from some search interval [ xmin , xmax ]
40
randomly. 6
20 4
2
0
Divide the population into M subpopulations 0 x
6
4 -2
simply according to the order of particles. Each 2
0 -4
y -2
subpopulation includes N/M particles. -4 -6
-6

Fig. 2. A single modal function f(x, y) = (x − 1)2 + (y − 2)2 .

in each subpopulation. Notice again that it is necessary to record

Is the number Yes the best particle of each subpopulation and the information will be
Stop utilized in the velocity formula to update all of particle positions in
of iterations
the same subpopulation. Besides, the design flow chart for the pro-
achieved? posed PSO method with multiple subpopulations is clearly listed
in Table 1 for the multimodal function optimization problem. The
following describes the complete design steps of the modified PSO
No algorithm.

Calculate the optimized function value for each I. Create a population consisting of N particles which are gener-
particle and enroll the individual best particle pbest ated from some search interval [xmin , xmax ] randomly.
and the best particle gbest in each subpopulation, II. Divide the population into M subpopulations simply accord-
respectively. ing to the order of particles. Each subpopulation includes N/M
particles.
III. If the number of iterations is achieved, then the algorithm
Execute the velocity updating formula of Eq. (1) in stops.
IV. Calculate the optimized function value for each particle and
which the global best is replaced by the best enroll the individual best particle pbest and the best particle
particle of each subpopulation and the position gbest in each subpopulation, respectively.
updating formula of Eq. (2), respectively, for all V. Execute the velocity updating formula of Eq. (1) in which the
particles of each subpopulation. global best is replaced by the best particle of each subpopula-
tion and the position updating formula of Eq. (2), respectively,
for all particles of each subpopulation.

Check the position of each particle by Eq. (3).

6

eleven to number twenty, and so forth. The best particle in each

subpopulation is necessary to be recorded to guide the moving of
0
y

all particles in the same subpopulation by a modified velocity for-

mula. Following this idea the gbest used in Eq. (1) is referred to as
the best particle of each subpopulation, not as the global best par- -2
ticle in the whole population. Moreover, in the proposed method
there are totally M best particles because of having M subpopula-
tions. It means that these M best particles are separately able to -4
catch different optima, at most M optima, for the multimodal func-
tion optimization because some of them probably search for the
same system optimum. Fig. 1 clearly explains the proposed con- -6
-6 -4 -2 0 2 4 6
cept of subpopulations, where N is the number of particles, i.e., x
population size, in the original single population, M is the num-
ber of subpopulations and, therefore, there contains N/M particles Fig. 3. Contour of a single modal function f(x, y) = (x − 1)2 + (y − 2)2 .
W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 173

Fig. 4. Case 1: single population.

Fig. 5. Case 2: two subpopulations.

174 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

Fig. 6. Case 3: four subpopulations.

VI. Check the position of each particle by w = 0.5, positive constants c1 = c2 = 0.5 in Eq. (1), and the search
⎧ interval [xmin , xmax ] = [−6, 6] in Eq. (3), respectively.
⎪ x if xi < xmin
⎨ min Three different examinations are considered including Case
xi = xi if xmin ≤ xi ≤ xmax for i = 1, 2, . . ., n . (3) 1: only one population, Case 2: two subpopulations, and Case
⎪
⎩ 3: four subpopulations for solving such a single modal func-
xmax if xi > xmax tion optimization. Convergence dynamics of all particles will be

VII. Go back to Step III.

2500
3. A single modal function optimization problem

In this section, a single modal function optimization problem 2000

is firstly illustrated to explain the concept of the proposed PSO
algorithm with multiple subpopulations. The optimized function
is given by 1500
f(x,y)

f (x, y) = (x − 1)2 + (y − 2)2 , (4)

1000
where the search space is defined by −6 ≤ x ≤ 6 and −6 ≤ y ≤ 6,
respectively. Fig. 2 shows the 3-D surface of the optimized func-
tion and its corresponding contour is then displayed in Fig. 3. It 500
can easily be seen from Eq. (4) or these figures that it has only one
global minimum occurred at (x, y) = (1, 2). In the following simu-
0
lations, the software of Borland C++ 5.02 under the environment 5
0 2 4 6
of Intel(R) Core(TM) i7-3770 [email protected] is utilized to program -5 -4 -2 0
y -6 x
the above design steps of the proposed PSO algorithm with mul-
tiple subpopulations. Furthermore, the common parameters used 2 2 2
Fig. 7. Multimodal function f (x, y) = (x2 + y − 11) + (x + y2 − 7) + 0.1[(x − 3) +
2
in the PSO are given by the population size N = 20, inertia weight (y − 2) ].
 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 175

6 single modal function, the green point means the best particle in
this single population, and the black point then represents the com-
mon particle. In the case, all particles quickly and fully converge to
4 the global minimum point of the optimized function after about 20
iterations. Fig. 5 displays the simulation results for Case 2 with two
subpopulations. It contains two best particles (two green points)
2 corresponding to two subpopulations. After a few iterations, the
other common black points as well as both two green points quickly
converge to the same point, red point (1, 2). In Case 3, the popula-
0
y

tion size is changed to N = 40 in order to meet the situation of using

four subpopulations. In Fig. 6, there are totally 40 points consisting
of 36 black points and 4 green points due to four subpopulations.
-2
Also, these four green points are successively converging to the
minimum point (1, 2) after some iterations. It is clearly seen from
-4 these figures that all best particles contained in the subpopulations
fully converge to the only one optimal solution of the optimized
function. It can be concluded that all numerical results derived by
-6 using single population and multiple subpopulations are the same
-6 -4 -2 0 2 4 6 for solving the single modal function optimization.
x
2
Fig. 8. Contour of multimodal function f (x, y) = (x2 + y − 11) + (x + y2 − 7) +
2 4. A multimodal function optimization problem
2 2
0.1[(x − 3) + (y − 2) ].
Consider the following multimodal function minimization prob-
displayed, respectively, when iteration = 0, iteration = 5, itera- lem [25]
tion = 10, and iteration = 20. Fig. 4 shows the convergence behavior 2
f (x, y) = (x2 + y − 11) + (x + y2 − 7) + 0.1[(x − 3)2 + (y − 2)2 ].
of particles converging to the global optimum (1, 2) for Case
1, where the red point represents the global minimum of the (5)

Fig. 9. Case 1: single population (converging to (3, 2)).

176 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

Fig. 10. Case 2: single population (converging to (−3.7634, −3.2660)).

Fig. 11. Case 3: single population (converging to (3.5814, −1.8208)).

 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 177

Fig. 12. Case 4: two subpopulations (converging to (−2.7870, 3.1282) and (3, 2), respectively).

Fig. 13. Case 5: two subpopulations (only converging to (−3.7634, −3.2660)).

178 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

Fig. 14. Case 6: two subpopulations (converging to (3.5814, −1.8208) and (3, 2), respectively).

Fig. 15. Case 7: four subpopulations (converging to (−2.7870, 3.1282) and (3, 2), respectively).
 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 179

Fig. 16. Case 8: four subpopulations (converging to (−3.7634, −3.2660), (3.5814, −1.8208), and (3, 2), respectively).

Fig. 17. Case 9: four subpopulations (converging to (−2.7870, 3.1282), (3.5814, −1.8208), and (3, 2), respectively).
180 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

Fig. 18. Case 10: two subpopulations and population size N = 20 (converging to (−2.7870, 3.1282) and (3, 2), respectively). Particles in the first subpopulation are denoted
by black points and particles in the second subpopulation are denoted by blue points.

Fig. 19. Case 11: two subpopulations and population size N = 30 (converging to (−3.7634, −3.2660) and (3, 2), respectively).
 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182 181

Fig. 20. Case 12: two subpopulations and population size N = 40 (converging to (−2.7870, 3.1282) and (3, 2), respectively).

This function is called the modified Himmelblau function which can be seen from Figs. 9–11, the PSO algorithm with a single pop-
is often used as a benchmark minimization problem. Fig. 7 dis- ulation is insufficient and not applicable to solving the complex
plays the 3-D surface of this multimodal function with the defined multimodal function problem. The solution derived may be a local
space −6 ≤ x ≤ 6 and −6 ≤ y ≤ 6, and its corresponding contour is minimum, not a global minimum of the optimized function due to
then shown in Fig. 8 Fig. 8. It can easily be found from Figs. 7 and 8 the effect of algorithm’s initial conditions.
that this optimized function has four minimum points includ- In the following, the population is further divided into a few
ing one global minimum occurred at f(3, 2) = 0 and another three subpopulations so that the best particle in each subpopulation can
local minimum points at f (−3.7634, −3.2660) ≈ 7.3673, f (3.5814, separately search for other minimum points. To explain this, Cases
−1.8208) ≈ 1.5043, and f (−2.7870, 3.1282) ≈ 3.4871, respectively. 4–6 are simulated for two subpopulations under different initial
In addition, the parameter settings employed in the algorithm are conditions. Fig. 12 displays the two best particles converging to
the same with those of the above example. two minimum points, one is the global minimum (3, 2), the other
For solving such a multimodal function optimization problem, a is the local minimum (−2.7870, 3.1282), respectively. Certainly, it
large number of different testing cases are provided to demonstrate maybe converges to only one minimum point (−3.7634, −3.2660)
the efficiency of the proposed method, including single popula- as shown in Fig. 13 or to another two minimum points (3.5814,
tion, two subpopulations, and four subpopulations. In Case 1, Fig. 9 −1.8208) and (3, 2) in Fig. 14. From these simulation results, it can
shows the convergence behavior of particles when iteration = 0, 5, be concluded that the PSO algorithm with two subpopulations is
10, and 20, respectively, where the four red points represent the better than a single population for solving the multimodal function
four minimization points and the green point denotes the global optimization. As for the algorithm with four subpopulations, Cases
best in this single population. In the case, the only one global best 7–9 are executed when number of particles is changed as N = 40.
(green point) also quickly converges to the global minimum (x, All simulation results are displayed in Figs. 15–17, respectively. In
y) = (3, 2), one of these four minimization points. However, the Fig. 15, two minimum points including the global minimum (3, 2)
global best particle probably converges to one of other local mini- and a local minimum (−2.7870, 3.1282) of the optimized function
mum points, for examples, converging to (x, y) = (−3.7634, −3.2660) are separately caught by these four best particles (green points). In
in Fig. 10 or to (x, y) = (3.5814, −1.8208) in Fig. 11, respectively, if Fig. 16, on the other hand, the four best particles then finally con-
the algorithm’s initial conditions are given by different values. As verge to points (3, 2), (−3.7634, −3.2660), and (3.5814, −1.8208),
182 W.-D. Chang / Applied Soft Computing 33 (2015) 170–182

respectively. In the case, three minimum points of the multimodal References

function are solved by using four subpopulations method. Fig. 17
[1] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceedings of the IEEE
shows another simulation result which another three minimum International Conference on Neural Networks, vol. IV, Perth, Australia, 1995, pp.
points (3, 2), (−2.7870, 3.1282), and (3.5814, −1.8208) are found. It 1942–1948.
is clearly shown from all simulations that the PSO algorithm with [2] D.Y. Sha, H.H. Lin, A multi-objective PSO for job-shop scheduling problems,
Expert Syst. Appl. 37 (2010) 1065–1070.
multiple subpopulations is rather effective and applicable in solv- [3] R. Zhang, S. Song, C. Wu, A two-stage hybrid particle swarm optimization algo-
ing the multimodal function optimization problem than the general rithm for the stochastic job shop scheduling problem, Knowl.-Based Syst. 27
PSO algorithm with a single population. (2012) 393–406.
[4] R. Kar, D. Mandal, S. Mondal, S.P. Ghoshal, Craziness based Particle Swarm
Finally, in order to clearly show all of the particle distributions
Optimization algorithm for FIR band stop filter design, Swarm Evol. Comput. 7
in different subpopulations over the search space and to clarity the (2012) 58–64.
effect of different population sizes on the convergence behavior, [5] W.D. Chang, D.M. Chang, Design of a higher-order digital differentiator using
the case of two subpopulations are reconsidered. First, the pop- a particle swarm optimization approach, Mech. Syst. Signal Process. 22 (2008)
233–247.
ulation size is assumed to be N = 20 and it is divided into two [6] S.M. Rafi, A. Kumar, G.K. Singh, An improved particle swarm optimization
subpopulations as well. Fig. 18 displays the simulation results, method for multirate filter bank design, J. Franklin Inst. 350 (2013) 757–769.
where particles belonging to the first subpopulation are denoted [7] A. Lari, A. Khosravi, F. Rajabi, Controller design based on ␮ analysis and PSO
algorithm, ISA Trans. 53 (2014) 517–523.
by black points and the blue particles then belong to the second [8] J.S. Chiou, S.H. Tsai, M.T.A. Liu, PSO-based adaptive fuzzy PID-controllers, Simul.
subpopulation. At the beginning, all particles are distributed over Model. Pract. Theory 26 (2012) 49–59.
the search space uniformly, and after a few iterations all particles in [9] L.K. Letting, J.L. Munda, Y. Hamam, Optimization of a fuzzy logic controller for
PV grid inverter control using S-function based PSO, Solar Energy 86 (2012)
each subpopulation are separately converging to two different opti- 1689–1700.
mum points (the black particles converging to (−2.7870, 3.1282) [10] S. Panda, B. Mohanty, P.K. Hota, Hybrid BFOA–PSO algorithm for automatic
and the blue particles converging to another point (3, 2), respec- generation control of linear and nonlinear interconnected power systems, Appl.
Soft Comput. 13 (2013) 4718–4730.
tively). This figure sufficiently reveals the convergence dynamics [11] M.R. Esmaili, R.A. Hooshmand, M. Parastegari, P. Ghaebi Panah, S. Azizkhani,
of all particles in each subpopulation. Furthermore, the optimum New coordinated design of SVC and PSS for multi-machine power system Using
points derived by two best particles may not be the same when BF-PSO algorithm, Procedia Technol. 11 (2013) 65–74.
[12] D. Mandal, H. Kumar, S.P. Ghoshal, R. Kar, Thinned concentric circular antenna
the population size is given by N = 30 and N = 40, as shown in
array synthesis using particle swarm optimization, Procedia Technol. 6 (2012)
Figs. 19 and 20, respectively. These show that the distribution of 848–855.
particles on the search space will significantly affect the solutions [13] D. Bratton, J. Kennedy, Defining a standard for particle swarm optimization, in:
obtained. Proceeding of the 2007 IEEE Swarm Intelligence Symposium (SIS 2007), 2007,
http://dx.doi.org/10.1109/SIS.2007.368035
[14] S. Yazdani, H. Nezamabadi-pour, S. Kamyab, A gravitational search algorithm
5. Conclusions and future work for multimodal optimization, Swarm Evol. Comput. 14 (2014) 1–14.
[15] T. Otani, R. Suzuki, T. Arita, DE/isolated/1: a new mutation operator for mul-
timodal optimization with differential evolution, Int. J. Mach. Learn. Cybern. 4
This paper has successfully developed a modified PSO algorithm (2013) 99–105.
with multiple subpopulations for solving the multimodal function [16] N.N. Glibovets, N.M. Gulayeva, A review of niching genetic algorithms for mul-
optimization problem. It begins with splitting up a single popula- timodal function optimization, Cybern. Syst. Anal. 49 (2013) 815–820.
[17] B.Y. Qu, J.J. Liang, P.N. Suganthan, Niching particle swarm optimization with
tion of the algorithm into several subpopulations by the order of local search for multi-modal optimization, Inform. Sci. 197 (2012) 131–143.
particles. In each subpopulation, the best particle with the best [18] K.N. Krishnanand, D. Ghose, Glowworm swarm optimization for simultaneous
cost function needs to be recorded, and this information is uti- capture of multiple local optima of multimodal functions, Swarm Intell. 3
(2009) 87–124.
lized in the modified velocity formula to replace the original global [19] H. Wang, I. Moon, S. Yang, D. Wang, A memetic particle swarm optimiza-
best in the single population. All particles within the same sub- tion algorithm for multimodal optimization problems, Inform. Sci. 197 (2012)
population are updated according to the modified formula. These 38–52.
[20] P.C. Chang, S.H. Chen, C.H. Liu, Sub-population genetic algorithm with mining
best particles can separately converge to various function optima gene structures for multiobjective flowshop scheduling problems, Expert Syst.
including the global and local solutions of the multimodal func- Appl. 33 (2007) 762–771.
tion. Simulation results sufficiently reveal the applicability of the [21] C.C. Chen, Two-layer particle swarm optimization for unstrained optimization
problems, Appl. Soft Comput. 11 (2011) 295–304.
proposed method. In the future work, the modified PSO algo-
[22] A. Kaveh, K. Laknejadi, A new multi-swarm multi-objective optimization
rithm with multiple subpopulations may be applied to solving the method for structural design, Adv. Eng. Softw. 58 (2013) 54–69.
multi-objective optimization problem and the MIMO control sys- [23] C.K. Goh, K.C. Tan, D.S. Liu, S.C. Chiam, A competitive and cooperative
tem designs. co-evolutionary approach to multi-objective particle swarm optimization algo-
rithm design, Eur. J. Oper. Res. 202 (2010) 42–54.
[24] Y. Peng, B.L. Lu, A hierarchical particle swarm optimizer with latin sampling
Acknowledgement based memetic algorithm for numerical optimization, Appl. Soft Comput. 13
(2013) 2823–2836.
[25] M. Jaberipour, E. Khorram, Two improved harmony search algorithms for solv-
This work was partially supported by the Ministry of Science and ing engineering optimization problems, Commun. Nonlinear Sci. Numer. Simul.
Technology of Taiwan under Grant MOST 103-2221-E-366-004. 15 (2010) 3316–3331.