See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266283526

Virus Optimization Algorithm (VOA): A Novel

Metaheuristic for Solving Continuous
Optimization Problems

Article

CITATIONS READS

3 330

5 authors, including:

Yun-Chia Liang
Yuan Ze University
73 PUBLICATIONS 2,236 CITATIONS

SEE PROFILE

All content following this page was uploaded by Yun-Chia Liang on 06 October 2015.

The user has requested enhancement of the downloaded file.

 APIEMS2009 Dec. 14-16, Kitakyushu

Virus Optimization Algorithm (VOA): A Novel Metaheuristic for

Solving Continuous Optimization Problems
Josue Rodolfo Cuevas Juarez1, Hung-Jie Wang2, Yi-Chieh Lai3, Yun-Chia Liang†4
Department of Industrial Engineering and Management
Yuan Ze University, Chungli, Taiwan 320, R.O.C.
Email: [email protected]
[email protected]
[email protected]
[email protected]

Abstract. The virus, an infectious agent that can reproduce only inside a host cell can apparently spread
without any control, but as we already know that the cells without any protection will tend to give better
chances to the virus in the reproduction activity. In this paper, we develop a novel metaheuristic, named Virus
Optimization Algorithm (VOA) which imitates the behavior of the virus. The host cell represents the entire
search space while the virus reproduction denotes the generation of new solutions. VOA is a population-based
method that begins the search with a small set of solutions and the solution size will grow at each iteration until
the stopping criterion is reached. We compare this new metaheuristic algorithm with some widely used
algorithms such as Genetic Algorithm (GA), Simulated Annealing (SA) and Particle Swarm Optimization (PSO).
As a conclusion from the test results show the proposed VOA to be a competitive and robust tool for solving
continuous optimization problems.

Keywords: Virus Optimization Algorithm, Continuous Optimization Problem, Metaheuristic.

1. INTRODUCTION whole solution space) and the exploitation (searching better

solutions in a specific region of the solution space). As well,
The word ‘ metaheuristic’combines the Greek prefix the level of a problem is very high constrained and difficulty
“ meta”(meaning “ higher level”) and “ heuristic”(meaning to solve, the efficiency of a metaheuristic may be
“ to discover”
). In academia, a metaheuristic often refers as a compromised. For that, metaheuristics often need to find
heuristic method for solving combinatorial optimization acceptable solutions within a reasonable amount of
problems. Since real-world optimization problems are computational time.
complex, metaheuristics stand out to be a promising approach Metaheuristic techniques have found wide applications
to solve computational complex problems. Among various in most areas of engineering. These techniques have also
metaheuristics been developed, some of them are population- been applied in business, financial and economic modeling.
based (multiple solutions are managed with the heuristic tool) Among these widely used approaches, all have shown their
and single solution based (only one solution is managed at a good performance working with high dimensional problems.
time by the algorithm). Genetic Algorithm (GA), Particle The key ingredients relevant to metaheuristics
Swarm Optimization (PSO), and Ant Colony Optimization development must be addressed. A suitable representation
(ACO) are considered population-based while Simulated for the candidate solution must be chosen, and a fitness (or
Annealing (SA) and Tabu Search are single solution based. cost) function must be derived to yield numeric values for the
Usually the answer from this kind of methods is close to the fitness (or cost) of each candidate solution. The notion of a
optimal solution (which means the optimal solution neighboring value must be defined to add new materials.
sometimes cannot be guaranteed.) Normally, the search space will be multi-dimensional and a
As commonly known, good metaheuristic tools need to similar arbitrary decision must be made about how many
balance among algorithm complexity, problem difficulty, values to change and in what way.
computation time, and solution quality. For instances, the The iterative nature of metaheuristic algorithms suggests
complexity of an algorithm sometimes causes the tradeoff that metaheuristics may be used to automate tedious and
between the exploration (searching for new solution in the difficult tasks thereby saving effort, time and money; they are
______________________
†Corresponding Author

2166
 APIEMS2009 Dec. 14-16, Kitakyushu

therefore potentially valuable and worthy. Also, the features At the first iteration the viruses (solutions) are created
which make metaheuristics applicable for real-life randomly. Viruses can only survive within the cell, i.e., only
applications suggest that there is great potential for feasible solutions that satisfy variable ranges will be
exploitation of new technologies. generated. Figure 1 illustrates the generation of initial
In this paper we introduce a new metaheuristic tool to solutions in VOA. The black dots in Figure 1 are the viruses
solve continuous optimization problems; the method has been generated and the white regions represent the good regions
named as Virus Optimization Algorithm (VOA) which (local or global optima). The yellow region is the host cell
simulates the behavior of the viruses when they attack a host and the blue region is the body liquid surrounding the cell
cell. The searching procedure of this method mainly relies where the virus will not survive.
on the generation of new solutions. The source and number Once the initial group of viruses is generated, their
of new solutions can be determined by the users. For ability of spreading will be determined where they are
example, the user may preset three new solutions coming located, i.e. the objective function value will be evaluated.
from the best member of the population and two new The viruses located in the “ good”regions (i.e., the weaker
solutions coming from the other members, etc. regions without antibodies or immunity) will replicate
VOA is compared with three widely used algorithms themselves at those regions trying to generate better viruses.
GA (Goldberg et al. 1989), PSO (Eberhart et al. 1995, The purpose of replication represents the intensive search
Kennedy et al. 1997) and SA (Kirkpatrick et al. 1983), and within the identified good regions. On the other hand, the
shows a better performance in terms of convergence at the virus located in “ bad”region (i.e., the healthy regions with
optimality. The benchmark functions tested consist of the the protection of antibodies or the immune system) will be
Zakharov, Rastrigin and Schwefel multimodal functions also reproduced. The difference between the replication on
(described in Section 4) which are well known because of “ good”and “ bad”regions is that spreads in bad regions will
their complexity when the dimension is increased. be quite limited because of the existence of antibodies or
The rest of this paper is arranged in the following order: immune systems.
Section 2 briefs on the concept of our Virus Optimization A maintenance mechanism named as antivirus is
Algorithm; Section 3 introduces the framework of this employed to limit the growing rate of viruses. The user can
methodology; Section 4 shows computational results; Section activate this mechanism to maintain a fixed number of
5 provides conclusions and suggestions. viruses created i.e. the population size will be same after the
replication process, where only the best viruses will survive
2. THE VIRUS OPTIMIZATION ALGORITHM (defined by their objective function value). On the other
(VOA) hand, if the antivirus is not activated the number of viruses
will grow after each replication; therefore after several
Inspired by the behavior of virus, our algorithm (VOA) iterations have been performed the area being covered by the
is a population-based approach and is designed to solve viruses will be very large (as we can observe in Figure 2).
continuous domain problems. In a cell, the condition of each Hence, the use of the antivirus mechanism is to control the
sub-region can be classified according to the health condition, reproduction so that the size of the virus matrix in the
i.e. the objective function value in the sense of optimization. computer won’ t be too large and too hard to manage.
The weak regions will then be good candidates for viruses to
grow (or spread) by generating more solutions (viruses).
Moreover, the range for such spread will be set by the user as
the maximum range of spreading.

Figure 2: Solutions generated after performing the replication

process

When the viruses are close to the optimal solution, it is

important to notice that the change in the values of the
variables cannot be too large, to avoid overshooting the
optimal solution. Therefore, the algorithm will automatically
Figure 1: Illustration of initial solution generations
decrease the range that will generate new solutions from
viruses at good regions in a small amount (defined by the

2167
 APIEMS2009 Dec. 14-16, Kitakyushu

user that could be 10% of the range amount), however, the In addition, several controllable parameters are
user is able to decide if reduce or not the range that will described as follows:
generate new solutions from the viruses at bad regions (this 1. The number of viruses at the very beginning of the
of course will affect the exploration, but, on the other hand program.
will create new viruses not too far away, intensifying the 2. Number of best solutions (viruses) to consider in
exploitation). After several replications the range(s) will be the replication at good regions.
small enough to generate solutions at the optima or near to it. 3. The growing rate or new solutions (viruses) to be
The reduction will depend on a “ Pe r
sistencec ounter” generated during the replication from the viruses
(variable defined by the user that will check for how long the found at good regions.
current solution has not been updated). 4. The range used to generate the new solutions,
determined by the range of the decision variables.
Say if X1 and X2 are between -5.12 and 5.12 the
3. METHODOLOGY
new solutions will be between those values, hence,
Figure 3 below shows the flowchart of the VOA where a range “ sma l
len ough ”toc rea t
et hen e ws ol
utions
h represents the number of iterations, H denotes the inside the good regions should be considered, of
maximum number of iterations, k is persistence counter of c ourset het erm“ sma l
l”de pendsont h epr oblem
the current solution i.e. the current solution has not been that we are trying to solve.
improved after replication, and K represents the maximum 5. The growing rate or new solutions to be generated
number of consecutive non-improving iterations (when the when the bad region has been found by the viruses.
persistence counter reaches this value the range(s) will be This value should be smaller in magnitude
decreased). compared with the growing rate of viruses at good
regions,s inc ewedon’ twa ntt os pe ndt oomu ch
S ta rt
effort generating new solution on non-promising
regions. However, the value should not be zero
Set
P a ra m e te rs because we still want to explore other areas to
know if there are better places in the solution space.
C re a te P o in ts
6. The range to generate new solutions at bad regions
h=1 should be a small amount of the variables domain
(or just defined by the domain of the decision
C a lc u la te th e
fitn e s s variables), this will allow us to explore in the
S o rtin g b y whole solution space (host cell). The value of this
fitn e s s v a lu e
range should be greater than the value of the range
used to generate new solution at good regions,
B e s t S o lu tio ns O th e r S o lu tio n s
since we want to explore using the viruses at bad
G e n e ra te n e w s o lu tio n fo r g o o d re g io n G e n e ra te n e w s o lu tio n fo r B a d
regions and exploit using the viruses at good
re g io n s
X i + - ran d () * ran g e o f viru se s at
X i + -ra n d () * ran g e o f v iru se s at
regions.
g o o d re g io n s
b a d re g io n s 7. Decrement percentage to be used when we reduce
the range to generate new solution coming from
K ill b a d s o lu tio n s
Is th e YES
a n d k e e p th e
viruses at good regions (this decrement could be
a n tiv iru s
p o p u la tio n s iz e
a c tiv a te d
fix e d
different if today we decide to reduce the range to
NO
generate new solutions coming from the viruses at
bad regions).
C a lc u la te fitn e s s
V a lu e fo r n e w 8. The following stopping criteria can be used in
viru se s
VOA:
S o rtin g b y
fitn e s s v a lu e
a. No improvement after several iterations,
Is c u rre n t YES
b. Maximum number of iterations,
s o lu tio n th e
s a m e a s la st
k = k+ 1 c. The optimal solution is found.
o n e?

NO
R e d u c e th e YES
k> K
4. COMPUTATIONAL RESULTS
ra n g e

NO The proposed VOA is coded in Visual studio C# and run

h =h+1
NO
h < H
on Intel (R) Core (TM)2 Duo CPU E6750@ 2.66Hz and 2
YES GB RAM. Our VOA is compared with other metaheuristic
End algorithms such as SA, PSO and GA. Then, the best setting
of each algorithm are first run under different parameter (see
Table 2 to Table 5).
Figure 3: The flowchart of VOA

2168
 APIEMS2009 Dec. 14-16, Kitakyushu

The functions tested in this study consist of three well- Table 3: The Best Setting of GA (arithmetic crossover,
known multimodal functions: Zakharov (Fan et al. 2004, random change for the mutation)
Chelouah et al. 2000), Rastrigin (Liang et al. 2006) and Function parameters Function
Schwefel (Liang et al. 2006). Their functions, range, and for GA Zakharov Zakharov Rastrigin Rastrigin Schwefel Schwefel
global optimum are as follows: Dimension 2 300 2 300 2 300
Probability of Crossover 0.85 0.95 0.95 0.95 0.95 0.85
Probability of Mutation 0.3 0.5 0.3 0.2 0.3 0.5
Zakharov function: Population Size 100 500 100 500 100 500
Number of Iterations 100 10000 100 10000 100 10000

Table 4: The Best Setting of VOA

Function parameters Function
(1) for VOA Zakharov Zakharov Rastrigin Rastrigin Schwefel Schwefel
Dimension 2 300 2 300 2 300
Initial Solutions 100 500 100 500 100 500
Global minimum Z* = 0 at point x* = (0, 0, …, 0) Solutions to Consider 8 40 7 40 7 45
The Rate of Good Solutions 7 15 7 20 7 15
The Rate of Bad Solutions 4 8 4 10 4 8
Rastrigin function: The Range of Good Solutions 2 1.5 2 2 2 2
The Range of Bad Solutions 4 3.5 4 4 4 4
Number of Iterations 100 10000 100 10000 100 10000
Antivirus (On/Off) On On On On On On
Persistence Counter 7 20 7 20 7 20
Reduction Rate of Good Solutions 0.7 0.7 0.7 0.7 0.7 0.7
Reduction Rate of Bad Solutions 0.9 1 0.9 0.9 0.9 1
(2)
Table 5, where we can observe that for low and high
Global minimum Z* = 0 at point x* = (0, 0, …, 0) dimensional problems our algorithm performed the best in
the 3 functions tested, showing the best objective function
Schwefel function: values.
Table 5: Output results for each algorithm
Problem definition Algorithm
Result
Function Dimension SA PSO VOA GA
Zakharov 2 0.00002 0.00000 0.00000 0.00000
(3) Average Rastrigin 2 0.00082 0.14544 0.00001 1.83979
objective Schwefel 2 830.07500 830.07500 0.00019 111.46679
Global minimum Z* = 0 at point x* = (420.9687, 420.9687, …, function Zakharov 300 4036.20000 13700.80000 1008.51500 1620.46009
value
Rastrigin 300 5103.27000 5932.02000 977.87720 1546.30502
420.9687)
Schwefel 300 125425.20000 124956.00000 51165.77000 68380.47020
Zakharov 2 0.00001 0.00002 0.00300 0.02000
Table 1: The Best Setting of SA Rastrigin 2 0.00010 0.00020 0.00330 0.02300
Function Average
Function parameters Schwefel 2 0.00011 0.00019 0.00350 0.02530
CPU time
for SA Zakharov Zakharov Rastrigin Rastrigin Schwefel Schwefel Zakharov 300 236.00000 170.00000 3300.00000 2700.00000
(seconds)
Dimension 2 300 2 300 2 300 Rastrigin 300 260.00000 379.00000 3420.00000 2880.00000
Number of state 100 10000 100 10000 100 10000 Schwefel 300 226.00000 348.00000 3300.00000 2940.00000
Number of movement 100 500 100 500 100 500
Initial Temperature 125 1000 120 1200 125 1100
Alpha 0.9 0.9 0.9 0.9 0.9 0.99 The behaviors in terms of the mean values for different
Table 2: The Best Setting of PSO, using the inertia weight objective functions are shown in Figures 4 to 9 below. We
rule for the update observed that our algorithm showed a very fast convergence
Function parameters Function at good objective function value (most of the time close or at
for PSO Zakharov Zakharov Rastrigin Rastrigin Schwefel Schwefel the optima) compared with the others, especially with high
Dimension 2 300 2 300 2 300
dimension problems.
Number of Iterations 100 10000 100 10000 100 10000
Population Size 100 500 100 500 100 500 When the parameters are set in a suitable value for a
Weight 0.8 0.1 0.1 0.3 0.2 0.3 rather large population size, setting larger number of
C1 2 2 1 0.5 2 0.5
C2 2 4 5 0.5 5 1
iterations to find optimal (or close to optimal) does not seem
necessary.

2169
 APIEMS2009 Dec. 14-16, Kitakyushu

In addition, the range of bad or good solutions will be

directly dependent on the objective function domain. The
probability to reach a better result by using viruses at good
regions to improve is higher compared with using those at
bad regions; hence the growing rate for viruses at good
regions should be higher than viruses at bad regions. If we
have more than one virus at good regions to consider, the
probability to get a better objective function value may be
increased,h owe v erifitistool argewedon’ th av eanyc han ce
to perform the exploration; and the probability to get stuck in
a local optimum is higher.
In terms of CPU time VOA is the worst compared with
others by using the same population size (or number of
movement in case of SA) and number of iterations (or Figure 6: Behavior comparison between algorithms for the
number of state for SA); however, VOA showed the best Schwefel Function (2-dimension problem).
performance in terms of the number of iterations needed to
reach the optimal value and average objective function value.

Figure 7: Behavior comparison between algorithms for the

Zakharov Function (300-dimension problem).
Figure 4: Behavior comparison between algorithms for the
Zakharov Function (2-dimension problem).

Figure 8: Behavior comparison between algorithms for the

Figure 5: Behavior comparison between algorithms for the Rastrigin Function (300-dimension problem).
Rastrigin Function (2-dimension problem).

2170
 APIEMS2009 Dec. 14-16, Kitakyushu

Zakharov function ( 2 dimensions )

0.00006

0.00005

Best objective function values

0.00004

0.00003

0.00002

0.00001

0.00000

SA PSO VOA GA

Figure 10: The Box Plot for Zakharov function

(2-dimension problem)
Figure 9: Behavior comparison between algorithms for the
Schwefel Function (300-dimension problem).
Table 7: 10-run results for Zakharov function
A box plot in terms of the average objective function (300-dimension problem)
value will help us to visualize how good VOA is solving Run number Algorithm
continuous domain problems, so the results can be seen at (Zakharov 300 dimensions) SA PSO VOA GA
Tables 6 to 11 and Figures 10 to 15, respectively. 1 3738.680000 13653.300000 1007.700000 1991.585000
2 4203.820000 14111.700000 1076.330000 1515.008000
VOA showed the best performance in terms of the mean
3 3981.180000 13903.300000 938.291000 1373.730000
objective function value, but, overall, the most interesting 4 3946.750000 14248.700000 1039.500000 1306.458000
part coming from the box plots is the quality of the solutions, 5 4011.180000 14074.400000 971.587000 1823.482000
since the variation in the objective function when VOA 6 4188.410000 13477.500000 913.088000 1266.663000
converges is very small, hence, from the results below we 7 4150.080000 13359.400000 925.310000 1783.877000
can conclude that VOA will not only give us a good objective 8 4146.870000 13733.900000 958.532000 1667.377000
9 3908.530000 14252.200000 979.572000 1750.408000
function value, also the robustness of the algorithm will
10 4138.520000 13335.300000 1275.240000 1522.482000
provide a good solution quality at the optima or near to it.

Table 6:10-run results for Zakharov function Zakharov function ( 300 dime nsions )
(2-dimension problem) 16000

14000
Run number Algorithm
(Zakharov 2 dimensions) SA PSO VOA GA 12000
Best objective function values

1 0.000029 0.000001 0.000005 0.000000 10000

2 0.000029 0.000001 0.000001 0.000000
8000
3 0.000019 0.000000 0.000000 0.000000
4 0.000058 0.000000 0.000001 0.000000 6000

5 0.000004 0.000001 0.000003 0.000000 4000

6 0.000041 0.000019 0.000000 0.000000
2000
7 0.000005 0.000001 0.000001 0.000000
8 0.000032 0.000000 0.000000 0.000000 0
SA PSO VOA GA
9 0.000034 0.000011 0.000000 0.000000
10 0.000027 0.000000 0.000000 0.000000
Figure 11: The Box Plot for Zakharov function
(300-dimension problem)

2171
 APIEMS2009 Dec. 14-16, Kitakyushu

Table 8:10-run results for Rastrigin function R astrigin ( 300 dime ns ions )
(2-dimension problem)
6000

Run number Algorithm

5000

Bes t objective function values

(Rastrigin 2 dimensions) SA PSO VOA GA
1 0.001771 0.000000 0.000000 4.322736 4000
2 0.003135 0.000776 0.000003 1.999665
3 0.000094 0.077928 0.000001 1.999665 3000
4 0.001096 0.000336 0.000001 1.999665
5 0.003074 0.798732 0.000001 3.522586 2000
6 0.000614 0.000016 0.000000 1.999665
7 0.001034 0.000000 0.000049 1.839785 1000

8 0.001551 0.995760 0.000001 2.039665

SA PSO VOA GA
9 0.000977 0.000000 0.000000 0.898529
10 0.000788 0.000718 0.000001 0.000000
Figure 13: The Box Plot for Rastrigin function
(300-dimension problem)
Rastrigin ( 2 dimensions )
Table 10: 10 times run for Schwefel function
(2-dimension problem)
Best objective function values

Run number Algorithm

(Schwefel 2 dimensions) SA PSO VOA GA
1 434.280000 830.075000 0.000002 0.000003
2 335.579000 830.075000 0.000107 157.487300
3 513.246000 830.075000 0.000001 7.647602
4 236.878000 830.075000 0.001545 231.829300
5 217.140000 830.075000 0.000083 241.4322
6 335.578000 830.075000 0.000005 73.965880
7 572.488000 830.075000 0.000000 129.281800
8 236.877000 830.075000 0.000000 119.082500
9 455.534000 830.075000 0.000002 273.841800
10 118.439000 830.075000 0.000030 84.009700
Figure 12: The Box Plot for Rastrigin function
(2-dimension problem)

Table 9:10-run results for Rastrigin function Schwe fe l ( 2 dime ns ions )

(300-dimension problem) 900

800

Run number Algorithm 700

Best objective function values

(Rastrigin 300 dimensions) SA PSO VOA GA

600
1 5108.550000 6011.150000 912.375000 1592.798000
500
2 5239.520000 5787.620000 865.612000 1499.956000
400
3 5389.160000 5859.320000 926.304000 1584.238000
300
4 5572.030000 6041.320000 907.400000 1503.855000
5 5502.400000 5837.490000 931.279000 1584.001000 200

6 5563.120000 5849.640000 986.001000 1510.376000 100

7 5498.080000 5709.990000 947.198000 1580.888000 0

8 5538.500000 6096.690000 828.798000 1510.951000 SA PSO VOA GA
9 5478.980000 5844.630000 944.213000 1578.840000
10 5486.160000 6078.570000 869.592000 1517.957000
Figure 14: The Box Plot for Schwefel function
(2-dimension problem)

2172
 APIEMS2009 Dec. 14-16, Kitakyushu

5.2 Suggestions
Table 11: 10 times run for Schwefel function
(300-dimension problem) The idea of using VOA to solve discrete domain
Run number Algorithm problems, such as TSP, MDVRP, and etc. is an issue to be
(Schwefel 300 dimensions) SA PSO VOA GA considered in the future.
1 155512.200000 124979.000000 56643.000000 68161.050000 Measure the performance of VOA by turning off the
2 165798.900000 124979.000000 52397.700000 67394.060000 antivirus, is another problem to be considered. Of course,
3 157596.000000 124979.000000 49980.500000 71045.020000
4 154571.700000 124979.000000 54512.400000 67946.590000
this could be a problem in terms of the computational effort;
5 148360.800000 124979.000000 49553.000000 68765.190000 however, the optimal solution (or near optima) may be
6 151243.200000 124979.000000 52692.700000 68006.350000 reached faster.
7 168358.500000 124979.000000 45645.200000 67526.230000 Lastly, reduction of the CPU time and number of
8 159698.100000 124979.000000 48706.900000 69715.500000
parameters are also topics to discuss in the future.
9 165914.700000 124979.000000 51369.900000 67171.190000
10 171054.000000 124979.000000 50246.500000 67614.470000
ACKNOWLEDGMENT

This paper was partially supported by the TaiwanICDF

Schwefel (300 dimensions)
(International Cooperation and Development Fund). The goal
175000
of ICDF is to strengthen international cooperation and
enhance foreign relations by promoting economic
Best objective function values

150000
development, social progress and the welfare of the people in
125000 partner nations around the world.
The TaiwanICDF's core competencies include technical
100000
assistance, investment and lending operations, education and
training, and humanitarian assistance. Technical cooperation
75000
programs managed by the TaiwanICDF are structured to
50000
ensure that their work can be extended during
implementation and replicated elsewhere upon completion.
SA PSO VOA GA
TaiwanICDF education and training programs, often
Figure 15: The Box Plot for Schwefel function developed in conjunction with government agencies,
(300-dimensions problem) nongovernment organizations and educational institutions,
have developed human resources in nations around the world.
5. CONCLUSIONS AND SUGGESTIONS Thanks to them we can give something back to our
society, a virus optimization algorithm that will help to solve
5.1 Conclusions
problems in the future and promises to be good enough to be
applied in several fields such as: transportation, production
In this study, we developed a new algorithm, called
and scheduling, economy, etc.
VOA, a population-based approach. The VOA simulates the
behavior of the viruses when they attack a host cell. Our
implementation shows that the VOA is sensitive to the
REFERENCES
parameter values.
As to the CPU time, for low dimensional problems the Chelouah, R. and Siarry, P. (2000) A Continuous
objective function value is reached very fast with a suitable Algorithm Designed for the Global Optimization of
setting of the parameters; however, if we cannot reach the Multimodal Functions, Journal of Heuristics, 6, 191-213.
value there are strong evidence to say that the output result Eberhart, R. C. and Kennedy, J. (1995) Particle Swarm
will be close to the optimal value. In high dimensional Theory. Proceedings of the IEEE International Conference
problems the results of VOA is much better compared with on Neural Networks, IV, 1942-1948.
others. Eberhart, R. C. and Kennedy, J. (1995) A New
Overall, VOA shows the best performance not only in Optimizer Using Particle Swarm Theory. Proceedings of the
low dimensional problems but also in high dimensional Sixth International Symposium on Micro and Human Science,
problems. The idea of the algorithm is very easy to IEEE Service Center, Piscataway, NJ, Nagoya, Japan, 39-43.
understand but, the value of the parameters is very hard to Fan, S. K., Liang, Y. C., and Zahara, E. (2004) Hybrid
decide since is highly dependent on the problem to be solved. Simplex Search and Particle Swarm Optimization for the
Global Optimization of Multimodal Functions. Engineering
Optimization, 36(4), 401-418.

2173
 APIEMS2009 Dec. 14-16, Kitakyushu

Goldberg, D. E. (1989) Genetic Algorithm in Search,

Optimization and Machine Learning. Addison-Wesley,
Reading, MA, 125-129.
Kennedy, J. and Eberhart, R. C. (1997) A Discrete
Binary Version of the Particle Swarm Optimization.
Proceedings of IEEE International Conference on Neural
Networks, V, 4104-4108.
Liang, J. J. and Qin, A. K. (2006) Comprehensive
Learning Particle Swarm Optimizer for Global Optimization
of Multimodal Functions. IEEE Transaction on Evolutionary
Computation, 10(3), 281-295.
Kirkpatrick, S. Gelatt, Jr. C. D., and Vecchi, M. P.
(1983) Optimization by Simulated Annealing, Science, New
Series, 220, 671-680

2174

View publication stats