Fuzzy Adaptive Particle Swarm Optimization

Yuhui Shi
EDS Embedded Systems Team
1401 E. Hoffer Street
Kokomo, IN 46902, USA
Y [email protected]
and

Russell C. Eberhart
Department of Electrical and Computer Engineering
Purdue School of Engineering and Technology
799 W. Michigan Street
Indianapolis, IN 46202
[email protected]

particles are manipulated according to the following

equation (Shi and Eberhart I998a, I998b):

Abstract
I n this paper, a fuzzy system is implemented to
dynamically adapt the inertia weight of the particle
swarm optimization algorithm (PSO).
Three
benchmark functions with asymmetric initial range
settings are selected as the test functions. The same
fuzzy system has been applied to all the three test
functions
with
different
dimensions.
The
experimental results illustrate that the fuzzy adaptive
PSO is a promising optimization method, which is
especially useful for optimization problems with a
dynamic environment.

where c I and c2 are positive constants, and rand() and

Rand() are two random functions in the range [O,I]. XI =
(xiI, xi?, ... , xlD)represents the ith particle. Pi= (pll,p12,
... , P , ~ )represents the best previous position (the
position giving the best fitness value) of the ith particle.
The symbol g represents the index of the best particle
among all the particles in the population. V, = (v,,, v , ~ ,
... , vlD) represents the rate of the position change
(velocity) for particle i . Variable w is the inertia weight.

Introduction

Particle s w a m optimization is a population-based

evolutionary algorithm. It is similar to other populationbased evolutionary algorithms in that the algorithm is
initialized with a population of random solutions. It is
unlike most of other population-based evolutionary
algorithms (Goldberg 1989, Fogel 1994, Rechenberg
1994, Koza 1992), however, in that PSO is motivated by
the simulation of social behavior instead of survival of
the fittest, and each candidate solution is associated with
a velocity (Eberhart and Kennedy 1995, Kennedy and
Eberhart 1995, Eberhart, Dobbins and Simpson 1996,
Kennedy 1997). The candidate solutions, called particles,
then fly through the search space. The velocity is
constantly adjusted according to the corresponding
particles experience and the particles companions
experience. It is expected that the particles will move
towards better solution areas. Mathematically, the

0-7803-6657-3/01/$10.00
02001 IEEE

The balance between global and local search throughout

the course of run is critical to the success of an
evolutionary algorithm. In some implementations of
evolutionary programming, balance between global and
local search is obtained through adapting the variance
(strategy parameter) of the Gaussian random function or
step size. Furthermore, in some implementations, even
the strategy parameter is encoded into the chromosomes
to undergo evolution itself (Rochenberg 1994, Saravanan
and Fogel 1996, Yao and Liu 1996 ). In PSO, it is the
inertia weight that is used to balance the global and local
search ability. The inertia weight has characteristics that
are reminiscent of the temperature parameter in the
simulated annealing (Eberhart and Shi 1998). A large
inertia weight facilitates a global search while a small
inertia weight facilitates a local search. By changing the

101

inertia weight dynamically, the search

dynamically adjusted.

ability

paper will be on the population level. To design a fuzzy

system to dynamically adapt the inertia weight, normally,
the inputs to the system are variables that measure the
performance of the PSO, and the output of the system is
the inertia weight or the change of the inertia weight (Shi
2000).

is

Since the search process of the PSO is non-linear and

very complicated, it is hard, if not impossible, to
mathematically model the search process to dynamically
adjust the inertia weight. Instead, a fixed inertia weight
(with or without additional noise) or a linearly
decreasing inertia weight is deployed (Shi and Eberhart
1999, Eberhart and Shi 200 I). By linearly decreasing the
inertia weight from a relatively large value to a small
value through the course of a PSO run, PSO tends to
have more global search ability at the beginning of the
run while having more local search ability near the end
of the run. The simulation results on the benchmark
problems illustrate that an inertia weight starting with a
value 0.9 and linearly decreasing to 0.4 through the
course of the run greatly improve the performance of
PSO (Shi and Eberhart 1998a, 1999).

As in (Shi and Eberhart 2000), two variables are selected

as inputs to the fuzzy system: the current best
performance evaluation and the current inertia weight;
the output variable is the change of the inertia weight.
The current best performance evaluation (CBPE)
measures the performance of the best candidate solution
found so far by the PSO.
Different optimization
problems have different ranges of performance
measurement values. To design a fuzzy system with the
CBPE as one of the inputs to be applicable to a wide
range of optimization problems, the CBPE has to be
converted into a normalized format.
Assume, the
optimization problems to be solved are minimization
problems, and the estimated (or real) minimum is
denoted as CBPE,,,, and the non-optimal CBPE is
denoted as CBPE,,,.
The non-optimal CBPE here
represents that any solution with CBPE greater or equal
to CBPE,,,
is not an acceptable solution to the
minimization problem. The normalized CBPE (NCBPE)
can be calculated as

As mentioned above, the PSO search process is a

nonlinear and complicated process and a linear
decreasing inertia weight approach has a linear transition
o f search ability from global to local search, which does
not truly reflect the actual search process required to find
the optimum. This especially is true for dynamic
optimization problems. For example, tracking problems
(Eberhart and Shi 2001), where the environment itself is
dynamically changed over the time, requires the search
algorithm to have a nonlinear search ability to
dynamically follow the changing environment.
Therefore, for better performance, the inertia weight
should be nonlinearly, dynamically changed to have
better dynamics of balance between global and local
search abilities. Due to the lack of knowledge of the
search process, it is difficult, if not impossible, to design
a mathematical model to adapt the inertia weight
dynamically.
Fortunately, over the years, some
understanding of the PSO search process has been
accumulated, and linguistic description of the search
process is available. This understanding and linguistic
description make a fuzzy system a good candidate for
dynamically tuning the inertia weight of PSO. In (Shi
and Eberhart 2000), a fuzzy system was designed to
adjust the inertia weight of PSO for the Rosenbrock
function. In this paper, the fuzzy system will be further
modified to fit a wider range of optimization problems.

(3)
All three fuzzy variables (two input variables and one
output variable) are defined to have three fuzzy sets:
LOW, MEDIUM and HIGH with associated membership
functions as leftTriangle, Triangle and right Triangle,
respectively. The definitions of these three membership
functions are:

Left-triangle membership function:

Fuzzy System

Triangle membership function:

PSO, like other evolutionary algorithms, can be

dynamically adapted on four different levels:
environment, population, individual, and component
level (Shi 2000). Since the inertia weight is a global
variable in the equation (1) and it is applied to the whole
population, the inertia weight adaptation discussed in this

102

if

x<x,

if

x>x2

List 1 : A description of a fuzzy system for adapting the

inertia weight of PSO.

ij-x < x ,

9
21
NCBPE 3 0 1
IeftTriangle 0 0.06
Triangle 0.05 0.4
rightTriangle 0.3 1

ij-x > x 2

weight 3 0.2 I . 1
IeftTriangle 0.2 0.6
Triangle 0.4 0.9
rightTriangle 0.6 1.1

Right-triangle membership,function:

Io
I'

if X < X I

w-change 3 -0.1 2 0.05

IeftTriangle -0.12 -0.02
Triangle -0.04 0.04
rightTriangle 0.0 0.05

if x > x 2

where x I and x2 are critical parameters which determine

the shape and location of the functions.
Other
membership function definitions are possible, of course,
but the authors have found these to be useful for a variety
of problems and to be easy to be implemented in
microcontrollers and microprocessors. The whole fuzzy
system for dynamically adapting the inertia weight can
be described as in List 1 (Eberhart, Dobbins, and
Simpson 1996, Shi, Eberhart and Chen 1999).

112
121
131
213
222
23 1
313
322
33 1

In List I , the 9 on the first line means that there are 9

rules in the fuzzy system; the second line means that
there are 2 inputs and 1 output in the fuzzy system.
Following the second line is the definitions for the three
variables. The first variable (input) is NCBPE, which
has 3 fuzzy sets with a dynamic range (0,l). Each fuzzy
set is associated with a membership function. The first
one is a left triangle function with two critical parameters
0, 0.06, the second is a triangle function with two critical
parameters 0.05, 0.4, and the final one is a right triangle
with two critical parameters 0.3, 1. These four lines
completely define the first input variable NCBPE (for a
detailed description refer to (Eberhart, Dobbins, and
Simpson 1996, Shi, Eberhart and Chen 1999)). The
second input variable is the current inertia weight named
weight and the output variable is the change (percentage)
of the inertia weight, which is called w-change.
Following the three variable definitions are the
definitions of nine rules in the fuzzy system where the
number 1 encodes fuzzy set LOW, 2 encodes MEDIUM
and 3 encodes HIGH. For example, the first one 1 1 2
represents a fuzzy rule i f NCBPE is Low, and weight is
Low, then the w-change is medium (for details refer to
(Eberhart, Dobbins, and Simpson 1996, Shi, Eberhart
and Chen 1999)).

3 Experimental Setting
For comparison, three benchmark functions reported in
(Angeline 1998a, 1998b, Shi and Eberhart 1999), are
used here. The first function is the Rosenbrock function
described by
n

2 2

f ( x ) l = ~ ( ~ O O ( -~x ij +) ~+ ( x i - 1 )

(4)

i=l

The second function is the generalized Rastrigrin

function described by equation (5):

f*(x) =

E(.: - IOcos(2m,) + 10)

(5)

,=I

The last function is the generalized Griewank function

described by equation (6):

'

103

--cx,'
= 4000

"

-ncos(++l
/=I

J;

The three are all minimization problems. The CBPE,,,

and CBPE,,, for the three functions are listed in Table I .

Table 4: Mean CBPE values for the Rosenbrock function

with linearly decreasing inertia weight.

Following the suggestion in (Fogel and Beyer 1995) and

for the purpose of comparison, the asymmetric
initialization method used in (Angeline 1998a) is adopted
here for population initialization. Table 2 lists the
initialization ranges of the three functions.
Table 1 : CBPE,,, and CBPE,,, values
Function CBPE,,,,
CBPE,,,

f3

' Function
fi

fi
f3

70
0.15

Table 5 : Mean CBPE values for the Rosenbrock function

Asymmetric Initialization Range

( I 5,30)"
(2.56,5.12)"
(300,600)"

As in (Angeline 1998a) , for all the three functions, three

different dimension sizes are tested: IO, 20 and 30. The
maximum number of generations is set as 1000, 1500
and 2000 corresponding to the dimensions I O , 20 and 30,
respectively. In order to investigate whether the PSO
algorithm scales well or not, different population sizes
are used for each function with different dimensions.
They are population sizes of 20, 40, and 80. The general
parameters of PSO are set as: c,=c2=2 for all the PSO
runs. The Vma,and X,,,are listed in Table 3. A total of
50 runs for each experimental setting were conducted.
The fuzzy system is designed to adapt the inertia weight
by the authors according to the authors' understanding
and experience with PSO which is given in List 1 .

Table 3: XmDV
and V,,,
Function
X,,,
100
fi
IO
f2
f?
600

Table 6: Mean CBPE values for the Rastrigrin function

with linearlv decreasing inertia weight.

values
Vmax
100
10

600
Table 7: Mean CBPE values for the Rasgtrigrin function

4 Experimental Results and Discussion

Tables 4, 6 and 8 list the mean CBPE values of the best
particle found for the 50 runs for the three benchmark
functions with a linearly decreasing inertia weight (Shi
and Eberhart 1999), respectively, and tables 5 , 7, and 9
list the mean CBPE values of the best particle found for
the 50 runs for the three benchmark functions using the
fuzzy system described in List I for tuning the inertia
weight.

23.27334
40

80

104

IO
20
30
IO
20
30

1000
1500
2000
1000
1500
2000

3.283368
15.04448
35.20146
2.328207
10.86099
22.52393

Popu. Size
20

40

80

Dim.
IO
20
30
IO
20
30
IO
20
30

Gene.
1000
1500
2000
1000
1500
2000
1000

1500
2000

Mean Best Fitness

0.0919
0.0303
0.0182
0.0862
0.0286
0.0127
0.0760
0.0288
0.0128

Table 9: Mean CBPE values for the Griewank function

with the fuzzy system tuning the inertia weight

I 30

2000

I 0.014945

By comparing the results in Tables 4, 6, and 8 with the

results in Tables 5 , 7, and 9, respectively, it is easy to
see that by using a fuzzy adaptive inertia weight the
performance of PSO can be improved and have similar
or better results than that of both PSO with a linearly
decreasing inertia weight and the evolutionary
programming method reported in (Angeline 1998a, Shi
and Eberhart 1999). From the tables, it is also clear that
PSO with different population sizes has reasonably
similar performance. This feature is extremely important
when the PSO is applied to real-time online problems
where the time is critical and the solution evaluation
consumes a significant amount of computation time.
Furthermore, it can be seen from the tables that the
performance of the PSO does not degrade a lot as the
problem dimension scales up.

5 Conclusion
In this paper, a fuzzy system is implemented to
dynamically adjust the inertia weight to improve the
performance of the PSO. Three benchmark functions
have been used for testing the performance of the fuzzy
adaptive PSO.
For comparison, simulations are
conducted for both the fuzzy adaptive PSO and the PSO
with a linearly decreasing inertia weight. The simulation
results illustrate the performance of PSO is not sensitive
to the population size, and the scalability of the PSO is
acceptable. The results also illustrate that PSO with a
fuzzy system tuning its inertia weight can improve its

performance to some extent, at least for the three

benchmark functions. More simulations are required to
be run on other dynamic functions to illustrate the
benefits and effectiveness of using a fuzzy adaptive PSO.

References
Angeline, P. J. ( 1 998a). Evolutionary optimization
versus particle swarm optimization: philosophy and
performance difference. 1 998 Annual Conference on
Evolutionary Programming, San Diego.
Angeline, P. J. (1998b). Using selection to improve
particle swarm optimization.
IEEE International
Conference on Evolutionary Computation, Anchorage,
Alaska, May 4-9, 1998.
Eberhart, R. C., Dobbins, R. W., and Simpson, P.
( 1 996), Computational Intelligence PC Tools, Boston:
Academic Press.
Eberhart, R. C., and Kennedy, J. (1995). A new
optimizer using particle swarm theory. Proc. Sixth
International Symposium on Micro Machine and Human
Science (Nagoya, Japan), IEEE Service Center,
Piscataway, NJ, 39-43.
Eberhart, R. C., Shi, Y. H. (1998). Comparison
between genetic algorithms and particle swarm
optimization. 1998 Annual Conference on Evolutionary
Programming, San Diego.
Eberhart, R. C., Shi, Y. (2001). Tracking and
optimizing dynamic systems with particle swarms. Proc.
Congress on Evolutionary Computation 2001, Seoul,
Korea. Piscataway, NJ: IEEE Service Center. (in press)
Fogel, D., Beyer H.-G. (1995). A note on the
empirical evaluation of intermediate recombination.
Evolutionary Computation, vol. 3, no. 4.
Fogel, L. J. ( I 994). Evolutionary Programming in
Perspective: the Top-down View, in Computational
Intelligence: Imitating Life, J.M. Zurada, R. J. Marks 11,
and C. J . Robinson, Eds., IEEE Press, Piscataway, NJ.
Goldberg, D. E. (1 989). Genetic Algorithms in Search,
Optimization, and Machine Learning, Reading MA:
Addison- Welsey.
Kennedy, J. (1997). The particle swarm: social
adaptation of knowledge.
Proc. IEEE International
Conference on Evolutionary Computation (Indianapolis,
Indiana), IEEE Service Center, Piscataway, NJ, 303-308.
Kennedy, J., Eberhart, R. C. (1995). Particle swarm
optimization. Proc. IEEE International Conference on
Neural Networks (Perth, Australia), lEEE Service Center,
Piscataway, NJ, pp. IV: 1942- 1948.
Koza, J. R. (1992). Genetic Programming: On the
Programming of Computers by Means of Natural
Selection, MIT Press, Cambridge, MA.
Rechenberg, I. (1 994). Evolution Strategy, in
Computational Intelligence: Imitating Life, J . M. Zurada,
R. J. Marks 11, and C. Robinson, Eds., IEEE Press,
Piscataway, NJ.

Saravanan, N., Fogel, D. (1996). An empirical

comparison of methods for correlated mutations under
self-adaptation. The Fifth Annual Conference on
Evolutionary Programming.
Shi, Y. H. (2000). Fuzzy Adaptive Evolutionary
Computation: A Review, The 4thWorld Multiconference
on Systemics, Cybernetics and Informatics, Orlando,
Florida, USA, July 23-26, 2000
Shi, Y. H., Eberhart, R. C. (1998a). Parameter
selection in particle swarm optimization. 1998 Annual
Conference on Evolutionary Programming, San Diego,
March 1998.
Shi, Y. H., Eberhart, R. C., (1998b). A modified
particle swarm optimizer.
IEEE International
Conference on Evolutionary Computation, Anchorage,
Alaska, May 4-9, 1998.
Shi, Y. H., Eberhart, R. C. (1999). Empirical study of
particle swarm optimization, I999 Congress on
Evolutionary Computation, Washington DC, USA, July
6-9.
Shi, Y. H., Eberhart, R. C. (2000) Experimental Study
of Particle Swarm Optimization, The 4'h World
Multiconference on Systemics, Cybemetics and
Informatics, Orlando, Florida, USA, July 23-26, 2000
Shi, Y. H., Eberhart, R. C., and Chen Y. (1999).
Implementation of Evolutionary Fuzzy Systems, IEEE
Transactions on Fuzzy Systems, Vol. 7, No. 2.
Yao, X., Liu, Y. (1996).
Fast evolutionary
programming.
The Fifth Annual Conference on
Evolutionary Programming.