Particle Swam Optimization with Particles
Having Quantum Behavior
Jun Sun Bin Feng Wenbo Xu
School of Information School of Information School of Information
Technology Technology Technology
Southem Yangtze University Southem Yangtze University SouthemYangtze University
Wuxi, Jiangsu 214036 China Wuxi, Jiangsu 214036 China Wuxi, Jiangsu 214036 China
Email: suniun wxE!hotmail.com Email: jsuCenu2003(daj.vahoo.com Email: xwbfilsvtu.edu
Abstrael-In this paper, inspired by the analysis of are described in Section 4. And the paper is concluded in
eonvergeoee of PSO, we study the individual particle of B Section 5.
Pso system moving in P quantum mullidimensional space
and establish a quantum Delta potential well model for PSO.
After that, a trial method O f parameter control and QDPSO is 11. QUANTUM DELTA POTENTIAL WELL MODEL
proposed. The experiment result shows much advantage of OF PSO
QDPSO to the traditional PSO.
A. Dynomics ofClussicul PSO
I. INTRODUCTION
In a classical PSO system with the population size of M,
Particle Swarm Optimization (PSO) algorithm, the position vector and velocity vector of particle i in
originally introduced by Kennedy and Eberhart in 1995 D-dimensional space are denoted as
[3], is a population-based evolutionary computation
technique. It is motivated by the behavior of organisms
such as fishing schooling and bird flock. In a PSO system, %(t) = ( X i , ( t ) , x i 2 ( t ) , " ' 1 XiD (t))
each particle corresponding to individual of the organism and
is a candidate solution to the problem at hand. Particles of
the population fly around in a multi-dimensional search "(t) = ( V j l ( ~ ) , " j ~ ( ~ ) , ", 'V i D ( t N T
space, to find out an optimal or sub-optimal solution by
competition as well as by cooperation among them.
Analyzing the trajectories of individual particles has And the trajectory of the particle is formulated by the
following equation
opened outthe mechanism of the PSO IS, 61. In terms of
classical mechanics, a particle is depicted by its position
vector ? . and velocity vector G I which determine the ?.(f +l)=wij(t)+qf@i-ij(f))t$qT@g -j5(f))
trajectory of the particle. The particle moves along a
(la)
determined trajectory in Newtonian mechanics, but this is
not the case in quantum mechanics. In quantum world, the
term trajectory is meaningless, because 2 and G of a
Xi(t + 1) = Zi(t) + v,(t +1) (Ih)
particle can not be determined simultaneously according (i = (1,2,...,M)
to uncertain@ principle. Therefore, if individual particles
in a PSO' system have quantum behavior, the PSO T T
algorithm is bound to work in a different fashion. where (Dl and p2 are random vectors denoted by
In this paper, keeping to philosophy of PSO, we present
a Delta potential well model of PSO in quantum world T
and show how it works. We also give a method of 471 = (4711.47,2>...>471D) (W
parameter control, and consequently propose Quantum 'T
Delta-Potential-Well-based Particle Swarm Optimization 472 = (47214722>...,47ZD) (2b)
(QDPSO) algorithm. For comparison with classical PSO (Pk,> 0, = 1,2)
algorithm, seven benchmark functions are used to test the
performance.of the QDPSO algorithm. The results show
that QDPSO supplies much advantage to classical PSO. In equation (la), vecto7 p i = (pil,pi2,
...,pa)
The paper is structured as follows. In next section, we is the hest position (the position giving the best fitness
will present the quantum Delta potential well model of value) of particle i, and vector p, = O.,,,pp2,. ..,pro) is
PSO. In Section 3, we propose a method of parameter
control andthe QDPSO algorithm. Experiment procedures the position of the best particle among all the particles in
the population. Parameter w is the inertia weight [I], 121,
0-7803-85 I5-2/04/$20.00 02004 IEEE 325
which does not appear in the original version of PSO [3]. This equation is called the time-dependent Schrodinger
In [5], M. Clerc and I. Kennedy prove that if the upper
equation. The operator H is the Hamiltonian operator.
limits of pu is properly selected, Zj will converge to For a single particle of mass m in a potential field V ( 2 ),
p = ( p , , p 2 ; . . p D )whosecoordinatesare
, it is given by
pd = ( P l d P i d + V ) Z d P g d ) / ( V ) l d + p 2 d ) (3) fi2
A
H = --V2 +V(2) (7)
2m
The particles moves around and careen toward point p
with its kinetic energy declining to zero, like a rehuning Constant fi in (9,(6) and (7) is Planck Constant.
satellite orbiting the earth. From the point view of
Now, we hypothesize that the PSO system is a quantum
dynamics, to avoid explosion and guarantee convergence,
system, each particle of which is of quantum state
the particle must be in bound state, moving in an
formulated by wavefunction. Inspired by analysis of
amaction potential field whose center is point p . In
convergence of the classical PSO in [ 5 ] , we assume that
parctice, in terms of classical dynamics, the individual
an individual particle moves in a Delta potential well in
particle of uaditional PSO system is fly in a attraction
search space, of which the center is point p .
potential field to guarantee its bound state.
For simplicity, we consider a particle in
Correspondingly, it is also indispensable for the particles
one-dimensional space firstly. With point p the center of
of PSO with quantum behavior to move in a potential field
potential, the potential energy of the particle in
that can ensure bound state of of the particle. There are
one-dimensional Delta potential well is represented as
many potential filed models in quantum mechanics. Delta
potential well and quantum oscillator are two paradigms
of such potential fields. In this paper, we employ the V ( x ) = -y6(x - p ) = -yS(y) (8)
former to bind the particle for the sake of convergence. In
subsection E, we will give reasons for OUT selection of letting ? = X - p , where m is the mass of the particle.
Delta potential well. Hence, the Hamiltonian operator is
E. Quanfum Delfa Potential Well Model o f P S 0
In quantum time-space framework, the quantum state
of a particle is depicted by wavefunction Y ( f , f ) ,
instead of x and <
. In 3-dimensional space, the
The Schrodinger equation for the model is
wavefunction for the particle y(?,t ) is such that
1Y12dxdydz = Qdxdydz (4)
where Qdrdydz is the probability that measurement of
the particle's position at the time t finds it in the volume
Do integral
L& dx ,and when E -+ 0
' ,we obtain
element about the point ( x , y , z ) . That is, IY 1' is the
y'(o+)-y'(o-)= --I-!P(o)
2mY (11)
probability density function satisfying
h
j[mlYI2dxdydz
m = PQdxdydz
m =1 (5) For y # 0,equation (IO) can be written as
The state function Y ( 2 , t ) for a system (e.g. a
single particle) develops in time according to equation
a - p =,/- (E <0) (12)
ifi-y(x,t) =f i ( i , t ) (6)
at To satisfy the b u n d condition
326
lyl-+Co, V + O , (13) ly ( y ) J 2 or Q ( y ) , the probability density function
that the particle appears at position y relative t o p . So
the solution ofequation (12) is bound to be we have to gauge the position of the panicle, which is
called collapsing the quantum state to the classical state.
V ( Y ) e -@lyl ( y # 0) (14) Monte Carlo Method can simulate the process of
measurement. The procedure of simulation is described as
Because it can be proved that only even wavefunction follows.
satisfies the hound condition (13), we consider the Let s be the random number uniformly distributed on
solution of (12) as ( O , l / L ) ,thatis
1 1
s =- ran40,l)=-U, U =run40,l) (21)
L L
where C i s normalization constant. According to condition
(11). we obtain
Substitutes for Q or I,V(y)(2in (21), we obtain
and
u = e -*IYl/L (23)
L
*
y = -ln(l/u)
2
(24)
Because y = x -p ,
L
Determining the constants so that the function V(Y) x = p -In(l/u)
+_ (25)
satisfy normalization condition
2
where U is random number uniformly distributed on (0,l).
Thus we can measure the position of particle accurately
by formula (25).
D.Convergence Conditions
Convergence condition for PSO, as formulated in
characteristic length of Delta potential well. Subsection A, is
et c = JB = I/JL , we can represent the x+p, when t ++m (26)
normalized wavefunction as Let the parameter L develop in time, L = L ( t ) . We
assume that
V’L
L(t) -+ 0 , when t -++a (27)
Hence, the probability density function Q is given by Thus, the limit of Q(0) will be that
lim Q L(0) = CC
L+O
Because Q(y) satisfies the normalization condition, it
C. Measurement ofthe Position can been seen from (18) and (28) that Q(y) has the
characteristics of Delta Function, that is
We have obtained the quantum bound state formulated
by (21) of the particle in a Delta potential well. To
evaluate the fitness value, we need to learn of precise
information of position of the particle. However, the
quantum state function ~ ( y )can only give
Due to the characteristics of Delta Function
327
where a in the formula is the characteristic length of
harmonic oscillator field. It is an important control
parameter like L in QDPSO. The difference between
formula (25) and (35) is that there is a square root
operator in (35). We can comprehend the influence of the
operator on the algorithm by looking at the curve diagram
[:6(x - p)& = c 6 ( x - p)& =1 (30b)
of the two probability density functions as shown below.
it can be inferred that the particle will appear
in E -neighbor of p with the probability of one, when
1+m
L + 0,which suggests X +p .
In a word, to guarantee the convergence of PSO, the
following condition must be satisfied
limL(t) = 0
”I
(31)
where
f r z , which is the vital parameter
L = 1/p = -
my
of QDPSO.
Fig. 1
E. Comparison of Delta potential well model with other
quantum potentialjield model
Besides the quantum Delta potential well model, there
are many other quantum potential models that can be
established for PSO. Let’s consider quantum harmonic
osicllator (QO) field for a PSO system with the same
methodology as adopted in previous sections. The wave
function of QO model is given by
I
Fig. 2
It is shown in figure (I) and (2)that the curve of (33) is
and the probability density function Q is relatively namower than that of (20), which means that the
particle converges much faster in QOPSO system than in
QDPSO system with the same parameter control method.
But such convergence speed is acquired in sacrifice of
global search ability of the algorithm. We test quatum
oscillator model of PSO on some benchmark fuctions and
Using Monte Carlo simulation, position of the the results tell us that QOPSO algorithm is prone to
particle can be measured by prematurity.
Other notential field models either have wavefuctions too
complex to be simulated, or is liable to prematurity.
Consequently, we select Delta potential well in our
proposed PSO algorithm.
Ill. THE PROPOSED QDPSO ALGORITHM
A. A Method of Parameter Contml
Given that the position of the particle at search step k is
328
x k , at step k i l , the particle may appear in the zone of xid = p + L * (ln(l/u))
Until termination criterion is met
(jy,l,\y,l) with the probability of 2, or out of the
zone with the probability I-r. To make x converge top, it
In the algorithm, only position vector ?
. is needed to
is absolutely necessary that Z > 0.5. That is to say, the depict a particle, and there is only one parameter g.
probability that y,+l is at the lei? side of IJkl must However, as we have emphasized, the given method of
parameter control and selection is trial.
be more than 0.75. So we can obtain
IV. EXPERIMENT RESULTS AND DISCUSSION
To test the performance of QDPSO, seven benchmark
functions are used here for comparison with SPSO[Z]. The
first fucntion is Sphere function described by
(39)
and i=I
The second function is the Rosenbrock function
described by
i=l
The third function is the generalized Rastrigrin function
We formulate L as L = (l/g)lXk -PI and g is a described by
parameter that is constrained by
g > In& (38)
f(x)] =c(x;- l O k o ~ ~ ~ ) + l O )
n
i=l
(41)
Ultimately, the control of parameter L can be reduced The fourth function is generalized Griewank function
to the control and selection of parameter g . described by
E . Qunntum Delta-Pofential- ell-bnsed PSO Algorithm
Having presented the dynamics of quantum Delta The fifth function is De long's function (no noise)
potential well model of PSO and the method of parameter described by
control, we describe the QDPSO algorithm as follows.
(43)
QDPSO Algorithm
Initialize population: random xi The sixth function is Rosenbrock variant function
Do described by
For I=I IOpopulation size M f ( x ) , =10qx: -x$ +(l-x$ (44)
'r/ .fGi1 < S ( p j1 hen p i = xi
P,, = min(p,) -
The seventh function is Shaffer's function described by
(sindx' + y 2 ) - 0 . 5
For d=l to dimension D f ( x ) , =0.5+ (1.0+0.001(x2 y 2 y + (45)
pt = rund(0,l) = rund(0,l) . .
These functions are all minimization problems with
*
P = (Pt Pid + Pz * P g d )/(P,+ 9 2 ) minimum value zero.
For the purpose of comparison, the asymmetric
u ~ =rand(0,l) initialization method used in [Z, 41 is adopted here for
L:= (I/g) * abs(x, - P ) population initialization. Tahle 1 lists the initialization
ifrand(0, I)>O. 5 ranges and table 2 lists the v, and x,,,values for
x i d = p - L * (W u ) ) all the functions, respectively. The fitness value is set as
function value. We had 50 trial NnS for every instance.
else
329
In order to investigate the scalability, different TABLE 3. THE MEAN FITNESS VALUE FOR
population sizes M are used for each function with SPHERE FUNCTION
different dimensions. The population sizes are 20,40 and
80. Generation is set as 1000,1500 and 2000 generations
corresponding to the dimensions IO, 20 and 30 for first
five functions, respectively, and the dimension of the last
two function is 2 as shown in table 7 and 8. The parameter
g is set to 0.96.
We record mean best fitness values for 50 m s of each
functions in table 3 to table 9. The value in column of
SPSO in from Table 4 to Table 8 is taken from [2]. By
compare the results, it is easy to see that QDPSO have
better result than SPSO when population size is larger
than 20. When there are 2 0 particles in population, the
result of QDPSO is not better than SPSO, for the QDPSO
is an entire probability searching technique and larger
population may show its better performance.
The results of experiment on Shaffer's fucntion in table 8
are worth noticing. The function has a suboptimal value
0.0097. The experiment results show that particles of TABLE 4. THE MEAN FITNESS VALUE FOR
SPSO system are able to escape trap of the suboptimal ROSENBROCK FUNCTION
more frequently than that of QDPSO, for the reason that QDPSO
the particles in SPSO search along relatively continuous
i 1 ;::: I 1
96.1715 14.2221
trajectories, while those in QDPSO fly discretly so that
they may miss the narrow zone where the optimal may 20 20 1500 214.6761 175.3186
lies. However, as the population size is enlarging, the
f: 316.4468 242.3770
defference between the two algorithms is shrinking.
70.2139 15.8623
4o 20 1500 180,9671 112.4612
Function Asymmetric Initialization Range
I 30 I 2000 I 299.7061 76.4273
80 20 [ 1500 [ 87.2802 23.5443
30 2000 205.5596 71.9221
(30, 100)
f6 (30, 100) TAI3LE 5 . THE MEAN FITNESSVALUE FOR
MSTRlCKlN FUSCTION
f, I (30, 100)
IO 1000 5.5572 4.9698
TABLE 2
I
20 __ I isno
20 ..-. I 22.8892 1I 17.0789
I
Function 2000 47.2941 48.6199
30
IO 1000 3.5623 2.0328
16.3504 10.9453
38.5250 21.3712
600 600
2.5379 0.9232
f. IO0 100
13.4263 6.9554
I f: I 100 I 100 I
30 [ 2000 I 29.3063 18.1305
330
l:\BLt 6 THE MEAT bINF.SS \‘.&I.UE FOR V. CONCLUSION
GKIIWAKK ti:sc riox
SPSO
In this paper, a quantum version of particle swarm
11 IO03 optimization algorithm was introduced. The algorithm is a
probability searching technique, or says it accurately, an
003U3 0 0086
uncertainty searching algorithm. Due to the characteristics
30 2000 0.0182 0.0544 of wavefunction o f quantum Delta potential well model,
there are some partilces appearing at far distance from
paint p. which is prohibited by classical PSO. These
particles like prophets in human society have chance to
find aut better positions in searching space.
30 I2000 0.0127 0.0009 Like classical PSO, QDPSO is sensitive to parameters.
IO I 1000 I 0.0760 0.0000 We present a method of parameter control, but it is trial.
The selection and control of parameters of QDPSO is an
open problem and an effective method of parameter
30 2000 0.0128 0.0000 control will improve the performance of QDPSO greatly.
The adaptive method of parameter control may lead to a
mare efficient algorithm and our current work is focusing
on this problem
TABLE 7. THE MEAN FIIXESS VALUE FOR
DE JONG’S FUNCTION
QDPSO
0.0000 0.0000 REFERENCES
1- 2000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
O.OO00
0.0000
Y. Shi and R. Eberhan, “Funy adaptive particle swarm
optimization,” IEEE int. Conf. on evolutionary
computation.2001, pp 102-106
Y. Shi and R. Eberhart, “Empirical study of panicle swarm
optimization,” Proc. Of Congress on Evolutionary
Computation, 1999, 1945-1950
I. Kennedv. and R. Eberhart. “Particle s w m ootimization”.
Pmc. IEEE int. Conf. O n Neural Network, 1995:’
1942-1948
0.0000 0.0000 1. Kennedy, “The particle swarm: social adaptation of
knowledge,” IEEE int. Conf. On Evolutionary
0.0000 0.0000
Computation, 1997,303-308
M. Clerc and J. Kennedy, “The particle swarm: explosion,
stability, and convergence in a multidimensional complex
space,” IEEE Transactions on Evolutionary Computation,
1ABI.CX 1 I I E \WAS FlTSESSVAI.UE tOR vol. 6, no. I ppS8-73, February 2002
~
R0SI:NUKOCK V A K I A K FLSCTION E. Ozcan and C. K. Mohan, “Particle swarm optimization:
surfing the waves,” Pmc. of congress an evolutionary
QDPSO
computation, 1999, 1939-1944
0.0oM) 0.0000 Y. Shi, and R. Eberhar, “Parameter selection in partkle
swarm optimization”, Proc. seventh annual canf. on
40 2000 0.0000 0.0000 evolutionaryprogra”ng, March 1998, pp. 591-M)l
80 2 2000 0.0000 0.0000 J. Kennedy, “The behavior of particles”, The seventh
annual emf. on evolutionaxy computation, 1991,
pp303-308
R. Eberhart, and Y. Shi, “Comparison between genetic
algorithms and particle swarm optimization,” Proc. seventh
annual oonf. on evolutionary programming, March 1995,
M D G, SPSO QDPSO pp.616-619
R. Eberhart and Y. Shi, “Panicle swarm optimlratoo:
20 2 2000 0.0012 0.OOSl developments,applications and resources: IEEE int. conf.
on evolutionary computation,2001 pp 81-86
40 2 2000 0.0006 0.0018
80 2 2000 0.0002 0.0004
33 1
