You are on page 1of 5

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO.

2, FEBRUARY 2010 141

A Quantum-Behaved Particle Swarm Optimization


With Diversity-Guided Mutation for the Design
of Two-Dimensional IIR Digital Filters
Jun Sun, Member, IEEE, Wei Fang, and Wenbo Xu, Member, IEEE

Abstract—This brief proposes quantum-behaved particle The brief is organized as follows. Section II describes
swarm optimization with diversity-guided mutation (QPSO- the problem. The proposed QPSO-DGM is presented in
DGM) to solve the problem of designing the optimal 2-D Section III. Section IV provides and discusses the experimental
zero-phase IIR digital filters. The new method integrates a
results on a 2-D filter design problem. Section V offers some
diversity control strategy into QPSO to guide the particle’s search
and thus improve the capabilities of exploration. Numerical results conclusions.
demonstrate that the design approach based on QPSO-DGM can
obtain better digital IIR filters than the existing methods.
II. P ROBLEM F ORMULATION
Index Terms—Digital IIR filters, heuristics, multidimensional
digital filters, particle swarm optimization (PSO). For design purposes, we consider the 2-D recursive filters
with the following transfer function:
I. I NTRODUCTION N N
i=0 j=0 aij z1i z2j
H(z1 , z2 ) = H0 N , a00 = 1
T WO-DIMENSIONAL (2-D) IIR filter design has attracted
growing attention due to its extensive applications in the
domain of denoising of digital images, biomedical imaging and
k=1 (1+qk z1 +rk z2 +sk z1 · z2 )
(1)
digital mammography, etc. [1]–[3]. The most popular design where N is the dimension of the filter. Let the frequencies
methods, which may result in an unstable filter, are based either ω1 , ω2 ∈ [−π, π], z1 = e−jω1 , and z2 = e−jω2 . The design task
on appropriate transformation of 1-D filters or on appropriate of 2-D filters amounts to finding a transfer function H(z1 , z2 )
optimization techniques such as linear programming, Remez as in (1) such that the function M (ω1 , ω2 ) = H(e−jω1 , e−jω2 )
exchange algorithm, and nonlinear programming. approximates the desired amplitude response of the 2-D
Modern heuristic methods have also been employed for 2-D filters Md (ω1 , ω2 ). The approximation can be achieved by
IIR filter design problems, such as genetic algorithm (GA) minimizing [4]–[7]
[4], neural network (NN) [5], particle swarm optimization
(PSO) [6], and the computer language GENETICA [7]. These J = J(aij , qk , rk , sk , H0 )
techniques were able to find out better solutions than those

N1  N2      p
mentioned in the previous paragraph.  πk1 πk2 πk1 πk2 

The PSO algorithm, which is motivated by social behavior of =  M N1 , N2 − M d N1 , N2 
k1 =0 k2 =0
bird flock or fish schooling, is a population-based optimization
technique [8]. In PSO, the potential solutions, called particles, (2)
fly toward the optima by following their own experiences
and the current global best particles. In this brief, based on where p is a positive even integer (p = 2, 4, or 8), and N1
a variation of PSO, namely, quantum-behaved PSO (QPSO) and N2 are positive integers. Thus, the design of 2-D filters is
[9], we propose QPSO with diversity-guided mutation (QPSO- equivalent to the following constrained minimization problem:
DGM) for the design of 2-D IIR digital filters. The method is
tested, and a performance comparison is made with some of Minimize J = J(aij , qk , rk , sk , H0 ) (3a)
other heuristics.
s.t.

Manuscript received June 25, 2009; revised October 10, 2009. First pub- |qk +rk |−1 < sk , sk < 1−|qk −r|, k = 1, 2, . . . , N.
lished January 22, 2010; current version published February 26, 2010. This
work was supported in part by the Innovative Research Team Project of (3b)
Jiangnan University under Contract JNIRT0702. This work was recommended
by Associate Editor M. Chakraborty.
The authors are with School of Information Technology, Jiangnan Uni- Here, the prime object is to reduce the difference between the
versity, Wuxi 214122, China (e-mail: sunjun_wx@hotmail.com; wxfangwei@ desired and actual amplitude responses of the filter at N1 × N2
163.com; xwb@jiangnan.edu.cn). points. Since the denominator only contains first-degree
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. factors, we can assert the stability conditions as the constraints
Digital Object Identifier 10.1109/TCSII.2009.2038514 (3b) [4]–[7].

1549-7747/$26.00 © 2010 IEEE


142 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO. 2, FEBRUARY 2010

III. P ROPOSED QPSO-DGM A LGORITHM TABLE I


PARAMETER C ONFIGURATIONS FOR THE C OMPETITOR A LGORITHMS
In QPSO with m particles in the n-dimensional space, for
particle i (1 ≤ i ≤ m), the jth (1 ≤ j ≤ n) component of its
position at iteration t + 1 is updated by
 t 

m
 t 
i,j = pi,j ± α Ci −xi,j · ln 1/ui,j
xt+1 t t
Cjt = (1/m) t
yi,j
i=1
(4)
where pti,j = ϕti,j yi,jt
+ (1 − ϕti,j )ŷjt , with yi,j
t
and ŷjt repre-
senting the jth components of the personal best (pbest) position
of the particle and the global best (gbest) position of the pop-
ulation. Here, C is called the mean best (mbest) position, and
ϕti,j , uti,j ∼ U (0, 1), where U (0, 1) is the uniform distribution
in [0, 1]. The parameter α is called the contraction–expansion
coefficient.
Although QPSO is efficient in solving global optimization
problems, it may also encounter premature convergence, which
is the result of constant declination of population diversity
during the search. In the proposed QPSO-DGM algorithm, we
set a lower bound dl for the diversity, which, as in [10], is
measured by

m 
 n
t 2
dt = [1/ (M · |A|)] ·  xi,j − xtj (5) To test QPSO-DGM, we use the same example of the design
i=1 j=1 problem as in [4]–[7], where Md (ω1 , ω2 ) is given by
 ⎧ 
where xtj = (1/m) m ⎨ 1, ω12 + ω2 ≤ 0.08π
2
i=1 xi,j , and |A| is the length of the
t

longest diagonal in the search space. If dt decreases to below Md (ω1 , ω2 ) = 0.5, 0.08π < ω12 + ω22 ≤ 0.12π (7)

dl , the following mutation is operated. That is, for each j, let 0, otherwise.

zjt = ŷjt + γ · |A| · ε, ε ∼ N (0, 1). (6) Since N = 2, each particle’s position has 15 real-valued
coordinates, with each representing a parameter (a00 = 1). To
Then set ŷjt = zjt and yg,j
t
= zjt , where yg,j
t
denotes the pbest handle the constraints, we also follow the methods in [4]–[7].
position of the current particle with the gbest position (called
the gbest particle). N (0, 1) denotes the standard normal distri- IV. E XPERIMENTAL R ESULTS AND D ISCUSSIONS
bution, and γ is a user-specified parameter, which should be
A. Experimental Results
much larger than dl so that dt can promptly increase to above
dl after mutation. We suggest that γ ≥ 10dl . z t is a temporary We ran QPSO, QPSO-DGM, PSO [6], [8], and GA [4] on
vector. When the mutation is operated, the displacement of the example design problem in [4]–[7]. The parameter configu-
the gbest particle increases the value of |ŷjt − yi,j
t
| and pulls rations for all algorithms are shown in Table I. The search scope
the position C away from its original position, making the for each coordinate is [−3, 3].
particles’ search scope expanded and, in turn, resulting in the The accuracy of the algorithms and their speed of conver-
gain of dt . Here, we outline the QPSO-DGM algorithm. gence were investigated through experiments. We used the
number of fitness function evaluations (FEs) as a measure of
Procedure of QPSO-DGM: time. To judge the accuracy of the algorithms, we ran all of
Step 0: Initialize particles with random position. them for 40 000 FEs. Each algorithm independently ran for
Step 1: For t = 1 to the maximum number of iterations, 100 times. The best values of filter parameters obtained, and
execute the following steps. the mean best J values, along with their standard deviations
Step 2: Calculate the mean best position C among the and their best values out of 100 runs, were reported in Tables II
particles. and III. Table II also lists the filter parameters yielded by NN
Step 3: For each particle, compute its fitness value f (xti ). If [5] and GENETICA [7] for comparison. The denotation Jp
f (xti ) < f (yit−1 ), then yit = xti . was used to denote the three sets of experiments performed
Step 4: Select the gbest position ŷ t among particles. with the value of J obtained using exponent p = 2, 4, and 8.
Step 5: Measure dt according to (5). If dt < dl , execute Table IV summarizes the results of the unpaired test on the J2
Step 6, or else go to Step 7. value (standard error of difference of the two means, 95% confi-
Step 6: Exert the mutation on each ŷjt and yg,jt
as (6) and dence interval of this difference, the t-value, and the two-tailed
update the fitness value of the gbest particle. P -value) between results of competitor algorithms in Table III.
Step 7: Update each component of the current position by (4) Figs. 1–10 show the ideal low-pass filter frequency response
and return to Step 1. and the frequency response of the filter designed using the
SUN et al.: PSO WITH DIVERSITY-GUIDED MUTATION FOR THE DESIGN OF IIR DIGITAL FILTERS 143

TABLE II
B EST VALUES OF F ILTER C OEFFICIENTS W ITH E XPONENT p = 2 A FTER 40 000 FE S G ENERATED BY THE C OMPETITOR A LGORITHMS

TABLE III
B EST ( IN B RACKETS ), M EAN VALUES , AND S TANDARD D EVIATIONS OF J (p = 2, 4, AND 8) A FTER
40 000 FE S FOR 100 I NDEPENDENT RUNS OF E ACH A LGORITHM

TABLE IV
R ESULTS OF U NPAIRED t-T ESTS ON THE DATA OF THE VALUES OF J W ITH E XPONENT p = 2 IN TABLE III

competitor algorithms. All these figures were drawn from the


results obtained using the minimization algorithm of J2 for
40 000 FEs and generated by NN [5] and GENETICA [7]
(in Table II). Figs. 2–10 reveal that QPSO-DGM, including
those with 40 (QPSO-DGM1) and 20 particles (QPSO-DGM2),
yielded better approximations of the desired response compared
with other algorithms and yielded a considerably good filter
response with a smaller number of FEs, as evident from Fig. 11.
The ripples in the stopband of Figs. 7–10 are much lesser than
those of Figs. 2–6. As evident from t-test results on the data
of Table IV, the QPSO-based methods (including QPSO and
QPSO-DGM with different population sizes) beat their nearest
competitor in a statistically significant manner. The comparison Fig. 1. Desired amplitude response |Md (ω1 , ω2 )| of 2-D filters.
also reflects the fact that GENETICA could generate better filter
coefficients than GA, NN, and PSO, as well as comparable
results with QPSO. The aforementioned comparison shows trapped into the worst solutions on average. It reflects that the
that, with different numbers of particles but equal numbers QPSO-DGM algorithms could converge to the better solutions
of FEs, the performance of PSO or QPSO has no statistically than PSO, GA, and QPSO in the 2-D digital filter design.
significant difference. It is, however, not the case for QPSO-
DGM, since the results in Table IV show that QPSO-DGM2 has
B. Discussions
better performance than QPSO-DGM1 in terms of statistical
significance, which will then be discussed. The whole search process of QPSO-DGM can be partitioned
Fig. 11 visualizes the convergence process of the algorithms into two stages. The first stage is the time interval from
toward the optima. Although GA converged fastest, it was initialization to the first time T , when QPSO-DGM reaches
144 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 57, NO. 2, FEBRUARY 2010

Fig. 2. Amplitude response |M (ω1 , ω2 )| of 2-D filters using the NN method


in [5].
Fig. 6. Amplitude response |M (ω1 , ω2 )| of 2-D filters using PSO2.

Fig. 3. Amplitude response |M (ω1 , ω2 )| of 2-D filters using GA in [4].


Fig. 7. Amplitude response |M (ω1 , ω2 )| of 2-D filters using QPSO1.

Fig. 4. Amplitude response |M (ω1 , ω2 )| of 2-D filters using GENETICA


in [7].
Fig. 8. Amplitude response |M (ω1 , ω2 )| of 2-D filters using QPSO2.

Fig. 5. Amplitude response |M (ω1 , ω2 )| of 2-D filters using PSO1. Fig. 9. Amplitude response |M (ω1 , ω2 )| of 2-D filters using QPSO-DGM1.

to below the low bound of diversity. The rest of the search maximum number of FEs. Table V records the statistical results
thenceforth is the second stage. If QPSO-DGM finishes the of T for both QPSO-DGM1 and QPSO-DGM2.
search before getting into the second stage, it essentially runs From Table V, one can see that T of QPSO-DGM2 is smaller
as the original QPSO without taking advantages of the diversity than that of QPSO-DGM1, on average, implying that the latter
control method. In this case, we assume that T is larger than the spends more time (number of FEs) on the first stage. Table V
SUN et al.: PSO WITH DIVERSITY-GUIDED MUTATION FOR THE DESIGN OF IIR DIGITAL FILTERS 145

TABLE V
S TATISTICAL R ESULTS OF T FOR QPSO-DGM1 AND QPSO-DGM2

Fig. 10. Amplitude response |M (ω1 , ω2 )| of 2-D filters using QPSO-DGM2.

guidance of diversity measure, and the mutation operation, in


turn, maintains the diversity at a favorable level for the search,
making QPSO-DGM a more adaptive and effective approach.

V. C ONCLUSION
QPSO-DGM has been proposed for the design of 2-D zero-
phase recursive filters. Compared with other methods described
in [4], [5], and [7], which, to our knowledge, are the most
recent and the best-known methods to date, the QPSO-DGM
Fig. 11. Convergence processes of tested algorithms on J2 for 40 000 FEs. algorithm used here yielded a better design. QPSO-DGM has
also shown to beat QPSO and PSO in terms of statistical
reveals that, out of 100 runs, there are 53 times for which significance for the design of 2-D IIR digital filters. Moreover,
QPSO-DGM1 does not enter the second stage. In other words, if the maximum number of function evaluations was given,
the 53 runs of QPSO-DGM1 are actually the 53 runs of QPSO1, fewer particles led QPSO-DGM to yield a better solution in a
since the diversity never reached to below dl during the search statistically significant manner.
of a single run. QPSO-DGM2, however, failed to get into the
second stage for only ten trial runs. This may be caused by R EFERENCES
combination of the following two reasons. The one is that more [1] W.-S. Lu, S.-C. Pei, and C.-C. Tseng, “A weighted least-squares method
particles diversify the population so that the diversity of QPSO- for the design of stable 1-D and 2-D IIR digital filters,” IEEE Trans. Signal
DGM1 more slowly declines. The other is that the value of α Process., vol. 46, no. 1, pp. 1–10, Jan. 1998.
more frequently decreases in QPSO-DGM2, resulting in faster [2] D. J. Krusienski and W. K. Jenkins, “Design and performance of adaptive
systems based on structured stochastic optimization strategies,” IEEE
convergence speed of the particle and, thus, more rapid decline Circuits Syst. Mag., vol. 5, no. 1, pp. 8–20, Feb. 2005.
of diversity. [3] B. Dumitrescu, “Optimization of two-dimensional IIR filters with non-
More rapid diversity decline helps the algorithm quickly separable and separable denominator,” IEEE Trans. Signal Process.,
vol. 53, no. 5, pp. 1768–1777, May 2005.
converge to find a relatively good solution at the early stage of [4] N. E. Mastorakis, I. F. Gonos, and M. N. S. Swamy, “Design of two-
the search. However, it is not desirable for the algorithm when dimensional recursive filters using genetic algorithms,” IEEE Trans.
it is at the later stage, since the algorithm may lose its global Circuits Syst. I, Fundam. Theory Appl., vol. 50, no. 5, pp. 634–639,
search ability if the diversity is too low. Therefore, exerting May 2003.
[5] V. M. Mladenov and N. E. Mastorakis, “Design of two-dimensional digital
mutation operation to maintain the diversity at the later stage recursive filters by using neural networks,” IEEE Trans. Neural Netw.,
of the search can be very helpful to the performance of QPSO- vol. 12, no. 3, pp. 585–590, May 2001.
DGM2. On the other hand, slower diversity decline maintains [6] S. Das, A. Konar, and U. K. Chakraborty, “An efficient evolutionary
algorithm applied to the design of two-dimensional IIR filters,” in Proc.
the global search ability but impairs the local search ability of GECCO, Washington, DC, 2005, pp. 2157–2163.
the algorithm, resulting in lower convergence speed. For QPSO- [7] I. F. Gonos, L. I. Virirakis, N. E. Mastorakis, and M. N. S. Swamy,
DGM1, in many runs, lower convergence leads the algorithm “Evolutionary design of 2-dimensional recursive filters via the computer
even not to converge to below dl by the end of the search, language GENETICA,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 53,
no. 4, pp. 254–258, Apr. 2006.
weakening the local search ability of QPSO-DGM1 at the later [8] Y. Shi and R. C. Eberhart, “A modified swarm optimizer,” in Proc. IEEE
stage. This is why the final solutions found by QPSO-DGM1 Int. Conf. Evol. Comput., 1998, pp. 1945–1950.
are, on average, poorer than those by QPSO-DGM2. [9] J. Sun, W. B. Xu, and B. Feng, “A global search strategy of quantum-
behaved particle swarm optimization,” in Proc. IEEE Conf. Cybern. Intell.
Furthermore, the proposed QPSO-DGM algorithm is very Syst., 2004, pp. 111–116.
different from similar methods, such as restarted PSO, in that [10] R. K. Ursem, “Diversity-guided evolutionary algorithms,” in Proc. Paral-
it explicitly exerts mutation on the global best position by the lel Problem Solving From Nature Conf., 2002, pp. 462–471.

You might also like