214 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO.

1, JANUARY 2011

Modification of Real-Number and Binary PSO

Algorithms for Accelerated Convergence
Arezoo Modiri, Student Member, IEEE, and Kamran Kiasaleh, Senior Member, IEEE

Abstract—Modifications in the velocity calculation of the par- different electromagnetic applications, a good list of which can
ticle swarm optimization (PSO) algorithm are proposed. The sug- be found in [8]. Although PSO has already shown considerable
gested modifications aim to arrive at a faster, more straightfor- advantages in electromagnetic designs [9]–[14], it should still
ward and still robust search procedure as compared to the con-
ventional method. Two main factors, i.e., personal best influence be noted that the EAs’ efficiencies are problem-dependent,
and initial velocity values, are evaluated. It is shown that in prob- and their performances are impacted significantly by the initial
lems with wide-range parameters, the effect of personal best loca- solutions [15]–[17].
tions is intrinsically encompassed by that of global best locations, In general, there are two approaches one may take for the pur-
thereby allowing for further simplification of the PSO algorithm pose of optimization. The first approach makes an attempt to
by eliminating the factor which accounts for the personal best so-
lutions in the velocity calculation. This simplification expedites the find a suitable solution as fast as possible, whereas the second
convergence procedure in real PSO. It is also shown that the ini- approach is concerned with the best possible solution regard-
tial velocity values can be modified to enhance the performance less of the number of iteration. Since the first approach is often
in terms of achieving better solution when compared with the ex- selected for a large number of electromagnetic problems, this
isting algorithms, particularly in binary PSO. In order to validate paper targets the same approach as well.
the viability of the proposed procedure, the performances of the
real-number and binary PSO algorithms with different velocity PSO has some deficiencies, which limit its performance. To
calculations are assessed in 1000-run sets, and pros and cons are elaborate, it is possible that particles get locked in a local best
studied. In particular, the performance of the proposed algorithm, solution, which prevents them from reaching the best solution;
when used to design software defined thinned array antennas, is or, the optimization may turn to be time consuming due to large
shown to be superior to those of the existing algorithms. number of required iterations and/or particles. The latter oc-
Index Terms—Evolutionary optimization, particle swarm opti- curs when the fitness evaluations go beyond simple mathemat-
mization, side lobe level, thinned antenna array. ical calculations (as is the case in a majority of electromag-
netic problems). Thus, PSO in its current form seems to be in-
appropriate for time-sensitive optimization problems, often en-
I. INTRODUCTION
countered in software-defined applications (such as smart beam-
forming of thinned array antennas). Hence, this study proposes
methods to enhance the performance (in terms of convergence
E VOLUTIONARY algorithms (EAs) are based on sto-
chastic variation of a number of particles (individuals)
in a solution space. The major difference between various EA
time) of PSO algorithm while preserving its robustness.
In the ensuing sections, after a brief description of PSO, two
algorithms resides in the variation operators. A comprehensive major modifications in velocity vector are introduced and their
comparison of EAs in electromagnetic can be found in [1] where performances are compared to that of the existing PSO.
three groups of EAs are introduced: evolutionary programming
(EP), evolution strategies (ES) and genetic algorithms (GAs).
After several assessments, the author concluded that the EP’s II. PSO ALGORITHM
simplicity, ease of implementation, and flexibility offer design
engineers with a valuable tool for efficient design of microwave The principles of PSO algorithm can be found in various ref-
devices. PSO algorithm falls within the EP group, and hence erences in literature (e.g., in [17], [18]), thus, this paper does not
go through basic descriptions of the algorithm.
the above conclusion also applies to the PSO algorithm.
Motivated by the behavior of groups of organisms during Since it is often possible to merge several fitness (objective)
their food-searching activities, Eberhart and Kennedy intro- functions into one, single-objective optimization is the focus of
duced the PSO algorithm in 1995 [2]. The idea was further most optimization problems. Single-objective PSO algorithms
have been classified in two groups, real-number (RPSO) and
developed in the ensuing years [3]–[7]. PSO has been used
along with a variety of analytical and numerical tools for binary (BPSO) [8], and both are considered in this study.
The only key operator in PSO is the velocity calculation
which has been investigated in several studies (e.g., [15], [8],
Manuscript received September 22, 2009; revised May 13, 2010; accepted
July 09, 2010. Date of publication November 01, 2010; date of current version
[19]). In this paper, benefiting from the results available in
January 04, 2011. This work was supported in part by National Institute of Jus- the literature, we propose and compare the performance of a
tice (NIJ) under Grant #2006IJCXK043. new velocity vector. To that end, let us examine the velocity
The authors are with the Department of Electrical Engineering, Uni- function of the PSO algorithm.
versity of Texas at Dallas, Richardson, TX 75080-3021 USA (e-mail:
[email protected]; [email protected]). It is generally assumed that there are particles and param-
Digital Object Identifier 10.1109/TAP.2010.2090460 eters (or dimensions) to be optimized. The particles’ positions
U.S. Government work not protected by U.S. copyright.
MODIRI AND KIASALEH: MODIFICATION OF REAL-NUMBER AND BINARY PSO ALGORITHMS 215

and velocities are stored in matrices. More precisely, the moving towards locations identified by which will never
velocity function is given by become one of the locations identified by, seems to be a mod-
ifiable step. Hence, in this study, we propose the following ve-
locity function:

(1)
(3)
where denotes the matrix of particles’ velocities in the
th iteration, is the position matrix in the th iteration, and The values assigned to velocity vector coefficients ( and
or “personal best” is the matrix of the locations in ) in the literature have been, in large part, obtained using ex-
the solution space, which are remembered by each particle as its perimentation [18]. Thus, they can be modified according to op-
most successful position discovered during its own exploration. timization goal.
The th row in corresponds to the n-dimensional location We modify the manner by which the initial velocity vector is
of the th particle. or “global best” is the matrix selected as well. The motivation behind this modification is sup-
of the best locations found (in terms of the fitness value) by ported by nature. For instance, bees start their flower-searching
all particles. Furthermore, and are two random variables movement with a speed that is higher than that of the subse-
uniformly distributed on [0, 1]; and finally, and are three quent steps. This characteristic has already been added to RPSO
coefficients which can be used to modify the algorithm in order by decreasing in each step of the iteration. However, such an
to achieve the desired performance. approach has shown no impact on BPSO [8]. In the proposed
In the conventional PSO, both RPSO and BPSO, the initial algorithm, the velocity vector is initialized using the following
velocity values are defined as it is shown in (2). statement:

(4)
(2)
The novelty of the equation is in its two-state discrete na-
Where is the matrix of initial velocities, and,
ture in compare to conventional velocity of (2). In this paper, as
and are the maximum and minimum allowed velocities in in conventional PSO, represents a velocity value equal to
the th dimension. A mapping function is used for BPSO to but in the opposite direction, and is equal to the pa-
quantize real values [8]. rameter range for RPSO. In BPSO, is selected with regards
In order to limit the search to the solution space, a boundary to the mapping function [8]. Since each parameter in BPSO can
condition should be introduced. In [15], several boundary con- either be 1 or 0, we propose to be 0. is assumed to be
ditions were assessed and invisible wall was claimed to show a 6 as in literature [8].
better performance. This study uses invisible wall as the main As it has been mentioned in [1], [8], [18] and [19], the sug-
boundary condition as well, which means that the fitness value gested values for , and are either 1.49 for both, or 2 for both.
calculation will be skipped for particles that fly out of the solu- is also suggested to decrease during iterations linearly from
tion space. Upon the return of the particle, the fitness value cal- 0.9 to 0.4 in RPSO, and have constant value of 1 in BPSO. Pre-
culation will resume. However, absorbing wall boundary condi- vious studies recommend for better convergence
tion is also considered yet as a second option. [8], and therefore, this paper focuses on as well.
III. PROPOSED VELOCITY VECTOR In Subsection A, it is illustrated that the proposed velocity
function of (3) reinforces a scenario in which the particles have
As mentioned in introduction, although PSO outperforms its collectively made a correct movement toward the global best
counterparts such as GA, in many ways, it has some deficien- solution on the average. However, the performance of this ap-
cies. Basic PSO has been modified in different ways to improve proach needs to be examined in the event where particles have
the drawbacks; in [20] various modified methods were intro- not taken the correct step toward the global best solution. Hence,
duced and their performances were compared. Furthermore, a in the ensuing sections, the performance of the proposed algo-
new method, named tabu-searching PSO (TS-PSO), was intro- rithm will be compared to the existing algorithm under realistic
duced. In TS-PSO a second swarm and a number of tabu con- scenarios.
ditions are introduced. Equations (5)–(7) in [20] show that the In Sections IV and V, the results of simplified velocity func-
solution is gained by leading each particle of the second swarm tion and the proposed initial values are studied in RPSO and
away from any dimension of the personal best and/or global best BPSO, respectively. It is shown that modifying PSO velocity
location associated to the same particle that enters the tabu con- vector as suggested, mainly increases the convergence speed but
dition. This method is claimed to prevent the algorithm from may yield (depending on the problem) a better solution than that
getting locked in local best solutions but optimization length obtained using the conventional method too.
increases.
Since a few rows of are qualified to become rows in A. Statistical Analysis
, the idea of ignoring unqualified locations in , as an al- The viability of (3) can even be proved mathematically by
ternate to the move-away procedure in TS-PSO, comes to mind. evaluating the average position of particles in each iteration
The goal of optimization is to move toward best solution, thus, cycle. Let and denote the position and velocity of
216 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO. 1, JANUARY 2011

the th particle in the th dimension during the th iteration re-

spectively. In PSO algorithm, this position is achieved by

(5)

Using the conventional velocity function (1), the average par-

ticle velocities are given by

(6)

where and denote the average personal best posi-

tion vector of all existing particles and the global best position
vector in the th iteration respectively. and are the position
and velocity vectors averaged over i.
Substituting in (6) and using (5), the average
position vector is simplified to

(7)

If the algorithm moves the particles in the correct direction,

should approach . Therefore, the average position
of the particles in a successful iteration cycle in which the ma-
jority of particles have moved to positions with better fitness
values (compared to previous iteration) can be approximated as

(8) Fig. 1. AIN behavior of the four methods over gradually increasing sample runs
in a 10D problem with 5, 10 and 20 particles for (a) Rastigrin, (b) Griewank,
One can observe that the average position of particles in (8) 0
and (c) Rosenbrock functions and parameter range of [ 500; 500].
may also be achieved by using the proposed velocity function,
given by (3) in above procedure.
• Method1: , and the initial velocity is given by
IV. EVALUATION OF THE PROPOSED VELOCITY (2) (conventional method).
FUNCTION IN RPSO • Method2: , and the initial velocity is given by
(4).
In order to have comparable results, the benchmark fitness
• Method3: , and the initial velocity is given
functions used in [1] and [20], shown in (9)–(11), are used. To
by (2).
elaborate, the functions shown in (9)–(11) all possess a global
• Method4: , and the initial velocity is given
minimum that occurs at the origin of the n-dimensional solution
by (4).
space. So the goal of optimization is to approach this zero value
In order to have comprehensive statistical study, the conver-
by moving particles in an -dimensional solution space.
gence behaviors in different cases are observed in 500 iterations
Rastigrin function:
during 1000 runs, and the average results associated with each
benchmark function are depicted in terms of the achieved av-
(9) erage final value (AFV) and the average iteration number (AIN)
required to attain a given AFV. The results are averaged over a
Griewank Function: set of runs. The number of sample runs in the set increases by
20, i.e., the results of sample runs are averaged
and plotted. In this manner, in addition to achieving consistent
(10) statistical results, we are also able to measure the number of
samples required for attaining reliable results.
Rosenbrock function: A 10D scenario is assumed with particle num-
bers of 5, 10, and 20. The parameters can vary in the ranges of
, and . Although a higher
(11) number of particles are generally considered in RPSO, two ob-
jectives are obtained here by these selections. First, the (m,n)
Where . To assess the impact of pro- values are chosen so that they match the parameters for the
posed modifications separately, the results of the following four ensuing binary PSO study, allowing for a comparative study.
methods are investigated (the initial solutions are identical for Second, since speed of convergence is a priority in this study,
all methods during each run throughout this paper). key solutions for accelerating the convergence while preserving
MODIRI AND KIASALEH: MODIFICATION OF REAL-NUMBER AND BINARY PSO ALGORITHMS 217

Fig. 3. AIN behavior of the four methods over gradually increasing sample
runs in a 10D problem with 10 particles for (a) Rastigrin, (b) Griewank,
0
and (c) Rosenbrock functions and parameter range of [ 5; 5], and
0
Fig. 2. AFV behavior of the four methods over gradually increasing sample
runs in a 10D problem with 5, 10 and 20 particles for (a) Rastigrin, (b) Griewank, [ 50000; 50000].

0
and (c) Rosenbrock functions and parameter range of [ 500; 500].

resulting figures are grouped together as in Figs. 1–4. For the

PSO’s intrinsic robustness are sought. In real cases, it is almost sake of clarity, the figure in the right column of Fig. 4(c) as well
impossible to reduce the number of dimensions in order to speed as some other figures which will appear later in this section are
up the solution. However, decreasing the number of particles shown in logarithmic scale.
can drastically shorten the optimization time. Thus, we do not To summarize the results, the modified methods outperform
step beyond particles in the first stage of our study. the conventional one in terms of convergence speed for Rasti-
AIN/AFV results are shown in Figs. 1–4. grin function better than the two others, and in terms of the best
From Figs. 1–4, two interesting trends are discovered. The final value for Rosenbrock function. However, the cost of final
first one is the high similarity between the results of the three value error is only acceptable for high parameter ranges.
modified methods, particularly in terms of convergence speed. Figs. 1–4 reveal the AFV and the required AIN, but they do
The second common trend is that the modified methods mostly not explain the gradual convergence behavior, which is of great
expedite the convergence by almost 18% at the expense of importance in this study. Fig. 5 demonstrates the fitness value
higher errors for lower parameter ranges. However, as the variation during optimization process, averaged over 1000 runs
parameter range increases, the modified methods begin to for 10D problem with 10 particles. It is obvious that modified
slow down, while getting much lower errors with respect to methods are superior for the typical range with termi-
Method1. For Rastigrin and Griewank, this transition happens nation criterion of 500 iterations. For the case of Griewank, all
somewhere between and , whereas for Rosen- methods perform almost similarly.
brock it occurs somewhere between and ; i.e., in Since average fitness trends may not be adequate for pre-
addition to exhibiting encouraging convergence behavior, the senting the success rate, in Fig. 6 the probability density func-
modified methods outperform the conventional one, both in tion of the final fitness value in 1000 runs for 500 iteration sce-
speed and error rate, provided that the proper parameter ranges nario are plotted for Rastigrin function. In this figure, MFFV
are selected. For Griewank, is such a range. denotes the maximum final fitness value. To elaborate, the four
In order to cope with the realistic computation limitations of bars at 0.1MFFV show the probability of the final fitness value’s
a standard 2.83 GHz and 3.23 GB RAM PC running MATLAB, becoming less than or equal to 0.1MFFV, for the four methods.
the procedure is performed for each function separately and the Similarly, the bars at 0.2MFFV show the probability of the final
218 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO. 1, JANUARY 2011

Fig. 5. Fitness value of the four methods averaged over 1000 runs in a 10D
problem with 10 particles for (a) Rastigrin, (b) Griewank, and (c) Rosenbrock
0
functions and parameter range of [ 50000; 50000].

Fig. 4. AFV behavior of the four methods over gradually increasing sample
runs in a 10D problem with 10 particles for (a) Rastigrin, (b) Griewank,
0
and (c) Rosenbrock functions and parameter range of [ 5; 5], and
0
[ 50000; 50000].

fitness value’s becoming less than or equal to 0.2MFFV and

more than 0.1MFFV. Note that the probability of ending with
smaller fitness values increases in modified methods as the pa-
rameter range increases. Similar trend is observed by studying Fig. 6. Probability density function of the final fitness value for four methods,
other functions. For the sake of conciseness, we have omitted with 1000 runs in a 10D problem for Rastigrin function with parameter range
those results.
0 0 0
of (a) [ 50000; 50000], (b) [ 500; 500], and (c) [ 5; 5].

In addition to convergence performance, the behavior of par-

ticles has also been studied in [20] as yet another performance quency, refining the solution to improve a reflection coefficient
criterion. Fig. 7 depicts how the average position of a sample of dB to, let’s say dB, although of some mathemat-
particle in a sample dimension varies. The results are averaged ical interest, is of little practical interest. In fact, the minimiza-
over 1000 runs for Rastigrin function assuming tion criterion in these types of problems is met far in advance
, and parameter range of . In order to have a better view, of the actual minimum [23]. Thus, convergence speed to meet
two areas in Fig. 7(a) are zoomed out in figures (b) and (c). It is a design criterion is a more meaningful approach than an ap-
noteworthy that the modifications offer the particles a slightly proach that strives to achieve the global optimum at all cost. To
greater degree of freedom in the initial phase of the exploration, illustrate this point, in Appendix A, a microstrip antenna design
but faster convergence in later iterations. Similar trend is ob- scenario is discussed.
served by studying other sample particles, other dimensions, To further elaborate on the tradeoff between the optimization
and other functions. Again for the sake of brevity, those results speed and the error rate, let us introduce error as
are not included.
As it has been declared earlier, this study emphasizes on time- (12)
sensitive optimizations, where rather than achieving optimum
solution in long optimization procedures, expeditious conver- In (12), regardless of the parameter range, error is defined
gence to an acceptable solution is targeted ([21], [22]). For in- as the difference between the final fitness value achieved by
stance, when an antenna is designed to resonate at a desired fre- the optimization with termination criterion of 500 iterations
MODIRI AND KIASALEH: MODIFICATION OF REAL-NUMBER AND BINARY PSO ALGORITHMS 219

Fig. 9. The success rate trend of the four methods in 1000 runs in a
0
10D problem for Griewank function with parameter range of (a) [ 5; 5],
0 0
(b) [ 500; 500], and (c) [ 50000; 50000] with respect to the error value.

Fig. 7. (a) Average sample particle position in an arbitrary dimension for the
four methods over 1000 runs with 10 particles in a 10D problem for Rastigrin
0
function with parameter range of [ 500; 500], (b) zoom of 1–100 iterations,
and (c) zoom of 200–300 iterations.

Fig. 10. The success rate trend of the four methods in 1000 runs in a
0
10D problem for Rosenbrock function with parameter range of (a) [ 5; 5],
0 0
(b) [ 500; 500], and (c) [ 50000; 50000] with respect to the error value.

Fig. 8. The success rate trend of the four methods in 1000 runs in a
0
10D problem for Rastigrin function with parameter range of (a) [ 5; 5],
0 0
(b) [ 500; 500], and (c) [ 50000; 50000] with respect to the error value.

, and the global minimum , which is

zero for the selected benchmark functions. Figs. 8–10 depict
the percentage of the 1000 runs for which error falls below
certain prescribed value. To clarify the parameter range effect,
10-dimensional Rastigrin function, for instance, has maximum
value of 250 when the three parameters are in the range of
, whereas the maximum value increases to 25e9 for the
parameter range of . Therefore, the difficulty
of approaching the global minimum of zero in a wider solution
Fig. 11. Fitness value of the four methods averaged over 500 runs in a 10D
space should be kept in mind while inspecting the curves. problem with 5 particles for Rosenbrock function with parameter range of
As expected, a gradual movement toward higher success rate 0
[ 50000; 50000] using absorbing boundary condition.

is observed in Figs. 8–10 for the curves related to the modified

methods versus those of the conventional one as one increases
the parameter range. However, as it is also observed in Fig. 6, the convergence behavior for Rosenbrock function in the same
the above trend is more evident in Rosenbrock and Rastigrin 10D problem with 5 particles. It is apparent that the results are
functions than in Griewank’s. similar to those of the invisible wall. Figures related to the other
Furthermore, absorbing wall boundary condition for a 10D cases and functions are eliminated since they also show analo-
problem with 5, 10 and 20 particles is examined. Fig. 11 shows gous outcomes. Since it was concluded from previous study that
220 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO. 1, JANUARY 2011

Fig. 13. Symmetric half-wavelength thinned linear array.

In this section, we first study 2n-element linear half-wave-

length thinned antenna array design, using BPSO, assuming
. It has been proven for BPSO that the gradual de-
crease of has no effect on performance, thus, as recommended
in literature, we assume to be constantly equal to “1” [8].
The optimization goal is to minimize the side lobe level (SLL)
in the antenna’s broadside. The lobes are assessed in de-
grees angular distance from main lobe centre. The fitness func-
tion, shown in (13), is minimized
Fig. 12. Fitness value of the four methods averaged over 500 runs with 30 par-
ticles for (a) Rastigrin, (b) Griewank, and (c) Rosenbrock functions with param-
0
eter range of [ 50000; 50000]. (13)

The array factor formula can be found in many references

such as (26) in [8]. Assuming the antenna in xy plane, in this
500 runs are enough for collecting reliable average results, the paper, the SLL is assessed in the plane including the array itself
sample number is reduced to 500 runs. and broadside for linear array, but for planar arrays a
Finally, to provide a comparable analysis with studies
number of -planes are considered.
available in the literature, the parameters of the search are
In order to have consistent study, first, the convergence be-
chosen similar to those presented in [1] and [20]. It is assumed
that PSO has a population size of particles, and the haviors in cases similar to those encountered in RPSO are ob-
fitness functions have three parameters ranging in served in 1000 runs, and the average results associated with each
. The termination criterion is defined by a case are gathered in two figures depicting AIN and AFV values.
maximum iteration number of 500. Fig. 12 depicts the conver- Then, planar array antenna cases are studied.
gence behavior averaged over 500 runs. As the complexity of fitness function evaluation increases,
It is observed in Fig. 12 that the modified methods yield much the necessity of achieving better results in a fewer number of
better performance in terms of the convergence speed and final iterations becomes more obvious. In addition, as the number
fitness value for Rastigrin and Rosenbrock, but for Griewank, of particles reduces, the calculation time reduces significantly,
the results of all methods are similar. In comparison, TS-PSO although the chance of convergence to appropriate results also
algorithm requires more than 5000 iterations to achieve similar drops. Thus, a problem-dependent trade-off should be consid-
results for parameters with shorter range [20]. ered. Since this study considers large number of runs, the di-
It was shown in this section that the two modifications pro- mensionality of the problems are selected according to address
posed for PSO yield similar improvements in the RPSO algo- memory and time limitations.
rithm behavior. The modifications generally speed up the algo-
rithm. As the parameter range increases, the superiority of the A. 2n-Element Symmetric Linear Array
proposed modifications improves greatly both in convergence The linear arrays are assumed to be symmetric arrays with 2n
speed and final value, particularly for Rastigrin and Rosenbrock elements, as shown in Fig. 13.
functions. It is also shown that run number of less than hundreds
A 10D scenario with particle numbers of 5, 10
is not sufficient for making statistical decision.
and 20 are considered and AIN and AFV curves are shown
in Figs. 14–15. The performance of Method2 and Method4
V. EVALUATION OF THE PROPOSED VELOCITY in terms of AIN and AFV for BPSO (similar to the RPSO
FUNCTION IN BPSO case) is superior to Method1. However, although Methods2 and
Method3 for RPSO yielded superior performance as compared
In this section, we assess the performance of the proposed al- to the conventional method, only Method2 yielded superior
gorithm in BPSO scenario, which has been focus of some recent performance consistently for BPSO case, while Method3 of-
studies [19]. BPSO is highly useful in problems with two-state fered inconsistent performance advantage over the conventional
parameters, such as irregular shaped planar antenna design and method in our studies. The major reason for this distinction
array antenna design [8], but to extend its appliance to smart between RPSO and BPSO is due to insertion of the two-level
type of applications requiring fast resulting, it is not appropriate (binary) quantization [7]. In fact, =0 (Method3) may even
yet. This paper studies thinned antenna arrays which are peri- degrade the convergence performance unless it is used together
odic arrays with a number of non-illuminating elements. In addi- with proposed initial velocity (as in Method4). The figures
tion to general array applications, these types of antennas can be clearly show that the convergence speed (AIN) is significantly
used where software-defined beam-forming or pattern-shaping improved by Method2 and Method4, whereas the AFV gain is
is required by using the currents of elements as tuning factors. just slightly enhanced.
MODIRI AND KIASALEH: MODIFICATION OF REAL-NUMBER AND BINARY PSO ALGORITHMS 221

Fig. 14. AIN behavior of the four methods over gradually increasing sample
runs in a 10D problem with 5, 10 and 20 particles.

Fig. 17. Fitness value trend of the four methods averaged over 700 runs in a
16D problem with 2 particles. Optimization is performed in (a) 6 '-planes, (b) 8
'-planes, and (c) 10 '-planes.

Fig. 15. AFV behavior of the four methods over gradually increasing sample
runs in a 10D problem with 5, 10 and 20 particles.

Fig. 16. Non-symmetric thinned planar antenna array. Fig. 18. AIN and AFV behavior of the four methods averaged over gradually
increasing sample runs in a 9D problem with 3 particles.

Also, it is important to mention that since in BPSO each pa- the fitness function introduced in (13), resulting in the following
rameter can either be 1 or 0, parameter range will not be a con- fitness function:
siderable factor.

B. Non-Symmetric Planar Array (14)

Similar study is repeated for the more complicated case of Figs. 18–20 show the AIN/AFV, density, and average fitness
planar half-wavelength, non-symmetric thinned array antennas trend for the 3 3 array, respectively. The results confirm the
(Fig. 16). A problem is first considered. In order previous conclusions. Method2 and Method4 are superior to
to shorten the total run time, 100 iterations are performed in 700 the conventional method and Method3, both in terms of con-
runs, and the swarm size is limited to two particles. The number vergence speed and final value. Fig. 19 shows that the prob-
of runs is reduced to 700 from 1000 according to the study re- ability of achieving a smaller final fitness values is higher for
sults in Section A, showing reliable run number. The pattern is Method2 and Method4 as compared to other methods. An im-
assessed in 6, 8 and 10 -planes ( portant point underscored by Fig. 20 is that wherever the opti-
, and , respectively). The superiority mization procedure is terminated, Method2 and Method4 out-
of Method2 and Method4 both in terms of convergence speed perform the other two techniques. Similar trend is observed by
and final result is evident in Fig. 17. studying other cases, but, for the sake of conciseness, the results
Similarly, 3 3, 4 4 and 5 5 planar arrays are studied are not included here. It is also notable in Fig. 18 that 500 runs
with 3, 4 and 5 particles, respectively. This time, the pattern is are sufficient for concluding statistical results.
assessed in no more than 6 -planes , 200 As in RPSO, absorbing-wall scenario is investigated, and
iterations are performed in each of 700 runs. In order to alter similar outcomes are achieved. In Fig. 21, the optimization
the fitness function, a success boundary of dB is added to results associated with the 3 3 array with three particles
222 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO. 1, JANUARY 2011

a realistic software-defined thinned array scenario, a substan-

tially smaller number of runs will be required. Thus, signifi-
cance of a method or methods with achieved satisfactory results
in a smaller number of iteration becomes quite evident.

VI. CONCLUSION
This paper has studied modified PSO algorithms used for op-
timizing various problems. As the first modification to the con-
ventional PSO, the effect of omitting personal best coefficient in
velocity function, both in RPSO and BPSO, was studied. As the
second suggested modification, the impact of modifying the ini-
tial velocity on performance of PSO algorithm was investigated.
The modifications were shown to give better results for
Fig. 19. Probability density function of the final fitness value for four methods, benchmark mathematical optimizations with wide-range pa-
with 700 runs in a 9D problem with 3 particles in 6 '-planes.
rameters. In particular, superior performances were observed
for Rastigrin and Rosenbrock functions, as well as thinned
antenna array design. More specifically, we have shown that
the use of the proposed algorithm will result in the reduction of
the number of iterations by almost 18% to achieve considerably
better final result in RPSO, while for BPSO, the proposed
algorithm resulted in a reduction in the optimized SLL value
of the antenna by 1–2 dB for non-symmetric thinned planar
array. Although the SLL gain remained small for linear thinned
arrays, the convergence speed was higher when compared
with the conventional method in both linear and planar arrays,
making the proposed algorithm suitable for implementation in
software-define radios.

APPENDIX A
Fig. 20. Fitness value trend of the four methods averaged over 700 runs in a
9D problem with 3 particles in 6 '-planes. In this Appendix, fitness function for designing a simple rect-
angular patch antenna is examined in order to demonstrate the
similarity between the studied benchmark functions and those
of some practical problems. In fact, due to computation com-
plexity, numerical tools are applied to execute the electromag-
netic simulations of the class of problems discussed below.
Considering cavity model, which is introduced in various ref-
erences, such as in [24], the fitness function shown in (15) can be
used for a rectangular patch antenna design on a substrate
with thickness of . The design objective is to minimize the re-
flection at the feeding point by making the input impedance
as close as possible to a defined line impedance of . Let
Fig. 21. Fitness value trend of the four methods averaged over 500 runs in a 9D
problem with 3 particles in 6 '-planes using absorbing wall boundary condition. (15)

be the impedance mismatch function. In this equation

averaged over 500 runs, using absorbing boundary condition,
are depicted. It is apparent that the conclusions of invisible
boundary condition can be extended to absorbing wall case. The
overall results confirm that the proposed methods (Method2 (16)
and Method4) yield superior performance for BPSO.
Finally, to underscore the time-consuming nature of the opti- where
mization process, 2.8, 3.4, and 5.5 hours were required for the
700 runs of simple 4 4 array with only 2 particles for pattern
assessment in 6, 8 and 10 -planes, respectively, and angular
resolution of degrees. To underscore the com-
puting requirement, it is notable that adding two more particles
to 6 -plane case converts 2.8 hours of computing on a stan- and
dard 2.83 GHz and 3.23GB RAM PC to 5.25 hours. Assuming
MODIRI AND KIASALEH: MODIFICATION OF REAL-NUMBER AND BINARY PSO ALGORITHMS 223

In the above, is the feed position. and denote [7] S. Janson and M. Middendorf, “On trajectory of particles in PSO,”
the feeding area dimensions in and , respectively. m and n in Proc. IEEE Swarm Intelligence Symp. (SIS 2007), Apr. 2007, pp.
150–155.
correspond to mode resonating inside the cavity with the [8] N. Jin and Y. Rahmat-Samii, “Advances in particle swarm optimization
propagation constant . Furthermore, and denote the for antenna designs: Real-number, binary, single-objective and multi-
angular frequency, permeability and permittivity of the medium, objective implementations,” IEEE Trans. Antennas Propag., vol. 55,
no. 3, Mar. 2007.
respectively. The ground plane size is assumed to be infinite. [9] D. Gies and Y. Rahmat-Samii, “Particle swarm optimization for recon-
To have a better understanding, let’s assume the simplified figurable phase-differentiated array design,” Microwave Opt. Technol.
case of a square patch , with very small feeding area Lett., vol. 38, no. 3, pp. 172–175, Aug. 2003.
[10] D. Gies and Y. Rahmat-Samii, “Particle swarm optimization (PSO) for
, for which one can rewrite (15) as reflector antenna shaping,” in IEEE Antennas Propag. Soc. Int. Symp.
Dig., Jun. 2004, vol. 3, pp. 2289–2293.
[11] W. Liu, “Design of multiband CPW-fed monopole antenna using a par-
ticle swarm optimization approach,” IEEE Trans. Antennas Propag.,
(17) vol. 53, no. 10, pp. 3273–3279, Oct. 2005.
[12] S. Cui and D. Weile, “Application of parallel particle swarm optimiza-
tion scheme to the design of electromagnetic absorbers,” IEEE Trans.
where Antennas Propag., vol. 53, no. 11, pp. 3616–3624, Nov. 2005.
[13] N. Jin and Y. Rahmat-Samii, “Particle swarm optimization for multi-
band handset antenna designs: A hybrid real-binary implementation,”
in Proc. IEEE Antennas Propag. Soc. Int. Symp., Jul. 2008, pp. 1–4.
[14] S. Mikki and A. Kishk, “Investigation of the quantum particle swarm
and optimization technique for electromagnetic applications,” in IEEE An-
tennas Propag. Soc. Int. Symp. Dig., Jul. 2005, vol. 2A, pp. 45–48.
[15] J. Robinson and Y. Rahmat-Samii, “Particle swarm optimization in
electromagnetics,” IEEE Trans. Antennas Propag., vol. 52, no. 2, Feb.
2004.
Considering as an optimization parameter, the simi- [16] D. W. Boeringer and D. H. Werner, “Particle swarm optimization
versus genetic algorithms for phased array synthesis,” IEEE Trans.
larity of to benchmark functions becomes apparent. Generally, Antennas Propag., vol. 52, no. 3, Mar. 2004.
it is enough to consider the three first modes, i.e., [17] N. Jin and Y. Rahmat-Samii, “Parallel particle swarm optimization and
and . Therefore, the problem transforms to a 3-dimen- finite difference time-domain (PSO/FDTD) algorithm for multiband
and wide-band patch antenna designs,” IEEE Trans. Antennas Propag.,
sional problem where three parameters are going to be discov- vol. 53, no. 11, Nov. 2005.
ered by the optimization. Having these three parameters, the real [18] R. C. Eberhart and Y. Shi, “Particle swarm optimization: Develop-
unknowns of the problem, , and , are calculated in a set of ments, applications and resources,” in Proc. Congr. Evol. Comput.
(CEC01), 2001, vol. 1, pp. 81–86.
three equations. [19] F. Camci, “Analysis of velocity calculation methods in binary PSO on
In this problem, the parameters can normally vary in a wide maintenance scheduling,” in Proc. Applications of Digital Information
range, especially for the higher order modes (any mode other and Web Technologies ICADWT 2008, Aug. 4–6, 2008, pp. 12–17.
[20] S. Nakano, A. Ishigame, and K. Yasuda, “Particle swarm optimization
than ) where increases in value because of the pos-
based on the concept of tabu search,” in Proc. Conf. Evol. Comput.,
sibility of dominator approaching zero. In addition, it is evident CEC 2007, Sept. 2007, pp. 258–3263.
that the input impedance of the patch increases greatly when [21] F. A. Ali and K. T. Selvan, “A study of PSO and its variants in respect
being fed close to the edges, which further confirms high pa- of microstrip antenna feed point optimization,” in Proc. Asia Pacific
Microwave Conf. APMC 2009, Dec. 2009, pp. 1817–1820.
rameter range in (17). In other words, this problem is an example [22] J. Lu, D. Ireland, and A. Lewis, “Multi-objective optimization in high
of a practical design scenario in which an expeditious search to- frequency electromagnetic—An effective technique for smart mobile
wards minimum in a wide solution area is more desirable than terminal antennas (SMTA) design,” IEEE Trans. Magn., vol. 45, no. 3,
pp. 1072–1075, Mar. 2009.
a search for global optimum minimum. [23] A. Modiri and K. Kiasaleh, “Efficient design of microstrip antennas
using modified PSO algorithm,” presented at the 14th Biennial IEEE
Conf. on Electromagnetic Field Computation, CEFC 2010, May 2010.
[24] K. Carver and J. Mink, “Microstrip antenna technology,” IEEE Trans.
REFERENCES Antennas Propag., vol. 29, no. 1, pp. 2–24, Jan. 1981.

[1] A. Hoorfar, “Evolutionary programming in electromagnetic optimiza-

tion: A review,” IEEE Trans. Antennas Propag., vol. 55, no. 3, pp.
523–537, Mar. 2007.
[2] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.
IEEE Int. Conf. Neural Networks, 1995, vol. 4, pp. 1942–1948.
[3] M. Clerc and J. Kennedy, “The particle swarm—Explosion, stability
and convergence in a multidimensional complex space,” IEEE Trans.
Evol. Comput., vol. 6, no. 1, pp. 58–73, Feb. 2002. Arezoo Modiri (S’09) received the B.S. degree from
[4] L. Matekovits, M. Mussetta, P. Pirinoli, S. Selleri, and R. E. Zich, University of Tehran, Tehran, Iran and the M.S. de-
“Improved PSO algorithms for electromagnetic optimization,” in Proc. gree from the Iran University of Science and Tech-
IEEE Antennas Propag. Soc. Int. Symp., Jul. 2005, vol. 2A, pp. 33–36. nology, both in electrical engineering, in 2001 and
[5] Y. Shi and R. Eberhart, “Empirical study of particle swarm optimiza- 2005, respectively. She is currently working toward
tion,” in Proc. Cong. Evol. Comput., 1999, vol. 3, pp. 1945–1950. the Ph.D. degree at the University of Texas at Dallas.
[6] M. A. Montes de Oca, T. Stuzle, M. Birattari, and M. Dorigo, Her research interests include antennas, radiation
“Frankenstein’s PSO: A composite particle swarm optimization al- analysis, optimization methods, body area networks,
gorithm,” IEEE Trans. Evol. Comput., vol. 13, no. 5, pp. 1120–1132, reflectometry, and computational electromagnetics.
Oct. 2009.
224 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 59, NO. 1, JANUARY 2011

Kamran Kiasaleh (SM’95) received the B.S. (cum SDR, synchronization, novel detection schemes for phase noise and fading
laude), M.S., and Ph.D. degrees from the University impaired digital communication systems, optical communications, and wireless
of Southern California (USC), Los Angeles, in as well as optical CDMA systems.
1981, 1982, and 1986, respectively, all in electrical Dr. Kiasaleh was the recipient of the Research Initiation Award (RIA) from
engineering. the National Science Foundation (NSF) and the NASA/ASEE faculty fellowship
In 1986, he joined Teknekron Communication award at the Jet Propulsion Laboratory (JPL) in 1992, where he participated in
System, Berkeley, CA, where he was responsible for the Galileo Optical Experiment (GOPEX) demonstration, the first successful
conducting research on the advanced telecommuni- demonstration of an optical communications link involving a deep-space ve-
cations IC products, ASIC design (DS-1 Controller), hicle. For his participation in this project, he received the NASA Group Achieve-
and was involved in the development of network ment Award. In 1993, he was the recipient of NASA/ASEE faculty fellowship
control management systems. He was also a Lecturer award at JPL where he participated in Compensated Earth-Moon-Earth Laser
in the Department of Electrical Engineering, San Francisco State University, Link (CEMERLL) demonstration. Since 2006, he has led a research and de-
San Francisco, CA, in 1986. Since 1987, he has been with the Erik Jonsson velopment effort focused on various issues related to software-defined-radios
School of Engineering and Computer Science, University of Texas at Dallas (SDR) for first responders. This effort has been funded by the National Institute
(UTD), Richardson, where he is currently a Full Professor. From Dec. 1996 of Justice (NIJ). He was an Associate Editor for IEEE Communications Letters
to Dec. 1997, He was on a special assignment with the DSP research and (1999–2006). He is also a member of Eta Kappa Nu and Sigma Xi. He was a
development center at Texas Instruments Inc., Dallas, where he conducted faculty advisor to the student chapter of the IEEE Society at the UTD and served
research on various aspects of the 3rd generation wireless communication as the Vice President and the President of the Optical Society of North Texas
systems. From Sep. 2000 to Jan. 2003, he was with Optical Crossing Inc. (OCI) (OSNT) in 1990–1991 and 1991–1992, respectively.
where he was responsible for the development of the state-of-the-art free-space
and millimeter-wave communications systems. His research interests include