A comparison of particle swarm optimization and the genetic algorithm by Rania Hassan

American Institute of Aeronautics and Astronautics
1
A COPMARISON OF PARTICLE SWARM
OPTIMIZATION AND THE GENETIC ALGORITHM
Rania Hassan*
Babak Cohanim†
Olivier de Weck‡
Massachusetts Institute of Technology, Cambridge, MA, 02139
Gerhard Venter§
Vanderplaats Research and Development, Inc., Colorado Springs, CO, 80906
Particle Swarm Optimization (PSO) is a relatively recent heuristic search method whose
mechanics are inspired by the swarming or collaborative behavior of biological populations.
PSO is similar to the Genetic Algorithm (GA) in the sense that these two evolutionary
heuristics are population-based search methods. In other words, PSO and the GA move
from a set of points (population) to another set of points in a single iteration with likely
improvement using a combination of deterministic and probabilistic rules. The GA and its
many versions have been popular in academia and the industry mainly because of its
intuitiveness, ease of implementation, and the ability to effectively solve highly nonlinear,
mixed integer optimization problems that are typical of complex engineering systems. The
drawback of the GA is its expensive computational cost. This paper attempts to examine the
claim that PSO has the same effectiveness (finding the true global optimal solution) as the
GA but with significantly better computational efficiency (less function evaluations) by
implementing statistical analysis and formal hypothesis testing. The performance
comparison of the GA and PSO is implemented using a set of benchmark test problems as
well as two space systems design optimization problems, namely, telescope array
configuration and spacecraft reliability-based design.
Nomenclature
1c = self confidence factor
2c = swarm confidence factor
f = fitness function
g = constraint function
oH = null hypothesis
aH = alternative hypothesis
n = sample size
fevalN = number of function evalutions
g
kp = position of the particle with best global fitness at current move k
i
p = best position of particle i in current and all previous moves
r = penalty multiplier
rand = uniformly distributed rand variable between 0 and 1
*
Postdoctoral Associate, Aeronautics & Astronautics and Engineering Systems, rhassan@mit.edu, AIAA Member.
†
Formerly, Graduate Student, Center for Space Research, Aeronautics & Astronautics. Currently, Systems
Engineer, The Jet Propulsion Laboratory, bec@alum.mit.edu, AIAA Member.
‡
Assistant Professor, Aeronautics & Astronautics and Engineering Systems, deweck@mit.edu, AIAA Member.
§
Senior R&D Engineer, gventer@vrand.com, AIAA Member.
Copyright © 2004 by Rania Hassan, published by AIAA, with permission.

2
solQ = solution quality
)(xs = estimate of the standard deviation of a population from a sample
)(xs = the standard deviation of a sample
t = t-statistic
uv = position in the ultra violet plane
i
kv = position of particle i in the design space at time k
w = inertia factor
x = sample mean
x = design vector
i
kx = position of particle i in the design space at time k
α = significance level of a test
β = power of the test
µ = mean
t∆ = time increment
φ = objective function
Introduction
article Swarm Optimization (PSO) was invented by Kennedy and Eberhart1
in the mid 1990s while attempting
to simulate the choreographed, graceful motion of swarms of birds as part of a sociocognitive study
investigating the notion of “collective intelligence” in biological populations. In PSO, a set of randomly generated
solutions (initial swarm) propagates in the design space towards the optimal solution over a number of iterations
(moves) based on large amount of information about the design space that is assimilated and shared by all members
of the swarm. PSO is inspired by the ability of flocks of birds, schools of fish, and herds of animals to adapt to their
environment, find rich sources of food, and avoid predators by implementing an “information sharing” approach,
hence, developing an evolutionary advantage. References 1 and 2 describe a complete chronicle of the development
of the PSO algorithm form merely a motion simulator to a heuristic optimization approach.
The Genetic Algorithm (GA) was introduced in the mid 1970s by John Holland and his colleagues and students
at the University of Michigan.3
The GA is inspired by the principles of genetics and evolution, and mimics the
reproduction behavior observed in biological populations. The GA employs the principal of “survival of the fittest”
in its search process to select and generate individuals (design solutions) that are adapted to their environment
(design objectives/constraints). Therefore, over a number of generations (iterations), desirable traits (design
characteristics) will evolve and remain in the genome composition of the population (set of design solutions
generated each iteration) over traits with weaker undesirable characteristics. The GA is well suited to and has been
extensively applied to solve complex design optimization problems because it can handle both discrete and
continuous variables, and nonlinear objective and constrain functions without requiring gradient information.
The major objective of this paper is to compare the computational effectiveness and efficiency of the GA and
PSO using a formal hypothesis testing approach. The motivation is to validate or refute the widely speculated
hypothesis that PSO has the same effectiveness as the GA (same rate of success in finding true global optimal
solutions) but with better computational efficiency. The results of this test could prove to be significant for the future
development of PSO.
This remainder of this paper is organized in five major sections. First, the PSO and GA versions that are
implemented in this comparative study are summarized and the hypothesis test procedure is formulated. Second,
three well-known benchmark problems that are used to compare the performance of the GA and PSO are presented.
Third, two space systems optimization problems that are used to test the performance of both algorithms with
respect to real life applications are presented. Results and conclusions are presented in the last two sections.
PSO versus GA
Particle Swarm Optimization
In this study, the basic PSO algorithm that is described in Reference 4 is implemented. The basic algorithm is
first described, followed by a discussion on side and functional constraint handling, and finally, a discrete version of
the algorithm is presented. It should be noted that while the GA is inherently discrete, i.e. it encodes the design
P

3
variables into bits of 0’s and 1’s, therefore it easily handles discrete design variables, PSO is inherently continuous
and must be modified to handle discrete design variables.
The basic PSO algorithm consists of three steps, namely, generating particles’ positions and velocities, velocity
update, and finally, position update. Here, a particle refers to a point in the design space that changes its position
from one move (iteration) to another based on velocity updates. First, the positions, i
kx , and velocities, i
kv , of the
initial swarm of particles are randomly generated using upper and lower bounds on the design variables values,
minx and maxx , as expressed in Equations 1 and 2. The positions and velocities are given in a vector format with
the superscript and subscript denoting the th
i particle at time k . In Equations 1 and 2, rand is a uniformly
distributed random variable that can take any value between 0 and 1. This initialization process allows the swarm
particles to be randomly distributed across the design space.
( )minmaxmin0 xxxx −+= randi
(1)
( )
time
position
∆
minmaxmin
0 =
−+
=
t
randi xxx
v (2)
The second step is to update the velocities of all particles at time 1+k using the particles objective or fitness
values which are functions of the particles current positions in the design space at time k . The fitness function value
of a particle determines which particle has the best global value in the current swarm, g
kp , and also determines the
best position of each particle over time, i
p , i.e. in current and all previous moves. The velocity update formula uses
these two pieces of information for each particle in the swarm along with the effect of current motion, i
kv , to
provide a search direction, i
k 1+v , for the next iteration. The velocity update formula includes some random
parameters, represented by the uniformly distributed variables, rand , to ensure good coverage of the design space
and avoid entrapment in local optima. The three values that effect the new search direction, namely, current motion,
particle own memory, and swarm influence, are incorporated via a summation approach as shown in Equation 3 with
three weight factors, namely, inertia factor, w , self confidence factor, 1c , and swarm confidence factor, 2c ,
respectively.
( ) ( )
t
randc
t
randcw
i
k
gi
k
i
i
k
i
k
k
∆
−
+
∆
−
+=+
xpxp
vv 211 (3)
The original PSO algorithm1
uses the values of 1, 2 and 2 for w , 1c , and 2c respectively, and suggests upper
and lower bounds on these values as shown in Equation 3 above. However, the research presented in this paper
found out that setting the three weight factors w , 1c , and 2c at 0.5, 1.5, and 1.5 respectively provides the best
convergence rate for all test problems considered. Other combinations of values usually lead to much slower
convergence or sometimes non-convergence at all. The tuning of the PSO algorithm weight factors is a topic that
warrants proper investigation but is outside the scope of this work. For all the problems investigated in this work, the
weight factors use the values of 0.5, 1.5 and 1.5 for w , 1c , and 2c respectively.
Position update is the last step in each iteration. The Position of each particle is updated using its velocity vector
as shown in Equation 4 and depicted in Figure 1.
tiii
k kk
∆+=
+
+ 1
1 vxx (4)
velocity of
particle i at time k+1
current
motion
particle memory
influence
swarm influence
inertia factor
range: 0.4 to 1.4
self confidence
range: 1.5 to 2
swarm confidence
range: 2 to 2.5

4
The three steps of velocity update, position update, and fitness calculations are repeated until a desired
convergence criterion is met. In the PSO algorithm implemented in this study, the stopping criteria is that the
maximum change in best fitness should be smaller than specified tolerance for a specified number of moves, S , as
shown in Equation 5. In this work, S is specified as ten moves and ε is specified as 10-5
for all test problems.
( ) ( ) ,..S,qff g
qk
g
k 21=≤− − εpp (5)
Unlike the GA with its binary encoding, in PSO, the design variables can take any values, even outside their side
constraints, based on their current position in the design space and the calculated velocity vector. This means that
the design variables can go outside their lower or upper limits, minx or maxx , which usually happens when the
velocity vector grows very rapidly; This phenomenon can lead to divergence. To avoid this problem, in this study,
whenever the design variables violate their upper or lower design bounds, they are artificially brought back to their
nearest side constraint. This approach of handling side constraints is recommended by Reference 4 and is believed to
avoid velocity “explosion”.2
Functional constraints on the other hand are handled in the same way they are handled
in the GA. In this comparative study, functional constraints are handled using a linear exterior penalty function
approach as shown in Equation 6.
( ) ( ) ( )[ ]∑
=
+=
conN
i
ii ,grf
1
0max xxx φ (6)
In addition to applying penalty functions to handle designs with violated functional constraints, in PSO, it is
recommended that the velocity vector of a particle with violated constraints be reset to zero in the velocity update
formula as shown in Equation 7.4
This is because if a particle is infeasible, there is a big chance that the last search
direction (velocity) was not feasible. This approach is implemented in this study.
( ) ( )
t
randc
t
randc
i
k
gi
k
i
i
k
k
∆
−
+
∆
−
=+
xpxp
v 211 (7)
i
k
x
g
k
p
swarm influence
i
p
particle memory
influence
i
k
v
current motion
influence
i
k 1+
x
i
k 1+
v
Figure 1. Depiction of the velocity and position updates in
Particle Swarm Optimization.

5
In many design applications, especially complex systems, the optimization problem statement usually includes
discrete design variables, such as technology choices. Reference 4 suggests a simple but effective way to implement
discrete design variables with PSO, that is to round particle position coordinates (design variable values) to the
nearest integers when the design variables being investigated are discrete. This is the strategy that is implemented in
this study.
There has been no recommendation in the literature regarding swarm size in PSO. Most researchers use a swarm
size of 10 to 50 but there is no well established guideline.2
For the purpose of comparing PSO and GA, the swarm
size that will be used for the PSO runs in all test problems will be the same as the population size in their equivalent
GA runs and is fixed at 40 particles in the PSO swarm and 40 chromosomes in GA population.
The Genetic Algorithm
The literature includes many versions of the Genetic Algorithm (GA). In this study, a basic binary encoded GA5
with tournament selection, uniform crossover and low probability mutation rate is employed to solve the benchmark
problems and the space systems design problems. The GA represents the design variables of each individual design
with binary strings of 0’s and 1’s that are referred to as chromosomes. It is important to note that the GA works with
a coding for the design parameters that allows for a combination of discrete and continuous parameters in one
problem statement. This encoding feature also forces the design variables to only take values that are within their
upper and lower bounds, i.e., no solutions will ever violate the side constraints and infeasibility can only occur
because of violation of functional constraints.
Like PSO, the GA begins its search from a randomly generated population of designs that evolve over successive
generations (iterations), eliminating the need for a user-supplied starting point. To perform its optimization-like
process, the GA employs three operators to propagate its population from one generation to another. The first
operator is the “Selection” operator that mimics the principal of “Survival of the Fittest”. The second operator is the
“Crossover” operator, which mimics mating in biological populations. The crossover operator propagates features of
good surviving designs from the current population into the future population, which will have better fitness value
on average. The last operator is “Mutation”, which promotes diversity in population characteristics. The mutation
operator allows for global search of the design space and prevents the algorithm from getting trapped in local
minima. Many references explain the details of the GA mechanics of evolution; see, for example, Reference 3. The
specifics of the GA version implemented in this study are partially based on empirical studies developed in
Reference 5, which suggests the combination of tournament selection with a 50% uniform crossover probability. For
the comparison with PSO, the population size is kept constant at 40 chromosomes for all problems under
consideration. A small, fixed mutation of 0.5% is implemented.
The basic GA version implemented in this study uses a linear exterior penalty function approach to handle
functional constraints similar to PSO as shown in Equation 6 above. The GA also uses the same convergence criteria
as PSO as shown in Equation 5.
Comparison Metrics and Hypothesis Testing
The objective of this research is to statistically compare the performance of the two heuristic search methods,
PSO and the GA, using a representative set of test problems that are of diverse properties. The t-test (hypothesis
testing) is used to assess and compare the effectiveness and efficiency of both search algorithms. In hypothesis
testing, a null hypothesis, oH will be correctly accepted with a significance (or confidence) level ( )α−1 and
falsely rejected with a type I error probability α . If the null hypothesis is false, it will be correctly rejected with a
power of the test ( )β−1 and will be falsely accepted with a type II error probability β . The decision options
summary is presented in Table 1. aH corresponds to an alternative hypothesis that is complimentary to oH .
Table 1. Possible decision outcomes in hypothesis testing (adapted from Reference 6).
The true situation may be
Action
oH is true oH is false
accept oH , reject aH ( )α−1 significance level β [type II error]
reject oH , accept aH α [type I error] ( )β−1 power of test
Sum 1 1

6
The t-test deals with the estimation of a true value from a sample and the establishing of confidence ranges
within which the true value can said to lie with a certain probability ( )α−1 . In hypothesis testing, increasing the
sample size, n, decreases type I and II error probabilities, α and β , based on the t-distribution. When n is very
large, the t-distribution approaches the normal distribution.7
Hypothesis testing involves five steps. First, the null hypothesis and the alternative hypothesis are defined. The
hypotheses could be two sided or single sided. For example, mH =µ:o (with mH a ≠µ: ) is a two sided
hypothesis testing whether an unknown population mean, µ , is equal to the value m or not. On the other hand,
mH >µ:o (with mH a ≤µ: ) is a single sided hypothesis. The second step is to decide on desired values for α ,
β , and n and to find out the corresponding criticalt value for the given α , β , and n parameters for the specific
type of test under consideration (single or two sided). The criticalt value could be obtained from t-distribution tables
that are available in many basic statistics textbook such as Reference 7. The third step is to evaluate n random
samples and evaluate the sample mean, x , an estimate of the standard deviation of a population from the sample,
)(xs , and finally, the sample mean, )(xs , as shown in Equations 8 through 10.
n
x
x
n
i
i∑
=
= 1
(8)
1
)(
2
11
2
−
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
=
∑∑
==
n
nxx
xs
n
i
i
n
i
i
(9)
n
xs
xs
)(
)( = (10)
The fourth step in the hypothesis testing process is to calculate the t-value of the null hypothesis as shown in
Equation 11. The formula for calculating the t-value differs according to the hypothesis being test. The form shown
in Equation 11 is equivalent to the two sided test mH =µ:o (with mH a ≠µ: ). Reference 7 lists the most
important t-value calculation formulas.
)(xs
mx
t
−
= (11)
The final step in the hypothesis testing process is to compare the calculated t-value to the tabulated criticalt
value. If criticaltt ≤ , the null hypothesis is then accepted with ( )α−1 confidence level.
In this study, two hypotheses are tested. The first test is related to the effectiveness (finding the true global
optimum) of the algorithms and the second is related to the efficiency (computational cost) of the algorithms. The
first test is the high effectiveness test. Effectiveness is defined as the ability of the algorithm to repeatedly find the
known global solution or arrive at sufficiently close solutions when the algorithm is started from many random
different points in the design space. In other words, effectiveness is defined as a high probability of finding a high
quality solution. Here, the quality of a solution is measured by how close the solution is to the known global solution
as shown in Equation (12).
%
solutionknown
solutionknown-solution
=solQ (12)

7
The solution quality metric described in Equation 12 could then be used to synthesize a meaningful hypothesis to
test the effectiveness of the search algorithms, PSO and the GA as shown below. It must be notes that the
effectiveness test must be carried out for each of the algorithms separately. In other words, this test measures the
effectiveness of each algorithm with respect to known solutions for the test problem rather than comparing the
effectiveness of the two algorithms to each others. This effectiveness test will be carried out for each algorithm with
respect to each of the test problems separately.
Effectiveness Test
Objective to test whether %99: solQ >µaH
%99: solQ ≤µoH
)(
%99
sol
sol
Qs
Q
t
−
=
taking %1=α , %1=β , and 10=n
This is a one sided test of significance of a mean (table 6.10, Reference 7) → 0.2=criticalt
The second hypothesis that is tested in this study, which is the important test from the authors’ point of view, is
the computational efficiency test. This test directly compares the computational effort required by PSO and the GA
to solve each of the test problems. The objective is to support or refute the widely speculated claim that PSO is
significantly more efficient than the GA. This direct comparison requires a t-test called “comparison of two means”.
For a test with the null hypothesis 21: µµ ≤oH , the t-value can be calculated as shown in Equation (13).
21
21
11)( nnxs
xx
t
+
−
= (13)
where
2
)()1()()1(
)(
21
2
2
21
2
1
−+
−+−
=
nn
xsnxsn
xs (14)
For the computational efficiency test, the metric that is implemented is the number of function evaluations,
fevalN , the algorithm performed until the convergence criteria (described in Equation 5) is met. The lower fevalN
is, the more efficient the algorithm. The efficiency test is summarized as follows.
Efficiency Test
Objective to test whether evaleval NNaH µµ GAPSO: <
evaleval NNH µµ GAPSO: ≥o
PSOGA 11)(
PSOGA
nnxs
t
fevalfeval NN
+
−
=
µµ
where
2
)1()1(
)(
PSOGA
2
PSOPSO
2
GAGA
−+
−+−
=
nn
snsn
xs
taking %1=α , %1=β , and 10GAPSO == nn
This is a one sided test of significance of comparison of two means (table 6.11, Reference 7) → 5.2=criticalt
Benchmark Test Problems
In this section, a set of well-known optimization benchmark test problems are presented. These problems will be
solved by both PSO and the GA, ten times each, to carry out the hypothesis testing procedure as discussed in the
previous section. This set of problems includes two simple, unconstrained continuous functions, namely, the Banana

8
(Rosenbrock) function and the Eggcrate function; each has only two design variables. A more complex and highly
constrained benchmark problem, namely, Golinski’s Speed Reducer, is also presented; it has seven design variables
and 11 functional constraints. The solutions for these test problems are known and will be used to evaluate the
effectiveness of the two algorithms, PSO and the GA.
The Banana (Rosenbrock) Function
This function is known as the “banana function” because of its shape; it is described mathematically in Equation
15. In this problem, there are two design variables with lower and upper limits of [-5 , 5]. The Rosenbrock function
has a known global minimum at [1 , 1] with an optimal function value of zero.
Minimize ( ) ( ) ( )2
1
22
12 1100 xxxf −+−=x (15)
The Eggcrate Function
This function is described mathematically in Equation 16. In this problem, there are two design variables with
lower and upper bounds of [-2π , 2π]. The Eggcrate function has a known global minimum at [0 , 0] with an optimal
function value of zero.
Minimize ( ) ( )2
2
1
22
2
2
1 sinsin25 xxxxf +++=x (16)
Golinski’s Speed Reducer
This problem represents the design of a simple gearbox such as might be used in a light airplane between the
engine and propeller to allow each to rotate at its most efficient speed. The gearbox is depicted in Figure 2 and its
seven design variables are labeled. The objective is to minimize the speed reducer’s weight while satisfying the 11
constraints imposed by gear and shaft design practices. Full problem description can be found in Reference 8. A
known feasible solution obtained by a sequential quadratic programming (SQP) approach (MATLAB’s fmincon) is
a 2994.34 kg gearbox with the following optimal values for the seven design variables: [3.5000 0.7000 17.0000
7.3000 7.7153 3.3502 5.2867 ]. This is a feasible solution with four active constraints. It is important to note
that the published solution in Reference 8 is not a feasible solution as is widely believed because it slightly violates
the 5th
, 6th
and 11th
functional constraints.9
Space Systems Problems
This section presents two space system design problems; the first problem has a scientific application while the
second focuses on commercial services. The two problems presented in this section, namely, ground-based radio
telescope array design, and communication satellite design, are used as test problems to compare the performance of
the two algorithms, PSO and GA, with respect to real life engineering problems.
Telescope Array Configuration
Ground-based radio telescope arrays use several distributed small telescopes to achieve the resolution of a costly
single large telescope by correlating data to create a synthesized beam. A measure of performance for correlation
Figure 2. Golinski’s Speed Reducer.

9
arrays is the degree of coverage in the instantaneous uv plane. The uv plane is defined by Equation 17 where x
and y are the ground positions of the stations measured in a convenient set of units such as kilometers.
ijij xxu −= , ijij yyv −= (17)
If autocorrelation points are removed (the correlation of data from one telescope station with itself), the number
of visibility points, uvN is given by Equation 18.
( )1−= stationsstationsuv NNN (18)
Each uv point samples the Fourier transform of the sky brightness.10
The more uniform the uv coverage is, the
lower the undesirable side lobes in the synthesized beam of the array.11, 12
Cornwell11
proposed a metric for the uv
coverage as the distance between all uv points as shown in Equation 19. This metric is implemented as the an
objective function in the
Maximize ( ) ( )∑ −+−=
lkji
klijklij vvuuuv
,,,
22
coverage (19)
Another objective function that is usually considered is the minimization of the distance between the different
stations as these distances represent the cost of the fiber optic cable needed to connect the stations. These are two
opposing objectives because maximizing the uv coverage requires the stations to be placed as far as possible from
each other. In this case study, only the uv coverage maximization metric is considered. Regardless of the number of
stations, the solution of this single objective, unconstrained problem is known to place all available stations on a
perimeter of a circle. The design variables then become the angles that describe the location the station on the circle.
Consequently, the number of design variables is equal to the number of stations in the telescope array to be
designed. In this research, four cases are investigated; these are: five, six, seven and eight station telescope arrays.
The solutions for these cases are given in Reference 11, and were verified using an SQP approach (MATLAB’s
fmincon) for use in the PSO and GA effectiveness tests. These known solutions are shown in Figure 3 below.
Figure 3. Optimal placement of telescope array stations and the coressponding uv coverage.
-4 -2 0 2 4
-4
-2
0
2
4
x
y
-5 0 5
-5
0
5
u
v
-4 -2 0 2 4
-4
-2
0
2
4
x
y
-5 0 5
-5
0
5
u
v
Nstations = 5, uv coverage = 2418.90 Nstations = 6, uv coverage = 5347.90
-4 -2 0 2 4
-4
-2
0
2
4
x
y
-5 0 5
-5
0
5
u
v
-4 -2 0 2 4
-4
-2
0
2
4
x
y
-5 0 5
-5
0
5
u
v
Nstations = 7, uv coverage = 10337.00 Nstations = 8, uv coverage = 18196.00

10
Communication Satellite Reliability-Based Design
This problem involves designing the payload and bus subsystems of a commercial communication
Geosynchronous satellite with given payload requirements. The design objective is to minimize the spacecraft
overall launch mass, which is a surrogate for cost, given design constraints on payload as well as overall system
reliability. The problem also involves geometrical constraints imposed by the choice of the launch vehicle. The
problem includes six functional constraints and 27 discrete design variables representing the technology choices and
redundancy levels of the satellite payload and bus subsystems. For a complete description of the problem, see
References 13, 14, and 15. The best solution found for this discrete problem is satellite with a an optimal mass of
3415.8 kg. The above references describe the technology choices and redundancy levels of the different subsystems
in this optimal spacecraft design.
Results and Discussions
Two performance tests were carried out for both algorithms under consideration, namely, PSO and the GA. The
performance tests were carried out for eight sample problems: the Banana function, the Eggcrate function,
Golinski’s Speed Reducer, Reliability-based Satellite Design, 5-station, 6-station, 7-station and 8-station Radio
Telescope Array Configuration. The first test is the effectiveness test, which measures the quality of the solutions
found by the heuristic algorithm with respect to known solutions for the test problems. This test investigates whether
the quality of the solutions obtained is greater than 99%. The second test is the efficiency test, which investigates
whether the computational effort of PSO is less than that of the GA for the sample problem set using the same
convergence criteria. These statements of the two tests were set as the alternative hypotheses while their
complimentary statements were set as the null hypotheses, which is a traditional hypothesis testing approach. Both
tests were conducted using acceptable Type I and Type II errors of 1% each. The t-statistic values that were
calculated for the PSO runs and GA runs for all eight sample problems are summarized in Table 2 below. The
sample problems below are organized in order of increasing number of design variables, a measure of complexity.
Table 2: Calculated t-values for the effectiveness and efficiency hypotheses tests
Effectiveness Test, 0.2=criticalt
Calculated t-value
Efficiency Test,
5.2=criticalt
Calculated t-value
PSO GA
1- Banana Function 8404.05 -2.49 1.9980
2- Eggcrate Function 312420.20 5214.34 7.2743
3- Telescope Array (5 Stations) ∞ ∞ 19.2995
7- Golinski’s Speed Reducer -0.39 1251.28 10.3079
8- Satellite Design -2.13 6.47 19.6993
The effectiveness tests for PSO and the GA show that criticaltt > in most sample test problems. This leads to
the rejection of the null hypothesis and the acceptance of the alternative hypothesis, that is the quality of the
solutions of both approaches is equal to or greater than 99% in most of the test problems. This alternative hypothesis
is accepted with a confidence level of 99%. The infinity t-values in the telescope array problems are obtained
because in each of the runs, both PSO and the GA consistently found the known solutions for the 5, 6, 7, and 8
stations telescope array problems. Therefore, the quality of the ten solutions in each case is 100% and the standard
deviation is zero.
In only one case, the Banana function, the null hypothesis is accepted for the GA runs, therefore it cannot be
proved that the GA solutions have high quality for this problem. By further investigating the data, it is found that
mean quality of the GA solutions for this problem is 91% with a standard deviation of 9.5%. This mean is much
lower than the 99% quality level implemented in the test. It is surprising that the GA could find high quality
solutions for the rest of the test problems with continuous design variables and with larger number of design

11
variables (problems 2 to 7) but were not able to obtain the same quality for the Banana problem. A larger sample
size might lead to improving the mean and standard deviation of GA solution quality for the Banana problem.
PSO solutions for all test problems exceeded the 99% quality limits except for the last two problems. Further
investigation of the data reveals that the mean quality of the PSO solutions for Golinski’s speed reducer is 98.8%
with a standard deviation of 1.7%. The scatter in the data of the sample does not allow for the quality limit of 99% to
be met with sufficient confidence. As for the satellite design problem, it is expected that PSO does not perform well
on this problem given the fact that all the design variables are discrete and that PSO is inherently continuous. Figure
4 below summarized the values implemented in the effectiveness test for both PSO and the GA.
80%
85%
90%
95%
100%
Banana
Function
Eggcrate
Function
Telescope
Array(5)
Telescope
Array(6)
Telescope
Array(7)
Telescope
Array(8)
Golinski's
Speed
Reducer
Satellite
Design
SolutionQuality
PSO mean +/- standard
deviation
GA mean +/- standard
deviation
Figure 4. The mean quality of the solutions obtained by PSO and the GA using 10 run samples
for eight test problems.
The efficiency test shows criticaltt > for all test problems except for the Banana function. These results lead
to the rejection of the null hypothesis and the acceptance of the alternative hypothesis with a confidence level of
99%. The interpretation of these results is that for all sample problems, except for the Banana function, the
computational effort required by PSO to converge to a solution is less than that of the GA 99% of the time using the
same convergence criteria. From Figure 5, it appears that the scatter of the GA runs for the Banana function is the
reason behind the acceptance of the null hypothesis that PSO does not outperform the GA for the Banana function as
it did in the rest of the test problems.
0
5000
10000
15000
20000
25000
30000
35000
40000
Banana
Function
Eggcrate
Function
Telescope
Array(5)
Telescope
Array(6)
Telescope
Array(7)
Telescope
Array(8)
Golinski's
Speed
Reducer
Satellite
Design
No.ofFunctionEvaluations
PSO mean +/- standard deviation
GA mean +/- standard deviation
Figure 5. The mean computational effort of the solutions obtained by PSO and the GA using 10 run
samples for eight test problems.

12
Figure 6 below couples the results of the effectiveness and efficiency tests. It shows the ratios of the mean values
of the solution qualities and computational effort of the GA and PSO. The insight learned from Figure 6 is that for a
range of test problems with variable complexity, PSO and the GA were able to arrive at solutions with the same
quality, which is indicated by the flat line of a ratio of 1 for the means of the solution qualities. However, the
varying ratios of the computational effort provides two conclusions: first, PSO does offer a less expensive approach
than the GA in general, second, the expected computational savings offered by PSO over the GA is problem
dependent. It appears that PSO offers more computational savings for unconstrained nonlinear problems with
continuous design variables whereas the computational savings is lower for constrained and mixed integer noninear
problems.
0
1
2
3
4
5
6
7
Banana
Function
Eggcrate
Function
Telescope
Array(5)
Telescope
Array(6)
Telescope
Array(7)
Telescope
Array(8)
Golinski's
Speed
Reducer
Satellite
Design
PerformanceRatio(GA/PSO)
Ratio of means of No. of function evalutions
Ratio of means of quality of solutions
Figure 6. Comparison of ratios of solution quality and computational effort of PSO and the GA.
Conclusions
Particle Swarm Optimization (PSO) is a relatively recent heuristic search method that is based on the idea of
collaborative behavior and swarming in biological populations. PSO is similar to the Genetic Algorithm (GA) in the
sense that they are both population-based search approaches and that they both depend on information sharing
among their population members to enhance their search processes using a combination of deterministic and
probabilistic rules. Conversely, the GA is a well established algorithm with many versions and many applications.
The objective of this research is to test the hypothesis that states that although PSO and the GA on average yield the
same effectiveness (solution quality), PSO is more computationally efficient (uses less number of function
evaluations) than the GA. To investigate this claim, two statistical tests were set to examine the two elements of this
claim, equal effectiveness but superior efficiency for PSO over the GA.
To carry out the t-tests, eight sample test problems were solved using both PSO and the GA over multiple runs.
The test problems includes three well-known benchmark test problems; these are: the Banana (Rosenbrock function,
the Eggcrate Function, and Golinski’s Speed Reducer. Two space systems design problems were also investigated to
test the algorithms on real-life engineering problems. The first problem is the configuration of a ground-based multi-
station radio telescope array. Four versions of this problem were investigated including 5, 6, 7 and 8-station arrays.
The second test problem involves the reliability-based design of a commercial communication satellite. All test
problems involve continuous design variables only except for the satellite design problem which contains discrete
design variables representing subsystems technology choices and redundancy levels. The eight test problems
represent a wide range of complexity, nonlinearity, and constraint levels.
Two metrics were identified for the two t-tests. The effectiveness test for both PSO and the GA uses a quality of
solution metric that measures the normalized difference between the solutions obtained by the heuristic approaches
and known solutions of the test problems. The efficiency test uses the number of function evaluations needed by the
heuristic approaches to reach convergence. The same convergence criterion is enforced on PSO and the GA.
The results of the t-tests support the hypothesis that while both PSO and the GA obtain high quality solutions,
with quality indices of 99% or more with a 99% confidence level for most test problems, the computational effort

13
required by PSO to arrive to such high quality solutions is less than the effort required to arrive at the same high
quality solutions by the GA. The results further show the computational efficiency superiority of PSO over the GA
is statically proven with a 99% confidence level in 7 out of the 8 test problems investigated. Further analysis shows
that the difference in computational effort between PSO and the GA is problem dependent. It appears that PSO
outperforms the GA with a larger differential in computational efficiency when used to solve unconstrained
nonlinear problems with continuous design variables and less efficiency differential when applied to constrained
nonlinear problems with continuous or discrete design variables. The authors encourage the readers to investigate
the results presented in this paper to help set the agenda for further PSO development efforts.
References
1
Kennedy, J. and Eberhart, R., “Particle Swarm Optimization,” Proceedings of the IEEE International Conference
on Neural Networks, Perth, Australia 1995, pp. 1942-1945.
2
Kennedy, J. and Eberhart, R., Swarm Intelligence, Academic Press, 1st ed., San Diego, CA, 2001.
3
Goldberg, D., Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, Reading, MA,
1989, pp. 1-25.
4
Venter, G. and Sobieski, J., “Particle Swarm Optimization,” AIAA 2002-1235, 43rd AIAA/ASME/ASCE/ AHS/ASC
Structures, Structural Dynamics, and Materials Conference, Denver, CO., April 2002.
5
Williams, E. A., and Crossley, W. A., “Empirically-Derived Population Size and Mutation Rate Guidelines for a
Genetic Algorithm with Uniform Crossover,” Soft Computing in Engineering Design and Manufacturing, P. K.
Chawdhry, R. Roy and R. K. Pant (editors), Springer-Verlag, 1998, pp. 163-172.
6
Harnett, D., Statistical Methods, 3rd
ed., Addison-Wesley, 1982, pp.247-297.
7
Volk, W., Applied Statistics for Engineers, 2nd
ed., McGraw-Hill, 1969, pp. 109-148.
8
Web site http://mdob.larc.nasa.gov/mdo.test/class2prob4.html cited August 16th, 2004.
9
Ray, T., “Golinski’s Speed Reducer Problem Revisited,” the AIAA Journal, Vol. 41, No. 3, 2003, pp. 556 -558.
10
Thompson, A., Moran, J., and Swenson, G., Interferometery and Synthesis in Radio Astronomy, Wiley
Interscience, New York, 1986.
11
Cornwell, T., “Novel Principle for Optimization of the Instantaneous Fourier Plane Coverage of Correlation
Arrays,” IEEE Transactions on Antennas and Propagation, AP, Vol. 36, 1988, pp. 1165.
12
Keto, E., “Shapes of Cross-Correlation Interferometers,” APJ, Vol. 475, 1997, pp. 843.
13
Hassan, R., Genetic Algorithm Approaches for Conceptual Design of Spacecraft Systems Including Multiobjective
Optimization and Design under Uncertainty, doctoral thesis, Purdue University, May 2004.
14
Hassan, R., and Crossley, W., “Multiobjective Optimization of Communication Satellites with a Two-Branch
Tournament Genetic Algorithm,” Journal of Spacecraft & Rockets, Vol. 40, No. 2, 2003, pp. 266-272.
15
Hassan, R., and Crossley, W., “Variable Population-Based Sampling for Probabilistic Design Optimization with a
Genetic Algorithm,” AIAA-2004-0452, 42nd Aerospace Sciences Meeting, Reno, NV, January 2004.

A comparison of particle swarm optimization and the genetic algorithm by Rania Hassan

More Related Content

Viewers also liked (15)

Similar to A comparison of particle swarm optimization and the genetic algorithm by Rania Hassan (20)

Recently uploaded (20)

A comparison of particle swarm optimization and the genetic algorithm by Rania Hassan