PSOt - a Particle Swarm Optimization Toolbox for use with Matlab

Brian Birge
NCSU, MAE Dept.
726 N. West St., #B
Raleigh, NC 27603
birgeb@bellsouth.net

Abstract - A Particle Swarm Optimization Toolbox (PSOt) for Computational Intelligence arena. With the Particle Swarm
use with the Matlab scientific programming environment has Optimization Toolbox (PSOt), users can apply the algorithm
been developed. PSO is introduced briefly and then the use of to their problem without having to write their own PSO code
the toolbox is explained with some examples. A link to or really even understand much about PSO. Users can train
downloadable code is provided.
an Artificial Neural Network (ANN) with PSO and/or apply
PSO to their more general problem.
I. GENERAL INFORMATION
This paper originally came out of an effort to replace standard
A. Particle Swarm Optimization Toolbox (PSOt), Summary of
backpropagation trained ANNs with PSO in order to make
Included Files
better ANNs and was inspired by the freely available Matlab
implementation of a Genetic Algorithm function optimizer by
Main files:
Houck, Joines, and Kay referenced at the end of this paper
1) PSO – finds min/max of arbitrary MISO functions
[3].
using PSO
2) trainPSO – Function to train neural nets using PSO
It is hoped this suite of programs will be useful to classroom
3) tpso1,2,3 – called by trainPSO for training 0 hidden
lecturers wishing to add a PSO unit as well as fellow
layer, 1 hidden layer, and 2 hidden layer ANNs
researchers in Computational Intelligence.
respectively.
C. Matlab
Support files:
4) wrapmat – convert vector(s) into matrix(ces)
Matlab is a scientific computing language developed by
5) unwrapmat – convert any 2d matrix(ces) into a row
Mathworks that is run in interpreter mode on a wide variety
vector(s)
of operating systems. It is extremely powerful, simple to use,
6) normalize – takes a 2D matrix and reformats it to
and can be found in most research and engineering
any specified range
environments. It gets more powerful and specialized with the
7) goplotpso – called by trainpso for graphical display
addition of ‘toolboxes’, additional functions added to the
during training
Matlab environment by various developers for specific tasks
or fie lds. For example, the Neural Net toolbox adds functions
Demo and miscellaneous files:
that allow one to develop ANNs and the Control System
8) DemoTrainPSO – shows an example of using PSO
toolbox includes many traditional engineering controls
to train a neural net to the XOR function
functions. The PSOt aims to add not only stand alone
9) f6 – Schaffer’s f6 function
function optimization using PSO but also the ability to
10) hypara – 4 dimensional hyperbolic paraboloid
interface easily with the Neural Net toolbox, training ANNs
function
with PSO rather than backpropagation. The reader is
11) hypara2 – 8 dimensional hyperbolic paraboloid
expected to have a basic understanding of Matlab to use the
function
PSOt. Please refer to the appropriate Matlab documentation
to understand the code syntax examples [6].
The toolbox may be downloaded freely as a self extracting
zipfile from: http://www4.ncsu.edu/~bkbirge/PSO/PSOt.exe
D. Particle Swarm Optimization (PSO)
B. Introduction
Inspired initially by flocking birds, Particle Swarm
Optimization (PSO) is another form of Evolutionary
With the growth of more mainstream interest in
Computation and is stochastic in nature much like Genetic
computational intelligence algorithms applied to real-world
Algorithms. Instead of a constantly dying and mutating GA
engineering problems there is a need to build tools that can be
population we have a set number of particles that fly through
used in a wide variety of environments, from classroom to
the hyperspace of the problem. A minimization (or
field. Particle Swarm Optimization (PSO) is one recent
maximization) of the problem topology is found both by a
technique that has been used with great success in the
particle remembering its own past best position and the entire
group’s (or flock’s, or swarm’s) best overall position. This B. Development/Program Usage
algorithm has been shown to have GA like advantages
without the big computational hit. The set of programs that comprise the PSOt are designed with
two purposes in mind. The first purpose is to act as a
The PSO algorithm is based on the concept that complex standalone function optimizer. The second purpose is to
behavior follows from a few simple rules [5]. replace backpropagation with PSO during neural net training.

The intricacies of the algorithm are briefly outlined here. For C. Finding global min/max of MISO function using PSO
more details and discussion see the references, especially [4],
[5], and [8]. The first goal is achieved with the Matlab script ‘PSO.m’. It
is very simple to use but can be very flexible as well. The
II. TECHNICAL INFORMATION function only works with single input single output (SISO) or
multi-input single output (MISO) functions that need all the
A. PSO Algorithm input space searched. It will work with any size dimension of
input space, the toolbox miscellaneous functions include 2D,
The Basic PSO algorithm consists of the velocity and 3D, and 7D functions to experiment with. The PSO function
position equations: can either minimize or maximize the target function. The
basic syntax can be found by typing ‘help PSO’.
vi ( k +1) = vi ( k ) +γ 1i ( p i − xi ( k ) ) +γ 2i ( G − xi ( k ) ) (1)
TABLE I
MATLAB OUTPUT FOR TYPING ‘HELP PSO’
xi ( k +1) = xi ( k ) +v i (k +1) (2)
%[optOUT]=PSO(functname,D), or:
%[optOUT,tr,te]=
i – particle index % PSO(functname,D,VarRange,minmax,PSOparams)
k – discrete time index %
v – velocity of ith particle %Inputs:
x – position of ith particle % functname–
% name (string) of matlab function to optimize
p – best position found by ith particle (personal best) % D-
G – best position found by swarm (global best, best of % # of inputs to the function (problem dimension)
personal bests ) %
γ1,2 – random numbers on the interval [0,1] applied to thi %Optional Inputs:
% VarRange-
particle % matrix of ranges for input variable,
% default -100 to 100, form:
The most used PSO form is to include an inertia term and % [ min1 max1
% min2 max2
acceleration constants , hence the Common PSO algorithm: % ...
% minD maxD ]
v ( k +1) = %
i % minmax-
φ ( k ) v ( k ) + α γ1i ( pi − xi ( k ) )  + α γ 2i ( G − xi ( k ) )  (3)
% if 0 then funct is minimized
i 1 2 % if 1 then funct maximized, default=0

φ - Inertia function The ‘PSOparams’ option is where the PSO specific

α1,2 – Acceleration constants parameters can be changed.

The inertia function is commonly either taken as a constant TABLE II

1.4 or as a decreasing linear function in k from 0.9 to 0.4. As ‘PSO.M’ PARAMETERS
training progresses using a linear inertia function, the
%PSOparams - PSO parameters
influence of past velocity becomes smaller.
%
%P(1)-
Reference [8] shows the acceleration constants are most % Epochs between updating display,
commonly set to both equal 2. % works with P(13), default = 25.
%
%P(2)-
There are also a few more important parameters. These are % Maximum number of iterations (epochs) to train,
initial positions of particles (init ial weights when training a % default = 2000.
neural net), the maximum velocity bounds, and number of %
%P(3)-
particles in the swarm. % population size, default = 20
%P(4)-
% maximum particle velocity, default = 4 A more interesting example is the Schaffer f6 function. The
%
%P(5)- function has 2 inputs and one output and looks like a still shot
% acceleration const #1 (local best influence), of a pebble dropped into water. It has circular ripples that die
% default = 2 out as the radius increases. It has a known global minimum at
% x,y = 0 and the maximum is the first ‘ripple’ around the
%P(6)-
% acceleration const #2 (global best influence), origin. It has lots of local minimums and maximums making
% default = 2 it a good test function for the algorithm.
%
%P(7)-
% Initial inertia weight, default = 0.9
Running the PSO function to find the minimum bound by -
% 100 = x,y = 100 gives the following table.
%P(8)-
% Final inertia weight, default = 0.2 TABLE III
%
%P(9)- FINDING THE MINIMUM OF THE ‘F6’ FUNCTION
% Epoch by which inertial weight = final value,
% default = 1500 >> pso('f6',2,[-100,100;-100,100],0)
% PSO: 1/2000 iterations, GBest = 0.499643.
%P(10)- PSO: 26/2000 iterations, GBest = 0.0858499.
% randomization flag, PSO: 51/2000 iterations, GBest = 0.00197516.
% = 0, random for each epoch PSO: 76/2000 iterations, GBest = 0.00197516.
% = 1, random for each particle at each epoch PSO: 89/2000 iterations, GBest = 0.00197516.
% = 2, random for each dimension of each particle
% at each epoch ***** global error gradient too small for too long
% ***** this means you've got the solution or it got
%P(11)- stuck
% minimum global error gradient, ans =
% if abs(Gbest(i+1)-Gbest(i)) < gradient over -4.400531404313024e-002
% certain length of epochs, terminate run, 4.527939759978006e-003
% default = 1e-9 1.975158000277422e-003
%
%P(12)-
% epochs before error gradient criterion FIGURE I
% terminates run, default = 50 USING ‘PSO.M’ TO FIND MINIMUM OF ‘F6’
%
%P(13)-
% plot flag, shows progress display if =1, 20
% default = 1
Pos dimension 2

0
The output of the function is returned as a column vector of
the final global best position. The last entry in the vector is -20
the global best value. Optional outputs are the # of epochs it
took to reach consensus that optimization had occurred and -40
-10 -5 0 5 10 15 20 25
the value of global best as a function of epoch. Pos dimension 1
0
PSO: 2 dimensional prob search, Gbestval=0.0019752
10
A simple example would be to find the minimum of the sine
function from 0 to π/2:
Gbest value

-1
10

-2
10
pso(‘sin’,1,[0,pi/2],0)
-3
10
The name of the function is surrounded by single quotes, the 0 10 20 30 40 50 60 70 80
epoch
dimension of the problem is 1 (only one input so only one
searchable dimension), the range is from 0 to π/2, and the
minmax flag is set to 0 indicating we want to find the Running the ‘PSO.m’ function to find the maximum of f6
minimum. with the same bounds yields the following table.

Naturally we expect it to return the column vector [0,0]’ TABLE IV

which it does. It finds it after 1 epoch (iteration) of the FINDING THE MAXIMUM OF THE ‘F6’ FUNCTION
algorithm but doesn’t stop looking until 50 epochs have gone
by with no change in the best global position. That can be >> pso('f6',2,[-100,100;-100,100],1)
PSO: 1/2000 iterations, GBest = 0.500179.
changed with P(12). PSO: 26/2000 iterations, GBest = 0.973786.
PSO: 51/2000 iterations, GBest = 0.976315. %
PSO: 76/2000 iterations, GBest = 0.976362. %TP(3)-
PSO: 101/2000 iterations, GBest = 0.976432. % Error goal, default=0.02.
PSO: 126/2000 iterations, GBest = 0.976432. %
PSO: 145/2000 iterations, GBest = 0.976432. %TP(4)-
% Population size, default = 20
***** global error gradient too small for too long %
***** this means you've got the solution or it got %TP(5)-
stuck % Maximum particle velocity, default= 4
ans = %
-1.399320469183342e+000 %TP(6)-
6.794290702799257e-001 % Acceleration constant #1, default = 2
9.764318246838789e-001 %
%TP(7)-
% Acceleration constant #2, default = 2
For a fun exercise, you can set the maximum velocity %
parameter to a small fraction, such as 0.04 and set the update %TP(8)-
display frequency to do so at every iteration. This will really % Initial inertia weight, default = 0.9
%
show the behavior of the PSO algorithm, showing the %TP(9)-
particles ‘fly’ towards the global best and explore new areas. % Final inertia weight(iwt),default=0.2
The reader is encouraged to change all of the parameters %
around and examine the changes in behavior of the PSO. %TP(10)-
% Epoch by which inertial weight = final value,
% default = 1500
D. Training Neural Net using PSO %
%TP(11)-
% Maximum initial network weight absolute value,
To train a neural net using the PSO algorithm the function % default = 100
‘trainPSO.m’ is used. This function requires that the Neural %
Net toolbox be installed for use with your copy of Matlab. %TP(12)-
TrainPSO is modeled after the structure of the Neural Net % Randomization flag, default = 2:
% = 0, random for each epoch
toolbox and can be used interchangeably with the other % = 1, random for each particle at each epoch
training functions such as TrainBP, TrainLM, etc. The only % = 2, random for each dimension of each particle
difference is in the training parameters passed to the training % at each epoch
algorithm. The format of the function is: %
%TP(13)-
% Minimum global error gradient
[W1,B1,W2,B2,...,TE,TR] = % (if SSE(i+1)-SSE(i) < gradient over certain
TRAINPSO(W1,B1,F1,W2,B2,F2,...,P,T,TP) % length of epochs, terminate run,
% default = 1e-9
%
W1, W2, W3, B1, B2, B3 are weights and biases for the %TP(14)-
neural net, initially made randomly or with the built in % Error gradient criterion terminates run here,
initialization function that comes with the Neural Net % default = 200
%
toolbox. F1, F2, F3, are activation function choices such as %TP(15)-
‘logsig’, ‘tansig’, etc. Refer to the Neural Net toolbox % Plot flag, if = 1, display is updated
documentation for more details on those and the other
parameters [6]. The PSO specific training parameters show Here is a run of ‘DemoTrainPSO’ where there is one hidden
up in TP. TP is a row vector of 15 values. They cover the layer of 2 neurons (9 dimensional optimization problem). The
same ground as the straight PSO function with a couple of net is trained to approximate the XOR function. It displays
parameters specific to neural net training. the evolution of the inertia weight (iwt) over time as well as
the training error and particle positions. The graph also shows
Typing ‘help trainpso’ yields the training parameter summary the trajectory of the global best position. Only a portion of the
and is shown in the following table. training is shown.

TABLE V TABLE VI
‘TRAINPSO.M’ PARAMETERS MATLAB OUTPUT FOR ‘DEMOTRAINPSO.M’

%Training parameters are: TRAINPSO: 100/1000 epochs, gbest SSE = 0.428059141

%TP(1)- mv = 4, iwt = 0.8537691795
% Epochs between updating display, default = 100. TRAINPSO: 125/1000 epochs, gbest SSE =
% 0.3870019042
%TP(2)- mv = 4, iwt = 0.8420947298
% Maximum number of iterations (epochs) to train,
% default = 4000.
TRAINPSO: 150/1000 epochs, gbest SSE = F. References
0.0943753488
mv = 4, iwt = 0.8304202802
TRAINPSO: 175/1000 epochs, gbest SSE = [1] Eberhart, R., Simpson, P., Dobbins, R., Computational
0.0943753488 Intelligence PC Tools , pp. 212-223, 1996, Academic
mv = 4, iwt = 0.8187458306 Press, Inc.
TRAINPSO: 200/1000 epochs, gbest SSE = [2] Haykin, S., Neural Networks, a Comprehensive
0.0943753488
mv = 4, iwt = 0.8070713809 Foundation, Second Edition, 1999, Prentice-Hall.
TRAINPSO: 225/1000 epochs, gbest SSE = [3] Houck, C., Joines, J., and Kay M., A Genetic Algorithm
0.04639357585 for Function Optimization: A Matlab Implementation,
mv = 4, iwt = 0.7953969313
TRAINPSO: 250/1000 epochs, gbest SSE =
ACM Transactions on Mathematical Software,
0.04454761798 Submitted 1996
mv = 4, iwt = 0.7837224817 [4] Kennedy, J., Eberhart, R., Particle Swarm Optimization,
***************************** Reached Goal Proc. IEEE Int’l. Conf. on Neural Networks (Perth,
******************
TRAINPSO: 256/1000 epochs, gbest SSE = Australia), IEEE Service Center, Piscataway, NJ, IV:
0.01641486174 1942-1948, 1995
mv = 4, iwt = 0.7809206137 [5] Kennedy, J., Eberhart, R., Shi, Y., Swarm Intelligence,
***************************** end of training 2001, Academic Press, Inc.
****************
[6] Matlab Online Help, http://www.MathWorks.com
[7] Parsopoulos, K.E., Plagianakos, V.P., Magoulas, G.D.,
FIGURE II Vrahatis, M.N., Stretching Technique for obtaining
USING ‘TRAINPSO.M’ TO APPROXIMATE XOR
global minimizers through Particle Swarm Optmization,
Proc. Particle Swarm Optimization Workshop,
mv=4 inertia wt=0.78092, 9 dimensions, GbestVal= 0.01641486174
10
1 (Indianapolis, IN, USA), pp. 22-29, 2001
[8] Shi, Y., Eberhart, R., Parameter Selection in Particle
10
0
Swarm Optimization, Proceedings of the Seventh Annual
gbest

Conf. on Evolutionary Programming, pp. 591-601, 1998

-1
10 [9] van den Bergh, F., Particle Swarm Weight Initialization
-2
In Multi-Layer Perceptron Artificial Neural Networks, In
10
0 50 100 150 200 250 300 Development and Practice of Artificial Intelligence
epoch Techniques (Durban, South Africa), pp. 41-45, Sept.
6 1999.
4
pos dim 9

-2

-4
-10 -8 -6 -4 -2 0 2 4
pos dim 1

E. Conclusion

The suite of functions and programs included with the

Particle Swarm Optimization Toolbox are useful both in
researching PSO behavior and in applying PSO to real world
optimization and computational intelligence problems. The
PSO algorithms implemented here are computationally
intensive and are missing some features that would improve
flexibility and work continues on improvement. As it stands,
the toolbox allows the researcher/engineer to utilize PSO
without having to write custom code from the ground up. The
included comments make the suite easily modified to fit more
specific types of problems.