Another Particle Swarm Toolbox

Implementation of a PSO algorithm with the same syntax as the Genetic Algorithm Toolbox.

Description

Particle swarm optimization is a derivative-free global optimum search algorithm based on the collective intelligence of a large group of intercommunicating entities. The individual particles are simple and primitive, knowing only their own current locations and fitness values, their personal best locations, and the swarm's best location. Each particle continually adjusts its trajectory based this information, moving towards the global optimum with each iteration. The swarm as a whole displays a remarkable level of coherence and coordination despite the simplicity of its individual particles. The coordinated behavior of the swarm has been compared with that of a flock of birds or a school of fish.

This is a Particle Swarm Optimization (PSO) algorithm which uses the same syntax as the Genetic Algorithm Toolbox, with some additional options specific to PSO. The idea is to allow some degree of code re-usability when trying different population-based optimization algorithms. Certain GA-specific parameters such as cross-over and mutation functions will obviously not be applicable to the PSO algorithm. However, many of the commonly used options and constraints for the Genetic Algorithm Toolbox may be used interchangeably with PSO since they are both iterative population-based solvers. See >> help pso (from the psopt directory) for more details.

Features:
* Modular and customizable
* Ability to solve problems consisting of binary variables. Type >> help psobinary (from the psopt directory) for details.
* Bounded, linear, and nonlinear constraints
* Support for vectorized fitness functions
* Algorithm parameters may be adjusted using an 'options' structure similar to existing MATLAB optimization solvers
* Custom plots may be written using same interface as the Genetic Algorithm plotting functions
* Another optimization solver may be called as a "hybrid function" to refine the results after the particle swarm terminates

A demo function (psodemo.m) is included, with a small library of test functions. To run the demo, from the psopt directory, call >> psodemo with no inputs or outputs.

New features and bug fixes will continue to be released until this is made redundant by the release of an official MATLAB PSO toolbox. Bug reports and feature requests are welcome.

Bibliography:
J Kennedy, RC Eberhart, YH Shi. Swarm Intelligence. Academic Press, 2001.

Particle Swarm Optimization. http://en.wikipedia.org/wiki/Particle_swarm_optimization

SM Mikki, AA Kishk. Particle Swarm Optimization: A Physics-Based Approach. Morgan & Claypool, 2008.

ANNEX:
Nonlinear inequality constraints in the form c(x) ≤ 0 and nonlinear equality constraints of the form ceq(x) = 0 have now been implemented. Either 'soft' or 'absorb' style boundaries may be used in problems with only nonlinear inequality constraints; calculations may be faster in some situations with 'soft' boundaries. However only 'absorb' style boundaries are usable for problems with nonlinear equality constraints.
See the following document for the proper syntax for defining nonlinear constraint functions: http://www.mathworks.com/access/helpdesk/help/toolbox/optim/ug/brg0p3g-1.html. To see a demonstration of nonlinear inequality constraints using a quadrifolium overlaid on Rosenbrock's function, run psodemo and choose 'nonlinearconstrdemo' as the test function.

Acknowledgements

The author wishes to acknowledge the following in the creation of this submission:
Particle Swarm Optimization Toolbox, Many testfunctions for global optimizers

• Required Products: Optimization Toolbox
• MATLAB release: MATLAB 7.7 (R2008b)
• Other requirements: Familiarity with the Genetic Algorithm and Direct Search Toolbox would help in using this package. Tested on Windows x86; will probably work for other platforms.

[Hanlin Zhang]

This PSO toolbox is very useful for solving constrained optimization problems. Thanks.

[Sam]

Glad you found it useful. FYI I just found a bug where the swarm doesn't actually comply with the imposed linear constraints. I'm working to fix it as soon as possible.

[tom]

 Hi, Sam,
   I am interested in PSO with both linear and nonlinear constraints for high-dimention problem, but I found this package can not handle it.
   Is there any other PSO toolbox to handle the problems with nonlinear and linear constraints? Please let me know: [e-mail]. I am waiting for your reply.

   Thanks,
Tommy
 

[Sam]

Hi Tom,

This webpage (http://www.particleswarm.info/Programs.html) gives a large list of available toolboxes, although I don't know if any of them can handle non-linear constraints, and most of them are not written in MATLAB. Since you're interested, I will work on implementing non-linear constraints for this toolbox in a future release, maybe in about two weeks' time?

Sam

[Mohammed Ahmed Saeed]

Dear sir ,How do you do ? i prepare for my Master in optimal relay coordination and if you please i need your help to provide me with a swarm code to determine the optimal relay settings ( i can not do the fitness function for each particle )
Thanks

[Sam]

Dear Mr Saeed,

Does your fitness function work with other MATLAB optimization solvers? Please see this document (http://www.mathworks.com/access/helpdesk/help/toolbox/gads/brdvur4.html) for how to write a fitness function for the Genetic Algorithm (which should also work for PSO). Note that this PSO code doesn't support binary (where the only possible values are 0 or 1) or discrete variables yet.

Sam

[karim]

Dear Mr sam

 Iam eng.karim and i want to thank you about this great tool box and i ask about any news for binary support for this tool box since i work in my Master's degree and i want to use binary pso and i didnt find any matlab code support this. finally thank you again.

[Sam]

Hi Karim,

Sorry about that, I know I mentioned the possibility that I might implement binary variables for this toolbox. Over the past few weeks my thesis work has taken me in another direction, so I don't think I'll have the time to do it. The first book listed in the bibliography section of my description: "Swarm Intelligence" by J Kennedy, RC Eberhart and YH Shi, describes in detail how to implement PSO with binary variables, if you're interested.

Sam

[Michael Johnston]

Sam -- Thanks very much for providing your code. You've clearly put a lot of work into it. The performance and stability on my end has been flawless thus far.

The only feature request I can think of is adding a field to the options structure to control the verbosity of the output to the command window? Sometimes it's nice to be able to kill this completely.

Best,

Mike

[Sam]

Mike -- that's a good idea, I can implement the "quiet" mode fairly quickly.

[Sam]

Mike -- I've uploaded a new version which should appear tomorrow. You can set options.Display to 'off' to kill all command line output (except for a few warnings and error messages that you'd probably want to know about anyway).

[Michael Johnston]

Thanks, Sam. Have you considered adding options for parallel processing (for those who have the Parallel Computing Toolbox)?

[Sam]

I've considered adding parallel processing features, but I don't have the toolbox myself so I'd have no way of testing it.

An alternative could be to use a vectorized fitness function, setting options.Vectorized to 'on', and do all the parallel computing tasks in the fitness function, independently of the PSO code.

If you're interested in collaborating to add Parallel Computing capability, I've got an SVN repository with this project going on Google Code Project Hosting, just search for "psomatlab".

I could easily set up PSOOPTIMSET to create the appropriate option, and somebody with the parallel computing toolbox could set up PSO itself to handle it. Let me know if you're interested.

[Albert Lee]

Can I know how to write for the Matlab code for the Particle Swarm Ant Colony Optimization? If the test function is the same as the PSO toolbox. Thank you.

[Sam]

Albert, Ant Colony algorithms are quite different from Particle Swarms, although there have been papers published proposing a hybrid of the two algorithms. My particle swarm code by itself does not do ant colony optimization, so I'm not sure what your question is -- are you trying to write an ant colony algorithm, or are you trying to learn how to use somebody else's ant colony toolbox?

The method for defining a fitness function for this PSO toolbox is the same as for other MATLAB optimizers such as GA, FMINCON, or FMINUNC. Any fitness function you write which works with those optimizers should also work for PSO. See this document http://www.mathworks.com/access/helpdesk/help/toolbox/optim/ug/brhkghv-3.html for instructions -- the only thing is, you don't need to provide the gradient/Jacobian or the Hessian to PSO. Note that PSO can only minimize fitness values; if you have a problem where you're trying to maximize f(x), just set g(x) = -f(x) and minimize g(x) with PSO.

[Albert Lee]

Hi Sam

I need to reply the question that you reply to me. Since I am doing a research about the hybrid PSO with the other algorithm, so I need to write out the program and testing on the test functions. One of my program is particle swarm ant colony optimization. Another one is evolutionary particle swarm optimization. So, I hope that some guidelines will be provided in order to solve the problem. thank you.

[Sam]

Sure, you may also be able to use the test functions provided with the two other files I listed under the "Acknowledgments" section.

[Uduakobong]

Hello sam, i am trying to write a PSO program to solve a multiobjective, nonlinear contriant problem. the problem has 3 varibles with 2 of the varibles being discret and one is continous. the thing is i have never used PSO before so i find it difficult ti understand what to do. Please could you give me some pointers thank you

[Sam]

Uduakobong, if you've never used PSO before then it's best to take some time to read through the three references listed under "bibliography" in this file's description. It's an interesting problem to mix continuous and discrete variables, but this toolbox isn't capable of that. As there has been some interest, I will implement the ability to solve problems with binary variables soon.

[Amaraporn]

Hi Sam,
I'm trying to use your toolbox solve one objective function subjecto lb and ub with 12 inputs but found it is pretty complicated to define suitable option to get the program run. Mostly, i found this error and been trying, now desperate to sort it out.

Swarming...??? Subscript indices must either be real positive integers or logicals.

Error in ==> pso at 258
        state.Score(setdiff(1:n,find(state.OutOfBounds))) = ...

Your comment would be very helpful

[Sam]

Hi Amaraporn, I'll investigate what's going on at line 258. Can I ask what is the exact code you use to set the options structure and then call pso?

I know it's a bit complicated to set all the options for this PSO code -- I wrote it to be very similar to the Genetic Algorithm that's already provided with a MATLAB toolbox, so that people who already know how to that toolbox can easily transfer their code to this PSO algorithm. If anything's unclear, the online documentation for the Genetic Algorithm Toolbox may help: http://www.mathworks.com/access/helpdesk/help/toolbox/gads/f6010dfi3.html

[Amaraporn]

Hi Sam,

Many thanks for your feedback:) Now I'm back to start from a test function which its answer is known. I found your toolbox easy to use with 2-11 inputs and suitable populationsize and generation are needed to help the finding of global min. Now, I got problem with 12 inputs, populationsize=50 and generation=from 2000 to 8000. If you could help to comment how to define suitable pso variables to achieve global min.
-----------------------------------
Eror:
The value of the fitness function did not improve in the last 50 generations and maximum constraint violation is less than 1e-006, after 61 generations.
 -------------------------------------------------------

The code for test function is below and subject to lb = -500 and ub=500 (as you know, in same dimension with input number).

%%%%%%%%%%%%%%%%%%%%%%%
function f = easyTest(x)
[xSize, Dim] = size(x);
persistent sumX
for j=1:Dim
     fx = (-x(j)) * sin(sqrt(abs(x(j))));
   sumX = sumX+fx;
end
f = sumX;
%%%%%%%%%%%%%%%%%%%%%%%%

[Sam]

Amaraporn, Can I ask why you've got sumX as a persistent variable in the test function? That means that its value is retained between calls to the function, so that the fitness value of every particle will be dependant on the fitness of all previously evaluated particles. This is probably not what you want PSO to do. What happens if you instead change that line to sumX = 0?

How does the swarm perform with the default options?

[Amaraporn]

Dear Sam

You are right about the persistent function, i have corrected it and try with the default options. I found the toolbox terminate with some local mins as follows.
[420.97 420.97 420.97 -302.53 420.97 420.97 420.97 -500 420.97 420.97 420.97 -302.52] ( Theoritical global min is 420.97).

So, I presume that the toolbox can works well with 12 inputs given that the suitable pso variables are defined.
Now I move to use it with my real obj function having single objective function subject to lb and ub for 12 inputs (this objective function already worked with fmincon (alone)). The error when used with you pso toolbox is as follows.

%%%%%%%%%%%%%%%%%%%%%%%%%
Swarming...??? Attempted to access Swarm(3); index out of bounds because numel(Swarm)=2.

Error in ==> vMarkSqr_spineCT_pso_mod at 37
vThoracicTransl=[Swarm(1);Swarm(2);Swarm(3)];

Error in ==> overlaysurface at 13
        ZZ(i,j) = fitnessfcn([XX(i,j) YY(i,j)]) ;

Error in ==> psoplotswarmsurf at 30
        overlaysurface(state.fitnessfcn,options) ;

Error in ==> pso at 334
            state = options.PlotFcns{i}(options,state,flag) ;

Error in ==> psodemo at 61
pso(problem);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
"Swarm" is my optimized parameterers.
The words "Error in ==> vMarkSqr_spineCT_pso_mod at 37
vThoracicTransl=[Swarm(1);Swarm(2);Swarm(3)];" is all about my objective codes and line 37 is very first line of my main calculation.
I think this is a problem between the default swarm dimension of the toolbox and the input dimension of my obj function. My function starts with simple head line as follows. Anything should be changed? It is not convenient to sent the whole function to you but I can explain how it works as follows.

%%%%%%%%%%%%%%%%%%
function LeastSqr = vMarkSqr_spineCT_pso_mod(Swarm)
%%%%%%%%%%%%%%%%%%

This obj function is the calculation based on multiple transformation matrices similar as the one from biomechanical application which you may have known it as "ankle joint parameters solving using parallel global optimization with particle swarm". In the main obj function, other 3 functions (created by my self) are called and all deals with transformation matrices with each matrix is calculated from (3-5) swarm inputs.

Could you please help comment about this?

[Sam]

Amaraporn, don't use PSODEMO to run your actual optimization. I included that function to provide an easy way to visualize how the swarm behaves, but it wasn't intended to be used to run actual optimizations.

Instead, you should call PSO directly using the syntax explained when you type >> help pso. It was designed to be very similar to the Genetic Algorithm (now called "Global Optimization") Toolbox, so it would help if you become familiar with its documentation (the link is provided in one of my previous posts). If you read the help provided with PSO, you'll see that there is no "default" dimensionality. The number of dimensions of the problem (nvars) must be provided by the user, i.e. >> pso(@fitnessfcn, nvars, ...).

[Amaraporn]

Sam

Many thanks for all of your feedbacks :) Now I can use your toolbox with actual function though spent almost a whole day for a clean termination. I'm not a computing guy so a bit wondering about its time consuming. Is it sounds resonable for the obj function which fits the 12 parameter dynamic model to the data of (5000 frames) time history of experimental dynamic motion? Anyway this is a good toolbox for everyone including students from out of field, I confirm :)

[Sam]

Glad it helped. 5000 frames sounds like a lot of data, so it might take a long time depending on your computer and the complexity of the calculations. See this document for tips on improving the performance of MATLAB code: http://www.mathworks.com/access/helpdesk/help/techdoc/matlab_env/f9-17018.html

[Amaraporn]

Dear Sam

Just one quick question about pso algorithm parameters used in your toolbox. I have tried to tune
problem.options.Generations eg. 10000
problem.options.CognitiveAttraction eg. 2
problem.options.SocialAttraction eg. 2
problem.options.StallGenLimit eg. 8000
problem.options.PopulationSize eg. 40
and initial inertia eg. 1

but global min could not been found for my objective function of 12 parameters. I'm seeking other important parameters and wondering if you got these following parameters somewhere in the algorithm. "inertia reduction parameter", "bound and velocity fraction","velocity reduction parameter"

[Sam]

Amaraporn, the swarm is only stable when the sum of the CognitiveAttraction and SocialAttraction parameters is less than 4; if they are 2 and 2 as you've got them, then 2+2 = 4 and the swarm will not converge. Try reducing one or both of them such that their sum is less than 4. I'll add a note (and paper reference) regarding this to the documentation.

Also, try using fewer generations; that will make the inertia reduction parameter scale better. 10,000 is a very large number of generations. Try setting "Generations" to a few hundred, at most. Same with "StallGenLimit" -- what happens when you use the default values?

[Amaraporn]

Sam, Many thanks for your kindly reply. When i used the default with test funcion of 12 inputs. The 12 output never reached the known global min. Then I tuned up all possible parameters in your toolbox and found the best parameter set comprising
CognitiveAttraction=1.5
 SocialAttraction parameters =1.5
Generation = 300
popsize= 50

which made the 11 out of 12 outputs meet the theoritical min. Till now I could not find any pso parameter set that can bring the output converge to the same point of every runs which is the general expectation for global min. Somehow making boundary more stricly may be helpful, so i'm trying it at the moment.

[chong]

hi,sam
       i am a new one. i can't run th psotoolbox successful ly. couldy you give me more details using the syntax.
        thanks a lot!

[loo cheng]

hi,sam
   Thanks for you hard working and selfless dedication fristly.
i have encountered some problems when i using psodemo, some errors emergence, i tried to figure out ,but failed lastly, the error indication is below
"??? Error using ==> strcmp
Inputs must be the same size or either one can be a scalar.

Error in ==> isfield at 12
  tf = any(strcmp(fieldnames(s),f));

Error in ==> psodemo at 38
if any(isfield(options,{'options','Aineq','Aeq','LB'}))"

[Sam]

Hey everyone, I've been extremely busy with my thesis so I won't be able to provide any technical support for this file in the foreseeable future. I'm glad that so many people have found it useful. If you have any questions, please refer to previous comments, the file description, as well as getting familiar with how to use the Genetic Algorithm included with MATLAB's Global optimization (http://www.mathworks.com/access/helpdesk/help/toolbox/gads/bqe0w5v.html#bqe0w6h-2).

[Samuel]

Hello Mr. Sam,

Much thanks for this excellent software. I just have a question about the specifics of your implementation. Are you using an inertia weight in the update velocity, and if so does that weight decrease? It is recommended by some of the original PSO guys in "Defining a Standard for Particle Swarm Optimization" - Daniel Bratton and James Kennedy. It can cause a good space search in beginning and fine tuning toward the end.

Also, how do you deal with particles going over the bound? Are you preventing them from going out entirely, or letting them go out without evaluation the cost function (which will make the particle eventually pull back into the allowable search space)? The reason I ask is that preventing the particles from going out entirely can cause some bias toward the center of the search space.

Thanks again! I have had great success using your implementation.

[Mike]

Hi Sam,
Your toolbox is prefect.
But now, i want to find the maximum of a function using PSO.
Would you mind helping me?

[George Evers]

Mike,

To maximize a function, simply minimize its additive inverse. In other words, maximizing f(x) is mathematically equivalent to minimizing -f(x).

One easy way to do this would be simply to add
"f = -f;" as the last line of your test function.

[Samuel]

Implementation details: This PSO version uses a static intertial weight. You can easily change the velocity update function in file "psoiterate.m" if you want to implement some dynamic change, or even using a different PSO technique such is a constriction factor instead of inertial weighting.

[Ben]

 

[Sam]

Good eye on psoiterate, Samuel. However the inertia weight does scale by default, from 0.9 at the beginning of the optimization to 0.4 as it reaches the maximum number of iterations. I've updated the toolbox so that the velocity update function can be changed to your own custom function by setting the 'AccelerationFcn' option to the appropriate function pointer. I haven't documented the syntax for this yet, so it might be best to use the default psoiterate function as a template for developing your own velocity update function. It's currently located in the /private directory, but it will be moved to the base directory in future releases.

Also you can now set a time limit (in seconds) for the solver using the 'TimeLimit' option. Default is infinity.

[Ben]

Hi Sam,

A minor suggestion for the plotting part. It would be more convenient, at least for me. Hope it is helpful.

if ~isempty(options.PlotFcns)
%%%% close(findobj('Tag', 'Swarm Plots', 'Type', 'figure'));
hFig = findobj('Tag', 'PSO_Plots', 'Type', 'figure');
if isempty(hFig)
state.hfigure = figure(...
'NumberTitle', 'off', ...
'Name', 'Particle Swarm Optimization', ...
'NextPlot', 'replacechildren', ...
'Tag', 'PSO_Plots' );
else
state.hfigure = hFig;
set(0, 'CurrentFigure', state.hfigure);
end;
end % if ~isempty

[kaz uki]

U can email me at [e-mail]...i try my best..

[kaz uki]

i'm newbie about PSO...can somebody help me how to minimize assembly sequence time(product) using Discrete PSO(DPSO) and implement it using matlab...i need this for my degree-final year project.....i dont where to start..so please someone guide me step by step...email me for best....

[Sam]

Ben, thanks for the suggestion. I'll have a look at it next week when things get a little less busy for me. kaz uki, I'm not familiar with discrete PSO myself so I can't be of much help to you. Try searching for papers about it on Google Scholar or Compendex. Hope your final year project goes well.

[Ben]

Hi Sam,

I using your code to solve a sub-problem in my algorithm. However, when doing intensive testing, I'v found some stability or repeatability problem of your code. In many cases (more than 10%) of a large number of testing, I got different result using the same configuration (same dataset, same objective function, same pso option).

From my observation, this is caused by the initial particles. It was trapped into some unreasonable local minima.

Could you tell me how you generate the initial particles?

Thanks

[Sam]

Reposting this on the public thread, in case others have the same issue:

Hey Ben,

The psocreationuniform function in the /private folder will generate an initial population using a uniform random distribution based on options.PopInitRange. If you've got linear or nonlinear (in)equality constraints, this initial population will then get passed on to psocheckinitialpopulation (again in the private folder), and it will ensure that all of the initial points are feasible, moving the ones which aren't. Anyway, if my following answer is unclear, let me know what kinds of constraints are in your problem and I'll see if I can get back to you. I'll be moving over the next couple of days, so it might be a while before I see your next reply:

One problem I did encounter was with the fact that the PopInitRange option is set from 0 to 1 in all dimensions, by default (i.e. repmat([0;1],1,nvars)). This is obviously not representative of all design spaces. This might be a problem if you haven't set any boundary constraints LB and UB. This behavior was in the original genetic algorithm code from MATLAB upon which I based my code, so I left it it as is, in case other people have already written GA code assuming this behavior and want to try it with PSO.

So basically, if you haven't tried this already, you can manually set the PopInitRange option to a reasonable range that fully encompasses your design space. The CognitiveAttraction and SocialAttraction options might also be adjusted, but first see the note about them provided in the psooptimset help. More drastically, you could try editing the psocreationuniform function.

If none of those work, I'd be inclined say that you've run into an inherent limitation of the PSO, i.e. that if you initialize it in a small enough domain of the design space, and the global optimum lies well outside of it, then the swarm is not guaranteed to find its way out of it.

Sam

[t g]

hi sam,
i am doing my university project on manufacturing cell design using PSO tool. The problem is defined as a part/machine incidence matrix which maps part and machine, and the clustering should be formed block diagonally in order to make the cells. the objective is to minimize exceptional element count (EE). the PSO particle string should contain the cell nos. and index of the string are the machine nos.
Since i am new to this field, facing problem to implement the logic and code. Can anyone help in this regards. Matlab is the interface of the program. the problem is shown in a link http://www.mypicx.com/uploadimg/53506343_09112010_1.jpg

if you can anyway help me out.

[WANG]

hi,sam
      Thanks for you hard working and selfless dedication fristly. i am a mater of electrical engineering,and want to use BPSO to optimize a function,but i can't run th psotoolbox successful ly.when i run pso.m,then
??? Function name must be a string.

Error in ==> psooptimset at 180
    idx = find(cellfun(@(varargin)strcmpi(varargin,requiredfields{i,1}),...

Error in ==> pso at 171
options = psooptimset(options) ;
couldy you give me more details using the syntax.
        thanks a lot!
best wishes!

[satish jain]

hi sam
i am using basic pso for my function minimization problem, it is working but gbest is going beyond range . could you pl let me know , how and where i can change my code .
thanks
sk jain
[e-mail]

[Sam]

WANG, the syntax for this PSO toolbox is described in the comments at the top of the file named pso.m. You can type >> help pso from the command line, with the current directory set to where the pso.m file is, to read it, or you could just open the file. As I mentioned before, it should be the same as the syntax for the Genetic Algorithm included with the Global Optimization toolbox, so you can also refer to them: http://www.mathworks.com/help/toolbox/gads/ga.html.

If by BPSO you mean binary PSO, make sure that you're not trying to impose any constraints. Type >> help psobinary to learn the syntax for binary pso, which is slightly different.

t g and satish, I'd like to be able to help you, but school has been very busy for me, and your questions are bigger than I can properly address at this time. Again, I recommend reading the documentation provided with my PSO toolbox, as well as MATLAB's Global Optimization Toolbox, which I recommended to WANG (it's probably better-written and better presented than mine). You could try some of their simple examples to get an idea of how to use the toolboxes. I hope your research goes well!

[Ben]

Hi Sam,

There is another problem: sometimes the result goes out of the box constraints (lower & upper boundaries) a lot. Did you meet it before?

Thx

[WANG]

thanks for your recommendation.i found that my release is lower than required ,so i updated it.then i run pso.m successfully,but when i run psobinary.m ,it appeared that
 ??? Input argument "fitnessfcn" is undefined.

Error in ==> psobinary at 37
[xOpt,fval,exitflag,output,population,scores] = ...
would you tell me what is the problem and how to solve it.
thank you .best wishes!

[Sam]

Ben, thanks for pointing that out. I was actually experimenting with a new default constraint enforcement method ('penalize'), but it doesn't seem to handle boundary constraints very well. You can fix that problem by setting options.ConstrBoundary to 'soft' or 'absorb'. I'll release a quick patch (should be up by tomorrow) to set the default back to 'soft' so that this doesn't happen to other people.

WANG, are you passing any input arguments to PSOBINARY? PSO will run a default demonstration case without any inputs, but I haven't provided a similar function for PSOBINARY yet. Anyway, the first argument to PSOBINARY should be a pointer to a fitness function that's written by you (it should not actually be 'fitnessfcn', that's just a placeholder used in the documentation for the aforementioned function point). It should be able to accept a 1xnvars vector of 0s and 1s, and returns a fitness value. Make sure that the m-file for the fitness function is in your path when you call PSOBINARY. Hope that helps.

[Sam]

Looks like it will take until Monday for the aforementioned bug fix to be posted.

[Ben]

Hi Sam,

Could you briefly explain how to set the options related to different constraints? Frankly, there is no much word in the .m file.

Thank you,
Ben

[Mark Shore]

A quick question - Another Particle Swarm Toolbox requires MATLAB's Optimization Toolbox, NOT the Genetic Algorithm (now Global Optimization) Toolbox, correct?

[Mark Shore]

...or I could have just downloaded APST, ran the included demo, and easily answered my own question. Yes - as listed - the requirement is the Optimization Toolbox.

[Sam]

Hey Ben, there's a brief note about that in the help provided for the PSOOPTIMSET m-file (I know, the documentation is scattered all over the place -- one of the things I was hoping to do was to create a more comprehensive help file).

Glad you found the demo helpful, Mark. I was thinking about including the Global Optimization toolbox in the list of requirements, since it would really help to be familiar with GA. But as you found, it's not strictly necessary to run the PSO code.

[Mark Shore]

Sam, I will be trying out your toolbox in the near future as time permits. As far as requirements for installed MATLAB toolboxes, as far as I'm concerned, the fewer the better.

As a single commercial-licence user, I have to justify each additional toolbox I purchase (to myself, but still...). The wavelet and signal processing toolboxes were a non-issue. Got the optimization toolbox expressly for John D'Errico's SLM tools, and still sitting on the fence for parallel processing, image processing, mapping toolbox and curve fitting toolboxes, among others. $1000 here, $1000 there, plus maintenance, eventually adds up and can cut into one's desire to test FEX submissions...

[Mike]

Dear Sam
i am writing a simple PSO function in matlab, i would like to ask:
if the particle go out of the boundary, is that i should reset the velocity to zero?

[Oliver]

Sam,

Thanks for your submission, it works wonderfully. And thank you for sticking to familiar syntax for those of us who have been using the optimization toolbox, this really helps with the learning curve.

I saw some comments about taking advantage of parallel processing, but it didn't look like anyone has done anything about it thus far. When I profiled pso using my objective function the big place that things got bogged down was obviously in the many objective function evaluations. As such I changed the code a little bit to evaluate all of the objective function calls in parallel. I changed the following lines from this:

        for i = setdiff(1:n,find(state.OutOfBounds))
            state.Score(i) = fitnessfcn(state.Population(i,:)) ;
        end % for i

to this:

tempstatepop = state.Population;
itinerary = setdiff(1:n,find(state.OutOfBounds));
temp = zeros(length(itinerary),1);
parfor i = 1:length(itinerary)
     temp(i) = fitnessfcn(tempstatepop(itinerary(i),:)) ;
end % for i
state.Score(itinerary) = temp;

This is probably a crude way of doing it, but even so I saw a speed up of just under 3 times. Perhaps something similar will be helpful in a future build. Thanks again.

-Oliver

[Erdal Bizkevelci]

 

[Sam]

Thanks Oliver, I'll see if I can implement what you suggested for a parallel processing option. I'll have no way of testing it, so I might add a little note saying that the feature is in beta. Glad you found the toolbox helpful!

[Nguyen]

Dear Sam!
I'm student and i do not know well pso. Can you tell me, how can i use pso from optimtool?
Thanks!

[Sam]

Hi Nguyen, I don't think that this toolbox can be used with optimtool -- that would require modifications to the optimtool code, which is beyond my capabilities at the moment. If you learn how to use the optimization toolbox functions from the command line, then you should be well-equipped to use this PSO toolbox. Hope your studies go well!

[Nguyen]

Dear Sam!
Thank you very much for your help and your efficient toolbox. It works very well for my problems now.

[Haydar Dag]

When using psobinary, the system crashes for not being able to find the function 'randi'.

[Sam]

Haydar, I think the RANDI function was introduced in MATLAB r2008b as discussed here: http://www.mathworks.com/matlabcentral/newsreader/view_thread/274544

If you're not able to get a more recent version of MATLAB, I can see if I can release a small update over the weekend that eliminates PSOBINARY's dependency on RANDI. Thanks for the feedback!

[besbesmany besbesmany]

Dear Sam
Thanks alot for your effort in this toolbox,
i've problem in constraint pso , lb=0 and UB= 1 but the result of pso is not restricted with this bounds

http://www.mathworks.com/matlabcentral/fileexchange/25986-another-particle-swarm-toolbox/content/psopt/psoboundsabsorb.m

also Aeq and beq is not restrict the result

i tried penalized and soft , penalized is out of range of lb, ub, Aeq, beq

soft is give me the same initial value with no change

can you help me in that

[Sam]

Sorry that it's not working for you! The 'penalize' constraint method is unfinished, and it doesn't quite work yet: I took that feature out of the documentation and put in a warning message in the last revision of this toolbox when I realized that it still had problems.

If the 'soft' constraint method is giving you the same initial value with no change, then it could be because of several reasons:

1. None of your initial particle positions are feasible (so they are all set to infinity, using the 'soft' constraint enforcement method)
2. Your objective function is throwing an error with the given input vector (could be related to reason #1 above)
3. The design space is flat in the region of the initial particle positions

Are you using discrete or binary variables? Right now my PSO toolbox doesn't support design vectors that have a combination of real and discrete components (but the GA toolbox that MATLAB comes with should be able to those with no problem).

[besbesmany besbesmany]

my variables is
LB vector of zeros
UB vector of ones

Aeq is matrix containing zeros and ones
Beq is vecotr of ones
my initial value is matrix of any number from 0 to 1

penalize is the only method give me output but not in range of LB, uB and Aeq, Beq
all other constraint method return the same initial value
is thiere any way to have correct result from pso, any updated toolbox you will release soon?

i've correct result from ga toolbox but i want to check other algorithms
is thier any matlab algorithm solve same problem other than ga, fmincon

any direction will be appreciated

Thanks so much Sam

[fenfen Xiong]

or some one who has code that can deal with optimization with non-linear constraints at hand, would u please send me a copy?
Much thanks!

my email: [e-mail]

[Sam]

Due to several requests, I will be looking at fixing some issues with the 'penalize' constraint method over the coming weeks, so that this algorithm will work properly for problems with nonlinear constraints. Stay tuned for updates.

[Rilin]

Hi,Sam. I have a problem to use PSO to find C=[c1,c2,c3,c4].The contraint is A*W=1,where A=[a b],W=[c1*exp(jc3),c2*exp(jc4)].How can I set Aeq. Thanks for your answer.

[Rilin]

Or, it can just be set in the nonlinear constraint fucntion file?

[troy lim]

hi Sam,

u have done a very good job for the pso toolbox.

i am a newbie in matlab dealing with a project in college.

i need to optimize the membership functions parameters(total 63 parameters)of flc in simulink by using pso according to observed data result from simulation of the developed system as the fitness function. i found that the most difficult part for me is to formulate the fitness function, or is it possible i can find the best solution without fitness function?
i wish u can help me out or give me some idea as this is the last part of my project towards the end....thx

[Joe Ajay]

Dear Sam, Thanks for this toolbox, It was really helpful in my project. and I've done 4 optimization problems with it. I'd like to work on discrete optimization using PSO. Do you have any update of this toolbox for discrete optimization? If not, what is the other option to go for discrete optimization using PSO.

[Rilin]

Hi,Sam.Thanks for your good job.
I have a problem when I use the nonlinear contraint.'Problem is infeasible due to nonlinear constraints' occured.But I don't know why. Can you help me? Or anybody knows it can give me a hand. Thanks.