Tolerance criteria for cost function obtained with PSO

Hi,

I have applied PSO for minimization of RMS value that is my cost function. I have used 150 iterations. How can I set up a tolerance criteria (e.g. 10^-12) for the cost function? So I want to run PSO until I get the cost function close to 10^-12. I call the PSO function within the script and the function outputs the best position, best cost, and cost convergence for 150 iterations. Should I set the tolerance criteria inside the script like a for loop or something else? Any advice would be very helfpul. Thanks.

Answer 1

[Walter Roberson]

https://www.mathworks.com/help/gads/particleswarm.html#inputarg_options

You can set FunctionTolerance in relative terms. You would want to set MaxIterations high.

[EB]

Can you give some more explanation. I went through the link you have provided but I can not find the solution to my problem.

I am not calling built in Matlab particleswarm solver for my optimization problem. I've built the PSO for my problem in a separate function and I call the function PSO_EBW_SD within a script. The function PSO_EBW_SD outputs a structure array out. This array contains two fields: best cost and best solution.

The PSO_EBW_SD function is set up that it's terminates when max number of iterations is achieved.

When I am within the script and call the PSO_EBW_SD I get one value for the best cost and best sol. If I call the PSO_EBW_SD again it will give me other value for the best cost and best sol. IS there a way how I can stop this? I thought that set up of a tolerance criteria for the objective function e.g. Tol = 1e-6it will help. So what I am trying to achieve is to run the PSO_EBW_SD as many times I need in order to get the best cost equal or less than Tol.

I don't know how to set up this criteria. I've read some example on output function http://uk.mathworks.com/help/optim/ug/output-functions.html Should I have this output function as well even though I am not using the built in solver particleswarm.

Suggestions are more than welcome. Thanks!