Particle Swarm Optimization info

I am trying to find the values of 19 free variables from my data, using the psw optimization algorithm. I got interest in seeing the evolution of the 19 variables together with the minimization of the objective function, over the iterations. I then want to compare this againnst the genetic algorithm.

with the genetic algorithm I am able to to so, and get information (see code below), but with the psw I am no cabable of doing it. With the ga, I call this function from gaoptimset('PlotFcns',{@gaplotbestf, @gapop_func},...

Can you help me with this? Thank you a lot

function [state,options,optchanged] = gapop_func(options,state,flag)
persistent   best r  history_min history_max %h1
optchanged = false;
switch flag
    case 'init'
        best = state.Population;
        assignin('base','gapopbestx',best);
    case 'iter'
        ibest = state.Best(end);
        ibest = find(state.Score == ibest,1,'last');
        bestx = state.Population(ibest,:);
        best = [best; bestx];
        assignin('base','gapop_max_best',history_max);
    case 'done'
       assignin('base','gapopbestx',best);
end

Answer 1

[Alan Weiss]

The syntaxes for output functions differ between ga and particleswarm. For an example of a particleswarm output function, see this example. For the correct way to set options for particleswarm, see the same example (you must use optimoptions to set options).

Alan Weiss

MATLAB mathematical toolbox documentation

[Marco Marchese]

Thank you very much. this was very useful!

My doubt is now, what if I wanted to extract the value of each dimension (a.k.a. each nvars) that I am looking at each iteration and see how that converges?

Marco

[Alan Weiss]

The optimValues.bestx field contains the current best point at each iteration. This is a 19-element vector for the problem you described.

Alan Weiss

MATLAB mathematical toolbox documentation

[Alan Weiss]

Allow me to make one further observation. Generally, patternsearch is the most reliable and fastest-converging solver for nonsmooth problems. Please be sure to compare with it. And, if your problem is smooth, you should use fmincon instead. To search for a global solution using fmincon, try MultiStart. To search for a global solution using patternsearch, try uniformly distributed initial points:

x0 = lb + rand(size(lb)).*(ub - lb);

Good luck,

Alan Weiss

MATLAB mathematical toolbox documentation

[Marco Marchese]

Thank you for the suggestion. The issue is that my function is quite complex and the computational time is very extended, thus I do not see much difference between G.A., PSO, Patternsearch. I have a non-smooth, non-linear function with a set of unknown values that i need to fit from experimental data.

[Marco Marchese]

Apparently both patternsearch and PSO work out fine.

Is there a way I can extract confidence intervals from here? I would want to avoid to pass through nlingit,lsqcurvefit or lsqnonlin.