Currently i am fightin the fitting the theoretical model (in LS sense) to the experimental data (5 params 0..1 valued) . I have tested different solvers in Matlab an they gave me reasonable results:
- Simulated annealing - fabulous (really)
- nonlinear LS (Levenberg-Marquardt) - not bad (could be better)
- genetic - so so (sometimes)
- pattern search - so so (sometimes)
However, particle swarm (default matlab settings) returns it completely ugly, stucks at lower boundary which is complete nonsense.
Thanks to SA i already have good result, nevertheless, I am still curious why the PSO results are poor when it is known to be very explorative. Or how to tune the options if possible.
I think you should change PSO parameters. Since this problem is in [0,1], PSO could be yielding solutions outside this interval and then reparing the solution by pushing it towards the boundaries.
Alternatively, if you dont want to change PSO default parameters you can try the following, see what happens:
K=10;
lb=zeros(1,5);
ub=K*ones(1,5);
x=particleswarm(fun,5,lb,ub)
% Make sure your function is evaluated as f(x/K), so that its argument is in [0,1]
You can also use QPSO algorithm.
https://www.sciencedirect.com/science/article/pii/S0957417409002978
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression