[galib] Comparison galib - Matlab genetic algorithm
Fábio Roberto Teodoro
fr.teodoro at gmail.com
Tue Apr 28 14:13:28 EDT 2009
Sorry for the self promotion here, but I usually test several
parameter combinations using this IDE:
http://sourceforge.net/projects/galib-ide, that I am the author. Its
beta (semi alpha) now, but I think its usefull.
One default parameter of galib that i noted that give poor results is
the RankSelector, almost always I change it to RouletteWheel and get
better results.
But it may vary according to the caracteristics of the problem, so to
use your understanding of the problem and to try several combinations
is the best advice, because of this I've recommended the above IDE as
I think it ease these trys.
Sorry for the bad english and hope this help
2009/4/28 gpipc <gpipc at cup.uni-muenchen.de>:
>
> I have compiled the galib genetic algorithm library together with some code
> that I wrote in order to run as a mex-file im Matlab. I hope that there is
> someone who knows enough about both galib and Matlab to give me a hint on
> how to proceed in the situation things are now.
>
> I am now in the following situation: the galib code evaluates chromosomes
> much faster than the Matlab genetic algorithm ('ga') (I would say at least
> 10 times faster), but it takes many more fitness function evaluations from
> galib in order to obtain the same level of optimization that I obtain with
> the Matlab ga; as a result I have not gained a lot in speed (maybe 30% as
> fast, maybe twice as fast, but not more).
>
> Part of the problem is that it not straightforward to test the settings of
> the two programs with fitness functions that evaluate more quickly - my
> impression is that I haven't yet found a simple test fitness function on
> which I can optimize the settings such that the optimized settings will be
> also optimal for the fitness function I am actually interested in. I am
> working on that but on the meantime I look here in the newsgroup for some
> hints.
>
> Does anyone know what are the galib settings which would give as a result
> an algorithm which is as close as possible as the Matlab ga algorithm?
> I have found that for my fitness function (with 32 variables) these ga
> settings start to give some reasonable result
>
> ga_options = gaoptimset(ga_options,'PopulationSize', 50);
> ga_options = gaoptimset(ga_options,'CrossoverFraction', 0.5);
> ga_options = gaoptimset(ga_options,'generations', 30);
>
> In the galib library I have tried both the GASteadyStateGA and the
> GASimpleGA. I realize that the question is a bit involved and that it will
> take some work on my part to get things to go well, but I hope to get some
> hints on which direction to look at.
>
> Thanks in advance for any replies.
>
> ----------------------------------------
> Message sent by Cup Webmail (Roundcube)
>
> _______________________________________________
> galib mailing list
> galib at mit.edu
> http://mailman.mit.edu/mailman/listinfo/galib
>
--
Fábio Roberto Teodoro
More information about the galib
mailing list