I'm expecting that the solution space is going to be very non-linear with each gain/gene being highly interrelated. That tends to rule out a single-threaded search in my opinion. I do think (fingers crossed) that the optimal solution will be robust, however, at this point in my project that is only speculation. So for the robust global minima, I would expect that a mutation only approach could lead to the solution quickly... I just have to get close enough to the solution for that to be the case.

One thing I may attempt is to initialize with a large population, then cut down the bottom half after each iteration till I get to some arbitrary smaller population. That way I can cover larger ground without too much recurring cost.

I am going to be performing analysis in parallel with the GA/Optimization function/Whatever, but if there is something that I am missing, or something fundamentally wrong with using the method on this size or this type of problem, I would rather find that early on.

I presume by your brief description that you have internal oscillatory loops of neurons comprising excitatory and inhibitory neurons?

Quote: Original post by birdtracker
I'm expecting that the solution space is going to be very non-linear with each gain/gene being highly interrelated. That tends to rule out a single-threaded search in my opinion. I do think (fingers crossed) that the optimal solution will be robust, however, at this point in my project that is only speculation. So for the robust global minima, I would expect that a mutation only approach could lead to the solution quickly... I just have to get close enough to the solution for that to be the case.

One thing I may attempt is to initialize with a large population, then cut down the bottom half after each iteration till I get to some arbitrary smaller population. That way I can cover larger ground without too much recurring cost.

I am going to be performing analysis in parallel with the GA/Optimization function/Whatever, but if there is something that I am missing, or something fundamentally wrong with using the method on this size or this type of problem, I would rather find that early on.

Given a good understanding of the search space, there will almost always be other specialized algorithms that perform better than GAs. However, in the current case, I would say that you should probably start by using the GA in an exploratory manner. Finding solutions that are pretty good may help you understand what it is that you are looking for to be exact. So, start with a relative small population size of about 50 to 100 and just let it run a while and monitor the fitnesses. When the best fitness gets to a point that you feel is pretty good, pull that solution. After you've pulled a few good solutions, you can then go back and see what it is you're looking for. Then you either tweak your GA in some special way or use something to target certain solution types. Personally, for very large problems, that would be my approach.