Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Web space for evolutionary computation project

Author: Andrew Wagner

Date: 15:36:00 12/07/05

Go up one level in this thread


Hi Mark, thanks for your comments. See my responses below.

On December 07, 2005 at 08:58:44, Marc Boulé wrote:

>Hi Andrew,
>
>Sorry, no webspace to offer. I just wanted to say that I find your project very
>interesting and I look forward to seeing some updates.
>
>At first, it might be best to only evolve heuristic parameters and/or >evaluation parameters.

Yes, but it's also been tried :)...yawn! To the best of my knowledge, nothing
like what I'm suggesting, where the acutal algorithms are selected by evolution,
has been tried quite like this.

>If the genome contains too many features (or bits), evolution might
>not converge quickly enough.

I'm not sure quite what you mean by this. My thinking is to have a computer
that's dedicated to a process that simply loops through the following steps,
starting with a set of engines that have their "genes" randomly selected:
1.) Randomly select two engines to play each other (initially this will be two
of my evolving engines, but eventually it will be one of mine and one other).
Randomly select a time control, and have them play.

2.) Repeat step 1 until each of my evolving engines has played at least X games.
I think I'll probably start with X at 20, and see how it goes.

3.) Select two engines, with the engines that have a better record in their 20
games having a higher likelihood of being selected than those with a worse
record. These two engines will be "parents". Somehow merge the genetic strings
of these two engines (haven't decided how to do that yet), to create a child.

4.) Repeat step 3 until there are as many engines in the new population as in
the old.

5.) Go back to step 1.

Note: Somewhere in there, I should probably allow for random mutations to occur,
though I haven't decided how yet.

>I am also curious to know how the fitness function
>will work: will you make it play entire games in order to evaluate the fitness
>of an individual, or will you perhaps make it search only a list of individual
>positions? If evaluating the fitness of an individual takes a lot of time, and
>you have thousands of individuals in the population, and you want to run
>thousands of cycles (or generations), this could take a long time :-(

I haven't decided about the population size yet, I'm thinking of keeping it
pretty small. I figure maybe 20 engines, 20 games per engine per generation.
Part of my goal in all this is to experiment with such generational parameters
and see what happens.
>
>Mutation regards,
>
>Marc Boulé
>

Thanks again for the interest. Andrew



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.