Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Research idea? re: weight optimization

Author: Sune Fischer

Date: 05:03:40 12/29/01

Go up one level in this thread


On December 29, 2001 at 07:48:38, Tom Kerrigan wrote:

>>Still you dimensional search space is very large, and with each concensus eval
>>that expensive, you can't really afford a sloppy approach.
>>Genetic algorithms might do the job if you can define a good consensus/scoring
>>function, which is no trivial thing to do in the first place.
>>
>>-S.
>
>I think we're talking about different algorithms. My idea is for making a
>program play a known-good move, not to find ideal values for all eval weights.

Well possibly, but when I read point number 3 I still see you comparing scoring
values (getting the deltas) and then apply a "sloppy" (no offense:) arbitrary
change in the weights. This can be done better I think, there is a lot of stuff
on tuning weights in the litterature, genetic algorithms is one of the really
big guns, and I believe this situation calls for some heavy artillery.

>The hope is that if you force the program to make enough good moves, the eval
>weights will end up being pretty good.

I'll buy that, this is one way of defining a consensus function.

-S.



>-Tom



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.