Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: any good experiences with genetic algos or temporal difference learning?

Author: Jay Scott

Date: 11:36:18 01/01/01

Go up one level in this thread


On January 01, 2001 at 10:14:18, Rafael Andrist wrote:

>The parameters of my evaluation function are at the moment not very good tuned.
>I've heard about genetic alogorithms and also temporal difference learning to
>tune these parameters. Does the values converge to the right values and how much
>time is needed?

For temporal difference learning, see KnightCap:

http://www.syseng.anu.edu.au/lsg/knightcap.html

Many people have had flings with genetic algorithms to tune chess
evaluators, but nobody that I know of has settled down. That may be
because it takes such a tremendous amount of CPU time that nobody
has been able to afford it--figure every new evaluator will need at
least a hundred games to find out how good it is (likely more), and
you'll have to go through thousands of evaluators (likely more). Here's
one report, from my Machine Learning in Games site:

http://satirist.org/learn-game/methods/ga/chess.html

I have proposed that these two methods could be combined:

http://satirist.org/learn-game/inspire/tdga.html

>Are manually tuned parameters better?

So far, for chess, yes. Not for othello, which has had a tradition since
the 80's of including large tables of patterns in the evaluator with
machine-generated values attached to each pattern.

In chess, hand-tuned evaluators have always been traditional.
Whether this tradition's successful because it's the best way, or only
because it's the way people have worked to make successful, is an open
question. :-)

  Jay



This page took 0.03 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

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