Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Local optima / genetic algorithms

Author: Jay Scott

Date: 04:10:33 08/04/98

Go up one level in this thread



On August 03, 1998 at 16:57:04, James Long wrote:

>Most programmers hand tune their eval functions, either
>themselves or with the help of a strong player.  Is
>this because the results are generally better, or
>because a good learning eval is difficult to write?

The learning is pretty easy to write. There are a lot of choices
of how to write it, and it helps to be knowledgeable about them.

Results have generally been better with hand-tuned evaluators.
A few learned evaluators have approached the strength of
hand-tuned evaluators. But in my view, most of the attempts
at learning have been weak in some way.

Also, learning takes a long, long time. You'll want a
fast computer, preferably many of them.

>For the program to "learn," the eval variables must
>mutate dependent on the outcome of a game.  One approach
>I read about is to set up a tournament of 100 "players,"
>each player having random values assigned to its
>eval terms (with the exception of the pawn score).
>When one player finally wins, the weakest player's
>terms are modified.  This requires thousands of games,
>but I'm in no hurry.  My question is - how to go about
>modifying the eval terms.

1. Fix the pawn score at 100 (or whatever). Let the other
terms vary. That eliminates a useless degree of freedom,
which has several good effects.

2. Learning will be much faster if the most important terms
(like material) start out in the right ballpark.

3. Modify the terms randomly. Crossover probably helps.

4. It's crucial to play enough games to tell the strong
evaluators from the weak ones. After a while, the differences
will become small and you'll need many games.

For more on genetic algorithms, here are a few starting points.
The first is the most suitable for beginners.

ftp://alife.santafe.edu/pub/USER-AREA/EC/Welcome.html
http://www.aracnet.com/~wwir/NovaGenetica/
http://www.aic.nrl.navy.mil/galist/

  Jay



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.