Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: learning by playing against yourself

Author: James Robertson

Date: 12:10:07 01/04/99

Go up one level in this thread


On January 04, 1999 at 14:21:31, Jay Scott wrote:

>On January 04, 1999 at 12:03:57, blass uri wrote:
>>I read that in backgammon the best program learned by playing against itself.
>>
>>Did someone try it in chess and what were the results?
>
>Yes. It doesn't work as well in chess. An issue that is not important
>in games between naive programs, as they are learning, won't be learned.
>If the program is not smart enough to notice that, say, attacking an
>isolated pawn ties down the enemy forces to defense, and recognize that
>that makes a difference, then it won't learn to avoid weak pawns.
>That's an example of an important but subtle idea that a program might
>have trouble learning by self-play.
>
>One idea about why it works so well in backgammon is that the
>in backgammon the dice force a self-play program to explore all the
>important parts of the game space. In chess, a learning program can make
>the same mistakes over and over and never correct them because it doesn't
>know how to punish them.

Well; they know how to punish them, but they don't know why. All they know is
that they won the game at the end. The problem with chess learning is that every
aspect learned must be put in by the program by the programmer. e.g. if the
programmer puts piece/square table learning in, the program learns piece/square
tables. If they put doubled pawn awareness in, the program fiddles with the
doubled pawn scoring. And so on...

James

>
>Machine Learning in Games:
>http://forum.swarthmore.edu/~jay/learn-game/index.html
>
>  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.