Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Automatic Learning

Author: Franck ZIBI

Date: 03:02:06 09/06/00

Go up one level in this thread



Hi Vladimir,

There are (at least) three kinds of Chess Learning:

1/ Book Learning
The goal is that the program updates its opening book after each game to
choose lines where it scores better.
An interesting program using book learning is WChess : it uses a very small
opening book, that is updated (and increased) after each game.

2/ Position Learning
Each time the program blunders, it stores its blunder in a file, so the program
will not make the same blunder in *exactly* the same position.
This is very useful when doing long test match between chess programs (Cf.
SSDF) or when playing on a chess server against the same human opponent.
For instance I remember a match between IGM Igor GLEK and ZChess on the ICS:
in the first game, ZChess (with black) played O-O-O?? and lost few moves later.
in the next game with black, it still played O-O-O?? but understood earlier that
it was lost. Eventually, in the third game it played the much stronger O-O.
and drew the game ... thanks to position learning.

3/ Evaluation Learning.
The idea is to automatically tune the parameters of the evaluation using
'Temporal Difference' (TD) method.
This will not add any new knowledge to the engine, but will only try
to improve the existing knowledge.
KnightCap is using TD Learning, and KnightCap authors wrote a nice article about
it in a previous issue of the ICCA.
Also the strong French program Capture is using it.

Regards.



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.