Computer Chess Club Archives


Search

Terms

Messages

Subject: Chessterfield: using a neural net to learn evaluation functions

Author: Chris Janeke

Date: 10:56:25 05/03/00


I recently downloaded the freeware program Chessterfield from Mathias Luscher's
site, http://www.luescher-online.com/computerchess.html.   The program is not
very strong.  It falls in the lower group of winboard engines, and is probably
comparable in strength to engines such as Faile  and Averno (I would estimate
its elo at slightly under 1900).   Despite its no more than moderate playing
strength,   the engine is special for two reasons:  Firstly its evaluation
functions were learned by means of a simple linear neural net.   In fact,  the
author claims that it learned all its positional knowledge from games played by
other winboard engines and by analysing some endgame databases.  The learning
technique used, is described in Michael Buro's paper: "From simple features to
sophisticated evaluation functions" – there is a link to the paper (in
postscript format) at  Luscher's site.     Secondly,  the winboard version does
not come with an opening book,  but the engine seems to run perfectly  under the
chessbase winboard adapter where it can use any  opening book created by a
chessbase program such as Hiarcs 7.32, etc.  In the Chesssbase  interface it is
actually quite an entertaining  opponent for any moderate level player, but be
warned, it's not in Hiarcs, Fritz or Crafty's league!



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.