Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Deep Blue eval function tuning technique

Author: Jay Scott

Date: 15:51:25 01/08/98

Go up one level in this thread


On January 08, 1998 at 14:43:03, Stuart Cracraft wrote:
>Has the method which was used to tune Deep Blue's eval function
>been described anywhere in depth?

I haven't seen a detailed description anywhere.

There was an article in Scientific American in October, 1990
about Deep Thought. It talks a little about the automatic
tuning algorithms they used back then. I don't know if
they used the same thing for Deep Blue, but either way I
can't tell the difference from what little I've read about DB.

  http://www.sciam.com/explorations/042197chess/042197hsu.html

Chris McConnell, from Hans Berliner's group at CMU, wrote
a short paper about estimating evaluation function weights
from human games.

  http://www.cs.cmu.edu/afs/cs.cmu.edu/user/ccm/www/home.html
  http://www.cs.cmu.edu/afs/cs.cmu.edu/user/ccm/www/papers/ml.ps

I keep track of this kind of thing at Machine Learning in Games.

  http://forum.swarthmore.edu/~jay/learn-game/index.html

I don't think it's such a smart idea to train a computer to
play moves that it thinks are similar to grandmaster moves.
Chess programs have such different reasoning abilities
than grandmasters that the computer won't see why the human
played the move, and it will have to fall back on superficial
similarity. Methods where the program learns from its own
games are likely to work better. For example, see KnightCap:

  http://wwwsyseng.anu.edu.au/lsg/knightcap.html

Jay



This page took 0.09 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.