Computer Chess Club Archives




Subject: Tuning evaluation function weights

Author: Don Beal

Date: 10:23:27 09/30/98

Go up one level in this thread

On September 29, 1998 at 21:42:15, Danniel Corbit wrote:
>Has anyone a pointer to research on tuning evaluation functions using least
>squares fitting or gradient methods?

You might like to look at Temporal Difference methods, in particular
the TD-lambda method proposed by Sutton in 1988.  This was used by
Tesauro to make the world-championship class backgammon player Neurogammon.
It's neither least-squares fitting, nor simple gradient descent, but it
does utilise gradients internally to make small adjustments to weights
after each move.  It can take thousands of games to reach good values,
but they can be speed games.

One paper describing the technique is:
"Learning Piece Values by Temporal Difference learning"
Beal & Smith, ICCA Journal, Vol 20, No 3, Sept 1997.

The method works for any weights, not just piece values, provided
that the evaluation consists of a sum of term*weight components.
Most chess evaluation functions have this structure.

This page took 0.01 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.