Computer Chess Club Archives




Subject: Re: Evaluation by neural network ?

Author: Jay Scott

Date: 12:00:04 11/10/97

Go up one level in this thread

On November 10, 1997 at 06:43:42, Mark Taylor wrote:
>Has anyone tried position evaluation by neural network?

This is a frequently asked question. It'd be nice
if somebody would put together a subject FAQ for
computer chess.

As a couple replies mention, the obvious ways of using
a neural net as an evaluator are not likely to work well,
at least not without hardware assist.

But the neural net field is huge and diverse. Besides
all the great ideas that haven't been had yet, there
are a few already known that might be good enough.

See the paper "Efficient neural net alpha-beta-evaluators"
by Alois Heinz for one idea. These networks
accept alpha and beta from the search, and can do
alpha-beta cutoffs in the middle of network
evaluation. This is a generalization of the
lazy evaluation trick that many chess programs
use: if the material score is so big that the
positional score can't bring it within the alpha-
beta window, then evaluation can be cut off.

The paper is available online. Here are two web
sites which link to it.

Alois Heinz's page is in German, but the papers
listed at the bottom are in English.

My web site, Machine Learning in Games, includes
information about learning game programs and
methods that can be used in learning programs.
chess programs. Follow the "online papers" link
for Alois Heinz's paper with my capsule review.


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.