Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: learning

Author: Tor Alexander Lattimore

Date: 00:13:35 04/09/05

Go up one level in this thread


Andrew Tridgell has written KnightCap, a program that uses temporal difference
learning to learn it's evaluation weights. He and a number of others researched
a variant of this that instead of looking at the actual game positions, it looks
at the leaf positions at the end of the programs PV. You can check out his paper
at http://cs.anu.edu.au/~Lex.Weaver/pub_sem/publications/ICML98-62.pdf

I dabbled with his method, but sort of gave up in favour of a different (and
easier to implement) method as below.

My new program learns from positional test suites using a genetic algorithm. So
far it's been fairly mixed success, but it can at least learn the positional
weights to reasonable values that easily beats a version with only piece
evaluations.

Cheers
Tor



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.