Newsgroups: rec.games.chess,comp.ai.games,rec.games.abstract
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!cis.ohio-state.edu!math.ohio-state.edu!howland.reston.ans.net!torn!watserv2.uwaterloo.ca!undergrad.math.uwaterloo.ca!eadengle
From: eadengle@cgl.uwaterloo.ca (Ed "Cynwrig" Dengler)
Subject: Re: Looking for Checkers program
Message-ID: <DAHC0r.KtB@undergrad.math.uwaterloo.ca>
Sender: news@undergrad.math.uwaterloo.ca (news spool owner)
Nntp-Posting-Host: cgl.uwaterloo.ca
Organization: University of Waterloo
References: <3s5r87$g87@nntp.Stanford.EDU>
Date: Tue, 20 Jun 1995 16:19:39 GMT
Lines: 31

Greetings!

In article <3s5r87$g87@nntp.Stanford.EDU>,
Kong Kritayakirana <kong@beren.Stanford.EDU> wrote:
>I would like to ask for information/pointer about some computer checkers
>playing program. I am now working on a neural net checkers project. The
>other computer checkers program is required because neural networks need
>to have a teacher. Right now the data we use to teach is from Dr.
>Samuel's
>original database, and hence there's no way to tell how well my program
>does in a quantifiable sense except how well it imitates the Samuel's
>data.

Rather than using a teacher, you can use time-differential reinforcement
learning so that your neural net is trained to predict the outcome of a
game.  See any paper by Tesauro on his backgammon program TD-Gammon
(an article recently appeared in the Communications of the ACM) for
a quick description and references, or papers by Richard Sutton for the
gory details.

The use of this method has the unique advantage that programs learn to
play into their strengths, rather than being trained to set up positions
they are weak at playing (most chess programs have run into this one).

>
>Thanks very much in advance.
>Kong
>kong@leland.stanford.edu

Ed

