Computer Chess Club Archives




Subject: Neural nets in backgammon

Author: Albert Silver

Date: 20:31:05 04/07/04

Go up one level in this thread

On April 06, 2004 at 20:16:25, Dann Corbit wrote:


Fascinating. I wonder how much the author might benefit from studying the way
neural net technology has been used to develop the latest crop of backgammon
programs: Snowie 4 and GNU Backgammon. Interestingly enough, GNU Backgammon is
actually the strongest one around. I even authored the tutorial on it at

The author of the neural nets is Joseph Heled, who keeps a detailed online
journal ( - at bottom left be
sure to change 'concise' for 'full') on the work he has done on them and his
methodlogy and results. For example, in backgammon, more than one neural net was
developed to cover different aspects of the game. The largest and most complex
one is certainly the middlegame neural net, aka the contact net. Unfortunately,
programs were absolutely abysmal when it came to races, where there wasn't any
contact between the checkers of the players, and precise maximalization of the
dice was necessary to achieve the best plays. Older programs were disastrous in
this, so a special neural net was developed to deal with such situations. The
programs also have look-ahead of course, and yes, it will consult the race net
if it sees it could encounter a race situation 2 plies ahead. Perhaps developing
some special nets for endgames, or even subsets of endgames, or whatnot, might
also be possible. Also note that the previous technology developed by Gerald
Tesauro ( presented
limitations in its progress, as documented by Heled, and special training was
done to further its development.


This page took 0.03 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.