Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!newsstand.cit.cornell.edu!intac!uunet!in1.uu.net!207.22.81.9!europa.clark.net!news.mathworks.com!howland.erols.net!ix.netcom.com!komodo
From: komodo@netcom.com (Tom Johnson)
Subject: Which type of network?
Message-ID: <komodoE53vtK.4Ct@netcom.com>
Organization: Netcom On-Line Services
X-Newsreader: TIN [version 1.2 PL2]
Date: Wed, 5 Feb 1997 01:22:32 GMT
Lines: 17
Sender: komodo@netcom18.netcom.com

I have been trying to get a network to predict the outcome of a holdem 
poker hand in terms of raw win chance disregarding betting. This is very 
easy to do by doing simulations, ie play the hand out 5000 times, score 
and then average. I have tried for months to do this with standard BP. I 
have used raw inputs, precomputed inputs and combinations with various 
topologies. The best that I can get is about 10% average error (absolute 
not relative), This is not very useful. I get a large clumping of samples 
with an expected output of 0.20 as there are 5 players with random 
hands.; and an occasional sample near the extremes of the  range. Scaling 
doesn't seem to help very much.

Any ideas on how to better attack this problem? Should I try a 'Boltzmann 
machine' or other technique instead of standard BP NN?

Thanks for any help.

Tom
