Newsgroups: comp.ai.games
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!travelers.mail.cornell.edu!news.tc.cornell.edu!news.cac.psu.edu!howland.reston.ans.net!swrinde!gatech!willis.cis.uab.edu!news.ecn.bgu.edu!siemens!darken
From: darken@scr.siemens.com (Christian Darken)
Subject: Monte Carlo-based solutions to games
Message-ID: <D9EKp8.MGn@scr.siemens.com>
Sender: news@scr.siemens.com (NeTnEwS)
Nntp-Posting-Host: avocet.scr.siemens.com
Organization: Siemens Corporate Research, Princeton NJ
Date: Tue, 30 May 1995 18:00:44 GMT
Lines: 33

To: Henrik@UniX11.com
Subject: Re: How to teach a computer to bluff?
FCC: doc/mail/mail.log
--text follows this line--

 >Trouble is, exact calculation of optimal strategies is intractable for
 >all but simple (and I do mean simple) games.  I agree with Bruno that
 >analysis for a game of the type and scale as Stratego is impractical.
 How about Monte Carlo simulation ?

 Cheers,
    Henrik

Yes, Monte Carlo simulation is always a possibility, and there are
existing (though hardly fully developed) approximate solution
techniques which rely on it (take Gerry Tesauro's backgammon program,
for example).  However, the problem with Monte Carlo is always that
some very important (i.e. very very good or very very bad) events are
also rare, and thus may be only very rarely seen (if seen at all) in a
set of Monte Carlo runs of practical size.  This is in stark contrast
to the human player who usually knows where the "pots of gold" in a
game are and will often design entire strategies around finding one of
these.  For games with important rare events much would seem to be
gained from augmenting (at the very least) Monte Carlo with some human
insight.

-----------------------------------------------------------------------
Christian J. Darken                  phone: (609)734-6536
Member of Technical Staff            fax:   (609)734-6565
Learning Systems Department          email: darken@scr.siemens.com   

Siemens Corporate Research, 755 College Road East, Princeton, NJ  08540

