From: cljones <cljones AT concentric DOT net>
Newsgroups: comp.os.msdos.djgpp
Subject: Re: [off topic] neural nets in games
Date: 03 Aug 1998 00:22:16 EDT
Organization: Concentric Internet Services
Lines: 37
Message-ID: <35C53B3A.52D0@concentric.net>
References: <35BF5909 DOT 3BFE AT cam DOT org>
Reply-To: cljones AT nospam DOT concentric DOT net
NNTP-Posting-Host: ts002d15.col-co.concentric.net
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
To: djgpp AT delorie DOT com
DJ-Gateway: from newsgroup comp.os.msdos.djgpp
Precedence: bulk

A better (faster) solution than NN for games like tic-tac-toe is
a backtracking algorithm, i.e. the computer tries a move, then
works backwards to see if any rules were violated (or if the move
puts him in 'check' or whatever).  NNs are more generally used
when there is plenty of time for the computer to learn what a
particular pattern looks like, then use some rule for the next move
by the computer.  With an unspecified number of rows and columns,
this becomes impracticable.  You might use a backpropogation
network if say you always used a 3x3 board - you then train the
network to recognize every possible x/o permutation on the board
and train it to some given rule for each permutation, i.e., if the
board looks like this, do this move.  You could do that also for nXn,
but realize there is some overhead training the network once he
knows what the board dimensions are.

For a 3x3 board, I would use 9 inputs into 9 rules, with perhaps an
intermediary layer of size {number of empty boxes}.

I've played a little with 3 layer backpropagation networks on a pc
and discovered I could really bog on anything greater than about
8x8.

-Chester

Vic wrote:
> 
> Hello. Among other things, I'm interested in AI.
> I wanted to do a neural net with a buddy but neither of us has any
> experience with NNs. We wanted to do a TIC-TAC-TOE like game, but with
> any number of columns and rows. I know it's easier with a more "normal"
> approach, but the point of this is not the tic-tac-toe, it's the neural
> net. So, the question is, what neural net can I use here , and where can
> I read more about it?
> the only way this is related to DJGPP is that I would use DJGPP/Allegro
> to do it :)
> 
> TIA,