Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!elroy.jpl.nasa.gov!jarthur!uunet!brunix!cs.brown.edu!cs196006
From: cs196006@cs.brown.edu (Josh Hendrix)
Newsgroups: comp.ai.neural-nets
Subject: Re: Chip vs Chess Master
Message-ID: <70216@brunix.UUCP>
Date: 29 Mar 91 07:28:07 GMT
References: <13807@ccncsu.ColoState.EDU>
Sender: news@brunix.UUCP
Distribution: usa
Organization: Brown Computer Science Dept.
Lines: 31

In article <13807@ccncsu.ColoState.EDU>, gordons@handel.cs.colostate.edu (vahl scott gordon) writes:

|> I am not aware of any neural programs, or even neural programs for
|> subsets of chess.  I think it would be a great area of research (for
|> SUBSETS of chess, that is).

I am in my first neural net course, and I have been thinking of trying to write something that would learn how to play tic-tac-toe or Qubic (tic-tac-toe with a four by four grid and four layers). The biggest problem that I see right now is the credit assignment problem (which I want to do through some neural net mechanism). I want to divide the problem into subproblems and use nets to solve them, and then find some process for communication between the sub-nets. For Qubic, I see the subproblems as:

	1. Making a legal move, no matter where the best move is. I want it to be 									        able to generalize from what it learns about legal moves in one square to
	all squares (i.e. don't place a piece in a square that already has one in 
	it, etc.)

	2. Identifying the best move to make in a given situation
		a) Always move into a blank cell where the opponent has 3 pieces 		                 in any row that intersects that cell.
		b) Try to make moves that block the opponent
		c) Try to get 4 in-a-row
		d) etc...
	I see this as mainly a credit assignment problem, because I don't want to 	
	give the net feedback for every move it makes. I want it to learn by 
	losing and winning games, which means the feedback is delayed, and
	credit must be assigned to the move that lost or won the game (perhaps
	others as well).

	3. I'm sure there are others that I haven't thought of yet...

Soooo, does anyone have any good pointers to literature in the above fields of interest, i.e. generalization and credit assignment in neural nets? I have read the Sutton & Barto pole-balancing paper, and I am thinking about ways to adapt their ideas, but I am definitely interested in others, also. Any references will be appreciated, and I will gladly summarize and post to the net.

Thanks in advance,

Josh Hendrix
cs196006@brownvm.brown.edu