NetNews Usenet Archive 1992 #20/NN_1992_20.iso/spool/comp/ai/neuraln/3472

Text File | 1992-09-12 | 3.2 KB | 56 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!brunix!cs.brown.edu!pcm
From: [e-mail] (Peter C. McCluskey)
Subject: Re: NNs in chess or other games
Message-ID: <[e-mail]>
Sender: [e-mail]
Organization: Brown Computer Science Dept.
References:  <[e-mail]>
Date: Sat, 12 Sep 1992 18:23:30 GMT
Lines: 44

In article <[e-mail]>, [e-mail] (Peter Bruhn) writes:
|> Does anyone of you know something about an application of neural nets
|> in strategic games (esp. chess). What I think about is that a NN could
|> be trained to evaluate a position or some aspects of a position (e.g.
|> the structure of your pawns, how safe is the position of your king,...).
|>
|> The idea behind is that evaluating a position is (as I see it) closely
|> related to pattern-recognition. An experienced human chess player is
|> able to identify a position as weak, even though there are no tactical
|> dangers to perceive and he/she has not seen the position before. So how
|> does he/she know? I think it is due to pattern recognition: He/She has
|> seen similiar positions before or has made the experience that a certain
|> pawn structure is a weak point. Once a chess master (I don't remember his
|> name) was asked how many moves he prefigures in advance and he answered
|> "normally: none". Even though this is an exageration, of course, his

 I think the quote you are thinking of was from Capablanca, and was about
how many different continuations he looked at for a given position, the
answer being "Only one, but it is the right one", the implication being that
a sufficiently powerfull move generator will generate the most important
moves first, and looking at the first few that it generates is usually
sufficient. This pruning of the search tree enables good players to look
farther ahead; they probably average 5-10 moves deep.
 A well trained NN should be able to take a representation of the current
position as input and produce one or more "important" moves as output, for
use in deciding what parts of the search tree to evaluate, and should also
be able to output an evaluation of the position for use in comparing the
results of moves.
 The standard NN algorithms could be trained on targets such as moves from
typical grandmaster games or expert evaluations of the most interesting
moves from a given position. For evaluating how good a position is, you
would probably need to get expert evaluations.
 The standard algorithms could be supplemented with reinforcement learning
algorithms, such as having 2 different networks play each other, and
rewarding the winning network, punishing the loser.
 The biggest problem is likely to be choosing a good input representation
of the current position. I suspect a good representation will compare to
a straightforward 8 x 8 array of squares in much the same way that the
output of the human optical nerves compares to bitmapped graphics.

----------------------------------------------------------------------
  \\ !Vote Libertarian! \\ Peter McCluskey  \\ !Vote Libertarian! \\
  // Vote Marrou in `92 // [e-mail] // Vote Marrou in `92 //
----------------------------------------------------------------------