Newsgroups: comp.ai.games
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!tank.news.pipex.net!pipex!in1.uu.net!ncrgw2.ncr.com!ncrhub6!daynews!falcon!news
From: Dick Menninger <Dick.Menninger@DaytonOH.ATTGIS.COM>
Subject: Re: do AI games exist ;-)
X-Nntp-Posting-Host: 149.25.99.44
Message-ID: <DDo3KM.ts@falcon.daytonoh.attgis.com>
Sender: news@falcon.daytonoh.attgis.com (News administrative Login)
Reply-To: Dick.Menninger@DaytonOH.ATTGIS.COM (mennid)
Organization: AT&T Global Information Solutions
X-Newsreader: DiscussIT 2.5.0.10 for MS Windows [AT&T Software Products Division]
References: <DDJ02w.CK2@actcom.co.il>
Date: Mon, 21 Aug 1995 15:41:10 GMT
Lines: 64


> ==========Uri Bruck, 8/18/95==========
> 
> Melle Koning (melle@rtbbs.iaf.nl) wrote:
> 
> : That is because computers still don't *understand* the game of
> chess. If good
> : chessplayers play very different (they use to call this
> "not-chessplaying"
> : when they do), they can win easily against any computerprogram.
> 
> : In other words, humans can learn to beat chessprograms by
> altering their
> : gameplay.
> 
> Game programs (such as programs) generally assume that the human oponent
> will make the best move possible and try to figure what is the
> best counter 
> move to that (minimax). I am not suggesting that one can win by making
> bad moves, one can't. This has been extensively tested, although not
> necessarily intentionaly. However, by constantly picking a good
> move, which 
> may not be the best, such as second best, according to the computer's 
> evaluation...
> The best way to test that would be to set up a program to play
> against itself
> when one side plays the bestmove and the other always plays the second
> best (some programs can be forced to play second best move, but this
> requires human intervention for every move)
> flame away people
> Uri
> 
Sometimes there is only one legal move: you make that one.
Sometimes there is only one winning (or non-losing) move: you make that one.
Sometimes one move is vastly superior: you make that one.
Sometimes there are several moves that evaluate as best:
you pick one at random (though you may consider more).

Sometimes there is some divergence in the value of the moves at top
where there is some overlap in their profiles (assuming the scoring
does that sensible thing): you pick at random but weight by some
likelihood of being the best.  Now minimax does not really work
well with this as it assumes absolutist scoring instead of
fuzzy scoring.  This is a weakness of minimax depth-first searching.
It leads to somewhat dumb scoring functions and the assumption
that searching deeply will make it less dumb.  That is
the major source of the "computer" style of play.

However, it is possible to do much smarter scoring functions
that understand the fuzziness of their evaluation of the current
positon (or moves from the current position).  Such a scoring function
goes naturally with a radically different tree search method
that selectively grows the tree and can reuse all previously scored
positions.  This is the approach that will more readily allow a
chess program to do its own opening theory.

So, picking the second best move is often inherently problematic.

Just a few observations.


Good Day
Dick

