Newsgroups: comp.ai
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!news.psc.edu!scramble.lm.com!news.math.psu.edu!chi-news.cic.net!newsfeed.internetmci.com!usenet.eel.ufl.edu!tank.news.pipex.net!pipex!dispatch.news.demon.net!demon!peer-news.britain.eu.net!bcc.ac.uk!link-1.ts.bcc.ac.uk!uceetwh
From: uceetwh@ucl.ac.uk (Theodore Wayne Hong)
Subject: Re: Kasparov vs. Deep Blue
Message-ID: <1996Feb22.003838.34074@ucl.ac.uk>
Date: Thu, 22 Feb 1996 00:38:38 GMT
References: <4g9b43$rtd@lilja.vtt.fi>
Organization: University College London
Lines: 31

dfo@vttoulu.tko.vtt.fi (Douglas Foxvog) writes:

>Does anyone have information on how Deep Blue operates.  

>Obviously pruning and selective tree extension are used.  How is this done?
>Where is the cut off?  How is the static position analysis done?  Are 
>positions rated as stable and non-stable and non-stable positions are 
>not scored without more in depth analysis?

Part of the answer is that DB uses an algorithm (alpha-beta?) where it
only has to refute a  possible move once, rather than refuting it all 
possible ways.  So if it's lucky and picks gstrong deciding moves to
look at first, it doesn't have to search all the remaining lines.  (not
sure how this works out, exactly).

Also, DB maintains a hash table of positions already analyzed, since
many lines will recur by simple transposition of moves.

>In game 2, Deep Blue allowed a rook and bishop to be totally boxed in, 
>obviously not deducting sufficient points from its static analysis function
>for such a situation.

This happened again in game 6, and was a major factor in causing 
Deep Blue to lose that game (apparently, its evaluation caused 
it to continue to think that it was winning many many moves after
its position was completely hopeless).  One of the programmers 
commented afterwards that they had not had time to factor trapped
pieces into the evaluation function.

theo


