Author: Roy Brunjes
Date: 00:22:06 11/06/99
Go up one level in this thread
On November 06, 1999 at 02:00:13, David Eppstein wrote:
>CCC readers may be interested in the article "Knowledge Discovery in Deep Blue",
>by Murray Campbell, in the latest Communications of the ACM. For ACM digital
>library subscribers, it can be found at
>http://www.acm.org/pubs/citations/journals/cacm/1999-42-11/p65-campbell/
>
>Anyway, it's about how DB plays in the opening phase of the game. According to
>Campbell, the actual book (by which he seems to mean, an explicit move tree
>generated by a human) is "only a few thousand positions". Apparently, though,
>it also has a database of millions of grandmaster game positions -- if the
>program reaches a position in this database but out of book, it does a search in
>which database information is used to award bonuses to certain moves. The basis
>for awarding these bonuses is apparently a complicated ("nonlinear") function of
>features such as game outcome, relative and absolute numbers of times the move
>has been played, how recently it was played, etc.
>
>The example he uses is DB-Kasparov rematch game #2 (the Ruy Lopez that Kasparov
>resigned even though he could have forced perpetual check). According to
>Campbell, the first nine moves were in book. Moves 10-17 were not in book, but
>the database bonus was so large that it just made the move without searching.
>Finally, on moves 18 and 19, it made a search in which the moves that ended up
>being made (18 Qd2 and 19 a4) were given bonuses of 17 and 9 centipawns
>respectively. After move 19, the game went over new ground and this database
>machinery stopped being useful.
>
>He also mentions the second game of the first match, in which the explicit book
>was apparently disabled, but the database extension to the book still worked,
>and allowed the program to follow theory for 13 moves, reaching (he claims) "a
>respectable opening position" although the game was later lost.
>
>The paper is only three pages long, though, so if you're looking for technical
>detail it won't be there.
Interesting. Sounds a LOT like what Ed Schroder has done with Rebel and the EOC
feature. If some still wonder just what EOC is useful for in Rebel, this
writeup by Murray Campbell may help demonstrate how useful this idea is. I'm
confident there are differences between the IBM team's approach and Ed's, but
the overall concept is very similar.
Roy
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.