Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: deep blue - hard facts from written paper

Author: Vincent Diepeveen

Date: 19:59:33 01/30/02

Go up one level in this thread


On January 30, 2002 at 20:10:57, Dann Corbit wrote:

>On January 30, 2002 at 20:01:04, walter irvin wrote:
>
>>it seems to me that deep blue was a great chess machine ,but i believe with deep
>>fritz and the right hardware a stronger machine could be had .i think it was
>>fritz 3 that proved that a pc could compete with the big boys .
>
>Now that Deep Blue is retired, all the PC's can thump their chest and call out:
>"I'm king of the jungle!"
>
>Who's to argue whether or not the current PC programs could compete effectively
>against Deep Blue?  We really don't know since we don't have a Deep Blue to play
>against.
>
>Really, I could claim that Deep Blue would never lose again.  Or I could claim
>that now it would lose 100% of its games against Diep.  But without anything to
>test again we are back to angels dancing on the head of a pin.

Anyway we know 2 things for sure now by a paper published in june 2001
by Hsu, Campbell, Hoane jr and that's that:

  a) iteration depth 12 means  8 ply software and 4 ply hardware
  b) against kasparov on average they searched 126 million nodes a second,
     and peaking at 200 million and minimum of around 100 million.
  c) they used *many* extensions:
       - singular extensions (even first ply in hardware)
       - a single legal move in a position was extended 2 ply even
       - DB didn't use nullmove pruning, but they used nullmove
         search to detect threats (also mate threats) and
         quickly extended upon this if previous move looked kind
         of forced.
       - check evasion
       - passed pawn extensions
       - in qsearch they extended one ply for hung material and such
       - some checks were tried in qsearch, especially the once that
         solve mating problems.
  d) their evaluation i count about 40 patterns, fromwhich most
     patterns are indexed to an array to give a bonus or penalty
     which results in 8000 adjustable parameters

Note that i miss recapture extension in their paper, though i don't
do it in DIEP either.

So in short i do not doubt with this they can solve quite some tactical
positions with this setup, but their minimum search depth is 12 ply
when it shows 12, for positional assessment this is trivial then.








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.