Author: Otello Gnaramori
Date: 09:31:33 12/29/01
Go up one level in this thread
On December 29, 2001 at 10:50:52, James T. Walker wrote:
>Back when Deep Blue was Deep Thought I think I read somewhere that it's eval was
>automatically tuned by an algorythm that took 1000 GM games and adjusted the
>eval to most closely play the GM moves. Maybe I'm wrong and somebody (BOB) can
>confirm/deny this. By the way I never thought this was a good idea when I read
>it.
>Jim
You're right , in fact I've found a doc. where Mr. A. Nowatzyk explains his
tuning algo , it's shareable as written in the end of the doc. :
"The files in this directory constitute the tuning program that was used by
Deep Thought to adjust its evaluation function parameters based on a set
of some 868 grand-master games. I forgot where these games came from, but
we did not type them in. It was last used in the summer of 1988 and it is
believed that this program might be of historical interest to some chess
programmers. The Deep Thought hardware is probably no longer functional
and without the actual Deep Thought program, this code can only show how
DT evaluated chess positions, but it can not play any chess.
This program was written by me during the Spring of 1988 and included
suggestions and feedback from the entire Deep Thought team (Feng H. Hsu,
Thomas S. Anantharaman, Murray S. Campbell, Mike C. Browne and myself). It
includes the DT evaluation function, which was developed by Murray.
This file gives a brief description how this tuning program worked and how
to use it. Expect some errors and omissions because I'm writing this from
memory, 12 years after I last touched this code.
The basic method used the mathematical concept of least square fitting.
This was hardly new and it had been applied to chess evaluation functions
before. However, there are plenty of details on exactly how to go about
this, and our approach likely differed from earlier work.
Let's suppose that the evaluation function is a weighted sum of positional
features (later referred to as a feature vector):
E(P) = SUM Ai * Fi(P)
i
For a given chess position <P> (= position of pieces on the board plus
castle and enpassant status), the evaluation <E(P)> is the sum of the
features recognized by Deep Thought <Fi(P)> times the weight given to each
feature <Ai>. For example, a feature may be the number of white pawns minus
the number of black pawns. The corresponding weight would be the value for
one pawn. There were roughly 100 features that Deep Though used. Some were
implemented via a piece-placement table that could give a different weight
for each piece depending on where it is on the board. For example, a
gradient in the pawn value could be used to add a bonus for advanced
pawns. King centrality was implemented likewise. There were five other
tables in the hardware for more complex features, that could detect open
files, passed/doubled pawns etc. (four pawn structure tables of 8192
entries each, a rook evaluation table with 2048 entries and the 1024 entry
piece/placement table along with a few special bonus registers made up the
DT evaluation hardware. While these nearly 40,000 programmable parameters
of the DT hardware could be regarded as the components of DT's evaluation
function, they all were derived from the 89 to ~100 parameters mentioned
earlier). The basic DT move cycle consisted of computing these tables
before every search so that the weights of the evaluation function could
be adjusted according to the overall situation of the game (opening,
mid-game, endgame, etc.). This took some time, so DT was not good at fast
games (keep in mind that DT used '88 hardware and the VME interface from
the host, a Sun 3 or 4, was not a spead deamon).
(....)
I would like to thank Feng H. Hsu and Murray Campbell for their permission
to release this code. Also let me add a note of appreciation for the
Computer Science Department at the Carnegie Mellon University, its open
and free-spirited environment made the Deep Thought project possible. If
you would like to know more about the roots of Chiptest and Deep Thought,
look out for Feng Hsu's upcoming book on his Deep Blue experience."
Share and enjoy,
-- A. Nowatzyk
January 2000
agn@acm.org
-----------------------------------------------------
The doc. named DT_eval_tune.txt and the files DT_eval_tune.tar.bz2 are at :
http://www.maths.uq.edu.au/~rwb/chess/crafty/
w.b.r.
Otello
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.