Author: José Antônio Fabiano Mendes
Date: 11:21:15 05/25/05
Please see ==> http://www.chesscafe.com/Reviews/books.htm
A Conversation with Team Hydra:
Howard Goldowsky: How does the strength of today's Hydra compare to the strength
of Deep Blue of 1997? Hydra computes 40 million positions per second on average,
where Deep Blue computed 200 million positions per second on average. Does
Hydra's evaluation function make up for the factor of five deficiency in speed?
Muhammad Nassir Ali: Deep Blue's rating was about 2800 Elo. The rating of
Hydra-Chimeira (the version which played in Bilbao, at the 2005 Computer Chess
Championships, in February of this year) is at least 2950. It has scored against
strong GMs 7 out of 8, and in the 9 direct matches with Shredder, 6.5 to 2.5. No
game was lost. Hydra should therefore win against Deep Blue by a wide margin.
There is an old rule in computer chess: The evaluation determines the playing
style, the search the playing strength. This does not mean that a smart
evaluation is worthless, but a 5-ply searcher can never be [as high as] 2200
Elo, a 20-ply searcher never below 2200. It is essentially the search which is
much more efficient in Hydra than in Deep Blue. This concerns every part: the
FPGA/Hardware search, the much more selective software search, and the better
parallelism. The mean search depth of Deep Blue was about 12, the mean search
depth of Hydra is 17-18. This is comparing apples with oranges, but it should
give a hint about the different search strategies. The evaluation is probably
more dynamic, less materialistic than in Deep Blue. This is related to some
design decisions. The Deep Blue evaluation was relatively slow. The full
evaluation took 11 hardware cycles. Hsu had to introduce a fast, mainly material
based evaluation. If, for example, one side was a minor-piece down, only the
fast evaluation was used. But this means that the full evaluation can never
compensate for a piece. In fact, the margin was set much smaller than the value
of a piece. In Hydra the full evaluation takes 3 cycles. It can therefore be
always used and the dynamic parts can easily exceed the value of a minor-piece.
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.