Author: Vincent Diepeveen
Date: 18:42:43 08/25/01
Go up one level in this thread
On August 25, 2001 at 20:47:44, Mig Greengard wrote:
>Sorry to dredge this up yet again, and ignore this rather than turn it into a
>flame war or something worse. I know feelings on this topic can run hot.
let's look to facts, ignoring it played kasparov
- deep blue searched between 11 and 13 ply, most likely 12 ply
- it's mainlines are complete nonsense
- it's very sure that kasparov played his worst chessgames ever
and initially still seemed to get away with it, until the disaster
happened as in game 6
- in 1997 i can't remember a single program searching 12 ply
without major forward pruning (i'm not referring to nullmove here
but way more dubious stuff)
- even in 1999 many programs could get away at the wcc with a very bad
endgame
- nowadays all programs seem to outperform any program from 2000 by a
huge margin, even if we only look to the improved endgame at most
programs, or the improved knowledge (see tiger or junior
for example) in engines.
- junior won this years wmcc in a way in which we would not have imagined
a few years ago, even from lost positions it recovered and still won,
idem for the escape from shredder against diep.
- in 1997 book from deep blue was said to be automatically generated and
the games provide us with no more info as than that was true
- in 2000 and 2001 WMCC definitely in many games the book completely
dominated the play
- most of todays engines are very well because of testing, look for example
to how Xinix did in the wmcc. In 1997 wmcc he definitely would have
scored way more than 0.5 points. He has in fact 0.5 points because of
poor testing. As far as facts are heart about deep blue, the hardware
chips only shipped a few weeks before the match. Also we know that
it only played 10 rapid games or so against a preprocessor program
that bigtime forward pruned (rebel8).
It's logical to conclude that deep blue wasn't tested very well. In
fact even its hardware hashtables Hsu didn't have time to design
RAM for.
I think we can with deduction to all aspects of a chess engine
conclude that deep blue would get crushed completely. Of course
if you put a todays engine into hardware or redesign an engine into
todays hardware standards, opinions might change again.
>Although we do not have enough of Deep Blue's games to make anywhere near an
>accurate assessment of its chess strength, I am requesting a summary of thoughts
>on how today's top programs measure up on a science level. In the past I've seen
>some admirably objective breakdowns on this topic from Bob Hyatt and a few
>others, but did not save them.
>
>Put Deep Fritz, or other top programs, on the best available platform on which
>they can run, and I imagine this is what they will have in Bahrain, and knowing
>what we do about DB, what comparisons can we make?
In short from my viewpoint deepfritz is better at all terrains than deep blue
with one exception: deepfritz engine itself i don't consider best engine,
but considering that the only weak point of it is its knowledge (like
it seems to know nothing from good/bad bishop for example), all other
terrains are definitely enough to kick the hell out of deep blue like it
was in 1997.
Don't compare that with Bob's statements, he is comparing with a machine
which would exist if a todays program would get made at todays hardware.
Note that making a hardware chess processor at 0.18 or 0.13 microns is
quite expensive. Just the machines needed to press such a processor might
all 5 machines together (at 12 million a machine or something) be
a pretty expensive thing.
The best combination to fight Kramnik would be a different engine as
deepfritz with the deepfritz kure book.
Note that deepfritz is a very well tested engine (one of the best
tested engines in the world) with a superb book and nowadays a strong
endgame and a huge search depth (like 17 plies or so).
hardware whether it's a 2 processor 1.5Ghz K7 or a 8 processor 700Mhz xeon,
that's not a big issue.
Search depth IMHO is never the most important thing when you get near
those search depths.
the chessknowledge from fritz/quest is in no respects bigger
as that what DB had, but it definitely is way better tested.
The openingsblunders of DB are well known (see icca june 1997, or
contact jaap v/d Herik or his secretary to get a copy of that
Seirawan article), but i don't expect fritz to
do better there. I mean it doesn't even know what a bad bishop is!
get a copy of the seirawan article and simply create a 100 or so
positions from the match where deep blue did weak moves.
Then run todays programs at it and you'll get amazed.
Like only 4 or 5 of those bad moves are getting done by todays chessprograms.
Posting those pos at your homepage would be the thing to do!
especially game 1 DB did some real pathetic things which none of todays
well tested programs is going to do wrong all at the sametime.
To compensate also put the ! moves on the homepae and you'll see that
from the strong moves made by DB, at least 95% of them are also getting
done by todays software, if not 100%.
>Subjective arguments (chess knowledge in particular) are also welcome, but
>should be concise as opposed to argumentative!
>Thanks, Mig
>Editor-in-chief
>http://www.kasparovchess.com
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.