[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] question regarding Hydra Chess PC computer
At 09:05 25-10-2004 +0100, chrilly wrote:
>>Don Dailey wrote:
>>> I agree with you, but according to Bob Deep-Blue had none of these
>>> issues, he claimed there were *no* compromises, that even given the
>>> same numer of nodes Deep Blue would blow everything else away.
>>
>>This is a pretty recent paper from the Deep Blue team:
>>
>>http://sjeng.org/ftp/deepblue.pdf
>>
>>It's quite revealing about what they actually did in search &
>>evaluation, or perhaps I should say, did not do...
>>
>The discussions in some forums are sometimes fascinating and remind me on
>discussions in the middleages. Nobody cares about the facts or the sources
>(in the middleages they cared at least about the sources).
>In the paper above there is e.g. a table were it is clearly stated: "No
>Hashtables" in the hardware search.
>
>Vincent Diepeveen said: Feng Hsu and myself have no knowledge of chess.
>Thats true. But both teams have engaged exactly for this reason a GM. In
Note that i'm a titled player myself and very easily have access because of
that to other titled players.
I speak daily with 6 or 7 GM's which all of them are pro players.
You give them a position.
"What do you think of it?"
Answer: "yes it feels fine for black"
question: "based upon what facts"
Answer: "i cannot factual say it's sure but i really have this feeling it
is ok".
a) they are not capable of expressing what is the actual problem
b) hundreds of years of chess history have brung next material values:
pawn = 1.0
knight = 3.0
bishop = 3.0
rook = 5.0
queen = 9.0
However in chessprograms that's not the case. Within a few years you find
out that the real values already are closer to, here is what they are in
DIEP :
pawn = 1.0
knight = 3.625
bishop = 3.675
rook = 5.8
queen = 11.75
With a lot of other rules to evaluate specific material situations.
It's very clear that deep blue had less evaluation than Crafty has.
It's very clear that Fritz from all commercial programs has very little
knowledge. It's very clear that hydra doesn't even have 12% of the
knowledge shredder has and it's clear hydra doesn't even have 4% of the
knowledge DIEP has.
In the go world there is also a lot of contradictory pro advice.
Here another pro help i was about to implement in my go program when
someone very careful told me it might not be true. For sure in chess it's
the opposite of the truth.
"engage the enemy where he is strong" 9-dan player advice.
I'm sure in go and chess and any board playing game you'll see similar
things. professional players are the most horrible source to 'improve' ones
evaluation function. They do not understand how to make an evaluation
function and will never understand what a computer is and how it works.
Both Deep Blue and Hydra/Brutus share in that the knowledge in those
programs is adviced by GM's who are known for not knowing how to explain
themselves and ever saw a chessbook written by any of those GM's that gave
those 'advices' ?
>case of Hydra GM Lutz works only since a few months for the team. Before I
>did not have the money to engage one.
Come here and grab my keyboard, i have 3 on the chat right now.
By the way earlier this year i played reigning world champion FIDE
Kasimdzhanov.
Of course i'm a titled player myself, i don't bug them with questions they
cannot answer, such as: "why is this position good for white?"
Their only answer will be: "i feel it is good".
Another great world champion advice for your chess engine: "a knight and a
queen are strong". Capablanca said that. Later repeated by another 100 GM's
in a big choir.
Hah, what a laymen.
Diep - Jonny world champs 2003 : 1-0
Jonny grabbed queen+knight, diep grabbed queen+bishop.
Those 512 processors of diep really didn't help beating it. Just the fact
that he had written down in the chessprogram Jonny the capablanca rule let
it commit suicide.
Let's face it. Your program will be 50 kyu strength when following such
advices.
The difference between you and me is that as a titled player i'm able to
throw out the 99% nonsense i hear there and i implement the 1% valid
knowledge. You on the other hand have no choice but to believe someone who
still must invent that a queen isn't 9.0, and who you will NEVER EVER be
able to convince that a queen is stronger than 2 rooks IN GENERAL and only
in some EXCEPTIONAL positions the 2 rooks are stronger.
Oh in case of chess i have another great quote for you. World champion Max
Euwe: "a king in the endgame is as much worth as a rook".
Make your king worth a rook and see how good your endgame play of your
engine is :)
Another quote from a top player alive, Jan Timman: "a knight pair is in
many cases stronger than a bishop pair".
In fact Jan Timman wrote a 200+ page book about it, just about how strong a
knight pair is.
Every day i used up a page of that book in the bathroom. After a year the
book ceased to exist.
>But that was not Dons question. He asked: Is it principially possilble to
>write a better evaluator in hardware than in software. It is possible. One
You don't even have enough transistor budget to put in hardware *current*
diep's evaluation.
>can of course always implement the same in software. But the C-simulator is
>by a factor of 1000 slower than the hardware. One could speed this up to a
>factor of 100. But it would nevertheless be a very inefficient software
>evaluation.
>
>There is also a big difference between having some application-specific
>knowledge and implementing one. Hans Berliner, the head of the Hitech
>project, was world-champion in correspondence chess. Feng Hsu had problems
>with the algabraic notation of the board. But there is no doubt that Hsu
>made the stronger chess programm. At least Deep-Blue II had also certainly
>more chess-knowledge than Hitech ever had.
Deep Blue was from a time that chessprograms played at a level that they
made simple tactical mistakes, in go equivalent by giving away a huge group
in 3 moves (6 ply).
Note that Deep Blue never was world champion. It made huge tactical
blunders against a program called Fritz in world champs 1995. Fritz won the
title.
I'm sure DIEP can beat a 460 processor deep blue even when playing with
diep on a wristwatch processor.
Simply because deep blue was positional a passive program.
Hydra isn't, in that respect hydra is 400 points better than deep blue.
A friend of mine works in VHDL hardware language and i asked him how a
simple chess pattern would look like there.
First shock i got is that every pattern needs to get programmed in hardware
360 times, in case of a GO program.
So your overhead for being 'faster' is roughly factor 360 for a go program
for a simple pattern. Including mirroring perhaps even a 1000 times.
A hardware go program with 5000 patterns will require you to type by HAND
for 50000 days non stop.
And dear oh dear when you make a typing mistake when mirrorring the right
square number by hand.
Your GM sure won't notice that bug.
He can't read hardware language.
Diep is 2.2 MB C source code now. That's just the engine. Not interfaces.
My text i/o interface in engine is :
14-10-2004 22:44 21.948 diepasci.c
It's clear your engines source should be at least a factor 16-64 bigger in
hardware source code, because
a) you need to write things down for black & white, where i just write it
usual down in a generic way
b) you need to write it down for either every square where the pattern
can happen, every row or every rank.
So a minimum is a factor 16 bigger source code to have the same size chess
knowledge.
I doubt you manage if you program for another 30 years Hydra that your
program will even have 20% of the chessknowledge diep has *now*.
No doubts that your chessknowledge is better tuned than mine.
Automatic tuning is not possible with more than a few parameters.
Deep Blue basically had 30 chesspatterns.
Of course every pattern was indexed by array, such as piece square tables.
The piece square tables for deep blue delivered directly 768 'adjustable
parameters'. If you count in that way in hardware, you look very great very
soon of course.
I'm sure hydra is better than that, but basically by some agressive
mobility and all parameters well HAND tuned by a professional chess
programmer and a very modern search.
Making knowledge in hardware is NOT so easy IMHO.
>Chrilly
>
>
>
>
>
>>--
>>GCP
>>_______________________________________________
>>computer-go mailing list
>>computer-go@xxxxxxxxxxxxxxxxx
>>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/