Author: Vincent Diepeveen
Date: 08:08:37 08/18/02
Go up one level in this thread
On August 18, 2002 at 09:06:02, Jorge Pichard wrote:
No for many reasons
a) deep blue is old and how many old programs are still considered
'good'?
Even programs that reigned the world for 10 years (genius) completely
get butchered with x-0 scores nowadays.
b) the quality of the games has increased, provable.
It is very provable that nowadays software is making less mistakes than
from around 1997. In the match kasparov-deep blue II we can clearly see
that deep blue made major pawn structure blunders. Kasparov managed to
waive this criticism against deep blue by murmuring something stupid about
'human help'. The few analysis of chessplayers regarding deep blue - kasparov
already says enough. It's bad analyzed, simply because the games were
so bad.
Even the random openings moves from Kasparov didn't hide the fact that
Deep Blue by todays standards would be definitely get wiped out completely.
Openingsbook from deep blue was 4000 moves made by hand. After that an
automatic book. In 1999 the computerchess insiders realized very well
that this was getting outdated. Nowadays all commercial engines have
hundreds of thousands of hand in given moves.
Of course those books are secret. Only at world championships and in
tournaments like CSVN champs, Paderborn (IPCCC) tournaments they get used.
Those books are *not* getting used at the internet, i repeat *not* getting
used there. The level difference between a program using such a book and
no book is about 600 points for its weakest link.
The endgame has improved tremendeously of software. Where deep blue can
be considered as one of the better endgame programs in 1997, by 2002
standards the endgames are *real good* from engines. Only titled players
can now profit sometimes from some mistakes here.
I remember how i could beat any engine in any endgame around 1997 without
any problem. Note that i wasn't fide master at the time.
In 2002 i must watch my steps in endgame carefully. Only in rook endgames
i make a good chance against any engine still, though some have improved
there also tremendeously.
the only weak chain is still openings play of engines. However as i said
a major handtuned book fixes that. The openingsbook of deep blue was just
4000 moves.
If modern main lines would get played against the thing, it would be under
the table within 25 moves. Under the table means: "resignable from titled
players standards". Don't take me wrong. Most computer chess 'guru's they
play on way too long. Like until they see -mate in 5 at their screen.
Of course with a piece less a game can take pretty long still, but who
wins it is out of the question :)
Regarding the math. That is hard to do. A node from deep blue is
not comparable with a very efficient software node.
Deep blue was hardware. In a time that the average program had a branching
factor of 10.0 it was awesome. But now the branching factors are much
better. Deep Blue searched about 11-12 ply fullwidth. That was very good
in 1997. I searched 8-9 ply at world championships 1997 (no forward pruning).
Sincethen hashtable management and efficiency of nullmove have improved
the branching factor a lot. Don't take me wrong. Nullmove is used pretty
long already, like Bob can explain to you. In 1997 it was considered dubious,
but a combination of better hashtable management and nullmove and a well
tuned evaluation proves to be a good combination.
Deep Junior at 20 ply (junior units) will obviously see most lines deeper
than deep blue. If you simply do a head to head compare you will see that
deep blue was simply previous generation. In hardware there is no killermoves
there is no hashtables. I don't say it is impossible to make in hardware.
In contradiction on paper it is easy. But reality is that it didn't use it.
Just like Brutus doesn't have either. If you have a 20Mhz hardware clock,
brutus by the way has 32 or something, then wasting a full clock at it is
just too much (killermoves). It means that instead of 8 hardware clocks
which at 20Mhz mean 20 / 8 = 2.5 million nodes a second, you need 10
hardware clocks which means 20Mhz / 10 = 2 million nodes a second.
So the nodes very very primitif from efficiency viewpoint. Despite that
getting 11-12 ply fullwidth with so many extensions was a very good
achievement from Hsu.
However lacking hashtable and nullmove simply makes that NPS a joke.
If i remember well last thing what bob claims was a 10% efficiency at
the 126MLN nodes a second deep blue got. Still getting 11-12 ply is
very good then with 12.6MLN nodes a second effectively.
Remember, we talk about software now that also gets in the million nodes
a second.
However i do not think it is search depth what is the biggest difference.
It is obviously a better evaluation. The pawn structure mistakes of deep
blue are just too pathetic to take serious in 2002.
Best regards,
Vincent
> Kasparov proved that he can defeat programs at fast time controls when he
>defeated Deep Thought in a game/90 two games match in 1989. This program was
>weaker than Deep Junior is today, as it searched well over 2,000,000 NPS, but
>didn't have as much chess knowledge as Deep Junior. He also defeated Deep Blue
>in 1996. This program is obviously much faster than Deep Junior is today, but in
>my opinion Deep Junior still has more chess knowledge than Deep Blue had back in
>1996.
>
>PS: It is hard to compare Deep Blue of 1997 vs Deep Junior of today, but in my
>opinion Deep Junior Chess Knowledge could make up for the difference of Deep
>Blue super calculating power of 1997.
>
>Pichard.
This page took 0.01 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.