Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Deep Blue----not true

Author: Robert Hyatt

Date: 16:08:08 06/14/03

Go up one level in this thread


On June 14, 2003 at 10:17:58, Vincent Diepeveen wrote:

>On June 13, 2003 at 00:08:32, Mike Byrne wrote:
>
>>On June 12, 2003 at 22:44:25, Vincent Diepeveen wrote:
>>
>>>On June 11, 2003 at 21:06:15, Mike Byrne wrote:
>>>
>>>
>>>
>>>>On June 11, 2003 at 00:28:29, K. Burcham wrote:
>>>>
>>>>>
>>>>>we can say that with todays hardware, todays programs wont play
>>>>>36.axb5 axb5 37.Be4 in tournament time control. there are several programs that
>>>>>will play these three moves if given enough time, for example, Deep Sjeng 1.5 in
>>>>>8 hours.
>>>>>
>>>>
>>>>
>>>>after I posted that, I realized I was thinking "normal time controls" but if
>>>>somebody left their program running a very long time -- it might find it.  I
>>>
>>>Let's be clear. Deep Blue just searched 10 or 11 ply there.
>>>
>>>Because of hashtables we get that very quick nowadays. At 30 seconds a move
>>>you already outsearch deep blue in these positions.
>>>
>>>That's why when Be4 showed up at 11 ply at the time in DIEP and in Crafty and
>>>also in Deep Blue is no big of a surprise.
>>>
>>>We got simply deeper with more time because of hashtables.
>>>
>>>So comparing something that searched 10-11 ply with a simplistic evaluation to
>>>modern chessprograms is simply pathetic.
>>
>>Deep Blue is simplistic evaluation - that's not true.  Why do you say that?  Is
>>this something Dr. Hsu told you?  Think about what you're saying - Deep blue
>>searched about 200M NPS- that is still 40-50x faster than any top of line PC SMP
>
>there is no proof it searched 200M NPS. All we know is that their logfiles do
>not report the number of searches a second even. Why did they remove this number
>from the 'logfiles' ?
>
>Everyone prints out how many searches a second or nodes a second you get.
>
>We know that their 200M nps number just like their 'parallel' speedup is
>extrapolated from single node tests.
>
>That is a very bad thing to do.
>
>>machine today.  Are you taking into the considerations of Deep blue extensions -
>
>Just look at what Uri posted here. What we know is that Diep is doing more
>singular extensions than Deep Blue did. I might kick them out though. They do
>not work positional and in future a good branching factor might be more
>important.
>
>Also note that they FORWARD PRUNED using a dubious algorithm called 'No
>Progress' pruning.
>
>You can count trivially how many patterns they had. They are described in their
>publication. Around 100 or so, in that region. Most had an 'array' of 64 values
>to lookup. That's how you quickly get a few thousands of 'adjustable'
>parameters.
>
>Best proof is of course the idiotic play it showed. Just take game 1.
>
>Which program is playing h6? Qa5? Bc7? g5?
>
>And which program is missing the move e5 when you can play it directly in the
>opening?
>
>>Dr Hsu in his book regarding this position said that the main line was at least
>
>If i say my go program has a lot of knowledge you must keep that very relative.
>
>>24 ply deep and maybe over 40 ply deep (pg 229) (accounting was not one of DB
>>strengths).
>
>I get sometimes crap lines from 70 ply.
>
>>You are right ->comparing PC programs to Deep Blue is like comparing you to Bob
>>Hyatt -> simply pathetic.
>
>Deep Blue is outdated hardware with outdated search with outdated evaluation,
>which for the year 1997 was pretty ok. See how programs at the time played. Very
>poor compared to 2003.
>
>Hsu was a hardware programmer, not a software chess programmer. He had enough
>problems at his head. What tells everything is that they only 2 weeks ahead
>received the processors.
>
>No way of course to bugfix them. No way to test it.
>
>You should really go ask yourself about their evaluation. The first deep blue
>version admitted by even Hyatt is a piece square table program. No
>chessknowledge at all.

Please don't quote me unless you quote me _correctly_.

"chiptest" _might_ have been a "piece-square evaluation only".

Deep Thought was most certainly something beyond that.  And deep blue
and deep blue 2 were beyond deep thought.

Your comment makes no sense since chiptest only played in one event and it
did badly.  It was just a "test".  Deep Thought was pretty devastating against
all oppositions when it came out.



> Then you believe in the fairy tale that something that
>makes simple mistakes with doubled pawns, which gnuchess doesn't even make (it
>knows f2 g2 g3 pattern to be good), that such a program suddenly from a piece
>square table program by a programmer who cannot play chess himself, suddenly
>within a few revisions has more knowledge than other programs?
>
>That's nonsense.
>
>They progressed (1 ply deeper search with a partly leaf eval) that's all.
>
>Still you can proof that deep blue is for a big part a preprocessor. That's
>clearly proven here. Or do you deny that?
>
>Now in 2003 you cannot call any top program which plays at the world champs a
>real preprocessor. Someone that goes there basically with a preprocessor can go
>home directly.
>
>Even the worst tuned program with a lot of chessknowledge will show potence by
>good positional moves.
>
>Deep Blue didn't. It only showed the opposite. Some trivial things:
>
>  a) big doubled pawn mistakes (this is really basics)
>  b) no bad/good bishop code (this is really basic knowledge)
>  c) very simplistic 'gnuchess' mobility
>  d) not knowing that g5 is bad for your king safety (2 mistakes, both game 1
>     and also later when it castled long with white and played
>     b4 which is the same mistake)


Kasparov said "g5 was the _only_ move to play."

Your comment there is also therefore wrong...


>
>In short it makes like 5 to 6 mistakes a game. very very poor.
>
>I'm not seeing how you can say that this would do well in 2003. It would not
>even get within top15. Also not when Hsu is going to work fulltime at it.
>
>He simply doesn't know how and what.
>
>The biggest crap of course is saying that he got 200M nps. His logfiles do not
>show *anything*. His parallel speedup is based upon extrapolation. Really poor
>science.
>
>Everyone who has made a parallel program knows the problems of getting a speedup
>at a bad-latency cluster.
>
>His algorithms were that outdated that in the 1999 IEEE publication he writes
>that he didn't see any noticable difference between normal alfabeta and other
>forms (like PVS and such).
>
>That really is big big big crap and says enough about how 'advanced' deep blue
>was.
>
>It is IBM that has created it to a myth by paying Kasparov some coins to give a
>show, which he regrets probably by now.


Of course they _won_.  Nobody has _repeated_ that yet.  They might not
repeat it for a while, so deep blue was more than "primitive"...



>
>If you want to i can create 150 ply mainlines btw, no problem. It is 1 minute
>work. But it is utter nonsense.
>
>>>>think that is true with many positions - obviously I meant somewhere in 20
>>>>minutes or less (or maybe not so obviously) -- but I agree with your point
>>>>--although the heading is a little harsh since I think you knew what I meant...



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.