Author: Vine Smith
Date: 04:04:12 04/18/02
Go up one level in this thread
On April 18, 2002 at 05:52:58, Uri Blass wrote: >On April 18, 2002 at 05:20:30, Vine Smith wrote: > >>On April 18, 2002 at 04:16:48, Uri Blass wrote: >> >>>On April 18, 2002 at 02:22:22, Slater Wold wrote: >>> >>>>On April 17, 2002 at 22:14:55, Vine Smith wrote: >>>> >>>>>Good Lord, your machine is fast, fast, fast! What is it -- I think I saw one of >>>>>your posts about this but can't quite remember, something like 2x1.7 GHz >Athlon? >>>> >>>>Dual AMD 1.73Ghz (Athlon XP+2100) - Asus A7M266-D Mobo - 1024MB RAM - 4x36.4GB >>>>SCSI - GeForce 4 >>>> >>>>>It seems roughly 6x as quick as PIII-850 with Junior 7 (assuming Deep Junior 7 >>>>>is much the same program), whereas I would have guessed only 4x. >>>> >>>>Deep Junior 7 is the same program. >>>> >>>>>The Shredder >>>>>analysis was strange -- I guess this is the only program that improves on >slower >>>>>hardware! Supporting this, there was a post by Nemeth that Shredder did not play >>>>>the awful moves that led to its dismal defeat by Smirin on his slower system. >>>> >>>>Well, I think there's more to it. If you look at position A with 2 computers, >>>>the evals will usually not be a *whole* lot different (granted you're using the >>>>SAME exact program). I've actually tested my machine's eval using Deep Fritz vs >>>>Deep Fritz on a P200. And the PVs are almost always exactly the same. However >>>>there is a LOT of randomness in a SMP search. Branching is almost completly >>>>random. >>>> >>>>>And then there's Crafty; every time I see it fail at a tactical problem, I gain >>>>>greater respect for its positional abilities, since this MUST be the way it >>>>>stays at the top (versus other non-professional programs, that is). >>>> >>>>Crafty is a terrific program. And I believe it has just as much positional >>>>understanding as any "commercial" program, perhaps even more than some. >>>>However, it does sometimes lack in tactics. >>>> >>>>>Rightly or >>>>>wrongly, DB rejected Deep Fritz's eventual choice of 36...b5 after reaching >>>>>depth 11(6) -- this had been DB's move at depth 10(6) [whatever that means; 16 >>>>>full-width? 10 full, 6 selective? 10 with a selective component plus 6 >>>>>full-width?]. >>>> >>>>Yes, I looked at the log. It looked at Rd7, b5, and eventually went with Kf8 >>>>after only looking at it for a short period of time. >>>> >>>>All of Deep Blue's searchs were full width. What I have come to understand is >>>>that 10(6) represents the (6) ply done in software and the 10 shows the ply in >>>>HW. The first 6 ply were always done in software, and the remainder was always >>>>done in HW. >>> >>>Not exactly and it was explained here that 10(6) means only 10 ply brute force. >>>6 plies in the hardware was based on selective search. >>> >>>Uri >> >>Was that a consensus opinion agreed to by all the parties who discussed this >>matter? I remember that Hyatt asserted the two depth numbers were additive to >>full-width, referring to a statement made by Hsu in a private e-mail. Then >>Diepeveen disputed this with some calculations intended to prove that even a >>machine as fast as DB could not have searched this deep full-width. The only >>thing I remember any agreement about was that the second depth number >>represented a search by the hardware nodes. Why the DB team never publicly >>explained what the log files they released actually meant is a mystery, as is >>the reason for keeping the code under wraps, since it seems that it will never >>be used again. The information vacuum keeps many aspects of various debates >>about DB alive, and there seems to be no rationale for this frustrating >>situation. >> >>Regards, >>Vine Smith > >I remember that somebody posted an old post from rgcc that proved that the >programmers of deeper blue did not see the last 6 plies as brute force search. > >Uri If the proof was solid, that just raises more questions. Why, for instance, would Hsu have told Hyatt the search was full-width? Perhaps DB was meant to search full-width in the hardware nodes, but wasn't working as intended. It seems plausible that something so complicated may not have been fully debugged before the match with Kasparov. Or perhaps terms like "full-width" and "brute force" have varying shades of meaning, depending on how implemented, just as the meaning of "search depth" under Junior differs from that of other programs. Most puzzling is the silence of the DB programming team about such issues. After all, the project had, at least initially, been portrayed as pure research with implications that the results would enhance computer science and thus be shared with the programming community. Of course, IBM was in no way obligated to disclose anything at all about Deep Blue, but they seemed to move from a very open approach to great secrecy at the end. Regards, Vine
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.