Author: Robert Hyatt
Date: 18:36:49 07/22/00
Go up one level in this thread
On July 22, 2000 at 19:55:29, blass uri wrote: >On July 22, 2000 at 19:18:23, Robert Hyatt wrote: > >>On July 22, 2000 at 18:53:39, blass uri wrote: >> >>>On July 22, 2000 at 17:38:49, Robert Hyatt wrote: >>> >>>>On July 22, 2000 at 10:20:58, Chris Carson wrote: >>>> >>>>>On July 22, 2000 at 07:46:00, Ed Schröder wrote: >>>>> >>>>>>On July 20, 2000 at 21:50:06, Robert Hyatt wrote: >>>>>> >>>>>>>On July 20, 2000 at 15:47:53, Ed Schröder wrote: >>>>>>> >>>>>>>>On July 20, 2000 at 14:50:48, Robert Hyatt wrote: >>>>>>>> >>>>>>>>>On July 20, 2000 at 14:26:16, Chris Carson wrote: >>>>>>>>> >>>>>>>>>>On July 20, 2000 at 13:03:42, Robert Hyatt wrote: >>>>>>>>>> >>>>>>>>>>>2 years ago a single DB chip played several matches with top commercial >>>>>>>>>>>programs. This DB chip was running at 1/10th of its normal speed, and yet >>>>>>>>>>>it won 36 out of 40 games. This has been reported several times here on CCC, >>>>>>>>>>>by several that have heard Hsu and Campbell give talks about the DB hardware. >>>>>>>>>>> >>>>>>>>>>>If it could win 90% of the games running at 1/10th the normal speed for one >>>>>>>>>>>chip, what does 480 chips at full speed get (hint: 4,800 times faster). Would >>>>>>>>>>>you think it might have a pretty easy time with today's programs? quads or >>>>>>>>>>>8-way boxes as you want? >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>>Games against P90 machines. What were the program versions? what were >>>>>>>>>>the settings? Who was the operator? Which book was used? Where are >>>>>>>>>>the games for inspection? Did the DB team get permission to perform >>>>>>>>>>this tournament and permission to report results? Were you there? >>>>>>>>> >>>>>>>>>I was there. The program authors were there. That was a requirement for the >>>>>>>>>ACM events. We all sat across the board from each other. Marty. Ed. Richard. >>>>>>>>>Hsu. Myself. anybody else you would care to name... >>>>>>>> >>>>>>>>Nah.... you are talking on late 80th's Rebel running on a 5 Mhz 6502 >>>>>>>>processor with 32 Kb Ram Rebel doing just 500 NPS. Your point again? >>>>>>>> >>>>>>>>Ed >>>>>>>> >>>>>>> >>>>>>> >>>>>>>I like the math game... >>>>>>> >>>>>>>so lets see.. you were doing 500NPS in 1988? And today you are doing maybe >>>>>>>500K? A factor of 1,000? >>>>>>> >>>>>>>They were doing 300K in 1988. They hit 1B in 1997. 1,000,000,000 divided by >>>>>>>300,000 is how many times faster? 3,333 X faster you say? >>>>>>> >>>>>>>Now do you get my point? They have widened the NPS gap by a factor of 3.5 >>>>>>>since 1988. >>>>>>> >>>>>>>_that_ is my point. The gap has continually _widened_. _not_ _narrowed_... >>>>>> >>>>>>Wrong math. >>>>>> >>>>>>From the IBM site (1988) >>>>>> >>>>>> "Deep Thought 0.01 becomes >>>>>> Deep Thought 0.02 and >>>>>> improves to 720,000 chess >>>>>> positions per second. The new >>>>>> program includes two >>>>>> customized VLSI chess >>>>>> processors. >>>>>> >>>>>>So 720,000 and not 300,000 >>>>>> >>>>>>Your second mistake is the 1B. Where did you read that? You know very >>>>>>well 200M is claimed, no quoting needed. >>>>>> >>>>>>So 200M / 720K = not even 300 times faster. >>>>>> >>>>>>As you say Rebel improved with a factor of 1000 (3½ times more than DB) >>>>>>nota bene the exact opposite you claim. >>>>>> >>>>>>Not that such math impresses me (as if computer chess is about hardware >>>>>>only) but I do you like you the fact you have an argument less. >>>>>> >>>>>>Ed >>>>> >>>>>Ed, >>>>> >>>>>Besides the impressive speed up that you point out, I would also >>>>>point out that Rebel improved at least as much from the SW side. >>>>> >>>>>You get a lot out of the HW you target for. :) >>>>> >>>>>Best Regards, >>>>>Chris Carson >>>> >>>> >>>>And the DT/DB program didn't improve? Have you read the JICCA paper on their >>>>search extensions? It is probably the _standard_ reference for search extensions >>>>and null-move testing. What says that their software didn't improve? a desire >>>>to paint them in the worst possible light??? >>> >>>I have no evidence that their software was improved. >>>They are not commercial so we cannot test it. >>>Their papers prove nothing and the only proof is in games. >>> >>>The fact that they did bad result in 1995 and did not play public games after it >>>to prove that they were unlucky is an evidence that their software did not >>>improve. >>> >> >> >> >>You keep saying "they did a bad result in 1995" when all they did was lose >>one game. Based on that criteria, _everybody_ sucks with two straws. I can >>look at the last 10 years of WMCCC (or WCCC) events and find a tournament >>where any program you chose lost at least one game, and probably more. Does >>that mean they did "bad"? >> >>It is interesting that the measuring stick for "good/bad" is way different >>for DB and the rest of the programs. Everybody else can lose 1-2 games in >>a single tournament and be happy. Shredder lost in the last WMCCC event >>yet it won the tournament. I assume that was a "bad" result? No? Then >>DB lost one game in 1995 and that was a bad result? Yes? > >DB lost one game and drew another game. >It is a bad result because people expected it to win because of their big >hardware advnatage. > So? Crafty plays "Hossa" a lot of games on ICC, because the programmer likes to use crafty for testing. I have a _huge_ hardware advantage, yet I still lose an occasional game or draw an occasional game with him. I don't consider that "bad". I consider that "statistical pay-back". Even rated 500 points above someone gives them a definite expectancy to win a game here and there. >If you got 90% in the past then less than 90% is a bad result so everything less >than 4.5 out of 5 for Deep blue was a bad result. That is statistically nonsense. If DB plays 5 rounds with a .9 probability of winning each game, it has a .9 ^ 5 probability of winning all 5 games. That turns out to be less than 60% chance. IE almost like flipping a coin. Heads, they win all 5 games. Tails they lose 1. So losing one there seems reasonable, as they had several events where they lost zero. > >Shredder did a good result in WCCC because people did not expect it to >win(Junior and Fritz and Ferret had hardware advantage and shredder had no >history of winning). But it did lose a game or two... > >> >>I guess I don't get it then... I'd love to win 100% of the games. That is >>not realistic. Which means I guess I am destined to have bad results all >>the time. >> >> >> >>>If they could get more than 90% against Fritz3(p90) they had no reason to avoid >>>proving it to the world before the match with kasparov. >>> >>>Uri >> >>Sure they did. They were working like mad to get the thing ready. Hsu's book >>is eye-opening on how close they cut things. They were even considering the >>possibility that they would have to postpone the first match, because the IC >>FAB shop screwed up badly on the first few batches of chips and none worked. >>They didn't have time to do _anything_ but get ready for Kasparov, and they >>_barely_ made it. Same for 1997... same sort of fab problems with the new >>chip, time-consuming test procedures for each and every chip and the DB team >>did _all_ the testing, etc... > >If this is the case then when did they have time to get more than 90% against >Fritz3? They had a chip lashed up to the test rig they used. I imagine that they are just like I am, curious as to how something totally new would work. Seems plausible. I played many games vs commercial programs with Cray Blitz to test how changes looked before a big event... I didn't report the results although I might occasionally mention them in casual comments as they did. I didn't consider it any big news when I blasted the (at the time) best micro around using a fraction of a Cray. As a fraction of a Cray was still a big hunk of computing. > >I think that they could do better by postponing the match and playing a match >against Fritz3(p90) first. > >Uri Right. With the marketing guys going crazy? With the potential advertising bonus with the interest in the match? That might be your choice, but it would get you fired at IBM...
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.