Author: Robert Hyatt
Date: 10:15:26 03/13/02
Go up one level in this thread
On March 13, 2002 at 12:31:54, Uri Blass wrote: >On March 13, 2002 at 12:13:49, Robert Hyatt wrote: > >>On March 13, 2002 at 11:41:42, Uri Blass wrote: >> >>>On March 13, 2002 at 10:16:56, Slater Wold wrote: >>> >>>>On March 13, 2002 at 07:26:08, Chris Carson wrote: >>>> >>>>>On March 13, 2002 at 04:09:54, Jerry Doby wrote: >>>>> >>>>>>It's hard to believe that anything can be that much strongeer then fritz7 on a >>>>>>fast platform. Is deepblue 100 elo or above deepfritz on an xp 2000 >>>>> >>>>>OK, I will bite and get a debate going most likely. First take a look at: >>>>>http://home.interact.se/~w100107/manmachine.htm >>>>> >>>>>Tony's page has the results for both Top programs today and Deep Blue. >>>>> >>>>>Here is a brief comparison: >>>>> >>>>>Deep Blue 97 2862 6 games >>>>>Chess Tiger 2788 11 games >>>>>Deep Junior 2702 9 games >>>>>Rebel Cen 2697 4 games >>>>>Deep Fritz 2678 12 games >>>>> >>>>>None of the Commercial programs are on fastest HW today. Deep Blue only played >>>>>6 games against one opponent that did not get to prepare (Rebel opponent played >>>>>100 games against Rebel before the match). My guess is that Deep Blue rating >>>>>would drop by 100 to 200 points if put to a serious test. The Commercial >>>>>programs would be 100 points stronger on fastest HW. So they are about the same >>>>>or slight favorite to the commercials. I think Rebel, Tiger on fastest single >>>>>processors and Deep F/J on fastest mps would beat DB 97 in a match. >>>>> >>>>>My conclusion is that 5 years after the match, the commercial programs rule. I >>>>>think that the gap was closed a couple of years ago. >>>> >>>>The thinking here just blows my mind. I cannot even begin to *imagine* why >>>>people would say something so silly. >>>> >>>>You're talking about a chess program, that used the _same_ exact search >>>>techniques that are used in 80% of the top engines today. While 5 years worth >>>>of research probably makes todays top commercial engines more "refined", but >>>>when it comes down to it, they are basically the same. >>>> >>>>With that said, now imagine your search is 100x faster. That has _GOT_ to be >>>>worth some ELO. 200M nps vs Fritz 7's 1M nps (on today's top HW) is hardly >>>>comparable. >>>> >>>>Just use the rule of HW speed. 2x the mhz is usually worth about 50 ELO. It >>>>wouldn't take much to get 250 ELO out of the speed of DB. >>> >>>You forget that programs got 200 elo only by software in the last years. >>>The best commercial program in 1997 is 200 elo weaker than the best program of >>>today in the same hardware. >>> >>>If you remember that there may be diminishing return at higher depthes then it >>>is not clear that the best programs of 1997 with 200M nodes per second are >>>better than the program of today with the hardware of today. >>> >>> >>>Another point is that I guess that deeper blue used some ideas that >>>are probably not good. >>> >>>Nobody use singular extensions in the way that deeper blue used them. >>>Ferret use them but not in the way that deeper blue used them. >>> >>>Crafty18.12 used the deep blue extension. >>>Crafty18.13 does not use it. >> >>This is incorrect. No published version of crafty has ever used singular >>extensions. > >The deep blue extension here is not singular extension but the extension >that was used by Crafty18.12 and I read in main.c of 18.13 that >it is not done in Crafty18.13. That was not an "extension". It was an "extension limit". Crafty has always had a limit of 1 ply of extensions for every ply of search. Deep Blue used a limit of "two plies of extensions for every two plies of search". Which is not the same thing. With Crafty, no ply could extend by more than one ply, while with the DB approach one ply could extend 2 plies if the next ply extends zero. I didn't like it although the difference was not very significant.. > >I meant to say that I believe that I guess that deeper blue did 2 mistakes >(one is singular extensions in the way that they used them and another one is >the deep blue extension that is used by Crafty18.12 but not by Crafty18.13). > >Uri I don't know whether their idea is a mistake or not. When you add SE, then their idea makes more sense. I used that approach always in Cray Blitz. But in the case of Crafty, I felt it wasted nodes here and there and took it out. Some thought it was better with the DB limit in, however, so it was not a clear win or lose situation... I am certain that SE is a good idea. I just haven't spent a lot of time trying to make it co-exist with the null-move R=2~3 search I now use. Cray Blitz had a much more restricted null-move search using R=1. And SE worked just fine there, as it did in HiTech, Deep Thought and others. The newer R=2 and R=2~3 and even R=3 for some, seemed to have a negative impact on the SE algorithm probably causing things to extend when they should not... This bears a lot of investigation, however, to see if it can work. It is working pretty well in Ferret it seems although his is much more restricted than what DB and Cray Blitz used....
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.