Author: Vincent Diepeveen
Date: 07:13:58 07/14/03
Go up one level in this thread
On July 13, 2003 at 21:18:23, Dave Gomboc wrote: >On July 10, 2003 at 14:46:16, Vincent Diepeveen wrote: > >>On July 10, 2003 at 13:13:00, Uri Blass wrote: >> >>>On July 10, 2003 at 12:40:38, Russell Reagan wrote: >>> >>>>On July 10, 2003 at 01:53:26, Derek Paquette wrote: >>>> >>>>>Just read a great post by the author of DIEP, and how he is getting an >>>>>Incredible machine, a godly machine I should say, now really, >>>>>Would a human even have a chance in hell? its going to be dozens of times faster >>>>>than Deep Blue >>>>>I know that speed isn't everything, but when you are looking 45ply ahead.... >>>>>I put all bets on the machine with 500 processors >>>>> >>>>>what do the rest of you think? >>>> >>>>I'd guess it will still lose to the top commercials, just like it did when he >>>>had a 1024 processor machine at the last WCCC. Since 1024 didn't work, what >>>>makes you think only 500 will be better? >>> >>>He had not a 1024 processor machine at the last WCCC. >>>I think that he tried to make Diep using them in a productive way but failed. >> >>I had 60 processors of a broken partition that was regurarly getting rebooted >>because of maintenance. In combination with a preparation time of 3 days it >>wasn't very good performing at it :) >> >>The machine indeed has 1024 processors in total. Biggest partition addressable >>is 512 from which you can use 500 processor maximum. >> >>>I think that less processors is an easier task. >>>if 1024 was too hard task for him then maybe 500 is going to be an easier task. >> >>>The first question is how much speed is he going to get from the 500 processors. >>>I do not predict nothing about it. >>>Uri >> >>Even a small speedup times 500 processors still is 10 times faster than any PC. >> >>Note that in 2002 i didn't lose from any commercial program except junior which >>was lost in a silly way (i had put the day before in order to test quicker the >>EGTBs to just 1 MB cache this at a very slow old harddisk; supercomputer i/o was >>broken at that time as it was getting upgraded and junior team had made it me >>impossible to use internet) as it just got 6 ply when i tried to prevent >>forfeiting and had put it to 1 minute rest of the game in the last 5 minutes i >>had left for the game. >> >>Then it played instantly a move with 6 ply search somewhere move 79 or something >>and that was losing move. Many others would have been simple repetition. >> >>DIEP nearly won from Fritz, Shredder and others in 2002. 3 games i played at >>supercomputer the others i had to play at dual or simply crashed at the >>supercomputer. Something that didn't help me either was what was going on at the >>big partition. There was some big program running at the supercomputer which >>eated all bandwidth away; it was using like 300 processors or so. In >>contradiction to most programs that are all running within L2 cache at each >>processor this software had allocated about 200GB memory. So it was eating from >>my 60 processors everything away too. Result was horrible latencies in a program >>not designed for NUMA. >> >>The combination of all that was disaster. >> >>In 2003 however i'll be running 500 processors and will have the partition for >>myself AFAIK. So no problems with other users at that partition. >> >>Then diep will be better tested for 2003 so it is impossible to compare the 2002 >>situation with 2003. >> >>Many try here it is not very smart to do so. >> >>Trivially others will be prepared very well too, like brutus and junior. >> >>Shredder perhaps will be unlucky and running perhaps at most at 2 processors. >> >>If we compare however then a lot of weak chains of DIEP will be a lot stronger >>in 2003 and one of its weakest chains in 2002 which was search depth, will be a >>lot different. >> >>Best regards, >>Vincent > >So how do you think Diep will place this year? >Dave With a very bad tested version it was dangerous in 2002. Even 2 rounds before the end had it won from Shredder (and remember it had a pawn up there for a while until shredder managed to save the endgame) it theoretically could still get world champ. However junior won both games and tied with shredder. For me i just look to what needs to be improved compared to worldchamps 2002. Second game against Lambchop it got itself a won position, but then to my horror i found out during the game that diep played without extensions there. it was a recompile of a supercomputer version which played at the pc. it really was intended for the supercomputer. So it also was 2 times slower. First thing that worried me was that it played without extensions. Because of that, diep in a won position made a tactical mistake. Some Rook move. Rd1 if i remember well. All other moves that do not tactical fail win there so to speak. But Rd1 loses. Then diep was lucky to get a draw there that game. Against junior i had a diep version with extensions turned on, but again for 128 processors. It played on 2 however so at a pc with just 256KB L2 cache instead of the 8MB L2 cache that the R14000 processors have at TERAs, that is horribly slow. It was a good recompile of a new version but a lot slower than it could have been. also its search was not optimal. For its time it was an ok version. Junior plays better in the middlegame and gets a won position. then near the endgame it makes some mistakes and diep can fight back sacraficing a queen which it didn't need to take. Lucky junior went into these lines and we get a drawn position. What loses for diep then is the fact that i was such an idiot to have the EGTB cache put at 1MB with a very very slow IDE disk (from a few years ago). Also the extra caching into the hashtable was turned off, because at the supercomputer the EGTBs are in the RAM, so no need to cache it in the hashtable. So when i put back diep to 1 minute whole game when having left 5 minutes on the clock at move 79 or something it plays horrible g3?? move and loses because of the 6 ply search causing this move. AMAI! Then diep plays a few very good games. Game against SOS is great. Game against isichess gets a draw quickly. Draws simply happens. Diep has a good game against Fritz initially. Then some stupid move gets played by a bug in its evaluation. These bugs are inside the diep evaluation only because of bad testing. the move f6?? really is horrible. Lucky fritz blunders again and diep gets a won rook endgame soon. DIEP pawn up. somehow fritz manages to draw that endgame. It wasn't easy to win, but it was a won rook endgame. Anyway, a draw is nothing to cry for. A deeper search would have fixed the 10 ply move that caused f6?? As i always said in the past. 12 ply is much better than 10. Above that you don't notice much. But this evaluation tuning mistake in half open file code was simply causing a bad move. That you do not win a won rook endgame then is not important. Computers will simply not manage to win them. Humans are superior there. Still i feel diep with a deeper search would have disabled fritz only advantage. Namely that it was outsearching diep by a ply or 4. We then get soon to the game against shredder. i was so stupid to force it to move 1.e4. this was my mistake. not of my openingsboo creator. let's be clear. perhaps diep would have won the game had it played 1.d4. We will never know. What we do know is that diep got a lost position against shredder thanks to being quickly out of book and a horrible nxc6?? move. Amazingly within 3 moves shredder blunders back and diep gets slowly a better position. good play by diep then gets a won position. When diep is won, shredder plays very good. Really very impressive defenses by Shredder. It clearly understood a thing better there and diep is missing a winning opportunity that shredder showed in its mainlines. Diep choses for a bishop pair and loses the pawn that it was up. So it is a drawn position then. Another good possibility to win a game against a world champion lost. But just marginally. The decision difference between the winning move and the drawing move there was like 0.02 in evaluation or so. Again a thing that might have been out of diep by good testing. A bigger depth also would have shown it. Not sure how many ply. But more than 2. Basically a bit bigger depth is needed when the evaluation is like 0.02 in evaluation wrong, it gets more lucky. Don't remember that diep was searching 9-11 ply at the world champs. The crucial moves that got itself a lost position against junior was 9 ply. The crucial moves that got itself from a lost position a won one against shredder was like 11 ply. Then i would need to lookup the mistake of diep. but from head i remember 10 ply or so. Then the only disaster happens to diep last round. it plays 1.e4 and simply doesn't get into the game against brutus. Good game from brutus. major mistake to play sicilian with white against kure of course. very dumb. any other move would have been better there. In short, the difference at world champs 2002 was not so big. A few small things. Even losing from brutus was no problem had it not blundered against fritz, warp (tactical blunder!), junior and won that game vs shredder. Enough to be world champ then instead of divided 5th together with lambchop. A 500 processor machine makes a big difference then. That the version will be hundreds of points better is trivial (some points of diep were weak, they won't be in world champs 2003; some competitors will be better too like Shredder). Basically i just fear Shredder. The brutus i do not fear at all. It is at fpga hardware and will be massively parallel probably, but if someone who has proven to know zero from parallel search in the past, now suddenly shows up in parallel i will be amazed if he gets actually dangerous because of that. Brutus will however be a program to beat in this sense that it will be good in winning from what we call 'amateurs' and the commercial programs that just search deep without much knowledge. Fritz doesn't need a victory as it already has played kasparov the first days of the month, so it will lose from Brutus of course. For their business that's the right decision to take and logical. So brutus is dangerous because it is going to score a big number of points in its home competition. Shredder is simply very good. And the junior team is always mercilous getting somehow points, but i don't fear them. You can't win world titles continuesly by just being mercilous and having a well debugged engine. The only unsure thing always in world champs is the role of the openingsbooks. In that sense Shredder is the program to beat there. Best regards, Vincent
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.