Author: Robert Hyatt
Date: 06:57:33 01/27/04
Go up one level in this thread
On January 27, 2004 at 03:37:23, Drexel,Michael wrote: >On January 26, 2004 at 22:09:22, Robert Hyatt wrote: > >>On January 26, 2004 at 17:00:07, Drexel,Michael wrote: >> >>>On January 26, 2004 at 11:49:28, Robert Hyatt wrote: >>> >>>>On January 26, 2004 at 11:24:58, Drexel,Michael wrote: >>>> >>>>>On January 26, 2004 at 09:33:41, Robert Hyatt wrote: >>>>> >>>>>>On January 26, 2004 at 02:14:39, ALI MIRAFZALI wrote: >>>>>> >>>>>>>On January 25, 2004 at 21:38:27, Robert Hyatt wrote: >>>>>>> >>>>>>>>On January 25, 2004 at 20:04:16, Rolf Tueschen wrote: >>>>>>>> >>>>>>>>>- - in a famous German forum the kids are on the streets and they shout: >>>>>>>>> >>>>>>>>>These old-fashioned Cray Blitz and Deep Blue monuments won't be "disqualified" >>>>>>>>>by their authors with actualized Elo numbers. >>>>>>>>> >>>>>>>>>Is that true? Would these legends lose badly against today's elite of >>>>>>>>>computerchess programs? >>>>>>>>> >>>>>>>>>I'm waiting! >>>>>>>>> >>>>>>>>>Rolf >>>>>>>> >>>>>>>> >>>>>>>>I don't believe _any_ of them would "lose badly". Any "super-program" from deep >>>>>>>>thought through Cray Blitz would be very tough opponents for today's programs. >>>>>>>>However, hardware is beginning to catch up. Someone just pointed out on a chess >>>>>>>>server last night that this quad opteron system I have is about the same speed >>>>>>>>as the Cray T90 I ran on in 1995, in terms of raw nodes per second (6-7M back >>>>>>>>then, 7-8M typically on the quad opteron). So it is now probable that Crafty >>>>>>>>could actually win a match from Cray Blitz on a T90 with 32 CPUs, assuming I use >>>>>>>>the quad opteron. My quad xeon 700 got ripped by the same machine a couple of >>>>>>>>years back, however, so it would still be dangerous. >>>>>>>> >>>>>>>>I can't say much about how it would compare to other commercial programs as I >>>>>>>>didn't run those tests with very little test time to play with the T90. >>>>>>>> >>>>>>>>The superiority of today's programs over the super-computers of 1995 are mainly >>>>>>>>mythical, IMHO. I suspect the games would be a _lot_ more interesting than some >>>>>>>>would believe. Of course, there is little chance to test such a hypothesis >>>>>>>>since most old programs are long-retired, and such hardware is not readily >>>>>>>>available today. >>>>>>>I disagree.DeepBlue would get slaughtered ;by todays top commercial programs. >>>>>> >>>>>>Fine. That is an opinion I don't agree with. But since there is no way to test >>>>>>the hypothesis, it is not worth the long argument. >>>>>> >>>>>>>It is known that standards in the midninties were not very high compared to >>>>>>>today.I think you over estimate Nodes per second for some reason.For instance >>>>>>>chess Tiger on Palm has a respectable SSDF rating of 2101 searching about >>>>>>>only 200 positions per second on the palm.A decade ago at such low NPS it was >>>>>>>inconceivable to get such rating. >>>>>> >>>>>> >>>>>>You _do_ know that deep thought, _not_ deep blue, but deep thought produced >>>>>>a 2650+ performance over 25 consecutive 40 moves in 2 hours games against GM >>>>>>competition to win the last second Fredkin prize? >>>>> >>>>>It was pretty weak nevertheless. It was not at all near 2650 level. >>>>>Many of those GMs didn't know anything about its significant weaknesses. >>>>>Deep Thought II played 1991 in Hannover against several German IMS and GMs. >>>>>It scored 3.5/7. >>>>>That was not even close to a 2650 performance. >>>> >>>>And your point would be? I have played games with versions of my program >>>>that were simply terrible, due to bugs and so forth. They had their share. >>>> >>>>But the Fredkin prize left little doubt, IMHO. 25 _consecutive_ games played >>>>over a year+, counting only games against GM players, games that were at least >>>>40 moves in two hours + 20 moves in one hour. Whatever happened in various >>>>things they did, you can _not_ play 25 consecutive games against varied GM >>>>players and pull off a 2650 without being _good_. Or do you really believe that >>>>a weak human player could do that also? I don't... >>> >>>Assume you could play a large number of games against strong human opponents >>>with >>>1. Crafty 19.09 on the Quad Opteron >>>2. Chess Genius 1.5 on a Pocket PC (400 Mhz) >>> >>>The humans don't get any information about program and hardware. >>> >>>I would expect both programs to score well. Maybe Craftys performance would be >>>50 ELO better. >> >>I would expect it to be much worse, myself. GM players are not to be laughed at >>for their tactics. I believe that then can quickly "feel" how well tactically >>the program is playing and adjust on-the-fly if they feel it is not very >>tactically aware... > >In a single game? Impossible >This might be true if they would play quite a few games in a row against the >same program. You might say impossible, but I have seen it happen. One case in point was a couple of years ago, I had a bug that on rare occasions would leave a Crafty process running after a game, even though a new instance of Crafty was fired off by xboard. 8 threads does not work well with the assumptions I have made in my parallel search, and crafty was playing very weakly due to some 3-4-5 ply searches. I watched a GM start a game, and by move 25 he was playing _way_ more aggressively than normal and he won, as he did in the next couple of games. I looked to see what was wrong, found the extra crafty running and killed it. In the next game he eased up on the wild stuff once the game was underway. I view it as trying to push an unknown object around. You can quickly tell how hard it is pushing back and adjust your strategy... > >> >>> >>>If you match both computers Chess Genius would lose badly of course. >>> >>>Performances against humans are a weak argument. >>>The initial posting was about computer-computer games. That is something >>>entirely different. >> >>One thing is pretty sure, a program that plays pretty equally with another >>program when they are pitted against GM players is _not_ going to smash the >>other program. It might _beat_ it, but not _smash_ it. Particuarly when we >>also have lots of data about how well DT did against computers up through 1994. >> > >Assume you would match an old crafty version on the 200 Mhz machine with Crafty >19.10 on the Quad. >Now that would be a pretty onesided match. I would call this a smash. >But Crafty scored already well against GM opponents back in 1996. Not the same thing. Current Crafty would score _significantly_ better against the same pool of GM players. And it would score significantly better against the old crafty/hardware. That was my point. I can't imagine a circumstance where two programs play humans and produce similar results but play each other and the result is way lop-sided. One could certainly be more anti-human, but it is not going to get killed by the other program if they both play GM players at a similar result level. > >> >> >> >>> >>>> >>>> >>>>> >>>>>[D] 1r1q1rk1/pp1b1ppp/3p4/2pBpP2/P2nP3/2NP2P1/1PP2R1P/R1Q3K1 b - - 0 15 >>>>> >>>>>Any strong program that does not avoid Bc6 in some ms? >>>> >>>> >>>>Again, the point would be? Pick _any_ program and look at 10 of its games. And >>>>pose the question "any program that would avoid this terrible move?" And you >>>>will get a dozen answers, including answers from programs that have no chance at >>>>all of beating the program in question... >>> >>>Well, King safety is very important. The most important factor in chess of >>>course. >>>Deep Thought would IMHO lose badly to Shredder 8 or Deep Junior 8 on a mere 2Ghz >>>PC. >> >>:) >> >>Remember that Hsu talked about the DB king safety advantage over the commercial >>programs of 1996/1997? He _specifically_ mentioned that their king safety was >>not up to the standard of DB and that led to the really bad results that all >>programs were reported to have against them in their lab, against a crippled >>processor version of DB, including yours truly... >> > >Have you ever seen one of those Deep Junior games where it spotted a decisive >King vulnerability 5 full moves before the opponent saw the 0.00 score (and >after another 3 full moves a losing score)? Yes, and I have seen games where it didn't see what was coming until it was too late, as well. It's a fine line for balancing something and it can fall either way... I can recall at least one game within the last year against Crafty, where Crafty was losing at around -3.5, but Junior (some copy on ICC) played a move and my score rose to 0.00 very quickly. And later Junior's dropped to 0.00 as it was definitely a deep repetition draw... That happens... both ways... > >> >> >> >> >>> >>>Deep Blue might be the sole exception. >>> >>>From a humans point of view todays top programs are strategical clearly >>>stronger. >>>Should be more important than Deep Blues possible tactical superiority. >>> >>>Michael >>> >> >>Except no one has said Deep Blue was strategically inferior either. The second >>game was uniformly praised (game 2 match 2) as the best strategic game ever >>played by a computer, and of such high quality as to be as good as any GM game >>ever played, strategically. >> > >This game is highly overrated. >The opening variation was ideal for any computer program. >Deep Blue just made normal moves that improved its position. >Black had no chance to get a dynamic counter play. >At the end Deep Blue even almost spoiled it. > >Michael I'll gently remind you it is the _only_ match against computers that K has lost. :) That's not a bad accomplishment under any circumstances... >> >> >>>> >>>> >>>>> >>>>>[Event "Hanover"] >>>>>[Site "Hanover"] >>>>>[Date "1991.??.??"] >>>>>[Round "7"] >>>>>[White "Tischbierek, Raj"] >>>>>[Black "Deep Thought II"] >>>>>[Result "1-0"] >>>>>[ECO "B23"] >>>>>[PlyCount "44"] >>>>>[EventDate "1991.05.??"] >>>>> >>>>>1. e4 c5 2. Nc3 Nc6 3. Nge2 e5 4. Nd5 d6 5. Nec3 Nge7 6. Bc4 Nxd5 7. Bxd5 Be7 >>>>>8. d3 Nd4 9. O-O Bh4 10. f4 O-O 11. f5 Rb8 12. a4 Bd7 13. g3 Bg5 14. Rf2 Bxc1 >>>>>15. Qxc1 Bc6 16. f6 gxf6 17. Qh6 Qb6 18. Qxf6 Be8 19. Raf1 Qxb2 20. Qg5+ Kh8 >>>>>21. Nd1 Qb4 22. c3 Qa3 1-0 >>>>> >>>>>[D] r1r3k1/1q1n1p1p/p2Q2pb/3Rp3/P1p1P3/1P3P2/5BPP/4KB1R w K - 0 22 >>>>> >>>>>How long does it take Crafty to avoid 22.bxc4? >>>> >>>>Perhaps more important: How long would it take Crafty to play all the other >>>>_good_ moves they played there? >>>> >>>> >>>> >>>>> >>>>>[Event "Hanover"] >>>>>[Site "Hanover"] >>>>>[Date "1991.??.??"] >>>>>[Round "6"] >>>>>[White "Deep Thought II"] >>>>>[Black "Wahls, Matthias"] >>>>>[Result "0-1"] >>>>>[ECO "E86"] >>>>>[PlyCount "56"] >>>>>[EventDate "1991.05.??"] >>>>> >>>>>1. d4 d6 2. c4 g6 3. Nc3 Bg7 4. e4 Nf6 5. f3 O-O 6. Be3 e5 7. Nge2 c6 8. Qd2 >>>>>Nbd7 9. d5 cxd5 10. Nxd5 Nxd5 11. Qxd5 Nb6 12. Qb5 Bh6 13. Bf2 Be6 14. Nc3 Qc7 >>>>>15. b3 Nd7 16. Qb4 a6 17. Rd1 Rfc8 18. Nd5 Bxd5 19. Rxd5 b5 20. a4 bxc4 21. >>>>>Qxd6 Qb7 22. bxc4 Bf8 23. Qxd7 Qb4+ 24. Rd2 Rd8 25. Qxd8 Rxd8 26. Be3 Bc5 27. >>>>>Bg5 Rd6 28. Ke2 Rxd2+ 0-1 >>>>> >>>>>Michael >>>>> >>>>> So thinking that its >>>>>>successor, which was 100x faster, would lose badly to today's programs is simply >>>>>>logic that I can't follow. There is absolutely _no_ basis to make such a wild >>>>>>leap of faith. >>>>>> >>>>>>For the record, Cray Blitz, in 1980, had a USCF rating of 2300. Running exactly >>>>>>one thousand nodes per second. Be careful of what you write if you are not sure >>>>>>of your facts. In this case you are simply wrong.
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.