Author: Robert Hyatt
Date: 08:03:09 06/01/00
Go up one level in this thread
On June 01, 2000 at 03:48:12, Ed Schröder wrote: >On May 31, 2000 at 02:29:21, blass uri wrote: > >>On May 31, 2000 at 01:40:08, blass uri wrote: >> >>>On May 30, 2000 at 21:43:12, Robert Hyatt wrote: >>> >>>>On May 30, 2000 at 17:54:45, Joshua Lee wrote: >>>> >>>>>On May 30, 2000 at 17:02:40, blass uri wrote: >>>>> >>>>>>On May 30, 2000 at 16:51:08, stuart taylor wrote: >>>>>> >>>>>>>I just noticed on ssdf rating list details, that hiarcs7.32 on 450mhz. beat >>>>>>>fritz 3 on 90mhz. 18.5 to 3.5. That fritz was very similar to the exact thing >>>>>>>which beat deep blue at the time.ah!!!! so what do you say to that? >>>>>>>S.Taylor >>>>>> >>>>>>I say that it is less than 90% and I read that Deep thought(not deep blue) got >>>>>>more than 90% against Fritz3(p90) >>>>>> >>>>>>I do not know if to believe to the last claim because they did not do the games >>>>>>public and I have no idea if the games are tournament time control or faster >>>>>>time control(I am interested only in tournament time control games). >>>>>> >>>>>>Uri >>>>>How about Hiarcs on my Athlon 800 Clocked to 880Mhz? >>>>>I think if it is still taking Hiarcs on Pos 3 of the LCTII test 49minutes to >>>>>solve it at 11ply and Deep Thought of 1989 which played Kasparov was searching >>>>>2M nps and 12Ply in 40/2 then Most computers would in fact win a game or two but >>>>>not a match. also Deep Thought of 1988 at 750,000 nodes per second would be >>>>>better but i have looked at games of the pre 1990 computers and can only say >>>>>that Hiarcs has to be better than some of those computers because it can spot >>>>>the mistakes right off the bat and wouldn't play the loosing move in the first >>>>>place. I'll find the game.... other than that my reasoning was just that the >>>>>opening books caused those programs to lose. >>>>> >>>>>maybe everyone interested should have their respective software analyze older >>>>>games like that of Cray Blitz and Hitech. >>>> >>>> >>>>I did this for the 1986 WCCC event (Cray Blitz only). I was amazed that Crafty >>>>did not find one single tactical blunder, even though Crafty of today is >>>>searching far faster than CB of 1986 (we were doing about 160K nodes per second >>>>back then on an 8 cpu YMP I believe). I used "annotate" for each game played. >>>> >>>>Chris whittington raised the question of a really ugly looking move Bh7 against >>>>Bobby I think. And he criticized it endlessly. And then we discovered that it >>>>was forced and CSTal also liked the _same_ move. :) >>>> >>>>That says a lot about the robustness of a good 1986 search on pretty good 1986 >>>>hardware. It is easy to reproduce the test since crafty will annotate a >>>>collection of PGN game scores (in a single file) at one batch run, >>>>automatically. >>>> >>>>I think you will find that the tactical mistakes of the 1986 supercomputers are >>>>_very_ hard to find with today's PC machines. >>> >>>Tactical mistakes of deep thought are not hard to find with today PC's program. >>>The last one was against Fritz3 but I found more mistakes in some games that >>>they lost or did not win. >>> >>>Uri >> >>I can add that I can see tactical mistake of Deeper blue in the game that it >>lost >> >>[D]4r3/8/2p2PPk/1p1r4/pP2p1R1/P1B5/2P2K2/8 b - - 0 1 >> >>Deeper blue played Rd1 instead of Rf5+ >> >>Rf5+ is losing but the logfiles of deeper blue showed that it did not see the >>line Rf5+ Ke3(chessmaster(ss=10) can see it after 82 seconds and >> >>Rd5f5 kf2e2 Re8g8 pg6g7 Kh6h5 rg4g1 Rf5f3 bc3b2 Kh5h6 rg1g4 Rf3f5 bb2d4 Pe4e3 >>ke2e3P Rf5f1 ke3e4 Rf1f6p was the main line with evaluation of -210 before Rf5+ >>failed low with evaluation of -260 >> >>The final main line is a bad main line: >> >>Rd5d1 pf6f7 Rd1f1 kf2f1R Kh6h5 pf7e8R/q Kh5g4r with evaluation of -180 when >>white has a simple mate after these stupid moves. >> >>Uri > >Interesting indeed... > >The log-file in question: > >#[Rf5](-210)[Rf5](-210) -210v T=73 >Rd5f5 kf2e2 Re8g8 pg6g7 Kh6h5 rg4g1 Rf5f3 bc3b2 Kh5h6 rg1g4 Rf3f5 bb2d4 Pe4e3 >ke2e3P Rf5f1 ke3e4 Rf1f6p > >11(6) #[Rf5](-260)v[find a move]#########[TIMEOUT][et3 1295 sec] > >#[Rg8](-183)[Rd1](-180) -180 T=204 >Rd5d1 pf6f7 Rd1f1 kf2f1R Kh6h5 pf7e8R/q Kh5g4r Note that DB used the hash table to construct the PV. The hardware processors could not supply _any_ pv moves, which means the way most of us construct a PV when we call Evaluate() and then pass it back ply by ply doesn't work. As a result, I have seen more than one bogus PV from them. The move at the root is always right of course, as that is how any search works. But the PV can (and does) have glitches when you have to depend on reconstruction after the fact. It is easy to fail low, search a long time trying to find a better move, totally blowing the hash table out, and then all you get for a PV are the few first moves, plus (maybe) a few bogus key matches. I tried this approach once up on a time and hated it. I would say one out of every 10-20 moves, I would get a good score at the root, play the move, and start pondering the second move in the 'reconstructed PV' only to discover it was preposterous. Score at root would be (say) +.3, score after pondering the supposedly best move in PV would be +9. I think you have to take their PV with a grain of salt, although I trust their score at the root pretty well since it was a 17 ply search in this case. > >Analysis of Rf5+ (using Analysis Include): > >00:00:06 9.00 -1.70 1..Rf5+ 2.Ke1 e3 3.f7 Rd8 4.Ke2 > Kh5 5.g7 Rxf7 6.g8=Q (1) > >00:00:23 10.00 -1.73 1..Rf5+ 2.Ke3 Rf3 3.Ke2 Rxc3 4.f7 > Rd8 5.g7 Rxc2 6.Ke1 Rc1 7.Kf2 e3 > 8.Kg2 Rc2 9.Kh3 e2 10.g8=Q (6) > >00:00:58 11.00 -1.73 1..Rf5+ 2.Ke3 Rf3 3.Ke2 Rxc3 4.f7 > Rd8 5.g7 Rxc2 6.Ke1 Rc1 7.Kf2 e3 > 8.Kg2 Rc2 9.Kh3 (23) > >00:02:41 12.00 -2.21 1..Rf5+ 2.Ke2 Rg8 3.g7 Kh5 4.Rg1 > >00:06:24 13.00 -2.37 1..Rf5+ 2.Ke3 Rf3 3.Ke2 Rg8 4.Bd2 > Kh5 5.Rg5 Kh4 6.f7 Rg7 7.Re5 Kg3 > 8.Be1 Kg2 (161) > >So a constantly dropping score for Rf5+ > >Then Analysis of Rd1 (using Analysis Include): > >00:00:04 9.00 -3.27 1..Rd1 2.f7 Rf8 3.Ke3 Rdd8 4.Bf6 > Ra8 5.Be7 Kg7 6.Bxf8 Rxf8 (1) > >00:00:11 10.00 -3.48 1..Rd1 2.g7 Rd5 3.Ke2 Rf5 4.Rh4 > Kg6 5.Rh8 Rg8 6.Rxg8 Rf3 (4) > >00:00:31 11.00 -3.71 1..Rd1 2.g7 Rd5 3.Ke2 Rf5 4.Rh4 > Kg6 5.Rh8 Rg8 6.Rxg8 Rxf6 (11) > >00:01:25 12.00 -4.03 1..Rd1 2.g7 Rd5 3.Ke2 Rd7 4.g8=Q > Rxg8 (31) > >00:03:49 13.00 -4.20 1..Rd1 2.g7 Rd5 3.Ke2 Rd7 4.g8=Q > Rxg8 (85) > >Rf5+ looks clearly better and the score of -1.80 for Rd1 given by >DB looks ridiculous and I don't understand it either. > >Ed
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.