Author: Robert Hyatt
Date: 10:09:13 10/11/02
Go up one level in this thread
On October 11, 2002 at 03:51:08, Uri Blass wrote: >On October 10, 2002 at 20:09:12, Robert Hyatt wrote: > >>On October 10, 2002 at 16:59:54, Jeremiah Penery wrote: >> >>>On October 10, 2002 at 15:55:32, Uri Blass wrote: >>> >>>>Deep blue did not see Qe3 in the main line based on the logfiles of IBM. >>>>A lot of programs of today can see it with no problem. >>>> >>>>Some programs can even see the move Kh1 instead of Kf1 of Deep blue(I did not >>>>check the last deep Fritz but previous Deep Fritz is one of them). >>>> >>>>It is not a proof that deep blue is weaker but it is an evidence that Deep >>>>blue's evaluation is inferior. >>>> >>>>Good evaluation is not about knowing the truth(that you cannot know) but about >>>>giving a better estimate and it seems that the top programs of today get better >>>>decisions based on positional reasons. >>> >>>In Hsu's book, there is a large section about the 1997 match, going through >>>game-by-game. It's very surprising how many evaluation bugs they found, and >>>there's no doubt there were many more. The full DB2 never played a serious game >>>before the 1997 match, either, so it's inevitable that things weren't tuned to >>>the best possible accuracy. It's hard to compare something like that to a >>>commercial program of today, which plays thousands of testing games and the >>>programmers have as long as they want to tune each parameter. >>> >>>DB2 only played 6 games in its lifetime. To compare its positional strength to >>>anything based on those games is meaningless. Or if you really want, I can >>>easily find 6 games where some commercial program was completely clueless >>>positionally. :) >> >> >>It simply shows that DB was very resiliant _and_ strong. > >The fact that the deep blue team found many evaluation >bugs in the match against kasparov does not show that >it was strong. > >Deeper blue was clearly the best computer >at the time of the match against kasparov. > >I believe that it could continue to be the best >in case that they continue to work on it but kasparov simply destoyed >it by losing against it. > >I believe that bugs in the evaluation are more important >in long time control. >In blitz you can compensate for wrong evaluation by taking advantage of a >tactical mistake. > >In long time control if you go for a bad position because of >a bug there is bigger probability that nothing is going to >help you and even if you see that you lose before >the opponent see the win you still lose the game. > > > > It was barely put >>together in time >>for the match, yet it managed to win. I did _exactly_ the same thing in 1983 >>when we won >>the WCCC event. We had not played a complete game prior to round 1. In fact, >>as I drove >>to New York, Harry Nelson was in Minneapolis at Cray headquarters helping them >>fix a >>multiprocessing bug in their operating system we had found. We started round 1 >>never >>having played more than 2-3 moves in a row. And the program performed >>flawlessly for >>five rounds and won outright... > > >> >>DB was very strong, regardless of what others say. To say Fritz is stronger is >>a joke, and >>a bad one at that. Based on what? DB _did_ beat Kasparov. Fritz is getting >>squeezed beyond >>imagining. > >You know that this does not prove nothing because >the conditions for the match Fritz-Kramnik are unfair. > > > >> I think that the DB accomplishment was remarkable, and all the more >>so after >>reading Hsu's book which accounts for the year prior to "the match" in 1997. > >I agree that it was a good result. > >I have one question about the depth. > >You say that 12 plies with no pruning in 3 minutes >is impossible for Fritz(p850) in normal >middle game position(I agree about it). > > >Do you think that 12 plies in software(no pruning >in software) >and 6 additional plies in hardware in 3 minutes was possible for >deep blue. > >Even if I assume branching factor of 3 to get >the last 6 plies then it means that the number of nodes >is only 729 times bigger. > >Deep blue is clearly faster than P850 but the number of >nodes is less than 729 times bigger (Fritz >get some hundreds of knodes per second on p850). > >Uri What does it show? That the program was strong enough to win in _spite_ of several evaluation bugs. Crafty has won games with terrible bugs in place. It simply shows that it, too, is very strong. Strong enough that a bug is not enough to wreck a game. Also comparing NPS is meaningless. Fritz has a known simple evaluation. DB appears to be not-so-simple, based on papers, Hsu's book, and comments by the team here and there.
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.