Author: Robert Hyatt
Date: 15:31:20 05/23/01
Go up one level in this thread
On May 23, 2001 at 15:48:27, Uri Blass wrote: >On May 23, 2001 at 13:46:30, Robert Hyatt wrote: > >>On May 23, 2001 at 11:55:29, Bruce Moreland wrote: >> >>>On May 22, 2001 at 13:07:29, Robert Hyatt wrote: >>> >>>>Rh5 seems better although I couldn't get DB Jr's score of +1.5 in any reasonable >>>>time (two hours or so). >>>> >>>>I guess this leads to Uri's question (again) of "does Bxg7 actually win or >>>>not?" >>> >>>My program liked Bxg7 in a minute or two, with a score of around +1, but it is >>>speculative. >>> >>>It has not been tuned for this position. It had Bg7 Kg7 Ne5 and I don't >>>remember what all after that. >>> >>>I went out for a while and when I came back it had Rh5 with a fail-high. This >>>morning it is at +1.24, same thing. >>> >>>I'll leave it running for a day or two and see what happens. >>> >>>bruce >> >> >>It seems obvious that all the programs like Rh5. The worry is that this is >>a possibly bad move on a deep enough search, although I don't believe this >>myself (yet). >> >>Kasparov seems to think that Bg7 is best. But when he says 'best' I am not >>sure what that means, exactly. IE 'forced win'? 'good prospects'? Etc. And >>of course he _could_ easily be wrong, although I would tend to not think this >>is the normal case... >> >>Your 1.24 is getting closer to DB Jr's score than mine reached, although I >>didn't let it run overnight. Think I will crank it up and let it burn for >>a while myself... > >Interesting also to know how much time per position does Crafty need to get 11 >out of 13(I say 11 out of 13 and not out of 14 because it seems that the Bxg7 is >not correct so I do not count this position as a good test suite unless the >solution is Rh5). I will try to crank it up and let it run. I could start at 1 hour per position, and then for the ones it misses, go to 4 or 8 to see what happens. No idea how hard this will be to do as I start at 6 after 20 minutes and there is a long way to go to reach 11. > >I believe that Deep Fritz on slower hardware can get something like 7 or 8 out >of 13 on PIII450(20 minutes per position). > >It could solve 1-5,7,8 but I did not give it enough time to find if it is going >to change it's mind in all of these cases. > >It could also find the right move in 9 but changed it's mind and it seem to fail >to solve it because of null move problems. > >It changed it's mind after more than 20 minutes on PIII450 so I may consider it >as solved on pIII450(20 minutes per position) Note that the original test was for 15 minutes per position if my instructions were correct... > >I did some analysis and found that later in the tree after Rd6 Rxd6 Nxd6+ Kd7 >Nb5 Ng7 h6 Deep Fritz has fail low,fail high, fail low, fail high... and can >never see things that Crafty has no problem to see. > >It is clearly a null move problem for Fritz because if I put selectivity=0 it >does not show the same problem. > > > > >The information of the results of deep blue Junior does not prove that Deep blue >was really better in tactics than the top programs of today because of the >following reasons: > >1)Tactics in games is different than tactics in test suites and a program can be >better at test suites when it fails to see tactics in games. Would never argue. I never used test suites to draw conclusions about them. But it is useful to have a well-known position that everybody can search and compare to their results, just for comparison's sake. > >2)It is possible that there are some software improvement in Deep blue Junior >and that the deep blue Junior that was tested is better than the deeper blue >that played against kasparov. > > No Idea there. Anything is possible. The hardware is certainly the DB2 stuff, but as Hsu said, they were using less than 50% of the new evaluation hardware he added in the last revision of the chip, so there is obviously lots of room to add stuff. Bob
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.