Author: Robert Hyatt
Date: 12:00:40 10/08/01
Go up one level in this thread
On October 07, 2001 at 14:47:48, Gareth McCaughan wrote: >Robert Hyatt wrote: > >[Slater Wold mentioned the following position and asked >how long programs take to see that Qxd4 draws:] >>>>> [D]8/1kbQ2NK/4p2p/8/1p1p1P2/8/4q3/8 w - - 0 1 >>>>> >>>>> Solution is Qxd4! and the game draws. >[I reported:] >>>> Crafty 18.11, Athlon 1GHz, takes about 15 1/2 hours >>>> to switch from Nxe6 to Qxd4. It doesn't think it's >>>> drawing (eval is about -2.6). >>>> >>>> At the moment (nearly 60 hours in) it's working on >>>> ply 20 and doesn't have an eval or a PV yet. >>>> The PV at the end of ply 19 is: >>>> >>>> 1. Qxd4 Qc2+ 2. Kxh6 b3 3. Nxe6 b2 >>>> 4. Nc5+ Kc6 5. Qd7+ Kb6 6. Na4+ Qxa4 >>>> 7. Qxa4 b1=Q 8. Qd4+ Kc6 9. Qc4+ Kd6 >>>> 10. Qd4+ Ke6 11. Qc4+ Kd7 12. Qf7+ >>>> Kc8 13. Qg8+ Bd8 14. Qe6+ Kc7 15. Qe5+ >>>> Kd7 16. Qd5+ Ke7 17. f5 >[Uri Blass said:] >>> The position after 7...b1Q should be evaluated as something >>> close to draw because queen and bishop usually does not win >>> against queen and here black has queen and bishop against >>> queen and pawn. >[Bob:] >> Nope. Never included all the special cases, just the ones that >> I have seen such as KR+minor + KR or KR+P. Lots of KQB vs KQP >> are winnable. enough that just saying "draw" would lead to >> embarassing results. This problem will become moot in another >> couple of years as more 6=piece EGTBs are done. I think it is >> dangerous to call KR vs KN a draw, because some are winnable, >> and I watched a game vs some computer on ICC last week where >> this came up. Crafty was in a KRP vs KBN ending and the opponent >> played BxP with an eval of 0.00, and crafty responded RxB with >> a score of mate in something. Apparently the opponent had >> that particular score hard-coded, and if there are exceptions, >> that becomes embarassing at a critical moment... > >Sure. But then, playing Nxe6 instead of Qxd4 in Slater's >position because you think the position after Qxd4 is -2.6 >is also embarrassing. Obviously it would be dangerous to >tell Crafty (or any other program) that KQB v KQP endings >are all drawn (i.e., score of 0 and stop searching), but >if you have the choice between being wrong 70% of the time >by saying -2.6 or being wrong 30% of the time by saying >-0.5 then the latter is better. (All the numbers in the >previous sentence were pulled out of a hat and could be >badly wrong, but the point I'm making is fairly stable >under such perturbations.) > >-- >g If I thought that 30%-70% split was correct, I would obviously do it. At the present, I don't have any idea whether the current code is correct more than it is wrong, or vice-versa...
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.