Author: Mike S.
Date: 21:02:50 12/07/02
Go up one level in this thread
On December 07, 2002 at 23:16:13, Robert Hyatt wrote: >On December 07, 2002 at 22:11:01, Mike S. wrote: >(...) >>[D]R7/3b1k2/8/6R1/4r2K/7P/3r2P1/8 w - - 0 92 >>Crafty, but also other engines, evaluate that position with a *big* advantage >>for black, sometimes up to -2.x. No combination involved (the material balance >>didn't change the next 80 moves). > >If Crafty says black is winning, something is broken as there is specific >code to handle that particular case (one side is a minor piece up but with >no pawns it has no chance of winning even if the opponent didn't have any >pawns left... > >I'll look to see if I broke something... although 18.15 might be old enough >that the fix was not in... Here is an analysis output of the cb-native Crafty 19.01 (2nd edition), from within the game score, P3/700 MHz 60 MB hash: Delfi 3.2 - Crafty-18.15 R7/3b1k2/8/6R1/4r2K/7P/3r2P1/8 w - - 0 1 Analysis by Crafty 19.01: 92.g4 Re3 93.Rh5 Kg7 94.Kg5 Rf3 95.Ra5 Re2 96.Rh6 Rg3 97.Rg6+ Kf7 98.Rf6+ Ke7 99.Kh4 -+ (-2.13) Depth: 1/5 00:00:00 -+ (-1.96) Depth: 13/23 00:01:39 36745kN, tb=1786 Then, I removed 2 rooks from the position (resulting in intense tbs. accesses): [D]8/3b1k2/8/6R1/4r2K/7P/6P1/8 w - - 0 1 92.g4 Re3 93.Rc5 Ke6 94.Rc7 Kd6 95.Rb7 Be8 96.Rb8 Kc5 97.Kg5 Re5+ 98.Kh6 Re6+ 99.Kh7 Re5 = (0.00) Depth: 1/4 00:00:00 = (0.00) Depth: 14/22 00:00:28 9343kN, tb=12478 Alternatively, I removed White's pawns only: [D]R7/3b1k2/8/6R1/4r2K/8/3r4/8 w - - 0 1 92.Kg3 Bc6 93.Rf5+ Ke6 94.Raf8 Re3+ 95.Kh4 Rh2+ 96.Kg4 Rg2+ 97.Kh4 Bd7 98.R8f6+ Ke7 99.Rf7+ Kd6 100.R7f6+ Be6 101.Kh5 -+ (-4.75) Depth: 1/3 00:00:00 -+ (-4.95) Depth: 13/24 00:01:35 51648kN, tb=26626 I played some moves forward, but it became not clear to me why the evaluation is so high in Black's favour. I ran a DOS file compare (COMP) utility to compare all my KR* tablebases with a CD which contains them too, and they were reported identical. Tablebase cache size was 8 MB both in WinBoard and in Fritz. First I thought, it may be a general evaluation problem of many engines, but especially the last Crafty result indeed looks as if something is broken. For comparison, here is the output of Shredder 5.32 which is usually good in such tbs.-near positions, for the 3 above: Delfi 3.2 - Crafty-18.15 R7/3b1k2/8/6R1/4r2K/7P/3r2P1/8 w - - 0 1 Analysis by Shredder 5.32: 92.g4 Re3 93.Rh5 Kg7 94.Rd8 Rdd3 95.Kg5 Re2 96.Rb8 ³ (-0.62) Depth: 1/2 00:00:00 ³ (-0.55) Depth: 12/24 00:02:36 29201kN, tb=41678 New game 8/3b1k2/8/6R1/4r2K/7P/6P1/8 w - - 0 1 Analysis by Shredder 5.32: 92.g4 Re3 93.Rd5 Be6 94.Ra5 Kg6 95.Ra6 Kh6 96.Rb6 Kg6 97.Rb7 ² (0.41) Depth: 1/2 00:00:00 ² (0.32) Depth: 15/27 00:00:15 2170kN, tb=26159 (White advantage!) New game R7/3b1k2/8/6R1/4r2K/8/3r4/8 w - - 0 1 Analysis by Shredder 5.32: 92.Rg4 Bxg4 -+ (-6.20) Depth: 1/2 00:00:00 -+ (-8.06) Depth: 1/2 00:00:00 92.Kh5 Rh2# -+ (-1.65) Depth: 1/2 00:00:00 -+ (-#1) Depth: 2/4 00:00:00 92.Rg4 Rxg4+ 93.Kh5 -+ (-#1) Depth: 2/4 00:00:00 -+ (-9.65) Depth: 2/4 00:00:00 92.Kg3 Rb4 93.Ra3 Kf6 94.Rc5 Rbb2 95.Kf4 Be6 96.Rc1 Rg2 97.Ke4 -+ (-5.69) Depth: 2/4 00:00:00 µ (-1.35) Depth: 11/22 00:01:14 13771kN, tb=45816 So far Shredder, which seems to evaluate ok as far as I can tell. It looks like I may really have spotted a Crafty bug this time... I'd be glad if I have, at least once, contributed something which can help to improve further (even if it leads only to a bugfix). Maybe this is even part of the reason, why the 19.01 version has been less successfull in comp-comp games than 17.x/18.x versions. Regards, Mike Scheidl
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.