Author: Ricardo Gibert
Date: 11:55:49 01/21/01
Go up one level in this thread
On January 21, 2001 at 11:02:05, Robert Hyatt wrote: >On January 21, 2001 at 09:12:59, Ricardo Gibert wrote: > >>[D]8/P6p/1P5P/8/8/4k3/5r2/4K3 b - - 0 0 >>[D]8/1P5p/P6P/8/8/4k3/5r2/4K3 b - - 0 0 >> >>If Crafty encounters these as leaf node positions, what evalution does it >>return? > > >It would think white is winning. However, if it encounters a position like >win at chess #141 at a leaf position it might think black is winning, when >white has a forced mate in 6. > >In this position I would agree with Crafty's evaluation. It seems to me that >if anybody wins at all, white wins. The non-ending checks/mate threats make >this an exception. But then _all_ chess programs have that exception for deep >checking lines. The variations feature mostly non-checks. White just keeps threatening mate. The first one is a draw and the 2nd one is a win for White. Note both positions can come from the same parent position: [D]8/7p/PP5P/8/8/4k3/5r2/4K3 w - - 0 0 1.b7 The best move [1.a7? Ra2 2.Kd1 Kd3 3.Kc1 Kc3 4.Kb1 Ra6 5.b7 Rb6+ 6.Kc1 Rxh6 7.Kd1 Kd3 8.Ke1 Ke3 9.Kf1 Kf3 10.Kg1 Rg6+ draw] [1.Kd1 Wastes time, but does not throw away the win A) 1...Ra2 2.b7 Kd3 3.Kc1 Kc3 A1) 4.Kb1? Rb2+ 5.Kc1 (5.Ka1 Rb6 6.a7 Ra6+ 7.Kb1 Rb6+ 8.Kc1 Rxh6 draw) 5...Rb6 6.Kd1 Kd3 7.Ke1 Ke3 8.Kf1 Kf3 9.Kg1 Kg3 draw; A2) 4.b8Q White wins B) 1...Kd3 2.Kc1 Kc3 and now not 3.Kb1? Rf1+ 4.Ka2 Rf2+ 5.Ka3 Rf1 6.Ka2 draws (6.Ka4? Kc4 7.Ka3 Ra1+ 8.Kb2 Rxa6 Black wins) but rather 3.Kd1 Kd3 4.Ke1 Ke3 5.b7 transposing back to 1.b7 that wins for White] 1...Rb2 [Neither 1...Rh2 2.Kf1 Rxh6 3.b8Q Rxa6 4.Qb3+ nor 1...Ra2 2.Kd1 Kd3 3.Kc1 Kc3 4.b8Q save Black] 2.Kd1 Kd3 3.Kc1 Kc3 4.a7 Rh2 5.Kd1 Kd3 6.Ke1 Ke3 7.Kf1 Kf3 8.Kg1 and White wins I just hope I didn't mangle the analysis in adapting it from the output Fritz generated in my interactive session with it. > >A quick check would be to leave white's two pawns where they are, but enumerate >_all_ the possible places you can put the white king and black king and rook, >and then compute the percentage of the time the current eval is wrong, and the >percent it is right. If it is right > 50% of the time (it will be) then this >is the right thing to do. If it is wrong > 50% of the time, this is bad. > >I would expect it to be right in > 90% of the cases. Which means not doing >this would be wrong in 90% of the cases. Which side of that fence would you >rather be on? I agree with this and I would go further and suggest it could be right better than 99% of the cases with connected passed pawns on the sixth or beyond. I think that 99% is a reasonable bound given the amount of garbage positions alpha-beta looks at even with perfect move ordering. My only point is Crafty must search it out or *guess*. Evidently, it guesses quite well and that is perfectly okay. Great even. > >I didn't do this eval term to find wac#2. I did it because in very fast games >on ICC I was getting whacked by two connected passers that reached the 6th, even >when it was possible to prevent them from getting there if it had only known >that preventing this would be worth giving up a pawn to accomplish. > >I haven't seen this term lose a single game that I can recall. At least not >since it was refined a year or two ago. That is the main way I evaluate the >effectiveness of such evaluation terms...
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.