Author: Uri Blass
Date: 12:57:08 09/19/01
Go up one level in this thread
On September 19, 2001 at 15:19:44, Robert Hyatt wrote: >On September 19, 2001 at 11:30:16, Uri Blass wrote: > >>On September 19, 2001 at 11:04:38, Robert Hyatt wrote: >> >>>On September 19, 2001 at 10:16:36, Uri Blass wrote: >>> >>>>On September 19, 2001 at 09:52:52, Robert Hyatt wrote: >>>> >>>>>On September 19, 2001 at 05:20:31, Bernhard Bauer wrote: >>>>> >>>>>>>>>>Here is a simple attempt: >>>>>>>>>> >>>>>>>>>>[D]2k5/1r6/3p1p2/n2p1p2/P2PpP2/R3P3/1BK5/8 b - - >>>>>>>>>> >>>>>>>>>>Here black has several moves to try, one which liquidates into a pawn up >>>>>>>>>>(but dead lost) ending. Rxb2 Kxb2 Nc4+ Ka2 Nxa3 Kxa3 and white is a pawn >>>>>>>>>>down, but winning easily. >>>>>>>>>> >>>>>>>>>>Once you start with Rxb2, you are committed. As if you try to back out and >>>>>>>>>>not play Nc4 and Nxa3, you are an exchange down. And if you do recover the >>>>>>>>>>material, you are dead lost. Add another such forced capture/recapture and >>>>>>>>>>you have burned 6 plies. You won't see white winning all the black pawns >>>>>>>>>>and winning. >>>>>>>>> >>>>>>>>> >>>>>>>>>Note that I don't say there are not better moves for black here. The point >>>>>>>>>was to show a move choice that commits you to a course of action that gets >>>>>>>>>worse and worse as you go deeper and deeper. >>>>>>>> >>>>>>>>I think that this is not a good example because white has an obvious positional >>>>>>>>advantage for programs(white has a passed pawn when black has 2 pawns on the >>>>>>>>same file for file d,f >>>>>>>> >>>>>>>>Uri >>>>>>> >>>>>>> >>>>>>>Pick any such position you want, where one side is a pawn up but the other is >>>>>>>winning. I have seen many. That is one example where if you trade, you lose. >>>>>>>And it is one example of where one extra pawn does _not_ mean you are winning. >>>>>>>Here it means you are losing and badly. >>>>>>> >>>>>> >>>>>>From a players point of view *white*is a pawn up, the a-pawn and therefor >>>>>>winning, just like Ed said. >>>>>>A player would not count the additional blocked black pawns. >>>>>>Kind regards >>>>>>Bernhard >>>>> >>>>> >>>>>Human, right. But set the position up and ask your favorite program which side >>>>>is winning with a simple static eval... >>>> >>>>I asked few program to give me their evaluation at depth 1 for this position >>>>[D]2k5/8/3p1p2/3p1p2/P2PpP2/n3P3/K7/8 w - - 0 3 >>>> >>>> >>>>CometB27 0.53 for white >>>>Junior7 0.20 for black. >>>>shredder5.32 0.18 for black. >>>> >>>>They clearly can see white's positional advantage by static evaluation and do >>>>not give +1 for black. >>> >>>The two commercial programs say black is better. White is winning. You don't >>>see a problem with that kind of incorrect evaluation? IE as black they would >>>go for this position rather than a repetition draw. >> >>there is a problem with the evaluation but it does not contradict my claim that >>it is usually safe to evaluate as +2 when there is no positional reason to >>prefer the weaker side in material. >> >>I meant that changing scores of 1.xx in pawn endgames to 2.xx if the side with >>material advantage has an 1.xx evaluation is usually safe >> >>I did not say that changing 0.xx to 1.xx is usually safe and I also did not say >>that improving the evaluation function is a bad idea. >>> >>>That was my point. If you say "a pawn ahead is winning" you are going to lose >>>many games by doing so. Particularly when your opponent notices that you are >>>doing this and steers the game into such positions. I will be happy to give >>>you a similar position where the pawns are not isolated if you think that is >>>the problem... >>> >>>Keep the pieces at the same squares. Pub black pawns at h7, g6, e7 and d6. >>>put white pawns at g5, d5 and a4. What does your engines think now? Black >>>is a pawn up. no isolated pawns... >> >>Shredder5.32 can see positive score at depthes that are at least 7 and >>positional advantage for white at depth 4. >> >> >>[D]2k5/4p2p/3p2p1/3P2P1/P7/n7/K7/8 w - - 0 1 >> >>Analysis by Shredder 5.32: >> >>1.Kxa3 h5 >> -+ (-1.84) Depth: 1/2 00:00:00 >>1.Kxa3 h5 2.gxh6 >> -+ (-2.88) Depth: 1/2 00:00:00 >>1.Kxa3 Kd7 2.Ka2 >> -+ (-1.84) Depth: 2/4 00:00:00 >>1.Kxa3 Kb7 2.Ka2 >> µ (-1.22) Depth: 2/4 00:00:00 >>1.Kxa3 Kb7 2.Kb4 >> ³ (-0.35) Depth: 2/4 00:00:00 >>1.Kxa3 Kb7 2.Kb4 h5 3.gxh6 >> -+ (-3.22) Depth: 3/6 00:00:00 >>1.Kxa3 Kd8 2.Ka2 h5 >> -+ (-1.87) Depth: 3/6 00:00:00 >>1.Kxa3 Kb7 2.Kb4 Ka8 3.Ka3 >> -+ (-1.56) Depth: 4/8 00:00:00 >>1.Kxa3 Kd8 2.Ka2 Kd7 3.Ka1 >> µ (-1.00) Depth: 4/8 00:00:00 >>1.Kxa3 Kc7 2.Kb4 Kb7 3.Kc4 >> = (-0.13) Depth: 4/8 00:00:00 >>1.Kxa3 e5 2.dxe6 Kd8 3.e7+ Kxe7 >> ³ (-0.43) Depth: 5/10 00:00:00 >>1.Kxa3 Kd8 2.Kb4 e5 3.Kc4 Ke7 >> ³ (-0.64) Depth: 5/10 00:00:00 >>1.Kxa3 Kd8 2.Kb4 Kc7 3.Kc4 Kb6 4.Kd3 >> ³ (-0.38) Depth: 6/12 00:00:00 2kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb3 Ka6 4.Kb4 Kb7 5.Kc4 >> = (-0.02) Depth: 6/12 00:00:00 3kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb3 Ka6 4.Kb4 Ka7 5.Kc4 h6 >> = (0.04) Depth: 7/14 00:00:00 6kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb3 Ka6 4.Kb4 Kb7 5.Kc4 Kc7 6.Kb4 Kd8 7.Kc4 Kc7 8.Kb4 >> = (0.04) Depth: 8/16 00:00:00 9kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Ka4 Ka6 4.Kb4 Kb7 5.Kb5 Kb8 6.Kc6 e6 >> ² (0.29) Depth: 9/18 00:00:00 14kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Ka4 Ka6 4.Kb4 Kb7 5.Kb5 Kb8 6.Kc6 e6 >> = (0.18) Depth: 9/18 00:00:00 15kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Ka4 Ka7 4.Kb3 Kb7 5.Ka4 Ka7 >> = (0.04) Depth: 10/20 00:00:00 25kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Ka4 Ka6 4.Kb4 h5 5.gxh6 e5 6.dxe6 d5 >> = (0.04) Depth: 11/22 00:00:01 45kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.a6+ Kb8 6.Kc4 Ka8 7.Kb5 e5 8.dxe6 h5 >>9.gxh6 >> ² (0.29) Depth: 12/24 00:00:01 66kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.a6+ Ka7 6.Ka5 Kb8 7.Kb6 Ka8 8.a7 >> ² (0.33) Depth: 12/24 00:00:01 73kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kc4 Ka6 5.Kb4 h5 6.gxh6 e5 7.dxe6 d5 >> ² (0.33) Depth: 13/25 00:00:02 113kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.Ka4 Ka6 6.Kb4 h5 7.gxh6 e5 8.dxe6 d5 >> ² (0.33) Depth: 14/26 00:00:03 154kN >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.Ka4 Ka6 6.Kb4 h5 7.gxh6 e5 8.dxe6 d5 >> ² (0.33) Depth: 15/27 00:00:04 203kN, tb=4 >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.Ka4 Ka6 6.Kb4 h5 7.gxh6 e5 8.dxe6 d5 >> ² (0.33) Depth: 16/28 00:00:05 277kN, tb=9 >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.Ka4 Ka6 6.Kb4 h5 7.gxh6 e5 8.dxe6 d5 >> ² (0.33) Depth: 17/29 00:00:08 379kN, tb=42 >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.a6+ Kc7 6.Ka5 Kc8 7.Kb6 e6 8.dxe6 >> ² (0.58) Depth: 18/30 00:00:16 805kN, tb=72 >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 e5 5.dxe6 h5 6.gxh6 Kb7 >> ± (1.08) Depth: 18/30 00:00:17 892kN, tb=72 >>1.Kxa3 Kc7 2.a5 Kb7 3.Kb4 Ka7 4.Kb5 Kb7 5.a6+ Ka7 6.Ka5 Ka8 7.Kb5 h5 8.gxh6 Kb8 >>9.h7 Kc7 >> +- (6.40) Depth: 18/30 00:00:39 1615kN, tb=380 >> >> >>Junior7 needs more time to see positive score but it still can see a positive >>positional score from the first iteration. >> >>Junior 7 - ,M >>2k5/4p2p/3p2p1/3P2P1/P7/n7/K7/8 w - - 0 1 >> >>Analysis by Junior 7: >> >>1.a5 Nc4 >> -+ (-2.98) Depth: 3 00:00:00 >>1.Kb3 Nb1 >> -+ (-2.24) Depth: 3 00:00:00 >>1.Kxa3 Kd7 2.Kb4 e5 3.dxe6+ >> ³ (-0.69) Depth: 3 00:00:00 >>1.Kxa3 Kb7 2.Kb3 Kb6 3.Kb4 >> ³ (-0.47) Depth: 6 00:00:00 1kN >>1.Kxa3 Kd7 2.Kb4 e5 3.Ka5 Ke7 4.Kb5 >> ³ (-0.29) Depth: 9 00:00:00 9kN >>1.Kxa3 Kb7 2.Kb3 Kb6 3.Kc4 Ka5 4.Kb3 Ka6 5.Kb4 >> = (-0.02) Depth: 12 00:00:01 50kN >>1.Kxa3 Kd7 2.Kb4 e5 3.dxe6+ Kxe6 4.a5 Kd7 5.Kb5 Kc7 6.a6 Kb8 7.Kb6 >> = (-0.22) Depth: 15 00:00:03 302kN >>1.Kxa3! >> = (0.08) Depth: 18 00:00:08 778kN >>1.Kxa3! Kd7 2.a5 Kc8 3.Kb4 Kc7 4.Kb5 Kb7 5.a6+ Ka7 6.Ka5 Ka8 7.Kb6 >> = (0.13) Depth: 18 00:00:11 1065kN >>1.Kxa3! >> ² (0.43) Depth: 20 00:00:17 1680kN >>1.Kxa3 Kd7 2.a5 Kc8 3.Kb4 Kc7 4.Kb5 Kb7 5.a6+ Ka7 6.Ka5 Ka8 7.Kb6 Kb8 8.a7+ >> ² (0.43) Depth: 22 00:00:35 3489kN, tb=2 >>1.Kxa3! >> ± (0.73) Depth: 23 00:00:52 4936kN, tb=8 >> >>Uri > > >I see a + score at depth=1. It drops slightly negative at depth=2, and >then goes positive from that point on: > 1 0.00 0.16 1. Kxa3 > 1-> 0.00 0.16 1. Kxa3 > 2 0.00 -0.14 1. Kxa3 Kd7 > 2-> 0.00 -0.14 1. Kxa3 Kd7 > 3 0.00 0.17 1. Kxa3 Kd7 2. a5 > 3-> 0.00 0.17 1. Kxa3 Kd7 2. a5 > 4 0.00 0.17 1. Kxa3 Kb7 2. a5 Ka6 > 4-> 0.00 0.17 1. Kxa3 Kb7 2. a5 Ka6 > 5 0.00 0.47 1. Kxa3 Kb7 2. Kb4 Kb6 3. a5+ Kc7 > 5-> 0.00 0.47 1. Kxa3 Kb7 2. Kb4 Kb6 3. a5+ Kc7 > 6 0.00 0.47 1. Kxa3 Kc7 2. a5 Kb7 3. Kb4 Kc7 > 6-> 0.00 0.47 1. Kxa3 Kc7 2. a5 Kb7 3. Kb4 Kc7 > 7 0.00 0.37 1. Kxa3 Kc7 2. a5 Kb7 3. Ka4 Ka6 4. > Kb4 > 7-> 0.00 0.37 1. Kxa3 Kc7 2. a5 Kb7 3. Ka4 Ka6 4. > Kb4 > 8 0.01 0.57 1. Kxa3 Kc7 2. a5 Kb7 3. Ka4 Ka6 4. > Kb4 Kb7 > 8-> 0.01 0.57 1. Kxa3 Kc7 2. a5 Kb7 3. Ka4 Ka6 4. > Kb4 Kb7 > >But note that my analysis is not based on "I am a pawn up, so I get a bonus. >It is based on "my distant passed pawn is a terrible advantage in a king and >pawn ending with no pieces left." > >But this wasn't about _my_ evaluation. It was about the idea of trading pieces >when a pawn up. Which is _obviously_ wrong in this position. And if you want >to get ripped by this, just let a GM find out you are doing this. cptnbluebear >found two holes in Crafty that he stomped me with over and over. One was to >trick me into winning a pawn but leaving him with a distant majority. Which, >at the time, Crafty didn't understand. And he won two of every three games >using this weakness... which was pretty amazing. There are a lot of computer-human games from the past and I almost never see humans trade all the pieces to win pawn endgames against programs. Junior played against GM's in serious tournaments and I never saw the GM's trade all the pieces in order to win a pawn endgame. I guess that if it happens then the problem is not only in evaluating pawn endgames but also in evaluating previous positions. I believe that programs can usually avoid a losing a pawn endgame by search or avoid doing an error because they plan a pawn endgame if they have better or equal endgames that is not a pawn endgames. There were serious tournament of human against computers in holland and even some years ago humans usually did not win pawn endgames against computers and they could know about this strategy because anand won a pawn endgame in 1994 against Genius3. The fact that the hardware get better does not help humans to win pawn endgames against computers so I need to see examples from games to see if there are programs that lose often pawn endgames against humans. If there is a program that is losing often pawn endgames against humans or losing games because of not evaluating pawn endgames correctly then I want to see the games. Uri
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.