Author: Robert Hyatt
Date: 11:23:16 10/01/01
Go up one level in this thread
On October 01, 2001 at 13:47:56, José Carlos wrote: >On October 01, 2001 at 12:53:09, Uri Blass wrote: > >>On October 01, 2001 at 12:37:09, José Carlos wrote: >> >>>On October 01, 2001 at 12:30:44, Uri Blass wrote: >>> >>>>On October 01, 2001 at 12:11:51, Robert Hyatt wrote: >>>> >>>>>On September 30, 2001 at 11:12:47, Uri Blass wrote: >>>>> >>>>>>On September 30, 2001 at 10:39:50, Miguel A. Ballicora wrote: >>>>>> >>>>>>>On September 30, 2001 at 06:15:44, José Carlos wrote: >>>>>>> >>>>>>>>On September 29, 2001 at 20:15:16, Gareth McCaughan wrote: >>>>>>>> >>>>>>>>>José Carlos wrote: >>>>>>>>> >>>>>>>>>> The point is not the move, but the eval. The program must >>>>>>>>>> know white is winning: >>>>>>>>>> >>>>>>>>>> [D]R4rk1/5pp1/5q1p/1p1Qp3/8/1B6/1PP2bPP/5K2 w - - 0 1 >>>>>>>>>> >>>>>>>>>> Qxf7+ and after the changes, the pawn ending is won. >>>>>>>>> >>>>>>>>>Crafty 18.11, Athlon 1GHz, has +0.8 after 0.4 seconds (8 ply), >>>>>>>>>rising to 0.93 after 11 seconds, 1.17 after 28 seconds, 1.36 >>>>>>>>>after 6 minutes. It plays Qxf7+ at all depths. >>>>>>>>> >>>>>>>>>-- >>>>>>>>>g >>>>>>>> >>>>>>>> It doesn't surprise me at all. Crafty is probably the best in the world >>>>>>>>evaluatiing pawn endgames. >>>>>>>> >>>>>>>> José C. >>>>>>> >>>>>>>Anyway, the position "as is" is not perfect for a test because white will want >>>>>>>to play Qxf7 even if doesn't understand the upcoming pawn endgame (because Qxf7 >>>>>>>just win a pawn). >>>>>> >>>>>>No >>>>>> >>>>>>Junior7 prefers to force a draw by perpetual check and >>>>>>cannot find Qxf7. >>>>>> >>>>>>If a program believes that black is better in the pawn endgame >>>>>>then it is not going to find Qxf7. >>>>> >>>>>If a program believes black is better in that position after the trades, >>>>>then it already has a serious endgame problem (lack of knowledge). Nothing >>>>>else matters in such cases... >>>>> >>>>> >>>>>> >>>>>> >>>>>> What is very good is the idea, so I try to improve the >>>>>>>original position in a way that the first move gives an idea that the program >>>>>>>understand the endgame. >>>>>>> >>>>>>>[D]8/4k1p1/1p2B2p/4p3/8/4P2P/1PP1KbP1/8 b - - am Bxe3; >>>>>>> >>>>>>>Crafty 17.14 avoids Bxe3 in an instant without any problem even in my K6-II 400 >>>>>>>mhz. How about other programs? The test is simple but I like it in the modified >>>>>>>version because it tells me fi a program understand a couple of concepts. >>>>>>>For instance, Gaviota wants to play Bxe3, So I know what to modify :-). >>>>>>> >>>>>>>Regards, >>>>>>>Miguel >>>>>> >>>>>>Deep Fritz also likes Bxe3 >>>>>>New position >>>>>>8/4k1p1/1p2B2p/4p3/8/4P2P/1PP1KbP1/8 b - - 0 1 >>>>>> >>>>>>Analysis by Deep Fritz: >>>>>> >>>>>>1...Kxe6-- >>>>>> ± (1.19) Depth: 1/3 00:00:00 >>>>>>1...Kxe6-- >>>>>> ± (1.19) Depth: 1/3 00:00:00 >>>>>>1...Bxe3! >>>>>> ² (0.44) Depth: 1/4 00:00:00 >>>>>>1...Bxe3! >>>>>> = (-0.13) Depth: 1/4 00:00:00 >>>>>>1...Bxe3 2.Bd5 >>>>>> = (-0.13) Depth: 2/5 00:00:00 >>>>>>1...Bxe3 2.Bd5 Kd6 >>>>>> = (-0.16) Depth: 3/6 00:00:00 >>>>>>1...Bxe3 2.Bf5 Kf6 3.Be4 >>>>>> = (-0.06) Depth: 4/7 00:00:00 >>>>>>1...Bxe3 2.Bd5 Kd6 3.c4 Bd4 >>>>>> = (-0.06) Depth: 5/8 00:00:00 1kN >>>>>>1...Bxe3 2.Bd5 Bc5 3.g3 g6 4.c4 >>>>>> = (-0.03) Depth: 6/10 00:00:00 3kN >>>>>>1...Bxe3 2.Bd5 Bc5 3.g3 g6 4.Be4 Kf6 >>>>>> = (-0.03) Depth: 7/11 00:00:00 7kN >>>>>>1...Bxe3 2.Bd5 Bc5 3.g3 g6 4.Be4 Kf6 5.c3 >>>>>> = (0.00) Depth: 8/12 00:00:00 16kN >>>>>>1...Bxe3 2.Bf5 Bc1 3.b4 Bb2 4.c4 Ba3 5.Be4 >>>>>> = (0.03) Depth: 9/15 00:00:00 50kN >>>>>>1...Bxe3 2.Bd5 Kd6 3.Be4 Bf4 4.b4 b5 5.Bd3 Kc6 6.c4 bxc4 >>>>>> = (0.09) Depth: 10/15 00:00:00 99kN >>>>>>1...Bxe3 2.Bf5 Bc1 3.b3 Bf4 4.Be4 Kd6 5.c4 Kc5 >>>>>> = (0.09) Depth: 11/17 00:00:00 231kN >>>>>>1...Bxe3 2.Bf5 Bc1 3.b3 Bf4 4.Be4 Kd6 5.b4 h5 6.g4 hxg4 7.hxg4 >>>>>> = (0.09) Depth: 12/19 00:00:01 483kN >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 h5 4.g4 g6 5.b4 h4 6.Ke4 g5 7.c5 >>>>>> = (0.19) Depth: 13/21 00:00:02 1039kN >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 Kd6 4.Ke4 h5 5.h4 >>>>>> = (0.25) Depth: 14/22 00:00:03 1611kN >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 Kd6 4.Ke4 Ke6 5.b3 h5 6.h4 >>>>>> = (0.25) Depth: 15/23 00:00:06 3076kN, tb=1 >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 Kd6 4.b4 Ke6 5.Ke4 >>>>>> ² (0.41) Depth: 16/24 00:00:12 5395kN, tb=6 >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 Kd6 4.b4 h5 5.g4 hxg4 6.hxg4 Kd7 7.Ke4 >>>>>> ² (0.50) Depth: 17/25 00:00:25 11346kN, tb=50 >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 Kd6 4.b4 h5 5.Ke4 Ke6 6.h4 Kd6 7.c5+ >>>>>> ² (0.56) Depth: 18/27 00:00:47 21088kN, tb=198 >>>>>>1...Bxe3 2.Kxe3 Kxe6 3.c4 h5 4.h4 Kf5 5.b4 Ke6 6.Ke4 Kf6 7.c5 >>>>>> ² (0.59) Depth: 19/30 00:01:54 49750kN, tb=993 >>>>>> >>>>>>(Blass, Tel-aviv 30.09.2001) >>>>>> >>>>>>Note that finding Kxe3 is also not an easy problem >>>>>>for part of the programs >>>>>> >>>>>>Deep Fritz has no problem but Junior7 cannot find it >>>>>> >>>>>>I remember that I told Amir Ban some years ago about >>>>>>pawn endgames problems and he told me that it is not important >>>>>>because Junior almost never does not get pawn endgames >>>>>>so it is more important for him to work about other problems >>>>>>in the evaluation. >>>>> >>>>>:) >>>>> >>>>>Funny comment. Which will change after playing a lot of GM players. :) >>>> >>>>I am not sure about it. >>>> >>>>I remmeber that Junior4.9 drew against 3 GM's at tournament time control >>>>at that time(1998 if I remember correctly). >>>> >>>>No game was pawn endgame at that time >>>> >>>>I also do not remember pawn endgames from the tournament in durtmond when Junior >>>>got 4.5 out of 9 against GM's so I doubt if GM's can practically often take >>>>advantage of the weakness of Junior. >>>> >>>>I did not watch GM's play against Junior in ICC but having a weakness does not >>>>mean that the GM's can often take advantage of it and the only proof is in games >>>>against Junior and not games against Crafty. >>>> >>>>Uri >>> >>> I believe Bob didn't mean 3 or 9 games, but hundreds. Let Junior play a lot of >>>games against GM's at ICC and, if it still has that weakness, they'll sure >>>discover it and win game after game. In 12 games against different players, the >>>cannot discover such a weakness, so it's not relevant. >>> >>> José C. >> >>I read that Slater claimed that Junior is good against humans >>See http://www.icdchess.com/forums/1/message.shtml?191120 >> >>It seems that it is not simple to take advantage of Junior's weaknesses. >> >>The GM's who played against Junior also could train at home against the >>commercial Junior,so the claim that they could not discover the weakness does >>not convince me. >> >>Uri > > And the GM's who played against Junior could also play a computer for the >first time in their lives. Both your and my statements are speculations. > The _facts_ are: 12 games against different opponents. Nothing more that I >know (if you know more facts, please tell us). With _those facts_, you cannot >conclude that the GM's did discover the weakness but failed to take advantage of >it. > Again, let a program with such weakness play hundreds of games against GM's on >ICC and you'll have facts. > > José C. I agree. And the _right_ way to do this is with an automatic interface. Then let the humans (IE Mecking) play 30 games in a row. And if he wins 15 endgames in a row, the problem is obvious. Many operators will play until the human "wises up" and starts to draw or win, and then they disconnect and come back another day. A program left on 24 hours a day, free to play as many games in a row as a human will play, is a _good_ testbed to see whether such a hypothesis (I won't improve endgame play because my program won't reach any endgames) is valid or not. I'll bet that in the case of Mecking it is not. Whomever cptnbluebear was, would be an even harder test. He was _very_ strong and just avoiding draws was a very difficult assignment against him. Many manually operated programs refused to play him for that reason... One had notes that said "for those of you that can draw at well against this program, go play somebody else..."
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.