Author: Uri Blass
Date: 02:13:25 04/22/03
Go up one level in this thread
On April 22, 2003 at 04:30:14, Sune Fischer wrote: >On April 22, 2003 at 04:14:44, Uri Blass wrote: > >>On April 22, 2003 at 03:13:41, Sune Fischer wrote: >> >>>On April 21, 2003 at 18:24:45, Drexel,Michael wrote: >>> >>>>On April 21, 2003 at 18:09:03, Uri Blass wrote: >>>> >>>>>On April 21, 2003 at 17:18:46, Robin Smith wrote: >>>>> >>>>>>On April 20, 2003 at 17:20:11, Uri Blass wrote: >>>>>> >>>>>>>It may be possible to use it in a productive way to avoid problems >>>>>>>that I mentioned(the program prefers loss by KR vs KPPP and not drawn KR vs KPP) >>>>>>>by having rules to trust tablebases scores only in part of the cases but I do >>>>>>>not like it. >>>>>> >>>>>>Uri, >>>>>> >>>>>>I know such things are possible in theory. Can you give an example of it >>>>>>actually happening? >>>>>> >>>>>>Robin >>>>> >>>>>No >>>>>If I remember correctly I saw a case when it happened but >>>>>I did not care to save the position and I do not plan >>>>>to look for it now. >>>>> >>>>>Uri >>>> >>>>[D] 8/8/8/7P/4K1P1/r6P/k7/8 b - - 0 1 >>>> >>>> >>>>Anaconda 1.0 lost this position with tablebases. >>>>It would take the pawn without. >>> >>>It's a fluke. >>> >>>I think what happens here is that table bases tell the engine it has a draw if >>>it wants it, however the engine is "stupid" and believes it is better and will >>>not be satisfied with a draw. >>> >>>So it's a simple horizon problem as far as I can tell, the TBs provide >>>"infinite" depth but the engine does not like the truth it sees and prefers to >>>kid itself. Without TBs it doesn't know it choosing a drawing line, so it >>>happily eats the pawn. >>> >>>If the engine was a little better it would know the draw was the best option >>>anyway. I don't know about Anaconda, but Ruffian and Frenzee both get a draw >>>score showing pretty fast with TBs, so maybe this is related to a bug in >>>Anaconda? >> >>I think that it is related to the knowledge that frenzee and Ruffian has and >>ankadota and movei have not. >>Movei does not use tablebases so it plays the right move but >>Movei has a negative score for white after Ra4+ even after a long time and I >>hope that it will be changed after adding score of pawns relative to kings. > >Frenzee doesn't have any knowledge of this, but it also chooses the wrong move >Ra4+, and so does Yace btw. I thought that it choose Rxh3 based on your previous post. The point is that tablebases can cause the program to choose the wrong move and I got the impression that frenzee does not choose the wrong move based on your previous post. I see now that I was wrong and tablebases also cause frenzee to choose the wrong move. The score is not important because it is possible that frenzee can find with tablebases that white can force a draw after the winning move. It's a little strange that they both show a draw >score for a non-drawing move, apparently they are content with a draw, so I'm >not sure why they don't chose Rxh4 immediatly. I think that it is clear. Tablebases tell frenzee that Rxh4 is a draw so it prefers another move at small depthes. At big depthes the score of the second move go down to draw because frenzee can find that white can get a draw by sacrificing a pawn but frenzee does not change it's mind because a draw by tablebases in the next move is not considered as better than the draw that it already has for Rxh4 >Ruffian must have special knowledge so it is a bad example, it shows a draw >score from ply 1 :) I think that intelligent engines should also see it by knowledge. > >>The point is that movei likes it's king in the centre in the middle game so it >>is not very happy with king supporting the passed pawns and 3 passed pawns are >>evaluated as clearly less than a rook and I cannot say that this knowledge is >>generally wrong because black can win the game if I change the place of the >>kings. >> >>I guess that frenzee see positive score for white after Ra4+ thanks to >>positional knowledge(king relative to the pawns) or thanks to overevaluating >>passed pawns(in case that it has not that knowledge) and not because of seeing >>the queen in the board. > >I have no idea why it's showing a draw score for Ra4+, but Yace does the same >and takes several minutes and 17 plies to fail low. > >Apparently it's tough to find the right move for the right reason, it is easier >to find it for the wrong reasons. I'd still call it a fluke though. > >Yace has now switched to Ka2b2 with a draw score, Ruffian agrees that that is >also a drawing move. I do not think that it is luck. The point is that without tablebases the program is better in giving bigger scores for better positions. Correct evaluation is not only about giving correct scores but also about giving better scores for better moves. 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.