Author: Uri Blass
Date: 19:01:10 04/12/03
Go up one level in this thread
On April 12, 2003 at 21:45:13, Koundinya Veluri wrote: >On April 12, 2003 at 19:58:09, Uri Blass wrote: > >>On April 12, 2003 at 19:45:22, Koundinya Veluri wrote: >> >>>On April 12, 2003 at 06:13:05, Uri Blass wrote: >>> >>>>On April 12, 2003 at 05:46:08, Jeremiah Penery wrote: >>>> >>>>>On April 12, 2003 at 04:22:57, Uri Blass wrote: >>>>> >>>>>>On April 12, 2003 at 01:44:26, Robert Hyatt wrote: >>>>>> >>>>>>>That has to be part of the evaluation. IE you have to know that you can >>>>>>>give the pawn up if your king is closer to the remaining pawns than the >>>>>>>opposing king is... >>>>>>> >>>>>>>I do that obviously... >>>>>> >>>>>>This has nothing to do with the pawn. >>>>>> >>>>>>You have to evaluate correctly the following position that can happen >>>>>>if you do not search deep enough >>>>>> >>>>>>[D]8/8/1K6/5p1p/4kP1P/6P1/8/8 w - - 0 6 >>>>>> >>>>>>I hope that movei will be able to see it after I add some knolwedge but the >>>>>>knowledge that is needed is not about passed pawns because there are no passed >>>>>>pawns in that position. >>>>> >>>>>Bob never said anything about passed pawns. >>>> >>>>He did in the post that started this thread: >>>> >>>>http://www.talkchess.com/forums/1/message.html?292975 >>>> >>>>"This seems to be an example of an engine that misses the power of the "distant >>>>passed pawn". >>>> >>>>I agree that a lot of engines have problems in the evaluation but the problem >>>>is about not evaluating correctly king relative to the pawns and has nothing to >>>>do with evaluation of the "distant passed pawns". >>>> >>>>Uri >>> >>>It may take several moves for the white king to capture the passed pawn in some >>>variations, so if the search can't see the capture from the initial position, >>>then the "king relative to pawns" evaluation might not be sufficient to solve >>>this. After the capture is made, the search can usually see the rest faily >>>easily so I think the "distant passed pawns" evaluation is more important to >>>solve these type of positions. >>> >>>Koundinya >> >>I disagree. >> >>A program without knowledge about distant passed pawns(but with some small bonus >>for passed pawns) and with knowledge about king relative to pawns will >>have no problem in that position. >> >>The *only* reason that a program with a bonus for passed pawns can fail here is >>lack of knowledge about king relative to the pawns. >> >>There may be other positions when knowledge about distant passed pawns is >>important but not the position that was posted. >> >>The point is that white has equality without trading rooks and can capture the >>black passed pawn without trading the rooks so even a small bonus for passed >>pawns is enough to avoid trading rooks unless the search can see that white can >>win a pawn after trading rooks and it is exactly what happens to program that >>trade the rooks. >> >>Uri > >For a program to see that Rd4 loses, it must also see the reply Rxd4. So >consider the position after the rooks have been captured. > >8/8/1p4k1/5p1p/3K1P1P/6P1/8/8 b - - 0 2 > >For the "king relative to pawns" evaluation to come into play, the black king >has to get closer to the white pawns than the white king. This doesn't happen >until the white king is pulled away by the passed pawn. Yes but if the pawn is not captured the program in the search the program has no reason to let the rook trade. simple evaluation that give bonus to every passed pawn is enough. This takes approximately >11 plies. Until that point, the "king relative to pawns" code won't even come >into play. 11 plies is a small depth in an endgame and program may get it even at blitz time control. If the program does not see the capture of the pawn then it is going to be right in giving black the advantage because of the passed pawn. Additionally, when the rooks are on the board, the program won't >search as deeply as after the trade. So from the initial position I think the >"distant passed pawn" knowledge is more important here. Of course you will >probably need both pieces of knowledge for the evaluation to be stable between >plies, but the "king relative to pawns" evaluation only may not be sufficient if >the program isn't a deep searcher. I do agree though that the evaluation you >suggested is also necessary. > >Koundinya I agree that evaluation of distant passed pawn is necessary but the position was not a good example because program with only small bonus for passed pawns and knowledge to evaluate pawns relative to the king can also solve it. It is possible to have a case when evaluation of distant passed pawn is important but I think that a case when both sides have passed pawns or a case that you need to sacrifice a pawn in order to create the distant passed pawn can be better. 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.