Author: Robert Hyatt
Date: 11:43:10 10/27/99
Go up one level in this thread
On October 27, 1999 at 11:36:25, Andrew Williams wrote: >On October 27, 1999 at 11:07:43, Robert Hyatt wrote: > >>On October 27, 1999 at 06:25:31, Andrew Williams wrote: >> >>>On October 27, 1999 at 02:59:52, James Robertson wrote: >>> >>>>I'm trying to write some code to evaluate passed pawns. Specifically, I want to >>>>know how people score pawns as they advance toward the 8th rank. How much bonus >>>>do you give each square the pawn advances? Does this increase with connected >>>>passers? I think Bob mentioned that if the King is defending the pawn or it is >>>>blocked by an enemy piece certian bonuses/penalties are given. What do other >>>>people do about this? >>>> >>>>Any help is appreciated. If I can get my program's passed pawn code in before >>>>Thursday when I leave for the Dutch Open, I will be very happy. :) >>>> >>>>James >>> >>>James, >>> >>>PostModernist considers: >>> >>>How far away from promotion is the pawn? >>> >>>Is the Pawn unblocked? >>> >>>Is the Pawn touching another friendly passer? >>> >>>Is there another friendly passer on an adjacent file (but not touching)? >>> >>>Is it an outside Pawn with very low enemy material? (This is very >>>unsophisticated at the moment in PM) >>> >>>Is it guaranteed to promote in a KPK endgame? (Rather crude, but reasonably >>>effective). >>> >>>A long time ago, I got PM to analyze thousands of games (TWIC?) to give me some >>>idea what these scores should be. Basically, I went through these games, looking >>>for a Passed Pawn. I then "evaluated" the PP, deciding which of PM's factors >>>matched the Pawn. By considering whether the game ended in a win, draw or loss >>>I generated probabilities, which I converted into scores for each factor. I fed >>>these scores into my program but they were no better than simply picking scores >>>and tuning them. The approach didn't work well I think, because in GM games, if >>>a decisive passed pawn is going to score, it rarely reaches even the seventh >>>rank; the GMs both see the outcome many moves before and the loser resigns. I >>>guess that apart from time-trouble cases, you generally see passers on the 7th >>>only if both sides have them and both will promote. I didn't factor that into >>>the assessment I was doing. Still, it was an interesting experiment. >>> >>>I've since changed my PP eval considerably, so the factors I now use are >>>somewhat different to then. >>> >>>Andrew >> >> >>beware of that "touching a friendly passer"code. IE set up a position with >>the kings wherever you want, and white with passers at d2/e2, black with passers >>at a7 and h7. Suddenly the connected passers lose in a big way. With pieces to >>help stop those 'split passers' this is a good way to evaluate, but you _must_ >>know enough to not trade your last piece (and hopefully it isn't a knight). > >That's a good tip. Thanks. I set up the position with White to move and it >saw the impending doom after about 5 seconds. But of course it won't get that >if the pos appears at the bottom of the tree. And what's worse, it thinks it >is a good idea to swap off a pair of Bishops if I introduce them. Yuk. > >Andrew that was an obvious one for explanation. try white pawns at d2/e2, black pawns at a7 and c7. This is a classic zug problem, because black will blockade the white pawns, and white finds himself strangely unable to capture either of the black pawns because of something like pa5/c5, and if white plays Ka3, c4 makes the king retreat (kxa5 c3 and it is over). The further apart the pawns are, the easier it is to win. But even separated by 1 file as above it is a trivial exercise. Roman watched "cptnbluebear" win several games with this theme, and pointed out the error. I fixed it and it hasn't lost like that yet.
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.