Author: Robert Hyatt
Date: 08:02:55 09/25/98
Go up one level in this thread
On September 25, 1998 at 05:17:49, Alessandro Damiani wrote: >On September 24, 1998 at 11:43:38, Robert Hyatt wrote: > >>In the ICCA Journal, Vol. 16, No. 4, GM Timoshchenko is comparing various >>chess programs on how they evaluate bishops vs knights. He gives several >>positions, with this one being the most interesting at present: >> >> >> +---+---+---+---+---+---+---+---+ >> 8 | | | | | | | *K| | >> +---+---+---+---+---+---+---+---+ >> 7 | *P| *B| | | *B| *P| *P| | >> +---+---+---+---+---+---+---+---+ >> 6 | | | | | *P| | | *P| >> +---+---+---+---+---+---+---+---+ >> 5 | | | | | | | | | >> +---+---+---+---+---+---+---+---+ >> 4 | | | | | | | | | >> +---+---+---+---+---+---+---+---+ >> 3 | | | N | | P | N | | P | >> +---+---+---+---+---+---+---+---+ >> 2 | P | | | | | P | P | | >> +---+---+---+---+---+---+---+---+ >> 1 | | | | | | | K | | >> +---+---+---+---+---+---+---+---+ >> a b c d e f g h >> >> >>fen: 6k/pb2bpp1/4p2p///2N1PN1P/P4PP/6K/ w - - >> >>Here is the gist... to feed this to a program and with no searching, >>see how it evaluates this. Chessmaster 2100 says -.26, Fritz-2 says -.47, >>and chess-genius 1.0 says -.48. I recently fixed two "holes" in Crafty's >>evaluation, one concerning "if you have one bishop in the ending, get your >>pawns on the *other* color squares" and the other "if in an endgame, with >>pawns on both wings, a bishop is significantly better than a knight." >> >>This last influences the evaluation of the above position quite markedly, >>because this is an endgame, and black is going to get the bishop-over- >>knight in endgame bonus two times since black has two bishops, white has >>none. Black also gets the traditional bishop-pair bonus as well, which >>might or might not be overkill here. In any case, I first gave this to a >>GM to look at and the response "black wins easily." Not surprising, the >>bishop pair vs the pair of knights should favor the bishops. I then told >>him that Crafty "statically" evaluates this position as "-.93" which is >>the raw score Evaluate() returns here. He said "hmmm... ". And we had >>an interesting conversation without concluding whether this is too large, >>too small, or "just right." >> > >I think, when adding knowledge to the current evaluator, we should not think >like humans: the program has only to know that one position is better than an >another. If the score of a position is higher than the score of another position >it doesn't matter how big the difference is, since the move choice doesn't >change. > >So, if we see the score "-.93" it says only that one side seems to be winning, >but nothing more. The program needs only to know *that* this position is better >than others. > >Ciao > >Alessandro For shallow searches you are right. But for deep searches, you are faced with choosing between "combinations" of things. IE the above B vs N bonus plus a rook on an open file, vs an opponents outpost knight and passed pawn. You had better have the weights ratio between each other right or you make the wrong decision...
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.