Author: Bob Durrett
Date: 18:43:50 11/14/00
Go up one level in this thread
On November 14, 2000 at 18:02:17, Severi Salminen wrote: >>I am working on my evaluation function. As many other ones, it includes material >>score and many bonus / penalty scores. I am considering how big the total of >>bonous and penalty scores should be, comparing with the material score: 1 pawn, >>1.5 pawn, 1 queen or bigger? Any comments would be appreciated. > >Hi! > >There is no one single answer to this. Every programmer has allways tried to >find the right weightings. Is a rook on open file worth a pawn? Maybe, maybe >not. Are doubled isolated pawns a pawn advantage to opponent? Maybe, maybe not. >You could try to make your program to play against itself so that one side has >one weighting and the other side has other weighting and see what happens. Try >to alter one evaluation term at a time - this way you'll know what caused the >difference...maybe. > >Severi If the programmer is a chess amateur [like myself], then it might help a LOT to solicit the help of the strongest chessplayer he/she can find to answer these "purely chess" quesions [as opposed to "purely programming" questions.] If you could be so lucky as to get a GM to help, you would have a great benefit. This would not eliminate the need for the type of experimentation suggested above, but would give you a huge head start, and surely would speed up your progress toward a competitive program. To use the cited example: A very strong chessplayer would be able to tell you, in a highly specific and definitive way, how to tell when doubled isolated pawns are a pawn advantage to the opponent and when they are not. You might be able to use that information in the design of your evaluation function(s). He/she could even go further by telling you how to come up with the appropriate plans for each side. Of course, if your computer relies on "purely brute force," maybe that knowledge might not help you. Or, maybe it would help. I would be very interested in knowing the extent to which such planning information is or is not utilized by the strongest programs. Do the top programs incorporate such knowledge into the search strategy or into the evaluation functions or in some other way into the operation of the program? Hypothetically, [in a futuristic program] one could evaluate a position by first identifying the correct plans for each side which are inherent in the position. The evaluation [which could be time-consuming] could then determine the liklihood of achieving those plans. The finding might then be converted into a numerical evaluation value. I doubt [but don't know] that any present-day programs utilize such a sophisticated evaluation function, based on what I have read. My perception is that most programs utilize relatively simple evaluation functions in order to save time. Certainly, to create a sophisticated complex evaluation function [or routine] would require that the programmer be very knowledgable about "purely chess" matters, or have some very good help.
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.