Author: Robert Hyatt
Date: 05:57:16 08/18/98
Go up one level in this thread
On August 18, 1998 at 05:18:16, Robert Hyatt wrote: >On August 17, 1998 at 21:17:59, Bruce Moreland wrote: > >> >>On August 17, 1998 at 07:41:52, Robert Henry Durrett wrote: >> >>>Is there something about chess computers/software that requires modification of >>>the values for human vs human so that the software will perform properly in >>>computer vs human and computer vs computer games? >>> >>>If so, what is it about the computer software that makes this necessary? >> >>There shouldn't be, since the programs are playing chess better and better, they >>should find the same rules of thumb work as work with humans. >> >>I think it might be useful to fudge occasionally though. You can take a >>position that a person would say is even, or perhaps worse for white, and yet a >>computer playing black against a human will be worse off. So perhaps you try to >>teach the computer to avoid these positions even if they are objectively better. >> >>The same could be true of material values, if a program is ham-handed about >>handling a bishop and a knight, maybe you would teach it to prefer a rook and a >>pawn. >> >>bruce > > >I think the most common reason for fudging the values is to avoid the BN for RP >trade... or N for PPP trade... I've done that for years myself, trying to just >offset positional benefits by increasing the value of minor pieces. I finally >got tired of it and added a simple bit of evaluation code that simply says if >one side has a rook, the other side has two minor pieces, the two minors get an >additional positional bonus (or, if you think about it, the rook gets a >positional penalty)... I do the same for a piece for 3 pawns... although the >three pawns may well make up the "penalty" if they are connected, and passed, >and advanced far enough that they might cause problems before the extra piece >can be used to pile up on them... I should add that I was somewhat embarassed that a few simple lines of code could fix something that had caused me grief for years... make the piece value too high and it screws up other things... make it too low and you trade it for a pair of pawns + compensation. I finally realized that when Crafty (or Cray Blitz in years gone by) did this, I just said "ugh, two minors for rook and pawn is almost always bad." Or "ugh, a minor for three pawns is almost always bad, especially earlier in the game." Once you think about it, this can become a simple evaluation term and then you don't have to dink around with piece values... and adjust them as you adjust your positional scores to keep things in sync...
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.