Author: Aaron Tay
Date: 11:54:58 10/28/02
Go up one level in this thread
On October 27, 2002 at 18:41:22, José Carlos wrote: >On October 27, 2002 at 18:25:32, Ingo Lindam wrote: > >>On October 27, 2002 at 18:01:30, José Carlos wrote: >> >>> There's a difference between: >>> a) my knowledge doesn't fit the position (ie no open lines where I have a >>>bonus for open lines). >>> b) the sum of weights is zero (I have one open line and my opponent has one >>>open line). >>> >>> In the first case, you're right that zero eval is misleading and dangerous. >>> In the second, the position is probably balanced (if this applies to many >>>parameters, of course, not just one), so the zero is correct. >> >> >>You are right about that difference... but I want to give you a realistic >>example for the programm evaluate near 0 for the few things he can >>evaluate...and another 0 for all the things it doesn't know anything about. >>My problem still isn't solved. You wouldn't play towards a position you just >>know it is ballanced in the aspect of open files, but in all the other aspects >>you can't judge it at all, would you? > > I know what you mean, but I'm afraid I wasn't clear myself. > Suppose I'm a program who has 50 eval terms (open lines, knights in the center >of the board, doubled pawns, pawn protection around the king...). > I look for those 50 patterns on the board: > a) I find them all there. The final sum is near zero. In this case, I can only >conclude that the position is balanced. > b) I only find 10 and they seem to be balanced. In this case, your idea is >totally correct: I might want to go to that position thinking it's drawn, but I >would be going into an unknown (for me) position. Why the cutoff at 10? > If b happens, the program should be able to assess a negative score for >itself, but: b happens very seldom, if ever; if a happens it is possible that >the program is simply not understanding the position, but that might be true no >matter what the score is. It also seems to me that it might not understand the position, but in fact might even be winning in a few moves! It would be sheer irony, for the program to avoid those positions. > So my point is that your idea is correct, but extremely difficult to detect. Maybe it can be studied, by logging positions where such occurances happen, and study the positions after and see whether it is true, the theory that such positions are bad. > José C.
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.