Author: Jay Scott
Date: 14:20:03 03/30/04
Go up one level in this thread
On March 29, 2004 at 19:03:22, Uri Blass wrote: >On March 29, 2004 at 18:52:52, Jay Scott wrote: > >>Chess programmers, I've found, are remarkably resistant to changing how they do >>certain traditional things. For example, every time I propose calibrating >>evaluation not in millipawns but by some standard with a sounder theoretical >>motivation, I'm shouted down by everyone who does not ignore me. I don't >>understand that. > >I do not know what you mean but if you mean to evaluation in terms of expected >result then >I am not against evaluation not by pawns but by expected result. Yes, that's an example. >The problem is that it is not easy for human to think about defining the >evaluation when it is in terms of expected result. Hmm, some humans, no doubt. Not the theoretically-minded, who are apparently a small minority. But especially not when the evaluation is tuned by machine learning, or even by a human using formal methods (imagine a test suite that estimates how accurate each factor in the evaluator is). Then a sound theoretical basis behind it all can help us come up with methods that are more powerful. This is especially true when it comes to interactions between evaluation features. Most chess programmers, apparently, find it easier or more satisfying to eyeball their results and tune by the seat of the pants. Jay
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.