Author: Amir Ban
Date: 15:09:58 02/25/98
Go up one level in this thread
On February 25, 1998 at 16:54:55, Jay Scott wrote: > >On February 25, 1998 at 05:09:41, Amir Ban wrote: > >>We all know what the BEST evaluation is. It's the one coming out of >>perfect knowledge of the game. But what is good evaluation ? More >>precisely, given two evaluation functions, how do you decide which is >>better ? > >The question "which one is better?" is meaningless by itself. Better >for what? > Better for Computer Chess. Since this is the name of the newsgroup, I thought I don't need to say that. >Label the set of all chess positions with the positions' game- >theoretic values, win=1 loss=0.5 draw=0. One thing you might like >your evaluation function to do is minimize, say, mean squared >error. You might like a good statistical fit to the truth, in >other words. > >But it's easy to construct an evaluation which has an excellent >statistical fit yet plays bad chess. For example, imagine an >evaluator which is perfect except that it thinks that white is >winning after 1. g4 e5 2. f4. The mean squared error is >negligible, since only a few positions are wrong, but a program >that relies on it is going to lose a lot of Fool's Mates. > >A program could also play perfectly with an evaluation which >had a poor statistical fit. All that's necessary is for one >of the optimal moves to be evaluated highest in any position >that the program can reach with optimal play up to that point. > Right. I'm not so much interested in ideas that are dismissed in the next sentence. I am looking for a formulation which at least someone considers to be correct. The fact that it's not easy to find one is surprising in itself. It shows that we don't have a good idea of what what we are doing when we are busy "improving" our evaluation". Amir
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.