Author: Dave Gomboc
Date: 16:35:47 07/25/00
Go up one level in this thread
On July 25, 2000 at 19:30:28, Tom Kerrigan wrote: >On July 25, 2000 at 19:08:57, Dann Corbit wrote: > >>On July 25, 2000 at 18:57:44, Tom Kerrigan wrote: >>[snip] >>>Throwing away what information? Ideally, chess positions are evaluated to one of >>>three values. If you have any more information than that, you can be sure it's >>>flawed. >> >>Really? You only score positions as +/-/0? >>Then how do you choose one when there are more than a single positive >>alternative? > >I didn't say that I only score positions as +/-/0. That would be ideal, but it's >just not practical. I use a flawed evaluation function, just like everybody >else. :) > >> >>>If I aim to duplicate Crafty's evaluation function scores, then I will end up >>>recreating Crafty's evaluation function. That's not my intention. >> >>I think you will find that for a given depth of evaluation, correct evaluation >>functions generally agree to within a pawn or better. >> >>>I would like to be able to tweak one of my evaluation function weights and see >>>if it helps the function predict the outcome of the game better. >> >>If your evaluation function misses the mark by several pawns, it won't help >>predict the outcome of the game better. > >If my evaluation function gave positive scores to winning positions, it would >win all of its games, period. If the positive scores were random, it would >probably win in stupid ways, but it would still win. > >-Tom Given that most annotated games use a few standard symbols for assessment (which I can't reproduce literally here, because the web board doesn't support unicode :-), but here's an approximation: + - +- += = =+ -+ - + ) I think it is reasonable to take such positions out of a massive game collection and do regression testing. There's bound to be a fair bit of noise, but that's unavoidable in any case, I'd imagine. Dave
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.