Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: knowing when you've improved your evaluation function

Author: Dave Gomboc

Date: 13:57:53 04/16/04

Go up one level in this thread


On April 15, 2004 at 14:51:26, Anthony Cozzie wrote:

>As I understand, your paper goes like this (omitting background, proofs,
>implementation details, results, little things like that :)
>
>1. Get 650,000 positions from Chess Informant with GM evals (e.g. +-, =, etc).
>
>2. Divide positions into sets of +-, +=, =, etc.
>
>3. Tune eval such that all the positions in the set '+-' evaluate higher than
>the positions in the set '+=' which all evaluate higher than positions in the
>set '=' (etc) using gradient search.
>
>Close?
>
>anthony

Yes, that's an amusing summary. :-)  Of course in general the position groups
won't actually stratify (even nearly) completely -- that's just in the ideal
case (e.g. they could stratify completely if you were tuning a special
evaluation function for KRPKR, then you could figure out the threshold points
(e.g. ">= is win, < is not win"; ">= is not loss, < is loss").

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.