Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: How I Learned to Stop Hating 141

Author: Stuart Cracraft

Date: 14:57:55 09/04/04

Go up one level in this thread


On September 03, 2004 at 16:54:42, Andrei Fortuna wrote:

>P.S. : I don't believe in auto-tuning for chess programs, except for that paper
>that I've seen in which they used a neural network to determine move generation
>order, but for eval I just don't see how to write such a bloody thing. No, I am
>in favour of manually tuning the eval and such.

It requires some type of mathematics/regression/operations-research that
a lot of us don't know.

Perhaps a book or two on regression would help?

Look at Andreas Nowatczyk of the Deep Thought team at CMU. He writes this
code (now at Tim Mann's WINBOARD/XBOARD site) that helps tune the evaluation
against master games.

Can anyone *explain* this code publically so that it can be understood?
I've looked at it and can't make heads nor tails of it. I just don't understand
the math. Is it a least-squares fit? How does that work for a large number
of coefficients that have to be tuned simultaneously, like in a chess program?
Any thought of adding a random perturbation?

I suppose this is a new threat but I am too lazy right now to start it.

Stuart



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.