Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Parameter Tuning

Author: jonathan Baxter

Date: 15:11:11 10/02/98

Go up one level in this thread


On October 02, 1998 at 10:48:46, James Robertson wrote:

>On October 01, 1998 at 20:28:55, jonathan Baxter wrote:
>
>[snip]
>>
>>Currently I am working on an improved version of KnightCap, with better
>>search and a much *simpler* evaluation function. Interesting preliminary results
>>show that with parameter learning and *just* piece/square tables we can get
>>a rating of around 2500 on ICC. This is without any opening learning (the
>>original version of KnightCap has an opening learning algorithm that is also
>>described in the above paper.)
>[snip]
>
>I don't think just piece/square tables are good enough for a program at all,
>even with parameter learning. For instance, how does it avoid making stupid
>doublings and isolations of it's pawns? My program without doubling and
>isolation penalties makes very bad choices about pawn structure, often
>sufficient to lose. Obviously your program is good enough tactically to win most
>games, but I think that it will do much worse against human opposition without
>the extra knowledge.

This performance id against all-comers on ICC, not just computers. But your
point is perfectly valid---higher order pawn features cannot be represented.
What surprised me was how well you can do (in particular against humans) with
just piece/square tables. It seems that tactics counts for an *awful* lot in
chess.

Cheers,

Jonathan Baxter



This page took 0.02 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.