Computer Chess Club Archives




Subject: Re: evaluationfunction tuning

Author: Andrew Williams

Date: 14:14:29 09/07/03

Go up one level in this thread

On September 07, 2003 at 11:35:48, JW de Kort wrote:

>Hi friends,
>I have recently spend a lot of time improving my evaluationfunction. I have
>added a lot of features and studied a number of sources (Crafy, Gerbil, TSCP,
>Phalanx, Beowulf etc.). Still my engine is not able to beat TCSP 181 in all
>games. Now i want to fine tune the evaluation features. Can anybody give me some
>advice how to ga ahead with this process? Is there a EPD set available of good
>testing positions for this purpose?
>Thnak in advance.
>Jan Willem

There are lots of ways to go about this. First of all, you should make sure that
your program is not being outsearched by TSCP. If TSCP is consistently searching
1 ply more than your program, you will want to look at your search control
first. Once you're happy that your search is comparable to TSCP's search, you
will want to make absolutely sure that you're not doing any pruning which is
making silly mistakes.

Assuming that you have a sensible search and that it's not significantly worse
than TSCP's search, you might gain something from working on the evaluation. I
would suggest playing a large number of long-ish games against TSCP. Then focus
on the defeats in the set. Play them over a few times and make a note of where
TSCP gets the upper hand. Play through them all first before trying to change
your program at all. If you find a common theme, it's easy to focus on that. Try
to get some test positions from your analysis; I find it's better to use test
positions that come from lost games.

If you find that you've made some changes, try another match with the same
conditions as the previous one. Unless you have some way of ensuring that you
always start with the same opening positions, I would not bother too much about
the score of the match; just focus again on the losses. Once again, make sure
you go through *all* the losses before making any changes. Try to find common
themes in the defeats. If there isn't a common theme, try working on the game
that you personally find the most embarassingly obvious. Try to fix it and

There are of course automatic pruning approaches, but I've never seriously tried
any of them. Try googling for "Temporal Difference Learning" with "knightcap".


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.