Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: testing static eval

Author: Tord Romstad

Date: 06:25:34 02/16/04

Go up one level in this thread


On February 16, 2004 at 08:51:50, Andrew Wagner wrote:

>On February 16, 2004 at 08:47:02, Matthias Gemuh wrote:
>
>>
>>
>>Evaluate mirror images of boards to see that scores don't change.
>>
>>/Matthias
>
>Great idea! Are there any particular positions or kinds of positions that are
>good to test? I'm thinking of something like
>http://homepages.caverock.net.nz/~peter/perft.htm ...except for eval() and not
>Move_Gen().

It is best to test with a huge set of randomly chosen positions from all
phases of the game.  In Gothmog, I have some code which reads an EPD file,
walks through all positions, and evaluates each position as well as its
mirror image.  Whenever the scores differ, the position is printed to a log
file.  Studying the log file helps you find asymmetry bugs.

Such tests should be done very frequently, even when you only make very
innocent-looking changes to the evaluation function.  It is surprising
how easy it is to introduce new asymmetry bugs.  I made the mistake of
doing some last-minute eval tweaks in my engine before I released the
last public version of my engine, and forgot to run the test.  Of course,
this resulted in a new asymmetry bug, which is there in the public version.

:-(

Tord






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.