Computer Chess Club Archives


Search

Terms

Messages

Subject: Constructive Testing Was Re: Root move ordering - an experiment

Author: Brian Richardson

Date: 15:14:41 09/28/04

Go up one level in this thread


I do a combination of testing these days.

First, I have hard coded in the program about a half dozen positions that
I "know" what Tinker's performance and results should look like.

If nothing there seems broken, then I run WAC, but at a maximum
of 60 seconds per position, but it only takes about a half an hour.
How?  I use Robert's suggestion.  In my case, starting after ply 6
for each position being searched, if the correct move is held for
3 ply, I stop early.  The majority of positions only take a few seconds.

On the topic of WAC, I have a pawn position and PCSQ and material only
evaluation version of Tinker.  It scores 300 on WAC (a few positions need more
than about 30 seconds), but plays poorly vs the "normal" Tinker evaluation,
or other opponents in real games.  You can also keep track of the total
WAC solved time and as someone else posted, check the actual positions not
solved.  More often than I would like, versions x and y both score say 296, but
the 4 unsolved are different!

I also put Tinker up on ICC where it gets quite a few games.
Yes, I pay for this service (actually for my human account), but
Tinker is entertainment for me.  Tinker is also on FICS from time to time, but
it gets about 10x more games on ICC.  I have practically no restrictions in the
formula so Tinker will play just about anybody.  I have also taken time to add a
reasonable accept/offer draw option, and resign code to make it more human
friendly.  Yes, whenever Tinker's rating creeps up, some computer accounts rated
400+ points higher come along and take it back down.  I don't care.  I look for
games against good opponents, and in particular, losses against weaker ones.

I also run Nunn position gauntlets against about a dozen other engines.
This takes a long time (2 days or so) at 15 10 time controls.  I have
tried faster times, and the results are highly non-deterministic.

Finally, I am thinking about the ICCA piece about random value evaluation
and its use in testing, along with temporal difference learning.

I have enjoyed your progress and it has gotten me to retry several things again.
Please keep at it.
Brian



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.