Computer Chess Club Archives




Subject: Re: Automatic Eval Tuning

Author: Landon Rabern

Date: 10:27:57 06/29/01

Go up one level in this thread

On June 29, 2001 at 11:18:48, Vincent Diepeveen wrote:

>On June 29, 2001 at 11:14:34, Artem Pyatakov wrote:
>>I am curious, have people here experimented or extensively used Eval Function
>>tuning based on GM games for example?
>>If so, is it effective to any extent?
>>I came across this page and it seemed kind of interesting:
>>Have others tried this too?
>>Thanks in advance.
>yes i tried some years ago automatic tuning.
>The bigger your evaluation is, the more problematic tuning it automatic
>is. Also automatic tuners don't have any chess knowledge, so they
>don't see the difference between tuning passed pawns negative if you happen
>to have a testset where a passer is bad now and then.
>Another problem for automatic tuners is that you tune for testposition set X,
>but that in reality it has to work well also for testset Y where it has
>not been tuned for.
>Evaluations hand tuned take into account testset Y, not only testset X.
>Anyway, when your number of parameters gets quite a big number then
>automatic tuning doesn't work anyway anymore.
>Of course it might beat random chosen parameters, but it'll never beat
>hand chosen parameters (unless a fool choses them).
>Best regards,

You are assuming that all you can do is supervised learning over a data set.
The method that shows the most promise is Reinforcement learning.  This allows
the learner to continuosly learn, if there is a hole in the evaluation it will
get fixed, because otherwise the program will lose.  You might want to try using
something like Q-learning or TD(lambda).  It might take a long time to get good
values from scratch, but you might have more success if you start from your
original hand coded numbers.



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.