A very simple question: since Eval and search are very strictly related what is your strategy when you try to tune Eval?
Tune the Eval with a simplified search (without pruning lmr and futility, maybe with null move pruning), or with a normal search algorithm?
The question is simple. Maybe the answer is not so simple.
Thank you for the reply
Eval tuning question
Moderators: hgm, Rebel, chrisw
-
Aleks Peshkov
- Posts: 892
- Joined: Sun Nov 19, 2006 9:16 pm
- Location: Russia
Re: Eval tuning question
AFAIK Rybka's author was the first who succeeded in tuning static evaluation in ultra-fast controls. When average search depth is very low search matters less.
-
tpetzke
- Posts: 686
- Joined: Thu Mar 03, 2011 4:57 pm
- Location: Germany
Re: Eval tuning question
Hi Marco,
I tune eval with full search enabled. Otherwise the tuner will produce values that don't harmonize with the pruning or reduction margins. You can of course say you tune first eval with a simpler search, which will give you a weaker engine (most likely) and then tune search to find new margins that fit to your new eval.
This can work but if it fails you wasted twice the CPU power. If you have enough CPU power it is worth a try but I have to be very careful where I spend mine.
I tune eval with full search enabled. Otherwise the tuner will produce values that don't harmonize with the pruning or reduction margins. You can of course say you tune first eval with a simpler search, which will give you a weaker engine (most likely) and then tune search to find new margins that fit to your new eval.
This can work but if it fails you wasted twice the CPU power. If you have enough CPU power it is worth a try but I have to be very careful where I spend mine.
-
jdart
- Posts: 4367
- Joined: Fri Mar 10, 2006 5:23 am
- Location: http://www.arasanchess.org
Re: Eval tuning question
Most people are using a regular search with all features enabled.
But I think theoretically some pruning techniques may be problematic when tuning, especially LMP, which just starts tossing out possible moves based on move count. That makes the search result not depend strictly on the eval function. However, I have not tried turning this off when tuning.
--Jon
But I think theoretically some pruning techniques may be problematic when tuning, especially LMP, which just starts tossing out possible moves based on move count. That makes the search result not depend strictly on the eval function. However, I have not tried turning this off when tuning.
--Jon