Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Limiting the number of qeval nodes

Author: Andrew Williams

Date: 14:08:23 08/04/99

Go up one level in this thread


On August 03, 1999 at 19:04:06, Scott Gasch wrote:

>>(1) Before you generate any moves at all, you call your evaluate() routine,
>>correct? If the score you get is >= beta, don't bother to search this position,
>>just return beta. (because whoseturn is already doing fine in this position,
>>and doesn't need to capture anything).
>
>Yes, I am doing this.
>
>>(2) MAKE SURE YOUR MOVE ORDERING IS SENSIBLE (This is of CRITICAL importance in
>
>I am using MVV-LVA and sorting the moves.  I also am using a history table to
>increase the value of moves that caused a cutoff earlier.
>
>>(3) Have you got a static exchange evaluator? If so, just chuck away any
>>captures that (in the opinion of the SEE) lose material (because static
>>exchange evaluators are faster than searching).
>
>What exactly is a SEE?  Does this code play out one trade fully and return a
>score based on it?  How does it work?

What the SEE does is to list all the pieces that are attacking a particular
piece. It then orders these pieces so that least valuable ones are considered
first. Then it "plays through" all the possible captures on the square, to see
whether the side on move can safely capture the piece. It's a bit like a
qsearch insofar as one side can "stand-pat" when it is its turn to capture.

However, with the techniques you are already using, I believe your qsearch
shouldn't be blowing up like it does. One test you can do is to see what
proportion of the cutoffs you are getting in the search come from the
*first* move you search at a node. This proportion should be *very* high,
well over 80%. You measure this by having two counters: Increment CUTOFFS
every time you get a cutoff. Increment FIRSTMOVE every time one of these
cutoffs is achieved after searching only one move at the node.
FIRSTMOVE/CUTOFFS needs to be a very high percentage, otherwise your tree
explodes.


Andrew Williams



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.