Page 1 of 1

Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 5:28 pm
by smatovic
I achieve the best nodes/seconds while using ascending MVV-LVA in AlhbaBeta and descending MVV-LVA in QSearch.

Is this common?

--
Srdja

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 6:03 pm
by bob
smatovic wrote:I achieve the best nodes/seconds while using ascending MVV-LVA in AlhbaBeta and descending MVV-LVA in QSearch.

Is this common?

--
Srdja
Ascending/descending doesn't make any sense to me. Do you mean maybe MVV/LVA in one and LVV/MVA in the other???

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 6:22 pm
by smatovic
Do you mean maybe MVV/LVA in one and LVV/MVA in the other???
Yes, thats what i mean.
Thank you for correcting my wording.

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 6:34 pm
by Houdini
smatovic wrote:I achieve the best nodes/seconds while using ascending MVV-LVA in AlhbaBeta and descending MVV-LVA in QSearch.

Is this common?

--
Srdja
Why is "nodes/second" a criterion for picking either choice?
It's quite possible that the lowest nodes/sec value actually corresponds to the better solution.

Robert

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 6:50 pm
by smatovic
Why is "nodes/second" a criterion for picking either choice?
It's quite possible that the lowest nodes/sec value actually corresponds to the better solution.
Maybe....my engine plays still Blitz on a search depth of 4 so a magnitude slower search algorithm is not applyable.

...i do not understand the background why the Qsearch behaves different than pure Alphabeta concerning the move ordering.

I observed the effect that using MVV-LVA instead of LVV-MVA in Qsearch leads to an tree-explosion.

So i want to ask if this is a common behavior or specific in my implementation (->bug).

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 6:59 pm
by hgm
The fastest nodes are those where you stand pat. No need to generate moves or do anything. So doing lots of futile moves (where the opponent makes you faillow by standing pat) speeds up nps. If you could cause a cutoff from current eval = alpha - 200 in QS by RxQ, where the Rook gets recaptured, that is not nearly as fast as capturing a Pawn and failing low on that. So you reduce the average time per node by first failing low on a number of Pawn captures, before trying the cut move that you could have tried immediately.

Re: Qsearch and MVV-LVA

Posted: Thu Feb 03, 2011 7:04 pm
by smatovic
thx, just found an older post from you about that:

http://www.talkchess.com/forum/viewtopi ... ht=lvv+mva