Winboard Forum

Hi guys. I've been a reader for a long time, but this is my first post. My engine is currently being shackled by extremely large quiescent searches. It averages around 10X more quiescent nodes than normal ones!

Currently I try all captures. I use a form of MVV/LVA for move ordering. What are other people doing to keep their searches down?

Thanks,
Jake

- Stop QS when not moving at all is sufficient for a beta cutoff.
- Do delta pruning. When eval<alpha, don't try moves that don't capture enough material to reach the search window.

3) don't do loosing captures

HJ.

It does depend a little on what exactly you consider a QS node. If you count the leaf nodes of the normal search (where depthLeft == 0) already as QS, this might be a quite normal ratio in opening/middle-game.

Onno - I'm already doing your first suggestion, but I hadn't heard of delta pruning. That sounds very logical and I bet it cuts the tree down a bit. Thanks.


Harald - Is there much of a reduction in playing strength by not searching all possible captures? i.e. If the last regular move made was PxP and the only capturing move in reply is QxP, then material is even. If I don't test the QxP because it's a losing capture, the score is off by a pawn. One could possibly only search good captures at deeper qsearch depths, while searching all captures early on?


H.G. - Right now I count all nodes visited, not just leaves. That's for both searches. You bring up a good point though in that I don't really know what a normal ratio is. What kinds of ratios are others getting?


Is it common to keep track of killer moves for quiescent search? That might help a bit.

Jake Keel wrote:Harald - Is there much of a reduction in playing strength by not searching all possible captures? i.e. If the last regular move made was PxP and the only capturing move in reply is QxP, then material is even. If I don't test the QxP because it's a losing capture, the score is off by a pawn. One could possibly only search good captures at deeper qsearch depths, while searching all captures early on?

Losing as in you know something will capture the queen afterwards. SEE here.

Is it common to keep track of killer moves for quiescent search? That might help a bit.

No. Killers are meant to be non-captures.

Harald Johnsen wrote:3) don't do loosing captures

HJ.

Loosing means (Attacker > Victim) && (See(move) < 0)
The first condition is redundant but it helps calling See less often.

HJ.

I assume use of transposition tables in quiescence searches by some people and not by others - could be (part of) the explanation for the original author's high node count.