Author: Robert Hyatt
Date: 18:48:34 11/27/00
Go up one level in this thread
On November 27, 2000 at 19:36:18, Scott Gasch wrote: >Hi, > >I recently attempted to change my move ordering so that "losing" captures >(MVV/LVA) were searched after everything else. This resulted in search trees >about 30% larger than when I searched all captures at once, best to worst >(MVV/LVA) before any non-capture moves. > >The reason I tried this was because both Prof. Hyatt and an article online by E. >Heinz suggest losing captures should be tried last for better move ordering. >Does this rule hold if I am doing a simple MVV/LVA move value scheme instead of >a more complicated SEE? How can I be sure a capture is "losing" just because >the piece I capture with is more valueable than the captured piece? > >Should qxn when n is hung be search last in a position just because it's >considered "losing" since the queen > knight? This seems wrong to me because >most likely such a hung-knight position would be a beta cutoff and we could have >saved all the work of considering all other non-capture moves in this position >had we looked at this "losing" capture first. IS this just the exception to the >rule? Don't get me wrong, I'm all for apparently "better" captures before >"worse" ones but I wonder if I am misunderstanding where to search the "worse" >capture moves. It would seem to be so since my tree expands after I make this >change. > >Scott MVV/LVA was developed by Ken Thompson. It makes a world of sense in hardware where trying to "sort" a move list would make the design 10 times more complex at least. But it fails in lots of places. QxN where the N is defended is obviously worse than NxP where the pawn is undefended. The advantage for MVV/LVA is speed. It is much easier/faster to compute than SEE scores for all the captures. SEE will make the tree smaller, but it will also drop your NPS. I believe SEE is better, but the last time I did a very careful comparison, I found that SEE was 10% more efficient in terms of number of nodes searched. However, if you use SEE to prune bad captures in the q-search, you will go 2x faster, which is a huge savings...
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.