Author: Andrew Williams
Date: 12:55:16 01/30/03
Go up one level in this thread
On January 30, 2003 at 12:33:10, Andrew Williams wrote: >On January 30, 2003 at 11:19:24, Russell Reagan wrote: > >>On January 30, 2003 at 09:42:01, Andrew Williams wrote: >> >>>To illustrate why SEE is better than MVV/LVA, suppose you have a position where >>>White attacks two of black's minor pieces (worth 300 points each). In each case, >>>White has one attacker on the piece. One of these pieces is attacked by a Rook >>>and is not defended. The other is attacked by a Pawn but is defended. SEE will >>>tell you that the rook capture is better, because it will calculate a material >>>gain of 300 points versus a material gain of 200 points. MVV/LVA will tell you >>>that the Pawn capture is better, because the score will be (300-1) versus >>>(300-5) for the Rook capture. The difference is that if the rook capture would >>>give you a beta cutoff and the pawn capture wouldn't, you'll do a lot of >>>unnecessary work by searching them in the wrong order >> >>Thanks for your comments, and I have a question. >> >>Are both SEE and MVV/LVA only concerned with captures on a single square? I have >>always thought of MVV/LVA to be a kind of qsearch. I think some people have >>different definitions of what SEE is. For instance, I recall Vincent saying that >>you can just call SEE() + Eval() straigth from your search function instead of >>calling QSearch(), so he obviously sees it as a whole board thing instead of >>something on a single square. I have always thought it was a whole board thing >>also, but I have never implemented it. > > >They're both concerned with activities on a single square. To be more precise, >as used in move-ordering, they're concerned with the consequences of making a >particular capture on a specific square. The MVV/LVA or SEE score will be used >later to order the available moves, with the intention that the best captures >are tried first. > I wasn't very happy with my response to this question, so here's another attempt. Suppose we have this position on the board: [D] r1bqkb1r/pppppppp/2n5/1P3n2/8/N4R2/P1PPPPP1/R1BQKBN1 w Qkq - 0 1 It is White's turn, and we have two possible captures, namely bxc6 and Rxf5. What I was describing was the process of ordering these captures, with the intention of searching the "best" one first. So we generate the two captures, and for each one, we try to ascribe a measure of how "good" it is likely to be. One way of doing this would be to use a Static Exchange Evaluator on both captures. The other way is to do what we call Most Valuable Victim, Least Valuable Atacker (MVV/LVA) analysis. As its name implies, you need to do the analysis for *every* available capture, in order to find out the MostVV which is attacked by the LeastVA. So in this sense, MVV/LVA *is* a whole-board technique. What I was describing (badly) was the process of analyzing each *individual* capture with a view to finding the MVV/LVA by comparing the available captures. That is now clearer to me; I hope it is clearer to you. :-) Andrew
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.