Author: Severi Salminen
Date: 14:47:28 01/28/01
Go up one level in this thread
>If the only difference is SEE vs MVV/LVA, I generally find the two work >about equally well. SEE will search a 10% smaller tree, but will use 10% >more time per node. Breaks even. MVV/LVA is the opposite. > >But if you use SEE to eliminate bad captures in the q-search, then you >can make the SEE program 2x faster than the non-SEE program. That should >begin to make a difference. Dang. My version without SEE is 2x as fast as the one with SEE (in NPS). It sure shrinks the trees but as said the speed drops a lot. In Qsearch I don't try moves if material+see_score(move)+PAWN_VALUE<=alpha or if see_score<0. For some reason the SEE is very slow to calculate. I do it in same fashion than in Crafty and it shouldn't be very slow. I assign SEE value to every capture at the same time they are generated and if the value of piece being captured<= value of capturing piece. Else I assign the difference. Then I sort them and start searching. There must be something I am missing, or maybe not. It seemed that at longer time controls the version with SEE played tactically better. How much does Crafty get speedup in NPS if you remove SEE? And what do you mean by 2x faster. Searches to same depth 2x faster or what? Should the SEEless version get beated by the one with it? Severi
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.