Author: Vincent Diepeveen
Date: 04:08:20 06/27/04
Go up one level in this thread
On June 26, 2004 at 00:32:01, Stuart Cracraft wrote: >On June 25, 2004 at 19:27:32, Andrew Williams wrote: > >>On June 25, 2004 at 16:15:44, Stuart Cracraft wrote: >> >>>So, normally in the literature I've read >>>(and code I've implemented, it's been to >>>do a null move search with R set to 2, >>>so search(...depth-1-R). >>> >>>But in large searches of 8, 9, 10, 11 and >>>beyond in full width searches, the reduction >>>of 2 does not seem to help as much as >>>larger reductions due to the much smaller >>>subtrees that the null move searches has >>>to search with an R of 2. >>> >>>My question is: what have people done >>>to experiment with larger figures of R and >>>verify the return value is effective >>>and horrid moves aren't produced? >>> >>>I've used R set to ply/2 and ply-2 >>>where ply is the original target depth >>>of the overall iteration. The savings >>>in time is substantial and the moves >>>look the same or as good, the tree searched is >>>drastically smaller of course, but >>>I am worried about quality. >>> >>>Is R of 2 or 3 a holdover from the slow >>>computing days in the literature and nowadays >>>you are using higher settings? >>> >>>Assume everything else about the null move >>>search is held the same (not done in endgames, >>>not done in the original position, no more >>>than 1 null move in a row during the search >>>without an intervening normal move, etc.) >>> >>>Thanks ahead, >>> >>>Stuart >> >>Ernst Heinz has published interesting papers on adaptive null move pruning >>(varying R with remaining depth and remaining material). I would guess that this >>is pretty standard now, at least among amateurs. Omid David Tabibi has published >>recently on verified null move pruning. >> > >Hi -- I implemented the Tabibi/Netanyahu scheme from their ICCA/ICGA >article today after reading your note and doing a google search for >their article (found it at www.cs.biu.ac.il). > >Early results are promising showing verified null move with R=3 almost >as good or better in most iterations than plain R=2. Why not try always R=3 then, gets you a ply deeper than R=2. Assuming of course you're doing checks in qsearch. >Thanks. >Stuart
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.