Author: Sune Fischer
Date: 12:31:45 03/16/04
Go up one level in this thread
On March 16, 2004 at 13:10:33, Robert Hyatt wrote: >On March 16, 2004 at 12:56:29, Sune Fischer wrote: > >>On March 16, 2004 at 12:33:30, Robert Hyatt wrote: >> >>>> 1> apply HH to the subtrees with that shallow (5-7 ply) remaining depth, and >>>> reset them for every other subtree, ofcourse >>>> 2> apply HH to the top (5-7 ply) only >>>> >>>>It did not have the effect I thought it would have... >>>> >>>>Any ideas on the topic? Which should be doing better anyway? Or apply them both! >>>> >>>>Renze >>> >>>Neither. At the root, before starting a brand new search (not a new iteration) >>>age the counters... IE in Crafty, in "main()" I shift the counters right 8 bits >>>(divide them by 256). This ages old history counters away over a few moves, >>>without losing important values instantly... >> >>I think it is better to rescale it once and a while if you analyze for a long >>time on each move. >> >>What happens is that the table will get filled with search information and as >>the search moves into a new branch (a branch which is largely uncorrelated with >>previous ones) the new search information will have to compete with the old one >>and that can/will result in decreased efficiency. >> >>If you rescale it every x nodes or so, the information keeps being fresh. >> >>Anyway, this is what seems to produce the lowest nodes to solution numbers in my >>tests. >>Every man is his own best tester :) >> >>-S. > > >I think we are different because I use _both_ approaches... killers and history >moves. Killers are definitely local. History is a "global killer". Perhaps if >you don't do both, your results could be different from mine??? I use killers also, but I don't consider HH to be "global" as such. I consider HH to be valid in the neighborhood of where it is constructed, globally I see no reason why it should work well. I think if it did work well globally there would be no need to build it on-the-fly, one could just seed the table before the search with precalculated data. My view is that the history tables provides a form of probability matrix. It tells us which type of move has the best probability of producing a cutoff. The the more local the matrix is, the more accurate does it describe the probabilities for the current position. As a global table I believe it will become quite smeared and lose some of its sting. Anyway it should be a quick test to see which (for Crafty) produces the smallest tree :) -S. >But in any case, I depend on killers for local cutoff moves, and I use history >when all else has failed...
This page took 0.01 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.