Author: Vincent Diepeveen
Date: 08:10:02 01/27/05
Go up one level in this thread
On January 27, 2005 at 04:38:04, Dr. Axel Steinhage wrote: >On January 26, 2005 at 12:59:42, Robert Hyatt wrote: > >>On January 26, 2005 at 08:50:03, Dr. Axel Steinhage wrote: > >>Your case of "fail high" seems to be terribly expensive. Doubling the size of >>the tree at a minimum, because to search the second move after you have a >>fail-high is pure overhead. DB did this as well, but they did do a search to a >>significantly reduced depth, rather than a normal search. Hsu called this "FH >>singular" (fail-high singular). I implemented it exactly as he spelled it out, > >Thanks for your comments! Just wanted to comment on this reply from you (above): >My "fail high" is not expensife, because I also go on searching with reduced >depth (I think I wrote that I use a reduction factor of R=2). >Maybe I did not make myself clear enough. I think the main difference between my >algorithm and the others I have seen so far is, that I NEVER touch a move twice >and I don't even have to generate all moves most of the time. What Hsu et Al did >was to FIRST find the best move and THEN test whether it is singular. What most >people do today is FIRST do a complete search (although at reduced depth) to >find a singular move and then search the whole branch again with the singular >move having higher depth, right? This also means to touch this branch twice. >What I do, however, is only to GO ON searching (with reduced depth!) after a >fail high. So no RESEARCH but a CONTINUATION of the search. This seems costly, At one point you MUST test whether the best move has a margin S or more to all other moves. Where do you plan to do that? >but it may not be! Most of the time the next move I search (after the fail high) >is already within the Singular window (or a fail high again). Plus, I have many >conditions where I don't have to go on searching anyway. >It seems that the test for singularity produces nearly no overhead. Only the >actual extension. But as you stated: this may already be too much overhead. I >can believe, that SE is somehow orthogonal to Nullmove. I will go on testing. >I did not do complete testing so far. I only checked the solution times for a >set of test positions. But that does not mean anything for the practical game >later on. I will keep you up to date on my findings, if you like. > >Thanks, Axel
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.