Author: Robert Hyatt
Date: 06:38:30 10/31/99
Go up one level in this thread
On October 31, 1999 at 02:26:29, Ed Schröder wrote: >>Posted by Robert Hyatt on October 30, 1999 at 10:08:51: > >>>What Bob describes over there is the OLD (original) approach of null-move see >>>my other posting on this topic. It is only about pruning on ply-1 situations. >>> >>>Nowadays null-move is done in the whole tree, quite a difference. >>> >>>Am I wrong about this? >>> >> >>Very definitely wrong. Let me give you the complete details: >> >>Cray Blitz (late 80's) used a null-move search, where a null-move could be >>played _one_ time in a single path. (this is called non-recursive null-move >>search since only one null move is allowed anywhere between the root and the >>tip.) I (and Campbell) also tried recursive null-move, which would allow >>multiple null-moves anywhere in the path (as we do today) but I could only >>test this on slow hardware (a Vax) and the 5 ply searches it could do really >>suffered to horizon effects dealing with mate threats, just like today's null- >>move programs suffer if the depth is limited or the time is short enough). >> >>I also (as did Campbell) try R=2 but at those speeds (It turns out it would >>have >>worked fine on the Cray since we were doing 9-10 plies then and we would have >>been maybe 50 or so points stronger based on testing on today's hardware which >>is fairly close to the 1988 hardware speed of CB) on the Vax, R=2 seemed to be >>terribly risky and lost more than it gained. >> >>Today I am using a dynamic null-move R factor, R=3 close to the root, R=2 >>closer >>to the tips. I use recursive null-move, but don't allow two null-moves in a >>row. >> >>The main difference between what I do now and what I did then was that in 1988 >>I wasn't doing recursive null-move, and in 1988 I used R=1. But I (and >>Campbell) had tested with code _identical_ to what I use today. > >You just described what I have been trying to point out all the time. The >evolution of a bright idea (do something with doing nothing) to a full >developed and powerful selective search good for hundreds of elo points >for many programs. > >Ed I don't follow. Murray defined R=2 and recursive null-move. I simply didn't like them early on. I tried them when I first started Crafty, as my intent was to try _everything_ again (using all my Cray Blitz 'notes', plus anything else I saw in the literature.) IE I looked at Gnuchess, and it was doing null-move R=2 recursive well prior to 95 when I started Crafty. That bumped it up in priority for me. In any case, I consider it far more important to publish an idea, than to later publish details about implementing an idea someone else had already tested. Chrilly didn't modify the original algorithm for null-move at all. His main contribution (IMHO) was the idea of useing the null-move result to detect threats. But 'null-move' itself was 'prior work'.
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.