Author: Fabien Letouzey
Date: 03:43:34 04/26/04
Go up one level in this thread
On April 26, 2004 at 06:31:58, Tord Romstad wrote: >In my tournament, I all four versions face each other 50 times. The >results: > Gothmog A Gothmog B Gothmog C Gothmog D Sum >Gothmog A XXXX 26.5 31.0 29.5 87.0/150 >Gothmog B 23.5 XXXX 30.5 26.0 80.0/150 >Gothmog C 19.0 19.5 XXXX 31.5 70.0/150 >Gothmog D 20.5 24.0 18.5 XXXX 63.0/150 >It is interesting that the version with the most restrictive criterion >achieves the highest score. This confirms my growing suspicion that >recursive null move pruning doesn't work nearly as well as most people >believe, and that it might be a good idea to make an effort to reduce >its use as much as possible. I will try to experiment with even more >restrictive criterions (perhaps static_eval-biggest_hanging >= beta+margin) >and see whether that improves the strength further. >The number of games is of course still not sufficiently big to make any >definitive conclusions. If I find the time, I will add more games and >more versions to the tournament. >Tord Could you add one with the even more restrictive condition that I use: eval() >= beta (aka fail-high nodes)? I expect it not to do as well, but a comparison is always interesting. One of the justifications for this choice is as follows. I consider null move as a "sit down" (my position is good enough, I challenge you), and the search stops if both players pass (eval() is then returned). The fact that I play Go is not unrelated :) Note that this has nothing to do with double null move, in which a search is done after two null moves. So the philosophy is different from normal null move (evaluate threats) and a small optimisation leads to the fail-high precondition. Fabien.
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.