Author: Tord Romstad
Date: 12:09:43 02/27/06
Go up one level in this thread
On February 27, 2006 at 12:23:29, Robert Hyatt wrote: >I ran with a flavor of this in CCT8 and noticed no problems whatsoever from a >tactical perspective. My original thinking when implementing this was that it >would hurt tactics, but improve positional play, since reducing the depth on the >oddball moves would tend to drive the tactics over the horizon. Interestingly, >in testing the idea prior to CCT8, I discovered the opposite. That's my experience, too. Tactically, my program is *much* stronger with late move reductions enabled. Horizon effects are rare. It does happen (but not very often) that some tactical shot is found one iteration later, but usually the time to solution is still shorter. I tried a quick run through ECMGCP at 10 seconds/position now. With late move reductions enabled, my program scored 166/183. When I disabled them, the score dropped to 150/183. This is a rather significant difference. There are also some rather cheap ways to improve the tactical accuracy. One simple trick which works surprisingly well for me is the following: If I reduce a move, and at the following node the null move search fails low, and furthermore the moving piece in the move that refuted the null move is the same as the moving piece in the reduced depth move, instantly stop the reduced depth search and re-search with full depth. Various forms of static threat detection could also help. It is probably also a good idea not to reduce anything too close to the horizon. >So, bottom line, the idea seems workable now. I've got a few other things to >test, and some are playing with the reduction amount (I used 1.0 plies for CCT8 >but Mike has been testing with 2.0 so there is room for improvement probably). I have never had success with reducing by more than one ply, but of course this doesn't mean that it is impossible. >There are also a few other types of moves that might be excluded, but I am not >doing this at present (yet). Passed pawn pushes for example probably should be >excluded, but presently are not. I do this, in an indirect way. I evaluate all internal nodes. If a reduced move seriously increases the passed pawn eval for the moving side, I cancel the reduction. I do something similar for king safety, and consider to add other evaluation components as well. There are plenty of other obvious ideas which are worth trying here. >I'll release all the details once the WCCC event is over, but the idea appears >to be fully workable despite what "some" will say. :) Well, let's rather say that it's fully workable in *some* programs. :-) Tord
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.