Author: Dan Homan
Date: 13:34:09 03/09/99
Go up one level in this thread
On March 09, 1999 at 16:26:39, Bruce Moreland wrote: > >On March 09, 1999 at 16:17:24, Dan Homan wrote: > >>On March 09, 1999 at 14:34:49, Will Singleton wrote: >> >>> >>>Using PVS, I can't seem to demonstrate a clear benefit to using a small window >>>around the root score, as opposed to +-mate. >>> >>>Using a window of, for example, one-third pawn, if a move fails-high or low, it >>>takes fewer nodes to ascertain the move that caused the change, because you >>>don't have to find the real score right away. But then you've got to resolve >>>the new score, and that will take more nodes since you've got to re-search with >>>the relaxed window. >>> >>>And if, after a fail-high or low, you attempt to delay score resolution until >>>the next ply (to avoid the re-search at the same ply), it seems you might have >>>the problem of finding a worse move after the fail at the same ply, since you >>>don't know the real score yet. This would result in even more nodes being used. >>> >>>So I don't see much benefit, unless I'm doing something wrong (likely). On a >>>normal search, without any fails, I see either more or less nodes (between >>>windowing and +-mate), depending on the position. But not really much change. >>> >>>Any comment would be apppreciated. >>> >>>Will >> >>I've gone back and forth on this issue myself. My testing seems to show >>that a small window (+,- 1/4 pawn) is slightly (a couple of percent) >>faster when measured over a large number of problems than (+,- mate), >>but the coding is a bit more complicated. >> >>I think I'll test this again.... >> >> - Dan > >I ran with a +- 32767 window for months and didn't notice. > >bruce I was thinking some more about this and the only place I can see a major improvement from using a window is on positions where you will fail high and change the move you will play... With the window, you can decide to play this move before you know how much better it is, sometimes it takes quite a while (x2 the time) to resolve the higher score. Of course, playing such a move without knowing how much better it is might be risky (sometimes they fail low on re-search!). - Dan
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.