Author: Georg v. Zimmermann
Date: 15:22:26 09/02/01
Go up one level in this thread
On September 02, 2001 at 15:48:26, Dieter Buerssner wrote: >On September 02, 2001 at 12:52:21, Georg v. Zimmermann wrote: > >>How do you prune window dependent ? > >Isn't (almost) all pruning window dependent? With null move, you compare with >beta. With futility pruning you compare with alpha. Hehe I am seriously braindead today. I was reading "pruning" and thinking "extending". What you say below is making a lot of sense. The >= I will need to think about when less tired. The NullMove thing might be not too much of a problem. Shouldn't it happen very very seldom ? It can be a problem in endgame we all know and know how to fix. It can also be a problem when there is a quiet threat, because of R being > 1. But with todays search depths, how often do you get a position where you get mated in 10 halfmoves if you pass, but don't lose material in 9 halfmoves if you do ? Regards, Georg >You have a lower bound score in the hash with enough depth. Say the score is +10 >(which means it was shown to be >= +10 in the earlier search). If you adjust >alpha to 10, and the exact score of this node would actually be +10 (the "=" in >">="), this score will not be inside the new (adjusted) alpha-beta window, >because the bounds are not inside the window (at least with the normal >formulation of alpha-beta search, that I am used to). I'll have to think about the >= thing a bit. This is interesting. >What I mean about the additional inconsistencies added by adjusting the bounds >due to window dependent pruning may be depicted by the following example: > >A position was searched to depth 10 with beta=0. The null move fails high. It >can happen, that if you would search this with no null move, that it really >would not fail high. It might be a score of -30 or something. But you don't know >this yet. >You store 0 as lower bound score in the hash table. You revisit this >node at depth 10. Say this time with a window of (-50,50); so beta is 50. Now >you adjust alpha to 0. Because of the higher beta, the null move shall not give >a cutoff this time. You search all moves with the window (0,50). And you can get >a fail low, because the former null move score was wrong. Without adjusting the >bound you may get a perfect exact score (-30) and a PV. With adjusting, you can >end up with a too high score, that may find its way back to the root, and >perhaps a real problem is only seen at one higher depth. > >All this said, I adjust the bounds. But I don't see a big difference by this. > >Regards, >Dieter
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.