Author: Ulrich Tuerke
Date: 15:56:41 05/25/02
Go up one level in this thread
On May 25, 2002 at 15:10:21, José Carlos wrote: >On May 25, 2002 at 07:29:30, Ulrich Tuerke wrote: > >>On May 24, 2002 at 19:41:52, José Carlos wrote: >> >>>On May 24, 2002 at 18:14:57, J. Wesley Cleveland wrote: >>> >>>> I was looking at the source of crafty and saw that on a fail low, the value >>>>returned is alpha, rather than Max of all the values from the level below. Is >>>>there some disadvantage in returning the real upper bound? I think it causes >>>>some problems in Crafty making it much slower resolving fail-highs and fail-lows >>>>at the root as these branches all need to be re-searched. >>> >>> I use fail soft because I think (not tested) that it generates smaller trees >>>generally. >>> The drawback is that what you call "real upper bound" is no always so. Null >>>move and hashing make the search return "not totally correct bounds" some times. >> >>What do you mean by "not totally correct bound" ? >>I don't care for truth or correctness in my chess program; the aim is just to >>find a good move. In case my fail high verification search is failing low, I >>just stick to the alpha of the verification search and go on searching the >>remaining moves. I think, that's not really a problem. > > No it isn't if you take care of it. But what I meant with "not totally correct >bound" is that, sometimes, the search says "this move is better than 0.54" and, >if you research with an open window you find that the move is worth 0.51. Yes, I guessed so. >I was >trying to answer the original question saying that the "possible problems" fail >soft can have come from this unexpected behaviour. I think that this "problem" is due to alpha-beta dependencies of the search. The verification search comes with a shifted aspiration window and may lead to a different result. I do not think that choosing "fail soft" can cure this because these dependencies are not related to this choice. IMHO, issues like null-move, evaluation window techniques, razoring and so on are the source of these search anomalies. I am convinced if your search has these dependencies then you will observe in both schemes cases where a verification search does not confirm the result of the preceeding search. c/u Uli > > Regards, > > José C.
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.