Author: Dieter Buerssner
Date: 15:16:29 10/28/02
Go up one level in this thread
On October 27, 2002 at 17:00:03, José Carlos wrote: >On October 27, 2002 at 14:48:58, Dieter Buerssner wrote: > >>On October 27, 2002 at 04:18:58, José Carlos wrote: >> >>>On October 26, 2002 at 20:10:38, Dieter Buerssner wrote: >>> >>>>On October 25, 2002 at 21:27:18, Nagendra Singh Tomar wrote: > To fight the problem, I don't allow null moves in the pv, and only use it in >refutation lines. I have two search functions: > - what I call AlphaBetaPV searches the pv nodes. I don't do null moves there >and search always open windows. > - my AlphaBeta only receives alpha (and has a local beta = alpha + 1) and does >all kind of dubious prunning. > When a move "appears" to be good enough, it's researched with AlphaBetaPV and >an open window. If I still find it good, I trust it. I cannot see, how the AlphaBetaPV can help much. Obviously there can never be a null move inside of the search window (whenever the search returns an "exact" score). I can see, that not allowing null moves while searching still for the PV will change the search tree - but this seems another problem. That exact score can (and will now and then) still be influenced by wrong pruning decisions in refutation lines. >>BTW. Some perhaps at the first look not too related things can influence the >>occurence of fail high/low pattern. Assume you have a lower bound hit with >>enough draft in TTs, but the score is not >= beta, it is > alpha however. > > So the window must be open, not a null window artifact... But such a null move artifact can fold in by the adjusting of the bound. >>If you now adjust alpha, this can avoid a following fail low in a >>research situation, that would have be seen, without adjusting alpha. > > By adjusting alpha you mean changing my current alpha for the found lower >bound, right? Yes. I tried with and without. It seems that adjusting the bound is better for my engine (smaller search trees in general, but now an then, a test position may be solved only at higer depth). Just another random thought, how such artifacts can be introduced. Assume you have a clever qsearch, that does not consider moves, that don't bring the material back close to alpha. In a research alpha may be lower, and you consider them now. Without any more rules, you might have pruned away earlier a move RxP in a KRNPKR position (material down a piece looked too bad). Now you consider that, and your eval returns a very drawish score (a piece more does not help here without a pawn). From the point of view of the other side, a line that looked very good (piece ahead) and failed high, now can fail low again to drawish score with the open window. Of course the fix here, is not to prune away a capture of a last pawn, even when it does not bring the material back close to alpha. 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.