Author: Vincent Diepeveen
Date: 09:04:17 04/10/01
Go up one level in this thread
On April 10, 2001 at 10:25:02, Dieter Buerssner wrote: >On April 10, 2001 at 08:58:14, Ulrich Tuerke wrote: > >>On April 10, 2001 at 07:05:59, Tony Werten wrote: >> >>>On April 09, 2001 at 23:40:21, Jon Dart wrote: >>> >>>>>On April 09, 2001 at 17:04:09, Vincent Diepeveen wrote: >>>> >>>>> An interesting thing is lazy evaluation, as the problems of it are >>>>> very similar to futility pruning. >[...] >>>>> My big question was: what score to return for example if evaluation in this >>>>> position is e and e+ 3.5 pawns <= alfa ? >>>> >>>>> Must one return alpha, estimated evaluation or evaluation+3.5 pawns, >>>>> when talking about e+margin <= alfa (idem story for e-margin >= beta) ? >>>> >>>>I return the estimated evaluation. But I fail to see that it makes a lot of >>>>difference. If it's below alpha, you're not going to propagate this evaluation >>>>up the tree, anyway. >>> >>>With failsoft alfabeta it easily could. > > >>I think that it is a bad idea to return the estimated eval here because it can >>result in some root-value far outside the aspiration window. It would be >>interpreted as a corresponding bound which may cause serious trouble (fail low >>in fail high verification search for instance). >>IMO, it's much safer to return the corresponding beta or alpha. In case of >>traversion of this value to the root this would cause a continous re-search >>(i.e. over a neighbour region of the original window). > >I also have thought about this problem. I am using fail-soft alpha-beta search. >I agree, that returning alpha or beta is safest. Nevertheless, some limited >testing has shown here, that returning an estimation solved the testpostions I >used with less nodes on average. > >From my understanding, returning an estimation should work and be theoratically >sound. In eval, I first check for cases, which I don't want to estimate, like >pawn endgames, endgames where one side has no pawns, etc. For other cases, I >more or less compute bound values. > >1) material score + highest positional advantage seen for this side >2) material score - highest positional advantage seen for the other side So i would need after a few ply a window materialscore + 20000 materialscore - 20000 pawn = 1000. I knew someone would again reinvent this old idea. Only works for simplistic evals. >where, with positional evaluation, I mean the overall positional score, that >includes both sides. > >1) is compared with alpha, if <= alpha, it is returned as estimation >2) is compared with beta, if >= beta, it is returned as estimation > >I also made some tests, to see how often this fails. IIRC this happened very >rarely. > >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.