Author: Dave Gomboc
Date: 12:11:30 08/23/00
Go up one level in this thread
On August 23, 2000 at 03:57:00, Christophe Theron wrote: >On August 23, 2000 at 03:42:09, Dieter Buerssner wrote: > >>On August 22, 2000 at 21:57:58, Antonio Dieguez wrote: >> >>>can a kind person explain me Negascout? >>>what it does in addition of pvsing? >> >>I think there are two differences between Negascout and PVS. >> >>Both use zero window searches with -alpha-1, -alpha for the recursive call. >>If the score returned from the recursive call is > alpha and < beta, Negascout >>will research with -beta, -score while PVS will research with -beta, -alpha. > > >This is a subtle difference and actually one should try both versions and see >which one works better. > > > >>Also Negascout will not research, when the depth is the final depth - 1 or >>the final depth - 2. >> >>I believe, this last point won't work, when you have pruning decisons in >>qsearch, >>that depend on the window (i.e. you skip capturing moves, that look, as if they >>could not bring back the score to alpha), or when you use futility pruning close >>to the leaves of the tree. > > >I have never understood this "not re-search in the last plies" trick. Obviously >it does not work for the reasons you have given, but I think it also simply does >not work if you use a QSearch. I suspect this optimization works only when you >do not have a QSearch, and you consider the nodes at the horizon to be terminal >nodes. Because in this case the score you have got with the null window cannot >change after the re-search with the full window. > >If somebody understands it in a different way, I'm interested to hear about the >explanation... > > > Christophe Yes, the "optimisation" is valid only if fixed-depth searches are being performed. I remember one day in Heuristic Search Jonathan criticised this "optimisation". He explained that it had cost him several hours of debugging once when he started using extensions. (Of course, he was able to solve the problem by getting rid of the subtraction.) More annoying was the fact that the "optimisation" wasn't worth much in practice even when it was theoretically sound, which is (in practical terms) never. Dave
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.