Author: Tony Werten
Date: 01:12:14 06/04/01
Go up one level in this thread
On June 03, 2001 at 14:59:13, Rafael Andrist wrote: >On June 02, 2001 at 15:29:36, Tony Werten wrote: > >>On June 02, 2001 at 14:52:56, Rafael Andrist wrote: >> >>>I've implemented a Nullmove-heuristic in my alpha-beta-search (full window). But >>>now the number of nodes in Q-search explodes. Before, the number of Q-nodes was >>>approximately the double of the normal nodes. Now it's 10x or more. This causes >>>the problem that I need the same time for the same search depth as before, but I >>>get a less accurate result. >>>Is maybe my Nullmove implementation incorrect or is this "normal" and should I >>>use futility pruning or something similar to reduce my Q-Search? >> >>No, there is something wrong. You should get at least 1 ply deeper in the same >>time. >> >>Did you reduce depth with 3 ? ( R=2 and the normal 1 reduction ) Or even 4 ? > >The total reduction of remaining depth is 3 (compared to 1 in a normal subtree >search) > >>Did you check not to do nullmove if you have no more depth left ? > >If there is no more depth left (remaining depth - 3 <= 0) a nullmove is made and >a qsearch applied. If I disabled this and do the nullmove-search only if there >is a positive remaining depth after the reduction, then I have practically >disabled the nullmove, because I let my program normally search 7 or 8 plies >deep and don't nullmove near the root (starting at depth 3). Aha, here we have it. If remaining depth<=0 jump to quiescence is OK. But start nullmove at depth=1 ( there's no reason to delay this to depth=3) The sooner you start the bigger the gain. Second, nullmove gaines more if the remaining depth is big. Searching 7 to 8 ply might be to shallow to see real gain. Try some endgame positions and disable nullmove if there is only one piece left. (not counting pawns ) Now you should see some bigger differences. cheers, Tony > >> >>cheers, >> >>Tony> > > >Rafael B. Andrist
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.