Author: Robert Hyatt
Date: 12:01:34 12/17/97
Go up one level in this thread
On December 17, 1997 at 13:51:12, Willie Wood wrote: > >On December 16, 1997 at 21:58:39, Dan Homan wrote: > >>I tried adding aspiration on top of >>PVS, but >>that seemed to add significant complication for little gain. Move >>ordering >>is hash moves, followed by captures, followed by history moves. I tried >>adding >>killer moves but this didn't seem to add much improvement over history >>(perhaps I did something wrong). >> > >I also took out my aspiration stuff, thinking it didn't add much. The >killer moves, as I understand it, are really a subset of the history >table, so it may not add much to use them in combination. I'm using >only killer moves now, which really cuts the tree down in deeper >searches. I will have a very good case for comparison, since in the >next few weeks I'll be replacing killers with history as an experiment. I use both, because I can try killers before I generate any moves. If I get a cutoff, it is really inexpensive. I didn't notice any particular tree size advantage when I added killers, but I noticed a 10% speed improvement because of trying them before generating any non-capture moves. > >BTW, have you tried eliminating the captures from your move-ordering? >Theoretically, I'd think you'd get more cutoffs if you did. > > >>The only forward pruning is by recursive null move. I recently >>understood >>something Bob Hyatt and Bruce Moreland had mentioned on occasion >>about allowing the null-move to collaspe straight into the capture >>search. >>I implemented this (it turned out to be a simplification over what I was >>doing >>before) and it was a big win. > > >Aha. I wonder if this idea is in the '93 Donniger article, or is it an >enhancement? just stock null-move, except that it is done even at depth=3, 2 or 1 positions, not just when the null move doesn't take you to a quiescence search position, such as those closer to the root... > > >>The capture search considers only captures that will bring the score >>above >>alpha - no checks are considered. > >I recently extended the quiesence to include checks to the extent that >checks occur in the full-width search. This is an attempt to reduce >horizon. It does slow it down, but seems to have improved it. > >Is your program playing on ICC or just fics? I'd like to play it a >couple of games. > >WW
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.