Author: Robert Hyatt
Date: 10:38:29 12/27/99
Go up one level in this thread
On December 27, 1999 at 12:48:51, Gian-Carlo Pascutto wrote: >On December 27, 1999 at 12:14:11, Robert Hyatt wrote: >> >>For me, adding aspiration search here helps significantly. IE start off >>with alfabeta(last-X, last+X); >> >>That (for me) reduces the size of the tree by 10% in the average case, >>much more in wild tactical positions. IE a position where there are >>zillions of mates, but none are forced. with -inf,+inf, you have to follow >>all those mates to their conclusion. With last-X, last+X, you don't... >> >>Harry had one problem position where -inf,+inf could not ever return a good >>score, yet last-X would find the solution very quickly (score was only +2 or >>something). > >I can second that ;) It's even much, *much* worse in bughouse games. > >There's definetly a gain in *not* researching with a full-window after a >fail-high or fail-low. Still, not many programs seem to do it. > >-- >GCP Here is what I think is the right way to do this, as was done in Cray Blitz and Belle: search the first move with a tight aspiration window. Search the rest with a null window. If one fails high, mark it as "best" but don't re-search it yet. If that is the only one that fails high, it is the best. If a second move fails high, re-search one with a wider window to get a real score, then search the second to see if if it fails high. If not, the first is the best, otherwise the fail-high move is the best. The only problem with this is that you can start a new iteration without a PV. If you use internal iterative deepening, this shouldn't be a problem. If you don't, it can cause problems... However, many hate getting a fail high without knowing whether the move is a centipawn better or a rook better. :) So nobody is doing this that I am aware of...
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.