Author: Robert Hyatt
Date: 06:56:26 12/15/03
Go up one level in this thread
On December 15, 2003 at 06:49:04, José Carlos wrote: >>>>I have seen worse. What happens is that the aspiration window initially cuts >>>>off all mate scores. But once it fails low, it can't just find any old mate >>>>and cut off analyzing that line. Now it has to follow _all_ mates and that >>>>kills performance. This also happens in endgames where you can potentially >>>>promote a pawn, but it always gets lost, so you can search to great depth. >>>>And eventually you find that you can promote it there, but now all those lines >>>>you got cutoffs on before (where the pawn promoted, but it wasn't forced) now >>>>explode and while the depth 25 search took 10 seconds, the depth 26 search might >>>>not take 10 days... >>>> >>>>It's a known problem with no known solution. >>>> >>>>Even not using aspiration search won't solve it. >>> >>>You can not lower the window to -INF, but rather on the initial fail low move it >>>down -1, then next do -3, and so on. Or just do it for fail-high and that >>>should affect fail-low also (since a fail-low on one side is a fail-high for the >>>other). >> >> >>Yes you can. We documented that in "using time wisely" and "using time wisely, >>revisited" which were published in the JICCA. But it only solves it for >>some cases, not the pawn promotion case. Because you _must_ drop it low enough >>to let the promotion stand, and that is low enough to let _all_ promotions >>stand, which blows the search right out of reality. > > Some time ago I read an idea I completely forgot to try later. It was about >starting the search again from iteration 1 when a fail low occurs at high depth, >so that new ordering information (plus actual hash info) could hopefully help in >finding a better move faster. Does this idea actually work? (yes, I know, I >should try it myself, but I can't until next weekend). > > José C. It can work. And it can work pretty well. The problem is that if you are not careful, and a key hash entry gets overwritten, then the new search may well find the old best move. It is possible that this will result in a loop...
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.