Author: Gerd Isenberg
Date: 01:25:29 08/07/04
Go up one level in this thread
On August 06, 2004 at 08:44:32, Tord Romstad wrote: >A basic question from a PVS newbie: > >How much of an improvement should I expect to see when I introduce aspiration >windows in a PVS search? When using an aspiration window of +- 0.5 pawns, >I don't see any improvement at all compared to a window of +- INFINITY. In >most positions, the node counts differ by just a couple of % (and the version >with aspiration is not always the fastest). > >Is this what I should expect, or should I start looking for a bug? > >Tord I use PVS with a small aspiration window of about +-1/3 pawn * some dynamic value based on recent fail low/highs and the absolute value of the score. I found it a bit faster and do probably some unsound things, to "gain" from instability. I do up to three fail-low re-searches and up to two fail-high re-searches with progressive enlarged windows. One bound is always the previous score. If i really have to solve a huge fail-low with an [-oo,score] window, i found a presearch (kind of IID) with depth-2 helpfull. If the re-search after a fail-low fails high or vice versa, i do no further re-search on the current depth, but go to depth N+1 instead. There are several boolean flags, informing the gameController to allocate more time until the "instability" is solved. Another, probably even more unsound "improvement" which seems to work somehow for me because i guees it reduces the "fliprate": If a fail-low occurs and there was none in the previous iteration, i immediatly terminate that root search after first move and do either a fail-low re-search or even go to next iteration for the case fail-low after fail-high. Gerd
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.