Author: Tony Werten
Date: 22:16:13 01/22/02
Go up one level in this thread
On January 22, 2002 at 10:46:40, Robert Hyatt wrote: >On January 22, 2002 at 04:43:30, Tony Werten wrote: > >>On January 22, 2002 at 00:46:01, Robert Hyatt wrote: >> >>>On January 21, 2002 at 10:55:50, Daniel Clausen wrote: >>> >>>>Hi >>>> >>>>On January 21, 2002 at 10:41:39, Robert Hyatt wrote: >>>>[snip] >>>>>1. If you get a fail high at the root on a zero-width window (any move after >>>>>the first move should be searched with a zero-width window) you can't trust it >>>>>unless you re-search it with a bigger beta bound and make _sure_ that it doesn't >>>>>then fail low. Such fail-high (zero window) fail-low (non-zero window) is an >>>>>artifact of null-move and if you play such a fail high move even if it fails low >>>>>on the re-search, you will die... >>>> >>>>Anyone has a trivial example at hand which demonstrates this behaviour? >>>> >>>>Why is it that a fail-high with a zero window can't be trusted but a fail-high >>>>with a non-zero window can? Is "non-zero window" enough to be trusted? Or does >>>>it have to be a certain minimum window? I'm sure that as soon as someone posts a >>>>mini-example which shows this behaviour, even I will understand it. :) >>>> >>>>Sargon >>> >>> >>>I don't have an example now as I found this bug several years ago and then >>>eliminated it as I explained. It is an artifact of mixing null-move with >>>the null-window PVS search. >> >>Did you find it to be related to the size of the window ? That's what I seem to >>get from some tests. >> >>BTW wouldn't that make MTD quite useless ? >> >>Tony > > >I think it is related to your evaluation code and search extensions. The size >of the original aspiration window had no effect for me, because the problematic >fail highs were _always_ on the null-window PVS search on a root move other than >the first one. Same for me. But I noticed that if I search the firstmove with a bigger window then the number of false failhighs drops. If I search the first move with (-inf,inf) the false failhighs even completely disappear. Tony >I found some positions where _every_ move would fail high on >the null-window search then fail low on the re-search. Turning off null-move >cleaned it up perfectly. Just ignoring the occasional erroneous fail-high has >worked well for me...
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.