Author: William Bryant
Date: 09:34:28 06/14/99
Go up one level in this thread
On June 14, 1999 at 11:41:58, Dave Gomboc wrote: >On June 14, 1999 at 09:30:07, William Bryant wrote: > >>On June 13, 1999 at 21:10:57, Pat King wrote: >> >>>Before implementing null move, I had a simple aspirated AB search, and would >>>research with a window of (Beta, Infinity) for fail highs and (-Inifinity Alpha) >>>for fail lows. And life was good. With null move, however, it's not uncommon for >>>my program to get stuck, alternately failing high and low, without ever properly >>>resolving the score. What's up with that? Must the research be done >>>(-Infinity Infinity)? Or is this a sign that I've a bug in the null move code? >> >>Pat, >> >> I just struggled with the same think, and I agree with Will that this should >>be part of a FAQ. It was explained as follows, when switching from a alpha-beta >>window to a beta-Infinity window, the search allows other possibilities that now >>return a fail low score. It happens. I track it and I get a fail low following >>a fail high about every 100 test positions or so. I can send you some positions >>that seem to cause this. >> >> Now for practical advice gleaned from this group. Beleive the initial fail >>high. If I fail low after a fail high, I ignore the fail low -- keep the fail >>high. Therefore the move that failed high should become the new PV move >>although you don't have a continuation and an exact score. > >I have seen the opposite advice as well (ignore the initial fail-high). I don't >think a definitive solution has been recognized. > >Dave Dave, What has you experience been. I ran a middle game mate position last night that failed high - then low with a score of ~ 4.5. On the next iteration, with a window of alpha = fail high score - window, beta = Infinity, it returned a score of 7.something and a PV, so the fail high was clearly correct. I can send you the position and or the analysis is interested.
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.