Author: Dan Honeycutt
Date: 07:58:40 09/13/04
Go up one level in this thread
On September 13, 2004 at 02:21:18, Scott Gasch wrote: >On September 13, 2004 at 00:16:55, Dan Honeycutt wrote: > >>On September 12, 2004 at 14:22:08, Scott Gasch wrote: >> >>[snip] >>>I tried this with mixed results. Overall my ECM/20 score dropped. Now I am >>>thinking about what is wrong with this idea. For one, I use PVS so doing >>>anything based on alpha in the search could cause problems. Imagine searching >>>some line with a=b-1, deciding to extend based on that alpha, finding something >>>great, researching with a real a..b window, deciding not to extend based on the >>>new alpha, and missing the "greatness". >>> >> >>I don't understand what "new alpha" you are talking about. I do the scout >>search with a window of alpha, alpha + 1. If the search fails I research with >>alpha, beta. If it's a reduction (which I expected to fail low but didn't) I >>research with normal depth. If it's an extension I leave depth alone - hoping >>to find that "greatness". >> >>Dan H. > >Ok so what I'm saying is this: you do a scout search where alpha = beta - 1. >Minimal window. Now, your search does something, anything, based on the current >value of alpha. Maybe it decides to extend or not. Maybe it does extended >futility pruning. Whatever. Based on this decision you search some tree and >find something that causes a fail high. > >Now your minimal window search has failed high, so you re-search the same tree. >This time say alpha = beta - 100. So again you search the same move with a >different window. And because your search is doing things based on the current >value of alpha... and because the current value of alpha is not the same as when >you were searching with a minimal window, this time you decide not to do >something. Like extend a critical position. Or prune a critical move. You >search a much different tree. Now your search returns failing low! > OK I see where the "new alpha" comes from - your fail hi/lo carries back to the root. >So from the caller's standpoint you had a non-first move... you did a MWS which >failed high. You researched it and it failed low. What do you do? > I think first you always assume your search is unstable so you never raise alpha on a fail high nor lower beta on a fail low. >My point is that doing things in search based on the current value of alpha is >dangerous and can hurt search stability. > >Scott If you are going to do extensions and reductions beyond basic stuff you need some idea of where you stand relative to alpha and beta. I don't know the answers to your concerns so my solution is simply to ignore it. Dan H.
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.