Author: Russell Reagan
Date: 20:12:57 02/10/03
Go up one level in this thread
On February 10, 2003 at 22:53:47, Tom Likens wrote: >A fail-low is the converse situation. It indicates that the true >value of the position is less than the lower bound we set (i.e. alpha). >This can be trouble, because the search is telling us that it sees a >problem, but it doesn't know know the true extent of the problem. >Usually, this situation needs to be resolved. Most programs will spend >extra search time trying to obtain the real score and hopefully find >a better move. I always hear people saying "Program X is failing low" in live computer chess events, or when a programmer is talking about a game, "My program failed low after move 31." Is the human just guessing that the program failed low, since it is taking an unusually long time? Or is there something in the program to tell them this? If so, is this only at the root where you detect a fail-low as a bad thing? It would seem ridiculous to double the time every time a node somewhere in the tree got a score back that was less than alpha, or maybe I overestimate how often it happens. So are we talking about "at the root" fail-lows, or interior fail-lows? >This can happen at the root if the program is searching the initial >position with a +/- window around a presumed valid score instead of ><-INF,+INF> (the so-called "aspiration search"). You could avoid this >problem altogether by using <-INF,+INF> as your initial alpha/beta >bound, but the total search time on average would take longer. When you say "You can avoid this problem..." above, do you mean that the program will search each root move with a window of <-inf, +inf>? That would seem to defeat the purpose of alphabeta and make your search go 36x as long on average. So, the other alternative I see is that you mean to do the initial search with something other than a <-inf, +inf> window, and I've always seen people use infinite windows when they start the root search. Could you clarify what you mean please? I'm confused :)
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.