Author: Uri Blass
Date: 08:08:54 11/30/02
Go up one level in this thread
On November 30, 2002 at 10:46:59, scott farrell wrote: >On November 30, 2002 at 10:39:45, Robert Hyatt wrote: > >>On November 30, 2002 at 09:27:20, scott farrell wrote: >> >>>This is a continuration from another thread, but I need some help in >>>understanding this. >>> >>>Some of the engines "failed-high" detecting the move, before they had fully >>>completed the ply. >>> >>>Uri and I cant understand how they detect or report this. >> >> >>Simple. When you start an iteration, you should use some sort of aspiration >>window, rather than setting alpha=-infinity and beta=+infinity. In the case >>of Crafty, if the score for iteration N-1 was +.30, then I might start the >>search at iteration N with alpha=0, beta=+60 to try to bracket the expected >>score. This makes the search tree smaller as you can prune lines that lose >>or win material quickly. However, if you really do suddenly see a path to >>win a pawn, that score will be >= +.60, which means the search returns beta >>rather than a true score. You have to re-set beta to something larger and >>search again to get the true score. That is a "fail high". > >mmmmmmmm ........ that's what I guessed. Is there anything tricky I need to do >with the research. I seem to get too many problems of search instability, and >end up getting different answers to a standard -INF,INF search. > >In my hashtable I have the score, and a/b/exact, and do your suggestion fail >high, fail low or return exact score (assuming enough depth). I am using >nullmove and pvs. It has nothing to do with null moves and hash tables I do not do -inf,inf searches. I do today x-30,x+30 search when the score from previous iteration is x and only if I get x+30 or x-30 I do research. If I am in the sam iteration I usually use x,x+1 and only in case of fail high I use a bigger window. I believe that using always 30 is a bad idea and it is better to use a varaible based on the changes on the search but I believe that I cannot get a big gain from it and my program is not at the level when it is worth to investigate it. Uri
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.