Author: Robert Hyatt
Date: 08:53:39 08/28/00
Go up one level in this thread
On August 28, 2000 at 05:07:37, Steve Maughan wrote: >Tom, > >It was good to meet you at the WMCCC - maybe next year I'll enter! > >What I do when I get a fail high / fail low is research with Alpha = -Inf and >Beta = +Inf. This seem to work since, as you rightly pointed out, the >phenomenon is rare and this gets the exact value for the move. > >Regards, > >Steve Maughan That works, but will blow up on critical positions. Imagine a position where you just won a pawn with a 10 ply search. But the 11 ply search is going to show that you can't hold it, so you fail low. If you relax alpha to -infinity, not only do you include all those positions where you don't win the pawn (which is correct) but you also have to search all those positions where you lose even more, which is not correct. Harry had a few positions where bumping beta too much or dropping alpha too much simply meant you would _never_ resolve the true score. IE it is similar to fine #70. You rip thru the plies like mad, until suddenly you find you can force a queen. Now all those lines you could prune don't get pruned any more, and your search simply hangs until time runs out... I _still_ like the aspiration window approach, and the fail high / fail low problem can be fixed easily as I have done...
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.