Author: Robert Hyatt
Date: 20:14:26 12/31/00
Go up one level in this thread
On December 31, 2000 at 10:56:07, Steve Maughan wrote: >Bob, > >>>Surely at least when alpha is raised this is equivalent to finding a new best >>>move and hence the PV needs to be cut off? >> >>I don't think so. I don't do this in Crafty, yet I raise alpha and lower beta >>just as you do. All the hash entry says is "this position should have a higher >>alpha bound (or lower beta bound)..." but it says nothing at all about a PV >>since you still have to keep searching... > >As an example, suppose that initially Alpha = 1.00, Beta = 5.00 . From the hash >we find a deeper search with a lower bound of 3.00. It is possible that had we >not changed alpha and after the node has been searched the score for this node >would be +2.00 and it could have formed part of the PV. However, if we raise >alpha to 3.00 the search will fail low and therefore the node cannot be part of >the PV. This doesn't seem right. Am I missing anything? > >I'll have a fiddle with my code and see if I come to any conclusions. > >Regards, > >Steve This isn't an issue. If it would have failed high on the original bound, but doesn't on the new/better bound, that would be the correct thing to do... think about what is going on... _anytime_ you can tighten the alpha/beta window, you should do so. This is what "fail-soft alpha/beta" does all the time, by allowing the search to return values outside the normal alpha/beta window to refine the window better... In the case you give, the so-called "PV" would _never_ have been backed all the way up to the root... It couldn't since it would have a score that is contradictory with a search you have _already_ done for that position... one of the two values would have to be wrong. And that would be inconsistent...
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.