Author: Andrew Walker
Date: 16:07:28 01/22/98
Go up one level in this thread
On January 22, 1998 at 14:08:34, Bruce Moreland wrote: > >On January 22, 1998 at 14:00:36, Robert Hyatt wrote: > >>The math part is simple. The PV takes longer to search than all the >>rest of the moves, in general. If you have 2 PV's, the math is a >>killer. >>Would be real easy to test, just skip the first move at the front of >>each >>iteration and then come back at the end and pick it up. > >He is searching the successors with a null window, Bob. > >burce Thanks for all the replys to my suggestion, I'm sorry if it was a bit hard to understand. I'd just like to make it clear that the alternate moves which are searched first use a cut-off value given by the previous depth plus a constant. As a result it is likely that none of the alternate moves are searched fully. The idea of the constant is that if an alternate move is seen to beat this and is searched fully, it is more likely than not that the original best move will not have to be searched fully. This means only one move is searched fully rather than two. As has been pointed out, we may have to search two moves fully rather than one (ie if the previous best move improves in value), the idea of the constant is to make 2->1 more likely than 1->2. A few more points on my original post: Firstly this is only meant to apply to the initial move in the tree search. It may be possible to apply it further down using hash table results but I'd like to see if it works in a simple version first! Also what I've seen with crafty and other programs is that at low depths, scores tend to fluctuate more and the best move tends to change more often. Therefore it may(?) be best to only start the above only once we have reached a certain depth, say 7 or 8. The best result from this idea is when there is an alternate move with a score of a pawn or more above the result from the previous depth. When the previous best move is searched, it is very likely that it will be cut off after searching only one second move. This will be a big time saving! Andrew Walker ajw01@uow.edu.au
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.