Author: Scott Gasch
Date: 17:52:47 04/02/01
Hi all, In an effort to tune my chess engine I've been paying close attention to first move beta cut percentage rates (which, in my engine, are consistently slightly lower than in other engines I test). I've added a debug mode where my code actually dumps the node position and moves in the order it considered them (along with their score) up to and including the move that caused a beta cutoff when that beta cutoff move was not the first move considered at node. Wading through this data I've seen a pattern that I wanted to ask the list about. Presently I search moves in the order that many others do: winning captures/promotions (by SEE), even captures, killer moves (which are explicitly non capturing / promoting), other non-captures ordered by history heuristic, and losing captures. Now, imagine a position where N takes (defended)p exposes the enemy q or k to discovered attack. Nxp is "losing" according to the SEE so it doesn't get considered until way down in the move list. However it's actually a really good move at this point and causes a beta cut. The problem is I do not use dynamic move ordering at all with captures / promotions. Only non-captures use history and killer heuristic. So with our hypothetical Nxp position we search a ton of recursive lines before seeing this one... which is ok because it's the exception not the rule. But later on in the search at the same tree depth we'll come to a position that is almost exactly the same. Again, Nxp is a great response... and again it's not searched until dead last because the SEE says its "losing". So the solution, I think, is to allow captures into the killer list. Maybe only if they are losing captures(?). Unfortunately because of the way my generator works this is a pain in the butt to do. Does anyone else have any thoughts on this matter? Has anyone else experimented with this stuff in reasonable detail? As always, thanks to all. I'm grateful for any advice. Scott
This page took 0.01 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.