Author: Gerd Isenberg
Date: 05:52:44 09/23/04
Go up one level in this thread
On September 22, 2004 at 17:15:38, Robert Hyatt wrote: >On September 22, 2004 at 15:57:34, martin fierz wrote: > >>i have a question about root move ordering: my engine currently generates a >>random root move order, and then orders every move that ever fails high to the >>top of the list, with previous best moves slowly going down in the list again. >> >>how do you order your root move list? by number of nodes to refute? or by >>something different? >> >>cheers >> martin > >Initially I do a q-search after making each root move, and use this score to >order the moves at the root. If I am pondering I make sure that the move from >the PV is put first of course. > >After each iteration, I re-order the move list keeping the best move first, and >ordering the rest based on the size of the sub-tree (number of nodes) each >produced when it was searched. I do that in a similar way. Imagine following scenario: There is a safe repetition line with move X with a shallow subtree. There are one or more alternatives Y,Z oscillating around zero with huge subtrees. At iteration i, X becomes best, because Y,Z and all others are < zero. Next iteration, Y becomes better. At i+2 Y is still > 0, but Z becomes better. Now at iteration i+3 Z fails low less zero after a huge time. What is better now? a) Try next moves only based on their sub-tree size, with the possible problem that all other moves except X fail to improve >= zero and X due to it's shallow subtree is searched lately - or even worse with rare search time - not at all. b) Trying Y which was previously best, and had a big subtree too. X as pre previous best is tried after Y and Z. c) To remember that there was a safe draw line two iterations before and to try X before Z and others with an improved window? I like option C). But of course one has to distinguish zero scores from real draw scores by backing up some flags.
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.