Author: Robert Hyatt
Date: 20:49:13 10/15/99
Go up one level in this thread
On October 15, 1999 at 15:51:13, KarinsDad wrote: >On October 15, 1999 at 15:31:21, Robert Hyatt wrote: > >[snip] >> >> >>Just remember that this is right in the heart of the 'chess loop'. Anything >>you do here, you do once for every node in the tree, which means _often_. >> >>you could set up 7 linked lists and add moves to the end of the right list. >>but when you analyze chess trees, you see why it doesn't work... because _so_ >>many positions in the tree only search one move and exit. Why do all that >>work? when 1/2 of the time it is absolutely wasted? > >Well, the simple answer is that you don't do it. > >The complex answer is that you could do it depending on the position and score. > >If you could predict with a high level of certainty whether you will get a >cutoff or not, you could reduce the amount of branching by creating a sorted >move order for those positions which should cutoff (note: I did not say that all >moves had to be sorted) and by not creating a move order at all for those that >should not cutoff. The problem boils down to coming up with a prediction >technique that is relatively accurate. > >Any ideas on how to do that Robert? > >KarinsDad :) It is partially doable. I did this pretty well in Cray Blitz, to predict where to do a parallel 'split' for the parallel search stuff. You want to split at ALL nodes (where all moves must be searched), not at a CUT node (where hopefully only one move is searched, but sometimes 2-3-4 before the cutoff is found). But in general, it is difficult. And when you are changing your mind anywhere in the tree, _all_ move ordering becomes backward as CUT nodes become ALL nodes and vice-versa. And that blows you out of the water.
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.