Author: Gian-Carlo Pascutto
Date: 15:01:14 09/04/03
Go up one level in this thread
On September 04, 2003 at 17:10:26, Robert Hyatt wrote: >Knuth and Moore _know_ that splitting at the root is the right thing to do. ...assuming perfect move ordering. >After you have searched the first move in parallel, of course... > >I do both. This is simply about getting a good splitpoint. Ideal splitpoints are ALL nodes, that have high certainty, are close to the root, and don't have any moves searched yet. The root itself doesn't fit that criteria as far as I'm concerned, and you messing around with 'exceptions' is an evidence of that. The root doesn't have a particularly high certainty and you're forced to serialize some moves often. Something funny is that you serialize if there's multiple big subtrees. You don't serialize when one tree is much bigger than the others. You're effectively getting exactly the opposite from what you'd like to have (big trees to search in parallel). That's a PV node for you. You preferably want to use splitpoints from ALL nodes with a high certainty. But there's a tradeoff between more splits and more overhead from that and the risk of wrong ordering at PV nodes generating wasted effort. Note that I'm working on the assumption that classifying nodes works more accurately than ordering moves. This should be true given some testing of the appropriate heuristics. I've been wondering why Crafty alledgedly goes so much faster when you split at the root. I'd assume it's because your simple design doesn't leave you much choice in getting a good splitpoint, i.e. you take the first thing that looks like an ALL node that you come by. Because of the same reason, you don't care much if you get a PV or an ALL node, since you don't split until you've serialized the first two moves anyway. I've looked what you wrote about this from Cray Blitz, and apparently it allowed split at the root also. I'm interested if you have performance results with and without splitting in the root (also for Crafty). By the nature of DTS, CB might have ended up not actually splitting in the root (in some cases) because there were simply better split points available. -- GCP
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.