Author: Robert Hyatt
Date: 16:51:32 09/29/99
Go up one level in this thread
On September 29, 1999 at 13:02:17, Gian-Carlo Pascutto wrote: >On September 29, 1999 at 09:34:45, Robert Hyatt wrote: >>1. I tried the R=2/R=1 trick... ie if depth > 2 use R=2, else use R=1, so >>that you always have a non-null on the end. That hurts performance badly. > >In what way ? Is it worse than using R=1 for everything ? It got the >impression it was ok, but maybe I didn't do enough testing. think about the shape of the tree. there are _many_ positions out at the depth=3, 2, 1 (remaining) plies. If you require a non-quiescence follow-up to a null move, at depth=3 (remaining) you won't use R=2, because that will drop you right into the q-search. At depth=2, you can't do a null-move at all as even with R=1 you end up in the capture search. And at depth=1 you go to the capture search anyway. But when you test, you discover that doing nulls at those positions is a big win, speed-wise... > >I'm unsure about R=2. It seems very risky when you get less than 5 ply >searches. it definitely is... > >>2. R=1 is safer, but it turns a 5 ply search into 4 plies or less (if you >>use recursive null-move). But on any reasonable hardware you should _never_ >>see 5 ply searches. IE on ICC in blitz games, I generally see at least 8 and >>normally 9-10 or more... > >Maybe I should have mentioned this: my program plays bughouse/crazyhouse, >mostly 3 0 timecontrols on a Cyrix150 machine. Most searches I see go to >4 or 5 ply, even less if there are lots of checking lines, and sometimes >more if there aren't much pieces in hand. I'm seeing a lot more moves per >position than the 38 average you guys have to cope with :) > >-- >GCP ugh... yes you are... :) keep us posted on how it works... that is an interesting game...
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.