Author: Steve Timson
Date: 12:02:00 01/24/02
Go up one level in this thread
Reading that again I realize the last paragraph is slightly unlcear, when talking about ecm gains and play vs humans/comps, etc, I am talking about singular extensions, not forward pruning (whose effects would be another conversation). On January 24, 2002 at 14:59:10, Steve Timson wrote: >I am not sure really :). The singular extensions I use are far simpler than a >classical deep blue style implementation. What I do was inspired by comments >Bruce made at CCT3 - basically, if a (significantly) reduced depth search of the >moves at a node reveal that only one move produces a score near alpha (the rest >are all worse), then extend it. Parameters to play with include - how much to >reduce depth for the reduced depth search, how much better must one move score >than the others to be considered singular, what part of the tree and what part >of the game to enable the code (e.g. I have had better luck restricting the >extensions to fairly near-root nodes), how it interacts with other extensions >(what extensions should be enabled at the same time, and what extensions should >be enabled when it is off - e.g. I don't do recap extensions when it is on, but >do in parts of the tree where it is off), etc. > >To me it has seemed worth it, but the cost is pretty steep. In the results >posted below you can see it roughly doubled the number of nodes needed to hit >ply 10 - it can be much worse too. I do some pretty significant forward >pruning, which helps keep it in check some - I am not sure I'd want to do SE >without it. On test suites it helps me quite a bit, I haven't run with it off >lately, but if I recall right, enabling it gained me roughly 40 ecm solutions at >20 seconds per move (chester currently scores 661 at 20secs). It also seems to >have helped a lot playing weaker humans on ICC - you know the type of people >that play 50 5 0 games in a row, looking to get that one draw. More frequently >now some tactical shot is discovered. Anyway, it seems this way. It seems to >harm performance in fast games vs comps though. > > - Steve > >On January 24, 2002 at 11:43:01, Will Singleton wrote: > >>On January 24, 2002 at 05:49:44, Steve Timson wrote: >> >>>Chester finds it in 2.17 seconds on my athlon 1200. >>> >> >>Well, pretty darn impressive. Where's the best info on singular extensions? >> >>Will
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.