Author: Vincent Diepeveen
Date: 13:04:31 07/19/99
Go up one level in this thread
On July 19, 1999 at 10:53:44, Robert Hyatt wrote: >On July 19, 1999 at 09:53:36, Vincent Diepeveen wrote: > >>On July 18, 1999 at 13:05:23, Francis Monkman wrote: >> >>> >>>On July 18, 1999 at 12:56:35, Robert Hyatt wrote: >>> >>>>The idea sounds attractive, until you realize that the game tree search is >>>>an exponential problem, not a linear one..... That makes it _very_ difficult >>>>for such a task to be done on computers... >>> >>>But why not treat the computers as nodes in a tree, with sub-delegation >>>software? (Just a thought -- but I built a working parallel audio synthesizer >>>out of multiple TMS99000s back in '84, so I've been 'thinking parallel' for a >>>while -- though not in chess. That's why I thought of you. Hope you don't mind) >>> >>>Say the machines are in a pool. Starting from root, one machine picks the next n >>>(=number of legal moves) machines' addresses. Then they in turn pick 'em off the >>>stack, and so on. Crazy? >>> >>>Francis >> >>Sounds like a deep blue parallellizing approach :) >> > >I don't see where. DB uses _all_ processors on _every_ move. For any ply-1 >move, the first 4 plies (one move each) are searched by one cpu. At this >point the tree is split among _all_ the SP processors. Each processor then >steps thru 4 plies (4 moves) and then uses its own private set of chess >processors to search stuff there in parallel. And it gets a lot more complex >after that as Hsu has a lot of 'pre-search' tricks to keep things busy... (See >his thesis for details). > >DB's search is as good as anything we are doing when you use that many >processors. It probably is better. Using 4 is not _nearly_ as hard as using >500. Right, that's why i'm using nullmove. > > > >>Well Francis, i fear it's a bit off reality. >> >>Deep Blue had a similar approach: >>first 4 ply: 1 SP processor >>ply 5..8 : 30 SP processors >>ply 8..12 : 480 hardware processors >> >>above is a similar idea, also not working that well. >> >>There are great dependancies: after first move has been searched one >>can efficiently give other processors a job. Till then they're doing >>nothing (assuming game start). >> >>example problem in your approach: if a processor finds somewhere that it >>wins material, then isn't it a shame in your approach that they all >>have done work for nothing, as they all are searching the same gamespace!
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.