On 10/30/07, [EMAIL PROTECTED] <[EMAIL PROTECTED]> wrote: > > > > -----Original Message----- > > From: Jason House <[EMAIL PROTECTED]> > > To: computer-go <[email protected]> > > Sent: Mon, 29 Oct 2007 3:00 pm > > Subject: Re: [computer-go] BOINC > > > > > On 10/29/07, [EMAIL PROTECTED] <[EMAIL PROTECTED] > wrote: > > > > > > > > milestone 1: All network-nodes compute pure Monte-Carlo (no search > > tree) scores for the possible moves, the > > > scores are combined centrally > > to pick the move. It's easy, it will wring out the system, and the bandwidth > > is low. > > The playing performance will always be poor because this > > algorithm doesn't scale well. > > > > > It scales, reasonably, but there's a maximum total work to do before any > extra becomes useless. > > > Both of our statements are so vague that I can't tell if we agree or > disagree. :) Here's what I meant. The consensus is that Monte-Carlo with UCT > converges in the limit, as time and memory approach infinity, while > Monte-Carlo by itself does not. In practice, Monte-Carlo by itself plateaus > out at around 5K playouts/move. Don's scalability study tested Monte-Carlo > with UCT out past 1 Million playouts/move and found no sign of a plateau. So > if a network of 1000 computers played pure Monte-Carlo go, I believe the > playing strength would still be weak. Not so for UCT.
I think we're in agreement. I didn't know about the 5k limit, but that's essentially what I was thinking. Having 1 computer do 5k sims is pretty quick already. Having 1000 computers doing 5 playouts each is just insanity. The overhead would probably make it take as long (or longer) as 1 computer doing all of it. This is of course true only for pure monte carlo or other 1-ply variants.
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
