> 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. > > milestone 2: Each network-node builds its own tree using UCT, but > information is only combined at the root. This version will play much > better because each node is smarter. The bandwidth will be higher. I > can only guess at the scaling behavior, but this milestone might be > the 80% solution. > > milestone 3: Information from the search-nodes is shared between > network-nodes, but only for search-nodes close to the root of the > tree. Sounds innocent enough. You just limit the shared nodes to the > first couple of plys. But it's a trap that will suck you in: best > scaling behavior requires too much communication-but what if you made > each Monte-Carlo simulation smarter...? Why not make each computer specialize in a certain branch of the tree? That way the (implicit) combination of the trees of all the computers could grow very large. Of course some central intelligence would need to allocate resources to different branches dynamically, so it would be more difficult to implement. In these three milestones, the different computers would do a lot of overlapping work with the deeper nodes, if I understood them correctly.
-- Tapani Raiko, <[EMAIL PROTECTED]>, +358 50 5225750 http://www.cis.hut.fi/praiko/ _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
