I actually tried leaf parallelization first, but after reading the mentioned paper I switched to an implementation of root parallelization (as described). I'm not sure if I implemented it correctly (like in my description), but after testing a 2-core-version against a single- core-version with a few games, I can say the dual core version wins at least 75% of the games, which I think would be an ELO difference of about 200. I have yet to switch colors but the results should be similar.
-ibd > There were a couple of papers [2] at CG2008 on this subject. The > consensus seemed to be that root parallelization [1] was best. In fact > Guillaume Chaslot's version got a strength speedup of 3.0 using 2 > threads, and 6.5 using 4 threads (dropping to 14.9 with 16 threads). > This is of course impossible, and implies the parallel version is > somehow doing MCTS better than the single thread algorithm! > > Darren > > [1]: Build multiple MCTS trees in parallel, one per thread. No > communication. At end add the trees together and use grand total to > select move. > > [2]: > http://www.cs.unimaas.nl/g.chaslot/papers/parallelMCTS.pdf > and > "A Parallel Monte-Carlo Tree Search Algorithm" by Tristan Cazenave (I > couldn't seem to find a PDF link.) -- Psssst! Schon vom neuen GMX MultiMessenger gehört? Der kann`s mit allen: http://www.gmx.net/de/go/multimessenger01 _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
