Thank you for the paper, Olivier. It's very interesting. Now I have now two questions.
(1) In Algorism 1, line 7 in the code part (below) is funny. >Function reward = PerformSimulation(T, s) I guess it's just "Function PerformSimulation(T, s)". (2) In Table 4, why the configurations are not 32 vs 1, 16 vs 1, 8 vs 1, ... but 32 vs 1, 32 vs 2, 32 vs 4, ...? Former is more common and easy to evaluate the scalability, I suppose. Don't you have those numbers? Hideki Olivier Teytaud: <[email protected]>: >> >> I'd like to know both numbers, Pasky. BTW, does pachi use root >> parallelisation on the cluster, ie, the same as of MoGo, Fuego >> and MFG? >> >> >In MoGo it's not root parallelization. We share the statistics in the tree, >e.g. once per second (depending on the time settings); >more precisely, we average >- the number of wins, >- the number of losses, >- the number of amaf-wins >- the number of amaf-losses >in all nodes in the tree with more than e.g. 5% of the number of simulations >at the root. > >This is not so different from averaging just at the root, but there's a >slight improvement. > (the difference is probably much higher when building strategies, as in >opening book building, instead of just choosing one move) > >It is, on the other hand, much better than averaging just before taking the >decision (i.e., roughly speaking, voting) as proposed in some papers. > >More details in http://hal.inria.fr/inria-00512854/. > >Best regards, >Olivier >---- inline file >_______________________________________________ >Computer-go mailing list >[email protected] >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
