Petr,
> Do all the games have to start synchonously? > > Games are started one clock at a time. > I think it would still be much better than alpha-beta with MC to do a > UCT search with m * n playouts where you run m parallel descends with > n playouts evaluated in the leaf. It is probably good idea to spread > the descends using virtual loss, not sure if adding 1 or n losses would > be better, perhaps a depth-dependent fraction rounded up would be > best... Of course, having strongly pre-biased node values would help > a lot here, I guess. > > m would be tuned so that you could pre-descend m times for next set of > playouts while FPGA would walk the first m. The disadvantage of course > is that you are descending on a tree that is missing (2*m-1)*n samples > (you are doing basically 2*m samples in parallel). If the FPGA supports > asynchronous game queuing, it would be of course much better as you > could queue next set of games right after the descent, so the > obsolescence factor would be just m*n (m+1 samples in parallel). > > I do not quite understand your description. I'm guessing that you mean to have n playouts at a time instead of one playout at a time like the regular UCT algorithm. Nonetheless, FPGA based method are very static in structure, and can not adapt easily to the dynamic changes of a game tree like a software implemented tree structure. So I don't know how to implement UCT type dynamic tree search algorithms in FPGA right now. Fuming
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
