Hi Edward and hi all,
> I truly believe that we can make big steps in this direction and getting > much better (real not simulation) players that way. I find MoGo already very impressive, so it surprised me that you think that more big steps are possible.
I do really think that it is only a first step, and the great improvements are "to do" and not "done". I am really optimistic for the community (and not for MoGo as I don't now the time I can spend on it for the next months).
The number of playouts per second for MoGo is very low (maybe the lowest of all MC players?). Is this caused by the large number of calculations necessary to get a good simulation policy? Or is it possible for MoGo to get a speedup like Lukasz Lew has accieved with his library?
The slowness of MoGo's simulations come from two parts. Of course improving the simulation policy make it slower. But also we designed MoGo so that we can add and test quickly new ideas. So there are a lot of unnecessary computations during the simulations. I believe that making specific optimisations by a good developer, like Lukasz Lew did, we could achieve a *2, *3 in the speed, perhaps even more? But I also think that it should be done "at the end", when we think that the simulation policy is the best we can get, and I don't think it is "time". To give a idea, the speed of the best simulation policy in MoGo is 0.6 * the speed of the uniform random one. Sylvain
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
