In my humble opinion, we need a change in the algorithm. The numbers are misleading - 95% of win of MoGo on 32 nodes against MoGo on 1 node (this is a real number for 19x19) certainly means that the parallel version is stronger than the sequential version, but not "much" better, far less than what suggests this 95%. MCTS algorithms adapt the beginning of simulations only, and for many cases we have to deal with predictions on the end of simulations: something like "if the opponent plays X, I'll reply Y". The bias on semeais is, in my humble opinion, equivalent to this fact that we learn only the beginning of the simulations (the tree part) and not the end.
I don't know if the good word is to say that it's a wall or a mountain, but I think the idea is that we need something really different - perhaps heavy playouts that solve some tactical elements, or perhaps some statistical trick for modifying the playouts depending on the simulations - I'd like to solve this with supervised learning like "when I reply X to move Y then I win with higher probability". It would be a nice solution, efficient beyond the game of Go. Well, as I've spent a lot of time on this idea without finding an implementation which works better than the baseline, perhaps my ideas are not very interesting :-) Regarding Moore's Law, I'd love to hear the Mogo team's perspective on this; > they have probably had more opportunity to test their algorithms extensively > on big-n-count computers than any of us. > >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
