Re:A way to deal with enormous branching factors in strategy...
by hyperbovine on Friday March 25, 2016 @04:06PM
(#11362301)
The insight was to use CNNs to prune the search space. In a sense, the AlphaGo algorithm is to MCTS as importance sampling is to regular Monte Carlo. However, even without the search part, AlphaGo crushes MCTS-based Go algorithms (see https://www.tastehit.com/blog/google-deepmind-alphago-how-it...). That in and of itself is quite extraordinary.
I didn't realize the go solution "just" used monte carlo branching. Still impressive, but does not seem as revolutionary as it's been portrayed.