I thought AlphaGo was still using Monte Carlo tree search, just with two neural networks: one to estimate the board values, and one to estimate likelihood of moves for sampling.