Training AlphaGo Zero

Since we don't use human-generated data for training or supervision, how does AlphaGo Zero learn at all? The novel reinforcement learning algorithm developed by DeepMind involves using MCTS as a teacher for the neural network, which represents both policy and value functions.

In particular, the outputs of MCTS are 1) probabilities, , for each selecting move during the simulation, and 2) the final outcome of the game, . The neural network, , takes in a board state, , and also outputs a tuple of , where  is a vector of move ...

Get Python Reinforcement Learning Projects now with the O’Reilly learning platform.

O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.