AlphaGo

AlphaGo's main innovation is how it combines deep learning and Monte Carlo tree search to play Go. The AlphaGo architecture consists of four neural networks: a small supervised learning policy network, a large supervised-learning policy network, a reinforcement learning policy network, and a value network. We train all four of these networks plus the MCTS tree. The following sections will cover each training step.

Supervised learning policy networks

The first step in training AlphaGo involves training policy networks on games played by two professionals (in board games such as chess and Go, it is common to keep records of historical games, the board state, and the moves made by each player at every turn). The main idea is to make AlphaGo learn and understand how human experts play Go. More formally, given a board state, 

, and set of actions, 

, we would like a policy network, 

, to predict the next move the human makes. The data consists of pairs of 

 sampled from over 30,000,000 historical...