Reinforcement Learning with TensorFlow

Monte Carlo Tree Search

In Chapter 5, Q-Learning and Deep Q Networks we studied the Monte Carlo Tree Search. Here, let's revise it again and see how it was used by AlphaGo to achieve better results.

Monte Carlo Tree Search is an alternative approach to game tree search. In this approach, we run many simulations of the game, where each simulation starts with the current game state and ends with one of the two players being the winner. At the start, simulations are random where actions are chosen randomly for both players. At each simulation, for each game state of that simulation, corresponding values are stored. This value of a game state (node) represents the frequency of occurrence of this node and frequency of how many of these occurrences ...