Reinforcement Learning with TensorFlow

The Monte Carlo Tree Search

So, how is Monte Carlo Tree Search different to minimax's approach, and how is it able to plan ahead in a highly complex game of Go, which has an immense number of potential counter moves? The MCTS builds a statistical tree that looks exactly like a game tree but whereas game trees or minimax have game positions, that is, game states in the node of the directed graph, in MCTS the node of the directed graph is the quantity for that game state that tells us the number of successful simulations (that is, moves that lead to a win at the end of the game) with respect to the total number of simulations that game state has been through. Therefore, the higher the number of simulations, the more nodes get the chance to ...