Introduction to Monte Carlo Tree Search

by Jeff Bradberry

The subject of game AI generally begins with so-called perfect information games. These are turn-based games where the players have no information hidden from each other and there is no element of chance in the game mechanics (such as by rolling dice or drawing cards from a shuffled deck). Tic Tac Toe, Connect 4, Checkers, Reversi, Chess, and Go are all games of this type. Because everything in this type of game is fully determined, a tree can, in theory, be constructed that contains all possible outcomes, and a value assigned corresponding to a win or a loss for one of the players. Finding the best possible play, then, is a matter of doing a search on the tree, with the method of choice at each level alternating between picking the maximum value and picking the minimum value, matching the different players' conflicting goals, as the search proceeds down the tree. This algorithm is called Minimax.

The problem with Minimax, though, is that it can take an impractical amount of time to do a full search of the game tree. This is particularly true for games with a high branching factor, or high average number of available moves per turn ...

Continue reading