The Application of Reinforcement Learning in Amazons

Computer game is a challenging and active research field. The success of AlphaGo has attracted people's attention and enthusiasm in this field, its following version, AlphaZero, provides a general solution for complete information game. However, different kinds of game have different rules, how to apply this framework to specific games is a problem worthy of study. This paper applies AlphaZero in Amazons game, using a convolutional neural network to provide evaluation and prior probability for current board which make up for the Monte Carlo tree search. Its improvement depends self-play entirely. Experiments show that the method is effective.