A game strategy model in the digital curling system based on NFSP

Abstract The digital curling game is a two-player zero-sum extensive in continuous action space. There are some challenging problems that still not solved well, such as the uncertainty of strategy, large tree searching, and use amounts supervised data, etc. In this work, we combine NFSP KR-UCT for games, where uses two adversary learning networks can automatically produce be used searching We propose reward mechanisms to make reinforcement converge quickly. Experimental results validate proposed method, show strategy model reach Nash equilibrium.