Offline PSRO with Max-Min Entropy for Multi-Player Imperfect Information Games

Game theoretic learning for the blueprint strategy is the central topic of offline game solving. Offline learning is one typical data-driven paradigm that learns from logged dataset. This paper provides one offline PSRO algorithms with max-min exploration that can solve extensive-form games with imperfect information. After constructing the offline dataset with expert interaction experience, the offline PSRO is trained with best response generated from the entropy regularized reinforcement learning method. This work will provide a reference for offline game solving, the experimental results demonstrated the effectiveness of our framework.