Master Multiple Real-Time Strategy Games with a Unified Learning Model Using Multi-agent Reinforcement Learning

Bo Ling, Xiang Liu, Jin Jiang, Weiwei Wu, Wanyuan Wang, Yan Lyu, Xueyong Xu

Abstract: General artificial intelligence requires an intelligent agent to understand or learn any intellectual tasks like a human being. Diverse and complex real-time strategy (RTS) game for artificial intelligence research is a promising stepping stone to achieve the goal. In the last decade, the strongest agents have either simplified the key elements of the game, or used expert rules with human knowledge, or focused on a specific environment. In this paper, we propose a unified learning model that can master various environments in RTS game without human knowledge. We use a multi-agent reinforcement learning algorithm that uses data from agents in a diverse league played on multiple maps to train the deep neural network model. We evaluate our model in microRTS, a simple real-time strategy game. The results show that the agent is competitive against the strong benchmarks in different environments.