An Investigation of Underground Rescue Scheduling with Multi-agent Reinforcement Learning

Underground rescue scheduling poses substantial challenges due to unknown environment and low visibility. The task inherently involve multi-agent systems and pose significant complexities. In this paper, we focus on simulating underground rescue missions by designing a novel simulator with multiple robots and survivors. To test the solubility of the environment, we investigate the performance of Multi-Agent Reinforcement Learning(MARL) algorithms in this new environment, a comprehensive evaluation of three widely used MARL algorithms, alongside an exploration of various reward structures. Also, we use a curiosity-driven QMIX algorithm to investigate the efficiency of curiosity mechanism on this task. The experimental results indicate that the reward fully represent rescuing reward yields the optimal performance, the three MARL algorithms can solve the rescue task successfully, and the curiosity mechanism can solve the task better.