Abstract: Efficient resource scheduling is critical for improving the effectiveness of emergency response in aerial flood disaster rescue. This paper proposes an innovative approach by developing a Multi-agent Reinforcement Learning with Artificial Intelligence Feedback (MA-RLAIF) model that incorporates feedback from a Large Language Model (LLM) to enhance decision-making capabilities during emergency operations. First, based on the characteristics of flood scenarios, the study defines specific rescue tasks and constructs a dynamic task allocation model using Multi-agent Reinforcement Learning (MARL). Then, to further optimize model performance, an automated feedback mechanism is designed using LLM and integrated into the MARL framework, resulting in the MA-RLAIF model. Experimental results demonstrate that, compared with traditional task allocation strategies, the MA-RLAIF model achieves significantly better performance in terms of task distribution efficiency and rescue effectiveness in flood scenarios, highlighting the critical role of LLMs in supporting complex management decision-making.

Key words: aerial emergency rescue, multi-agent reinforcement learning, feedback optimization, large language model, dynamic task allocation