Hierarchical decision-making for UAVs’ game via LLM enhanced multi-agent reinforcement learning

For the multi-UAV game missions, intelligent agents face challenges in high-dimensional state processing, dynamic situational awareness, and cooperative decision-making that conventional methods cannot address effectively. This paper proposes a novel Large Language Model-enhanced Multi-agent Deep Reinforcement Learning scheme (LLM-MATD3) with a hierarchical decision-making architecture comprising an environment modeling layer, LLM task planner, and task-enhanced MATD3 network. The scheme first employs a hierarchical task decomposition mechanism that transforms high-level game directives from LLMs into continuous embedding vectors through task embedding techniques. Then, the LLM task planner analyzes global game situations and assigns differentiated tasks to UAVs, overcoming the limitation of traditional MATD3 frameworks that rely solely on local observations. Simulation results show significant improvements in win rate, remaining rate, and collaborative efficiency for UAVs, providing a innovative reference approach for multi-agent systems in complex game environments.