关键词: 移动边缘计算, 多智能体强化学习, 无人机避障, 任务卸载, 三维轨迹优化

Abstract: Abstract: The rapid growth of Internet of Things (IoT) and mobile terminal devices has led to significant challenges in network latency and energy consumption due to the massive computational tasks. To address this, a multi-unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) system was studied, where UAVs provide efficient computational offloading services for ground users. In the system, multiple UAVs collaborate to process computational tasks in a three-dimensional space with complex obstacles. To minimize the weighted sum of the maximum task completion latency for all users and the total system energy consumption, a joint optimization of users' discrete offloading decisions and UAVs' continuous three-dimensional trajectories is performed. To solve this mixed (discrete-continuous) optimization problem, a heterogeneous multi-agent deep reinforcement learning-based UOUM (User Offloading and UAV Mobility Co-optimization) algorithm was proposed. The algorithm constructs a heterogeneous multi-agent framework, where user offloading decisions are designed for discrete action space, and UAV trajectory optimization is designed for continuous action space, addressing the optimization challenge of mixed action spaces. A differential reward mechanism is introduced to precisely quantify the marginal contributions of each agent's strategy and allocate rewards, solving the multi-agent credit assignment problem. Additionally, artificial potential field constraints are integrated to transform obstacle avoidance requirements into differentiable safety potential functions, ensuring UAVs avoid obstacles and improving training efficiency. Simulation results show that, in various test scenarios, UOUM outperforms three benchmark algorithms (user offloading optimization only, UAV trajectory optimization only, and heterogeneous multi-agent reinforcement learning) in terms of latency, energy consumption, and system cost, validating its effectiveness and reliability. The UOUM algorithm achieves significant improvements in latency optimization, energy control, and obstacle avoidance safety, demonstrating strong environmental adaptability.

Key words: Mobile edge computing, multi-agent reinforcement learning, UAV obstacle avoidance, task offloading, three-dimensional trajectory optimization