Energy-Efficient UAV-Assisted MEC With Dual-Priority Scheduling and Multi-Agent Reinforcement Learning

Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) enhances access to terrestrial MEC infrastructure when direct user-to-base-station connectivity is limited, congested, or unevenly distributed. In the proposed architecture, UAVs do not replace terrestrial infrastructure; instead, they operate as energy-aware aerial relays that collect tasks from user equipment (UEs) and forward them to terrestrial base stations (BSs) equipped with MEC servers, where computation is ultimately performed. Two challenges then become critical: the stringent energy limits of UAVs and the need to prioritize tasks according to both application urgency and geographical relevance. Existing studies typically address energy efficiency and task prioritization separately, leaving their joint impact underexplored. This work introduces a hierarchical MEC architecture where UAVs act as energy-aware relays operating under a dual-priority mechanism combining application type and geographical zone. We formulate a coupled mobility–offloading optimization problem encompassing propulsion and communication energy, multi-tier BS queues, offloading decisions, and UAV flight constraints. The problem's nonlinear and multi-agent nature motivates a reformulation as a Multi-Agent Markov Decision Process (MAMDP). To solve it, we employ Multi-Agent Proximal Policy Optimization (MAPPO) within a Centralized Training and Decentralized Execution (CTDE) framework, enabling cooperative, energy-efficient UAV mobility policies while BSs enforce priority-driven service differentiation. Simulation results show that the proposed approach considerably reduces UAV energy consumption and markedly lowers high-priority deadline violations compared to state-of-the-art Deep Reinforcement Learning (DRL) baselines, demonstrating its suitability for scalable and QoS-aware MEC deployments.