Joint trajectory and offloading optimization in UAV-assisted MEC via federated multi-agent reinforcement learning and potential fields

Abstract

Unmanned Aerial Vehicles (UAVs) assisted mobile edge computing (MEC) is characterized by flexible deployment, high mobility, and dynamic coverage. It facilitates an efficient execution of latency-sensitive tasks in scenarios such as emergency rescue and dynamic computility support, thereby demonstrating significant application prospect. However, joint scheduling of computility and task is still an open issue in optimizing task efficiency and UAVs’ energy consumption. To address this problem, we propose an UAV-assisted MEC framework based on federated multi-agent reinforcement learning (MARL) and potential fields (PF), which jointly optimizes UAV trajectories and task offloading strategies to minimize age of information (AoI) under latency and energy constraints. The decision-making process of multiple UAVs is to be modeled as a Partially Observable Markov Decision Process (POMDP) and to be solved by using a distributed federated MARL architecture. An adaptive federated collaboration model is designed for periodic parameter sharing based on credit allocation to enhance UAV collaboration and to alleviate partial observability. Additionally, a deep reinforcement learning (DRL) trajectory planning algorithm based on PF to enhance agents’ environment perception and decision-making ability. Experimental results show the effectiveness and feasibility of our proposed framework. It outperforms several existing RL-based approaches in terms of data freshness, task efficiency, and other key metrics while demonstrating strong adaptability in dynamic and complex MEC environments.