Hybrid-Action MADRL-Based Scheduling for Task Offloading and Trajectory Planning in Multi-UAV Systems With Latency Minimization

Abstract

Considering the limitations of a single uncrewed aerial vehicle (UAV) in capacity and coverage, cooperative UAVs assisted mobile edge computing (MEC) architecture has been popular for large-scale mobile terminals (MTs). However, the high mobility of MTs and the limited resources of UAVs lead to challenges such as imbalanced resource allocation and high task latency. Therefore, this paper proposes a hybrid-action MADRL-based scheduling framework for task offloading and trajectory planning in multi-UAV systems. Specifically, considering the predictable mobility of MTs and the limited computing resources of UAVs, a cooperative offloading framework is proposed through alternative MEC based on UAV tracking. To minimize the long-term total delay through the joint optimization of UAV trajectory control, computing resource allocation, and task migration decisions, a non-convex and high-dimensional coupled problem is formulated with constraints on the resources of UAVs and service requirements. To solve this problem, we propose an enhanced hybrid-action multi-agent deep deterministic policy gradient (HA-MADDPG) algorithm, built upon centralized training and decentralized execution (CTDE) paradigm. The developed algorithm incorporates the Gumbel-Softmax mechanism for differentiable discrete action modeling and adopts soft policy updating for convergence stability. Simulation results demonstrate that the HA-MADDPG algorithm based on multi-UAV cooperative MEC system significantly reduces system task delay, improves edge resource utilization and enhances training convergence compared to several mainstream benchmark schemes.