Edge-Coordinated Multi-Head PPO for Multi-UAV Sensing-Assisted Computation Systems

Abstract

This paper investigates a phased sensing-assisted mobile edge computing system composed of multiple unmanned aerial vehicles (UAVs). A framework is proposed to operate in three sequential phases: local user sensing, global state aggregation, and centralized decision making for distributed offloading. To achieve efficient load balancing across these phases, a mixed-integer nonlinear programming problem is formulated to minimize the maximum bottleneck delay. To address the tight coupling between continuous spatial resource variables and discrete task assignments, an edge-coordinated multi-head proximal policy optimization (ECMH-PPO) scheme based on deep reinforcement learning is proposed to simultaneously optimize 3D trajectories, offloading frequencies, and task associations. A custom action transformation mechanism strictly enforces physical constraints, eliminating invalid exploration. Simulations demonstrate the ECMH-PPO scheme significantly outperforms baselines by circumventing local optima and mitigating the straggler effect. It reduces the maximum execution delay to 27.07s under baseline settings and to 22.00s with increased resources.