Multi-agent TD3 for HAPS-assisted Multi-UAV aerial edge computing

Aerial edge computing serves areas without terrestrial infrastructure. We study a high-altitude platform station (HAPS)-assisted network of unmanned aerial vehicles (UAVs), where ground devices partially offload tasks. To minimize a latency-energy-deadline cost under mobility and capacity constraints, we adopt multi-agent twin delayed deep deterministic policy gradient (MATD3): decentralized UAV actors set trajectories, offloading ratios, and computing priorities, and a decision map derives feasible computing and bandwidth allocations. Under nominal and hard-task loads, MATD3 has the lowest mean cost, 15.4%–16.9% below multi-agent proximal policy optimization and 2.3%–7.5% below multi-agent deep deterministic policy gradient, the latter within run-to-run variation.