A two-layer system architecture for Unmanned Aerial Vehicle (UAV)-assisted offloading: Game-theoretic allocation and CTDE-MADDPG control

Abstract

UAV-assisted offloading presents an effective solution for areas with limited infrastructure or temporary congestion. Yet jointly optimising task allocation and trajectory control is difficult due to heterogeneous GU demands and UAV energy/kinematic limits. We propose a two-layer architecture that decouples discrete allocation from continuous control. In the allocation layer, a responder-only external incentive gates accept/reject decisions in an auction–predation–trade scheme, without altering the potential, learning targets, or metrics. This mechanism induces monotone ascent of the potential and terminates in finitely many steps at a pairwise-stable (Nash) allocation. The control layer adopts centralised training with decentralised execution (MADDPG) with a centralised critic and task-specific actor heads to generate 3D trajectories subject to flight and safety constraints. This partition confines the combinatorial complexity to the outer allocation, resulting in a lighter on-board inference than a monolithic policy that jointly handles both allocation and planning. In a 3-UAV/7-GU simulator with obstacles and heterogeneous task sizes, the proposed method achieves higher global reward (2.02 vs 1.91) and shorter completion time (92 s vs 109 s) than a reproduced game-theoretic baseline. Compared with end-to-end MADDPG without prior allocation, it also converges faster and more stably.