Communication-Assisted Multi-Agent Reinforcement Learning Improves Task-Offloading in UAV-Aided Edge-Computing Networks

Abstract

Equipping unmanned aerial vehicles (UAVs) with computing servers allows the ground-users to offload complex tasks to the UAVs, but the trajectory optimization of UAVs is critical for fully exploiting their maneuverability. Existing studies either employ a centralized controller having prohibitive communication overhead, or fail to glean the benefits of interaction and coordination among agents. To circumvent this impediment, we propose to intelligently exchange critical information among agents for assisting their decision-making. We first formulate a problem for maximizing the number of offloaded tasks and the offloading fairness by optimizing the trajectory of UAVs. We then conceive a multi-agent deep reinforcement learning (DRL) framework by harnessing communication among agents, and design a communication-assisted decentralized trajectory control algorithm based on value-decomposition networks (VDN) for fully exploiting the benefits of messages exchange among agents. Simulation results demonstrate the superiority of the proposed<br/>algorithm over the state-of-the-art DRL-based algorithms.