Competition-Awareness Partial Task Offloading and UAV Deployment for Multitier Parallel Computational Internet of Vehicles

Vehicular edge computing is poised to meet the requirements of emerging applications in Internet of Vehicles (IoV) by offloading computation tasks from resource-limited vehicles to edge. However, the space-time-dynamic offloading demands of vehicle users (VUs) can hardly be satisfied only by road side units (RSUs) due to their fixed resource deployment and incomplete coverage. To this end, in this article, we design a multitier IoV system, where RSU, parked cars, and unmanned aerial vehicles (UAVs) serve as edge platforms to offer computing services. To fully utilize the multitier resources, the tasks generated by VUs can be split into multiple parts and executed in parallel on local processors as well as edge servers. Under this arrangement, we formulate a joint UAV deployment and partial task offloading problem to minimize the system cost, which includes processing delay, energy consumption, and rental price. We then develop a heuristic UAV deployment method to optimize the coverage of multitier network. Moreover, a distributed task offloading approach based on multiagent deep reinforcement learning is proposed to achieve cooperative decision makings and load balancing, thereby overcoming the adversarial competition among VUs. Experimental evaluations reveal that compared to state-of-the-art schemes that rely on a centralized controller, the proposed approach achieves superior performance with higher implementation efficiency while avoiding extra information exchange overhead.

I. Introduction

By Promising efficient vehicle-to-everything connection, real-time information processing, and enhanced traffic safety, Internet of Vehicles (IoV) is surfaced as a vital enabler for intelligent transportation services in the forthcoming 6G era [1], [2]. To realize this, vehicles are outfitted with communication transceivers as well as computation units, with the aim of supporting emerging vehicular applications, e.g., autonomous driving [3]. However, these applications generate heterogeneous tasks with stringent latency requirements and diverse resource demands. It is challenging for vehicle users (VUs) to promptly process the arriving tasks only utilizing their on-board computing capability.