Multi-Agent Deep Reinforcement Learning based Collaborative Computation Offloading in Vehicular Edge Networks

Recently, to cope with the long communication distance and unreliability of cloud-based computing architectures, mobile edge computing has emerged as a solution with great promise. This pattern extends cloud-based services towards the vehicular edge network and enables vehicular tasks to be offloaded to intermediate Roadside Units (RSUs) directly. However, as more and more tasks are offloaded to RSUs, the computation capacity of a single RSU becomes insufficient. Without edge cooperation, overall resource utilization and effectiveness are prone to being underutilized. To address this issue, this paper investigates a collaborative computation offloading scheme where adjacent RSUs can process offloaded tasks collaboratively rather than individually. First, we explore a vehicular edge network where the bilateral synergy between RSUs is leveraged. In particular, we incorporate a price-based incentive mechanism into the resource allocation process to promote overall resource utilization. Second, considering the time-varying system conditions and uncertain resource requirements, the optimization problem is approximated as a Markov Decision Process (MDP) and extended to a multi-agent system. Finally, we propose a Multi-agent Deep deterministic policy gradient-based computation Offloading and resource Allocation scheme (MDOA) to solve the corresponding problem. Simulation results show that the proposed MDOA can not only achieve a higher long-term utility of RSUs but also have better performance than other baselines in different scenarios.