A bandwidth-fair migration-enabled task offloading for vehicular edge computing: a deep reinforcement learning approach

Vehicular edge computing (VEC), which extends the computing, storage, and networking resources from the cloud center to the logical network edge through the deployment of edge servers at the road-side unit (RSU), has aroused extensive attention in recent years, by virtue of the advantages in meeting the stringent latency requirements of vehicular applications. VEC enables the tasks and data to be processed and analyzed in close proximity to data sources (i.e., vehicles). VEC reduces the response latency for vehicular tasks, but also mitigates the burdens over the backhaul networks. However, how to achieve cost-effective task offloading in VEC remains a challenging problem, owing to the fact that the computing capabilities of the edge server are not sufficient enough compared to the cloud center and the uneven distribution of computing resources among RSUs. In this paper, we consider an urban VEC scenario and model the VEC system in terms of delay and cost. The goal of this paper is to minimize the weighted total latency and vehicle cost by balancing the bandwidth and migrating tasks while satisfying multiple constraint conditions. Specifically, we model the task offloading problem as a weighted bipartite graph matching problem and propose a Kuhn-Munkres (KM) based Task Matching Offloading scheme (KTMO) to determine the optimal offloading strategy. Furthermore, considering the dynamic time-varying features of the VEC environment, we model the task migration problem as a Markov Decision Process (MDP) and propose a Deep Reinforcement Learning (DRL) based online learning method to explore optimal migration decisions. The experimental results demonstrate that our strategy has better performance compared to other methods.