CSO-DRL: A Collaborative Service Offloading Approach with Deep Reinforcement Learning in Vehicular Edge Computing

Abstract

In vehicular edge computing, vehicles move along the road and request the services from the nearest edge servers with low latency. Due to the limitation of computation capacity of vehicular devices, the services should be offloaded on RSUs equipped with edge servers to provide service with low latency. Noticed that the location of service offloading may affect the service requesting delay directly, and it may exist some interrelationship between interacting services; all of these are rarely considered in recent studies. To address such problems, we propose a collaborative service offloading approach with deep reinforcement learning in vehicular edge computing named CSO-DRL. Our approach first divides the road segments by k-means-based algorithm through analyzing the trajectory data of vehicles, and then the offloading location is determined by observing the vehicle running status. Secondly, the interacting services are discovered by a parallel frequent pattern-based algorithm efficiently. Furthermore, the collaborative service offloading algorithm is presented by the DDPG model for offloading the interacting services, which can minimize the service requesting delay and data communication delay between interacting services. Finally, the efficiency of the algorithm is evaluated by real-world data-based simulation experimental evaluations. The results show our algorithm can obtain a lower delay than other baseline algorithms in searching for the optimal service offloading strategy.