| With the continuous development of automotive industry and Internet of Things,intelligent connected vehicle is not just a construct.Highly intelligent vehicles are connected to networks,resulting in a series of emerging applications and services,such as automatic driving,intelligent traffic and entertainment office,gradually participating in people’s travel.Although it makes people’s life more convenient,the appearance of these applications inevitably arouses some challenges.New intelligent applications require enormous computation resources.The growing computational ability of on-board chips is still insufficient to deal with a mass of computation-intensive vehicular applications.Now,it is an acceptable solution to offload vehicular applications to servers with abundant computation resource via wireless link.On this occasion,it is imperative to decide which servers should applications offload to and compute on,and how will the computed results be delivered back to vehicles.However,inherent mobility of vehicles and latency constraint for applications put a strain on the offloading service.Results are no longer time-efficient if applications go through a long offloading and computation.Meanwhile,vehicles may be far away from their original locations at the completion of applications,thus,it is difficult to track the current vehicular locations to deliver the computed results.Based on the above discussion,mobile edge computing（MEC）has drawn attention to researchers for providing computation ability in proximity to vehicles.This paper investigates the offloading decision about computation-intensive and latencysensitive vehicular applications on top of MEC.Main works of this paper are summarized as follows:First,a collaborative computing scheme for Internet of Vehicle via cooperation of MEC node is proposed.This paper focuses on partial offloading and vehicular applications are divided into several mutually dependent sub-tasks.Further,applications are modeled as a general task graph which is an integration of the mesh and the radial structure.To compute such applications,one application will be fully offloaded to a roadside unit（RSU）that exhibits acceptable channel condition.Then,the RSU can process all its sub-tasks or forward them to the MEC servers co-located with other RSUs for collaborative computing.At last,an RSU that can connect with the vehicle currently upon completion of computation is selected to deliver the computed results.Second,an optimization problem is formulated.According to the dependency among sub-tasks,the corresponding communication and computation models are established,obtaining the total latency of an application during the whole process of offloading.In order to increase reliability and reduce latency,this paper aims to minimize the weighted sum of the latency during the whole process of offloading and delivery failure penalty for all vehicles by jointly optimizing RSU/MEC server selection for offloading,computing,and delivery.Third,a reinforcement learning（RL）algorithm is adopted to solve the problem.Besides the problem is an integer programming problem,the dynamic nature of Internet-of-Vehicle and sub-task dependency empower the use of RL.To implement RL,the problem is transferred into a Markov decision process（MDP）.Subsequently,state space,action space and reward function are defined.To handle the high-dimensional state and action space,multi-agent deep deterministic policy gradient（MADDPG）,one of RL algorithms in multi-agent setting,is eventually adopted.Finally,the effectiveness of the proposed scheme and algorithm is verified through simulation.To explore the feasibility of the scheme,this paper compares the proposed scheme over various situations with a scheme that all sub-tasks of one application will be computed on the same MEC server.Further,comparison results for three algorithms within the proposed scheme are shown.Numerical results demonstrate that the proposed scheme together with MADDPG can gain the better performance. |
PDF Full Text Request