Digital Twin-Assisted Vehicles Edge Network Computing Offloading Scheme

Abstract

Traditional vehicle edge computing research has not fully considered the differences between vehicle tasks and edge server computing resources while often ignoring the use of deep reinforcement learning (DRL), which requires a large amount of training data, and DRL is easy to fall into the problem of local optimality. Therefore, this paper proposed a digital twin (DT) -assisted vehicle edge network computation offloading method. For the problem of DRL requiring a large amount of training data, a real-time data acquisition method based on DT is proposed for vehicle tasks and edge nodes. The differentiability of computing resources proposed a computing model based on deep neural network (DNN) partitioning and early exit mechanism, making full use of the computing resources of edge servers and vehicles, reducing the computation delay and transmission delay of tasks, Aiming at the problem that the asynchronous dominant actor-critic (A3C) algorithm in DRL is easy to fall into the local optimal solution, the dynamic baseline and ε-greedy policy are introduced based on A3C and used to select the channel to reduce the transmission delay further. Experimental results show that this scheme outperforms other comparative methods and can effectively reduce the delay of task offloading in vehicular edge computing.