Research On Cooperative Offloading Algorithm Of DNN Model In Vehicular Edge Computing

Vehicular edge computing(VEC)is a new computing paradigm that has great potential to improve the computing capability of intelligent connected vehicles,and can support computation-intensive,energy-consuming and latency-sensitive intelligent on-board applications based on deep neural network(DNN).Existing research results show that jointing devices and edge nodes to offload inference tasks through DNN task partitioning is an effective way to reduce latency and energy consumption.However,due to edge nodes in the VEC have limited computing resources and need to provide computing resources for multiple vehicles at the same time,the computing resources allocated to each task may be different,which will lead to a different optimal partition point of the DNN inference task.Design an efficient VEC on-board DNN task collaborative offloading algorithm,so that the vehicle can dynamically and adaptively select the optimal partition point of the DNN task for computation offloading according to the computing resources of edge nodes,thereby further reducing the processing latency of the task.In view of the challenges encountered in the collaborative offloading of DNN tasks in VEC scenarios,this paper proposes two collaborative offloading algorithms of on-board DNN tasks based on edge computing to accelerate the processing speed of tasks.The specific content is as follows:1.Dynamic resource allocation algorithm for jointing vehicle-edge cooperative offloading in DNN task.The algorithm considers that the MEC server dynamically allocates computing resources to multiple vehicles within the coverage of RSU,and a jointing vehicle-edge DNN inference offloading scheme Jo VEO is proposed.For the on-board DNN task with delay constraints,the NP-hard optimization problem of dynamic allocation of computing resources in MEC server and selection of optimal partition point for DNN task is defined.The goal of this problem is to minimize the overall latency of all vehicles in the system.Subsequently,a fast convergence algorithm based on chemical reaction optimization was designed,and the proposed scheme was simulated and evaluated.The experimental results show that the Jo VEO scheme is superior to the other three benchmark schemes in terms of overall latency and failure rate,and is close to the optimal scheme.2.Autonomous learning based V2 V collaborative offloading algorithm of DNN task.This algorithm proposes a V2 V collaborative scheme for DNN tasks in the dynamic and changing network environment,which makes full use of the computing resources of resource idle vehicles and reduces the average processing latency of DNN tasks.Aiming at the challenge of obtaining global information in a dynamic V2 V environment,an dynamic adaptive offloading decision algorithm based on online learning is proposed,which is called OLAO,and an improved UCB algorithm is used to improve the efficiency of online learning.A large number of simulation experiments have evaluated the proposed scheme and algorithm.The experimental results show that OLAO is superior to the other three benchmark schemes in terms of average processing latency for DNN tasks.At the same time,the learning efficiency of the scheme based on UCB algorithm is much higher than that based on ε-Greedy algorithm scheme.