Research On Edge Computing Task Offloading For Internet Of Vehicles

As an important technology of intelligent transportation,the Internet of Vehicles has been strongly supported by national policies in recent years and developed rapidly.The on-board applications of computing intensive and timely extension and sensitivity are emerging one after another.However,it is difficult for a single vehicle with limited computing resources to meet the delay requirements of on-board applications.Combining the Internet of vehicles with edge computing can effectively solve this problem.The edge computing is used to realize task unloading.The tasks beyond the computing capacity of the vehicle are unloaded to the edge server equipped with the roadside unit to assist the vehicle calculation,which can effectively meet the delay requirements of on-board applications.However,due to the dynamic and uncertain environment of the Internet of Vehicles,the traditional heuristic algorithm cannot fully adapt to the changes of the environment.Therefore,this dissertation adopts the deep reinforcement learning algorithm to design and calculate the task unloading scheme,so that the vehicle can make the unloading decision adaptively and judge the task unloading node and task transmission volume intelligently,effectively reduce the task delay of the vehicle and obtain the best long-term return.Further,considering that the unloading scheme of static optimization tasks often requires the acquisition of all task information,but due to the unknowability,variability and complexity of the network transmission environment,the real bandwidth and task quantity information cannot be obtained in advance until the unloading task is completed,resulting in the inaccurate decision scheme.Therefore,this dissertation predicts the next task in advance.To make efficient unloading decision,the response delay of the task is further reduced,and the unloading performance of the system is enhanced.The research work and corresponding achievements of this dissertation are summarized as follows:Firstly,aiming at how to make optimal unloading decision in dynamic environment,this dissertation proposes a task unloading architecture based on edge intelligence.In order to reduce the system delay,this dissertation establishes the communication and calculation model between the autonomous vehicle and the edge server,and transforms the problem into a long-term total task processing delay minimization problem.To solve this problem,an MDPCO algorithm was designed to minimize the processing delay of vehicle computing tasks.In this dissertation,the unloading process is described as Markov decision process,and the corresponding incentive mechanism is developed to obtain the optimal unloading decision under the minimum task processing delay.The simulation results showed that compared with other baseline schemes,the MDPCO algorithm could achieve the lowest total task processing delay in different scenarios.Secondly,to improve the unloading efficiency by predicting the next task in advance,this dissertation proposes a TP-DRL system based on deep reinforcement learning and task prediction.The system includes prediction module and decision module.The prediction module uses SRU algorithm to predict the task quantity at the next moment by collecting the historical task quantity information.The decision module uses TD3 algorithm to predict the task quantity as the input of the model and makes unloading strategy in advance according to the prediction results.The simulation results show that,under different task quantity,the delay of the task with prediction is lower than that without prediction.Compared with other baseline schemes,the TP-DRL scheme has better optimization effect on reducing the total delay of task processing.