Research On Machine Learning Based Vehicular Edge Computing

With the development of vehicular technology and Internet of Vehicles(IoV),many vehicular intelligent applications have emerged,which have stringent requirements on service quality,user experience and system overhead.As an effective way to improve the performance of on-board intelligent applications,Vehicular Edge Computing(VEC)has been highly anticipated.By applying Mobile Edge Computing(MEC)technology to the IoV scenario,VEC can effectively reduce the delay and energy consumption,which improves the user experience of on-board intelligent applications.However,the mobility of vehicles will cause frequent handovers between vehicles and computing servers,which increases extra delay and affects user experience.At the same time,the non-IID local data will also affect distributed model training efficiency in IoV scenario.In addition,the false messages uploaded by unreliable or malicious vehicles will influence the service quality of on-board intelligent applications,and the uneven distribution of vehicles will aggravate the difficulty of screening false messages.This thesis focuses on the difficulties in IoV scenario and carries out the following three aspects:In order to tackle the longer computing unloading delay caused by the unstable wireless connection between moving vehicles and computing servers in IoV scenario,based on the analysis of the unloading,processing and return process of computation tasks during the movement of vehicles,the expression of the total delay of computation tasks is derived.Then,the average offloading delay of computation tasks minimization problem is proposed.Secondly,a reinforcement learning computation offloading scheme is proposed to minimize the average offloading delay.The road side unit(RSU)is regarded as the learning agent and makes computation offloading decisions.Compared with vehicle-centric agent,RSU can be identified as agent to collect interactive information in static manner during the learning process.In addition,the moving buses are selected as mobile computing server.The mobility of buses is regular and predictable,since they follow the prescribed routes and timetables.In the absence of prior knowledge,the utility functions of different buses are constructed according to the observed offloading delay sequence,and the sub-optimal unloading strategy is learned.Simulation results show that,compared with the existing computation offloading schemes,the average offloading delay of the proposed scheme can be reduced by at least 11%.Aiming at the problem that the distributed model training efficiency is reduced due to the non-IID local data in IoV scenario,the similarity between the non-IID problem of vehicular local data and the ZSL problem is analyzed.Then,a federated learning vehicular edge model training scheme is proposed.According to the EMD(earth mover’s distance)distance derivation,the effectiveness of the proposed scheme is proved.By introducing a set of high-dimensional semantic attributes,each class label is represented as a certain attribute vector.The original multi-class classification model is decomposed into multiple attribute classifiers.According to the EMD distance between local data distribution and global data distribution,the cloud server will select the appropriate model weights of vehicular clients for global aggregation.Simulation results show that,compared with the existing model training schemes,the proposed scheme can significantly improve the model training efficiency of FL on non-IID local data and the average prediction accuracy can be improved by about 30%.To address the low map update accuracy problem when only unreliable private vehicles are used for crowd sourcing map updating in IoV scenario,a high-precision map updating scheme assisted by mapping vehicles is proposed.In this scheme,to improve the updating accuracy of maps,a reasonable scheduling strategy is developed for reliable mapping vehicles.Firstly,based on the analysis of movement of vehicles and wireless transmission process,the expression of map update accuracy is derived.Then,a sequential decision making problem is formulated to maximize the map update accuracy under the condition that the system overhead is less than a preset threshold.In order to obtain the optimal mapping vehicle scheduling strategy,the algorithm based on deep reinforcement learning(DRL)is proposed.The proposed algorithm can not only solve the difficult problems caused by the large dimension of system state and action space in the VEC system,but also obtain the approximate optimal strategy by learning the historical experience without prior knowledge of the environment.Simulation results show that,compared with the existing schemes,the proposed scheme can effectively improve the map updating accuracy by about 20%.