| With the development of the Internet of Things technology,the Internet of Vehicles,as an important part of it,has developed a large number of intelligent applications.Autonomous driving is one of the key applications.Autonomous driving includes two core issues:path planning and driving decision.At present,there are many path planning methods for vehicles,which have been applied into various online map platforms.However,driving decision scheme for vehicles still need to be further studied.Therefore,this thesis proposes a driving decision scheme for autonomous driving vehicles based on the architecture of the Internet of Vehicles.Meanwhile,considering the influence of the limit of wireless channel capacity on driving decisions,some optimization schemes are proposed.The main contents of this thesis include:Based on the distributed driving scenario which is composed of central server,roadside units(RSUs)and vehicles,a hierarchical driving decision scheme is proposed.Firstly,at the basic layer,the central server and RSUs build the model for driving decision problem and solve the problem based on the sampled data of all roads.A basic decision model is obtained to ensure driving safety and efficiency.Then basic decision model,the model for driving decision problem and the method to solve the problem are submerged into the enhancement layer.On a considered road,the RSU and the vehicles driving on the covered road further optimize the decision model based on the real driving data.Finally a local enhanced decision model is obtained based on the features of the current road.Besides,the modeling of driving problem is based on reinforcement learning method,and the solution is based on the improved Asynchronous Advantage Actor-Critic(A3C)algorithm.The simulation results show that the basic decision model can preliminary guarantee driving security and efficiency,and the local enhanced decision model can further improve the decision performance.Based on the distributed architecture,vehicles need to communicate with RSU through vehicle to infrastructure(V2I)communications in the enhancement layer,in which the RSU transmits the parameters of the global model downlink and the vehicles transmit the update gradients of local model parameters uplink.However,due to the limitation of wireless communication capacity,to transmit the gradients within a certain period,the vehicles need to quantify the gradients before transmission,and the quantization error will affect the performance of the driving decision model.Therefore,two optimization methods are proposed to reduce the impacts.The two methods both improve the model by changing the combining strategies:when the RSU combines the uploaded gradients of each vehicle,the combining weight of each gradient is determined according to the position of the vehicle when it is uploaded.In the first method,the channel capacity and the number of quantization bits are firstly calculated according to the position of vehicles.Then the quantitative signal-to-noise ratio is obtained and the combining weights are obtained according to it.In the second method,the merge gradient under the ideal communication circumstance that there is no limitation of communication capacity or quantization error is taken as real gradients.Then based on the least square method,the optimization objective is to minimize the error between the weighted merging gradients and the real gradients.Then the optimal merging weight is obtained.Finally,through simulation,the ability of the proposed methods mitigating the impact of limited communication resources is verified by comparing the results of the following four situations:non-quantization error,using traditional method to combine gradients,and using the proposed two methods to combine gradients. |
