Research On Predictive Resource Allocation Scheme For On-road Driving Scenarios

With the development of intelligent transportation systems,a variety of innovative in-vehicle services have emerged.In order to enhance network capacity,a number of “stacking” technologies,such as large-scale multiple-in multiple-out,and air-space integration,have received widespread attention.Nonetheless,such technologies need to be implemented at the high cost of energy consumption and infrastructure deployment.To this end,a natural question arises as to how to save radio resources while improving the network performance based on the existing vehicular communication infrastructure.In addition,for vehicular networks,the high mobility of vehicles leads to significant changes in channel gain on both large and small scales.Due to the random distribution of vehicles,the combinations of the channel gains experienced by individual vehicular users sharing radio resources are more diverse.This poses a great challenge to efficient radio resource management in vehicular networks.Leveraging big data analysis technology,vehicle trajectory prediction has been achieved with an accuracy of 97%.In this sense,the future large-scale channel gain to be experienced by the vehicular users can also be accurately predicted with the aid of the radio environment map.As such,the predictive resource allocation(PRA)algorithm based on long-term channel prediction has gained the attention of national and international scholars.However,existing work only demonstrates the potential benefits of PRA in terms of energy saving and quality of service(Qo S)improvement.The inherent characteristics of the services as well as the differentiated requirements on Qo S and quality of experience(Qo E)evaluation methods are not fully considered in the algorithm design.This means that existing PRA algorithms are open to further optimization.To this end,in order to cater to the 6G vision of customized service,this research aims to address three typical types of services in vehicular networks,combining the technologies of network slicing,video adaptive coding,and semantic communication,formulates the PRA algorithms into different mathematical problems considering their individual characteristics,and proposes the corresponding solutions based on different underlying theories and research methods including Martingalebased statistical delay analysis method,deep reinforcement learning,and Markov approximation.The major contributions of this research are as follows.(1)Taking the Video on Demand(Vo D)service as an example,this research proposes a double-subslice parallel-transmission PRA algorithm for the services with non-real-time nature and real-time data transmission rate requirement.In this algorithm,to boost the average spectral efficiency of the system,some packets are scheduled to be transmitted in advance when the users are experiencing good channel conditions.Given that the delay requirement of such pre-cached packets is greatly relaxed compared with that of online watched packets,the Vo D slice is further divided into two subslices to transmit online video contents and pre-cached video contents,respectively,and parallelly.Moreover,with consideration of the random small-scale fading,this research proposes a Martingale-based estimation method for a specified video stream with a specific statistical delay requirement.In this method,the randomness of the wireless channel and scheduling process is embedded into the required minimum time-frequency resources under a given average channel gain.Based on this method,the research designs a three-dimensional video transmission reconfiguration algorithm associated with users,prediction time slot,and subslice,which is used to guarantee Qo S while optimizing spectrum efficiency based on the long-time channel prediction in a large time scale.To instantiate this three-dimensional transport strategy,the research designs a small-time-scale resource allocation algorithm based on utility theory,which is responsible for completing the subslice scheduling and physical resource block allocation in each transmission time interval.(2)For the live streaming service with real-time nature and real-time data transmission rate requirement,this research proposes a Qo E-oriented predictive adaptive video transmission algorithm.In view of the fact that Qo E is a metric evaluated from a long term.multi-user adaptive video transmission algorithm is essentially a high-dimensional matrix optimization problem.Such a problem is hard to solve via numerical analysis methods when considering the three mutually coupled metrics in the Qo E evaluation.In this sense,this research resorts to reinforcement learning(RL),where the problem is first transformed into a Markov Decision Process(MDP),and the optimal long-term video quality allocation strategy is determined step by step.Explicitly,in order to reduce unnecessary video quality switching,while optimizing the users’ total video quality and guaranteeing the smoothness of video playback,the agent needs to be aware of the future channel gain variation of the users.With this in mind,this research introduces an auxiliary time scale to record the large-scale channel gain of driving vehicular users so as to enable the agent to capture user mobility information.Given that the partial dynamics of the considered MDP model can be determined immediately by the current action,the post-decision state(PDS)is introduced in RL.Then,this research derives the Bellman equation about PDS value and designs the neural network accordingly to improve the training efficiency and performance of the strategy.(3)For download services where only the end-to-end delay is required,this research proposes a task-oriented semantic-aware predictive cooperative multi-relay transmission algorithm that follows the future network vision of the semantic and green communication.Given that existing deep learning-based semantic communication,which only considers simple channel fading models,is difficult to apply to complex dynamic vehicular networks,this research introduces a knowledge graph(KG)-assisted semantic communication system.Therein,semantic coding and decoding are performed based on KG generation and KG inference techniques,respectively.In addition,from the perspective of transmission,this research models the knowledge graph as a weighted directed graph,which records the semantic importance and data size of each semantic element viewed from the semantic level and technical level,respectively.Meanwhile,this research identifies three predictable parameters about individual vehicle trajectories to perform the following predictive analysis and decision-making.First,to achieve proper semantic extraction,a closed-form expression of the achievable throughput within the maximum acceptable delay supported by the current vehicular network is derived considering inter-vehicle interference,mutual constraints between cascaded forwarding paths,and other factors.Then,for the selected semantic representation,this research jointly optimizes energy efficiency and semantic transmission reliability.In particular,mutually coupled predictive relay selection and semantic unit assignment are jointly formulated as a combinatorial optimization problem,and a low-complexity algorithm is designed based on Markov approximation to provide a real-time solution.