Theories And Methods Of Resource Allocation Optimization In Fog Radio Access Networks For Vehicular Scenarios

To improve traffic efficiency and road safety,reliable interaction and efficient processing should be enabled among vehicles,road infrastructures,and cloud platforms through wireless networks.Existing cellular-based vehicular networks with cloud/edge computing are limited in terms of the network capacity and edge computation resource.As a result,they are unable to meet the performance requirements of vehicle-infrastructure cooperative message distribution,multi-vehicle cooperative perception,and edge computation offloading.By utilizing centralized/distributed communication collaboration and cloud-edge-end computing collaboration,fog radio access networks(F-RANs)are able to effectively enhance the communication capacity and reduce the endto-end delay.However,in the vehicular scenarios with dynamic wireless environments and service loads,how to design efficient resource allocation methods to fully capture the potentials of F-RANs is still a hot and difficult topic in academia.Hence,in this dissertation,the theories and methods of resource allocation optimization in F-RANs arc investigated to guarantee the performance requirements of vehicle-infrastructure cooperative message distribution,multi-vehicle cooperative perception,and edge computation offloading for vehicular scenarios.Firstly,considering the reliability requirement of vehicleinfrastructure cooperative message distribution service is hard to be guaranteed in a dynamic environment,a distributed mode selection and communication resource allocation method is proposed based on federated deep reinforcement learning(DRL).Secondly,to achieve a tradeoff between the sensing performance and the service load for multi-vehicle cooperative perception,a communication and computation resource allocation optimization method is proposed by jointly optimizing vehicle clustering and sensing message selection.Thirdly,to address the challenges of time-varying load and constrained edge computation resource for edge computation offloading,two-timescale end-edge-cloud orchestrated communication and computation resource allocation is investigated.The main contents and contributions are summarized below.1.Communication Resource Allocation Optimization for Vehicle Infrastructure Cooperative Message DistributionConsidering the fast-varying channel state information between vehicles is unavailable for centralized resource allocation schemes,the reliability requirement of vehicle-infrastructure message distribution service is difficult to be guaranteed in a dynamic environment.In this dissertation,a distributed mode selection and communication resource allocation optimization scheme is proposed based on federated DRL,which improves the reliability of vehicleinfrastructure cooperative message distribution through autonomous decision of vehicles.Firstly,due to its sequential nature,the joint mode selection and communication resource allocation problem is modeled as a Markov decision process and then solved in a distributed manner with DRL.Secondly,a clusterbased federated learning scheme is proposed to address the limited training data in the proposed distributed algorithm.In this scheme,vehicles are clustered using the graph partitioning method,and individual vehicles inside a cluster perform federated learning with the assistance of fog access points for periodic model aggregation.Simulation results show that when the line-of-sight state of the channel between vehicles is time-varying,the proposed distributed method can improve the reliability by 7%～20%.2.Communication and Computation Resource Allocation Optimization for Multi-Vehicle Cooperative PerceptionConsidering the high service load of multi-vehicle cooperative perception,it is difficult to balance the sensing performance and the service load with existing communication and computation resource allocation methods.This dissertation proposes a joint vehicle clustering,sensing message selection,communication,and computation resource allocation method,that effectively reduces the service load while improving the sensing performance.Firstly,the sensing performance is defined as the combination of messages’spatial-temporal values and service delay.A joint cluster formulation,message selection,communication,and computation resource allocation problem is formulated to maximize the vehicles’ sensing performance,which is further decoupled into two subproblems based on the different timescales.Secondly,for the joint vehicle clustering and message selection sub-problem,on account of unpredictable spatialtemporal value and cooperation of vehicles,it is reformulated as a multi-agent stochastic game and then solved with an attention-aided multi-agent deep reinforcement learning algorithm.Thirdly,the swap matching and interior-point methods are utilized to iteratively optimize the communication and computation resources.Simulation results show that when the vehicle density is high,the proposed method can effectively reduce the service load and improve the sensing performance,and the proposed joint communication and computation resource allocation algorithm can reduce the delay by 20%.3.Communication and Computation Resource Allocation Optimization for Edge Computation OffloadingConsidering the existing edge computation offloading is limited by the edge computation resource,it is difficult to guarantee the delay requirements with dynamic load.In this dissertation,a two-timescale cloud-edge-end orchestrated communication and computation resource allocation optimization method is proposed,which combines the dynamic vehicular incentives and the cloud-edge-end collaborative computation offloading to reduce the service delay with dynamic load.Specifically,the two-timescale cloud-edge-end communication and computation resource allocation optimization problem is established,based on the fluctuating timescale differences between the service load and the channel state information.On a large timescale,to extend the edge computation resource,the network incentivizes some vehicles to offer their computation resource based on the long-term service load.A Stackelberg game-based dynamic vehicle incentive problem is formulated with the cloud server as the leader and multiple vehicles as the followers,and then an iterative algorithm is proposed to achieve the Stackelberg equilibrium of computation resource pricing and reservation.On a small timescale,given the cloud-edge-end resource in F-RANs,the joint offloading mode selection,communication,and computation resource allocation problem is transformed into a multi-agent stochastic game,and then solved by lenient multi-agent DRL.Simulation results show that the proposed method can reduce the delay by 23%,when the service load is high.In order to efficiently support the vehicle-infrastructure cooperative message distribution,multi-vehicle cooperative perception,and edge computation offloading,in this dissertation,the theories and methods of resource allocation optimization in F-RANs for vehicular scenarios are investigated.This research can provide theoretical guidance for the planning,deployment,and performance optimization of future vehicular networks.