Research On Intelligent Access Control And Scheduling Mechanism For Wireless Networks

The number of mobile network applications and access devices have been growing at an exponential rate in recent years due to the rapid development of communication technology.According to statistics and forecasts given by Cisco,global mobile devices will grow from 880 million in 2018 to 1.31 billion in 2023.Moreover,5G devices and connections will account for more than 10%of global mobile devices and connections.In addition,with the wide application of emerging vertical industries such as Smart City,Internet of Things,Mobile Edge Computing,user requirements become more diversi-fied.Therefore,in order to meet the large-scale growth of mobile traffic and diversified user requirements,several novel techniques,for example,heterogeneous cellular network(HetNets),multi radio access technology,network slicing technology,etc.,have been in-troduced into 5G and beyond.Although beneficial for the improvement of network per-formance,the introduction of the novel techniques makes the mobile networks become much complicated and more heterogeneous.With the increasing complexity and heterogeneity of wireless network environment,research on access control and scheduling mechanism in HetNets becomes more diffi-cult.Due to the time-varying nature of heterogeneous networks and the randomness of user behavior,traditional model-based resource allocation methods(such as optimization methods and game theory)lack certain adaptive and generalization capabilities.In het-erogeneous networks,dense deployment of cellular small base station(SBS)can improve the network capacity but is constrained by the severe co-frequency interferences between SBSs.Moreover,when multi-RATs(Radio Access Technologies)are introduced to ex-pand network capacity,the cross-domain resource management is more difficult due to differences in transmission protocol and performance.In addition,the introduction of Radio Access Network(RAN)slicing also needs to reconsider the scheduling of wire-less resources.Overall,in the future wireless networks,due to the time-varying nature of the network environment(e.g.,channel state,user behavior,etc.)and the multi-level overlapped coverage of heterogeneous networks,great challenges have taken place to the resource allocation of wireless networks.Therefore,it is necessary to adopt more intelli-gent and adaptive solutions to solve the resource allocation problem in the future complex HetNets,so as to better cope with the explosive growth of mobile traffic and diversified user needs.Machine learning(ML)is an efficient and intelligent method to solve dynamic decision-making problems in complex network environment.ML which has powerful data process-ing and decision-making abilities is an effective tool to make decisions in uncertain envi-ronment.ML method,such as deep learning(DL)and reinforcement learning(RL),can be used to design appropriate access strategies by analyzing the characteristics of user behav-ior and service in HetNets,with the aim to optimize the long-term network performance.Therefore,in this dissertation,we exploit ML and multi-agent game theory to solve access control,resource scheduling,power allocation in 5G networks,including the following is-sues:(1)Self-imitation learning based inter-cell interference coordination in autonomous HetNets;(2)Smart multi-RAT access based on multi-agent reinforcement learning;(3)Intelligent resource scheduling for 5G radio access network slicing;(4)Federated coop-eration and augmentation for power allocation in decentralized wireless networks.The main contributions and innovations of this dissertation are summarized as follows:We first investigate the inter-cell interference coordination(ICIC)for plug-and-play SBSs in autonomous HetNets,where the SBSs agilely schedule sub-channels to individual users at each Transmit Time Interval(TTI)with aim of mitigating interferences and maxi-mizing long-term throughput by sensing the environment.We model the decision process of SBSs as partially observable Markov decision process(POMD),and then formulate the interference coordination problem as a model-free based inverse reinforcement learning(IRL)problem.In the following,we propose a self-imitation learning(SIL)method to solve the IRL problem from behavior imitation and reward function approximation.In addition,in order to cater for the plug-and-play operation mode of indoor SBSs,the SIL is initialized according to the SINR,and a nested training scheme is adopted to overcome the“slow-start" problem of the learning process.Numerical results reveal that the overall network throughput of SIL can be improved by up to 19.8%when compared with other known benchmark algorithms.Then,we focus on user association and resource scheduling in multi-RAT HetNets.We propose a Smart Aggregated RAT Access(SARA)strategy with aim of maximizing the long-term network throughput while meeting diverse traffic Quality of Service(QoS)re-quirements.As RAT selection and resource allocation in multi-RAT access can be deemed as two sequential processes,we propose an SMDP based hierarchical decision framework(HDF)to solve the j oint RAT selection and resource allocation problem.Numerical results reveal that the network throughput can be maximized while meeting various traffic QoS requirements with limited number of searches by using our proposed SARA algorithm.Another attractive feature of SARA is that a trade-off between the solution optimality and learning time can be readily made by terminating the search of MCTS according to the time constraint.Next,we investigate resource scheduling in RAN slicing.In this dissertation,RAN slicing schedules resources to customized services with the aim of maximizing the re-source multiplexing gains while guaranteeing a certain degree of isolation among slices.Resource scheduling strategy(RSS)is the key issue of RAN slicing.RSS needs to allocate limited spectrum resources to users with different QoS requirements in dynamic network environment.On this basis,we propose an intelligent-RSS(iRSS)for RAN slicing,where DL and RL respectively perform large time-scale resource allocation and on-line resource scheduling in a collaborative way.Finally,the performance advantages of iRSS are veri-fied by simulation experiments in different scenarios.Finally,we study the power allocation problem in distributed HetNets based on user privacy protection.As a distributed ML algorithm,federated learning(FL)enables users to update the shared decision model by uploading weight coefficients or gradients,while keeping their dataset on the local device.In this dissertation,we study the power alloca-tion problem in distributed HetNets,and propose a FL based cooperation and augmenta-tion(FL-CA)method.In addition,in order to solve the over-fitting problem caused by data leakage,FL-CA uses WGANs(Wasserstein Generative Adversarial Networks)for data augmentation,so as to reduce the correlation between different user datasets.The simulation results show that FL-CA has better robustness and power allocation accuracy than other benchmark algorithms.