Distributed Machine Learning In Radio Access Networks-Modeling And Optimization

6G(The Sixth-generation Mobile Communication Network)intends to introduce endogenous intelligence capabilities and multi-node collaboration to achieve ubiquitous intelligence provision.Artificial intelligence(AI)has been widely regarded as a key technology to promote the development of 6G intelligent networks.On the one hand,the popularization of graphics processing units and dedicated AI chips has promoted the reduction of the AI computing power cost,which is of importance for developing endogenous AI technology in 6G networks.On the other hand,the continuous deployment and application of AI algorithms such as reinforcement learning and deep learning in wireless networks have demonstrated their advantages in decision-making,learning,and predicting under dynamic network environments.Nonetheless,due to the rapid rise of 6G services and the explosive growth of smart devices,integrating AI into 6G networks is still confronted with numerous challenges,particularly when deployed in radio access networks.Notably,running AI technologies and processing AI tasks consume significant network resources,which are scarce in the radio access network.Furthermore,wireless environment factors such as the instability of the wireless environment,vulnerability of the wireless channel,and resource heterogeneity in the radio access networks,may impact the model learning performance.Therefore,it is crucial to deeply explore optimization theoretical frameworks and key technologies by integrating multi-dimensional network resources to support the 6G intelligent network.Distributed machine learning(DML)has been widely regarded as one of the most promising AI technologies to support 6G network intelligence.In DML,agents cooperate to build consensus by sharing lightweight experiences.Compared with traditional centralized machine learning methods,DML can not only maintain a globally shared learning model to solve large-scale machine learning problems but also reduce the network burden.Although DML offers these attractive and valuable advantages,it still faces many challenges in pushing forward intelligent radio access networks.On the one hand,when the user equipment(UE)transmits the local models,the transmission resource consumption can be quite large.On the other hand,the heterogeneity of wireless communication systems,wireless channel impairments,and dynamic nature of the wireless environment,siginificantly degrade learning performance.In addition,traditional DML is not suitable for 6G networks where both the privacy concerns and heterogeneous data are highly concerned,as:(1)the central server controls the computing nodes as well as the data within them,and(2)the data on different computing nodes are usually evenly,independently,and identically distributed.Therefore,to develop intelligent radio access networks,it is necessary to consider whether the resource-limited radio access networks can support network intelligence and achieve the required learning performance.Meanwhile,resource scheduling mechanisms are also needed to cope with the substantial growth of UEs and services.As a special DML,federated learning(FL)can effectively solve the problems of data heterogeneity,resource heterogeneity,privacy and security issues in traditional DML.Therefore,this dissertation takes FL as an example,based on tools such as stochastic process,ensemble distillation,and deep reinforcement learning,to study the modeling and optimization for distributed learning in radio access networks.The main contributions include the following four parts:(1)Modeling and analysis of distributed learning in radio access network: This part studies the relationship between the model accuracy of distributed learning models and the network resources consumed in radio access networks,which theoretically answers whether radio access networks can support network intelligence.Namely,how many resources are needed to support FL-enabled radio access networks,and how much learning performance can be achieved when given specific resources.Specifically,an FL analysis model is constructed by assuming the time-space domain Poisson distribution.Based on the analytical model,the mathematical relationship between model accuracy,available computing resources,communication resources,and channel quality is clarified.Numerical results validate the theoretical modeling and analysis of the work.(2)An adaptive quantization scheme for distributed learning in wireless networks:This part studies an adaptive quantization scheme based on ensemble distillation.The scheme allows UE to learn personalized quantization models with different quantization levels,structures,and tasks,where UEs can adjust their quantization levels according to instantaneous radio resource constraints and channel impairments.The convergence of the proposed adaptive quantization scheme is analyzed theoretically,and its effectiveness in term of learning accuracy is demonstrated through simulation.(3)A dynamic clustering scheme for distributed learning in wireless networks: This part investigates a dynamic straggler-aware clustering scheme based on distributed learning by improving the balanced iterative reduction and clustering algorithm.Specifically,each cluster model sent to the central server undergoes a global aggregation,after which the aggregated global model is transmitted to the user device.Numerical results show that the proposed Fe DSC scheme significantly outperforms some conventional solutions in terms of accuracy and convergence time in typical cases.(4)Case study: hybrid federated learning based access control for radio access networks.This part studies a case for applying distributed learning in radio access network(RAN)slicing,where an access control mechanism based on hybrid federated deep reinforcement learning(HDRL)is presented,aiming at improving network throughput and reducing hand-off costs.Specifically,HDRL is composed of two-layer model aggregation including horizontal aggregation and vertical aggregation.Numerical results show that,in typical cases,the proposed HDRL scheme for access control significantly outperforms some conventional solutions in terms of network throughput and communication efficiency.The evolution of 6G intelligent networks represents an inevitable progression toward constructing a digital China.Distributed learning has been widely regarded as a fundamental technological approach for facilitating the deployment of 6G intelligent networks,where the modeling and optimization in radio access networks hold paramount importance.This dissertation establishes a valid theoretical foundation for the utilization and implementation of distributed learning in radio access networks by modeling a theoretical analysis model of distributed learning-enabled radio access networks.Additionally,effective distributed learning algorithms are designed and optimized to account for the dynamic characteristics of wireless networks.Furthermore,based on the theoretical analysis model,this dissertation addresses the control access problem under the three-layer association architecture of UE-base station-slice,where the results demonstrate the effectiveness of applying distributed learning to radio access networks.