| With the increasing demand for digitization and intelligence in social life,the sixth generation wireless communication system(6G)needs to be deeply integrated with artificial intelligence technology to usher in a new era of "Internet of Intelligence".In emerging application scenarios such as smart cities,autonomous driving,and industrial Internet of Things(IoT),traditional centralized machine learning requires a large amount of communication resources to upload a massive amount of local data samples to a single edge server or cloud computing center for model training,which not only brings extremely high transmission delay but also poses a risk of leaking user personal privacy.With the enhancement of the computing power of IoT devices,a more practical approach is to allow clients to use local data for distributed learning,and then collaboratively train a powerful global model by exchanging local parameters or gradients.This not only achieves real-time data processing but also effectively solves the problem of data islands while protecting user privacy.This distributed machine learning method of“data does not move but model moves" is also called federated learning(FL).However,when applying FL in resource-constrained 6G IoT systems,it is important to consider how to ensure the effectiveness and accuracy of model transmission,and how to improve the adaptability and scalability to heterogeneous data and devices.Aiming at the above research problems,this thesis conducts studies from the aspects of network architecture design,transmission mechanism adaptation,and radio resource management.This thesis optimizes the transmission capacity,aggregation accuracy,convergence speed,and energy consumption of the proposed FL system.The main contributions and novelties of this thesis are summarized as follows:Firstly,during the model training process of FL,local devices need to ex’change model parameters frequently with the edge server.However,in IoT systems,the limited time-frequency resources and the unreliability of wireless channels may have a great impact on the aggregation and accuracy of the global model.Therefore,this thesis proposes an FL framework based on over-theair computing(AirComp)and reconfigurable intelligent surface(RIS).Specifically,by using the superposition characteristics of wireless channels to complete the weighted aggregation of multiple local models,this thesis presents an AirComp-based model uploading and aggregation scheme.In order to reduce the aggregation distortion of the global model and improve its prediction accuracy,this thesis uses multiple RISs to adjust the uplink wireless channel.To solve the dual-criteria optimization problem of minimizing the mean square error and maximizing the number of participating devices,this thesis derives the optimal solutions of user transmit power,base station(BS)receive coefficients,and RIS reflection matrix,and designs a client selection algorithm based on the difference-of-convex programming(DCP).Simulation results show that,compared with benchmark schemes without RIS or with one RIS,the proposed solutions effectively reduce the aggregation distortion of the global model and accelerate the convergence rate of FL.Secondly,considering that future 6G networks may have both communication users focused on data transmission and learning users focused on model computation,this thesis proposes a RIS-assisted communication-learning integrated network architecture that integrates non-orthogonal multiple access(NOMA)-based wireless communication users and over-the-air federated learning(AirFL)users into a unified framework.It is worth noting that the RIS in the proposed architecture plays an important role in flexibly adjusting the signal decoding order of heterogeneous users,which is the key to improving the spectral efficiency and maintaining the stable operation of the considered system.For NOMA users,their communication performance can be measured by traditional data rates.For AirFL users,this thesis defines a performance metric to effectively measure the AirComp efficiency.Then,by jointly optimizing user transmit power,BS receive scaling,and RIS discrete phase shifts,this thesis formulates a mixed-integer programming problem to maximize the hybrid rate(a weighted sum of communication and computing rates).For the subproblems of transmit power allocation and receive coefficient control,this thesis proposes iterative algorithms based on the DCP and successive convex approximation(SCA),respectively,and derives the optimal closed-form solutions under simplified conditions.For the subproblem of RIS discrete phase shift design,this thesis uses a“relaxation-then-quantization"approach to first convert the non-convex optimization problem into a semi-definite programming problem,and then proposes a RIS discrete phase shift design algorithm based on semidefinite relaxation.Simulation results show that the proposed communicationlearning integrated network can effectively serve two different types of users using shared time-frequency resources with the help of RIS.When the interference coefficient is small,the proposed joint design framework achieves higher spectrum efficiency than independent design schemes.Thirdly,to solve the problem of limited coverage of existing RIS,this thesis leverages a simultaneous transmitting and reflecting reconfigurable intelligent surface(STAR-RIS)that supports both signal reflection and refraction to further improve the performance of the propossed communication-learning integrated network.Compared to reflection-only RIS,STAR-RIS provides omnidirectional coverage,relieving the geographical restriction that users and the BS need to be located on the same side of the RIS.This thesis quantifies the optimality gap between the actual loss value and the optimal loss value of AirFL after multiple communication rounds,and discusses the specific effects of these factors such as diminishing learning rate and gradient aggregation error.Then,this thesis minimizes the optimality gap by jointly designing the transmit power of users and the coefficient of STAR-RIS to accelerate model convergence.For the mixed-integer non-linear programming(MINLP)problem,a trust regionbased SCA method is first proposed to determine users’ transmit power,and then a penalty function-based semi-definite relaxation method is used to optimize the passive beamforming of STAR-RIS.Simulation results show that,compared with existing schemes,the proposed scheme can improve the spectrum efficiency and learning performance of the communication-learning integrated network under different data distributions.Finally,although the communication-learning integrated network proposed has good compatibility with computation-heterogeneous IoT devices,the computing and storage resources at the BS have not been fully utilized.To address this issue,this thesis proposes a semi-federated learning(SemiFL)framework that combines centralized learning(CL)and FL to fully leverage all computing resources of the entire network(including local devices and the BS)to collaboratively train a powerful global model.By allowing some devices to upload privacy-insensitive data,the performance of the global model can be effectively improved.Subject to the quality-of-service requirements of CL users and the aggregation distortion threshold of FL users,this thesis formulates an MINLP problem that minimizes the total transmit power of all users,so as to reduce energy consumption of IoT devices and extend network lifetime.To solve this non-convex optimization problem,this thesis provides the optimal solution for the transmit power,and designs a penalty function-based SCA algorithm for the STAR-RIS configuration.Simulation results show that:1)the proposed resource management algorithm can effectively reduce users’ communication power consumption,and 2)the proposed SemiFL scheme requires less communication overhead than that of traditional CL scheme,and has faster convergence speed and higher prediction accuracy than that of traditional FL scheme. |
