| With the rapid development of wireless communication network,the number of mobile users and data traffic both maintain a sustained and rapid growth in recent years,which makes machine learning technique be increasingly integrated with wireless network.On the one hand,machine learning can be applied to optimize wireless network:for instance,with machine learning technique,operators can establish a prediction model by efficiently and accurately extracting the relationship between network service measurements and Quality of Experience(QoE),so as to optimize the resource allocation.On the other hand,the optimization of wireless network can greatly improve the performance of machine learning system:for instance,in federated learning system,the global model aggregation suffers from the limited communication bandwidth,while the unfavorable environment in the signal propagation process also has a significant effect.This thesis starts from the bidirectional relationship between machine learning and wireless network,and the main work and contributions are as follows:(1)We propose a resource allocation model for the dynamic network based on supervised learning classification.Specifically,we use logistic regression classifier to guide the resource allocation and construct the optimization problem by minimizing the time-average probability of new user complaints.We then leverage the Lyapunov optimization technique to transform the stochastic optimization problem and propose an online resource allocation algorithm.Simulation results verify the effectiveness of our proposed resource allocation algorithm.(2)Based on over-the-air computation,we further present an intelligent reflecting surface(IRS)aided federated learning system for efficient fast model aggregation.This is achieved by jointly optimizing the device selection,receiver beamforming vector at the base station,and the phase shift matrix at the IRS.This problem turns out to be a sparse and low-rank optimization problem,for which we propose an alternating optimization framework supported by a novel difference-of-convex-functions(DC)programming algorithm.Simulation results demonstrate the algorithm’s superiority in achieving desirable learning performance by selecting more admitted devices. |
