Online Learning and Optimization in Communication Networks

Abstract

In this thesis, I propose new online learning and optimization approaches to evaluate and design communication networks by investigating unknown system variation, feedback delay, and communication efficiency over time. The results of this thesis provide new analytical insights and design guidelines that will help to improve future communication networks. In the first part of this thesis, we consider periodic decision updates for constrained Online Convex Optimization (OCO). This is motivated by many practical wireless communication systems, which only permit a fixed decision update over multiple time slots, while the environment changes between the decision epochs. We propose an efficient online algorithm, which employs a periodic virtual queue together with aggregated gradient descent for decision updates. We evaluate the performance of theproposed algorithm in a large-scale multi-antenna system shared by multiple wireless service providers. In the second part of this thesis, we study OCO with long-term constraints in the presence of multi-slot feedback delay. We propose an efficient online algorithm, which uses a double regularization together with a penalty mechanism on the long-term constraint violation, to tackle the asynchrony between information feedback and decision updates. We apply the proposed algorithm to solve a general network resource allocation problem. In the third part of this thesis, we exploit over-the-air computation to jointly optimize the training of the global model and the analog aggregation of the local models over time for federated learning (FL). We propose an efficient algorithm to adaptively update the local and global models based on the time-varying communication environment. The trained model is both channel- and power-aware, and it is in closed form incurring low computational complexity. We derive performance bounds on both the computation and communication performance metrics. In the last part of this thesis, we encourage temporal similarity in the decision sequence over time to control the communication overhead in online distributed optimization. We propose an efficient algorithm, which uses a tunable virtual queue together with a modified Lyapunov drift analysis to jointly consider computation and communication over time. We apply the proposed algorithm to enable communication-efficient FL.