The Interplay between Computation and Communication

Abstract

In this thesis, a comprehensive exploration into the integration of communication and learning within the massive Internet of Things (mIoT) is undertaken. Addressing one of the fundamental challenges of mIoT, where traditional channel estimation methods prove inefficient due to high device density and short packets; initially, a novel approach leveraging unsupervised machine learning for joint channel estimation and signal detection is proposed. This technique utilizes the Gaussian mixture model (GMM) clustering of received signals, thereby reducing the necessity for exhaustive channel estimation, decreasing the number of required pilot symbols, and enhancing symbol error rate (SER) performance. Building on this foundation, an innovative method is proposed that eliminates the need for pilot symbols entirely. By coupling GMM clustering with rotational invariant (RI) coding, the model maintains robust performance against the effects of channel rotation, thereby improving the efficiency of mIoT systems.

This research delves further into integrating communication and learning in mIoT, specifically focusing on federated learning (FL) convergence under error-prone conditions. It carefully analyzes the impact of factors like block length, coding rate, and signal-to-noise ratio on FL's accuracy and convergence. A novel approach is proposed to address communication error challenges, where the base station (BS) uses memory to cache key parameters.

Closing the thesis, an extensive simulation of a real-world mIoT system, integrating previously developed techniques, such as the innovative channel estimation method, RI coding, and the introduced FL model. It notably demonstrates that optimal learning outcomes can be achieved even without stringent communication reliability. Thus, this work not only achieves comparable or superior performance to traditional methods with fewer pilot symbols but also provides valuable insights for optimizing mIoT systems within the FL framework.