Chapter 31: Machine Learning for Wireless Communications

Learning Objectives

• Formulate channel estimation and signal detection as supervised learning problems, and design neural-network architectures that map pilot observations to channel estimates or transmitted symbols
• Explain the end-to-end autoencoder paradigm in which transmitter and receiver are jointly optimised as encoder-decoder networks, and analyse the learned constellations that emerge under various channel models
• Describe the deep unfolding methodology that converts iterative algorithms (ISTA, ADMM, gradient descent) into trainable neural networks with per-layer learnable parameters, and quantify the convergence and sample-efficiency advantages of LISTA over classical sparse recovery
• Formulate wireless resource allocation (power control, scheduling) as a Markov decision process and apply tabular Q-learning and deep reinforcement learning to maximise network utility
• Explain the federated learning framework (FedAvg), analyse its convergence under non-IID data across base stations, and discuss privacy and communication-efficiency trade-offs
• Compare model-based (deep unfolding, algorithm unrolling) and data-driven (black-box NN) approaches in terms of sample efficiency, generalisation, interpretability, and computational cost, and formulate design guidelines for selecting the appropriate paradigm