Model-based learning for high-dimensional wireless systems

The field of wireless communications has traditionally relied on signal processing techniques, based on mathematical models, to describe and optimize wireless transmission of information. However, when the underlying assumptions of these models are not fully satisfied, conventional methods often exhibit suboptimal performance. In recent years, machine learning techniques have emerged as a promising alternative to address these limitations, owing to their adaptability and data-driven nature. Nevertheless, machine learning-based approaches also entail additional challenges, particularly in terms of interpretability and computational complexity.

This thesis investigates the use of the model-based machine learning paradigm in wireless communication or sensing systems. This paradigm seeks to combine the strengths of both model-based and machine learning approaches, while mitigating their respective weaknesses. In particular, it leverages the mathematical models used in model-based approaches to structure, initialize, or optimize learning methods. 

Three application areas are explored within this thesis. First, it is proposed to use a physical propagation channel model to structure a neural architecture that learns the location-to-channel mapping. This neural network is subsequently used to improve the performance of traditional radio-localization techniques. Second, as practical wireless systems are inherently affected by hardware impairments, it is proposed to apply the model-based machine learning paradigm to mitigate their impact. Third, the growing number of antennas in modern cellular networks leads to a rapid increase in the dimensionality of the channel, thereby significantly increasing computational complexity. To mitigate this issue, this thesis investigates the use of a model-based dimensionality reduction technique known as channel charting.