Physics-Based Machine Learning for Wireless Communications

In recent years, Machine Learning (ML) approaches, have been largely applied to deal with scenarios which are not completely understood or cannot be analytically modelled. The success of these data-driven techniques lies in their ability to learn a general patter from many known examples. However, some drawbacks are present, and they may limit the performance of purely data-driven strategies. First, a large amount of training data is required to adapt to all possible complex scenarios, and often such amount of information is not available. Furthermore, the learned patterns are not interpretable and cannot be used to develop further knowledge about the examined context. Finally, they may learn spurious relationships between variables since the physical properties of the system are not considered.
To face these issues, the so-called Physics-Based ML paradigm has recently emerged. It aims to combine the potential of both data-driven and model-driven approaches, learning from data while considering also relevant prior knowledge. In this way, the search space of the learning algorithm is reduced to physically consistent models. Consequently, better generalizability of the models is ensured with a lower amount of data required for the training.
Our project aims to increase the performance of wireless communications systems with the help of Physics-Based ML techniques. In many scenarios, in fact, theory-based approaches do not lead to satisfactory results in telecommunications. Optimization problems may be too complex to be solved analytically, and Information Theory does not offer a complete description when multi-connected network topologies are considered. Thus, Physics-Based ML can offer the suitable tools to extend the limits of communication systems by building on top of the well-established knowledge. Two potential scenarios where Physics-Based ML could play an interesting role are identified.
In wireless communications, channel estimation is the process of identifying the properties of the wireless link between a transmitter and a receiver. The knowledge of the channel, i.e. the Channel State Information (CSI), plays a crucial role for achieving efficient and reliable transmissions. However, the propagation of the electromagnetic fields in practical contexts is too complex to be deterministically described, and statistical model may not be sufficient. Thus, the physics behind the electromagnetic propagation, and information-theoretical notions can be used to build efficient Physics-Based ML models, able to learn the channel properties, aware of the underlying physics.
The second scenario regards Intelligent Reflecting Surfaces (IRSs) communications, where reconfigurable passive elements are deployed to properly reflect the signals for improving spectrum and energy efficiency. The fundamental challenge is to configure the reflecting elements to maximize the signal power delivered at the receiver. In this context, it is often difficult to find the optimal design choice by solving optimization problems due to the complexity offered by the network. Thus, Physics-Based ML are critical in answering to this challenge.
Relevant EPSRC research areas: "Artificial intelligence technologies", "Digital signal processing", "RF and microwave communications".