Multiple access and interference management for 5G-and-beyond wireless networks

Abstract

Wireless networks are shifting from the rate-centric paradigm in 1G-4G towards a new one in 5G-and-beyond, where rate-centric applications (e.g., broadband mobile access and multimedia broadcasting) and computation-driven applications [e.g., edge artificial intelligence (AI) and Internet-of-things] coexist with equal importance. The paradigm shift revolutionizes the multiple-access techniques and divides the design principle into two, which differ from the perspective of interference management. The classic rate-centric multiple access views interference as being harmful and thus avoids it by orthogonalizing simultaneous transmissions. In contrast, the emerging computation-driven multiple access sees merits in “interference” and exploits superimposed concurrent transmissions for over-the-air function computation (e.g., averaging), called AirComp. By revisiting the classic principle in new system setups and developing new solutions under the emerging principle, this dissertation contributes a set of novel multiple access techniques enabling diverse applications in 5G-and-beyond networks.

For rate-centric applications, two use cases involving both cellular and sensor networks are considered, in which a set of novel signal-processing techniques are proposed to allow large-scale multiple access under hardware constraints. Specifically, to avoid costly digital implementation of the massive multiple-input multiple- out (MIMO) in a cellular network, a hybrid (analog and digital) beamforming design framework that allows hardware-efficient space division multiple access is first proposed. Underpinning the design is an innovative idea of Kronecker decomposing the analog beamformer vector. The decomposition allows a decoupled design of the analog beamformer components for either signal enhancement or interference cancellation, while respecting the analog-circuit constraints on the beamforming coefficients. A matching channel estimation scheme under the hybrid hardware is also developed to acquire the needed channel state information (CSI). Next, consider the energy-efficient backscatter sensor networks, where the backscatter sensors with limited signal processing capabilities cannot support the algorithms for conventional multiple access and channel estimation. A so-called BackSense framework is proposed to enable CSI-free simultaneous sensing data uploading and detection. The framework features the application of advanced tools from statistical learning at the reader to resolve collided signals, and simple randomized on-off transmissions at sensors overcoming the hardware constraints.

For computation-driven applications, we developed enabling techniques for two typical use cases, i.e., wireless data aggregation for multi-modal sensor networks and federated edge learning (FEEL) for edge AI. For the first case, a framework of MIMO-AirComp is proposed for achieving vector-valued function computation over multi-modal sensing values. The framework consists of a low-complexity solution for AirComp beamforming using the theory of manifold optimization, and an efficient channel feedback scheme that enables the beamforming without es- calating overhead w.r.t. sensor population. For the second case, namely FEEL, where a global AI-model at an edge server is updated by aggregating (averaging) local models trained at edge devices. To enable fast model-aggregation and thus communication-efficient FEEL, a broadband AirComp solution, called broadband analog aggregation (BAA), is thus proposed. An in-depth analysis of the solution unveils two key fundamental tradeoffs between communication efficiency and learning performance in FEEL. Characterizing the tradeoffs yields practical insights on the design of FEEL network.