UC Davis

Explosive growth of mobile networking and IoT demands efficient and reliable service for massive wireless systems. With limited radio resources, multi-input-multi-output (MIMO) technologies are successfully utilizing spatial diversity to substantially improve spectral efficiency. When considering large scale deployments, managing radio resource is more important than ever to service all these devices with appropriate quality of service. Moreover, in the context of performance-constrained, low-complexity devices, there is a clear need for new approaches that yield good performance with appropriate computational complexity.

In this dissertation, we study such large scale networks from a geometric perspective, in order to better manage the networks' limited resources and mitigate co-channel interference in two key scenarios: when the multi-antenna servicing node is unaware of which devices are active (uplink access control); and when it does know all active devices (user scheduling).

In the first part of this dissertation, we tackle the problem of uplink grant-based access via blind signal recovery. Different from traditional grant-free access mechanisms that use pilot signals for signal separation, we propose two blind approaches based on the Constant Modulus Algorithm (CMA) for simultaneous multiple signal recovery: a regularized CMA cost function, and a Riemannian manifold optimization framework. By characterizing the underlying geometry of these formulations, we provide theoretical convergence guarantees for CMA-based signal recovery with limited data samples. The resulting algorithms provide successful signal recovery with high probability and reasonable computational load.

On the other hand, user scheduling is a combinatorial, NP-hard problem that has been long eluded optimal solutions. In MIMO networks, users groups with low co-channel interference correspond to groups that show high spatial channel diversity. In the second part of this dissertation, we propose a new two-step paradigm for MIMO user scheduling. First, unsupervised learning identifies which devices experience similar channel conditions (i.e., low spatial diversity) and would incur high co-channel interference if they were to share resources. By clustering in the Grassmannian manifold, spatial similarity is inherent to the geometry and is easily computed in a global sense. We then leverage these learned features to assign users into low CCI groups that avoid pairings of users from the same cluster. The resulting similarity-assisted scheduling yields increased spectral efficiency and better user quality of service across design parameters for large number of users, compared to a direct scheduling mechanism.