Mobility and Resource Management in O-RAN with Online Meta-Learning

Modern cellular networks are tasked to deliver guaranteed performance for a wide array of users that increasingly demand higher throughput and reliability, lower latency, seamless connectivity, ubiquitous coverage, energy efficiency, fairness, and security, to name a few. To meet these demands, networks are becoming increasingly complex, combining diverse deployments and multiple radio access technologies that are envisioned to extend beyond 5G. At the same time, the resources (e.g., spectrum, energy, capacity) needed to serve all users are limited and expensive; and control decisions, such as mobility and resource allocation, often require trading throughput with other user-perceived performance metrics such as lower delays, signaling/communication costs, and failure risks.

In these environments where traffic patterns change rapidly, signal qualities fluctuate unpredictably and cost/availability of resources is uncertain, it becomes apparent that static control rules and legacy mechanisms built on heuristics are poorly suited. In this context, the evolution of mobile network architectures, particularly the emergence of open Radio Access Network (RAN), represents a necessary and enabling change. The O-RAN Alliance, for example, is a global initiative aimed at softwarizing and standardizing RANs to improve their performance, reduce costs, and lower the entry barrier for a broader vendor ecosystem. It enables scalable, datadriven control loops that can be implemented centrally by intelligent controllers and enforced at different time scales, namely, near-real-time (near-RT) and non-real-time (non-RT). In this way, it becomes possible to embed online learning solutions in the RAN itself, where data are collected and used for effective and robust learning.

This dissertation responds to these challenges by developing online (meta-) learning algorithms for two coupled control layers in O-RAN: (i) mobility management (via user-cell association and traditional/conditional handovers) and (ii) resource allocation (via threshold, non-RT policies) for virtualized base stations. Online learning provides a principled way to make sequential decisions under uncertainty, and online meta-learning enables the system to combine various (online) learners, each tailored for different environments, achieving both effectiveness, which translates to high performance under all conditions, as well as robustness, which ensures this high performance without knowing precisely the conditions. All proposed methods deliver operation guarantees under all conditions (from stationary to even adversarial dynamics), as well as practical gains on country-scale operator data and O-RAN-compatible testbeds.