Machine Learning-based Applications for Open Radio Access Networks in B5G

Abstract

The evolution of mobile networks toward open and disaggregated architectures has been accelerated by the introduction of Open RAN. This paradigm shift enables network operators to integrate components from multiple vendors, leading to increased flexibility and innovation. However, significant challenges in resource management and service quality optimization have emerged due to the decentralized nature of these networks.

In this thesis, three novel machine learning approaches are proposed to address fundamental challenges in Open RAN systems. The theoretical foundations of distributed learning frameworks are investigated and extended to enhance network performance while maintaining vendor interoperability.

A FRL framework for indoor power control is first presented. Through this approach, optimal policies are collaboratively learned by multiple network controllers. The framework is developed with particular attention to shared spectrum scenarios, where traditional orthogonal resource allocation becomes impractical due to increasing network density. A novel reward mechanism based on first-quartile throughput is implemented to ensure fair service distribution. Significant improvements in user throughput are demonstrated compared to traditional methods, while the system's ability to adapt to various indoor propagation environments is enhanced through federated learning.

A federated meta-learning system for traffic steering between LTE and 5G NR technologies is then developed. The proposed framework is designed to handle the complex trade-offs between coverage, capacity, and service requirements in multi-RAT environments. The system's effectiveness is demonstrated through its handling of diverse traffic patterns, including time-critical applications and high-throughput services, while maintaining the quality of service requirements across different network conditions.

Finally, a xApp conflict resolution method utilizing policy distillation is introduced. A novel distillation framework is developed to consolidate knowledge from multiple vendor-specific xApps while preserving their unique optimization objectives. The method successfully addresses the challenge of conflicting network management decisions through a unified policy approach. Substantial reduction in network outages is achieved compared to traditional O-RAN conflict mitigation schemes, while Open RAN's architectural principles of openness and interoperability are preserved.

Extensive simulations and comparative analyses are conducted to evaluate the proposed frameworks against both conventional and machine learning-based solutions. Through these contributions, the theoretical foundations of distributed learning in wireless networks are not only advanced, but practical implementations addressing real-world deployment challenges in Open RAN systems are also provided. The results demonstrate that machine learning techniques, when properly adapted, enhance network performance while preserving Open RAN's fundamental principles of openness and interoperability in next-generation mobile networks.

Date of Award	4 Feb 2025
Original language	English
Awarding Institution	University of Bristol
Sponsors	Turkish Ministry of National Education
Supervisor	Robert J Piechocki (Supervisor) & George Oikonomou (Supervisor)