Navarro, Annalisa (2025) Autonomous Orchestration of Next-Generation Networks Using Reinforcement Learning. [Tesi di dottorato]

Documento PDF navarro_annalisa_38.pdf Download (3MB)

Abstract

Next-Generation Networks (NGNs), including 5G and the upcoming 6G, mark a paradigm shift in telecommunications by introducing unprecedented requirements in terms of throughput, latency, scalability, and reliability. To meet these challenges, networks must evolve into programmable, adaptive, and autonomous infrastructures, enabled by paradigms such as Software-Defined Networking (SDN), Network Function Virtualization (NFV), and the integration of Artificial Intelligence (AI) and Machine Learning (ML). Among ML approaches, Reinforcement Learning (RL) has emerged as particularly promising for dynamic network resource management. However, the inherent complexity and lack of trustworthiness have hindered its widespread adoption in real-world telecom systems. This thesis addresses the challenge of designing scalable, trustworthy, and safe RL-based frameworks for the automated management and control of advanced communication networks. This thesis proposes an RL-powered SD-WAN traffic engineering framework that dynamically enforces network policies over heterogeneous overlays. Then this approach is extended by introducing a decentralized Multi-Agent RL (MARL) architecture to tackle scalability in multi-site scenarios, and further enhanced with cooperative strategies to improve overall policy compliance. Furthermore, an explainability layer for RL-based network control is introduced by employing surrogate models to provide interpretable insights into RL decisions, thereby fostering trust and operational adoption. Finally, a novel framework for safe RL deployment in the Radio Access Network (RAN) is presented that, by leveraging offline training and evaluation, ensures performance improvements in handover management while mitigating risks associated with online learning and testing.

Downloads

Actions (login required)

Modifica documento