Intent-driven Radio Access Network Management Using Advanced Machine Learning Techniques

Abstract

Future Sixth-Generation (6G) Radio Access Networks (RANs) are expected to support an unprecedented scale of heterogeneous services, massive device connectivity, and diverse Quality-of-Service (QoS) requirements. It will significantly increase traffic diversity and operational complexity. Managing such large-scale, dynamic RAN environments through manual configuration or rule-based automation is neither scalable nor sustainable. In this context, intent-driven RAN management emerges as a paradigm shift, where high- level human objectives are translated into machine-executable control policies. However, realizing this translation in practice requires learning-based intelligence, as the mapping from intents to control policies must adapt to non-stationary traffic conditions, multi-objective QoS constraints, and uncertain network dynamics. As AI-native networking becomes a foundational principle of 6G, advanced machine learning techniques are expected to enable intelligent, autonomous, and adaptive control in RAN infrastructure. In this thesis, we aim to develop a scalable, intent-driven, AI-native RAN management framework that translates high-level operator objectives into safe and effective control actions in dynamic multi-service environments. First, we develop a Hierarchical Reinforcement Learning (HRL)-based intent-driven RAN management scheme that adapts to dynamic network conditions by selecting and sequencing appropriate RAN applications for execution. Even though later in this thesis we move towards Large Language Model (LLM)-based intent processing and attention- based HRL for better performance optimization, value-based HRL remains constrained by reward engineering complexity, exploration inefficiency, and sensitivity to distributional shifts in dynamic RAN environments. This motivates a shift to a sequence modeling paradigm through hierarchical Decision Transformers (DTs), which generates goal-based policies rather than explicit reward maximization. The LLM component is further enhanced through Quantized Low-Rank Adapter (QLORA)-based fine-tuning to improve domain adaptation and intent interpretation accuracy. A key limitation of the hierarchical DT-based RAN controller is its reliance on high- quality offline Reinforcement Learning (RL) trajectories for training. As a result, performance may degrade under distribution shifts and previously unseen RAN conditions. To overcome these constraints, the thesis advances toward a hierarchical Online Decision Transformer (ODT) integrated within an Agentic AI architecture. By utilizing online updates and closed-loop feedback, the proposed Agentic architecture supports autonomous policy refinement and QoS drift mitigation during off-peak periods to establish a practical foundation for zero-touch RAN management. Finally, to reduce the computational overhead associated with transformer-based attention mechanisms, the thesis adopts state-space sequence modeling through Decision Mamba and its hierarchical extension, Hierarchical Decision Mamba (HDM). The resulting Agentic framework integrates HDM and a fine-tuned LLM to coordinate inter-slice provisioning, intra-slice scheduling, and self-healing within a unified, computationally efficient control hierarchy for 6G RAN systems. The key findings of this thesis are summarized as follows. First, while a range of machine learning algorithms can optimize RAN performance, hierarchically organized and logically structured learning frameworks are particularly well-suited for intent-driven RAN management. Second, and most importantly, this thesis demonstrates that Generative Artificial Intelligence (GenAI)-based control paradigms can outperform conventional RL- based approaches in intent-driven RAN automation. Extensive system-level simulations show that the proposed hierarchical GenAI frameworks consistently improve throughput, latency compliance, energy efficiency, and fairness across diverse traffic conditions when compared with baseline algorithms.