Novel AI Architectures for Intelligent Telecom Network Control

Sammanfattning

The evolution of mobile communication networks toward 6G is characterized by a massive growth in connectivity scale and complexity. Traditional network management, which relies on static and manual rules, is increasingly inadequate for handling the interactions of billions of devices. Large Language Models (LLMs) offer promising higher-level reasoning and planning capabilities, but their integration into mission-critical telecommunications infrastructure is hindered by a governance gap, as the risk of hallucinations and the probabilistic nature of LLMs conflict with the deterministic requirements of network protocols. This thesis proposes a hybrid architecture in which an LLM functions as a procedural interface between human operators and the management of a network. The architecture employs a control loop where sub-symbolic models handle anomaly detection and a symbolic grounding layer consisting of an RDF-based Knowledge Base and a deterministic Procedure Catalog ensures that an LLM's outputs remain within safe and policy-compliant boundaries. A central contribution of this work is the implementation of an episodic memory loop based on the Reflexion paradigm. This allows the Cognitive agent to generate verbal reflections of failed network interventions and adapt its strategy for subsequent anomalies, enabling the system to exhibit emergent problem-solving behaviours not explicitly encoded in the defined static rule sets. The architecture was evaluated using an an ns-3 O-RAN simulation environment across four experimental phases, including a baseline and an ablation study of the episodic memory. The evaluation involved deterministic traffic surges and signal blockages in a simulated network. The results demonstrate that the full Reflexion-based architecture reduced the number of latency anomalies, defined as downlink latency exceeding a 40 ms threshold, by approximately 33% compared to an ablated configuration without episodic memory. The system also successfully demonstrated that it was capable of intent merging, where it balanced long-term administrative goals (such as energy efficiency) with the immediate requirements of reactive anomaly recovery. These results indicate that grounding an LLM in a symbolic, deterministic layer can support safer, auditable autonomous network control while still permitting emergent adaptation.