Abstract
The increasing complexity of modern networking systems, alongside the rapid integration of artificial intelligence (AI), has led to a transition from AI-driven to AI-native networking. This PhD aims to optimize intelligent networking systems from an AI-native perspective through lifecycle-aware methodologies, where AI is embedded and coordinated across all stages of the solution lifecycle.In traditional AI-driven systems, AI is treated as an add-on tool, often disconnected from the overall lifecycle of the networking solution. In contrast, AI-native networking systems embed AI as a core and intrinsic component. AI-native networking systems coordinate data collection, data transmission, model design/selection, model chaining/placement, and performance maintenance so that interdependent components can be optimized end-to-end for concrete system objectives (e.g., lower SLA violations, faster recovery, and higher diagnosis accuracy). The work is motivated by three challenges: (1) selecting and fusing contextual data so models act responsibly and effectively; (2) placing models under uncertain resource consumption and performance fluctuations; and (3) coordinating components for reactive recovery and proactive maintenance at the system level.
The thesis develops four lifecycle-aware methods that address these challenges and yield measurable improvements: (1) a context-aware and ethics-guided model design that integrates environmental/contextual data under principled constraints (Non-Maleficence, Utilitarianism) to reduce SLA violations of critical services while maintaining overall efficiency; (2) an uncertainty-aware model placement method using fuzzy logic to learn and accommodate stochastic compute, storage, and inference performance fluctuations, improving placement robustness and resource utilization; (3) an LLM-based system-level recovery approach that performs real-time reasoning, root-cause diagnosis, and autonomous repair, achieving a 94.5% recovery success rate with an average correction time of 410.8 s across 200 systems in a NeRF-over-networking task; and (4) a proactive maintenance pipeline coupling a Large Forecasting Model (Timer-XL) for segment-level predictive monitoring with an LLM for root-cause reasoning, reaching 98.8% degradation-detection accuracy, 98.2% diagnostic accuracy, and 95.4% end-to-end maintenance success over 1200 cases.
Embedding AI across the lifecycle yields quantifiable gains, including higher recovery success, faster correction, and more accurate proactive prevention, thereby laying a practical foundation for context-aware, adaptive, and autonomous AI-native networking systems.
| Date of Award | 30 Sept 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Yulei Wu (Supervisor) |