Machine learning for network slice resource management

Network softwarization through Network Function Virtualization (NFV) and Software-Defined Networking (SDN) enables flexible service deployment and cost-efficient operation in modern communication systems. Network slicing, a key feature of 5G and beyond networks, allows a shared physical infrastructure to be partitioned into multiple isolated end-to-end (E2E) virtual networks tailored to diverse application requirements. However, efficient slice management remains challenging due to multi-domain orchestration, dynamic traffic conditions, limited infrastructure visibility, and strict Service Level Agreement (SLA) guarantees.
This thesis investigates scalable E2E network slice management in multi-domain environments using Artificial Intelligence and Machine Learning (AI/ML). First, an adaptive SLA decomposition framework is proposed to translate E2E requirements into domain-specific SLAs under uncertainty. The framework employs machine learning-based risk models to improve service request acceptance while handling incomplete state information. Second, the thesis introduces sequence-aware orchestration for cross-domain Service Function Chain (SFC) deployment. Transformer-inspired models capture dependencies among Virtual Network Functions (VNFs), enabling efficient partitioning and placement decisions that enhance long-term service quality and resource utilization. Third, an integrated learning-driven edge resource management mechanism jointly optimizes task placement and resource scaling to support latency-sensitive services under dynamic workloads.
Overall, the proposed AI-driven framework improves automation, scalability, and reliability of E2E network slice lifecycle management, advancing efficient operation of future 5G and beyond communication infrastructures.