Network slicing, a cornerstone technology for 5G and 6G networks, enables multiple end-to-end logically isolated network slices to operate on shared infrastructure, with each slice optimised to meet specific quality of service (QoS) requirements across distinct service types, from ultra-reliable low-latency communications to massive machine-type communications. while 5G’s transition to cloud-native architectures promises elastic scaling and operational flexibility, current resource management approaches fail to realise these benefits: they either over-provision resources, leading to excessive energy consumption and costs, or under-provision, causing service degradation. Industry standard autoscalers, designed primarily for web applications, cannot accommodate the unique characteristics of network functions where microsecond delays translate to dropped connections and traffic bursts necessitate instantaneous scaling decisions.

This research addresses these fundamental challenges through the development of intelligent, adaptive resource management systems specifically designed for cloud-native network functions. In contrast to generic autoscalers that employ simplistic threshold-based heuristics, the proposed artificial intelligence (AI)-based solutions handle the complex dynamics of telecommunications workloads and autonomously optimise the trade-off between QoS guarantees and resource efficiency. This work is essential for enabling operators to realise the benefits of cloud-native transformation while meeting stringent service level agreements (SLAs) across diverse use cases and achieving substantial reductions in both operational expenditure and carbon emissions.

This thesis presents three novel AI-powered resource management solutions that advance beyond existing approaches. First, NFScaler introduces a zero-shot simulation-to-reality (sim-to-real) transfer learning framework that eliminates the costly and risky training phase of deep reinforcement learning (DRL) in production networks. Through domain randomisation techniques, NFScaler learns robust auto-scaling policies in simulation that transfer directly to production networks, achieving 40% improvement in QoS performance compared to the industry standard autoscaler while reducing resource utilisation by 75% and power consumption by 45% relative to traditional dedicated server deployments, thereby demonstrating that intelligent auto-scaling is essential for cloud-native 5G networks.

Second, multi-agent recurrent soft actor-critic (MARSAC) framework addresses multi-slice resource management where services, including enhanced mobile broadband (eMBB), ultra-reliable and low-latency communications (URLLC), massive machine-type communications (mMTC), and vehicle-to-everything (V2X), with conflicting requirements compete for shared resources. MARSAC deploys decentralised autonomous agents that independently optimise their respective slices while observing system-wide constraints, preventing both local failures and global resource waste. This decentralised architecture achieves linear scalability as the number of network slices increases, outperforming centralised multi-agent reinforcement learning approaches that suffer from computational bottlenecks. Experimental results demonstrate that MARSAC delivers 25.9% superior resource efficiency and 20.3% improved energy efficiency compared to the industry standard autoscaler, while consistently satisfying the distinct QoS requirements of each slice type.

Third, proactive priority-based resource allocation and admission control (PPRAAC) algorithm designed to guarantee QoS for network slices in resource- and energy-constrained 6G non-terrestrial networks (NTNs). PPRAAC employs a hybrid machine learning framework combining MARSAC for resource auto-scaling, gated recurrent units (GRUs) for satellite power consumption forecasting, and priority-aware decision making for both resource allocation and admission control. Experimental results demonstrate that PPRAAC outperforms benchmark approaches in guaranteeing QoS for high-priority critical slices. Specifically, for the emergency services slice, PPRAAC achieves substantial improvements over benchmark methods, with reliability gains exceeding 61% and latency improvements above 14%. These improvements are critical for emergency services, where reliable communication and rapid response times directly impact operational effectiveness.

Experimental validation in a 5G testbed and emulated 6G NTN environment demonstrates that these solutions not only outperform current industry standards but also provide the adaptability and efficiency necessary for sustainable network operations. This work establishes a foundation for autonomous network management systems capable of supporting the diverse and demanding applications of future networks while minimising operational costs and environmental impact.