Please note: This PhD defence will take place online.
Muhammad Sulaiman, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Raouf Boutaba
Unlike legacy mobile networks, which largely followed a one-size-fits-all architecture, 5G networks enable multiple logically isolated virtual networks to coexist over a shared physical infrastructure. The resulting virtual networks, or network slices, can be customized to support the heterogeneous requirements of different applications such as cloud gaming and remote surgery. Autonomous management and orchestration (MANO) of these network slices involves several interdependent problems that must be solved jointly.
This thesis focuses on three closely related MANO problems: (i) slice admission control (SAC), (ii) resource allocation (RA), and (iii) virtual network embedding (VNE). Under the Network Slice as a Service (NSaaS) model, service providers (SPs) submit slice requests (SRs) to an infrastructure provider (InP), which owns the 5G substrate network. Accepting an SR generates revenue but consumes scarce resources that could otherwise support future requests. SAC therefore determines whether an SR should be admitted by balancing its offered revenue against its resource commitments and their implications for future admission opportunities. RA determines the amount of resources the slice requires to meet its service requirements without unnecessary provisioning. VNE determines where these resources should be allocated in the substrate network.
Although artificial intelligence (AI)-based approaches have enabled the automation of these decisions, a number of technical challenges and operational constraints remain overlooked. This thesis identifies these gaps and addresses them by developing a set of AI- and optimization-driven methods across three research directions: network slice modeling, resource allocation, and joint SAC and VNE. For network slice modeling, we show how probabilistic, composable slice models can effectively balance predictive accuracy, training-data requirements, and inference efficiency while supporting modularity and seamless integration with MANO algorithms. For RA, we develop complementary deep reinforcement learning (DRL)- and optimization-based methods that enable resource efficiency under explicit quality-of-service (QoS) degradation constraints while accounting for uncertainty and different runtime computational budgets. For joint SAC and VNE, we develop multi-agent DRL (MADRL)- and pricing-based methods. Our MADRL-based approach utilizes separate but coordinated policies to enable faster learning and improve long-term InP revenue while producing fast and feasible online decisions. The pricing-based approach instead leverages utilization-dependent resource prices to capture the opportunity cost of committing resources to each request, yielding explainable admission and embedding decisions that scale to metropolitan networks without requiring expensive training. The proposed approaches are evaluated through testbed- and simulation-based experiments, demonstrating their effectiveness in addressing the MANO problems of SAC, RA, and VNE in next-generation mobile networks.