Recent advancements in Software-Defined Networking (SDN), Network Function Virtualization (NFV), Open Radio Access Network (O-RAN), and 6G technologies have transformed communication networks into programmable, virtualized, and service-aware infrastructures. However, this transformation has also made end-to-end (E2E) network and service management significantly more complex. In modern 5G/6G systems, operators must instantiate, monitor, assure, adapt, and optimize network slices and services across the Radio Access Network (RAN), transport, core, and edge while satisfying heterogeneous Key Performance Indicators (KPIs) and Service Level Agreements (SLAs). Traditional human-in-the-loop management and orchestration approaches cannot react at the speed and scale required by these environments, motivating the vision of Zero-Touch Network and Service Management, in which E2E service lifecycle management is achieved through autonomous, closed-loop, policy-driven, and cross-domain control with minimal human intervention.
Network slicing is one of the central mechanisms through which this vision can be realized. By composing Virtual Network Functions (VNFs) into isolated end-to-end slices, operators can tailor these slices to the needs of diverse services. Yet achieving zero-touch slicing remains difficult because slice creation, assurance, scaling, and adaptation must be coordinated across multiple domains and adapting to the dynamics of the service. Some of the main open challenges are: 1) Deciding how many resources to allocate to each service with their hetrogeneous KPIs while avoiding over- and under-provisioning, 2) adapting slices to extreme and unpredictable conditions such as node failures and traffic surges, and 3) learning to provision resources efficiently for emerging workloads with highly dynamic behavior, such as AI training and immersive Metaverse services.
In this thesis, we study intelligent network slicing and resource provisioning as core enablers of zero-touch, E2E, and cross-domain network and service management. This thesis can be summarised into two segments as follows:
The first segment establishes a foundation for zero-touch provisioning by addressing the fundamental problem of creating and optimizing cost-effective network slices through proper VNF placement and scaling under load uncertainty and multiple KPI constraints. In this section, we introduce the Error-Aware, Cost-Effective, and Proactive Network Slicing Framework (ECP), which facilitates proactive and predictive network slicing in 5G and 6G environments. This framework leverages AI-based forecasting, along with Deep Reinforcement Learning (DRL), to create cost-effective and accurate network slices ahead of time. ECP effectively addresses practical challenges such as load prediction errors and diverse KPI requirements, thereby advancing E2E slice commissioning and adaptive lifecycle management. Our results demonstrate significant cost reductions and improvements in resource usage, validating the framework against various baselines and state-of-the-art solutions.
The second segment of the thesis builds upon the intelligent resource provisioning foundation established in the first part, extending the network slicing and resource allocation problem to two highly challenging and advanced network environments as follows:
The first problem is tailoring network slicing for AI training, where we explore the slicing of resources to support the training of different AI models (often offloaded from computationally limited devices) with varied goals and requirements. This involves the novel joint allocation and tuning of not only computing and network resources but also the hyperparameters of the AI models themselves. This represents a key advancement toward AI-native 6G, enabling the network to automate and support the training of AI models beyond general services.
The second problem extends intelligent network slicing into complex, multi-domain environments where slice allocation alone is insufficient, creating the need for joint resource allocation and cooperation. We use the Metaverse as a representative use case for this setting, given its highly dynamic and immersive nature, where ultra-low latency, high-fidelity rendering, and real-time synchronization of Digital Twins (DTs) are essential for user immersion. In this context, we investigate and build two complementary immersion-aware cooperative frameworks across multiple Metaverse Service Providers (MSPs). First, we develop CIVIC, which transforms traditional network slicing into a cooperative orchestration system by introducing a General Credit Pool (GCP) that enables dynamic, incentive-compatible resource sharing through a standardized cloud credit; leveraging DRL, CIVIC learns optimal policies for joint resource provisioning and cooperative credit donation to in-demand MSPs. Second, we introduce ZTCI (Zero-Touch Cooperative Immersion), a plug-and-deploy DRL-based allocation-and-cooperation framework that enables autonomous, scalable cooperation among distributed Metaverse slicing systems at deployment and runtime. Together, these solutions enhance Quality of Experience (QoE) and user immersion and advance the vision of zero-touch, cross-domain, and E2E network and service management, where AI enables scalable, fair, and resilient resource orchestration in complex, distributed 6G ecosystems.
| Date of Award | 2026 |
|---|
| Original language | American English |
|---|
| Awarding Institution | - HBKU College of Science and Engineering
|
|---|