Mohammad Asif Habibi

• The next-generation radio access network (NG-RAN) architecture has been standardized by the Third Generation Partnership Project (3GPP) as the radio access network (RAN) architecture for fifth-generation (5G) mobile networks. It consists of a set of next-generation NodeBs (gNBs). Each gNB is composed of a centralized unit (CU), at least one distributed unit (DU), and at least one radio unit (RU). These components of a gNB can be deployed as virtual network functions (VNFs) and/or physical network functions (PNFs). In this thesis, we consider the CU and the DU as VNFs and the RU as a PNF. The CU, DU, and RU can be mapped onto the underlying aggregation data center, edge data center, and cellular network site in the NG-RAN architecture, respectively. The Open Radio Access Network (O-RAN) Alliance has redefined the NG-RAN architecture by establishing open and standard-compliant interfaces. The architecture defined by the O-RAN Alliance, known as the O-RAN architecture, comprises several components that interoperate to create a flexible, cloud-based RAN for 5G and beyond. In addition to the components specified by 3GPP for NG-RAN, the O-RAN Alliance introduces a new component for the management and orchestration of O-RAN elements. This component is referred to as the service management and orchestration (SMO) framework. The SMO framework is responsible for deploying and operating RAN services and coordinating the various O-RAN components. It includes a non-real-time RAN intelligent controller (Non-RT RIC) and can incorporate components from other standards-developing organizations (SDOs). Moreover, the O-RAN architecture features the near-real-time RAN intelligent controller (Near-RT RIC), which performs tasks within a near-real-time time frame. The Near-RT RIC, along with other gNB functions, can be mapped onto the underlying infrastructure. This infrastructure includes open-cloud (O-Cloud) sites, which provide the cloud environment for hosting these functions. The O-Cloud also features a notification interface for receiving relevant events. Overall, the O-RAN architecture aims to enhance flexibility, interoperability, and cloud-native capabilities in the deployment and operation of RAN systems, thereby supporting the evolution of 5G and beyond. The mapping of VNFs onto the underlying physical network infrastructure at the edge of a cellular network is a challenging task due to the joint allocation of virtual compute, storage, and networking resources across nodes and links, the diverse technical requirements of end users, and the need for coordination across multiple host domains. This issue is further complicated by the provisioning of RAN slicing, given the varying characteristics of wireless communication channels. To this end, this thesis addresses the mapping and virtual resource allocation problems of the VNFs of RAN slice subnets onto the underlying intelligent network infrastructure in NG-RAN. In this context, unlike most prior proposals that often fail to meet performance objectives and overlook resource allocation constraints, this thesis introduces and employs automation and intelligent techniques to map VNFs onto their corresponding physical nodes, with the aim of achieving improved efficiency in virtual resource utilization while ensuring the performance of RAN slice subnets in NG-RAN. Adopting a top-down approach, the key contributions of this thesis are as follows: • extend the framework of network slicing, as defined by the Next Generation Mobile Networks (NGMN) Alliance, to the NG-RAN architecture, and provide a critical analysis and overview of the components and functionalities of different types of RAN slices; • integrate the Experiential Network Intelligence (ENI) framework, as proposed by the European Telecommunications Standards Institute (ETSI), into a joint architecture of network functions virtualization--management and orchestration (NFV--MANO), also proposed by ETSI, and the Third Generation Partnership Project network slicing management system (3GPP-NSMS), in order to introduce automation and intelligence into the management and orchestration of different types of RAN slices in NG-RAN; • propose a learning-assisted solution (consisting of three phases: virtual resource automation, virtual resource management, and virtual resource allocation) for mapping the VNFs of a RAN slice subnet onto the underlying intelligent data centers in NG-RAN; • unify the management and orchestration components of 3GPP, ETSI, and the O-RAN Alliance within the SMO framework to enhance the unification and interoperability of various standard-compliant interfaces within the O-RAN architecture; • integrate management data analytics (MDA) into the management systems of 3GPP, ETSI, and the O-RAN Alliance within the SMO framework, with the goal of introducing intelligence and automation into the functionalities of the SMO framework for managing and orchestrating O-RAN components; • explore several deployment scenarios for integrating MDA and automation into the SMO framework, thereby providing network operators with multiple deployment options for enabling intelligence within the SMO framework; • present a comprehensive system model based on which the mapping problem of the CU and DU, internal and external VLs, and the VNFs of a RAN slice subnet onto the underlying infrastructure is mathematically formulated; • discuss various types of machine learning (ML)-assisted algorithms, such as supervised, unsupervised, and reinforcement learning, with a particular focus on the mathematical background and application of reinforcement learning, and select the Q-learning algorithm to solve the VNF mapping problem; • describe the simulation environment (including the simulation setup, network topology, and simulation parameters) considered for simulating the virtual components involved in mapping different types of RAN slices in NG-RAN; • obtain simulation results using the Q-learning algorithm for the mapping of RAN slice subnets, demonstrating significant performance improvements in mapping various types of RAN slice subnets onto the underlying infrastructure under different conditions; • evaluate the performance objectives achieved using the Q-learning algorithm against the service level agreement (SLA); and • provide findings on the global resource allocation required to host a large number of RAN slice subnets in the underlying infrastructure, as well as highlight the advantages of employing Q-learning for VNF mapping compared to other state-of-the-art algorithms.