Network Slicing: Joint Optimization of Admission Control, Resource Attribution and Traffic Routing

• Main Author: Gil, Santiago Fidalgo Riveiro
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10362/173164
• topic: Network Slicing; DRL; Admission Control; Traffic Routing; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

The emergence of 5G networks and networks beyond 5G has brought about the need to dynamically and flexibly share physical infrastructure among multiple virtual networks supporting various services. This requirement, coupled with the development of technolo- gies such as Software-Defined Networks (SDNs) and Network Function Virtualization (NFV), paved the way for the realization of the concept of network slicing, which involves segmenting the physical network into multiple virtual networks. Effective mechanisms for slice admission and traffic routing within the physical network are essential to op- timize resource utilization, meet quality of service (QoS) requirements, and ensure the efficient operation of various slices. This thesis presents a study of the application of Deep Reinforcement Learning (DRL) algorithms, specifically the Deep Q-Network (DQN) and Dueling Deep Q-Network (Dueling DQN) algorithms, to address the joint problem of optimizing slice admission and traffic routing in the network. Our research encompasses the development and evaluation of DQN-based solutions, focusing primarily on three developed algorithms, with a fourth serving as a control algorithm. The first is a DQN algorithm with a single neural network for joint action (routing and admission), the second is a DQN algorithm with two neural networks, considering routing and admission actions separately, and the third is the Dueling DQN algorithm with joint action. Finally, we have a random choice algorithm primarily to assess the efficiency of DRL in these scenarios. To provide context for the results, a comparative analysis is conducted between the proposed methods and a solution that considers only the admission problem with simple routing. Our results reveal that there is a benefit in considering routing alongside slice ad- mission, with DRL agents that consider joint action yielding better results than those considering two independent actions. Among the algorithms used, the Dueling DQN, with its separation of value and advantage networks, exhibited superior performance compared to the simple DQN approach.