Network slice admission using reinforcement learning and information-centric networking for mobile networks

Abstract

The evolution of the current most popular mobile network (4G), the so-called 5G, is targeting an increased traffic load at a lower cost. Thus, optimization of the delivery network plays an essential role at 5G; another aspect of the evolution is that 5G has the ambition to be highly customized, e.g., reliable enough to be used in industrial automation and cheap enough to be used for mobile broadband services. In this context, this thesis assesses two aspects of 5G: the first is to use information-centric networking (ICN) to improve the efficiency of multimedia delivery in mobile broadband services; and the second is the application of a reinforcement learning strategy as an enabler for the highly configurable network, which could pose a challenge to be understood and configured manually. ICN aims at circumventing several issues of current internet protocol, among them, achieving a more efficient multimedia distribution. Given the significant growth rate of video transmission over mobile networks, it is sensible to consider how mobile networks can leverage ICN. There is a substantial body of work considering ICN for fixed networks and also for the core of mobile networks. Less attention has been dedicated to ICN on the radio access network (RAN) or ICN-RAN, which has currently a user plane based on many connection-oriented protocols. To fully benefit from ICN, mobile networks must enable it on the RAN, not only on the core. This work details an ICN deployment on the RAN of the fourth and fifth generation of mobile networks and also presents a testbed that enables proofs of concept of this ICN-RAN using 4G. The results indicate, for example, that evolving ICN features can be tested with currently available tools, but the lack of hardware accelerators and optimized code limit the bit rate that can be achieved in real-time processing. In the context of network customization, the most prominent enablers are the so-called network slices. Slices can be understood as a part of the network that is customized to deliver certain services. The service requirements are imposed by the tenant, which acquire slices from an infrastructure provider. The 5G infrastructure provider must optimize the infrastructure resource utilization, usually admitting as many slices as possible. However, infrastructure resources are finite and admitting all the slices could increase the probability of service level agreement violation. This thesis investigates the application of reinforcement learning agents that learn how to increase the infrastructure provider revenue by intelligently admitting network slices that bring the most revenue to the system. We present a neural networks-driven agent for network slice admission that learns the characteristics of the slices deployed by the tenants from their resource requirements profile and balances the benefits of slice admission against orchestration and resource management costs.