Game Theoretic and Machine Learning Techniques for Efficient Resource Allocation in Next Generation Wireless Networks

Abstract

The rationale behind the next generation wireless networks is the handling of the recent massive surge in wireless traffic, especially due to the advent of the Internet of Things (IoT) ecosystem. Tremendously high data rates, extremely low latency, and significantly high Quality of Service (QoS) are among the key objectives of the forthcoming fifth generation (5G) standard. Some of the concepts which act as the driving forces behind realizing these goals are network virtualization, fog computing, heterogeneous networks, and spectrum sharing. Taking these into account, a few efficient resource allocation frameworks for these techniques are proposed in this dissertation. Considering the distributed behaviors of the different sets of entities involved and their interrelationships, we incorporate the potentials of game theory and Machine Learning (ML) as powerful mathematical tools for strategic decision making. Firstly, two resource allocation frameworks for network virtualization based on matching theory are proposed: a three-sided matching based model involving radio resources, physical infrastructure, and mobile users for wireless network virtualization, and a similar model involving Tracking Areas (TAs), Virtual Network Function (VNF) instances, and Cloud Networks (CNs) for Network Function Virtualization (NFV). Secondly, an Equilibrium Problem with Equilibrium Constraints (EPEC) and a many-to-many matching based framework is proposed for NFV integrated IoT fog computing: a large-scale model for the optimization of resource pricing for the Data Service Operators (DSOs), as well as for the optimization of resource allocation from the Fog Nodes (FNs) as per the requirements of the Authorized Data Service Subscribers (ADSSs). Thirdly, a resource allocation framework for heterogeneous networks based on Reinforcement Learning (RL) and EPEC is proposed: a multi-hop data transmission route determination model for an indoor Visible Light Communication (VLC) and Device-to-Device (D2D) heterogeneous network. Finally, a framework to enhance the spectrum utilization of a Cognitive Radio Network (CRN) is proposed: a classification approach to detect Primary User Emulation (PUE) attacks using Generative Adversarial Networks (GANs), which are effective ML models to train classifiers in a semi-supervised manner. In this dissertation, a comprehensive discussion of these frameworks is performed, followed by the validation of their effectiveness through extensive simulations.