Online learning for self-optimization in heterogeneous networks

Resumo

Resumen de la tesis: The ever-increasing growth of the mobile data demands combined with the new application-specific network requirements has triggered the development of the fifth-generation (5G) of the mobile network technology. The deployment of low power base stations (small cells) is one of the key improvements in network architecture aimed at enabling 5G to meet its requirements. The network architecture resulting from the combination of legacy macro base stations and small cells is referred to as Heterogeneous Network (HetNet). This architecture presents many advantages such as frequency reuse, load balancing, and lower latency. However, HetNets also presents two main technical problems: inter-cell interference associated with the massive frequency reuse performed by the small cells, and the higher power consumption introduced by an increasing number of base stations. These problems have been addressed by means of interference coordination (IC) and energy saving (ES) mechanisms. Although the configuration of these two mechanisms has been addressed separately so far, we show in this thesis that they are highly coupled. Moreover, the configuration of IC and ES is commonly addressed using network models, which presents several limitations. In this thesis, we consider the self-optimization functionality within the Self-Organizing Networks (SON) paradigm, which is intended to address these problems by allowing the network to autonomously configure its parameters while it is operating. To implement the self-optimization functionality, we propose the use of online learning algorithms, which learn efficient network configurations from experience without explicitly knowing the accurate mathematical model of the network beforehand. The first part of this thesis addresses the configuration of the IC mechanism in HetNets. We propose several online learning model-free solutions based on different techniques such as Response Surface Method (RSM) and Multi-Armed Bandit (MAB) algorithms. We also consider stochastic constraints in the learning process. In the second part, we address the joint problem of IC and ES in HetNets proposing several solutions based on Dynamic Programming, Contextual Multi-Armed Bandit algorithms, and Machine Learning tools such as Neural Networks and Gaussian processes.