Bayesian Network Approach for Modelling and Inference of Communication Networks
Examensarbete för masterexamen
Engineering mathematics and computational science (MPENM), MSc
Mouliakos, Themistoklis Christos
A collection of radio base stations defines the radio access network which is responsible for connecting the user equipment with the core network which in turn connects to the internet through a gateway. The radio base station provides the uplink and downlink communication of the user equipment and its cover region is divided in cells. There can be thousands of parameters affecting the functionality of a radio base station and a primary concern of this thesis is to estimate probabilistic relations between those parameters and the key performance indicators. Bayesian Networks is a powerful mathematical tool which can model complex systems and present possible co-influences between variables. In the last decades there was a great advancement on creating Bayesian Networks from data, mainly because of new algorithms and the increase of computational power. The power of the Bayesian Networks to represent in a compact and visually easy way the joint probability of a set of variables, make it ideal candidate for the complex data of a communication network. In this thesis work, most of the focus will be put on learning an acceptable structure by a combination of expert’s knowledge and appropriate learning methods. Two main approaches were investigated by using a Hybrid Network(mix of continuous and discrete data) and a full discretization of the variables. Although both networks managed to capture expected associations between the variables, the hybrid network poses serious restrictions that are not well supported by the data. In order to investigate more approaches to structure learning, the Hill Climbing algorithm was enhanced with random restarts. As a third option the bootstrap method was tested also as a different way to construct more robust models and estimate the strength confidence of both the arcs and their direction. Finally a comparison of the predictive power for the different learning techniques was evaluated through cross validation.