Stochastic methods for measurement-based network control

The main task of network administrators is to ensure that their network functions properly. Whether they manage a telecommunication or a road network, they generally base their decisions on the analysis of measurement data. Inspired by such network control applications, this dissertation investigates several stochastic modelling techniques for data analysis. The focus is on two areas within the field of stochastic processes: change point detection and queueing theory.
Part I deals with statistical methods for the automatic detection of change points, being changes in the probability distribution underlying a data sequence. This part starts with a review of existing change point detection methods for data sequences consisting of independent observations. The main contribution of this part is the generalisation of the classic cusum method to account for dependence within data sequences. We analyse the false alarm probability of the resulting methods using a large deviations approach. The part also discusses numerical tests of the new methods and a cyber attack detection application, in which we investigate how to detect dns tunnels.
The main contribution of Part II is the application of queueing models (probabilistic models for waiting lines) to situations in which the system to be controlled can only be observed partially. We consider two types of partial information. Firstly, we develop a procedure to get insight into the performance of queueing systems between consecutive system-state measurements and apply it in a numerical study, which was motivated by capacity management in cable access networks. Secondly, inspired by dynamic road control applications, we study routing policies in a queueing system for which just part of the jobs are observable and controllable.