András Horváth
The material presented in the dissertation is about approximating
non-Markovian behavior by Markovian models.
Introduction to various aspects of the field is provided in Part I
consisting of three chapters. Chapter 1 provides the background for
discrete-time Phase-type (DPH) distributions. A short introduction to
Markovian modeling for traffic engineering is given in Chapter
2. Characteristics of random quantities and random processes of
telecommunication networks with related statistical tests are introduced in
Chapter 3.
Part II investigates DPH distributions. It is shown that the use of DPH
distributions, which received much less attention than the use of
continuous-time Phase-type (CPH) distributions, can be beneficial in
numerous modeling situations. We show that the minimal coefficient of
variation of the acyclic DPH (ADPH) class is lower than that of the CPH
class. Also the structure of the ADPH distribution exhibiting minimal
coefficient of variation is provided. We present a fitting procedure as
well and make use of it to perform and analyze several fitting experiments.
Chapters of Part III present algorithms that result in Markovian models capturing phenomena whose existence in telecommunication networks was reported recently and, initially, it was agreed that Markovian models are not appropriate to model them. In Chapter 7 we concentrate on random durations with heavy tail and give a Phase-type fitting procedure that handles separately the fitting of the body and the tail of the distribution. In the last two chapters of the dissertation, a step further is taken by fitting point processes instead of distributions. In Chapter 8 a heuristic fitting method is presented that, by treating separately the short- and long-range behavior, captures important features of real traffic sources. In the subsequent chapter of the dissertation, we introduce a set of Markovian arrival processes with special structure that exhibit multifractal behavior (as demonstrated by multifractal analysis) and also a fitting procedure to capture the multifractal properties of data sets.