Modular specification of

This thesis concentrates on specifying dynamic probabilistic models and their application in the field of discrete time series analysis. At first, the basics of discrete probabilistic models and traditional clique tree propagation inference is summarized to name the concepts used further in this work. Then, the idea of using a stream of small Bayesian networks for dynamic time series modeling is presented. Previous work on dynamic bayesian networks, like model representation and the interface algorithm used for inference, are discussed. Their implications are subsequently taken into account in the appli-cation presented in this work. A way to specify DBNs in practice is introduced and algorithmic use of the models is studied. An EM algorithm is developed for estimating the quantitative parameters of a specified model when partial data is given. Properties of the learning algo-rithm are also studied. In the experimental part of this thesis, two different settings are demonstrated. First, an idealized and controlled laboratory setting is considered by generat-ing artificial data and using it for demonstrating the discussed methods. The another case is a brief study of modeling human DNA copy number amplifi-cations in a real world setting