Bayesian networks have become a popular probabilistic model for coping with uncertainty. When learning their structure from given data the constraint-based approach has proven to be very efficient in many experiments. However, this approach is only well-understood under certain conditions, typically only when infinite data sets are provided. This thesis focuses on the constraint-based approach for those cases where only a finite amount of data is available, as typical in practical applications. Among various extensions to this approach the so-called necessary path condition is presented. While its use does not notably increase computation time, it entails considerable improvements in the quality of the induced network structures. In particular, uncertainty regarding the presence of edges can be discovered by this extension: instead of a single network structure, as typical for state-of-the-art constraint-based approaches, it can find several graphs. As the induced graphs usually have many common edges and differ only in the presence of a few edges, they can be visualized by means of a single graph, thus easing the interpretation.     «

Bayesian networks have become a popular probabilistic model for coping with uncertainty. When learning their structure from given data the constraint-based approach has proven to be very efficient in many experiments. However, this approach is only well-understood under certain conditions, typically only when infinite data sets are provided. This thesis focuses on the constraint-based approach for those cases where only a finite amount of data is available, as typical in practical applications....     »

Während sich der testbasierte Ansatz zum Strukturlernen in Bayes' schen Netzen in der Praxis als sehr effizient erwiesen hat, kann sein Verhalten bisher nur im asymptotischen Grenzfall verstanden werden. Diese Arbeit charakterisiert die Eigenschaften dieser Lernmethode bei endlichen Datenmengen. Darauf aufbauend werden diverse Verbesserungen dieses Ansatzes entwickelt, insbesondere wie die in endlichen Daten vorhandene Unsicherheit bzgl. der gelernten Strukturen ohne merklich längere Rechenzeit entdeckt und mittels eines einzigen Graphen übersichtlich dargestellt werden kann.     «

Während sich der testbasierte Ansatz zum Strukturlernen in Bayes' schen Netzen in der Praxis als sehr effizient erwiesen hat, kann sein Verhalten bisher nur im asymptotischen Grenzfall verstanden werden. Diese Arbeit charakterisiert die Eigenschaften dieser Lernmethode bei endlichen Datenmengen. Darauf aufbauend werden diverse Verbesserungen dieses Ansatzes entwickelt, insbesondere wie die in endlichen Daten vorhandene Unsicherheit bzgl. der gelernten Strukturen ohne merklich längere Rechenzeit...     »