Abstract

Although a Bayesian network is widely accepted as a sound and intuitive for- malism for reasoning under uncertainty in arti .cial intelligence,their use in diagnostic expert systems has been limited.The primary goal within these di- agnostic systems is to determine the most probable cause given a set of evidence and to suggest what additional information is best to collect.The framework of a Bayesian network supports this goal by providing various reasoning algo- rithms for the calculations of the e .ect of new information.However,for the support of practical models the networks are often accompanied by restrictions. One such restriction is that only one cause can be present since the support for multiple causes becomes computationally challenging.Another restriction is the limited support for user interaction.In most systems the user has nothing to say about which causes are investigated,instead the system always investigates all the causes.
In this thesis I aim to improve the functionality of Bayesian networks by pro- viding approximation approaches that support the diagnosis of multiple causes. At the same time I try to improve the interactivity with the user by supporting the ability to pursue and di .erentiate between any possible set of causes.The foundation of the approximation approaches is the relation between the proba- bility of causes separately and the probability of a combination of those.The ability to pursue and di .erentiate between any possible set of causes is a gener- alization of current possibilities to perform diagnosis,e.g.,the pursuit of one or all possible causes.I believe that these improvements will have a positive e .ect on the user acceptance of Bayesian networks in modelling complex diagnostic systems.