Probabilistic model-based diagnostics

This thesis presents the concept of general argumentation systems, a framework for representing uncertain knowledge using information algebras and information systems as well as probability algebras. Argumentation systems are a generalization of assumption-based systems and propositional argumentation systems and can deal with very general formalisms. We show also that an argumentation system itself is a special case of an information system. We focus on argumentation systems as a tool for doing model-based diagnostics of complex systems built of components. Given a system and observations, this knowledge is modeled in the framework of an argumentation system. If the observations are not as predicted from the specification of the system, we have a diagnostic problem. Then, using concepts which are built on top of the argumentation system (such as generalized hints and allocations of support), conflicts and diagnoses, arguments, etc. for the system can be computed. Diagnoses of a system which does not work as it is supposed to, can then be used to define repair or replacement strategies for the components of the system. The set of diagnoses is often too big to be computed explicitly, but we define a logical framework for the description of sets of diagnoses. A main ingredient of an argumentation system is probabilistic knowledge of some parts of the available information, essentially knowledge about the modes of the components. This probabilistic knowledge allows to define allocations of probability and allocations of belief on top of the argumentation systems. Using these allocations, the symbolical results are weighed and discrimination between them is possible. Further, we present algorithms for efficient computation of the probability of logical representations of sets of arguments without explicitly computing the sets. The local computation framework of Shenoy & Shafer can be applied to argumentation systems. We show also how additional information can be added to an argumentation system, i.e. how argumentation systems are combined. Combination of argumentation systems can also be replaced by combination of the generalized hints, the allocations of arguments or belief defined on top of them. Further, we address the problem of how additional information about the system can be obtained, especially where further measurements should take place in order to obtain a maximal expected gain of information in the argumentation system; this is useful for a sequential process of discrimination of diagnoses in the system. Finally, we present the language ABEL, an implementation of a special type of argumentation system. Several examples are presented in order to show the strength of argumentation systems.