Syntax-based Default Reasoning as Probabilistic Model-based Diagnosis
This work addresses the challenge of handling uncertain and conflicting information in AI reasoning systems, offering a probabilistic framework that is incremental, building on existing ATMS and model-based diagnosis methods.
The paper tackles the problem of default reasoning by framing syntax-based approaches as a model-based diagnosis problem, where each information source is treated as an independent component with a small failure probability, resulting in a Dempster-Shafer belief function that defines non-monotonic consequence relations, which are then studied and compared.
We view the syntax-based approaches to default reasoning as a model-based diagnosis problem, where each source giving a piece of information is considered as a component. It is formalized in the ATMS framework (each source corresponds to an assumption). We assume then that all sources are independent and "fail" with a very small probability. This leads to a probability assignment on the set of candidates, or equivalently on the set of consistent environments. This probability assignment induces a Dempster-Shafer belief function which measures the probability that a proposition can be deduced from the evidence. This belief function can be used in several different ways to define a non-monotonic consequence relation. We study and compare these consequence relations. The -case of prioritized knowledge bases is briefly considered.