Measuring Inconsistency in Probabilistic Knowledge Bases

摘要

In terms of standard probabilistic reasoning, in order to perform inference from a knowledge base, it is normally necessary to guarantee the consistency of such base. When we come across an inconsistent set of probabilistic assessments, it interests us to know where the inconsistency is, how severe it is, and how to correct it. Inconsistency measures have recently been put forward as a tool to address these issues in the Artificial Intelligence community. This work investigates the problem of measuring inconsistency in probabilistic knowledge bases. Basic rationality postulates have driven the formulation of inconsistency measures within classical propositional logic. In the probabilistic case, the quantitative character of probabilities yielded an extra desirable property: that inconsistency measures should be continuous. To attend this requirement, inconsistency in probabilistic knowledge bases have been measured via distance minimisation. In this thesis, we prove that the continuity postulate is incompatible with basic desirable properties inherited from classical logic. Since minimal inconsistent sets are the basis for some desiderata, we look for more suitable ways of localising the inconsistency in probabilistic logic, while we analyse the underlying consolidation processes. The AGM theory of belief revision is extended to encompass consolidation via probabilities adjustment. The new forms of characterising the inconsistency we propose are employed to weaken some postulates, restoring the compatibility of the whole set of desirable properties. Investigations in Bayesian statistics and formal epistemology have been interested in measuring an agent's degree of incoherence. In these fields, probabilities are usually construed as an agent's degrees of belief, determining her gambling behaviour. Incoherent agents hold inconsistent degrees of beliefs, which expose them to disadvantageous bet transactions - also known as Dutch books. Statisticians and philosophers suggest measuring an agent's incoherence through the guaranteed loss she is vulnerable to. We prove that these incoherence measures via Dutch book are equivalent to inconsistency measures via distance minimisation from the AI community.