Maintenance in Probabilistic Knowledge-Based Systems

Recent developments using directed acyclical graphs (i.e., influence diagrams and Bayesian networks) for knowledge representation have lessened the problems of using probability in knowledge-based systems (KBS). Most current research involves the efficient propagation of new evidence, but little has been done concerning the maintenance of domain-specific knowledge, which includes the probabilistic information about the problem domain. By making use of conditional independencies represented in she graphs, however, probability assessments are required only for certain variables when the knowledge base is updated. The purpose of this study was to investigate, for those variables which require probability assessments, ways to reduce the amount of new knowledge required from the expert when updating probabilistic information in a probabilistic knowledge-based system. Three special cases (ignored outcome, split outcome, and assumed constraint outcome) were identified under which many of the original probabilities (those already in the knowledge-base) do not need to be reassessed when maintenance is required.