Kathryn Blackmond Laskey
George Mason University

Bayesian networks have become a common structure for representing probabilistic models in statistics and artificial intelligence. A Bayesian network uses a directed graph to represent the dependence structure
among a set of uncertain hypotheses. Nodes in the graph represent hypotheses and arcs represent conditional dependencies among the hypotheses. Expert systems based on Bayesian networks are developed either by eliciting structure and parameters from domain experts or by estimating a Bayesian network from a database of cases. In more complex problems arising in artificial intelligence, the number of objects to be reasoned about and their relationships to each other varies from problem instance to problem instance. For this reason, it is not possible to use a single, fixed Bayesian network to encompass all problem instances. Partially specified Bayesian networks, or network fragments, can be used to represent uncertain relationships among related hypotheses encompassing a limited subset of the domain. A multi-entity Bayesian network is a collection of network fragments that together represents a coherent probabilistic domain model. At run time, a probabilistic query triggers construction of a situation-specific Bayesian network can be constructed from a multi-entity Bayesian network knowledge base. This talk describes the multi-entity Bayesian network knowledge representation, an algorithm for constructing a situation-specific Bayesian network that yields provably correct query responses relative to the knowledge base, and methods for constructing approximately correct situation-specific networks.