Extensible Multi-Entity Models: A Framework for Incremental Model Construction

Graphical models have become common for representing probabilistic models in statistics and artificial intelligence. A Bayesian network is a graphical model which encodes a probability model as a directed graph in which nodes correspond to random variables, together with a set of conditional distributions of nodes given their parents. In most current applications of Bayesian networks, a fixed network is specified to apply to all problem instances. Inference consists of conditioning on certain random variables, called evidence variables, and inferring the distribution of others, called target variables. In more complex problems arising in artificial intelligence, it is useful to use the belief network formalism to represent uncertain relationships among variables in the domain, but it is not possible to use a single, fixed belief network to encompass all problem instances. This is because the number of entities to be reasoned about and their relationships to each other varies from problem instance to problem instance. This paper describes a framework for representing probabilistic knowledge as fragments of belief networks and a method for constructing situation-specific belief networks for particular problem instances.