Using Domain Knowledge to Aid Scientific Theory Revision

Abstract

Discovery systems must often face the task of revising an initially held theory in order to account for new information. To this end, the REVOLVER system was constructed, employing a set of heuristics to find models of objects (i.e., theories) consistent with initial beliefs (i.e., data) that contain them. When inconsistencies arise, the program performs a hill-climbing search for a new consistent solution. While the program was initially designed to use domain-independent heuristics to evaluate potential revisions and construct consistent theories, scientists often employ knowledge or assumptions of a specific domain in order to help constrain the revision process. This paper describes ways in which domain knowledge has been used to aid hill climbing in REVOLVER. First, new domain assumptions can help improve the search for theories. To illustrate this, I present an example from the domain of particle physics; in this domain and others (e.g., genetics), the addition of a new domain-specific heuristic to the system's evaluation function leads to convergence on a single set of models that replicates historical results. Second, since the program uses previous inference episodes when evaluating revisions, its current theory also influences search. To illustrate this concept, I present new experiments showing how the system's ability to predict new beliefs improves with increasing knowledge.