Least exception logic
Stephen D. Post, Andrew P. Sage · 1989
Default reasoning is a central component of research in artificial intelligence as well as of our everyday reasoning. This dissertation puts forth a new model for making assumptions and judgements toward the solution of real-world problems such as natural language understanding, diagnosis, and robotics. The method, least exception logic, is based on both logic and measures. It decomposes logical resolution into unification and solution, and performs the solution as an integer linear program, where the constraints are ground clauses drawn from the first-order rules, and the objective function minimizes the weighted sum of exceptions to the default rules subject to simultaneous solution of the logical constraints. The method is nonmonotonic, provides truth maintenance, orders the multiple extensions, and is relatively practical. The dissertation first establishes the requirements for default reasoning. It then defines principles of least exception logic, shows how the model satisfies the requirements, and compares the model to prominent paradigms, such as nonmonotonic logics, circumscription, truth maintenance systems, Dempster Shafer theory, Bayesian networks, and neural networks. The dissertation then describes a fully developed shell that embodies the model, and recounts experiments conducted with the shell, including a fairly extensive exercise in story understanding.