A system for reasoning by minimizing exception

The author describes least exception logic (LEL), a new paradigm for default reasoning synthesized from the first-order predicate calculus and integer linear programming. Whereas logical inference couples unification with Boolean constraint satisfaction and draws inferences in a piecewise process, LEL decouples unification from solution and draws inferences by global optimization. This approach breaks down the distinction between logic-based and measure-based methods for default reasoning and advances each. Logic-based methods are advanced through LEL's ability to choose the best combination of default rules to apply when various feasible combinations are in contention. Measure-based methods are advanced through LEL's ability to focus on the best overall solution rather than becoming muddled by propagation of deteriorating measures of certainty for individual propositions. LEL has been implemented and used on several moderate-sized problems in natural language understanding.<>