A probabilistic commonsense reasoner

Abstract

Most commonsense reasoning formalisms are based on first order logic: default logic, predicate circumscription, autoepistemic logic, theory formation, variable conditional logic, and nonmonotonic logic. Probability has long been thought to be “epistomologically inadequate” for representing and reasoning in the commonsense domains partly because probability distributions are not ordinarily available. We claim that probability is indeed epistomologically adequate and describe a reasoning formalism based on the probability calculus and conditional independence that requires only a knowledge base of probabilistic inequalities. Numerical distributions are not required, but we depart in a significant way from the “logicist” school of AI: rather than “accept” defeasible conclusions and reject them on receipt of new evidence, we reason about how observations cause beliefs to shift. When reasoning forwards, the system produces answers expected for many of the problems we have encountered in the nonmonotonic reasoning literature. When reasoning backwards, the system produces the answers expected for many diagnostic problems. Unexpected answers can be explained in a statistical sense. This is important for it provides a testable account of commonsense reasoning.