Toward Robust Real-World Inference: A New Perspective on Explanation-Based Learning

Abstract

Over the last twenty years AI has undergone a sea change. The once-dominant paradigm of logical inference over symbolic knowledge representations has largely been supplanted by statistical methods. The statistical paradigm affords a robustness in the real-world that has eluded symbolic logic. But statistics sacrifices much in expressiveness and inferential richness, which is achieved by first-order logic through the nonlinear interaction and combinatorial interplay among quantified component sentences. We present a new form of Explanation Based Learning in which inference results from two forms of evidence: analytic (support via sound derivation from first-order representations of an expert’s conceptualization of a domain) and empirical (corroboration derived from consistency with real-world observations). A simple algorithm provides a first illustration of the approach. Some important properties are proven including tractability and robustness with respect to the real world.