Scaling Challenges in Explanatory Reasoning

Despite well-known limitations, human cognition exhibits remarkable abilities for scaling to factors like task complexity and knowledge base size. In this paper, we revisit a recently proposed theory of explanatory inference and its implementation in the PENUMBRA system, which we hypothesize will support similar properties. We examine – analytically and empirically – the computational costs associated with the architecture’s basic inference cycle, which alternates between selecting a focus belief, elaborating current explanations, and repairing violated constraints. At a higher level, we study PENUMBRA’s effectiveness at searching the space of alternative explanations for a set of observations. We conclude with comments on related work and proposals for future research.