Forming Concepts for Fast Inference

Knowledge compilation speeds inference by creating tractable approximations of a knowledge base, but this advantage is lost if the approximations are too large. We show how learning concept generalizations can allow for a more compact representation of the tractable theory. We also give a general induction rule for generating such concept generalizations. Finally, we prove that unless NP non-uniform P, not all theories have small Horn least upper-bound approximations.