Learning Numeric Action Models with Anytime Guarantees

Abstract

Model-based planning requires a predictive model of the world, yet such models are often hand-crafted. We address this by presenting a framework that learns PDDL action models with numeric fluents automatically from demonstrations. Our approach compactly represents the infinite space of linear numeric preconditions and effects consistent with the data, learning each action modularly and exploiting simpler structures when present. The framework provides anytime soundness and completeness guarantees with respect to the true hidden model, yielding conservative under-approximations and optimistic over-approximations at every stage of learning. Our findings across standard benchmark domains shows the advantages of our framework over existing approaches.