Can I do that? Discovering domain axioms using declarative programming and relational reinforcement learning

Abstract

Robots deployed to assist humans in complex, dynamic domains need the ability to represent, reason with, and learn from, different descriptions of incomplete domain knowledge and uncertainty. This paper presents an architecture that integrates declarative programming and relational reinforcement learning to support cumulative and interactive discovery of previously unknown axioms governing domain dynamics. Specifically, Answer Set Prolog (ASP), a declarative programming paradigm, is used to represent and reason with incomplete commonsense domain knowledge. For any given goal, unexplained failure of plans created by inference in the ASP program is taken to indicate the existence of unknown domain axioms. The task of learning these axioms is formulated as a Reinforcement Learning problem, and decision-tree regression with a relational representation is used to generalize from specific axioms identified over time. The new axioms are added to the ASP-based representation for subsequent inference. We demonstrate and evaluate the capabilities of our architecture in two simulated domains: Blocks World and Simple Mario.