Integrating relational reinforcement learning with reasoning about actions and change

Abstract

This paper presents an approach to the integration of Relational Reinforcement Learning with Answer Set Programming and the Event Calculus. Our framework allows for background and prior knowledge formulated in a semantically expressive formal language and facilitates the computationally efficient constraining of the learning process by means of soft as well as compulsive (sub-)policies and (sub-)plans generated by an ASP-solver. As part of this, a new planning-based approach to Relational Instance-Based Learning is proposed. An empirical evaluation of our approach shows a significant improvement of learning efficiency and learning results in various benchmark settings. © 2012 Springer-Verlag Berlin Heidelberg.