Algebraic reinforcement learning hypothesis induction for relational reinforcement learning using term generalization

Abstract

The TG relational reinforcement learning algorithm builds first-order decision trees from perception samples. To this end, it statistically checks the significance of hypotheses about state properties possibly relevant for decision making. The generation of hypotheses is restricted by constraints manually specified a priori. In this paper we propose Algebraic Reinforcement Learning (ARL) for eliminating this condition by employing rewrite theories for state representation, enabling induction of hypotheses from perception samples directly via term generalization with the ACUOS system. We compare experimental results for ARL with and without generalization, and show that generalization positively influences convergence rates and reduces complexity of learned trees in comparison to trees learned without generalization.