Learning State-Action Basis Functions for Hierarchical MDPs

Learning State-Action Basis Functions for Hierarchical MDPs
Sarah Osentoski - University of Massachusetts Amherst, USA Sridhar Mahadevan - University of Massachusetts Amherst, USA
This paper introduces a new approach to actionvalue function approximation by learning basis functions from a spectral decomposition of the state-action manifold. This paper extends previous work on using Laplacian bases for value function approximation by using the actions of the agent as part of the representation when creating basis functions. The approach results in a nonlinear learned representation particularly suited to approximating action-value functions, without incurring the wasteful duplication of state bases in previous work. We discuss two techniques to create state-action graphs: offpolicy and on-policy. We show that these graphs have a greater expressive power and have better performance over state-based Laplacian basis functions in domains modeled as Semi-Markov Decision Processes (SMDPs). We present a simple graph partitioning method to scale the approach to large discrete MDPs.

Sarah Osentoski - University of Massachusetts Amherst, USA
Sridhar Mahadevan - University of Massachusetts Amherst, USA

This paper introduces a new approach to actionvalue function approximation by learning basis functions from a spectral decomposition of the state-action manifold. This paper extends previous work on using Laplacian bases for value function approximation by using the actions of the agent as part of the representation when creating basis functions. The approach results in a nonlinear learned representation particularly suited to approximating action-value functions, without incurring the wasteful duplication of state bases in previous work. We discuss two techniques to create state-action graphs: offpolicy and on-policy. We show that these graphs have a greater expressive power and have better performance over state-based Laplacian basis functions in domains modeled as Semi-Markov Decision Processes (SMDPs). We present a simple graph partitioning method to scale the approach to large discrete MDPs.