Supervised Learning Combined with an Actor-Critic Architecture TITLE2:

Abstract

To address the shortcomings of reinforcement learning (RL) a number of researchers have focused recently on ways to take advantage of structure in RL problems and on ways to make domain knowledge part of RL algorithms. In this paper we examine a {\em supervised} actor-critic architecture, whereby a supervisor adds structure to a learning problem and supervised learning makes that structure part of an actor-critic framework for reinforcement learning. We provide a steepest descent algorithm for real-valued actions such that the actor adjusts its policy in accordance with gradient information from both supervisor and critic. We also illustrate the approach with two kinds of supervisors: a feedback controller that is easily designed yet sub-optimal, and a human operator providing intermittent control of a simulated robotic arm.