Environment Guided Interactive Reinforcement Learning: Learning from Binary Feedback in High-Dimensional Robot Task Environments

Isaac Sheidlower (Tufts University), Elaine Schaertl Short (Tufts University), Allison Moore (Tufts University)

Abstract

Continuous Action-space Interactive Reinforcement learning (CAIR) is the first continuous action-space interactive reinforcement learning algorithm that can out-preform state-of-the-art reinforcement learning algorithms early on in training. We test CAIR in two simulated robotics environments with intuitive and easy to design heuristic teachers.