On-line EM Algorithm and Reinforcement Learning

We previously proposed an on-line EM algorithm for Normalized Gaussian Network (NGnet), which is a network of local linear regression units. In this article, we will apply our approach based on the on-line EM algorithm to reinforcement learning problems. We will examine a task for swinging-up and stabilizing a single pendulum with a limited torque, and a task for stabilizing a double pendulum. As a result, our approach is much more efficient than that based on the gradient descent algorithm.