Deep Deterministic Policy Gradient for Magnetic Levitation Control

Abstract

MagLev ball motion control is a challenging problem because of its nonlinear and unstable behaviour. It also requires to have a large operating range. In this work, we have extended the current success of a deep reinforcement learning algorithm in continuous control using the Deep Deterministic Policy Gradient (DDPG), to magnetic levitation (MagLev) control. In reinforcement learning, the reward function is very critical for learning the optimal policy. We propose a novel and simple reward function for DDPG and evaluate our method on MagLev control. Our DDPG-based controller shows the capability to learn from the system with comparable performance to the previously proposed reward function. It can also find the solution to control the ball at all positions, including the ones that were not shown during the learning stage.