Reinforcement Learning in Industrial Applications

Abstract: Although reinforcement learning has gained great success in computer games, there are only few yet known implementations in ndustrial applications. This despite the fact that reinforcement learning offers interesting methods to optimise the control of nonlinear processes. In this thesis we have used two model free reinforcement learning algorithms (PPO and DDPG) to control three different simulations of industrial processes, the simplified Tennessee Eastman, original Tennessee Eastman and the Haldex brake valve. Both reinforcement learning algorithms could in almost all cases learn to reach a set point. In addition, hyperparameters were found to have a high impact on training performance. In conclusion, our tests indicate that the model free reinforcement learning algorithms are basically capable of controlling industrial processes. Python code for the PPO algorithm applied to the Original Tennesse Eastman process can be found at Github. 1 1 https://github.com/Heigke/Reinforcement-Learning-In-Industrial-Applications