Reinforcement Learning for Controlling a Coupled Tank System Based on the Scheduling of Different Controllers

摘要

Reinforcement Learning has been an approach successfully applied for solving several problems available in literature. It is usually employed for solving complex problems, such as the ones involving systems with incomplete knowledge, time variant systems, non-linear systems, etc., but it does not mean that it cannot be applied for solving simple problems. Therefore, this paper proposes an alternative application, where RL could be applied to switch among controllers with a fixed tuning, in a system with a known non-linear dynamics, aiming to optimize its time response. It was shown, after the online training and test of the RL agent that it could take advantage of the best characteristics of the available controllers to improve the response of the coupled tank system.