A novel reinforcement learning-based approach for optimal control: An application to multi-tank water systems
Eva Masero, Giacomo Mussita, Alessio La Bella, Riccardo Scattolini · Engineering Applications of Artificial Intelligence · 2025
The optimization of control actions is a critical challenge in industrial systems, especially when dealing with complex and unknown dynamics. Data collected from the environment enables the application of reinforcement learning techniques, which let the controller learn a policy based on data. This work proposes a novel model-free reinforcement learning approach that consists of a value iteration algorithm based on separate policy evaluation and policy improvement phases to provide an accurate control policy estimation. The proposed approach addresses tracking control for quadruple-tank water systems while obtaining minor tracking errors and faster transient responses. The results from the case study reveal better accurate estimation of the value function, up to 86.13% mean improvement in tracking accuracy and faster responses compared to existing methods. Therefore, the proposed approach demonstrates advantages in optimizing control performance and stands as a promising control method for industrial applications.