A Reinforcement Learning Approach to Robust Control in an Industrial Application

The objective of this study was to design and implement a reinforcement learning-based controller for a nonlinear industrial system, specifically a liquid water tank controlled via a programmable logic controller to achieve robust control in the presence of disturbances from the outlet drain valve at various ratios. Initially, the system’s model parameters were determined, and a mathematical model was developed using the OpenAI Gym open-source platform. Subsequently, multilayer perceptron-based reinforcement learning (RL), adaptive proportional integral (A-PI), and reinforcement learning-integral (RL-I) controllers were trained and validated using the developed software model. The designed controllers were then implemented on the real system both fixed and variable drain valve ratios. Tests conducted with a fixed drain valve ratio revealed that the proposed RL-I controller outperformed the RL and A-PI controllers in terms of transient and steady-state responses. The error values of the RL-I controller were significantly lower than those of the other algorithms (p = 0.000). In the final test, where the drain valve was adjusted to different ratios, the RL-I controller demonstrated robust performance. This study successfully developed a novel, robust controller for nonlinear systems commonly encountered in industrial applications.