Learning to Tune a Class of Controllers with Deep Reinforcement Learning

Control systems require maintenance in the form of tuning their parameters order to maximize performance face process changes minerals processing circuits. This work focuses on using deep reinforcement learning train an agent perform this continuously. A generic simulation a first-order with time delay, controlled by proportional-integral controller, was used as training environment. Domain randomization environment aid generalizing unseen conditions physical circuit. Proximal policy optimization agent, and hyper-parameter performed select optimal neural network size algorithm parameters. Two agents were tested, examining impact observation space concluding that best consists auto-regressive exogenous input model fitted measurements variable. The trained deployed at industrial comminution circuit where it tested two flow rate control loops. improved one these loops but decreased other loop. While does show promise controller tuning, several challenges directions for further study have been identified.