A Comparison between Process Control Strategies: Reinforcement Learning with RBF and NMPC coupled with EKF

Reinforcement Learning (RL) arises from the set of Machine Learning techniques that are interesting for data-based process control purposes. Many authors discuss the advantages of coupling RL techniques to classical control frameworks, such as model predictive control (MPC), to tackle their drawbacks. In this work, an RL based on radial basis function (RBF) approximation is compared to nonlinear Model Predictive Control (NMPC) coupled with an extended Kalman filter (EKF) for the control and optimization of the Van de Vusse reactor, a multivariable benchmark characterized by its nonlinear dynamics. The algorithms are tested in the same simulation scenario, in which a setpoint-tracking objective is imposed to a variable while an extremum-seeking objective is imposed to another, and an unmeasured disturbance is applied to the plant. It is shown that both controllers can attend all the goals with some minor performance differences, and an RL coupled with RBF approximation is a promising framework for multi-objective, multivariable and data-based controller design of nonlinear systems.