(363ae) Online Data-Driven Closed-Loop Model Predictive Control of Nonlinear Systems Using Artificial Neural Networks

Standard Model Predictive Control (MPC) methodology for nonlinear dynamical systems design follows a two step process: (1) offline development of a system model from the open-loop data, and (2) use of the developed model in design of the MPC policy. These offline-developed models are relatively unchanged during the application of MPC in the closed-loop setting, which could potentially limit the closed-loop performance of the system. In this abstract, we present a novel online data-driven closed-loop MPC approach which combines the two steps mentioned above into a single step, thus allowing for simultaneous model development and design of control actions in a closed-loop MPC framework. Specifically, our approach uses a generic Long Short-term Memory (LSTM) network with pre-specified hyperparameters and trains the LSTM in a closed-loop MPC framework by using the measured output obtained in response to the applied control actions. Consequentially, the LSTM predictive performance is improved in real-time by applying online weight updates using the system’s state feedback. In this presentation, we will first discuss our online data-driven closed-loop MPC algorithm. Then, we will demonstrate the feasibility of our approach in controlling the 2-dimensional Lotka-Volterra predator-prey system, the 3-dimensional chaotic Lorentz system, and a rat cardiovascular system subject to input constraints. Our control scheme successfully achieves set-point tracking for each system, despite no prior knowledge of the system dynamics. Additionally, our approach is relatively robust to both noisy dynamics and hyperparameter selection for the systems considered. Taken together, we have developed a control scheme that harmonizes the flexibility of MPC with the adaptability of neural networks to yield a real-time data-driven control scheme for nonlinear dynamical systems. We will conclude the presentation by discussing the potential limitations of our approach, and ongoing efforts to address stability and convergence within this control framework.