Active Set Prediction using Machine Learning Methods for Complexity Reduction in Nonlinear MPC

We use Machine Learning (ML) methods to simplify the Nonlinear Program (NLP) arising in Nonlinear Model Predictive Control (nonlinear MPC, NMPC). Since every solution to an NMPC problem defines a set of active and a set of inactive constraints, we propose to use predictions about these sets to reduce the complexity of the NLP to be solved. Specifically, we use ML methods to predict active sets for NMPC problems. The required classification networks are simple enough to be evaluated online, i.e., during the runtime of the controller. They can be trained to a high accuracy, qualifying as suitable candidates for an application to NMPC. The results are evaluated using numerical simulations for a model of a continuous stirred-tank reactor.