Constrained learning for model predictive control in asymptotically constant reference tracking tasks

摘要

There is a steadily increasing demand for full and partial autonomous operation of systems. One way to achieve autonomy for systems is the fusion of classical control approaches with methods from machine learning and artificial intelligence. We consider machine learning approaches to learn unknown or partially known references to increase the autonomy and performance of control systems for reference trajectory tracking. To improve learning and provide guarantees, we incorporate system properties such as constraints and the system dynamics in the learning algorithm. In particular, Gaussian processes are used to support a model predictive control scheme that exploits the predicted—learned—reference. Recursive feasibility and stability is established, and improved performance is illustrated considering a chemical process and asymptotically constant references.