The key feature of MPC is the knowledge of a model of the process to be control, which allows the controller to predict its future behavior and compute the optimal control moves accordingly.
The assumption of perfect knowledge of the model is however practically wrong since, as they say, all models are wrong (even though some of them are useful). Moreover, deriving a prediction model from first principle is in general difficulty and sometimes impossible.
To overcome this problem, one can try to derive models directly from input-output measurements. Even though data-driven modelling is not a new field, in the last decade, a cross-fertilization occurred between the system identiﬁcation and the statistical/machine learning communities, leading to the successful application of kernel methods to data-based learning of system dynamics.
This research line aims to study new methods for developing a learning-based MPC.
This study has direct implications to the Artificial Pancreas research line, since it opens the door to personalization and adaptation of MPC.