Section: Systems and Control
Tutor: FAGIANO LORENZO MARIO
Advisor: SCATTOLINI RICCARDO Major Research topic
:Machine learning based data-driven control of nonlinear systemsAbstract:
One of the main factors preventing the application of advanced control techniques to real systems is the lack of an adequate model. When a physical model of the system is not available, e.g. when deriving it from first-principle equations is cumbersome, it is possible to resort to data-driven control strategies, i.e. to exploit the data collected from the field to infer information on the system and synthesize a controller. The goal of the research project is to investigate the application of machine-learning techniques in the context of data-driven control strategies applied to nonlinear systems. The first part of the project is devoted to indirect data-driven control strategies, and in particular to the employment of Neural Networks as black-box models of nonlinear dynamical systems. In order to effectively adopt such models for control purposes, several concerns must be addressed, namely:
- The choice of the structure of the Neural Network
- The definition of an appropriate training strategy
- The estimation of a bound on the prediction error
- The assessment of theoretical properties of the trained Neural Network, such as the Input-to-State Stability, which can be useful both during the controller design phase and to certify the representational capability of the network.
In the latter part of the project, the possible applications of machine-learning techniques to direct data-driven control strategies will be investigated, comparing the performances and advantages of direct methods to those of indirect ones.