Data-Driven (DD) control has emerged as a discipline in the first half of the 1990’s. DD control consists of methodologies that provide safe auto-tuning controllers and adaptive controllers, which have found a wide range of applications. Our book, published by Springer in 2011, presents an unifying theory for DD control design methods and an overall account of their properties and commissioning principles, based on theory as well as on experimental case studies: Data-Driven Controller Design: the H2 Approach
DD design methods can be either iterative or in “one-shot”. Iterative methods provide sequential improvements in the tuning of a controller that is already operating in a closed loop. Each improvement is based on a batch of data collected from the closed-loop operation, exploiting the information that these data contain about how well the current controller is performing and about how the performance can be improved. “One-shot” methods provide the controller’s tuning based on only one batch of data, usually (but not necessarily) obtained in closed-loop operation. Both approaches have, of course, advantages and disadvantages.
It is probably fair to say that the most well known and most widely applied iterative methodology is IFT (Iterative Feedback Tuning). Regarding “one-shot” methods, VRFT (for Virtual reference Feedback Tuning) is probably the most common choice. It is very instructive to read the fundamental references on each method, namely the classical IFT paper by Hjalmarsson, Gevers, Gunnarsson and Lequin published in the Control Systems Magazine in 1998 and the also classical 2002 paper on VRFT by Campi, Lecchini and Savaresi, that appeared in Automatica. Some additional fundamental references that we like to recommend as initial reading material are: the FDT (Frequency Domain Tuning) paper by Kammer, Bitmead and Bartlett; the CbT (Correlation based Tuning) paper by Karimi, Miskovic and Bonvin; the data-based LQG control problem, by Skelton and Gi.
Though DD control has matured considerably as a scientific discipline, there is still plenty of work to be done to improve the methodologies and to extend their application to wider ranges of processes. The use of DD approaches for the tuning of nonlinear controllers, for instance, is an important subject yet to be developed. The inclusion of additional features in the design, such as guaranteed robustness and respect to state and control constraints, is equally relevant and is mostly an open field. Numerous descriptions of new applications appear constantly, showing the potential of DD methods but also highlighting application-specific problems waiting to be solved. In our research group we investigate all these problems, searching for practical solutions and tackling the theoretical challenges involved.
Black-box system identification is also often referred to as DD, particularly when the identification is aimed at obtaining a model for model-based control design. This is also a focus of our research, and thus of this page, and will be the subject of a future post.
