Alicia Esparza

Technical communique: Extensions to virtual reference feedback tuning: A direct method for the design of feedback controllers

The papers [Campi, Lecchini & Savaresi (2002). Automatica, 38(8), 1337-1346; (2003). European Journal of Control, 9(1), 66-76] present a direct controller synthesis procedure that uses identification algorithms applied to filtered input-output plant data. This contribution discusses variations that, in some cases, may alleviate noise-induced correlation (in the open-loop case) and allow the applicability of the approach to unstable plants. Importantly, it also introduces an invalidation test step based on the available data (i.e., prior to experimental controller testing), to check if the flexibility of the controller parameterisation and the approximations involved are suitable for the design objectives or, on the contrary, the resulting closed loop may be unstable.

Neural networks in virtual reference tuning

Abstract This paper discusses the application of the virtual reference tuning (VRT) techniques to tune neural controllers from batch input–output data, by particularising nonlinear VRT and suitably computing gradients backpropagating in time. The flexibility of gradient computation with neural networks also allows alternative block diagrams with extra inputs to be considered. The neural approach to VRT in a closed-loop setup is compared to the linear VRFT one in a simulated crane example.

VIRTUAL REFERENCE FEEDBACK TUNING IN RESTRICTED COMPLEXITY CONTROLLER DESIGN OF NON-MINIMUM PHASE SYSTEMS

Abstract This paper discusses the applicability of the identification-based Virtual Reference Feedback Tuning scheme in (Campi et al. , 2002; Campi et al. , 2003; Sala and Esparza, 2005) to reduced-order controller design. As the presence of zeros outside the unit circle is quite usual in sampled-data systems, a particular discussion on the topic is carried out. Also, reduced-order controllers identified with the VRFT scheme may not meet the specifications (or even be unstable) so some invalidation tests are needed prior to experiment.

Reduced-Order Controller Design via Iterative Identification and Control

In this paper, a customized version of iterative identification and control algorithms is presented, oriented to the design of reduced-order controllers with increasing performance (in terms of bandwidth for reference tracking). Although a particular set of identification and control design methodologies is chosen due to its simplicity and wide availability, some of the ideas can be applied to alternative methods. The procedure is based on estimating reduced-order models on narrow control-oriented frequency bands when the closed loop performs badly, and interpolating those models and previous ones allowing larger errors at frequencies unimportant for control purposes, so that reduced-order models can still be used. The allowable error for the interpolated model will be determined by commonly used small-gain inequalities.

APPLICATION OF NEURAL NETWORKS TO VIRTUAL REFERENCE FEEDBACK TUNING CONTROLLER DESIGN

Abstract This paper discusses the application of the Virtual Reference Feedback Tuning technique to tune neural controllers from experimental data, by particularising nonlinear VRFT and suitably computing gradients backpropagating in time. Alternative block diagrams with extra inputs have also been considered. The neural approach to VRFT is compared to the linear one in a simulated crane example.

Brief paper: Asymptotic statistical analysis for model-based control design strategies

In this paper, we generalize existing fundamental limitations on the accuracy of the estimation of dynamic models. In addition, we study the large sample statistical behavior of different estimation-based controller design strategies. In particular, fundamental limitations on the closed-loop performance using a controller obtained by Virtual Reference Feedback Tuning (VRFT) are studied. We also extend our results to more general estimation-based control design strategies. We present numerical examples to show the application of our results.

Iterative Identification and Control Design: Methodology and Applications

When using model-based controller design methodologies in order to control a plant whose model is not a priori known, there are several practical issues to be taken into account. Theoretically, the procedure is to identify an accurate enough model of the plant and, then, apply any model-based controller design technique. But the real world is much more complex and issues as the presence of nonlinearities in the plant or/and in the actuators, the noise that always corrupts the data, the existing actuator constraints or computational limitations, etc. arise, imposing natural limitations over the achievable performance of the plant to be controlled as well as in the possible model to be experimentally obtained. In this contribution, an iterative framework has been used to overcome some of these problems. The proposed algorithm starts by using a very rough plant model and low-demanding specifications. These and the plant model are progressively improved only as needed, i.e. until the final desired performance is achieved or no improvements are attained in consecutive iterations. This procedure somehow avoids useless effort: on the one hand, the identification stage is only a tool for controller design, not aiming at achieving a very accurate model at any working condition but just at those frequencies that are interesting for control purposes; on the other hand, the bandwidth is increased as long as the plant allows it, avoiding undesirable experimental results trying to achieve what is not possible at all.

Encoding Fuzzy Diagnosis Rules as Optimisation Problems

This paper discusses how to encode fuzzy knowledge bases for diagnostic tasks (i.e., list of symptoms produced by each fault, in linguistic terms described by fuzzy sets) as constrained optimisation problems. The proposed setting allows more flexibility than some fuzzy-logic inference rulebases in the specification of the diagnostic rules in a transparent, user-understandable way (in a first approximation, rules map to zeros and ones in a matrix), using widely-known techniques such as linear and quadratic programming.