A neural network approach to direct data-driven control design

The aim of the thesis is to investigate the use of a particular class of recurrent neural network (RNN), called RNN-OE, in the context of direct data-driven controller (DDDC) design. We consider the problem of controlling both linear and non-linear systems. The peculiar feature of the DDDC approach is that no information is known about the mathematical model of the plant to be controlled; the only available information are the input and the output data collected by performing an open loop experiment on the plant. The controller is designed in such a way that the closed loop response of the system behaves like a linear model reference, designed to satisfy specific transient and steady-state quantitativerequirements. In the thesis two approaches are investigated for the design of RNN-OE controller: one is based on a straightforward extension of the already available DDDC method for linear systems; the other is inspired by the nonlinear virtual reference feedback tuning approach. The proposed RNN-based DDDC approach enjoy some peculiar feautures that make it substantially different from others neural networks-based approaches. In contrast to what is usually done in the classical approach to neural-control, the methodology proposed in this thesis only requires to select the structure of the controller (complexity of the RNN-OE) and to use the collected data to train the network, without building any explicit mathematical model for the plant. As far as the comparison with the reinforcement learning approach is concerned, the proposed approach requires less data in the training stage, it is more robust to parameter variations and does not required to be trained again when the control objective changes. The performance of the newly proposed RNN-OE design approaches are also compared to the ones of a linear controller design with the standard DDDC approach and pros and cons are discussed.