Learning control for a class of discrete-time, nonlinear systems

Abstract

Over the last few decades, control theory has developed to the level where reliable methods exist to achieve satisfactory performance on even the largest and most complex of dynamical systems. The application of these control methods, though, often require extensive modelling and design effort. Recent techniques to alleviate the strain on modellers use various schemes which allow a particular system to learn about itself by measuring and storing a large, arbitrary collection of data in compact structures such as neural networks, and then using the data to augment a controller. Although many such techniques have demonstrated their capabilities in simulation, performance guarantees are rare. This thesis proposes an alternate learning technique, where a controller, based on minimal initial knowledge of system dynamics, acquires a prescribed data set on which a new controller, with guaranteed performance improvements, is based.