Connectionist Feedforward Networks for Control of Nonlinear Systems

Computer science, Artificial intelligence

Sbarbaro Hofer, Daniel G

University:University of Glasgow

关键词: Computer science, Artificial intelligence;   

Others  :  http://theses.gla.ac.uk/75258/1/13815422.pdf

来源: University of Glasgow

【 摘 要 】

The control of nonlinear systems is addressed from a new perspective, which makes use of several concepts and techniques developed in the area of Artificial Neural Networks. In particular, this work explores the potential of connectionist representations which consist of only feedforward connections of sigmoids or gaussian units. A unified review demonstrating the capabilities of these structures to approximate continuous nonlinear functions is given. The use of the Fourier transform and the properties of kernel functions are exploited in order to demonstrate some properties of gaussian networks. The adjustment of the different parameters plays a significant role in the relevance of these structures in control. For sigmoid networks a new learning algorithm is proposed, its main feature being the use of the forgetting factor and pseudoinverse. For gaussian networks several algorithms using different techniques, such as: the Fourier Transform, gradient approach, and/or clustering algorithm, are explored and compared. The representation of a dynamic system by means of a static nonlinear function, and consequently, by a connectionist representation, is addressed. Concepts such as controllability, observability, and invertibility, which are needed to develop any control structures, are put forward for the nonlinear case. Four control structures using connectionist models to generate the control signal are proposed, and its potential analysed. For each approach a simple example is presented to illustrate their performance. A summary of their main characteristics is also given. Two industrial applications have been tackled, and solutions developed, illustrating not only the differences between different techniques, but also the potential and limitations of the ideas pursued in this work. Finally, some suggestions are given, which may engender further research in the field of control and artificial neural systems.