Font Size: a A A

Optimal feedback controller approximation via neural and fuzzy-neural networks

Posted on:1998-11-29Degree:Ph.DType:Dissertation
University:The University of Texas at AustinCandidate:Niestroy, Michael AnthonyFull Text:PDF
GTID:1468390014976997Subject:Engineering
Abstract/Summary:
This dissertation presents three methods for the training of a neural or fuzzy-neural network to approximate an optimal feedback controller for nonlinear dynamical systems with state, control and terminal constraints. The controller is trained to operate over a subspace of initial conditions selected from the normal or design operating range of the dynamical system. The first of these design methods, referred to as the indirect method, relies upon existing optimal control generation techniques to produce a training set needed to tune the network parameters to approximate the optimal control. The second, direct training method uses a nonlinear programming technique to tune the network parameters. The method is direct in the sense that the system performance index and constraints are satisfied explicitly in the nonlinear programming process, in contrast to the indirect method which attempts to satisfy these requirements implicitly. The hybrid training method is then proposed which takes the network resulting from indirect training as a starting point for the direct training method. This hybrid method incorporates information provided by conventional optimal control techniques into a controller which is then tuned by the nonlinear programming algorithm to enforce state, control and terminal constraints. Demonstrations of these methods for three example problems show that both the neural and fuzzy-neural networks can be trained to act as approximate nonlinear optimal feedback controllers for nonlinear systems.
Keywords/Search Tags:Optimal feedback, Controller, Network, Neural, Method, Nonlinear, Training, Approximate
Related items