Neural networks for system identification and state observation of dynamical systems
Abstract
The resurgence in neural network research in the mid-eighties provides alternative approaches to tackle problems in system identification and control of dynamic systems. In the investigation conducted here, applications of feedforward networks and recurrent networks are developed. In the case of feedforward networks, the structure of a three-layer network with Gaussian hidden units is used to approximate system nonlinearity. A procedure of determining the centers and standard deviation is detailed. The unknown cocfficients of Gaussian units are then trained by the method of Recursive Least-Squares. The method will expedite the learning compared to the more traditional method of gradient descent, although there is a mild increase in computation complexity. In the case of recurrent networks, the structure of the Hopfield network is adopted. Instead of tracking the output of the dynamic system, the parameters of the state space model of the system are identified. In addition, the Hopfield network is integrated in a process of adaptive state observation of an unknown system. The result is both the identification of the system parameters and the reconstruction of the state of the system. The convergence of Hopfield-based procedures and the speed of convergence will be discussed. Furthermore, the effect of noise on the parameter estimation by adaptive observer will also be investigated. As a collateral study of exploring neural networks for optimization, a conceptual design of learning the optimal control law of a dynamic system is proposed where two neural networks are involved in an iterative process to approximate the result of dynamic programming. The learning process is model based, therefore, system identification is needed. The convergence of the learning process is investigated for the case of a Linear Quadratic Regulator (LQR).
Degree
Ph.D.
Advisors
Shoureshi, Purdue University.
Subject Area
Computer science|Artificial intelligence|Mechanical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.