Stochastic approximation for neural network weight estimation in the control of uncertain nonlinear systems

The use of neural networks for controlling a nonlinear system with unknown process equations is considered. To make such an approach practical, it is necessary that connection weights in the neural network be estimated. The use of a new stochastic approximation algorithm for this weight estimation that is based on a simultaneous perturbation gradient approximation is considered. It is shown that this algorithm can greatly improve on the efficiency of more standard stochastic approximation algorithms based on finite-difference gradient approximations