Neurocontrol design using ordered networks

Abstract

The problem of optimal control approximation for discrete time nonlinear dynamic systems using ordered neural network architectures is considered. Ordered networks generalize the notion of traditional feedforward layered networks by allowing inner-layer connections as well as forward connections between the layers. The aim of the paper is to present a self-contained description of the neurocontrol design process using ordered neural networks in the closed-loop nonlinear discrete time dynamical systems. This is accomplished by, first, developing closed-loop system sensitivity equations, and, second, incorporating them into the neurocontrol training process. The results reported here can be viewed as the extension of the work presented in Reference 1 to the class of fully forward connected neural network architectures. A neurocontrol design example is presented, and three aerospace applications of the proposed method are discussed.