Multivariable output probability density function control for non-gaussian stochastic systems using simple MLP neural networks

This paper presents a new algorithm for the modelling and control of the shape of the output probability density functions (PDF) for general multivariable non-Gaussian dynamic stochastic systems. The purpose of such a control requirement is to design a control input vector so that the shape of the output PDF is made to follow a given target PDF function. In this context, the instant output PDF is approximated by a standard two-layers Multi-layer Perceptron (MLP) neural network subjected to some constraints. An alternative control algorithm is obtained by asymptotically minimizing a functional distance between the target and the output PDFs. In the case that the target PDF cannot be given, a minimized output entropy control algorithm is derived. It has been shown that both control algorithms can stabilize the closed loop system. Copyright (C) 2003 IFAC.