An Adaptive System Identification Algorithm with a General Performance Index Based on Entropy Optimization

This paper presents an entropy minimization algorithm for nonlinear system identification based on the information theory. The Parzen windowing estimator is used to approximate the entropy when the probability density functions of the variances can not be known as a priori or the variances are not realistically expressed with the traditional probability density functions. A general performance index based on the information entropy is discussed in this paper. Minimizing the performance index adopted can make the desired output of the adaptive system being tracked directly by the output of the neural network identifier. Furthermore, this performance index can be easily extended when treating other control problems. The performance of the entropy optimal algorithm is shown by several simulations with backpropagation neural networks.