An improved PE algorithm for continuous nonlinear system identification with application to learning in neural networks

A new approximating sensitivity algorithm in prediction error method is derived for a class continuous nonlinear dynamic system identification including training multi-layer neural networks. The algorithm can be used to approximate the gradient of output with respect to unknown parameter in a wic class of continuous-discrete nonlinear systems. The comparison between new and convention algorithm, and simulated example is included to demonstrate the effectiveness of the new algorithm.