Author(s)

H.G. Zimmermann & R. Neuneier

Abstract

We present our experiences of time series modeling by finite unfolding in time. The advantage of this approach is that the set of learnable neural network functions is restricted by a set of regularization methods which do not constrain the essential dynamics. Keywords in this section are over- and undershooting, the analysis of cause and effect, and the estimation of the embedding dimension in a partially externally driven dynamic system. 1 Introduction Standard analysis of dynamic systems by feedforward neural networks translates the time series identification problem into a pattern recognition approach. Typi- cally, the first step in such an analysis looks for an appropriate description of the present time state which can be used as an input vector. This is usually very tricky because one has to consider many pre

Keywords