Reconstructed Dynamics and Chaotic Signal Modeling.pdfVIP

Reconstructed Dynamics and Chaotic Signal Modeling Jyh-Ming Kuo and Jose C. Principe Computational NeuoEngineering Lab. University of Florida, Gainesville, FL 32611 kuo@synapse.ee.ufl.edu ABSTRACT: A nonlinear AR model is derived from the trons (MLPs) can become an accurate nonlinear mapper reconstructed dynamics of a signal. The underlying system [Hornik][Hecht-Nielsen]. In this work we train MLPs as is assumed to be nonlinear, autonomous, and determinstic. NAR models of some test signals, which were produced In this formulation, the output error scheme is shown to be by nonlinear, deterministic systems. We compare the more suitable than the equation error scheme in training a results of two adaptation schemes in AR modeling, network as a NAR model of the signal. A method to incor- namely equation error and output error schemes porate the information of the dynamical invariants in sig- [Shynk]. If the length of training sequences can be prop- nal modeling is proposed. erly chosen, the output error scheme is shown to be more suitable in modeling signal dynamics than equa- 1. INTRODUCTION tion error scheme which has been adopted by most research groups [Lapedes][Casdagli][Mead]. In this Signal modeling in either time domain or frequency pa