Dynamics approximation and change point retrieval from a neural network model

Discrete dynamical systems and nonlinear response models contaminated with random noise are considered. The objective of the paper is to propose a method of extraction of change-points of a mapping (system) from a neural network approximation. The method is based on the comparison of normalized Taylor series coefficients of the estimator and on the consistency of the estimator and its derivatives. The proposed algorithm is illustrated on a simple piecewise polynomial mapping.