Approximate models for nonlinear dynamical systems and their generalization properties (Q5936775)

A methodology to obtain approximate models from input-output data for nonlinear, causal, time-invariant discrete systems, having a certain type of continuity condition called fading memory, is presented. Besides, the region of input space, where the model can be applicable, the so-called generalization domain, is studied in this context, too. The method used consists in using a finite set of discrete Kautz systems, followed by a single hidden layer perceptron. Some strategies are considered to determine the parameters of this structure, i.e., the number of Kautz functions and the number of neurons of the hidden layer. The number of Kautz functions is evaluated by means of the Lipschitz coefficients, and the number of neurons is evaluated by pruning techniques. Then the proposed method of approximation of nonlinear dynamical systems and its generalization properties are applied to two examples: The approximation of a laboratory pendulum and of a simple discrete dynamical system. It is argued that the results of simulation show an improvement in accuracy of approximation and a reduction of the model order complexity. To this end not so bad. But the point is that the method was applied to known systems. So the question is: How successful this method would be in the case of real world data?