Modeling of a Highly Nonlinear Stochastic Process by Neural Networks

Methods of identifying of highly nonlinear processes defined by a small data set are presented in this paper. Neural networks of different structures are implemented on two types of data set in order to get corresponding nonlinear models. Two layers NN based on Levenberg-Marquardt algorithm is used, in the first part of the paper, in process of prediction. Trend of the considered highly nonlinear process is extracted by the Elman recurrent neural network. In the second part of the paper the bootstrap technique is used in order to get model with good generalization capability of the one very typical biological process. The individual bootstrap models of the mentioned process are implemented by two paralel radial basis function neural networks. The 5 bootstrap patterns are used to get the resulting bootstrap model. In the following of the paper the roughly approximative relation for computing the upper limit value of a reduction of standard deviation of bootstrap technique based model is evaluated.