Neural networks with polynomial piecewise-continuous activation functions for the problems of finding data patterns

Annotation: This paper investigates the application of neural network models using polynomial piecewise-continuous activation function (PPCAF) to solve the problems of finding patterns in data sets. The authors apply multi-layered feed forward neural networks (NNs) with four types of PPCAF and use the sliding window method to predict a time series and determine the amplitude of a given signal on the white noise background. The training process of the multi-layered feed forward NN is conducted with the Levenberg-Marquardt back-propagation algorithm. From the testing results, a comparison is made between a known multi-layered feed forward NN and four multi-layered feed forward NNs using various types of PPCAF. Based on the result of the comparison analysis, recommendation is made for the use of multi-layered feed forward NNs with PPCAF for the problems of finding patterns in data sets.

Keywords: fuzzy activation functions, fuzzy neural networks, neural network training, prediction, determining signal parameters, finding data patterns