Time Series Analysis Using Unsupervised Construction of Hierarchical Classifiers

May 2001

Proceedings of the Fourteenth International Florida Artificial Intelligence Research Society Conference (FLAIRS 2001)

Volume

Proceedings of the Fourteenth International Florida Artificial Intelligence Research Society Conference (FLAIRS 2001)

All Papers

Recently we have proposed an algorithm of constructing hierarchical neural network classifiers (HNNC), that is based on a modification of error back-propagation. This algorithm combines supervised learning with self-organisation. Recursive use of the algorithm results in creation of compact and computationally effective self-organised structures of neural classifiers. The above algorithm was expanded for unsupervised analysis of dynamic objects, described by time series. It performs segmentation of the analysed time series into parts characterised by different types of dynamics. This paper presents the latest successful results of testing the algorithm of time series analysis on pseudo-chaotic maps. Keywords: self-organisation, adaptive hierarchical classifiers, time series analysis, error back-propagation