Data allocation for large unsupervised decomposition of switiching time series by predictive modular neural networks

摘要

A switching time series is the outcome of several different sources, the sources being activated in succession. The problem of on-line, unsupervised learning of switching time series, consits in separating incoming data and corresponding a separate class corresponding to each source; then developing a distinct model for each class of data. This is solved by a predictive modular neural network, each module being trained on data from a particular source, rejecting all other data. In this context, data allocation consists in separating data from different classes in an on-line, unsupervised manner. We present a mathematical analysis regarding the convergence of a quite general calss of data allocation schemes based on predictive modular neural networks. The theoretical conclusions are supported by numerical experiments on several problems of on-line time series decomposition.