Analysis of switching dynamics with competing neural networks
K. Robert Müller, Jens Kohlmorgen, Klaus Richard Pawelzik · 1995
Introduction Neural networks have been broadly used as approximators of nonlinear functions and as tools for classification and prediction. Trained from input and output examples, they provide a powerful structure for the representation of relations present in data [1, 2]. An important prerequisite for the successful application of such systems, however, is a certain uniformity of the data. In most cases of temporally ordered data, a stationary process is assumed, i.e. the relations remain constant over time. If, on the contrary, the data originate from different sources, the assumption of stationarity has to be discarded. In principle, we can think of three different kinds of non-stationarities: (a) a superposition of many sources or the case, where the underlying system (b) drifts or (c) switches between different dynamics. In all cases standard approaches like simple multi-layer perceptrons are likely to fail to represent the underlying input-output relations. In the presen