ANNEALED COMPETITION OF EXPERTS FOR A SEGMENTATION AND CLASSIFICATIONOF SWITCHING DYNAMICS

We present a method for the unsupervised segmentation of data streams originating from different unknown sources that alternate in time. We use an architecture consisting of competing neural networks. Memory is included to resolve ambiguities of input-output relations. To obtain maximal specialization, the competition is adiabatically increased dur ing training. Our method achieves almost perfect identification and se gmentation in the case of switching chaotic dynamics where input manif olds overlap and input-output relations are ambiguous. Only a small da taset is needed for the training procedure. Applications to time serie s from complex systems demonstrate the potential relevance of our appr oach for time series analysis and short-term prediction.