Improving Short-term Prediction with Competing Experts

We show, how competing neural networks can improve short-term prediction of time series which originate from systems with multiple modes of behaviour. With the presented method, each expert network specializes on a diierent dynamical mode and the time series will be segmented accordingly. In order to obtain a maximal specialization, the competition is adi-abatically increased during training. Memory is included in order to resolve ambiguities of input-output relations. We illustrate the properties of the method in the case of switching chaotic dynamics. The application to Data Set D from the Santa Fe Time Series Prediction Competition demonstrates the potential relevance of this approach for time series analysis and short-term prediction.