A Selective Committee Architecture for Time Series Prediction and Pattern Classification

In this report we describe a novel technique to generate a committee architecture for time series predic~ion. The algorithm, here named Selective Multiple Prediction Network, consists of three steps: a systematic partition of the input hyperspace, a selective training of many agents and a flexible combining strategy. Potencially uncorrelated agents are generated which improves the combination process. The proposed architecture is easily extended to the class of classification problems.