A divide-and-conquer methodology for modular supervised neural network design

A novel learning strategy based on the divide-and-conquer concept is proposed to effectively overcome the slow learning speed and hard-determined network size problems in supervised learning neural networks. The proposed method first partitions the whole complex training set into several manageable subsets and then generates small size networks to 'conquer' (or learn) all these training subsets. In order to achieve efficient partition on a train set, we have proposed an error correlation partitioning (ECP) scheme such that sub-training-sets are formed with small (acceptable) training error. Based on this learning strategy, a self-growing modular neural network system can be developed. By applying the proposed learning strategy, a neural network is not only useful for pattern classification problems but also for continuous valued function approximation problems.<<ETX>>