Pruning strategies for constructive neural network learning algorithms