Extracting Symbolic Rules from Trained Neural Network Ensembles