This dissertation addresses the classification problem for applications with extensive amounts of data and complicated features. The learning system developed utilizes a hierarchical multiple classifier learning scheme and is flexible, efficient, highly accurate and of low cost.; The learning system achieves the goals of reducing the training cost and increasing the prediction accuracy compared to other multiple classifier algorithms. Various clustering algorithms are utilized to partition the training data to decrease the training burdens of component classifiers, while the use of subclass labeling improves the estimation of decision boundaries. The combination of component classifiers is achieved by the training of super-classifier with cross-validation.; The system was tested on three large sets of data in which two of them are for the medical diagnoses application from NeoPath Inc. and one is for the prediction of forest cover type from UCI Repository. The results obtained are superior compared to other learning algorithms. |