An Experimental Evaluation of the Cascade-Correlation Network in Pattern Recognition Problems

An experimental investigation of the cascade-correlation network (CC) is carried out in diierent benchmarking pattern recognition problems. An extensive experimental framework is developed to establish a comparison between the CC network and the more traditional multilayer perceptron (MLP) and radial basis function models (RBF). The diierent networks are evaluated with respect to generalization performance in three real-world tasks: the diagnosis of coronary diseases, the credit screening problem and the recognition of handwritten characters. In addition to some advantages of the cascade-correlation network such as the dynamic deenition of the number of hidden units and relatively fast training, the practical results obtained suggest that the CC model may represent in some situations an alternative to the MLP and RBF networks.