Boundary search procedure for the design and training of an ontogenic neural network classifier

The article presents a new heuristic for the construction and training of an ontogenic neural network for classification. It uses a boundary search procedure that incrementally generates the hidden layers in a neural network. This classifier is a heuristic designed to create hyper-rectangular shaped surfaces to mask the separate classes. The resulting structure is then used to determine the structure and weights of the neural network. The Wisconsin Breast Cancer Data is used as a test set to compare this new method against back-propagation neural networks and against well established linear programming based methods designed to train a neural network. The results show that this new method is able to produce correctness rates that are competitive to the other classification methods while training in a much more efficient manner.