A learning algorithm for multilayer perceptron as classifier

Multilayer perceptron can be trained with empirical data to estimate general real-valued functions or to be used as a pattern classifier to estimate indicator functions. The typical backpropagation learning algorithm and its variations do not distinguish the training of an MLP as a pattern classifier from that of a general function estimator. In this paper, we present a learning algorithm based on an optimization layer by layer (OLL) procedure. Its main difference from previously reported OLL-type learning algorithms is that the weights between the last hidden layer and the output layer are determined through optimization of a piecewise linear objective function subject to constraints designed specifically for training an MLP to be a pattern classifier. The performance of the proposed learning algorithm is compared with that of the backpropagation algorithm, the modified Newton´s method and the improved descending epsilon algorithm over multiple training sessions using both simulated and real data classification problems