Back propagation is a steepest descent type algorithm that normally has slow learning rate and the search for the global minimum often becomes trapped at poor local minima. This paper proposes an algorithm called modified recursive prediction error (MRPE) algorithm for training multilayered perceptron networks. MRPE is a modified version of recursive prediction error (RPE) algorithm. RPE and MRPE are based on Gaussian-Newton type algorithm that generally provides better performance than a steepest type algorithm such as back propagation. The current study investigates the performance of MRPE algorithm to train MLP networks and compares its performance to the famous back propagation algorithm. Three data sets were used for the comparison. It is found that the proposed MRPE is much better than back propagation algorithm.
© 2001-2024 Fundación Dialnet · Todos los derechos reservados