CONSTRUCTIVE TRAINING ALGORITHM FOR DESIGNING FEEDFORWARD NEURAL NETWORKS: A REVIEW

Abstract

In this paper, we review neural networks, models of neural networks, methods for selecting neural network architecture and constructive algorithms for regression problems. Cascade correlation algorithm is the most suitable for solving regression problems. Dynamic Node Creation (DNC) algorithm is a method which automatically grows backpropogation networks until the target problem is solved. Cascade 2 algorithm is a variant of Cascade correlation algorithm. Recurrent CBPmethod addresses the construction of recurrent networks by the use of constructive backpropogation. Casper is known to produce more compact networks with very promising results. Adaptivesigmoidal activation function can be used for better generalization performance and training time can be reduced.