A comparative evaluation of sequential constructive methods

A general treatment of a particular class of learning techniques for neural networks, called sequential constructive methods, is proposed. They subsequently add units to the hidden layer until all the input-output relations contained in a given training set are satisfied. Every addition involves the update of a small portion of the whole weight matrix and depends on a subset of samples whose size decreases with time. In most cases this leads to a large reduction of the computational cost. General convergence theorems are presented that ensure the achievement of a good multilayer perceptron within a finite execution time. The output weights need not to be trained but are obtained by the application of simple algebraic equations.