A self-organizing system for the development of neural network parameter estimators

The design an optimal neural network estimator from training data is difficult because: 1) the required complexity of the estimation network is unknown, 2) existing training algorithms for multilayer perceptrons (MLPs) are inefficient, in terms of training time and use of free parameters, 3) existing bounds on neural network estimation error assume noiseless inputs and are not practical to calculate, 4) there is no generally accepted procedure for finding the best subset of input features to be used in optimal estimation, and 5) a method for automatically developing optimal estimators from training data is not available. In this paper, we present a methodology for attacking these problems. We describe three separate processing blocks which attempt to solve problems (1), (2), and (3). These blocks are then assembled into larger compound systems or blocks which attempt to solve the remaining problems. Examples of multilayer perceptron (MLP) estimators, designed using the proposed system, are given