Multilayer Perceptrons for Classification.

摘要

Techniques for training, testing, and validating multilayer perceptrons are thoroughly examined. Results obtained using perceptrons are compared and contrasted with two multivariate discriminant analysis techniques-logistic regression and k neighbor. Methods for determining significant input features are investigated and a procedure for examining the confidence to place in the significance of these features is developed. Procedures to evaluate the applicability of high-order feature inputs are examined. These methods and procedures are applied to two very different applications. The first application concerns the prediction of Air Force pilot retention/separation rates for input to force projection models. The second application concerns the classification of Armor Piercing Incendiary (API) projectiles based on firing parameters. Results showed that none of the classification methods considered was able to accurately classify individual pilot's retention decisions, however, multi perceptrons were judged to be the superior discriminator for the classification of API projectiles. For the API projectile analysis, a procedure to determine which input features are no more significant than noise was demonstrated. The resulting salient set of feature inputs was shown to train quicker and decrease the output error. A method to identify appropriate high-order inputs was also demonstrated.