Prediction intervals and lack-of-fit tests for neural networks.

摘要

This purpose of this dissertation is to determine if a typical lack-of-fit test used in regression modeling where the data contain true replicates or near replicates can be used to select an adequate model for a model created with a neural network program and to determine the reliability of the predictions from a model created by a neural network program. Also provided is a brief overview of artificial neural networks and some interesting aspects of neural network programs discovered during a simulation study involving three different neural network programs namely Braincel, Ripley's S+ program, and Nychka's S+ program.