Criteria for Choosing the Best Neural Network: Part 1
Abstract
An investigation into the problem of determining a parsimonious neural network for use in prediction/generalization based on a given fixed learning sample was undertaken. Both the classification and nonlinear regression contexts were addressed. An exposition and survey of the problem and past research on model selection techniques in other statistical settings was compiled, and algorithms for selecting the number of hidden layer nodes in a three layer, feedforward neural network were developed. The selection criteria developed attempt to grow the networks beginning with a small initial number of hidden layer nodes (as opposed to pruning a relatively large network). For the nonlinear regression problem, the method is based on cross-validation estimates of the prediction mean squared error for the candidate networks. For the classification problem, the method is based on a cost complexity measure of the candidate networks based on resubstitution estimates of the probability of misclassification and a penalty function of the number of hidden layer nodes. Also considered was the use of principal.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 24, 1991
- Accession Number
- ADA247725
Entities
People
- J. E. Angus
Organizations
- Naval Health Research Center