Accession Number:

ADA247941

Title:

Stretch and Hammer Neural Networks for N-Dimensional Data Generalization

Descriptive Note:

Final rept. May-Oct 1991

Corporate Author:

WRIGHT LAB WRIGHT-PATTERSON AFB OH

Personal Author(s):

• Raeth, Peter G.
• Gustafson, Steven C.
• Little, Gordon R.
• Puterbaugh, Todd S.

Report Date:

1992-01-15

Pagination or Media Count:

102.0

Abstract:

A hypersurface stretch and hammer neural network has been developed that generalize data from processes that have one output variable and one or more input variables. This network achieves several desirable properties through a novel combination of standard methods. The methods incorporate principal components, linear least squares, Gaussian radial basis functions, and diagonnally dominant matrices. An easily visualized physical model of network function ensures that the combination of methods is appropriate and practical. The model has natural potential for parallel implementation and for n- dimensional classification and other pattern recognition tasks. These tasks include smoothing interpolation, filtering, and prediction extrapolation. The model can be extended to accommodate multiple outputs. Unlike many other neural networks such as backpropagation-trained networks, the training and performance characteristics of the stretch and hammer neural network. The trials on three-dimensional surface interpolation are also presented, as are notes on other potential applications.

Descriptors:

• *NEURAL NETS
• *PROTOTYPES
• *COMPUTER APPLICATIONS
• *COMPUTER AIDED MANUFACTURING
• INPUT
• OUTPUT
• FUNCTIONS
• MODELS
• TRAINING
• NETWORKS
• VARIABLES
• SURFACES
• THREE DIMENSIONAL
• INTERPOLATION
• CLASSIFICATION
• STANDARDS
• PATTERNS
• RECOGNITION
• PATTERN RECOGNITION
• FILTRATION
• HAMMERS

Subject Categories:

• Computer Systems
• Manufacturing and Industrial Engineering and Control of Production Systems

Distribution Statement:

APPROVED FOR PUBLIC RELEASE