Using 3-D Surface Maps to Illustrate Neural Network Performance
Abstract
This paper suggests a possible means for evaluating the performance of neural networks from a global perspective in parameter-space. Traditional evaluations tend to focus on performance in weight-space or on overall output error during one training session. However, a global perspective of performance in parameter-space may be of primary importance during the initial stages of problem solution. During these stages, the researcher is typically trying to determine a network configuration and suitable values for its training equation parameters. Instead of a hit-or-miss approach, this paper describes an organized experimental method that identifies network configuration and parameter value choices which are not sensitive to minor variations for a standard training metric. The technique is illustrated for the network used by Hopfield and Tank to solve a traveling salesman problem and with traditional Backpropagation as described by Lippmann.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1990
- Accession Number
- ADA578295
Entities
People
- Peter G. Raeth
Organizations
- Wright Laboratory