Neural Network Performance on the Stochastic Exclusive-or Problem
Abstract
In this report the application of neural networks to the detection of variance transitions in Gaussian noise is considered. The problem, which consists of transition detection between a pair of input sample variances, is a benchmark example of hypothesis testing on a nonstationary stochastic process. In the case of neural net algorithms the testing of hypotheses results from the decision space output of the last layer of neurons. Variance transition detection in a Gaussian random process is probably the most tractable example upon which to study machine analysis of a nonstationary stochastic process. It is an obvious test bed for the analysis of neural network decisioning of stochastic data. In this report the modeling of the process as Gaussian is dictated by the desire for a comparison of neural network and classical detection techniques. For specific applications, other data-derived parameters, such as correlation length, may be more appropriate for the analysis of transitions. It is expected, however, that neural network structures required for hypothesis testing are independent of the nature of the sufficient statistics; dependent instead upon the pattern of mean values. The purpose of this research is to provide a theoretical foundation for work currently being done on satellite maneuver detection. The detection of maneuvers is performed by neural networks which have been trained upon signal variances, spectral widths, and autoregressive model coefficients from radar cross section data. This research is also relevant to the data fusion effort, in which these parameters are obtained from different sensors.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 08, 1988
- Accession Number
- ADA197789
Entities
People
- R. Y. Levine
Organizations
- Massachusetts Institute of Technology