Statistical Decisions Utilizing Neural Nets

Abstract

Neural networks were developed that accurately determine the statistical characteristics: modality and number fo stochastic components of underlying probability distribution(s) for sample data. Sample data examples, used to teach the neural nets were generated utilizing either a single beta distribution or a mixture of beta distributions. Once the neural net learned to distinguish between unimodal and multimodal examples and also between unimodal and mixture densities, they were challenged with unknown test cases. The test cases were also generated from either a single beta distribution or a mixture of beta distributions. Therefore the initial test results apply to a restricted class of distributions having bounded domains. The initial testing of the neural networks consisted of 40 unknown sample data examples generated utilizing beta distributions. The results of these tests are: (1) correctly identified 39 out of the 40 as being either unimodal or multimodal, an accuracy of 97.5 percent. This exceeds the accuracy of currently available statistical methods; (2) correctly identified 36 out of 40 as having either one component or more than one component, an accuracy of 90 percent.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
May 30, 1990
Accession Number
ADA224752

Entities

People

  • George Schlenker
  • Jack Manata

Tags

Communities of Interest

  • Human Systems
  • Space
  • Weapons Technologies

DTIC Thesaurus Topics

  • Availability
  • Classification
  • Computer Programs
  • Computers
  • Computing System Architectures
  • Economic Analysis
  • Expert Systems
  • Histograms
  • Network Architecture
  • Neural Networks
  • Operations Research
  • Probability
  • Probability Distributions
  • Random Variables
  • Security
  • Test Sets
  • Training

Readers

  • Neural Network Machine Learning.
  • Statistical inference.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Neural Networks