Threshold Dependent Robust Discrimination for Convex Probability Uncertainty Classes.
Abstract
A methodology for finding robust discriminators for composite binary hypotheses defined for uncertainty classes which are not necessarily 2-alternating capacitable is developed. Past robust discrimination schemes have been threshold independent. In this paper, we present a methodology for finding robust detection structures which are threshold dependent and which sharply upper-bound the Bayes risk over a specified input uncertainty class and the chosen detector threshold. The support of the random variables is assumed to have a finite number of elements. A robust detection structure results from solving an associated limiting minimization problem. Results on the existence of these solutions are presented and conditional solutions for the divergence and divergence/linear uncertainty classes are formulated.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 31, 1998
- Accession Number
- ADA352446
Entities
People
- K. R. Gerlach
Organizations
- United States Naval Research Laboratory