Statistical Decision making with Uncertain and Conflicting Data
Abstract
Automated decision-making techniques are investigated for use in situations where incoming tactical reports are frequently inaccurate or incorrect and where information about probability models is limited. The first problem addressed is to find a decision method for deciding among a set of hypothesis (H sub i) based on a set of evidence (E sub j), given data quality factors (W sub j) and numerical probability judgements such as (P(H sub i/E sub j) or (P(E sub j/H sub i)). Four candidate decision methods are compared: the linear opinion pool, the logarithmic opinion pool, Dempster's rule, and a fuzzy-logic algorithm. Comparisons include the MYCIN certainty-factor calculus for the two-hypothesis examples. The decision methods give similar results, but none is fully satisfactory. At least one should perform adequately if selected carefully for the decision problem, but an effort should be made to find a fully suitable method. Several statistical techniques are investigated for measuring the relevance of evidence to a decision problem and for identifying suspicious or conflicting evidence. An approach to implementing the decision process in a knowledge-based system is outlined. The problem of eliciting and aggregating numerical probability judgements is also discussed.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1991
- Accession Number
- ADA243853
Entities
People
- R. A. Dillard