Distributed Inferencing for Classification
Abstract
In the first year of funding our research focused on classification tasks under uncertainty, with special emphasis on propagating beliefs in a system containing a mixture of categorical and probabilistic relationships. Previously available techniques were able to handle taxonomic trees, where the main relationship is ISA--i.e., class membership. The knowledge involved in object recognition contains non-binary relationships (e.g., IN-BETWEEN) arranged in non-decomposable structures. The difficulties encountered stem from the incompleteness of the model. In other words, we normally have information about the relationship between a variable and each of its neighbors but not between a variable and all of its neighbors taken together. This precludes the construction of a complete Bayesian model. To overcome this difficulty, we have developed a new formulation of the Dempster-Shafer theory in terms of a static constraint-networks (representing the stable knowledge) bombarded by randomly fluctuating constraints, (representing uncertain evidence) (Appendix I). We have also devised a scheme for computing Dempster-Shafer belief functions in that model, using Assumption-Based Truth Maintenance Systems (ATMS) and Incidence Calculus (Appendix I).
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 30, 1988
- Accession Number
- ADA208182
Entities
People
- Lashon B. Booker
Organizations
- University of Southern California