Distributed Inferencing for Classification
Abstract
This objective of this project was to develop algorithms, tools, and architectures that will improve the classification of complex systems. The approach was based on representing the system as a network of parallel, autonomous units, communicating via messages that summarize the state of neighboring subsystems. This approach offers reduction in complexity, ease of programming, universality, and potential for parallel hardware implementation. A major part of the effort focused on augmenting the traditional belief-network representation with non-Bayesian methods, qualitative information, variable- strength defaults statements, and casual interactions. Our main effort has concentrated on seeking non-Bayesian methods for classification, where we have explored the applicability of Ordinal Conditional Functions and the Dempster- Shafer belief functions. The need for such formalisms stems from the fact that we often do not possess the probabilistic knowledge required for full Bayesian analysis. For example, we may not know the components' failure rates and, which is more often the case, we may not be able to rate which among several possible modes of failure is likely to be realized when a given component fails
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 30, 1991
- Accession Number
- ADA267034
Entities
People
- Judea Pearl
Organizations
- University of California, Los Angeles