Automating Rule Strengths in Expert Systems.
Abstract
Automating rule strengths in expert systems is a way to alleviate the knowledge acquisition bottleneck. It is assumed that rules are fixed, except for the values of their strengths, which are computed or adjusted from initial values given by experts. A model of expert systems is proposed, in which rules have the form IF (P sub 1 & P sub 2 & . . . & P sub n) THEN C WITH ATTENUATION a, where P sub 1, P sub 2 , . . ., P sub n, and C are weighted propositions, i.e., statements with a certainty factor (CF), and a, the strength of the rule, is a number between 0 and 1. To compute rule attenuations, two problem settings are considered. In the first, an oracle is given, that can provide the CFs of the conclusions of the entire rule-based system, given any assignment of certainty factors to the premises of the entire system (complete case). In the second, a fixed set of cases is available (incomplete case). A fast algorithm for synthesis in the complete case for simple rule bases is given both for MAX and probabilistic sum. In the incomplete case, the synthesis of attenuations is shown to be NP-Complete, even for very shallow rule bases with only two propositions in the premise of each rule, both for MAX and probabilistic sum. The refinement of attenuations from expert-given estimates is shown to be NP-Hard, no matter how close to the correct value the estimates are and how small an improvement towards the correct value is desired. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1987
- Accession Number
- ADA185626
Entities
People
- Marco Valtorta
Organizations
- Duke University