Guidelines for the verification and validation of expert system software and conventional software: Evaluation of knowledge base certification methods. Volume 4

This report presents the results of the Knowledge Base Certification activity of the expert systems verification and validation (V ampersand V) guideline development project which is jointly funded by the US Nuclear Regulatory Commission and the Electric Power Research Institute. The ultimate objective is the formulation of guidelines for the V ampersand V of expert systems for use in nuclear power applications. This activity is concerned with the development and testing of various methods for assuring the quality of knowledge bases. The testing procedure used was that of behavioral experiment, the first known such evaluation of any type of V ampersand V activity. The value of such experimentation is its capability to provide empirical evidence for -- or against -- the effectiveness of plausible methods in helping people find problems in knowledge bases. The three-day experiment included 20 participants from three nuclear utilities, the Nuclear Regulatory Commission's Technical training Center, the University of Maryland, EG ampersand G Idaho, and SAIC. The study used two real nuclear expert systems: a boiling water reactor emergency operating procedures tracking system and a pressurized water reactor safety assessment systems. Ten participants were assigned to each of the expert systems. All participants were trained in and then used a sequence of four different V ampersand V methods selected as being the best and most appropriate for study on the basis of prior evaluation activities. These methods either involved the analysis and tracing of requirements to elements in the knowledge base (requirements grouping and requirements tracing) or else involved direct inspection of the knowledge base for various kinds of errors. Half of the subjects within each system group used the best manual variant of the V ampersand V methods (the control group), while the other half were supported by the results of applying real or simulated automated tools to the knowledge bases (the experimental group)