. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1 II. Problem Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1 2.1 Verification & Validation Testing . . . . . . . . . . . . . . . . 2-1 2.2 Methods of Verification & Validation . . . . . . . . . . . . . . 2-3 2.3 V & V of knowledge based systems versus conventional software 2-4 2.4 Knowledge Acquisition . . . . . . . . . . . . . . . . . . . . . 2-5 2.5 Knowledge Representation . . . . . . . . . . . . . . . . . . . 2-6 2.6 The Bayesian Knowledge Base representation . . . . . . . . . 2-7 2.7 Knowledge Base Errors . . . . . . . . . . . . . . . . . . . . . 2-9 III. Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1 3.1 PESKI Validation . . . . . . . . . . . . . . . . . . . . . . . . 3-2 3.1.1 Test Cases . . . . . . . . . . . . . . . . . . . . . . . 3-2 3.1.2 Direct Dependency R...

Frequently expert systems are being developed to operate in dynamic environments where they must reason about time-varying information and generate hypotheses, conclusions, and process inputs that can drastically influence the environment within which they operate. For instance, expert systems used for fault diagnosis and fault accomodation in nuclear power plants reason over sensor data and operator inputs, form fault hypotheses, make recommendations pertaining to safe process operation, and in crisis situations could generate command inputs to the process to help maintain safe operation. Clearly, there is a pressing need to verify and certify that such expert systems are dependable in their operation and can reliably maintain adequate performance levels. In this paper we develop a mathematical approach to verifying qualitative properties of rule-based expert systems that operate in dynamic and uncertain environments. First, we provide mathematical models for the expert system (including the knowledge-base and inference engine) and for the mechanism for interfacing to the user inputs and the dynamic process. Next, using these mathematical models we show that while the structure and interconnection of information in the knowledgebase influence the expert system’s ability to react appropriately in a dynamic environment, the qualitative properties of the full knowledge-base/inference engine loop must be considered to fully characterize an expert system’s dynamic behavior. To illustrate the verification approach we show how to model and analyze the qualitative properties of rule-based expert systems that solve a water-jug filling problem and a simple process control problem. Finally, in our concluding remarks we highlight some limitations of our approach and provide some future directions for research.

1 In most real-world applications, knowledge bases grow and change over time. This paper examines two solutions to the problem of knowledge base consistency maintenance as new knowledge is acquired and assimilated in an evolving knowledge base. One of the proposed solutions makes use of symbolic inference methods to identify potential inconsistencies that might result from the addition of new knowledge in the form of rules; the other uses a hybrid symbolic-connectionist system that can be re-trained using a database of examples whenever new knowledge needs to be assimilated. The proposed solutions are discussed in the context of an intelligent advisory system designed to help identify probable cases of discrimination in public housing in Iowa and several other states in the midwestern United States. Keywords: consistency checking, expert systems, machine learning, knowledge acquisition, neural networks. 1 Introduction In the context of this paper, we use the term knowledgebased syst...