Automating Knowledge Acquisition for Expert Systems

In June of 1983, our expert systems research group at Carnegie Mellon University began to work actively on automating knowledge acquisition for expert systems. In the last five years, we have developed several tools under the pressure and influence of building expert systems for business and industry. These tools include the five described in chapters 2 through 6 -­ MORE, MOLE, SALT, KNACK and SIZZLE. One experiment, conducted jointly by developers at Digital Equipment Corporation, the Soar research group at Carnegie Mellon, and members of our group, explored automation of knowledge acquisition and code development for XCON (also known as R1), a production-level expert system for configuring DEC computer systems. This work influenced the development of RIME, a programming methodology developed at Digital which is the subject of chapter 7. This book describes the principles that guided our work, looks in detail at the design and operation of each tool or methodology, and reports some lessons learned from the enterprise. of the work, brought out in the introductory chapter, is A common theme that much power can be gained by understanding the roles that domain knowledge plays in problem solving. Each tool can exploit such an understanding because it focuses on a well defined problem-solving method used by the expert systems it builds. Each tool chapter describes the basic problem-solving method assumed by the tool and the leverage provided by committing to the method.

1. Introduction.- 2. MORE: From Observing Knowledge Engineers to Automating Knowledge Acquisition.- 2.1. Introduction.- 2.1.1. Goals for MORE.- 2.1.2. MORE’s Problem-Solving Strategy.- 2.1.3. An Overview.- 2.2. Strategies for Knowledge Acquisition.- 2.2.1. The Genesis of the Strategies.- 2.2.2. Summary.- 2.3. Knowledge Representation.- 2.3.1. The Event Model.- 2.3.2. An Example Representation.- 2.4. From Event Model to Rules.- 2.4.1. Generating Rules.- 2.4.2. Advice about Confidence Factors.- 2.5. Strategy Evocation and Implementation — Advice for Improving the Knowledge Base.- 2.6. The Problem-Solver Revisited.- 2.7. Learning from MORE.- 2.7.1. MORE’s Problems.- 2.7.2. Improving on MORE with TDE.- 2.7.3. User Interface.- 2.8. Conclusion.- 3. MOLE: A Knowledge-Acquisition Tool for Cover-and-Differentiate Systems.- 3.1. Introduction.- 3.2. MOLE’s Problem-Solving Method and Knowledge Roles.- 3.2.1. The Cover-and-Differentiate Problem-Solving Method.- 3.2.2. Diagnosing Inefficiencies in a Power Plant.- 3.3. Acquiring the Knowledge Base.- 3.3.1. Acquiring the Initial Symptoms.- 3.3.2. Acquiring Covering Knowledge.- 3.3.3. Acquiring Differentiating Knowledge.- 3.4. Handling Uncertainty.- 3.5. Identifying Weaknesses in the Knowledge Base.- 3.5.1. Refining Covering Knowledge.- 3.5.2. Refining Differentiating Knowledge.- 3.6. MOLE’s Scope.- 3.7. Conclusion.- 4. SALT: A Knowledge-Acquisition Tool for Propose-and-Revise Systems.- 4.1. Introduction.- 4.2. Acquiring Relevant Knowledge Pieces.- 4.3. Analyzing How the Pieces Fit Together.- 4.3.1. General Completeness.- 4.3.2. Compilability.- 4.3.3. Convergence.- 4.4. Compiling the Knowledge Base.- 4.5. Explaining Problem-Solving Decisions.- 4.6. Evaluating Test Case Coverage.- 4.7. Understanding SALT’s Scope.- 4.7.1. Acquiring Relevant Knowledge Pieces.- 4.7.2. Compiling the Knowledge Base.- 4.8. Conclusion.- 5. KNACK: Sample-Driven Knowledge Acquisition for Reporting Systems.- 5.1. Introduction.- 5.2. The Presupposed Problem-Solving Method and Its Knowledge Roles.- 5.3. Acquiring Knowledge.- 5.3.1. Acquiring the Sample Report.- 5.3.2. Acquiring the Domain Model.- 5.3.3. Generalizing the Sample Report.- 5.3.4. Demonstrating Understanding of the Sample Report.- 5.3.5. Defining, Generalizing, and Correcting Strategies.- 5.4. Analyzing the Knowledge Base.- 5.5. Rule Generation.- 5.6. Combining Problem-Solving Methods.- 5.6.1. The Combined Method.- 5.6.2. Acquiring Additional Knowledge.- 5.6.3. Generating Additional Rules.- 5.7. KNACK’s Scope.- 5.7.1. KNACK Tasks.- 5.7.2. Some Performance Data.- 5.8. Conclusion.- 6. SIZZLE: A Knowledge-Acquisition Tool Specialized for the Sizing Task.- 6.1. Introduction.- 6.2. Problem-Solving Strategies for Sizing.- 6.2.1. Sizing Knowledge.- 6.2.2. Methods for Computer Sizing and Knowledge Acquisition.- 6.2.3. The Choice: Extrapolation from a Similar Case.- 6.3. Using SIZZLE.- 6.4. Knowledge Representation and Proceduralization.- 6.5. The Scope of the Knowledge-Acquisition Tool.- 6.6. Conclusion.- 7. RIME: Preliminary Work Toward a Knowledge-Acquisition Tool.- 7.1. Introduction.- 7.2. A Knowledge-Acquisition Tool for XCON?.- 7.2.1. Acquiring Knowledge.- 7.2.2. Generating an Application.- 7.2.3. Issues addressed by RIME.- 7.3. What is RIME?.- 7.3.1. Control.- 7.3.2. Focus of Attention.- 7.3.3. Organizational Structures.- 7.3.4. Programming Conventions.- 7.4. Scope of Applicability.- 7.5. Future Directions.- 8. Preliminary Steps Toward a Taxonomy of Problem-Solving Methods.- 8.1. Introduction.- 8.2. A Few Data Points.- 8.2.1. MOLE.- 8.2.2. YAKA.- 8.2.3. SALT.- 8.2.4. KNACK.- 8.2.5. SIZZLE.- 8.2.6. SEAR.- 8.3. Conclusions.- References.