| In fuzzy domains where precise and well-defined problem solving domain models do not exist or are hard to derive, two important issues dominate knowledge-based systems development: efficiency and reliability. Moreover, for complex applications, as the size of knowledge bases increases, the search for specialized knowledge in particular problem solving situations becomes the computational bottleneck.;In this work, we focus on an approach to improve efficiency that directly exploits domain structure present in the problem solving knowledge encoded as rules in the knowledge base. Domain structure refers to inherent regularities and differences that can be observed in the expert-supplied knowledge. This structure should have a direct bearing on problem solving tasks, and it can be exploited to reorganize the expert-supplied problem solving knowledge and to make access of problem solving knowledge faster and more reliable. In this framework, we introduce the concept of rule models and build a rule model hierarchy using ITERATE, an unsupervised conceptual clustering algorithm. The advantage of using the rule model hierarchy is that it suggests the most relevant form of knowledge at an appropriate level of detail to best suit the problem solving task.;The lack of precise domain models makes it difficult to develop or apply systematic knowledge acquisition methodologies. A practical solution is to adopt incremental knowledge acquisition methodologies, and try and maintain consistency using rule models and knowledge refinement techniques. We have developed a framework in which the expert interacts with the system and performs a four-step knowledge refinement task: discrepancy detection, fault location, repair suggestion, and repair validation.;These research contributions have been put together to build a general-purpose knowledge-based system construction tool called MIDST (Mixed-Inferencing Dempster Shafer Tool), which contains the following primary modules: a knowledge acquisition tool for building initial knowledge bases, reasoning system using rule models and problem-solving primitives for efficient reasoning, an efficient Dempster-Shafer combination scheme for evidence combination in uncertain reasoning environments, and knowledge refinement tool for maintaining the knowledge base. MIDST is used to build PLAYMAKER, a geological expert-system for characterizing hydrocarbon plays, as the focus of our expert system development research. |
