Heuristic , Qualitative , and Quantitative Reasoning

The growing complexity, sophistication, and bulk of engineering knowledge has lead to situations where incorrect or sub-optimal decisions are made because the appropriate knowledge cannot be accessed and brought to bear in a timely fashion. Just as the number-crunching power of computers several decades ago made possible the routine application of analysis methods too time consuming to use by hand, so Artificial Intelligence (AI) computer techniques promise the power to access and apply specialized knowledge. Solving an engineering problem requires use of knowledge about: how to gather data to define the problem, how to structure the data into an engineering model, and how to analyze the model to get numeric data. Searching for a solution by reasoning with a simplified model and then verifying, revising, and refining the rough model lies at the heart of engineering problem solving. To effectively capture this behavior requires more than compiled empirical knowledge in the form of antecedent-consequent rules. Analysis and simulation include large chunks of procedural knowledge, not easily accommodated by a rule-based paradigm. The definition, modeling, and interpretation steps are largely symbolic while the analysis step is largely numeric. Qualitative reasoning is one mechanism that can be used to make the transformation from heuristic knowledge to an engineering model suitable for mathematical manipulation. The strategy of using an intermediate qualitative simulation layer manipulating first order engineering models to connect a predominantly heuristic and symbolic rule-based top layer with a largely procedural and numeric quantitative root layer is applicable to a wide variety of engineering problems. Rules are good for generating hypotheses to define the initial areas to search. Qualitative methods can be used to enumerate possible behaviors and focus on promising models. Quantitative analysis resolves ambiguous behavior and provides quantified answers. Each method is a better tool for a different problem solving phase. This thesis presents such a unified tool implemented with a three reasoning level plus one shared communication level architecture. To demonstrate the validity of the approach, the specific domain of fatigue and fracture in steel bridges is addressed in CRACK (Consultant Reasoning About Cracking Knowledge). CRACK performance was verified for solving failure analysis, predictive modeling, and design critique tasks for welded plate girder and rolled beam bridges. The results of the three plus one system architecture was satisfactory but this experiment was not an unalloyed success. In particular, the qualitative level was found to be over-engineered and under-utilized. Limitations of the extablished AI techniques for qualitative reasoning must be overcome for these methods to be useful in solving realistic engineering problems. Recent work in the field of common sense reasoning has promise for overcoming some of these technologic short-comings.