An expert system for X-ray rocking curve analysis using analogical reasoning
Richard William Henson · Warwick Research Archive Portal (University of Warwick) · 1993
This thesis examines the contribution of an adaptive expert system architecture to the field of X-ray Rocking Curve Analysis. The domain of X-ray Rocking Curve Analysis is used as an example to illustrate how a formal computer architecture can be enhanced through the principles of analogical reasoning to provide a deep knowledge of the domain. A conventional expert system core holds knowledge of a target problem in the form of frames, production rules and confidence factors. Through logical inference and demon logic, a reasoning cycle instantiates the frame structure and creates a new set of classes that represents the solution to the problem. The solved problem is a linked set of data held within the frame structure and complete knowledge across all domains held in a common on-line datastore. Knowledge elicitation reveals X-ray Rocking Curve Analysis to be a strongly visual task, which cannot be completely encoded within the expert system core. Through the application of the concept formation methodology, a set of key visual features (peak density, peak count, peak type) have been elicited from the X-ray Rocking Curve domain. The key features are used as a probability index for referencing previously solved problems. Structurally, the expert system core is embedded in a five staged analogical problem solving cycle consisting of: Targetting - building a description of the current problem; Source Selection - selecting a problem from a set of previously solved problems; Mapping - adding additional reasoning from the source to the target; Evaluation - a mathematical evaluation of the closeness of fit between the selected source and target; and Consolidation - the modification of the source based on the results of the evaluation. The key features provide the link between the target and source problems, providing a practical solution to isomorphic comparisons from inexact mapping. Statistical inference is used to enumerate between problems and allow the analogical inferencing to operate without exhaustive computation. A set of consolidative algorithms have been implemented for modifying the source data. It is these algorithms that give the expert system its adaptive characteristics. The analogical cycle provides a way of both guiding problem solving, and adding and adapting examples of previous cases. The expert system no longer behaves in the same manner each time it operates, but adapts its solutions by modifying the locatability of source information within a 3D probability array. These locations and the data held within is the deep knowledge of the domain that is achieved as the expert system is used to solve problems. Solutions are thereby not fixed, but evolve.