Reconciling Data-derived Knowledge with Expert Rules Using Clustering

Successful construction of intelligent systems to automate human reasoning processes often requires the use of a variety of techniques, including symbolic rule-based representations of reasoning and data-driven reenement of the knowledge. Connicts may arise between knowledge derived from the data and expert rules, as evidenced by misdiagnoses which are not correctable using neural network reene-ment training. Problems in the underlying knowledge base can be investigated using clustering to, for example, nd similar patterns of misdiagnoses. In this paper, we expose some of the problems in reconciling expert knowledge with data-derived knowledge, and present a solution based on neural network clustering techniques. The techniques are demonstrated in a elded system in gas chromatograph instrument fault diagnosis. OVERVIEW Constructing intelligent systems to automate human reasoning processes often involves symbolic representations of the reasoning process and data-driven re-nement of the knowledge. Combining techniques often leads to the discovery of diierences in human expert reasoning and information extracted from data about the decision process being studied. Some of the connicts which may arise in data-derived knowledge and expert rules result in consistent misdiagnoses or the inability to discriminate between two similar diagnoses based on the rules given. In each case, expert knowledge can be reened with neural network style learning using data (Lacher, 1993), but performance of the system may still be hampered by problems with the underlying structure of the knowledge base, problems which can be illuminated by use of neural network clustering. In the following sections, we expose some of the problems in reconciling expert knowledge with data-driven knowledge and present a solution based on clustering techniques. To demonstrate the developed clustering techniques, we use a elded system in gas chromatograph (GC) instrument fault diagnosis (Adair et al, 1996). A GC instrument produces chromatograms which, for example, can be used to determine the makeup of a processed soil sample. However, if the instrument