Generation of qualitative descriptions of complex objects

Humans develop useful insights into the structure and behavior of physical systems by discovery of relations of dependence between major subsystems or between their descriptive variable. Commonly, the nature of these relations is approximate and qualitative rather than precise and quantitative. By constrast, the representation methods employed in modern data repositories-largely based on underlying measurement systems-stress precision and accuracy of detail. Furthermore, their reliance on techniques that facilitate computational manipulation rather than knowledge discovery makes those repositories ill suited for the purpose of system analysis and understanding. We present details of a program of research for the generation of qualitative descriptions of complex objects, such as large biological molecules, multidimensional time series, or statistical surveys. The production of these representations is the initial step toward augmenting the utility of data repositories by facilitating retrieval and discovery processes based on perceptual features that are closer to the experience of those seeking to understand the underlying systems. Our approach to the qualitative description of complex objects is illustrated by presentation of a genetic-algorithm based methodology for the identification of qualitative features that match approximately a collection of prototypical models of interesting features. We present also heuristics for the summarization and interrelation of discovered features, providing also results of the application of our methods to financial time series.