A Prolog system for case-based classification (abstract)

摘要

It is becoming apparent from the research of Kolodner, Schank, and others that case-based reasoning is an important tool for artificial intelligence. We have implemented a case-based system in Prolog that classifies objects described by attribute-value pairs. It uses Prolog's facilities to manage a case database and to encode features. The basis for classification is a "nearest neighbor" measure on a universal scale of similarity. The user specifies a scaling factor for naturally numeric data, thus allowing comparisons across dimensions. For non-numeric data, the user can specify numeric values on the standard scale for each value of the feature. For features that take on unordered values, the user can specify a standard difference to use in the case of inequality between values. Cases can have missing values for any feature, a circumstance that arises often in real applications. In a comparison between cases where one has a missing value for a feature, a standard difference for that feature isused. The user can control the importance of features in classification by manipulating scale factors and the numbers assigned to non-numeric values. If the system misclassifies an object, it saves its description with the correct classification. As the user interacts with the program, new features can be introduced. The program has been tested successfully with numeric audiology data and with non-numeric descriptors. This program should make case-based classification more accessible to potential users.