Induction of Partial Orders Beats Classification: Improvement to the Cirrus Protocol Analysis System
Bernadette Kowalski, Kurt A. VanLehn · 1988
Abstract : In this paper, we discuss Cirrus, a semi-automated data analysis and model formation system. Cirrus starts with a half-order model for a data set, and induces the remaining knowledge necessary to turn the model into one which can reproduce subjects' Protocols. This knowledge takes the form of scheduling strategies for selecting tasks from an unordered agenda. The first implementation of the system, Cirrus-I, used a standard concept formation approach, with ID3 as the induction algorithm. Drawbacks of Cirrus-I led to the development of Cirrus-II, which is based on induction of a partial order. This approach appears to be applicable to many standard classification problems. Moreover, results from computational experiments on a pilot data set indicate that Cirrus-II, outperforms human analysts. Keywords: Machine learning; Artificial intelligence; Protocol analysis.