Classification Problem Solving: A Tutorial from an AI Perspective

Abstract

This paper is a tutorial discussion on classification problem solving, especially hierarchical classification. First we compare how pattern recognition and AI approaches to classification differ by pointing out how knowledge provides leverage against complexity, and thus point to the rationale behind knowledge-based systems. But we critique much of the work in knowledge-based systems, showing that important distinctions between various generic problem solving activities are often obscured by concentration on the implementation level of abstraction, such as “rules”, “logic”, or “frames”. We then argue that a generic task approach facilitates problem analysis, system design, knowledge acquisition and explanation of problem solving. We describe MDX, a medical diagnosis system that performs knowledge-based hierarchical classification, and motivate a number of issues in classification from that perspective. We also describe a high-level language called CSRL that is specially designed for hierarchical classification problem solving and show its power and utility.