Induction of qualitative trees

Abstract

We consider the problem of automatic construction of qualitative models by inductive learning from quantitative examples. We present an algorithm QUIN (QUalitative INduction) that learns qualitative trees from a set of examples described with numerical attributes. At difference with decision trees, the leaves of qualitative trees contain qualitative functional constraints as used in qualitative reasoning. A qualitative tree defines a partition of the attribute space into the areas with common qualitative behaviour of the chosen class variable. We describe a basic algorithm for induction of qualitative trees, improve it to the heuristic QUIN algorithm, and give experimental evaluation of the algorithms on a set of artificial domains. QUIN has already been used to induce qualitative control strategies in dynamic domains such as controlling a crane or riding a bicycle (described elsewhere) and can be applied to other domains as a general tool for qualitative system identification.