Incremental Concept Formation with Composite Objects

Abstract

This chapter discusses incremental concept formation involving composite objects in a system called LABYRINTH. Most research on incremental concept formation has focused on attribute-value representations. Fisher's (1987) COBWEB system integrated classification and learning, sorting each instance through its concept hierarchy and concurrently updating memory. It modified the concept's probability and the conditional probabilities for its attribute values; it then recurses to the children of that concept. LABYRINTH borrows from COBWEB the basic principle of probabilistic concepts organized in a disjoint hierarchy, but it extends the representation language for instances and concepts. A composite instance is as a set of components, which may themselves be composite objects. Like COBWEB, LABYRINTH represents concepts with an associated set of attributes and their values. However, the values associated with these attributes refer to other, more primitive nodes in the concept hierarchy, giving an interleaved memory structure. Thus, multilevel concepts are stored not as a single, monolithic structure, but as many concepts distributed through memory. Each composite concept is grounded in a composite concept whose attributes refer to primitive concepts, which are represented as was in COBWEB. LABYRINTH differs from most earlier approaches in that it is constantly revising the structure of these attributes. This results from the fact that the descriptions of composite concepts refer to other nodes in the concept hierarchy, which LABYRINTH itself defines. In effect, LABYRINTH is dynamically changing the representation used to describe composite concepts.