Donald Bren School of Information and Computer Sciences

Concept learning and organization are much studied in artificial intelligence and cognitive psychology. Computational models of learning and memory that hope to be flexibly applied in real-world settings need to be incremental and improve an agent's ability to make predictions about the environment. While these are useful properties for purely artificial organisms, they also characterize much of human learning and memory.

This dissertation describes COBWEB, an incremental method of conceptual clustering that builds a classification hierarchy over a sequence of observations. These hierarchies are characterized in terms of their ability to improve prediction of unknown object properties. Computer experimentation and comparisons with alternate methods of classification show that COBWEB's approach effectively improves prediction ability. More generally, prediction of unknown object properties is forwarded as a performance task for all conceptual clustering systems. This opens the way for objective, not anecdotal, characterizations of and comparisons between concept formation systems.

A fundamental bias of this dissertation is that research on human learning and memory can usefully inspire directions for work on artificially intelligent systems and vice versa. Concept representations and measures of concept quality used by COBWEB are inspired by work in cognitive psychology on typicality and basic level effects. Conversely, COBWEB is the basis for a second system, COBWEB/2, that accounts for typicality and basic level effects in humans. Apparently, this is the first computational model that accounts for basic level effects. The account of typicality effects stresses the need to consider concepts in the context of a larger memory structure. This approach also facilitates speculation on possible interactions between basic level and typicality effects.

In summary, the dissertation presents an incremental method of conceptual clustering that is evaluated with respect to a prediction task. Concept representations and heuristics are borrowed from cognitive psychology, with repayment in the form of a cognitive model of basic level and typicality effects.