Abstract. Learning by observation involves automatic creation of categories that summarize experience. In this paper we present UNIMEM, an artificial intelligence system that learns by observation. UNIMEM is a robust program that can be run on many domains with real-world problem characteristics such as uncertainty, incompleteness, and large numbers of examples. We give an overview of the program that illustrates several key elements, including the automatic creation of non-disjoint concept hierarchies that are evaluated over time. We then describe several experiments that we have carried out with UNIMEM, including tests on different domains (universities, Congressional voting records, and terrorist events) and an examination of the effect of varying UNIMEM's parameters on the resulting concept hierarchies. Finally we discuss future directions for our work with the program.
IntroductionLearning from observation is a task that is important in domains where examples are not pre-classified, but where one still wishes to detect general rules and intelligently organize examples. In this paper we discuss UNIMEM, a system that learns from observation by noticing regularities among examples and organizing them into a generalization hierarchy. We view UNIMEM both as implementing an algorithm for concept formation and as a prototype intelligent information system that can incorporate large amounts of data into memory and retrieve appropriate information in response to user queries. UNIMEM is not intended to be a psychological model per se, since it deals with a task more data-intensive than people are likely to perform. However, in developing the program we have made use of techniques derived by observing human behavior.The task of UNIMEM is to take a series of examples (or instances) that are expressed as collections of features and build up a generaliza-104 M. LEBOWITZ tion hierarchy of concepts. For example, UNIMEM might use information about a collection of universities to inductively determine the concepts of Ivy League universities, European technical universities, and so forth, and determine which examples are described by which concepts. The point of creating such concept descriptions is that they allow a performance element using the output of the program to make inferences about new examples based on partial information.Successful learning from real-world input must deal with several constraints. The key features that characterize the operation of UNIMEM are:• It learns by observation] it is not explicitly told how examples should be grouped into categories;• It is incremental] output must be available after processing each example; it cannot wait for all the input;• It must handle examples in large numbers (currently hundreds, eventually more);• Its generalizations are pragmatic] they need not perfectly describe all the instances they cover.1Although certain learning systems have dealt with tasks having some of these characteristics, little work has been concerned with all of them. However, all seem to characte...