Building machines that learn from experience is an important research goal of artificial intelligence. Recently, considerable research attention has been devoted to the theoretical study of machine learning. This project focuses on several new directions designed to enhance the current learning models to more accurately model real-life learning situations. Most work in the area of concept learning assumes there is a well-defined border that divides all objects into those that are instances of the concept and those that are not. In reality, though, categorization of objects is often not so clear cut: an algorithm designed to read handwritten cheques will likely encounter many handwritten characters that look somewhat like a `4,` and somewhat like a `9.` In these situations, one possible goal for the learner is to determine which objects are unclassifiable (i.e., the classification is not clear cut) as well as determining the classifications of objects which are classifiable. Another possibility is to require only that the learner determine a categorization so that no object is incorrectly categorized. Thus, for this second goal, the learner can arbitrarily categorize those objects that can go either way. This project defines and studies learning models for both possibilities. Also this project continues the PI's work initiated on developing and studying formal models of teaching to understand how a teacher with knowledge of the target concept and the learner can reduce the training time needed by the learner. As well as being of theoretical interest, there are potential applications of the research to improving automated manufacturing environments.
