Extraction and use of contextual attributes for theory completion: an integration of explanation-based and similarity-based learning

This research investigates the use of contextual cues to address problems in machine learning that arise from assumptions about the initial knowledge that is necessary for the acquisition of new information. Machine learning approaches may be placed along a spectrum describing purely inductive to purely deductive techniques. Inductive systems possess essentially no explicit knowledge that can be used in acquiring new facts, while deductive systems are assumed to contain a complete theory of the domain. Most work in machine learning has concentrated on approaches at the two ends of the spectrum. This dissertation describes an approach that integrates inductive and deductive methods. It provides a mechanism by which induction can be used in order to detect and acquire knowledge missing from the domain theory of a deductive system, while extracting information to guide the inductive search for missing knowledge. This information may be termed contextual. Much knowledge can be brought to bear in filling gaps in a deductive system's domain theory. This might include information about the domain theory itself or perhaps the history of the learning system. I show that a domain-independent set of attributes can be used to describe contextual information available in the deductions made by a system. Furthermore, I present an algorithm for using that information to generate domain-independent heuristics that guide induction. The heuristics use context information to select examples for induction of missing information and guide the way the examples are interpreted. My empirical investigations have yielded a characterization of the efficacy of various subsets of attributes. The algorithm for extracting and applying contextual knowledge is implemented in the Gemini system. Gemini combines similarity-based learning, an inductive technique, and explanation-based learning, a deductive method, in an architecture of mutual dependence. Similarity-based learning is used to fill gaps detected in the domain theory used by explanation-based learning. I have developed Gemini as a general learning system and have tested it in four domains. It has been the testbed for my investigation of contextual attributes and confirms that there exists a set of domain-independent attributes that can effectively guide an inductive search for concept definitions.