COMBINATION AND ACQUISITION OF CONDITIONAL KNOWLEDGE (ARTIFICIAL INTELLIGENCE)
Abstract
This research tries to solve two problems in knowledge engineering. The first problem is that of choosing proper rules for combining evidence or other uncertainty combination. The second problem is to accelerate the process of knowledge acquisition by knowledge compression and learning. Both axiomatical and probabilistic approaches are used to construct right types of formulas for combining evidence. The axiomatical approach requires the formula satisfy a set of final properties and the probabilistic approach requires the formula satisfies a consistent set of assumptions in probability theory. Point-valued, interval-valued, and quantized belief representations are studied. Knowledge compression and learning are accomplished by the simplification of a network of conditionals. A new conceptual clustering method "defocussing" is proposed to compress and enlarge the knowledge. An experiment on Chinese medicine shows that this clustering method generates similar intermediate concepts as those generated by generations of experts. A tool for knowledge acquisition, the Relation Organizer, is built. It demonstrates the power of learning by simplification. It checks redundancy and inconsistency in the knowledge base. It raise questionnaires about possible relations between concepts, for example, synonyms, belonging relations, and other default rules. This tool can also generate frame and relational database structures from a network representation. This is a demonstration of a possible unification of common knowledge representation schemes.
Degree
Ph.D.
Subject Area
Computer science|Artificial intelligence
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.