| This dissertation addresses learning ontologies in a multiagent system where individual learning agents lack a commitment to a common, pre-defined ontology but share a distributed, collective memory of semantic objects. In this system each agent creates and learn conceptualizations, or ontologies, which can be useful for its individual purposes but will also share its knowledge in order to improve group problem solving performance. My thesis states that multiagent learning of ontologies among individual agents with diverse conceptualizations is feasible and these learned ontologies can be used by the agents to improve group search performance for related semantic concepts through experience in the problem domain. My approach addresses the current weaknesses to sharing knowledge among distributed agents by introducing a theory for learning ontologies based on combining agent communication and machine learning with two novel methodologies: (a) recursive semantic context rule learning and (b) concept cluster integration. I have implemented and used a proof of concept, DOGGIE (Distributed Ontology Gathering Group Integration Environment), to perform my experiments and demonstrate my theory. The results of my evaluation of this novel multiagent learning approach and algorithms contained in this dissertation demonstrate that my thesis is supported. |
