Dynamic ontology refinement

Human communication is highly fault tolerant: although words and phrases that are not jointly understood are often used, a shared common language can be used to explain these differences. In agent communication, however, mutual comprehension usually depends on a perfect matching of messages to internal ontologies. Thus any kind of ontological mismatch will lead to communication failure, even though large sections of the ontologies may be common to both parties. Ontologies, once envisaged as definitive descriptions of what exists in a domain, are commonly not static but are continually updated and altered, both centrally and by individual users. As the environments in which agents interact become increasingly diverse and distributed, with agents being designed by a large number of different users, ontology mismatch becomes increasingly common. Standard approaches to resolving this problem assume that the mismatched ontologies can be fully observed and often assume that it is desirable to match large sections, or even all, of the ontologies. However, this is not always a reasonable assumption, as many of these changes are not made public, and the computational cost of mapping entire ontologies is often prohibitive. We believe that it is more appropriate to assume that the ontologies of external agents are not available for observation, except for the specific parts of their ontologies revealed through normal agent communication. Consequently, a real-world solution, which we propose, is to patch specific instances of ontology mismatch when these particular mismatches lead to communication problems. This thesis describes the development of ORS (Ontology Refinement System), a system designed to dynamically refine ontologies whenever mismatches lead to communication problems during agent interaction. ORS contains a framework for agents to diagnose and refine ontological mismatch, integrated within an environment where iii planning agents can use this ability to achieve goals that would otherwise have been unreachable. These abilities are evaluated against genuine examples of ontological mismatch to demonstrate that they are useful and can be successfully performed.