Architecture for Belief Revision in Multi-Agent Intelligent Systems

Abstract

The Semantic Web provides a framework for agents to easily publish and consume structured knowledge suitable for automatic reasoning. However, the open and distributed nature of the Web causes information of dubious quality to be published by sinister or incompetent agents. At the same time, Web accessible knowledge can change often. As a result, agents are forced to function in an environment of unreliable, incomplete, and contradictory information. They require mechanisms to change their beliefs dynamically and remove contradictions while maintaining the quality of the knowledge base. Computational agents might rely on social networks to judge information quality. Relationships of trust help determine the degree of belief an agent may ascribe to a given proposition. A previous work has proposed a formal logic for agent-oriented reasoning about belief and trust. This logic allows agents to deduce their confidence levels on both received and inferred knowledge. The present paper shows how to employ this logic in a collaborative document repository for scientific literature. Such a system relies on user-supplied classification data to construct comprehensive, personalized taxonomies for document browsing and search. It is based on a system of multiple agents representing users, constantly involved in knowledge acquisition and belief revision to provide the users with the best data available. The paper describes the theoretical foundations and briefly outlines a software implementation.