Date of this Version
Our system, based on a multiagent framework called collaborative understanding of distributed knowledge (CUDK), is designed with the overall goal of balancing agents’ conceptual learning and task accomplishment. The tradeoff between the two is that while conceptual learning allows an agent to improve its own concept base, it could be counter-productive: conceptual learning is time consuming and requires processing resources necessary for the agent to accomplish its tasks. In our current phase of research, we investigate the roles of resource and knowledge constraints, environmental factors (such as the frequency of queries), and learning mechanisms in a CUDK-based distributed information retrieval (DIR) application. In this application, an agent is motivated to learn about its neighbors’ concept base so it can collaborate to satisfy queries that it cannot satisfy alone. Similarly, to conserve resources, an agent is motivated not to learn from neighbors that have been unhelpful in the past. As a result, it is possible for an agent to learn from a helpful neighbor that is not the authoritative expert in the system. The agents use neighborhood profiling to learn about other agents’ helpfulness and conceptual inferencing to learn about other agents’ known concepts. The helpfulness measure defines a metric called collaboration utility, and the inferencing results are stored in a translation table in which each entry is a mapping between two concepts plus an associated credibility score. The experiments investigate how operational and conceptual factors impact the DIR application’s performance.