Philip Hendrix, Barbara J. Grosz
In many settings agents need to identify competent partners to assist them in accomplishing tasks. Direct experience may not provide sufficient data to learn the competence of other agents. Reputation---a community-based assessment of agent competence---can augment direct experience, but is prone to error. This work provides a systematic study of how the utility of reputation varies by group size, group competency, competency distribution, and reputation error. It provides a systematic study of how the utility of reputation varies by group size, group competency, and reputation uses. Results demonstrate that the relative usefulness of direct experience and reputation depends on the particular multi-agent setting. They confirm that the utility received from reputation increases as group size increases. Interestingly, results include a ``pigeonholing phenomenon" in which reputation wrongly identifies some agents as having sub-par competence, based on early random sequences, even when agents are generally highly capable. This effect can be countered by introducing systematic positive bias to the system.
Subjects: 7.1 Multi-Agent Systems; 15.8 Simulation
Submitted: Apr 10, 2007