Preparing Semantic Agents for an Unsuspecting and Unreliable World.

摘要

Semantic web-enabled agents offer tremendous promise for enabling automated coordination and negotiation amongst diverse participants. Ideally, a user should be able to quickly instruct his personal agent to achieve some goal, e.g., to schedule a meeting so that at least one representative of each division of the company would attend. This agent would then negotiate with the agents of all the invited participants, receiving prompt and definitive replies, and quickly reach consensus on an acceptable time. The original user would be involved again only when the goal was achieved. In reality, many problems are likely to arise: (1) Goal Specification: The user may know precisely what he wishes to have done, but instructing a general-purpose software agent to achieve this goal can be very complex; (2) Agent Proliferation: In a typical organization today, very few participants will know what a software agent is, much less have one that could act on their behalf; (3) Participant Reliability: Even if the user's agent is able to directly interact with the participants, many of them are likely to not respond, due to confusion about how to do so or general busyness. Moreover, those that do properly respond may later need to change their response, a situation that the user's agent may not be prepared to handle; (4) Evolving Goals: Finally, even the originating user may be uncertain about his precise goals. For instance, he may decide to invite new participants later, or to modify the goals after seeing the initial responses. Such changes pose technical challenges for the agent and may render its earlier actions irrelevant or even counter-productive. Collectively, these issues pose serious problems to the wide-scale deployment of useful, flexible semantic agents. In previous work, we examined the first issue by considering how ordinary people could specify flexible, explainable semantic agents (McDowell, Etzioni, & Halevy 2004).