Designing Reusable Behaviors for Information Agents

Abstract

An important issue in multi-agent systems is the specification and effective implementation of various classes of computational agent behaviors. One such class of behaviors involves collecting and supplying information to other computational agents or humans. We call an agent that exhibits such behaviors an information agent. Information agents can play an important role in many larger mixed human- and computational- agent organizations. Much previous work has focussed on the language one might use to communicate with such an agent (e.g., KQML), but not on the set of behaviors the agent needs in order to constructively respond to such communications -- reusable behaviors that are independent of the particular problem-solving domain. This paper discusses a set of architectural building blocks that support the specification of reusable behaviors for information agents. We present an initial set of information agent behaviors, including responding to one-shot or repetitive queries, proactive monitoring of changing information sources for new occurrences of given patterns, notification of relevant agents, and automatic self cloning of an agent to achieve increased levels of service and efficient use of system resources. We have implemented these reusable information agent behaviors and tested them experimentally on the World Wide Web in several domains: tracking stock prices, extracting news stories, and monitoring airfares.