Thomas Wagner, John Phelps, Yuhui
Qian, Erik Albert, and Glen Bean
Software-agents and AI Lab (MaineSAIL)
University of Maine
wagner@umcs.maine.edu
Our previous work in sophisticated information gathering agents utilized a large set of research-grade AI technologies, including blackboard problem solving, information extraction, and Design-to-Criteria scheduling for real-time agent control. These efforts culminated in the BIG information gathering agent [AIJ2000, AAAI99, etc.] that could recommend software products to human clients without much knowledge a priori. BIG is an aggressive research artifact. However, BIG requires significant computational resources, produces uncertain results, and is not easily adapted to new problem domains.
Based on experiences learned from the BIG project, we have generated a new approach to sophisticated information gathering agent construction. The new architecture removes some AI components, replaces others, and reframes the problem space to support stronger domain problem solving. The end result is an architecture that will support crisis response information gathering, travel planning, and music-centered digital library information gathering. In this paper we discuss the new architecture, provide rationale for change and restructuring, and identify the new technologies. Examples are framed in the travel planning domain as this application is the closest to actual deployment. The advantage of using agent technology in these applications is that it enables the software to meet real-time deadlines, respond to the dynamics of the Internet environment, reason about client preference, and plan to achieve the objectives. An agent approach will also facilitate future work in distributing the computation to multiple agents.