Cooperating Agents for Information Retrieval

capabilities for locating, combining, processing, and organizing information. With the vast number of information resources A promising approach to this problem is to proaviilable today, a critical problem is how to lovide access to the large number of information cate. retrieve and process information. It would sources by organizing them into a network of inbe impractical to build a single unified system f rmation agents [Papazoglou ef al., 1992]. The that combines all of these information resources. A more promising approach is to build specialgoal of each agent is to provide information and ized information retrieval agents that provide acexpertise on a specific topic by drawing on relecess to a subset of the information resources vant information from other information agents. and can send requests to other information reSimilar to the way current information sources trieval agents when appropriate. In this paper are independently constructed, these information we present the architecture of the individual inagents can be developed and maintained sepaformation retrieval agents and describe how this rately, drawing on the other available information architecture supports a network of cooperating agents and providing a new information source information agents. We describe how these inforagen s an prov id a Ew information mation agents represent their knowledge, commuthat others can then build upon. Each information nicate with other agents, dynamically construct agent is another information source, but provides information retrieval plans, and learn about other a desperately needed abstraction of the many inagents to improve efficiency. We have already formation sources available. Similarly, an informabuilt a small network of agents that have these tion source, such as a database or program, can be capabilities and provide access to information for turned into a simple information agent by buildtransportation planning. ing the appropriate interface code around the information source. Given this simple mapping beIntroduction tween information agent and information source, we will use these terms interchangeable throughWith the expanding amount of information availout the rest of this paper. able, the problem of how to combine distributed, Figure 1 shows an examT+l0. network of inforheterogeneous information sources becomes more mation retrieval agents. This network includes and more critical. The available information agents such as a USC cc ,nputer science technical sources include traditional databases, flat files, report agent, which is an agent that only provides knowledge bases, programs, etc. Traditional apinformation about and access to the department proaches to building distributed or federated systechnical reports. This agent is in turn used to tems do not scale well to the large, diverse, and construct both a computer science technical regrowing number of information sources. Recent port agent that spans multiple universities as well Internet systems such as Mosaic, WAIS, and Goas USC technical report agent that spans departpher allow users to search through large numberv ments within the university. These agents could in of information sources, but provide very limited turn he used to construct other agents, and so on. An important feature of this organization is that "The research reported here was supported in part the individual agents can be independently built by Rome Laboratory of the Air Force Systems Comand maintained. This makes it possible to scale mand and the Advanced Research Projects Agency the architecture to large numbers of information under contract no. F30602-91-C-0081, and in part by sources. the National Science Foundation under grant number Tources. IRI-9313993. The views and conclusions contained in To build a network of specialized informathis report are those of the authors and should not be tion agents, we need an architecture for a single interpreted as representing the official opinion or polagent that can be instantiated to provide multiple icy of RL, ARPA, NSF, the U.S. Government, or any agents. In previous work we developed an informaperson or agency connected with them. tion server, called SIMS [Arens et al., 1993], which