Summary and discussion Session 6: Applications in Information Retrieval Victor Lesser University of Massachusetts - Amherst No slides available.

Summary

Victor Lesser discussed applying mobile agents to cooperative information gathering. To automate information gathering, the agents should extract and fuse information from multiple sites. We are at a point where the Web can be used to do this. Text extraction technology has come a long way. Sites will cost money in the future and we'll have to pay, so we need to worry about the time and money cost. This suggests an important research direction: resource-bounded agents. One proposed solution is to break the information gathering task into subproblems, to assign each subproblem to a different agent, and to develop algorithms for piecing the solution together. For example, the Retsina architecture (by Decker and Sycara) is a layered system in which agents are introduced in a task-specific way: there are interface agents, task agents, and information agents. Such a system will spawn and send agents to process information. These agents have multiple goals and their activities define a complex interdepended search process.

In information gathering, a key issue is to identify the resource contention in such systems and to design algorithms for detecting resource contention and resolving it through load balancing algorithm. Another important issue is the level of autonomy associated with agents. In a one-level hierarchy of agents, the agents pass all the filtered data to a centralized location, so there is not much decision making autonomy in this process. In a multi-agent community, agents have to make autonomous decisions, move to other sites, spawn agents, and revise and generate new processing goals. Mobile agents are suited for such situations.

After arguing that information gathering can be structured naturally as a multi-agent activity Victor Lesser discussed the implications of this approach to mobile agents. Specific questions include: (1) how do we organize the task to make it most suitable for mobile agents?; (2) how much processing/abstraction/filtering can be done by a mobile agent based on local information at the site?; (3) how much of the mobile agent can be statically reside at the site (for example, as in the docking system of the Lockheed Martin DAIS system)?; (4) what is the cost for transmitting information across boundaries?; (5) is there a different cost for transmitting information to a user site than to transmitting information to an agent residing on the network; (6) do we need parallelism to meet the user performance criteria?; (7) how data fusion and correlation needs to be done by the agents?; (8) can we devise incremental algorithms for the mobile agents that exploit parallelism effectively?

Victor discussed a concrete example of information gathering for distributed situation assessment. He argued that the decision to transmit data or the agent itself is a complex issue. Information from other sites is often needed to resolve uncertainties. Decisions need to be made about: when and where to send solutions, at what level of detail should the solutions be sent, how to react to inconsistent information, and when to terminate processing.

Discussion

The discussion focused on the benefits of mobility for distributed information gathering. Before choosing the mobile agent paradigm to implement an application, the tradeoffs between distributing the application and the mobility overhead should be understood. Victor said that the first thing he learned in computer science was the space-time tradeoff and asked the audience if the network technology is at a place where operating in a local network is much faster than going outside the network. One argument for mobility was that in wireless networks it is advantageous to do work remotely, especially when the connection is lost. George supported this comment by observing that in some environments mobility is a powerful thing: otherwise we would all be trees. He speculated that in 100 years' time, when you buy a computer, it will immediately install and run the programs at your old computer. Victor noted that there will be control abstractions designed and specified at a different level than they are today, but that it is hard to see now where the ease of use of mobility will be. He continued by saying that agents will need to be aware of their environment and task and they will need to change and adapt to the state of this environment.