Information theoretic principles of agents

George Cybenko, Robert Gray, Alexy Khrabrov and Yunxin Wu. Information theoretic principles of agents. In Yannis Labrou and Tim Finin, editors, Proceedings of the CIKM Workshop on Intelligent Information Agents, Third International Conference on Information and Knowledge Management (CIKM 94), Gaithersburg, Maryland, December 1994.

* Contents

* Summary

This paper describes some formal models that capture many of the properties and activities of information agents. Our view is that agents are assigned tasks that serve to organize and reduce uncertainty in the information resources available to the user. This leads to a notion of entropy for information resources which agents seek to reduce in some way. The process of reducing uncertainty in an information corpus translates into elementary operations such as pattern matching, correlation and association. The model suggests algorithms appropriate for monitoring redundancy in information retrieval applications. Examples of implementations will be presented.

* Detail

Most current implementations of intelligent information agents are ad hoc and based on simple descriptions of tasks and actions. In many cases, an agent's operation is defined in terms of first matching patterns specified by a user and then performing appropriate actions when patterns are closely matched. Pattern association can in this way be described in terms of reducing uncertainty as measured by conditional entropies and mutual information. These quantities are intrinsic because they measure changes in uncertainty and have been successfully used for many years in other applications areas.

One of the consequences of this model is a precise notion of redundancy. That is, consider an information agent that seeks to collect information relevant to a specification provided by the user. The issue of redundancy plays a role when an ensemble of objects has been collected and new objects must be evaluated not only on the basis of their relevance to the original user specification but also to the corpus of objects successfully retrieved. The issue then becomes one of balancing relevance against novelty. Such considerations are extremely important when information resources charge for object retrieval or when a {\em representative} sample of items is sought. Examples will be given from applications such as medical records monitoring and inventory reuse.

We believe that these ideas can serve as a basis for building an analytic foundation for information agents, extending their domain of application and improving algorithms that are used in their implementation.

Back Back to the transportable agent page

The D'Agents Project ended in approximately 2003. The software, and this web site, is no longer maintained. For questions please contact David Kotz.