IMPACT: the interactive Maryland platform for agents collaborating together

We describe a platform called IMPACT to support multiagent interactions. The platform provides a set of servers (yellow pages, thesaurus, registration, type, and interface) that facilitate agent interoperability in an application independent manner. In IMPACT agents have an associated set of service descriptions, specifying the services they provide. We develop an HTML-like language for such service descriptions. When an agent wishes to identify another agent that provides a service, the requested service must be matched, using a metric approach, against existing service descriptions. We provide a formal framework within which this may be done, and develop algorithms to compute the k nearest matches, as well as all matches within a given distance from the requested service. We report on experiments evaluating our algorithms with large data sets.