Coordination Strategies for Geographically-Dispersed Multiple Agents

This white paper explores coordination strategies for Internet-based cooperating agents which must interact with a number of geographically dispersed users. There are many dimension to be investigated; I examine some of those dimensions here.

For example, there are many degrees of decentralization of the agent (one single central server; multiple mirrored servers; multiple loosely-consistent servers; a large number of servers with no consistency guarantees; an undetermined number of peers with no server at all and no global knowledge of all participants). What at the advantages and disadvantages of each approach, and how are those domain-specific or not? What sorts of organization strategies are appropriate in the various regimes? What can we saw about the capabilities that could be expected of each of the agent in these regimes (e.g., in a regime where no agent can know about all others, what do we have to give up & what advantages might we obtain)? Which organization is more trustable (security, completeness, privacy, performance, robustness, etc) for some possible applications?

I will place some of our own systems in this classification (e.g., Yenta, Couch Potato, Firefly, Webhound) and some others (Julia, Lycos, etc) and speculate on how their operation might change if they were distributed differently.

This white paper is intended as an evolving document. I encourage those with examples of other agent systems with distributed users to contact me with their evaluation where they fit in the classification systems I lay down here. I'll use this input to keep these pages updated as new systems become available.

Lenny Foner
Last modified: Fri Dec 15 06:54:47 1995