The Web of Communication and Cooperation in Multi-Agent Systems,
Talk, Universtiy of Edinburgh, Scotland, March 2002.
Eric Werner, Ph.D.
President, Cellnomica, Inc.
Abstract
When describing, designing and constructing any multi-agent system of agents there are fundamental problems that one encounters. What is an agent? What does the architecture of agent look like? What makes an agent social? What is cooperation? How can agents communicate and coordinate their activities? What are agent intentions? How does one define agent abilities? How is cooperative group activity generated? What are the limits of reactive agents versus more complex agents? In this introductory talk, a brief historical introduction to intelligent agents will lead naturally into a discussion of these basic questions about agents and agent societies.
Background:
In the past we have investigated the foundations of inter-agent communication in multi-agent systems [Werner 89]. The intentional states of agents were defined and we showed how pragmatic operators, initiated by linguistic or gestural communicative acts, can transform the informational and intentional states of agents. This gave us the foundations for a theory of multi-agent cooperation. Such a theory is necessary for implementing agents that can communicate and cooperate in a social way. As a part of any theory of cooperation, the logic of plans and ability in multi-agent contexts must be fully understood and given a semantics. We described this semantics in [Werner 90, 91]. A formal understanding of the semantics of 'Agent A plans to do X' and 'Agent A can do X' in a multi-agent context is necessary to understand multi-agent intention formation and their interactions in multi-agent social action.
A full account of cooperation will involve a theory of intentional states, minimally, plan states. A theory of communication will have to give an account of how the intentional states of the agent is transformed by the communication. However, this theory of operators on intentions must be consistent with a more fundamental theory of rationality based not just on deduction of truths from given truths, but instead must go beyond traditional theories of logic and deal with agent interests and values. Such a theory of social rationality must provide a way of determining how the intentions of an agent change given a change in the agent's interests. The logic of intention is a step in this direction. However, we still need a more detailed unified theory of an agent's social rational action. Traditional economic theory is perhaps too weak since that theory does not consider the fine structure of an agent's intentions. Thus a key area of future research will be to integrate the theory of communication (operators on representational states) and agent social rationality.
Algorithms that generate cooperative and noncooperative behavior already integrate the agent's intentions in a multi-agent intentional context with evaluation and ranking [Werner & Reineld 90]. Still we need to go a step further in describing the communicative process in more detail.
Finally, social structure and organization must be integrated with local agent social rationality. This can be accomplished in part by further developing theories of power based on interests. Such a theory of power will provide the basis for a theory of how agent organizations form, continue in a steady state, and, ultimately, transform.
All this is based on furthering our understanding of the general structure of agents, their interactions with other agents, and agent social structure. We envision a general theory of agents that includes simple agents, to more and more complex agents with invariant principles providing a unified theory of their nature.