First-class agent interaction protocols

Tim Miller

Many current approaches to multi-agent interaction involve specifying protocols as sets of possible interactions, and hard-coding decision mechanisms into agent programs in order to decide which path an interaction will take. This leads to several problems, three of which are particularly notable: hard-coding the decisions about interaction within an agent strongly couples the agent and the protocols it uses, which means a change to a protocol involves a changes in any agent that uses such a protocol; agents can use only the protocols that are coded into them at design time; and protocols cannot be composed at runtime to bring about more complex interactions. These problems are out of place with the goals of agents being intelligent and adaptive.

To achieve the full potential of multi-agent systems, it is important that multi-agent interaction protocols exist at runtime in systems as entities that can be referenced, inspected, composed, and shared, rather than as abstractions that emerge from the behaviour of the participants. We use the term ``first-class protocol'' to refer to such protocols. We propose a framework, called RASA, which regards protocols as first-class. In this talk, I will present the initial parts of this framework, including a language for protocol specification that is built on constraint languages, and a logic for reasoning about protocol outcomes.

About the speaker: http://www.csc.liv.ac.uk/~tim/