What Is a Conversation Policy?

In this paper we define the concept of conversation policies: declarative specifications that govern communications between software agents using an agent communication language. We discuss the role that conversation policies play in agent communication, and suggest several subtypes of conversation policy. Our reasoning suggests, contrary to current transition net approaches to specifying conversation policies that conversation policies are best modeled as sets of fine-grained constraints on ACL usage. These constraints then define the computational process models that are implemented in agents. The Roots of Conversation Policies The dream of agent interoperability is commonly thought to rest on three main characteristics shared by the interoperating agents: 1. They would be able to access a set of shared infrastructure services for registration, reliable message delivery, agent naming, and so forth (i.e., there must be structural interoperability); 2. They would share (possibly through translation) a common content ontology, truth theory, and method of binding objects to variables (i.e., there must be logical interoperability); and 3. They would agree on the syntax and semantics of a common agent communication language (ACL) in which to express themselves (i.e., there must be language interoperability). In the last few years, international standards bodies (e.g., OMG, FIPA) and government-sponsored research efforts (ESPRIT, DARPA CoABS) have attempted to address these three aspects of agent interoperability. A surprising thing that all of this work has shown is the incompleteness of this list of interoperability characteristics: it is not difficult to construct a group of agents which satisfies all of them, and yet which cannot usefully interoperate. One common problem occurs because the above characterization of language interoperability is not broad enough. Specifically, for logically powerful and expressive ACLs like KQML [6] and the FIPA ACL [14], language interoperability requires more than simply that the agents agree on the format and meaning of the various primitive ACL messages. As a practical matter, agents must also agree on the range of possible sequences and contents of messages when they are interpreted in the context of larger goaldirected interagent dialogues, or conversations. Why aren’t existing ACL specification techniques sufficient for this task? Current methods for defining ACLs are built around complex and technically arcane methods for making precise the syntax and semantics of each message type. In the most advanced of these, the semantics are also formally compositional, in that there are well-defined ways to derive the meaning of a sequence of two or more messages from the meanings of the constituents [5]. However, compositional semantic theories for ACLs often do not uniquely specify the actual content and sequencing of agent messages needed to achieve to a given communicative goal. This gives rise to a significant ambiguity problem for agents that need to interact using a powerful ACL, which we will call the Basic Problem: Modern ACLs, especially those based on logic, are frequently powerful enough to encompass several different semantically coherent ways to achieve the same communicative goal, and inversely, also powerful enough to achieve several different communicative goals with the same ACL message. Put another way, the Basic Problem states that for powerful ACLs, there is a many-to-many mapping between the externally visible messages an agent produces and the possible internal states of the agent that would result in the production of the message. This would be a significant but manageable problem, except that agent interaction does not consist of agents lobbing isolated and context-free directives to one another in the dark. Rather, the fact that problems of high communicational complexity may be delegated to agents dictates that those agents must participate in extended interactions. Because agents are autonomous, they will need to independently optimize their own utility functions in these interactions, and hence they must use their beliefs about the probable goals of the other participants in order to determine the best next message to generate. And, because knowledge of the semantics of ACL