Example Agent Policies

Previous: References Up: DRAFT

Specification of the KQML

Agent-Communication Language

plus example agent policies and architectures

The DARPA Knowledge Sharing Initiative

External Interfaces Working Group Next: Example Agent Architectures and Implementations

Example Agent Policies

Agent-based software needs more than just a language for agents to describe their belief and wants. Agents need motivation for performing these communicative acts in terms of expectations about a helpful response. The shared expectations about message-passing behavior, e.g., helpfulness, responsiveness, commitment, etc., comprise the agents' protocols.

There is no single collection of protocols necessary for agenthood. The protocols of a particular system should be optimized for the constituent programs and the task at hand. In this specification, we merely list several protocols that may be useful in many applications. Other protocols, say for skepticism, bidding, reimbursement, and security, should be defined in this manner.

honesty
a message's KQML semantics apply to the sender.

gullibility
agents adopt the beliefs of others that are consistent with their own.
helpfulness
agents adopt the goals of others that are consistent with their own

responsiveness
agents will eventually respond to every received performative for which a response is expected

NOTE: this protocol folds in two important constraints: that an agent will eventually process every performative, and that it will generate some sort of response whenever responses are expected. The purpose of the latter constraint is to force a response like ``sorry'' to performatives that just happen to not produce any other responses. Of course, the meaning of this is totally wrapped-up in the word ``expected''; the intent is that response(s) are expected from a performative like ``ask'', but not ``tell''. ``advertise'' is trickier, but even though responses are possible, or even commonplace, they are not ``expected''.
empathy
agents have a built-in way of determining what performatives are needed by others (i.e., without needing an explicit performative to which to respond)

pertinence
agents will not send performatives that they believe will not benefit others

identity
agents will never register a networking name that is identical with the name of another agent on the same network

finin@cmsc