Previous: KQML String Syntax 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: Reserved Performative Parameters

KQML Semantics

The semantic model underlying KQML is a simple, uniform context for agents to view each others' capabilities. Each agent appears, on the outside, as if it manages a knowledge base (KB). That is, communication with the agent is with regard to this KB base, e.g., questions about what a KB contains, statements about what a KB contains, requests to add or delete statements from the KB, or requests to use knowledge in the KB to route messages to appropriate other agents.

The implementation of an agent is not necessarily structured as a knowledge base. The implementation may use a simpler database system, or a program using a special datastructure, as long as wrapper code translates that representation into a knowledge-based abstraction for the benefit of other agents. Thus we say that each agent manages a virtual knowledge base (VKB).

When defining performatives, it is useful to classify the statements in a VKB into two categories: beliefs and goals. An agent's beliefs encode information it has about itself and its external environment, including the VKBs of other agents. An agent's goals encode states of its external environment that the agent will act to achieve. Performative definitions make reference to either or both of an agent's goals and beliefs, e.g., that the agent wants another agent to send it a certain class of information. The English-prose performatives in this document make reference to these terms, but this view of the VKB is especially important in the formal semantics of KQML [SEMANTICS].

Agents talk about the contents of theirs and other's VKBs using KQML, but the encoding of statements in VKBs can use a variety of representation languages. That is, the KQML performative tell is used to specify that a particular string is contained in an agent's belief store, but the encoding of that string can be a representation language other than KQML.

The only restrictions on such a representation is that it be sentential, so that expressions using that representation can be viewed as entries in a VKB, and that sentences have an encoding as an ascii string, so that sentences can be embedded in KQML messages. Fortunately, these restrictions appear to hold for the representations of interest to KQML users, including AI languages, database languages, object-oriented representations, and many CAD formats.