Agent Communication Languages: Rethinking the Principles

The author proposes that no current agent communication language (ACL) truly lets heterogeneous agents communicate, largely due to proprietary languages. Despite two ACL standards being formalised, the author posits neither are particularly suited for general use, due to emphasis on mental agency and relying on the beliefs, desires and intentions model to communicate. The author argues that a new standard should be created, easing the tension between standardisation and use of dialects in specific agents.

The author discusses the evolution of ACLs, focusing on KQML and Arcol. KQML provides only assertive and directive primitives, and presupposes a knowledge base for each agent with each statement corresponding to a query to that database. Arcol provides fewer primitives than KQML, however it allows them to be composed. Arcol's primitives are also all assertive or directive. Arcol has the drawback that, by the standard, an agent can only inform another agent of something that it believes, and that it believes the other agent does not believe. As the author points out, this is a problem with agents that do not use a belief model.

The author proceeds to discuss features that a proposed new ACL should have; it should treat the sender and receiver of a message as equal partners as opposed to KQML and Arcol favoring the sender, it should be normative such that a breakdown in communication can be examined for where compliance was broken, and should cover all communicative acts, beyond assertive and directive statements and including declaratives and permissives, for example, as primitives. The author argues that current ACL approaches leave compliance to the decision of the designer, who may declare it complies without oversight.

The author proposes that an ACL written in this way would not have a problem of dialects; an agent who encounters a dialect it doesn't understand will be able to tell that this is the case in a correct implementation of the ACL. The author does not however present such a language, due to the challenges in finding a normative approach at the society level that preserves the intuitions behind, without being beholden, to the BDI model. He proposes three approaches; a behaviour based approach that may limit the ability to describe complx states, a mentalist approach, that reduces design autonomy, and a social approach focusing on social commitments and public perspective on mental states that is being investigated.