Introduction

Traditional computational approaches have treated  language as something for the machine to crunch. It is only in the recent past that researchers have started looking at language as a design tool for modeling communication between agents. John McCarthy’s Elephant project and the KQML effort have drawn “inspiration” from Speech Act theory, and has used it to develop frameworks for agent communication.

Both these approaches have concentrated on a single aspect of language, namely its role as a system for communication. According to McCarthy, he uses Speech Act theory in Elephant because I/O streams can be considered similar to speech acts. But language is more than just an I/O stream or a system of communication. It is also an organising system for concepts, society and organisations. Hence cognitive models of  language can act as more than frameworks for agent communication; it can help model full-fledged multiagent systems.

The Design Problem

Most models for multi-agent systems treat the agents’ status – the bundle of beliefs, desires and intentions (BDI) – and communication as different modules. They are modeled separately, taking inspiration from a related field of social science or humanities.

One of the problems with such a segregated approach is that it becomes extremely difficult to model communication between agents, though individual agents get described in terms of BDI. The problem arises from the fact that the properties of agents are disparate and there is no uniform framework or property which envelopes both the status and communication aspects.

 A major reason for the lack of a uniform framework is that there is no common conceptual ground to think about all the abstractions related to agents. This proposal suggests language, specifically, one of its aspects, — force — as a conceptual basis for thinking about multiagent systems.

Force Dynamics

The concept of force is an old one in linguistics. In fact, Speech Act theory uses it extensively. Most models of agent systems also use the concept, largely for classifying different kinds of  performatives and the acts they perform. While Speech Act theory uses the idea of force as something like the “power” of an utterance, the concept of Force Dynamics (FD) goes beyond that and describes the role of force in cognition. Considered part of the emerging discipline of Cognitive Semantics, Force Dynamics, put forward by Leonard Talmy, captures the mechanisms used by language to represent how entities interact with respect to force. Talmy develops the model as a generalisation over the traditional notion of  “causative”. According to Talmy, it

analyses “causing into finer primitives and sets it naturally within a framework that also includes “letting”, “hindering”, “helping”, and still further notions.

Most significantly, Talmy projects FD as the semantic category that uniquely characterises modals as a whole. Modals act as the syntactic category for the expression of Force Dynamics.  This paper suggests that this property of modals can be extended to structure multiagent systems and coordinate communication between them.

Though Talmy defines an entire range of modals using the concept of force, for illustrative purposes let’s take three – CAN, MAY and MUST.

CAN: She can go to the playground.

A force dynamic definition of the above sentence would say that though the subject has a force towards doing the action, there is a force which is preventing her from doing it. But the CAN also states implicitly that she is capable of overcoming the force.

In contrast,

MAY: She may go to the playground

suggests that she doesn’t have this overcoming capability, but an authority has enabled her to do the action. She has been permitted to do something.

In the case of
MUST: She must go to the playground;

the force is on the subject to do the action. She is compelled to do it.

Now, if the three modals are considered to form a hierarchy, starting from CAN, going on to MAY and then to MUST, they can form a minimal framework for a multiagent system. CAN represents empowerment, MAY represents enablement and MUST represents commitment. In other words, the CAN-agent is the boss, while the MAY-agent is the manager and the MUST-agent the worker. Let’s see how such a hierarchy could work for the Contract Net, a protocol for multiagent systems.

The Contract Net & Force Dynamics

A typical instance of a Contract Net would have a system of several agents. One of them has a task to perform, which s/he cannot do entirely locally. S/he then splits the task into a number of sub-tasks, taking on the role of a manger, and sends out a call for bids to a subset of other agents, describing a subtask. Of the other agents, the ones who can, and are willing to perform the subtask, respond by sending a bid to the manager. The manager evaluates the bids received and selects one of them. S/he then sends a message assigning the subtask to that agent, who then becomes the contractor for that task. The contractor performs the task, possibly invoking other agents in the process. Finally s/he communicates the result of  performing the assigned task  to the manager. The manager collects the results of all the subtasks and computes its results.

In the regular frameworks, the messages passed between the agents are classified using speech act primitives. The message making the bid would be a commissive, the assigning one a directive, the one with the result of the task an assertive etc.

A drawback of  this categorisation is that the status of the agent does not come into the picture while modeling communication. This distinction makes the Contract Net unlike a human situation, where an organisation is defined both by the status of its members and the communication channels.

In a human situation similar to the Contract Net, having a repertoire of declaratives should get an executive’s job done. But in a human situation, just having declaratives won’t do. It has to be backed up with an institutional mechanism that ensures compliance. It’s this mechanism that gives force to the declaratives, and perhaps coordinates activity in the organisation. The institutional mechanism, which can be defined and formalised as a system of forces, also “create” the organisation. The organisation is the force hierarchy.

It’s hypothesised that a model based on force dynamics can effect the transfer of  something like the human situation to machinedom. For instance, if the manager agent gets an empowered status (CAN), and a force is associated with it, the subsequent levels of the organisation would form a hierarchy. Thus the contractor agent (MAY) would get a lesser force value, and the worker agent (MUST) the least one. This means that once the task is distributed, the organisation is also formed automatically, with a systematic flow of control. The messages need not be categorised. They will automatically form part of the agents’ status. The MUST-agent can send only commissives, the CAN-agent only directives, while the MAY-agent can send and receive both.

This in itself does not create a system more distributed than the Contract Net. But such a hierarchy, with associated force values, can be extended. An organisation can be formed using the primitives and the force values. (Which is not possible if one goes by bare performatives.) This arises from the fact that once the element of force is brought in, there can be degrees of force, which would enable the structuring of different levels of organisation, thus facilitating the setting up of well-defined communication channels. Who talks to whom, and who has authority over what, would be a matter of settling things by force, literally.

Another advantage of such a force-based system is that it can be linked up to learning software,  since the concept of force-values can be made compatible with the idea of “weights” in neural networks. Also, further enlargement of the model is possible if other FD variables are brought in.

The segmentation using CAN, MAY and MUST need not be absolute. It can vary from task to task. There could also be a process of promotion to, and demotion from, a status (Affirmative action!). This could depend on addition and deletion of information and the resulting knowledge level. Such a structure would lessen complications arising out of belief revision. The agent is limited by his/her force and the worlds accessible by it.

A very important outcome of a force dynamic categorisation is that a division using natural language modals would have more intuitive appeal for programmers. The language primitives would be more user-friendly than the performatives and their associated definitions. In a sense, language would be entering the machine through the back door.

The status/message demarcation in current systems is a natural extension of the idea that language is an independent entity, with which we communicate. The idea that language can simultaneously act as an organising system has not been accommodated in agent theory. This is a design flaw and can complicate things while organising agents into larger systems. The nature and destination of messages would get more and more complex and the associated scheduling problems would become hard to tackle.

Also, a bare speech-act model for agent organisation would be limited by its one-way flow of command and the centralised structure that results. The agents wouldn’t be autonomous in any strict sense of the term. In short, a performative-based approach would hinder the extent of distribution possible.

Discussion

The above proposal is based on the assumption that language creates a cognitive and communicative status, which can be extended to model multiagent systems. This can be done through allocation of force to syntactic elements (not just utterances) and using them to allot status for agents. Since most social relations are maintained through language, and the hierarchies are maintained through force parameters manifested through linguistic elements, the model can be extended further, to form agent organisations.

The idea of force is suggested as some kind of  “conceptual glue” while thinking about multi-agent systems. For instance, forces can be associated with other properties of agents, like belief, desire, intention, knowledge etc. This could lessen complications arising out of belief and knowledge revision.

Again, the categories defining the hierarchy can be linguistic. This is important since multiagent systems, in the long run, are intended to interact with both humans and machines. The natural language primitives, being more transparent to humans, would help broaden human-machine interaction.

Among the hazy notions on further applications of FD is its advantage of being an extension over a causative. This could help in understanding the system from the agent’s perspective. Also, the limiting nature of a status definition could help in overcoming the logical omniscience problem.

References

A Language and Protocol to Support Intelligent Agent Interoperability
Tim Finin,  Rich Fritzson, Don Mckay

Force Dynamics in Language and Cognition
Leonard Talmy

Foundations of Artificial Intelligence
Special Edition, Artificial Intelligence, Volume 47, January 1991
Elsevier

Intelligent Agents. Lecture notes in Artificial Intelligence 890
Proceedings of ECAI-94 workshop on agent theories, architectures and languages.
Springer-Verlag

Multiagent systems: A Theoretical Framework for Intentions, Know-how and Communications Lecture Notes in Artificial Intelligence 799
Munindar P Singh
Springer-Verlag

Pragmatics: an Introduction
Jacob L Mey
Oxford: Blackwell 1993