Towards the Standardization of Multi-Agent System Architectures: An Overview

Roberto A. Flores-Mendez
robertof (at) cpsc.ucalgary.ca
Computer Science Department,
University of Calgary, Canada
 
 
(In ACM Crossroads, Special Issue on Intelligent Agents, Association for Computer Machinery, Issue 5.4, pp. 18-24, Summer 1999

Introduction

Agents are everywhere. One can find intelligent agents, information agents, mobile agents, personal assistant agents, and so on. At this point, one could start asking if it is possible to make some sense of this apparent anarchy: What is it that makes an agent? Is there something that agents have in common? How can one organize them to carry out tasks?

Fortunately, we are not the first to ask these questions, since researchers have already defined some helpful notions. Firstly, authors have acknowledged that agents are entities within an environment, and that they can sense and act (not necessarily in that order). This means that agents are not isolated entities, and that they are able to communicate and collaborate with other entities. Simply put, agents that are not able to work together with other agents (humans included) are destined to become virtually useless.

Once agents are ready for collaboration, they will need to find other agents they need to collaborate with. Such a task is easy if they know exactly which agents to contact at which location. However, our everyday human experience has shown us that such a static setting is very unlikely to exist: people are usually on the move and they are not always readily available to interact with others. The same holds true for dynamic multi-agent systems: agents need support to find other agents.

Such are the motivations pursued by research groups working on the standardization of dynamic collaborative multi-agent systems. Some of these groups are the Foundation for Intelligent Physical Agents (FIPA), the Object Management Group (OMG), the Knowledge-able Agent-oriented System (KAoS), and the General Magic group.

This article briefly describes these groups' efforts toward the standardization of multi-agent systems architectures, and sketches early works to define a multi-agent systems architecture at the University of Calgary. However, the main objective of this article is to give the reader a basic overview of the background and terminology in this exciting area of research.

Background

Although agent researchers come from a variety of backgrounds, the Distributed Artificial Intelligence (DAI) and the Distributed Computing (DC) communities stand out as traditional agent research areas.

During the mid 70's, researchers from DAI began to formulate some of the basic theories, architectures and experiments that showed (computationally speaking) how interaction and division of labor could be effectively applied to problem solving [15]. Experiments showed that intelligent, rational behaviour is not an attribute of isolated components, but rather an outcome that emerges from the interaction of entities with simpler behaviours [4 , 10].

More recently, DC became an active discipline in agent research. DC is challenged to integrate heterogeneous, largely autonomous and often legacy computer components as part of collaborative environments. From this perspective, agents are applied as interaction entities  to mediate differences among components, while providing a syntactically uniform and semantically consistent intermediary role [21].

Agents Overview

The term "agent" is an elusive one to define. Agents are often described as entities with attributes considered useful in a particular domain. This is the case with intelligent agents, where agents are seen as entities that emulate mental processes or simulate rational behaviour; personal assistant agents, where agents are entities that assist users to perform a task; mobile agents, where entities are able to roam networking environments to fulfill their goals; information agents, where agents filter and coherently organize unrelated and scattered data; and autonomous agents, where agents are able to accomplish unsupervised actions, for example.

Attributes versus Attributions

Several researchers have attempted to provide a meaningful classification of the attributes that agents recognized as such might have. A list of common agent attributes is shown below [2]. According to their attributes, agents could be classified as showing weak or strong notions of agency [24]. The weak notion of agency, which comes from DC and DAI, sees agents as a paradigm of network based cooperative automation. The strong notion of agency, from Artificial Intelligence (AI), leads towards an anthropomorphic view where agents are seen as conscious, cognitive entities that have feelings, perceptions and emotions just like humans [26].

Nevertheless, it has been asserted that agents cannot be only characterized by their attributes, but also they need to be analyzed by their attributions as perceived by humans. A simple categorization of behaviour can be achieved by applying the following three predictive stances [9]:

It is interesting to notice how human perception of intelligent behaviour influences the definition of agency:
``[It] helps us understand why coming up with a once-and-for-all definition of agenthood is so difficult: one person's 'intelligent agent' is another person's 'smart object'; and today's 'smart object' is tomorrow's 'dumb program.' The key distinction is in our expectations and our point of view ''. [2]

What Do We Mean By Agent?

One aspect of agents that is broadly mentioned in the literature is the notion of agents as interactive entities that exist as part of an environment shared with other agents. This definition of agenthood is taken from descriptions given by several authors, who describe agents as conceptual entities that perceive and act [4 , 33] in a proactive or reactive manner [24] within an environment where other agents exist and interact with each other [34] based on shared communicational and representational knowledge [14].

Objects versus Agents

Agents and objects share some characteristics that sometimes make them hard to differentiate. For example, agent-oriented programming (AOP) could be considered a specialization of the object-oriented programming (OOP) paradigm [34]. On the one hand, OOP views systems as consisting of objects communicating with one another to perform internal computations, whereas AOP specializes this view to have agents (instead of objects), whose internal computations are based on beliefs, capabilities, choices and so forth, and which communicate with each other using messages adopted from speech-act theory.

Although this conception allows one to appreciate the similarities between agent and objects, their differences are less obvious. There are three main differences that have been identified between agents and objects:

``The first is in the degree to which agents and objects are autonomous. We thus do not think of agents as invoking methods upon one-another, but rather as requesting actions to be performed. In the object-oriented case, the decision lies with the object that invokes the method. In the agent case, the decision lies with the agent that receives the request.

``The second important distinction ... is with respect to the notion of flexible (reactive, pro-active, social) autonomous behaviour.

``The third important distinction ... is that agents are each considered to have their own thread of control ... in the standard object model, there is a single thread of control in the system.'' [23]

Multi-agent Systems Overview

Terminology

As part of their study of agent systems, researchers began to coin (purposely or otherwise) a terminology for agents. Although there is no complete agreement among researchers on the exact meaning of these terms they more or less resemble each other. Some of these terms are described as follows [21]:

Agent Architectures analyze agents as independent reactive/proactive entities. Agent architectures conceptualize agents as being made of perception, action and reasoning components. Perception feeds the reasoning component, which governs the agents' actions, including what to perceive next.

Agent System Architectures analyze agents as interacting service provider/consumer entities. System architectures facilitate agent operations and interactions under environmental constraints, and allows them to take advantage of available services and facilities.

Agent Frameworks are programming tools for constructing agents. Examples of these are Voyager [30], Aglets [22], and Odyssey [16].

Agent Infrastructures provide the regulations that agents follow to communicate and to understand each other, thereby enabling knowledge sharing. Agent Infrastructures deal with the following aspects:

What is a Multi-agent System?

Various definitions from different disciplines have been proposed for the term multi-agent system. As seen from DAI, a multi-agent system is a loosely coupled network of problem-solver entities that work together to find answers to problems that are beyond the individual capabilities or knowledge of each entity [11]. More recently, the term multi-agent system has been given a more general meaning, and it is now used for all types of systems composed of multiple autonomous components showing the following characteristics [23]: One of the current factors (and arguably one of the more important ones) fostering MAS development is the increasing popularity of the Internet, which provides the basis for an open environment where agents interact with each other to reach their individual or shared goals. To interact in such an environment, agents need to overcome two problems: first, they must be able to find each other (since agents might appear, disappear and move at any time), and once they have done that, they must be able to interact [23].

Finding Agents

There is a need for mechanisms for advertising, finding, fusing, using, presenting, managing and updating agent services and information. To address these issues, the notion of middle agents [7] was proposed. Middle agents are entities to which other agents advertise their capabilities, and which are neither requesters nor providers from the standpoint of the transaction under consideration. The advantage of middle agents is that they allow a MAS to operate robustly in the face of agent appearance, disappearance and mobility.

There are several types of agents that lie in the definition of middle agents. Note that these types of agents, which are described below, are defined so vaguely that sometimes it is difficult to mark a clear differentiation between them.

Agent Interaction

Interaction is one of the most important features of an agent [29]. In other words, it is in the nature of agents to recurrently interact to share information, knowledge and tasks to achieve their goals. Researchers investigating agent communication languages mention three key elements to achieve multi-agent interaction [14 , 21 , 32]: Subsections below explain concepts on agent communication languages and describe one of the most recognized agent languages: KQML. This subsection on agent languages is followed by a brief description of ontologies, which are described as structures of concepts and their relationships defining a context of reference.

Agent Communication Language

There are two main approaches to designing an agent communication language [17]. The first approach is procedural, where communication is based on executable content. This could be accomplished using programming languages such as Java [1] or Tcl [30], for example. The second approach is declarative, where communication is based on declarative statements, such as definitions, assumptions, and the like. Probably because of the limitations on procedural approaches (e.g., executable content is difficult to control, coordinate and merge), declarative languages have been preferred for the design of agent communication languages. Most declarative language implementations are based on illocutionary acts, such as requesting or commanding, actions that are commonly called performatives. One of the more popular declarative agent languages is KQML.
KQML
KQML, which is an acronym for Knowledge Query and Manipulation Language [18],  was conceived both as a message format and a message handling protocol to support run-time knowledge sharing among agents [14]. This language can be thought of as consisting of three layers: a communication layer (which describes low level communication parameters, such as sender, recipient and communication identifiers), a message layer (which contains a performative and indicates the protocol of interpretation); and a content layer (which contains information pertaining to the performative submitted). 
(register
      :sender       agentA
      :receiver     agentB
      :reply-with   message2
      :language     common_language
      :ontology     common_ontology
      :content      ``(ServiceProvision Manufacturing:TaskDecomposition)''
)
Figure 1. Example of a KQML message.
The format of a KQML message is shown in Figure 1. The message in the example starts with the word ``register'', which is the action (performative) intended for the message. The remainder of the message contains keywords needed for the message and communication layers. Keywords used in KQML messages are defined as follows [25]:

Ontologies

Ontologies are defined as specification schemes for describing concepts and their relationships in a domain of discourse [14]. It is important that agents not only have ontologies to conceptualize a domain, but also that they have ontologies with similar -- if not identical -- constructions. Such ontologies, when they exist, are called common ontologies. Once interacting agents have committed to a common ontology, it is expected that they will use this ontology to interpret communication interactions, thereby leading to mutual understanding and (ultimately) to predictable behaviours. Ontolingua [20] is often mentioned in the literature as a system that provides a vocabulary for the definition of reusable, portable and shareable ontologies. Ontolingua definitions are described using syntax and semantics similar to those of the Knowledge Interchange Format [19], also known as KIF, which is a format attempting to standardize knowledge representation schemes based on first-order logic.

MAS Architectures Review

The design of computer programs as multi-agent systems presents a useful software engineering paradigm where systems are described as populated by individual problem-solving agents pursuing high-level goals. Although having such an abstraction seems promising, its widespread adoption among system designers has not materialized yet. One reason is that MAS development is a technically difficult task. Efforts are challenged not only by known distributed programming issues, but also by the complexities associated with supporting agent collaboration.

If the agent-oriented paradigm is to succeed, systematic methodologies will be required for specifying and structuring applications as multi-agent systems. Once these methodologies are agreed upon, it would only be a matter of time until commercial MAS development toolkits emerge, and agent technology becomes accessible to a wide variety of software developers.

MAS Architectures Standardization

Although agent research started more than two decades ago, few efforts have been directed toward a consensual definition of an acceptable MAS architecture. It is possible that one of the reasons for such an absence of consensus might be the common misconception in research circles that MAS architectures and frameworks need to be designed from first principles to match project requirements [39]. Although this approach might prove time efficient for individual projects, it certainly creates incompatible systems that are difficult to reuse from project to project. Therefore, one might expect that the widespread adoption of MAS technology will only begin after the formalization and standardization of architectures, mechanisms and protocols supporting distributed interoperation of agents [35].

It was not until recent years that several independent industrial and research groups started to pursue the standardization of multi-agent technology. Prominent efforts, such as those of the Object Manager Group (OMG), the Foundation for Physical Agents (FIPA), the Knowledge-able Agent-oriented System (KAoS) group, and the General Magic group are briefly described bellow.

OMG's Model

The OMG group proposes a reference model as a guideline for the development of agent technologies [35]. This model outlines the characteristics of an agent environment composed of agents (i.e., components) and agencies (i.e., places) as entities that collaborate using general patterns and policies of interaction. Under this model, agents are characterized by their capabilities (e.g., inferencing, planning, and so on), type of interactions (e.g., synchronous, asynchronous), and mobility (e.g., static, movable with or without state). Agencies, on the other hand, support concurrent agent execution, security and agent mobility, among others.

FIPA's Model

The Foundation for Intelligent Physical Agents (FIPA) is a multi-disciplinary group pursuing the standardization of agent technology. This organization has made available a series of specifications to direct the development of multi-agent systems. Of particular importance are their Agent Management [12] and Agent Communication Language [13] specifications. FIPA's approach to MAS development is based on a ``minimal framework for the management of agents in an open environment.'' This framework is described using a reference model (which specifies the normative environment within which agents exist and operate), and an agent platform (which specifies an infrastructure for the deployment and interaction of agents).

KAoS' Model

Another important standardization effort is pursued by researchers of the Knowledge-able Agent-oriented System [3] architecture. This system, which is also known as KAoS, is described as ``an open distributed architecture for software agents.'' The KAoS architecture describes agent implementations (starting from the notion of a simple generic agent, to role-oriented agents such as mediators and matchmakers), and elaborates on the interactive dynamics of agent-to-agent messaging communication by using conversation policies.

General Magic's Model

General Magic is a commercial endeavor researching mobile agent technology for electronic commerce. Conceptually, this technology models a MAS as an electronic marketplace that lets providers and consumers of goods and services find one another and transact business. This marketplace is modeled as a network of computers supporting a collection of places that offer services to mobile agents. Mobile agents, which are pictured as entities that reside in one particular place at a time, are described as showing the following capabilities [36]:

Design of a MAS Architecture

The Computer Science and Mechanical Engineering departments at University of Calgary have recently assembled a research group to study multi-agent systems. This group, called the Collaborative Agents Group [5], has adopted notions from the models described in the preceding sections to define a dynamic collaborative multi-agent systems architecture. This architecture will be put to test the mapping higher-level multi-agent manufacturing architectures, such as MetaMorph [27], into a dynamic collaborative agent environment.

The envisioned architecture, which is still in an early stage of development, is briefly described as follows (publications forthcoming):
The architecture models open environments composed of logically distributed areas where agents exist. Basic agents in this architecture are minimal agents, local area coordinators, yellow page servers, and cooperation domain servers. These agents are described as:

Simply described, a MAS is an environment consisting of areas. Areas are required to have exactly one local area coordinator, which is an agent that acts as a facilitator for other agents within its area. Agents can be identified as being inside an area if they have registered with the area's local area coordinator. Agents will use the services of local area coordinators to access other agents in the system. Agents can advertise services and find out about other agents' services by means of yellow page servers. Agents requiring data sharing with other agents can join virtual environments called cooperation domains, which are supported by cooperation domain server agents.

Conclusions

The multi-agent systems paradigm promises to be a valuable software engineering abstraction for the development of computer systems. In addition, the wide adoption of the Internet as an open environment and the increasing popularity of computer-independent programming languages, such as Java, provide the grounds to believe that multi-agent technology is a feasible endeavor.

Consequently, it might pay off to invest some time reading and understanding key concepts on this area. If one sets aside the hype literature about agents, it can be noticed that most serious readings in this topic are basically compilations of research papers previously published in specialized conferences and workshops. The depth in which these articles cover their topics makes it difficult for the uninitiated reader to coherently integrate this information up to a higher level of abstraction. Therefore, the motivation for this article was to provide readers with a global perspective on the research literature on multi-agent systems.

In addition, this article briefly introduced the very basic notions in the design of a multi-agent systems architecture being developed at the University of Calgary, an effort that promises to become a fruitful contribution for the area of agent research.

Acknowledgments

I would like to thank the Collaborative Agents Group (CAG) at University of Calgary, especially my supervisor Rob Kremer, Niek Wijngaards and Douglas Norrie for their invaluable comments, suggestions and support helping me to understand better the intricacies of multi-agent research.

Partial support for this work was provided by a grant from Smart Technologies, Inc.

References

1
Arnold, K. and Gosling, J. The Java Programming Language. Addison-Wesley Publishing Co., second edition. 1998.
2
Bradshaw, J.M. An Introduction to Software Agents. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 3-46.
3
Bradshaw, J.M., Dutfield, S., Benoit, P. and Woolley, J.D. KAoS: Toward An Industrial-Strength Open Agent Architecture. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 375-418.
4
Brooks, R. A. Intelligence Without Reason. Massachusetts Institute of Technology, Artificial Intelligence Laboratory, A.I. Memo Number 1293, April, 1991.
5
CAG Collaborative Agents Group. 1998. Web site: http://sern.ucalgary.ca/cag/  (publications forthcoming)
6
Cohen, P.R., Cheyer, A., Wang, M., and Baeg, S.C. An open agent architecture. In: Proceedings of the AAAI Spring Symposium. 1994.
7
Decker, K., Sycara, K. and Williamson, M. Middle-Agents for the Internet. In: Proceedings of the International Joint Conferences on Artificial Intelligence (IJCAI-97), January, 1997.
8
Decker, K., Williamson, M. and Sycara, K. Matchmaking and Brokering. In: Proceedings of the Second International Conference on Multi-Agent Systems (ICMAS-96), December, 1996.
9
Dennett, D.C. The Intentional Stance. MIT Press, Cambridge, Mass. 1987.
10
Durfee, E.H. What Your Computer Really Needs to Know, You Learned in Kindergarten. In: Proceedings of the Tenth National Conference on Artificial Intelligence, July 1992, pages 858-864.
11
Durfee, E.H., Lesser, V.R. and Corkill, D.D. Trends in Cooperative Distributed Problem Solving. In: IEEE Transactions on Knowledge and Data Engineering, March 1989, KDE-1(1), pages 63-83.
12
FIPA FIPA 97 Specification, Part 1: Agent Management. Foundation for Intelligent Physical Agents, Version 1.2, October 10, 1997.
13
FIPA FIPA 97 Specification, Part 2: Agent Communication Language. Foundation for Intelligent Physical Agents, Version 1.2, October 10, 1997.
14
Finin, T., Labrou, Y. and Mayfield, J. KQML as an Agent Communication Language. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 291-316.
15
Gasser, L. Foreword. In: Readings in Agents, Huhns, M.N. and Singh, M.P. (Eds.), San Francisco, Calif., Morgan Kaufmann Publishers, 1998, pages v-vi.
16
General Magic Odyssey. 1998. Web site: http://www.genmagic.com/technology/odyssey.html
17
Genesereth, M. An Agent-based Framework for Interoperability. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 317-345.
18
Genesereth, M. and Fikes, R. Knowledge Interchange Format, Version 3.0 Reference Manual, Technical Report, Computer Science Department, Stanford University, USA., 1992.
19
Ginsberg, M. The Knowledge Interchange Format: The KIF of Death. In: AAAI Magazine, Fall 1991, Volume 12, Issue 3, pages 57-63.
20
Gruber, T.R. A Translation Approach to Portable Ontology Specifications. In: Proceedings of the Knowledge Acquisition for Knowledge-Based Systems (KAW'93), Gaines, B.R. and Musen, M. (Eds.), Banff, Canada, 1993, pages 199-220.
21
Huhns, M.N. and Singh, M.P. Agents and Multi-agent Systems: Themes, Approaches, and Challenges. In: Readings in Agents, Huhns, M.N. and Singh, M.P. (Eds.), San Francisco, Calif., Morgan Kaufmann Publishers, 1998, pages 1-23.
22
IBM Aglets. 1998. Web page: http://www.trl.ibm.co.jp/aglets/
23
Jennings, N.R., Sycara, K. and Wooldridge, M. A Roadmap of Agent Research and Development. In: Autonomous Agents and Multi-Agent Systems Journal, N.R. Jennings, K. Sycara and M. Georgeff (Eds.), Kluwer Academic Publishers, Boston, 1998, Volume 1, Issue 1, pages 738.
24
Jennings, N.R. and Wooldridge, M. Intelligent Agents: Theory and Practice. In: The Knowledge Engineering Review, 1995, Volume 10, Number 2, pages 115-152.
25
Labrou, Y. and Finin, T. A Proposal for a new KQML Specification. TR CS-97-03, Computer Science and Electrical Engineering Department, University of Maryland Baltimore County, Baltimore, February 1997.
26
Mamdani, A. The Social Impact of Software Agents. In: Proceedings of the Workshop on The Impact of Agents on Communications and Ethics: What do and don't we know?, Program presentation, Foundation for Intelligent Physical Agents (FIPA), Dublin, July 15, 1998.
27
Maturana, F., Shen, W. and Norrie, D.H. MetaMorph: An Adaptive Agent-Based Architecture for Intelligent Manufacturing. In: International Journal of Production Research, 1998. (in press)
28
Nii, H.P. Blackboard Systems. In: The Handbook of Artificial Intelligence, A. Barr, P.R. Cohen and E.A. Feingenbaum (Eds.), Addison-Wesley, New York, 1989, Volume IV, chapter XVI, pages 1-82.
29
Nwana, H.S. Software Agents: An Overview. In: The Knowledge Engineering Review, October/November 1996, Volume 11, Number 3, pages 205-244.
30
ObjectSpace Voyager 2.0, User Guide. 1998. Web page: http://www.objectspace.com/developers/voyager/
31
Ousterhout J.K. Tcl: An Embedded Command Language. In: Proceedings of the USENIX Conference, Winter 1990, pages 133-146.
32
Peng, Y., Finin, T., Labrou, Y., Chu, B., Long, J., Tolone, W.J. and Boughannam, A. A Multi-Agent System for Enterprise Integration. In: Proceedings of the Third International Conference and Exhibition on the Practical Application of Intelligent Agents and Multi-Agent Technology, H.S. Nwana and D.T. Ndumu (Eds.), London, UK, March, 1998, pages 155-169.
33
Russell S.J. and Norvik, P. Artificial Intelligence: A Modern Approach, Prentice Hall, Englewood Cliffs, N.J., 1995.
34
Shoham, Y. An Overview on Agent-oriented Programming. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 271-290.
35
Virdhagriswaran, S., Osisek, D. and O'Connor, P. Standardizing Agent Technology. In: ACM StandardView, 1995, Volume 3, Number 3, pages 96-101.
36
White, J.E. Mobile Agents. In: Software Agents, J.M. Bradshaw (Ed.), Menlo Park, Calif., AAAI Press, 1997, pages 437-472.
37
Wiederhold, G. Mediators in the Architecture of Future Information Systems. In: IEEE Computer, March 1992, pages 38-49.
38
Wiederhold, G. and Genesereth, M. Basis for Mediation. In: Proceedings of the Third International Conference on Cooperative Information Systems (COOPIS'95), Vienna, Austria, May 1995, pages 138-155.
39
Wooldridge, M.J. and Jennings, N.R. Pitfalls of Agent-Oriented Development. In: Proceedings of the Second International Conference on Autonomous Agents (Agents '98), K. P. Sycara and M. Wooldridge (Eds.), ACM Press, May 1998.

About the author...

Roberto is a student at the University of Calgary, Canada, where he is pursuing a Ph.D. in Computer Science. His interests include software agents, the design of multi-agent systems and issues in distributed computing. Some of his previous achievements include a B.Sc. in Computer Systems Engineering from the Instituto Tecnologico y de Estudios Superiores de Monterrey (Mexico) and an M.Sc. in Software Engineering from the University of Calgary (Canada). In 1997, Roberto was awarded as a first prize winner in the ACM/IBM Quest for Java contest for his jKSImapper concept mapping tool. He can be contacted by email at robertof (at) cpsc.ucalgary.ca , or by snail mail at: 
University of Calgary
Computer Science Department
2500 University Drive NW
Calgary, AB, Canada T2N 1N4
Last Modified on December, 1998