agent base system for manufacturing.pdf

7/27/2019 agent base system for manufacturing.pdf

1/24

Agent-Based Systems for Manufacturing

L. Monostori1,2 (1), J. Vncza1,2 (2), S.R.T. Kumara3 (1)

1Computer and Automation Research Institute, Hungarian Academy of Sciences, Budapest, Hungary2Department of Production Informatics, Management and Control,

Budapest University of Technology and Economics, Budapest, Hungary3The Pennsylvania State University, Industrial and Manufacturing Engineering,

University Park, PA 16802, USA

AbstractThe emerging paradigm of agent-based computation has revolutionized the building of intelligent and de-centralized systems. The new technologies met well the requirements in all domains of manufacturingwhere problems of uncertainty and temporal dynamics, information sharing and distributed operation, or co-ordination and cooperation of autonomous entities had to be tackled. In the paper software agents andmulti-agent systems are introduced and through a comprehensive survey, their potential manufacturing ap-plications are outlined. Special emphasis is laid on methodological issues and deployed industrial systems.

After discussing open issues and strategic research directions, we conclude that the evolution of agenttechnologies and manufacturing will probably proceed hand in hand. The former can receive real challengesfrom the latter, which, in turn, will have more and more benefits in applying agent technologies, presumablytogether with well-established or emerging approaches of other disciplines.

Keywords : Agent, Multi-agent systems, Manufacturing

1 INTRODUCTIONThe past decade has witnessed an explosive growth incomputer, communication, and information technologies.High-performance computing, the world-wide web, uni-

versal access and connectivity, virtual reality, and enter-prise integration are but a sampling of this revolutionsmany facets. At the same time, organizations and mar-kets have also changed dramatically, represented bydevelopments such as virtual organizations, businessprocess integration, customer-centric supply chains, andelectronic commerce. And, although industry strategistsand academics continue to debate the precise futuretrajectory of changes in technologies and organizations,they agree that information - its availability, and theability to exchange it seamlessly and process it quickly -lies at the core of organizations abilities of meeting esca-lating customer expectations in global markets.Agent-based computation is a new paradigm of infor-mation and communication technology that largelyshapes and, at the same time, provides supporting tech-nology to the above trends [1] [2] [3]. Agent theories andapplications have appeared in many scientific and engi-neering disciplines. Agents address autonomy and com-plexity: they are adaptive to changes and disruptions,exhibit intelligence and are distributed in nature. In thissetting computation is a kind of social activity. Agents canhelp in self-recovery, and react to real-time perturbations. Agents are vital in the globalization context, as globaliza-tion refers to an inherently distributed world both fromgeographical and information processing perspectives. Agents and similar concepts were welcome in manu-facturing because they helped to realize important prop-erties as autonomy, responsiveness, redundancy, distrib-utedness, and openness. Agents could be designed towork with uncertain and/or incomplete information andknowledge. Hence, many tasks related to manufacturing -from engineering design to supply chain management -could be conducted by agents, small and large, simple

and sophisticated, fine- and coarse-grained that wereenabled and empowered to communicate and cooperatewith each other.The goal of the keynote paper is to explore the software

agent technology and clearly show the applications andopportunities for agent technology in manufacturing.Though we survey the field from both manufacturing andcomputing perspectives, we pay special attention to thework done by the CIRP community. For other reviewsabout the application of agent technology in manufactur-ing, see, e.g., [4] [5] [6].The structure of the paper is as follows: We start with thenew emphasis in manufacturing, and how informationtechnology is changing the face of manufacturing. This isfollowed with the definitions of software agents and multi-agent systems in Section 3. Section 4 introduces basicagent technologies concerning individual agents, agentinteractions as well as organization models of multi-agentsystems. In Section 5 we undertake a fairly exhaustiveand rigorous survey of applications of agent technology invarious domains of manufacturing. Departing from thisdetailed survey, next (in Section 6) we sum up the meth-odological issues of agent applications, with a specialemphasis on running industrial applications in manufac-turing. Section 7 deals with open issues and strategicdirections. We conclude the paper in Section 8.

2 PARADIGM SHIFTS IN MANUFACTURINGIn discrete manufacturing, developments in informationtechnology led to the realization of computer integratedmanufacturing systems. With all of its merits, integrationresulted in rigid, hierarchical control architectures whosestructural complexity grew rapidly with the size and the

scope of the systems. Moreover, integration resulted alsoin complex decision problems [7] [8]. Disturbances inmanufacturing, as well as changing market demand pro-vided an unstable environment. By now it has becomeclear that increased volatility of market conditions disfa-

Annals of the CIRP Vol. 55/2/2006


2/24

vors these rigid, hierarchical architectures. No matter how, it is next to impossible to be prepared with pre-programmed, top-down responses to abrupt changes andto complete real-time computations on sophisticateddecision models before the results are invalidated.Long ago, manufacturing system theory suggested cau-tious and rather pessimistic organizing principles for sys-tems exposed to substantial internal and external uncer-tainties [9] [10]. These principles say that it is better torecognize ignorance than to presume knowledge; that it isbetter to regard the future as unpredictable than only tobe prepared for expected events. From this stance, therefollows a system structure with distributed responsibilities,tasks and resources. Accordingly, manufacturing systemsshould be organized as loosely coupled networks of communicating and cooperating components [11] or agents [12]. Redundant functions and responsibilities arenecessary for absorbing sudden changes, regardless of whether they stem from internal disturbances or fromvolatile market conditions. When doing so, there is anincreasing need for giving real-time responses [13].Decision rights should be co-located with the pertinent

information; time should be seriously considered as alimiting resource of decision-making, and the systemsshould have changeable, easy-to-reconfigure organiza-tional structures. Growing complexity is another featureshowing up in production processes and systems, fur-thermore, in enterprise structures as well [14].The concepts of process modelling [15], virtual manufac-turing [16], the digital enterprise [17], i.e., the mapping of the key processes of an enterprise to digital structures bymeans of information and communication technologies(ICT), give a unique way of managing the above prob-lems. By using recent advances in ICT, theoretically, allthe important production-related information is availableand manageable in a controlled, user-dependent way[18]. Product-related information can follow the productsthroughout their life-cycles [19] [20] [21].Earlier, it became evident that the management and ex-ploitation of this huge amount of information cannot beimagined without the effective application of the methodsand tools of artificial intelligence [22] [22] [24], or morespecifically, machine learning techniques [25]. In thecurrent age of globalization, virtual enterprises and pro-duction networks [26], some other requirements are morerelevant than ever to be considered: we have to solve thecoordination of various production entities that may havevery different objectives, goals, capabilities, and evencultures. New organizing principles and methods areneeded for supporting the interoperability of dynamicvirtual alliances of agile and networked enterprises which

- when acting together - can make use of opportunitieswithout suffering from diseconomies of scale.Various solution proposals unanimously imply that thefuture of manufacturing lies in the loose and temporalfederations of cooperative autonomous production enti-ties. However, nodes of a network organization can never be completely aware of each others goals and objectives,so they have almost necessarily adverse incentives.Hence, coordination, collaborative work and cooperationmay play an important role in managing these complexstructures [27]. Since the new information technologiesallow members of a network to widen their span of inter-est and control, the distribution of decision rights andactions introduces some new elements of uncertainty thatcan be resolved by communication (i.e., information shar-ing) and cooperation only. The interaction of individualsmay lead to the emergence of complex system-level be-havior [28]. Evolutionary system design relies on emer-

gence when modeling and analyzing complex manufac-turing and, in a wider context, eco-technical systems.Under the pressure of the above challenges, the trans-formations of manufacturing systems and organizationsare already underway [29]. The need for novel organiza-tional principles, structures and method has called forthvarious approaches [30] in the past decade, such asholonic [31] [32], fractal [33], random [34], biological [35],and multi-agent manufacturing systems [36], to name butthose investigated by CIRP colleagues. All the above approaches are similar in that they assumenetwork-like, dynamic, open and reconfigurable systemswhere decisions are made and production is carried outby more or less independent and cooperative partners. Inthe next two sections we present the information technol-ogy background of these recent developments.

3 AGENTS AND MULTI-AGENT SYSTEMSThe theory of computational agents goes back at least aquarter of a century when research in distributed artificialintelligence (DAI) had been initiated [37]. In the early 90sthe notion of agents appeared simultaneously also ininformation and communication technology (mobile, inter-face and information agents). Agents made the realbreakthrough a decade ago or so when the emphasis inthe mainstream AI research shifted: the focus on logicwas extended and attention changed from goal-seeking torational behavior; from ideal to resource-bound reasoning;from capturing expertise in narrow domains to re-usableand sharable knowledge repositories; from the single tomultiple cognitive entities acting in communities [38].These developments also coincided with the evolution of network-based computing technology, the internet, mobilecomputing, the ubiquity of computing as well as novel,human-oriented software engineering methodologies. All these achievements led to what is considered now the

agent paradigm of computing [1] [2] [39]. While thisnovel paradigm has several roots as far as theory, ena-bling technologies and applications are concerned, thereis a general consensus about its two main abstractions: An agent is a computational system that is situated

in a dynamic environment and is capable of exhibit-ing autonomous and intelligent behavior.

An agent may have an environment that includesother agents. The community of interacting agents,as a whole, operates as a multi-agent system .

In the coming paragraphs we give a concise characteriza-tion of agents and multi-agent systems.

3.1 Basic properties of agents

An agent operates in an environment from which it isclearly separated (Figure 1). Hence, an agent (1) makesobservations about its environment, (2) has its ownknowledge and beliefs about its environment, (3) haspreferences regarding the states of the environment, andfinally, (4) initiates and executes actions to change theenvironment. Agents operate typically in environments that are onlypartly known, observable and predictable. Autonomous agents have the opportunity and ability to make decisionsof their own. Rational agents act in the manner mostappropriate for the situation at hand and do the best theycan do for themselves. Hence, they maximize their ex-pected utility given their own local goals and knowledge.Rationality can be bound by the computational complexityof a decision problem, the limitation of computing re-sources, or by both. An agent with optimization objectivesbut with limited means is abounded rational agent.


3/24

actions

evaluation

agent

observations

utility

goals, objectivespreferences

prior knowledge

Figure 1: The agent and its environment [40].

The most important common properties of computationalagents are as follows: Agents act on behalf of their designer or the user

they represent in order to meet a particular purpose . Agents are autonomous in the sense that they con-

trol both their internal state and behavior in the envi-

ronment. Agents exhibit some kind of intelligence , from apply-

ing fixed rules to reasoning, planning and learningcapabilities.

Agents interact with their environment, and in acommunity, with other agents.

Agents are ideally adaptive , i.e., capable of tailoringtheir behavior to the changes of the environmentwithout the intervention of their designer.

Further agent properties, characteristic in particular do-mains and applications, are mobility (when an agent cantransport itself to another environment to access remoteresources or to meet other agents), genuineness (whenit does not falsify its identity),transparency , and credibil-ity or trustworthiness (when it does not communicatefalse information willfully).Even though they exhibit only some of the above proper-ties, agents relax several strong assumptions of classicalcomputational intelligence: they typically have incompleteand inconsistent knowledge as well as limited reasoningcapabilities and resources.

3.2 Multi-agent systems A multi-agent system (MAS) is formed by a network of computational agents that interact and typically communi-cate with each other.The agents may have only a partial (and, in a sense,distorted) model of their environment: they may possess alimited set of means for the acquisition and integration of new knowledge into their models and for pushing thesystem's state towards their own goals. The knowledge of two agents, referring to the same things, is not necessar-ily commensurate and may have different representa-tions. No closed-system assumption has to be main-tained: the MAS is submerged into and interacts with itsenvironment, which is not described completely (or isdifficult to describe) by formal means. Whenever novelkinds of interaction with the environment may occur, theMAS should be open and able to evolve.In a multi-agent system the decisions and actions of vari-ous agents do necessarily interact. However, just due tointeraction , a multi-agent system can occasionally solve

problems that are beyond the limits of the competence of the individual agents and/or may exhibit emergent be-havior that cannot be derived from the internal mecha-nisms of the components.

In a community an agent has to coordinate its actionswith those of the other agents; i.e., to take the effects of other agents' actions into account when deciding what todo. Coordination models provide both media (such aschannels, blackboards, pheromones, market, etc.) andrules for managing the interactions and dependencies of agents. Coordination requires some regulated flow of information between the agent and its surrounding envi-ronment in other words, communication . Note that in aMAS coordination is possible both by indirect communica-tion via the environment, or by direct information ex-change between specific agents. In any case, communi-cation needs some language(s) with syntax and seman-tics, at least partially known for each communicatingagent.Collaboration means carrying out concerted activities soas to achieve some shared goal(s). For instance, in ascheduling domain machine agents may agree on execut-ing each task of a job with the aim of completing an order by the given due date. The shared goal (completing anorder) can be achieved only if all agents commit them-selves to carrying out the actions they have agreed upon.

In general, meeting high-level objectives and satisfyingsystem-wide constraints need cooperation in a multi-agent system where agents are self-interested andautonomous.The overall operation of a multi-agent system is affectedby an organization that is imposed on the individualagents. Even though there may be no global control or centralized data and the computations are asynchronous,some organizational rules always exist. The organizationdetermines the sphere of the activity of agents, as wellas their potential interactions (see Figure 2).

environment

agent organizationalrelation

interaction influence sphere

Figure 2: Generic scheme of a multi-agent system [41].

The various organization patterns of multi-agent systems,such as teams, coalitions, markets, as well as hierarchicand heterarchic (including holonic) architectures provide

different ways to achieve system-wide design objectivesand/or to facilitate the emergence of desired properties inmulti-agent systems.

4 AGENT TECHNOLOGIES

4.1 Agent levelSumming up, agents are individual problem-solvers withsome capability of sensing and acting upon their envi-ronment, for deciding their own course of action, as wellas for communicating with other agents. Depending onthe actual problem and available technology at hand,agents can apply various faculties of problem solving ,including searching, reasoning, planning, and learning.The notion of agents has a strong synthesizing power;

hence the applied techniques may include both symbolicand sub-symbolic methods, classical and quantitativedecision theory, as well as knowledge-based reasoningand sophisticated belief-desire-intention (BDI) models [42][43].


4/24

There are several approaches for realizing rationalagents. Following the principles of classical decision the-ory, an agent makes choices from a set of alternativeactions so that it can maximize the expected utility of itsdecisions. If a utility function cannot be expressed, or notall the required input data are available, qualitative mod-els are applied that work with preferences. An agent may have explicit knowledge of how its actionscan change the states of its environment. Given somestates to achieve so-called goals the reasoning over future courses of actions is a key component of rationalbehavior. Artificial intelligence (AI) provided a host of planning methods to solve this problem under variousassumptions; many of them already have been proved tobe applicable in real-life domains [44].The most expressive model of an agent and its knowl-edge about the surrounding environment and itself is theso-called BDI model . This assumes the agent has bothcertain and uncertain knowledge beliefs (B) regardingthe states of its environment. States to achieve are ex-pressed in terms of goals, and states preferred in thelong-term are represented by desires (D). Decisions

concerning future events have motivations and pre-arrangements in the past: these are expressed by the so-called intentions (I) that represent the commitments of the agent made previously. Intentions can be thought of as past decisions behind a yet uncompleted plan. Stickingto intentions stabilizes the behavior of the agent. Agentswho are situated in a dynamic environment can benefitfrom having plans which are collections of actions on atime line. Plans, on the one hand, can constrain theamount of reasoning, and, on the other hand, can makecoordination possible [2] [45].The BDI model is used for describing an agents cogni-tive state and casts its decision problem in terms of cog-nitive concepts, such as knowledge and belief, desires

and goals, plans and intentions [43] [40]. A BDI agent iscontinually updating its beliefs based on perceptions,using its beliefs to reason about possible plans, commit-ting to certain intentions based on its beliefs and desires,and realizing these intentions by acting. The BDI formal-ism offers means to add and retract goals and generateactions.

4.2 Interaction Agents necessarily interact with each other either indi-rectly, via the environment, or by direct communication.The various coordination and cooperation mechanismsrange from emergent methods (without explicit communi-cation among agents) to coordination protocols, coordina-tion media and distributed planning [1].

Coordination protocols control the interactions of agents in order to reach common decisions. For instance,a widely used protocol is the contract net protocol (CNP) [46]. Here, a manager agent wants to have sometask performed by one or more other agents. CNP ismodeled on the contracting mechanism used by busi-nesses to govern the exchange of goods and services.The manager agent announces the task to other agentsand requests bids for the execution of the task. Thesebids are evaluated according to specified criteria (price,quantity, due date, distribution of tasks), and the winningbidder gets the contract. A coordination media providesa shared memory space for communicating data in anasynchronous way. Typical examples are blackboards [3]or the pheromones in stigmergy-based coordination [47][45]. Goal-oriented agents forming a community mayhave disparate and conflicting goals. For resolving conflictsituations, various negotiation mechanisms were devel-

oped, including auctions, one-to-one negotiation, bargain-ing, and argumentation-based negotiation [2].If coordination is based on communication, agents mustbe able to talk to each other through some agent commu-nication language. Collaborative acting and planninginvolve the intentions of multiple agents. Since the pres-ence of other agents is always a source of uncertainty(beyond other possible sources), collaboration requiresan integrated treatment of the beliefs and intentions of theagents who may take part in a collaborative act. That iswhy the BDI model provided the theoretical basis for agent communication languages (ACL), including thewidely used Foundation for Intelligent Physical Agents(FIPA) standard [48]. ACL was developed based on thespeech act theory [49]. Speech act theory views humannatural language as actions, such as requests, sugges-tions, commitments, and replies. It uses the term perfor-mative to identify the intended meaning of utterances.Examples of performative verbs are request , tell , insist and promise to name few. The first ACL was the Knowl-edge Query and Manipulation Language (KQML) thatincluded many performatives, assertives and directiveswhich agents use for telling facts, asking queries, sub-scribing to services and/or finding other agents.Communication and interoperability requires consensualknowledge of a community. A so-calledontology is anexplicit specification of the conceptual structures of agiven domain. It is usually expressed in a logic-basedlanguage that makes it possible to distinguish classes,instances, properties, relation and functions in a clear-cut,consistent way. Consensual means that the whole com-munity has a common understanding both on the contentand form of the expressed knowledge. Ontologies alsocan facilitate machine processing: automated reasoning,as well as the inter-operability of different agents. Notethat current agent representation and communicationtechniques are presented in Section 6.2.

4.3 OrganizationIn any human community, goals are achieved by thedivision of labor and coordination . There are a number of organization structures that define various patterns of decomposing work, assigning responsibilities to thosewho do the work, as well as collecting and combiningresults. Like any community, a MAS is formed by agentsthat are aimed at achieving some purposes, be themindividual, system-wide or both. It is no wonder that multi-agent systems adapt all the basic human organizationpatterns. An organization of a MAS assignsroles to the agents andspecifies their relations . The role identifies the agent and

specifies its function, its rights and obligations, and therules of its interaction with other agents and the environ-ment. There are two basic types of roles: the operator accomplishes a task a piece of work usually, bound bysome resource and time constraints. The manager , onthe other hand, exercises control: assigns tasks, monitorstask execution, collects and combines results. Dependingon the actual organization scheme, agents may performdifferent roles at the same time. The position of an agentin an organization is given by the role(s) it fulfills and itsrelations to the other agents. Finally, the admissible inter-actions of agents are defined by ordered sets of mes-sages, i.e., by protocols . A particular organization struc-ture comes together with rules concerning the conditionsof participating in a MAS, the assignment of the roles andrelations, as well as the use of protocols - all of whichtogether realize a particular coordination mechanism .Organizational structures and coordination mechanismsof most MAS systems are adapted from some form of


5/24

human organizations. Below, following [50] and [51], wepresent basic coordination mechanisms prevalent inmulti-agent systems. In Figure 3 and Figure 4, requestsfrom the environment (or other parts of the MAS) aretransmitted by a requesting agent (R), handled by a man-ager agent (M), if any, and processed by operator agents(O). Agents without prior role are depicted as (A). Thefigures show the organizational relations and the main

messages of the protocols.In direct supervision a manager has control over someoperator agents (see Figure 3). The manager takes re-sponsibility for the work of the operators and arrangestheir coordination. This pattern may be repeated at sev-eral levels of a hierarchy. Typically, one way to take careof global objectives and system-wide constraints is tochannel all/some interactions through a central coordina-tor (manager) agent. Due to its position, this agent ob-tains information essential to the operation of the com-plete system. Direct supervision reduces the number of inter-agent relations, but also makes the MAS very de-pendent on the central coordinator.

R

M

O O O

o u t p u t o b j e c t

i n p u t

t a s k

a c t i v i t y

o b j e c t

external

request

Figure 3: Direct supervision [51].

Standardization of work and/or skill delegates part of the control to the operators (see Figure 4): the manager instructs the operators with some procedures (i.e., taskspecification), but leaves the responsibility of execution intheir hands. The operators work in a decentralized way.Consequently, they control the flow of objects. Operatorsare selected for taking part in a task execution on thebasis of their competence and current load. A typicalexample of this coordination mechanism is when a man-ager agent arranges the work of a community of experts.

R

M

O O O

p r o c e d u r e

t a s k

obje ct

o u t p u t o b j e c t

i n p u

t o

b j e

c t

p r o c e

d u r e

exte rnalrequest

Figure 4: Standardization of work/skill [51].

Mutual adjustment is the coordination mechanism of aMAS without a priori organizational structure. Control andwork are not separated; manager and operator roles arenot distinguished any longer. The agent who gets a re-quest cannot but communicate (by passing objects and/or tasks) with the other agents, though cannot instruct them(see Figure 5). This is the simplest coordination method,however, with no (or very limited) guarantee that the re-quest will really be responded appropriately by the MAS,as a whole. One should not expect agents to behave in a

cooperative, altruistic way. The power and benefits of individual control can be destroyed by poor cooperation atthe community level. Trade-offs can be found bynegotia-tion methods. A widely used mutual adjustment coordina-tion technique is the formation and dissolution of teams .

R

A A

o b j e c t / t a s k

ob ject/task

o b j e c

t /

t a s k

e x t e r n a l r e q u e s t

A

Figure 5: Mutual agreement [51].

In a MAS, however, human organizational models can besurpassed. The ideas of decentralized problem solving,including the resolution of the conflict between individualand collective good, are widely studied in a number of other disciplines, such as economics, game theory, politi-cal science, biology, and ecology. Computational modelsof agency borrowing analogies from these fields are simi-lar in that they rely on some form of self-organization .No central control is exercised, and the system adapts itsstructure and functionality to the changing requirementsand environmental conditions. Typically, members of thesystems are individually able to achieve simple tasks, buttheir interactions lead to the emergence of complex col-lective behavior [28] [52]. In self-organizing systems,agents also communicate with each other indirectly: theyuse stigmergy (following ant communities) [45] [47] [53]

or attraction-repulsion fields [54] [55]. Though, currentlythere is no generic coordination method that could beapplied in engineering self-organizing systems [1].Organization patterns in manufacturing Numerous variants of the above organization models andcooperation patterns have appeared throughout the his-tory of manufacturing. In fact, manufacturing called for new, more robust, adaptable, fault-tolerant, decentralizedand open organizational structures even before the ideaof agents and MAS emerged from artificial intelligenceand computer science [10] [27]. Departing from the ge-neric ideas of Koestler [56], these requirements led to theholonic organization model developed and applied par-ticularly in manufacturing [45] [32] [57]. A holonic system

consists of so-called holons all of which are assumed tobe both autonomous and cooperative. The control isdistributed: based on their own situation assessment andlocal knowledge, individual holons decide over their ac-tions. On the other hand, holons are expected also tocooperate, i.e., to coordinate their actions in order to meetglobal goals and to satisfy system-wide constraints. Thesystem is open and no explicit control structure is im-posed on it, but thanks to cooperation, the members mayform also temporal federations. Thus, a holonic organiza-tion tries to combine the responsiveness and robustnessof decentralized, network-like organizations, and the sta-bility and efficiency of hierarchical control architectures. Aparallel trend was making some explored mechanisms of Nature operational in the organization and control of manufacturing systems. This approach resulted in the so-called biological manufacturing systems [35] [58] [59] andlater on, led to the engineerings concept of emergence[28].


6/24

5 APPLICATION DOMAINS IN MANUFACTURINGIn this section we present characteristic applications of agent technology in the main manufacturing domains.The survey below is far from being exhaustive; it focuseson pioneering works, as well as on developments wherean agents-based approach seems to be inevitable for meeting the requirements of modern manufacturing.

5.1 Engineering designIn the past decade, both the scale and the scope of engi-neering design had been changed and much enlarged.Design activities in various branches of engineering (me-chanical, electrical, control, etc.) are now being inte-grated . Furthermore, acknowledging that engineeringdesign must take into account the intrinsic requirementsand properties of the processes that bring to life, create,maintain and re-cycle artifacts, concurrent engineering includes all the main life-cycle activities such as market-ing, design, manufacturing, distribution, sales, operation,maintenance, disposal and re-cycling. Participants in theproduct life-cycle can interact in parallel.Collaborativeengineering (or design) transcends the above ap-

proaches: it emphasizes interaction instead of iteration,makes the conflicts between different stakeholders in thedesign process explicit and strives to achieve acceptabletrade-offs via negotiation [60] [61] [62] [63]. Negotiation isextended towards customers in a framework presented by[64] where design specifications of customized productsare developed in the course of an iterative give-and-takeprocess.The above developments posed new requirements for thecomputational support of design. Decomposition andparallel execution in collaborative design, naturally, lendthemselves to an agent-based approach. Beyond theusual advantages of having a modular system structure,additional merits are as follows: the agents need not beco-located, and they can form wrappers around and pro-vide interface for existing legacy systems (e.g., analyticaltools, simulators, CAD systems from various fields of engineering). Agents embodying knowledge and encap-sulating tools of different engineering domains can com-municate and work together if they have a set of sharedconcepts and terminology, a common language for ex-pressing this knowledge, and a communication and con-trol protocol for requesting information and services. Anearly infrastructure like this was demonstrated through thePACT system [65] which was built upon the early resultsof ontology design and knowledge interchange , andset a quasi-standard for a knowledge query and manipu-lation language. Since then, KQML, a typed messagelanguage with application-independent semantics andprotocol has become widely popular in other MAS appli-cations [66]. Ontologies also provided the basis for aknowledge-intensive design approach where relationshipsof different models were handled by a so-called meta-model mechanism [67] [68]. For instance, when design-ing mechatronic artifacts, different physical phenomenarelated to geometry and kinematics, force and distortion,electricity, friction, sound and heat have been representedby different, but re-usable and sharable aspect models.Interrelations of various kinds (causality, abstraction,approximation, aggregation, etc.) among the models weremaintained by the metamodel. Recent approaches sug-gest solutions for situations where a priori models do nothave sufficient representation power [69]. E.g., in [70], anagent system supporting the design of intelligent machin-

ing centres is proposed that concerns also the thermalbehaviour of machines. Notwithstanding the general as-sumptions and layered architecture, it was very difficult totransfer this custom-tailored system to another applica-tion.

All the agent-based approaches to design are similar inthat they represent partial, dynamically evolving designobjects (beyond the well-proven object-oriented represen-tation) means of constraints . Constraints are declarative,non-directional, additive, mutually dependent and canexpress partial information. Hence, they state clearly whathas to be satisfied without stating how. Constraints aresuitable for maintaining distributed, locally incompleterepresentations of design objects as well. That is thereason why constraints provide a key technology for multi-agent design. A couple of design specialists may work in a fixed organi-zation [65], or, alternatively, even thousands of agents,each responsible for a particular constraint, may delimitthe design space of the feasible solutions. In the widespectrum of possible organizations, knowledge-intensivemulti-agent design systems tend to apply the organizationpattern of fixed hierarchy or standardization of work. For some notable examples, see the descendants of PACT(SHADE [71], First-Link [72]), as well as the Agent-BasedDesign Concurrent Design Environment (ABCDE) [73],RAPPID [74], and Facilitator [75].

The RAPPID (Responsible Agents for Product ProcessIntegrated Design) system relies on three strategicmechanisms: autonomous agents as a way to distributedecision-making among a community of human beingsand computers, market-based control as a mechanism for coordinating distributed decision-making, and set-baseddesign as a means of making decisions in parallel, re-garding partial designs. [75] presented a MAS with afacilitator agent, a console agent and some serviceagents. The facilitator is responsible for the decomposi-tion and dispatching of tasks, and for resolving conflicts of poor designs. The console agent acts as an interfacebetween designers and the system. Each service agent isused for modeling different product development phases.One expects that multi-agent design cannot be easier than single-agent design. In principle, it may be so, how-ever, in a MAS design problems and the available knowl-edge can be structured in novel ways. Collaborative andlife-cycle approaches to design are successful in particu-lar because they are utterly based on interaction . Inter-action helps to harness external knowledge that could nothave been captured and internally represented. Further-more, collaborative engineering (CE) makes explicit thedisparate goals, objectives, priorities and concerns inshort, interests behind the various activities related to aproduct's life-cycle. These interests manifest themselvesas conflicts just in the early phase of design when deci-sions with far-reaching effects are made. The well-knownmerits of CE are due to the early detection and negotia-

tion-based resolution of such conflicts. Negotiation over conflicts can drive the design process towards innovativesolutions [62]. From a wider perspective, one can see CEas trading incomplete knowledge against interest. A tradelike this can be very fruitful: rational, interest-seekingbehavior on the part of autonomous fagents can result insuccessful overall performance in cases when the agentshave limited capabilities (knowledge and/or resources). Atthe same time, collaboration rests upon interaction, whichis still the key to creative design.

5.2 Process planningProcess planning is aimed at creating plans for discretemanufacturing operations that are executable in resource-constrained production environments and produce the

designed artifacts. Hence, computer-aided process plan-ning (CAPP) incorporates both design and productionrelated concepts: geometry, tolerances, surface quality,material properties, manufacturing processes, machines,tools and holding equipment (fixtures, grippers and ro-


7/24

bots). Useful domain knowledge varies also with the ac-tual technologies such as machining, sheet metal bend-ing, inspection, or (dis)assembly. There are two usualways of handing the complexity of CAPP problems: De-structuring its world into manageable micro-

worlds these are the so-called features. Decomposing the planning problem into sub-

problems such as process and resource selection,setup planning, sequencing, path planning, and NCprogramming.

The idea of a planning engine that could synthesize theresults of several domain-specific agents appeared in [65]and [76]. This process planner worked with a kind of hier-archical task network following the standardization-of-work coordination pattern. A similar solution strategy withsophisticated geometric reasoning appeared in [77].Human interaction is emphasized in [78] where specificagents, working on the exact geometrical representationof the part, available machines and tools, build up thespace of the process plans. Then, an optimization engineextracts solutions from this space. In fact, full-fledgedCAPP systems are concurrent engineering systems (seethe above section) with a special emphasis on processmodeling. However, making a consistent synthesis out of results generated on partial models proved to be prob-lematic. The crux of all but the simplest CAPP problemsis to reconcile the logical and optimization aspects of planning and to handle inconsistent pieces of design andtechnology related knowledge. The resolution of conflictslike this calls for negotiation no wonder that theseissues of CAPP and concurrent engineering prompted anovel socio-technical concept that considers engineering as collaborative negotiation (ECN) [61] [79] [62]. Alter-natively, in this ill-defined area one can rely on method-ologies of emergent synthesis [28] [80]. The sharing of tasks in multi-agent process planners followed some

traditional decomposition of the CAPP problem, with sub-problems calling for knowledge-intensive solutions.Hence, their characteristic organization patterns are fixedhierarchy and standardization of work.In recent works, CAPP has been extended towards theexecution of plans: planning is aware of available re-sources and takes actual lot-sizes or due dates into con-sideration. On the other hand, process planning knowl-edge is used for short-term scheduling decisions at theshop floor. Multi-agent system concepts are particularlywell suited to this integration [80] [81]. For further details,see Section 5.6.Finally, note that plans are blueprints of behavior com-posed of pre-defined actions. Hence, process planning

as a large body of planning problems in general can beconsidered a problem of configuration . Configurationaldesign and CAPP tend to apply similar representation andsolution techniques that enable representing, solving andrelaxingdistributed constraint systems .

5.3 Production planning and resource allocationProduction planning is the process of selecting and se-quencing activities so that they should achieve one or more goals of an enterprise and satisfy a set of domainconstraints [82]. At the strategic and long-term level, topmanagers try to allocate available resources based ontheir experience, intuition and computer support, if avail-able.Resource-aware aggregate planning is addressed in [80]where a library of generic resource classes is used for describing the enterprises resources. Resource-awareness means the creation of a dynamic inter-relationship between the planning entities and the enter-

prise resources, human beings and machines in a distrib-uted manufacturing enterprise. A market-based negotiation mechanism, called prece-dence cost ttonement (P-TTO), was reported on in[114]. The system is composed of a project manager agent, task agents, resource manager agents, resourceagents, and coordinator agents. The project manager agent maintains the project milestones, the project activitynetwork and each tasks resource allocation information. A task agent is in charge of its own single task. A re-source manager agent is in charge of monitoring andcoordinating a set of resources. A coordinator agent isresponsible for coordinating multiple resource allocationmarkets in the virtual market model. The simulation re-sults based on P-TTO indicated high levels of solutionquality and computational efficiency.Sequential allocation of resources can be viewed as adigraph where each vertex represents resources and thearcs represent the allocation [83]. In this method agentsare associated with actors, resources and groups of re-sources. The method has been applied to an automatic-guided vehicle (AGV) system. The resources identified for

this system are the segments and points that form thefree space world model utilized by the AGVs for navigat-ing. Agents are utilized to manage these resources (andgroups of resources) to maintain the flow of AGVs.The exploitation of agent-based technology in productionplanning was addressed in the ExPlanTech project [84].There, one of the 5 intra-enterprise agents introduced isthe production planning agent which is in charge of con-structing an exhaustive, partially ordered set of tasks tobe carried out in order to accomplish the given project.

Corporate agent

Plant agents

Order agent

Job agent

High levelnegotiation

Medium levelnegotiation

Low levelnegotiation

Shop-floor level

negotiation

Very High levelnegotiation

Group agents

Group agents

Group agents

Plant agents

Resource agents

Figure 6: Negotiation models for different planning levels

of the enterprise [86].Five levels of production planning in a reconfigurableenterprise (RE) are distinguished in [85] [86] (seeFigure 6). The authors claim that the successful tools for operation management in RE need to be based on thedecentralization of the decisions where each entity incharge of specific planning decisions makes its own deci-sion autonomously, while global planning decisions areachieved by means of coordination and negotiationamong them. The various levels differ from each other onthe planning horizons, the planning issues, producers andmechanisms. The negotiation model was compared witha centralized solution, showing the benefits of the agent-based approach [86].


8/24


9/24

ing blocks, such as order and machine agents in [101], job and resource agents in [102], and order and machine(resource) holons in [103]. A common feature of theseapproaches is that the functions of the order and productholons are integrated in one basic type.Based on thePROSA architecture, an agile holonic multi-cell controlsystem (HoMuCS) was described in [104]. The holonicbehavior of machine tools CNC was highlighted in [105].

Market-based approachesCooperation and conflict resolution are the main issues inagent-based scheduling systems. For solving these prob-lems, the idea of negotiated factory scheduling ap-peared some time ago. Early attempts implemented dis-patching mechanisms [107] [108] but did not concernadvance scheduling. Conversely, other works realizedpredictive schedulers with no reactive capabilities. For negotiation, versions of the contract net protocol [109]were used. Market-oriented programming based on gen-eral equilibrium theory has been first applied to resourceallocation problems with no temporal aspect [109]. Dy-namic reconfigurability, which asserts itself again in theholonic concept, appeared first in [108] that suggestediterated bidding for selecting cells that complete a job.This approach was developed further into random manu-facturing [34] where temporal coalitions of machinescompeted for incoming orders.Nowadays, negotiation based algorithms have been usedalmost without exception. In this way, schedule genera-tion is a recursive, iterative process with announce-bid-award cycles where market mechanisms [34] [102] [106]can be advantageously used. Order (or part) driven andresource (machine, cell) driven techniques may be distin-guished, based on who makes the announcements. Theextent of coordination can be limited to given (e.g. pairsof) agents. Tasks are, usually, announced one by one,mostly after the preceding operation has been completed

which causes decision myopia , a common drawback of distributed approaches. More advanced systems supportlook ahead scheduling with a longer, sometimes varyinghorizon.In multi-agent scheduling , agents manipulate resourceand/or order variables under their own authority. Potential job interactions are handled by agents telling each other their aggregate workloads and free capacities [110] [111].Since global consistency can hardly be guaranteed, con-straint checking or simulation is needed. Since constraintrelaxation and backtracking may require enormous com-putational overhead, their use is limited. As an alternativeto backtracking, heuristic tinkering of almost conflict-freeschedules can be applied.In [106] and [112] a market mechanism was proposedwhich made cooperation in the control of distributedmanufacturing systems possible. Particularly, an inte-grated production planning and distributed schedulingproblem was investigated. The system included self-interested, autonomous agents whose goal was to maxi-mize their own profit. While the production planning prob-lem was handled by a central management agent, ma-chine agents were responsible for building up and execut-ing local schedules at machines with different technologi-cal capabilities. Congestion control (via order selection)was the responsibility of the management alone, butscheduling decisions were reached by a partially commit-ting bidding mechanism between the management agentand the machine agents. The bidding protocol is pre-

sented in Figure 8. A market mechanism has been applied to solve dynamicmulti-project resource-constrained scheduling problems:[113] [114] handle the decentralized scheduling of re-sources, which are shared by multiple projects.

OrderAgent Resource

Agent

bidding(s)

bid evaluation

task announcement

accept, reject decisions

confirmationtask earlieststarting time

start

awarded starting time

Figure 8: Agents and their relations according to [106].

In a virtual economy where agents act as buyers andsellers of resources, resource-time slots (i.e., availabilityover a given time) are traded as goods. Due to the dy-namic and distributed nature of the economy, the ap-proach achieved higher levels of flexibility and scalability.In the AARIA (Autonomous Agents for Rock Island Arse-nal) system unit process, resource, manager, part, cus-tomer, and supplier are identified as agent classes [115]. A collaborative control system for mass customizationmanufacturing is described in [102]. A market mechanism has, however, some drawbacks:since it is a nondeterministic, utility-based method, thesystem's worst case behavior and even the avoidance of

extreme situations are hard to guarantee. No more thanthe average behavior can be predicted; the significance of the results should be based on careful simulation experi-ments. Moreover, it is doubtful whether small artificialmarkets with a low number of participants and limitednumber of encounters could produce the effects we meetin real markets of large size. It is hard to define where theborderlines of a model like this should lay. Note that in thereal world, production elements with monetary terms dohave counterparts in consumption. Money that can buynothing is good for nothing.

Stigmergy-based coordination and control A relative novel approach for coordination in multi-agentsystems is stigmergy which belongs to mechanisms

which mimic animal-animal interactions [45]. Stigmergy isan indirect coordination tool within an insect society whereparts of global information is made available locally bypheromones, e.g., as in the case of ant colonies. Thisway, individual agents are not exposed to the complexityand dynamics of the situation, and the communicationburden in the computer realization is significantly lower than in, e.g., market-based solutions. A holonic manufacturing execution system (MES) is pre-sented in [45] that preserves the advantages of heterar-chical designs and predicts the near future, while ac-counting for changes and disturbances. In order toachieve this, the software agents reflect the underlyingmanufacturing system (e.g., order agents reflect tasks)and delegate mobile (ant) agents consistently. For in-stance, exploring ant agents query, the associated re-source agent about processing times; they do not havetheir own model of the resource and, therefore, make noassumptions that can be faulty. The forget-and-refresh


10/24

mechanism of the stigmergy infrastructure ensures thatinformation remains up-to-date. As for the realization of stigmergy-based systems, virtualants can be realized by mobile software agents. A mes-sage-based realization can be conceived, as well. Market-and stigmergy-based approaches do not represent twototally different ways of multi-agent coordination and con-trol: they can be nicely combined in complex societies.

Adaptation and learning in multi-agent production control Early investigations implied that the efficiency and qualityof multi-agent scheduling depended on variable, adaptivecommitment strategies. This is especially so in dynamicdomains where commitment may hinder an agent fromresponding to new contingencies. The main issues arewhen, what and to what a degree to commit.In [116] an agent-based system is described which isused for resource allocation and production control. Agents assigned to all orders and resources are adap-tively conditioned according to the logistic situation. Simu-lation experiments proved that the approach deliverednearly the same logistic results as established ERP

methods did. However, the proposed agent-based systemreacted on disturbances and unexpected events better than other ERP systems.Learning and other forms of adaptation are essential inmulti-agent systems [25] [117] and can be categorized asfollows: Centralized learning (or isolated learning) refers to

learning approaches which are entirely executed bysingle agents, with no regard to interaction with other agents.

Decentralized learning (or interactive learning)involves several agents which require a joint and co-ordinated interaction among them.

The adaptation procedure described in [117] and [118] isa centralized approach in which each resource agentlocally adapts its behavior in order to achieve a moreprofitable position in the agent society. The feedbacks arerepresented by changes in local utilization parametersand bid awarding and/or rejection reactions issued by theorder agent. Each resource agent incorporates a rule-base by which it can locally decide on the cost factor tobe applied in a task announcement. The preconditions of these rules are the utilization of the resource and the ratiobetween the won and lost bids which are stored locally for each agent in the table of machine abilities and history.Simulation results demonstrate that the major advantageof the above solution is a more balanced usage of re-sources. Moreover, several performance measures, such

as maximum tardiness and makespan proved to be better with cost factor adaptation [117].In [119] a market-based distributed production controlsystem with learning and cooperative agents is described.The adaptive scheduling is done by a triple-level learningmechanism. The top level of learning consists of a simu-lated annealing algorithm, the middle (and the most im-portant) level contains a reinforcement learning (RL) [120]system, while the bottom level is done by numerical func-tion approximators, such as artificial neural networks[121]. The proposed neurodynamics-based system canbe used for solving the general dynamic job-shop sched-uling problem in a distributed, iterative and robust way.The time and space complexities of the solution are com-pared with those of classical approaches.Finally, note that adaptive agents represent one of themost promising approaches towards Class II and Class IIIproblems of emergent synthesis, i.e., problems with in-complete environment description [56] [59] [122].

5.5 Process control, monitoring and diagnosisProcess control, monitoring and diagnosis are closelyrelated, partly overlapping fields. Monitoring involvesobserving, recording, and processing signals, and detect-ing abnormal conditions of the controlled process. Diag-nosis is the process of generating plausible hypotheseson the causes that led to the current (abnormal) state of asystem. Prognosis refers to predicting when a particular state will occur next or what state the system will reach ata specified future time.Monitoring, diagnosis and prognosis apply both to physi-cal processes (e.g., at the machine level) and businessprocesses (e.g., material and work flows). Signals in thephysical process correspond to process parameters suchas force, vibration, temperature, pressure. For businessprocesses, material movements (e.g., from one locationto another one), process completion times, and other transactional data associated with information or materialflow serve as the signals. The frequency and range of physical process signals are much higher. While auto-matic (feedback) control is quite common for physicalprocesses, business processes will likely require humanintervention. Despite these seemingly wide differences,the scientific principles and even the techniques underly-ing monitoring, diagnosis and prognosis are similar inboth the physical and business settings.The application of pattern recognition (PR) techniques,expert systems (ESs), artificial neural networks (ANNs),fuzzy systems (FSs) and nowadays hybrid artificial intelli-gence (AI) techniques in manufacturing can be regardedas some consecutive elements of a process started morethan two decades ago [123]. Extensive background onmonitoring can be found, e.g., in [124] [125] [126] [127]and [128].In the CIRP-survey on developments and trends in controland monitoring of machining processes, the importance of sensor integration/fusion , sophisticated models, multi-model systems, and learning ability was also outlined[129]. The development of user-programmable, multi-purpose, modular monitoring systems in manufacturingstarted in the early eighties [130] [131]. In these systemsthe main goals of modularity were to accomplish sensor integration and achieve appropriate computing power.Modularity can be regarded as a step towards agent-based systems. In [132] a multi-agent approach waspresented for the selection of grinding conditions. Theagentification was made according to the different reason-ing approaches used, i.e., case-based reasoning for se-lecting the combinations of the grinding wheel and valuesof control parameters, rule-based reasoning for caseswhere no data were available in the case-base, and neu-ral networks for selecting the grinding wheel if required.The final decision was made by the operator.In [133] an approach for the modeling, monitoring andoptimization of manufacturing processes and processchains was introduced. A significant feature of the pro-posed technique was that it could handle models of dif-ferent types at the same time.By taking an agent-based approach, sub-systems andsub-system components are mapped to agents and agentorganizations; interaction between sub-systems and sub-system components are mapped to cooperation, coordi-nation and negotiation mechanisms; and relations be-tween them are mapped to explicit mechanisms for repre-senting organizational relationships [134].The distributed architecture of the MAGIC multi-agentdiagnostic system is based on the Multi-Agents-Multi-Level (MAML) concept as depicted in Figure 9. The ideais that the tasks of the complex embedded system's diag-


11/24

nosis and operator support are distributed over a number of intelligent agents which perform their individual tasksnearly autonomously and communicate via the MAGICarchitecture [135]. An architecture like this can be readilydistributed on and adapted to existing monitoring andcontrol systems of large-scale plants.The holonic control system described in [136] specifies amulti-robot control for surface treatment, especially for shot-blasting. The planning activities of the robots aresupported by product models. Task-sharing for the robotsproceeds through blackboard or contract net based nego-tiation, including task allocation, detailed motion planning(in time and space). The robots support each other byproviding sensor information to locate work pieces and bysharing the workpiece to gradually generate the motionplans for the shot blasting paths.

End User SystemEngineer (EUSE)

IndustrialProcess

Process Specifica-tion Agent (PSA)

DataAcquisitionAgent (DAA)

DataBase

Diagnostic Agents (DA)

DiagnosticAgent 1

SignalBased

ic2

Graph

ic..

...

icn

Neuro

DiagnosticDecision

Agents (DDA)

DecisionSupport

Agents (DSA)

OperatorInterface

Agents (OIA)

Operator

Figure 9: Concept of the MAGIC multi-agent-based diag-nostic system [135]. An agent-based diagnostic system was developed for aPLC-controlled low-volume assembly line for an Austra-lian automotive manufacturer [137]. The line consists of three assembly stations linked by a transfer line. Theholonic, model-based diagnostic system represents aholarchy of four diagnostic holons representing the lineand the three stations, respectively. The on-line test dem-onstrated the applicability of the approach, by giving 95%fault coverage and 20-40 s/fault execution time.IT systems (digital factory tools) on the operating levels,usually, are not yet integrated and thus, they supportseparate tasks such as production order control, produc-

tion monitoring, sequence planning, vehicle identification,quality management, maintenance management, materialcontrol and others [138]. An agent-based productionmonitoring system (ProVis.Agent) was developed for aleading German car factory for monitoring and controlling

the body, paint and assembly shops, either from a centralcontrol room, or from decentralized control panels on theshop floor [138]. System modules were wrapped andenabled to communicate via a standard agent platform.Later, it was connected to an automatic object identifica-tion and localization system, closing the gap betweenproduction monitoring and sequence setup [139].The Multi-Agent Tool Management System (MATMS) for automatic tool procurement in a supply network was in-troduced in [140]. MATMS operates in the framework of anegotiation-based, multiple-supplier network where aturbine blade producer requires dressing jobs on worn-outgrinding wheels from external tool manufacturers. Animportant element of MATMS is the Dressing Time Pre-diction Agent for predicting grinding wheel dressing cycletimes founded on historical data, under both supplier independent and dependent bases.Modern communication technologies have made dramaticimpacts on remote monitoring and maintenance. The e-maintenance platform named POMAESS (Problem-Oriented Multi-Agent-based E-Service System) intercon-nects separated (reactive and cognitive) agents for col-

laborating in an open and time-constrained environment[141]. The platform supports the cooperation of differentproblem solving experts (human or autonomous machine)such as production management expert, maintenanceexpert and control expert. An important feature of thedeveloped system is the integration of case-based rea-soning. A so-called watchdog agent is described in [142]which - in a remote monitoring and e-maintenance setting- provides continuous monitoring and prognostics of assetdegradation and also evaluates asset performance.

5.6 Enterprise organization and integrationEnterprise integration is aimed at providing an informa-tion technology infrastructure for all business, engineer-ing, operational and administrative functions of an enter-prise that can be used for information exchange, decisionmaking, coordination and collaboration. The integrationefforts, usually, come together with organizational re-design and the re-engineering of business processes bothwithin and between the main functional entities. A sum-mary of practical and research issues is given in [143]. Itis a general observation that collaboration among theentities of an enterprise could help increasing the sys-tems overall flexibility, reliability and performance.Key ideas related to enterprise organization and integra-tion are enterprise modeling, distributed planning andcontrol, and information system modeling. Primarily dueto globalization, the nature of businesses tends to bedistributed making modeling, monitoring and control of business processes critical. Applying decentralizedagents in enterprise integration permits the local parts of an enterprise to continue operation during temporarylapses in connectivity. Modeling formalisms include Petrinets, finite state machines, holons, and software agents.The most important questions here are: how to achieve the appropriate representation of

process models, how to model various constraints within and between

business functions such as marketing, design, plan-ning, manufacturing, and material supply, and howto use them in order to find the best-of-practice proc-ess, and

how to maintain interdependencies within the networkof organizational entities.

Furthermore, distributed modeling and control of ex-tended enterprises need to consider scaling and temporalissues, as well.


12/24

In an early work [144] a multi-agent framework was pro-posed for coordination among the different components of the manufacturing system on system level, process leveland decision level. Each agent has been modeled asconsisting of three main parts: a knowledge base, a con-trol unit and a functional component. Agents representedboth physical and decision making entities: hence lot-sizing decisions, production scheduling and resource

assignment could be coordinated. A number of current approaches for agent-based enter-prise organization and integration can be found in theliterature. For instance, the principles of the ADACORadaptive holonic production control architecture are ap-plied in the context of a global value network [145]. TheExPlanTech framework [146] covers the functional di-mension: it provides specialized modules to support com-plex planning, resource allocation and scheduling prob-lems (so-called heavy duty planners). The system has itsown ontology and, being implemented on the top of JADE, runs on various platforms. A knowledge-intensivemulti-agent cooperative framework for concurrent intelli-gent design and assembly planning, utilizing a Petri netas a core representation system, is proposed in [147].The Collaborative Agent System Architecture (CASA)system for multi-plant production planning in Internet-based collaborative enterprises is presented in [148]. Thisenterprise model includes the related functional compo-nents, communication infrastructure and coordinationmechanisms. The multi-agent cooperative scheduling(MACS) system maintains a dynamic "envelope" com-posed of capacity and temporal constraints for delimitingthe solution space for various agents of an extendedenterprise [149]. A recent work [150] emphasizes the communication andtemporal aspects of enterprise integration and, for real-time information exchange and decision making pro-poses, a web-based multi-agent language is used that isgrounded in the Extensible Markup Language (XML) andJava.One specific and recurrent concern of todays enterprisesis the separation between planning and executing activi-ties, which results in loss of time and information [151].Concurrent process planning and scheduling bymeans of intelligent software agents are addressed in[152] [153]. There, a layered system architecture is pro-posed, which is based on the application of cooperativeagents for optimizing information logistics between proc-ess planning and production control. Product- and pro-duction capacity-related information are taken into con-sideration on the level of centralized process planning. Onthe other hand, process planning knowledge is used for

short-term decentralized scheduling decisions on theshop floor level [153]. Therefore, the planning and sched-uling functions on the different levels can be carried out inrelative independence. Problems which result from time-delayed return of manufacturing knowledge and capacitydata, or other lacks of information flows (e.g., from theuse of static process plans) are eliminated (see Figure10).Finally, an important practical question is how to integrateagent-based solutions into existing systems. In [154]approaches to agentification of resources or even wholemanufacturing systems were introduced partly as anextension of the Virtual Manufacturing (VM) concept [16].The ability of dealing with so-called legacy software pack-ages in a complex system is of high importance. Agentscan provide a wrapper both for legacy systems and hu-man participants, as shown in [101].

Figure 10: Electronic marketplaces for cooperative agents[153].

5.7 Production in networks A production (or supply ) network is a net of suppliers,factories, warehouses, distribution centers and retailersthrough which raw materials are acquired, transformedand delivered to customers. In a supply network, thetraditional boundaries of firms are dissolved: decisions onthe use of resources should concern both internal and

external capacities, and the internal flow of materialsshould be synchronized with the incoming and outgoingflows. There exist a number of Supply Chain Manage-ment (SCM) systems for integrating data of all major business functions at the nodes of a supply network, butthese systems are rather transactional: they provide tech-nology for information storing, retrieval and sharing, butdo not really support decision making [26].SCM is an approach satisfying the demands of customersfor products and services via the integrated managementof the flow of materials, information and financial assetsbetween autonomous business partners. Since supplynetworks are unique and complex, local planning at thenodes is done in many different ways: there is no one-size-fits-all solution. Instead, there is a need for a portfo-lio of coordination mechanisms where relations betweenthe partners can be represented on a range of tempera-tures: from cold (competitive auctions, single businessrelations), through warm (cooperative planning), to hot(full integration). Basically, there are two types of re-search in applying multi-agents to supply chain manage-ment: The general approach handles SCM as a problem of

designing and operating a multi-agent system. Specific problems in the supply chain management,

such as collaborative inventory management [155],bidding decision [156], and material handling and in-ventory planning in warehouses [157] are solved byborrowing some MAS concepts and methods.

The majority of the literature has been focusing on thegeneral application of agent-based supply chain man-agement [96] [158] [159] [160] [161] [162] [163].


13/24

In [157] the AWAS (Agent-based model for Warehousesystem) systems was proposed. AWAS consists of threesub-systems: an agent-based communication system(ACS), an agent-based material handling system(AMATH), and an agent-based inventory planning andcontrol system (AIPCON). These sub-systems are de-signed to cooperate together to facilitate just-in-time ex-change of orders and materials. Under these sub-

systems, seven kinds of basic agents are defined includ-ing customer, supplier, order, inventory, product, supplier-order, and automatic-guided vehicle (AGV) agents.In [96] the problems based on a representation of obligedand forbidden behavior among agents were analyzed inan organizational framework, together with an inferencemethod that decides which obligations to break in conflict-ing situations. These are integrated into an operational,practically useful agent development language. The major strength of the approach is the way it supports coordina-tion by exchanging constraints about obliged and forbid-den behavior among agents.Issues and solutions for the construction of an agent-oriented software architecture for managing the supply

chain on the tactical and operational level are presentedin [158]. Among other things, a communication protocolfor coordination between supply chain agents is pro-posed. The approach relies on the use of an agent build-ing shell (ABS), providing generic, reusable, and guaran-teed behavior components and services for speech actbased communication, conversational coordination, androle-based organization modeling. The ABS is a collectionof reusable software components and interfaces thatsupport application-independent agent services. Usingthese services, developers can build on a high-level infra-structure, whose abstractions provide a conceptualframework that helps in designing and understandingagent systems, eliminate work duplication, and offer guarantees about the services provided by the tool. For illustration, six types of agents are used: order acquisition,logistics, transportation, scheduling, resource, as well asdispatching agents. A framework for computer support of supply chain man-agement in chemical and process industries by usingJATLite (Java agent template, lite) and the standardknowledge query message language (KQML) was pre-sented in [159]. The architecture of the multi-agent supplychain support system is shown in Figure 11. The systemis defined through retailer, logistics, warehouse, plant,purchasing and raw material supplier agents.

Retailer AgentRetailer

Delivery Reply Enquiry

Warehouse

Warehouse

Warehouse

LogisticsData

MiningTools

Plant

Plant

Plant

Operation

SchedulingResourceManagement

Purchase

Reply Enquiry

Raw Materials

Delivery

Supplier

Figure 11: The multi-agent-based supply chain [159].

The integration of partners who use heterogeneous in-formation systems was discussed in [164]. To support thecommunication of different partners and enable the infor-mation flow within the value added chain, XML neutraldata format was used. Data are exchanged among thecompanies via a multi-agent software system that handlesthe execution of the business process. It is combined witha generic hierarchical planning model, to improve the

performance of the supply chain utilizing the productionresources, while in parallel it enables the user to custom-ize the resource allocation method via the selection froma set of dispatching algorithms. Additionally, the application of web-based software tech-nologies for supporting the planning and monitoring a shiprepair contract was presented in [165]. The discussionwas based on the transactions among the ship-owner, theshipyard, the material suppliers and the shipyards sub-contractors in the maritime supply chain. Each companywas represented by a set of software modules implement-ing an 'agent-like' behavior and controlling the executionof critical supply chain functions. A business model wasimplemented in a software package, representing a costeffective, simple to configure and use, platform independ-ent, integrated environment. The system also allows smallenterprises with moderate information technology infra-structure to participate in the supply network. The adapta-tion of an approach like this is feasible and reduces com-munication efforts, while it improves the communicationamong the cooperating companies in a supply chain con-tract. A Multi-Agent Tool Management System (MATMS) for automatic tool procurement in a supply network was in-troduced in [166]. MATMS has been developed for amultiple-supplier network where a turbine blade producer (customer) requires from external tool manufacturers(suppliers) the purchase of new grinding wheels and theperformance of dressing operations on worn-out grindingwheels for turbine blade fabrication. The system has alayered architecture [167]: the Supplier Network Levelwhich comprises of supplier type agents is responsible for carrying out the actual dressing jobs. The EnterpriseLevel coordinates MATMS activities to achieve the bestpossible results in terms of on-time delivery, minimalinventories and costs. Agents on the Plant Level repre-sent production lines of the customer factory.The unified framework for supply chain decision supportsystem proposed in [160] integrates the various elementsof a supply chain such as enterprises, their productionprocesses, and their associated business data andknowledge and represents them in a unified, object-oriented way. Supply chain elements are classified as

entities, flows and relationships. Software agents areused to emulate the entities i.e., various enterprises andtheir internal departments. Flows of material and informa-tion are modeled as objects. A virtual market solves the allocation of resources in adynamic environment as described in [161]. Efficientcoordination among agents is very important for realizingthe agility of supply chains. Typically, there are two typesof agents in the architecture of multi-agent-based agilesupply chains: functional agents and mediator agents.Functional agents perform planning and/or controllingactivities in the supply chains, while mediators play asystem coordinator role by prompting cooperation amongagents and providing message transfer services [162].This organization is realized according to the standardiza-tion of work or skill pattern (see Section 4.3).Negotiation via constraint evaluation is defined to supportsupplier selection in a hybrid supply chain [168]. Strate-gies are defined via constraint logic programming predi-


14/24

cates for proposal request, generation, evaluation, andcounter-proposal generation, as well as for order negotia-tion. The approach has been implemented as a set of negotiating processes operating across the internet.The use of mobile agents in a Brazilian automobile sup-ply network was presented by [163]. The main motivationof this work was to design and implement an environmentconductive for flexibility, adaptiveness and efficiency. Thesystem uses the Kernel Language for Agent Interactionand Mobility (KLAIN), a language that realizes a pro-gramming paradigm where processes can be transportedacross different computing environments.Members of a supply network may take part incoopera-tive planning [169]. In this case, partners have a definiteincentive and commitment to cooperate and share boththeir risks and benefits. The main driver for cooperation isuncertainty, which has its roots in market demand, manu-facturing and supply. Uncertainty can be managed only if powerful planner systems fill in the various planner roleslocally. Plans which are really executable and cost effi-cient make the future - even market demand - more pre-dictable, and the actual production more profitable. Novel

information channels have to be established between thepartners so as to share the results of local planning, fromdetailed production schedules up to demand plans, on allthe horizons and levels.Finally, in a broader setting, the phenomena of emer-gence of products and their underlying supply networksare demonstrated in [170] by simulation, showing thepotential of emergent synthesis in this complex field.One of the recurrent issues related to supply chains ishow to handle complexity. There has been a renewedinterest in this problem, and now, multi-agent modelingand analysis is providing new approaches. We discussrelated issues in the strategic directions section.

5.8 Assembly and life-cycle management Assembly, usually, represents the last technical step inthe product creation process; however, according to theup-to-date organization principles, the ready-made prod-ucts are to be followed throughout their life cycles. Hierar-chical, heterarchical and holonic control structures for anassembly cell are compared in [31]. Holonic systems werefound to deliver better performance in a wider range of situations than their more conventional counterparts. For instance, the holonic concept demonstrated improve-ments in the robustness and volume flexibility in an en-gine assembly system in the automotive industry [172].

The concept of plug & produce was introduced in [173]as analogy of plug & play in computer world. The viabil-ity of the approach was demonstrated by using a holonicassembly cell installed by the authors.The value-adding chain of a modern enterprise is a world-wide network of suppliers, transport agencies, manufac-turers, retailers, distributors and recyclers. In [171] anagent-based approach is presented which integrates thesupplier and transportation agency into assembly control.The objective of agent-based control is to avoid large leadtimes and capital-intensive stock, and to make the as-sembly process as refractory as possible to disturbances.The agent society, organized on three levels, is shown inFigure 12. The first level incorporates the negotiationagents: Assembly Scheduler (AS), Assembly Transporta-tion Scheduler (ATS), Transportation Scheduler (TS), theSupplier Scheduler (SS) and the Master AssemblyScheduler (MAS). The example given has only one MAS, ATS and TS, four SS and AS. The second level belongsto the Carrier Agent (C) and the Assembly Agent (A). Thelatter tries to ensure assembling of the right parts. TheCarrier Agent supervises the complete tour of delivery.

On the third level the stock agents represented by theSupplier Buffer (SB) and the Assembly Buffer (AB) passinformation on the actual inventory levels to agents re-questing them [171]. The authors outline that by using theproposed architecture the network became more trans-parent and can be operated with smaller inventory.The life of the product holons is not ended with the ac-complishing of the manufacturing process, rather, it existsuntil the product is disposed of. According to this idea, theproduct holon proposed in [174], besides containing dataconcerning the information on the product life cycle, user requirements, design, process plans, bill of materials,quality assurance procedures, and the process and prod-uct knowledge, also contains the information related to itsend-of-life. A framework based on autonomous, cooperative agentsfor life-cycle oriented data support is presented in [175].The framework identifies internal and external datasources for product optimization. Depending on their types, the data are stored and permanently updated indifferent decentralized locations (web, company and ma-chine). Initiated by the product manufacturer, the data-bases will be permanently updated with changes or ex-periences from current operations. The data can be ac-quired from machine control, Auto-ID or other sources.

ASi

ATS j MAS

ManufacturingScheduling

TransportationScheduling

AssemblyScheduling

Supplier Scheduler

TransportationScheduler

Assembly TransportationScheduler

Assembler

Carrier

Supplier Buffer Assembly Buffer

Agent Negotiation Information

SSi

SB1 SBk...

... ...

... ... ...

C1 C2 Cn

TS j

A1 A2 Am

Master Assembly Scheduler

AB11 AB1n AB21 AB2n ABm1 ABmn

Assembly Scheduler

Figure 12: Society of agents for assembly control [171].


15/24

6 USE OF AGENTS Agent technology and multi-agent systems have becomeprevalent in the past decade, enabled by a wide spectrumof information and control technology (such as network-ing, software engineering, distributed and concurrentsystems, mobile technology, electronics commerce, inter-faces, semantic web). As reviewed above, developmentslike this found their way to application and deployment in

all fields of manufacturing. How could the idea of agentsbecome so powerful? Below, also taking the perspectivesof [1] into consideration, some explanations are given.

6.1 Agents as a design metaphor As computer systems become ever more complex, wealso need more powerful abstractions and metaphors todesign and explain their operation. The concepts of agents enables us to structure our knowledge (and sys-tem design, accordingly) around components that haveautonomy and capability to communicate. This decompo-sition may happen both along functional or physical di-mensions. Objects that earlier had complex propertiescan now be personified. We can demarcate agents fromthe environment, and then attach intentional notions such

as goal, objective, belief, intention and plan to them. Thisso-called intentional stance is an abstraction that helpsus in a useful and familiar way to describe, explain, andpredict the behavior of complex systems [176]. Note thatmost important developments in computing have beenbased on new abstractions, such as procedures, abstractdata types, objects. Now, agents as intentional systemsrepresent a further and increasingly powerful abstraction. Any kind of intelligence requires the handling of conflicts rooted in disparate interests. Efficient operation, adapta-bility, and occasionally even the survival of a complexsystem hinges on whether it is able to resolve conflictsituations. The conflicts and their resolutions constitutethe base of the so-called social knowledge a kind of

knowledge which emerges from the collective action of the individuals and, therefore, belongs to the system as awhole. Instead of trying to eliminate conflicts (which isimpossible), conflicts are rather to be exploited. Conflictsbecome explicit if the system is modeled like a communityof self-interested, rational agents. Hence, the agent-based approach forces the systems designer to find waysfor managing conflicts.

6.2 Agents as source of software technology As far as software technology is concerned, agents andMAS provide a wide array of models, techniques, formalmodeling approaches and development methodologiesthat all together shape the general-purpose paradigm of agent-oriented software engineering (AOSE) [41]

[177]. There have been developed several programminglanguages and software development environments whichnot only support MAS programming, but also implementkey concepts of MAS in a unified framework. Now con-solidated FIPA standards are available which are in-tended to promote the interoperation of heterogeneousagents and the services they represent. The FIPA specifi-cations include Agent Communication Language mes-sages, message exchange interaction protocols, speechact theory-based communicative acts and content lan-guage representations [48]. The agent unified modelinglanguage (AUML) [178] and softwareplatforms such asthe IBMs Java Agent DEvelopment Framework (JADE)[179], IAIs CyebelePro [180], the BDI agent frameworksJadex or JACK [181] [182] the Zeus Agent Toolkit, andthe Cougaar Agent Architecture [183] are all available for the specification and realization of multi-agent systems.

CougaarJADE

Jack

ZeusJadex

FIPA ACL

OIL

DAML+OIL

OWL

XML 1.0

RDF

FIPA 00 AUML

Swarm

GaiaMaSE

Tropos

KQML

REPAST

2000 2005

Standards

Platforms

Methodologies

Figure 13: Current standards, platforms and methodolo-

gies of agent-oriented software engineering.These platforms reduce the developmental time of agent-based systems by providing the decomposition and com-munication infrastructure. For instance, JADE and Cye-bele are general-purpose, while Couggar is readily appli-cable to logistics applications. Once the designer definesthe agents and their interconnections, all of these plat-forms allow the designer to build the agent system withrelative ease. However, it must be noted that buildingagent functionalities is left to the designer. All the aboveplatforms give application protocol interface facilities,making the integration of other programs and systemsrelatively less cumbersome. The platforms communica-tion is based on KQML structure and are FIPA compati-ble. Furthermore, they allow for agent mobility and sincethey are based on Java, they are independent of theactual hardware/software platforms. There are frame-works directly developed to support simulation, such asSWARM or its successor, the Recursive Porous AgentSimulation Toolkit (REPAST) [184] .Recent developments have provided infrastructure for

open agent communities. The DARPA Agent MarkupLanguage (DAML) project was aimed at developing alanguage and appropriate tools to facilitate semanticinteroperability of agents [185], just as parallel efforts of the W3C consortium have resulted in the standards of Extensible Markup Language (XML) [186], and the so-called Resource Description Format (RDF) [187]. Theserepresentation technologies set the base for the recentlydeveloped Web Ontology Language (OWL) [188].However, it is a wide-spr

agent base system for manufacturing.pdf

Documents