Content area
Agent technology derived from Distributed Artificial Intelligence is increasingly being considered for next generation computer-integrated manufacturing systems, to satisfy new requirements for increased integrability, configurability, adaptability, extendibility, agility, and reliability. This paper reviews our previous research on the application of the agent-based technology to intelligent design and manufacturing and describes the current research project MetaMorph II (an agent-based architecture for distributed intelligent design and manufacturing). [PUBLICATION ABSTRACT]
*u*n*s*t*u*r*t*u*r*e**t*e*x*t*
Journal of Intelligent Manufacturing (2000) 11, 237s'251 MetaMorph II: an agent-based architecture for distributed intelligent design and manufacturing
W E I M I N G S H E N , * F R A N C I S C O M AT U R A N A - and D O U G L A S H . N O R R I E
Division of Manufacturing Engineering, The University of Calgary, Calgary, Alberta, Canada T2N 1N4 E-mail: [wshen -- Maturana -- Norrie]@enme.ucalgary.ca
Received July 1997 and accepted April 1999
Agent technology derived from Distributed Artio""cial Intelligence is increasingly being considered for next generation computer-integrated manufacturing systems, to satisfy new requirements for increased integrability, cono""gurability, adaptability, extendibility, agility, and reliability. This paper reviews our previous research on the application of the agent-based technology to intelligent design and manufacturing and describes the current research project MetaMorph II (an agent-based architecture for distributed intelligent design and manufacturing).
Keywords: Agent, multi-agent systems, distributed systems, intelligent design and manufacturing, shop r'oor planning and scheduling, concurrent engineering
1. Introduction New architectures for next-generation of computer-
integrated manufacturing (CIM) systems need to be distributed, intelligent, and to incorporate concurrent design and manufacturing (concurrent engineering) principles. Multi-agent system architectures offer a promising solution with their novel approaches for dynamically creating and managing agent commu- nities in widely distributed and ever-changing manufacturing environments (Pan and Tenenbaum, 1991; Norrie and Gaines, 1996). Recently, a number of researchers have attempted to apply agent technology to concurrent engineering, manufacturing enterprise integration, supply chain management, manufacturing scheduling and control. We review brier'y some related projects in the following paragraphs.
SHADE (McGuire et al., 1993) was primarily concerned with the information sharing aspect of the concurrent engineering problem. It demonstrated a r'exible infrastructure for anticipated knowledge- based, machine-mediated collaboration between dis- parate engineering tools. PACT (Cutkosky et al., 1993) was a landmark demonstration of both collaborative research efforts and agent-based tech- nology. SHARE (Toye et al., 1993) was concerned with developing open, heterogeneous, network- oriented environments for concurrent engineering. FIRST-LINK (Park et al., 1994) was a system of semi- autonomous agents helping specialists to work on one aspect of the design problem. NEXT-LINK (Petrie et al., 1994) was a continuation of the previous project for testing agent coordination. Process-Link (Goldmann, 1996) followed on from Next-Link and provides for the integration, coordination, and project management of distributed interacting CAD tools and services in a large project. SIFA (Brown et al., 1995) was intended to address the issues of patterns of interaction, communication, and conr'ict resolution. RAPPID (Responsible Agents for Product-Process
0956-5515 # 2000 Kluwer Academic Publishers
*Currently at IMTI, National Research Council of Canada, London, ON, Canada. E-mail: [email protected] -Currently at Rockwell Automation, Cleveland, OH, USA. E-mail: [email protected]
Integrated Design) (Parunak et al., 1997b) was proposed to develop agent-based software tools and methods for using market place dynamics among members of a distributed design team to coordinate set-based design of a discrete manufactured product. MADEFAST (Cutkosky et al., 1996) was a DARPA DSO-sponsored project to demonstrate the techno- logies developed under the ARPA MADE (Manufacturing Automation and Design Engineer- ing) program. It was an ambitious experiment in collaborative engineering over the Internet.
Pan and Tenenbaum (1991) described a software Intelligent Agent (IA) framework for integrating people and computer systems in large, geographically dispersed manufacturing enterprises. This framework is based on the vision of a very large number (e.g., 10,000) computerized assistants, known as Intelligent Agents (IAs) and human participants are encapsulated as Personal Assistants (PAs), a special type of IA. Roboam and Fox (1992) presented an Enterprise Management Network (EMN) to support the integ- ration of activities of the manufacturing enterprise throughout the production life cycle with six levels: Network Layer, Data Layer, Information Layer, Organization Layer, Coordination Layer and Market Layer. Barbuceanu and Fox (1997) suggested organ- izing the supply chain as a network of cooperating agents, each performing one or more supply chain functions, and each coordinating their actions with other agents. A similar proposal had been made by Swaminathan et al. (1996) for modeling supply chains, with a multi-agent framework being devel- oped. The AIMS (Agile Infrastructure for Manufacturing Systems) team proposed to create an open, scalable infrastructure for agile manufacturing, and to demonstrate its effectiveness in pilot produc- tion (Park et al., 1993). The infrastructure was intended to provide standardized ways of accessing a wide variety of agile production services over local networks as well as the Internet.
Butler and Ohtsubo (1992) described a distributed architecture for dynamic scheduling in a manufac- turing environment. Fischer (1994) proposed a hierarchical planning structure consisting of six layers: the layer of the production planning and control system, the layer of the shop r'oor control system, the task coordination layer, the task planning layer, the task execution layer and the machine control layer. AARIA (Autonomous Agents for Rock Island Arsenal) (Parunak et al., 1997a) was developed as an
industrial-strength agent-based shop-r'oor scheduling and control system.
The above is only a short overview of related research literature and several research projects related to agent-based engineering design and manufacturing. A more detailed review of cooperative environments for engineering design can be found in Shen and BartheA's (1996a) and an extensive survey of research projects related to agent-based manufac- turing in Shen and Norrie (1998). A comprehensive review of agent theories, architectures, and languages, can be found in Wooldridge and Jennings (1995).
This paper now presents our previous work on the application of the agent-based technology to intelli- gent design and manufacturing and describes the current research project MetaMorph IIDHan agent- based architecture for distributed intelligent design and manufacturing. Section 2 presents the DIDE (Distributed Intelligent Design Environment) archi- tecture and its implementation; Section 3 describes the main features of the Metamorph I architecture and its prototype implementation; Section 4 introduces a new MetaMorph II architecture and several related issues, and Section 5 gives some concluding remarks.
2. DIDE (distributed intelligent design environment)
DIDE was a multi-agent experimental environment
developed at the University of Technology of CompieA'gne for large engineering projects, originally for engineering design, but its general architecture can also be used to develop advanced distributed manufacturing systems or distributed integrated CAD/CAM systems. Its objective was to integrate existing engineering tools, like CAD/CAM tools, database systems, or knowledge-based systems, into a truly open system, that is, a system for which users can freely add or remove tools without having to halt or to reinitialize the work in progress (Shen and BartheA's, 1996b).
2.1. General architecture The general architecture of DIDE is organized as a
population of asynchronous cognitive agents for integrating engineering tools and human specialists in an open environment. Each tool (or interface for human specialist) can be encapsulated as an agent.
238 Shen, Maturana and Norrie Engineering tools and human specialists are con- nected by a local network and communicate via this network. Each can also communicate directly with other agents located in the other local networks using the Internet. All agents are independent and autono- mous. There is no facilitator structure as in PACT (Cutkosky et al., 1993). The agents exchange design data and knowledge via a local network or the Internet.
Several issues need to be noted. Firstly, the system is not intended to run automatically. On the contrary, human beings are part of the system. The system cannot be organized independently of the company structure. Thus, we assume that the project will be run by a project manager, and that each local group will in turn have a local project manager. Note, that in this type of system, human specialists do not have direct control either of the other agents or of the whole system. Each human specialist works with an inter- face which is encapsulated as an agent connected to the system. For example, the project manager can start or stop the design process, may take some decisions when replying to the requests of other agents, but has no control over the other agents or on the whole working environment.
A second issue relates to the control of the design task. Apart from the project manager, we assume no central control of the design task. Subtasks are created for answering requests as needed, once the work is initiated. There is no global planning, neither central nor distributed. However, some subtasks may have predeo""ned sequences (e.g., scenarios, amounting to local plans). When this is the case, there is no reason why such sequences should not be used.
A third issue concerns the global consistency of the design. Consistency is kept at a minimum during each stage of design process, considering that each agent has a local model of the designed device. However, in mechanical design it is sometimes possible to show a model of the designed product. In such a case, each subgroup is allowed to modify the model indepen- dently, using its own design space. Results are kept in different versions of a model database. Then, at each review meeting, it is the responsibility of the product manager to merge all the proposed versions into a unique design. Of course, the environment has to support the merging process (reconciliation) efo""ciently (BartheA's, 1994b).
A fourth issue concerns the agent behavior. We assume that all agents are connected by means of a
network, local network or the Internet. Locally, each agent can reach any other active agent by means of a broadcast message. All agents receive messages. They may or may not understand such messages. When they do not understand a message, they simply do nothing. Whenever they understand the message, they start working on it, provided the priority of the message is higher than the current work they were doing. Thus, agents are multi-threaded. When a new agent is introduced and connected to the system, it informs other working agents, builds its own representation of the task being solved and acquires information about other agents. Thereafter, it receives messages like any other agents. The situation becomes more interesting when some agents are actively offering services knowing the o""nal goal, i.e., the agents not only reply to the requests of the other agents, but also can give some suggestions or advice to the current design project according to their knowledge.
A o""fth issue is related to the legacy problem. In our design environment, an agent offers some specio""c service, usually by encapsulating a particular en- gineering tool. Agent interaction relies on shared concepts and terminology for communicating know- ledge across disciplines, an interlingua for transferring knowledge among agents, and a commu- nication and control language that enables agents to request information and services. This technology allows agents working on different aspects of a design to interact at the knowledge level, sharing and exchanging information about the design indepen- dently of the format in which the information is encoded internally
2.2. Internal structure of an agent In this environment, agents are autonomous cognitive
entities, with deductive, storage, and communication capabilities. Autonomous here means that an agent can function independently of any other agent.
A DIDE agent is composed of o""ve parts (as shown in Fig. 1): (i) a network interface; (ii) a communica- tion interface; (iii) symbolic models of the other agents, and associated methods to use them; (iv) a model of its own expertise with its internal knowledge bases; (v) a model of the task to be performed, or of the current context (local knowledge).
The network interface simply couples the agent to the network. The communication interface of an agent is composed of several methods or functions for
MetaMorph II 239 treating all incoming and outgoing messages. A message box is used to temporarily store all received messages. Processing incoming messages requires two steps: (i) receiving, storing, and sorting messages; (ii) encoding a message content for further processing by the agent in the context of a particular task. Processing an outgoing message similarly requires encoding of the information to be transmitted, and mailing it according to the exchange protocol. The symbolic models of the other agents are constructed using a knowledge base containing information about the other agents, obtained during interaction. The model information includes the agent's name, address, and skills or competencies. This knowledge helps the current agent to select one or more agents as subcontractors for processing tasks. The model of its ``own expertise'' is also a knowledge base composed of self information such as the name, address and competencies or skills. The latter may be methods for activating tasks corresponding to the received requests. The local knowledge is a knowledge base composed of the information about the working project and the subtasks. It contains also the historic information of the agent. It is implemented by a cache mechanism. Each agent always records important information (e.g., key words and key parameters, etc. . .) extracted from incoming messages and out- going messages in its knowledge bases. Such information can be reused in the future using case-
based reasoning techniques so as to increase agent's responsiveness, reduce communication overhead, and therefore improve the system's performance. Such a mechanism is particularly useful for database agents.
At o""rst, when a new agent is connected to a group of active agents, only its communication interface and its own expertise contain information. The part which records facts about the work to be done, or the capabilities of the other agents, is empty. In the case of slave agents (e.g., a local database), this will not change, i.e., it will remain empty. In other cases, each agent must build its own image of both the work to be done and the capabilities of the other active agents, by extracting information from the various messages it receives.
2.3. Inter-agent communication In general, communication can be synchronous or
asynchronous, and the communication mode can be point-to-point (between two agents), broadcast (one to all agents), or multicast (to a selected group of agents). DIDE uses the Inter-Agent Communication Language (IACL) developed under the OSACA project at the University of Technology of CompieA'gne (Scalabrin et al., 1996). IACL has two layers: a message layer and a content layer. The message layer encodes a set of communication
Fig. 1. Internal structure of an agent in DIDE. 240 Shen, Maturana and Norrie features that describe the parameters at the lower level of the message, such as the identio""ers of the sender and the receiver, a unique identio""er and the used protocol, etc. In addition, the message layer contains the minimal vocabulary necessary to overcome the problems concerning a common language for inter- agent communication. This type of communication language that is independent of the internal language used by the agents is the o""rst step to making possible interactions among heterogeneous systems. The content layer is the actual content of a message, in the agent's own language representation. DIDE allows for o""ve simple actions, request, inform, announce, bid and notice. They are grouped into two categories requests and assertions (REQUEST, INFORM, NOTICE); call for bids and offers (ANNOUNCE, BID). In DIDE, the messages are formatted in an extended KQML format (Finin et al., 1993).
2.4. Implementation An experimental prototype was developed for testing
the basic feasibility of the DIDE approach, including encapsulating traditional tools and bringing agents in/ out without halting the working system. The current version of the prototype was developed on a network of SUN and VAX workstations, using the OSACA platform (Scalabrin et al., 1996). Although developed as a local group, it has the capability of inter-group communication using the Internet. Some agents are implemented as MOSS objects (BartheA's, 1994a), a system of recursive frames capable of modeling objects with versions and well adapted to design activities, constructed on top of Common LISP. Some other agents are developed in C/CG*G* . Some graphic Interfaces have been developed on CLIM and LispView for showing the system information, design results, and for user manipulations. AutoCADTM is encapsulated as an agent for showing the design results. Web browser Netscape or HotJava is used to show the information about the active agents and the information about the current situation of the design project. Two databases MATISSETM (an object-oriented database which is commercial product of ADB) and EDBMS (an extended relational database developed in the Chinese Academy of Sciences) are used for storing design data and knowledge.
The test case for the DIDE system used a small
design exampleDHa simple gear box design. A detailed description of this test case can be found in Shen and BartheA's (1996b).
3. MetaMorph I approach MetaMorph I was an agent based architecture for
intelligent manufacturing developed at The University of Calgary to address system adaptation and extended-enterprise issues at four fundamental levels: virtual enterprise, distributed intelligent systems, concurrent engineering, and agent architec- tures. The architecture has been named MetaMorphic, since a primary characteristic is its changing form, structure, and activity as it dynamically adapts to emerging tasks and changing environment.
3.1. Hybrid agent architecture In the MetaMorph I architecture, hybrid agent models
have been used to build both resource and mediator agents which can be classio""ed as soft-hybrid agent models because none of the reactive-agent levels are strongly implemented in the agent structures. Resource agents are used to represent manufacturing devices and operations, while mediator agents are used to coordinate resource agents. Different levels of intelligence and behavior are associated with these two different types of agents.
3.2. Mediator-centric organization In this system, manufacturing agents are software or
hardware entities that interact, communicate, coop- erate, coordinate and negotiate with one another, and learn on behalf of the users. These agents may work in isolation or in virtual negotiation clusters for planning, scheduling, and control during the life cycle of the product. To create the virtual clusters, agents are dynamically organized into groups. Such virtual clusters are coordinated by both static and dynamic mediator agents (as deo""ned in Section 3.3) and carry out negotiation and control activities. From this point of view, the overall manufacturing system and its virtual coordination clusters are mediator- centric organizations.
MetaMorph II 241 3.3. Physical modeling Each mediator agent encapsulates functionality to
allow local coordination and interaction with other dissimilar mediator agents. Mediator agents respon- sible for mapping system entities are classio""ed as static mediator agents. In addition, there are dynamic mediator agents for coordinating dynamic interaction among agents. Dynamic mediator agents are also classio""ed into inter- and intra-cluster mediator agent categories.
Each manufacturing enterprise needs at least one high-level Enterprise Mediator agent to act as the system's integrator. This Enterprise Mediator agent is able to recognize all sub-level mediator agents, platforms, and resources in the enterprise. The Enterprise Mediator agent supplies a global view of a system during integration of plans.
3.4. Mediator agent modeling Mediator agents are intended to encapsulate various
manufacturing behaviors to facilitate the coordina- tion of heterogeneous agents. A generic model for the design of mediator agents, based on the specio""cation of various meta-level activities, has been created which can enable different types of mediator agents cope with various activities in the factory.
The generic model for mediator agents includes the following seven meta-level activities: Enterprise, Product Specio""cation and Design, Virtual Organizations, Planning and Scheduling, Execution, Communication, and Learning. Every mediator agent model or template includes some or all of these activities. The conceptual domains of these activities can be described as follows:
ffl The enterprise domain involves global know- ledge of the system and representation of goals through objectives. Enterprise knowledge enables environment recognition and maintenance of organi- zational associations.ffl
The product specio""cations and design domain includes encoding data for the manufacturing task to enable mediator agents to recognize the tasks to be coordinated.ffl
The virtual organization domain resembles the enterprise domain in some of its activities, but its scope is detailed knowledge of resource behavior at
the shop-r'oor level. Activities in this domain dynamically establish and recognize relationships between dissimilar resources and agents.ffl
The planning and scheduling domain plays an important role in integrating technological constraints with time-dependent constraints into a concurrent information-processing model.ffl
The execution domain facilitates transactions among physical devices. During the execution of tasks, it coordinates various transactions between manufacturing devices and between the devices and other domains to complete the information require- ments.ffl
The communication domain provides a common communication language based on the KQML used to wrap the message content.ffl
The learning domain incorporates the resource capacity planning activity, which involves repetitive reasoning and message exchange and which can be learned and automated.
3.5. Adaptation through coalitions In MetaMorph I, the core coalition mechanism is
based on task decomposition and dynamically-formed agent groups. High-level tasks are initially decom- posed by mediator agents acting at the corresponding information level. Each subtask is subsequently distributed among the multi-agent platforms to determine the best solution plan. Mediator agents dynamically learn from the agent interactions and identify both macro- and micro-agent coalitions1 that can be used to establish distributed searches for the resolution of tasks.
Agent coalition is incorporated in the main problem-solving mechanism in MetaMorph I. Agents are dynamically contracted to participate in a solving group. Mediators are used for coalition coordination and to create dynamic mediators to partition coordina- tion activities into inter- and intra-agent coordination levels. Learning is incorporated into some or all of the mediators based on positive agent interactions ( positive-interaction-driven learning).
3.6. Coordination level Coordination is initially involved in two main phases
of: subtasking; and creation of virtual coordination clusters. These activities are supported by the
242 Shen, Maturana and Norrie Resource Mediator, the Data-Agent Manager (DAM), and the Active Mediator (AM) agents, each of which coordinate their specio""c levels in the overall coordination task. The Resource Mediator is a static mediator, but the Data-Agent Manager and the Active Mediator are dynamic mediators created and destroyed as necessary.
At the resource community level, shop r'oors are provided with high-level static mediator agents (e.g., Resource Mediators), which collectively constitute the high-level enterprise model of the system. A high- level task initially passes through such a static mediator agent for recognition and decomposition. According to shop-r'oor capability, the subtasks are then each assigned to separate coordination clusters. Each coordination cluster commonly also incorpor- ates clone agents derived from active manufacturing resource agents (such as machine or tool agents). These entity interrelationships form a concurrent coordination framework.
3.7. Learning Two fundamental learning mechanisms have been
implemented in the MetaMorph I architecture to enhance the system's performance and responsive- ness. First, a mechanism that allows mediator agents to learn from history is placed ``on top'' of every multi-agent resource-related grouping to capture signio""cant multi-agent interactions and behaviors. Second, a mechanism for propagating the system's behavior into the future is implemented to help mediator agents ``learn from the future''. The combined action of these two learning approaches provides a basic mechanism for multi-agent manu- facturing learning.
In order to reduce the ``real-world knowledge gap,'' a forecasting simulation has been developed to support planning and scheduling in MetaMorph I. By partially projecting ``unpredictable behaviors'' and agent interactions, the multi-agent system can thus ``learn from the future'' and correct its current real-world models and provide more accurate plans. More information can be found in Shen et al. (1998b).
3.8. Implementation The MetaMorph I architecture and coordination
protocols, described previously, have been used for
implementing a distributed concurrent design and manufacturing system in simulated form. This system includes the following high-level mediators: Enterprise Mediator, Design Mediator (CAD Mediator), Resource Mediators, and Simulation/ Execution Mediator (AGV Mediator), as shown in Fig. 2. Each such mediator provides high-level coordination for the multi-agent systems within its ``domain'', as well as with the other mediators shown. Thus, each Resource Mediator provides high-level coordination for a manufacturing shop r'oor of machines, tools, and other entities. Similarly, the Design Mediator coordinates one or more design systems. Common enterprise integration protocols are used to facilitate agent interoperability.
The system is implemented within a distributed computing platform consisting of four HP Apollo 715/ 50 workstations, each running an HP-UX 9.0 operating system. The workstations communicate with each other through a Local Area Network (LAN) and TCP/IP protocol. The graphical interfaces for each high-level mediator are created in VisualWorks 2.5 (Smalltalk) programming language, which is also used for programming the mediators. The KQML (Finin et al., 1993) is used as high-level agent communication language. The whole system is coordinated by the high-level mediator agents shown in Fig. 2. This structure can be extended for realizing an enterprise of wider scope (Maturana and Norrie, 1996).
The Design Mediator provides a graphical inter- face for retrieving design information and requesting manufacturability evaluations through the Enterprise Mediator. Designs are created in a separate intelligent design system named the Agent-Based Concurrent Design Environment (ABCDE), devel- oped in the same research group (Balasubramanian et al., 1996). The ABCDE architecture includes an environment manager, feature agents, part agents, and CAD physical layers. The CAD layers are implemented in AutoCADTM with Advanced Modeling Extensions (AME) version 2.0. The multi-agent system for the CAD application is written in CG*G* . ABCDE agents interact (through the Design Mediator) with the Resource Mediators to obtain manufacturability assessments during the product design process. Human production man- agers may request manufacturability evaluations using either the CAD system or the design system interface.
MetaMorph II 243 4. MetaMorph II architecture 4.1. Conclusions and experiences from previous research work
DIDE demonstrated that the agent-based approach for
advanced CAD/CAM systems have the following advantages: (1) a multi-agent system can be devel- oped into a real open and dynamic design system; (2) autonomous and independent agents make the design system less complex and allow an easy integration of the existing engineering tools; (3) It is easy to integrate human elements; (4) It provides efo""ciency through parallel and distributed processing; (5) it is r'exible and scalable. However, there are also some disadvantages: (1) it is not easy to test the overall behavior of the multi-agent system; (2) It is not easy to integrate interactive engineering tools; (3) there are some difo""cult problems still to be resolved (such as a shared ontology, and reestablishing temporary con- sistency in the global system).
The MetaMorph I architecture facilitates integra- tion of planning and scheduling, plan reo""nement, and learning. Integration of planning and scheduling has been identio""ed as a difo""cult task in manufacturing. In the virtual manufacturing model used, intelligent agents (e.g., machine, tool, material-handling, part, and coordinator agents) are dynamically organized into coordination groups to solve these two manu- facturing issues simultaneously. Plan reo""nement enhances accuracy by using simulated real-world
information, in each manufacturing agent domain. Learning allows capture of the most signio""cant interactions among agents during problem-solving processes. Agent activity is learned, stored, and reused, thereby establishing associations between input tasks and feasible solutions obtained from agent interactions.
The experimental results of DIDE and MetaMorph I have shown the potential of the agent-based approach for advanced CIM systems. The DIDE architecture with its autonomous agents integrates existing engineering tools and human specialists into a multi-agent engineering design system, and the MetaMorph I mediator architecture can be used to provide agent-based intelligent manufacturing sys- tems with adaptive virtual organization.
4.2. MetaMorph II architecture The MetaMorph II project commenced in early 1997.
Its objective is to integrate the manufacturing enterprise's activities such as design, planning, scheduling, simulation, execution, and so on, with those of its suppliers, customers and partners within a distributed intelligent open environment. For this purpose, we propose a hybrid agent-based architecture combining and extending the architectures used in our previous projects. In this architecture, the manufac- turing system is organized at the highest level through ``subsystem'' mediators (see Fig. 3). Each subsystem
Fig. 2. MetaMorph I distributed concurrent design and manufacturing system. 244 Shen, Maturana and Norrie
is connected (integrated) to the system through a special mediator. Each subsystem can be an agent- based system (such as an agent-based manufacturing scheduling system), or any other type of system (such as a feature-based design system, or a knowledge- based material management system). Agents in a subsystem may also be autonomous agents at the subsystem level. Some of these agents may also be able to communicate directly with other subsystems or agents in other subsystems
4.3. Main characteristics of the MetaMorph II architecture
4.3.1. Integration of design and manufacturing Several design subsystems can interact within the
whole system simultaneously. Such design subsys- tems may be either agent-based intelligent design systems like DIDE and ABCDE, or other types of design systems like feature-based design systems, or knowledge-based multi-expert design systems. Each design subsystem interfaces with the manufacturing system through a Design Mediator which also serves as the coordinator of this subsystem.
Ensuring the manufacturability of the product constitutes the o""rst step in implementing concurrent engineering. Geometric and functional specio""cations,
availability of raw materials, and the capability and availability of shop r'oor resources each has a major inr'uence on manufacturability. A design may be manufacturable under one combination of product requirements and shop r'oor resources, but not under another. The integration of design and manufacturing in the MetaMorph architecture allows immediate and progressive manufacturability assessments through- out the design process. As a product part is pro- gressively designed by repeated instantiation of features, manufacturability at every level of instantia- tion is evaluated by resource agents. Design Mediator and Resource Mediators ensure the coordination among design parts and resource agents. Thus, design subsystems (or design agents) interact with resource agents via a Resource Mediator to obtain manufacturability assessments during the product design process. This process not only ensures the manufacturability of a product, but also results in incremental identio""cation of process plans for subsequent use.
4.3.2. Integration of planning and scheduling Traditional approaches to planning and scheduling do
not consider the constraints of both domains simul- taneously. In spite of being sub-optimal, these approaches have been in vogue due to the non- availability of a unio""ed framework. The proposed
Fig. 3. Functional architecture of MetaMorph II. MetaMorph II 245 architecture allows integration of planning and scheduling activities through manufacturing enter- prise level coordination between Design Mediator and Resource Mediators who in turn coordinate resource agents at the shop r'oor level.
4.3.3. Integration of manufacturing resources Manufacturing resource agents are coordinated by
appropriate mediators at all levels of the system. For example, as shown in Fig. 4, a Machine Mediator is used to coordinate all the machines in a shop r'oor; a Tool Mediator is used to coordinate all tools, and so on. A high level Resource Mediator coordinates lower-level mediators such as Machine Mediators, Tool Mediators, Worker Mediators, Transportation Mediators. In this type of architecture, the system level organization of mediators may appear to be hierarchical, but there is neither a hierarchical control structure nor a hierarchical coordination mechanism. A machine agent can also communicate and negotiate directly with a Worker Mediator, worker agents, a Tool Mediator and tool agents, and so on. In other words, the communication between two different types of resource agents (e.g., between a machine agent and a worker agent) can be direct, rather than
always through facilitators or mediators as in some other multi-agent systems (Cutkosky et al., 1993).
4.3.4. Integration of simulation and execution control Simulation Mediators will be developed to carry out
production simulation and forecasting. Each Simulation Mediator corresponds to one Resource Mediator and therefore to one shop r'oor. Execution Mediators will be developed to coordinate the execution of the machines, AGVs, and workers as necessary. Each shop r'oor is, in general, assigned with one Execution Mediator. Execution Mediators can also be considered as its interfaces with related hardware. This is related to the intelligent control project of our research group (Brennan et al., 1997).
4.3.5. Integration with customer services This will be assisted through easy-to-use interfaces for
marketing engineers and end customers to request product information ( performance, price, manufac- turing period, etc), select products, request modio""cations to a particular product specio""cation, and send feedback to the enterprise.
Fig. 4. Organization of resource agents.
246 Shen, Maturana and Norrie 4.3.6. Integration of material supply and manage- ment
A Material Mediator will be developed to coordinate a
special subsystem for material handling, supply, stock management, and related areas.
4.3.7. Integration with partners' activities Manufacturing enterprises need always to collaborate
with their partners for parts fabrication, raw materials supply etc. In this architecture, each partner has an Enterprise Mediator (e.g., Enterprise Mediator 1 and Enterprise Mediator 2 in Fig. 3) as their interface to the system. Each partner's Enterprise Mediator registers with the Enterprise Mediator of the principal manufacturing enterprise in a similar way to other subsystem mediators. A subsystem mediator may also cooperate directly with other Enterprise Mediators, e.g., a Material Mediator may communicate directly with the Enterprise Mediator of a raw material supply company.
4.3.8. Integration of human agents As in DIDE, special graphic interfaces will be
developed as necessary to facilitate the participation of the human specialists. For example, at the Marketing Mediator level, a graphic interface is needed for allowing customers to view and select products; the Enterprise mediator needs an interface for the enterprise manager; the Material Mediator needs an interface for the material supply manager; and the Resource Mediator needs an interface for the production manager.
However, in normal operation, humans will not have direct control over other mediators and agents. What human specialists may do is to start and stop the system, to input and receive information, and to response to the requests of the system for decision making or conr'ict resolution. For maintenance and other system modio""cations, the affected agent or entity may need to be temporarily deactivated. ``Hot- pluggability'', though desirable, may be quite difo""cult to achieve in all situations.
The implementation environment of the proposed architecture can be heterogeneous provided standard communication protocols are used. TCP/IP was used and tested under MetaMorph I on a network of HP workstations, SUN workstations and PCs, and could be used for MetaMorph II. It should be noted that a particular real implementation in a manufacturing enterprise may not need all components of this
architecture. Some mechanisms may be simplio""ed also in particular situations.
4.4. Internal structure of a mediator Mediators are also agents, so can also be called
mediator agents. They are different from facilitators as used in PACT (Cutkosky et al., 1993). The main difference between a facilitator and a mediator is that while a facilitator primarily provides the message services, a mediator additionally assumes the role of a system coordinator. Specio""c responsibilities include ensuring cooperation among intelligent agents and learning from the agents' behavior.
If it is to have above mentioned functionality, a mediator needs the following components (Fig. 5):
ffl a network interface for connecting it to the system;ffl
a communication interface for treating incoming and outgoing messages;ffl
knowledge about itself;ffl detailed knowledge about the agents coordinated by itself;ffl
knowledge about other mediators;ffl knowledge about its environment;ffl knowledge about the working projects or products to be manufactured;ffl
reasoning mechanisms for using its knowledge;ffl learning mechanisms for updating its knowledge;ffl control mechanisms for controlling its actions and events.
A mediator is an agent. A resource agent is also an agent. Some basic aspects of their architectures will therefore be similar. A resource agent, however, does not need all of the detailed knowledge about other agents and mediators, and about its environment and the whole project or parts to be manufactured, that a mediator may need.
4.5. Prototype implementation The current implementation of the proposed archi-
tecture consists of four mediators. Enterprise Mediator, Design Mediator, Resource Mediator and Marketing Mediator.
The Enterprise Mediator can be considered as the administration center of the manufacturing enterprise. Other mediators register with this mediator. The Design Mediator is being used to integrate a feature-
MetaMorph II 247 based intelligent design system. The functions of the design module include: (1) generating design candi- dates from the design functional requirements; (2) modeling design geometry; and (3) representing the design using manufacturing features. The Resource Mediator is used to coordinate an agent-based dynamic manufacturing scheduling subsystem. In the present prototype implementation, at the resource level, only machine agents and worker agents have so far been implemented. The Marketing Mediator is used to integrate the customer services (marketing subsystem) into the system. The functions of the marketing subsystem include: (1) using its class libraries ( product catalogue) to show different product models on a 3-D graphical interface according to the customers' requests; (2) if the requested product cannot be found in its libraries ( product catalogue), sending design (functional) requirements to the Design Mediator asking the design subsystem to design an appropriate new or modio""ed product; (3) requesting product price for any order with special product specio""cations or due time.
In this implementation, all scheduling requests derive from a feature-based intelligent design system via a Design Mediator. Each product corresponding to a customer order has been decomposed into related manufacturing features. Each manufacturing feature can be realized by a manufacturing task. These
manufacturing tasks are organized in a graph data structure representing the sequence to produce the corresponding manufacturing features. Each product is modeled as a part agent containing the information about this product including the feature data, due time, and its plan (which will be assigned with times and resources during the scheduling process).
The current prototype is being developed on a network of PCs. The main development language is VisualWorksTM (Smalltalk). The graphical interface for customer services is being developed using 3-D studioTM and Visual CG*G* TM. Communication among mediators is realized using the TCP/IP protocol and the inter-mediator messages are formatted in the KQML format (Finin et al., 1993). The current implementation is being developed intially as a simulation.
4.6. A scenario for a furniture fabrication factory The present prototype is being developed for a
furniture fabrication enterprise. Here we depict a scenario showing the system's operation.
Consider a customer wanting to buy a set of furniture. He or she comes to an agency of the furniture fabrication enterprise which is equipped with the MetaMorph II system. Since the Marketing Mediator can be connected to the Enterprise Mediator
Fig. 5. Internal structure of a mediator. 248 Shen, Maturana and Norrie and other mediators through the Internet, such an agency may not be in the same city as the enterprise. The customer can use the special 3-D graphical interface of the Marketing Mediator to look at items of furniture from various directions under selected color light, and navigate and compare among different sets of furniture. He or she may request about a desired or customized set of furniture concerning performance, price, and manufacturing and delivery time etc. The customer may propose specio""cations for a particular set of furniture, and again ask for price and manufacturing time. The MetaMorph II system will reply to the customer's questions immediately if the information exists in its database system (the local database of the Marketing Mediator or the remote database of the Enterprise Mediator). If the informa- tion cannot be found in this way, MetaMorph II will start a product scheduling/rescheduling procedure as detailed in Shen et al. (1998a) for obtaining manufacturing time and an estimated price for this special product, and then send a message to the Marketing Mediator to answer the customer's request.
During the manufacturing scheduling process, the Resource Mediator may communicate and negotiate with the Material Mediator for raw material supply, and with other partners' Enterprise Mediators for parts fabrication.
If the customer is not satiso""ed with the quoted fabrication date, he or she may be allowed to request a special delivery date for the particular product. When MetaMorph II receives this request, it starts a rescheduling procedure to negotiate with other products in progress, to try to o""nd a possible solution for product delivery within the requested period, and then o""nally will send a message to the Marketing Mediator with a positive (or negative) response.
5. Concluding remarks Different agent-based architectures for manufacturing
systems have been proposed in the literature. In our experience, the autonomous agent architecture as used in DIDE is well suited for developing distributed intelligent design systems when existing engineering tools are encapsulated as agents, and the system consists of a small number of agents. The federation architecture with facilitators or mediators as used in MetaMorph I provides computational simplicity and manageability for more sophisticated systems. It is
quite suitable for developing distributed manufac- turing systems which are complex, dynamic, and composed of a large number of resource agents. This paper proposes a hybrid architecture combining the main features of the above architectures. Such a hybrid architecture has the following interesting features:
(1) Knowledge capitalization is realized primarily at the mediator level. When the system cono""guration changes, only the mediators need to be informed and updated, which obviously reduces communications among subsystems and agents;
(2) The bottleneck problem may be mitigated through direct communication among the agents located in different subsystems and coordinated by different mediators;
(3) It is relatively easy to integrate existing software and hardware, so as to resolve the legacy problem;
(4) Maintenance of the system is simpler because of its modular architecture;
(5) Remote services such as marketing and material supply can be interconnected through standard communication systems such as the Internet;
(6) The architecture should provide the system with considerable r'exibility and scalability.
However, adopting a hybrid architecture in a heterogeneous environment may increase the system complexity and make system implementation and integration more difo""cult. Also, some drawbacks of centralized systems, such as the bottleneck problem cannot be completely overcome in a mediator-centric organization.
The proposed MetaMorph II architecture has some similarities to the supply chain network architecture proposed by Barbuceanu and Fox (1997). However, the agents in their architecture are primarily autono- mous agents and there is more emphasis on developing a coordination language. The MetaMorph II architecture is primarily mediator- centric, and the objective of this current research project is to implement a practical agent-based manufacturing system using existing tools where possible. The hybrid architecture of MetaMorph II allows an agent in one subsystem to communicate directly with other subsystems or agents in other subsystems so as to mitigate the bottleneck problem.
Our research experience in developing agent-based technology for distributed intelligent design and
MetaMorph II 249 manufacturing suggests that agent technology will be widely used in developing next generation manufac- turing systems. However, the manufacturing industry involves very complex systems. So its adoption of agent technology may initially be in ``islands of automation''. To test the MetaMorph II architecture in a real world manufacturing application will require further research and development. A MetaMorph- based industry-oriented application was recently begun for furniture fabrication. Other collaborative projects with industrial partners are under considera- tion.
Notes 1. In MetaMorph I, the organization of resource commu-
nities occurs at two levels: macro and micro levels. Macro-level organization is static, based on knowledge about closely related agents that are distinct from others and can be physically separated. Such communities can be made of part, machine, tool, and transportation agents. Within a macro-level community, heterogeneous agents form several micro-level communities whose composi- tion is not static and therefore unknown. Such micro- level communities are determined by predominant behavioral characteristics (grouped in classes). These communities, for example, may contain primary or secondary process machine agents, design feature agents, and tool agents
References Balasubramanian, S., Maturana, F. and Norrie, D. (1996)
Multi-agent planning and coordination for distributed concurrent engineering. International Journal of Cooperative Information Systems, 5(2s'3), 153s'179. Barbuceanu, M. and Fox, M. (1997) Integrating
communicative action, conversations and decision theory to coordinate agents, Proceedings of Autonomous Agents 97, Marina del Rey, CA, ACM Press. BartheA's, J-P. (1994a) Developing integrated object environ-
ment for building large knowledge-based systems. Int Journal of Human-Computer Studies, 41, 33s'58. BartheA's, J. P. (1994b) La probleA^matique de reA^conciliation en
ingeA^nierie simultaneA^e, Organization de la conception, K. Zreik and B. Trousse, (eds). Europia, Paris, pp. 275s' 288. Brennan, R. W., Balasubramanian, S. and Norrie, D. H.
(1997) Dynamic control architecture for advanced
manufacturing systems, Proceedings of International Conference on Intelligent Systems for Advanced Manufacturing, Pittsburgh, PA. Brown, D., Dunskus, B., Grecu, D. and Berker, I. (1995)
SINE Support for single function agents, Proceedings of AIENG'95, Applications of AI in Engineering, Udine, Italy. Butler, J. and Ohtsubo, H. (1992) ADDYMS Architecture
for distributed dynamic manufacturing scheduling. In Artio""cial Intelligence Applications in Manufacturing, A. Famili, D. S. Nau and S. H. Kim (eds), AAAI Press, pp. 199s'214. Cutkosky, M. R., Engelmore, R. S., Fikes, R. E., Genesereth,
M. R., Gruber, T. R., Mark, W. S., Tenenbaum, J. M. and Weber, J. C. (1993) PACT: An experiment in integrating concurrent engineering systems. IEEE Computer, 26(1), 28s'37. Cutkosky, M. R., Tenenbaum, J. M. and Glicksman, J.
(1996) Madefast collaborative engineering over the internet. Communication of the ACM, 39(9), 78s'87. Finin, T., Fritzon, R., McKay, D. and McEntire, R. (1993)
KQMLDHA language and protocol for knowledge and information exchange Tech. Report, University of Maryland, Baltimore. Fischer, K., Muller, J. P., Pischel, M. and Schier, D. (1995)
A model for cooperative transportation scheduling, Proceedings of the First International Conference on Multi-Agent Systems, AAAI Press/MIT Press, San Francisco, CA, pp. 109s'116. Goldmann, S. (1996) Procura: a project management model
of concurrent planning and design, Proceeding of WET ICE'96, Stanford, CA. Maturana, F. and Norrie, F. (1996) Multi-agent mediator
architecture for distributed manufacturing. Journal of Intelligent Manufacturing, 7, 257s'270. McGuire, J., Kuokka, D., Weber, J., Tenenbaum, J., Gruber,
T. and Olsen, G. (1993) SHADE: technology for knowledge-based collaborative engineering. Journal of Concurrent Engineering Application and Research, 1(3). Norrie, D. and Gaines, B. (1996) Distributed agents systems
for intelligent manufacturing. Canadian Artio""cial Intelligence, 40, 31s'33. Pan, J. Y. C. and Tenenbaum, M. J. (1991) An intelligent
agent framework for enterprise integration. IEEE Transactions on Systems, Man, and Cybernetics, 21(6), 1391s'1408. Park, H., Tenenbaum, J. and Dove, R. (1993) Agile
infrastructure for manufacturing systems (AIMS): A vision for transforming the us manufacturing base, Defense Manufacturing Conference. Park, H., Cutkosky, M., Conru, A. and Lee, S. H . (1994) An
agent-based approach to concurrent cable harness design. AIEDAM, 8(1) 45s'62.
250 Shen, Maturana and Norrie Parunak, H. V. D., Baker, A. D. and Clark, S. J. (1997a) The
ARRIA agent architecture: an example of require- ments-driven agent-based system design, The First International Conference on Autonomous Agents, Marina del Rey, CA, ACM Press. Parunak, H. V. D., Ward, A., Fleischer, M. and Sauter, J.
(1997b) A marketplace of design agents for distributed concurrent set-based design, Proceedings of the Fourth ISPE International Conference on Concurrent Engineering Research and Applications, Troy, MI. Petrie, C., Cutkosky, M., Webster, T., Conru, A. and Park, H.
(1994) Next-Link: an experiment in coordination of distributed agents. Position paper for the AID-94 Workshop on Conr'ict Resolution, Lausanne. Roboam, M. and Fox, M. S. (1992) Enterprise management
network architecture, In Artio""cial Intelligence Applications in Manufacturing, A. Famili, D. S. Nau, S. H. Kim (eds). The AAAI Press, pp. 401s'432. Scalabrin, E., Vandenberghe, H., De Azevedo, H. and
BartheA's, J. P. (1996) A generic model of cognitive agent to develop open systems, Proceeding of 18th Brazilian Symposium on Artio""cial Intelligence, Curitiba, Brazil. Shen, W. and BartheA's, J. P. (1996a) Computer supported
cooperative environments for engineering design, Tech. Report 96s'122, CNRS URA 817 Heudiasyc, University of Technology of CompieA'gne, 96 pages. Shen, W. and BartheA's, J. P. (1996b) An experimental multi-
agent environment for engineering design. International Journal of Cooperative Information Systems, 5(2s'3), 131s'151.
Shen, W. and Norrie, D. H. (1998) Agent-based approaches
for intelligent manufacturing: A State-of-the-Art Survey, Proceedings of the 4th Australia Workshop on Distributed Artio""cial Intelligence, Brisbane, Australia, pp. 1s'29. Shen, W., Xue, D. and Norrie, D. H. (1998a) An agent-based
manufacturing enterprise infrastructure for distributed integrated intelligent manufacturing systems, Proceedings of the Third International Conference on the Practical Application of Intelligent Agents and Multi-Agents (PAAM'98), London, UK, pp. 533s'548. Shen, W., Maturana, F and Norrie, D. H. (1998b) Learning
in agent-based manufacturing systems, Proceedings of the AAAI's Special Interest Group in Manufacturing Workshop on Artio""cial Intelligence and Manufacturing, The AAAI Press, Albuquerque, NM, pp. 177s'183. Swaminathan, J. M., Smith, S. F. and Sadeh, N. M. (1996) A
multi-agent framework for modeling supply chain dynamics, Proceedings of NSF Research Planning Workshop on Artio""cial Intelligence and Manufacturing, Albuquerque, NM. Toye, G., Cutkosky, M., Leifer, L., Tenenbaum, J. and
Glicksman, J. (1993) SHARE: A methodology and environment for collaborative product development, Proceeding of 2nd Workshop on Enabling Technologies Infrastructure for Collaborative Enterprises, IEEE Computer Press. Wooldridge, M. and Jennings, N. R. (1995) Intelligent
agents: Theory and Practice. Knowledge Engineering Review, 10(2), 115s'152.
MetaMorph II 251
Copyright (c) 2000 Kluwer Academic Publishers