CHANG-OUK KIM [1,*]

SHIMON Y. NOF [2]

Together with the development of computer network technology, the automatic collaboration of CIM data activities is becoming an important research issue. Two models are proposed for the research issue: a data activity flow network model called DAF-Net for coordinating interdependent data activities and an agent-based integration model of information systems called AIMIS for integrating distributed and heterogeneous information systems on demand. Additionally, this paper addresses the key design factors of the DAF-Net and also the AIMIS: the structural type of the DAF-Net and the agent interaction protocol of the AIMIS. Statistical methods and PVM (Parallel Virtual Machine) software are employed for the design of the DAF-Net and the AIMIS.

1. Introduction

Computer Integrated Manufacturing (CIM) is the ostensible evolutionary outcome of computed-aided design, computer-aided manufacturing, computer-aided production and business management. To operate the functional units in a factory as a cohesive unit, automation technology must be applied to both material processing activities (machining, material handling, warehousing, etc.) and data processing activities (inventory transaction, file transfer, alarm message by e-mail, CAD transaction, generation of material requirement plan, etc.).

Over the past 20 years, research on the automation of material processing activity has been extensively exploited (Bedworth et al., 1991; Chang et al., 1991; Viswanadham and Narahari, 1992). Research concerning the automation of data processing activity has also been studied for the past 10 years. However, previous research on data processing automation has focused on either static integration or relatively simple coordination of CIM databases (Libes and Barkmeyer, 1988; Vernadat, 1988; Weber, 1992; Veeramani et al., 1993; Hsu et al., 1994; Urban et al., 1994).

The information structures of most CIM enterprises characterized by: (i) heterogeneity of information; (ii) distribution of information systems; (iii) ad-hoc installation of new information systems; and (iv) many interdependent data processing activities. More specifically, even in a CIM enterprise, heterogeneous information systems, such as databases, files, application software, e-mail, and human experts, are distributed. Additionally, new heterogeneous information systems are installed in the CIM enterprise as needed because of technological and financial limits (island of automation). Furthermore, due to frequent changes in market conditions, product specifications, and manufacturing processes, today's CIM enterprises are evolving toward a more agile framework for which interdependent data processing activities should be automatically controlled.

This demand is driving CIM enterprises to establish an automatic, flexible collaboration framework for data processing activities. In this framework, new information systems are integrated by demand, upon which interdependent data processing activities in response to an event of interest should be automatically coordinated. Since numerous events occur in a CIM enterprise, such automatic coordination of data processing activities would be a complex task.

There exist two dependency constraints imposed on the interdependent data processing activities: (i) procedural dependency; and (ii) data dependency (Khanna et al., 1993; Kim, 1996). For instance, rescheduling activity after a machine breakdown is allowed only after the shop-floor manager is alerted (procedural dependency), and input data for the rescheduling activity should be retrieved from several information systems (data dependency) before the activity is processed.

Considering the above requirements, it is essential to develop: (i) a flexible information integration model of heterogeneous information systems; and (ii) an automatic coordination model of interdependent data processing activities. In our previous research (Kim, 1996), DAF-Net (Data Activity Flow-Net) and AIMIS (Agent-based Integration Model of Information Systems) were proposed for the automatic coordination model and the flexible integration model, respectively.

The DAF-Net is an active coordination model of interdependent data processing activities. Upon the occurrence of events, pre-defined DAF-Nets are automatically triggered that coordinate interdependent data activities. Hence, it can be considered as an advanced workflow model (Georgakopoulos et al., 1995; Khoshafian and Buckiewicz, 1995; Sheth, 1995; Cichocki et al., 1998). Informally, the DAF-Net is defined as a set of interdependent data processing activities called component data activities, together with an execution order imposed among them. The execution order is specified by a set of irreflexive and transitive precedence relations. The AIMIS is a multi-agent-based integration model of heterogeneous and distributed information systems. It consists of GCAs (Global Coordination Agents), LCAs (Local Coordination Agents), and an agent interaction protocol called the allocation protocol of component data activities. The AIMIS is designed to add or release different information systems as needed, thereby achi eving the concept of integration by demand.

The DAF-Net and AIMIS are conceptually different from contemporary enterprise integration frameworks, which can be defined as a collection of modeling principles, methods, or tools relevant to developing models for various aspects of an enterprise (Vernadat, 1996). Common modeling aspects of an enterprise integration framework are function flows, information flows, business processes, resource structures, and organization structures.

Representative enterprise integration frameworks include CIM-OSA (Jorysz and Vernadat, 1996a,b), PERA (Williams, 1994), GRAI (Doumeingts et al., 1987), ARIS (Scheer, 1999), IDEF (Anon, 2000), and TOVE (Fox and Gruninger, 1998). They describe the particular or integrated aspect of an enterprise expressed in terms of some formalism (or language) defined by their own modeling constructs. Therefore, they can be considered as enterprise modeling methods. Whereas, the DAF-Net and AIMIS are a multi-agent-based workflow execution framework on which business process models derived by the enterprise integration frameworks can be realized.

The objective of this research is to experimentally determine the key design factors of the DAF-Net and AIMIS: the structural type of the DAF-Net and the allocation protocol of component data activity. In Section 2, the DAF-Net and AIMIS are briefly introduced together with the design factors. Section 3 presents a statistical experiment by which the design factors are decided. Finally, the conclusion of this research and relevant future research work are described in Section 4.

2. DAF-Net and AIMIS

2.1. The DAF-Net

The DAF-Net is defined for each event [e.sub.i] (i = 1, 2,...,L) where L is the number of events. It is constructed to achieve a pre-defined goal which compensates the event. Therefore, the relation between events and DAF-Nets is one-to-one mapping. Formally, a DAF-Net is defined as follows:

DAF-Net, [D.sub.i] (i = 1, 2,...,L), is a coordination model of interdependent data processing activities and is defined as a five-tuple:

= [D.sub.i] = ([C.sub.i], [TR.sub.i], [IN.sub.i], [OUT.sub.i], [MR.sub.i]).

The elements of a DAF-Net are defined as follows:

1. Set [C.sub.i] of component data activities. Each component data activity [c.sub.ij] (j = 1, 2,..., \[C.sub.i]\) is restricted to accessing at most one information system. Typical examples are database transaction, file copy, file transfer, e-mail, the execution of an application program, internal computation, and condition evaluation.

2. Set [TR.sub.i] of transitions. Transitions represent the precedence relations among the component data activities. There are two types of transitions: immediate transitions and delayed transitions.

3. Input function [IN.sub.i] defines directed arcs from [C.sub.i] to [TR.sub.i], and the output function [OUT.sub.i] defines directed arcs from [TR.sub.i] to [C.sub.i].

4. Marking [MR.sub.i] of DAF-Net [D.sub.i] is a 1 x \[C.sub.i]vector. Each element represents the execution status of component data activity [c.sub.ij] of DAF-Net [D.sub.i] and can be in one of three states: N for not being executed, E for being executed, and C for commitment (successfully completed activity).

Structurally the DAF-Net model follows the Petri net model. However, compared with traditional applications of a Petri net such as performance evaluation and the control of a manufacturing cell (Viswanadham and Narahari, 1992), a DAF-Net coordinates a set of interdependent CIM data processing activities that correspond to an event in real-time mode. In other words, in the traditional applications of a Petri net, each place denotes a certain state of a given system. But, in a DAF-Net, each place represents a component data activity which should be delegated to an information system for its execution. Figure I shows a DAF-Net for the change of CAD and Bill-Of-Material (BOM) data.

2.2. Design factor of DAF-Net: structural type of DAF-Net

One important characteristic of the DAF-Net is that it is designed to be automatically triggered whenever a related event occurs. Based on this characteristic, two methods for executing a given DAF-Net are developed.

One is to execute the DAF-Net as given, and the other is to decompose the DAF-Net into several small ones, such that each of them is triggered sequentially upon the completion of its precedent one. The former is defined here as a composite DAF-Net and the latter is defined as a decomposite DAF-Net. For example, the DAF-Net [D.sub.i] in Fig. 1 can be decomposed into three small DAF-Nets (DAF-Net [D.sub.i1], DAF-Net [D.sub.i2], DAF-Net [D.sub.i3]). They can be separately triggered whenever their precedent ones are completed. Under unreliable information systems, the structural type of a DAF-Net would influence the number of committed (successfully completed) DAF-Nets. This conjecture is verified in Section 3.

2.3. Agent-based integration model of information systems

The architecture of the AIMIS is based on the following three functional modules: (i) the functions GCAs; (ii) the functions of LCAs; and (iii) an allocation protocol of component data activity. Figure 2 shows the organizational view of agents in the AIMIS model. GCAs are centrally located in a computer. They are responsible for triggering the DAF-Nets and executing them concurrently on heterogeneous information systems. Three types of GCAs are employed in the AIMIS architecture: (i) DAF-Net triggering agent [G.sub.T] for retrieving [D.sub.i] from the DAF-Net library in response to event [e.sub.i] and for creating a DAF-Net execution agent [[g.sup.E].sub.i]; (ii) set [G.sub.E] = {[[g.sup.E].sub.1], [[g.sup.E].sub.2], ..., [[g.sup.E].sub.n]} of DAF-Net execution agents each of which is responsible for controlling the execution order of component data activities of a DAF-Net and dispatching them to corresponding LCAs (n is the number of DAF-Nets being executed); and (iii) DAF-Net coordination agent [G.sub.C] for controlling the execution order of DAF-Nets concurrently being executed.

Relative to GCAs that are centrally located in a computer, LCA [L.sub.j] (j = 1,2,..., m) is distributed on an information system (m is the number of information systems). Its role is to interface between GCAs and local information systems. Specifically, three main functions are performed by LCAs. First, each LCA [L.sub.j] monitors every local data processing activity coming into an information system j. As soon as [L.sub.j] detects an event, it notifies the DAF-Net triggering agent of the occurrence of the event. Second, in order to select an LCA which can process a component data activity in terms of data availability of its information system, every LCA [L.sub.j] cooperates with the DAF-Net execution agents by using the allocation protocol of component data activity. Third, to execute a component data activity transmitted by a DAF-Net execution agent, the selected LCA, say [L.sub.j] is responsible for submitting the component data activity to the information system.

The operational paradigm of AIMIS requires a number of interactions between GCAs and LCAs for executing DAF-Nets concurrently. From the viewpoint of the DAF-Net execution agents, two types of interactions are required: (i) searching LCAs which can execute given component data activities; and (ii) transferring component data activities to the selected LCAs and receiving the execution results of the component data activities from the LCAs. In this research, the first interaction is called the allocation protocol of component data activity. The protocol is analogous to a task allocation protocol in distributed agent systems.

2.4. Design factor of AIMIS: allocation protocol of component data activity

If information systems are unreliable and many duplicates of data are spread over several information systems, searching LCAs that can access input/output data of component data activities becomes a critical factor for successfully completing the DAF-Net. Such effort may increase the number of messages passed between DAF-Net execution agents and LCAs, resulting in heavy network traffic. Therefore, the allocation protocol of data activity significantly affects the throughput of the DAF-Net. In this research, two different protocols are proposed: Asynchronous Request and Response protocol (ARR protocol) and Negotiation with Case-based Learning protocol (NCL protocol). The first protocol is designed to minimize the number of messages passed, while the second protocol is designed to be resilient to the failures of information systems. Each protocol is explained as follows.

2.4.1. The asynchronous request and response protocol

The ARR protocol is an efficient communication protocol in a client-server model because of the small number of messages passed between DAF-Net execution agents and LCAs for completing DAF-Net. The logic of the ARR protocol is as follows:

Step 1. Using the address of the receiver (LCA), the field of component data activity specification, [[g.sup.E].sub.i] sends the request to the LCA and waits for a response (acknowledgment) from the LCA.

Step 2. If the LCA scans its request queue and finds the arrival of the request, then it sends [[g.sup.E].sub.i] an acknowledgment signal.

The main advantage of the ARR protocol is that only one round-trip message is required to find the location of the LCA. The major drawbacks of the protocol are that: (i) the address of the receiver must be specified for every component data activity; and (ii) there is a chance that [[g.sup.E].sub.i] infinitely waits to receive an acknowledgment from an LCA that has broken down. Figure 3 depicts the operational logic of the ARR protocol.

2.4.2. Negotiation with the case-based learning protocol

The NCL protocol relaxes the assumption of a priori knowledge about the addresses of the LCAs. Instead, this protocol adopts a negotiation mechanism between [[g.sup.E].sub.i] and the LCAs to find the most appropriate ones that can execute component data activities efficiently.

One remarkable advantage of the NCL protocol is its learning ability. After each negotiation, [[g.sup.E].sub.i] accumulates the processing capability of each LCA in the LCA capability table. That is, each negotiation makes [[g.sup.E].sub.i] learn which LCA stores what type of data. The LCA capability table is a triple (LCAidentifier, LCAaddress, list of data objects) where LCAidentifier is a unique identifier of LCA, LCAaddress is the physical address of the LCA on the computer network, and list of data objects is the list of data objects that are accessible by the LCA.

As the interaction between [[g.sup.E].sub.i] and the LCAs evolves, [[g.sup.E].sub.i] gains accurate knowledge of the LCAs' capability and can send limited broadcasting messages, which will reduce the communication traffic between [[g.sup.E].sub.i] and the LCAs. The main disadvantage of the NCL protocol is that it can generate too many communication messages when the data stored in each information system dynamically change. This protocol can be explained procedurally as follows:

Step 1. [[g.sup.E].sub.i] searches the LCA capability table to find LCAs that are able to execute a given component data activity. The selected LCAs are called candidate LCAs. If only one candidate LCA exists, go to Step 5. If no candidate LCA exists, then go to Step 2b.

Step 2a. [[g.sup.E].sub.i] broadcasts a component data activity to all candidate LCAs and waits for their response. Go to Step 3.

Step 2b. [[g.sup.E].sub.i] broadcasts a component data activity to all LCAs and waits for their responses.

Step 3. Each LCA makes a bid. It consists of: (i) the possibility of the execution of the component data activity based on the availability of data required for the execution; and (ii) the number of component data activities waiting in its request queue.

Step 4. LCAs send their bids to [[g.sup.E].sub.i].

Step 5. [[g.sup.E].sub.i] selects a winning LCA that can not only execute the component data activity but also has the smallest request queue. In addition, [[g.sup.E].sub.i] updates the processing capability of the winning LCA.

Step 6. After the negotiation procedure, [[g.sup.E].sub.i] sends an awarding message to the winning LCA. It also sends a revoking message to other candidate LCAs to cancel the execution of the data component activity.

Figure 4 graphically illustrates the operational logic of the NCL protocol.

3. Experiment

3.1. The objective

In general, the high completion rate of distributed tasks with a limited or moderate amount of message passing among distributed agents would be a main design goal for the successful applications of multi-agent models. This goal is also applied to the DAF-Net and the AIMIS. Therefore, it is necessary to decide the optimal level of the design factors of the DAF-Net and the AIMIS. As explained in the previous section, the design factors are: (i) the structural type of the DAF-Net (composite versus decomposite); and (ii) the allocation protocol of the component data activity (ARR protocol versus NCL protocol). Throughout this experiment, the two design factors are varied under various information environments.

3.2. Experimental scenario

For this experiment, a virtual Manufacturing Resource Planning (MRPII) scenario was chosen. From the viewpoint of information processing, the core function of MRPII is the interoperability of heterogeneous local information systems. Thus, a typical MRPII system must be supported by an integrated information framework. Within this integrated information framework, interdependent data processing activities associated with the MRPII system should be coordinated automatically in response to the events of interest. In this experiment, the AIMIS model is selected for integrating the local information systems, and the DAF-Net is applied to model the coordination of the interdependent data processing activities.

Figure 5 shows the MRPII scenario. It consists of five departments. Two heterogeneous information systems and a manager are assumed to be associated with each department, except for the case of the business management department, where only one information system is assumed. All LCAs are connected to GCAs, and each LCA is linked to each information system or to the manager. In addition, a shop-floor module is set up to create external events randomly. In the same way, each information system generates internal events randomly. Eight event types are chosen as follows. Corresponding to each type, a compensating DAF-Net is constructed.

a. Inventory update.

b. Change of CAD and BOM data.

c. Change of cost and lot-size data.

d. Change of lead-time.

e. Shop-floor order completion.

f. Machine breakdown.

g. Due-date change of scheduled receipt.

h. Daily production progress report to production manager.

3.3. Software tool for the experiment

The PVM software of Geist et al. (1994) and the C Programming language are used to simulate the DAF-Net and the AIMIS under the MRPII scenario. PYM was developed to facilitate distributed computing on a computer network. It is an appropriate tool to stimulate the interactions among the agents in the AIMIS. For this purpose, PVM supports a variety of communication functions such as asynchronous and synchronous message passing mechanisms. These communication functions are implemented by the address keeping method of PVM. That is, each process which emulates an agent (GCA or LCA) stores the logical addresses of other processors. Every agent can communicate with the others by referring to their logical addresses. The physical communications are undertaken by PVM.

To use the PVM software for the MRPII simulation, 33 processes were run simultaneously on six Unix-based workstations that were interconnected in the Engineering Computer Network at Purdue University. Each process is an independent program. It emulates the behavior of either an information system, a user (manager), or an agent (GCA or LCA) of the AIMIS. Some processes are loaded on the same workstation (shown as a box) due to the limited number of workstations available.

3.4. Performance measures

Three performance measures are defined: (i) the number of messages passed; (ii) the number of committed DAF-Nets; and (iii) the DAF-Net processing quality. The significance of each performance measure is explained as follows.

3.4.1. Number of messages passed ([N.sub.m])

The total number of messages passed between GCAs and LCAs is counted. So as to act as a measure the effectiveness of the allocation protocol of component data activity and also of the structural type of the DAF-Net. For this part of the experiment to be meaningful, it is assumed that the information systems are reliable. This assumption is reasonable because the number of messages per DAF-Net is only significant when the DAF-Net is successfully completed.

3.4.2. Number of committed DAF-Nets ([N.sub.c])

The number of committed DAF-Nets, which is equivalent to the number successfully completed, is counted as the failure rates of information systems vary.

3.4.3. DAF-Net processing quality (PQ)

DAF-Net processing quality is defined here as the sum of the number of messages passed divided by the number of committed DAF-Nets. It is a measure to determine which pair of the allocation protocol of component data activity and the structural type of the DAF-Net is the most efficient in terms of the average number of messages passed per completed DAF-Net, Since a small number of messages passed per DAF-Net is more efficient than a large number of messages, a low value of the DAF-Net processing quality is more efficient than a high value.

3.5. Experimental design

This experiment consists of three sub-experiments: (i) an experiment on the efficiency; (ii) an experiment on the failure resilience; and (iii) an experiment on the DAF-Net processing quality. For each sub-experiment, an ANOVA test and a Student--Newman--Keuls (SNK) test are conducted to statistically analyze the experimental results (Hicks, 1993). Table 1 shows a summary of the experimental design. In this table, the data replication factor is the number of data duplications. Two levels of data replication factor are chosen: one and two. For example, if the data replication factor of a BOM data set is set to two, then two copies of the same BOM data are stored in two different information systems. Each sub-experiment is illustrated in detail as follows.

3.5.1. The experiment on the efficiency

The experiment on the efficiency is conducted to reveal: (i) whether the allocation protocol of component data activity affects the number of messages passed; (ii) whether there is an interaction effect between the allocation protocol of component data activity and the number of events; and (iii) which of the two allocation protocols of component data activity is efficient in terms of messages passed. By applying the NCL protocol, it is expected that the DAF-Net execution agents accumulate sufficient knowledge about the processing capability of LCAs as the number of event increases. In this sub-experiment, the information systems are assumed to be reliable and the data replication factor is fixed at one. Also, the time interval between two events is assumed to follow an exponential distribution with a mean of 2.2 mm.

3.5.2. The experiment on the failure resilience

The purpose of this sub-experiment is to identify: (i) whether the allocation protocol of component data activity affects the number of committed DAF-Nets; (ii) whether the structural type of the DAF-Net affects the number of committed DAF-Nets; and (iii) which pair of the two allocation protocols of component data activity and the structural type of the DAF-Net is more resilient to information system failure.

3.5.3. The experiment on the DAF-Net processing quality

The purpose of this sub-experiment is to identify: (i) whether the allocation protocol of component data activity affects the DAF-Net processing quality; (ii) whether the structural type of the DAF-Net affects the DAF-Net processing quality; and (iii) which pair of the two allocation protocols of component data activity and the structural type of the DAF-Net is effective in terms of the processing quality of the DAF-Net. In this sub-experiment, the information systems are assumed to be unreliable and the data replication factor varies from one to two. The formal statements of the research hypotheses are provided in Table 2.

3.6. Results and analysis

To accurately estimate the mean of each treatment, five simulation outputs are collected by varying the random number seed. For each sub-experiment, an ANOVA test and a SNK test are performed using SAS software.

3.6.1. The experiment on the efficiency

The experimental results for this sub-experiment are presented in Table 3. The experimental result of the ANOVA test are listed in Table 4, where the effects of both the allocation protocol of component data activity, and of the number of events are significant for the number of messages passed. However, it is observed that there is no interaction effect between the allocation protocol of component data activity and the number of events.

The result of the SNK test is shown in Table 5. The upper part of the table shows that the ARR protocol is better than the NCL protocol in terms of the number of messages passed. However, the lower part of the table shows an interesting observation in that the ARR protocol is superior to the NCL protocol until the number of events reaches 100, but thereafter, there is no significant difference between the two protocols. This result can be explained by reference the nature of the NCL protocol. Since the NCL protocol has the ability to learn the location of necessary data, and this capability is enhanced as the number of events increases, the NCL protocol reaches a point in time at which it has as much knowledge on the data location as does the ARR protocol. From that point, the learning capability of the NCL protocol led it to reduce the number of negotiations. From this sub-experiment, it can be inferred that intelligent protocols that are able to learn from past experience, such as the NCL protocol, are val uable when the integrated system is operated for a longer time period.

3.6.2. The experiment on the failure resilience

Table 6 shows the results of this sub-experiment. The observed data of each treatment are the number of committed DAF-Nets. Table 7 presents the ANOVA results where the single effects significantly influence the performance measure at a 0.05 significance level. That is, the allocation protocol of component data activity and the structural types of the DAF-Net significantly influence the number of committed DAF-Nets.

An SNK test is conducted to judge the optimal pair of the allocation protocol of component data activity and the structural type of the DAF-Net. Its results arc shown in Table 8. From this table, it is observed that (NCL, Decomposite) is superior to the other pairs. In particular, (ARR, Composite) turns out to be the worst pair.

3.6.3. The experiment on the DAF-Net processing quality

The experimental results shown in Table 9 are obtained after 200 events are generated, so that the NCL protocol establishes sufficient knowledge about the data location. From the results of the ANOVA test, as shown Table 10, it is observed that at a 0.05 significance level the single effects significantly influence the DAF-Net processing quality.

Table 11 shows the results of the SNK ranking test. The results indicates that (NCL, Decomposite) is better than the other combinations. This result implies that (NCL, Decomposite) guarantees a better DAF-Net processing quality under the assumption that the information systems are subject to failure.

In summary, the experimental results suggest the following findings:

(1) From the experiment on the efficiency, it is observed that the ARR protocol is efficient in terms of the number of messages passed for a short operation interval of the system. Whereas the NCL protocol becomes as efficient as the ARR protocol when the system is operated for a time long enough for it to accumulate sufficient knowledge about the data locations. Therefore, it is suggested that if the configuration of the integrated system changes frequently, such as in a CIM information system, the application of the NCL protocol is better than the ARR protocol. The reason for this is that the NCL protocol can automatically detect changes of data location.

(2) From the experiments on the failure resilience and DAF-Net processing quality, it is verified that the (NCL, Decomposite) pair is the best choice when the information systems are unreliable. The (NCL, Decomposite) pair guarantees not only the maximum number of committed DAF-Nets but also the highest DAF-Net processing quality.

4. Conclusion

For CIM enterprises, the automatic coordination of interdependent data activities and the integration of heterogeneous information systems are essential to survive in the competitive world-market. In this research, a DAF-Net and an AIMIS are briefly introduced as an automatic coordination model of interdependent data processing activities and an integration model of heterogeneous CIM information systems. The main objective of this research is to decide the key design factors of the DAF-Net and the AIMIS -- the structural type of the DAF-Net and the allocation protocol of component data activity. Statistical testing methods (ANOVA and SNK test) are applied to design the factors under various CIM information environments. The test results indicate that an intelligent protocol with decomposite coordination of component data activities is effective when the information systems are unreliable.

For a future research topic, it is necessary to develop a distributed simulation tool which can measure time-related performance measures, such as the mean completion time of the DAF-Net.

(1.) Department of Industrial Engineering, Myong Ji University, Yongin, Kyungki-Do, 449-728, Korea E-mail: kimco@wh.mju.ac.kr

(2.) School of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA E-mail: nof@ecn.purdue.edu

(*.) Corresponding author

Biographies

Dr. Chang-Ouk Kim is an Assistant Professor in the Industrial Engineering Department at Myong Ji University in Korea. He received his B.S. and M.S. degrees in Industrial Engineering from Korea University, and a Ph.D. degree in Industrial Engineering from Purdue University. He has published several research papers in refereed journals including International Journal of Production Research, International Journal of Computer Integrated Manufacturing, and Journal of Object-Oriented Programming. His current research interests are concerned with the design of distributed information systems for CIM, development of object-oriented software, and intelligent production control.

Dr. Shimon Y. Nof is a Professor of Industrial Engineering at Purdue University. His scholarly and consulting activities are focused on the areas of computer integrated manufacturing, applications of information technologies, and industrial robotics. He is the Director of the NSF-industry supported PRISM Program (Production Robotics and Integration Software for Mfg.) and co-director of Purdue's ITS (Intelligent Transportation Systems) Program. In addition, he is a Fellow of IIE, Secretary General of IFPR, and a member of ACM, IFIP, IFAC, and SME. Dr. Nof has published over 200 articles on production engineering and information technology, and is the author/editor of seven books. His current research focus is on computer-supported integration and collaboration of distributed work.

Contributed by the Engineering Design Department

References

Anon (2000) White papers on IDEF series. KBSI, Available at http.//idef.com.

Bedworth, D.D., Henderson, M.R. and Wolfe, P.M. (1991) Computer-Integrated Design and Manufacturing, McGraw-Hill.

Chang, T.C. and Wysk, R.A. (1991) Computer-aided Manufacturing, Prentice-Hall, Englewood Cliffs, NJ.

Cichocki, A., Helal, A., Rusinkiewicz, M. and Woelk, D. (1998) Workflow and Process Automation. Concepts and Technology, Kluwer, Norwell, MA.

Doumeingts, G., Vallespir, B., Darricar, D. and Roboam, M. (1987) Design methodology for advanced manufacturing systems. Computers in Industry, 9, 271-296.

Fox, M.S. and Gruninger, M. (1998) Enterprise modeling. AI Magazine, Fall, 109-121.

Geist, A., Beguelin, A., Dongarra, J., Jiang, W., Manchek, R. and Sunderam, V. (1994) PVM: Parallel Virtual Machine. A User's Guide and Tutorial for Networked Parallel Computing, MIT Press, Cambridge, MA.

Georgakopoulos, D., Hornick, M. and Sheth, A. (1995) An overview of workflow management: from process modeling to workflow automation infrastructure. Distributed and Parallel Databases, 3, 119-153.

Hicks, C.R. (1993) Fundamental Concepts in the Design of Experiments, Oxford University Press, Oxford, UK.

Hsu, C., Gerhardt, L., Spooner, D. and Rubenstein, A. (1994) Adaptive integrated manufacturing enterprises: information technology for the next decade. IEEE Transactions on Systems, Man and Cybernetics, 24, 828-837.

Jorysz, H.R. and Vernadat, F.B. (1990a) CIM-OSA part 1: total enterprise modelling and function view. International Journal of Computer Integrated Manufacturing, 3, 144-156.

Jorysz, H.R. and Vernadat, F.B. (1990b) CIM-OSA part 2: information view. International Journal of Computer Integrated Manufacturing, 3, 157-167.

Khanna, N., Fortcs, J.A.B. and Nof, S.Y. (1994) A formalism to structure and parallelize the integration of cooperative engineering design tasks. Research Memorandum No. 94-37, School of Industrial Engineering, Purdue University, West Lafayette, IN 47907.

Khoshafian, S. and Buckiewicz, M. (1995) Introduction to Groupware, Workflow and Workgroup Computing, John Wiley and Sons, New York.

Kim, C.-O. (1996) DAF-net and multi-agent based integration approach for heterogeneous CIM information systems. Unpublished Ph.D. thesis, Purdue University, West Lafayette, IN 47907.

Libes, D. and Barkmeyer, E.D. (1988) The integrated manufacturing data administration system (IMDAS) -- overview. International Journal of Computer Integrated manufacturing, 1, 44-49.

Scheer, A.W. (1999) ARIS -- Business Process Modeling, Springer-Verlag, New York.

Sheth, A. (1995) Workflow automation: applications, technology and research: tutorial notes, in SIGMOD Conference, ACM Press.

Urban, S.D., Shah, J.J., Rogers, M., Jeon, D.K., Ravi, P. and Bliznakov, P.A. (1994) Heterogeneous, active database architecture for engineering data management. International Journal of Computer Integrated Manufacturing, 7, 276-293.

Veeramani, D., Bhargava, B. and Barash, M.M. (1993) Information system architecture for heterarchical control of large FMSs. Computer Integrated Manufacturing Systems, 6, 76-92.

Vernadat, F.B. (1988) Artificial intelligence in CIM databases, in Artificial Intelligence: Implications for CIM, Kusiak, A. (ed.), Springer-Verlag, New York.

Vernadat, F.B. (1996) Enterprise Modeling and Integration. Principles and Application, Chapman and Hall, London, UK.

Viswanadham, N. and Narahari, Y. (1992) Performance Modeling of Automated Manufacturing Systems, Prentice Hall, Englewood Cliffs, NJ.

Weber, D.M. (2992) Information architecture for an advanced integrated manufacturing enterprise. Unpublished Ph.D. thesis, Purdue University, West Lafayette, IN 47907.

Williams, T.J. (1994) The Purdue enterprise reference architecture, Computers in Industry, 24, 141-158.

                   Summary of the experimental design
                         of DAF-Net processing
Sub-experiment      Independent variables       Level
Efficiency          a. allocation protocol of   ARR, NCL
                       component data activity
                    b. number of events         50,100,150,200
                       generated during
                       simulation
Failure resilience  a. allocation protocol of   ARR, NCL
                       component data activity
                    b. Poisson failure rate of  0.08, 0.16
                       information system
                       (no./min)
                    c. structural types of      composite, decomposite
                       DAF-Net
                    d. Data replication factor  1, 2
Processing quality  a. allocation protocol of   ARR, NCL
                       component data activity
                    b. Poisson failure rate of  0.00, 0.08, 0.16
                       information system
                       (no./min)
                    c. structural types of      composite, decomposite
                       DAF-Net
Sub-experiment      Dependent variables
                    (performance measures)
Efficiency          number of messages
                      passed ([N.sub.m])
Failure resilience  number of commited
                      DAF-Nets ([N.sub.c])
Processing quality  DAF-Net processing
                      quality (PQ)
                       Hypotheses of the DAF-Net
                         processing experiment
Sub-experiment      Research hypotheses
Efficiency          hypothesis 1A: for any given allocation
                     protocol of component data activity,
                     there is no significant impact on the
                     number of message passed
                    hypothesis 1B: there is no interaction
                     effect between the allocation protocol
                     of component data activity and the
                     number of events on the number of
                     message passed
Failure resilience  hypothesis 2A: for any given allocation
                     protocol of component data activity,
                     there is no significant impact on the
                     number of committed DAF-Nets
                    hypothesis 2B: for any given structural
                     type of the DAF-Net, there is no
                     significant impact on the number of
                     committed DAF-Nets
                    hypothesis 2C: for any given combination
                     of the allocation protocol and the
                     structural type of the DAF-Net, there
                     is no significant impact on the number
                     of committed DAF-Nets
Processing quality  hypothesis 3A: for any given allocation
                     protocol of component data activity,
                     there is no significant impact on DAF-
                     Net processing quality under an
                     unreliable information system
                    hypothesis 3B: for any given structural
                     type of the DAF-Net, there is no
                     significant impact on DAF-Net
                     processing quality under an unreliable
                     information system
                    hypothesis 3C: for any given combination
                     of the allocation protocol and the
                     structural type of the DAF-Net, there
                     is no significant impact on the DAF-Net
                     processing quality
                Experimental results for the efficiency
                 experiment - number of messages passed
                              ([N.sub.m])
Allocation protocol of component
data activity
Number of events
50                                 100   150   200
ARR
486                               1023  1586  2081
518                               1047  1612  2183
527                                902  1487  2098
511                               1027  1671  2085
508                                984  1544  1906
NCL
758                               1074  1665  2102
611                               1141  1638  2216
685                               1118  1561  2106
613                               1171  1688  2097
584                               1122  1590  1925
                    ANOVA results for the efficiency
                               experiment
Source of variation        Degree of freedom  Sum of squares
                                 (df)              (SS)
Allocation protocol of             1               70 476.0
  component data activity
Number of events                   3           12 722 939.7
Interaction                        3               26 730.3
Error                             32              153 996.0
Source of variation        Mean square  Computed
                              (MS)      F-value
Allocation protocol of       70 476.0    14.64 [*]
  component data activity
Number of events           4240 979.9   881.27 [*]
Interaction                    8910.1     1.85
Error                          4812.4
(*.)denotes that the result is significant
to a significance level of 0.05.
                  SNK test results for the efficiency
                 experiment -- number of messages passed
                              ([N.sub.m])
Effect                      Ranking      Mean
Allocation protocol of      NCL         1373.25
  component data activity   ARR         1289.30
Combination of the          (NCL, 200)  2089.20 [I]
  allocation protocol of    (ARR, 200)  2070.60 [I]
  component data activity   (NCL, 150)  1628.40 [I]
  and the number of events  (ARR, 150)  1580.00 [I]
                            (NCL, 100)  1125.20
                            (ARR, 100)   996.60
                            (NCL, 50)    650.20
                            (ARR, 50)    510.00
Significance level = 0.05.
Bar (I.)indicates that values within the
bar are not statistically different.
                  Experimental results for the failure
                   resilience experiment -- number of
                     committed DAF-Nets (N.sub.c])
Structural type of the
DAF-Net
Composite               data replication factor  1
                                                 2
Decomposite             data replication factor  1
                                                 2
Structural type of the  Allocation protocol of component data
DAF-Net                               activity
                                    Failure rate
                                         ARR
                                        0.08                   0.16
Composite                                176                    132
                                         171                    153
                                         182                    138
                                         164                    142
                                         152                    139
                                         174                    148
                                         181                    144
                                         166                    158
                                         172                    139
                                         158                    146
Decomposite                              182                    163
                                         181                    158
                                         192                    168
                                         185                    189
                                         187                    171
                                         181                    168
                                         182                    171
                                         188                    166
                                         178                    172
                                         160                    165
Structural type of the
DAF-Net
                         NCL
                        0.08  0.16
Composite                175   142
                         174   138
                         157   149
                         184   142
                         178   142
                         183   171
                         188   168
                         193   175
                         176   178
                         184   170
Decomposite              189   175
                         183   163
                         201   172
                         187   183
                         195   187
                         189   183
                         199   193
                         211   188
                         218   182
                         206   195
                     ANOVA results for the failure
                         resilience experiment
Source of variation                     Degree of freedom
                                              (df)
Allocation protocol of component                1
  data activity
Structural type of the DAF-Net                  1
Failure rate                                    1
Data replication factor                         1
Combination of the allocation protocol          3
  of component data activity and
  structural type of the DAF-Net
Error                                          64
Source of variation                     Swn of squares  Mean square
                                             (SS)          (MS)
Allocation protocol of component           3380.00        3380.00
  data activity
Structural type of the DAF-Net             8080.20        8080.20
Failure rate                               7144.20        7144.20
Data replication factor                    1344.80        1344.80
Combination of the allocation protocol   11 499.40        3833.13
  of component data activity and
  structural type of the DAF-Net
Error                                      4298.40          67.16
Source of variation                     Computed
                                         F-value
Allocation protocol of component          50.33 [*]
  data activity
Structural type of the DAF-Net           120.31 [*]
Failure rate                             106.7 [*]
Data replication factor                   20.02 [*]
Combination of the allocation protocol    57.07 [*]
  of component data activity and
  structural type of the DAF-Net
Error
(*.)denotes that the result is
significant to a significance level of 0.5.
                    SNK test results for the failure
                   resilience experiment -- number of
                     committed DAF-Nets ([N.sub.c])
Effect                          Ranking             Mean
Combination of the allocatin    (NCL, Decomposite)  189.8
  protocol of component data    (ARR, Decomposite)  175.4
  activity and structural type  (NCL, Composite)    165.9
  of the DAF-Net                (ARR, Composite)    156.7
Significance level = 0.05.
                Experimental results for the processing
                quality experiment -- DAF-Net processing
                              quality (PQ)
Structural type of the DAF-Net  Allocation protocol of component data
                                              activity
                                            Failure rate
                                                 ARR
                                                 0.0
Composite                                       12.50
                                                12.81
                                                11.93
                                                13.02
                                                11.75
Decomposite                                     13.52
                                                13.24
                                                14.01
                                                13.45
                                                12.89
Structural type of the DAF-Net
                                              NCL
                                 0.08  0.16   0.0    0.08  0.16
Composite                       10.18  7.52  12.62  10.06  7.75
                                 9.84  7.32  12.21  10.52  7.23
                                11.03  7.12  12.91   9.85  7.82
                                10.74  6.97  11.87   9.98  8,12
                                10.42  7.99  12.06  10.92  7.36
Decomposite                      9.84  7.67  16.00  10.69  7.66
                                 9.34  7.61  15.92   9.84  7.32
                                 9.69  7.78  16.02  10.30  7.91
                                10.21  7.32  15.48  10.46  7.82
                                 9.01  7.73  16.37  10.71  7.14
                    ANOVA results for the processing
                           quality experiment
Source of variation                     Degree of freedom
                                              (df)
Allocation protocol of                          1
  component data activity
Structural type of the DAF-Net                  1
Failure rate                                    2
Combination of the allocation protocol          3
  of component data activity and the
  structural type of the DAF-Net
Error                                          48
Source of variation                     Sum of squares  Mean square
                                             (SS)          (MS)
Allocation protocol of                        4.52          4.52
  component data activity
Structural type of the DAF-Net                7.03          7.03
Failure rate                                358.27        179.13
Combination of the allocation protocol       15.91          5.30
  of component data activity and the
  structural type of the DAF-Net
Error                                         7.73          0.16
Source of variation                      Computed
                                          F-value
Allocation protocol of                    28.06 [*]
  component data activity
Structural type of the DAF-Net            43.61 [*]
Failure rate                            1112.00 [*]
Combination of the allocation protocol    32.93
  of component data activity and the
  structural type of the DAF-Net
Error
(*.)denotes that the result is
significant to a significance level of
0.05.
                  SNK test results for the processing
                quality experiment -- DAF-Net processing
                              quality (PQ)
Effect                          Ranking             Mean
Combination of the allocation   (NCL, Decomposite)  11.31
  protocol of component data    (ARR, Decomposite)  10.22 [I]
  activity and structural type  (NCL, Composite)    10.09 [I]
  of the DAF-Net                (ARR, Composite)    10.08 [I]
Significant level = 0.05.
Bar (I.)indicates that values within the
bar are not statistically different.
Activity     Type                  Information system
[C.sub.is]   Null                  Null
[C.sub.i1]   File copy             BOM file in CAD dept.
[C.sub.i2]   File transfer         BOM file in CAM dept.
[C.sub.i3]   Alarm message         E-Mail
[C.sub.i4]   File transfer         BOM file in CAPP dept.
[C.sub.i5]   Alarm message         E-Mail
[C.sub.i6]   File transfer         BOM file in CAP dept.
[C.sub.i7]   Alarm message         E-Mail
[C.sub.i8]   DB transaction        Production DB
[C.sub.i9]   DB transaction        Inventory DB
[C.sub.i10]  DB transaction        Production DB
[C.sub.i11]  Condition evaluation
[C.sub.i12]  Application Software  MRP software
[C.sub.i13]  Alarm message         E-Mail
[C.sub.i14]  Application software  MRP software
[C.sub.i15]  Alarm message         E-Mail
[C.sub.ie]   Null                  Null
Activity     Description
[C.sub.is]   Starting dummy node
[C.sub.i1]   Copy BOM data in CAD dept.
[C.sub.i2]   Send CAD data to CAM department
[C.sub.i3]   Notify manufacturing manager of "CAD
             data transfer"
[C.sub.i4]   Send BOM data to CAPP department
[C.sub.i5]   Notify CAPP manager of "BOM data
             transfer"
[C.sub.i6]   Send BOM data to CAP department
[C.sub.i7]   Notify CAP manager of "BOM data
             transfer"
[C.sub.i8]   Retrieve MPS data
[C.sub.i9]   Retrieve inventory data
[C.sub.i10]  Retrieve item master data
[C.sub.i11]  If no data is found for new part, then
             take transition [t.sub.6].
             Otherwise take transition [t.sub.5].
[C.sub.i12]  Execute MRP software
[C.sub.i13]  Send warning message to CAP manager
[C.sub.i14]  Execute MRP report generation program
[C.sub.i15]  Notify CAP manager of "MRP update"
[C.sub.ie]   Ending dummy node