International Journal of Production Research, Vol. 46, No.

5, 1 March 2008, 12971313

Game theoretic enterprise management in industrial collaborative networks with multi-agent systems
T. KAIHARA*y and S. FUJIIz
yFaculty of Engineering, Department of Computer and Systems Engineering, Kobe University, Kobe, Japan zFaculty of Science and Technology, Department of Mechanical Engineering, Sophia University, Tokyo, Japan

(Revision received January 2007) Nowadays, virtual enterprise (VE) is an important paradigm of business management in an agile environment. VE exists in several kinds of business organization through multi-layered product flows. It is obvious that a mechanism, through which these different enterprises can be integrated together, is required for the better management of VE. In this paper, we focus on the negotiation process in VE formulation as a basic research to clarify its effective management. Each enterprise in VE is defined as a software agent with multiutilities, and three types of primitive business models are targeted, such as the vertically integrated business model, horizontally specialized business model and hybrid business model. We develop a three-layered VE model for computer simulation so as to clarify VE formulation dynamism with the proposed negotiation mechanism. Keywords: Virtual enterprise; Negotiation; Multi-agent system; Game theory; Collaborative network

1. Introduction With the globalization of commerce, distribution and manufacturing, cooperation between enterprises of different sectors and cultures is significantly increasing. It is not only limited to sub-contracting and cooperation with suppliers and customers better known as supply chain or extended enterprise, but also concerned with virtual enterprise (VE). VE is an important concept of business management in an agile environment. VE is a temporary alliance of enterprises that come together to share skills or core competencies and resources in order to better respond to business opportunities in an agile environment and whose cooperation is supported by global computer networks (Camarinha-Matos et al. 1999), shown in figure 1. VE is a relatively new concept and different from previous cooperative business models (Putnik et al. 2005). Several kinds of organization are included in VE, and it is not easy to establish an appropriate collaboration amongst a large number
*Corresponding author. Email: kaihara@cs.kobe-u.ac.jp
International Journal of Production Research ISSN 00207543 print/ISSN 1366588X online 2008 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/00207540701224400

1298

T. Kaihara and S. Fujii

Figure 1.

VE and other co-operative enterprises.

of enterprises, such as in supply chain management (SCM) (Goldratt 1983, Fisher 1994, Chang et al. 2004). They normally have different objectives and these are often conflicting. Typical purchasing and manufacturing departments views, such as large product lot size, are completely against marketing or distribution views (Kovacs et al. 2001, Shen 2006). That might cause the separation of enterprises and destroy VE relationships. A new mechanism is obviously required to integrate each enterprise in VE so as to establish effective collaborative VE (Camarinha-Matos et al. 2006). In this paper, we focus on the negotiation process in VE formulation as a basic research to clarify its effective and concurrent management. Each enterprise is defined as an agent (Fox et al. 2000, Conen and Sandholm 2002), and a framework of multi-agent programming with marketing science (Katahira 1987) and game theoretic approach (Von Neumann and Morgenstern 1947) is newly proposed as a negotiation algorithm among the agents. Concretely, each unit is defined as a software agent in our VE model, and their decision making is formulated as marketing science models and N-person game theoretic methodology. We first classify business models in VE into three types, such as the vertically integrated business model, horizontally specialized business model and hybrid business model. Then we propose a contract net protocol (CNP)-based negotiation protocol amongst enterprises with marketing science models, such as the lexicographic model and maximum likelihood hierarchical (MLH) model, and N-person game theoretic approach (Smith 1980, Durfee et al. 1987, Katahira 1987, Sandholm 1993). CNP models transfer of control in a distributed system with the metaphor of negotiation among autonomous intelligent beings (Conen and Sandholm 2002). CNP consists of a set of nodes that negotiate with one another through a set of messages (Kaihara and Fujii 2002, 2003). Nodes generally represent the distributed computing resources to be managed and correspond to enterprises in this paper. Marketing science-based negotiation is applied into the vertically integrated business model considering a realistic enterprise management strategy. The horizontally specialized business model includes N-person game theoretic negotiations to realize the coordination amongst enterprises in the same business segment. We develop a computer simulation model to form VE through multiple negotiations

Game theoretic enterprise management in industrial collaborative networks

1299

amongst several potential members in the negotiation domain, and finally clarify the formation dynamism about three types of business model with the negotiation process. The contribution of this paper is to facilitate comprehensive and rational management amongst all the collaborative enterprises not only for VE but also SCM.

2. Business model in VE 2.1 Virtual enterprise model A large number of diversified networked organizations of enterprises fall under the general definition of VE. We assumed our VE model in the simplest possible definition as a basic research, as follows: 1. Duration: Single business An alliance of the enterprises is established towards a single business opportunity, and is dissolved at the end of such process. 2. Topology: Fixed structure There exist established supply chains with an almost fixed structure. 3. Participation: Single alliance All the enterprises are participating int only a single alliance at the same time. 4. Coordination: Democratic alliance A different organization can be found in some supply chains without a dominant company. All the enterprises cooperate on an equal basis, preserving their autonomy. 5. Visibility scope: Single level All the enterprises in VE communicate only to their direct neighbours in their architecture (figure 2). That is the case observed in most supply chains. Needless to say, it is a very important and difficult activity when forming a virtual enterprise to select appropriate business partners, i.e. partnering, because

Figure 2.

VE model.

1300

T. Kaihara and S. Fujii

each enterprise must consider not only the pursuit of profit but also sharing the risk to join the virtual enterprise. The partnering is described as coordination activity amongst the enterprises, and some sophisticated coordination mechanism is required to realize efficient interactions. The development of a coordination mechanism in computer-science can be found in the area of workflow management system, computer-supported cooperative work (CSCW), and multi-agent systems. The area of multi-agent systems, especially when involving intelligent autonomous agents, has been the discussion of coordination issues and supporting mechanism (Kaihara and Fujii 2002, 2003). The interaction capability, both amongst agents and between agents and their environment, is one of the basic characteristics of an agent (Fox 2000). In this paper we focus on the contract net protocol (CNP), which is one of the mechanisms coming from early work on multi-agent systems (Smith 1980), as the coordination and negotiation mechanism amongst business units in VE. Figure 2 shows the assumed VE model in this paper. We call an enterprise as unit, and there exist m layers, which have mn units in the model. The lowest level corresponds to consumers who create original task requests to the VE. At first, the customer dispatches new orders to all the units in layer m, and then several units, which are satisfied with the order, respond and circulate the order toward upper units in the VE model. Finally a VE with single supply chain will be established as a temporary alliance for the order as a consequence of their negotiations through all the layers. 2.2 Business model in VE environment As described in the previous section, there exist many business models in VE environment. We classify three types of business model in this study: the vertically integrated business model, the horizontally specialized business model and the hybrid business model as basic study. VE is regarded as an aggregation or a combination of these three business models, no matter how messy and complicated their structure is in reality. We try to simplify the VE structure from the systems engineering viewpoint, and analyze the formation dynamism in VE with the negotiation mechanism in this paper. Figure 1 shows the hypothetical VE model in this paper. We call an enterprise a unit, and there exist n layers, which have mn units in the VE model. The right end corresponds to consumers who can create original task requests to the VE. As the layer number increases, we describe it lower based on the product flow order in this paper. Generally we can observe vertical business models in traditional industries. In this business model the manufacturing business processes top-down as well as bottom-up from the process requirements, resulting in an integrated approach to overall business requirements. We will call this business model, which is based on end-to-end proprietary solutions that lock a customer to a manufacturer, a vertically integrated business model. Each unit at the same layer in figure 2 never tries to cooperate in our vertically integrated business model, because each unit is keen to find its appropriate partner just in vertical directions. For example, Unitij tries to find an alliance unit at each neighbouring layer: layer (i 1) and (i 1) in figure 2. Consequently only one SC can be formed in this business model. We introduce the

Game theoretic enterprise management in industrial collaborative networks

1301

consumers behaviour in marketing science into unit behaviour, and that makes our VE model more practical. On the other hand, enterprise relationships may represent a forerunning pattern of the learning alliance, whereby ongoing close interaction of horizontal alliance partners at single level or multiple hierarchical levels can be used to facilitate the mutual accumulation of superior organizational capabilities within the alliance firms. We will call this business model a horizontally specialized business model. All the units at the same layer try to cooperate to maximise their profit in total in our horizontally specialized business model. They behave as if they are under joint management, and we can see a kind of this style of management in industrial clusters in Japan (Kansai Bureau of Economy, Trade and Industry, available online). We apply N-persons game theory to formulate their behaviour in cooperation, because it describes a decision-making process to find their equilibrium solution in a social manner. Finally, a hybrid business model combines the above-mentioned two business models. There exists a parentchild relationship amongst units in VE, and the parent company has the advantage in their contract. Units in the child companies try to keep harmony as a group by sharing a common destiny in their management. We can see this type of business model sometimes in the Japanese automobile industry, such as Toyota or Mazda. We propose an alliance strategy based on cooperative game theory in this business model. Detailed architecture of the three business models is described in section 4.

3. Enterprise agent 3.1 Unit structure Each unit (enterprise) is defined as agent in our VE model, and its structure is described in figure 2. We adopt CNP as the coordination and negotiation mechanism amongst the units. CNP models transfer control in a distributed system with the metaphor of negotiation between autonomous intelligent beings. CNP consists in three interaction phases, involving two roles (manager and contractor). A manager announces a task to a set of contractors, each contractor bids for the task, and the manager awards the task (i.e. reward) to the contractor with the best bid. Any agent can start such a protocol by endorsing the adequate role. Nodes generally represent the distributed computing resources to be managed, corresponding to units in this paper. The unit structure proposed in this paper is described in figure 3. An agent (unit) can act both as a manager and a contractor of a delivery sets. When a unit receives new order (task announcement) i, it creates a contractor/ manager set (manager i/contractor i) for task i inside. Manager i creates a new order towards the lower units to secure the contract with the upper layer. There exist several situations in partnering amongst enterprise agents. In this paper it is assumed that the product demand is predictable in the negotiation under multipurpose criterion as basic study. That means order patterns are previously given and the negotiations start after the order reaches each enterprise agent. They should prepare robust solutions with maximum utilities against the order.

1302

T. Kaihara and S. Fujii

Figure 3.

Unit structure.

Figure 4.

Negotiation flow.

We propose several agent behaviours including game theoretic approach according to this assumption. 3.2 Negotiation algorithm The timeline of the proposed negotiation mechanism in this paper is shown in figure 4 (m-contractors n-managers model). Negotiation steps according to agent roles are described as follows: Manager (in layer x) Step M1: Create a new task based on the received bid information. Step M2: Task announcement (TA) to the upper units. Step M3: After the bidding period (Bidding period) expires, check all the acquired bids according to its standard. If there exists no bid to select, go to M4. Otherwise go to M5. Step M4: Modify the task and go to M2. Step M5: Select the task and send reward (Reward) to the corresponding unit.

Game theoretic enterprise management in industrial collaborative networks

1303

Contractor (in layer x 1) Step C1: Receive TAs. Step C2: Create an estimated bid according to its own capability. Step C3: Send the bid to the manager (Bid). Step C4: Request task announcement to the manager. We prepare several parameters to define unit behaviours as follows: Pij TA Qij TA Lij TA
i1j Q BID i1j LBID

TA price of Unitij TA quantity of Unitij TA lead time of Unitij Bid quantity of Unit(i1)j Bid lead time of Unit(i1)j Bid price of Unitij Bid quantity of Unitij Bid lead time of Unitij TA price of Unit(i1) j TA quantity of Unit(i1)
j

Pij BID Qij BID Lij BID i1j P TA i1j Q TA i1j L TA costij profitij process timeij plusij procure timeij

TA lead time of Unit(i1) j Process cost of Unitij per product Profit rate of Unitij Process time of Unitij per product Quantity increase rate of Unitij Estimated procure time of Unitij per product

Unit formulation in each business model is described in the following sections.

3.3 Vertically integrated VE model CNP is just the skeleton in the negotiation algorithm, and it is necessary to define how to select the appropriate bid (Step M5) to establish profitable contract. Marketing science-based negotiation is applied into the managers decision making to realize sophisticated bid selection based on consumers behaviours in the vertically integrated business model. First of all, each attribute of managerij in the TA producing process (Step M1) is as follows: Pij TA costij 1 profitij
i1j Qij TA QBID i1j i1j Lij TA LBID QBID processing timeij

1 2 3

1304

T. Kaihara and S. Fujii

Then contractorij tries to create its bid (Step C1) against TA from manager in layer (i 1) by the following equations:   1 profitij ij PBID costij 4 2
i1j Qij BID QTA 1 plusij ij Lij BID QBID processing timeij procure timeij

5 6

In this model, the contractor democratically tries to share their profit with a lower agent (i.e. manager) equally, shown in equation (4). The marketing science approach is applied in the bid selection mechanism (Step M5) in the vertically integrated VE model. We consider two types of marketing science model, named lexicographic model and MLH model (Katahira 1987). Bid selection mechanism in manager agent is described in each model as follows: Lexicographic model Step L1: Set priorities on all the attributes. Step L2: Select a bid with the highest value in the attribute. In case of a tie, move to the next attribute and check the highest value. This routine is continued until only one bid is selected. MLH model Step MLH1: Step MLH2: Set priorities on all the attributes. Standardize the highest attribute with the following equations: ~ij z highest zij amongst unselected j zij highest zij amongst unselected j 7

~ij is the standardized zij. where zij is the evaluation value of task j on ith attribute; z ~ij is within the tolerable amount (i.e. tolerance error:  i%), Step MLH3: If z then keep this attribute for the calculation in the next step. Step MLH4: After all the attributes are evaluated, then the preference index Vj of task j is calculated: ~1j 2 z ~2j i z ~ij Vj 1 z Finally selection probability of taskj is calculated by the following equation: Vj Pj Pn Vk 9 8

k1

3.4 Horizontally specialized VE model The horizontally specialized business model includes N-person game theoretic negotiations to realize the coordination amongst enterprises in the same business

Game theoretic enterprise management in industrial collaborative networks

1305

segment. In this model we consider there is a coordinator to manage the negotiation process amongst all the enterprises (i.e. contractors) in each layer. The coordinator in layer i manages all the units in the layer. It receives all TAs from lower layer (i 1) on behalf of coordinators in the layer i, and replies to the manager. The N-person cooperative game approach is applied to share its profit amongst all the units in the layer based on their attributes. So it is regarded as a kind of joint order organization. Firstly, the TA producing process (Step M1) is in common with the vertically integrated business model. Then the Bidding process is automatically replied based on the received TA, unless it yields no profit in any cooperation of contractor units (Step C2). So the attributes in the bid are defined as follows:
i1j Pij BID PTA i1j Qij BID QTA i1j Lij BID LTA

10 11 12

Finally the order sharing process is formulated with axioms called Shapley value. We applied the Shapley value because our VE cooperation model doesnt always satisfy super-additivity condition. This horizontally specialized VE corresponds to a kind of enterprise union in our VE model, and it sometimes requires some enterprise to make a sacrifice so as to obtain global profit for their union. That means super-additivity condition isnt always satisfied in this horizontally specialized VE model. The Shapley value describes one approach to the fair allocation of gains obtained by cooperation among several actors even in non-super-additivity conditions. The setup is as follows: a coalition of actors cooperates, and obtains a certain overall gain from that cooperation. Since some actors may contribute more to the coalition than others, the question arises of how to distribute fairly the gains among the actors; in other words, how important is each actor to the overall operation, and what payoff can they reasonably expect? The Shapley value is one way to distribute the total gains to the actors, assuming that they all collaborate. At first we define characteristic function (V(S)) as follows: VS Uprofit S Uleadtime S 13

where Uprofit(S ) is the profit in cooperation S, Uleadtime(S ) is the Boolean of lead time constraint (1: OK, 0: NG). Then the task is divided under the cooperation S 0 , which maximizes V(S). The contribution of unit i in cooperation S is calculated as follows:   X jSj 1!n jSj!  S v S v 14 i n ! f ig S:i2SN where N is a set of all units, v is all the profit in cooperation S, jSj is the number of members in cooperation S, (S\{i}) is the cooperation without unit i. Then naturally the following equation is acquired: X i v vN 15
i2N

1306

T. Kaihara and S. Fujii

So the profit, which gained in the coordinator, is divided into each unit in the layer according to the Shapley value in equation (14).

3.5 Hybrid VE model The hybrid business model combines the two above-mentioned business models. There exists a parentchild relationship amongst units in VE, and the parent company has the advantage in their contract, such as in automobile or space industries. Units in the child companies try to keep harmony as a group in their management. The TA producing process (Step M1) has no difference in the two abovementioned business models. Then contractorij creates its bid (Step C1) against TA from manager in layer (i 1) by the following equations: Pij BID costij 1 profitij Qij BID
i1j Q TA Ni

16 17 18

ij Lij BID QTA processing timeij procure timeij

where N is the number of cooperate units in layer i. Finally the bid selection process by manager (Step M5) is formulated by the cooperative game theory under super-additivity condition. We define characteristic function (V(S)) as follows: VS U0profit S Uleadtime S 19

0 (S) is the managers profit in cooperation S. where Uprofit Then the characteristic function values in all the received bids are calculated and the manager finally selects the bid with the highest value amongst them.

4. Experimental results 4.1 Simulation parameters As a basic study, a three-layered VE model for computer simulation was developed to clarify VE formulation dynamism with the proposed negotiation mechanism. Each layer consists of three enterprises in this simulation model described in figure 2. Simulation parameters are shown in table 1. All the results are the average of 500 trials in each simulation scenario.
Table 1. m n i cost0j cost1j cost2j Simulation parameters. profitij plusij processing timeij procure timeij 0.080.12* 0.100.20*

3 3 0.4 1525* 4555* 8595* 0.170.23* 0.080.12*

*Followed by uniformed random distribution.

Game theoretic enterprise management in industrial collaborative networks

1307

4.2 Vertically integrated VE model We applied two types of marketing science model, named lexicographic model and MLH model, as described previously. The priority of attributes plays an important role in marketing science. Several simulation patterns in different priority, shown in table 2, were carried out so as to analyze the performance of proposed negotiation mechanisms. The results of two types of marketing science model, lexicographic model and MLH model, are shown in tables 3 and 4, respectively. The parameter  is set as 0.4 in MLH model (table 4) after several preliminary experiments.

Table 2. Pattern Pattern Pattern Pattern Pattern Pattern 1 2 3 4 5 6

Priority in negotiation attributes. price4quantity4leadtime price4leadtime4quantity quantity4price4leadtime quantity4leadtime4price leadtime4price4quantity leadtime4quantity4price

Table 3. Average

Total results in lexicographic model (n 3). Standard deviation Average Standard deviation Average Standard deviation

Pattern 1 Profit Stock (WIP) Lead time 8076.05 947.00 21.14 2.50 24.10 2.82 Pattern 4 8086.16 20.51 22.34 947.69 2.47 2.66

Pattern 2 8076.05 947.00 21.14 2.50 24.10 2.82 Pattern 5 7506.56 23.56 21.17 879.68 2.82 2.45

Pattern 3 8086.16 947.69 20.51 2.47 22.97 2.73 Pattern 6 7506.56 23.56 21.17 879.68 2.83 2.45

Profit Stock (WIP) Lead time

Table 4. Average

Total results in MLH model (n 3,  0.4). Standard deviation Average Standard deviation Average Standard deviation

Pattern 1 Profit Stock (WIP) Lead time 7513.20 879.24 23.56 2.81 21.18 2.44 Pattern 4 8082.60 21.14 24.11 946.19 2.50 2.81

Pattern 2 8092.63 946.84 20.51 2.47 22.35 2.65 Pattern 5 8092.68 20.51 22.98 946.84 2.47 2.73

Pattern 3 7513.20 879.25 23.56 2.82 21.18 2.45 Pattern 6 8082.60 21.14 24.11 946.19 2.50 2.81

Profit Stock (WIP) Lead time

1308

T. Kaihara and S. Fujii

It is obvious that the results depend only on the highest prioritized attribute in lexicographic model. The bid selection process is carried out according to the direct managers preferences in this model, and that causes only the highest attribute to affect the bid selection in the negotiation domain. In other words, the negotiation with lexicographic model tends to establish the manager-driven negotiation, i.e. lower layer-driven negotiation environment, in supply chains. We observed the similarity between Patterns 1 and 3, Patterns 2 and 5, and Patterns 3 and 6, in table 4. They are all classified into the same group in terms of attributes within the second highest. MLH model has a tolerance ( ) in its selection attributes, and it has been confirmed that the bid selection process is carried out within the second highest attributes in ( 0.4) in the simulation scenario. As a consequence, it has been confirmed that the bid selection based on MLH model emerges comprehensive and well-balanced tradings in terms of managers attributes, although it also has a tendency to establish the manager-driven negotiation amongst all the layers in VE. We increased the number of agents to 10 in all the layers so as to investigate the scalability of the above mentioned observations. As an example, the results in MLH model is shown in table 5, and it is obvious that the large VE model has the similar tendency as the small model shown in table 4. We have confirmed that our considerations mentioned before are validated even in larger VE model.

4.3 Horizontally specialized VE model Figure 5 illustrates an experimental model of the horizontally specialized VE model in three layers. It is assumed that the middle layer has a coordinator to realize collaborations amongst all the enterprise inside the layer. The cooperation is carried out according to the negotiation mechanism described in section 3.4. Simulation results from each enterprise in layer 1 and 2 are shown in table 6. It is obvious that the total profit of layer 1 is greater than that of layer 2 in this table. It has been clarified that coordinator increases the efficiency of layer 1 by conducting collaboration amongst enterprises in the layer, and that means it plays an important role in establishing a contractor-driven negotiation environment, i.e. upper layer-

Table 5. Average

Total results in MLH model (n 10,  0.4). Standard deviation Average Standard deviation Average Standard deviation

Pattern 1 Profit Stock (WIP) Lead time 7365.45 881.08 22.37 2.83 20.73 2.86 Pattern 4 7692.87 21.71 23.65 916.56 2.45 3.16

Pattern 2 7390.09 884.49 21.14 2.50 21.24 2.92 Pattern 5 7189.76 20.61 23.65 853.37 2.44 3.16

Pattern 3 7360.69 880.00 22.61 2.83 20.68 2.82 Pattern 6 7692.87 21.71 23.65 916.56 2.45 3.16

Profit Stock (WIP) Lead time

Game theoretic enterprise management in industrial collaborative networks

Unit00 Unit01 Unit10 Unit20 Unit21

1309

Unit11 Unit0j Unit12 Unit0n Upper Stream Co-ordinator Unit2n Customer Lower Stream Unit2j

Figure 5. Table 6. Unit20 Profit Stock (WIP) Lead time 1024.94 10.08 8.78

Experimental model (n 3).

Unit results in horizontal VE model (n 3). Unit21 31.54 0.26 0.225 Unit22 14.292 0.10 0.087 Unit10 2858.94 6.39 8.36 Unit11 2250.02 4.44 4.42 Unit12 0 0 0

Table 7.

Total results in horizontal VE model. n3 n 10 Average 6200.25 21.13 16.57 Standard deviation 765.00 2.50 2.16

Average Profit Stock (WIP) Lead time 6179.73 21.28 16.43

Standard deviation 730.06 2.56 2.14

driven negotiation environment, compared with vertically integrated VE models in section 4.2. Total results of horizontally specialized VE model in n 3, 10 are described in table 7. It has been observed that this model has better performance in lead time but less in total profit due to the collaborative negotiation in layer 1. The difference between n 3, and 10 is very small, and that means the horizontal VE model is stable and robust in terms of VE scalability. It has been proved that their collaborative behaviour enhances their stability against structural change, but sacrifices their profit for the robustness in this experiment.

4.4 Hybrid VE model An experimental model of hybrid VE model is shown in figure 6. It combines vertical and horizontal models, and establishes complex negotiation under the superadditivity condition described in section 3.5.

1310
Unit00 Unit01

T. Kaihara and S. Fujii

Unit20 Unit10 Unit21

Unit11 Unit0j Unit12 Unit0n Upper Stream Unit2n Customer Lower Stream Unit2j

Figure 6. Table 8. Unit20 Profit Stock (WIP) Lead time 3968.46 10.31 8.79

Experimental model (n 3).

Unit results in hybrid VE model (n 3). Unit21 91.86 0.22 0.19 Unit22 54.67 0.12 0.10 Unit10 1959.81 5.11 6.85 Unit11 1853.74 6.53 7.55 Unit12 0 0 0

Simulation results about individual enterprise in layer 1 and 2 are shown in table 8. Compared with table 6, it is observed that the total profit of layer 1 is much closer to that of layer 2 in this table. Spontaneous collaborated behaviour between Unit10 and Unit11 is also observed without coordinator in horizontal model. It has been confirmed that the hybrid model obtains the characteristics observed in the vertical and horizontal models. Total results of hybrid VE model in n 3,10 are shown in table 9. They are also almost equivalent, but the stability is slightly weaker in terms of scalability compared with the horizontal VE model shown in table 7. They maintain, however, higher profit with relatively short lead time. That means the hybrid VE model is wellbalanced in terms of efficiency and robustness because of its spontaneous collaborative negotiations.

4.5 Comparison of three business models in VE The performances of finally acquired VE in three business models are compared in table 10 (Model 1: Vertically integrated business model), table 11 (Model 2: Horizontally specialized business model) and table 12 (Model 3: Hybrid model) in n 3. We tried to analyze the VE robustness of each business model against due date change and production volume change in this experiment and 500 trials are examined in each business model. The default due date and production volume are set to 30 and 100, respectively, and (0.5, 1.0), for example, means the due date is shortened to 1/2 (0.5) and production volume is equivalent (1.0) to the default in these tables. Only the results of the lexicographic model are described in table 10.

Game theoretic enterprise management in industrial collaborative networks Table 9. Total results in hybrid VE model. n3 Average Profit Stock (WIP) Lead time 7928.55 22.32 16.68 Standard deviation 930.86 2.70 2.07 Average 7904.90 21.80 17.68 n 10 Standard deviation 973.48 2.71 2.55

1311

Table 10. Experimental results of vertically integrated business model (Model 1). Average Standard deviation (1.0, 1.0) Profit Stock (WIP) Lead time 8076.05 21.14 24.10 947.00 2.50 2.82 (1.0, 1.5) 1413.53 3.79 4.21 8076.05 21.14 24.10 Average Standard deviation (0.5, 1.0) 947.00 2.50 2.82 (0.5, 1.5) 1413.53 3.79 4.21

Profit Stock (WIP) Lead time

12098.30 31.64 36.10

12098.30 31.64 36.10

Table 11. Experimental results of horizontally specialized business model (Model 2). Average Standard deviation (1.0, 1.0) Profit Stock(WIP) Lead time 6179.73 21.28 16.43 730.06 2.56 2.14 (1.0, 1.5) Profit Stock(WIP) Lead time 9263.05 32.12 25.49 1084.19 3.82 2.99 8889.12 33.04 20.67 5919.06 21.82 13.47 Average Standard deviation (0.5, 1.0) 708.20 2.68 1.81 (0.5, 1.5) 1047.26 3.81 2.59

Table 12. Experimental results of hybrid model (Model 3). Average Standard deviation (1.0, 1.0) Profit Stock(WIP) Lead time 7928.55 22.32 16.69 930.86 2.70 2.07 (1.0, 1.5) 1390.56 4.02 2.99 7628.55 22.32 12.98 Average Standard deviation (0.5, 1.0) 930.86 2.70 1.62 (0.5, 1.5) 1390.56 4.02 2.31

Profit Stock(WIP) Lead time

11877.20 33.49 24.95

11877.20 33.49 19.41

1312

T. Kaihara and S. Fujii

The following points have been observed in those experiments: . Any models dont satisfy lead time (15.00) in (0.5, 1.5), because the requirement change is too heavy to handle. . Lead time is satisfied at Model 2 and 3 both in (0.5, 1.0) and (1.0, 1.5). These relatively slight changes are manageable in those business models except Model 1. . Total profit is the highest at Model 1, because only the contributed units can take direct profit. Additionally stock is the least at this model. . Model 3 attains the shortest lead time in most cases. These results have been summarized as follows: . It is difficult for the vertically integrated business model to adapt due date change and production volume change, because it does not include any unit cooperation mechanisms. However, it performs best in profit, and this model is efficient in case there are enough margins in lead time (i.e. stable situations). . The horizontally specialized business model is robust against due date change and production volume change, although the profit is less than the vertically integrated business model. This business model is suitable for agile manufacturing situations with autonomous cooperation alliance. . The hybrid model takes a middle position between vertically integrated business model and horizontally specialized business model. Although it takes advantages of both models, it has been observed that the shared profit is inclined into lower layer units in VE compared with horizontally specialized business model.

5. Conclusions In this paper, we focused on the negotiation process in VE formation to clarify its effective management. We firstly classified the business model into three types: the vertically integrated business model, horizontally specialized business model and hybrid business model. Then we proposed a CNP-based negotiation protocol amongst enterprises with marketing science models, such as lexicographic model and MLH model, and N-person game theoretic approach. Marketing science-based negotiation was applied into the vertically integrated business model considering a realistic enterprise management strategy. The horizontally specialized business model included N-person game theoretic negotiations to realize the coordination amongst enterprises in the same business segment. We developed a computer simulation model to form VE through multiple negotiations amongst several potential members in the negotiation domain, and finally clarified the formation dynamism with the negotiation process. It has been confirmed that the vertically integrated business model is profit-oriented and it is the best in relatively stable business situations. On the contrary, the horizontally specialized business model is robust against the order change, and it suits agile manufacturing situations. The hybrid business model is moderate characteristic between them, and it seems useful practically as often shown in real situations.

Game theoretic enterprise management in industrial collaborative networks

1313

The contribution of this paper lies in the idea of a multi-agent-based VE negotiation mechanism combined with marketing science and cooperative game, which take a metaphor of decision making processes in social activities. This paper gives an initial illustration of the approach into a primitive VE structure. Effective enterprise partnering in global industrial collaborative networks is expected by this research. There is one obvious extension, which is to elaborate the negotiation protocol, possibly by exploiting some adaptive decision processes. It is expected that the extension makes the enterprise agent collaborate more rationally and that leads to more effective VE formulation with robustness.

References
Camarinha-Matos, L.M. and Afsarmanesh, H., The Virtual Enterprise Concept. Infrastructures for Virtual Enterprises, pp. 314, 1999 (Kluwer Academic Publishers: Boston, MA). Camarinha-Matos, L.M., and Afsarmanesh, H., Modeling Framework for Collaborative Networked Organizations. Network-Centric Collaboration and Supporting Frameworks, 224, pp. 314, 2006 (Springer, Boston, MA). Chang, Y.S., Makatsoris, H.C. and Richards, H.D., Evolution of Supply Chain Management, 2004 (Kluwer: Boston, MA). Conen, W. and Sandholm, T., Differential-Revelation VCG Mechanisms for Combinatorial AuctionsAgent-Mediated Electronic Commerce IV, pp. 6986, 2002 (Springer: Berlin). Durfee, E., Lesser, V. and Corkill, D., Coherent cooperation among communication problem solvers. IEEE Trans. Comp., 1987, 36, 12751291. Fisher, M.L., Making supply meet demand in uncertain world. Harv. Bus. Rev., 1994, May/ June, pp. 8393. Fox, M.S., Barbuceanu, M. and Teigen, R., Agent-oriented supply chain management. Int. J. FMS, 2000, 12, 165188. Goldratt, E.M., The GOAL, 1983 (North River Press: Great Barrington). Kaihara, T. and Fujii, S., A Proposal on Negotiation Methodology in Virtual Enterprise. Collaborative Business Ecosystems and Virtual Enterprises, pp. 125132, 2002 (Kluwer Academic Publishers: Boston, MA). Kaihara, T. and Fujii, S., Partnering Mechanism with Adaptive Multi-agent Protocol for Virtual EnterpriseProcess and Foundations for Virtual Organisations, pp. 407414, 2003 (Kluwer Academic Publishers: Boston, MA). Kansai Bureau of Economy, Trade and Industry. Available online at: http://www.kansai. meti.go.jp/3-2sanki/cluster-beam/english.html (accessed 16 February 2006). Katahira, H., Marketing Science [in Japanese], 1987 (Tokyo University Press: Tokyo). Kovacs, G.L., Bertok, P. and Haidegger, G., Digital Enterprise Challenges, 2001 (Kluwer: Boston, MA). Putnik, G.D. and Cunha, M.M., Virtual Enterprise Integration, 2005 (Idea Group: Hershey). Sandholm, T., An implication of the contact net protocol based on marginal cost calculations, in Proceedings of the National Conference on AI, 1993, pp. 256262. Shen, W., Information Technology for Balanced Manufacturing Systems, 2006 (Springer: New York, NY). Smith, R., The contract net protocol. IEEE Trans. Comp., 1980, C-29, 1041113. Von Neumann, J. and Morgenstern, O., Theory of Games and Economic Behavior, 1947 (Princeton University Press: Princeton).