Computers & Industrial Engineering 84 (2015) 43–55

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Optimal configuration of cluster supply chains with augmented Lagrange

coordination
T. Qu a, D.X. Nie a,b,⇑, X. Chen a, X.D. Chen a, Q.Y. Dai a, George Q. Huang c
a
Guangdong CIMS Provincial Key Lab, Guangdong University of Technology, Guangzhou 510006, China
b
The College of Science, South China Agricultural University, Guangzhou 510642, China
c
Department of Industrial and Manufacturing Systems Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong, China

a r t i c l e i n f o a b s t r a c t

Article history: Enterprises in an industrial cluster could dynamically alliance in the form of cluster supply chains to
Received 4 September 2014 share inner-cluster resources and services, and respond to the ever-fluctuating customer demands in a
Received in revised form 29 November 2014 cost-effective way. However, an effective and feasible method enabling such dynamic cluster supply
Accepted 22 December 2014
chain configuration (CSCC) lags behind practice due to the conflict of interests. Researchers are designing
Available online 3 January 2015
All-in-One theoretic models to optimize CSCC with the assumed decision details of all enterprises, while
in fact clustered enterprises are seeking effective decentralized decision mechanisms which protect their
Keywords:
decision autonomy in the frequently re-configured CSC. A newly emerged multi-disciplinary optimization
Cluster supply chain
Supply chain configuration
method, Augmented Lagrangian Coordination (ALC), which supports the open-structure collaboration
Supplier selection with strict optimization convergence, is thoroughly investigated in this paper and applied to solve the
Multidisciplinary design optimization conflict. Through a complete analysis of CSC’s configuration policies in typical stages, a generic CSCC
Augmented Lagrangian coordination model is proposed and then partitioned into an ALC-based decentralized decision model by the typical
decision autonomy distribution in clusters. Clustered enterprises collaborate vertically and laterally along
the ALC model through multi-dimensional couplings to achieve the overall consistency and optimality.
Results have proved the effectiveness of ALC for CSCC problem. A set of sensitivity analysis is also con-
ducted to find out the condition in which an order has to be fulfilled in a CSC and the most appropriate
configuration.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction approaches in the CSC, either horizontally subcontracting the order

Industrial cluster is a group of all-sized enterprises of a special-
ization division located in a certain area. They collaborate to gain
both cost and time advantages through convenient and effective
resource and service sharing, and cultivate higher industrial pres-
tige to achieve better regional competitiveness and market oppor-
tunity (Beaudry & Breshi, 2003; Pandilt, Cook, & Peter Swann,
2002). Cluster-based production is normally initiated by a leading
enterprise which owns an independent product brand, and is oper-
ated in the form of cluster supply chain (CSC) which comprises sev-
eral coupled single supply chains and a set of independent
suppliers (Li, Xiong, & Park, 2012). Literature concerning with
industrial clusters normally refers to the CSC as an alliance and
the leading enterprise as alliance leader (Huang, Qu, Zhang, &
Yang, 2012; Li, Huang, Fang, & Qu, 2013), and this terms will be fol-
lowed in this paper. Large and urgent orders will adopt outsourcing
⇑ Corresponding author at: Guangdong CIMS Provincial Key Lab, Guangdong
University of Technology, Guangzhou 510006, China.
to other single supply chains or vertically sourcing components
from independent suppliers or in a combined way, to guarantee
the service time and quality requirement of the order with cluster
advantages (Kawtummachai & Hop, 2005). Industrial cluster is
becoming an important cost-effective industry development mode
for enterprises to respond to the frequently changed market
demand and thus being promoted by more and more countries.
Cluster supply chain is essentially an order-based production
alliance system, which is subject to the naturally existed short life-
cycle, especially in the trend of ever-flourished product varieties
and frequently changed customer demands. Therefore, to effec-
tively and efficiently configure and reconfigure a CSC has become
a key stage for the CSC operation. Supply chain configuration
(SCC) is a complex decision process aiming to optimize certain per-
formance indicators of the supply chain through making decisions
such as selecting suitable suppliers for each stage, assigning values
to characteristics parameters of each stage and setting operation
policies for governing the interrelationships among these stages
(Huang & Qu, 2008). Following this definition, cluster supply chain

http://dx.doi.org/10.1016/j.cie.2014.12.026
0360-8352/Ó 2015 Elsevier Ltd. All rights reserved.
44 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55
configuration (CSCC) is responsible for selecting suitable single
supply chains and independent suppliers to collaborate a given
customer order in an appropriate way so that certain cluster pro-
duction measures could be optimized. Due to the higher configura-
tion frequency brought by the shorter supply chain lifecycle and
the multiplied number of supply chain participants, the CSCC deci-
sion process becomes largely complicated.
Since MIT researchers proposed the SCC problem (Graves &
Willems, 2001), abundant works have been conducted in this area,
including the SCC modeling toward different product structures
(Huang, Zhang, & Liang, 2005), sourcing policies (Amini & Li,
2011; Li & Amini, 2012), supplier constraints (Li & Womer, 2008),
etc. Various All-in-One (AIO) optimization methods are also pro-
posed for SCC solution, including DP (Dynamic Programming)
(Graves & Willems, 2005), GA (Genetic Algorithm) (Huang et al.,
2005), etc. During the increase of supply chain scale which con-
tains dynamically changed supplier base, maintaining the decision
autonomy of suppliers while enhancing the overall SCC optimiza-
tion efficiency have become a major concern of SCC (Fan,
Stallaert, & Whinston, 2003; Lee & Whang, 1999). Therefore,
decentralized SCC with MDO (Multidisciplinary Optimization) appli-
cation has attracted the interests from researchers (Walsh &
Wellman, 2003). The authors of this paper have attempted to use
ATC (analytical target cascading) to solve the configuration prob-
lem of assembly supply chains and obtained satisfactory results
(Huang & Qu, 2008; Qu, Huang, Chen, & Chen, 2009; Qu, Huang,
Cung, & Mangione, 2010).
The current research outputs concerning with CSCC are scarce,
with limited research width and depth. Most of researchers mainly
focus on investigating coordination of inventory and lateral trans-
shipments (Lorenzo & Stefano, 2011; Yang & Qin, 2007). A few
researchers use simplified models to deal with the configuration,
which discuss either horizontal subcontracting to sibling supply
chains (Bikram, Bahinipati, & Deshmukh, 2009; Xiang, Faishuai, &
Feifan, 2014) or vertical sourcing from independent suppliers
(Camarinha-Matos, Afsarmanesh, & Galeano, 2009). A few CSCC
works consider both yet only take a two-stage approach (Li et al.,
2012), i.e. deal with the two-way collaborations in different stages,
and restrict the number of configurable cross-chains. Such simpli-
fications largely hinder the obtaining of the global optimal CSCC in
the cluster. While for the optimization methods, most literatures
assume the supreme decision right of the alliance leader for the
CSCC and apply AIO optimization models. In fact, these two limita-
tions of CSCC models and solution methods are twined together,
i.e. it is difficult to establish the optimization models for AIO meth-
ods such as DP and GA if the CSC structure and the number of sin-
gle supply chains cannot be fixed. So far, we have not found any
literature talking about MDO approach applied for CSCC. This situ-
ation prevents the application of research output from solving
practical problems, especially with the ever-demanding CSCC
requirements of supporting larger scale, shorter lifecycle, and
higher-level privacy protection. Without an effective and adapt-
able MDO method supporting decentralized and collaborative
CSCC decision making, the integration of the actual decisions and
operations of cluster enterprises in a practical CSCC system is
difficult.
ALC (Augmented Lagrange Coordination) is a newly emerged
decomposition-based MDO method with strict convergence proof
and supports collaborative optimization (Tosserams, Etman, &
Rooda, 2008, 2010). ALC is put forward by Tosserams (2008) and
used to deal with the optimal design problem of large complex sys-
tem (Allison & Papalambros, 2010). The basic principle of ALC is to
partition the system into a decentralized decision structure com-
posed of a set of independent decision elements based on certain
partition rules, e.g. decision autonomy (Zhang, Huang, Sun, &
Yang, 2014; Zhang, Zhang, et al., 2014), and then to coordinate
the elements’ local decision process by ALC to obtain the global
optimal solution. As compared to other MDO methods, ALC offers
higher flexibility to the system’s coordination. First, ALC supports
open collaboration structure allowing for both horizontal and verti-
cal interaction among decision elements. Second, ALC supports not
only quasi-separable coupling (Tosserams, Etman, & Rooda, 2007)
among elements, but also coupling function including objective
coupling function and coupling constraint function. Such flexible
features make ALC applicable for CSCC problem with complex and
dynamic collaboration relationships among cluster enterprises.
In order to maintain simplicity without losing generality, this
research designates the supply chain of a given product with one
key/bottleneck component as a research problem. As the supply
chain sourcing process of different components is usually indepen-
dent for an enterprise, this research is extensible to the model con-
sidering multiple components. In case of the coming of a large
order with urgent service time, the order owner wishes to config-
ure a CSC to collaborate with other homogeneous manufacturers
(single supply chains) and independent suppliers to solve the
capacity limitation problem. The major research questions of this
paper are as follows. First, the mechanism of cluster supply chain
formation will be investigated, including the general policies of
order subcontracting and component sourcing in the case of lim-
ited production capacity. Second, a general modeling and solution
strategy of a decentralized ALC process will be established for CSCC
in specific and for all supply chain related problems in general.
Third, the optimization effectiveness and efficiency of ALC and
other optimization methods will be compared to provide a useful
reference for researchers. Fourth, the condition that an enterprise
should resort to cluster supply chain instead of normal supply
chain will be investigated, and the corresponding CSCC is
compared.
The remainder of this paper is structured as follows. Section 2
will investigate the formation mechanism of cluster supply chain
and analyze the current solution challenges. Section 3 presents
the ALC principles with a reference problem modeling procedure,
which is then applied to the CSCC modelling in Section 4. The
detailed ALC solution procedure is given in Section 5, and results
are analyzed with sensitivity analysis. Finally, the paper concludes
in Section 6, where managerial implications and future research
directions are outlined.
2. Problem description
2.1. Supply chains clustering modes
There are three types of enterprises normally existed in an
industrial cluster, namely product manufacturers (M), manufactur-
ers’ private suppliers which have long-term stable relationships
(S), and other independent suppliers dispersed in the cluster (O),
as shown by the three grey dotted blocks in Fig. 1(a). Literature
has reported various collaboration modes among these enterprises,
mainly falling into the two categorizes of horizontal order-subcon-
tracting among sibling supply chains and vertical component
sourcing among independent suppliers. The former means, a man-
ufacturer – normally a brand owner – will take orders from market
and split and subcontract a portion of the order to other manufac-
turers, i.e. sibling supply chains (also referred to as single supply
chains in literature) (Li et al., 2012). Some highly specialized indus-
trial clusters also have a special kind of so-called brand operators
who dedicate to the marketing and a fixed group of OEM (Original
Equipment Manufacturers) for order fulfilling, i.e. subcontract all the
orders. The latter means, after a manufacturer takes an order, it
could source components from both its private suppliers and mul-
tiple independent suppliers in the consideration of cost optimiza-
tion, safe supply, price control and workload control, etc.
 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55
Industrial Cluster
Cluster supply chain
L Leader
Manufacturers
Horizontal
M1 Subcontract
Mn
CSCC
Policies
Independent Suppliers Private Suppliers
O1 … …Om S1 … … Sn
(a) Operational concept of cluster supply chain
Fig. 1. Operation modes and configuration model of cluster supply chains.
2.2. Configuration model of cluster supply chains
The operational concept of a cluster supply chain is shown in
Fig. 1(a). Before an optimal configuration method is applied, the
concept model should be converted into a configurable model
based on the corresponding CSCC policies. In the following, the
typical configuration policies in the three key CSC levels, i.e. alli-
ance leader level, manufacturer level, and supplier level, will be
discussed respectively. First of all, the assumptions and the corre-
sponding generality explanations are summarized in Table 1.
The above assumptions will lead to three important and feasible
CSCC policies which will be adopted in the following discussion of
this paper. These policies are also applicable and extensible to
most of the cluster supply chains.
 Alliance Policy. Node L represents the alliance leader. Based on
the assumption 1 and 2, the leader is the manufacturer who
owns the original order. It could be a specialized brander oper-
ator or the sales department of the manufacturer.
 Subcontracting Policy. Based on the assumption 3, 9 and 10, an
alliance leader will remain enough orders to its private manu-
facturer ML based on the capacity and then horizontally subcon-
tract the rest of the order to no more than two sibling
manufacturers, i.e. other single supply chains. Orders will be
Table 1
Problem assumptions and generality explanation.
No Problem assumptions
1 Manufacturing taken orders will serve as the alliance leader. Customer order
competition among manufacturers is not considered
2 Alliance leader releases orders to its own manufacture in the first priority, while
a manufacturer sources component from its private suppliers in the first
priority. Sourcing order competition among suppliers is not considered
3 Maximum number of sibling supply chains to be subcontracted is 2
4 Only one critical component is to be sourced
5 The BOM ratio of the critical component is 1:1
6 There is only one private supplier of the critical component
7 Will not source component from other manufacturers’ private suppliers
8 Multiple sourcing will be applied for independent suppliers without number
limitation
9 Both manufacturers and suppliers have capacity limitation
10 Both the order subcontracting and component sourcing follow the rule of step
price, i.e. the higher quantity, the lower cost
45
CSCC Model
L
ORM
M1 … … Mn
ORS
S1 O1 … …Om Sn
(b) Coniguration model of cluster supply chain
allocated among sibling manufacturers according to the pricing
and profit sharing mechanisms, in the consideration of all man-
ufacturers’ capacity constraints.
 Sourcing Policy. Based on the assumption 4–10, a manufacturer
will source from its private suppliers first, and then source from
other independent suppliers with multiple sourcing strategy.
Suppliers with lower prices will be preferred, in the consider-
ation of capacity constraint.
2.3. Challenges of cluster supply chain configuration
After the overall configuration policies are established, the
operational concept shown in Fig. 1(a) will be converted into a
cluster supply chain configuration (CSCC) model shown in
Fig. 1(b). The alliance leader L in the top level will select no more
than 3 manufacturers from all the cluster manufacturers in the sec-
ond level through ORM node, with its own manufacturer as a fixed
option. Each manufacturer will then select suppliers from level
three through ORS, with its private supplier as a fixed option.
Although the authors have solved supply chain configuration prob-
lem with capacity constraint and multiple sourcing strategies, sev-
eral challenges listed in the following still exist when solving CSCC
problems.
Generality explanation
It is a generally accepted cluster policy which encourages enterprises to strive
for more orders, and is equal and beneficial to all the cluster enterprises
Private subsidies normally have lower costs. Even in the case that outsourcing
more could bring higher profit, the dynamic allying mechanism requires each
enterprise to maintain enough order/market share for their private subsidies
A general requirement of subcontracting enterprises to reduce interaction
complexity
Different components normally have different supplier bases, and the sourcing
process normally has independent decision and audit
Linear changes of BOM ratio will not affect the overall optimization result
If the total supply capacity of private suppliers remains unchanged, the
configuration results are the same
The private suppliers normally serve only for the host manufacturer
Most of the independent suppliers are small and micro business
The most important feature of a cluster is to share capability and capacity
among the specialized cluster enterprises
A general market rule
46 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55

(1) Dynamic Structure. Although this paper considers a three- 3.2. ALC solution
echelon structure topped by a single alliance leader, inde-
pendent suppliers are open to all manufacturers and each 3.2.1. Objective system model
of them has the chance to be shared and thus results in net- ALC which has more flexible optimization structure is proposed
worked structure. Therefore, the overall structure will jump by Tosserams et al. (2008). ALC is to solve problem with the follow-
between hierarchical and networked structures in the con- ing mathematical formulation.
figuration processes.
X
M
(2) Uncertain branches. Unlike most of the literatures which min f 0 ðy; x1 ; . . . ; xM Þ þ f j ðy; xj Þ
restrict the maximum number of suppliers, this paper allows z¼½yT ;xT1 ;...;xTM 
T
j¼1
random number suppliers. This will result in uncertain num-
s:t: g 0 ðy; x1 ; . . . ; xM Þ 6 0
ber of branches below those multiple selection nodes, i.e.
h0 ðy; x1 ; . . . ; xM Þ ¼ 0 ð1Þ
ORM and ORS in Fig. 1(b).
(3) Complex function coupling. Due to the consistent cluster g j ðy; xj Þ 6 0; j ¼ 1; . . . ; M
policy and the common characteristics resulted from the hj ðy; xj Þ ¼ 0; j ¼ 1; . . . ; M
long-term collaborations among cluster enterprises, some T
coupling functions will be shared among certain enterprises, The vector of design variables z ¼ ½yT ; xT1 ; . . . ; xTM  2 Rn consists
ny nx
such as the same objective function (e.g. maximum cluster of a number of linking variables y 2 R , local variables xj 2 R j
P
profit) or the coupling constraint function (e.g. share the alli- associated exclusively with subsystem j, where ny þ M x
j¼1 nj ¼ n.
ance leader’s order). Therefore, the quasi-separable model The coupling objective function f 0 : Rn ! R, and coupling con-
(Tosserams et al., 2007) adopted by traditional MDO meth- g h
straints g 0 : Rn ! Rm0 and h0 : Rn ! Rm0 are non-separable and
ods will have to be substituted by the coordination methods
may depend on all design variables z. Local objective functions
which support multi-dimensional couplings with both the g
m mh
linking variable and coupling function. f j : Rnj ! R and local constrains g j : Rnj ! R j and hj : Rnj ! R j
are associated exclusively with subsystem j.
P PM h
The above challenges make it difficult for AIO method to solve ny þ nxj ¼ nj ; mg0 þ M m
j¼1 j
g
¼ m g
; m h
0 þ m
j¼1 j ¼ mh
.
CSCC problems, e.g. it is difficult to determine the number of genes
to create the chromosome with uncertain fitness function structure
3.2.2. Auxiliary variables and consistency constraints
when using GA method, and most of the traditional MDO methods
The first step of the problem decomposition is to introduce aux-
only support quasi-separable models. ALC not only supports open
iliary variables and consistency constraints. Auxiliary linking vari-
coordination structure, but also accommodates mixed coupling of y
ables yj 2 Rn are introduced to each subsystem to separate the
both variable and function, and therefore provides a possible solu-
tion measure for solving the above CSCC problem. local constraints g i and hi . Consistency constraint c (if c satisfies
y1 ¼ y2 ¼    ¼ yM , all auxiliary linking variables yi are associated
with linking variables y) is a set of consistency cjn . cjn is the consis-
3. Augmented Lagrange coordination tency between the sub-problem j and its neighbour n 2 N j .

cjn ¼ yj  yn ¼ 0 fn 2 Nj jn > jg; j ¼ 1; . . . ; M ð2Þ

3.1. ALC solution framework
MDO methods are used for solving decomposable complex sys-
tems. A system should be generally decomposed into smaller or
simpler sub-systems according to the principle of model indepen-
dency, such as decision autonomy. A suitable MDO method will
then be chosen for system coordination and solution, e.g. CO (col-
laborative optimization) for parallel decision structure and ATC
(analytical target cascading) for hierarchical decision structure.
ALC also supports hierarchical structure as ATC, although it is nor-
mally applied to accommodate more complex open structure.
ALC method to solve the complex system problem takes the fol-
lowing steps: (1) Formulate the objective system model with ALC.
(2) Introduce auxiliary variables and consistency constraints to
indicate the coupling relationship among decision elements which
serves as a base for the coordination between adjacent elements.
(3) Relax the coupling constraints and consistency constraints.
(4) Formulate the decomposed problem. (5) Determine the ALC
coordination type for solution. Based on whether the problem
has a master problem, the ALC model could be classified into two
types as centralized coordination (Fig. 2(a)) and distributed coordi-
nation (Fig. 2(b)). The first four steps are concerning with problem
transformation, while the last step is responsible for the coordina-
tion-based problem solution.
The centralized decomposition of ALC has a similar structure as
some bi-level MOD methods such as CO, while differs in the cou-
pling ways between the decomposed sub-problems. The structure
of distributed decomposition could be either hierarchical or net-
worked, and the coupling ways are also flexible as the centralized
decomposition.
N j is defined as the number of subsystems which are linked to sub-
problem j through consistency constraint. n > j ensures that only
one of cjn and cnj which are linearly correlated is in a consistency
constraint.
3.2.3. Constraints relaxation
The second step is to relax the coupling constraints h0 ; g 0 and
consistency constraint c. q ¼ ½c; g 0 þ x0 ; h0  is defined as the vector
of constraints. x0 P 0 is the relaxing variable which turn the
inequality constraints g 0 into equality constraints. Lagrange pen-
alty function is used to relax constraints q:
/ðqÞ ¼ v T q þ jjw  qjj22 ð3Þ
v ; w are the vector of penalty parameters which should be suitably
selected,  represents Hadamard Product.
3.2.4. Sub-problem formulation
The third step is to formulate the decomposed problem. The for-
mulation of sub-problem P j is as follows:
min f j ðyj ; xj Þ þ f 0 ðy1 ; x1 ; . . . ; yM ; xM Þ þ /ðqðy1 ; x1 ; . . . ; yM ; xM ; x0 ÞÞ ð4Þ
xj
s:t g j ðyj ; xj Þ 6 0
hj ðyj ; xj Þ ¼ 0
Here we define xj ¼ ½yj ; xj ; x0 . For the centralized decomposition
which has a master problem, the original problem is decomposed
and relaxed into a master problem P0 and a number of sub-
problems Pj ; j ¼ 1; . . . ; M. In the master problem P0 , only the
 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55 47

Centralized coordination Start Distributed coordination Start

Parameter Initialization Parameter Initialization

Solve Master Problem P0

Solve … Solve … Solve
Subproblem P1 Subproblem Pi Subproblem PM

Solve … Solve … Solve

Subproblem P1 Subproblem Pi Subproblem PM
No Inner loop
converged?

Update penalty No Iteration

Update penalty No Outer loop
parameters Converged?
parameters converged?

Output optimal solution Output optimal solution

Stop Stop

(a) Centralized coordination (b) Distributed coordination

Fig. 2. ALC coordination types.

penalty terms depend on the variables y and x0 in the master j and its neighbors that have a lower index n < j; n 2 Nj . T j is the
problem. The master problem P 0 is given by: selection matrix, which assigns a sub-problem to each part of y.
cjn ¼ Sjn yj  Sjn yn ¼ 0; Sjn is selection matrix, Sjn – Snj .
X
M
min f 0 ðy; x1 ; . . . ; xM Þ þ /c;j ðcj ðy; yj ÞÞ
T
x0 ¼½yT ;xT0  j¼1
3.2.5. ALC coordination
þ /g ðg 0 ðy; x1 ; . . . xM Þ; x0 Þ þ /h ðh0 ðy; x1 ; . . . xM ÞÞ ð5Þ Coordination algorithm is used to solve the problem. It has two
tasks. The first one is to select the appropriate penalty parameters
Among them:
v ; w, while the second one it to coordinate the coupling of sub-
X
M X
M problems. To fulfill the tasks, two nonlinear programming tech-
/c;j ðcj Þ ¼ /c ðcÞ ¼ v Tc c þ jjwc  cjj22 ¼ /c;j ðy; yj Þ niques are adopted. Multiplier method is used to set penalty
j¼1 j¼1
parameters for the outer loop, while alternating optimization
X
M method (also called BCD, block coordinate descent algorithm) is
¼ v Tc;j ðSj y  yj Þ þ jjwc;j  ðSj y  yj Þjj22 used to solve a series of independent sub-problems (possibly in
j¼1
parallel) in inner loops. Usually, the above coordination algorithm
Sj is selection matric. Each sub-problem P j is linked with the master is called IM (inexact inner loop method). The specific steps are as
problem P0 , with the following form: follows:

min f 0 ðy; x1 ; . . . ; xM Þ þ f j ðyj ; xj Þ þ /c;j ðcj ðy; yj ÞÞ Step1: Set initial values to the penalty parameters.
T
xj ¼½yTj ;xTj 
Step2: For inner loops, solve the decomposed problem with fixed
þ /g ðg 0 ðy; x1 ; . . . xM Þ; x0 Þ þ /h ðh0 ðy; x1 ; . . . xM ÞÞ ð6Þ penalty parameters by BCD algorithm.
s:t: g j ðyj ; xj Þ 6 0; j ¼ 1; . . . ; M Step3: Stop the algorithm if the convergence conditions are satis-
fied. Otherwise go to Step 4.
hj ðyj ; xj Þ ¼ 0; j ¼ 1; . . . ; M
Step4: For outer loops, update the penalty parameters with mul-
For the distributed decomposition, the original problem is tiplier method and return to Step 2.
decomposed and relaxed into M sub-problems P j ; j ¼ 1; . . . ; M,
and P j is given by: In the outer loop, the method of multipliers updates the esti-
mates of the Lagrange multipliers using (8) and (9) for iteration
min f 0 ðT 1 y1 ; x1 ; . . . ; T M yM ; xM Þ þ f j ðyj ; xj Þ k + 1.
T
xj ¼½yTj ;xTj ð;xT0 Þ
X X
þ /c;jn ðcjn ðyj ; yn ÞÞ þ /c;jn ðcjn ðyj ; yn ÞÞ v kþ1 ¼ v k þ 2wk  wk  qk ð8Þ
fn2Nj jn>jg fn2Nj jj<ng
(
þ /g ðg 0 ðT 1 y1 ; x1 ; . . . ; T M yM ; xM Þ; x0 Þ ð7Þ wki jqki j 6 cjqk1 j
i
wkþ1 ¼ ð9Þ
þ /h ðh0 ðT 1 y1 ; x1 ; . . . ; T M yM ; xM ÞÞ i
bwki jqki j > cjqk1 j
i
s:t: g j ðyj ; xj Þ 6 0; j ¼ 1; . . . ; M
hj ðyj ; xj Þ ¼ 0; j ¼ 1; . . . ; M Here, k is the iterations. qk is the value of the linking constraint q at
the termination of the kth inner loop. wi is the penalty weights of
where the relaxing variables x0 are included only in sub-problem the ith linking constraint qi . b > 1 is the amplification coefficient
PM . The consistency constraint penalty of Pj includes only terms that of penalty weight w. Normally, 0 < c < 1, while 2 < b < 3. The outer
depend on yj and hence consists of two parts. The first part is asso- loop has two convergence conditions. Formula (10) means the
ciated with the consistency constraints between subsystem j and its difference of the maximal linking constraint values between two
neighbors that have a higher sub-system index n > j. The second consecutive outer loop iterations should be smaller than certain
part accounts for the consistency constraints between sub-system user-defined termination tolerance e; e > 0. Formula (11) means
48 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55

that the maximal linking constraint violation must also be smaller multiplier method to coordinate the master problem and the
than the tolerance e. sub-problems.

jjqk  qk1 jj1 < e ð10Þ

4. ALC modeling for cluster supply chain
k
jjq jj1 < e ð11Þ
4.1. Model description
For the fixed penalty parameters v ; w, the alternating optimi-
zation method in the inner loop continuously solves sub-problems. Without losing generality, this paper considers a single-period
The convergence condition is that the difference of two consecu- CSCC problem initiated by M3 with the assumptions given in
tive inner loop iterations is smaller than a user-defined termina- Table 1, as shown in Fig. 3. The objective is to maximize the total
tion tolerance einner > 0. profit of the whole CSC through configuring a cluster supply chain
with both the single supply chains (i.e. SSC1 to SSCn) and indepen-
jF n  F n1 j dent suppliers (O1 to Om) from the cluster, through simultaneously
< einner ð12Þ
jF n j þ 1 considering horizontal order subcontracting and vertical compo-
nent sourcing. Due to the competitive nature among single supply
where chains, they require decision autonomy to protect the enterprise
X
M privacy. The alliance leader L i.e. ðM 3 Þ owns the highest level deci-
Fðy1 ; x1 ; . . . ; yM ; xM ; x0 Þ ¼ f ðxj ; yj Þ þ f 0 ðy1 ; x1 ; . . . ; yM ; xM Þ sion right, to determine the way of order subcontracting among
j¼1 single supply chains. Each single supply chain has the independent
þ /ðqðy1 ; x1 ; . . . ; yM ; xM ; x0 ÞÞ ð13Þ decision right, which is owned by the manufacturer fM 1 ; . . . ; M n g.
Each independent supplier has its independent decision right. In
einner > 0 is tolerance limits. Normally, we set einner ¼ e=100. n Fig. 3, all the real-line circles represent enterprise owning decision
denotes the inner loop iteration number. rights, and thus form the decision elements in the ALC decision
Although the above algorithm could converge with any initial structure, while ORM and ORO are configuration nodes representing
weight of positive values, the performance of the multiplier multiple selection (e.g. subcontracting and sourcing) relationships
method of outer loop depends on the choice of the initial weights between downstream and upstream levels.
w. In literature (Tosserams et al., 2008) the initial weights are Notations of the problem are given in Table 2.
v ¼ 0, and for all the weights w ¼ x; / ¼ x2 qT q. The order being subcontracted to the ith single supply
sffiffiffiffiffiffiffiffiffiffi chain may result in a component demand exceeding the
a j~f j production capacity of the chain’s private supplier, i.e.
x ¼ ~0T ~ ð14Þ Q i < D  PQ L 6 PQ i ði – LÞ. Therefore, the manufacturer of the
q q
single chain will sourcing component from other independent
where ~f and q
~ are the estimates of the objective function and the suppliers, resulting in a complex collaboration network. As the
linking constraint respectively. Where 0 < a0 < 1, for example, component sourcing follows step prices, i.e. the higher the
a0 ¼ 0:1. quantity, the lower the cost. We have the sales prices function
As a special case of inexact inner loop method, when the inner of ith single supply chain as Pri ¼ bPri ð1  di Þ þ 0:5c, and that of
loop has only one iteration (BCD is executed one time), the method component price of the kth independent supplier as
is called ADMOM (alternating direction method of multipliers) Prsk ¼ bPrsk ð1  hk Þ þ 0:5c, where hk is the ratio of the sourcing
(Tosserams, Etman, & Papalambros, 2006; Li, Lu, & Michalek, quantity from the kth independent supplier to its production
2008). The coordination algorithm utilizes alternating direction capacity Q k , and bxc is the largest integer smaller than x.

OR M

Single Supply Chain Alliance Supply Chain Single Supply Chain

(SSC1) (SSCL) (SSCn)

M1 ML Mn
…… dL …… dn
d1

S1 SL Sn
yL1…yLm
y11…y1m yn1…ynm
OR O

O1 …… Om
Independent Suppliers

Fig. 3. Decentralized CSCC model.

 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55 49

Table 2
Notation and definition.

i Order of single supply chains Qi Capacity constraint of the private supplier in the ith single supply chain
L Alliance leader SC i Unit component cost of the private supplier in the ith single supply chain
di Order proportion subcontracted to the ith single supply chain Pi Unit production cost of the manufacturer in the ith single supply chain
D Total order quantity taken by the alliance leader yLk Sourcing quantity from the kth independent supplier by alliance leader’s
manufacturer
Pri Unit product sales price of the ith single supply chain Prsk The unit component sourcing price from the kth independent supplier
Prsi Unit component sourcing price from the private supplier of the ith single SC k Unit component cost of the kth independent supplier
supply chain
mpi Total sales profit of the ith single supply chain Si The private supplier in the ith single supply chain
Ok The kth independent supplier Mi The manufacturer in the ith single supply chain
k Order of independent supplier yik Sourcing quantity from the kth independent supplier by the ith single
supply chain
Qk Capacity constraint of the kth independent supplier mspk Total sales profit of the kth independent supplier
PQ i Capacity constraint of the kth manufacturer mpL The total sales profit of the alliance leader
n Total number of the single supply chains in the cluster m The total number of independent suppliers

4.2. ALC modeling supply chains respectively. Constraint (19) defines the order picking
requirement of single supply chains. Constraint (20) defines the
Following the ALC solution procedure given in Section 2 as well remained order by the alliance leader for itself.
as the assumptions, a profit maximizing model of a cluster supply The coupling relationships among the three types of decision
chain could be established as follows. elements in the ALC model are depicted in Fig. 4.
X
n X
m As can be seen from Fig. 4, the alliance leader element and a sin-
max ðmpL Þ þ max ðmpi Þ þ max ðmspk Þ ð15Þ gle supply chain’s manufacturer element are not only coupled by
d1 ;d2 ;...;dn di ;yi1 ;yi2 ;...;yim y1k ;y2k ;...;ynk
yL1 ;yL2 ;...;yLm i–L k¼1 the linking variable di , but also by the constraint
P
1  ni–L di  dL ¼ 0. Similarly, there’s also a coupling constraint
where P
Ddi  Q i ¼ m k¼1;i–L yik shared between a single supply chain’s man-
X
n
ufacturer element and an independent supplier element, while alli-
mpL ¼ DdL PrL þ DðPrL  bPri ð1  di Þ þ 0:5cÞdi  ½PrsL Q L þ PL DdL
i–L
ance leader element and the independent supplier element are
Xm     coupled by the linking variable yLk and linking constraint
y P P
þ Prsk 1  Lk þ 0:5 yLk  Dð1  ni–L;i¼1 di Þ  Q L ¼ m k¼1 yLk . Based on these coupling relation-
k¼1
Qk
ships, the local decision model of the three-level CSC enterprises
could be formulated with ALC coordination in the following
mpi ¼ Ddi bPri ð1  di Þ þ 0:5c  ½Prsi Q i þ Pi Ddi sections.
Xm     In the following discussion, in order to offer adequate autonomy
y
þ Prsk 1  ik þ 0:5 yik  for all the suppliers so as to maintain enough flexibility for the
k¼1
Qk overall decision CSCC structure, we separate the decision models
    X of private supplier from the manufacturer, although the latter in
n    
y y this paper takes over the decision right of the former.
mspk ¼ yLk Prsk 1  Lk þ 0:5 þ Prsk 1  ik þ 0:5 yik
Qk i–L
Qk
! 4.2.1. ALC formulation for alliance leader decision element
X
n
 SC k yLk þ yik The objective of the alliance leader is shown by function (21).
i–L The total profit is composed of the sales profit from both its own
supply chain and the order subcontracting to other single supply
With the constraints,
chains, i.e. the first two items in the function. The other items rep-
X
n resent the costs of production and sourcing, as well as the relaxed
di þ dL ¼ 1; 0 6 di < 1 ð16Þ constraints.
i–L
X
n

X
m max mpL ¼ dL DPrL þ DðPrL  bPri ð1  di Þ þ 0:5cÞdi
d1 ;d2 ;...;dn
DdL  Q L ¼ yLk ; 0 6 yLk 6 Q k ð17Þ yL1 ;yL2 ;...;yLm i–L
k¼1
"    #
Xm
yLk
 PrsL Q L þ PL DdL þ yLk Prsk 1  þ 0:5
X
m
k¼1
Qk
Ddi  Q i ¼ yik ; 0 6 yik 6 Q k ð18Þ
X
m X
n
k¼1;i–L
 uðyLk Þ  uðdi Þ  uðh0 Þ
k¼1 i–L
Q i < Ddi 6 PQ i ; ði – LÞ ð19Þ ð21Þ

DdL ¼ PQ L ð20Þ The constraints are as follows.

The objective function (15) comprises three parts. The first is the X
m

maximum profit of alliance leader, and the second is the maximum DdL  Q L ¼ yLk ð22Þ
k¼1
total profit of all the single supply chains, and the third is the
maximum total profit of all the independent suppliers. Constraint
0 6 yLk 6 Q k ð23Þ
(16) gives the order subcontracting scope. Constraints (17) and
(18) define the components’ multiple sourcing from independent
DdL ¼ PQ L ð24Þ
suppliers by the alliance leader’s supply chain and other single
50 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55

X
n
4.2.3. ALC formulation for supplier decision element
h0 ¼ 1  di  dL ¼ 0 ð25Þ
The objective of the manufacturer is shown by function (38).
i–L
The first item is the sales profit of the independent supplier, while
other items represent the sourcing cost and the constraints
0 6 di < 1 ð26Þ
relaxation.
Couplings are relaxed as follows.    
y
max mspk ¼ yLk Prsk 1  Lk þ 0:5
uðyLk Þ ¼ myLk ðyLk  ylLk Þ þ jjwyLk  ðyLk  ylLk Þjj22 ð27Þ y1k ;y2k ;...;ynk Qk
Xn    
y
þ yik Prsk 1  ik þ 0:5
uðdi Þ ¼ mdi ðdi  dli Þ þ jjwdi  ðdi  dli Þjj22 ð28Þ
i–L
Qk
!
X n Xn
uðh0 Þ ¼ mh0 h0 þ jjwh0  h0 jj22 ð29Þ  SC k yik þ yLk  uðyik Þ  uðyLk Þ ð38Þ
i–L i–L
The meanings of constraints (22)–(26) are the same as in the system
model. Function (27) is the relaxation of sourcing constraint, ylLk is The constraints are as follows.
the backtrack value from upstream independent suppliers to the alli- 0 6 yLk 6 Q k ð39Þ
ance leader, and myLk ; wyLk are the relaxation parameters. Function
(28) represents the constraint coupling relaxation of order subcon- 0 6 yik 6 Q k ð40Þ
l
tracting ratio, where di is the backtrack value from subcontracting
single supply chains to the alliance leader, and mdi ; wdi are the relax- X
n
ation parameters. Function (29) is the coupling relaxation of equality 0 6 yLk þ yik 6 Q k ð41Þ
constraint, and mh0 ; wh0 are the relaxation parameters. i–L

Couplings are relaxed as follows.

4.2.2. ALC formulation for manufacturer decision element
The objective of the manufacturer is shown by function (30). uðyik Þ ¼ myik ðyik  yuik Þ þ jjwyik  ðyik  yuik Þjj22 ð42Þ
The first item is the sales profit of the single supply chain, while
other items represent the costs of production and sourcing, as well uðyLk Þ ¼ myLk ðyLk  yuLk Þ þ jjwyLk  ðyLk  yuLk Þjj22 ð43Þ
as the constraints relaxation.
Constraints (39) and (40) represent the scopes of backtrack values
max mpi ¼ Ddi bPri ð1  di Þ þ 0:5c from independent suppliers to the alliance leader and subcon-
di ;yi1 ;yi2 ;...;yim
" tracted single supply chains respectively, while function (41) repre-
Xm    #
yik sents corresponding constraint. Functions (42) and (43) represent
 Prsi Q i þ P i Ddi þ yik Prsk 1  þ 0:5 the coupling relaxation of the sourcing quantity, where yuik and yuLk
k¼1
Qk
are the sourcing quantity from independent suppliers by the single
X
m
 uðyik Þ  uðdi Þ  uðh0 Þ supply chain’s manufacturer and alliance leader respectively.
k¼1 myik ; wyik ; myLk ; wyLk are the relaxation parameters.
ð30Þ
5. ALC solution
The constraints are as follows.
X
n This section will apply ALC for two kinds of CSCC problems
h0 ¼ 1  di  dL ¼ 0 ð31Þ based on the solution framework given in Section 3.1. The first
i–L one is an empirical CSC design problem borrowed from Li et al.
(2012), aiming to demonstrate the ALC effectiveness for accommo-
X
m
Ddi  Q i ¼ yik ð32Þ dating CSC model, i.e. horizontal and vertical collaboration among
k¼1;i–L supply chain nodes. The proved ALC method will then be applied to
the second problem which has been discussing in this paper.
Q i < Ddi 6 PQ i ; ði – LÞ ð33Þ
5.1. Empirical problem considering cross-chain cooperation
0 6 yik 6 Q k ð34Þ
5.1.1. ALC modeling
Couplings are relaxed as follows. This problem is given in Li et al. (2012), which is a cluster
 2 supply chain with cross-chain horizontal cooperation. To apply
uðyik Þ ¼ v yik ðyik  ylik Þ þ wyik  ðyik  ylik Þ2 ð35Þ
ALC solution, we assume the two single supply chains in the CSC
own their autonomous decision rights, xjti and yjti are product i’s
uðdi Þ ¼ v di ðdi  dui Þ þ jjwdi  ðdi  dui Þjj22 ð36Þ coupling linked variables between single supply chains j and
tðj ¼ 1; 2; t ¼ 1; 2; j – tÞ, and d1i ; d2i represent product i’s market
uðh0 Þ ¼ v h0 h0 þ jjwh0  h0 jj22 ð37Þ demands for single supply chains 1 and 2 respectively. Other
assumptions and notations remain the same as in the original
The meanings of constraints (31)–(33) are the same as in the system problem. The resulted ALC model of the original problem is given
model. Constraint (35) is the relaxation of sourcing quantity, where as follows.
ylik is the backtrack value from upstream independent suppliers to
the manufacturer, myik ; wyik are the relaxation parameters. Function (1) ALC model of the whole CSC
(36) represents the constraint coupling relaxation of order subcon-
u X
2 X
2
tract ratio, where di is the backtrack value from subcontracting sin-
gle supply chains to the manufacturer, and mdi ; wdi are the
min TC þ uðTC j Þ þ uðdji Þ ð44Þ
TC 1 ;TC 2 ;d1i ;d2i
j¼1 j¼1
relaxation parameters. Function (37) is the coupling relaxation of
equality constraint, and mh0 ; wh0 are the relaxation parameters. where
 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55 51

X
I

L capji xji 6 M j ð53Þ

i¼1
n

∑d
i ≠L
i
+ dL = 1
X
I
yji 6 Pj ð54Þ
OR M i¼1
d1 dn
m

M n Ddn- Q n = ∑ ynk
m X
I
Dd1 - Q1 = ∑ y1k M1 … … k=1
hji yji 6 INV j ð55Þ
k=1
i¼1

yn1 OR S y
yi1 nm X
I
y11 yim uji P 1 ð56Þ
y1m
S1 O1 Sn i¼1
…… Om
X
I
zji P 1 ð57Þ
Fig. 4. Coupling relationships among decision elements. i¼1

di ¼ d1i þ d2i ð45Þ yji  axji uji  xjti 6 0 ð58Þ

TC ¼ TC 1 þ TC 2 ð46Þ xji P 0; yji P 0; uji 2 f0; 1g; v ji 2 f0; 1g;

Couplings are relaxed as follows: zji 2 f0; 1g; wji 2 f0; 1g ð59Þ

uðTC j Þ ¼ v TC j ðTC j  TC Lj Þ þ jjwTCj  ðTC j  TC Lj Þjj22 ð47Þ Couplings are relaxed as follows.

 2 uðdji Þ ¼ v dji ðdji  dUji Þ þ jjwdji  ðdji  dUji Þjj22 ð60Þ

L  L 
uðdji Þ ¼ v dji ðdji  dji Þ þ wdji  ðdji  dji Þ ð48Þ
2
uðTC j Þ ¼ v TCj ðTC j  TC Uj Þ þ jjwTC j  ðTC j  TC Uj Þjj22 ð61Þ
Eq. (45) means the demand to product i which includes the
demands of single supply chains 1 and 2, i.e. d1i and d2i . Eq.
xjji ¼ ð1  aÞv ji xji ð62Þ
(46) means that the CSC cost is equal to the sum of the costs
of single supply chain 1 and 2, i.e. TC 1 and TC 2 . TC L1 ; TC L2 rep-
resent the backtrack values from single supply chains 1 and 2 yjji ¼ ð1  aÞwji yji ð63Þ
respectively. Function (47) represents the coupling relaxation  
of the costs of the single supply chain, while mTC j ; wTCj are the aj ¼ xjji ; yjji ; xjti ; yjti  xtji ; ytji ; xtti ; ytti ð64Þ
relaxation parameters. Function (48) represents the coupling
L
relaxation of the demand dji . dji is the backtrack demand uðaj Þ ¼ v aj aj þ jjwaj  aj jj22 ð65Þ
value from the single supply chain jðj ¼ 1; 2Þ, while mdji ; wdji
are the relaxation parameters.
(2) ALC model of single supply chain j The meaning of (50) remains the same as in the original litera-
ture. The meaning of (51) represents the outer demand quantity for
min TC j þ uðTC j Þ þ uðdji Þ þ uðaj Þ ð49Þ the single supply chain j product i. Constraints (52)–(59) remain
j j
xji ;yji ;x ;y
ti ti
uji ;zji ;v ji; wji the same meanings as the original literature. The meaning of (60)
is the coupling relaxation of outer demand quantity ðdji Þ for the
where U
single supply chain j product i. dji represents the cascading value
X
I X
I from the cluster supply chain system, and mdji ; wdji are the relaxa-
TC j ¼ axji direct cji þ ð1  aÞxji cross cji tion parameters. Function (61) represents the coupling relaxation
i¼1 i¼1 of the cost TC j of chain j. TC Uj represents the cascading value from
X
I X
I XI the cluster supply chain system, and mTC j ; wTC j are the relaxation
þ mji xji þ ayji direct pji þ ð1 parameters. Eqs. (62) and (63) represent the transportation vol-
i¼1 i¼1 i¼1
umes of shipping product i from supplier to manufacturer and
X
I
from manufacturer to retailer at single supply chain j at the other
 aÞyji cross pji þ uji direct f ji
i¼1
single supply chain respectively. Eq. (64) represents the inconsis-
tency constraints of linked variables between single supply chains
X
I X
I
þ v ji cross f ji þ zji direct g ji j and t ðj – tÞ. Function (65) represents the coupling relaxation of
i¼1 i¼1 linked variables, maj ; waj are the relaxation parameter vectors.
X
I
þ wji cross g ji ð50Þ 5.1.2. ALC solution
i¼1 We have b ¼ 2:2; c ¼ 0:4; a0 ¼ 0:1, tolerance e ¼ 0:001, the
maximum numbers of outer loop are set to 1000. The initial values
dji ¼ ayji zji þ yjti ð51Þ of penalty factors mdji ; wdji ; mTC j ; wTC j ; maj ; waj are all set to 1.
The constraints are as follows. Parameters TC L1 ; TC L2 are both initialized by 20. The initial values
L L
of TC U2 ; TC U1 are set to 0. Parameters d1i ; d2i are both initialized
X
I
U U
xji 6 cj ð52Þ by 0. Parameters d1i ; d2i are initialized by di =2. Parameters
i¼1 x21 ; x22 ; y21 ; y22 x11 ; x12 ; y211 ; y212 are set to
1 1 1 1 2 2
68, 5, 54, 15, 1, 69,
52 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55

55, 13 respectively. The initial values of other parameters are the Table 4
same as in the original literature. The inner sub-problem ALC mod- Parameters setting.

els for single supply chain 1 and single supply chain 2 are solved by Parameters Parameters values Parameters Parameters values
the genetic algorithm (GA) used in the original literature respec- PQ i 720, 980, 790, 960, 840 Qk 700, 250, 540
tively. The simulation is conducted by Matlab 7.11 in a PC with Qi 400, 250, 170, 240, 330 Prsi 30, 25, 32, 36, 40
2.5 GHz CPU and 4 GB RAM. The comparative results with AIO Pri 340, 450, 540, 320, 410 Prsk 25, 21, 24
(All-in-One) model in the original literature are given in Table 3 Pi 110, 200, 212, 230, 260 SC k 12, 10, 13

if the allocation proportion of cooperation a is 0.8.

Comparative results from Table 3 show that, when a is set to the independent suppliers are selected through multiple sourcing, with
other value which is less than or equal to 1, ALC solution for solving sourcing quantity as 222, 72 and 326. Single supply chain 1 and 4
CSC model with vertical and horizontal cooperation is as good as also employ multiple sourcing from independent suppliers, and
AIO method, e.g. GA in the original literature. the sourcing quantity is 276, 42, 2 and 102, 137, 10 respectively.
Second, the ALC method proposed by this paper offers the identical
5.2. CSCC problem computational accuracy as by Lingo 11, while offers much shorter
optimization time. Even considering the network transmission time
5.2.1. Parameter setting when the distributed ALC model is put into practice, ALC is still com-
Likewise we have b ¼ 2:2; c ¼ 0:4; a0 ¼ 0:1, tolerance petitive not to say the advantage of decision autonomy protection.
e ¼ 0:001, the maximum iteration of outer and inner loop are set
to 1000 and 10 respectively. The initial values of penalty factors 5.2.4. Sensitivity analysis
v ; w are both set to 1. The inner iteration sub-problem uses inte- For the alliance leader, along with the order’s quantity increase,
rior point methods for calculation. the formation of the optimal CSCC will change accordingly.
Obviously, in case of a small order, M3’s single supply chain (i.e.
5.2.2. Experiment simulation M3–S3) is enough. While for a large order, the alliance leader has
We set the number of single supply chains in the cluster as to collaborate with other sibling supply chains and/or independent
n = 5, and the number of independent suppliers m = 3, and the alli- suppliers to configure a CSCC to fulfill. In order to find out the con-
ance leader is M3. The alliance leader takes the order as D = 2000, dition in which a manufacturer has to resort to CSC and what the
while other parameters are shown in Table 4, where most appropriate CSCC is, a set of sensitivity analysis will be
i ¼ 1; 2; 3; 4; 5; k ¼ 1; 2; 3. conducted.
The initial values set for other parameters in the distributed ALC In the experiment, we will increase the order quantity from 0 to
coordination are as follows. a very large quantity (exceeds the summation of all the indepen-
l l l l l l
dent suppliers’ production capacities and three private suppliers’
d ¼ ½d1 ; d2 ; d3 ; d4 ; d5  ¼ ½470=D; 0; 790=D; 340=D; 0; production capacities). The observed formation change of CSCC
ylL ¼ ½ylL1 ; ylL2 ; ylL3  ¼ ½400; 0; 220; yl1 ¼ ½yl11 ; yl12 ; yl13  ¼ ½100; 50; 40; could be described as follows.
yl2 ¼ ½yl21 ; yl22 ; yl23  ¼ ½0; 0; 0; yl4 ¼ ½yl41 ; yl42 ; yl43  ¼ ½70; 10; 8;
(1) If the order quantity is less than M3’s production capacity
yl5 ¼ ½yl51 ; yl52 ; yl53  ¼ ½0; 0; 0; and the required component quantity is less than the private
myLk ¼ myik ¼ mdi ¼ mh0 ¼ v ¼ 1; wyLk ¼ wyik ¼ wdi ¼ wh0 ¼ w ¼ 1: supplier’s capacity, i.e. D 6 170, M3 own supply chain is
enough to fulfill the order. As can be seen in the first left
The decision values are represented by d and y in the model. block in Fig. 5(a).
2 3 (2) If the order quantity is less than M3’s production capacity,
yL1 yL2 yL3
6 7 while the private supplier cannot satisfy the resulted com-
6 y11 y12 y13 7
6 7 ponent requirement, i.e. 170 < D 6 790, M3 will become an
d ¼ ½d1 ; d2 ; dL ; d4 ; d5 ; y¼6
6 y21 y22 y23 7
7 alliance leader, and source component from independent
6 7
4 y41 y42 y43 5 suppliers through multiple sourcing strategy. The CSCC
y51 y52 y53 strategy changes are shown in Fig. 5(a), while the sourcing
quantity from independent suppliers is shown in Fig. 5(b).
The simulation is also conducted by Matlab 7.11 in a PC with
(1) If 170 < D 6 420; O2 will be selected as the only inde-
2.5 GHz CPU and 4 GB RAM. The comparative results with Lingo
pendent supplier.
11 are given in Table 5.
(2) If 420 < D 6 560, M3will conduct multiple sourcing from
both O2 and O3 .
5.2.3. Result analysis
(3) If 560 < D 6 710; O3 will replace O2 to be selected as the
Two aspects of results can be analyzed out of Table 5. First, single
only independent supplier. This is because the larger
supply chain 1 and 4 are selected for subcontracting, while three
supply capacity of O3 with the step-pricing function
Table 3 which results in lower total sourcing cost.
Results comparison between ALC and AIO in the original literature (Li et al., 2012). (4) If 710 < D < 750, M3will conduct multiple sourcing from
both O2 and O3 .
Optimization Objective function Optimization variables
method value (TC) (5) If 750 6 D 6 770, M3will transfer from multiple sourcing
2 3 from single sourcing from O1 , due to the same reason as
x11 ; x12 ; x21 ; x22 ; y11 ; y12 ; y21 ; y22
4 u11 ; u12 ; u21 ; u22 ; z11 ; z12 ; z21 ; z22 5 explained in the above (3).
v 11 ; v 12 ; v 21 ; v 22 ; w11 ; w12 ; w21 ; w22 (6) If 770 < D 6 790, M3will conduct multiple sourcing from
2 3 both O1 and O3 .
ALC 44605.597 0; 350; 350; 30; 70; 286; 280; 94
4 0; 1; 1; 1; 1; 1; 1; 1 5 (3) If the order quantity exceeds M3’s production capacity while
0; 1; 1; 1; 1; 1; 1; 1
still falls into the scope of the cluster’s total capacity, i.e.
2 3
AIO (Li et al., 44605.600 0; 350; 350; 30; 70; 286; 280; 94 790 < D 6 2290, M3 will become an alliance leader, and sub-
2012) 4 0; 1; 1; 1; 1; 1; 1; 1 5
0; 1; 1; 1; 1; 1; 1; 1
contracts a portion of the order to other single supply
chain(s). According to the subcontracting quantity, the other
 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55 53

Table 5
Results comparison between ALC and Lingo 11.

Methods Optimization variable d Optimization variable y Objective function value Comput speed
2 3
ALC [0.360, 0, 0.395, 0.245, 0] 222; 72; 326 mpL ¼ 424386:3 5 outer iterations in 3 min
6 276; 42; 2 7 P5
6
6 0; 0; 0
7
7 i–L maxdi ;yi1 ;yi2 ;...;yim ðmpi Þ ¼ 35143:87
6 7 Pm
4 102; 137; 10 5 k¼1 maxyk1 ;yk2 ;...;ykn ðmspk Þ ¼ 1849:784

0; 0; 0
2 3
Lingo11 [0.360, 0, 0.395, 0.245, 0] 222:15; 71:42; 326:41 The maximum profit of the cluster supply chain is 461380.0 431,127 Iterations in 11 min
6 276; 41:66; 2:1 7
6 7
6 0; 0; 0 7
6 7
4 101:9; 136:88; 10:3 5
0; 0; 0

(a) Configuration strategies with order quantity (b) Sourcing quantity of independent suppliers

Fig. 5. CSCC formations with the alliance leader’s order quantity.

Fig. 6. Changes of maximal profits of cluster participants with order quantity.

single supply chain(s) will determine whether to source
from its private supplier or to adopt multiple sourcing from
other independent suppliers. The changes could be
described as follows.
(1) If 0 < D  790 6 240, M3 will first of all source compo-
nents from its private supplier and independent suppli-
ers O1 and O2 , and then collaborates with M4 for order
subcontracting, d4 ¼ D790
D
. As 240 is the M4’s private
supplier’s capacity, M4 does not need to source from
independent suppliers.
(2) If 240 < D  790 6 960, the subcontracted quantity to
M4 will exceed the private supplier’s capacity, and thus
M4 will adopt multiple sourcing from independent sup-
pliers O1 ; O2 and O3 . As these independent suppliers are
also serving M3, their capacities will be shared between
M3 and M4.
(3) If 960 < D  790 6 1410, M3 will collaborate with M1
and M4 for order subcontracting, both of which will
adopt multiple sourcing from independent suppliers
O1 ; O2 and O3 .
54 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55
(4) If 1410 < D  790 6 1500, M3 will collaborate with M1
and M5 instead for subcontracting order, both of which
will adopt multiple sourcing from independent suppliers
O1 ; O2 and O3 .
(4) If 1500 < D  790, the cluster cannot fulfill the complete
order, and only portion of the order could be digested by
the available enterprises. The customer will have to wait
for the capacity releasing, or the cluster enterprises have
to expand their current capacities if such cases happen
frequently.
In the cluster, if the order quantity exceeds the production
capacity of the alliance leader, the changing curves of the maximal
profits of (1) the whole CSC, (2) the alliance leader, (3) the subcon-
tracted single supply chains, and (4) the independent suppliers are
given in Fig. 6. It shows that, with the typical configuration policies
and the generally accepted assumptions in the cluster, the profits
changing of all the cluster participants are consistent, which con-
tributes to the positive and sustainable collaboration in the cluster.
6. Conclusions
This paper has conducted a complete analysis for the overall
operations of cluster supply chains, especially the cluster supply
chain configuration (CSCC). A generic CSCC model comprising hor-
izontal order subcontracting with parallel single/sibling supply
chains and vertical component sourcing with upstream suppliers
has been established, which is applicable to most of the CSCC prob-
lem. Based on the analysis of the commonly required decision
autonomy of cluster enterprises, a distributed CSCC model which
enables loosely coupled decentralized decision mechanism has
been established, and Augmented Lagrangian Coordination (ALC)
has been adapted for such distributed CSCC solution. A typical
CSCC problem is adopted to demonstrate the proposed ALC solu-
tion framework and procedure, and its overall effectiveness and
efficiency are proved by comparative experiment simulations with
the classical results of Lingo. This paper has extended ALC from
theoretical level to an application level, and provides a systematic
approach comprising analysis, model, and solution techniques of
applying ALC for the optimization of real engineering and manage-
ment problems.
The major contributions of this research are as follows. First, the
typical relationships, collaboration mechanisms, and sourcing
strategies of clustered enterprises are systematically formulated
following the mature supply chain configuration model, which
provides an enabling framework and mechanism for optimal CSCC
toward dynamic customer orders. Second, dynamic collaboration
relationships among clustered enterprises are investigated with
the consideration of their pressing demands of local decision
autonomy protection. This leads to the achievement of a decentral-
ized CSCC decision model which serves as an enabling condition for
clustered enterprises to practically conduct CSCC toward order
optimization. Third, ALC is for the first time introduced to the sup-
ply chain management field to coordinate the multi-enterprise-
involved collaborative decision making process. An integral CSCC
solution procedure comprising objective model definition, system
decomposition, coupling identification and relaxation has been
proposed for a feasible coordination. Last, sensitivity analysis is
conducted to master the changing rules of the optimal CSCC along
with the dynamic customer orders, so as to provide managerial
implications for cluster participants especial leading enterprises
to manage and extend appropriate collaborations with other clus-
ter partners in a forward-looking way.
The future works of this research could be summarized as
follows. First, this paper investigated the CSCC under typical
configuration policies at current stage. However, due to the rapidly
emerged new formations of industrial clusters, various configura-
tion policies will be produced during the development of industrial
clusters, and thus a more comprehensive investigation to the var-
ious CSCC model is to be furthered and deepened in the future. Sec-
ond, this paper only considered single-period CSCC with
capacitated subcontracting and sourcing. Yet cluster supply chains
will also be possibly operated in a multi-period situation. There-
fore, it is worthwhile to study the CSCC with optimal setting of
multi-period parameters such as safety stocks. This is also a direc-
tion obtained inadequate research so far. Third, different sizes and
configuration policies will result in different CSCC decision struc-
tures, which may be solved by other MDO methods with better
performance, e.g. hierarchical structure is suitable for ATC solution.
In such cases, whether ALC still offers better performance deserves
further research.
Acknowledgements
This work was supported by the National Natural Science Foun-
dation of China (51105081, 61074146), Guangdong Natural Sci-
ence Foundation (S2012010010016), Guangdong College Talent
Import Scheme (11ZK0066), 2014 ‘‘Thousand-Hundred-Ten’’
Scheme of Guangdong Education Department, Guangzhou Pearl
River New Star Fund Science and Technology Planning Project
(2011J2200017), National Science and Technology Ministry of
China (2012BAF12B10).
References
Allison, J., & Papalambros, P. Y. (2010). Consistency constraint allocation in
augmented Lagrangian coordination. Journal of Mechanical Design, 132(7), 1–8.
Amini, M., & Li, H. T. (2011). Supply chain configuration for diffusion of new
products: An integrated optimization approach. Omega, 39, 313–322.
Beaudry, C., & Breshi, S. (2003). Are firms in clusters really more innovative.
Economics of Innovation and New Technology, 12, 325–342.
Bikram, K., Bahinipati, A. K., & Deshmukh, S. G. (2009). Horizontal collaboration in
semiconductor manufacturing industry supply chain: An evaluation of
collaboration intensity index. Computers and Industrial Engineering, 57, 880–895.
Camarinha-Matos, L. M., Afsarmanesh, H., Galeano, N., et al. (2009). Collaborative
networked organizations – Concepts and practice in manufacturing enterprises.
Computers and Industrial Engineering, 1(57), 46–60.
Fan, M., Stallaert, J., & Whinston, A. B. (2003). Decentralized mechanism design for
supply chain organizations using an auction market. Information System
Research, 14(1), 1–22.
Graves, S. C., & Willems, S. P. (2001). Optimizing the supply chain configuration for
new products. Working paper. Leaders for Management Program and A.P., Sloan
School of Management, MIT.
Graves, S. C., & Willems, S. P. (2005). Optimizing the supply chain configuration for
new products. Management Science, 51(8), 1165–1180.
Huang, G. Q., & Qu, T. (2008). Extending analytical target cascading for optimal
configuration of supply chains with alternative autonomous suppliers.
International Journal of Production Economics, 115, 39–54.
Huang, G. Q., Qu, T., Zhang, Y. F., & Yang, H. D. (2012). RFID-enabled product-service
system for automotive part and accessory manufacturing alliances. International
Journal of Production Research, 50(14), 3821–3840.
Huang, G. Q., Zhang, X. Y., & Liang, L. (2005). Towards integrated optimal
configuration of platform products, manufacturing processes, and supply
chains[J]. Journal of Production Economics, 23(2–4), 267–290.
Kawtummachai, B., & Hop, N. V. (2005). Order allocation in a multiple-supplier
environment. International Journal of Production Economics, 231–238.
Lee, H., & Whang, S. (1999). Decentralized multi-echelon supply chains: Incentives
and information. Management Science, 45(5), 633–641.
Li, H. T., & Amini, M. (2012). A hybrid optimisation approach to configure a supply
chain for new product diffusion: A case study of multiple-sourcing strategy.
International Journal of Production Research, 50(11), 3152–3171.
Li, Z., Huang, G. Q., Fang, J., & Qu, T. (2013). Ontology-based dynamic alliance
services (ODAS) in production service system. International Journal of Computer
Integrated Manufacturing, 27(2), 148–164.
Li, Y. J., Lu, Z. S., & Michalek, J. J. (2008). Diagonal quadratic approximation for
parallelization of analytical target cascading. Journal of Mechanical Design, 130,
1–11.
Li, H. T., & Womer, K. (2008). Modeling the supply chain configuration problem with
resource constraints. International Journal of Project Management, 26(6),
646–654.
 T. Qu et al. / Computers & Industrial Engineering 84 (2015) 43–55
Li, J. Z., Xiong, N. X., Park, J. H., et al. (2012). Intelligent model design of cluster
supply chain with horizontal cooperation. Journal of Intelligent Manufacturing,
23, 917–931.
Lorenzo, T., & Stefano, S. (2011). A heuristic for balancing the inventory level of
different locations through lateral shipments. International Journal of Production
Economics, 131(1), 87–95.
Pandilt, N. R., Cook, G. A. S., & Peter Swann, G. M. (2002). Comparison of clustering
dynamics in the British broadcasting and financial services industries.
International Journal of the Economics of Business, 9(2), 195–224.
Qu, T., Huang, G. Q., Chen, X., & Chen, H. P. (2009). Extending analytical target
cascading for optimal supply chain network configuration of a product family.
International Journal of Computer Integrated Manufacturing, 22(11), 1012–1023.
Qu, T., Huang, G. Q., Cung, V.-D., & Mangione, F. (2010). Optimal configuration of
assembly supply chains using analytical target cascading. International Journal
of Production Research, 48(23), 6883–6907.
Tosserams, S., Etman, L. F. P., Papalambros, P. Y., et al. (2006). An augmented
lagrangian relaxation for analytical target cascading using the alternating
direction method of multipliers. Structural and Multidisciplinary Optimization,
31(3), 176–189.
Tosserams, S., Etman, L. F. P., & Rooda, J. E. (2007). An augmented Lagrangian
decomposition method for quasi-separable problems in MDO. Structural and
Multidisciplinary Optimization, 34(3), 211–227. http://dx.doi.org/10.1007/
s00158-006-0077–z.
55
Tosserams, S., Etman, L. F. P., & Rooda, J. E. (2008). Augmented Lagrangian
coordination for distributed optimal design in MDO. International Journal for
Numerical Methods in Engineering, 73(13), 1885–1910.
Tosserams, S., Etman, L. F. P., & Rooda, J. E. (2010). Multi-modality in augmented
Lagrangian coordination for distributed optimal design. Structural and
Multidisciplinary Optimization, 40(1–6), 329–352.
Walsh, W. E., & Wellman, M. P. (2003). Decentralized supply chain formation: A
market protocol and competitive equilibrium analysis. Journal of Artificial
Intelligence Research (19), 513–567.
Xiang, Wei, Faishuai, Song, & Feifan, Ye (2014). Order allocation for multiple supply-
demand networks within a cluster. Journal of Intelligent Manufacturing, 25,
1367–1376.
Yang, J., & Qin, Z. (2007). Capacitated production control with virtual lateral
transshipments. Operations Research, 55(6), 1104–1119.
Zhang, Y. F., Huang, G. Q., Sun, S. D., & Yang, T. (2014). Multi-agent based real-time
production scheduling method for radio frequency identification enabled
ubiquitous shop floor environment. International Journal of Computers &
Industrial Engineering, 76, 89–97.
Zhang, Y. F., Zhang, G., Wang, J. Q., Sun, S. D., Si, S. B., & Yang, T. (2014). Real-time
information capturing and integration framework of the internet of
manufacturing things. International Journal of Computer Integrated
Manufacturing. http://dx.doi.org/10.1080/0951192X.2014.900874.