You are on page 1of 20

International Journal of Production Research, 2019

Vol. 57, No. 24, 7567–7585, https://doi.org/10.1080/00207543.2019.1587189

Facility location and scale optimisation in closed-loop supply chain


a
Lu Zhen , Qiuji Suna , Kai Wang a,b∗
and Xiaotian Zhanga,c
a School of Management, Shanghai University, Shanghai, People’s Republic of China; b Department of Logistics and Maritime Studies,
The Hong Kong Polytechnic University, Kowloon, Hong Kong; c Oulu Business School, University of Oulu, Oulu, Finland
(Received 25 April 2018; accepted 18 February 2019)

This paper solves a strategic-level decision problem on determining the optimal location of (re)manufacturers and logistics
centres with the consideration of facility scales in an integrated closed-loop supply chain network. A two-stage stochastic
mixed-integer non-linear programming model is established to minimise the fixed cost and the expected operation costs
under uncertain demand and return. We develop an improved tabu search heuristic algorithm to solve the model. We also
design a distance-based decision rule to validate the effectiveness of the proposed model. Numerical experiments are con-
ducted to test the performance of the proposed model and the solution method. In addition, sensitivity analysis is provided
to investigate the influences of varying inspection locations and recovery rates on the final performance.
Keywords: Facility location and scale; multi-product; uncertainty; reverse logistics; remanufacturing

1. Introduction
Closed-loop supply chain (CLSC) has attracted a great deal of academic and practical attention for the social responsibility,
economic and global environmental concerns during recent years. For some companies that manufacture electronics, auto-
mobiles and large home appliances, the value of returned products that can be recovered after simple inspection/repackaging
can be staggeringly high. Reuse through simple remanufacturing/refurbishing can bring significant economic benefits to
companies (Guide and Wassenhove 2001). Thus, many companies have implemented reuse activities and achieved success,
such as Dell, GM, HP, Kodak and Xerox (Khatami, Mahootchi, and Farahani 2015). In China, with the rapid development
of the electronic commerce, increasing people are used to shopping online, the annual revenue of online retail sales reaches
1.15 trillion USD in 2016. Meanwhile, the return rate of products is greatly increased due to the quality of products, delivery
errors and distribution delays. For large retailers, such as JD (i.e. one of the largest online shopping malls in China), it is
common that their return rates are larger than 10%. The annual value of returns in JD is usually about hundreds of billions
of USD. Therefore, it is necessary for corporations to design a CLSC network which integrates the forward supply chain
with the reverse supply chain for minimising the total cost of the system.
CLSC network design problem plays a key role in timely and efficient product delivery and collection, it typically
involves determining the location of facilities and the number of product shipments between facilities. When designing a
CLSC network, companies often need to determine the scale of facilities because it greatly affects not only the fixed cost of
locating facilities, but also the unit production cost and the long-term transportation cost (Wu et al. 2017). There have been
a lot of studies on the decision problems for locating facilities in a CLSC network. However, most of the studies have not
considered the decision of the scale of facilities.
The main contributions of this paper can be summarised as follows:
(1) This paper extends the research on CLSC network design by jointly considering the location and scale of facili-
ties, capacitated facilities, uncertain customers’ demand and return, multiple products, production/processing costs,
transportation costs and holding costs.
(2) An improved tabu search (ITS) heuristic algorithm is developed to solve the challenging problem. This paper
designs a release criterion to enhance the global searching ability of the algorithm.
(3) From our computational studies, we obtain interesting managerial insights. For example, sensitivity analysis indi-
cates that inspection at logistics centre or customer locations is better than inspection at producer locations by
saving unnecessary resources and costs.

*Corresponding author. Email: chwangkai@gmail.com

© 2019 Informa UK Limited, trading as Taylor & Francis Group


7568 L. Zhen et al.

The remainder of the paper is arranged as follows. The related literature reviews are presented in Section 2. Section 3
elaborates on the problem background and description. A model is formulated in Section 4. A solution method is described
in Section 5. Computational results are reported in Section 6. Section 7 conducts the sensitivity analysis. The conclusions
and some future research directions are put forward in the last section.

2. Related works
This paper is related to a facility location problem of a CLSC network with considering the uncertainty of demand and
return. The literature on the CLSC network design problem and its variants is rich. For comprehensive reviews, we refer to
the books by (Flapper, van Nunen, and Van Wassenhove 2006; Lebreton 2006; Ferguson and Souza 2016), and the survey by
(Srivastava 2007). In this section, we summarise the related literature into two aspects: the CLSC models under uncertainty
and the TS algorithm used in CLSC network design.

2.1. CLSC models under uncertainty


Many researchers have proposed various models to assist decisions on the structure of a CLSC logistics network. The
mixed-integer linear programming (MILP) models have been widely formulated in this area for optimising decisions on
CLSC network design, such as Salema, Barbosa-Povoa, and Novais (2007), El-Sayed, Afia, and El-Kharbotly (2010), Amin
and Zhang (2013), Jindal and Sangwan (2014), Chen, Chan, and Chung (2015), Üster and Hwang (2016), and Srinivasan
and Khan (2018). The methodology of MINLP modelling has also been used for optimising the CLSC network decisions.
For example, studies by Abdallah, Diabat, and Simchi-Levi (2012), Zhang, Berenguer, and Shen (2014), Diabat, Abdallah,
and Henschel (2015), and Radhi and Zhang (2016). In addition, researchers have established other models to study the
CLSC network design problems. For example, dynamic CLSC model (Huang, Yan, and Qiu 2009), non-linear integer
programming (NIP) model (Zhang and Unnikrishnan 2016; Patne et al. 2018), and stochastic programming model (Badri,
Ghomi, and Hejazi 2017).
The solution methodologies for the uncertainty model can be summarised as robust optimisation (Pishvaee, Rabbani,
and Torabi 2011; Mirmajlesi and Shafaei 2016; Yousefi-Babadi et al. 2017), fuzzy optimisation (Ramezani et al. 2014;
Zhalechian et al. 2016; Kang, Wang, and Ma 2017), stochastic optimisation (Listeş and Dekker 2005; El-Sayed, Afia, and
El-Kharbotly 2010; Abdallah, Diabat, and Simchi-Levi 2012), and scenario-based optimisation (Laporte, Louveaux, and van
Hamme 1994; Pishvaee, Jolai, and Razmi 2009; Amin and Zhang 2013; Özceylan and Paksoy 2013; Khatami, Mahootchi,
and Farahani 2015; Üster and Hwang 2016; Jerbia et al. 2018). A more detailed review of recent literature in the area of
CLSC network design under uncertainty is illustrated in Table 1.
Based on the above studies, it can be observed that although many CLSC models have been studied, few of them have
considered the facility scale decision. This paper makes contribution to literature by optimising locations and scales of
facilities simultaneously in a CLSC network. A two-stage MINLP model is established in which the first stage corresponds
to determining the optimal location of (re)manufacturers and logistics centres with the consideration of facility scales. The
second stage addresses optimisation of expected transportation, processing and holding costs over a set of scenarios that cap-
ture uncertainty in demand and return. Unlike other related studies (Easwaran and Üster 2010; Keyvanshokooh et al. 2013;
Zhang, Berenguer, and Shen 2014), which usually force each demand location (or logistics centre) to interact with single
logistics centre (or (re)manufacturer) per product, this paper considers multi-source for both logistics centres and customers
for potential improvements on operation efficiency. Also, the proposed model is an extension of the previous models by
jointly considering uncertain customers’ demand and return, multi-product, multi-echelon, and capacitated facilities. Fur-
thermore, the proposed model allows us to eliminate the requirement for co-location of (re)manufacturing facilities (since
we model co-location by a constraint, rather than a variable definition) and to model alternative layers for return inspection
locations (by modifying the formulation accordingly).

2.2. TS algorithm with CLSC models


CLSC literature based on TS algorithm mainly focuses on problems without considering uncertainty (Lee and Dong 2008;
Easwaran and Üster 2009; Eskandarpour et al. 2014; Yuchi et al. 2016). Mirmajlesi and Shafaei (2016) studied a multi-
period, multi-product, capacitated CLSC network design problem under uncertain demand. The problem was formulated as
location and allocation models. After solving the location model, the location of facilities was obtained and used as input in
the allocation model. They used TS algorithm to solve the allocation model to make decisions on product distribution.
The above literature shows that TS algorithm has been successfully applied to CLSC network design problems. As this
paper considers uncertain demand and return, facility scale selection, multi-product, and multi-source, the proposed model
International Journal of Production Research 7569

Table 1. CLSC models under uncertainty.


Multi-source Uncertain factors
Multiple Uncertainty Solution
Article products Center Customer Demand Return method Outputs Modeling method
Pishvaee, Rabbani, and     Robust L,AL,ND,QSP,TA MILP E
Torabi (2011)
Abdallah, Diabat, and  Stochastic L,AL MINLP E
Simchi-Levi (2012)
Özceylan and     Scenario L,AL,TA MILP E
Paksoy (2013)
Amin and Zhang (2013)     Scenario L,AL,QP,PA,TA MILP E
Zhang, Berenguer, and    Stochastic L,AL,TA MINLP B&C
Shen (2014)
Ramezani et al. (2014)    Fuzzy S,L,AL,TA,I,DC FP E
Diabat, Abdallah, and  Stochastic L,AL MINLP LR
Henschel (2015)
Khatami, Mahootchi,     Scenario L,FC,AL,TA,I,QSP SMILP BD
and Farahani (2015)
Zhang and   Stochastic L,AL,RI,OL,SC NIP E,OA
Unnikrishnan (2016)
Üster and Hwang (2016)     Scenario L,FC,AL,TA MILP BD
Mirmajlesi and     Robust L,AL,QP,PA,ND MILP TS,DE
Shafaei (2016)
Zhalechian et al. (2016)     Fuzzy L,FC,AL,OS,VA,WT SPP GA,VNS
Yousefi-Babadi     Robust L,AL,TA,PD MINLP LR
et al. (2017)
Kang, Wang, and   Fuzzy L,AL FP P-PSO
Ma (2017)
Jerbia et al. (2018)    Scenario L,AL,TA MILP E
Our paper      Scenario L,FS,AL,TA MINLP ITS
Notes: S: Supplier; L: Facility location; FC: Facility capacity; FS: Facility scale; AL: Allocation; DC: Discount; PA: Production assign-
ment; TA: Transportation amount; I: Inventory; US: Amount of used products which are processed; QP: Quantity of production; ND:
Quantity of non-satisfied demand; QSP: Quantity of scrapped products; RI: Review interval; OL: Order-up-to-level; OS: Order size; SC:
Storage capacity; VA: Vehicles assignment; WT: Waiting time; PD: Product delivery time; FP: Fuzzy programming; SPP: Stochastic-
possibilistic programming; E: Exact; B&C: Branch and Cut; OA: Outer approximation; BD: Benders decomposition; LR: Lagrangian
relaxation-based; DE: Differential evolution; GA: Genetic algorithm; VNS: Variable neighbourhood search; P-PSO: Priority-based
global-local-neighbour particle swarm optimisation.

is non-linear and much harder to solve. This study explores the potential of using TS algorithm to solve the non-linear
optimisation model.

3. Problem background and description


The problem studied in this paper originates from the GREE Electric Appliances Corporation, which is a global home
appliance enterprise founded in 1991. GREE achieved total revenue of 17.56 billion USD in 2016. GREE air conditioners
are famous in the world, the annual productivity of home air conditioners exceeds 60 million units and the annual throughput
of commercial air conditioners reaches 5.5 million units. The company is headquartered in Zhuhai, it has facilities for
manufacturing and remanufacturing located in Guangdong, Anhui, Henan, Wuhan, Zhejiang, Hunan and other provinces
in China. The main products made by the corporation are home air conditioners, central air conditioners, water heaters,
refrigerators, mobile phones and small appliances. In the supply chain of GREE, thousands of millions of products are
shipped from producers to customers through logistics centres each year. Meanwhile, hundreds of millions of returned
products are transported to remanufacturing facilities for remanufacturing/refurbing. In practice, many companies, such
as GREE, did not conduct in-depth analysis of customers’ demand change trend and transportation volume when building
supply chain network, causing considerable waste of money on unnecessary location and transportation costs. Moreover, the
competition becomes fiercer in recent years. So it is necessary for companies to optimise the location and scale of facilities
in the CLSC network. Figure 1 shows the geographic distribution of customers of GREE and the possible candidate location
of producers and logistics centres.
7570 L. Zhen et al.

Figure 1. Locations of producers, logistics centres and customers.

Under the above background, Figure 2 shows an example of a CLSC network structure for some companies, such as
GREE. The CLSC network refers to a complete supply chain which comprises two parts: forward and reverse logistics.
The role of forward supply chain is to distribute products from producers (for manufacturing/remanufacturing) to customers
through logistics facilities. While the role of reverse supply chain is to collect returned products from customers and transport
them to producers (for remanufacturing).
This paper considers a multi-product, multi-facility, three-layer CLSC design problem under uncertainty. Since the
network system needs to satisfy the customers’ demand and return, the most important decision is to determine the optimal
location of (re)manufacturers and logistics centres as well as their scales, so as to minimise the total cost.
Before modelling the CLSC network design problem, some assumptions are listed as follows:

• Remanufactured products are perfect substitutes for newly manufactured products.


• For each product, we use a random recovery fraction to denote the percentage of returned products which can be
recovered.
• A producer location is selected for remanufacturing only if it has been selected as a manufacturing site. This
assumption is made for reducing the response time and improving operation efficiency. Because we use a constraint
rather than a variable definition to ensure the co-location, this assumption can easily be undone.
• In our initial model M0, we assume that the product inspection to identify recoverable returns is conducted at the
remanufacturing locations. In Section 7, we show how our model can be modified to obtain alternative models for
inspection in other layers of the network.
International Journal of Production Research 7571

Figure 2. CLSC network structure.

4. Model formulation
In this section, a stochastic mathematical model is formulated for solving the CLSC network design problem. The total cost
objective includes two components: the fixed installation cost and the expected operation costs (i.e. production/processing
cost, transportation cost and holding cost). This paper considers the uncertain demand and return of customers. Since the
scenario-based programming has been widely used to deal with the uncertainty in parameters, we use a set of scenarios to
represent the uncertainty. The objective is to minimise the total cost of the system under the scenarios. The model in this
paper enables multi-source for each customer and allows each logistics centre to interact with multiple (re)manufacturers.
For the assignment variables, some papers defined the assignment variables as binary variables and restricted that a demand
location (or a logistics centre) is assigned to only one logistics centre (or one producer) in the forward and reverse flows.
However, our model considers multi-source for each customer and each logistics centre for potential operation efficiency
and capacity utilisation. Thus, we set the assignment variables θω,l,k,p , λω,k,l,p , μω,s,l,p , νω,l,s,p to be continuous and do not
enforce each facility interacts with single sourcing.
Notation of sets
S Set of candidate producer sites indexed by s.
L Set of candidate logistics centre sites indexed by l.
K Set of customers indexed by k.
P Set of products indexed by p.
A Set of capacity options of manufacturing sites indexed by a.
B Set of capacity options of remanufacturing sites indexed by b.
C Set of capacity options of logistics centre sites indexed by c.
 Set of scenarios indexed by ω.
Notation of inputs
dω,k,p Demand of product p in customer k under scenario ω.
sω,k,p Return of product p in customer k under scenario ω.
gm,n Cost of transportation one truckload of products from a location m to a location n.
F
fs,a Fixed cost of opening a producer at location s with capacity option a.
R
fs,b Fixed cost of selecting producer s as remanufacturing site with capacity option b.
fl,c Fixed cost of opening a logistics centre at location l with capacity option c.
mFs,a,p Manufacturing capacity of product p at manufacturer s with capacity option a.
mRs,b,p Remanufacturing capacity of product p at remanufacturer s with capacity option b.
mFl,c Processing capacity of forward logistics facility l with capacity option c.
mRl,c Processing capacity of reverse logistics facility l with capacity option c.
F
ks,a,p Unit manufacturing cost of product p at manufacturer s with capacity option a.
R
ks,b,p Unit remanufacturing cost of product p at remanufacturer s with capacity option b.
7572 L. Zhen et al.

nFl,c,p Unit processing cost of product p at forward logistics facility l with capacity option c.
nRl,c,p Unit processing cost of product p at reverse logistics facility l with capacity option c.
rs,p Recovery rate of product p at producer s.
hp Unite holding cost of product p per period.
ep Fraction of truck capacity (in terms of one truck) required by a unit of product p.
vp Volume of a unit of product p.
tm,n Shipment time from a location m to a location n.
probω , Probability of scenario ω.
Notation of decision variables
αs,a Binary variable, equals to one if a producer is built at location s with capacity option a, a ∈ A, otherwise,
equals to zero.
βs,b Binary variable, equals to one if a producer s is selected as remanufacturing site with capacity option b,b ∈ B,
otherwise, equals to zero.
γl,c Binary variable, equals to one if a logistics centre is built at location l with capacity option c, c ∈ C, otherwise,
equals to zero.
θω,l,k,p fraction of product p in customer k’s demands is satisfied by logistics centre l under scenario ω.
λω,k,l,p fraction of product p in customer k’s return sent to logistics centre l under scenario ω.
μω,s,l,p amount of product p transported from producer s to logistics centre l under scenario ω.
νω,l,s,p amount of product p transported from logistics centre l to producer s under scenario ω.
Mathematical model
  
[M0] min TC = F
fs,a αs,a + fsbR βs,b + fl,c γl,c
s∈S,a∈A s∈S,b∈B l∈L,c∈C
  
Total fixed facility cost
 

  
+ probω μω,s,l,p gs,l ep + F
ks,a,p αs,a + ts,l hp − F
rs,p ks,a,p αs,a νω,l,s,p
ω∈ s∈S l∈L p∈P a∈A a∈A
  
Total cost from producers to logistics centres in forward chain


 
+ probω νω,l,s,p gl,s ep + R
rs,p ks,b,p βs,b + tl,s hp
ω∈ l∈L s∈S p∈P b∈B
  
Total cost from logistics centres to producers in reverse chain


 
+ probω θω,l,k,p gl,k ep dω,k,p + nFl,c,p γl,c dω,k,p + tl,k hp dω,k,p
ω∈ l∈L k∈K p∈P c∈C
  
Total cost from logistics centres to customers in forward chain


 
+ probω λω,k,l,p gk,l ep sω,k,p + nRl,c,p γl,c sω,k,p + tk,l hp sω,k,p (1)
ω∈ k∈K l∈L p∈P c∈C
  
Total cost from customers to logistics centres in reverse chain

subject to:

αs,a ≤ 1, ∀ s ∈ S. (2)
a∈A

βs,b ≤ 1, ∀ s ∈ S. (3)
b∈B

γl,c ≤ 1, ∀ l ∈ L. (4)
c∈C
 
βs,b ≤ αsa , ∀ s ∈ S. (5)
b∈B a∈A
International Journal of Production Research 7573

θω,l,k,p = 1, ∀ k ∈ K, ∀ p ∈ P, ∀ ω ∈ . (6)
l∈L

λω,k,l,p = 1, ∀ k ∈ K, ∀ p ∈ P, ∀ ω ∈ . (7)
l∈L
 
μω,s,l,p = dω,k,p θω,l,k,p , ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (8)
s∈S k∈K
 
νω,l,s,p = sω,k,p λω,k,l,p , ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (9)
s∈S k∈K
 
vp dω,k,p θω,l,k,p ≤ mFl,c γl,c , ∀ l ∈ L, ∀ ω ∈ . (10)
k∈K,p∈P c∈C
 
vp sω,k,p λω,k,l,p ≤ mRl,c γl,c , ∀ l ∈ L, ∀ ω ∈ . (11)
k∈K,p∈P c∈C
 
μω,s,l,p ≤ mFs,a,p αs,a , ∀ s ∈ S, ∀ p ∈ P, ∀ ω ∈ . (12)
l∈L a∈A
 
νω,l,s,p ≤ mRs,b,p βs,b , ∀ s ∈ S, ∀ p ∈ P, ∀ ω ∈ . (13)
l∈L b∈B

θω,l,k,p , λω,k,l,p , μω,s,l,p , νω,l,s,p ≥ 0, ∀ s ∈ S, ∀l ∈ L, ∀ k ∈ K, ∀ p ∈ P, ∀ ω ∈ . (14)


αs,a , βs,b , γl,c ∈ {0, 1}, ∀ s ∈ S, ∀l ∈ L, ∀ a ∈ A, ∀b ∈ B, ∀ c ∈ C. (15)

In the objective function, the first term shows the total fixed location costs of the (re)manufacturers and logis-
tics centres. The second term represents the total expected transportation, manufacturing and holding costs between
manufacturers and logistics facilities in the forward direction. Note that, since remanufactured products are perfect
substitutes for new products, flow from manufacturers to logistics facilities, μω,s,l,p includes both new and remanu-
factured
products. Therefore, for a given scenario ω ∈ , the manufacturing cost at a producer s ∈ S is given by
(k F
α μ ) − (r k F
α νω,l,s,p ) and the remanufacturing cost at a producer s ∈ S is calcu-
s,a,p
l∈L,a∈A,p∈P s,a w,s,l,p l∈L,a∈A,p∈P s,p s,a,p s,a
lated as l∈L,b∈B,p∈P (rs,p ks,b,p
R
βs,b νω,l,s,p ). The third term is the total cost between remanufacturers and logistics facilities in
the reverse direction. Finally, the fourth and fifth term represent total expected transportation, processing and holding costs
associated with logistics centres and customers in forward and reverse directions, respectively.
Constraints (2)–(4) elaborate that at most one manufacturer (remanufacturer or logistics centre) can be established at
a candidate location of manufacturer(remanufacturer or logistics centre). Constraints (5) guarantee that only a candidate
producer location which has been selected as a manufacturing site can be selected for remanufacturing. Constraints (6)
and (7) ensure that each customer is assigned one logistics centre for receiving products and sending returned products.
Constraints (8) and (9) ensure the flow conservation at logistics centres in forward and reverse directions, respectively.
Constraints (10) and (11) limit the total volume of products shipped to a logistics centre do not exceed the centres’ respective
capacity. Similarly, constraints (12) and (13) are capacity restrictions for manufacturers and remanufacturers, respectively.
Finally, constraints (14) and (15) define the domain of decisions variables.
The non-linear parts (μω,s,l,p αs,a ), (νω,l,s,p αs,a ), (νω,l,s,p βs,b ), (θω,l,k,p γl,c ) and (λω,k,l,p γl,c ) in objective (1) are the multipli-
cation of a continuous variable and a binary variable. The method of linearising the non-linear parts is straightforward, and
for the sake of simplicity, a general procedure for the linearizion is given as follows.
For an instance of z = x · δ where x is a continuous variable and δ is a binary variable, we assume x has lower bound x
and upper bound x, then we can form the linearised inequalities:

min{0, x} ≤ z ≤ max{0, x} (16)

x·δ ≤z≤x·δ (17)

x − x · (1 − δ) ≤ z ≤ x − x · (1 − δ) (18)
7574 L. Zhen et al.

5. Solution method
The model in this paper is formulated for problem on facility location under uncertainty considering the facility size options.
Since the facility location and size decision problem under known demand is NP-hard (Wu et al. 2015), our model is
also NP-hard; and its computational complexity increases by a growth in the number of customers, producers, logistic
centres and products. As a result, exact optimisation tools and softwares may be no longer effective, while meta-heuristic
approaches have been quite successful. For solving facility location problems, some meta-heuristics have been widely used,
such as TS, GA, SA, and neural networks. Among these, TS is a powerful approach for solving many difficult combinatorial
optimisation problems. It was suggested primarily by Glover, McMillan, and Novick (1985) and has been successful in a
variety of problem contexts like scheduling, transportation, layout design, and graphs. The reason for adopting TS to solve
the model lies on that TS has been widely used to optimise the facility location problems (Arostegui, Kadipasaoglu, and
Khumawala 2006; Prins et al. 2007; Ross and Jayaraman 2008; Punyim et al. 2018), and the results are encouraging. Thus,
this paper designs an improved TS (ITS) approach to solve the large-scale instances.
The basic TS (BTS) approach is to search for the best solution with assistance of a memory procedure that proceeds
as follows. At each iteration, candidate neighbourhood moves are evaluated, which leads to the evolution of the current
solution to a new solution. Restrictions are imposed to tag certain moves with ‘tabu move’, and then these moves are
forbidden in selection. Conventionally, a non-tabu move with the highest evaluation is selected, although aspiration criteria
permits sufficiently attractive moves to be selected in spite of their tabu status. Intensification and diversification strategies
are applied to accentuate and broaden the search in the solution space. The algorithm terminates after a pre-defined number
of iteration. Glover (1997) provided a detailed survey and description of TS.
In the BTS algorithm, only the moves that are not forbidden (not in the tabu list) or forbidden but profitable are accepted;
while this rule may make the algorithm get stuck into the local optimum in some cases. In this paper, in order to jump out
of the local optimum and search for the best global solution, we design a release criterion, that is, we accept a forbidden but
nonprofitable move with a probability ε (0 < ε < 1).
The details of the ITS algorithm are defined in the following subsections.

5.1. Solution representation


This paper determines the optimal location of (re)manufacturers and logistics centres with considering their scale decisions.
Thus we characterise a solution by three core variables (i.e. αs,a , βs,b , γl,c ). We use a N1 + N2 + N3 dimensional sequence
to represent a solution, where N1 =| S | × | A |, N2 =| S | × | B |, N3 =| L | × | C |. The N1 dimensions indicate the set
of candidate manufacturer sites and their capacity options. The value in each of the dimensions is a random real number
between zero and two. The integer part of the value in each dimension represents the value of variable αs,a . More specifically,
we introduce s to represent the integer part of i/(| S | −1); a to represent the remainder part of i/(| S | −1), if the integer
part of the value in dimension i is equal to one, we set αs,a = 1. That is, we build a manufacturer at location s with capacity
a. Otherwise, we set αs,a = 0. Similarly, for the N2 and N3 dimensions, we use the same method to generate the value of the
other two variables (i.e. βs,b , γl,c ). Note that in order to avoid generating some infeasible solutions, solutions are iteratively
fine-tuned by some procedures for holding the validity of the related constraints in the model.

5.2. Neighbourhood searching strategy and tabu list


We use the same method to generate the neighbourhood Nα , Nβ and Nγ of variable αs,a , βs,b and γl,c , respectively. Thus,
we introduce the method which is used to generate the neighbourhood of variable αs,a as an example. For a given variable
αs,a in a feasible solution, we define K as the set of αs,a that is equal to one, H as the set of αs,a that is equal to zero. Then
we randomly choose an element from the set K and an element from the set H to form the exchange pairs (moves). That is
an exchange simply referring to the closing of an opened manufacturer and the opening of a new one. This choice is made
from the set of nodes that satisfy the capacity requirement of customers. The neighbourhoods Nβ and Nγ are generated by
the same approach.
The elements in the tabu list are set according to the neighbourhood searching strategy. Let Tα , Tβ and Tγ be the tabu list
of variable αs,a , βs,b and γl,c , respectively. Z ∗ is defined as the best global fitness value, Vα , Vβ and V γ represent the moves
in the neighbourhood Nα , Nβ and Nγ , respectively. In this paper, we classify these moves into three categories: unforbidden
(U), forbidden but profitable (FP), forbidden and nonprofitable (FN). Moves from set (Vα Tα ) ∪ (Vβ Tβ ) ∪ (Vγ Tγ ) are
U-moves. Whereas moves from set (Vα Tα ) ∩ (Vβ Tβ ) ∩ (Vγ Tγ ) are forbidden. A forbidden move is profitable if it leads
to a better solution, which is also defined as the aspiration criterion. That is, FP-moves are defined by set

A = {v ∈ Vα ∩ Tα ∩ Vβ ∩ Tβ ∩ Vγ ∩ Tγ : Z(θ,λ,μ,ν|α,β,γ ) < Z ∗ }
International Journal of Production Research 7575

The remaining moves are FN-moves.


It is quite natural that a move to be performed should be selected among U-moves and FP-moves. Note that our neigh-
bourhood is smaller than those defined by other studies, to avoid the algorithm falling into local optimisation quickly, we
design a release criterion, that is, we accept a FN-move with a probability ε (0 < ε < 1). More specifically, we randomly
generate parameter ξ between zero to one, we accept the FN-move if ξ > ε, otherwise the move will be dropped.

5.3. Diversification strategy


Diversification strategy is an algorithmic mechanism that aims to avoid the search space is too ‘local’. In this study, when
the best-known fitness has not being improved for a certain number of iterations, several feasible solutions are generated
randomly by using the same method that generates the initial solutions. Then we select the best solution as the current
solution.

5.4. Tabu search procedure


Based on the above introductions, the algorithm framework for solving our model is shown as Algorithm 1.

Algorithm 1: Algorithm framework of ITS meta-heuristic


1 Initial the current solution (α n , β n , γ n ) and evaluate its fitness value by using off-the-shelf solvers (e.g. CPLEX);
2 Set the iteration number n = 1; for the case that the best-known solution has remained unchanged, set its iterations
number m = 0; initial the tabu list Tα = ∅, Tβ = ∅, Tγ = ∅ and set each list length to be l;
3 while the end condition is not satisfied (the iteration number n does not exceed the maximum value N) do
4 Generate neighbourhoods of the current solution (α n , β n , γ n ) and calculate their fitness values;
5 Sort neighbourhood solutions in an ascending order according to their fitness value and select the solutions
according to the sequence, identify the best solution;
6 if the move of a neighbourhood solution is U-move then
7 Substitute this neighbourhood solution for the current solution (α n , β n , γ n );
8 else if the move of a neighbourhood solution is FP-move then
9 Substitute this neighbourhood solution for the current solution (α n , β n , γ n );
10 else
11 For the FN-move of a solution, apply the release criterion to the neighbourhood solution;
12 Update the tabu list Tα , Tβ and Tγ ;
13 if the new current solution (α n , β n , γ n ) is better than the best known solution (α ∗ ,β ∗ , γ ∗ ) then
14 (α ∗ , β ∗ , γ ∗ ) ← (α n , β n , γ n ), and set m=0;
15 else
16 Set m=m+1;
17 if m>M then
18 Regenerate the current solution (α n , β n , γ n ) according to diversification strategy and set m=0;
19 Set n=n+1;
20 return the best efficient solution(α ∗ , β ∗ , γ ∗ ).

6. Numerical experiments
In this section, in order to validate the effectiveness of the proposed model and the efficiency of the algorithm, we have
carried out several numerical experiments on a LENOVO P910 workstation with 28 cores of CPUs, 2.4 GHZ processing
speed and 256 GB of memory. The mathematical models, embedded in the solution algorithms, were implemented by
CPLEX 12.5.1 (Visual Studio 2015, C #).
7576 L. Zhen et al.

Table 2. Experiments under different numbers of scenarios.


Number of scenarios Max. Min. Range Avg. S.D. Agv.CPU time(s)
10 2.97×1011 2.72×1011 2.51×1010 2.84×1011 7.03×109 37
20 2.91×1011 2.81×1011 9.34×109 2.85×1011 2.87×109 98
50 2.85×1011 2.80×1011 5.44×109 2.83×1011 1.71×109 376
80 2.86×1011 2.82×1011 4.35×109 2.84×1011 1.47×109 823
100 2.85×1011 2.81×1011 4.75×109 2.83×1011 1.22×109 1399
Notes: (1) Range = Max − Min. (2) S.D. denotes the standard deviation of each group.

6.1. Experimental settings


In order to illustrate the wide applicability of the proposed model and method, a number of problem instances are needed to
validate the effectiveness of the model and the efficiency of the solution method. In addition, for finding more managerial
implications based on our model, a lot of instances with different settings are also needed for the numerical experiments.
Thus we generate sets of parameters according to some probability distributions that follow the real-world patterns, which
is explained as follows.
Since the customers’ demand and return change along with different scenarios, we set the random demand dw,k,p to
follow uniform distribution (i.e. dw,k,p ∼ U(800, 5000)), then we multiply the value of demand by a randomly gener-
ated fraction δk,p ∼ U(0.5, 0.8), and the returns are calculated as sw,k,p = dw,k,p × δk,p . We define basemFs,a,p as the basic
manufacturing capacity of product p at manufacturer s with capacity option a, it is calculated as basemFs,a,p = η1 × TDp ,
where η1 ∼ U(0.3, 0.5), TDp represents the maximum total demands of product p among || scenarios (i.e. TDp =
maxω∈ { k∈K dw,k,p }), their specific values are proportional to their scale options a. Similarly, we obtained the basic value
basemRs,b,p = η1 × TRp , TRp represents the maximum total returns of product p among || scenarios. For the capacity of a
l,c = η2 × TD, where η2 ∼ U(0.2, 0.4),
F
forward logistics centre l with capacity option c, its basic value is calculated as basem
TD represents the maximum volume of all products among || scenarios, TD = p∈P ep × TDp . Meanwhile, the basic value

mRl,c = η2 × TR, TR = p∈P ep × TRp . Their specific values are proportional to their scale options c. This paper considers
the effect of the economic of scale on unit (re)manufacturing and processing cost, the value of unit (re)manufacturing and
F
forward/reverse processing cost decreases with the expansion of scales. Firstly, we define baseks,a,p as the basic unit man-
ufacturing cost, which follows the uniform distribution baseks,a,p ∼ U(50, 3000), then we calculate the unit manufacturing
F

cost of each product p at manufacturer s with capacity option a as ks,a,p F


= (1 − 0.05 × a) × baseks,a,p
F
, the remanufacturing
R
cost is calculated as ks,b,p = 0.6 × ks,a,p
F
. Similarly, we generate the basic unit forward processing cost basenFl,c,p ∼ U(10, 30)
and we calculate nFl,c,p = (1 − 0.1 × c) × basenFl,c,p , the unit reverse processing cost are generated by the similar method.
For the transportation cost gm,n , it is set as gm,n = uc × lm,n , where uc is the unit cost which equals to 1.5 RMB/km, lm,n
denotes the distance between location m and location n which follows the uniform distribution lm,n ∼ U(0, 3000). Similarly,
we generate the transportation time parameter tm,n . For the fixed construction cost, the basic value of fixed cost fs,a of a
manufacturer at location s with capacity a is set as 6 × 108 , the basic value fs,b is set as 2 × 108 , and the basic value fl,c is
set as 4 × 107 , their specific values are proportional to their scale options (i.e. a, b and c), respectively. The unit inventory
holding cost hp of product p follows the uniform distribution hp ∼ (900, 2000). The recovery rate of product p at producer
s follows the uniform distribution rs,p ∼ (0.5, 0.8).

6.2. Testing of different numbers of scenarios


The number of scenarios has a significant influence on the proposed stochastic programming model, so it is important
to determine the suitable number of scenarios. Five experiments with different scenarios under the scale 6-8-10-6 (i.e. 6
candidate producers, 8 candidate logistics centres, 10 customers and 6 products) are conducted, namely 10, 20, 50, 80 and
100. Ten different set of cases are generated randomly and solved in each group. The maximum, minimum, range, average
objective value and standard deviation of the five experiments are recorded, as shown in each row of Table 2.
Table 2 demonstrates that the S.D. decreases evidently when the number of scenarios increases. It can be observed
that when the number of scenarios exceeds 50, the range between the maximum and minimum values does not decrease
evidently and the standard deviation decreases slowly, but the right most column of Table 2 shows that the CPU time
increases evidently with the number of scenarios increasing. It suggests that the setting of 50 scenarios is appropriate for the
model. Hence, in the following experiments, the number of scenarios is set as 50.
International Journal of Production Research 7577

Table 3. Comparison between the TS methods and the optimal results solved by CPLEX solver.
OBJ Time (s) GAPOBJ
Case ID OBJ CPLEX Time (s) CPLEX ITS BTS ITS BTS ITS BTS
5-7-7-5-1 1.57×1011 57 1.60×1011 1.61×1011 169 142 1.78% 2.61%
5-7-7-5-2 1.71×1011 105 1.73×1011 1.74×1011 159 141 0.85% 1.44%
5-7-7-5-3 1.13×1011 111 1.15×1011 1.16×1011 149 137 1.81% 3.41%
6-8-8-6-1 1.82×1011 249 1.84×1011 1.84×1011 225 208 1.21% 1.27%
6-8-8-6-2 1.73×1011 299 1.76×1011 1.78×1011 267 242 1.61% 2.76%
6-8-8-6-3 2.81×1011 561 2.84×1011 2.85×1011 256 239 0.97% 1.33%
6-8-15-6-1 4.95×1011 976 5.02×1011 5.02×1011 663 670 1.49% 1.51%
6-8-15-6-2 4.19×1011 1053 4.21×1011 4.23×1011 380 539 0.45% 0.93%
6-8-15-6-2 4.01×1011 1113 4.08×1011 4.08×1011 419 577 1.72% 1.71%
6-10-10-8-1 3.89×1011 1008 3.91×1011 3.93×1011 458 445 0.43% 0.95%
6-10-10-8-2 2.70×1011 1066 2.73×1011 2.74×1011 519 558 0.93% 1.41%
6-10-10-8-3 2.74×1011 1804 2.76×1011 2.77×1011 360 342 0.94% 1.30%
8-10-10-8-1 2.68×1011 1116 2.72×1011 2.73×1011 484 561 1.74% 2.04%
8-10-10-8-2 3.21×1011 1765 3.26×1011 3.26×1011 466 672 1.70% 1.73%
8-10-10-8-3 2.59×1011 1852 2.61×1011 2.63×1011 522 540 0.87% 1.64%
8-10-12-8-1 cannot solve N.A. 4.38×1011 4.38×1011 706 793 N.A. N.A.
8-10-12-8-2 by CPLEX N.A. 5.13×1011 5.15×1011 642 702 N.A. N.A.
8-10-12-8-3 in two hours N.A. 5.10×1011 5.12×1011 539 538 N.A. N.A.
Avg. 2.78×1011 876 2.81×1011 2.83×1011 366 383 1.23% 1.73%
Notes: (1) The numbers in each case ID (e.g. ‘5-7-7-5-1’) denote the number of producers (5 producers), logistics centres (7 centres),
customers (7 customers) and products (5 products), respectively, the fifth value denotes the index of the cases with the same scales. (2)
‘OBJ’ denotes the objective value, ‘Time’ denotes the computation time, and the unit is the second. (3) GAPITS = (OBJITS − OBJCPLEX )
/ OBJCPLEX ; GAPBTS = (OBJBTS − OBJCPLEX ) /OBJCPLEX .

6.3. Experiments of proposed solution method


To verify the optimisation quality and efficiency of the proposed algorithm, we conduct a series of experiments in different
scales and make a comparison between BTS and ITS algorithm about CPU time and accuracy. Table 3 shows that when
the network exceeds a certain scale, the ITS can obtain the results quickly than the CPLEX solver, for some large-scale
instances, the CPLEX solver cannot obtain the results within a reasonable CPU time, but the algorithm can obtain a good
solution in a short CPU time and the average optimality gap of the results is 1.23%, which indicates the results obtained by
our proposed method are quite near to the optimal solutions. Moreover, it can also be observed that compared to the BTS,
the ITS has a better performance with respect to the average gap from the optimality, while their computation times are
similar.
In the literature, we observe that some papers set the customer demand to follow uniform distribution (Pishvaee, Fara-
hani, and Dullaert 2010; Torres-Soto and Üster 2011; Özceylan and Paksoy 2014; Keyvanshokooh, Ryan, and Kabir 2016),
while some papers generated the customer demand by following normal distribution (Javid and Azad 2010; Devika, Jafar-
ian, and Nourbakhsh 2014; Nekooghadirli et al. 2014). Therefore, to investigate the performance of our algorithm in more
comprehensive contexts with respect to the distribution of customer demands, we also conduct sets of comparative experi-
ments where the quantity of customer demands is generated by following normal distribution (i.e. dw,k,p ∼ N(3000, 500)).
The result in Table 4 is consistent with that in Table 3. The average optimality gap of the results of ITS and BTS is 0.89%
and 1.31%, respectively, which validates the good performance of the TS method in the context of normal distribution and
also illustrates the ITS outperforms the BTS.

6.4. Comparison between the proposed model and the distance-based decision rule
In reality, when designing the CLSC network, decision makers tend to adopt some intuitive decision rules. For example, they
prefer to choose the nearest (re)manufacturer for (re)manufacturing and the nearest centre for distribution or collection. In
this paper, to validate the effectiveness of our proposed model, we develop an intuitive decision rule, namely, the distance-
based decision rule. The purpose of the distance-based decision rule is to find out the shortest distance when selecting the
(re)manufacturers and logistics centres for each customer. We implement the distance-based decision rule as follows: Since
we need to satisfy the customers’ demand and return, we need to determine the location and scale of logistics centres at
first. We choose the nearest candidate logistics centre for each customer, then for the forward flow, we select the nearest
candidate producer site as the manufacturer for each centre which has been chosen. For the reverse flow, we select the nearest
7578 L. Zhen et al.

Table 4. Comparison of the two TS methods under normal-distribution customer demand.


OBJ Time (s) GAPOBJ
Case ID OBJ CPLEX Time (s) CPLEX ITS BTS ITS BTS ITS BTS
5-7-7-5-1 6.28×1010 30 6.34×1010 6.41×1010 145 137 0.97% 1.99%
5-7-7-5-2 1.65×1011 59 1.66×1011 1.67×1011 160 145 1.00% 1.40%
5-7-7-5-3 1.29×1011 103 1.30×1011 1.30×1011 158 127 0.90% 1.01%
6-8-8-6-1 2.25×1011 212 2.26×1011 2.28×1011 204 168 0.66% 1.20%
6-8-8-6-2 1.54×1011 264 1.54×1011 1.55×1011 196 186 0.49% 0.60%
6-8-8-6-3 2.69×1011 545 2.70×1011 2.71×1011 198 192 0.42% 0.92%
6-8-15-6-1 3.01×1011 979 3.05×1011 3.06×1011 644 645 1.17% 1.44%
6-8-15-6-2 4.81×1011 1215 4.84×1011 4.84×1011 693 687 0.55% 0.58%
6-8-15-6-3 5.15×1011 958 5.17×1011 5.17×1011 586 602 0.47% 0.48%
6-10-10-8-1 2.88×1011 1167 2.90×1011 2.92×1011 560 539 0.78% 1.17%
6-10-10-8-2 3.90×1011 822 2.94×1011 3.94×1011 459 502 0.88% 1.00%
6-10-10-8-3 4.01×1011 1070 4.03×1011 4.04×1011 578 551 0.44% 0.76%
8-10-10-8-1 2.77×1011 1176 2.82×1011 2.84×1011 481 587 1.68% 2.62%
8-10-10-8-2 2.42×1011 1228 2.46×1011 2.47×1011 493 662 1.80% 2.36%
8-10-10-8-3 3.06×1011 1826 3.11×1011 3.11×1011 503 673 1.68% 1.75%
8-10-12-8-1 4.23×1011 2791 4.26×1011 4.26×1011 682 739 0.68% 0.78%
8-10-12-8-2 4.00×1011 3486 4.04×1011 4.10×1011 770 890 0.86% 2.32%
8-10-12-8-3 3.66×1011 2155 3.69×1011 3.70×1011 559 621 0.68% 1.15%
Avg. 3.00×1011 1116 3.02×1011 3.03×1011 448 481 0.89% 1.31%

Table 5. Comparison between the results obtained by the proposed method and the optimal results.
Case ID Proposed model Distance-based decision rule Difference Rate of difference
5-7-7-5-1 1.57×1011 2.88×1011 1.17×1011 68.07%
5-7-7-5-2 1.71×1011 2.50×1011 7.81×1010 45.57%
5-7-7-5-3 1.13×1011 1.46×1011 3.31×1010 29.37%
6-8-8-6-1 1.82×1011 2.28×1011 4.57×1010 25.14%
6-8-8-6-2 1.73×1011 1.97×1011 2.35×1010 13.57%
6-8-8-6-3 2.81×1011 3.71×1011 8.98×1010 31.97%
6-8-15-6-1 4.95×1011 5.52×1011 5.71×1010 11.55%
6-8-15-6-2 4.19×1011 6.50×1011 2.31×1011 55.07%
6-8-15-6-3 4.01×1011 5.19×1011 1.18×1011 29.31%
6-10-10-8-1 3.89×1011 5.30×1011 1.40×1011 36.09%
6-10-10-8-2 2.70×1011 4.26×1011 1.55×1011 57.46%
6-10-10-8-3 2.74×1011 4.32×1011 1.58×1011 57.79%
8-10-10-8-1 2.68×1011 3.64×1011 9.63×1010 36.01%
8-10-10-8-2 3.21×1011 5.18×1011 1.97×1011 61.35%
8-10-10-8-3 2.59×1011 3.92×1011 1.33×1011 51.22%
Avg. 2.78×1011 3.91×1011 1.12×1011 40.64%
Notes: Difference = the objective value of the distance-based decision rule − the objective value of
the proposed model. Rate of difference = (the objective value of the distance-based decision rule −
the objective value of the proposed model) / the objective value of the proposed model.

remanufacturing site from the set of sites which have been selected for manufacturing. If customers’ demand and return
are both satisfied, the location of the (re)manufacturers and logistics centres as well as the product flow are determined;
otherwise, the remaining customers’ demand or return will be satisfied and transported by the second minimum distance
road. Finally, we calculate the total costs (i.e. the fixed installation cost, production/processing cost, transportation cost and
holding cost). We make a comparison between the total cost obtained by the proposed model and the total cost that obtained
from the above decision rule, which is shown in Table 5.
From Table 5, we can observe that the average rate of difference is about 40.64%, which validates the effectiveness of
developing our model.
International Journal of Production Research 7579

6.5. Comparison between the proposed model and the model with ‘single-source assignment for each customer and
each logistic centre’ constraints
Among the related studies, most papers usually force each demand location to be assigned to only one logistics centre or
force each centre to interact with only one producer per product (Easwaran and Üster 2010; Keyvanshokooh et al. 2013;
Zhang and Unnikrishnan 2016). For example, Easwaran and Üster (2010) considered a multi-product CLSC network design
problem with hybrid (re)manufacturing facilities and finite-capacity hybrid logistic centres to serve a set of retail loca-
tions. They proposed an MILP model to determine the optimal solution that characterises facility locations, along with the
integrated forward and reverse flows such that the total cost of facility construction, processing, and transportation in the
network is minimised. In their model, they considered a single sourcing strategy. More specifically, they set two binary
decision variables, one represents if a retailer is served by a distribution centre; the other one represents if a retailer is served
by a collection centre. Then they used two constraints to ensure that a retailer is assigned to exactly one logistics centre for
the forward and reverse flow of products, respectively.
Our study considers multi-source for each customer and each logistics centre for potential improvements in operation
efficiency. For comparison, we modify our proposed model according to Easwaran and Üster (2010) to force each customer
and each logistic centre to interact with single sourcing per product. Instead of setting the assignment variables θω,l,k,p and
λω,k,l,p to be continuous, we redefine them as binary variables so that constraints (6) and constraints (7) ensure each customer
is satisfied by one logistics centre in forward and reverse directions, respectively. Then we add two decision variables τω,s,l,p
and ϕω,l,s,p . τω,s,l,p is a binary variable, it equals to one if product p in logistics centre l is supplied by manufacturer s under
scenario ω in the forward flow; otherwise, it equals to zero. Also, ϕω,l,s,p is a binary variable, it equals to one if product p in
logistics centre l is transported to remanufacturer s under scenario ω in the reverse flow; otherwise, it equals to zero. Other
parameters and decision variables are the same with that of the model M0, then the modified model with ‘single-source
assignment for each customer and each logistic centre’ constraints is defined as follows:
objective (1)
subject to: constraints (2)–(15),

τω,s,l,p = 1, ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (19)
s∈S

ϕω,l,s,p = 1, ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (20)
s∈S

μω,s,l,p ≤ M τω,s,l,p , ∀ s ∈ S, ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (21)


νω,l,s,p ≤ M ϕω,l,s,p , ∀ s ∈ S, ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (22)

where M is a sufficiently large positive number. Constraints (19) and (20) ensure that each logistics centre interacts with
only one (re)manufacturer per product in the forward and reverse flows, respectively. Constraints (21) and (22) force that
only when the logistics centre interacts with the (re)manufacturer, can products be transported between the arc in the forward
and reverse flows, respectively.
To verify the superiority of our proposed model, we conduct some large-scale instances and use ITS algorithm to solve
them. Table 6 shows the total cost of the model with the ‘single-source assignment for each customer and each logistics
centre’ constraints and the total cost of the proposed model in Section 4.
Table 6 demonstrates that compared to the model with one-source constraints, our proposed model can reduce the total
cost for 4.10% on average. Moreover, with the instance scale growing, the cost savings of our proposed model increase,
which validates the significant benefit of allowing each customer to interact with multiple logistics centres and considering
multiple sourcing for each logistics centre.

7. Analysis on recovery location and rate


In the model M0, we assume that the returned products are inspected at producer locations, after being inspected, some
returns are disposed of and the implementation of remanufacturing for product p is operated on the rs,p fraction (i.e. recovery
fraction) of all the returns to producer s. However, in reality, the inspection of products can also be performed at logistics
centres and customer locations before being transported to producer locations.
In this section, we consider three possible stages (i.e. producer, logistic centre and customer) for inspections and analyse
their effects on the total system cost and the structure of the network. We assume that the inspection costs are different at
three stages, more specifically, the unit inspection cost is the lowest at the producer stage, and the highest at the customer
7580 L. Zhen et al.

Table 6. Comparison between the proposed model and the model with one-source constraints.
Case ID Proposed model Model with one-source constraints Difference Rate of difference
10-15-20-8-1 4.85×1011 4.91×1011 6.09×109 1.26%
10-15-20-8-2 5.42×1011 5.48×1011 6.38×109 1.18%
10-15-20-8-3 4.95×1011 5.05×1011 1.00×1010 2.02%
15-20-25-10-1 5.95×1011 6.22×1011 2.64×1010 4.43%
15-20-25-10-2 5.97×1011 6.18×1011 2.10×1010 3.52%
15-20-25-10-3 4.03×1011 4.21×1011 1.83×1010 4.54%
20-25-30-12-1 6.35×1011 6.80×1011 4.48×1010 7.06%
20-25-30-12-2 6.29×1011 6.70×1011 4.10×1010 6.51%
20-25-30-12-3 6.24×1011 6.64×1011 4.00×1010 6.41%
Avg. 5.56×1011 5.80×1011 2.38×1010 4.10%
Notes: Difference = the objective value of the model with one-source constraints − the objective value of
the proposed model. Rate of difference = (the objective value of the model with one-source constraints − the
objective value of the proposed model) / the objective value of the proposed model.

stage. We first revise the model M0 to investigate the above issues on determining the suitable stage for conducting the
inspection of returns.
Option 1: Inspection at producer. This is the case wherein the inspection is performed at producer locations, this has
been solved in the basic model M0. After being inspected, the fraction rs,p of returns are remanufactured and the rest are
disposed of. We introduce parameter ζs,p to represent the unit inspection cost of each product p at producer s. Then a modified
model (denoted by M1) considering the inspection cost is defined as follows:

[M1] min TC + probω ζs,p βs,b νω,l,s,p (23)
ω∈ s∈S l∈L p∈P b∈B

subject to: constraints (2)–(15).


In the objective function of model M 1, the first term TC is the objective function of model M 0, which represents the
fixed installation costs and the expected variable costs, namely, production/processing cost, transportation cost and holding
cost. The second term shows the expected inspection cost.
Option 2: Inspection at logistics centre. This is the case wherein all returns are inspected at logistics centres and only
the products that can be recovered are transported to producer locations for remanufacturing. The recovery rate of each
product p at centre l is denoted as rl,p , we set rs,p in M0 to one in the modified model. Similarly, we set ζl,p to represent the
unit inspection cost of each product p at logistics centre l. Then a model (denoted by M2) considering the inspection cost is
defined as follows:

[M2] min TC + probω ζl,p γl,c sω,k,p λω,k,l,p (24)
ω∈ l∈L k∈K p∈P c∈C

subject to: constraints (2)–(8), (10)–(15),


 
νω,l,s,p = rl,p sω,k,p λω,k,l,p , ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (25)
s∈S k∈K

In the objective function of model M 2, the first term TC is the objective function of model M 0, which represents the fixed
installation costs and the expected production/processing, transportation and holding costs. The second term shows the
expected inspection cost. Constraints (25) ensure the flow conservation at logistics centres in the reverse direction.
Option 3: Inspection at customer. This is the case wherein all returns are inspected at customer stage. Therefore, a
fraction of returns rk,p are transported to producer sites for remanufacturing through logistics centres. Similar to the above
case, we set the original recovery rate rs,p to one. The unit inspection cost of each product p at customer k is denoted as ζk,p .
Then a modified model (denoted by M3) considering the inspection cost is defined as follows:

[M3] min TC + probω ζk,p sω,k,p (26)
ω∈ k∈K p∈P

subject to: constraints (2)–(8), (10), (12)–(15),


 
νω,l,s,p = rk,p sω,k,p λω,k,l,p , ∀ l ∈ L, ∀ p ∈ P, ∀ ω ∈ . (27)
s∈S k∈K
International Journal of Production Research 7581
 
rk,p vp sω,k,p λω,k,l,p ≤ mRl,c γl,c , ∀ l ∈ L, ∀ ω ∈ . (28)
k∈K,p∈P c∈C

In the objective function of model M 3, the first term TC is the objective function of model M 0, which represents the
fixed installation costs and the expected production/processing, transportation and holding costs. The second term shows
the expected inspection cost. Constraints (27) ensure the flow conservation at logistics centres in the reverse direction.
Constraints (28) limit the total volume of products shipped from the customers to a logistics centre do not exceed the
centres’ capacity.
In addition to investigating three possible inspection stages, we consider two levels recovery rates for each product:
High (H) and Low (L). The value of ‘H’ recovery rate is generated within a range [0.7, 0.8] and the value of ‘L’ recovery
rate is generated within a range [0.2, 0.3]. Therefore, we obtain six different problem settings and represent them using the
notation ‘Recovery rate (H or L) − Inspection stage(producer, logistics centre and customer)’.
Based on the location of producers, logistics centres and customers in Figure 2, we further conduct a large-scale exper-
iment to compare the results. We select 10 cities as candidate producer sites, 12 cities as candidate logistics centre sites
and 15 large cities as customer locations. The geographic distributions of the candidate locations are shown in Figure 3.
Other parameters are set as same as the instance defined in Subsection 6.1. To obtain the results of the large-scale instance
(10-12-15-8), we use our ITS algorithm to solve each setting of the instance.

7.1. Comparison of total cost


We set the inspection cost of each product p at producer, logistics centre and customer locations to follow the uniform
distribution (i.e. ζsp ∼ U(35, 45), ζlp ∼ U(75, 85) and ζkp ∼ U(115, 125)), respectively. The results of the two experiments
with different scales are shown in Figure 4.
From Figure 4, it can be observed that when the recovery rate is high, inspection at logistics centres and customers save
less cost than the cases with low recovery rate. This can be explained that when the recovery rate is high, a large number
of returns are shipped to producer locations for inspection, but when the recovery rate is low, a large number of returns
which cannot be recovered are deposed of before being transported to producers, which eliminates unnecessary logistics
costs. Moreover, it indicates that the transportation cost, holding cost and fixed facility cost have a significant impact on the
total cost of the system. For the small-scale instance (6-8-10-6), it can be observed that inspection at logistics centres or
customers is more cost-effective than inspection at producers. We obtain similar value of total cost for inspection at logistics
centres and customers. For example, when the recovery rate is high, the total cost decreases by both 5.96% for inspection at
logistics centres and customers comparing to inspection at producers. When the recovery rate is low, the cost savings become
larger, which is about 20.65% when inspection takes place at logistics centres and 20.76% when inspection is performed at
customer locations compared to inspection at producers, respectively. On the other hand, for the large-scale instance, when
the recovery rate is high, the outperformance rates of inspection at logistics centres and inspection at customer locations
are 4.35% and 3.88% comparing to inspection at producers, respectively. But under low recovery rate, the cost savings by
performing inspection at customers are considerable greater than inspection at other two locations, which is about 12.27%
compared to inspection at producers, but less savings, 1.19% compared to inspection at logistics centres. This is due to that
with the expansion of the size of the instance, a relatively large number of returned products will be disposed of at customers
stage rather than being transported further back in the chain, which saves considerable transportation and holding costs.

Figure 3. Locations of customers, producers and logistics centres for large-scale instance.
7582 L. Zhen et al.

Figure 4. Total CLSC cost for six case settings for different size instances.

Figure 5. Active locations of producers and logistics centres as well as their scales for six case settings.
International Journal of Production Research 7583

In summary, it is suitable to perform inspections at logistics centre or customer locations. It should be mentioned that
when the recovery rate is high, inspection at logistics centres is more preferable. When the recovery rate is low, it is more
beneficial to operate inspections at customer locations.

7.2. Comparison of locations and scales


We also compare the producers and logistics centres’ locations as well as their scale options for all six settings under the
large-size instance. Figure 5 shows the distribution of active producers and logistics centres as well as their scale decisions in
the solutions. It can be observed that the location of active manufacturers and logistics centres is similar in all six settings,
but their capacity options are different for different recovery rates and inspection stages. Another finding is that under
high recovery rate, almost all of the producers serve for both manufacturing and remanufacturing, this is not the case with
low recovery rate, especially when the inspection is operated at logistics centre and customer locations; only six of nine
manufacturing sites are selected for remanufacturing. Inspection at logistics centre and customer locations require fewer
hybrid producer locations, which leads to a significant saving in fixed facility costs. What is more, it indicates that when the
inspection takes place at producer locations under low recovery rate, the producers which are selected for remanufacturing
are used mainly for inspections.

8. Conclusions
This paper studies an optimisation problem on deciding the suitable locations and scales of facilities in an integrated CLSC
network under uncertain environment. For solving the problem, a stochastic MINLP model is formulated and an ITS heuris-
tic algorithm is developed. Numerical experiments on various comparisons and instances validate the effectiveness of the
proposed model and the efficiency of the proposed solution method. Specifically, to validate the effectiveness of the pro-
posed model, an intuitive decision rule based on the distance is proposed and an evident outperformance (40.64% average
improvement rate) of the proposed model is observed in the numerical experiments. Also, to verify the superiority of the
proposed model, we compare the proposed model and the model with one-source constraints. Computational results show
that our model can reduce the total cost for 4.10% on average. In addition, to validate the efficiency of the proposed method,
numerical experiments are conducted to compare CPLEX, ITS and BTS. The computational results illustrate the superior
performance of the ITS algorithm (1.23% average optimality gap under the uniform-distribution customer demand and
0.89% average optimality gap under the normal-distribution customer demand). Finally, our sensitivity analysis indicates
that early product inspection has a positive effect by saving unnecessary resources and costs, and parameters such as recov-
ery rate, transportation cost and holding cost can be instrumental in deciding where the return product inspection should
take place, the overall cost and the structure of the CLSC network.
However, there are still some limitations for the current model. For example, this paper considers the uncertainty of cus-
tomers’ demand and return in the CLSC network, but the dynamic nature of the problem which is mainly due to fluctuations
in demand and return is not taken into account, which will further complicate the original model. Moreover, the current
solution approach is still a bit time consuming. All of these issues will form the directions for our future studies.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This work was supported by the National Natural Science Foundation of China [grant numbers 71831008, 71671107, 71422007].

ORCID
Lu Zhen http://orcid.org/0000-0001-5209-1109
Kai Wang http://orcid.org/0000-0002-3040-0422

References
Abdallah, T., A. Diabat, and D. Simchi-Levi. 2012. “Sustainable Supply Chain Design: A Closed-loop Formulation and Sensitivity
Analysis.” Production Planning & Control 23: 120–133.
Amin, S. H., and G. Zhang. 2013. “A Multi-objective Facility Location Model for Closed-loop Supply Chain Network Under Uncertain
Demand and Return.” Applied Mathematical Modelling 37 (6): 4165–4176.
7584 L. Zhen et al.

Arostegui Jr, M. A., S. N. Kadipasaoglu, and B. M. Khumawala. 2006. “An Empirical Comparison of Tabu Search, Simulated Annealing,
and Genetic Algorithms for Facilities Location Problems.” International Journal of Production Economics 103 (2): 742–754.
Badri, H., S. F. Ghomi, and T.-H. Hejazi. 2017. “A Two-stage Stochastic Programming Approach for Value-based Closed-loop Supply
Chain Network Design.” Transportation Research Part E: Logistics and Transportation Review 105: 1–17.
Chen, Y., F. Chan, and S. Chung. 2015. “An Integrated Closed-loop Supply Chain Model with Location Allocation Problem and Product
Recycling Decisions.” International Journal of Production Research 53 (10): 3120–3140.
Devika, K., A. Jafarian, and V. Nourbakhsh. 2014. “Designing a Sustainable Closed-loop Supply Chain Network Based on Triple Bottom
Line Approach: A Comparison of Metaheuristics Hybridization Techniques.” European Journal of Operational Research 235 (3):
594–615.
Diabat, A., T. Abdallah, and A. Henschel. 2015. “A Closed-loop Location-inventory Problem with Spare Parts Consideration.” Computers
& Operations Research 54: 245–256.
Easwaran, G., and H. Üster. 2009. “Tabu Search and Benders Decomposition Approaches for a Capacitated Closed-loop Supply Chain
Network Design Problem.” Transportation Science 43 (3): 301–320.
Easwaran, G., and H. Üster. 2010. “A Closed-loop Supply Chain Network Design Problem with Integrated Forward and Reverse Channel
Decisions.” IIE Transactions 42 (11): 779–792.
El-Sayed, M., N. Afia, and A. El-Kharbotly. 2010. “A Stochastic Model for Forward–reverse Logistics Network Design Under Risk.”
Computers & Industrial Engineering 58 (3): 423–431.
Eskandarpour, M., E. Masehian, R. Soltani, and A. Khosrojerdi. 2014. “A Reverse Logistics Network for Recovery Systems and a Robust
Metaheuristic Solution Approach.” The International Journal of Advanced Manufacturing Technology 74 (9–12): 1393–1406.
Ferguson, M. E., and G. C. Souza. 2016. Closed-loop Supply Chains: New Developments to Improve the Sustainability of Business
Practices. Roca Raton, London and New York: CRC Press.
Flapper, S. D., J. van Nunen, and L. N. Van Wassenhove. 2006. Managing Closed-loop Supply Chains. Berlin: Springer Science &
Business Media.
Glover, F. 1997. “Tabu search and adaptive memory programming—advances, applications and challenges.” In Interfaces in computer
science and operations research (pp. 1-75). Springer, Boston, MA.
Glover, F., C. McMillan, and B. Novick. 1985. “Interactive Decision Software and Computer Graphics for Architectural and Space
Planning.” Annals of Operations Research 5 (3): 557–573.
Guide, V. D. R., and L. N. Wassenhove. 2001. “Managing Product Returns for Remanufacturing.” Production and Operations
Management 10 (2): 142–155.
Huang, X.-Y., N.-N. Yan, and R.-Z. Qiu. 2009. “Dynamic Models of Closed-loop Supply Chain and Robust H∞ Control Strategies.”
International Journal of Production Research 47 (9): 2279–2300.
Javid, A. A., and N. Azad. 2010. “Incorporating Location, Routing and Inventory Decisions in Supply Chain Network Design.”
Transportation Research Part E: Logistics and Transportation Review 46 (5): 582–597.
Jerbia, R., M. K. Boujelben, M. A. Sehli, and Z. Jemai. 2018. “A Stochastic Closed-loop Supply Chain Network Design Problem with
Multiple Recovery Options.” Computers & Industrial Engineering 118: 23–32.
Jindal, A., and K. S. Sangwan. 2014. “Closed Loop Supply Chain Network Design and Optimisation Using Fuzzy Mixed Integer Linear
Programming Model.” International Journal of Production Research 52 (14): 4156–4173.
Kang, K., X. Wang, and Y. Ma. 2017. “A Collection-distribution Center Location and Allocation Optimization Model in Closed-loop
Supply Chain for Chinese Beer Industry.” Mathematical Problems in Engineering, 2017, (pp. 1-15)
Keyvanshokooh, E., M. Fattahi, S. Seyed-Hosseini, and R. Tavakkoli-Moghaddam. 2013. “A Dynamic Pricing Approach for Returned
Products in Integrated Forward/reverse Logistics Network Design.” Applied Mathematical Modelling 37 (24): 10182–10202.
Keyvanshokooh, E., S. M. Ryan, and E. Kabir. 2016. “Hybrid Robust and Stochastic Optimization for Closed-loop Supply Chain Network
Design Using Accelerated Benders Decomposition.” European Journal of Operational Research 249 (1): 76–92.
Khatami, M., M. Mahootchi, and R. Z. Farahani. 2015. “Benders’ Decomposition for Concurrent Redesign of Forward and Closed-loop
Supply Chain Network with Demand and Return Uncertainties.” Transportation Research Part E: Logistics and Transportation
Review 79: 1–21.
Laporte, G., F. V. Louveaux, and L. van Hamme. 1994. “Exact Solution to a Location Problem with Stochastic Demands.” Transportation
Science 28 (2): 95–103.
Lebreton, B. 2006. Strategic Closed-loop Supply Chain Management. Berlin: Springer Science & Business Media.
Lee, D.-H., and M. Dong. 2008. “A Heuristic Approach to Logistics Network Design for End-of-lease Computer Products Recovery.”
Transportation Research Part E: Logistics and Transportation Review 44 (3): 455–474.
Listeş, O., and R. Dekker. 2005. “A Stochastic Approach to a Case Study for Product Recovery Network Design.” European Journal of
Operational Research 160 (1): 268–287.
Mirmajlesi, S. R., and R. Shafaei. 2016. “An Integrated Approach to Solve a Robust Forward/reverse Supply Chain for Short Lifetime
Products.” Computers & Industrial Engineering 97: 222–239.
Nekooghadirli, N., R. Tavakkoli-Moghaddam, V. R. Ghezavati, and S. Javanmard. 2014. “Solving a New Bi-objective Location-routing-
inventory Problem in a Distribution Network by Meta-heuristics.” Computers & Industrial Engineering 76: 204–221.
Özceylan, E., and T. Paksoy. 2013. “A Mixed Integer Programming Model for a Closed-loop Supply-chain Network.” International
Journal of Production Research 51 (3): 718–734.
International Journal of Production Research 7585

Özceylan, E., and T. Paksoy. 2014. “Interactive Fuzzy Programming Approaches to the Strategic and Tactical Planning of a Closed-loop
Supply Chain Under Uncertainty.” International Journal of Production Research 52 (8): 2363–2387.
Patne, K., N. Shukla, S. Kiridena, and M. K. Tiwari. 2018. “Solving Closed-loop Supply Chain Problems Using Game Theoretic Particle
Swarm Optimisation.” International Journal of Production Research 56 (17): 5836–5853.
Pishvaee, M. S., R. Z. Farahani, and W. Dullaert. 2010. “A Memetic Algorithm for Bi-objective Integrated Forward/reverse Logistics
Network Design.” Computers & Operations Research 37 (6): 1100–1112.
Pishvaee, M. S., F. Jolai, and J. Razmi. 2009. “A Stochastic Optimization Model for Integrated Forward/reverse Logistics Network
Design.” Journal of Manufacturing Systems 28 (4): 107–114.
Pishvaee, M. S., M. Rabbani, and S. A. Torabi. 2011. “A Robust Optimization Approach to Closed-loop Supply Chain Network Design
Under Uncertainty.” Applied Mathematical Modelling 35 (2): 637–649.
Prins, C., C. Prodhon, A. Ruiz, P. Soriano, and R. Wolfler Calvo. 2007. “Solving the Capacitated Location-routing Problem by a
Cooperative Lagrangean Relaxation-granular Tabu Search Heuristic.” Transportation Science 41 (4): 470–483.
Punyim, P., A. Karoonsoontawong, A. Unnikrishnan, and C. Xie. 2018. “Tabu Search Heuristic for Joint Location-inventory Problem
with Stochastic Inventory Capacity and Practicality Constraints.” Networks and Spatial Economics 18 (1): 51–84.
Radhi, M., and G. Zhang. 2016. “Optimal Configuration of Remanufacturing Supply Network with Return Quality Decision.”
International Journal of Production Research 54 (5): 1487–1502.
Ramezani, M., A. M. Kimiagari, B. Karimi, and T. H. Hejazi. 2014. “Closed-loop Supply Chain Network Design Under a Fuzzy
Environment.” Knowledge-Based Systems 59: 108–120.
Ross, A., and V. Jayaraman. 2008. “An Evaluation of New Heuristics for the Location of Cross-docks Distribution Centers in Supply
Chain Network Design.” Computers & Industrial Engineering 55 (1): 64–79.
Salema, M. I. G., A. P. Barbosa-Povoa, and A. Q. Novais. 2007. “An Optimization Model for the Design of a Capacitated Multi-product
Reverse Logistics Network with Uncertainty.” European Journal of Operational Research 179 (3): 1063–1077.
Srinivasan, S., and S. H. Khan. 2018. “Multi-stage Manufacturing/re-manufacturing Facility Location and Allocation Model Under
Uncertain Demand and Return.” The International Journal of Advanced Manufacturing Technology 94 (5–8): 2847–2860.
Srivastava, S. K. 2007. “Green Supply-chain Management: A State-of-the-art Literature Review.” International Journal of Management
Reviews 9 (1): 53–80.
Torres-Soto, J. E., and H. Üster. 2011. “Dynamic-demand Capacitated Facility Location Problems with and Without Relocation.”
International Journal of Production Research 49 (13): 3979–4005.
Üster, H., and S. O. Hwang. 2016. “Closed-loop Supply Chain Network Design Under Demand and Return Uncertainty.” Transportation
Science 51 (4): 1063–1085.
Wu, T., F. Chu, Z. Yang, Z. Zhou, and W. Zhou. 2017. “Lagrangean Relaxation and Hybrid Simulated Annealing Tabu Search Procedure
for a Two-echelon Capacitated Facility Location Problem with Plant Size Selection.” International Journal of Production Research
55 (9): 2540–2555.
Wu, T., Z. Yang, F. Chu, and Z. Zhou. 2015. “A Lagrangean Relaxation Approach for a Two-stage Capacitated Facility Location Problem
with Choice of Depot Size.” In 2015 IEEE 12th International Conference on Networking, Sensing and Control (ICNSC), Taipei,
Taiwan, 39–44. IEEE
Yousefi-Babadi, A., R. Tavakkoli-Moghaddam, A. Bozorgi-Amiri, and S. Seifi. 2017. “Designing a Reliable Multi-objective Queuing
Model of a Petrochemical Supply Chain Network Under Uncertainty: A Case Study.” Computers & Chemical Engineering 100:
177–197.
Yuchi, Q., Z. He, Z. Yang, and N. Wang. 2016. “A Location-inventory-routing Problem in Forward and Reverse Logistics Network
Design.” Discrete Dynamics in Nature and Society 2016: Article ID 3475369, 2016, (pp. 1-18)
Zhalechian, M., R. Tavakkoli-Moghaddam, B. Zahiri, and M. Mohammadi. 2016. “Sustainable Design of a Closed-loop Location-routing-
inventory Supply Chain Network Under Mixed Uncertainty.” Transportation Research Part E: Logistics and Transportation
Review 89: 182–214.
Zhang, Z.-H., G. Berenguer, and Z.-J. Shen. 2014. “A Capacitated Facility Location Model with Bidirectional Flows.” Transportation
Science 49 (1): 114–129.
Zhang, Z.-H., and A. Unnikrishnan. 2016. “A Coordinated Location-inventory Problem in Closed-loop Supply Chain.” Transportation
Research Part B: Methodological 89: 127–148.
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.

You might also like