Journal of Cleaner Production 162 (2017) 717e742

Contents lists available at ScienceDirect

Journal of Cleaner Production

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

Close loop supply chain network problem with uncertainty in demand

and returned products: Genetic articial bee colony algorithm
approach
Yan Yan Cui a, Zailin Guan a, Ullah Saif a, b, *, Li Zhang a, Fei Zhang a, Jahanzaib Mirza a
a
State Key Lab of Digital Manufacturing Equipment and Technology, HUST-SANY Joint Lab of Advanced Manufacturing, Huazhong University of Science and
Technology, Wuhan, 430074, PR China
b
Department of Industrial Engineering, University of Engineering and Technology, Taxila, Pakistan

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

Article history: In recent years, due to environmental concerns, remanufacturing of products is practiced in different
Received 25 July 2016 companies and closed loop supply chain network in these companies is signicant to optimize. There-
Received in revised form fore, current study is aimed to determine an optimal closed loop supply chain network, which is
7 June 2017
composed of multiple producers, remanufacturers, intermediate centers and customer centers.
Accepted 9 June 2017
Available online 10 June 2017
Furthermore, uncertainty in the demand and uncertainty in the quantity of returned products is
considered simultaneously in the network to make it signicantly useful in the uncertain environment. A
novel genetic articial bee colony (GABC) algorithm is introduced with a new food source representation
Keywords:
Close loop supply chain network
for the current problem. The proposed GABC algorithm considered neighbor food sources for local search
Remanufacturing and used crossover and mutation operations of genetic algorithm to enhance the exploration ability of
Demand uncertainty the proposed algorithm. Taguchi method is employed to compute the optimal parameters of GABC for
Genetic articial bee colony algorithm two different size test problems taken from literature and a Case problem which are modied according
to the current research problem. The performance of presented GABC algorithm is tested by comparing
the results of considered test problems with the results obtained from the original articial bee colony
(ABC) algorithm and genetic algorithm (GA). Moreover, to test the robustness of the proposed GABC
algorithm, different scenarios of small and large size problems based on the quantity of demand and
variations in demand are made to perform the experiments. Results indicate that proposed GABC out-
performs standard ABC and GA in different scenarios to give smaller value of the total cost of network
and gives more robust results to give smaller variations in the total cost of network due to uncertain
variations in the demand, as compared to original ABC and GA.d
2017 Elsevier Ltd. All rights reserved.

1. Introduction
Close loop supply chain optimization problem is gaining popu-
larity in recent years due to environmental concerns of different
manufacturing industries. Close loop supply chain consists of for-
ward and reverse loop of supply chain, which works simulta-
neously. In the forward loop, the newly manufactured products are
moved to the customers in different areas. While in the reverse
* Corresponding author. State Key Lab of Digital Manufacturing Equipment and
Technology, HUST-SANY Joint Lab of Advanced Manufacturing, Huazhong Univer-
sity of Science and Technology, Wuhan, 430074, PR China; Department of Industrial
Engineering, University of Engineering and Technology, Taxila, Pakistan.
E-mail address: saifullah47@yahoo.com (U. Saif).
loop, used products are collected from customers and moved to
intermediate centers for cleaning and disassembly operations. Later
they are moved to the remanufacturing facilities for remanu-
facturing. Close loop of supply chain is signicant to reduce envi-
ronmental pollution due to reuse of scrap and used products from
customers. Many companies now a days are using reverse ow of
their used products from customers to recovery centers. The re-
covery options of the used products includes reuse of the collected
products, repairing of the returned items, remanufacturing or the
combination of these options (Thierry et al., 1993; Zhiqiang and
Nathalie, 2007). The reverse supply chain network system can be
designed for direct reusable network, repairing network, recycling
network and remanufacturing network (Fleischmann et al., 1997;
Thierry et al., 1993; Zhiqiang and Nathalie, 2007). In reusable

http://dx.doi.org/10.1016/j.jclepro.2017.06.079
0959-6526/ 2017 Elsevier Ltd. All rights reserved.
718 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
network, the products are directly reused, for example, soft drink
bottles, pallets or containers etc. The repairing logistics network
restores the failed products to their working conditions. For
example, reconditioning of the domestic appliances including,
washing machines, electric equipment, refrigerators etc, are
involved in the repairing network. In recycling process, material is
recovered without converting product structure, for example, steel
scrap is used to manufacture different steel products, glass, papers
etc. In remanufacturing, a product is reformed into its original
shape by several operations which can include disassembly, over-
hauling cleaning and replacement operation etc. Remanufacturing
systems are used to remanufacture several kind of machineries like,
air craft, automobiles etc. Current research is focused on remanu-
facturing, as recovery option, for the used products in close loop
supply chain network.
Closed loop supply chain network problem is the focus of many
researchers due to the necessity of reducing the use of raw material,
controlling the environmental pollution, recycle management and
due to the social responsibility (Zohal and Soleimani, 2016). The
supply chain network design is a strategic decision problem which
is aimed to decide number of different facilities required to make in
the network. Furthermore, supply chain network optimization is
important to identify the ideal location of these facilities in the
network and identify the effective connections of these facilities
with each other to satisfy the overall demand of the customers with
minimum total cost of the network. The location and position of the
facilities in the network can affect, the investment cost, service
level of the customers and has its much practical implication in the
real remanufacturing industries now a days. Aras et al. (2007)
determined an optimal locations of the collection centers and
maximize the purchase price of the products used in reverse lo-
gistics network. Zohal and Soleimani (2016) presented a closed
loop supply chain problem for a gold industry to minimize the cost
and carbon emissions. Yi et al. (2016) presented a closed loop
supply chain network design problem for a construction machinery
remanufacturing company to minimize total cost of the network
which includes, xed cost, transportation cost, processing cost,
revenue cost and cost of new components. In closed loop supply
chain network, the location decision of all facilities signicantly
affects the remanufacturing process. For example, the location
position of the customer centers need to be optimized to satisfy the
demand of customers, located in the areas near the location posi-
tion of the customer centers. In addition, the location and number
of intermediate facilities is substantial to know while making the
optimal network because in some areas, there is more population of
customers who are willing to return the used products to the
company and in some other areas, the population of customers is
less prepared to return the used products for remanufacturing.
Therefore, the location decision of different facilities is signicant in
the network design and optimization problem for different rema-
nufacturing industries. Furthermore, there are some capacity con-
straints related to each location position of the customer centers
which are required to be considered while making the network.
The capacity of the customer centers can affect the service level of
the customers and affects the quantity of the returned products
that can be collected from different customers. In addition, there is
transportation cost which can occur in the network and the
transportation cost depends on the decision of the connections
between different facilities in the network. The closed loop supply
chain network is concerned to determine the relationship between
all the facilities in the closed loop supply chain, including customer
centers, the collection centers, the manufacturers and re-
manufacturers so that the network is optimized and it can satisfy
the overall demand of different customers located in different
areas. However, in real environment, the closed loop supply chain
has uncertain environment and the demand of the customers and
quantity of the returned products is uncertain. The need to account
uncertainty in the supply chain network is essential to design
closed loop supply chain network model in strategic view (Ozkir
and Basligil, 2013). The demand of the new products from
different customer centers is varying and it is not constant in most
of the real cases. Furthermore, the quantity of the returned prod-
ucts from different customer centers is also uncertain and varies
due to several factors. Therefore, it is noteworthy to consider un-
certain demand of new products and uncertainty in the quantity of
returned products while making strategic decision on the closed
loop supply chain network and therefore considered in the current
research.
The closed loop supply chain network problem is combinatorial
optimization problem and it can take long computational time if it
is solved with exact methods. Therefore, a novel metaheuristic
method which has not been used in literature to solve closed loop
network optimization problem is introduced in the current
research. The performance of the proposed algorithm is tested on
different size problems under several scenarios of demand and
returned products and the corresponding results are compared
with the performance of other algorithms.
The structure of the current research paper is as follows: Section
2 shows the review of literature work on closed loop supply chain
problem. Section 3 shows a mathematical model for the considered
closed loop supply chain network problem. Section 4 illustrates
proposed novel solution approach based on articial bee colony
algorithm. Section 5 indicates design of experiments and tuning of
parameters of the proposed algorithm on two different size prob-
lems to get optimal values of algorithm parameters for these
problems. Furthermore, it also shows the results of considered two
problems. Section 6 illustrate the implementation of the proposed
GABC to solve the small and large size problems in detail with
considering uncertainty in demand and in the returned products.
Moreover, scenario planning is described in this section to test
different category of problems based on different values of the
variations in demand and returned products and different ranges of
the uncertain demand and returned products. Furthermore,
different scenarios are considered in each category of test problem
and the results are explained in this section. Section 7 shows the
conclusion of the current research and future recommendations.
2. Literature review
In literature, closed loop supply chain network optimization
problem has been investigated by several researchers. Fleischmann
et al. (1997) presented a comprehensive literature review on the
application of mathematical modeling in reverse logistics man-
agement and shows different models which have been presented in
literature. Krikke et al. (1999) proposed a mixed integer program-
ming model (MILP) for reverse logistics network for photocopier
manufacturer and considered minimization of supply chain cost,
energy and residual waste in close loop supply chain. Uster et al.
(2007) developed a closed loop supply chain network model to
optimize the ow of products in the network. They used collection
centers, remanufacturing facilities in their proposed network. Their
study focused to minimize the processing cost, transportation and
xed costs of facilities in the close loop supply chain network. Lu
and Bostel (2007) proposed a mixed integer programming model
containing three types of facilities in the network. They considered
customers, intermediate centers, remanufacturing centers and
producers in the close loop supply chain network. Du and Evans
(2008) studied close loop supply chain problem to minimize
overall cost and total tardiness. They supposed manufacturers,
collection sites i.e., third party logistics, repairing facilities in the
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
closed loop network. Gupta and Evans (2009) presented a non-
preemptive goal programming approach to model closed loop
supply chain network. They used, manufacturing cost, refurbishing
cost, disassembly cost, testing cost, remanufacturing cost, disposal
cost and transportation cost in their developed model. Fleischmann
et al. (2001) reported a generic model for the design of closed loop
supply chain and observed that considerable savings can be made
in cost of production due to simultaneous forward and reverse
network integration. They included customer centers, warehouses
and production plants in the close loop supply chain network.
Nevertheless, in most of the above mentioned literature, the
customer demand and returned quantity of used products is
considered as deterministic. However, in real environment, de-
mand and quantity of returned products is uncertain. The closed
loop supply chain network which is optimal for a xed value of
demand may not be suitable to use for different values of the de-
mands. This may give different value of cost of transportation,
storage cost or may not deliver products in time to the customers
when the demand of new products and quantity of returned
products in the network varies. Furthermore, the closed loop
network which is based on deterministic demand and returns, can
give local optimal solution of the network and it cannot be used for
large range of uncertain values of the demand of products. There-
fore, it is signicant to use uncertainty in the demand of new
products and in the quantity of returned products simultaneously
in closed loop supply chain network problem. In literature, few
studies has considered uncertainty in the demand and returned
products for different customers in closed loop supply chain
network problem. For example, Salema et al. (2007) extended the
model proposed by Fleischmann et al. (2001) and incorporated
uncertainty in demand and returned products using mixed integer
programming approach. They considered scenario planning
approach to consider the uncertainty in demand of products.
Francas and Minner (2009) proposed two stage stochastic model to
design closed loop supply chain network considering uncertain
demand and returns. Shi et al. (2010) considered uncertain demand
and returns and proposed stochastic programming approach for
closed loop supply chain network optimization. Pishvaee et al.
(2011) considered uncertainty in demand in closed loop supply
chain and proposed a robust approach to deal with the uncertainty.
Vahdani et al. (2012) applied fuzzy approach for robust optimiza-
tion of closed loop supply chain network optimization problem.
They considered suppliers, manufacturing facilities, customers,
some bidirectional facilities in their considered closed loop supply
chain network. They aimed to minimize total cost which includes
the xed cost, transportation cost in forward and reverse ow in
the network. Recently, Saif et al. (2015) presented a brief descrip-
tion of a model for facility location problem in closed loop supply
chain network problem but have not solved it. Saman and Zhang
(2013) proposed a facility location model in closed loop supply
chain network considering uncertainty in demand and returns.
They considered stochastic programming and scenario planning
approach to include uncertainty in the demand and returned
products. They used multi plants, collection centers, demand
markets and products. They proposed a mixed integer program-
ming model to minimize total cost. However, they assumed a xed
location of the plants and collection centers which is known in
advance in their model. However, in closed loop supply chain
network problems, the xed location for certain facilities can limit
the diversity of solution of network and cannot give global optimal
solution. In the network optimization problems, if the location sites
are not xed for the facilities, then it may give a network solution
which can consider all possibilities of placing facilities in all loca-
tion sites and can give more global search for network solution.
Therefore, current research considers a closed loop supply chain
719
network composed of customer centers, intermediate centers,
producers and remanufacturers or a facility containing both pro-
ducer and remanufacturer. The location sites for any facility is not
xed and each facility has possibility to be allocated to any site in
the network while making the optimal network. This can give more
diversity and can give global solution of the network in the current
research.
Moreover, in literature, demand uncertainty is dealt with sto-
chastic programming, or scenario planning approach and in real
environment, the value of demand in different scenarios depends
on several factors which has not been included in previous re-
searches. The demand of new products and quantity of returned
products varies randomly and it depends on different factors of the
market. Furthermore, each factor has different weightage on
changing the demand and returned quantity in different facility
locations. For example, demand of returned products in certain
areas depends on unexpected events occurring in market, uncer-
tain changes in season, uncertain fall in the prices of competitors or
uncertain changes in the taxes or other price. These demand related
factors can cause uncertain variations in the demand of the prod-
ucts. Product demand uncertainty occurs due to several factors and
each factor has its own probability to occur. Furthermore, each
factor has its own impact in changing the value of the demand.
Similarly, there are several factors which can cause uncertain
change in the quantity of the returned products. For example, a
different product model is introduced by some other manufac-
turers to the market, then there is possibility that customers may
return some products to the company and this may change the
quantity of returned products randomly. These factors have
different weightage and each factor has different range to cause
change in quantity of the returned products. The uncertain demand
of new products and uncertain quantity of the returned products
has not been investigated in literature based on these factors.
Moreover, the scenarios for demand and uncertain quantity of
returned products based on these factors has not been investigated
in literature. Current research is novel to include demand uncer-
tainty in new products and uncertainty in the quantity of the
returned products based on several factors for making different
scenarios. Furthermore, in literature, holding cost in the network
has not been considered during optimizing the closed loop supply
chain network (Salema et al., 2007; Pishvaee et al., 2011; Vahdani
et al., 2012; Zohal and Soleimani, 2016). However, the quantity of
returned products and quantity of products manufactured in the
plants also incurs holding costs but it has not been investigated in
literature in closed loop supply chain network optimization prob-
lems (Salema et al., 2007; Pishvaee et al., 2011). However, the
holding cost of returned products in the collection centers and the
holding cost of the new and remanufactured products in the
manufacturing facilities are required to be considered during
optimization of the network and therefore considered in the cur-
rent research. In addition, the safety stock of new products and
returned products for remanufacturing in the remanufacturing
centers is required to be included and is considered in the current
study. Current research is novel to include holding cost in the in-
termediate centers and remanufacturers in the closed loop supply
chain network. Moreover, safety stock of the returned products in
intermediate centers, safety stock of the products in the producers,
safety stock of remanufactured products in the remanufacturing
centers is rst time considered simultaneously in the current study.
In addition, several additional constraints in the closed loop supply
chain network problem are used in the current research which
includes the safety stock constraints, storage capacity constraints
and service level constraints to store the remanufacturing products,
returned products in the producers and remanufacturing centers.
In literature, different solution methods has been used to solve
720 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
closed loop supply chain problem. For example, Salema et al. (2009)
presented a mixed integer programming model and proposed
branch and bound method to solve their presented problem. Other
methods which are used in literature include, nested diagonaliza-
tion solution algorithm (Chen et al. (2007), an OR opt algorithm
procedure (Alshamrani et al., 2007), scatter search algorithm (Du
and Evans, 2008), two phase heuristic method based on greedy
construction (Wollenweber, 2008), heuristic algorithm based on
imperialist competitive algorithm and variable neighborhood
search (Devika et al., 2014; Fakhrzad and Moobed, 2010) etc. In
addition to these methods, metaheuristic methods including ge-
netic algorithm (Lee and Chan, 2009; Aydin et al., 2016; Yi et al.,
2016), genetic algorithm with priority based encoding method
(Lee and Chan, 2009), Simulated annealing with dynamic neigh-
borhood search (Pishvaee et al., 2010), Ant colony algorithm (Zohal
and Soleimani, 2016) etc, have also been used to solve closed loop
supply chain network optimization problem. Furthermore, in
literature there are some studies which used simulation method for
closed loop supply chain network optimization problem (Tonanont
et al., 2009; Suryabatmaz et al., 2004). Recently, Karaboga (2005)
introduced an algorithm called articial bee colony algorithm
(ABC) which is based on the foraging behavior of bees is an excel-
lent algorithm. The original structure of ABC is useful for contin-
uous problems. However, ABC algorithm has also been successfully
applied in discrete nature of problems including, jobshop (Li et al.,
2011), and owshop problems (Fatih Tasgetiren et al. (2011). ABC
algorithm needs less control parameters and is easy to implement
and to the best of authors knowledge, it has not been used to solve
closed loop supply chain network optimization problem. Current
research is novel to present genetic articial bee colony (GABC)
algorithm to solve the considered closed loop supply chain network
optimization problem.
Current research is focused to develop a model for closed loop
supply chain network considering uncertainty in demand and
quantity of the returned products in an automobile manufacturing
company. The main questions which current research seeks to
address is
How to develop closed loop supply chain network model
considering uncertainty in demand and in the quantity of the
returned products?
How to solve it using a novel metaheuristic algorithm?
Current research is aimed:
To identify the optimal location of different facilities in the
closed loop supply chain network
To optimize and decide the connections between different fa-
cilities in the closed loop supply chain network.
To consider production cost, transportation cost, holding cost of
inventory in facilities, xed cost, capacity constraints of the fa-
cilities, the service level constraints and some additional con-
straints in the closed loop supply chain simultaneously.
To consider uncertainty in the demand of new products and
quantity of the returned products by incorporating change in
demand and returned products considering different factors
which are causing uncertainty in making the scenarios
simultaneously.
To develop a novel genetic articial bee colony algorithm to
search an optimal solution of the considered closed loop supply
chain network optimization problem.
In the optimized network, the total cost of investment including
the total cost of the network is minimized. Furthermore, the un-
certain demand of customers is also fullled with the given service
level and optimal inventory levels. In addition, the quantity of the
returned products from different customers is optimally used
during remanufacturing.
3. Problem description
In this section the proposed facility location problem in closed
loop supply chain network is presented. In the considered closed
loop supply chain network, four kinds of facilities are considered.
These facilities includes customer centers to collect the used
products from customers, intermediate centers which are used for
cleaning, disassembly, checking, and ranking the returned products
before they enter into the remanufacturing centers. The third kind
of facility used in the proposed network is remanufacturing centers
where the returned products from intermediate centers are
checked again and they are remanufactured. The fourth type of
facility considered in the current research problem is producers
which are used to make new products according to the demand of
products. The produced products and the remanufactured products
are directly transported to the customer centers for sale. There are
two transportation roots in the considered network problem. One
root is used to move returned products from customer centers to
intermediate centers and then to the remanufacturing centers or
producers, called reverse ow. The other root is used to move
products from remanufacturing facilities and producers to the
customer centers called forward ow. Fig. 1 illustrates closed loop
supply chain network with forward and reverse ows of products
between different facilities. The notations and abbreviations used
for the current problem are described as under.
Indices
C f1; 2; :::; i; ::::; nC g: Set of different locations of customer
centers
I f1; 2; ::::; j; ::::; nI g: Set of intermediate center
P f1; 2; ::; k; :::::; nR g: Set of different producers or remanu-
facturing centers at location kVi fvi1 ; vi2 ; :::::; viM g: Set of different
factors which may cause variance in demand at customer center i
Parameters
Fixed Costs:
fkP : Fixed cost to setup a producer center at location kfkR : Fixed
cost to setup a remanufacturing center at location kfjI : Fixed cost to
setup an intermediate center at location jfiC : Fixed cost to setup
customer center at location i
Unit Costs:
UCkP : Unit cost at producer center at location kUCkR : Unit rema-
nufacturing cost at remanufacturing center at location kUCjI : Unit
reprocessing cost at intermediate center at location j
Shipping Costs:
SCki : Shipping cost from producer or remanufacturing center k
to customer center iSCij : Shipping cost from customer center i to
intermediate center at location jSCjk : Shipping cost from interme-
diate center j to producer or remanufacturing center k
Disposal Costs:
DCjI : Disposal cost at intermediate center jDCkR : Disposal cost at
remanufacturing center k
Holding Costs:
Hk : Holding cost of returned and produced products at rema-
nufacturing center or producer kHjI : Holding cost of returned pro-
duced products at intermediate center j
Demand:
di  Ndi ; dsi : Normal distributed demand of the product at
m
customer center idri  Ndi r ; disr : Normal distributed returned
m
m
products at customer center idi : Mean value of demand of the
s
product at customer centeridi : Variance value of demand of the
product at customer center i
Returns:
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 721

Producers or Intermediate Customer

Remanufacturers Centers Centers

Produced or
Remanufactured
Products Flow

Disposed Products in Returning

Remanufacturing Centers Products
Flow

Disposed Products in
Intermediate Centers

Fig. 1. Closed loop supply chain network.

m
di r : Mean value of quantity of the returned product at customer
center idsi r : Variance value of quantity of the returned product at
m
customer center iedi : Estimated average value of demand of the
m
product at customer center iedi r : Estimated average value of
quantity of the returned product at customer center i
Demand Variations:
bim : The weight for the factor m for customer center i which
causes variance in the demand
Number of Facilities:
NjC : Maximum number of customers which are allowed to
deliver their returned products to intermediate center jNkI :
Maximum number of intermediate centers which are allowed to
deliver their returned products to remanufacturing or producer at
locationkNkC : Maximum number of customers which are allowed to
get delivery from products or remanufacturing center k
Service Level:
drP
k : Service level for the returned product at remanufacturing
center or producer kdPk : Service level for the product at remanu-
facturing center or producer k
Disposals:
b : Percentage at which returned products can be disposed of at
the intermediate centers (Suppose it is equal on all intermediate
centers)
g : Percentage at which checked returned products can be
disposed of at the remanufacturing centers (Suppose it is equal on
all remanufacturing centers)
Safety Stock:
SPk : Safety stock of the product at producer kSrP k
: Safety stock of
the returned products at producer or remanufacturing center kSrI j :
Safety stock of the returned products at intermediate center i
Capacity of Facilities:
SCapPk : Storage capacity of products at producer or remanu-
facturing center kSCaprP k
: Storage capacity of the returned products
at producer kSCaprI j : Storage capacity of the returned products at
intermediate center j
Decision variables
Continuous variables:
XPik : Fraction of demand from customer centeri to remanufac-
turer or producer at location kXRijk : Fraction of returned products at
customer centeri taken through intermediate center j to remanu-
facturer at location k
Binary Variables:


1 if ;
YPk Producer is at position k

0 Otherwise
1 if ;
YRk Remanufact uring center is at position k

0 Otherwise
1 if ;
YIj Intermediate center is at position j

0 Otherwise
1 if;
YCj Customer center is at position i
0 Otherwise
Consider there are known number of locations sites for making
facilities in the closed loop supply chain network. From these
location sites, there are known number of producers, re-
manufacturers, intermediate centers and customer centers which
are desired to be located on known number of site location points to
make a closed loop supply chain network. The number of site lo-
cations is known and it is required to decide, at which location
point, which facility is to be build and with which facilities it can be
connected in the optimized network. The location decision of these
facilities depends on the initial cost called, xed cost of these fa-
cilities at different location sites. Moreover, the location decision of
facilities also depends on their connections with other facilities so
that the transportation cost between them is minimum. For
example, the location of customer centers not only depends on its
xed cost, but it also depends on its possible connections with
intermediate centers so that it can transport the returned products
to the intermediate center with smaller transportation cost. It is
possible that a customer center can have possibility to get con-
nected with different intermediate centers. However, the decision
of its connection also depends on the distance of the intermediate
center and its storage capacity so that it can economically stores the
returned products and the transportation cost is also desired to be
minimum. Furthermore, there is possibility that the demand of the
considered customer center might not be fullled with one
manufacturer, so it can be connected with more than one pro-
ducers. The considered network problem is to assign optimal
722 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

location sites to the facilities in the network so that the demand of

customer centers is fullled, the returned products are moved X
nR X
nR X
nI X
nC
Fixed:cos t fkP YPk fkR YPk fjI $YIj fiC $YCi
optimally to the intermediate centers and from intermediate cen-
k1 k1 j1 i1
ters they can be moved optimally to the remanufacturing facilities
keeping transportation cost and holding cost as minimal. Moreover, (2)
the production cost are also different for manufacturers and in-
termediate centers located at different location sites. In addition to
these, the returned products are moved from different customer
3.1.2. Shipping cost
centers and some of these products are disposed at the interme-
Shipping cost in the network is a part of collection cost in the
diate centers while remaining products are moved to the re-
current research problem. The overall collection cost consists of the
manufacturers. The remanufacturers and producers have two
cost of purchase of the returned products and the transportation
choices to be located in the network. For example, the remanu-
cost involved in the network. However, in the current research, the
facturer is allocated to a site location or producer is located on that
purchase cost of returned products is considered as constant in
site location, or both the remanufacturer and producer are located
different location sites in the network and it do not affect the
at the same site location in the proposed network, as can be seen
network and the transportation cost is considered as shipping cost.
from Fig. 1. The considered problem is to optimally allocate the
This shipping cost is the cost required to move the products from
facilities in the closed loop supply chain network. The objective
one facility to some other facility in the closed loop supply chain
function for the network problem and constraints are presented in
network. Shipping cost in the network depends on the location
this section.
position of different facilities in the network between which the
products are required to be transported. Therefore, shipping cost is
variable in different networks and it is considered here to optimize
3.1. Objective function
the network. The shipping cost of the network is illustrated in
Equation (3). It can be seen from Equation (3) that the rst term in
The objective function of the considered problem is to minimize
Equation (3) shows the shipping cost to move the demand of
the total cost of the network, as indicated in Equation (1). It is
products from remanufacturing centers or producers to the
assumed that the purchase cost of the returned products is con-
required customer centers. The second and third term in Equation
stant in different customer centers and only the costs which are
(3) shows the transportation cost of the returned products from
variable in the network are considered for optimization. It can be
different customer centers to the intermediate centers and trans-
seen from Equation (1) that the total cost depends on the xed cost
portation cost of the checked returned products from intermediate
of the facilities, shipping costs in forward and reverse ows be-
centers to the remanufacturing centers or producers, respectively.
tween different facilities, total production cost and holding cost of
the products in inventory. In the current research, the cost of !
X
nR X
nC
m
X
M
collection in the network includes the purchase cost of the returned Shipping:cos t SCki edi bim vim XPik  YCi YPk
products and the shipping cost of the products in the network. k1 i1 m1
However, in the current problem it is assumed that the purchase !
X
nR X
nI X
nC
mr
X
M
cost of the returned product is constant in different locations of the SCij edi bim vim XRijk
collection centers and it is not affecting the network therefore the k1 j1 i1 m1
collection cost of the network depends on the shipping cost of the
  mr
network in the current research. Moreover, the refurbish cost of  YCi YIj 1  bSCjk edi
products consists of the production cost of new products and the
!
remanufacturing cost of the products in the remanufacturing cen- X
M  
ters. The storage cost is considered here as holding cost of the bim vim XRijk  YIj YPk
returned products in different facilities. m1
(3)
Objective Minffixed:cos t Shipping:cos t Prod:cos t
Holding:cos tg where,
(1) m m P
di edi M m1 bim vim represents mean value of demand of
products at customer centeri which is equal to the sum of the
average value of estimated demand in certain customer centeri and
change in demand caused by different factors.
3.1.1. Fixed cost
Fixed cost is the sum of costs of creating facilities including, 3.1.3. Production cost
producers, remanufacturers, intermediate centers and customer The refurbishing cost in the current research consists of pro-
centers on different location sites, as shown in Equation (2). It can duction cost of new products and remanufacturing cost of the
be seen from Equation (2) that the xed cost is the sum of the xed remanufactured products. Total production cost is composed of the
cost of all facilities in their location sites in the network. The rst unit production cost times the net demand of products that are
term in Equation (2) illustrates the sum of xed cost of certain required to produce, as indicated in Equation (4). It can be seen
number of producers in their different location sites in the network, form Equation (4) that the total production cost consists of the sum
the second term indicates the sum of xed cost of certain number of of three terms. The rst term in Equation (4) shows the total cost of
remanufacturing facilities in their different location sites of the manufacture of new products in the producers which are needed to
network, the third term in Equation (2) shows the xed cost of manufacture to full the demand. The second term in Equation (4)
certain number of intermediate centers in their different location illustrates the total cost of remanufacturing in the remanufacturing
sites in the network and the fourth term in Equation (2) shows the facilities in the network. The third term in Equation (4) illustrated
sum of xed cost of the certain number of customer centers which the total disposal cost of the returned products in the intermediate
are positioned in different location sites in the network. centers and in the remanufacturing facilities respectively.
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 723

" ! rm
X
nR X
nC X
M SrR  1  bdi
Prod:cos t UCkP 
m
edi bim vim XPki  YCi YPk ZkrR k
(8)
1  bdri s
k1 i1 m1
# 2
nR X
X nI X
nC
rm
 SPk UCkR  4SrP
k 1  g1  b
SrI  di
ZkrI k
(9)
k1 j1 i1 dri s
! 3
nI X
X nC X
M
bim vim XRijk YCi YIj 5
rm
edi
j1 i1 m1
X
nR X nC hn
nI X o 3.2. Constraints
1  bUCjI b:DCjI  YIj
k1 j1 i1
n o i 3.2.1. Production and demand constraints
1  b1  gUCkR g1  b  DCkR  YPk The new products and returned products are required to fully
! meet their demand. Equation (10) shows that, for each customer
X
M
 edi
mr
bim vim XRijk YCi center, the demand is collectively satised from all producers or
m1 remanufacturing centers connected with it. The constraint shown
in Equation (11) indicates that for each customer, the quantity of
(4)
the returned products are moved to all intermediate centers con-
nected with it which are then moved to all of the producers or
remanufacturing centers connected with each intermediate
3.1.4. Holding cost centers.
Holding cost or the storage cost is incurred due to inventory of
P
nR
the returned products in the remanufacturing centers and the in- XPik ; 1 ci (10)
ventory in the producers. It is the sum of total holding cost of the k1
new products in the producers, holding cost of returned products in
the producers, holding cost of returned products in the remanu- P
nR P
nI
XRijk 1; ci (11)
facturing centers and holding cost of the returned products in the k1 j1
intermediate centers, as illustrated in Equation (5). Equation (6) to
Equation (9) are used to calculate the probability to ensure that the The constraint shown in Equation (12) shows that the sum of
safety stock levels of new products in producers, safety stock of safety stock of new products, safety stock of the returned products
remanufactured products in the producers, safety stock of rema- in remanufacturing centers and quantity of the returned products
nufactured products in the remanufacturing canters and safety entering in the remanufacturing centers should meet the total
stock level of returned products in the intermediate centers can demand of new products.
satisfy their corresponding demands in the network, respectively.

" !# " !# "

X
nR X
nC X
nR X
nC X
nR X
nC
Holding:cos t hk max 0; SPk  di  XPik YPk max 0; SrP
k  1  bdri  XRijk YPk max 0; SrR
k  1
k1 i1 k1 i1 k1 i1
!# " !#
X
nI X
nC
 bdri  XRijk YRk max 0; SrI
j  dri  XRijk YIj
j1 i1
2 v 3 2 v 3
u nC u nC
X
nR uX  s 2 X
nR uX  2  2
hk 4ZkP t d i  XPik YPk 25
hk 4Z t
rP
k 1  bdsr  XRijk YPk 5
i
k1 i1 k1 i1
2 v 3 2 v 3
u nC  2 u nC
X
nR uX 
rR t sr 2
X
nI uX  sr 2  2
4
hk Zk 1  bdi  XRijk YRk 5 hj 4ZjrI t di  XRijk YIj 5
k1 i1 j1 i1

(5)

!
X
nI X
nC
rm
X
M
SPk SrP
k 1  g1  b edi bim vim XRijk YCi YIj
j1 i1 m1
m
SPk  di !
ZkP (6) X
nC X
M
dsi  edi
m
bim vim Xik YCi YPk ; ck
i1 m1
rm (12)
SrP  1  bdi
ZkrP k
(7)
1  bdri s The difference between the net quantities of the new products
required with the net quantity of the returned products can give
724 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

the quantity of products that are needed to produce, if there is a ! !

producer at position k. The constraint indicated in Equation (13), X
nI
rm
X
M
1  b edi bim vim XRijk  YIj  YPk SrP P
k Sk
Equation (14) and Equation (15) are signicant to link the location
j1 m1
and allocation variables in which three different type of facilities !
are related to three different location variables and forward and X
nI
m
X
M
 edi bim vim Xik  YIj  YPk
backward ows are linked to the three location variables, i.e.,
j1 m1
customer centers, intermediate centers and producers or remanu-
facturing centers.  SCapPk ; ck
" ! # 2 (20)
X
nC X
M
edi
m
bim vim Xik YCi  SPk  4SrP
k
i1 m1
! 3
X
nC X
nI X
M
3.2.4. Service level constraints
bim vim XRijk YCi YIj 5
rm
1  g1  b edi
i1 j1 m1
In the proposed problem, the demand of new products and
quantity of returned products is uncertain so there is possibility of
 YPk M; ck losing service level due to late delivery of products. Therefore,
(13) service level constraints are presented here which are similar to the
service level constraints suggested by Glpinar et al. (2003) for
XRijk  YCj ; ci; j; k (14) supply chain. These constraints can dene the probability of
ensuring that the safety stock of the produced products and the
! safety stock of the returned products in the remanufacturing cen-
X
M
1  b edi
rm
bim vim XRijk  YRk M ; ci; j; k (15) ters or producers must be greater than their required or demanded
m1 quantity. The constraint shown in Equation (21) tells that at any
demand of the product, there is a positive probability that the
safety stock of new products is greater than the required demand,
or the safety stock is sufcient to fulll the required demand of
3.2.2. Non negativity constraints products. However, there is possibility that all of the demand of
Non negativity constraints are indicated in Equation (16) and products cannot be fullled by using these returned products. So
Equation (17), which are used for the integration of constraints and the constraint indicated in Equation (22) explains that the safety
non-negativity constraints respectively. stock of new products can have such a level that it can fulll the
remaining demand of products. The constraint indicated in Equa-
YPk ; YRk ; YIj 0; 1; cj; k (16)
tion (22) tells that there is positive probability that the sum of
safety stock of the returned products in the producers or in the
Xik ; XRijk  0; ci; j; k (17) remanufacturing centers and the newly entered returned products
is greater than demand of products. This constraint can ensure that
the existing returned products in stock and the newly entered
returned products can easily be utilized in fullling the demand of
3.2.3. Storage capacity constraints
new products.
There are some storage capacity constraints that can explain the
storage capacity of returned products in intermediate centers,
remanufacturing centers, and storage capacity constraint in the ( ! ! )
X
nC
m
X
M
P
producers to store returned products and new products in it ProbP SPk  edi bim vim Xik 0  dk ; ck
respectively. The storage capacity constraints are indicated in i1 m1
Equation (18), Equation (19) and Equation (20). (21)
!
P
nC
rm
X
M
edi bim vim XRijk  YCi  YIj SrI rI
j  SCapj cj
i1 m1
(18)

80 ! 1 ! ! 9
< X
nC X
nI X
M X
nC X
M =
ProbrP @SrP bim vim XRijk A 
r m m rP
1  g1  b edi edi bim vim Xik 0  dk ; ck (22)
: k ;
i1 j1 m1 i1 m1

3.2.5. Maximum number of facilities constraint connected to a

! facility
P
nI
rm
X
M
There are some other constraints which can satisfy the condition
1  b edi bim vim XRijk  YIj  YRk SrP rP
k  SCapk ; ck
j1
that each intermediate center can be assigned to some specic
m1
number of customer centers. Similarly, each remanufacturing
(19)
center or producers is connected with some specic number of
intermediate centers. Moreover, each producer or remanufacturing
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 725

center is responsible to deliver their products to some predened 4.2. Genetic articial bee colony algorithm
number of customer centers. These constraints are dened in
Equation (23) to Equation (27). In the original structure of ABC algorithm, each employee bee is
assigned only one food source which can be signicant to explore
P
nC the number of solutions simultaneously in the employee bee phase.
YCi  YIj  NjC ; cj (23)
i1
However, this exploration can further be improved with incorpo-
rating local search scheme for each employee bee. Due to the
complex nature of the current problem, the original ABC is modi-
P
nI
YIj  YRk  NkI ; ck (24) ed here by adding some steps of genetic algorithm in different
j1 phases of ABC algorithm and it is named as genetic articial bee
colony algorithm (GABC). The proposed GABC is designed for the
P
nI current close loop supply chain network optimization problem and
YIj  YPk  NkI ; ck (25) it is explained in this section.
j1

P
nC
YRk  YCi  NkC ; ck
i1
P
nC
YPk  YCi  NkC ; ck
i1
4. Method
4.1. Articial bee colony algorithm
Articial bee colony (ABC) algorithm is rst time proposed be
Karaboga (2005). This algorithm is inspired by foraging behavior of
honey bees. The algorithm is used to explore good quality solutions
which are represented as a food sources for the employee bees. The
nectar amount of a food source describes the quality of solution
which can be considered as the tness of the solution candidate.
Three kinds of bees are used in this algorithm. The rst half of the
population are employee bees and remaining half is considered as
onlooker bees. In the rst stage of ABC algorithm, employee bees
are sent to the food sources and each employee bee is assigned one
food source. Employee bees nd out the nectar amount of the food
sources assigned to them. The onlooker bees wait in the hive and
get the signicant information about the food sources with better
nectar amounts from the employee bees. Onlooker bees can decide
the employee bees to move on which direction for more food
source search. This is decided by onlooker bees on the basis of
nectar amount of the food source. The employee bees which have
no more improvement in their nectar amount can become scouts
bees and they can nd a new food source randomly.
4.2.1. Encoding scheme for solution representation
(26) The encoding scheme introduced for the current problem is
indicated in Fig. 2. The solution representation of a food source is
presented in Fig. 2 which gives the complete information of the
location sites for customer centers from the known number of total
(27)
site locations. Furthermore, the food source also explains the sites
allocated for intermediate centers, manufacturers or producers.
Moreover, the network information to identify the connection of
customer centers with intermediate centers, connection of inter-
mediate centers with remanufacturers or producers and the
connection of producers and remanufacturers with different cus-
tomers is also clearly indicated in Fig. 2. As an example, total 10 site
locations are available from which 6 customer centers, two inter-
mediate centers and two locations are used for remanufacturing
centers, or producers or both can be made on a single location. It
can be seen from Fig. 2 that there are two rows in the food source
vector. The numbers in the rst 6 elements of the food source
vector in the rst row indicate the location points to which the
customer centers are assigned. For example, the numbers in the
rst 6 elements of the rst row of food source are 1, 10, 5, 2, 6 and 9.
These numbers describes the location site from the available loca-
tion sites which are allocated for customer centers. The numbers in
the next two elements (7 and 8 element) in the rst row describe
the location sites assigned for intermediate centers, i.e., 4 and 7 are
allocated to intermediate centers. Furthermore, the numbers in
element 9 and 10 in the rst row of the food source vector shows
the location sites assigned to the remanufacturer centers or pro-
ducers. The remaining 6 elements in the rst row shows the 6
customer centers. The second row of the food source vector in-
dicates the information of the connections between different fa-
cilities. For example, the numbers in the rst 6 elements in the
second row shows the two intermediate centers with which the
corresponding customer center is connected. For example, the

Fig. 2. Solution representation for closed loop supply chain netwrok problem.
726 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

Initializing Food Sources

Send Employee Bee Onlooker Bee

Create Neighbors of Each Employee Bee

If No Improvement in Food Source

Select Best Neighbor of Each Employee Bee NO

YES
Update Archive
Send Scout Bee to Create New

Select Parent Food Source

Sorting Food Sources and Update Archive
Crossover

Mutation Solutions in Archive

Employee Bee Population
YES
New Food Source Population
NO

If Termination Criteria is Satisfied

NO

YES
Stop

Fig. 3. Flowchart of proposed GABC algorithm.

numbers 1, 2, 1, 1, 2, 1 in the rst 6 elements of the second row of
food source vector indicates that the customer centers allocated at
the location sites 1, 10, 5, 2, 6 and 9 are connected with the inter-
mediate center 1, 2, 1, 1, 2 and 1 respectively. The numbers shown in
element 7 and 8 in the second row of food source vector shows the
connection of the corresponding intermediate centers with the
producer or remanufacturers, i.e., intermediate center at site 4 and
site 7 are connected with the remanufacturer or producer 1 and 2
respectively. Moreover, the numbers shown in the element 9 and 10
facilities in neighbors of a food sources
of the second row in the food source shows that at their corre-
sponding location sites, which facility is build, the numbers can be
0, 1 and 2 which indicates producers, remanufacturers and both.
For example, for the elements 9 and 10, the location sites as
described in the rst row for produces and remanufacturers are 8
and 3. The corresponding numbers in the second row, i.e., 0 and 1
shows that at location sites 8 and 3 there is producer and a rema-
nufacturing centers respectively. The numbers 1 and 2 presented in
the last 6 elements of the second row of the food source vector

1 10 5 2 6 9 4 7 8 3
Different schemes of connection
Same location points for different

1 1 2 2 1 1 2 1 1 0
between facilities

1 10 5 2 6 9 4 7 8 3

1 2 1 1 2 1 1 2 0 1

1 10 5 2 6 9 4 7 8 3

2 1 2 1 1 2 1 1 1 0

Fig. 4. Neighbor food sources for an employee bee food source.

 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
shows rst and second facilities from the facilities described in the
element 9 and 10 of the second row. Furthermore, the numbers
shown in last 6 elements of the second row of the food source gives
information of the connection of the corresponding facilities with
the respective customers. For example, 1, 2, 2, 1, 2, 1 indicates that
the 6 customer centers i.e., 1, 2, 3, 4, 5, 6 are connected with the
producer or remanufacturer, (1 and 2 shows the rst and second
facility as described in the element 9 and 10, i.e., producer and
remanufacturer respectively). This representation of food sources
for the current problem is signicant and can completely dene the
solution of a problem. The ow chart of the proposed GABC is
shown in Fig. 3 and the step by step procedure is explained in this
section. The owchart of the proposed GABC is indicated in Fig. 3
and the algorithm is explained in this section.
4.2.2. Initialization of food source population
In GABC algorithm, the number of food sources initially created
is equal to the number of employee bees. The initial population of
food sources is randomly generated in GABC algorithm to induce
diversied solutions in the search space.
4.2.3. Sending employee bees
In this step, each employee bee is sent to its corresponding food
source and the corresponding step wise procedure is explained in
this section.
4.2.3.1. Create neighbors of food source of each employee bee.
In the proposed GABC, each food source assigned to an employee
bee creates its E number of food source neighbors. The neighbor of
an employee bee changes the information of connection schemes of
various facilities while the location sites positions are same for the
neighbors of an employee bee. For example, for each employee bee
the location sites assigned are same in all neighbors of an employee
bee and the difference only exists in the connections of the facil-
ities. Fig. 4 illustrates an example of the neighbor food sources of an
employee bee. This method of creating neighbor food sources is
used for local search of the food sources for each solution of
assigned locations to the facilities. Food source assigned to each
employee bee have different location points for the facilities while
for each employee bee, the neighbor food sources of the corre-
sponding employee bee have same location points of facilities and
the difference exists in the connection schemes of the facilities.
4.2.3.2. Select best neighbor of each employee bee. In this step, each
neighbor of food source of each employee bee is observed and its
nectar value is computed. The nectar amount of each food source r
is obtained from the relation shown in Equation (28). The neighbor
food source of each employee bee which gives maximum nectar is
considered as best food source neighbor of the corresponding
employee bee.
Parent 1 Food
1 10 5 2 6 9 4 7
Source
Parent 2 Food
4 6 1 10 5 3 7 2
Source
2 9 8 4 6 1 10 5 3
Offspring 1
5 3 7 2 6 9 4 8
Food Source
7 8 3 1 10 5 2 6 9
Offspring 2
6 9 4 10 5 3 7 8
Food Source
Fig. 5. Step 1 Order Crossover between selected parent food sources.
727
1
NECr
Fixed:cos t Shipping:cos t Prod:cos t Holding:cos tr
(28)
4.2.3.3. Update the archive. In proposed GABC, there is a food
source archive which stores one best neighbor of each employee
bee. The archive size is equal to the employee bee population. After
every cycle, the best neighbor food source of each employee bee
obtained is compared with the food source stored in archive and
the best one among them is stored in the archive to update the
archive.
4.2.3.4. Select parent food source. Two of the best food sources from
the archive are taken from the archive and considered as parent
food sources. Tournament selection is used for the selection of
parent food sources in GABC algorithm.
4.2.3.5. Crossover. In this step, crossover operation is performed
between the two parents food sources selected. The crossover
operation is performed in two steps in GABC algorithm.
Step 1: In this step the crossover only occurs in the portion of
food source where the information of location points of the facil-
ities is shown. Order crossover is performed in rst step of cross-
over operation in GABC. As an example, Fig. 5 illustrates the rst
step of crossover.
Step 2: In this step, two point crossover operations is performed
in the offspring food sources obtained in step 1. The crossover only
occurs in the portion of food source where the information of
connections between the facilities is described. Fig. 6 shows the
two point crossover in the offspring food sources obtained in step 1.
Two point crossover is performed in two different positions in the
food source as shown in Fig. 6.
4.2.3.6. Mutation. In this step, two point swap mutation is per-
formed in the offspring food source. The swap mutation is sepa-
rately performed in the rst and second row in the food source to
introduce the diversity in the food sources in both location points
assigned to facilities and the connection schemes of the facilities.
4.2.3.7. New food source population. New population of food
sources is formed by replacing the worst food sources with the
offspring food sources formed. The new population of food sour-
ces with their nectar amount information is sent to the onlooker
bees.
4.2.4. Sending onlooker bees
In this step, onlooker bee observers the nectar of each food
sources received from employee bee phase. The food sources which
do not give improvement in the nectar amount are sent to scout
8 3
bees. The food sources are sorted in decreasing order of their nectar
amount. The food sources along with the food sources in the
9 8 archive are sorted and stored in archive from top of the sorted list to
update the archive. The archive size is equal to the number of food
7 employee bees. The food sources in the updated archive are sent to
the next cycle of the algorithm.
1 10
4
4.2.5. Scout bee
1 2 If there is no improvement in a food source, the corresponding
bee is considered as a scout bee and its food source is created
randomly.
728 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

5 3 7 2 6 9 4 8 1 10 1 2 3 4 5 6

1 2 1 1 2 1 1 2 0 1 1 2 2 1 2 1

6 9 4 10 5 3 7 8 1 2 1 2 3 4 5 6
2 1 2 1 1 2 1 1 1 0 2 1 1 2 1 2

5 3 7 2 6 9 4 8 1 10 1 2 3 4 5 6

1 2 2 1 1 1 1 2 0 1 1 1 1 2 2 1

6 9 4 10 5 3 7 8 1 2 1 2 3 4 5 6

2 1 1 1 2 2 1 1 1 0 2 2 2 1 1 2

Fig. 6. Step 2 Two point crossover in the offspring of Step 1.

Table 1
Effective parameters of the proposed algorithm.

Levels Number of food sources Maximum number of algorithm Cycles Limit number of Cycles for food sources

1 200 400 10
2 300 600 15
3 400 800 20
4 500 1000 25

Table 2
Orthogonal array (OA) for Taguchi design of experiments for the proposed algorithm.

Experiment Level of Parameters

Population Size or Number of food sources Maximum number of algorithm Cycles Limit number of Cycles for food sources

1 1 1 1
2 1 1 2
3 1 2 3
4 1 2 4
5 2 3 1
6 2 3 2
7 2 4 3
8 2 4 4
9 3 1 1
10 3 1 2
11 3 2 3
12 3 2 4
13 4 3 1
14 4 3 2
15 4 4 3
16 4 4 4

5. Experiments and results
5.1. Tuning of algorithm parameters
Experiments are performed to investigate the performance of
the proposed GABC algorithm in this section. The performance of
an algorithm can be observed from the quality of results and vari-
ations in the results it can produce for various problems. Taguchi
method (Taguchi et al., 2005) is signicant to improve the quality of
the proposed algorithm by minimizing the effects of causes which
can produce variation in the results without eliminating those
causes. Set of experiments is used which contains different levels of
the algorithm parameters. An orthogonal array (OA) of different
level of parameter values is used in Taguchi method to determine
the signal to noise ratio (S/N). S/N ratio is the ratio of the value of
objective function with the variations in its values. In OA the
number of columns represents the parameters and the number of
rows describes the number of experiments. Taguchi experiment is
designed to determine the best level of parameters at which the S/N
is maximum i.e., quality of optimizing objective is better with less
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 729

variance. value of objective at level 1 for parameter population size. Simi-

In the Taguchi design of experiments, the algorithm parameters
which can inuence the performance of algorithm are determined
and four levels of each parameter against certain values of the
parameter are assigned. In the proposed GABC algorithm, the
population size of food sources, number of cycles and limit cycles
are considered as algorithm parameters.
Table 1 illustrates the effective algorithm parameters with their
different levels for the proposed GABC algorithm. It can be seen
from Table 1 that each column of the table indicates different values
of the parameters and their corresponding levels for each value. For
example, level 1 represents that the number of food sources in the
experiment is 200, maximum number of cycles of the algorithm to
run is 400 and the limit cycles is 10 after which an employee bee
can become a scout bee if there is no improvement in its nectar
value.
In the current study, 4 levels for each parameter are considered
and for each level, 4 numbers of experiments are performed. The
current set of experiments is used to test 16 experiments contain-
ing 4 levels and 3 different parameters. Table 2 indicates orthogonal
array (OA) for the Taguchi design of experiments for the proposed
algorithm. In Table 2, each row indicates the level of factors which
are considered in each experiment and each column represent a
specic level for that parameter and it has different value in each
experiment.
The proposed GABC is tested on two problems categorized on
different sizes on the basis of the number of facilities and demand
of new and quantity of returned products in these problems. This
categorization of problems into different sizes might be signicant
to accurately measure noise (or variation in the quality of results) in
each category of problems.
In the current experiment, each problem is tested according to
different level of parameter as mentioned in the proposed OA and
the corresponding values of the objective function is computed.
Once each problem is tested according to OA set of parameters, the
mean value of the objective, for each level of each parameter is
computed for each problem. For example, the mean value of
objective for the parameter population size is obtained from rst
four experiments of the OA matrix because they can give mean
100/50/25/20 200/100/50/40 130/65/32/30 150/75/37/30
15/20 10/5 12/40 18/10
A B C D
225/112/56/60
E 5/2 175/87/43/40
larly, mean value for parameter population size for level 2 is ob-
tained by taking the average of objective values obtained from the
next four experiments. Similar procedure is adopted to get mean of
each parameter for each level. Later mean of mean objective values
(called mean of means) for each level of each category of problems
is computed. Furthermore, the corresponding value of signal to
noise ratio (S/N) for each objective is also computed using the
relation shown in Equation (29)
!
mean2
S=NNominal 10 log (29)
Variance2
where, mean2 indicates the mean value of the optimizing objec-
tive and Variance2 is the variance value in the optimizing objec-
tive. S/N values of different problems for each objective is
computed according to the OA and then mean value of S/N of each
objective for each level of parameter is computed. Later, mean value
of S/N values (called mean of S/N) for each level of each category of
problem is computed. The data for small and large problem is
presented here. The data of the small category of facility location
problem is obtained from the facility location problem proposed by
Daskin (1995) with some modications in their input data to use in
the current problem. The problem description of the small and
large size problem are indicated in this section.
5.1.1. Small problem
For small size problem, 12 nodes are considered as possible
location sites for the customer centers, intermediate centers, pro-
ducers and remanufacturers for the considered problem. The small
problem considered in the current research is used from the
example of network problem proposed by Daskin (1995). The same
problem data has been used by Zhiqiang and Nathalie (2007) for
analyzing network design problem in literature. However, they only
considered the location points for producers, remanufacturers and
intermediate centers while customer center locations are not
considered in their data. Therefore, in the current research, the
same example problem and its corresponding data is used and the
data which is missing in the original problem given by Daskin
(1995), is assumed in the current research problem. The data they
used for reverse ows at each facility site including the quantity of
returned items, demand of products and xed costs of facilities is
used here as such as indicated in Fig. 7. The same problem is
considered here with some additional information including the
xed cost of the customer center which is assumed and presented
in Fig. 7 for the current problem. The units of cost are not given in
the problem analyzed by Zhiqiang and Nathalie (2007) and in the
current research CNY (Chinese Yuan) are assumed as the unit of all

F 24/15 Table 3
Unit transportation cost of products and returned products in CNY (Chinese Yuan)
between different facilities.
G H I
Location Sites A B C D E F G H I J K L
190/95/47/30 210/105/52/40 165/82/41/35
A 0 5 15 20 15 30 15 18 27 30 32 37
11/5 16/8 13/7 B 5 0 5 15 10 25 13 15 16 17 18 20
C 15 5 0 10 5 15 20 15 7 18 15 10
D 20 15 10 0 7 5 22 17 10 17 13 8
E 15 10 5 7 0 8 10 5 10 12 14 16
J K L F 30 25 15 5 8 0 20 15 5 18 16 8
G 15 13 20 22 10 20 0 5 10 8 11 20
230/115/57/25 125/62/31/20 215/108/53/40 H 18 15 15 17 5 15 5 0 6 5 8 13
I 27 16 7 10 10 5 10 6 0 8 6 8
22/17 19/10 20/8 J 30 17 18 17 12 18 8 5 8 0 5 10
K 32 18 15 13 14 16 11 8 6 5 0 6
L 37 20 10 8 16 8 20 13 8 10 6 0
Fig. 7. Small example problem data for facility location problem.
730 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

Table 4 four different values of xed costs and they are underlined, i.e., 100,
Storage capacities of the intermediate centers, producers and remanufacturers. 50, 25 and 20. These costs represents the xed cost of producers,
Location Sites H I J K L remanufacturers, intermediate centers and customer centers
Capacity of intermediate centers 20 25 30 40 50
respectively if they are located on position A in the network.
Capacity of producers or remanufacturing centers 10 15 20 10 13 Moreover, it is assumed that all the costs included in the current

Table 5
Unit processing cost in different facilities in CNY (Chinese Yuan).

Unit Costs at Facilities Cost CNY (Chinese Yuan)

UCkP unit production cost at producers 25

UCkR unit remanufacturing cost at remanufacturers 10
UCjI unit reprocessing cost in intermediate centers 4
DCjI unit disposal cost in intermediate centers 0.5
DCkR unit remanufacturing disposal cost in remanufacturers 0.5
hk unit holding cost of returned and produced products 0.2
b percentage at which returned products can be disposed of at the intermediate centers 40%
g percentage at which checked returned products can be disposed of at the remanufacturing centers 10%

the costs included in the problem. The numbers underlined above example problem are in CNY (Chinese Yuan). Furthermore, in Fig. 7
the boxes in Fig. 7 are the xed cost respectively for the four the numbers shown in the boxes represents the estimated average
m
different type of facilities. For instance, the location position A has demand edi of new products and the estimated average quantity of

130/65/32/30 150/75/37/30
100/50/25/20 T
120/140 108/90
225/140 215/108/53/40

120/85
C D
M

100/50/25/20 200/100/50/40 130/65/32/30 150/75/37/30

150/120 180/250 225/112/56/60 175/87/43/40 180/140 330/210

E 145/120 F 315/255
A B P Q

G H I
L
225/112/56/60
190/95/47/30 210/105/52/40 165/82/41/35
215/108/53/40
R 95/46
110/205 110/210 200/300
200/120

J K
200/100/50/40

230/115/57/25 125/62/31/20 300/180

230/120 200/230
O
N
175/87/43/40
190/95/47/30
S 204/105
180/90

Fig. 8. Large size example problem data for facility location problem.
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 731

m
the returned products edri from the customer, if there is a customer
center in that location point. For example, the numbers mentioned
in the box which is made at location position A are 15 and 20 which
shows the estimated average demand of the new products and the
estimated average demand of the returned products if the customer
center is made at the location A in the network.
In the current small example problem there are total 12 sites
and it is assumed that there are total 7 customer centers, 2 inter-
mediate centers and 3 location points are required to assign to
locate the considered facilities. The unit transportation cost of the
new and returned products from one facility location to other fa-
cility location is described in Table 3 and the unit of cost is
considered in CNY (Chinese Yuan). The demand of new products
and returned products is uncertain in the current example problem
and it depends on several factors. It is assumed that there are ve
factorsbim with their values randomly taken in the range of [5, 9]
and the corresponding weightages vim for each factor is randomly
taken from the range of [1.5, 3]. The mean value of the demand of
new products and mean value of the quantity of the returned
m m P make total 12 customer centers, 4 intermediate centers and 4
products is calculated from di edi M m1 bim vim
m m P location points are required to assign to locate the producers or
anddri edri M b v
m1 im im respectively.
The variance in demand of new products and variance in the
quantity of the returned products is randomly taken from a uniform
 
m m m m
distribution of 18di ; 14di and 18dri ; 14dri respectively. 90% service
level is required for all facilities which contains inventory and the
corresponding safety stock of the facility is calculated
fromZ Saftystockmean
variance
. mand of new products and returned products is uncertain in the
The storage capacity of intermediate centers and remanufac-
turer centers are assumed as 1:5  Safetystock of the facility as
shown in Table 4 for each location point, if there is an intermediate
center or remanufacturer center on that location point respectively.
Table 5 indicates the other costs related with the product in
different facilities and values percentage of disposal of returned
products in intermediate centers and remanufacturing centers in
CNY (Chinese Yuan).
5.1.2. Large problem
The large problem is made from the supply chain of a small
remanufacturing Case company in China. There are 20 different
towns which are considered in their network and therefore, in the
considered problem 20 nodes are considered as possible location
points for the customer centers, intermediate centers, producers
and remanufacturers for the considered problem. The data used for
reverse ows at each facility site including the quantity of returned
items, demand of products and xed costs of facilities is used here
based on the Case company information and it is indicated in Fig. 8.
In Fig. 8, the numbers shown underlined and are written outside
the box represents the xed cost of facilities. For example, there are
four numbers, i.e., 130, 66, 32 and 30 written and underlined
outside of the location site C. These represents the xed cost of
producer, remanufacturer, intermediate center and customer cen-
ter respectively. Similarly, the numbers written in the box repre-
m
sents the estimated average demand edi of new products and the
m
estimated average quantity of the returned products edri from the
customer, if there is a customer center in that location point. For
example, there is average estimated demand of 120 products and
average estimated quantity of the returned products on a customer
center if it is located on the location site C in the network.
There are total 20 sites and the considered company needs to
remanufacturers. The unit transportation cost of the new and
returned products from one facility location to other facility loca-
tion is described in Table 6 and the costs are assumed in CNY
(Chinese Yuan). In the current problem, the costs included in the
network are assumed in the proportion of the costs of the case
company. The transportation cost includes the shipping cost of
parts which are moved using Trucks and Trailers by road. The de-
current example problem and it depends on several factors. It is
assumed that there are ve factors bim with their values randomly
taken in the range of [10, 18] and the corresponding weightages vim
for each factor is randomly taken from [1, 3]. The mean value of the
demand of new products and mean value of the quantity of the
m m P
returned products is calculated from di edi M m1 bim vim
rm rm P M
anddi edi m1 bim vim respectively.
The variance in demand of new products and variance in the
quantity of the returned products is randomly taken from a uniform
 
m m m m
distribution of 18di ; 14di and 18dri ; 14dri respectively. 90% service
level is required for all facilities which contains inventory and the
corresponding safety stock of the facility is calculated

Table 6
Unit transportation cost of products and returned products between different facilities in CNY (Chinese Yuan).

Location Sites A B C D E F G H I J K L M N O P Q R S T

A 0 10 20 25 20 35 10 14 20 20 30 35 40 38 45 45 50 50 53 58
B 10 0 15 17 13 25 12 10 16 25 27 30 30 26 35 40 45 42 48 54
C 20 15 0 10 10 15 22 20 18 30 35 28 24 38 32 32 40 36 45 38
D 25 17 10 0 12 13 30 28 20 32 30 20 10 39 36 24 35 36 40 37
E 20 13 10 12 0 12 15 12 10 28 23 16 15 27 26 27 30 29 36 35
F 35 25 15 13 12 0 27 24 15 34 20 10 10 30 27 20 28 20 30 25
G 10 12 22 30 15 27 0 10 15 10 18 28 35 20 35 53 58 52 58 62
H 14 10 20 28 12 24 10 0 13 12 15 20 30 15 24 32 36 38 40 45
I 20 16 18 20 10 15 15 13 0 15 10 12 20 20 18 27 30 26 30 35
J 20 25 30 32 28 34 10 12 15 0 13 27 50 10 29 54 60 54 58 65
K 30 27 35 30 23 20 18 15 10 13 0 14 35 10 12 30 36 30 25 40
L 35 30 28 20 16 10 28 20 12 27 14 0 20 25 10 20 25 15 25 30
M 40 30 24 10 15 10 35 30 20 50 35 20 0 48 40 14 20 25 37 16
N 38 26 38 39 27 30 20 15 20 10 10 25 48 0 16 40 45 35 28 58
O 45 35 32 36 26 27 35 24 18 29 12 10 40 16 0 30 35 15 10 40
P 45 40 32 24 27 20 53 32 27 54 30 20 14 40 30 0 10 10 25 15
Q 50 45 40 35 30 28 58 36 30 60 36 25 20 45 35 10 0 13 27 13
R 50 42 36 36 29 20 52 38 26 54 30 15 25 35 15 10 13 0 17 20
S 53 48 45 40 36 30 58 40 30 58 25 25 37 28 10 25 27 17 0 40
T 58 54 38 37 35 25 62 45 35 65 40 30 16 58 40 15 13 20 40 0
732 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

Table 7
Storage capacities of the intermediate centers, producers and remanufacturers.

Location Sites M N O P Q R S T

Capacity of intermediate centers 20 25 30 40 50 35 45 23

Capacity of producers or remanufacturing centers 10 15 20 10 13 33 51 24

11800
40
11600 38
Mean of Means

36
11400 34

Mean of S/N
11200 32
30
11000 28
10800 26
24
10600 22
1 2 3 4 20
Level of Parameters 1 2 3 4
Level of Parameters
population size Cycles Limit Cycles

Fig. 9. Mean of means for small size problem. Population size Cycles Limit Cycles

Fig. 11. Mean of S/N for small size problem.

50000
Mean of Means

49500
41
49000 39
37
Mean of S/N

48500 35
33
31
48000
29
27
47500 25
1 2 3 4 1 2 3 4
Level of Parameters Level of Parameters
Population size Cycles Limit Cycles Population size

Fig. 12. Mean of S/N for large size problem.

Fig. 10. Mean of means for large size problem.

Table 8
fromZ Saftystockmean
variance
. Optimum level of parameters used for each category of problem.

The storage capacity of intermediate centers and remanufac- Problem Category Optimal level of parameters
turer centers are assumed as 3  Safetystock of the facility as shown Population size Cycles Limit Cycles
in Table 7 for each location point, if there is an intermediate center
Small Size Problem 2 4 4
or remanufacturer center on that location point respectively. The Large Size Problem 3 4 4
other costs related with the product in different facilities and values
percentage of disposal of returned products in intermediate centers
and remanufacturing centers is same as the costs used for large size
level, as indicated in Fig. 10. The mean of S/N values of objective for
problem are same as used for small size problems and presented in
small problem and large problem are indicated in Figs. 11 and 12
Table 5.
respectively. The optimum level of parameters obtained after
Graphical method is used to identify the optimal level of algo-
observing the mean of means and mean of S/N values for each
rithm parameters for the small and large size problem. The opti-
category of problem is illustrated in Table 8.
mum level of parameters for small problem is obtained by
observing mean value of means at each parameter level, as indi-
cated in Fig. 9. The optimum level of parameters for large problem 5.2. Results
is obtained by observing mean value of means at each parameter
In this section, small and large sizes of network optimization
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 733

Close Loop Supply chain Network obtained from GABC

Customer Centers The cost objective and computational time taken by each algorithm
is indicated in Table 9.
B C
Intermediate Centers
D

A
Remanufacturers 5.2.1. Results for small problem
It can be seen from Table 9 that proposed GABC algorithm gives
E
Cost 10900
a close loop supply chain network which have total cost of 10900
F

Time 1870
CNY (Chinese Yuan) which is less as compared to the close loop
I
supply chain networks obtained from ABC or GA. This indicates that
H
the proposed GABC algorithm gives better results in terms of
G
quality. These results indicate that the close loop supply chain
network which is obtained by GABC require less amount of the total
J K L cost. Moreover, the proposed GABC gives the optimal network by
considering the overall effect of all costs including the xed cost of
Fig. 13. Closed loop supply chain network for small size problem obtained from GABC. facilities, shipping cost, holding cost and production costs etc. For
instance, the cost of network is 10900 CNY (Chinese Yuan) which is
obtained from the GABC algorithm, while the close loop network
problems are analyzed with the proposed GABC, ABC and GA using formed by ABC and GA algorithms have total costs of 11855 CNY and
the optimal parameter values obtained in the previous section. The 11267 CNY. These results indicate that the close loop network ob-
network solution of the small size problem obtained from GABC, tained from GABC is economical as compared to the networks
ABC and GA algorithm is indicated in Figs. 13, 14 and 15 respectively. formed by ABC and GA algorithm. However, the computational

Close Loop Supply chain Network obtained from ABC Customer Centers

B C D Intermediate Centers
A

Remanufacturers

F
Cost 11855
Time 1092

H E
G I

J L
K

Fig. 14. Closed loop supply chain network for small size problem obtained from ABC.

Close Loop Supply chain Network obtained from GA Customer Centers

A B C D Intermediate Centers

Remanufacturers

Producer
E
F

I
G H Cost 11267
Time 1817

J K L

Fig. 15. Closed loop supply chain network for small size problem obtained from GA.
734 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

Table 9 difference is signicant. These results indicate that the proposed

Cost of the network solutions and time taken by algorithms for solution for small GABC is economical to give good quality results as compared to ABC
problem in CNY (Chinese Yuan).
and GA.
GABC ABC GA

Cost (CNY Yuan) 10900 11855 11267 6. Result implementation

Time (Sec) 1870 1092 1817

In this section, the proposed GABC is applied to the closed loop

time is a little more as compared to ABC and GA. The cost of the
network has a big difference from the cost of the networks obtained
from ABC and GA.
5.2.2. Results for large problem
The closed loop supply chain network solution for the large
problem obtained from proposed GABC, ABC and GA algorithms is
indicated in Figs. 16, 17 and 18 respectively. Moreover, the results of
cost and computational time taken by each algorithm for large
problem is indicated in Table 10.
It can be seen from Table 10 that the close loop network ob-
tained from proposed GABC is economical as compared to the close
loop network obtained from ABC and GA. The network obtained
from GABC is needs 48469 CNY to make an optimal close loop
network based on the considered input data of the problem. The
cost of network obtained from GABC is less as compared to cost
needed to build the close loop supply chain network which are
obtained from ABC and GA. These results indicate that in large
problem size, the network obtained from GABC is more economical
to make as compared to the close loop supply chain network ob-
tained from ABC and GA. However, the computational time taken
by proposed GABC is little more as compared to ABC and GA. The
difference in computational time is not signicant but the cost
supply chain problems considering uncertainty in the demand and
quantity of returned products under different scenarios. For small
problem the range of [5, 9] for ve factors bim with their corre-
sponding weightages vim for each factor is randomly taken from
[1.5, 3]. Similarly, for large problems the value of ve factors bim are
randomly taken in the range of [10, 18] and the corresponding
weightages vim for each factor is randomly taken from [1, 3]. The
problems tested as based on one particular instant of the demand
and returned products uncertainty. However, demand and quantity
of returned products is considered randomly in the current
research, therefore to increase the reliability of the network solu-
tions for small and large size problem, scenario planning is used.
Different scenarios of demand and returned products are made and
the considered test problems are solved for each scenario to get a
robust solution. The network solution which can give minimum
cost of the network and shows less variations in the objective
function values against all scenarios is considered as robust solu-
tion in the current research. Table 11 illustrates the experiments in
which different scenarios for demand and returned product un-
certainty are prepared for each test problem. The range of demand
and returned products is considered here into three different cases
of values of the variations, i.e., small, medium and large values of
variations. Moreover, in each size of variations, there are three
different values of the range of variations, i.e., small range of

Close Loop Supply chain Network obtained from GABC

Customer Centers
Cost 48469
Time 1857
Intermediate Centers

Remanufacturers

C D M T
Producer

A B

H F P
G

I L R

J K

N O S

Fig. 16. Closed loop supply chain network for large size problem obtained from GABC.
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 735

Close Loop Supply chain Network obtained from ABC

Customer Centers Cost 50363

Time 1453
Intermediate Centers
M

Remanufacturers D T
C

Producer

A B Q

F
E

P
I
L
G H

J O
N S

Fig. 17. Closed loop supply chain network for large size problem obtained from ABC.

variations, medium range of variations and large range of varia-
tions. Furthermore, for each category of problem, ten different
scenarios of each small and large size problem are tested.
The results of the objective function values and CPU time for
each category of problems for small problem among ten scenarios
is obtained and illustrated in box plots as indicated in Figs. 19, 20
and 21 respectively. Furthermore, the results of objective function
values and CPU time for each category of problems for large
problem among ten scenarios is obtained and illustrated in box plot
in Figs. 22, 23 and 24 respectively.
6.1. Small problem with small demand
The small size problem with small size of demand with different
ranges of demand variations is considered here and the total cost of
the network against different demand scenarios is indicated in
Fig. 19. It can be seen from Fig. 19 that for small problem, when the
variations in demand increases from small to medium and from
medium to large variation, the objective function value i.e., total
cost of the network, increases. Moreover, the objective value ob-
tained from GABC is smaller as compared to the objective function
values obtained from ABC and GA when there is small and medium
variations in demand. However, for large variations in demand, the
objective function obtained from GA is less as compared to GABC
and ABC but their difference is very small. Moreover, the CPU time
taken by GA is less than both GABC and ABC for small range, me-
dium range and large range of variations in demand but the dif-
ference in CPU time is not signicant. For small range of variations
in demand, GABC performs better as compared to both ABC and GA
in terms of solution quality and computation i.e., the cost of
network which is obtained from GABC is less than the cost obtained
from ABC and GA. The total cost of network includes the xed costs
of facilities, the holding costs, shipping costs and production costs.
Therefore, the network obtained from GABC is better in quality as
compared to the network obtained from ABC and GA. Moreover, the
CPU time difference is also small and not much signicant. The
difference between the objective function and computation time
increases as the variations in the demand increases from small to
medium and from medium to large. Furthermore, for small and
medium variations in demand, the objective function value of the
small problem is not much variable and variations in the objective
function values is small as compared to the ABC and GA algorithm.
These small variations can show that the optimal network is ob-
tained from GABC remains better as compared to the network ob-
tained from ABC and GA even there are changes in the demand
occurs in the network. Therefore, it can show that the network
obtained from GABC algorithm can cope with the demand varia-
tions better as compared to the network obtained from ABC and GA
and this shows the better robust performance of the network ob-
tained form GABC as compared to ABC and GA. However, for large
variations in demand, there is more variations in the objective
function obtained from GABC as compared to ABC and GA. These
results indicate that when there is small variation and medium size
variation in demand, the proposed GABC algorithm can give the
closed loop supply chain network with consistence value of the
total cost. These results indicate that the proposed GABC algorithm
gives more robust solution results as compared to ABC and GA,
when there are small and medium size variations in demand.
However, when the variations in demand is large, the proposed
GABC algorithms is not signicantly robust.
736 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

Close Loop Supply chain Network obtained from GA

Customer Centers Cost 50725

Time 1663
Intermediate Centers

Remanufacturers
C D M T
Producer

A B
F
P
Q
E

G L
H

R
N
K
S
J O

Fig. 18. Closed loop supply chain network for large size problem obtained from GA.

Table 10 problem, when the variation in demand increases from small to

Cost of the network solution and time taken by algorithms for solution for large medium and from medium to large variation, the objective function
problem CNY (Chinese Yuan).
value is increases, this indicates that the cost of the network in-
GABC ABC GA creases as the variation in demand increases. Moreover, the
Cost (CNY Yuan) 48469 50363 50725 objective value obtained from GABC is smaller as compared to the
Time (Sec) 1857 1453 1663 values obtained from ABC and GA when there are small variation,
medium variation and large variation in demand. The network
obtained from GABC gives smaller value of the total cost of the
6.2. Small problem with medium demand network as compared to ABC and GA for different range of varia-
tions in demand. These results indicate that the GABC gives optimal
The small size problem with medium size of demand with network with smaller value of total cost. Moreover, the network
different ranges of demand variations is considered here and the obtained from GABC is better even the demand variations consid-
total cost of the network against different demand scenarios is ered are in small, medium or in large range. These results indicate
indicated in Fig. 20. It can be seen from Fig. 20 that for small that the network obtained from GABC is robust to give consistent

Table 11
Scenario of different problems containing different value and range of uncertain demand and returned products.

Problem Category Number of Test Scenarios Small Problem Large Problem

Range of variables Range of variables

vim bim vim bim

Small demand variation with small range 10 [5, 10] [1.5, 3] [10, 12] [1, 3]
Small demand variation with medium range 10 [5, 20] [1.5, 6] [10, 24] [1, 6]
Small demand variation with large range 10 [5, 30] [1.5, 9] [10, 36] [1, 9]
Medium demand variation with small range 10 [15, 10] [2, 3] [20, 12] [1.5, 3]
Medium demand variation with medium range 10 [15, 20] [2, 6] [20, 24] [1.5, 6]
Medium demand variation with large range 10 [15, 30] [2, 9] [20, 36] [1.5, 9]
Large demand variation with small range 10 [25, 10] [2.5, 3] [30, 12] [2, 3]
Large demand variation with medium range 10 [25, 20] [2.5, 6] [30, 24] [2, 6]
Large demand variation with large range 10 [25, 30] [2.5, 9] [30, 36] [2, 9]
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 737

4
Small problem& Small demand
x 10
28
GABC GABC GA
GABC GA
ABC ABC
ABC
4 GA 26

3.5
24
Objective Value

3
22
ABC

CPU Time
2.5 GABC
GA 20

2
18

1.5 GABC ABC GA

16

1
14
ABC GA
0.5
GABC
12

Small Medium Large Small Medium Large

variation variation variation variation variation variation

Fig. 19. Total cost of the network against different demand scenarios for small problem with small demand.

Small problem & Medium demand

4
x 10

10 ABC GABC GABC

GA 30
GABC
ABC GA ABC
9 GA
28

8
26
Objective Value

GA
ABC
7 GABC
CPU Time

24

6
22

GABC ABC
5 GA
20

4
18

3 16

GABC ABC
2 GA
14

1 12

Small Medium Large Small Medium Large

variation variation variation variation variation variation

Fig. 20. Total cost of the network against different demand scenarios of small problem with medium demand.

performance based on the total cost of the network as compared to comparable with the CPU time taken by ABC and GA. For small
ABC and GA. Furthermore, the CPU time taken by GABC for small range of variations in demand, medium range and large range of
range of demand, medium range and large range of demand, is demand variations GABC performs better as compared to both ABC
738 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742

4
Small problem& Large demand
x 10

6 ABC 32

30
5.5

GABC GA
28
5 ABC
GABC
GA
26
Objective Value

4.5
24

CPU Time
4
22

GABC ABC GA
3.5
20

3 18

16
2.5

14
2 GABC ABC GA GABC
ABC GA GABC
ABC GA
12

1.5

Small Medium Large Small Medium Large

variation variation variation variation variation variation

Fig. 21. Total cost of the network against different demand scenarios of small problem with large demand.

4
Large problem &Small demand
x 10

ABC GABC
GABC
ABC GA
12
28

GA

26
10

ABC
GA
Objective Value

24
GABC
8
CPU Time

22

GA
6 GABC ABC
20

18
4

16

14
GABC ABC GABC
GA ABC
GA
0
12

Small Medium Large Small Medium Large

variation variation variation variation variation variation
Fig. 22. Total cost of the network against different demand scenarios of large problem with small demand.

and GA in terms of solution quality, i.e., the total cost of the network the better stability of the results obtained from GABC. Furthermore,
obtained from GABC is smaller as compared to the network ob- it can be seen from Fig. 20 that the proposed GABC algorithm gives
tained from ABC and GA. Furthermore, there is less variations in the less variations in the objective function value when there is change
CPU time and objective values obtained from GABC which indicates in the range of variations in demand. These less variations in the
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742 739

4 Large problem & Medium demand

x 10
ABC GA
7 12.1

6.5 12
GABC

6 11.9

GABC
5.5 GABC GABC
Objective Value

11.8

5 ABC

CPU Time
11.7

4.5 GABC ABC ABC

ABC 11.6
GA
4
GA
11.5
3.5 GA
GA
11.4
3

11.3
2.5 GABC ABC GA

11.2
2

Small Medium Large Small Medium Large

variation variation variation variation variation variation

Fig. 23. Total cost of the network against different demand scenarios of large problem with medium demand.

4
x 10
Large problem & Large demand
10 ABC
GABC
13.8
GA
9

13.6

8
13.4

7
Objective Value

13.2
GABC ABC GA
6 GABC
GABC
CPU Time

13
GABC
5 ABC ABC
12.8
GA
4 ABC
12.6
GA GA

3
12.4

2 12.2
GABC ABC GA

1 12

Small Medium Large Small Medium Large

variation variation variation variation variation variation
Fig. 24. Total cost of the network against different demand scenarios of large problem with large demand.

objective function, i.e., total cost of network, indicate that the consistently gives the optimal network even there are changes in
network obtained from GABC gives consistent results and the demand. Moreover, the network obtained from GABC can cope
740 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
with the uncertain variations in demand with total network cost as
minimum as compared to network obtained from ABC and GA.
These results indicate that the proposed GABC algorithm gives
more robust solutions as compared to both ABC and GA and gives
less variations in the total cost of the network.
6.3. Small problem with large demand
The small size problem with large size of demand with different
ranges of demand variations is considered here and the total cost of
the network against different demand scenarios of the problem are
indicated in Fig. 21. It can be seen from Fig. 21 that for small
problem, when the variation in demand increases from small to
medium and from medium to large variation, the objective function
value is increases, this indicates that the cost of the network in-
creases as the variation in demand increases. Moreover, the
objective value obtained from GABC is smaller as compared to the
values obtained from ABC and GA when there are small variations
in demand. Whereas, the objective function value obtained from
GABC is comparable with ABC and GA for medium variation and
large variation in demand. The total cost of network obtained from
GABC is smaller as compared to the network obtained from ABC
and GA when there are small variations in the demand, which in-
dicates that GABC gives better results in terms of solution quality as
compared to ABC and GA. Furthermore, the CPU time taken by
GABC for small range of demand, medium range and large range of
demand, is comparable with the CPU time taken by ABC and GA and
there is no much difference in CPU time for GABC, ABC and GA for
small, medium and large variations in demand problem scenarios.
It can be seen from Fig. 21 that there is less variations in the
objective function value obtained from GABC algorithm when there
is small and medium size variations in the demand as compared to
ABC and GA. These results indicate that the proposed GABC algo-
rithm gives the network which has smaller value of the total
network cost as compared to ABC and GA even there are changes in
the demand. This shows that the network obtained from GABC
gives more robust solutions as compared to both ABC and the total
cost of the network is not changing even there is variations in the
demand in small or medium range. However, as the size of varia-
tions increases to large size variations, the objective function values
obtained from GABC, ABC and GA shows large variations. These
results indicate that the total cost of the network obtained from
GABC, ABC and GA varies in large range as the demand varies in
large range. The robustness of the results obtained from GABC is
better than ABC and GA in terms of the total cost of network when
there is small size and medium size variations in the demand.
6.4. Large problem with small demand
The large size problem with small size of demand with different
ranges of demand variations is considered and the total cost of the
network against different scenarios of demand are indicated in
Fig. 22. It can be seen from Fig. 22 that for large problem, when the
variations in demand increases from small to medium and from
medium to large variation, the objective function value is increases,
this indicates that the cost of the network increases as the variation
in demand increases. Moreover, the objective value, i.e., total cost of
the network, obtained from GABC is smaller as compared to the
values obtained from ABC and GA for problems with small variation
and medium range of variations. However, there is small difference
in the objective value obtained from GABC, ABC and GA for the
scenario with large variations in demand. Furthermore, there is
small variations in the objective function obtained from GABC al-
gorithm when there is small and medium size variations in demand
compared to the objective function values obtained from ABC and
GA. These results indicate the robustness of the results obtained
from GABC is better as compared to ABC and GA. When the varia-
tions in demand increases to large, the variations in the objective
function values obtained from GABC, ABC and GA also increases.
These results indicate that for large variations in demand, the so-
lutions given by GABC, ABC and GA are not signicantly robust but
compared to ABC and GA, GABC is relatively gives less variations in
objective function when there is large variations in demand.
Furthermore, the CPU time taken by GABC for small range of de-
mand, medium range and large range of demand, is comparable
with the CPU time taken by ABC and GA and there is no much
difference in CPU time for GABC, ABC and GA for small, medium and
large variations in demand problem scenarios.
6.5. Large problem with medium demand
The large size problem with medium size of demand with
different ranges of demand variations is considered and the total
cost of the network against different demand scenarios of the
problem are indicated in Fig. 23. It can be seen from Fig. 23 that for
large problem, when the variation in demand increases from small
to medium and from medium to large variation, the objective
function value is increases, this indicates that the cost of the
network increases as the variation in demand increases. Moreover,
the objective value obtained from GABC is smaller as compared to
the values obtained from ABC and GA for problems with small
variation. However, there is small difference in the objective value
obtained from GABC, ABC and GA for the scenario with medium and
large variations in demand. Moreover, the objective function values
obtained from GABC gives less variations as compared to the vari-
ations in total cost of network obtained from ABC and GA when the
variations in demand increases from small, medium and large
variations. These results indicate better robustness of the solutions
in terms of the total cost of network obtained from GABC as
compared to ABC and GA. Furthermore, the CPU time taken by GA
for small range of demand, medium range and large range of de-
mand, is smaller but comparable with the CPU time taken by GABC
and ABC and there is no much difference in CPU time for GABC, ABC
and GA for small, medium and large variations in demand problem
scenarios.
6.6. Large problem with large demand
The large size problem with large size of demand with different
ranges of demand variations is considered and the total cost of the
network against different demand scenarios of the problem are
indicated in Fig. 24. It can be seen from Fig. 24 the objective value,
i.e., total cost of the network, obtained from GABC is smaller as
compared to the values obtained from ABC and GA for problems
with the problem scenarios with small variation. The objective
function value obtained from GABC is smaller and also there are less
variations in the objective function values for the solutions given by
GABC as compared to the objective function values obtained from
ABC and GA for small, medium and large size variations in demand.
These results indicate that the results obtained from GABC gives
more optimal closed loop supply chain network and the total
network cost is robust, ie., total network cost is not changing in
large range when there is uncertain change in demand. Further-
more, the CPU time taken by GA for small range of demand, me-
dium range and large range of demand, is smaller but comparable
with the CPU time taken by GABC and ABC and there is no much
difference in CPU time for GABC, ABC and GA for small, medium and
large variations in demand problem scenarios.
 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
7. Conclusion
In recent years, the manufacturing companies are trying to
reuse and remanufacture their products due to environmental
concerns to reduce pollution due to end of life products. Current
research proposed a facility location optimization problem of
closed loop supply chain. The proposed closed loop supply chain
network is composed of customer centers, intermediate centers,
producers and remanufacturers. In the current research, the de-
mand of new products and returned products from different cus-
tomers is considered uncertain and assumed to vary randomly. A
mixed integer programing model is proposed with an aim to
minimize total cost of the network including xed cost of facilities,
shipping cost (i.e., the collection cost) of the products between
different facilities, holding cost (i.e., storage cost) of inventory of
new and returned products in different facilities and production
cost (i.e., manufacturing and remanufacturing cost). Moreover,
safety stock levels of new and returned products and the service
level and some additional constraints are considered in the pro-
posed model.
A new genetic articial bee colony (GABC) algorithm is pre-
sented for the current research problem and a new food source
representation for the current problem is introduced. The proposed
GABC algorithm considered neighbor food source creation for the
local search and used crossover and mutation concepts of genetic
algorithm to improve the exploration ability of the proposed GABC
algorithm. Two different size of test problems i.e., small and large,
are used to test the performance of GABC algorithm. Small problem
is taken from literature which is modied according to the current
problem while the second problem is made from a Case company
data in China. The algorithm parameters are tuned using Taguchi
method and the considered problems are solved using proposed
GABC algorithm. The performance of proposed GABC algorithm is
compared with the performance of standard, ABC and GA by
comparing objective function values (i.e., total cost of the network)
obtained for each test problem and the computation time taken by
each algorithm. Moreover, to increase the application ability of the
considered problem and proposed GABC algorithm, the considered
two problems are solved on different demand values, i.e., small
demand, medium demand and large demand. Furthermore, for
each demand value, different range of variations in demand,
including small range variations, medium range variations and
large range of variations are considered. This can consider the real
situation of the remanufacturing of the products which have small,
medium or large demand with different levels of variations. For
each problem category, 10 different scenarios are tested and the
corresponding results are analyzed for small and large size prob-
lem. Results indicate that the proposed GABC gives better results in
terms of both quality and computation for different problems in
different demand range and different size of demand variations. For
example, the total cost of network obtained from GABC is less as
compared to the cost of network obtained from ABC and GA.
Moreover, the difference in CPU time taken by GABC is not much
higher than the CPU time taken by ABC and GA. Moreover, less
variations in the objective values, i.e., total cost, for different sce-
narios are obtained from GABC algorithm as compared to ABC and
GA. The total cost of the network obtained from GABC gives less
variations as compared to the variations in the total cost observed
from the network using ABC and GA. This shows better robustness
of the results obtained from GABC algorithm based on the varia-
tions in the total cost of the network. Current research is signicant
to apply to get the optimized closed loop supply chain network of
some product which have certain value of demand and it variations
in certain range. The implementation of the proposed GABC on
different demand scenarios of the small and large problem indicate
741
the robustness of the proposed GABC and its signicance to
implement to determine optimal closed loop supply chain network
for the products which have different range of demand variations.
The proposed GABC can give robust results of the optimized
network which can be implemented under different demand
variations.
In the current research, the transportation capacity is not
considered which can also inuence the total cost of network. In
future, the closed loop supply chain model integrating the trans-
portation times with transportation capacity in facilities can be
incorporated and to deal with uncertainty in demand of new
products and uncertainty in quantity and quality of the returned
products using robust optimization methods. Furthermore, the
reliability of the closed loop supply chain network in terms of better
service level and the total cost of network can be simultaneously
considered to optimize using Pareto optimization concepts in
future.
Acknowledgements
This work has been supported by MOST (Ministry of Science &
Technology of China) the Funds for International Cooperation and
Exchange of the National Natural Science Foundation of China
(No.51561125002), the National Natural Science Foundation of
China (No. 51275190, 51575211, 71131004, 71471071), and the
Fundamental Research Funds for the Central Universities (HUST:
2014TS038).
References
Alshamrani, Ahmad, Mathur, Kamlesh, Ballou, Ronald H., 2007. Reverse logistics:
simultaneous design of delivery routes and returns strategies. Comput. Oper.
Res. 34 (2), 595e619.
Saman, Hassanzadeh Amin, Zhang, Guoqing, 2013. A multi-objective facility loca-
tion model for closed-loop supply chain network under uncertain demand and
return. Appl. Math. Model. 37, 4165e4176.
Aras, N., Aksen, D., Tanugur, A.G., 2007. Locating collection centers for incentive
dependent returns under a pick-up policy with capacitated vehicles. Eur. J.
Oper. Res. 191 (3), 1223e1240.
Aydin, R., Kwong, C.K., Ji, P., 2016. Coordination of the closed-loop supply chain for
product line design with consideration of remanufactured products. J. Clean.
Prod. 114, 286e298. http://dx.doi.org/10.1016/j.jclepro.2015.05.116.
Chen, H.K., Chou, H.W., Chiu, Y.C., 2007. On the modeling and solution algorithm for
the reverse logistics recycling ow equilibrium problem. Transp. Res. Part C 15,
218e234.
Daskin, M.S., 1995. Network and Discrete Location, Models, Algorithms, and
Application. Wiley: Wiley-Inter science Publication.
Devika, K., Jafarian, A., Nourbakhsh, V., 2014. Designing a sustainable closed-loop
supply chain network based on triple bottom line approach: a comparison of
metaheuristics hybridization techniques. Eur. J. Oper. Res. 235 (3), 594e615.
Du, Feng, Evans, Gerald W., 2008. A bi-objective reverse logistics network analysis
for post-sale service. Comput. Oper. Res. 35 (8), 2617e2634.
Fakhrzad, M.B., Moobed, M., 2010. A GA model development for decision making
under reverse logistics. Int. J. Indus. Eng. Prod. Res. 21, 211e220.
Fatih Tasgetiren, M., Suganthan, P.N., Angela, H-L Chen, 2011. A discrete artificial bee
colony algorithm for the total owtime minimization in permutation ow
shops. Inf. Sci. 181, 3459e3475.
Fleischmann, M., Bloemhof-Ruwaard, J.M., Dekker, R., Van Der Laan, E., Van
Nunen, J.A.E.E., Van Wassenhove, L.N., 1997. Invited review, quantitative models
for reverse logistics: a review. Eur. J. Oper. Res. 103, 1e17.
Fleischmann, M., Beullens, P., Bloemhof-ruwaard, J.M., Wassenhove, L., 2001. The
impact of product recovery on logistics network design. Prod. Oper. Manag. 10,
156e173.
Francas, D., Minner, S., 2009. Manufacturing network configuration in supply chains
with product recovery. Omega 37 (4), 757e769.
Glpinar, N., Pachamanova, D., anakoglu, E., 2003. Robust strategies for facility
location under uncertainty. Eur. J. Oper. Res. 225 (1), 21e35.
Gupta, Aman, Evans, Gerald W., 2009. A goal programming model for the operation
of closed-loop supply chains. Eng. Optim. 41 (8), 713e735.
Karaboga, D., 2005. An Idea Based on Honeybee Swarm for Numerical Optimization.
Technical Report TR06. Erciyes University Engineering Faculty. Computer En-
gineering Department.
Krikke, H.R., van Harten, A., Schuur, P.C., 1999. Business case Oce : reverse logistic
network re-design for copiers. OR Spectr. 21 (3), 381e409.
Lee, C.K.M., Chan, T.M., 2009. Development of RFID-based reverse logistics system.
742 Y.Y. Cui et al. / Journal of Cleaner Production 162 (2017) 717e742
Expert Syst. Appl. 36, 9299e9307. http://dx.doi.org/10.1016/j.eswa.2008.12.002.
Li, Jun-Qing, Pan, Quan-Ke, Gao, Kai-Zhou, 2011. Pareto-based discrete articial bee
colony algorithm for multi-objective exible job shop scheduling problems. Int.
J. Adv. Manuf. Technol. 55 (9e12), 1159e1169.
Lu, Z., Bostel, N., 2007. A facility location model for logistics systems including
reverse ows: the case of remanufacturing activities. Comput. Oper. Res. 34,
299e323.
Ozkir, Vildan, Basligil, Huseyin, 2013. Multi-objective optimization of closed-loop
supply chains in uncertain environment. J. Clean. Prod. 41, 114e125.
Pishvaee, Mir Saman, Kianfar, Kamran, Karimi, Behrooz, 2010. Reverse logistics
network design using simulated annealing. Int. J. Adv. Manuf. Technol. 47 (1e4),
269e281.
Pishvaee, Mir Saman, Rabbani, Masoud, Torabi, Seyed Ali, 2011. A robust optimi-
zation approach to closed-loop supply chain network design under uncertainty.
Appl. Math. Model. 35 (2), 637e649.
Salema, Maria Isabel Gomes, Barbosa-Povoa, Ana Paula, Novais, Augusto Q., 2007.
An optimization model for the design of a capacitated multi-product reverse
logistics network with uncertainty. Eur. J. Oper. Res. 179 (3), 1063e1077.
Salema, Maria Isabel Gomes, Po  voa, Ana Paula Barbosa, Novais, Augusto Q., 2009.
A strategic and tactical model for closed-loop supply chains. OR Spectr. 31 (3),
573e599.
Shi, Jianmai, Zhang, Guoqing, Sha, Jichang, Amin, Saman Hassanzadeh, 2010.
Coordinating production and recycling decisions with stochastic demand and
return. J. Syst. Sci. Syst. Eng. 19 (4), 385e407.
Suryabatmaz, A.C., Altekin, F.T., Sahin, G., 2014. Hybrid simulation-analytical
modeling approaches for the reverse logistics network design of a third-party
logistics provider. Comput. Indus. Eng. 70, 74e89.
Taguchi, G., Chowdhury, S., Wu, Y., 2005. Taguchi Quality Engineering Handbook.
John Wiley and Sons, New York.
Thierry, M.C., Salomon, M., Van Nunen, J.A.E.E., Van Wassenhove, L.N., 1993.
Strategic Production and Operations Management Issues in Product Recovery
Management. Management Report Series No. 145. Erasmus Universiteit/Rot-
terdam School of Management.
Tonanont, A., Yimsiri, S., Rogers, K.J., 2009. Reverse logistics optimization with data
envelopment analysis. In: Proceedings of 2009 Industrial Engineering Research
Conference, Miami, 30 May-3 June 2009, pp. 1268e1273.
Saif, Ullah, Guan, Zailin, Xianhao, Xu, Zongdong, He, Baoxi, Wang, Jahanzeb, Mirza,
2015. Modeling of facility location problem for reverse logistics system with
uncertainty in forward and reverse ows of products. Int. J. Mech. Eng. Robot.
Res. 4 (3), 242e246.
ster, Halit, Easwaran, Gopalakrishnan, Akali, Elif, etinkaya, Sila, 2007. Benders
decomposition with alternative multiple cuts for a multi-product closed-loop
supply chain network design model. Nav. Res. Logist. 54 (8), 890e907.
Vahdani, Behnam, Tavakkoli-Moghaddam, Reza, Modarres, Mohammad,
Baboli, Armand, 2012. Reliable design of a forward/reverse logistics network
under uncertainty: a robust-M/M/c queuing model. Transp. Res. Part E Logist.
Transp. Rev. 48 (6), 1152e1168.
Wollenweber, Jens, 2008. A multi-stage facility location problem with staircase
costs and splitting of commodities: model, heuristic approach and application.
OR Spectr. 30 (4), 655e673.
Yi, Pengxing, Huang, Min, Guo, Lijun, Shi, Tielin, 2016. A retailer oriented closed-
loop supply chain network design for end of life construction machinery
remanufacturing. J. Clean. Prod. 124, 191e203. http://dx.doi.org/10.1016/
j.jclepro.2016.02.070.
Zhiqiang, Lu, Nathalie, Bostel, 2007. A facility location model for logistics systems
including reverse ows: the case of remanufacturing activities. Comput. Oper.
Res. 34, 299e323.
Zohal, Mostafa, Soleimani, Hamed, 2016. Developing an ant colony approach for
green closed-loop supply chain network design: a case study in gold industry.
J. Clean. Prod. 133, 314e337.