Computers & Operations Research 65 (2016) 93–103

Contents lists available at ScienceDirect

Computers & Operations Research

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

A mixed integer nonlinear programming model and heuristic solutions

for location, inventory and pricing decisions in a closed loop
supply chain
Onur Kaya a,n, Busra Urek b
a
Department of Industrial Engineering, Anadolu University, End 201, Eskisehir, Turkey
b
Department of Industrial Engineering, Koc University, Sariyer, Istanbul, Turkey

art ic l e i nf o a b s t r a c t

Available online 17 July 2015 We analyze a network design problem for a closed-loop supply chain that integrates the collection of the
Keywords: used products with the distribution of the new products. We present a mixed integer nonlinear facility
Closed-loop supply chain location-inventory-pricing model to decide on the optimal locations of the facilities, inventory amounts,
Facility location prices for new products and incentive values for the collection of right amount of used products in order
Pricing to maximize the total supply chain profit. We develop heuristics for the solution of this model and
Inventory analyze the effectiveness of these heuristics and the effects of the parameters on this system through
Mixed integer nonlinear programming numerical experiments.
Heuristics & 2015 Elsevier Ltd. All rights reserved.

1. Introduction incentives, and growing awareness for environment. Thus, the

companies need to adjust their supply chain to integrate the
Closed loop supply chain management (CLSCM) includes all collection activities of end-of-life products into their traditional
reverse logistics activities as collection, remanufacturing and distribution channels and design an efficient CLSC.
refurbishing processes in addition to all traditional forward In this study, considering the inventory, pricing and incentive
logistics activities from supplier to customer. CLSCM does not only determination problems in the supply chain, we analyze the
gain increasing attention among society, government and practi- network design problem for a closed-loop supply chain that
tioners of operations but also gains the attentions of researchers integrates the collection of the used products with the distribution
(see [28] for a review on CLSCM). The growing concern for of the new products. In today's world, in which the companies are
environment leads integrating environmentally sound solutions required to collect a portion of their used products due to
into traditional supply chain management in practice. Besides the legislative issues or cost concerns, they need to integrate their
environmental concerns, recent legislations in many countries collection network for used products with the distribution net-
force the companies to undertake take-back responsibilities to work of new products. In this study, we present a mixed integer
reduce waste. The governmental concerns create strong legislative nonlinear facility location-inventory-pricing model to decide on
incentives on why and how take back activities are organized for both the optimal locations of the collection and distribution
different classes of products. WEEE directives in Europe will centers ðCDCÞ, optimal inventory amounts to be carried at these
require all large electronic retailers to take back their old equip- centers, optimal prices for new products and the values of
ment. Companies can also benefit from these used products to incentives that need to be offered for the collection of right
decrease their cost of production. End-of-life products that nor- amount of used products, in order to maximize the total supply
mally go to waste might actually have some remaining value chain profit.
which will reduce consumption of the raw materials and provide Logistics network design is one of the most strategic decisions
cost reduction on waste treatment. In other words, CLSCM is for CLSCM because opening or closing a facility is very expensive,
driven by high profitability, increasing number of legislative time consuming and infeasible to change in a short time. Many
tactical and operational decisions are limited by this strategic
decision because logistic network configuration is directly related
n
to facility locations. Facility location problems have been studied
Corresponding author: Tel.: +90 222 3350580 / 6439.
E-mail addresses: [email protected] (O. Kaya),
for more than a century, starting from Weber's location problem in
[email protected] (B. Urek). 1909 (see Brandeu and Chiu [5] for a review of facility location

http://dx.doi.org/10.1016/j.cor.2015.07.005
0305-0548/& 2015 Elsevier Ltd. All rights reserved.
94 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103

problems). The simplest and one of the most studied location demand and a deterministic rate of collection, we use the economic
problem is p-median problem which is introduced by Hakimi [12]. order quantity (EOQ) formulation which has been widely used in the
This problem analyzes which p facility are to be selected among network design literature. For example, Camm et al. [6] develop an
the potential locations and Mladenovic et al. [22] give a complete uncapacitated facility location formulation for Procter and Gamble
survey about metaheuristic approaches applied to p-median Company to locate the DCs and assign the selected DC's to customer
problem which is classified as NP-hard. Another type of the facility zones with the objective of minimizing the total cost, composed of
location models, which is more in line with our study, considers material handling costs, inventory costs and transportation costs,
fixed facility location costs, therefore number of facilities to be while maintaining a certain customer service level. Daskin et al. [8]
established becomes an important decision. This problem is propose a location inventory model and its solution methodology,
known as uncapacitated facility location problem (UFLP) and is considering the locations of DCs with the working inventory and
extensively studied in literature. Revelle et al. [23] survey a safety stock costs as well as the economies of scale that exist in the
number of important problems in facility location including UFLP. transportation costs from suppliers to DCs.
Moreover, an extensive study can be reached in Mirchandani and Reverse and closed loop logistics network design problems are
Francis's book [21]. widely studied in literature. Louwers et al. [19] study a facility
Collecting the right amount of products is one of the main location allocation model for reusing of carpet waste as a case study
concerns in a CLSCM and there are two main factors that effect the and analyze its two applications one in Europe and the other in the
collection amount; the incentives offered to the customers and the United States of America. Jayaraman et al. [14] examine the electronic
proximity of collection centers to customers. Firstly, some compa- equipment recovery closed loop network design with multi-product
nies offer financial incentives to end users to return their products capacitated warehouse location MILP to seek the optimal number
and as the amount of this incentive increases, more products will and location of remanufacturing facilities while minimizing the total
be returned. Amount of this incentive plays a major role in the costs. Fleischmann [9] emphasizes a continuous network design
collection activity and it must be critically analyzed to maximize model using a MILP model for product recovery and develops a
the profitability of the companies. Secondly, accessibility of the heuristic algorithm. Shih [27] develops a MILP model to design the
collection centers (CC) is another important issue for the custo- reverse network flow of computers and home appliances in Taiwan.
mers to return their products. As the number of CCs increase, they Jayaramann et al. [15] study on a MILP reverse distribution model to
will be closer to the customers and more products will be minimize the total transportation and fixed costs of opening facilities
returned, however, opening more CCs might be too costly for the and develop a heuristic procedure for this problem. Beamon and
companies. Thus, opening the right number of CCs at the right Fernandes [2] develop a closed loop supply chain model to determine
places is a critical factor for a CLSCM. In addition, in a society, since the location and sorting capability of warehouses and collection
the end users are generally located at the same place as the new centers as well as deciding the quantity of flow between the sites.
potential buyers for these products, the CCs can be integrated with Schultmann et al. [26], Biehl et al. [4], Lu et al. [20], Ko et al. [16] and
the companies’ traditional distribution centers (DC) to carry out Lee et al. [17] are some of the other studies related to reverse logistics
both distribution and collection activities to decrease both facility network design problems. Listes [18], Salema et al. [24] and Inder-
opening and operating costs. Thus, we aim to determine the best furth [13] consider randomness in reverse logistics problems. Aras
locations for these collection and distribution centers (CDC) in et al. [1] develop a discrete facility location–allocation model to find
order to maximize the total supply chain profit considering both the predetermined number of collection centers and the optimal
the forward and the backward flow of products. financial incentive values for different return types and propose a
In addition to the incentives offered to the end users, we also focus heuristic procedure using two nested loops to solve this problem. In a
on determining the sales price of new products in order to maximize recent study, Salema et al. [25] formulate a multi-period, multi-
total profit, which will also depend on the locations of the CDCs. In product network MILP model for the simultaneous design of a
literature, there are only a few studies that consider pricing with Portuguese glass company's forward and reverse networks. Fuente
facility location decisions and they mainly focus on pricing issue while et al. [10] deal with an integrated supply chain management which
designing an optimal drop-off facility network. Wojanowski et al. [29] combines the new processes taken from reverse supply chain into
concentrate on the forward supply chain and using continuous existing processes of forward supply chain.
modeling approaches, they determine the sales price of the product Different from the literature, we develop a model that con-
by focusing on the deposit-refund policy to maximize firm's profit. In siders a closed loop supply chain in which we analyze the location,
this policy, the sales price is known by the customers including the inventory and pricing decisions simultaneously. We build a MINLP
deposit amount which will be paid back when customer takes back model for this problem, however due to the complexity of the
the used product at a collection facility. In that model, customers' problem, it is not possible to solve this model using commercial
willingness with regards to purchasing and returning the product is solvers such as BARON for even moderately sized problems (when
related to a stochastic utility choice model. On the other hand, Aras the number of potential facility locations are 20 or higher). Thus,
et al. [1] concentrate on reverse supply chain and try to optimize only we develop three hybrid metaheuristic algorithms to solve our
return decision of product holders rather than the purchasing decision model for medium or large sized instances. Simulated Annealing
under deposit-refund requirements. Another difference between these (SA), Tabu Search (TS) and Genetic Algorithms (GA), all hybridized
studies arises from their proposed pick-up policies. In Wojanowski with Variable Neighborhood Search (VNS) are compared in terms
et al. [29], customers have to bring their used products to the of both solution quality and computational time. We also present
collection centers, however, Aras et al. [1] develops a pick-up policy an upper bound for the objective of our problem based on
with vehicles of limited capacity. piecewise linear approximation of a part of the nonlinear objective
In this study, we also consider the inventory costs of the CDCs since function and compare the results of the heuristics with this upper
inventory costs also play an important role in the management of bound through an extensive computational study. We also use this
these centers. The collected products from the customers should be piecewise linear approximation technique to develop a heuristic
sent to the manufacturing facility and new products should be for our problem. In the next section, we define our problem and
brought from the manufacturing facility at certain time intervals. present our model in detail. Then, in Section 3, we describe the
Therefore, the inventory system should be organized and controlled solution methods and the heuristics that we develop for this
appropriately to provide integrity between the forward and backward model and then in Section 4, we present our computational
flows of the materials. Since we assume a deterministic rate of studies. Finally in Section 5, we conclude our study.
 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103 95

2. Model end-of-life products that are accumulated at that CDC are collected
and taken back to the production facility. We let Ai denote the cost
In this study, we consider a company that has a production of transportation of the products from the facility at node i to or
facility at a certain location and have customers at N different from the production facility. The interarrival times of the ship-
customer zones. The company needs to decide on where to open ments made to a CDC at zone i, Ti, is effected by the demand and
retail facilities (stores) to sell their products and these facilities collection rates at that CDC as well as the fixed shipment cost for
will also act as collection centers of the used products. The used the CDC at zone i, Ai, and inventory holding costs, hpi and hri, for the
products will be collected at these stores and sent back to the new and used products, respectively. We use the well known EOQ
production facility for remanufacturing. We do not differentiate model to determine the optimal Ti values such that the total fixed
between the remanufactured and original products that are sold at shipment costs and total inventory holding costs are minimized.
these stores. We consider M different potential store locations and The notation that we use in this study is summarized below with
assume that each opened facility acts both as a collection and a the MINLP formulation to maximize the total profit, by determin-
distribution center. V1 denotes the set of potential facility locations ing the optimal number and location of the CDCs as well as the
and V2 denotes the set of the customer zones. We let c denote the amount of price and incentive offered.
production cost of a new product, p denote the sales price for new
Parameters:
products and r denote the financial incentive (refund) offered for
Kj: number of people in zone j
each returned product. Even though the returned products can be
Dij: demand of customers in zone jA V2 that go to the facility
of different quality levels, we let s denote the expected value of a
i A V1
returned product for the company, which will be used in the
Bij: amount of returned products from zone j AV2 to the facility
remanufacturing process. In our model the used products are
iA V1
collected by the opened facilities and transported back to the
s: amount of gained from unit used product
production facility for remanufacturing/recycling.
c: cost of producing a new product
We let Kj denote the number of people in zone j A V 2 and
Fi: fixed cost of a CDC located at candidate site iA V1
depending on price and accessibility of the facilities, we assume
k: parameter denoting the price sensitivity of demand
that a portion of these people will buy a new product from an
b: parameter denoting the incentive sensitivity of return
open facility. In our model, we assume that the customers go to
amounts
the retail stores themselves to buy the product and if the price of
Ai: fixed ordering cost of a CDC located at candidate site iA V1
the product decreases or if a facility is opened at a closer location
from the factory
to the customers in a certain zone j, more of the customers in that p
zone j will buy the product. Then, the demand from zone j to a hi : unit holding cost of new product in the CDC located at
facility at zone i will be Dij ¼ K je kp xij αij where k is the coefficient candidate site iA V1
r
for price sensitivity of demand, xij is a binary variable such that it hi : unit holding cost of used product in the CDC located at
takes the value 1 if the customers in zone j go to the facility in candidate site iA V1
zone i A V 1 , and 0 otherwise; and αij is a parameter between 0 and Decision variables:
1, depending on the distance between zones i and j, denoting the In the model, we have financial, inventorial and location
portion of people who will go to zone i from zone j to buy a decision variables as follows:
product. Note that in this model for every j, Dij will be positive for p: optimum price value offered for unit product
only at most one i and Dij will be zero for all other i since we use a r: optimum incentive value offered for unit used product
binary variable xij. Thus, for all j, Dij will be zero if the customers in Ti: inventory cycle time for CDC located at candidate site i A V1
zone j do not go to zone i to buy the product (i.e. if xij ¼ 0). In our

1 if the CDC is opened at candidate site iA V1
model, the end of life product owners (ELPO), who are the yi ¼
0 otherwise (2.1)
customers of the forward supply chain, also transport the products
from their locations to their assigned facility. Therefore, the

1 if customer jϵV2 is served by the CDC located at candidate site i A V1
tendency of customers returning the used products is also asso- xij ¼
0 otherwise (2.2)
ciated with the distance between the facility and the ELPO's
location. Similar to the new product demand, the number of
returned products will be Bij ¼ K j ð1  e  br Þxij β ij where b is the
Then, we can write the mathematical model (named Model
incentive sensitivity of the collected amount and βij is a parameter
1 afterwards) as follows:
between 0 and 1, depending on the distance between zones i and j,
Model 1:
denoting the portion of customers who will go to zone i from zone XX
j to return a product. Note that, in our model, the customers will Π ¼ max ðp cÞK j e  kp xij αij
go to the open facilities themselves in order to buy or return a iϵV jϵV
X1X2 X
product and thus we do not consider any cost of transportation for þ ðs  rÞK j ð1  e  br Þxij β ij  F i yi
the company between facility nodes and customer nodes in the iϵV 1 jϵV 2 iϵ V 1
2 0 " # 13
objective function of our model. Instead, the decision regarding X Ai X h
p
h
r
whether to open a facility or not will affect the accessibility of the  4 þ@ K je  kp
xij αij T i þ K j ð1  e
i  br
Þxij β ij T i i A5yi
iϵV
Ti jϵ V
2 2
stores and thus affects the demand for new products or the 1 2

collected amounts of end-of-life products at any customer node i ð2:3Þ

through the values of xij αij and xij β ij since an open facility at a
subject to
closer location to a customer node will have a higherαij or βij value X
with the corresponding xij ¼ 1. xij ¼ 1 8jAV2 ð2:4Þ
In our model, when a returned product is dropped to a CDC, it i A V1

will not be economical to send that product immediately back to

the production facility. To integrate the collection and distribution xij r yi 8 iA V 1 ; 8 jϵV 2 ð2:5Þ
networks more efficiently, we assume that when a shipment of The objective function (2.3), labeled as Π, consists of mainly
new products is made to a CDC from the production facility, the two parts to maximize the total profit. The first part calculates the
96 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103

sum of the net profit from selling new products and the value of that the optimal p and r values cannot be written explicitly in
collected end of life products. The net profit is found by multi- terms of the other variables as we did for Ti due to the structure of
plying the amount of total customers’ demand with profit per unit the objective function (3.2).
ðp  cÞ and the amount of total return with net unit gain. The Exact solution algorithms for this type of problems are ana-
amount of payment given by the company as incentive is dis- lyzed in some papers (e.g. Conn and Cournejols [7]). However,
counted from the value of used products and net unit gain of solving this problem with exact methods, even for moderately
company becomes ðs  rÞ. In our model, we use the function e  kp to sized problem instances is seen to be very difficult. We first try to
denote the effect of price on demand, however, other types of solve this model using a commercial MINLP solver, BARON 9.3.1 in
functions such as linear functions can also be used to model the GAMS and we could obtain the optimal results when the number
effect of price on demand and similar results can be obtained. of potential facility locations are small as given in Table 1 in the
Likewise, the proportion of end-of-life products that is collected is numerical results section. However, when we increase the number
modeled using a function ð1  e  br Þ as the collection increases of potential facility locations, it became impossible to obtain
with r. Besides the pricing outcome, considering the proximity of acceptable results with the commercial solver in a reasonable
the customers in zone i to a CDC at zone j, demand and return time amount.
amounts are affected by αij and βij parameters. We think of αij and Since it is very difficult to solve this model with exact methods,
βij parameters as a decreasing function of the distances between we develop heuristics to find near optimal solutions in a reason-
the zones i and j that will denote the portion of customers that will able time. In order to analyze the effectiveness of our heuristics,
go to zone j from zone i to buy or return a product. we also find an upper bound for our problem using a relaxation of
The second part of Π (2.3) denotes the opening, operating and the original problem. We first describe our approach to obtain the
inventory costs of the facilities. F iyi denotes the opening and upper bound and our first heuristic named Piecewise Linearization
operating cost of facility i per unit time and the remaining part (PL) heuristic in the next subsection. Then, we describe three
denotes the fixed order and inventory holding cost of facility i per different metaheuristics developed for this problem in the later
unit time. We note that the classical EOQ model and its related subsections.
cost function is used in this model. Ai =T i denotes the fixed
shipment cost per unit time and the remaining portion denotes 3.1. Development of an upper bound and piecewise linearization
the inventory holding cost per unit time at zone i if a CDC is heuristic
opened there.
After the optimum cycle time of each candidate CDC are restruc-
tured in terms of price and incentive values as stated in Proposition 1,
3. Solution methodologies we observe that the part of the objective function related to the
P pffiffiffiffiffiffiffi
inventory decisions, which is the part stated as W i where
The proposed model can be considered as a combination of P iϵV 1
three subproblems: location, pricing and inventory. Since the basic W i ¼ 2Ai jϵV 2 ðK j e  kp
x α h
ij ij i
p
þK j ð1  e  br
Þx β h
r
ij ij i Þ , is one of the
facility location problem is known to be NP-hard, with additional main parts that makes the problem difficult to solve. The square root
inventory and pricing decisions, our model is also NP-hard. In function together with the minus sign in front of it disturbs the
order to solve this problem, we first analyze the inventory concavity of the objective function making the problem difficult to
decisions in this model. We observe that for any given p, r, yi solve. For this reason we propose applying a linearization of the
and xij decisions, the optimal value of Ti can be found through the square root function in order to obtain a concave objective function
classical economic order quantity (EOQ) results in inventory at the end.
pffiffiffi
theory as stated in the following proposition. Since x is a concave function, we use piecewise linear lines
that lie below this function, so that the resulting value will be a
Proposition 1. If a facility i is opened, i.e. if yi ¼1, then the optimal
lower bound for the inventory cost and the objective function will
cycle time between inventory replenishments at this facility will be as
be higher than the original one leading to an upper bound for our
follows:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi problem. For piecewise linearization of the square root function,
u we define a new index set t ¼ 0; 1; 2; …:T to define the break
u 2Ai
Ti ¼ u
t P ½ðK j e  kp Þxij αij hp þ ðK j ð1  e  br ÞÞxij β hr  ð3:1Þ points. Then, slope of the line in every segment is calculated and
i ij i
jϵV 2 using these slopes we find an approximate value for the function
inside the square root term using the model below. We use the
following new parameters and variables for the piecewise linear-
We omit the proof for the above proposition since it is ization model of the square root function.
straightforward EOQ result. We determine the optimum cycle
time Ti of each facility using Proposition 1 in terms of other
parameters as given above. Proposition 1 allows us to reduce the
number of decisions variables in our problem making the resulting
problem easier to solve. After substituting the Ti values with the Table 1
above equation in the model, the objective function for our second Results of the mathematical models obtained through the solver BARON.
model (called Model 2 afterwards) will be as follows:
XX XX X Size Model 1 Model 2 Model PL
max ðp  cÞK j e  kp xij αij þ ðs  rÞK j ð1 e  br Þxij β ij  F i yi
iϵ V 1 jϵ V 2 iϵV 1 jϵV 2 iϵ V 1 % Opt. CPU % Opt. CPU % Opt. CPU
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Gap (seconds) Gap (seconds) Gap (seconds)
u 0 1
Xu
u @X
ðK j e  kp xij αij h þ K j ð1  e  br Þxij β h ÞA 10  10
p r
 t2Ai i ij i ð3:2Þ 40.4 28800 0 189.2 0 74.8
20  20 76.2 28800 16.3 28800 9.11 28800
iϵ V 1 jϵV 2
50  50 99.5 28800 55.9 28800 47.6 28800
100  100 – 28800 – 28800 - 28800
Observe that the above objective function is still nonlinear and this 200  200 – 28800 – 28800 - 28800
problem is a mixed integer nonlinear program (MINLP). We note
 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103 97

Parameters: resulting solution and use the values of the decision variables as a
P
ai ¼ 2Ai jϵV 2 ðK j αij hi þK j βij hi Þ: maximum possible value of the
p r
heuristic solution for our problem, naming this heuristic as
function inside Piecewise Linearization (PL) heuristic. We recalculate the original
the square root term, also denoting the last break point. objective function value by plugging in these decision variables
ait ¼ ai t=T: the break points for piecewise linearization, for all into the original objective function since this solution is a feasible
t ¼0,1...T solution for our problem and we call the resulting objective
pffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi function value ZPL, which is equal to
Δit ¼ ð ait  ai;t  1 Þ=ðait  ai;t  1 Þ: the slope of the linear
XX
approximation function in segment t Z PL ¼ max ðp  cÞK j e  kp xij αij
Variables: iϵV 1 jϵV 2
XX X
ωit: non-negative variable denoting the amount of the function þ ðs  rÞK j ð1  e  br Þxij βij  F i yi
value that lies inside iϵ V 1 jϵ V 2 iϵV 1
P vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
segment t¼1,2...T. Note that Tt ¼ 1 ωit denotes the whole u 0 1
function value as stated in Xu
u @X r A
xij αij hi þ K j ð1  e Þxij β ij hi Þ
p
 t2Ai ðK j e  kp  br
constraint (3.6) below. iϵ V 1 jϵV 2
γ it ¼ 1 if segment t is totally covered by the function value, and
0 otherwise ð3:9Þ
i.e. γ it ¼ 1 if ωit ¼ ait  ai;t  1 and γ it ¼ 0 otherwise. We observe that if p and r values are given as parameters into the
model, this model turns into a mixed integer linear program
(MILP), which is much easier to solve than the MINLP model
Using the above variables and parameters, the resulting model
which includes nonlinear terms. In addition, we can use the solver
(called Model PL afterwards) will be as follows:
XX XX CPLEX to solve this MILP model instead of using BARON. However,
max ðp  cÞK j e  kp xij αij þ ðs  rÞK j ð1  e  br Þxij β ij note that the previous models without the piecewise linearization
iϵ V 1 jϵ V 2 iϵV 1 jϵV 2 (Models 1 and 2) are still nonlinear programs even for given p and
X XX
T r values due to the square root term in the objective function and
 F i yi  ωit Δit ð3:3Þ thus we cannot use the CPLEX solver for those models. Thus, as an
iϵV 1 iϵV 1 t ¼ 1
alternative solution methodology in this case, we do a two-
subject to dimensional search over p and r values and for each given
X combination of (p,r), we solve the above model as a MILP model
xij ¼ 1 8jAV2 ð3:4Þ using the solver CPLEX in GAMS. We use the two dimensional
i A V1
golden section search method (see Giordano et. al [11], p. 289–295,
xij ryi 8 i A V 1; 8 jϵV 2 ð3:5Þ for the implementation of the golden section search method) to
search for the best p and r combination since we observe through
X
T X our initial experiments that our objective function is unimodal in
ωit ¼ 2Ai ðK j e  kp xij αij hi þK j ð1  e  br Þxij β ij hi Þ
p r
8 iA V 1 terms of p and r. In this algorithm, for any given p and r
t¼1 j A V2 combination, we run the above model and determine the optimal
ð3:6Þ yi and xij values for that combination. At the end, we pick the best
p and r combination with the resulting yi and xij values found
ωit r ðait  ai;t  1 Þγ i;t  1 8 i A V 1; 8t ¼ 2 .. T ð3:7Þ through the model. We know that r o s and p 4 c, thus s forms an
upper bound for r and c forms a lower bound for p in our search
ωit Z ðait  ai;t  1 Þγ it 8 iA V 1 ; 8t ¼ 2 .. T ð3:8Þ for the optimal p and r values and the use of these bounds
In this model, we approximate the term inside the square root decreases the solution time.
function with piecewise linear lines. The last term in the objective
function denotes the approximate value for the function inside the 3.2. Metaheuristics
square root term in the original model. Constraint (3.6) is used to
calculate the term for which the square root will be found. Constraint In addition to the PL heuristic, we also develop three heuristics
(3.7) satisfies the condition that if γ i;t  1 ¼ 0, meaning that the value for the solution of this problem. All the heuristics that we use in
of the function does not cover segment t 1 totally, then ωit should this study are composed of two nested loops. In the outer loop in
also be 0. Constraint (3.7) also gives an upper bound for ωit. our heuristics, we use metaheuristic methods to find the best
Constraint (3.8) satisfies the condition that if γ it ¼ 1, meaning that locations of CDCs that will be opened to maximize the profit and in
the value of the function covers segment t totally, then ωit should be the inner loop, the best price and incentive values are searched
equal to the whole length of segment t, which is ait ai;t  1 . through a gradient search algorithm. In the outer loop, we
After the piecewise linearization of the square root function, hybridize three well-known metaheuristic methods which are
the objective function becomes a jointly concave function of p and Simulated Annealing (SA), Tabu Search (TS) and Genetic Algorithm
r given the other variables. The approximation with this piecewise (GA) with Variable Neighborhood Search (VNS) as a solution
linearization depends on the number of segments, T, used in the procedure and name them as SAVNS, TSVNS and GAVNS. We apply
piecewise linearization procedure. As the number of segments all three heuristics in our study and compare their performances
increases, the error resulting from the approximation decreases, in our test instances. Since the gradient search algorithm that we
but the computational complexity increases with the number of use for the price and incentive values in the inner loop is
segments used. Therefore, we consider the trade off between straightforward and can be commonly used in many textbooks
complexity and allowable error range and according to our (e.g. see Bertsekas [3]), we focus on the metaheuristics SAVNS,
preliminary experiments, we use T¼ 10 equally spaced segments TSVNS and GAVNS, that we use in the outer loop in the next
in the piecewise linearization procedure. subsections.
We solve this problem using the solver BARON in GAMS after In all the proposed algorithms, the optimum locations of
the piecewise linearization of the model and we obtain an upper opened CDCs yi are represented with a binary vector in which
bound for our problem and call it ZUB. In addition, we look at the the ith entry shows whether facility i is open or not. For example,
98 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103

the representation of a solution with the vector [1, 0, 0, 1, 0, 0, 1, 0, maximum number of allowable iterations that incumbent solution
1, 0] denotes that the CDCs located at 1, 4, 7 and 9 are selected to does not improve. Both criteria are used at the same time and
open. Then, using these yi values, the assignments of customer reaching one of them is enough to terminate the current neigh-
zones to open facilities are represented with another vector x. For borhood and to move to the next one. In each neighborhood, the
the same example, the vector x ¼[1 1 7 4 4 9 7 9 9 9] denotes that allowable number of non-improved solutions are determined
the customer zones 1 and 2 are assigned to facility 1, 4 and 5 are again with respect to the size of the neighborhood. Non-
assigned to facility 4, 3 and 7 are assigned to facility 7 and 6, 8, improved solutions are accepted to allow diversification and
9 and 10 are assigned to facility 9. We also note that, for each escape from local maxima. However, new neighborhood begins
solution found by these metaheuristics for yi and xij values, we with a solution that is the best solution in the previous neighbor-
then apply the gradient search algorithm at each iteration to hood up to that time. In other words, best found solution is also
determine the best p and r values for those solutions and calculate kept in each neighborhood when allowing to move towards a
the objective function value accordingly. The implementations of worse solution.
the metaheuristics are detailed below.

3.2.2. Proposed TSVNS algorithm

3.2.1. Proposed SAVNS algorithm We combine TS algorithm with VNS in our proposed algorithm
In this study, we hybridize simulated annealing with a variable to help the solutions not to get trapped at a local optima and allow
neighborhood search (VNS) algorithm. At each iteration, within a it to visit larger neighborhoods with systematical change. We use
defined neighborhood structure, the best solution is sought and deterministic search to expand neighborhoods by finding the best
when the algorithm stops for that neighborhood, i.e termination neighbor of incumbent solution in a sequential order. Basically,
conditions are met, another neighborhood structure is created. The VNS jumps from the incumbent solution to a new one if and only if
implementation of the algorithm is detailed as below. a better solution has been found and does not forbid any moves.
Initial Solution: For the proposed model, to obtain an initial However, Tabu Search employs tabu restrictions to prevent visiting
solution to start the algorithm, we generate m different solutions for earlier selected solutions throughout the iterations. At each itera-
facility location array y randomly and then pick the best resulting tion, the best solution which has the maximum profit is sought by
one as our initial solution. We assign the customers to their nearest applying local search within the allowed set in the current
open facilities to determine the initial vector x for each y. The value neighborhood. Then, the best solution is added to the tabu list
of m is determined after several trials in our numerical experiments and one of the solutions that were already in the tabu list is
and we use m¼50 in our computational study. removed in a FIFO order. Thereafter, the search continues around
Neighborhood Structure: The neighboring solutions are gener- the best solution. When the termination criteria are met, the
ated by applying combinations of add, drop or swap moves to the algorithm stops and the best solution found so far within the
current solution. The proposed SAVNS has three neighborhoods to previous neighborhood is moved to the next neighborhood in the
visit different solution spaces. In the first neighborhood, as a move same way as SAVNS.
operator, we apply 1-0 exchange to current solution. In other Initial Solution: For TSVNS, initial solution is again constructed
words, an add move which opens node i if the node is not opened as in SAVNS.
or a drop move if the node is opened is assigned to the current Neighborhood Structure: To find the best neighborhood struc-
solution. By changing the status of ith coordinate, n different ture, we compare four different neighborhood structures for the
solutions are visited. The second neighborhood is composed of array y in this case. In the first neighborhood, as in SAVNS
changing status of two randomly selected coordinates simulta- algorithm, 1-0 exchange is applied to current solution and visits
neously, i.e applying twice 1-0 exchange in the same iteration. all possible neighbors. The opened facility number increases or
Neighborhood is generated by applying predetermined number of decreases by one with the help of 1-0 exchange. The second
1-0 exchanges to two randomly selected nodes. Finally, in the third neighborhood structure is composed of 1-swap moves. 1-swap
neighborhood, status of randomly selected three nodes are chan- move relocates a closed facility with an opened one in the current
ged by applying 1-0 exchange to them. solution. All pairs of candidate locations that generates a different
We apply similar neighborhood structures for the array x. Every solution from current one is visited through the local search to
time the vector y is changed, we initially assign the customers in change the sequence of the facility positions. The possible neigh-
zone j to the nearest open facility i and create an initial vector x for bors that provide the same order, i.e. if both of the candidate
that y. Then, as a first neighborhood structure, for each j, we locations in a pair are opened or closed, are prevented to visit in
change its assignment value to another open facility i. In the the proposed algorithm. By this way, each solution in a neighbor-
second neighborhood structure, we pick two different customer hood assures a different combination of the opened CDCs. In the
zones j and change their assignments with each other. Addition- third neighborhood, 3-swap moves relocates the 3 consecutive
ally, in the third neighborhood structure we randomly select three facilities. All 3 combinations of current solution are visited in this
customer zones j and change their assignments with each other. neighborhood, therefore this neighborhood size is much bigger
Cooling Schedule: We tried various cooling schedules such as than the second neighborhood. By the 3-swap moves, the number
linear, exponential and hyperbolic ones. In the equations below, i of opened facilities remains constant, however, locations of the
denotes the cycle that the algorithm runs within, ti denotes the opened facilities change. In the last neighborhood structure, 1-0
temperature at cycle i, to denotes the initial temperature and N exchange is applied twice to the current solution. By this way,
denotes the maximum cycle number. For linear cooling schedule, unlike the others, number of opened facilities can increase or
the temperatures are updated according to the relation decrease by two. We also note that for the vector x, we apply the
t i ¼ t o  iðt o  t N Þ=N. Similarly, for exponential cooling schedule same neighborhood structures as we use in SAVNS.
t i ¼ t o  ilnðto  tN Þ=lnðNÞ and for hyperbolic cooling schedule, the Tabu definition: When a facility is involved in obtaining a new
relation t i ¼ to 2 tN ð1  tanhðð10ni=NÞ  5ÞÞ þt N is used. solution which can either be added or removed, named as facility i,
Termination Criteria: In the proposed algorithm, we use two ith node position cannot be changed again for a certain number of
different stopping criteria. The first one is the maximum number iterations, i.e. labeled as tabu moves and will not be allowed for a
of iterations allowed in a neighborhood. This number is adjusted certain number of iterations such as the size of tabu tenure. Using
according to the size of the neighborhood. The second one is the dissimilar neighborhood structures result in keeping tabu
 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103 99

restrictions in different ways. In the 1-0 exchange kind of moves, Mutation Operators: Like crossover operators, we apply differ-
keeping the one node position, the exchanged, is sufficient as tabu ent mutation operators in arrays y and x. After applying crossover,
move. In the second kind of neighborhood, a pair of different some genes of offsprings are altered, with small probability (i.e.
positioned coordinates are swapped. The coordinates resulting in mutation probability), in order to diversify the search space. For
the best solution in the whole neighborhood are labelled as tabu each offspring, a mutation probability is created randomly and we
moves and in tabu list, the pair of coordinates are kept. Therefore, determine whether it undergoes a mutation or not. In the first
the 1-swap move using the coordinates used previously cannot be neighborhood, which was described above for crossover operators,
applied until they are non-active in tabu list. We use the recency- a random number is generated between 1 to n and that much of
based memory in the TS heuristic which forbids moves towards randomly selected gene status are changed to another feasible
the most recent visited solution. Size of tabu tenure is also selected value. In the second neighborhood, inversion mutation technique
according to the problem size. is applied to a chromosome which will be mutated. In the third
Aspiration Criteria: We use the classical aspiration criteria of neighborhood, head and tail of the chromosome is changed from a
allowing a move even if it is tabu, if it results in a solution which randomly selected gene position. The mutation probability of the
provides a profit value better than the incumbent one. neighborhoods increases so as to diversify the search space by the
Termination Criteria: Like SAVNS, we use two termination end of algorithm to escape from local maximum.
criteria at the same time as maximum number of iterations and Formation of the next generation: After the generation of mating
the maximum number of allowable non-improved solution. pool, the chromosomes to compose the next generation are
selected according to their fitness values. We use the roulette
wheel selection (RWS) method for the selection of chromosomes
3.2.3. Proposed GAVNS algorithm in the next generation. In this process, fittest individuals will tend
In this algorithm, each parent, or actually named as chromo- to have a greater chance of survival than weaker ones and form the
some, represents a solution of the problem. Combination of the mating pool for the next generation. In other words, fittest
genes form the chromosomes and N combinations of the chromo- individuals has larger share of the roulette wheel. The individuals
somes generate a population. Biological evolution operators, are sorted according to their fitness values and based on the
mainly crossover and mutation are used to create a new genera- randomly selected point, an individual is selected to form the next
tion from current population. Selection of new population generation. After each selection, the probabilities are calculated
depends on their fitness values which are calculated upon the again in order to provide greatest chance to fittest individuals
objective values of the chromosomes. Integrating the VNS into GA among the remaining ones.
also causes diversification in search space of each neighborhood by
changing the formation of new generations while affecting the
crossover and mutation structures. As mentioned in previous 4. Computational studies
algorithms, in each iteration within a defined neighborhood
structure, we firstly generate the offsprings (i.e new generation) In this section, we analyze the performances of proposed
using genetic operators. Then, using fitness values of the chromo- models and heuristics in terms of the solution quality and
somes a new population is selected deterministically. When the computation time through numerical experiments. All of the runs
termination criterion of one neighborhood is met, the algorithm throughout the computational experiments are performed on a
jumps to next neighborhood to diversify and explore the search workstation with an Intel(R) Core(TM)2 Duo processor, 2.53 GHz
space. Neighborhoods are designed in nested strategy and enlarges speed, and 4 GB of RAM. The nonlinear mathematical models are
as proceeded to the end of algorithm. solved with the solver BARON 9.3.1 and the mixed integer linear
Initial Population Generation: A population of arrays y denoting models are solved with CPLEX 12.2 with GAMS 23.6. All of the
the open facilities are created randomly from candidate solutions. metaheuristic algorithms are coded in C. In Section 4.1 we describe
For each y, we also create a population of matrices x denoting the how the test instances are generated and parameters are set, in
assignments of customers to open facilities. Population sizes for x Section 4.2 we present the results of the mathematical models and
and y are determined according to the size of the problem at hand. then in Section 4.3, we present the results of the heuristics and
Crossover Operators: Instead of move operators in SA and TS, GA discuss them in detail.
uses crossover and mutation operators. We use three different
crossover operators to visit different solution spaces and diversify 4.1. Data generation
the search for the arrays y and x. In the first neighborhood,
uniform crossover mask is generated randomly as the same length For computational analysis, we first create a base case instance,
of the chromosomes. Parts of chromosome or namely parents, to considering an electronics manufacturer with the manufacturing
be exchanged is determined by the mask. The offspring 1 is facility at Izmit, Turkey. Fig. 1 shows the map of the Anatolian side
produced by taking the bit from parent 1, if the corresponding of Istanbul and we use N ¼10 different districts on this map as the
mask bit is equal to 1, or take from parent 2 if the corresponding customer zones in our initial base case experiment. In all our
mask is equal to 0. After the crossover, two offsprings will be experiments, the customer zones are also used as candidate
created from mated parents. In the second neighborhood, 1-point locations for potential CDCs (i.e. M ¼N). We try to use real life
crossover is applied to mated parents in which parents are cut data for the parameters in our model, however it is not possible to
from a randomly selected position and their chromosome strings determine the exact values for some of the parameters. Thus, for
are swapped from their tails. By this way, two offsprings are those parameters, we make realistic assumptions about their
generated and partly similar to their parents. In the third neigh- values as detailed below.
borhood, 2-point crossover is used. The pair of parents’ chromo- For the parameter Kj, which denotes the number of potential
somes are cut from the randomly selected two points and the customers in each zone, we use the populations of the districts
middle segments of the parents’ strings are swapped. Different given in Fig. 1, according to 2013 survey, multiplied by 0.02
crossover probabilities are tried during the tuning. If crossover assuming 2% of the population are potential customers of this
probability of the mated parents is less than the set value, in other company. We let αij be a decreasing function of the distances
words, if crossover is not applied, the selected parents are between nodes i and j, tij, such that αij ¼ 1=1 þ 0:1t ij , where the
duplicated in order to generate the two offsprings. distances between different zones (tij) are obtained from Google
100 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103

Fig. 1. Districts from the Anatolian side of Istanbul used as N¼ 10 customer zones.

Maps. Similarly, for the return coefficient βij, we assume that half very poorly even for small sized problems. It results in an
of the sales are returned as used products such that βij ¼ 0:5αij . optimality gap of around 40% in a time limit of 8 h for N ¼ 10
Fixed ordering and operation costs for each candidate CDCs are customer zones and the optimality gap gets even worse as N
generated considering the rental costs at these different districts increases. When we look at Model 2, in which the variables Tij in
and these values range between 5000 and 20,000. The fixed Model 1 are eliminated by writing them in terms of the other
ordering costs, Ai, are generated considering the transportation variables, the performance of the model gets better and we can
costs and the distances of these zones to the manufacturing facility find the optimal solution for N ¼ 10 customer zones using Model
and these values range between 500 and 1500 in our study. The 2 in 437 s in our test instance. However, for N ¼20 and N ¼50,
cost of producing a product, c, is assumed to be 100 and the value Model 2 also fails to find the optimal solution in 8 h and results in
of a returned product for the company, s, is assumed to be 50. We an optimality gap of 16.3% and 55.9% at the end of 8 h, respectively.
let the holding costs, hpi and hri, to be 1% of c and s, respectively, When we consider Model PL, this model can also be solved to
p r
such that hi ¼ 1 and hi ¼ 0:5 for all i. Finally, we assume that the optimality in 74.8 s for N ¼ 10 customer zones, but we cannot find
parameters for price sensitivity of demand, k, and incentive the optimal solution of this model, either, for N ¼20 and N ¼50
sensitivity of return amounts, b, are both 0.01. customer zones in 8 h. Model PL results in an optimality gap of
In addition to the base case instance with N ¼ 10 customer 9.11% and 47.6% at the end of 8 h, respectively, which are better
zones and parameters defined as above, we also generate addi- than the other two models in terms of the optimality gaps
tional instances in varying problem sizes (with N ¼20, N ¼50, obtained. We note that the solver BARON cannot find any feasible
N ¼100 and N ¼200 customer zones). For increasing values of N, solutions for any of the models in 8 h when the number of
we use additional districts from different cities of Turkey and use customer zones becomes 100 or 200, thus we leave those cells
the populations of these districts as Kj and the distances between empty in Table 1.
these districts as tij to calculate the values of αij and βij according to Since the mathematical models currently do not perform well
the equations stated above. due to the nonlinearities in the models, we try to separate the
non-negative variables p and r from the binary location variables
4.2. Analysis of the mathematical models (yi) and assignment variables (xij). We run the models for given p
and r values and then choose the best p and r combination that
In this section, we analyze and compare the performances of results in the maximum profit. We observe through our initial
the mathematical models that we develop in this paper. In Table 1, experiments that our objective function is unimodal in terms of p
we present the results of the mathematical models and their and r, thus we use the two dimensional golden section search
computational times for varying problem sizes. In all our models, method, with a tolerance of 0.01, to search for the best p and r
we use the solver BARON 9.3.1 in GAMS and run the models with a combination. Since r o s and p 4 c, we use these values as upper
time limit of 8 h. Since we know that r o s and p 4 c, we use these and lower bounds in our search for the variables r and p,
values as upper and lower bounds for the variables r and p, respectively. We use the bounds 0 and s for r and we do our
respectively, in all our models. We observe that Model 1 performs search for p between c and a large enough upper bound that we
 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103 101

choose depending on the problem parameters. We note that a of best p and r combination and such algorithms can decrease the
smaller tolerance value can be used if more sensitive results are total run time. However, in this study we focus on the models and
needed. In this algorithm, for the searched p and r values, we run the heuristics, and the development of the best algorithm for the
the models and determine the optimal yi and xij values for that search of best p and r values is out of the scope of this paper.
combination. Then, we pick the best p and r combination with the
resulting yi and xij values found through the models as the final 4.3. Analysis of the performances of the heuristics
solution.
Table 2 shows the average results of 10 runs of the models for In this section, we explain and compare the results of the
10 different combinations of given p and r values. We observe that proposed algorithms in terms of solution quality and tabulate run
if p and r is given, Model 1 can find the optimal solution for yi, xij times. The first row in Table 3 shows the results of our initial
and Ti values in 12 s on average for N ¼ 10 customer zones. experiment with the base case data as explained above. In order to
However, when N ¼20 or more, Model 1 fails to find the optimal understand the effect of the parameters in our model, we also do
solution in 8 h even when p and r are given. Also note that, the additional experiments by modifying the base case instance by
CPU times given in Table 1 are the average values for a single (p,r) doubling only one parameter value at a time as stated in the tables.
combination, but we need to run the models for many times to Note that in each row of Table 3 all the remaining parameters that
search for the best p and r combination. Thus, using Model 1 with are not stated in that row, are used as in the base case setting. We
this approach does not help us find the optimal solution. However, also analyze problems of different sizes as N  M¼ (10  10),
when we consider Model 2, in which Ti values are eliminated, it (50  50), (100  100) and (200  200). We note that since there
can find the optimal solution even when N ¼200 for a given p and exists some probabilistic structure in the implementation of the
r combination and the run times decrease significantly. However, metaheuristics, we run those heuristics for 10 times and report
for N ¼200, the run time for a single run of the model is around the best result. The CPU times reported in the tables below denote
40 min on average. Thus, it does not seem possible to use this the total time for those 10 runs. We also tried running the heuristics
model for the search of the best p and r combination since the run for more than 10 times but we observe that it did not improve the
time of the whole algorithm will need days or maybe weeks. results significantly and it is not worth the additional time.
Finally, when we look at Model PL, we observe that for given p and In order to obtain the best performances of the metaheuristics,
r values, this model turns into an MILP model while the other two we make some initial experiments for parameter tuning of the
models were still nonlinear programs (MINLP). As a result of this, heuristics by using the base case test instance for each problem
we use CPLEX 12.2 to solve the MILP model and we observe that size. Initial and final temperature values, cooling schedule, max-
the optimal solution for Model PL can be found in all cases and the imum iteration and maximum non-improved iteration numbers
run times are significantly less than the other models since the are the main parameters in the simulated annealing algorithm.
model does not include any nonlinearity anymore. It takes less After trying various values on different sized problems, we decide
than 10 s on average to find the optimal solution of this model to set To as 5000 and Tf as 0.01 in this heuristic, since they give the
even when N ¼200 and the runtime is less than a second on best results in our initial experiments. As termination criteria, the
average when N is less than 100. The last column in Table 2 shows most efficient and effective results are gained with max iteration
the total run time of the algorithm including the search for the number as 100 and max nonimproved iteration number as 20.
best p and r values. We observe that the run time of the whole For the tabu search heuristic, we analyze the effect of different
algorithm is less than 3 h when N ¼200 and it is much less for tabu tenures on solution quality and run time using 30 experi-
smaller sized instances. We also note that better algorithms than ments for each problem size and we observe that the best
the golden section search method can be developed for the search performances of tenure sizes are achieved when they are 7 for
50  50 problem size and 15 and 30 for 100  100 and 200  200
problem sizes, respectively. For the 10  10 problems, all tabu
Table 2
Results of the mathematical models with given p and r values.
tenures give the same solution, so we set the tenure as 7 for this
case. We observe that increasing tabu tenure size after a value
Size Model 1 Model 2 Model PL does not produce better results and increase the CPU time
redundantly. For the termination criteria, the maximum number
% Opt. CPU % Opt. CPU % Opt. CPU Total
of iterations is set as 50 and the maximum number of nonim-
Gap (seconds) Gap (seconds) Gap (seconds) CPU
proved solutions are set as 5, 5, 10 and 10 for 10  10, 50  50,
10  10 0 12.1 0 0.08 0 0.06 57.3 100  100 and 200  200 problem sizes, respectively.
20  20 0.47 28800 0 1.22 0 0.16 147.2 For the genetic algorithm, to decide on the best value of
50  50 1.43 28800 0 158 0 1.57 1316 crossover probability, pc, the values of 0,50, 0.65, 0.8, 0.95 and
100  100 2.54 28800 0 1203 0 3.65 3422
200  200 15.8 28800 0 2498 0 9.78 9183
1.0 are tried for all problem sizes and to decide on the best
mutation probability, pm, values of 0.001, 0.01, 0.1, 0.2 and 0.5 are

Table 3
Results of the optimal solution and the heuristics for N ¼ 10 customer locations.

Size Test instance Optimal PL SAVNS TSVNS GAVNS

Result CPU Result CPU Result CPU Result CPU Result CPU

10  10 Original 62507 189.2 62502 74.8 62507 0.68 62507 0.35 62507 5.46
Doubled Kj 167010 158.7 167007 51.5 167010 0.39 167010 0.19 167010 4.79
Doubled Ai 61465 186.6 61456 85 61465 0.55 61465 0.26 61465 5.02
Doubled hpi 61724 105.2 61715 90.1 61724 0.46 61724 0.23 61724 3.97
Doubled hri 62178 146.3 62176 100.9 62178 0.42 62178 0.29 62178 4.54
Doubled Fi 41262 143.7 41261 46.6 41262 0.24 41262 0.17 41262 5.37
Doubled tij 51315 85.5 51183 129.4 51313 0.45 51315 0.24 51315 5.42
102 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103

analyzed for all neighborhoods. We observe that the best perfor- doubled, we observe that the profits decrease slightly and the
mances are achieved in terms of solution quality and CPU times effect of the holding cost of new products are higher than the
when pc is 0.95 and pm is 0.2 and thus we use those values in our effect of the holding cost of returns. When the fixed operation
subsequent numerical analysis. In addition, maximum iteration costs (Fi) are increased, less number of facilities are opened and
number is set as 20 and maximum non-improved solution number the profits decrease substantially. Finally, when the distances, tij,
as 10 for 10  10 problem sizes. For other sizes, maximum iteration are increased, it leads to less demand and less returns and as a
number is set as 100 and maximum non-improved solution result more facilities are opened and the profits decrease
number is set as 20. For all problem sizes, we use 3 different substantially.
population sizes with 20, 40 and 80 chromosomes. Due to the failure of exact computational methods when
We first analyze a problem with 10 candidate locations for the solving large sized instances, we use the heuristics to find
facilities and 10 customer zones as stated as the base case instance approximate solutions. We also use Model PL with given p and r
above. For this 10  10 problem, we were able to find the optimal values and the golden section search method to determine the
solution by solving Model 2 with BARON 9.3.1. In Table 3, we best p and r values as explained above to obtain an upper bound
present the result of the optimal solution as well as the results of for the problem. We analyze the effectiveness of the heuristics by
the heuristics. We also present the running times of the mathe- comparing the results found by these heuristics with the upper
matical models and the running times of the heuristics. The first bound value. The deviations of the results obtained by each
row in Table 3 shows the results for our base case and in the other heuristic from the upper bound and the run times of the heuristics
rows we present the results when one of the parameters are are stated in Table 4.
doubled while all others remain the same as in the base case. We observe from the results presented in Table 4 that all the
We observe that all the metaheuristics can find the optimal metaheuristics perform pretty well and they all find the optimal
solution in a very short time in this case when N ¼ 10 and SAVNS result in our trials when N ¼10. When we look at larger sized
has the smallest runtime. When we look at the effects of the instances, we observe that the deviations increase with the
parameters on the results, we observe from Table 3 that when the problem size but TSVNS gives the best results on average. The
populations (Kj) are doubled, the profits become almost three deviation between the profits obtained with the best performing
times the base case. In addition, we observe that more facilities are metaheuristic and the upper bounds are around 1.58% when
opened as demand increases to serve the increased number of N ¼50, 2.65% when N ¼100 and 4.5% when N ¼200, on average.
customers. When the fixed ordering costs (Ai) are increased, we We observe that the heuristic PL does not perform as good as the
observe that the profits are slightly decreased. Similarly, when the metahueristics and the percentage deviation between the results
holding cost of products and returns (hpi and hri, respectively) are obtained by PL and the upper bound reach to almost 10% when

Table 4
Analysis of the performances of the heuristics.

Size Test instance PL SAVNS TSVNS GAVNS

% Deviation CPU % Deviation CPU % Deviation CPU % Deviation CPU

10  10 Original 0.008 74.8 0 0.68 0 0.35 0 5.46

Doubled Kj 0.002 51.5 0 0.39 0 0.19 0 4.79
Doubled Ai 0.015 85.1 0 0.55 0 0.26 0 5.02
Doubled hpi 0.014 90.1 0 0.46 0 0.23 0 3.97
Doubled hri 0.003 100.9 0 0.42 0 0.29 0 4.54
Doubled Fi 0.002 46.6 0 0.24 0 0.17 0 5.37
Doubled tij 0.257 129.4 0.004 0.45 0 0.24 0 5.42
Average 0.043 82.6 0,0006 0.46 0 0.25 0 4.94

50  50 Original 2.43 1316 1.67 4.24 1.73 24.22 1.36 52.63

Doubled Kj 2.60 1174 2.39 4.33 2.08 21.33 2.11 41.09
Doubled Ai 3.01 1445 2.89 4.58 2.18 25.62 1.87 60.43
Doubled hpi 4.36 1514 1.82 5.14 0.93 27.56 1.36 62.01
Doubled hri 4.08 1589 1.74 5.33 1.32 28.12 1.73 59.34
Doubled Fi 2.22 983 1.06 4.16 1.06 20.23 0.84 45.79
Doubled tij 5.47 2034 1.84 5.42 1.79 29.49 1.96 67.44
Average 3.59 1436 1.92 4.74 1.58 25.22 1.60 55.53

100  100 Original 5.24 3422 2.63 19.57 2.23 182.63 2.34 554.62
Doubled Kj 3.85 2897 2.47 18.34 2.31 173.57 2.43 533.24
Doubled Ai 6.45 3769 3.34 20.79 3.21 187.72 3.27 546.52
Doubled hpi 7.26 3913 3.21 19.63 3.09 182.53 3.16 595.91
Doubled hri 6.34 4245 3.17 21.07 2.14 194.33 2.81 502.48
Doubled Fi 3.35 2279 2.79 16.22 2.12 164.23 2.53 511.27
Doubled tij 6.98 4683 4.52 22.38 3.36 208.77 3.21 736.45
Average 5.64 3601 3.16 19.71 2.64 184.82 2.82 568.64

200  200 Original 9.09 9183 4.07 99.12 3.56 1322.76 4.01 2245.65
Doubled Kj 8.76 7943 4.28 83.46 3.82 1369.55 3.79 1875.33
Doubled Ai 10.92 9786 5.65 76.65 4.02 1443.14 5.04 2039.54
Doubled hpi 11.12 10023 5.74 80.32 5.33 1266.28 5.42 2122.55
Doubled hri 11.04 10467 5.87 92.03 5.42 1346.94 5.89 2334.42
Doubled Fi 7.22 8499 3.93 71.62 3.71 1108.53 3.71 1763.64
Doubled tij 11.13 12348 5.79 119.34 5.63 1569.31 5.92 2486.07
Average 9.89 9749 5.05 88.93 4.50 1346.64 4.83 2123.88
 O. Kaya, B. Urek / Computers & Operations Research 65 (2016) 93–103 103

N ¼200. When we look at the CPU times, we observe that PL has [6] Camm JD, Chorman TE, Dull FA, Evans JR, Sweeney DJ, Wegryn GW. Blending
the highest and SAVNS has the smallest run times. However, OR/MS, judgement, and GIS: restructuring Procter and Gamble's supply chain.
Interfaces 1997;27:128–42.
TSVNS gives better results in profits than SAVNS and we believe [7] Conn R, Cornuejols G. A projection method for the uncapacited facility location
that the increase in profits is worth the small increase in run times. problem. Math Program 1990;46:273–98.
When we look at the performance of GAVNS, we observe that it [8] Daskin MS, Coullard CR, Shen Z-J. An inventory-location model: formulation,
solution algorithm and computational results. Ann Oper Res 2002;110:83–106.
has the highest run times among our metaheuristics but the [9] Fleischmann M. Reverse logistics network structures and design, ERIM Report
additional run times do not grant a higher performance than the Series Research In Management, ERS-2001-52-LIS, Erasmus University Rotter-
others. dam, 2001.
[10] Fuente MV, Rosa L, Cardo M. Integrating forward and reverse supply chains:
application to a metal-mechanic company. Int J Prod Econ 2008;111
(2):782–92.
5. Conclusion [11] Giordano FR, Fox WP, Horton SB. A first course in mathematical modeling, 5th
ed., Pacific Grove, California: Brooks Cole; 1999, pp. 289–295.
In this study, we analyze a closed loop supply chain manage- [12] Hakimi SL. Optimum locations of switching centers and the absolute centers
and medians of a graph. Oper Res 1964;12:450–9.
ment problem focusing on the facility location, inventory and [13] Inderfurth K. Impact of uncertainties on recovery behavior in a remanufactur-
pricing issues. We develop a mixed integer nonlinear mathema- ing environment: a numerical analysis. Int J Phys Distrib Logist Manag
tical model for this problem and propose three different meta- 2005;35(5):318–36.
[14] Jayaraman V, Guide Jr. VDR, Srivastava R. A closed-loop logistics model for
heuristics for the solution of this model. We also find an upper remanufacturing. J Oper Res Soc 1999;50(5):497–508.
bound for the objective function in this model and analyze the [15] Jayaraman V, Patterson RA, Rolland E. The design of reverse distribution
effectiveness of these heuristics through an extensive numerical networks: models and solution procedures. Eur J Oper Res 2003;150:128–49.
[16] Ko HJ, Evans GW. A genetic algorithm-based heuristic for the dynamic
study. We observe that our heuristics perform pretty well in integrated forward/reverse logistics network for 3PLs. Comput Oper Res
general. We observe that all our heuristics can find the optimal 2007;34(2):346–66.
solution in our trials when there are N ¼10 customer locations and [17] Lee DH, Dong M. A heuristic approach to logistics network design for end-of-
lease computer products recovery. Transp Res Part E: Logist Transp Rev
the percentage deviation of the profits obtained by the heuristics
2008;44(3):455–74.
from the upper bound are less than 5% for N ¼200 customer zones. [18] Listes O. A generic stochastic model for supply-and-return network design.
We also analyze the effects of the parameters on the system. Our Comput Oper Res 2007;34(2):417–42.
computational study shows that TSVNS gives better results than [19] Louwers D, Kip BJ, Peters E, Souren F, Flapper SDP. A facility location allocation
model for reusing carpet materials. Comput Ind Eng 1999;36(4):855–69.
SAVNS and GAVNS in terms of the profits obtained. [20] Lu Z, Bostel N. A facility location model for logistics systems including reverse
In the future, an exact algorithm may be proposed for this flows: the case of remanufacturing activities. Comput Oper Res 2007;34
problem to find the optimal solution or more effective heuristics (2):232–99.
[21] Mirchandani PB, Francis RL. Discrete location theory. New York: Wiley; 1990.
can be developed. In addition, capacity limits, multi product [22] Mladenovic N, Brimberg J, Hansen P, Moreno-Pérez JA. The p-median
systems, quality differentiation for returns and uncertainty in problem: a survey of metaheuristic approaches. Eur J Oper Res 2007;179
demand and return rate can be considered in the future as an (3):927–39.
[23] ReVelle CS, Eiselt HA, Daskin MS. A bibliography for some fundamental
extension of this study. problem categories in discrete location science. Eur J Oper Res
2008;184:817–48.
References [24] Salema MIG, Barbosa-Povoa AP, Novais AQ. An optimization model for the
design of a capacitated multi-product reverse logistics network with uncer-
tainty. Eur J Oper Res 2007;179(3):1063–77.
[1] Aras N, Aksen D, Tanugur AG. Locating collection centers for incentive- [25] Salema MIG, Barbosa-Povoa AP, Novais AQ. Simultaneous design and planning
dependent returns under a pick-up policy with capacitated vehicles. Eur J of supply chains with reverse flows: a generic modelling framework. Eur J
Oper Res 2008;191(3):1223–40. Oper Res 2010;203(2):336–49.
[2] Beamon BM, Fernandes C. Supply-chain network configuration for product [26] Schultmann F, Engels B, Rentz O. Closed-loop supply chains for spent
recovery. Prod Plan Control: Manag Oper 2004;15(3):270–81. batteries. Informs 2003;33(6):57–71.
[3] Bertsekas D. Nonlinear programming. 2nd ed.Athena Scientific; 1999. [27] Shih LH. Reverse logistics system planning for recycling electrical appliances
[4] Biehl M, Prater E, Realff MJ. Assessing performance and uncertainty in and computers in Taiwan. Conserv Recycl 2001;32(1):55–72.
developing carpet reverse logistics systems. Comput Oper Res 2007;34 [28] Srivastava K. Green supply-chain management: a state of the art literature
(2):443–63. review. Int J Manag Rev 2007;9(1):53–80.
[5] Brandeau ML, Chiu SS. An overview of representative problems in location [29] Wojanowski R, Boyaci T, Verter V. Retail-collection network design under
research. Manag Sci 1989;35:645–74. deposit-refund. Comput Oper Res 2007;34:324–45.