Author's Accepted Manuscript

A Mixed Integer Nonlinear Programming

Model and Heuristic Solutions for Location,
Inventory and Pricing Decisions in a Closed
Loop Supply Chain
Onur Kaya, Busra Urek

www.elsevier.com/locate/caor

PII: S0305-0548(15)00172-0
DOI: http://dx.doi.org/10.1016/j.cor.2015.07.005
Reference: CAOR3821

To appear in: Computers & Operations Research

Received date: 3 December 2013

Revised date: 27 August 2014
Accepted date: 8 July 2015

Cite this article as: Onur Kaya, Busra Urek, A Mixed Integer Nonlinear
Programming Model and Heuristic Solutions for Location, Inventory and
Pricing Decisions in a Closed Loop Supply Chain, Computers & Operations
Research, http://dx.doi.org/10.1016/j.cor.2015.07.005

This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal
pertain.
 A Mixed Integer Nonlinear Programming Model and
Heuristic Solutions for Location, Inventory and Pricing
Decisions in a Closed Loop Supply Chain
1 2
Onur Kaya Busra Urek
Department of Industrial Engineering, Koc University, Turkey

Abstract

We analyze a network design problem for a closed-loop supply chain that integrates
the collection of the used products with the distribution of the new products. We
present a mixed integer nonlinear facility location-inventory-pricing model to decide
on the optimal locations of the facilities, inventory amounts, prices for new products
and incentive values for the collection of right amount of used products in order to
maximize the total supply chain profit. We develop heuristics for the solution of this
model and analyze the effectiveness of these heuristics and the effects of the parameters
on this system through numerical experiments.

Keywords: Closed-loop supply chain, facility location, pricing, inventory, mixed integer
nonlinear programming, heuristics

1 Introduction

Closed loop supply chain management (CLSCM) includes all reverse logistics activities as
collection, remanufacturing and refurbishing processes in addition to all traditional forward
logistics activities from supplier to customer. CLSCM does not only gain increasing atten-
tion among society, government and practitioners of operations but also gains the attentions
of researchers (see [28] for a review on CLSCM). The growing concern for environment leads
1
Corresponding author: Onur Kaya, Koc University, Eng 206, Sariyer, Istanbul, 34450, Turkey, Tel:
90-212-3381583, Fax:90-212-3381548, okaya@ku.edu.tr
2
Koc University, Sariyer, Istanbul, 34450, Turkey, burek@ku.edu.tr

1
integrating environmentally sound solutions into traditional supply chain management in
practice. Besides the environmental concerns, recent legislations in many countries force
the companies to undertake take-back responsibilities to reduce waste. The governmental
concerns create strong legislative incentives on why and how take back activities are or-
ganized for different classes of products. WEEE directives in Europe will require all large
electronic retailers to take back their old equipment. Companies can also benefit from these
used products to decrease their cost of production. End-of-life products that normally go to
waste might actually have some remaining value which will reduce consumption of the raw
materials and provide cost reduction on waste treatment. In other words, CLSCM is driven
by high profitability, increasing number of legislative incentives, and growing awareness for
environment. Thus, the companies need to adjust their supply chain to integrate the collec-
tion activities of end-of-life products into their traditional distribution channels and design
an efficient CLSC.

In this study, considering the inventory, pricing and incentive determination problems
in the supply chain, we analyze the network design problem for a closed-loop supply chain
that integrates the collection of the used products with the distribution of the new products.
In today’s world, in which the companies are required to collect a portion of their used
products due to legislative issues or cost concerns, they need to integrate their collection
network for used products with the distribution network of new products. In this study, we
present a mixed integer nonlinear facility location-inventory-pricing model to decide on both
the optimal locations of the collection and distribution centers (CDC), optimal inventory
amounts to be carried at these centers, optimal prices for new products and the values of
incentives that need to be offered for the collection of right amount of used products, in
order to maximize the total supply chain profit.

Logistics network design is one of the most strategic decisions for CLSCM because opening
or closing a facility is very expensive, 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 to facility locations. Facility location problems have
been studied for more than a century, starting from Weber’s location problem in 1909 (see

2
Brandeu and Chiu [5] for a review of facility location problems). The simplest and one
of the most studied location problem is p-median problem which is introduced by Hakimi
[12]. This problem analyzes which p facility are to be selected among the potential locations
and Mladenovic et al. [22] give a complete survey about metaheuristic approaches applied
to p-median problem which is classified as NP-hard. Another type of the facility location
models, which is more in line with our study, considers fixed facility location costs, therefore
number of facilities to be established becomes an important decision. This problem is known
as uncapacitated facility location problem (UFLP) and is extensively studied in literature.
Revelle et al. [23] survey a number of the important problems in facility location including
UFLP. Moreover, an extensive study can be reached in Mirchandani and Francis’s book [21].

Collecting the right amount of products is one of the main concerns in a CLSCM and
there are two main factors that effect the collection amount; the incentives offered to the
customers and the proximity of collection centers to customers. Firstly, some companies
offer financial incentives to end users to return their products and as the amount of this
incentive increases, more products will be returned. Amount of this incentive plays a major
role in the collection activity and it must be critically analyzed to maximize the profitability
of the companies. Secondly, accessibility of the collection centers (CC) is another important
issue for the customers to return their products. As the number of CCs increase, they will
be closer to the customers and more products will be returned, however, opening more CCs
might be too costly for the companies. Thus, opening the right number of CCs at the right
places is a critical factor for a CLSCM. In addition, in a society, since the end users are
generally located at the same place as the new potential buyers for these products, the CCs
can be integrated with the companies’ traditional distribution centers (DC) to carry out both
distribution and collection activities to decrease both facility opening and operating costs.
Thus, we aim to determine the best locations for these collection and distribution centers
(CDC) in order to maximize the total supply chain profit considering both the forward and
the backward flow of products.

In addition to the incentives offered to the end users, we also focus on determining the
sales price of new products in order to maximize total profit, which will also depend on the

3
locations of the CDCs. In literature, there are only a few studies that consider pricing with
facility location decisions and they mainly focus on pricing issue while designing an optimal
drop-off facility network. Wojanowski et al.[29] concentrate on the forward supply chain
and using continuous modeling approaches, they determine the sales price of the product
by focusing on the deposit-refund policy to maximize firm’s profit. In this policy, the sales
price is known by the customers including the deposit amount which will be paid back when
customer takes back the used product at a collection facility. In that model, customers’
willingness with regards to purchasing and returning the product is related to a stochastic
utility choice model. On the other hand, Aras et al.[1] concentrate on reverse supply chain
and try to optimize only return decision of product holders rather than the purchasing
decision under deposit-refund requirements. Another difference between these studies arises
from their proposed pick-up policies. In Wojanowski et al. [29], customers have to bring
their used products to the collection centers, however, Aras et al. [1] develops a pick-up
policy with vehicles of limited capacity.

In this study, we also consider the inventory costs of the CDCs since inventory costs also
play an important role in the management of these centers. The collected products from the
customers should be sent to the manufacturing facility and new products should be brought
from the manufacturing facility at certain time intervals. Therefore, the inventory system
should be organized and controlled appropriately to provide integrity between the forward
and backward flows of the materials. Since we assume a deterministic rate of demand and a
deterministic rate of collection, we use the economic order quantity (EOQ) formulation which
has been widely used in the network design literature. For example, Camm et al. [6] develop
an uncapacitated facility location formulation for Procter and Gamble Company to locate
the DCs and assign the selected DC’s to customer zones with the objective of minimizing the
total cost, composed of material handling costs, inventory costs and transportation costs,
while maintaining a certain customer service level. Daskin et al. [8] propose a location
inventory model and its solution methodology, considering the locations of DCs with the
working inventory and safety stock costs as well as the economies of scale that exist in the
transportation costs from suppliers to DCs.

4
 Reverse and closed loop logistics network design problems are widely studied in literature.
Louwers et al. [19] study a facility location allocation model for reusing of carpet waste as
a case study and analyze its two applications one in Europe and the other in the United
States of America. Jayaraman et al. [14] examine the electronic equipment recovery closed
loop network design with multi-product capacitated warehouse location MILP to seek the
optimal number and location of remanufacturing facilities while minimizing the total costs.
Fleischmann [9] emphasizes a continuous network design model using a MILP model for
product recovery and develops a heuristic algorithm. Shih [27] develops a MILP model to
design the reverse network flow of computers and home appliances in Taiwan. Jayaramann et
al. [15] study on a MILP reverse distribution model to minimize the total transportation and
fixed costs of opening facilities and develop a heuristic procedure for this problem. Beamon
and Fernandes [2] develop a closed loop supply chain model to determine the location and
sorting capability of warehouses and collection centers as well as deciding the quantity of
flow between the sites. Schultmann et al. [26], Biehl et al. [4], Lu et al. [20], Ko et al. [16]
and Lee et al. [17] are some of the other studies related to reverse logistics network design
problems. Listes[18], Salema et al. [24] and Inderfurth [13] consider randomness in reverse
logistics problems. Aras et al. [1] develop a discrete facility location-allocation model to find
the predetermined number of collection centers and the optimal financial incentive values
for different return types and propose a heuristic procedure using two nested loops to solve
this problem. In a recent study, Salema et al. [25] formulate a multi-period, multi-product
network MILP model for the simultaneous design of a Portuguese glass company’s forward
and reverse networks. Fuente et al.[10] deal with an integrated supply chain management
which combines the new processes taken from reverse supply chain into existing processes of
forward supply chain.

Different from the literature, we develop a model that considers a closed loop supply chain
in which we analyze the location, inventory and pricing decisions simultaneously. We build
a MINLP model for this problem, however due to the complexity of the problem, it is not
possible to solve this model using commercial solvers such as BARON for even moderately
sized problems (when the number of potential facility locations are 20 or higher). Thus,
we develop three hybrid metaheuristic algorithms to solve our model for medium or large

5
sized instances. Simulated Annealing(SA), Tabu Search(TS) and Genetic Algorithms(GA),
all hybridized with Variable Neighborhood Search (VNS) are compared in terms of both
solution quality and computational time. We also present an upper bound for the objective
of our problem based on piecewise linear approximation of a part of the nonlinear objective
function and compare the results of the heuristics with this upper bound through an extensive
computational study. We also use this piecewise linear approximation technique to develop a
heuristic for our problem. In the next section, we define our problem and present our model
in detail. Then, in section 3, we describe the solution methods and the heuristics that we
develop for this model and then in section 4, we present our computational studies. Finally
in section 5, we conclude our study.

2 Model

In this study, we consider a company that has a production facility at a certain location and
have customers at N different customer zones. The company needs to decide on where to
open retail facilities (stores) to sell their products and these facilities will also act as collection
centers of the used products. The used products will be collected at these stores and sent
back to the production facility for remanufacturing. We do not differentiate between the
remanufactured and original products that are sold at these stores. We consider M different
potential store locations and assume that each opened facility acts both as a collection and
a distribution center. V1 denotes the set of potential facility locations and V2 denotes the set
of the customer zones. We let c denote the production cost of a new product, p denote the
sales price for new products and r denote the financial incentive (refund) offered for each
returned product. Even though the returned products can be of different quality levels, we
let s denote the expected value of a returned product for the company, which will be used
in the remanufacturing process. In our model the used products are collected by the opened
facilities and transported back to the production facility for remanufacturing/recycling.

We let Kj denote the number of people in zone j ∈ V2 and depending on price and
accessibility of the facilities, we assume that a portion of these people will buy a new product

6
from an open facility. In our model, we assume that the customers go to the retail stores
themselves to buy the product and if the price of the product decreases or if a facility is
opened at a closer location to the customers in a certain zone j, more of the customers in
that zone j will buy the product. Then, the demand from zone j to a facility at zone i
will be Dij = Kj e−kp xij αij where k is the coefficient for price sensitivity of demand, xij is
a binary variable such that it takes the value 1 if the customers in zone j go to the facility
in zone i ∈ V1 , and 0 otherwise; and αij is a parameter between 0 and 1, depending on
the distance between zones i and j, denoting the portion of people who will go to zone i
from zone j to buy a product. Note that in this model for every j, Dij will be positive for
only at most one i and Dij will be zero for all other i since we use a binary variable xij .
Thus, for all j, Dij will be zero if the customers in zone j do not go to zone i to buy the
product (i.e. if xij = 0). In our model, the end of life product owners (ELPO), who are the
customers of the forward supply chain, also transport the products from their locations to
their assigned facility. Therefore, the tendency of customers returning the used products is
also associated 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 = Kj (1 − e−br )xij βij
where b is the incentive sensitivity of the collected amount and βij is a parameter between 0
and 1, depending on the distance between zones i and j, denoting the portion of customers
who will go to zone i from zone j to return a product. Note that, in our model, the customers
will go to the open facilities themselves in order to buy or return a product and thus we do
not consider any cost of transportation for the company between facility nodes and customer
nodes in the objective function of our model. Instead, the decision regarding whether to open
a facility or not will affect the accessibility of the stores and thus affects the demand for new
products or the collected amounts of end-of-life products at any customer node i through
the values of xij αij and xij βij since an open facility at a closer location to a customer node
will have a higher αij or βij value with the corresponding xij = 1.

In our model, when a returned product is dropped to a CDC, it will not be economical
to send that product immediately back to the production facility. To integrate the collection
and distribution networks more efficiently, we assume that when a shipment of new products
is made to a CDC from the production facility, the end-of-life products that are accumulated

7
at that CDC are collected and taken back to the production facility. We let Ai denote the
cost of transportation of the products from the facility at node i to or from the production
facility. The interarrival times of the shipments made to a CDC at zone i, Ti , is effected
by the demand and collection rates at that CDC as well as the fixed shipment cost for the
CDC at zone i, Ai , and inventory holding costs, hpi and hri , for the new and used products,
respectively. We use the well known EOQ model to determine the optimal Ti values such
that the total fixed shipment costs and total inventory holding costs are minimized. The
notation that we use in this study is summarized below with the MINLP formulation to
maximize the total profit, by determining the optimal number and location of the CDCs as
well as the amount of price and incentive offered.
Parameters:
Kj : number of people in zone j
Dij : demand of customers in zone j∈ V2 that go to the facility i∈ V1
Bij : amount of returned products from zone j∈ V2 to the facility i∈ V1
s : amount of gained from unit used product
c : cost of producing a new product
Fi : fixed cost of a CDC located at candidate site i∈ V1
k : parameter denoting the price sensitivity of demand
b : parameter denoting the incentive sensitivity of return amounts
Ai : fixed ordering cost of a CDC located at candidate site i∈ V1 from the factory
hpi : unit holding cost of new product in the CDC located at candidate site i∈ V1
hri : unit holding cost of used product in the CDC located at candidate site i∈ V1
Decision variables:
In the model, we have financial, inventorial and location decision variables as follows:
p : optimum price value offered for unit product
r : optimum incentive value offered for unit used product
Ti : inventory cycle time for CDC located at candidate site i∈ V1
⎧
⎨ 1 if the CDC is opened at candidate site i∈ V
1
yi = (2.1)
⎩ 0 otherwise

8
 ⎧
⎨ 1 if customer j V is served by the CDC located at candidate site i∈ V
2 1
xij = (2.2)
⎩ 0 otherwise

Then, we can write the mathematical model (named Model 1 afterwards) as follows:
Model 1:
  
Π = max (p − c)Kj e−kp xij αij + (s − r)Kj (1 − e−br )xij βij − Fi yi
iV1 jV2 iV1 jV2 iV1
 Ai  (2.3)
−kp hpi −br hri
− [ + ( [Kj e xij αij Ti + Kj (1 − e )xij βij Ti ])]yi
iV
Ti jV
2 2
1 2

subject to:

xij = 1 ∀j ∈ V2 (2.4)
i∈V1

xij ≤ yi ∀i ∈ V1 , ∀jV2 (2.5)

The objective function (2.3), labeled as Π, consists of mainly two parts to maximize the
total profit. The first part calculates the sum of the net profit from selling new products and
the value of collected end of life products. The net profit is found by multiplying the amount
of total customers’ demand with profit per unit (p − c) and amount of total return with net
unit gain. The amount of payment given by the company as incentive is discounted from
the value of used products and net unit gain of company becomes (s − r). In our model,
we use the function e−kp to denote the effect of price on demand, however, other types of
functions such as linear functions can also be used to model the effect of price on demand
and similar results can be obtained. Likewise, the proportion of end-of-life products that is
collected is modeled using a function (1 − e−br ) as the collection increases with r. Besides
the pricing outcome, considering the proximity of the customers in zone i to a CDC at zone
j, demand and return amounts are affected by αij and βij parameters. We think of αij and
βij parameters as a decreasing function of the distances between the zones i and j that will
denote the portion of customers that will go to zone j from zone i to buy or return a product.

The second part of Π (2.3) denotes the opening, operating and inventory costs of the
facilities. Fi yi denotes the opening and operating cost of facility i per unit time and the
remaining part denotes the fixed order and inventory holding cost of facility i per unit time.

9
We note that the classical EOQ model and its related cost function is used in this model.
Ai
Ti
denotes the fixed shipment cost per unit time and the remaining portion denotes the
inventory holding cost per unit time at zone i if a CDC is opened there.

3 Solution Methodologies

The proposed model can be considered as a combination of three subproblems: location,

pricing and inventory. Since the basic facility location problem is known to be NP-hard,
with additional inventory and pricing decisions, our model is also NP-hard. In order to
solve this problem, we first analyze the inventory decisions in this model. We observe that
for any given p, r, yi and xij decisions, the optimal value of Ti can be found through the
classical economic order quantity (EOQ) results in inventory theory as stated in the following
proposition.

Proposition 1. If a facility i is opened, i.e. if yi = 1, then the optimal cycle time between
inventory replenishments at this facility will be as follows.

2Ai
Ti =  p (3.1)
jV2 [(Kj e
−kp )xij αij hi + (Kj (1 − e−br ))xij βij hri ]

We omit the proof for the above proposition since it is 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 above equation in the model, the objective function for our second model
(called Model 2 afterwards) will be as follows:
  
max (p − c)Kj e−kp xij αij + (s − r)Kj (1 − e−br )xij βij − Fi yi
iV1 jV2 iV1 jV2 iV1
  

p
(3.2)
−
2Ai r
(Kj e−kp xij αij hi + Kj (1 − e−br )xij βij hi )
iV1 jV2

Observe that the above objective function is still nonlinear and this problem is a mixed
integer nonlinear program (MINLP). We note that the optimal p and r values can not be

10
written explicitly in terms of the other variables as we did for Ti due to the structure of the
objective function 3.2.

Exact solution algorithms for this type of problems are analyzed in some papers (e.g.
Conn and Cournejols [7]). However, solving this problem with exact methods, even for
moderately sized problem instances is seen to be very difficult. We first try to solve this
model using a commercial MINLP solver, BARON 9.3.1 in GAMS and we could obtain the
optimal results when the number of potential facility locations are small as given in Table 1
in the numerical results section. However, when we increase the number of potential facility
locations, it became impossible to obtain acceptable results with the commercial solver in a
reasonable time amount.

Since it is very difficult to solve this model with exact methods, we develop heuristics to
find near optimal solutions in a reasonable time. In order to analyze the effectiveness of our
heuristics, we also find an upper bound for our problem using a relaxation of the original
problem. We first describe our approach to obtain the upper bound and our first heuristic
named Piecewise Linearization (PL) heuristic in the next subsection. Then, we describe
three different metaheuristics developed for this problem in the later subsections.

3.1 Development of an Upper Bound and Piecewise Linearization

Heuristic

After the optimum cycle time of each candidate CDC are restructured in terms of price
and incentive values as stated in Proposition 1, we observe that the part of the objective
 √
function related to the inventory decisions, which is the part stated as iV1 Wi where
 
 −kp p −br r
Wi = 2Ai jV2 (Kj e xij αij hi + Kj (1 − e )xij βij hi ) , is one of the main parts that
makes the problem difficult to solve. The square root function together with the minus sign
in front of it disturbs the concavity of the objective function making the problem difficult
to solve. For this reason we propose applying a linearization of the square root function in
order to obtain a concave objective function at the end.
√
Since x is a concave function, we use piecewise linear lines that lie below this function,

11
so that the resulting value will be a lower bound for the inventory cost and the objective
function will be higher than the original one leading to an upper bound for our problem. For
piecewise linearization of the square root function, we define a new index set t = 0, 1, 2, ....T
to define the break points. Then, slope of the line in every segment is calculated and 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
linearization model of the square root function.

Parameters:

āi = 2Ai jV2 (Kj αij hpi + Kj βij hri ): maximum possible value of the function inside
the square root term, also denoting the last break point
ait = āi t/T : the break points for piecewise linearization, for all t=0,1..T
√ √
ait − ai,t−1
Δit = ait −ai,t−1
: the slope of the linear approximation function in segment t
Variables:
ωit : non-negative variable denoting the amount of the function value that lies inside

segment t=1,2..T. Note that Tt=1 ωit denotes the whole function value as stated in
constraint 3.6 below.
γit = 1 if segment t is totally covered by the function value, and 0 otherwise
i.e. γit = 1 if ωit = ait − ai,t−1 and γit = 0 otherwise.
Using the above variables and parameters, the resulting model (called Model PL after-
wards) will be as follows:

   
T
−kp −br
max (p − c)Kj e xij αij + (s − r)Kj (1 − e )xij βij − Fi yi − ωit Δit
iV1 jV2 iV1 jV2 iV1 iV1 t=1
(3.3)
subject to:

xij = 1 ∀j ∈ V2 (3.4)
i∈V1

xij ≤ yi ∀i ∈ V1 , ∀jV2 (3.5)


T 
ωit = 2Ai (Kj e−kp xij αij hpi + Kj (1 − e−br )xij βij hri ) ∀i ∈ V1 (3.6)
t=1 j∈V2

ωit ≤ (ait − ai,t−1 )γi,t−1 ∀i ∈ V1 , ∀t = 2..T (3.7)

12
 ωit ≥ (ait − ai,t−1 )γit ∀i ∈ V1 , ∀t = 2..T (3.8)

In this model, we approximate the term inside the square root function with piecewise
linear lines. The last term in the objective function denotes the approximate value for the
function inside the 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 3.7 satisfies the condition that
if γi,t−1 = 0, meaning that the value of the function does not cover segment t − 1 totally,
then ωit should also be 0. Constraint 3.7 also gives an upper bound for ωit . Constraint 3.8
satisfies the condition that if γit = 1, meaning that the value of the function covers segment
t totally, then ωit should be equal to the whole length of segment t, which is ait − ai,t−1 .

After the piecewise linearization of the square root function, the objective function be-
comes a jointly concave function of p and r given the other variables. The approximation
with this piecewise linearization depends on the number of segments, T , used in the piecewise
linearization procedure. As the number of segments increases, the error resulting from the
approximation decreases, but the computational complexity increases with the number of
segments used. Therefore, we consider the trade off between complexity and allowable error
range and according to our preliminary experiments, we use T = 10 equally spaced segments
in the piecewise linearization procedure.

We solve this problem using the solver BARON in GAMS after the piecewise linearization
of the model and we obtain an upper bound for our problem and call it ZU B . In addition,
we look at the resulting solution and use the values of the decision variables as a heuristic
solution for our problem, naming this heuristic as Piecewise Linearization (PL) heuristic.
We recalculate the original objective function value by plugging in these decision variables
into the original objective function since this solution is a feasible solution for our problem
and we call the resulting objective function value ZP L , which is equal to:
  
ZP L = max (p − c)Kj e−kp xij αij + (s − r)Kj (1 − e−br )xij βij − Fi yi
iV1 jV2 iV1 jV2 iV1
   (3.9)

p
−
2Ai r
(Kj e xij αij hi + Kj (1 − e )xij βij hi )
−kp −br

iV1 jV2

We observe that if p and r values are given as parameters into the model, this model turns

13
into a mixed integer linear program (MILP), which is much easier to solve than the MINLP
model which includes nonlinear terms. In addition, we can use the solver CPLEX to solve
this MILP model instead of using BARON. However, note that the previous models without
the piecewise linearization (Models 1 and 2) are still nonlinear programs even for given p
and r values due to the square root term in the objective function and thus we can not use
the CPLEX solver for those models. Thus, as an alternative solution methodology in this
case, we do a two-dimensional search over p and r values and for each given combination of
(p, r), we solve the above model as a MILP model using the solver CPLEX in GAMS. We use
the two dimensional golden section search method (see Giordano et. al [11], pages 289-295,
for the implementation of the golden section search method) to search for the best p and r
combination since we observe through our initial experiments that our objective function is
unimodal in terms of p and r. In this algorithm, for any given p and r combination, we run
the above model and determine the optimal 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 through
the model. We know that r < s and p > c, thus s forms an upper bound for r and c forms a
lower bound for p in our search for the optimal p and r values and the use of these bounds
decreases the solution time.

3.2 Metaheuristics

In addition to the P L heuristic, we also develop three heuristics for the solution of this
problem. All the heuristics that we use in this study are composed of two nested loops. In
the outer loop in our heuristics, we use metaheuristic methods to find the best locations
of CDCs that will be opened to maximize the profit and in the inner loop, the best price
and incentive values are searched through a gradient search algorithm. In the outer loop,
we hybridize three well-known metaheuristic methods which are Simulated Annealing (SA),
Tabu Search (TS) and Genetic Algorithm (GA) with Variable Neighborhood Search (VNS)
as a solution procedure and name them as SAVNS, TSVNS and GAVNS. We apply all three
heuristics in our study and compare their performances in our test instances. Since the
gradient search algorithm that we use for the price and incentive values in the inner loop is

14
straightforward and can be commonly used in many textbooks (e.g. see Bertsekas [3]), we
focus on the metaheuristics SAVNS, TSVNS and GAVNS, that we use in the outer loop in
the next subsections.

In all the proposed algorithms, the optimum locations of opened CDCs (yi ) are repre-
sented with a binary vector in which the ith entry shows whether facility i is open or not. For
example, the representation of a solution with the vector [1, 0, 0, 1, 0, 0, 1, 0, 1, 0] denotes
that the CDCs located at 1, 4, 7 and 9 are selected to open. Then, using these yi values, the
assignments of customer zones to open facilities are represented with another vector x. For
the same example, the vector x=[1 1 7 4 4 9 7 9 9 9] denotes that the customer zones 1 and
2 are assigned to facility 1, 4 and 5 are assigned to facility 4, 3 and 7 are assigned to facility
7 and 6, 8, 9 and 10 are assigned to facility 9. We also note that, for each solution found
by these metaheuristics for yi and xij values, we then apply the gradient search algorithm
at each iteration to determine the best p and r values for those solutions and calculate the
objective function value accordingly. The implementations of the metaheuristics are detailed
below.

3.2.1 Proposed SAVNS Algorithm

In this study, we hybridize simulated annealing with a variable neighborhood search (VNS)
algorithm. At each iteration, within a defined neighborhood structure, the best solution is
sought and when the algorithm stops for that neighborhood, i.e termination conditions are
met, another neighborhood structure is created. The implementation of the algorithm is
detailed as below.

Initial Solution

For the proposed model, to obtain an initial solution to start the algorithm, we generate m
different solutions for facility location array y randomly and then pick the best resulting one
as our initial solution. We assign the customers to their nearest open facilities to determine
the initial vector x for each y. The value of m is determined after several trials in our
numerical experiments and we use m = 50 in our computational study.

15
 Neighborhood Structure

The neighboring solutions are generated by applying combinations of add, drop or swap
moves to the current solution. The proposed SAVNS has three neighborhoods to visit differ-
ent solution spaces. In the first neighborhood, as a move operator, we apply 1-0 exchange to
current solution. In other words, an add move which opens node i if the node is not opened
or a drop move if the node is opened is assigned to the current solution. By changing the sta-
tus of ith coordinate, n different solutions are visited. The second neighborhood is composed
of changing status of two randomly selected coordinates simultaneously, i.e applying twice
1-0 exchange in the same iteration. Neighborhood is generated by applying predetermined
number of 1-0 exchanges to two randomly selected nodes. Finally, in the third neighborhood,
status of randomly selected three nodes are changed by applying 1-0 exchange to them.

We apply similar neighborhood structures for the array x. Every time the vector y is
changed, we initially assign the customers in zone j to the nearest open facility i and create
an initial vector x for that y. Then, as a first neighborhood structure, for each j, we change
its assignment value to another open facility i. In the second neighborhood structure, we pick
two different customer zones j and change their assignments with each other. Additionally,
in the third neighborhood structure we randomly select three customer zones j and change
their assignments with each other.

Cooling Schedule

We tried various cooling schedules such as linear, exponential and hyperbolic ones. In
the equations below, i denotes the cycle that the algorithm runs within, ti denotes the
temperature at cycle i, to denotes the initial temperature and N denotes the maximum cycle
number. For linear cooling schedule, the temperatures are updated according to the relation
ln(to −tN )
i(to −tN )
ti = to − N
. Similarly, for exponential cooling schedule ti = to − i ln(N) and for
to −tN
hyperbolic cooling schedule, the relation ti = 2
(1 − tanh( 10∗i
N
− 5)) + tN is used.

Termination Criteria

In the proposed algorithm, we use two different stopping criteria. The first one is the
maximum number of iterations allowed in a neighborhood. This number is adjusted accord-

16
ing to the size of the neighborhood. The second one is the maximum number of allowable
iterations that incumbent solution does not improve. Both criteria are used at the same time
and reaching one of them is enough to terminate the current neighborhood and to move to
the next one. In each neighborhood, the allowable number of non-improved solutions are
determined again with respect to the size of the neighborhood. Non-improved solutions are
accepted to allow diversification and escape from local maxima. However, new neighborhood
begins with a solution that is the best solution in the previous neighborhood up to that time.
In other words, best found solution is also kept in each neighborhood when allowing to move
towards a worse solution.

3.2.2 Proposed TSVNS Algorithm

We combine TS algorithm with VNS in our proposed algorithm to help the solutions not to
get trapped at a local optima and allow it to visit larger neighborhoods with systematical
change. We use deterministic search to expand neighborhoods by finding the best neighbor of
incumbent solution in a sequential order. Basically, VNS jumps from the incumbent solution
to a new one if and only if a better solution has been found and does not forbid any moves.
However, Tabu Search employs tabu restrictions to prevent visiting earlier selected solutions
throughout the iterations. At each iteration, the best solution which has the maximum profit
is sought by applying local search within allowed set in the current neighborhood. Then, the
best solution is added to the tabu list and one of the solutions that were already in the tabu
list is removed in a FIFO order. Thereafter, the search continues around the best solution.
When the termination criteria are met, the algorithm stops and the best solution found so
far within the previous neighborhood is moved to the next neighborhood in the same way
as SAVNS.

Initial Solution

For TSVNS, initial solution is again constructed as in SAVNS.

Neighborhood Structure

To find the best neighborhood structure, we compare four different neighborhood struc-

17
tures for the array y in this case. In the first neighborhood, as in SAVNS algorithm, 1-0
exchange is applied to current solution and visits all possible neighbors. The opened facility
number increases or decreases by one with the help of 1-0 exchange. The second neighbor-
hood structure is composed of 1-swap moves. 1-swap move relocates a closed facility with an
opened one in the current solution. All pairs of candidate locations that generates a different
solution from current one is visited through the local search to change the sequence of the
facility positions. The possible neighbors that provide the same order, i.e. if both of the
candidate locations in a pair are opened or closed, are prevented to visit in the proposed
algorithm. By this way, each solution in a neighborhood assures a different combination
of the opened CDCs. In the third neighborhood, 3-swap moves relocates the 3 consecutive
facilities. All 3 combinations of current solution are visited in this neighborhood, therefore
this neighborhood size is much bigger than the second neighborhood. By the 3-swap moves,
the number of opened facilities remains constant, however, locations of the opened facilities
change. In the last neighborhood structure, 1-0 exchange is applied twice to the current
solution. By this way, unlike the others, number of opened facilities can increase or decrease
by two. We also note that for the vector x, we apply the same neighborhood structures as
we use in SAVNS.

Tabu definition
When a facility is involved in obtaining a new solution which can either be added or removed,
named as facility i, ith node position can not be changed again for a certain number of
iterations, i.e. labeled as tabu moves and will not be allowed for a certain number of iterations
such as the size of tabu tenure. Using dissimilar neighborhood structures result in keeping
tabu restrictions in different ways. In the 1-0 exchange kind of moves, keeping the one node
position, the exchanged, is sufficient as tabu move. In the second kind of neighborhood, a
pair of different positioned coordinates are swapped. The coordinates resulting in the best
solution in the whole neighborhood are labelled as tabu moves and in tabu list, the pair of
coordinates are kept. Therefore, the 1-swap move using the coordinates used previously can
not be applied until they are non-active in tabu list. We use the recency-based memory in
the TS heuristic which forbids moves towards the most recent visited solution. Size of tabu
tenure is also selected according to the problem size.

18
 Aspiration Criteria
We use the classical aspiration criteria of allowing a move even if it is tabu, if it results in a
solution which provides a profit value better than the incumbent one.

Termination Criteria
Like SAVNS, we use two termination criteria at same time as maximum number of iterations
and the maximum number of allowable non-improved solution.

3.2.3 Proposed GAVNS Algorithm

In this algorithm, each parent, or actually named as chromosome, represents a solution of

the problem. Combination of the genes form the chromosomes and N combinations of the
chromosomes generate a population. Biological evolution operators, mainly crossover and
mutation are used to create a new generation from current population. Selection of new
population depends on their fitness values which are calculated upon the objective values of
the chromosomes. Integrating the VNS into GA 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 algorithms, in each iteration within a
defined neighborhood structure, we firstly generate the offsprings (i.e new generation) using
genetic operators. Then, using fitness values of the chromosomes a new population is selected
deterministically. When the termination criterion of one neighborhood is met, the algorithm
jumps to next neighborhood to diversify and explore the search space. Neighborhoods are
designed in nested strategy and enlarges as proceeded to the end of algorithm.

Initial Population Generation

A population of arrays y denoting the open facilities are created randomly from candidate
solutions. For each y, we also create a population of matrices x denoting the assignments of
customers to open facilities. Population sizes for x and y are determined according to the
size of the problem at hand.

Crossover Operators

Instead of move operators in SA and TS, GA uses crossover and mutation operators.

19
We use three different crossover operators to visit different solution spaces and diversify the
search for the arrays y and x. In the first neighborhood, uniform crossover mask is generated
randomly as the same length of the chromosomes. Parts of chromosome or namely parents,
to be exchanged is determined by the mask. The offspring 1 is produced by taking the bit
from parent 1, if the corresponding mask bit is equal to 1, or take from parent 2 if the
corresponding mask is equal to 0. After the crossover, two offsprings will be created from
mated parents. In the second neighborhood, 1-point crossover is applied to mated parents
in which parents are cut from a randomly selected position and their chromosome strings
are swapped from their tails. By this way, two offsprings are generated and partly similar
to their parents. In the third neighborhood, 2-point crossover is used. The pair of parents’
chromosomes are cut from the randomly selected two points and the middle segments of the
parents’ strings are swapped. Different crossover probabilities are tried during the tuning.
If crossover probability of the mated parents is less then the set value, in other words, if
crossover is not applied, the selected parents are duplicated in order to generate the two
offsprings.

Mutation Operators

Like crossover operators, we apply different mutation operators in arrays y and x. After
applying crossover, some genes of offsprings are altered, with small probability (i.e.mutation
probability), in order to diversify the search space. For each offspring, a mutation probability
is created randomly and we determine whether it undergoes a mutation or not. In the
first neighborhood, which was described above for crossover operators, a random number is
generated between 1 to n and that much of randomly selected gene status are changed to
another feasible value. In the second neighborhood, inversion mutation technique is applied
to a chromosome which will be mutated. In the third neighborhood, head and tail of the
chromosome is changed from a randomly selected gene position. The mutation probability
of the neighborhoods increases so as to diversify the search space by the end of algorithm to
escape from local maximum.

Formation of the next generation

After the generation of mating pool, the chromosomes to compose the next generation

20
are selected according to their fitness values. We use the roulette wheel selection(RWS)
method for the selection of chromosomes in the next generation. In this process, fittest
individuals will tend to have a greater chance of survival than weaker ones and form the
mating pool for the next generation. In other words, fittest individuals has larger share of
the roulette wheel. The individuals are sorted according to their fitness values and based
on the randomly selected point, an individual is selected to form the next generation. After
each selection, the probabilities are calculated again in order to provide greatest chance to
fittest individuals among the remaining ones.

4 Computational Studies

In this section, we analyze the performances of proposed models and heuristics in terms of
the solution quality and computation time through numerical experiments. All of the runs
throughout the computational experiments are performed on a workstation with an Intel(R)
Core(TM)2 Duo processor, 2.53 GHz speed, and 4GB of RAM. The nonlinear mathematical
models are solved with the solver BARON 9.3.1 and the mixed integer linear models are
solved with CPLEX 12.2 with GAMS 23.6. All of the metaheuristic algorithms are coded in
C. In section (4.1) we describe how the test instances are generated and parameters are set,
in section (4.2) we present the results of the mathematical models and then in section(4.3),
we present the results of the heuristics and discuss them in detail.

4.1 Data Generation

For computational analysis, we first create a base case instance, considering an electronics
manufacturer with the manufacturing facility at Izmit, Turkey. Figure 1 shows the map of
the Anatolian side of Istanbul and we use N = 10 different districts on this map as the
customer zones in our initial base case experiment. In all our experiments, the customer
zones are also used as candidate locations for potential CDCs (i.e. M = N). We try to
use real life data for the parameters in our model, however it is not possible to determine

21
the exact values for some of the parameters. Thus, for those parameters, we make realistic
assumptions about their values as detailed below.

Figure 1: Districts from the Anatolian side of Istanbul used as N = 10 customer zones

For the parameter Kj , which denotes the number of potential customers in each zone, we
use the populations of the districts given in Figure 1, according to 2013 survey, multiplied by
0.02 assuming 2% of the population are potential customers of this company. We let αij be a
1
decreasing function of the distances between nodes i and j, tij , such that αij = 1+0.1tij
, where
the distances between different zones (tij ) are obtained from Google Maps. Similarly, for the
return coefficient βij , we assume that half of the sales are returned as used products such
that βij = 0.5αij . Fixed ordering and operation costs for each candidate CDCs are generated
considering the rental costs at these different districts and these values range between 5000
and 20000. The fixed ordering costs, Ai , are generated considering the transportation costs
and the distances of these zones to the manufacturing facility and these values range between
500 and 1500 in our study. The cost of producing a product, c, is assumed to be 100 and
the value of a returned product for the company, s, is assumed to be 50. We let the holding
costs, hpi and hri , to be 1% of c and s, respectively, such that hpi = 1 and hri = 0.5 for all
i. Finally, we assume that the parameters for price sensitivity of demand, k, and incentive

22
sensitivity of return amounts, b, are both 0.01.

In addition to the base case instance with N = 10 customer zones and parameters defined
as above, we also generate additional instances in varying problem sizes (with N = 20,
N = 50, N = 100 and N = 200 customer zones). For increasing values of N, we use
additional districts from different cities of Turkey and use the populations of these districts
as Kj and the distances between these districts as tij to calculate the values of αij and βij
according to the equations stated above.

4.2 Analysis of the Mathematical Models

In this section, we analyze and compare the performances of the mathematical models that
we develop in this paper. In Table 1, we present the results of the mathematical models and
their computational times for varying problem sizes. In all our models, we use the solver
BARON 9.3.1 in GAMS and run the models with a time limit of 8 hours. Since we know
that r < s and p > c, we use these values as upper and lower bounds for the variables r and
p, respectively, in all our models. We observe that Model 1 performs very poorly even for
small sized problems. It results in an optimality gap of around 40% in a time limit of 8 hours
for N = 10 customer zones and the optimality gap gets even worse as N increases. When
we look at Model 2, in which the variables Tij in Model 1 are eliminated by writing them in
terms of the other variables, the performance of the model gets better and we can find the
optimal solution for N = 10 customer zones using Model 2 in 437 seconds in our test instance.
However, for N = 20 and N = 50, Model 2 also fails to find the optimal solution in 8 hours
and results in an optimality gap of 16.3% and 55.9% at the end of 8 hours, respectively.
When we consider Model PL, this model can also be solved to optimality in 74.8 seconds
for N = 10 customer zones, but we can not find the optimal solution of this model, either,
for N = 20 and N = 50 customer zones in 8 hours. Model PL results in an optimality gap
of 9.11% and 47.6% at the end of 8 hours, respectively, which are better than the other two
models in terms of the optimality gaps obtained. We note that the solver BARON can not
find any feasible solutions for any of the models in 8 hours when the number of customer
zones becomes 100 or 200, thus we leave those cells empty in Table 1.

23
 Since the mathematical models currently do not perform well due to the nonlinearities in
the models, we try to separate the non-negative variables p and r from the binary location
variables (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 results in the maximum profit. We observe
through our initial experiments that our objective function is unimodal in terms of p and r,
thus we use the two dimensional golden section search method, with a tolerance of 0.01, to
search for the best p and r combination. Since r < s and p > c, we use these values as upper
and lower bounds in our search for the variables r and p, respectively. We use the bounds 0
and s for r and we do our search for p between c and a large enough upper bound that we
choose depending on the problem parameters. We note that a smaller tolerance value can be
used if more sensitive results are needed. In this algorithm, for the searched p and r values,
we run the models and determine the optimal yi and xij values for that combination. Then,
we pick the best p and r combination with the resulting yi and xij values found through the
models as the final solution.

Table 2 shows the average results of 10 runs of the models for 10 different combinations
of given p and r values. We observe that if p and r is given, Model 1 can find the optimal
solution for yi , xij and Ti values in 12 seconds on average for N = 10 customer zones.
However, when N = 20 or more, Model 1 fails to find the optimal solution in 8 hours even
when p and r are given. Also note that, the CPU times given in Table 1 are the average
values for a single (p, r) combination, but we need to run the models for many times to
search for the best p and r combination. Thus, using Model 1 with this approach does not
help us find the optimal solution. However, when we consider Model 2, in which Ti values
are eliminated, it can find the optimal solution even when N = 200 for a given p and r
combination and the run times decrease significantly. However, for N = 200, the run time
for a single run of the model is around 40 minutes on average. Thus, it does not seem
possible to use this model for the search of the best p and r combination since the run time
of the whole algorithm will need days or maybe weeks. Finally, when we look at Model PL,
we observe that for given p and r values, this model turns into an MILP model while the
other two models were still nonlinear programs (MINLP). As a result of this, we use CPLEX
12.2 to solve the MILP model and we observe that the optimal solution for Model PL can

24
be found in all cases and the run times are significantly less than the other models since the
model does not include any nonlinearity anymore. It takes less than 10 seconds on average
to find the optimal solution of this model even when N = 200 and the runtime is less than a
second on average when N is less than 100. The last column in Table 2 shows the total run
time of the algorithm including the search for the best p and r values. We observe that the
run time of the whole algorithm is less than 3 hours when N = 200 and it is much less for
smaller sized instances. We also note that better algorithms than the golden section search
method can be developed for the search of best p and r combination and such algorithms
can decrease the total run time. However, in this study we focus on the models and the
heuristics, and the development of the best algorithm for the search of best p and r values
is out of the scope of this paper.

Table 1: Results of the mathematical models obtained through the solver BARON
Model 1 Model 2 Model PL
Size % Opt. Gap CPU(seconds) % Opt. Gap CPU(seconds) % Opt. Gap CPU(seconds)
10x10 40.4% 28800 0% 189.2 0% 74.8
20x20 76.2% 28800 16.3% 28800 9.11% 28800
50x50 99.5% 28800 55.9% 28800 47.6% 28800
100x100 - 28800 - 28800 - 28800
200x200 - 28800 - 28800 - 28800

Table 2: Results of the mathematical models with given p and r values

Model 1 Model 2 Model PL
Size % Opt. Gap CPU(seconds) % Opt. Gap CPU(seconds) % Opt. Gap CPU(seconds) Total CPU
10x10 0% 12.1 0% 0.08 0% 0.06 57.3
20x20 0.47% 28800 0% 1.22 0% 0.16 147.2
50x50 1.43% 28800 0% 158 0% 1.57 1316
100x100 2.54% 28800 0% 1203 0% 3.65 3422
200x200 15.8% 28800 0% 2498 0% 9.78 9183

4.3 Analysis of the Performances of the Heuristics

In this section, we explain and compare the results of the proposed algorithms in terms
of solution quality and tabulate run times. The first row in Table 3 shows the results of
our initial experiment with the base case data as explained above. In order to understand

25
the effect of the parameters in our model, we also do additional experiments by modifying
the base case instance by doubling only one parameter value at a time as stated in the
tables. Note that in each row of Table 3 all the remaining parameters that are not stated
in that row, are used as in the base case setting. We also analyze problems of different sizes
as NxM=(10x10),(50x50),(100x100) and (200x200). We note that since there exists some
probabilistic structure in the implementation of the metaheuristics, we run those heuristics
for 10 times and report the best result. The CPU times reported in the tables below denote
the total time for those 10 runs. We also tried running the heuristics for more than 10
times but we observe that it did not improve the results significantly and it is not worth the
additional time.

In order to obtain the best performances of the metaheuristics, we make some initial
experiments for parameter tuning of the heuristics by using the base case test instance for
each problem size. Initial and final temperature values, cooling schedule, maximum iteration
and maximum non-improved iteration numbers are the main parameters in the simulated
annealing algorithm. After trying various values on different sized problems, we decide to
set To as 5000 and Tf as 0.01 in this heuristic, since they give the best results in our initial
experiments. As termination criteria, the most efficient and effective results are gained with
max iteration number as 100 and max nonimproved iteration number as 20.

For the tabu search heuristic, we analyze the effect of different tabu tenures on solution
quality and run time using 30 experiments for each problem size and we observe that the
best performances of tenure sizes are achieved when they are 7 for 50x50 problem size and
15 and 30 for 100x100 and 200x200 problem sizes, respectively. For the 10x10 problems, all
tabu 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 does not produce better results and increase the
CPU time redundantly. For the termination criteria, the maximum number of iterations is
set as 50 and the maximum number of nonimproved solutions are set as 5, 5, 10 and 10 for
10x10, 50x50, 100x100 and 200x200 problem sizes, respectively.

For the genetic algorithm, to decide on the best value of crossover probability, pc , the
values of 0,50, 0.65, 0.8, 0.95 and 1.0 are tried for all problem sizes and to decide on the

26
best mutation probability, pm , values of 0.001, 0.01, 0.1, 0.2 and 0.5 are analyzed for all
neighborhoods. We observe that the best performances are achieved in terms of solution
quality and CPU times when pc is 0.95 and pm is 0.2 and thus we use those values in our
subsequent numerical analysis. In addition, maximum iteration number is set as 20 and
maximum non-improved solution number as 10 for 10x10 problem sizes. For other sizes,
maximum iteration number is set as 100 and maximum non-improved solution number is
set as 20. For all problem sizes, we use 3 different population sizes with 20, 40 and 80
chromosomes.

We first analyze a problem with 10 candidate locations for the facilities and 10 customer
zones as stated as the base case instance above. For this 10x10 problem, we were able to
find the optimal solution by solving Model 2 with BARON 9.3.1. In Table 3, we present
the result of the optimal solution as well as the results of the heuristics. We also present
the running times of the mathematical models and the running times of the heuristics. The
first row in Table 3 shows the results for our base case and in the other rows we present the
results when one of the parameters are doubled while all others remain the same as in the
base case.

Table 3: Results of the optimal solution and the heuristics for N = 10 customer locations
Optimal PL SAVNS TSVNS GAVNS
Size test instance Result CPU Result CPU Result CPU Result CPU Result CPU
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
10x10 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

We observe that all the metaheuristics can find the optimal solution in a very short time
in this case when N = 10 and SAVNS has the smallest runtime. When we look at the
effects of the parameters on the results, we observe from Table 3 that when the populations
(Kj ) are doubled, the profits become almost three times the base case. In addition, we
observe that more facilities are opened as demand increases to serve the increased number
of customers. When the fixed ordering costs (Ai ) are increased, we observe that the profits

27
are slightly decreased. Similarly, when the holding cost of products and returns (hpi and hri ,
respectively) are doubled, we observe that the profits decrease slightly and the effect of the
holding cost of new products are higher than the effect of the holding cost of returns. When
the fixed operation costs (Fi ) are increased, less number of facilities are opened and the
profits decrease substantially. Finally, when the distances, tij , are increased, it leads to less
demand and less returns and as a result more facilities are opened and the profits decrease
substantially.

Due to the failure of exact computational methods when solving large sized instances,
we use the heuristics to find approximate solutions. We also use Model PL with given p
and r values and the golden section search method to determine the best p and r values as
explained above to obtain an upper bound for the problem. We analyze the effectiveness
of the heuristics by comparing the results found by these heuristics with the upper bound
value. The deviations of the results obtained by each heuristic from the upper bound and
the run times of the heuristics are stated in Table 4.

We observe from the results presented in Table 4 that all the metaheuristics perform
pretty well and they all find the optimal result in our trials when N = 10. When we look
at larger sized instances, we observe that the deviations increase with the problem size but
TSVNS gives the best results on average. The deviation between the profits obtained with
the best performing metaheuristic and the upper bounds are around 1.58% when N = 50,
2.65% when N = 100 and 4.5% when N = 200, on average. We observe that the heuristic
P L does not perform as good as the metahueristics and the percentage deviation between the
results obtained by P L and the upper bound reach to almost 10% when N = 200. When we
look at the CPU times, we observe that PL has the highest and SAVNS has the smallest run
times. However, TSVNS gives better results in profits than SAVNS and we believe that the
increase in profits is worth the small increase in run times. When we look at the performance
of GAVNS, we observe that it has the highest run times among our metaheuristics but the
additional run times do not grant a higher performance than the others.

28
5 Conclusion

In this study, we analyze a closed loop supply chain management problem focusing on the
facility location, inventory and pricing issues. We develop a mixed integer nonlinear mathe-
matical model for this problem and propose three different metaheuristics for the solution of
this model. We also find an upper bound for the objective function in this model and analyze
the effectiveness of these heuristics through an extensive numerical study. We observe that
our heuristics perform pretty well in general. We observe that all our heuristics can find the
optimal solution in our trials when there are N = 10 customer locations and the percentage
deviation of the profits obtained by the heuristics from the upper bound are less than 5%
for N = 200 customer zones. We also analyze the effects of the parameters on the system.
Our computational study shows that TSVNS gives better results than SAVNS and GAVNS
in terms of the profits obtained.

In the future, an exact algorithm may be proposed for this problem to find the optimal
solution or more effective heuristics can be developed. In addition, capacity limits, multi
product systems, quality differentiation for returns and uncertainty in demand and return
rate can be considered in the future as an extension of this study.

References
[1] Aras N., Aksen D., Tanugur A. G., Locating collection centers for incentive-dependent returns under
a pick-up policy with capacitated vehicles. European Journal of Operational Research, 191(3), pp.
1223-1240, 2008.

[2] Beamon B.M., Fernandes C., Supply-chain network configuration for product recovery. Production
Planning & Control: The Management of Operations, 15(3), pp. 270-281, 2004.

[3] Bertsekas D., Nonlinear Programming, Athena Scientific, 2nd edition, 1999.

[4] Biehl M., Prater E., Realff M.J., Assessing performance and uncertainty in developing carpet reverse
logistics systems. Computers & Operations Research, 34(2), pp. 443-463, 2007

[5] Brandeau M.L., Chiu S.S., An Overview of Representative Problems in Location Research. Management
Science, 35, pp. 645-674, 1989.

29
 [6] Camm J.D., Chorman T.E., Dull F.A., Evans J.R., Sweeney D.J., Wegryn G.W., Blending OR/MS,
judgement, and GIS: Restructuring Procter and Gamble’s supply chain, Interfaces, 27, pp.128-142,
1997.

[7] Conn R., Cornuejols G., A projection method for the uncapacited facility location problem, Mathe-
matical Programming, 46, 273-298, 1990.

[8] Daskin M.S., Coullard C.R., Shen Z-J., An Inventory-Location Model: Formulation, Solution Algorithm
and Computational Results. Annals of Operations Research, 110, pp. 83-106, 2002.

[9] Fleischmann M., Reverse logistics network structures and design, ERIM Report Series Research In
Management, ERS-2001-52-LIS, Erasmus University Rotterdam, 2001.

[10] Fuente M.V., Rosa L., Cardo M., Integrating Forward and Reverse Supply Chains: Application to a
metal-mechanic company. International Journal of Production Economics, 111(2), pp. 782-792, 2008.

[11] Giordano F.R., Fox W.P., Horton S.B., A First Course in Mathematical Modeling, 5th edition, pp.
289-295, 1999.

[12] Hakimi S.L., Optimum locations of switching centers and the absolute centers and medians of a graph.
Operations Research, 12, pp. 450-459, 1964.

[13] Inderfurth K., Impact of uncertainties on recovery behavior in a remanufacturing environment: A

numerical analysis. International Journal of Physical Distribution and Logistics Management, 35(5),
pp.318-336, 2005.

[14] Jayaraman V. , Guide Jr. V. D. R., Srivastava R., A Closed-Loop Logistics Model for Remanufacturing.
The Journal of the Operational Research Society, 50(5), pp. 497-508, 1999.

[15] Jayaraman V., Patterson R.A., Rolland E., The design of reverse distribution networks: Models and
solution procedures, European Journal of Operational Research, 150, pp. 128-149, 2003

[16] Ko H.J., Evans G. W., A genetic algorithm-based heuristic for the dynamic integrated forward/reverse
logistics network for 3PLs . Computers & Operations Research, 34(2), pp. 346-366, 2007.

[17] Lee D.H., Dong M., A heuristic approach to logistics network design for end-of-lease computer products
recovery. Transportation Research Part E: Logistics and Transportation Review, 44(3), pp. 455-474,
2008.

[18] Listes O., A generic stochastic model for supply-and-return network design. Computers & Operations
Research, 34(2), pp. 417-442, 2007.

30
[19] Louwers D., Kip B.J., Peters E., Souren F., Flapper S.D.P., A facility location allocation model for
reusing carpet materials. Computers & Industrial Engineering, 36(4), pp. 855-869, 1999.

[20] Lu Z., Bostel N., A facility location model for logistics systems including reverse flows: The case of
remanufacturing activities. Computers & Operations Research, 34(2), pp. 299-232, 2007.

[21] Mirchandani P.B., Francis R.L., Discrete Location Theory, Wiley, New York, 1990.

[22] Mladenovic N.,Brimberg J., Hansen P., Moreno-Prez J.A., The p-median problem: A survey of meta-
heuristic approaches. European Journal of Operational Research, 179(3), pp. 927-939, 2007.

[23] ReVelle C.S., Eiselt H.A., Daskin M.S., A bibliography for some fundamental problem categories in
discrete location science. European Journal of Operational Research, 184, pp. 817-848, 2008.

[24] Salema M.I.G., Barbosa-Povoa A.P., Novais A.Q., An optimization model for the design of a capacitated
multi-product reverse logistics network with uncertainty. European Journal of Operational Research,
179(3), pp. 1063-1077, 2007.

[25] Salema M.I.G., Barbosa-Povoa A.P., Novais A.Q., Simultaneous design and planning of supply chains
with reverse flows: A generic modelling framework. European Journal of Operational Research, 203(2),
pp 336-349, 2010.

[26] Schultmann F., Engels B.,and Rentz O., Closed-Loop Supply Chains for Spent Batteries. Informs,
33(6), pp. 57-71, 2003.

[27] Shih L.H, Reverse logistics system planning for recycling electrical appliances and computers in Taiwan,
Conservation and Recycling, 32(1), pp. 55-72, 2001

[28] Srivastava K., Green supply-chain management: A state of the art literature review.International
Journal of Management Reviews, 9(1), pp. 53-80, 2007.

[29] Wojanowski R., Boyaci T., Verter V., Retail-collection network design under deposit-refund. Computers
& Operations Research, 34, pp.324-345, 2007.

31
 Table 4: Analysis of the performances of the heuristics
PL SAVNS TSVNS GAVNS
Size test instance %deviation CPU %deviation CPU %deviation CPU %deviation CPU
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
10x10 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
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
50x50 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
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
100x100 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
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
200x200 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

32