Computers & Industrial Engineering 90 (2015) 381–389

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

A location–inventory supply chain problem: Reformulation

and piecewise linearization
Ali Diabat ⇑, Effrosyni Theodorou
Department of Engineering Systems & Management, Masdar Institute of Science & Technology, Abu Dhabi, United Arab Emirates

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

Article history: In this paper, we study a two-echelon inventory management problem with multiple warehouses and
Received 10 May 2014 retailers. The problem is a natural extension to the well-known one-warehouse multi-retailer inventory
Received in revised form 16 May 2015 problem. The problem is formulated as a mixed integer non-linear program such that its continuous
Accepted 18 May 2015
relaxation is non-convex. We propose an equivalent formulation with fewer non-linear terms in the
Available online 27 May 2015
objective function so that the continuous relaxation of the new model is a convex optimization problem.
We use piecewise linearization to transform the resulting MINLP to a mixed integer program and we
Keywords:
solve it using CPLEX. Through numerical experiments, we compare the solutions obtained by solving
Supply chain
Location–inventory
the new formulation using CPLEX with two previously published Lagrangian relaxation based heuristics
Integer programming to solve the original mixed integer non-linear program. We demonstrate that the new approach is cap-
Piecewise linearization able of providing almost the same solutions without the need of using specialized algorithms. This impor-
tant contribution further implies that additional variants of the problem, such as multiple products,
capacitated warehouses and routing, can be added to result in a problem that will again be solvable by
commercial optimization software, while the respective Lagrangian heuristics will fail to solve such vari-
ants or extended problems.
Ó 2015 Published by Elsevier Ltd.

1. Introduction problems is the facility location problem, while amongst

In a highly competitive market, it becomes a necessity for com-
panies and organizations to optimize their supply chain in order to
ensure efficient operations, which will lead to lower costs and
higher customer satisfaction. In order to reduce costs and increase
service levels, effective supply chain strategies should be
implemented, that take into consideration the interactions at the
different stages of the supply chain. Therefore, supply chain man-
agement involves decisions to be made on various levels: strategic
decisions, tactical decisions, and operational decisions (Dolgui &
Proth, 2010; Simchi-Levi, Kaminsky, & Simchi-Levi, 2003).
Strategic-level decisions refer to planning that impacts the
company over a long period of time, such as the location of facili-
ties (warehouses), while tactical-level decisions have a short-term
impact, relative to the strategic decisions, as they determine
factors more frequently, such as transportation and inventory poli-
cies. Finally, the operational-level decisions are performed on a
weekly or daily basis in order to plan the scheduling and routing.
In terms of strategic-level decisions, one of the most important
⇑ Corresponding author. Tel.: +971 2 810 9101; fax: +971 2 810 9901.
E-mail address: adiabat@masdar.ac.ae (A. Diabat).
tactical-level decisions one of the most important is the inventory
policy problem. These problems are interdependent, in the sense
that modifying the number or location of warehouses has an
impact on inventory costs, and similarly a change in inventory
policy impacts assignment decisions and thus also affects
location-related costs. However, these two problems are usually
dealt with independently, which leads to sub-optimal solutions.
Considering the two decision levels in an integrated manner
results in the joint location–inventory problem, which can lead
to improved management of the supply chain, as more information
is utilized and the important interactions between the problems
are captured (Diabat, Richard, & Codrington, 2013). The basic
model addresses the delivery of a single product from a manufac-
turer to warehouses, also named distribution centers (DCs), and
from there to multiple retailers. The DCs can be opened in multiple
locations, which is something that is decided through the model.
DCs and retailers hold working inventory, based on product that
has not yet been requested by retailers or end-customers, respec-
tively. The purpose of the model is to determine the optimal num-
ber of open DCs, the optimal allocation of retailers to DCs, and the
optimal inventory strategies for both DCs and retailers. The objec-
tive is to minimize the fixed ordering, holding, transportation and
facility-related costs. This problem was addressed by Teo and Shu

http://dx.doi.org/10.1016/j.cie.2015.05.021
0360-8352/Ó 2015 Published by Elsevier Ltd.
382 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389
(2004), Shu (2010), Diabat et al. (2013), Diabat, Battaïa, and Nazzal
(2015).
Integration, however, of the location and inventory problem
comes at the cost of higher model complexity, leading to the need
for use of heuristics. Such is the case in the works of Chan, Chung,
and Wadhwa (2005), Jayaraman and Ross (2003), Diabat, Aouam,
and Ozsen (2009), Shiguemoto and Armentano (2010), and
Diabat (2014) who developed metaheuristics such as genetic algo-
rithms, simulated annealing and tabu search to solve similar prob-
lems. Furthermore, Lagrangian relaxation has been implemented
by Chen and Chu (2003), Diabat and Richard (2015), and Eskigun
et al. (2005). Benders decomposition method is used by
Santibanez-Gonzalez and Diabat (2013) for solving a reverse logis-
tics problem.
In the current paper, we focus on the multi-echelon joint loca-
tion–inventory problem, which decides on the locations of ware-
houses, through which a single product will be distributed from
the manufacturer to the retailers. The model assigns retailers to
warehouses and determines the times between orders at the ware-
houses and retailers in order to minimize the operating cost of the
supply chain. We evaluate the performance of the developed tech-
nique by comparing with the results obtained by previous works
by Diabat et al. (2013) and Diabat et al. (2015), who develop
Lagrangian heuristics for the problem. The objective of this paper
is to address the integrated problem in a way that allows for
solving it with the use of commercial software, rather than through
the development of complicated heuristic techniques, which is
prevalent in the literature. The additional benefit of this is that
possible extensions to the model can be made so that the model
remains solvable by commercial software. On the other hand, the
Lagrangian heuristics with which we compare the particular
model, will more likely fail to solve the extended problems.
2. Literature review
Most literature has traditionally considered facility location
decisions and inventory management decisions independently.
The motivation to solve integrated strategic and tactical level prob-
lems in order to improve supply chain management is reflected in
more recent streams of research.
A study of a multi-echelon join inventory–location model that
produces the simultaneous decision of warehouse location and
inventory policies at both the warehouses and retailers was
performed by Diabat et al. (2013). They formulate the model as
a non-linear mixed-integer program and solve it using a
Lagrangian relaxation-based heuristic. An important observation
the authors make is that for certain regions of the parameter
space there is an undoubtable benefit for the integration. Later
work by Diabat et al. (2015) develops an improved Lagrangian
relaxation-based heuristic for the same problem. The approach is
based on reducing the solution space by removing feasible solu-
tions that can’t be optimal before implementing the Lagrangian
relaxation heuristic.
Another interesting approach is that of Teo and Shu (2004), who
structure the location–inventory design problem as a
set-partitioning integer programming model and solve it using col-
umn generation, in order to determine the number of open ware-
houses, their location, the allocation of retailers to these
warehouses and finally the optimal inventory policies for both
entities. Their results demonstrate that moderate size problems
can be solved within reasonable time. In later work, Shu (2010)
develops a simple greedy algorithm for the set-covering problem,
which selects the best set at each iteration, until all elements are
covered by the selected sets. Because of the large number of sets,
the author reduces the sub-problem to a sub-modular function
minimization problem. Results show that the greedy algorithm
can solve large scale problems with satisfactory performance, since
solution errors are within 3–4%.
Daskin, Coullard, and Shen (2002) and Shen, Coullard, and
Daskin (2003) developed a location–inventory model with risk
pooling (LRMP), which is formulated as a non-linear integer pro-
gramming problem that incorporates inventory costs at the DCs.
Risk pooling was also addressed by Vidyarthi, Çelebi, Elhedhli,
and Jewkes (2007), who consider a multi-product two-echelon pro
duction–inventory–distribution system. Their model incorporates
risk-pooling effects by consolidating the safety-stock inventory of
the retailers at DCs. It is formulated as a non-linear mixed-integer
program and it is linearized using piecewise-linear functions. The
authors decompose the problem by echelon using Lagrangian relax-
ation, which provides a lower bound and a heuristic is proposed to
produce overall feasible solutions. Their results prove that the
Lagrangian relaxation provides a tight lower bound, while the
heuristic solution is within 5% of the optimal.
As a final note in this section, one might argue that since there
are several developed heuristic techniques for addressing this prob-
lem, which perform extremely well, why an approach such as the
current one, that allows for solving through commercial software,
may be useful. The reason is that in potential extensions of the
model, techniques such as Lagrangian relaxation will more likely
fail to provide a solution, whereas with the current approach it is
easy and straightforward to incorporate the extension and again
solve the problem with the help of commercial software. Once
again, we highlight this important contribution of the current work.
3. Problem description and reformulation
3.1. Problem description
Given a set of retailers I, and a set of warehouses that can be
located at certain predetermined sites J, the aim of the
multi-echelon joint inventory–location problem is to distribute a
single commodity from a single manufacturer to the warehouses
J and from there to the retailers I. The retailers deal with determin-
istic demands and they hold working inventory, which is the pro-
duct that has been ordered from a warehouse but has not yet been
requested by end-customers. As far as the warehouses are con-
cerned, they order a single commodity from the manufacturer at
regular time intervals and they distribute the product to retailers.
For the warehouses, the working inventory represents product that
has been ordered from the manufacturer, but has not yet been
requested by retailers. In this model, lateral supply among ware-
houses is prohibited, as warehouses are only supplied by the man-
ufacturer, and shortages are not permitted. Note also, that the
existence of the production unit does not imply a three-echelon
model. Rather, it demonstrates that the shipping cost between pro-
duction and retailers is accounted for, in order to aid in the deci-
sion of which warehouses will be opened, as will become evident
in subsequent sections through the model formulation. A graphic
illustration of the described network can be seen in Fig. 1.
At this point it is important to discuss the adopted assumption of
the power-of-two inventory policy (Roundy, 1985; Simchi-Levi,
Chen, & Bramel, 2007). We begin by defining the cycle time for each
echelon, which is the time starting when the work begins upon a
specific request and ends when the item is ready for delivery. Thus,
for example, the cycle time for distribution commences when they
request an item from the manufacturer and ends when this item is
ready to be delivered to the retailer. Applying common production
cycle policy implies that the production cycle times of all end items
are identical. However, adopting a power-of-two policy assumes
that the ratio of the cycle times is a power-of-two number. This
 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389 383

confines the solution space in order to implement effective solution

methods to this problem (Kuhn & Liske, 2014).
As far as cost is concerned, we can identify the following major
components: (i) fixed order cost, the cost of placing an order, (ii)
unit-inventory cost, the cost of holding one unit of product for
one unit of time, (iii) unit-shipping cost, the cost of shipping one
unit of product between facilities, and (iv) fixed location cost, the
cost of establishing and operating a warehouse. The objective of
the model is to decide upon the number of warehouses to open,
the location of these warehouses, the set of retailers assigned to
each warehouse and finally the size and timing of orders for each
facility, with the goal of minimizing the sum of inventory, shipping,
ordering, and location costs and with the condition of satisfying
end-customer demand.

3.2. Problem formulation Fig. 1. Supply chain network.

At this point, before introducing the formulation to the current

problem, we will discuss the single warehouse multi-retailer
1 1 0 ^
model. This comprises of a distribution system with a single hi  di  T i ¼ ðhi þ h 0 Þ  di  T i
warehouse, responsible for providing retailers with a single 2 2
product. The demand of each retailer is known and must be met. 2. T i < T 0 . In this case, there is a specific portion of the warehouse
The warehouse receives orders from the retailers, while it places inventory that is intended for delivery to retailer i. Considering
0
orders to an external supplier. As far as cost is concerned, there that the inventory at retailer i costs a rate of hi and the system
is a fixed cost each time an order is received as well as an inventory inventory a rate of h ^ 0 , we obtain:
0
carrying cost. Both are facility dependent. The objective is to decide
the timing and quantity of retailer deliveries and the replenish-
1 0 1 ^0
ment strategy at the warehouse, in order to minimize the total Hi ðT 0 ; T i Þ ¼ h  di  T i þ h  di  T i
2 i 2 0
costs. Before developing the formulation, we will first present the
notation:
Therefore, the average cost of a power-of-two policy is
Sets given by:
I: set of retailers, indexed by i

Parameters X ki X
þ Hi ðT 0 ; T i Þ
di : demand rate of retailer i i
Ti i
si : unit-shipping cost from the warehouse to retailer i
hi : unit-inventory cost per unit of time at retailer i Thus, the formulation of the single warehouse multi-retailer
ki : fixed-order cost at retailer i problem is as follows:
^ :
h 0
unit-inventory cost per unit of time at the warehouse
X ki X1 
btrn : weight factor associated with transportation costs
min þ ^0  d  T þ 1 h
hi  h ^0  d  maxðT 0 ; T Þ ð1Þ
i i i i
binv : weight factor associated with inventory costs i
Ti i
2 2
Decision variables
Ti: cycle-time of retailer i s:t: T 0 =T i ¼ 2Ni 8i 2 I ð2Þ
T 0: cycle-time of warehouse
Ni 2 Z 8i 2 I ð3Þ

In order to determine the cost of and arbitrary power-of-two

T0 > 0 ð4Þ
policy we define the system inventory at retailer i as the inventory
T i > 0 8i 2 I ð5Þ
at retailer i plus the inventory at the warehouse that will be
delivered to retailer i. Then, echelon holding cost rates will be
As we have already discussed the objective function, we can
h ^ 0 and h0 ¼ hi  h
^0 ¼ h ^ 0 , which essentially means that for the hold-
0 i move on to explaining the constraints. Constraints (2) enforce
ing cost at each echelon we need to subtract the cost of the previ- the power-of-two policy, meaning that the ratio of the cycle times
ous echelon. We will prove that the following holds for the average of the warehouse and each retailer will be a power of two. Con-
cost of holding inventory for retailer i at the warehouse and at straint (3) ensures that the power is an integer number, while con-
retailer i, assume Hi ðT 0 ; T i Þ: straints (4) and (5) simply guarantee positive cycle times.
1 
^0  d  T þ 1 h
^0  d  maxðT 0 ; T Þ
Now we can extend the previous problem and introduce the
H i ðT 0 ; T i Þ ¼ hi  h i i i i problem addressed in the current paper, which is the multiple
2 2
warehouse multi-retailer problem. In this problem, there are sev-
There are two cases to be considered at this point: eral warehouses which can be opened at a number of candidate
locations. Thus, we have two additional decisions to make. The first
1. T i P T 0 . In the power-of-two policy, this inequality practically one is where to position the warehouses and the second one is,
means that the warehouse places an order every time the since there are now more than one warehouses, which retailer will
warehouse does. Thus, the warehouse does not hold inventory be supplied by which warehouse. The following notation is used
for the particular retailer and the average holding cost is: for the formulation.
384 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389

In the above formulation, note that we have defined

Sets    
^ d and e ¼ 1 h
bij ¼ sij þ ^sj di ; cij ¼ 12 hi  h ^ d . The objective
I: set of retailers, indexed by i j i ij 2 j i
J: set of possible warehouse locations, indexed by j function, as explained, aims to minimize the sum of inventory, ship-
ping, ordering as well as location costs, while at the same time
Parameters
meeting end-customer demand. The first set of constraints (2)
di : demand rate of retailer i
enforce that every retailer be assigned to precisely one warehouse.
f j: fixed cost of establishing and operating a warehouse at
Constraints (3) ensure that retailers will not be assigned to closed
location j
warehouses, while constraints (4) and (5) are the integrity
sij : unit-shipping cost from warehouse j to retailer i
constraints. Finally, constraints (6) and (7) ensure that a
^sj : unit-shipping cost from manufacturer to warehouse j
power-of-two inventory policy is maintained for retailers and ware-
hi : unit-inventory cost per unit of time at retailer i houses, and constraints (8) and (9) ensure positive cycle times.
ki : fixed-order cost at retailer i The reason for which we choose to employ the transportation
^:
h j
unit-inventory cost per unit of time at warehouse j and inventory weight factors btrn and binv , respectively, is so that
^:
kj
fixed-order cost at warehouse j we can easily test our algorithm for various operational circum-
btrn : weight factor associated with transportation costs stances. For example, assuming a weight factor of binv ¼ 0 implies
binv : weight factor associated with inventory costs that we have an uncapacitated problem, since the inventory cost
becomes zero. On the other hand, having a value of binv ¼ 100,
Decision variables implies that we place the highest possible importance on inventory
T ij : cycle-time of retailer i when served by warehouse j cost. Thus, the two solutions generated for these two instances
T^ j : cycle-time of warehouse j would be very different and it is in our interest to observe the
changes for a combination of various values for these weight factors.
Binary decision variables

1 if a warehouse is opened at location j 3.3. Problem reformulation
Xj ¼
0 otherwise

1 if a retailer i is served by warehouse at location j In this section the process of linearizing the non-linear elements
Y ij ¼
0 otherwise in the objective function is described, in order to produce a formu-
lation that is easier to solve. The objective function can be rewrit-
Based on the previous problem, we can derive the average cost of ten as follows, with its various terms numbered:
the power-of-two inventory policy for the multiple warehouse 0 1 1 0
zfflfflfflfflffl}|fflfflfflfflffl{ 2
multi-retailer problem. We need to additionally consider the zfflfflfflfflfflffl}|fflfflfflfflfflffl{
XB B b
kj C C X X B
B ki
fixed-order cost at warehouse j, as well as the binary variables min Bf j X j þ binv X j C þ Bbinv Y ij þ btrn bij Y ij
which decide whether or not a warehouse will be opened at a given j2J
@ Tbj
A j2J i2I
@ T ij

location and whether a retailer will be supplied by a given ware-

1
house. Also, we add the inventory weight factor to obtain the final 4
3
zfflfflfflfflfflfflffl}|fflfflfflfflfflfflffl{ zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
n o
form of the cost: C
þbinv cij T ij Y ij þ binv eij max Tb| ; T ij Y ij C A ð15Þ
X k^j XX ki XX 1 h 
^ j T ij þ h
i
^ j maxðT^ j ;T ij Þ di Y ij
binv X j þ binv Y ij þ binv hi  h
j
^
Tj i j
T ij i j
2
The initial step is the linearization of non-linear term 1, which is
Aside from this cost, there is also the fixed cost of establishing a
essentially a piecewise linearization of kb| = Tb| . Assuming we divide
facility:
X Tb| into n ¼ 1; . . . ; N linear pieces, we will have N + 1 break points,
f jXj
which we denote t1ij ; t2ij ; . . . ; t Nþ1
ij . Then we can implement the
j
following approximation:
Finally, there is the shipping cost between the assigned !
warehouses and retailers: kb| kb| n kb| kb| kb| kb|
XX  ’ Pj þ ð1  P nj Þ ¼ þ  Pnj ; ð16Þ
 Tb| tb| n tb| nþ1 tb| nþ1 tb| n tb| nþ1
btrn sij þ ^sj di Y ij
i j for some Pnj 2 ½0; 1. Then, multiplying by X j the following equation
Now we can present the formulation of the joint inventory– is obtained:
!
location problem: kb| kb| kb| kb|
! Xj ’ Xj þ  Pnj X j ð17Þ
X kbj b
T| b
t | nþ1 tb| n tb| nþ1
minX;Y;T;T
^ f j þ binv Xj
Tbj
j2J
Now, let Anj ¼ Pnj X j , which transforms the non-linear term into:
XX ki
þ binv þ btrn bij þ binv cij T ij !
j2J i2I
T ij kb| kb| kb| kb|
n o binv X j ’ binv þ binv  Anj ð18Þ
b
þbinv eij max T j ;T ij Y ij ð6Þ Tb| tb| nþ1 tb| n tb| nþ1
X
s:t: Y ij ¼ 1 8i 2 I ð7Þ subject to the following constraints, where M is a sufficiently large
j2J
number:
Y ij 6 X j 8i 2 I; 8j 2 J ð8Þ
Y ij 2 f0;1g 8j 2 J ð9Þ Anj 6 P nj ð19Þ
X j 2 f0;1g 8j 2 J ð10Þ Anj P P nj þ MðX j  1Þ ð20Þ
c
T | =T ij ¼ 2Nij 8i 2 I; 8j 2 J ð11Þ Anj 6 MX j ð21Þ
X n
Nij 2 Z 8i 2 I; 8j 2 J ð12Þ Pj ¼ 1 ð22Þ
Tbj > 0 8j 2 J ð13Þ n

T ij > 0 8i 2 I; 8j 2 J ð14Þ Pnj P0 ð23Þ

 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389 385

Similarly, term 2 can be linearized as follows, where Bnij ¼ Q nij Y ij : wij Y ij ð36Þ
!
ki ki ki ki With the following constraints:
binv Y ij ’ binv nþ1 þ binv n  nþ1 Bnij ð24Þ
T ij tij t ij t ij wij P Tbj 8i; j ð37Þ
subject to the following constraints:
wij P T ij 8i; j ð38Þ

Bnij 6 Q nij ð25Þ Then 40 can be modelled by letting Eij ¼ wij Y ij which makes term
4 become:
Bnij P Q nij þ MðY ij  1Þ ð26Þ
eij Eij ð39Þ
Bnij 6 MY ij ð27Þ
X n
Q ij ¼ 1 ð28Þ subject to:
n
Eij 6 W ij ð40Þ
Q nij P0 ð29Þ
Eij P W ij þ MðY ij  1Þ ð41Þ
At this point it is worth noting that the reason that a binary Eij 6 MY ij ð42Þ
variable is not required for this linearization process is that the
W ij P Tb| ð43Þ
function 1/T is convex. This can be demonstrated with the help
of Figs. 2 and 3. W ij P T ij ð44Þ
In Fig. 2, the convexity of the function means that the minimum Tb| ; T ij P 0 ð45Þ
two points that are required to describe a linear piece with the Y ij 2 f0; 1g ð46Þ
lowest value will always be on the period. Such is the case in
Fig. 2, where the two points with the lowest value are n and Putting together all the equations for the reformulation of the
n + 1 and they lie on the same line. On the other hand, in Fig. 3, model we obtain the following linear program:
the two lowest points n and n + 4 do not belong to the same line, ! !
X kbJ kbJ kbJ
as the function is non-convex. This creates the need for the use minX;Y;T;T
^ f j X j þ binv þ binv  Anj
of binary variables in order to ensure that the two points that will j2J tbJ nþ1 tbJ n tbJ nþ1
! !
be used lie on the same line. XX ki ki ki n
For term 3, if we let C ij ¼ T ij Y ij , we obtain the following: þ binv nþ1 þ binv n  nþ1 Bij þ btrn bij Y ij þ binv cij C ij þ binv eij Eij
j2J i2I
tij tij tij
cij C ij ð30Þ ! !
X X b
kJ kbJ kbJ X ki ki ki
n
¼ ðf j X j Þ þ binv  Aj þ þ n  nþ1 Bnij
subject to: j2J j2J tbJ nþ1 tbJ n tbJ nþ1 i2I
tijnþ1 tij tij
 XX
C ij 6 T ij ð31Þ þcij C ij þ eij Eij þ btrn ðbij Y ij Þ ð47Þ
j2J i2I
C ij P T ij þ MðY ij  1Þ ð32Þ
C ij 6 MY ij ð33Þ X
Y ij ¼ 1 8i 2 I ð48Þ
T ij P 0 ð34Þ j2J

Y ij ¼ f0; 1g ð35Þ Y ij 6 X j 8i 2 I; 8j 2 J ð49Þ

  Anj 6 Pnj ð50Þ
Finally, for term 4, initially we let wij ¼ max Tbj ; T ij ; 8i; j.
Anj P Pnj þ MðX j  1Þ ð51Þ
Then, one part of the term, namely 40 , becomes:
Anj
6 MX j ð52Þ
X n
Pj ¼ 1 ð53Þ
n

Bnij 6 Q nij ð54Þ

Bnij P Q nij þ MðY ij  1Þ ð55Þ
Bnij
6 MY ij ð56Þ
X n
Q ij ¼ 1 ð57Þ
n

Q nij P0 ð58Þ
C ij 6 T ij ð59Þ
C ij P T ij þ MðY ij  1Þ ð60Þ
Fig. 2. Convex function. C ij 6 MY ij ð61Þ
Eij 6 W ij ð62Þ
Eij P W ij þ MðY ij  1Þ ð63Þ
Eij 6 MY ij ð64Þ
W ij P Tb| ð65Þ
W ij P T ij ð66Þ
Tb| ; T ij P 0 ð67Þ
Nij 2 Z 8i 2 I; 8j 2 J ð68Þ
Y ij 2 f0; 1g 8j 2 J; 8i 2 I ð69Þ
Fig. 3. Non-convex function.
X j 2 f0; 1g 8j 2 J ð70Þ
386 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389

Table 1
Computational gap for varying number of approximation segments.

Problem size 1 Segments 2 Segments 3 Segments 4 Segments 5 Segments 6 Segments

I J GAP GAP GAP GAP GAP GAP
Beta(inv) = 1, Beta(trn) = 1
50 10 58.752 19.188 6.462 0.330 0.317 0.317
50 30 28.336 7.875 2.317 0.157 0.152 0.151
50 50 21.488 7.623 1.265 0.117 0.113 0.112
100 50 48.473 28.851 5.427 0.307 0.289 0.287
100 75 41.886 19.043 4.797 0.244 0.235 0.234
100 100 34.481 19.290 4.392 0.226 0.219 0.219
150 100 34.758 18.946 3.781 0.209 0.198 0.197
150 125 44.686 17.437 4.374 0.288 0.283 0.281
150 150 78.792 38.015 6.661 0.406 0.393 0.391
250 200 44.052 13.138 2.998 0.234 0.231 0.230

Table 2
Results for ðbinv ; btrn Þ ¼ ð0:01; 1Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 0.01, Beta(trn) = 1
50 10 56.6 0.196 1.7 0.359 53.5 0.217 63.7 0.212
50 30 174.6 0.256 3.7 0.355 140.4 0.278 161.5 0.271
50 50 285.0 0.132 4.7 0.177 262.0 0.141 296.1 0.138
100 50 517.7 0.142 9.8 0.238 557.8 0.155 624.7 0.151
100 75 875.4 0.147 14.0 0.215 690.3 0.167 766.2 0.161
100 100 1100.8 0.126 14.3 0.160 815.6 0.130 962.4 0.125
150 100 1683.9 0.180 15.0 0.328 1260.6 0.201 1462.3 0.198
150 125 2372.1 0.209 20.1 0.267 2077.7 0.215 2430.9 0.206
150 150 2625.6 0.137 28.5 0.216 1941.5 0.156 2291.0 0.150
250 200 5124.6 0.264 32.9 0.433 4939.0 0.296 5432.9 0.281

By observing the new formulation it is possible to see that it is
now a mixed integer program (MIP), since the objective function
and constraints are all linear. Furthermore, we notice that the same
number of binary variables remain, but an additional number of
continuous variables are employed. However, this does not pose
a big challenge in terms of solving complexity for the model and
therefore overall the new formulation is easier and more efficient
to solve. As a final remark, note how the only actual approximation
was that of the first term, while the rest are exact reformulations.
4. Computational analysis
In this section we describe the characteristics of the problems
which were solved using ILOG CPLEX 12.5 (IBM, 2013) for the
reformulated problem and we compare the computational results
with those obtained by Diabat et al. (2013) and Diabat et al.
(2015). Tables 2–8 summarize the computational results, where
tests were run for several combinations of the weight factors
associated with transportation and inventory costs, and specifically
for certain combinations of btrn 2 f0:001; 0:01; 0:1; 1; 10; 100g and
binv 2 f0:001; 0:01; 0:1; 1; 10; 100g. The reason why we have
selected these values is to provide a large range of trade-offs
between transportation and inventory costs.
Firstly, it is essential to decide on the number of segments to be
used for the approximation. Table 1 summarizes the computa-
tional gaps for a varying number of segments from 1 to 6. It is pos-
sible to observe that the difference in the gap is huge between 1
and 2 linear segments, while it gradually diminishes between 2
and 3 segments and so on. By comparing the results obtained for
4 and 5 segments we note that there is not a great difference in
the gap, but the computational effort for 5 segments is substan-
tially higher, as depicted in Table 2. As for the difference in the gaps
obtained for 5 and 6 segments it is practically negligible. Based on
these observations, we run our tests for 4 and 5 segments.
In these tables, the first set of columns state the size of the prob-
lem, by providing the number of retailers ðIÞ and the number of
warehouses ðJÞ. The second and third set of columns refer to the
results obtained by Diabat et al. (2013) and Diabat et al. (2015),
respectively, for which the required CPU time and the gap are
reported. The final sets of columns refer to the current approach
when the approximation process employs 4 and 5 linear pieces,
respectively.
As far as the reported gaps are concerned, for Diabat et al.
(2013) the gap is calculated as the difference between the best
upper bound and best lower bound over the best lower bound.
For the work of Diabat et al. (2015) the gap is calculated as the dif-
ference between the newly obtained upper bound minus the lower
bound of Diabat et al. (2013) over this lower bound. The reason for
this is that the sub-problem by Diabat et al. (2013) was solved to
optimality, which means that we are certain that the obtained
bound constitutes a lower bound. On the other hand, the lower
bound by Diabat et al. (2015) was obtained heuristically and this
is why we use the lower bound of the former work instead. For
the current approach, if we observe the piecewise linearization it
is possible to see that the first non-linear term is approximated
from the top, meaning that the result will constitute an upper
bound to the problem. Therefore, the gap is calculated as the differ-
ence between this upper bound and the lower bound obtained by
Diabat et al. (2013) over the lower bound.
As far as the inventory and transportation factors are concerned,
we generally note a declining trend in the gap as the percentage of
the inventory for the transportation factor decreases. For example,
in the first tested problem, depicted in Table 2, the inventory factor
has a value of 0.01, while the transportation factor has a value of 1,
meaning that the inventory factor is 1% of the transportation
 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389 387

Table 3
Results for ðbinv ; btrn Þ ¼ ð0:01; 100Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 0.01, Beta(trn) = 100
50 10 56.4 0.284 0.8 0.417 56.2 0.272 62.4 0.265
50 30 176.0 0.088 2.4 0.176 164.2 0.080 187.2 0.079
50 50 320.2 0.156 4.2 0.200 296.5 0.149 329.1 0.147
100 50 595.7 0.165 4.2 0.273 467.4 0.154 551.5 0.152
100 75 829.1 0.203 7.1 0.319 727.5 0.192 843.9 0.180
100 100 1241.9 0.129 9.1 0.232 859.9 0.120 945.9 0.116
150 100 1590.7 0.177 9.2 0.303 1327.9 0.168 1566.9 0.164
150 125 1895.8 0.221 11.0 0.356 1870.1 0.207 2075.8 0.200
150 150 2391.7 0.114 13.2 0.220 1927.3 0.104 2139.3 0.100
250 200 5041.2 0.093 16.1 0.163 4125.7 0.088 4868.3 0.085

Table 4
Results for ðbinv ; btrn Þ ¼ ð1; 0:01Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 1, Beta(trn) = 0.01
50 10 61.5 0.208 1.0 0.367 48.6 0.332 54.9 0.319
50 30 190.6 0.264 2.5 0.412 152.2 0.392 176.6 0.383
50 50 265.4 0.110 4.9 0.152 221.0 0.163 263.0 0.158
100 50 530.5 0.203 4.6 0.331 489.8 0.315 553.5 0.299
100 75 820.3 0.256 7.1 0.359 779.9 0.398 912.5 0.382
100 100 1078.8 0.270 9.4 0.429 1168.4 0.457 1402.1 0.452
150 100 1740.4 0.189 8.3 0.307 1499.3 0.313 1649.2 0.299
150 125 2186.8 0.154 12.0 0.209 2007.5 0.245 2248.4 0.234
150 150 2848.2 0.241 12.5 0.362 2568.6 0.359 3005.3 0.348
250 200 5267.6 0.116 19.5 0.213 4155.7 0.206 4862.2 0.195

Table 5
Results for ðbinv ; btrn Þ ¼ ð1; 1Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 1, Beta(trn) = 1
50 10 63.5 0.233 0.8 0.366 53.4 0.330 63.5 0.317
50 30 168.6 0.105 3.1 0.176 122.5 0.157 145.8 0.152
50 50 305.2 0.087 5.2 0.160 202.0 0.117 230.3 0.113
100 50 636.1 0.210 5.9 0.347 436.7 0.307 506.6 0.289
100 75 815.9 0.166 8.4 0.261 787.2 0.244 881.7 0.235
100 100 1120.3 0.171 11.6 0.313 1145.7 0.226 1260.3 0.219
150 100 1559.0 0.156 10.8 0.309 1599.1 0.209 1823.0 0.198
150 125 2165.4 0.217 12.9 0.368 2173.1 0.288 2499.1 0.283
150 150 2444.3 0.304 14.6 0.385 2172.5 0.406 2520.1 0.393
250 200 5913.8 0.168 19.8 0.243 4050.6 0.234 4779.7 0.231

factor. In the second problem, illustrated in Table 3, the inventory
factor has a value of 0.01 while the transportation factor has a
value of 100, meaning that the inventory factor is 0.01% of the
transportation factor. Between these two problems, we observe a
general decrease in the reported gap, for most cases.
On the other hand, Table 4 represents the case where the inven-
tory factor is 10000% of the transportation factor and in this case
the gap is generally higher. This can be explained by the fact that
during the reformulation of the problem, the only approximation
is that of 1/T, the first non-linear term in the objective function.
The rest, as mentioned in the previous section, are exact reformu-
lations and not approximations of the non-linear terms. The
approximated term 1/T is multiplied by the inventory factor, which
explains why reducing the inventory factor compared to the trans-
portation factor yields a lower gap. However, this reduction in gap
is not great because 1/T merely constitutes one of the terms, as the
rest also influence the solution. This means that even if our approx-
imation is not strong, this does not greatly affect the solution, fur-
ther contributing to the strength of the approach.
Another observation is that the current approach outperforms
the Lagrangian heuristic of Diabat et al. (2013) in terms of compu-
tational time, but underperforms with respect to the gap. On the
other hand, it achieves a better gap than the heuristic Diabat
et al. (2015), but it requires higher computational time. In this
trade-off however, it should further be considered that the current
approach is solved using commercial software, which means that
possible extensions to the model can be easily implemented and
solved, which is not the case for either of the Lagrangian heuristics.
388 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389

Table 6
Results for ðbinv ; btrn Þ ¼ ð1; 100Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 1, Beta(trn) = 100
50 10 61.6 0.144 1.1 0.214 45.9 0.154 54.6 0.146
50 30 172.1 0.159 2.5 0.236 133.6 0.175 153.6 0.167
50 50 300.8 0.186 4.8 0.266 208.8 0.190 250.6 0.183
100 50 566.6 0.335 6.0 0.446 568.0 0.350 664.6 0.341
100 75 844.6 0.321 8.9 0.417 703.7 0.334 809.3 0.323
100 100 1236.6 0.085 10.3 0.166 1013.6 0.092 1196.0 0.089
150 100 1779.8 0.127 8.1 0.199 1457.0 0.133 1704.7 0.130
150 125 2389.8 0.305 12.6 0.398 1905.6 0.333 2172.4 0.319
150 150 2782.3 0.119 15.1 0.221 2063.7 0.123 2476.4 0.119
250 200 6006.0 0.172 23.0 0.219 4313.5 0.194 5046.8 0.183

Table 7
Results for ðbinv ; btrn Þ ¼ ð100; 0:01Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 100, Beta(trn) = .01
50 10 57.7 0.125 0.9 0.162 58.2 0.233 64.6 0.229
50 30 175.2 0.286 2.9 0.392 162.3 0.587 181.8 0.554
50 50 288.5 0.307 4.3 0.446 220.6 0.785 247.1 0.769
100 50 576.3 0.136 4.1 0.196 552.0 0.333 645.8 0.316
100 75 887.4 0.190 6.0 0.325 827.4 0.472 959.8 0.452
100 100 1033.5 0.245 10.0 0.408 1199.6 0.698 1391.5 0.662
150 100 1670.0 0.223 9.5 0.417 1669.8 0.612 1853.5 0.589
150 125 2349.2 0.243 11.2 0.413 1890.5 0.596 2117.4 0.570
150 150 2782.0 0.204 13.0 0.380 2669.8 0.608 3177.1 0.595
250 200 6283.6 0.291 18.7 0.421 4578.2 0.721 5127.6 0.685

Table 8
Results for ðbinv ; btrn Þ ¼ ð100; 1Þ.

Problem size Diabat et al. (2013) Diabat et al. (2015) Current approach, 4 Current approach, 5
segments segments
I J CPU GAP CPU GAP CPU GAP CPU GAP
Beta(inv) = 100, Beta(trn) = 1
50 10 57.7 0.319 0.9 0.404 42.5 0.472 50.6 0.450
50 30 175.2 0.111 2.9 0.187 138.6 0.168 160.8 0.166
50 50 288.5 0.221 4.3 0.277 233.6 0.337 268.6 0.328
100 50 576.3 0.156 4.1 0.284 520.3 0.262 582.7 0.251
100 75 887.4 0.107 6.0 0.172 632.7 0.192 752.9 0.186
100 100 1033.5 0.197 10.0 0.289 920.9 0.354 1049.8 0.333
150 100 1670.0 0.236 9.5 0.448 1528.3 0.424 1757.5 0.419
150 125 2349.2 0.215 11.2 0.413 2096.8 0.381 2474.2 0.375
150 150 2782.0 0.115 13.0 0.209 2108.7 0.172 2403.9 0.163
250 200 6283.6 0.258 18.7 0.450 4250.0 0.404 4930.0 0.387

Finally, we can overall notice that in every case the calculated
gap for the current approach remains less than 1%, even though
in certain cases such as in that of Table 7, where the inventory
factor is 10,000 times greater than the transportation factor, we
would expect a great increase in the gap. However, this is not
the case and this guarantees that our reformulation is robust. Fur-
thermore, observing the CPU time for the Lagrangian heuristics it is
possible to see that it increases not only when the problem size is
greater, but also when the inventory factor increases. The reason
for this is that the problem is non-linear with respect to the inven-
tory terms. However, in the current approach we observe that the
CPU time increases, only with the increase in problem size and it is
not sensitive to increases in the inventory factor. Overall, the
results are satisfactory and the aforementioned points prove that
the algorithm is potent and reliable.
5. Conclusions
The current paper considered the two-echelon inventory prob-
lem with multiple retailers, some of which can also act as ware-
houses. While the problem was originally formulated as a mixed
integer non-linear program, the purpose of the current paper was
to implement a piecewise linearization in order to transform this
initially non-convex problem into a convex discrete optimization
problem. After obtaining the mixed integer program, the problem
was solved using CPLEX and the results were compared against
 A. Diabat, E. Theodorou / Computers & Industrial Engineering 90 (2015) 381–389
those obtained by two recently published algorithms, which were
developed to solve the same problem.
The computational analysis clearly demonstrates that the cur-
rent approach is robust and reliable and in certain cases outper-
forms the existing approaches. However, the important
contribution is not found in the performance of the developed
algorithm, but in its flexibility to be extended and still remain solv-
able by commercial software. Even though the Lagrangian relax-
ation based heuristics with which we compare the current
approach yield very good results, they are incapable of handling
potential extensions to the problem.
Finally, the current work can be extended in a number of impor-
tant ways. It can focus on incorporating further assumptions, such
as multiple products or a capacitated warehouse (Diabat, in press).
Furthermore, routing decisions can also be integrated in order to
examine a joint location, inventory, and routing problem (Le,
Diabat, Richard, & Yih, 2013; Diabat, Abdallah, & Le, 2014).
References
Chan, F. T. S., Chung, S. H., & Wadhwa, S. (2005). A hybrid genetic algorithm for
production and distribution. Omega, 33(4), 345–355.
Chen, H., & Chu, C. (2003). Supply chain planning with order/setup costs and
capacity constraints a new Lagrangian relaxation approach. In 2003 IEEE
international conference on robotics and automation (Cat. no. 03CH37422) (Vol. 2).
Daskin, M. S., Coullard, C. R., & Shen, Z.-J. M. (2002). An inventory–location model:
Formulation, solution algorithm and computational results. Annals of Operations
Research, 110(1-4), 83–106.
Diabat, A. (2014). Hybrid algorithm for a vendor managed inventory system in a
two-echelon supply chain. European Journal of Operational Research, 238(1),
114–121.
Diabat, A., Aouam, T., & Ozsen, L. (2009). An evolutionary programming approach
for solving the capacitated facility location problem with risk pooling.
International Journal of Applied Decision Sciences, 2(4), 389–405.
Diabat, A., Battaïa, O., & Nazzal, D. (2015). An improved Lagrangian relaxation-based
heuristic for a joint location–inventory problem. Computers & Operations
Research, 61(1), 170–178.
Diabat, A., Abdallah, T., & Le, T. (2014). A hybrid tabu search based heuristic for the
periodic distribution inventory problem with perishable goods. Annals of
Operations Research, 1–26.
389
Diabat, A. (in press). A capacitated facility location and inventory management
problem with single sourcing. Optimization Letters. http://dx.doi.org/10.1007/
s11590-015-0950-z.
Diabat, A., & Richard, J.-P. (2015). An integrated supply chain problem: A nested
Lagrangian relaxation approach. Annals of Operations Research, 229(1), 303–323.
Diabat, A., Richard, J.-P., & Codrington, C. W. (2013). A Lagrangian relaxation
approach to simultaneous strategic and tactical planning in supply chain
design. Annals of Operations Research, 203(1), 55–80.
Dolgui, A., & Proth, J.-M. (2010). Supply chain engineering: Useful methods and
techniques. Springer.
Eskigun, Erdem et al. (2005). Outbound supply chain network design with mode
selection, lead times and capacitated vehicle distribution centers. European
Journal of Operational Research, 165(1), 182–206.
IBM (2013). IBM ILOG CPLEX optimization studio. <www.cplex.com>.
Jayaraman, V., & Ross, A. (2003). A simulated annealing methodology to distribution
network design and management. European Journal of Operational Research,
144(3), 629–645.
Kuhn, H., & Liske, T. (2014). An exact algorithm for solving the economic lot and
supply scheduling problem using a power-of-two policy. Computers &
Operations Research, 51(0), 30–40.
Le, T., Diabat, A., Richard, J.-P., & Yih, Y. (2013). A column generation-based heuristic
algorithm for an inventory routing problem with perishable goods. Optimization
Letters, 7(7), 1481–1502.
Roundy, Robin (1985). 98%-Effective integer-ratio lot-sizing for one-warehouse
multi-retailer systems. Management Science, 31(11), 1416–1430.
Santibanez-Gonzalez, E. D. R., & Diabat, A. (2013). Solving a reverse supply chain
design problem by improved benders decomposition schemes. Computers &
Industrial Engineering, 66(4), 889–898.
Shen, Z.-J. M., Coullard, C., & Daskin, M. S. (2003). A joint location–inventory model.
Transportation Science, 37(1), 40–55.
Shiguemoto, A. L., & Armentano, V. A. (2010). A tabu search procedure for
coordinating production, inventory and distribution routing problems.
International Transactions in Operational Research, 17(2), 179–195.
Shu, J. (2010). An efficient greedy heuristic for warehouse-retailer network design
optimization. Transportation Science, 44(2), 183–192.
Simchi-Levi, D., Chen, X., & Bramel, J. (2007). The logic of logistics: Theory, algorithms,
and applications for logistics and supply chain management. Springer.
Simchi-Levi, D., Kaminsky, P., & Simchi-Levi, E. (2003). Designing and managing the
supply chain: Concepts, strategies, and case studies.
Teo, C.-P., & Shu, J. (2004). Warehouse-retailer network design problem. Operations
Research, 52(3), 396–408.
Vidyarthi, N., Çelebi, E., Elhedhli, S., & Jewkes, E. (2007). Integrated production–
inventory–distribution system design with risk pooling: Model formulation and
heuristic solution. Transportation Science, 41(3), 392–408.