You are on page 1of 15

Applied Mathematical Modelling 37 (2013) 6400–6414

Contents lists available at SciVerse ScienceDirect

Applied Mathematical Modelling


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

A combined facility location and network design problem with


multi-type of capacitated links
Ragheb Rahmaniani a, Abdolsalam Ghaderi b,⇑
a
Department of Industrial Engineering, Iran University of Science and Technology, P.C. 1684613114, Narmak, Tehran, Iran
b
Department of Industrial Engineering, University of Kurdistan, P.C. 66177-15177, Sanandaj, Iran

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

Article history: This article presents a mixed-integer model to optimize the location of facilities and the
Received 5 July 2011 underlying transportation network at the same time to minimize the total transportation
Received in revised form 29 November 2012 and operating costs. In this problem, it is assumed that for connecting two nodes, there
Accepted 9 January 2013
are several types of links in which their capacity, transportation and construction costs
Available online 23 January 2013
are different. The developed model has various applications in telecommunication, emer-
gency, regional planning, pipeline network, energy management, distribution, to just name
Keywords:
a few. To solve the model effectively, this paper also proposes a fix-and-optimize heuristic
Facility location
Network design
based on the evolutionary fire-fly algorithm. Finally, to validate the model and evaluate the
Multi-type capacitated links algorithm’s performance, a series of test instances with up to 100 nodes and 600 candidate
Fix-and-optimize heuristic links with three different levels of quality are reported.
Ó 2013 Elsevier Inc. All rights reserved.

1. Introduction

Facility location models with ever growing body of literature have been applied extensively to solve real-world applica-
tions such as transportation, emergency, distribution, and so forth [1,2]. In general, facility location problems consist of a
number of facilities to be located on a given network to satisfy a set of customers in respect to some constraints. Moreover,
these problems also optimize different objectives such as transportation costs [3], operating costs [4], the number of covered
demand points [5], maximum travel time [6], and so on.
On the other hand, network design models have been used widely to optimally construct a network that enables some
kind of flow between nodes. Generally, network design problems consist of selecting a number of candidate links between
network’s nodes to minimize the sum of construction and travel costs [7,8]. For further details on network design problems,
interested readers are referred to the work provided by Magnanti and Wong [9].
Individually, facility location and network design problems have been studied extensively. However, in the recent years
many authors have pointed out an important missing link between these two problems. The reason is that, in many cases
ignoring the restriction on connection between nodes is meaningless. More importantly, the configuration of the underlying
network has a significant impact on the optimal facilities location [10]. In fact, it may be more cost-effective to change the
underlying transportation network than to add extra facilities [11]. In line with this subject, Berman et al. [12] demonstrated
how the accessibility to the facilities can be improved by modifying the underlying network. Likewise, Peeters and Thomas
[13,14] showed that the underlying network has a significant impact on optimal solutions to the p-median and

⇑ Corresponding author. Tel.: +98 9128085385.


E-mail addresses: ab.ghadri@gmail.com, ab.ghaderi@uok.ac.ir (A. Ghaderi).

0307-904X/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.apm.2013.01.001
R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6401

location-allocation problems. In view of these observations, Daskin et al. [11] in an attempt to answer this question that
‘‘Should we always assume that the underlying network is given’’? integrated decisions of both facilities location with net-
work design problems so as to introduce Uncapacitated Facility Location-Network Design Problem (UFLNDP). In this prob-
lem, we have a graph composed of a number of nodes and edges. The nodes of the graph represent customers and also the
possible locations for facilities and the edges correspond to the possibility of travelling from one node to another. Worth not-
ing that, they also showed tradeoffs between locating facilities and constructing links. This work is followed up by the doc-
toral thesis of Mlekote [15]. In this thesis, he provided stronger formulation for UFLNDP than that of Daskin et al. [11]. He
introduced the capacitated facility location-network design problem (CFLNDP), and the maximum covering location-
network design problem (MCLNDP), as well.
Another related work in this line is that of Drezner and Wesolowsky [16]. They proposed a model to optimize the location
of a single facility on a network with a set of candidate links. Meanwhile, each link can be constructed either as a one-way or
two-way link. In this paper, solution algorithms based on descent algorithm, simulated annealing, tabu search, and genetic
algorithm are put forward to solve the model. In another doctoral thesis, Cocking [7] proposed both heuristic and exact ap-
proaches to solve the budget constrained UFLND problem. The developed heuristics were: simple greedy heuristics, a local
search heuristic, and meta-heuristic algorithms including: simulated annealing and variable neighborhood search, as well as
a custom heuristic based on the problem-specific structure of FLND. Furthermore, a branch-and-cut method that uses heu-
ristic solutions as upper bounds, and cutting planes for increasing the lower bound at each node of the problem tree were
developed.
Murawski and Church [17] introduced the maximal covering network improvement problem in which it deals with exist-
ing facilities location as fixed and improves accessibility to health services by upgrading the transportation network. They
also presented and solved a real-world problem as an application of the model. A mixed-integer optimization model for inte-
grated urban hierarchy and transportation network planning was presented by Bigotte et al. [18]. In this paper, the promo-
tion of the urban centers and network links to a new level of hierarchy so as to maximize accessibility to all classes of
facilities was investigated.
Contreras et al. [19] proposed a new facility location and network design model to minimize the maximum travel time in
the network. They also reported computational results for instances with up to 100 nodes and 500 candidate links. In another
work, Contreras and Fernández [20] provided relevant modeling aspects, alternative formulations and possible algorithmic
strategies for combined facility and network design problems. Ghaderi and Jabalameli [21] presented a model for the budget-
constrained dynamic UFLNDP. They assumed a budget constraint on investment for opening the facilities and constructing
(activating) links for each time period during the planning horizon. Furthermore, a greedy heuristic and a fix-and-optimize
heuristic based on simulated annealing and exact methods (Branch & Bound and cutting methods) were proposed to solve
the model. A case study of health care was also studied in this paper.
To our best knowledge, among all the papers on network location models, no one has considered capacitated links in his/
her model. Even though, some authors have emphasized an unfulfilled requirement for such model [10,19]. However, ignor-
ing capacity on links is only valid when we are sure that links operate below their capacity level. Otherwise, any obtained
solution would be quite useless. Moreover, we are not aware of any work that considers multi-type links. In a better word, all
combined location and network models provide only one option for each possible connection in the network. However, in
reality, depends on the situation it might be more economically attractive to connect big cities with highways and small cit-
ies or rural areas with dirt roads. As a result, to build a more flexible network, we also generalize FLND problems by provid-
ing several alternative links for each possible connection and let the model to select the best one. We assume that the links
with better qualities have more capacities and entail less transportation cost while their configuration cost is much more
expensive. Nevertheless, our experiments indicted that the type of links also has a tremendous impact on the facilities opti-
mal location.
The developed model covers a wide range of application such as pipeline distribution systems, transportation systems,
power transmission network, health care accessibility, regional planning, telecommunication, distribution systems, and
can be used to solve the emergency problems with a simple modification. Additionally, the developed model can be signif-
icantly used to improve the existing networks.
As suggested from its name, FLND is combined of two well-known classes of NP-Hard combinatorial optimization prob-
lems [20]. To our best knowledge, the largest test instance of FLND problems which has been solved (even not to optimality)
is a network with 100 nodes and 500 links [19]. We should note that our model due to the links’ capacity and existing multi
type of links is more complex than the classical FLND problems. For example, we attempted to solve a network with 100
nodes, 200 links and three different types of quality (i.e. 600 links) by means of CPLEX. Unsurprisingly, it ran out of memory.
As a result, we provide some solution strategies to solve the model with less computational efforts.
The rest of this paper is structured as follows. In the following section, definition, assumptions and notations of the ad-
dressed problem are formally stated. After that, a mixed integer programming formulation is proposed to model the problem
mathematically. In Section 3, we introduce and develop a fix-and-optimize heuristic based on the firefly algorithm to solve
the developed model effectively. In Section 4, in advance of comparing performance of the proposed algorithm to that of
CPLEX, we conduct some analysis to get further insight into the model’s behavior. Finally, we summarize our work and pro-
vide some lines for future research, in Section 5.
6402 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

2. Problem statements and mathematical formulation

2.1. Problem definition and assumptions

The considered problem is defined as a network with a set of nodes and a set of candidate capacitated links. Each node
stands for both clients and a potential facility site. The underlying network should be configured in such a way that makes a
possible flow between customers and facilities. The objective function is to minimize the sum of transportation and link con-
struction costs.
This problem entails the following assumptions: (1) there are no existing links, (2) for each possible connection between
two nodes in the network, there are several types of links with the different level of quality, (3) between each pair of nodes at
most one link can be constructed, (4) links are capacitated, (5) facilities are uncapacitated, (6) each node hosts at most one
facility, (7) the network is a customer-to-server system, (8) all network links are directed, (9) links can be constructed only in
one direction and, (10) the amount of budget for opening facilities is limited.
To better understand assumption (2), consider the way between two cities that can be constructed as either dirt or paved
road. Moreover, in many cases, two cities may be connected by sea or air ways rather than ground ways. Therefore, in this
problem, we assume that there are several possible links to connect two points in which their capacity, construction and
transportation costs are different. Similar to the real world, in this problem if we want to construct a road with better quality
(paved road against dirt road), construction cost increases while the transportation cost per a unit of flow decreases. Note
from assumption (2) that the problem is multiple source one.
We refer to this problem as Capacitated Facility Location and Network Design Problem with Multi-type of Links, or
CFLNDML as an abbreviation.

2.2. Notations

To simplify the presentation of the proposed mathematical formulation for the problem, the following notations shown in
Table 1 are used throughout of this paper.
The decisions of the considered facility location and network design configuration include decisions regarding the facil-
ities that are to be located at the potential locations and links that are to be selected from the candidate links as well as the
quantities of clients’ demand that are to travel on the transportation links and fraction of demand which will be satisfied at
each facility site. These decision variables are shown in Table 1.

2.3. Model formulation

According to the aforementioned assumptions, between each pair of nodes we should have at most one connecting link.
This can be translated into mathematical equations as follows:
X t
X ij 6 1; ði; jÞ 2 L: ð1Þ
t2T

On the other hand, links should be constructed only in one direction.

Table 1
The set of notations and decision variables used in the proposed model.

Symbol Description
Sets
N Set of network nodes, i, j e {1, 2, ..., |N|} and set of clients, k e {1, 2, ..., |N|},
L Set of candidate links, ði; j; tÞ 2 L,
T Set of different types of links, t e {1, 2, ..., |T|},
Parameters
Dk Demand of client k,
fi Fixed cost of opening a facility on node i,
ctij Fixed cost of constructing link (i, j) of type t,
tr tij Cost of travelling on link (i, j) of type t per a unit of demand,
dij Length of link (i, j),
V tij Capacity of link (i, j) of type t,
B Maximum available budget for opening facilities,
Decision variables
Zi If facility at node i is open (1), otherwise (0),
X tij If link (i, j) of type t is constructed (1), otherwise (0),
Y kt Fraction of the client’s demand k traveling on link (i, j) of type t,
ij

W ki Fraction of the client’s demand k served by facility i


R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6403

X tij þ X tji 6 1; ði; jÞ 2 L; t 2 T: ð2Þ

By combining these two inequalities, we can write a single set of inequality which covers both of these assumptions as
follows:
X t
X ij þ X tji 6 1; ði; jÞ 2 L: ð3Þ
t2T

To build a more flexible network, we let the model determines the optimal number of opened facilities within a pre-spec-
ified budget rather than fixing a priori number of them.
X
fi Z i 6 B: ð4Þ
i2N

According to the aforementioned notations and definitions, the proposed mixed-integer model for CFLNDML is formu-
lated as follows:
!
X XX X
Min Dk tr tij Y kt
ij þ ctij X tij : ð5Þ
t2T k ði;jÞ2L ði;jÞ2L

Subject to:
XX
Zi þ X tij P 1; 8i 2 N; ð6Þ
t2T j2N

X t X t
X ij þ X ji 6 1; ði; jÞ 2 L; ð7Þ
t2T t2T

XX kt XX kt
Y ji ¼ W ki þ Y ij ; 8i; k 2 N : i–k; ð8Þ
t2T j2N t2T j2N

X
Zk þ W ki ¼ 1; 8k 2 N; ð9Þ
i2N:i–k

XX
Zi þ Y itij ¼ 1 8i 2 N; ð10Þ
t2T j2N

W ki 6 Z i ; 8i; k 2 N; ð11Þ

Y kt t
ij 6 X ij ; 8i; j; k 2 N; ði; jÞ 2 L; t 2 T; ð12Þ
X
Dk Y kt t t
ij 6 V ij X ij ; 8ði; jÞ 2 L; t 2 T; ð13Þ
k

X
fi Z i 6 B; ð14Þ
i2N

X tij 2 f0; 1g; 8ði; jÞ 2 L; t 2 T; ð15Þ

Z i 2 f0; 1g; 8i 2 N; ð16Þ

Y kt
ij P 0; 8ði; jÞ 2 L; 8k 2 N; 8t 2 T; ð17Þ

W ki P 0; 8i; k 2 N : k–i; ð18Þ


In the proposed model, the objective function (Eq. (5)) minimizes the total costs of transportation and links construction.
Worth noting that by minimizing transportation costs, we aim at increasing accessibility to the facilities. Eq. (6) ensure that if
node i is not a facility node, it has to be allocated to the located facilities at the other nodes (via constructing outbound links
to transship its demand). In this equation, since the links are capacitated, each customer may need to transship its demand
via constructing several outbound links. On the other hand, the set of inequalities (7) states that at most a one-way-link can
be constructed between two nodes. Therefore, according to the capacity, transportation and construction cost of each type of
links between two nodes, the model determines the best type of link (if required) to construct. Thus the constraints (6) and
(7) are related to link construction.
6404 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

Equalities (8)–(10) are flow conservation constraints. In constraints (8) total flow that belongs to client k and comes into
node i, should be equal to amount of satisfied demand at this node and amount of demand which traverse it. Constraints (9)
indicate that the demand of each customer should be fully satisfied and if node k is a facility node, it can satisfy its own de-
mand. In other words, each customer must find a destination, whether it be the facility located at that node or at the other
nodes. As a result, constraints (10) state if node k was a facility node, it should not construct any flow on the network’s links;
otherwise, it should find a route to transship its demand.
The assignment and routing conditions are given in Eqs. (11)–(13). Inequalities (11) ensure that nodes can satisfy demand
for customers if and only if they are a facility node. Likewise, inequalities (12) guarantee that flow can only exist on the con-
structed links, whereas constraints (13) make sure that the total flow on each link does not exceed its capacity. As usual,
constraints (12) are redundant and can be eliminated. However, they can significantly improve the quality of solutions in
relaxation based solution algorithms. The budget limitation on the opening facilities is imposed in constraints (14). Finally,
equations (15)–(18) state the aforementioned decision variables.

3. Evolutionary discrete firefly

The problem belongs to the NP-hard class of combinatorial optimization problems [20] and practically, it is very hard to
solve even medium-size test instances with the conventional methods. On the other hand, the model can be decomposed
into two sub problems: (a) location sub-problem and (b) network design sub-problem. The former sub-problem can be easily
handled with meta-heuristic algorithms and the latter one with local search algorithms. Therefore, in this part by dividing
the problem into two sub-problems, we are going to adopt and develop a fix-and-optimize algorithm to solve the model in
hand efficiently.
Firefly algorithm (FA) is one of the most recent nature-inspired meta-heuristics presented by Yang [22]. This algorithm is
originally introduced to solve continuous optimization problems and there have been written fairly few articles concerning it
in discrete space. FA mimics the social behavior of fireflies by idealizing the following rules: (a) all the fireflies are unisex, (b)
each firefly is attracted only to the brighter fireflies, and the brightest firefly moves randomly, (c) strength of the attractive-
ness is proportional to the firefly’s brightness and decreases over the distance and, (d) brightness of every firefly is propor-
tional to the objective function [22]. Based on [23], FA is very efficient in finding the global optima with high success rates.
The performed simulation by Yang shows that FA is superior to both PSO and GA in terms of both efficiency and success rate
[23]. Lukasik and Zak [24] also study FA for a continuous constrained optimization task. Their experiment demonstrates the
efficiency of FA.
Using the metaheuristic algorithms to solve the facility location problems was very common during recent years. The effi-
ciency of the hybrid PSO algorithm in comparison with genetic algorithm and variable neighborhood search was proven to
solve the uncapacitated location-allocation problem with the extensive experiments by Ghaderi et al. [25]. However, to our
best knowledge, no one has applied the FA algorithm to the location problems, so far. For that reason, we mainly draw infor-
mation from [24,26,27]. In short, the challenges in discretization of this algorithm are calculating the distance between two
fireflies and how define the social interaction of fireflies in movement. The pseudo-code of the proposed firefly algorithm is
given in Fig. 1.

3.1. Initial population of fireflies

Similar to other evolutionary meta-heuristics, the initial locations of fireflies are scattered over the solution space ran-
domly. Size of this population is m and needs to be experimentally tuned. However, the initial location of each individual
should satisfy Eq. (14) in order to be feasible.

3.2. The representation of firefly

Each firefly represents one solution to the location sub-problem. Representation of each firefly is just like a chromosome
that represents an individual in genetic algorithm as illustrated by Fig. 2. In this kind of representation, each element of array
stands for a node and indicates whether a node is a facility node (1), or not (0).

3.3. Distance of two fireflies

Distance between two fireflies i and j can be defined in two possible ways: (a) Hamming’s distance and (b) the minimum
number of required swaps of firefly i in order to get firefly j [26]. In this paper, we use Hamming’s distance because of two
main reasons. First, it is more convenient to implement and second, the difference in the objective function value of two fire-
flies decreases with smaller Hamming’s distance, not the other one. Therefore, distance between any two fireflies i and firefly
j is defined as the number of different elements between their permutations. From now on, r symbolizes the distance be-
tween two fireflies.
R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6405

Fig. 1. The proposed firefly algorithm pseudo-code.

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

Fig. 2. The representation of a firefly for a test instance with 10 nodes and 3 opened facilities.

3.4. Attractiveness of a firefly

Since the model is a minimization problem, firefly i will be attracted to firefly j if and only if its objective function be
greater than the objective function of firefly j. On the other hand, the degree of attractiveness of firefly j on firefly i (i.e.
bj!i ) can be defined as any decreasing function of distance between fireflies (i.e. rij ) and increasing function of attractiveness
of the firefly j at r ij ¼ 0 (i.e. b0). In this paper, the following equation is used.

b0
bj!i ¼ ð19Þ
1 þ cr2ij ;

where c is the light absorption coefficient. Although in theory c 2 ½0; 1Þ, we recommend setting this parameter in the
interval [0.01, 0.20].

3.5. Movement toward attractive firefly

In the original form of FA for continuous problems, each firefly is attracted and moved toward brighter fireflies (if any)
and then takes a random step. This issue is shown in Eq. (20) and (21) [22].

xi xi þ bj!i ðxj  xi Þ þ aðRandðÞ  0:5Þor ð20Þ

xi ð1  bj!i Þxi þ bj!i xj þ aðRandðÞ  0:5Þ ð21Þ

where a is the user-defined parameter that adjusts the maximal random step. Note from Eq. (21) that parameters bj?i and
a determines the movement of the firefly in two ways: (1) bj?i determines the movement of firefly i toward firefly j and, (2) a
determines the random movement of firefly i in the solution space. Worth noting that the order of these two steps is not
6406 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

interchangeable, since it results in farther or closer distance between fireflies and consequently, changes the value of bj?i.
Therefore, movement in this algorithm can be made in two consecutive steps: (a) b-step and, (b) a-step. Accordingly, in order
to adopt this algorithm and solve discrete problems, these two steps should be modified in an appropriate way [26].
Although there have been some attempts to develop the discrete version of this algorithm, none of them can be applied
to the problem in hand (or other similar location problems) as they are. Therefore, following sub-sections illustrate how
these two steps can be modified properly to solve location problems.
b-Step
This step determines how to move the less attractive firefly toward the more attractive one. The movement should bring
two fireflies closer to each other; which means the amount of their common elements in permutation should increase. In line
with this subject, the common elements of these two fireflies will be extracted. Afterward, gaps in the permutation should be
filled with probability bj!i .
To understand this, assume we are filling 3rd position in Fig. 3. The value of this cell after movement would change to 1 or
remain 0. To do so, a random number is generated, and if it is smaller than bj!i , then an element from permutation of firefly j
will be inserted into the permutation of firefly i; otherwise, the value of this element would not change and remain zero.
Calculating bj!i for the fireflies i and j in Fig. 3 with c = 0.1 and b0 = 1 results in bj!i = 0.38. Now, assume that the generated
random number has been 0.12. As a result, the value of the third position in permutation of firefly i (the less attractive firefly)
becomes 1 instead of 0.
The former Hamming’s distance was 4 while the new Hamming’s distance is 1; therefore, firefly i has approached towards
firefly j with three distance unit closer. However, the new position might be infeasible to the model, if the cost of opened
facilities is more than the considered budget. The next step illustrates how to repair infeasible solutions, as well.
a-Step
This step is somehow easier than the b-step and it only does two things: (a) make infeasible solutions to feasible and, (b)
randomize the movement of fireflies. Therefore, if a permutation does not satisfy the constraint (14), we randomly close a
number of facilities in that permutation to satisfy this constraint. Otherwise, we only swap elements of two randomly se-
lected cells. This procedure can cause either increase or decrease in distance between two fireflies. By applying this step
to the produced position in Fig. 3, the distance between fireflies i and j increase by one unit as given in Fig. 4.

3.6. Local search

The location of facilities is determined at each iteration of the described algorithm.. After that, by fixing the facility nodes
and the rest of the problem is solved with the CPLEX solver. To do so, the algorithm is coded in MATLAB and a software that
makes an interface between MATLAB and GAMS [28] is used. Moreover, since the local optimal is determined by an exact
method, we have designed a long-term memory to prevent from the generated moves to being visited again.

3.7. Stopping criteria

Finally, the algorithm will stop if it meets any of the following criteria: (I) reach to the optimal solution, (II) elapse the
considered maximum CPU run time, and (III) after |N| non-improvement iterations for all fireflies.

Firefly i 1 0 0 0 1 0 0 1 0 0
Firefly j 0 0 1 0 0 0 0 1 0 1
Extracting common elements
? 0 ? 0 ? 0 0 1 0 ?
Filling gaps
Filling 1st position: 0 0 ? 0 ? 0 0 1 0 ?
Filling 3rd position: 0 0 1 0 ? 0 0 1 0 ?
Filling 5 th position: 0 0 1 0 1 0 0 1 0 ?
Filling 10 th position: 0 0 1 0 1 0 0 1 0 1
The updated position
0 0 1 0 1 0 0 1 0 1

Fig. 3. Movement of firefly i toward firefly j.

0 0 0 0 1 0 0 1 0 1

Fig. 4. The position of firefly i after performing a-step.


R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6407

4. Results and discussions

For our experiments, all test problems have been solved with the standard mathematical programming software GAMS
23.3.3, namely with the branch-and-bound algorithm of CPLEX 12.1. Moreover, the algorithm has been coded in MATLAB 7.6
and run under Windows XP on a Personal DELL computer with 2.2 GHz processor and 2 GB of RAM. Finally, CPU time was
limited to |N|2  |T| seconds.

4.1. Test instances

The procedure of our test problem generation is basically based on that of Contreras et al. [19] and Melkote and Daskin
[29]. Therefore, we generated various test networks with a number of nodes n e {10, 20, 40, 60, 80, 100} and links L e
{2n, 4n, 6n} according to the following procedure. For the location of the clients (network nodes), we randomly distributed
nodes over a 100 ⁄ 100 area. Then, the desired number of candidate links was randomly selected with a bias towards shorter
links to imitate real-world transportation networks. In addition, the required parameters for each test network are generated
according to the following procedure. Budget, fixed costs and demands are randomly generated from uniform distributions
in [100n, 200n], [1000, 1500] and [10, 40], respectively.
For each link, we assume three different levels of quality (i.e. |T| = 3) which can be interpreted as the dirt road (type 3), the
paved road with low quality (type 2) and the paved road with the standard quality (type 1). As it should be expected, if we
construct a link of type 1, its construction cost should be more than the other types while, its capacity is more and its trans-
portation cost is lower than the other types of connecting links. Therefore, we first generate capacities and construction costs
for the links type 1 and then reduce it by 5 to 15 percentages to obtain the data for links type 2 and by 15 to 30 percentages
to obtain the data for links type 3. On the other hand, we increase transportation costs on links type 1 by 10 to 20 and 15 to
30 percentages to obtain transportation costs on links type 2 and 3, respectively.
The capacities of the links type 1 are uniformly drawn from interval [40, 120]. The construction costs of the links type 1
are calculated by c1ij ¼ udij ð1 þ rÞ in which, u is the unit link construction cost and varies in the interval [2,10] and r is a ran-
dom number from a uniform distribution in [0.2, 0.2]. Likewise, the transportation cost per a unit of flow on link (i, j) of type
1 is given by t1ij ¼ dij ð1 þ rÞ.
By this procedure, we generated 18 asymmetric and 18 symmetric test instances to evaluate the algorithm’s performance
in both cases.

4.2. An example application

To study the behavior of the model, we carry out the sensitivity analysis on a widely used test network in the literature
([29–31]). This network and its data which is taken from Melkote and Daskin [29], is depicted in Fig. 5. Notice, other required
data for the analysis are generated according to the discussed procedure in Sections 4-1. We also assume that the capacity of
all links is identical, so all V ti j are replaced with V. This parameter varies from 75 to 195 in increments of 10. Additionally, to
assess the sensitivity to the budget, we assume four different values for the amount of budget. To conduct this analysis, we
generated both symmetric and asymmetric data and solved all cases to optimality.
Fig. 6 shows the optimal structure of the facilities location and the constructed transportation network as model’s output
in four cases. In all cases, we fixed the budget at $3000 and in the capacitated cases, we set the capacity of each link at 85. We
clearly observe that how the structure of data and capacity of links can affect the underlying network and the optimal facil-
ities location. In uncapacitated cases, all demand will be sourced by a single facility, and each node has at most one outbound
link. In a better word, in uncapacitated cases of the problem, if the network contains P facilities, the underlying transporta-
tion network will be exactly made of |N|-|P| links. Moreover, in this case, the network is partitioned into P separated net-
works. On the other hand, in the capacitated cases, demands may be served by multiple sources. Therefore, further links
need to be constructed, and customers have to travel the longer path to meet a facility node. As a result, both transportation
and link investment costs increase. Furthermore, it seems that in problems with asymmetric data, both transportation and
link investment costs are greater than the symmetric cases.
Another interesting part of our observations is about considering different types of links. If we only consider one type of
links (for example, the first type), though the quality of designed network improves, on the average, the total cost increases
by 6.93%. Notice, this percentage greatly depends on the difference between the transportation and construction costs of dif-
ferent types of links. Moreover, it is far away from the reality if we design a network with the same type of links. For exam-
ple, in the reality, small cities will be connected with one-band low-quality roads while the big cities will be connected via
high-quality highways. More importantly, type of links greatly affects the optimal facilities location. For instance, if we only
consider one type of links, the facilities optimal location is {4, 13 and 17} while, the optimal facilities location as an output of
the model is {4, 16 and 20}. Therefore, the designed network can match to the real-world problems much more appropriately
than traditional FLND models.
Nevertheless, although uncapacitated cases have much cheaper design, we believe that the obtained designs for the
capacitated cases are much more effective. The reason is that, in many real world problems (e.g. pipeline network, distribu-
tion system, telecommunication, regional planning, and so forth) ignoring links capacity is meaningless. Meanwhile, the
6408 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

Fig. 5. A 21-node test network [29].

other benefits except accessibility to the facilities could be concluded from the designed network. In a better word, con-
structing additional links increase, for example, accessibility to the job centers, rural resources, and etc. However, such ben-
efits cannot be easily measured, and we have to ignore less important or hard-to-model parts of the problem.
However what is more interesting is studying the behavior of the key parameters of the model. In line with this subject,
Figs. 7–10 respectively display the transportation costs, links investment costs, the number of multi-sourced demands, and
the number of constructed links versus the links capacity for each level of the considered budget. Needless to say, as the
amount of budget increases (decreases), the number of opened facilities also increases (decreases).
Fig. 7 depicts changes in the transportation costs for each level of the considered budget by increasing the links capacity.
This figure clearly shows, as the capacity and/or budget slightly increases, transportation costs decrease. The reason is that,
by increasing the budget, more facilities will be located and by increasing the capacity, fewer demand points will be satisfied
by multi sources (see Fig. 8). As a result, by increasing the budget and capacity the configuration becomes more efficient in
terms of the transportation cost. However, for small values of capacity, the model is very sensitive to the budget. As the
capacity for the first time increase, transportation costs sharply drop, and if we continue to increase the problem become
uncapacitated in which the greater values of capacity cannot diminish transportation costs any more.
We can observe similar phenomena between link investment costs, capacity and the budget in Fig. 8. As we slowly in-
crease links capacity, number of multi-sourced demands reduce (see Fig. 8) and in turn, it allows us to invest less in links
construction to transship the same amount of demand. However, this decrement is not monotonically. It appears that links
investment costs fall sharply and then fluctuate until capacity has no longer the effect on the links construction costs. In fact,
the amount of this fluctuation is more evident for a problem with B = 6000. The reason is that, in this problem 5 or 6 facilities
are located and by changing the capacity, the model shifts some of these facilities to more expensive locations (which also
R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6409

Capacitated Problem with Symmetric Data Capacitated Problem with Asymmetric Data
Transportation cost: 53021 Transportation cost: 53295
Link investment cost: 5453 Link investment cost: 5745

Uncapacitated Problem with Symmetric Data Uncapacitated Problem with Asymmetric Data
Transportation cost: 35866 Transportation cost: 37695
Link investment cost: 3305 Link investment cost: 3388

Fig. 6. The topology of the constructed network in four cases.

contains more demand) and constructs more expensive (direct) links so as to reduce the transportation cost and hence, total
costs.
Fig. 9 shows that, larger values of capacity and/or budget corresponds to fewer multi-sourced demands. As we observe,
the model is not affected by this parameter after a specific level of the links capacity.
As we would have thought, for a given budget, as we further increase the capacity of links, fewer links are required to
configure the underlying network. If we keep increasing the capacity, eventually, the problem becomes uncapacitated which
consists of |N|-|P| links. Further reduction can be only achieved by increasing the amount of budget. Notice, fluctuation in the
number of links should be expected, because as the capacity increases, facilities location and in consequence, the underlying
network also changes.

4.3. Algorithm’s performance

In this part, we compare the performance of the proposed FA based algorithm to that of CPLEX. The algorithm’s param-
eters are experimentally tuned and set as follows: b0 = 1, c = 0.1 and m = 20. Computational results are summarized in Tables
2 and 3. Note that, we ran the proposed algorithm five times and report the average and maximum CPU time requirement
6410 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

Fig. 7. Transportation cost vs. links capacity by the budget.

Fig. 8. Link investment cost vs. links capacity by budget.

Fig. 9. Number of multi-sourced demands vs. links capacity by budget.


R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6411

Fig. 10. Number of links vs. links capacity by budget.

Table 2
Computational results for test instances with asymmetric data.

Test problem CPLEX FA


N L L.B Time (s) Error (%) Average time (s) Max time (s) Average error (%) Max error (%)
TP1 10 2n 11991.14 4.51 0.00 6.94 10.01 0.00 0.00
4n 8987.65 4.34 0.00 4.56 7.73 0.14 1.68
6n 7813.09 3.87 0.00 5.75 21.09 0.00 0.00
TP2 20 2n 44396.23 5.57 0.00 5.03 13.00 0.00 0.00
4n 73276.00 10.27 0.00 6.39 8.04 0.00 0.00
6n 11968.92 27.77 0.00 11.62 17.16 0.00 0.00
TP3 40 2n 40498.17 2451.17 0.00 189.63 256.89 0.48 2.38
4n 27238.59 794.69 0.00 173.15 313.33 0.00 0.00
6n 23553.12 263.75 0.00 107.22 160.41 0.00 0.00
TP4 60 2n 22843.51 10,800 11.85 3190.14 4855.52 3.03 5.08
4n 16805.64 3391.87 0 407.42 601.42 0.37 1.52
6n 15542.85 1162.11 0 381.06 498.36 0.00 0.00
TP5 80 2n 28308.03 19,200 N/A 2153.87 3087.50 7.89 13.89
4n 19470.27 16680.07 3.13a 5176.92 7111.03 0.12 0.81
6n 13529.97 19,200 2.32 3503.06 4808.52 0.76 2.41
2n O/M – – 5176.04 8050.11 62490.48 70103.63
TP6b 100 4n O/M – – 6074.31 10389.49 38364.02 39783.42
6n O/M – – 3103.97 4234.37 10029.59 12307.44
Ave. 4933.33 1.09 1021.52 1451.33 0.35 0.99

The bold values are the average amount of results of the different solved test problems in each column
a
CPLEX runs out of memory after given elapsed time.
b
The reported value of FA in the last two columns instead of error.

and the optimality gap of these experiments in the last four columns. Moreover, the column labeled by ‘‘L.B’’ reports the ob-
tained lower bounds by CPLEX. The error percentage of the heuristics is computed as:

Algorithms objective value  Lower bound obtained by CPLEX


Error ¼ 100 
Algorithms objective value
Finally, those cells that are filled with ‘‘O/M’’ and ‘‘N/A’’ respectively indicate that the algorithm ran out of memory and
has not been able to find a feasible solution. Note that, the values reported in the last row of these tables indicate the average
of those results that CPLEX has also obtained a feasible solution for.
Table 2 reports the obtained computational results for the generated test network with asymmetric data. As we clearly
observe, CPLEX has been able to solve all test networks with up to 40 nodes while, the algorithm in a few cases has a slight
error. However, this is due to one or two runs out of five runs in which the algorithm fails to obtain the optimal solutions.
Nevertheless, for this set of instances, on the average, the algorithm performs seven times and at the worst case, four times
faster than CPLEX.
6412 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

Table 3
Computational results for test instances with symmetric data.

Test problem CPLEX FA


N L L.B Time (s) Error (%) Average time (s) Max time (s) Average error (%) Max error(%)
TP1 10 2n 12831.57 4.21 0.00 4.45 9.79 0.00 0.00
4n 10089.47 4.43 0.00 4.66 6.06 0.59 2.25
6n 9247.19 6.01 0.00 8.74 18.18 0.00 0.00
TP2 20 2n 41340.97 6.56 0.00 5.95 11.10 0.00 0.00
4n 58459.02 19.64 0.00 7.11 21.09 0.00 0.00
6n 9931.99 34.47 0.00 7.94 14.56 0.00 0.00
TP3 40 2n 49086.92 2750.56 0.00 1103.01 1509.47 0.00 0.00
4n 27978.60 802.65 0.00 362.42 480.13 0.71 3.57
6n 24155.64 751.91 0.00 166.14 298.05 0.00 0.00
TP4 60 2n 21551.49 3180.51 2.07 325.73 380.98 0.10 0.51
4n 19423.19 2040.06 0.00 1210.34 1887.34 0.31 1.35
6n 13136.25 900.57 0.00 710.50 979.11 0.00 0.00
TP5 80 2n 28193.16 19,200 N/A 4531.81 7839.00 5.82 7.75
4n 17503.09 12540.05 2.11a 9010.39 13267.57 0.62 2.88
6n 12313.84 5040.56 0.89a 5058.40 6782.11 1.19 4.31
2n O/M – – 5020.10 8093.22 54316.40 68372.31
TP6b 100 4n O/M – – 12300.28 14067.75 38924.52 50031.63
6n O/M – – 3856.90 7008.12 30389.59 37022.10
Ave. 3152.15 0.36 1501.17 2233.64 0.25 1.06

The bold values are the average amount of results of the different solved test problems in each column.
a
CPLEX runs out of memory after given elapsed time.
b
The reported value of FA in the last two columns instead of error.

Moreover, as the scale of the problems increases, the performance of CPLEX greatly reduces. To be more precise, in four test
instances it cannot even find a feasible solution and ran out of memory. In addition, after approximately three and half hours
CPLEX runs out of memory and cannot find the optimal solution for TP5 with 4n links. On the average, for TP4 and TP5, the algo-
rithm improves error and CPU time requirement by CPLEX respectively more than 75 and 78 percentages. Furthermore, the
algorithm finds a solution for TP5 with 2n links that has less than 8% error even though CPLEX cannot find a feasible solution
for it. For TP6, we cannot report error percentage for the algorithm’s results, since CPLEX runs out of memory and cannot even
reach to a lower bound.
Table 3 summarizes performance comparison for the symmetric test instances. CPLEX finds the optimal solution for 11
out of 18 test instances. Whereas, the algorithm only solves 8 test instances optimally. Moreover, CPLEX cannot even find
a feasible solution for TP5 with 2n links after 19,200 s and runs out of memory for all test instances with 100 nodes. Note
that, after elapsing some time and finding a reasonable optimality gap, it also runs out of memory for two test instance with
80 nodes. All in all, the algorithm can improve respectively error and time requirement by roughly 31% and 52%.
According to Table 2 and 3, we can point out that the problem becomes much harder to solve as the number of candidate
links decrease. This issue also affects the algorithm’s performance. The reason is that, the local optimums are extracted by

Fig. 11. Convergence of both CPLEX and FA over time.


R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414 6413

CPLEX. Especially for TP6, an iteration may take hundreds of seconds. Therefore, for this instance, we terminate CPLEX when
its solution is within 2% of error. However, developing an efficient local search for the problem requires conducting further
studies in the future. Roughly speaking, although the algorithm is well suited to solve the asymmetric problems, it seems
that symmetric problems are easier to solve. Generally, as the number of links increases, total costs decrease because the
model can construct more direct links and consequently reduces both terms of the objective function.
In order to further demonstrate superiority of the proposed algorithm to CPLEX, the convergence of both approaches over
time is studied. For this experiment, TP3 with 2n links and symmetric data is taken as representative and the percentage of
error versus time for solving this problem is depicted in Fig. 11. As we clearly observe, by increasing the running time, dis-
tance of the obtained solutions from the best-known solution decrease. It might be interesting to note that, CPLEX found its
first feasible solution at a time of 320 s with an error of 91% and it takes 2750 s to find the optimal solution. On the other
hand, the algorithm found its first feasible solution approximately after 6 s with an error of 48% and it takes around
1200 s to find the optimal solution.
All these experiments demonstrate that the proposed fix-and-optimize algorithm which is based on FA outperforms
CPLEX.

5. Conclusions and remarks for future work

In this paper, we have introduced and investigated a budget constrained facility location and network design problem in
which for each possible connection between two nodes, there are different type of capacitated links. Broadly speaking, the
developed model has a number of important applications in regional planning, distribution, telecommunications, pipeline
network, energy management, to just name a few. To validate and get insight into the model, we have conducted some com-
putational experiments on a well-known 21-node test network.
Furthermore, a fix-and-optimize heuristic has been developed to solve the proposed model effectively. In this heuristic,
we developed and implemented a discrete version of one of the newest meta-heuristics known as the firefly algorithm.
Numerical studies have been carried out to examine the quality of the heuristic, and the results are compared to that of
CPLEX. Based on the results of numerical experiments, we can say that the performance of the proposed algorithm is much
better than CPLEX in terms both CPU time requirement and the optimality gap.
For the future research, the model can be extended to the situation in which facilities are capacitated. Likewise, other
appropriate and important avenues for forthcoming studies would be developing the model with uncertain parameters
either as the static or dynamic model. Another topic of interest is considering congestion effects on both facilities and links.
In addition, future studies may take into account the other benefits from constructing a transportation network other than
accessibility to the facilities. For example, they might introduce a second objective to maximize accessibility of people to
other nodes. Finally, future study is needed to develop more efficient local search and compare the proposed heuristic to
other solution algorithms.

Acknowledgements

The authors wish to thank anonymous referees for their helpful and constructive comments on earlier versions of this
paper.

References

[1] C.S. ReVelle, A.H. Eiselt, Location analysis: A synthesis and survey, Eur. J. Oper. Res. 165 (2005) 1–19.
[2] J.R. Current, M.S. Daskin, D.A. Schilling, Discrete network location models, in: Drezner, Hamacher (eds.) Facility Location, Springer, 2002, pp. 81–118.
[3] S.L. Hakimi, Optimum locations of switching centers and the absolute centers and medians of a graph, Oper. Res. 12 (1964) 450–459.
[4] A.A. Kuehn, M.J. Hamburger, A heuristic program for locating warehouses, Manag. Sci. 9 (1963) 643–666.
[5] Church, C. ReVelle, The maximal covering location problem, Regional Science Association Papers, vol. 32, 1974, pp. 101–118.
[6] S. Elloumi, M. Labbe, Y. Pochet, New formulation and resolution method for the p-center problem, INFORMS J. Comput. 16 (2004) 84–94.
[7] C. Cocking, Solutions to facility location–network design problems, Germany. University of Heidelberg, 2008.
[8] A. Balakrishnan, T.L. Magnanti, R.T. Wong, A dual-ascent procedure for large-scale uncapacitated network design, Oper. Res. 37 (1989) 716–740.
[9] T.L. Magnanti, R.T. Wong, Network design and transportation planning: models and algorithms, Transport. Sci. 18 (1984) 1–55.
[10] S. Melkote, M.S. Daskin, An integrated model of facility location and transportation network design, Transp. Res. 35 (2001) 515–538.
[11] M.S. Daskin, A.P. Hurter, M.G. Van Buer, Toward an integrated model of facility location and transportation network design, The Transportation Center,
Northwestern University, Working Paper. Evanston, IL, USA, 1993.
[12] O. Berman, D.I. Ingco, A.R. Odoni, Improving the location of mini-sum facilities through network modification, Ann. Oper. Res. 40 (1992) 1–16.
[13] D. Peeters, I. Thomas, The effect of spatial structure on location-allocation results, in: ISOLDE VI Conference Lesvos, Greece, 1993.
[14] D. Peeters, I. Thomas, The effect of spatial structure on p-median results, Transport. Sci. 29 (1995) 366–373.
[15] S. Melkote, Integrated models of facility location and network design, Industrial Engineering and Management Science, Northwestern University,
Evaston, Illinois, 1996.
[16] Z. Drezner, G.O. Wesolowsky, Network design: selection and design of links and facility location, Transp. Res. 37 (2003) 241–256.
[17] L. Murawski, R.L. Church, Improving accessibility to rural health services: The maximal covering network improvement problem, Socio-Econom. Plan.
Sci. 43 (2009) 102–110.
[18] J.F. Bigotte, D. Krass, A.P. Antunes, O. Berman, Integrated modeling of urban hierarchy and transportation network planning, Transp. Res. Part A 44
(2010) 506–522.
[19] I. Contreras, E. Fernandez, G. Reinelt, Minimizing the maximum travel time in a combined model of facility location and network design, Omega 40
(2012) 847–860.
6414 R. Rahmaniani, A. Ghaderi / Applied Mathematical Modelling 37 (2013) 6400–6414

[20] I. Contreras, E. Fernández, General network design: a unified view of combined location and network design problems, Eur. J. Oper. Res. 219 (2012)
680–697.
[21] A. Ghaderi, M.S. Jabalameli, Modeling the budget-constrained dynamic uncapacitated facility location-network design problem and solving it via two
efficient heuristics: A case study of health care, Math. Comput. Model. 57 (2013) 382–400.
[22] X.S. Yang, Nature-inspired Metaheuristic Algorithm. Luniver Press, 2008.
[23] X.S. Yang, Firefly algorithms for multimodal optimization, in: Watanabe, O., Zeugmann, T. (eds.) SAGA 2009. LNCS, Springer, Heidelberg, vol. 5792,
2009, pp. 169–178.
_
[24] S. Lukasik, S. Zak, Firefly algorithm for continuous constrained optimisation tasks, Systems Research Institute, Polish, Academy of Sciences, 2010, p. 1–
10.
[25] A. Ghaderi, M.S. Jabalameli, F. Barzinpour, R. Rahmaniani, An efficient hybrid particle swarm optimization algorithm for solving the uncapacitated
continuous location-allocation problem, Netw. Spat. Econom. 12 (2011) 421–439.
[26] K. Durkota, Implementation of a discrete firefly algorithm for the QAP problem within the SEAGE framework Electrical engineering, Czech Technical
University, 2011.
[27] A. Yousif, A.H. Abdullah, S.M. Nor, A.A. Abdelaziz, Scheduling jobs on grid computing using firefly algorithm, J. Theor. Appl. Inform. Technol. 33 (2011)
155–164.
[28] M.C. Ferris, MATLAB and GAMS: interfacing optimization and visualization software, University of Wisconsin, 2005. http://www.cs.wisc.edu/math-
prog/matlab.html.
[29] S. Melkote, M.S. Daskin, Capacitated facility location/network design problems, Eur. J. Oper. Res. 129 (2001) 481–495.
[30] O. Berman, M.R. Rahnama, Optimal path of a single service unit on a network to a ‘‘nonemergency’’ destination, Transport. Sci. 17 (1983) 218–232.
[31] M.J. Hodgson, K.E. Rosing, A network location–allocation model trading off flow capturing and p-median objectives, Ann. Oper. Res. 40 (1992) 247–260.

You might also like