Transportation Letters

The International Journal of Transportation Research

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/ytrl20

A bi-objective transportation-location arc routing

problem

Alireza Amini, Reza Tavakkoli-Moghaddam & Sadoullah Ebrahimnejad

To cite this article: Alireza Amini, Reza Tavakkoli-Moghaddam & Sadoullah Ebrahimnejad
(2020) A bi-objective transportation-location arc routing problem, Transportation Letters, 12:9,
623-637, DOI: 10.1080/19427867.2019.1679405

To link to this article: https://doi.org/10.1080/19427867.2019.1679405

Published online: 18 Oct 2019.

Submit your article to this journal

Article views: 315

View related articles

View Crossmark data

Citing articles: 10 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ytrl20
TRANSPORTATION LETTERS
2020, VOL. 12, NO. 9, 623–637
https://doi.org/10.1080/19427867.2019.1679405

A bi-objective transportation-location arc routing problem

Alireza Aminia, Reza Tavakkoli-Moghaddam b,c
and Sadoullah Ebrahimnejadd
a
School of Industrial Engineering, South Tehran Branch, Islamic Azad University, Tehran, Iran; bSchool of Industrial Engineering, College of Engineering,
University of Tehran, Tehran, Iran; cArts et Métiers ParisTech, LCFC, Metz, France; dDepartment of Industrial Engineering, Karaj Branch, Islamic Azad
University, Karaj, Iran

ABSTRACT KEYWORDS
Location and routing problems belong to well-known optimization ones, whose combinations have merged Location-arc routing;
as a location-routing problem (LRP). Arc routing has recently become a popular interest of researchers. transportation; bi-objective
Accordingly, the location-arc routing problem (LARP) is defined. This paper deals with a location-arc routing model; augmented ε-
problem when there are transportation decisions from suppliers to established depots (TLARP). It is constraint; genetic
algorithm; late acceptance
addressed by developing a bi-objective mathematical model to optimize the total costs and makespan. hill-climbing algorithm
An augmented ε-constraint algorithm is employed to find the true Pareto solutions. Then two meta-
heuristics are considered: non-dominated sorting genetic algorithm II (NSGA-II) and multi-objective late
acceptance hill-climbing (MOLAHC) algorithm. There are four heuristics regarding NSGA-II, its combination
with MOLAHC, and having/not having a local search (LS). Results demonstrate that NSGA-II+LS is the best
and hybrid+LS is the second best heuristics while the solving time of hybrid+LS is significantly less than the
time of NSGA-II+LS.

Introduction presented a mixed-integer non-linear programming (MINLP)

model for a two-echelon fleet multi-period location-routing pro-
Supply chains exist for all kinds of business in the world.
blem with time windows. They used a Lagrangean decomposition
Organizations are expected to recognize the supply chains they
method to solve the problem. Given that the LRP is derived from
are involved in. Thus, most of the business models include logistics,
the integration of the two NP-hard issues, it is also NP-hard.
which may be highly important to a business. Related costs can
In general, routing issues may be defined on a graph of a number
extremely affect the business model of each organization. It can be
of nodes and arcs. The LRP deals with routing problems, whose
optimized by designing proper distribution networks (Prodhon and
demands are on nodes while it can be assumed on arcs. Researchers
Prins 2014). In fact, managing distribution systems have become
have recently focused on the routing problem when demands
the researchers’ interests in the last decades (Azadeh and Farrokhi-
belong to arcs instead of nodes. This kind of routing problem is
Asl 2019). Location problems are among these problems and find
called an arc routing problem (ARP). In this content, researchers
optimum locations of facilities or nodes of a supply chain.
have tried taking into account the conditions and restrictions in real
Routing vehicles, on the other hand, is an old challenging issue
applications, a variety of models, and methods for solving these
in supply chain management (SCM). The vehicle routing problem
problems. Ghiani and Laporte (1999) employed the concept of the
(VRP) was introduced by Dantzig and Ramser (1959), which is
rural postman problem (RPP) as one of the initial studies on ARP.
a well-known problem in optimization and finds the optimum
Also, Pia and Filippi (2006) considered a capacitated ARP (CARP)
routing of vehicles to fulfill the demands of customers that can be
for modeling a waste collection problem and Beullens et al. (2003)
on nodes or arcs. These two issues affect the cost of supply networks
employed a number of heuristics for a periodic CARP. In fact, there
and supply chains. Fallah et al. (2019) presented a new mathema-
are numerous studies on ARP with different assumptions and
tical model for a green periodic competitive VRP under uncertainty
conditions.
to minimize fuel consumption and maximize commercial profit-
The location-arc routing problem (LARP) is merged as the result
ability. They used differential evolution and particle swarm optimi-
of two different location and arc routing problems. There are some
zation algorithms to solve the problem. The VRP has a rich
real applications for the LARP, such as waste collection, mail
background. It has been considered in different categories (e.g.,
delivery, snow removal, road maintenance, and telecommunication
see Yao et al. 2016). The interested readers may consider different
network design (Levy and Bodin 1989; Albareda-Sambola 2015;
types of the routing problem.
Quirion-Blais, Trépanier, and Langevin 2015; Amini, Tavakkoli-
The suitable decisions on both location and routing problems
Moghaddam, and Ebrahimnejad 2017). Please note that all pro-
are vital for managers (Nadizadeh and Kafash 2019). Hence,
blems in mentioned areas are not the LARP, necessarily. For
a location-routing problem (LRP) is appeared and has been widely
instance, Rabbani et al. (2018) considered an incompatible-waste
considered by researchers. Location is a long-term and vehicle
collection LRP and developed meta-heuristics to cope with the
routing is short-term planning (Calvete, Gal, and Polo 2016).
complexity of the developed mathematical model. Moreover,
Therefore, it is necessary to consider both in an integrated model
Farrokhi-Asl et al. (2017) dealt with another waste collection pro-
where they exist simultaneously. Considering the LRP, logistics can
blem as LRP. Definitely, these problems may be converted to the
be better designed and optimized (Drexl and Schneider 2014).
LARP. It means that some node routing problems can be ARP;
Amiri, Hassanzadeh Amin and Tavakkoli-Moghaddam (2019)

CONTACT Reza Tavakkoli-Moghaddam [email protected] School of Industrial Engineering, College of Engineering, University of Tehran, Tehran, Iran
© 2019 Informa UK Limited, trading as Taylor & Francis Group
624 A. AMINI ET AL.

because it may help in reducing the complexity of some problems Martínez-Salazar et al. (2014) addressed a bi-objective transpor-
(Amini, Tavakkoli-Moghaddam, and Ebrahimnejad 2017). tation LRP, called TLRP, through developing a mathematical
As mentioned, the ARP has been addressed by examples, such as model. Considered objective functions are costs minimization and
snow removal of streets and roads, garbage collection and delivery time minimization. Time minimization aims at minimizing the
of mails. Nevertheless, in the real world applications, other issues maximum workload time among those for drivers of vehicles. As
can be considered with this approach. It occurs when some demand a result, it can be interpreted to makespan minimization. They
points can be considered as an arc of demand at the same time. In employed a weighted-sum approach to obtain exact Pareto fronts
fact, some factors (e.g., the number of demand nodes, the number and two meta-heuristics called scatter tabu search procedure for
of arcs, the configuration of streets and roads, and the distance nonlinear multi-objective optimization (SSPMO) and non-
between demand nodes) have a direct impact on the choice of the dominated sorting genetic algorithm II (NSGA-II) to obtain an
type of the problem. The low distance between demand nodes from excellent Pareto set. It is clear that they proposed meta-heuristics
each other in a joint arc suggests that these nodes should be to solve large-sized problems as well.
supplied in the row. Therefore, it is more logical to deal with such There are numerous supply chains around the world and the
problems as the ARP. LARP is likely an important part of them. So, making transporta-
One of the most common problems in this area is the supply of tion decisions along with the LARP is vital. According to the
supermarkets and restaurants. In fact, these centers are demand literature, separate optimizations are not efficient. Thus, it is
nodes. On many streets and roads, there are a lot of these nodes, required to have a comprehensive model and optimize it. This
which can equivalently form some arcs. The demand of a given arc paper addresses a transportation-location arc routing problem
equals the total demand its nodes. In fact, the nature and size of (TLARP) by developing a bi-objective mathematical model
those routing problems will determine if arc routing decreases their where objectives are costs and makespan minimization. As
complexities or not. Thus, the LARPs have a lot of real applications a matter of fact, the cost is not the only important criterion in
and should be more considered in studies. real problems. Lopes et al. (2013) categorized objective functions
Firstly, Levy and Bodin (1989) considered the LARP by which in the LRP in which costs and times are being mentioned as two
a mail delivery problem is taken into account. They used the important objectives. It means that times considerations are
location-allocation-routing (LAR) algorithm to handle the com- important. So, routing problems are always potential to consider
plexity of the problem. The LAR finds the location of depots as its the time as an important factor. For instance, Musolino, Polimeni,
first step. Then it searches for finding best arc’s allocations to and Vitetta (2016) addressed a routing problem in which link
established depots. Finally, routing decisions are taken into travel times are key factors. In this matter, Gendreau, Ghiani,
account. Then, Ghiani and Laporte (2001) used LAR and and Guerriero (2015) provided a review of time-dependent rout-
Allocation-Routing-Location (ARL) heuristics to deal with the ing problems.
LARP. Similar to the concept of LAR, decisions in ARL will be Obviously, in the field of routing problems, various studies have
made as to the sequence of its name: 1) Allocation, 2) Routing, been conducted from the viewpoint of the VRP, LRP and ARP.
and 3) Location. While there are very few studies in the LARP. On the other hand,
Lopes et al. (2014) developed a mathematical model and some the issue of supplying is an important one. In fact, the optimum
heuristics for the LARP. The authors employed tabu search (TS), deliveries from each supplier to each depot should be found at the
greedy randomized adaptive search procedure (GRASP), and same time as other decisions are taken. Ignoring this major issue,
variable neighborhood search (VNS) meta-heuristics and used may undermine the optimality of the overall process and signifi-
their proposed heuristics to enhance the meta-heuristics. In addi- cantly reduce its effectiveness. The main contributions of this paper
tion, Hashemi Doulabi and Seifi (2013) developed two mathema- are addressing the transportation issues along with the LARP,
tical models for single-depot and multiple-depot LARPs, developing a bi-objective mathematical model with cost and time
respectively. They took the advantages of meta-heuristics, as considerations, and developing some meta-heuristics for solving
well, along with some heuristics as the local search algorithms the given problems.
for those meta-heuristics. Also, Riquelme-Rodríguez, Gamache, The LARP’s graphs can be undirected, directed, or mixed
and Langevin (2016) addressed an inventory-based LARP. In one (Albareda-Sambola 2015). This study takes a mixed graph into
of the latest studies, Amini, Tavakkoli-Moghaddam, and account. According to the respective literature, the LARP belongs
Ebrahimnejad (2017) employed scenario-based approaches to to NP-hard problems. So, the TLARP will also become NP-hard. So,
study uncertain LARPs (SLARP). In addition, Tavakkoli- two meta-heuristics are employed: NSGA-II and also its combina-
Moghaddam, Amini, and Ebrahimnejad (2018) developed two tion with multi-objective late acceptance hill-climbing (MOLAHC)
types of mathematical models for a multi-product LARP, algorithm. Local search algorithms will also be implemented to
MPLARP. enhance the performance level of the algorithms.
Transportation in the industrialized world plays an important The remainder of the paper is as follows. Section 2 provides the
role in the economic development of countries by enabling the description of the problem and Section 3 develops the bi-objective
consumption of products at distant locations. Delivering a final mathematical model. Then, Section 4 provides readers with our
product to a customer involves the transfer of raw materials from solving approaches. Section 5 presents the numerical results and
suppliers to manufacturers, the transfer of semi-finished products Section 6 concludes the paper.
between the factories, and the final delivery of the final product to
customers and markets. According to transportation activities,
Problem description
shipping costs are among the main ones in logistics costs.
Therefore, efficient transportation throughout the supply chain is There is a complete mixed graph G ¼ ðN; V; AÞ where N is the set
of great importance. Using efficient transportation systems current of suppliers and V and A denotes the set of vertices and set of
markets might be expanded, new markets might be created, and arcs, respectively. V includes K as the set of potential depots and
business owners are better able to compete with their rivals. In our R as the set of the extremities arcs, V ¼ R [ K. Each arc denoted
best of knowledge, transportation issues have not been considered by ði; jÞ which means that the arc begins from vertex i and ends
in the LARP yet. at j.
 TRANSPORTATION LETTERS 625

Suppliers are supposed to supply opened depots. It is assumed ● Breakdown of vehicles is not considered
that all suppliers can supply all depots regarding their capacities. ● All required arcs must be served by one tour
Please note that the type and specifications of the vehicles can ● All arcs have allowed directions (one-way/two-way)
highly affect the problem. Angueira et al. (2017) dealt about the ● The number of deadheaded traverses in each arc is not
vehicle type choice and the traveled distance. In this paper, there is limited.
a given number of homogeneous vehicles for each potential depot ● Tours must begin from a depot and end to the same one.
in a set called P. Then, each depot is able to have jPj separate tours.
There are two kinds of traverses for arc: servicing and dead- Mathematical model
headed. A servicing traverse is one which supplies the demand of an
arc. A deadheaded one, on the other hand, is a traverse without This section aims at developing a linear bi-objective mathematical
supplying. It happens when the vehicle needs to pass an arc to reach model. The notations used by the paper are given below:
another one.
Some arcs have positive demands and the demand for an arc Parameters:
ði; jÞ equals the demand for an arc ðj; iÞ. Of course, arcs have
directions. Some of them can be traversed on both sides and some Fk Establishment cost of depot k 2 K
of them in one given side. In two-way arcs, both of them are not Cijk Servicing cost of arc ði; jÞ 2 R by depot k 2 K
required to be serviced and it is sufficient to meet the demand of Hijk Deadheading cost of arcði; jÞ 2 V by depot k 2 K
one of them. The problem is finding optimal depots among the W Hiring cost of each vehicle
potential ones, transportation decisions from supplier to opened Dij Demand of arc ði; jÞ 2 R
depots, vehicles and arcs allocations to those depots, and traverses Q Capacity of vehicles
on those arcs in an optimal routing. Optimum decisions are found Bk Capacity of depot k 2 K
regarding two objective functions: Ln Capacity of supplier n 2 N
(1). Minimizing total costs, including transportation costs Tij Transportation time of arc ði; jÞ 2 V
between suppliers and depots, depot establishment, vehicle employ- Unk Transportation time from supplier n 2 N to
ment, and required and deadheaded traversing. deport k 2 K
(2). Minimizing makespan. Makespan is the maximum tour Gnk Unit transportation cost from supplier n 2 N to
cycle among employed vehicles. depot k 2 K
Makespan minimization results in having a balanced distribu- M A sufficiently big number
tion network. Also, it can be considered as a social responsibility
concept. Vehicles’ drivers work similarly through a balanced dis- Variables:
tribution network. A tour cycle time is equal to the difference
between finishing and starting time of the vehicle at its depot in pk
xij Binary variable, which equals 1 if required arc
addition to the maximum transportation times elapsed from sup- ði; jÞ 2 R is served by vehicle p 2 P of depot k 2 K
pk
pliers to the given depot. Figure 1 presents a simple view of the yij Number of deadheaded traverses in an arc ði; jÞ 2
TLARP. All assumptions are as below: V by vehicle p 2 P of depot k 2 K
pk
fij Departed flow from the arc ði; jÞ 2 V
● Split delivery is not allowed qpk Binary variable, which equals 1 if the tour ðp; kÞ is
● There is only one type of products selected and vehicle p 2 P of depot k 2 K is
● Vehicles are homogenous established
● Service times are assumed zero zk Binary variable, which equals 1 if depot k 2 K is
● All vehicles are ready at time zeros established

Figure 1. Scheme of transportation-location arc routing problem.

626 A. AMINI ET AL.

X X X
rnk Binary variable, which equals 1 if supplier n 2 N fji 
pk pk
fij ¼
pk
Dji xji "i 2 V; "p 2 P; "k 2 K (15)
supplies depot k 2 K j2V j2V j2V
lnk Transported units supplier n 2 N to deport k 2 K
uNK
k Time by which all demands of depot k are met  
pk pk pk
u Maximum time of tours fij  Q xij þ yij "ði; jÞ 2 V; "p 2 P; "k 2 K (16)

Regarding the assumptions and notations, the bi-objective X pk

mathematical model for the TLARP is developed. fik ¼ 0 "p 2 P; "k 2 K (17)
X X X pk
i2V
Minθ1 ¼ lnk Gnk þ zk Fk þ xij Cijk
n2N;k2K k2K ði; jÞ2V;p2P;k2K
X X pk
X pk
X lnk  fki "k 2 K (18)
þ yij Hijk þ qpk W (1) n2N p2P;j2V
ði; jÞ2V;p2P;k2K p2P; k2K
X
lnk  Ln "n 2 N (19)
Min θ2 ¼ u (2) k2K

s.t.
X   lnk  Ln rnk "k 2 K; "n 2 N (20)
pk pk
xij þ xji ¼1 "ði; jÞ 2 R (3)
k2K;p2P
k  rnk Unk
uNK "p 2 P; "k 2 K; "n 2 N (21)
X  
pk
xij þ yij
pk
¼0 "ði; jÞ 2 ; (4) X  pk pk

k þ
u  uNK xij þ yij Tij "p 2 P; "k 2 K; "n 2 N (22)
k2K;p2P
ði; jÞ2V

X  pk pk

xij þ yij qpk "ði; jÞ 2 V (5) pk
xij ; zk ; qpk ; rnk 2 f0; 1g "ði; jÞ 2 V; "p 2 P; "k 2 K; "n 2 N
k2K;p2P
(23)
X pk pk
 X
pk pk

xij þ yij  xji þ yji ¼ 0 "j 2 V; "p 2 P; "k 2 K pk
i2V i2V
yij  0 & Integer "ði; jÞ 2 V; "p 2 P; "k 2 K (24)
(6)
pk
fij ; lnk ; u  0 "ði; jÞ 2 V; "p 2 P; "k 2 K; "n 2 N (25)
qpk  zk "p 2 P; "k 2 K
X pk  Objective function (1) minimizes total costs, including units’ trans-
pk
xkj þ ykj ¼ qpk "p 2 P; "k 2 K (7) portation from suppliers to depots, the opening of depots, traver-
j2R sing, and hiring vehicles. Objective function (2) minimizes the
maximum time of tours which is already called makespan.
X pk pk
 Constraint (3) enforces the model to find a tour for each required
xkj þ ykj ¼ qpk "p 2 P; "k 2 K (8) arc in an allowed direction and constraint (4) guarantees that
j2R
prohibited directions are not selected. Note that [ is the set of
pk pk
prohibited directions. Constraint (5) declares that xij and yij must
X pk pk

xjk þ yjk ¼ qpk "p 2 P; "k 2 K (9) equal zero if the respective tour is not selected.
j2R One vital condition of such problems is about the continuity of
tours. It means that if there is an entry for a node there must be an
X exit for it. Hence, Constraint (6) guarantees the continuity of tours.
qpk  jPj "k 2 K (10) Both kinds of arcs (servicing and deadheaded) can keep the con-
p2P
tinuity of tours. According to Constraint (7), tour ðp; kÞ is allowed
to be established when depot k is opened. It means that an unse-
X X  
pk
fkj 
pk
Dij xij  M 1  qpk "p 2 P; "k 2 K (11) lected depot does not have any assigned tour. Constraint (8) states
j2R "ði; jÞ2V
that established tours must begin from a depot and Constraint (9)
emphasizes that those tours must finish their service in a depot. It is
X X   clear that the beginning and ending depots of a tour are the same.
pk pk
fkj  Dij xij þ M 1  qpk "p 2 P; "k 2 K (12) Considering Constraint (10), a number of tours for each depot
j2R "ði; jÞ2V should not exceed the maximum number of its vehicles.
By Constraints (11) and (12), output units of each tour of depots
X pk
must be the total demands of that tour. Also, total units transported
fkj  Q "p 2 P; "k 2 K (13) by a vehicle should not exceed its capacity and total units sent by
j2R
a depot should not exceed its capacity, which is clear in Constraints
(13) and (14), respectively. By Constraint (15), it is guaranteed that
X pk the difference between an entry flow and its departure equals the
fkj  Bk "k 2 K (14)
j2R;p2P
demands of the arc if there is a servicing traverse. Of course,
regarding constraint (16), the flow of an arc equals zero if there is
 TRANSPORTATION LETTERS 627

not any traverse. Also, returning flow of tour to its depot must be ● Elements 1 to jN jjK j: There is a sequencing rule in these cells:
zero. It is shown by Constraint (17). the higher value the sooner turn to transport items from
There must be some constraints on the supply phase of the a supplier to a depot.
problem. Constraint (18) displays that transported units to each ● Elements jN jjK j þ 1 to jN jjK j þ jAR j: for a given required arc
depot cannot be lower than its demand. In addition, due to there is a depot by rounding up the result of multiplying its
Constraint (19) total sent units of a supplier should not exceed its value to |K|.
capacity and regarding Constraint (20) transportation between ● Elements jN jjK j þ jAR j þ 1 to jN jjK j þ 2jAR j: for a given
supplier n and depot k is allowed when rnk ¼ 1. Finally, in order required arc there is a direction by rounding up the result of
to calculate the makespan, Constraint (21) measures the time by multiplying its value to its number of possible directions (one/
which the demand of depot k is met and Constraint (22) measures two).
the time of tours due to the servicing and deadheaded traverses. ● Elements jN jjK j þ 2jAR j þ 1 to jN jjK j þ 3jAR j: for a given
Finally, Constraints (23) to (25) define variables. required arc, there is a vehicle by rounding up the result of
multiplying its value to |P|.
● Elements jN jjK j þ 3jAR j þ 1 to jN jjK j þ 4jAR j: final routing
Solving approaches of tours is determined by knowing the turn of required arcs to
be assigned to their respective depots. Higher values deter-
The mathematical model is considered for solving and achieving
mine the higher priorities of required arcs for sequencing.
optimum results. This paper employs the augmented ε-constraint
Considering the turns and routings, deadheaded traverses
algorithm in order to find the true Pareto frontiers. In addition, two
will be released as well.
meta-heuristics are used with and without a local search algorithm.
Then, there are four meta-heuristics along with the augmented ε-
Figure 2 provides an example of the solution structure. In this
constraint algorithm.
example, there are two suppliers (S1 and S2), two potential depots
(D1 and D2), and five required arcs (A1 to A5). Therefore, the
solution’s matrix has 24 elements. According to the presented
Augmented ε-constraint descriptions, elements 1 to 4 indicate the priorities in sending the
Let O denote the set of objective functions with jOj members. goods from the suppliers to the depots. The first two elements are
In ε-constraint method jOj  1 objectives appear in constraints for the S1. In fact, it has an element for each depot. Similarly,
and model is optimized regarding the remaining the second two elements belong to the S2. With the descending
  objective
function. For example, fmax f1 ðxÞ; f2 ðxÞ . . . fjOj ðxÞ jxSg con- order of the elements mentioned, we conclude that the priorities for
verts to fmaxðf1 ðxÞÞj; f2 ðxÞ  ε2 ; f3 ðxÞ  ε3 . . . fjOj ðxÞ  εjOj ; xSg. sending goods are S1-D1, S2-D1, S1-D2, and S2-D2, respectively. In
There is an extension of this method, called augmented ε- fact, the first priority is supplying is with link S1-D1. If all demands
constraint, provided by Mavrotas (2009) in which the final of D1 are provided by S1, then other links to D1 are removed from
! the solution. In this way, all links from suppliers and depots are
P
jOj
model is fmax f1 ðxÞ þ eps  si =ri j; f2 ðxÞ  s2 ¼ ε2 ; f3 ðxÞ  found and the transportation problem is solved.
i¼2 Then, 20 elements will remain. These elements are arranged in
s3 ¼ ε2 . . . fjOj ðxÞ  sjOj ¼ εjOj ; xS; si  0g. In this model eps is different phases for A1 to A5 arcs. In fact, elements 5 to 9 belong to
a very small number and ri is the range of the i-th objective arcs A1 to A5, respectively. It is also true for elements 10 to 14, 15 to
function. 19, and 20 to 24. Referring to the previous descriptions and the
The range of a given objective function is measured by finding example:
the difference between its optimal level and its worst level obtained
during finding optimal values of other objective functions. This ● Elements 5 to 9 allocate the required arcs to the depots. With
range is divided into Lε þ 1 sections, where Lε is a given integer two depots in this example, arcs A1 to A5 are assigned to D1,
number determined by the decision makers. Then, the model con- D1, D2, D2, and D1, respectively.
siders an iterative procedure to find the optimal Pareto set. ● Elements 10 through 14 determine the direction of the traver-
Interested readers may refer to Mavrotas (2009) to find the details sing of each arc in order to supply them. Each of them can
of the augmented ε-constraint method. have one or two allowed directions. Assuming that there are
two directions for each arc in this example, the demand for
arcs A1 to A5 are provided in directions 1, 1, 1, 2, and 1,
Meta-heuristics respectively. It should be noted that as indicated, each arc
can be represented by its vertices. We assume that in the two-
The paper on hand develops the NSGA-II and also its combination way arc (i, j), with the condition j > i, directions 1 and 2 are
with MOLAHC (hybrid) in order to find a suitable Pareto frontier i-j and j-i, respectively.
in a reasonable time. Also, there is a local search (LS) algorithm to ● Elements 15 to 19 assign the required arc to the vehicles.
improve the quality of meta-heuristics. Therefore, there are totally Assuming two available vehicles (V1 and V2) in this example,
four meta-heuristics: NSGA-II without LS (H1), NSGA-II with LS the arcs A1 to A5 are allocated to vehicles V1, V1, V2, V1, and
(H2), hybrid without LS (H3), and hybrid with LS (H4). Following V2, respectively.
subsections provide explanations about solution representation, ● Elements 20 to 24 determine the priority of the required arcs
introduced meta-heuristics, and local search. to be placed in the tours. With the found solutions, it is clear
that arcs A1 and A2 are assigned to depot D1 and vehicle V1. As
Solution representation a result, these two arcs are on a common tour. To find out the
There are jN jjK j þ 4jAR j elements in the matrix presenting solu- first one in the routing, we look at the value of elements
tions between 0 and 1. Please note that AR is the set of required arcs equivalent to these arcs among elements 20 to 24. Because
with |AR| members. In this matrix: the value for A1 is greater than the value of A2 one, this tour
628 A. AMINI ET AL.

Figure 2. Solution’s matrix structure.

starts with A1 and then goes to A2. As a result, the tour of solutions in each iteration. Also, the termination criterion of
vehicle V1 of this depot is defined as D1-A1-A2-D1. Similarly, the algorithm is reaching the final iteration, maxitI. The pro-
other tours are formed. posed NSGA-II uses the ranking and crowding distance con-
cepts. The members of a given front do not dominate each
● With the given solution, the opened depots are automatically other and at least one of them dominates a member of a front
identified. In addition, deadheaded traverses are found. For with a worse rank.
example, in the same tour D1-A1-A2-D1, if the arcs A1 and A2 The crowding distance is used to maintain a diversity of
are represented by (i, j) and (i’, j’), this tour is D1-i-j-i’-j’-D1 responses on the Pareto-optimal front. To calculate this value
and deadheaded traverses D1-i, j-i’, and j’-D1 are formed. for each point on a certain front, the points before and after the
target point are selected using and the crowding distance of it is
P jþ1 j1
The most important advantages of this kind of solution repre- measured using cdj ¼ 2o¼1 θθmax
o θo
θmin
according to the two objec-
sentation are simply to generate it and no need to modify it in o o

different operators. It is also possible to easily define various opera- tive functions of the TLARP. In this equation, θjo is the value of
tors on this matrix. For this reason, this paper employs it. the o-th objective function of point j, θmin
o is the minimum value
One point in developing these algorithms is how to deal with the of the o-th objective function, and θmax
o is also its maximum one
constraints on the capacity of depots and vehicles and the allowed on the Pareto front. In addition, given that the first point of
directions of the arcs. In fact, with the answer given by a matrix, the a front has the best value in an objective function and the last
capacity limitations may be violated or forbidden directions are point is the best solution in another objective function their
determined for deadheaded traverses. This paper solves this pro- crowding distances are assumed infinite. It should be noted
blem by adding penalties on the first and the second objective that more value in this criterion is better.
function value (OFV) of such solutions. In this method, by multi- Two main operators of the genetic algorithms are crossover
plying the amount of violation of the capacities in a positive num- and mutation. There are two rates for crossover (pcr) and muta-
ber and adding it to OFVs the penalty is implemented. In addition, tion (pmu), which should be valued by the decision maker. This
for each forbidden direction, we add a positive number to the paper employs three types of crossovers and four types of muta-
OFVs. Keeping the infeasible solution in the algorithm may help tions. One-point, two-point, and weighted mean are the consid-
it in finding ones that are more suitable. ered crossover operators, Figure 4. One of them is randomly
chosen for each operation per iteration. In the weighted mean
NSGA-II operator, a random coefficient, named wt, is generated in the
Deb et al. (2002) developed a multi-objective evolutionary algo- range [0 1]. Regarding wt, Child1 ¼ wt  Paretnt1 þ ð1  wtÞ 
rithm on the basis of the genetic algorithm, called NSGA-II. It is Parent2 and Child2 ¼ ð1  wtÞ  Paretnt1 þ wt  Parent2.
a well-known population-based algorithm. Figure 3 presents the Figure 4 presents an example by wt ¼ 0:8 for this operator. Also,
Pseudo code of NSGA-II. It begins and continues with npop there are four mutation operators in our NSGA-II, Figure 5:

Figure 3. Pseudo code of NSGA-II.

 TRANSPORTATION LETTERS 629

Figure 4. Employed crossover operators.

Figure 5. Employed mutation operators.

● Regeneration: it regenerates a randomly selected range of the ● Replacement: it replaces the value of two randomly selected
solution. elements with each other.
● Complement: it selects a random range of the solution and set ● The algorithm chooses one of the above-mentioned operators
the value of those elements equal to 1  valueP where valueP for each solution per iteration randomly.
are the previous values of those elements.
● Mirror: it perform
● s like a mirror. For instance, if elements 3 to 6 are randomly MOLAHC
selected, valueN3 ¼ valueP6 ,valueN4 ¼ valueP5 , valueN5 ¼ valueP4 , Burke and Bykov (2017) proposed late acceptance hill-climbing
and valueN6 ¼ valueP3 where valueNi is the new value and (LAHC) algorithm as a simple local search method which works
valuePi is the previous one of i-th element. with one solution. It accepts movements by introducing a length
630 A. AMINI ET AL.

(Lh). In fact, a movement is compared with the solution of Lh-th significantly less amount of time to be completed in comparison
earlier iteration. Then it is accepted if it wins in this comparison. with NSGA-II. On the other hand, its simple procedure might cause
The authors provided the method for single-objective problems. a low-quality solution. Then, this paper proposes a hybrid algo-
This paper considers a multi-objective concept in LAHC, called rithm by which the initial solution of MOLAHC is driven from
MOLAHC. NSGA-II. In fact, NSGA-II will provide a set of solutions as its best
In the MOLAHC, we are looking for appropriate Pareto fron- Pareto solutions at the end of its procedure. One of these solutions
tiers. Therefore, to evaluate the new solution in each iteration, we is randomly selected for the initial solution of MOLAHC.
compare it with the Pareto archive, which is created and updated Therefore, a quality initial solution is provided and we take advan-
during the process. In the algorithm, we also accept non-improved tage of MOLAHC in solving time, simultaneously.
solutions with a certain probability that decreases in each iteration.
In fact, in the initial iterations, we can probably make better Local search
searches using weaker solutions. Then in higher iterations, this Generally, meta-heuristics need to be enhanced via local search
chance is reduced. The initial probability of accepting a non- algorithms. It may take more time in solving procedures. A local
improved solution is 0.5 in this paper, which is reduced in each search method is employed in this paper, in which the value of two
iteration by dividing it into the index of the iteration. The termina- elements of the solution’s matrix is replaced with each other. This
tion condition is as same as what is proposed for the NGSA-II. The procedure is done only once for each solution; because an iterative
maximum number of iterations is called maxitII in the MOLAHC. local search may considerably increase the solving time. If the
In each iteration using the current solution (sol), a new one modified solution dominates the previous one, the modification is
(solnew) is constructed and compared with the sol and the v-th accepted; else, we keep the previous solution.
member of set H. It should be noted that v ¼ mod ðLh; itÞ
where it expresses the iteration’s index. In addition, H is a set
including Lh solutions that is constructed during the process. We Numerical experiments
consider ½pnc  NC elements out of NC elements of a solution to This section deals with analyzing the developed model and meta-
change and create the solnew. to If solnew dominates sol or Hv, they heuristics. Accordingly, 20 test problems are randomly generated
will be replaced by solnew. Otherwise, depending on whether they using the pattern and specifications presented by Tables 1 and 2.
dominate solnew or not, they are replaced with a probability or by We generate transportation times by calculating the Euclidean
comparing their crowding distances. Figure 6 fully describes the distance between the nodes. The coordinates, called X and Y, are
steps of this algorithm. Idle is defined as the number of unsuccessful generated by the presented pattern by Table 1. Parameter ρ is the
consecutive iterations in construction a solnew that dominates sol. total number of directions in A and ρR is the same one in AR. Note
This parameter is used along with the given maximum iterations as that one-way and two-way arcs have one and two directions,
the stopping criteria of the algorithm. respectively. In fact, the problem is analyzed for a different number
Figure 6 presents the Pseudo code of the proposed MOLAHC. of arcs from 8 to 103 arcs. Each test problem is considered to have 2
Please note that our MOLAHC employs the mutation operators and 4 available vehicles for each potential depot, separately. It
which are already explained to create a new solution. It is a simple means that there are forty test problems overall. Note that all
single-solution-based algorithm and is expected to need problems are solved on a desktop Core i7 2600K PC, 3.40 GHz

Figure 6. Pseudo code of MOLAHC.

 TRANSPORTATION LETTERS 631

Table 1. Data generation pattern. Table 4. Obtained variables of E6.

Parameter Generation Pattern Level 1 Level 2
Fk Uniformð5000; 50000Þ jPj ¼ 2 θ1 2–1 & 1–3 1-7-6-5-4-5-1 (1)
Cijk Uniformð100; 300Þ 1-5-7-6-4-8-5-1 (2)
Hijk Uniformð10; 100Þ 3-4-8-5-3 (1)
3-4-7-8-5-3 (2)
Dij Uniformð100; 1000Þ
 

    θ2 1-1 & 2-2 1-4-7-8-1 (1)
Q
Max max Dij ; Unifrom max Dij =2jK j; max Dij =jK j 1-5-7-6-4-1 (2)
i; j i; j i; j
 2-6-5-4-2 (1)
Bk     2-5-8-4-2 (2)
Unifrom max Dij =jK j; 1:5 max Dij
i; j i; j jPj ¼ 4 θ1 2-1 1-6-7-4-1 (1)
W 1000 1-5-7-8-5-1 (2)
Gnk Uniformð0; 20Þ 1-8-4-5-1 (3)
Ln     1-5-6-7-1 (4)
Unifrom max Dij =jNj; 1:5 max Dij θ2 1-1 & 1-2 1-4-5-6-4-1 (1)
i; j i; j
X; Y Uniformð0; 1000Þ 1-5-7-1 (2)
1-4-7-8-1 (3)
1-4-7-6-4-1 (4)
2-5-8-4-2 (1)
Table 2. Characteristics of examples. 2-5-4-2 (1)
Examples jNj jK j jV j jAj ρ AR ρR
E1 2 2 5 9 16 3 4
E2 2 2 6 14 26 6 10
E3 2 2 7 20 36 10 16 establish hub. Thus, less number of hubs will likely result in having
E4 2 3 6 12 23 3 5 lower costs, because the cost of depot establishment is typically
E5 2 3 7 18 34 6 10 much more than employing a vehicle. In addition, more numbers
E6 2 3 8 24 46 9 16
E7 3 2 5 8 15 2 3 of tours will likely cause a lower makespan.
E8 3 2 6 14 26 6 10 There are two levels in the supply chain of the TLARP: 1.
E9 3 2 7 20 38 10 18 Transportation from suppliers to depots; 2. Routings. Table 4 pre-
E10 3 3 6 12 23 3 5 sents the effect of the number of vehicles and objective functions on
E11 3 3 7 18 36 6 12
E12 3 3 8 25 47 10 17
these levels in E6 as a randomly selected example. Indeed, it reports
E13 5 5 13 67 128 27 48 the extremities of the optimal front of E6 for jPj ¼ 2 and jPj ¼ 4.
E14 5 5 15 91 164 41 64 Consider θ1 as the only objective function of the optimization. As
E15 5 6 14 75 138 27 42 already mentioned it is expected to have less opened depots when
E16 5 6 16 101 186 41 66 there are more vehicles available. Table 4 shows that depots 1 and 3
E17 6 5 13 67 126 27 46
E18 6 5 15 91 170 41 70 are established when jPj ¼ 2, while only depot 1 is enough when
E19 6 6 14 74 132 26 36 jPj ¼ 4. Of course, it depends on the capacity of depots. For exam-
E20 6 6 16 103 192 43 72 ple, in E1 there is no difference for the decision makers between
having 2 vehicles for each depot or 4. Here the column ‘Level 1ʹ
displays the links between suppliers and depots with the pattern n 
with 16 GB RAM. Each problem is optimally solved by GAMS k for n 2 N and k 2 K. Also, the column ‘Level 2ʹ shows the routing
optimization software. Regarding the ε-constraint method para- of arcs by the nodes with the pattern k  i . . . j  kð pÞ, where k 2 K,
meters, eps and Lε are assumed 0.001 and 25, respectively. It is i . . . j 2 V, and p 2 P.
worth mentioning that E13 to E20 cannot be handled by GAMS in Now consider θ2 as the only objective function. Table 4 demon-
a significant amount of time. strates that decision variables are optimally found differently.
Table 3 presents the final results of E1 to E12 in the aspects of Although there are the same tours when jPj ¼ 2 in both objective
optimum objective functions values and the number of true Pareto functions considerations, there are modifications in both decision
solutions (NTPS). It is expected to have higher costs and times levels. In jPj ¼ 4 it is found that the number of tours increases along
when there is a more strict limitation on the number of vehicles. with other modifications when θ2 is optimized.
Having more vehicles available, decision makers do not have to E10 is randomly selected as another example to elaborate. Figure 7
establish extra hubs and are able to assign more tours to each displays the true Pareto solutions (TP) of E10 for jPj ¼ 2 and jPj ¼ 4.
One can see the trend of the OFVs schematically. It is obvious that the
decision maker needs to identify her/his priorities to make a true
Table 3. Optimum value of each objective function and number of true Pareto
solutions.
decision. She/he has a set of solutions in the Pareto-optimal front.
Consider two first solutions in jPj ¼ 2 of E10. The coordinates are
jPj ¼ 2 jPj ¼ 4
(36,048, 3382.856) and (38,834, 2809.255), respectively. Although
Example θ1 θ2 NTPS θ1 θ2 NTPS choosing the second one means of decreasing the makespan by almost
E1 20121 2245.685 3 20121 2245.685 4 17%, total costs increase by almost 7.7%. It is a very significant decrease
E2 46463 2110.963 3 22653 2095.124 3 in the makespan. In a condition, in which the decision maker puts her/
E3 90326 2727.647 4 78472 2493.543 9
E4 31295 1882.163 7 31295 1882.163 4 his best efforts in having a low makespan, this transition makes a good
E5 108538 2155.717 4 60373 1832.749 5 sense. On the other hand, it can be more expensive for her/him to have
E6 61198 2329.641 5 31984 2289.86 6 such a solution. Maybe the decision maker defines a threshold for such
E7 13001 2232.719 4 13001 2232.719 4 situations. For instance, she/he can tolerate up to 10% increase in the
E8 85509 2458.665 7 85509 2375.543 7
E9 53280 2858.576 12 50654 2693.833 5
optimum costs and a more increase cannot be acceptable. This analysis
E10 36048 1846.174 11 15481 1846.174 9 is valid for all pair of solutions in the TP.
E11 76860 2487.373 7 42168 2362.034 8 Table 5 reports four criteria of the solutions of the TP for E10,
E12 136735 2350.477 8 99082 2084.493 9 where jPj ¼ 2 and jPj ¼ 4: the number of links from suppliers to
632 A. AMINI ET AL.

Figure 7. TP of E10.

Table 5. Optimum values of NoL, NoD, NoT, and NoDT results of E10. Table 6. RSM levels and results.
jPj ¼ 2 jPj ¼ 4 NSGA-II MOLAHC
Solution No. NoL NoD NoT NoDT NoL NoD NoT NoDT Level npop maxitI pcr pmu Lh maxitII pnc
1 2 2 3 7 2 1 3 6 Minimum (−1) 50 100 0.6 0.05 2 5000 0.01
2 3 2 3 6 2 1 3 6 Maximum (1) 200 500 0.8 0.15 10 10,000 0.5
3 4 2 3 6 2 2 3 6 Optimum 145 294 0.79 0.06 10 5000 0.5
4 3 2 3 6 3 2 3 6
5 3 2 3 6 3 2 3 6
6 3 2 3 6 2 1 3 6
7 2 2 3 6 2 1 3 6 higher NoDT in comparison with its lower bound is determined
8 2 2 3 6 2 2 3 6 for the first solution of the TP of E10 where jPj ¼ 2. It means that
9 3 2 3 6 2 2 3 6
10 2 2 3 6 nobody can guarantee that there are minimum NoDT for all
11 2 2 3 6 optimal solutions.
Regarding the complexity of the model, an NSGA-II and
a hybrid NISGAII-MOLAHC are developed in Section 4. In addi-
tion, a local search method is introduced in order to enhance the
depots (NoL), the number of established depots (NoD), the number methods. So, there are totally four meta-heuristics, namely NSGA-
of tours (NoT), and the number of deadheaded traverses (NoDT). II (H1), NSGA-II+LS (H2), hybrid (H3), and hybrid+LS (H4).
Here the first solution (solution 1) is one with the best possible value A response surface methodology (RSM) is a method in the design
for OFV1 and the last one (solution 11 for |P| = 2 and 9 for |P| = 4) is of experiments (DOE) employed to find suitable levels for parameters
one with the best possible value for OFV2. The maximum value of of meta-heuristics. Considering the RSM, Table 6 presents the given
NoL is jN j  jK j. The higher value of NoL means there are more links and optimal levels of NSGA-II and MOLAHC. Note that npop and
between the suppliers and depots and results in having more trans- maxitI of NSGA-II in the hybrid are assumed half of and pcr and
portation costs. However, the answers demonstrate that there is not pmu are similar to what is found for the NSGA-II in the RSM.
any specific trend from the first to the last solution of a TP. This paper measures the performance of meta-heuristics
In jPj ¼ 2 one can see that NoD ¼ 2 for all solutions, while for using nine multi-objective metrics, in which jOj is the number
jPj ¼ 4 it is not same for all ones. It is also obvious that we need 3 of objective functions, jOPj is the number of obtained Pareto
tours regardless of the NoD. Each tour ate least needs two dead- solutions by the respective method, jTPj is the number of true
headed traverses: the outbound arc from the depot at the first of Pareto solutions by the ε-constraint algorithm, di is the
the tour and the inbound one, which occurs at the end of the Euclidean distance between the i-th solution of the obtained
tour. As a result, NoDT  NoT  2. Although deadheaded tra- Pareto set and the closest solution to it in the true Pareto set,
verses are not desired for the proposed objective functions, the  is the average of di for i 2 OP:
and d
 TRANSPORTATION LETTERS 633

Figure 8. Obtained Pareto solutions of E1 to E12 where |P| = 2.

● Spacing (SP): It represents the performance of the model distance between the first and last points of the obtained
in the aspect of convergence and measured by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pareto solutions. A higher value of MS is preferred.
PjOPj  2 ● Delta (Δ): This metric measures the diversity of a method. It is
i¼1 d  di . A lower value of SP is preferred.
1
jOPj1 PjOPj1
df þdl þ di d
calculated by d þd þði¼1jOPj1Þd
where df and dl are Euclidean
● Maximum Spread (MS): It is a kind of diversity metric calculated f l

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi distances between boundary solutions of OP and the extreme

PjOj solutions (Deb et al. 2002). A higher value of Δ is preferred.
by i¼1 maxfdðai ; bi Þg where dðai ; bi Þ is the Euclidean
634 A. AMINI ET AL.

Figure 9. Obtained Pareto solutions for E20.

● Generational Distance (GD): It determines the performance of ● Error Ratio (ER): It is a metric checking the success of a given
PjOPj
a method from the convergence point of view and calculated
qffiffiffiffiffiffiffiffiffiffiffiffiffi
PjOPj 2ffi
e
algorithm and calculated by jOP i¼1 i
j where ei ¼ 1 if the i-th
di
by
i¼1
. Lower value of GD is preferred. solution of the obtained Pareto solution is belong to the true
jOPj
one and ei ¼ 0, otherwise. A lower value of ER is preferred.
● Inverted Generational Distance (IGD): It employs the concept
of GD from another point of view. Then, it belongs to the Comparing obtained Pareto solutions of heuristics with the true one
metrics which q
report the convergence of a solving method and
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi for E1 to E12 when jPj ¼ 2, Figure 8, it is shown that heuristics were
PjTPj 0 2 able to find exact solutions in some problems and very close ones in
i¼1
ðdi Þ 0
calculated by jTPj where di is the Euclidean distance others. This finding is still valid when jPj ¼ 4. In larger examples, on
between the i-th solution in TP and the closest solution to it in the other hand, different performances appear. Figure 9 compares the
OP. A lower value of IGD is preferred. findings of heuristics for E20 for both cases jPj ¼ 2 and jPj ¼ 4. The
● Hypervolume (HV): It measures the covered solution space by impact of local search is clear in these comparisons.
a given algorithm according to a reference point. Readers are All nine metrics are measured for heuristics. Figure 10 displays
referred to (Zitzler and Thiele 1999) in order to know more the boxplot of each one for each heuristic separately. In order to
about this metric. A higher value of HV is preferred. compare the heuristics statistically, it is required to employ a true
● Maximum Pareto Fron Error (MPFE): It belongs to conver- test. One-way ANOVA is a test by which the equality of means of
gence metrics and calculated by maxfdi g. A lower value of the a number of groups is tested. One important criterion for using this
i2OP
MPFE is preferred. test is having Normally-distributed data. Since our data are rejected
● Number of Pareto Solutions (NPS): it counts the number of in the normality test, Kruskal-Wallis test, as a nonparametric one, is
obtained Pareto solutions. A higher value of the NPS is preferred. selected. This test is performed for each metric, independently.
 TRANSPORTATION LETTERS 635

Figure 10. Boxplot of performance metrics.

Table 7. Difference analysis of the heuristics.

SP M Δ GD IGD HV MPFE NPS ER
Kruskal-Wallis 0.474 0.543 0.641 0.000 0.013 0.760 0.000 0.061 0.000
Wilcoxon H1 vs. H2 - - - 0.001 0.001 - 0.000 - 0.013
rank-sum H1 vs. H3 - - - 0.299 0.567 - 0.001 - 0.000
H1 vs. H4 - - - 0.473 0.324 - 0.000 - 0.000
H2 vs. H3 - - - 0.000 0.004 - 0.300 - 0.000
H2 vs. H4 - - - 0.005 0.114 - 0.858 - 0.000
H3 vs. H4 - - - 0.089 0.209 - 0.000 - 0.000

Obtained P-Value of this test under 0.95 confidence level is shown in a kind of quality metric. Considering Figure 10, H2 and H4 are the
Table 7. These values prove that our heuristics do not perform two best algorithms in this metric. Although all heuristics are not
equally in metrics GD, IGD, MPFE, and ER. In fact, for the rest different from each other in GD, IGD, and MPFE, this conclusion is
five metrics, the hypothesis of having equal heuristics is rejected. also valid for them; because in case of comparison between two
Wilcoxon rank-sum test, a nonparametric pairwise comparison significantly different heuristics in these metrics, it is found that
test, is employed for those four metrics. Table 7 presents the final NSGA-II+LS and hybrid+LS are two best algorithms.
results and shows all found different heuristics. For instance, these This preference is found regardless of heuristics’ solving times.
values confirm that there is not equality between heuristics in ER as Table 8 presents solving times of GAMS and heuristics. Firstly,
636 A. AMINI ET AL.

Table 8. Solving times (in seconds).

P=2 P=4
GAMS H1 H2 H3 H4 GAMS H1 H2 H3 H4
E1 2.76 40.01 62.26 14.40 24.79 5.22 38.03 53.77 11.87 17.83
E2 10.25 48.52 83.67 15.86 33.08 18.81 44.28 64.05 16.69 26.62
E3 24.58 39.46 55.60 13.55 22.58 295.98 37.73 59.58 13.95 18.79
E4 11.78 35.51 52.05 11.05 16.44 23.03 36.82 49.30 9.04 13.53
E5 38.39 41.98 59.99 12.17 24.62 67.88 37.80 52.70 11.19 19.22
E6 197.44 38.76 60.18 12.29 22.18 1743.55 38.02 56.43 12.41 16.83
E7 3.92 33.44 45.32 8.22 12.13 4.31 35.36 44.83 8.68 11.83
E8 24.95 31.95 49.48 11.40 20.04 360.67 33.70 48.42 12.28 18.65
E9 257.05 41.09 56.08 11.45 18.02 3401.05 37.19 49.75 11.22 15.33
E10 11.19 36.39 50.54 13.32 20.84 32.91 35.40 47.97 11.82 15.79
E11 120.02 49.00 76.92 18.16 35.35 1436.37 39.94 58.52 14.49 22.83
E12 1729.90 40.38 67.19 18.16 30.91 90242.36 46.19 58.45 23.41 21.97

GAMS’s solving times prove the need for other methods. Then, it H4 needs significantly less time to find OP. It proves that H4 take
can be found that hybrid and hybrid+LS need remarkably lower the advantages of NSGA-II (in the aspect of quality) and MOLAHC
times to find appropriate solutions. Regarding this fact and few (in the aspect of speed), simultaneously. An important suggestion
differences between them and NSGA-IIs in Figure 9, it is shown that for the research in this problem is to deal with some methods in
developed hybrid heuristics are suitable methods in order to find order to find a solution among those ones. In addition, exact
quality solutions in a significant amount of time. methods and uncertainty studies are strongly recommended as
future research in this area.
Conclusion
Disclosure statement
This paper on hand addressed a transportation-location-arc routing
problem (TLARP) by developing a bi-objective mathematical No potential conflict of interest was reported by the authors.
model that was developed regarding the previous ones in the
relative literature. It contains two objective functions: total cost ORCID
minimization and makespan minimization. It is worth mentioning
that makespan minimization deals with controlling the time of Reza Tavakkoli-Moghaddam http://orcid.org/0000-0002-6757-926X
distribution while it can be even considered as a social responsi-
bility criterion. Considering the bi-objective model, an augmented References
ε-constraint algorithm was employed to find the optimal and true
Pareto solutions. According to the background of the LARP, it Albareda-Sambola, M. 2015. Location Science. Edited by, G. Laporte, S. Nickel,
belongs to NP-hard problems. Its combination of transportation and F. Saldanha Da Gama. Cham: Springer International Publishing. Print
ISBN: 978-3-319-13110-8. Online ISBN: 978-3-319-13111-5.
problems makes it more complex. In addition, this paper employed Amini, A., R. Tavakkoli-Moghaddam, and S. Ebrahimnejad (2017) “Scenario-
an NSGA-II as a population-based method and its combination Based Location Arc Routing Problems: Introducing Mathematical Models.”
with a MOLAHC algorithm. MOLAHC is a single-solution-based in Proceedings of the Eleventh International Conference on Management
method and is expected to be a less time-consuming algorithm. So, Science and Engineering Management, 511–521. Japan: Springer.
Amiri, M., S. Hassanzadeh Amin, and R. Tavakkoli-Moghaddam. 2019. “A
in the hybrid method, NSGA-II builds a quality initial solution, Lagrangean Decomposition Approach for a Novel Two-echelon Node-
then MOLAHC takes its procedure into account and finds a Pareto- based Location-routing Problem in an Offshore Oil and Gas Supply
solutions set. These two methods and having/not having a local Chain.” Transportation Research Part E: Logistics and Transportation
search algorithm (LS) result in four heuristics, namely NSGA-II Review 128: 96-114. doi: 10.1016/j.tre.2019.05.014.
(H1), NSGA-II+LS (H2), hybrid (H3), and hybrid+LS (H4). The Angueira, J., K. C. Konduri, V. Chakour, and N. Eluru. 2017. “Exploring the
Relationship between Vehicle Type Choice and Distance Traveled: A Latent
parameters of these heuristics were estimated using RSM. Then, 40 Segmentation Approach.” Transportation Letters 11 (3): 1–12.
test problems were generated to evaluate the mathematical model Azadeh, A., and H. Farrokhi-Asl. 2019. “The Close – Open Mixed Multi Depot
and proposed heuristics. Vehicle Routing Problem considering Internal and External Fleet of
Numerical results are divided into two sections: true Pareto (TP) Vehicles.” Transportation Letters 11 (2): 78–92. doi:10.1080/
19427867.2016.1274468.
set resulted by the augmented ε-constraint algorithm and the Beullens, P., L. Muyldermans, D. Cattrysse, and D. Van Oudheusden. 2003.
obtained Pareto (OP) set resulted by heuristics. TP of the test “A Guided Local Search Heuristic for the Capacitated Arc Routing Problem.”
problems displayed the contrast existed between objective functions European Journal of Operational Research 147 (3): 629–643. doi:10.1016/
and performance of the developed mathematical model. However, S0377-2217(02)00334-X.
it was impossible to find the TP of problems in some large test Burke, E. K., and Y. Bykov. 2017. “The Late Acceptance Hill-Climbing
Heuristic.” European Journal of Operational Research 258 (1): 70–78.
instances. OP of examples was considered to compare the perfor- doi:10.1016/j.ejor.2016.07.012.
mance of heuristics. Nine multi-objective performance metrics Calvete, H. I., C. Gal, and L. Polo. 2016. Modeling and Simulation in Engineering,
have been employed to do this: SP, M, Δ, GD, IGD, HV, MPFE, Economics and Management. Edited by, R. León, M. J. Muñoz-Torres, and
NPS, and ER. OPs were not proven to be Normally distributed. So, J. M. Moneva. Cham: Springer International Publishing (Lecture Notes in
Business Information Processing). ISBN: 978-3-319-40506-3.
a Kruskal-Wallis test was employed to find if all heuristics are the Dantzig, G. B., and J. H. Ramser. 1959. “The Truck Dispatching Problem.”
same. This test was done for each metric, separately. Final results of Management Science 6 (1): 80–91. doi:10.1287/mnsc.6.1.80.
it and also pairwise comparisons by the Wilcoxon rank-sum test Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan. 2002. “A Fast and Elitist
shown that H2 is the first and H4 is the second best methods while Multiobjective Genetic Algorithm: NSGA-II.” IEEE Transactions on
 TRANSPORTATION LETTERS 637

Evolutionary Computation 6 (2): 182–197. doi:10.1109/4235.996017. Musolino, G., A. Polimeni, and A. Vitetta. 2016. “Freight Vehicle Routing
Drexl, M., and M. Schneider. 2014. “A Survey of the Standard Location-Routing with Reliable Link Travel Times: A Method Based on Network
Problem.” Annals of Operations Research 259 (1–2): 389–414. Fundamental Diagram.” Transportation Letters 10 (3): 159–171.
Fallah, M., R. Tavakkoli-Moghaddam, M. Alinaghian, and A. Salamatbakhsh- doi:10.1080/19427867.2016.1241040.
Varjovi. 2019. “A Robust Approach for a Green Periodic Competitive VRP Nadizadeh, A., and B. Kafash. 2019. “Fuzzy Capacitated Location-routing
under Uncertainty: DE and PSO Algorithms.” Journal of Intelligent & Fuzzy Problem with Simultaneous Pickup and Delivery Demands.” Transportation
Systems 36 (6): 5213–5225. doi: 0.3233/JIFS-179323. Letters 11 (1): 1–19. doi:10.1080/19427867.2016.1270798.
Farrokhi-Asl, H., R. Tavakkoli-Moghaddam, B. Asgarian, and E. Sangari. 2017. Pia, A., and C. Filippi. 2006. “A Variable Neighborhood Descent Algorithm for
“Metaheuristics for a Bi-objective Location-routing-problem in Waste A Real Waste Collection Problem with Mobile Depots.” International
Collection Management.” Journal of Industrial and Production Engineering Transactions in Operational Research 13 (2): 125–141. doi:10.1111/
34 (4): 239–252. doi:10.1080/21681015.2016.1253619. itor.2006.13.issue-2.
Gendreau, M., G. Ghiani, and E. Guerriero. 2015. “Time-dependent Routing Prodhon, C., and C. Prins. 2014. “A Survey of Recent Research on
Problems: A Review.” Computers & Operations Research 64: 189–197. Location-routing Problems.” European Journal of Operational Research 238
doi:10.1016/j.cor.2015.06.001. (1): 1–17. doi:10.1016/j.ejor.2014.01.005.
Ghiani, G., and G. Laporte. 1999. “Eulerian Location Problems.” Networks 34 Quirion-Blais, O., M. Trépanier, and A. Langevin. 2015. “A Case Study of
(4): 291–302. doi:10.1002/(ISSN)1097-0037. Snow Plow Routing Using an Adaptive Large Hood Search
Ghiani, G., and G. Laporte. 2001. “Location-arc Routing Problems.” Opsearch 38 Metaheuristic.” Transportation Letters 7 (4): 201–209. doi:10.1179/
(2): 151–159. doi:10.1007/BF03399222. 1942787514Y.0000000042.
Hashemi Doulabi, S. H., and A. Seifi. 2013. “Lower and Upper Bounds for Rabbani, M., R. Heidari, H. Farrokhi-Asl, and N. Rahimi. 2018. “Using
Location-arc Routing Problems with Vehicle Capacity Constraints.” European Metaheuristic Algorithms to Solve a Multi-objective Industrial
Journal of Operational Research 224 (1): 189–208. doi:10.1016/j.ejor.2012.06.015. Hazardous Waste Location-routing Problem considering Incompatible
Levy, L., and L. Bodin. 1989. “The Arc Oriented Location Routing Problem.” Waste Types.” Journal of Cleaner Production 170: 227–241. doi:10.1016/
INFOR: Information Systems and Operational Research 27 (1): 74–94. j.jclepro.2017.09.029.
Lopes, R. B., C. Carlos Ferreira, B. Sousa Santos, and S. Barreto. 2013. Riquelme-Rodríguez, J. P., M. Gamache, and A. Langevin. 2016. “Location Arc
“A Taxonomical Analysis, Current Methods and Objectives on Routing Problem with Inventory Constraints.” Computers and Operations
Location-routing Problems.” International Transactions in Operational Research 76: 84–94. doi:10.1016/j.cor.2016.06.012.
Research 20 (6): 795–822. Tavakkoli-Moghaddam, R., A. Amini, and S. Ebrahimnejad. 2018. “A New
Lopes, R. B., F. Plastria, C. Ferreira, and B. Sousa Santos. 2014. “Location-arc Mathematical Model for A Multi-product Location-arc Routing Problem.”
Routing Problem: Heuristic Approaches and Test Instances.” Computers and in 2018 4th International Conference on Optimization and Applications
Operations Research 43: 309–317. doi:10.1016/j.cor.2013.10.003. (ICOA). Morocco: IEEE.
Martínez-Salazar, I. A., J. Molina, F. Angel-Bello, T. Gomez, and R. Caballero. Yao, B., Y. Cao, Z. Wang, P. Hu, M. Zhang, and B. Yu. 2016. “Routing Problem
2014. “Solving a Bi-objective Transportation Location Routing Problem by with Mixed Load Plan A Two-stage Heuristic Algorithm for the School Bus
Metaheuristic Algorithms.” European Journal of Operational Research 234 Routing Problem with Mixed Load Plan.” Transportation Letters 8 (4):
(1): 25–36. doi:10.1016/j.ejor.2013.09.008. 205–219. doi:10.1080/19427867.2015.1110953.
Mavrotas, G. 2009. “Effective Implementation of the ε-constraint Method in Multi- Zitzler, E., and L. Thiele. 1999. “Multiobjective Evolutionary Algorithms:
Objective Mathematical Programming Problems.” Applied Mathematics and A Comparative Case Study and the Strength Pareto Approach.” IEEE Trans.
Computation 213 (2): 455–465. doi:10.1016/j.amc.2009.03.037. On Evolutionary Computation 3 (4): 257–271. doi:10.1109/4235.797969.