European Journal of Operational Research 230 (2013) 539–550

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research

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

Production, Manufacturing and Logistics

Two-stage vehicle routing problem with arc time windows: A mixed

integer programming formulation and a heuristic approach
Cihan Çetinkaya a, Ismail Karaoglan b,⇑, Hadi Gökçen c
a
Logistics Command of Gendarmerie, Guvercinlik, Ankara, Turkey
b
Department of Industrial Engineering, Faculty of Engineering and Architecture, Selcuk University, 42075 Selcuklu, Konya, Turkey
c
Department of Industrial Engineering, Faculty of Engineering, Gazi University, 06570 Maltepe, Ankara, Turkey

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

Article history: In this paper, we introduce a new variant of the Vehicle Routing Problem (VRP), namely the Two-Stage
Received 26 August 2011 Vehicle Routing Problem with Arc Time Windows (TS_VRP_ATWs) which generally emerges from both
Accepted 2 May 2013 military and civilian transportation. The TS_VRP_ATW is defined as finding the vehicle routes in such a
Available online 11 May 2013
way that each arc of the routes is available only during a predefined time interval with the objective
of overall cost minimization. We propose a Mixed Integer Programming (MIP) formulation and a heuristic
Keywords: approach based on Memetic Algorithm (MA) to solve the TS_VRP_ATW. The qualities of both solution
Logistics
approaches are measured by using the test problems in the literature. Experimental results show that
Vehicle routing problem with arc time
windows
the proposed MIP formulation provides the optimal solutions for the test problems with 25 and 50 nodes,
Memetic algorithm and some test problems with 100 nodes. Results also show that the proposed MA is promising quality
Mathematical formulation solutions in a short computation time.
 2013 Elsevier B.V. All rights reserved.

1. Introduction local search algorithms, and more recently (Doerner and Schmid,
2010) for meta-heuristic algorithms.
The design of distribution systems is one of the key components The vehicle routing problem with time windows (VRP_TW) is a
of logistics systems as it offers a great potential in order to reduce natural extension of the VRP, motivated by the availability of nodes
costs and to improve service quality. Therefore, huge number of for a specific time windows. In general form of the VRP_TW, each
studies have been devoted to find good solutions for these systems. node must be visited in a predetermined time interval. However,
In this research area, one of the main topics is the vehicle routing in practice, time window restrictions for node routing problems
problem (VRP) which involves the design of a set of minimum cost (i.e. VRPs) can also be appeared on arcs in addition to nodes. Two
routes serving a set of customers with known demands. In classical of them can be given as follows:
VRPs, each customer is served exactly once, each vehicle originates
and terminates at the main depot, and all customers must be as-  In military logistic system of Turkey, some roads can only be
signed to only one vehicle such that the vehicle capacities are used in predetermined time windows. For instance, especially
not exceeded. Following the introduction of the VRP by Dantzig in eastern cities of Turkey, the roads between city centers and
and Ramser (1959), several mathematical models and exact solu- districts generally pass through highlands which are potential
tion procedures developed for small and medium-size VRPs in ambush locations. Therefore, these roads can only be used in
the literature (Achuthan et al., 2003; Araque et al., 1994; Augerat daylights to avoid terrorist attack.
et al., 1998; Baldacci et al., 2012; Lysgaard et al., 2004; Ralphs,  In addition to the military applications, time window restriction
2003). It is commonly known that VRP belongs to the class of can also be appeared in civilian transportation systems. For
NP-hard problems. Because of its intractability manner, a great example, the trucks cannot enter the city center in rush hours
deal of heuristic/meta-heuristic approaches have been successfully (i.e. between 7–10 am and between 4–7 pm), because these
implemented for the problem. We refer the readers to the review hours includes the beginning and ending time of shifts and
of (Toth and Vigo, 2002) for VRPs and their applications, (Eksioglu the traffic density reaches at the extreme level. Therefore, the
et al., 2009) for taxonomic review of VRPs, (Funke et al., 2005) for entrance of trucks to the city center in these hours is forbidden
by local government to facilitate the traffic flow.
⇑ Corresponding author. Tel.: +90 332 223 3751.
Although, there exist many studies in the node routing
E-mail addresses: cihancetinkaya1903@hotmail.com (C. Çetinkaya),
ikaraoglan@selcuk.edu.tr (I. Karaoglan), hgokcen@gazi.edu.tr (H. Gökçen). literature considering the node time windows, to the best of

0377-2217/$ - see front matter  2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2013.05.001
540 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

our knowledge, there is no previous study considering the arc Mullaseril (1997). In this paper, a heuristic algorithm and a trans-
time windows. Because of the increasing importance of these formation approach into a VRP_TW have been developed. Wøhlk
restrictions for the military and civilian transportation, this paper (2005) has also considered the CARP_TW and presented two math-
considers the Vehicle Routing Problem with Arc Time Windows ematical models, several heuristics, a column generation technique
(VRP_ATWs). Two-stage version of VRP_ATW (namely to obtain the lower bounds and a dynamic programming approach
TS_VRP_ATW) is discussed as well, which is the distribution net- for the problem. Reghioui et al. (2007) have proposed a hybrid heu-
work structure of the military logistic system in Turkey. In this ristic technique combining greedy randomized adaptive search
problem, the network is divided into three layers (facility, depots procedure (GRASP) and path relinking. The authors have shown
and customers) and routing operations encountered in both that proposed heuristic can compete with the best existing heuris-
stages each consisting successive layers (i.e. between facility-de- tic in the literature. Vansteenwegen et al. (2010) have considered
pots, and depots-customers). We propose Mixed Integer Program- the mobile mapping van problem equivalent to CARP with time
ming (MIP) formulations for both VRP_ATW and TS_VRP_ATW. soft windows and proposed a two-phase heuristic search algorithm
We also develop a meta-heuristic approach called Memetic Algo- for the problem. The authors have proven the strength of the algo-
rithm (MA) to solve the large-size problem instances. In order to rithm on the test problems derived from the literature, and several
evaluate the performance of the proposed MIP formulation and real-life examples. For reference, Corbeŕn and Prins (2010) can be
MA, we carry out experimental study by using various size test reviewed.
problems adapted from the literature. Experimental results show
that the proposed MIP formulation gives the optimal solutions for 3. Vehicle routing problem with arc time windows
the test problems with 25 and 50 nodes, and some test problems
with 100 nodes. The proposed MA also achieves good quality In this section, we first define the VRP_ATW and give the rela-
solutions especially for the large-size test problems. The average tionship with other related problems. We then present a Mixed
percentage gap values of the proposed MIP formulation and MA Integer Programming (MIP) formulation for the problem.
are 1.05 and 2.26, while the average solution times are 984.81
and 11.20 s, respectively. 3.1. Problem definition
The main contribution of this paper is threefold. First, we define
the VRP_ATW and emphasize the relationship with other node The VRP_ATW can be defined as follows: Let G = (N, A) be a com-
routing problems. Second, we propose a MIP formulation for plete directed graph where N = {0} [ NC is a set of nodes in which
VRP_ATW, and extend this formulation to TS_VRP_ATW. Although, ‘‘0’’ and NC represent the depot and customers, respectively, and
there are several papers considering complex routing operations A = {(i, j)ji – j, "i, j 2 N} is the set of arcs. A fleet of homogeneous
among stages (i.e. (Ambrosino and Graziascutella, 2005; Ambrosin- vehicles with known capacity is available at the depot. Each cus-
o et al., 2009)), to the best of our knowledge, this is the first paper tomer has a nonnegative demand. Each arc has a nonnegative cost,
dealing with multi stage VRP under time window restrictions. Fi- a transportation time, and a time window which identifies the
nally, we propose a meta-heuristic approach (MA) to solve the availability of arc. The transition for all arcs must begin within
large-size problem instances. its time window, if used. The problem is to determine the vehicle
The rest of this paper is organized as follows. Related literature routes with minimum total cost under following constraints: (i)
for the problem given in Section 2. The definition of the each vehicle is used at most one route, (ii) each customer is served
VRP_ATW, the mathematical formulation and the relationship exactly once, (iii) each route begins and ends at the depot, (iv) the
with other routing problems are shown in Section 3. The defini- total vehicle load at any point of the route does not exceed the
tion and the mathematical formulation for the TS_VRP_ATW are vehicle capacity, and (v) the transition for each arc on any route be-
given in Section 4. The detailed description of the proposed algo- gins within its time window.
rithm is given in Section 5. Section 6 reports computational re-
sults, and conclusion and future directions are discussed in the 3.2. Mixed integer programming formulation for the VRP_ATW
last section.
We next introduce sets, notations and decision variables then
proceed by representing the MIP formulation (referred as F1).
2. Literature review
3.2.1. Sets and notations
Hadjar and Soumis (2009) have studied the closely related
problem, namely multiple depot vehicle scheduling problem with
time windows, and proposed a branch-and-price algorithm for NC: set of all customers
the problem. In this paper, the time windows have been trans- N: set of all nodes (N = {0} [ NC)
ferred from nodes to the arcs for reducing the dimension of huge Q: vehicle capacity
number of variables, and thus speed up the solution procedure. di: demand of customer i 2 NC with di 6 Q
However, such a problem involves node time windows rather than cij: cost of an arc (i, j) 2 A
arc time windows under different structure. tij: transportation time of an arc (i, j) 2 A
The concept of ‘‘time window restrictions on arcs’’ has been aij: earliest available time of an arc (i, j) 2 A
well documented in the capacitated arc routing problem (CARP) bij: latest available time of an arc (i, j) 2 A
literature. In these problems, the main objective is to find a least
cost traversal of all required edges in the network in such a way
that service must begin and end at the same node (i.e. depot node),
and the service must start and finish in a predetermined time 3.2.2. Binary decision variables
interval for all required edges. These time window restrictions
may be encountered in areas such as urban waste collection, postal 
deliveries, treatment of roads to remove snow or to prevent ice 1 if a vehicle travels directly from node i to node j ð8i;j 2 NÞ
xij ¼
forming, or taking pictures of street and roads for digital maps. 0 otherwise
CARP with time windows (CARP_TW) have been introduced by
 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550 541

3.2.3. Additional decision variables as possible to obtain good lower bounds, which shorten the opti-
mization process. Therefore, we set M1ij ¼ Si  Sj þ tij and
M2ij ¼ Si  bij where Si ¼ maxj2N ðbij þ tij Þ and Sj = min(a0j + t0j).
Ui; Load on a vehicle just before having serviced to customer i The last way is to include the valid inequalities, one of the most
("i 2 NC) practical ways to strengthen the LP relaxations of the formulations.
Si; Departure time of a vehicle from node i (" i 2 NC) We therefore utilize the following valid inequality for (F1).
X
x0j P rVRP ATW ðN C Þ ð13Þ
j2N C
P
where, r VRP ATW ðN C Þ ¼ d i2NC di =Q e and de is the smallest integer
3.2.4. Model (F1)
bigger than . This constraint bounds below the number of routes
XX originating from the depot.
min cij xij ð1Þ In this paper, we used all of these rules to strengthen the LP
i2N j2N relaxation of our formulation.
Subject to
X 3.3. Relationships between VRP_ATW and other routing problems
xij ¼ 1 8j 2 NC ð2Þ
i2N
X X The VRP_ATW can be considered as the classical VRP when
xji  xij ¼ 0 8i 2 N ð3Þ
aij = 0 and bij = 1. Since the VRP belongs to the class of NP-hard
j2N j2N
problems, the VRP_ATW is also NP-hard.
U j  U i þ Qxij þ ðQ  di  dj Þxji 6 Q  di 8i; j 2 NC ; i – j ð4Þ Any VRP_TW instance can easily be converted into a VRP_ATW
X
U i P di þ dj xij 8i 2 NC ð5Þ by replacing the time window constraints on nodes with time win-
j2N0 ;j–i dow constraints on arcs, such that aij = ei and bij = min{li,lj  tij}
U i 6 Q  ðQ  di Þxi0 8i 2 NC ð6Þ where ei(li) is the earliest (latest) available time of nodes for the
Si  Sj þ M1ij xij 6 M1ij  tij 8i; j 2 NC ; i – j ð7Þ corresponding VRP_TW instance. This transformation technique
has also been used by Hadjar and Soumis (2009) for the multiple
Sj P ða0j þ t 0j Þx0j 8j 2 NC ð8Þ
depot vehicle scheduling problem with time windows. Therefore,
Si P aij xij 8i 2 NC ; 8j 2 N; i – j ð9Þ any solution procedure developed for the VRP_ATW can also be
Si 6 bij þ M2ij ð1  xij Þ 8i 2 NC ; 8j 2 N; i – j ð10Þ used directly to solve the VRP_TW after this transformation. More-
xij 2 f0; 1g 8i; j 2 N ð11Þ over, time window constraints can be appeared on both nodes and
U j ; Sj P 0 8 j 2 N C ð12Þ arcs. In this situation, the problem can be converted to the equiv-
alent VRP_ATW by setting a0 ¼ maxfei ; aij g and
where M1ij and M2ij are sufficiently large numbers satisfying con-   ij
0 0 0
bij ¼ minfbij ; minfli ; lj  tij gg where aij bij is the earliest (latest)
straint sets (7) and (10) when xij = 0, respectively. Furthermore,
we set xij = 0 whenever (di + dj > Q) or (aij + tij > max(k2Njk–i)(bjk)). available time of arcs for the transformed VRP_ATW instance.
In (F1), objective function (1) minimizes the total transportation Please note that, after these transformations any arc satisfying
cost. Constraint sets (2) and (3) are known as degree constraints. aij > bij can be removed from the problem instance.
While constraint set (2) ensures that each customer must be visited However, as far as we search the literature, the reverse relation-
exactly once, constraint set (3) guarantees that the number of enter- ship (i.e. transformation from VRP_ATW to VRP_TW) does not ex-
ing and leaving arcs to each node are equal. Constraint sets (4)–(6) ist, since there could be different time windows of all outgoing
are well known Miller–Tucker–Zemlin (MTZ) constraints eliminat- arcs for all nodes, and aggregating them to a single time window
ing the subtours, subject to the capacity restrictions. Constraint sets for origin node is not possible. Instead, we can transform any
(7)–(10) define the time windows restrictions. While constraint set VRP_ATW instance to the Generalized Vehicle Routing Problem
(7) determines the departure time of each customer depending on with Time Windows (GVRP_TW) which provides a useful modeling
the previous node on a route, constraint sets (8) and (9) guarantee framework for a wide variety of applications (Baldacci et al., 2010;
that the departure time of each node must be greater than the ear- Moccia et al., 2011). The detailed definition of the GVRP_TW and
liest available time of used arc. Similarly, constraint set (10) ensures the transformation steps are given in Appendix A.
that the departure time of each node must be smaller than the latest
available time of used arc. And finally, constraint sets (11) and (12) 4. Two-stage vehicle routing problem with arc time windows
are integrality and non-negativity constraints, respectively.
In most combinatorial optimization problems, the efficiency of Our interest in considering two-stage version of VRP_ATW actu-
any MIP formulation depends on the strength of the linear pro- ally stems from the logistic system of the General Commandership
gramming (LP) relaxation of a given formulation. There are several of Gendarmerie of Turkey. In this system, ammunitions (spare
ways to strengthen the LP relaxations of the formulations. One of parts) stored (produced) in a main depot (factory) which is located
them is to lift the current constraints. Please note that constraint in Ankara are firstly dispatched to the provincial gendarmerie units
sets (4)–(6) are lifted version of the classical MTZ constraints. which are located in city centers all around country by using large
These constraints have been proposed by Miller et al. (1960) for trucks. Then these parts routed from the provincial gendarmerie
the Traveling Salesman Problem (TSP). Kulkarni and Bhave units to the local/border units by using small pickup trucks. How-
(1985) have adapted them to the VRP and Desrochers and Laporte ever, due to the ‘‘safe day’’ implementation which is applied for the
(1991) have lifted them and introduced new bounding constraints distribution to the eastern cities in order to avoid the terrorist at-
on additional variables. Finally, Kara et al. (2004) have corrected tack, the roads are available only during a predetermined time
the lifted version of these constraints. Meanwhile, exact meanings windows. The problem under investigation can be described as
to additional variables have been given by Kara (2008). Two-Stage VRP_ATW (TS_VRP_ATW) where the network is divided
Another way to strengthen the LP relaxations is improving the into three layers (i.e. facility, depots and customers) and routing
constraints including big M constants (i.e. M1ij and M2ij in operations is encountered in both stages each consisting successive
constraint sets (7) and (10)). These numbers should be as small layers (i.e. between facility-depots, and depots-customers).
542 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

TS_VRP_ATW can be defined as follows: Let G = (N, A) be a tripar- 4.4. Model (F2)
tite complete directed graph where N = {0} [ J [ K is a set of nodes
in which ‘‘0’’, J and K represent facility, depots and customers,
respectively, and A = {(i, j)ji – j, "i, j 2 N} is the set of arcs. Each X X X X X
arc has a nonnegative cost, a transportation time, and a time win- min cij xij þ ckl yjkl ð14Þ
i2ð0[JÞj2ð0[JÞ j2J k2ðj[K j Þl2ðj[K j Þ
dow which identifies the availability of arc. Each customer is served
Subject to
from a predefined depot. Facility and depots have enough capacity X
to meet their demands. Two types of vehicle fleet with known xij ¼ 1 8j 2 J ð15Þ
i2ð0[JÞ
capacities are available to serve between facility-depots, and be- X X
xji  xij ¼ 0 8i 2 ð0 [ JÞ ð16Þ
tween depots-customers, respectively. Each customer has a non- j2ð0[JÞ j2ð0[JÞ
negative demand, The problem is to determine the vehicle routes U j  U i þ Q fd xij þ ðQ fd  Di  Dj Þxji 6 Q fd  Di 8i;j 2 J; i – j ð17Þ
with minimum total cost under following constraints: (i) commod- X
U i P Di þ Dj xij 8i 2 J ð18Þ
ities are transported from facility to depots on a route, and then j2J;j–i

from depots to the customers. The transportation from facility to U i 6 Q fd  ðQ fd  Di Þxi0 8i 2 J ð19Þ
customers is not permitted, (ii) each vehicle is used at most one Sj  Sm þ M1jm xjm 6 M1jm  t jm 8j;m 2 J; j – m ð20Þ
route in only one stage, (iii) each customer/depot is served exactly Sj P ða0j þ t 0j Þx0j 8j 2 J ð21Þ
once, (iv) each route begins and ends at the node where they started Sj P ajm xjm 8j 2 J; 8m 2 ð0 [ JÞ; j – m ð22Þ
(i.e. facility or depot), (v) the total vehicle load at any point of the Sj 6 bjm þ M2jm ð1  xjm Þ 8j 2 J; 8m 2 ð0 [ JÞ; j – m ð23Þ
X
route does not exceed the vehicle capacity, and (vi) the transition yjkl ¼ 1 8j 2 J; 8l 2 K j ð24Þ
for each arc on any route begins within its time window, k2ðj[K j Þ
X X
The TS_VRP_ATW can be considered as the VRP_ATW when the yjkl  yjlk ¼ 0 8j 2 J; 8l 2 ðj [ K j Þ ð25Þ
k2ðj[K j Þ k2ðj[K j Þ
depots intermediate between facilities and customers are re-
moved. Since the VRP_ATW belongs to the class of NP-hard prob- V l  V k þ Q dc yjkl þ ðQ dc  dk  dl Þyjlk 6 Q dc  dk 8j 2 J; 8k;l 2 K j ; k – l ð26Þ
X
lems, the TS_VRP_ATW is also NP-hard. V k P dk þ dl yjkl 8j 2 J; 8k 2 K j ð27Þ
l2K j ;k–l
According to these definitions, the sets, notations, decision vari-
ables and MIP formulation (referred as F2) are given as follows (cij, - V k 6 Q dc  ðQ dc  dk Þyjkj 8j 2 J; 8k 2 K j ð28Þ

tij, aij and bij are as given in Section 3.2): Sk  Sl þ M1kl yjkl 6 M1kl  t kl 8j 2 J; 8k;l 2 K j ; k – l ð29Þ
Sk P Sj þ ðajk þ t jk Þyjjk 8j 2 J; 8k 2 K j ð30Þ
4.1. Sets and notations Sk P akl yjkl 8j 2 J; 8k 2 K j ; 8l 2 ðj [ K j Þ; k – l ð31Þ
Sk 6 bkl þ M2kl ð1  yjkl Þ 8j 2 J; 8k 2 K j ; 8l 2 ðj [ K j Þ; k – l ð32Þ
xij 2 f0;1g 8i;j 2 ð0 [ JÞ ð33Þ
J Set of all depots yjkl 2 f0;1g 8j 2 J; 8k;l 2 ðj [ K j Þ ð34Þ
K Set of all customers U j P 0 8j 2 J ð35Þ
N Set of all nodes (N = {0} [ J [ K) V k P 0 8k 2 K ð36Þ
Kj Set of customers assigned to depot j 2 J Sj P 0 8j 2 ðJ [ KÞ ð37Þ
Qfd Capacity of a vehicle available to serve between facility
and depots where, M1ij and M2ij are defined as previously, and
P
Qdc Capacity of a vehicle available to serve between depots Dj ¼ k2K j dj : 8j 2 J. Moreover, we set xij = 0 whenever (Di + Dj > Qfd)
and customers or (aij + tij > maxk2(0[J)jk–i(bjk)), and yjkl = 0 whenever (dk + dl > Qdc)
dk Demand of customer k 2 K with dk 6 Qdc and or (akl + tkl > maxm2(0[J)jm–k (bkm)).
P
k2K j dk 6 Q fd : 8j 2 J; 8k 2 K j
It is easy to see that (F2) is an extended version of (F1). In (F2),
objective function (14) minimizes the total transportation cost.
The objective function includes two parts where the former re-
4.2. Binary decision variables lated to the first stage and the latter related to the second stage
transportation operations, respectively. Constraint sets (15) and
(16) are known as degree constraints. Constraint sets (17)–(19)
8 are MTZ capacity/subtour elimination constraints for the first
< 1 if a vehicle travels directly from node i to node j
>
stage. Constraint sets (20)–(23) define the time windows restric-
xij ¼ in the first stage ð8i; j 2 ð0 [ JÞÞ
>
: tions. While constraint set (20) determines the departure time
0 otherwise of each depot depending on the previous node on a route, con-
8 straint sets (22) and (21) guarantee that the departure time of
< 1 if a vehicle; originating from depot j; travels directly
> each node must be greater than the earliest available time of used
yjkl ¼ from node k to node l in the second stage ð8j 2 J; 8k;l 2 ðj [ K j ÞÞ arc. Similarly, constraint set (23) ensures that the departure time
>
:
0 otherwise of each node must be smaller than the latest available time of
used arc. Constraint sets (15)–(23) are used for modeling the rout-
ing operation in the first stage. Similarly, constraint set (24)–(32)
4.3. Additional decision variables are used for the second stage. And finally, constraint sets (33) and
(34) are integrality constraints, and (35)–(37) are non-negativity
constraints.
Uj Load on a vehicle just before having serviced to depot j in It is worthy to note that, in the absence of the time window
the first stage ("j 2 J) restrictions, the second stage routes are not affected from the ar-
Vk Load on a vehicle just before having serviced to customer rival time to the depot of the first stage routes (please see con-
k in the second stage ("k 2 K) straint set (30)). Under this assumption, the TS_VRP_ATW
Sk Departure time of a vehicle from node k ("k 2 (J [ K)) reduces to (J + 1) independent VRPs where the one is related to
the routing operation between factory and depots (j 2 J), and the
 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550 543

Depots Customers

[ 10 1 2
3
3 4
7 12 6 14 11 4
5 6 7 8 9
8
10
9
11
2
12 13
1 13 5
14
]
(a)
0 2 1

1 5 3 7 4 6

2 12 11 8 14 9 10 13
(b)
2-1

0 2 2 1 1

5-3 3-7 7-4 4-6

1 5 5 3 3 7 7 4 4 6 6

5-3-7 3-7-4

12-11 11-8 8-14 14-9 9-10 10-13

2 12 12 11 11 8 8 14 14 9 9 10 10 13 13

12-11-8 11-8-14

(c)
(20)

14
(20) (10)
(10) (20)
(5) 3
11 7
8 9
(10) 4
10 (30)

12 6 5 (10)
13 (5) (25)
(20)

2 1
[110] [75]

A A: Facility D D: Customers

(E) E: Demand of customers

B B: Depots

[C] C: Total demand of customers which are assigned to corresponding depot

(d)
Fig. 1. An illustrative example for chromosome.

X
others are related to each depot (j 2 J) and corresponding custom- x0j P r TS VRP ATW ðJÞ ð38Þ
ers (k 2 Kj). j2J
X
As explained in previous section, we utilize the following valid yjjk P r TS VRP ATW ðK j Þ 8j 2 J ð39Þ
inequalities to strengthen the LP relaxation of model (F2). k2K j
544 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

Cut Points

P1 [ 1
10 3
2 3 4 5
7 12 6 14 11 4
6 7 8 9
8
10
9
11
2
12 13
1 13 5 ]
14

6
Map
2

P2 [ 1 2
1 14 12 4
3 4 5
2
6
9
7
3 10 8
8 9 10
6
11
7
12
5 11 13 ]
13 14 14
11
9
3
4 10
C [ 1
1
2 3 4 5
9 12 10 6 14 11 4
6 7 8 9
8
10
2
11
7
12
5 3 13 ]
13 14

Fig. 2. An illustrative example for PMX operator.

Fig. 3. Steps of the proposed memetic algorithm.

Table 1
Computational results for C-type test problems with tight time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 C101_25⁄ 234.79 0.00 0.37 0.00 1.82
C102_25⁄ 234.77 0.00 0.42 0.00 1.48
C103_25⁄ 234.76 0.00 0.55 0.00 1.47
C104_25⁄ 234.74 0.00 0.47 0.00 2.30
C105_25⁄ 234.79 0.00 0.44 0.00 1.51
C106_25⁄ 234.79 0.00 0.45 0.00 1.49
C107_25⁄ 234.78 0.00 0.55 0.00 1.49
C108_25⁄ 234.78 0.00 0.43 0.00 1.51
C109_25⁄ 234.77 0.00 0.50 0.00 2.01
Average 234.77 0.00 0.46 0.00 1.68

40 10 C101_50⁄ 426.41 0.00 10.09 0.00 5.91

C102_50⁄ 423.26 0.00 3.39 0.00 5.89
C103_50⁄ 423.22 0.00 2.26 0.00 6.29
C104_50⁄ 423.19 0.00 2.65 0.00 6.40
C105_50⁄ 426.40 0.00 11.14 0.00 6.27
C106_50⁄ 426.40 0.00 17.14 0.00 5.99
C107_50⁄ 426.32 0.00 10.67 0.00 6.46
C108_50⁄ 423.85 0.00 4.07 0.00 7.17
C109_50⁄ 423.25 0.00 2.92 0.00 6.18
Average 424.70 0.00 7.15 0.04 6.28

90 10 C101_100 871.12 8.40 3600.00 10.70 24.49

C102_100 869.26 8.50 3600.00 10.70 22.68
C103_100 865.58 8.70 3600.00 11.00 22.98
C104_100 873.29 7.70 3600.00 10.00 22.53
C105_100 860.73 9.70 3600.00 12.00 23.49
C106_100 881.41 7.00 3600.00 9.30 22.73
C107_100 864.48 8.80 3600.00 11.20 21.36
C108_100 882.54 6.60 3600.00 8.90 23.22
C109_100 871.52 8.00 3600.00 10.20 24.74
Average 871.10 8.16 3600.00 10.44 23.14

Overall average 510.19 2.72 1202.54 3.48 10.37

 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550 545

Table 2
Computational results for R type test problems with tight time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 R101_25⁄ 485.57 0.00 0.31 0.00 1.50
R102_25⁄ 485.57 0.00 0.19 0.00 1.24
R103_25⁄ 485.56 0.00 0.20 0.00 1.23
R104_25⁄ 485.56 0.00 0.12 0.00 1.48
R105_25⁄ 485.57 0.00 0.20 0.00 1.45
R106_25⁄ 485.57 0.00 0.22 0.00 1.28
R107_25⁄ 485.56 0.00 0.11 0.00 1.75
R108_25⁄ 485.56 0.00 0.14 0.00 1.43
R109_25⁄ 485.57 0.00 0.33 0.00 1.79
R110_25⁄ 485.56 0.00 0.36 0.00 1.27
R111_25⁄ 485.56 0.00 0.31 0.00 1.49
R112_25⁄ 485.56 0.00 0.30 0.00 1.23
Average 485.56 0.00 0.23 0.00 1.43

40 10 R101_50⁄ 830.15 0.00 5.13 0.00 5.30

R102_50⁄ 825.04 0.00 2.61 0.00 5.29
R103_50⁄ 825.04 0.00 2.40 0.00 5.02
R104_50⁄ 825.04 0.00 2.92 0.00 5.26
R105_50⁄ 825.05 0.00 2.34 0.00 5.34
R106_50⁄ 825.04 0.00 2.37 0.00 4.94
R107_50⁄ 825.04 0.00 2.68 0.00 5.33
R108_50⁄ 825.04 0.00 2.32 0.00 5.50
R109_50⁄ 825.04 0.00 3.04 0.00 6.16
R110_50⁄ 825.04 0.00 2.67 0.00 6.58
R111_50⁄ 825.04 0.00 3.14 0.00 5.63
R112_50⁄ 825.04 0.00 3.07 0.00 6.74
Average 825.47 0.00 2.89 0.00 5.59

90 10 R101_100 1021.96 1.60 3600.00 6.70 23.20

R102_100 1021.68 0.70 3600.00 4.30 22.88
R103_100 1021.09 0.70 3600.00 4.20 24.19
R104_100 1021.08 0.70 3600.00 4.20 23.52
R105_100 1021.01 0.90 3600.00 5.30 21.63
R106_100 1020.60 0.80 3600.00 4.60 26.42
R107_100 1019.14 0.90 3600.00 4.40 26.76
R108_100 1021.41 0.70 3600.00 4.20 24.14
R109_100 1019.73 0.90 3600.00 4.30 27.50
R110_100 1021.85 0.60 3600.00 4.10 28.26
R111_100 1021.07 0.70 3600.00 4.30 28.21
R112_100 1021.11 0.80 3600.00 4.20 25.32
Average 1020.98 0.83 3600.00 4.57 25.17

Overall average 777.34 0.28 1201.04 1.52 10.73

P P
where r TS VRP ATW ðJÞ ¼ d j2J Dj =Q fd e, rTS VRP ATW ðK j Þ ¼ d k2K j dk =Q dc e mechanisms (i.e. local search). This section presents the detailed
and de is the smallest integer bigger than . While, constraint description of components of the proposed MA for TS_VRP_ATW.
(38) bounds below the number of routes originating from the facil-
ity, constraint set (39) bounds below the number of routes originat-
ing from each depot, respectively. 5.1. Representation

The performance of MA depends heavily on the chromosome

5. Proposed heuristic approach for the TS_VRP_ATW structure that is used to represent solutions, and recombination
operators that are used for this chromosome structure. Therefore,
As mentioned above, TS_VRP_ATW belongs to the class of NP- an appropriate chromosome structure should be developed for
hard problems. Therefore, the proposed formulation is not directly problem. When the related literature is examined, it is seen that
applicable to find optimal solutions for the large-size problems. good results were obtained by permutation encoding scheme orig-
Hence, heuristic solution procedures are necessary to obtain good inally proposed by Prins (2004) for the VRPs. For this reason, we
quality solutions quickly for these problems. developed an encoding scheme for our problem based on permuta-
Motivated by the successful applications of Memetic Algo- tion encoding. In our structure, the length of the chromosome is
rithms (MAs) in solving various vehicle routing problems (Boudia equal to the sum of number of depots and customers (i.e., jJj + jKj).
et al., 2007; Labadi et al., 2008; Liu and Jiang, 2012; Ngueveu Each gene gives information about the relative sequence of corre-
et al., 2010; Prins, 2009), we propose a MA to solve realistic size sponding node with respect to other nodes. The decoding of a chro-
of the TS_VRP_ATW. mosome includes two main steps. Firstly, all nodes are grouped
MA is an evolutionary search procedure that use local search into their assigned depot. Please note that, for the first stage, the
within a classical genetic algorithm framework. MA offer signifi- depots are considered as the customers of the facility. Then, the
cant advantages over conventional methods by using simulta- optimal routes for each group are obtained by using following
neously diversification (i.e. recombination) and intensification splitting procedure which is modified version of route partitioning
546 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

Table 3
Computational results for RC type test problems with tight time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 RC101_25⁄ 367.41 0.00 0.31 0.00 2.03
RC102_25⁄ 367.41 0.00 0.33 0.00 1.59
RC103_25⁄ 367.40 0.00 0.36 0.00 1.58
RC104_25⁄ 367.40 0.00 0.20 0.00 1.80
RC105_25⁄ 367.41 0.00 0.22 0.00 1.68
RC106_25⁄ 367.41 0.00 0.30 0.00 1.54
RC107_25⁄ 367.40 0.00 0.27 0.00 1.66
RC108_25⁄ 367.40 0.00 0.37 0.00 1.94
Average 367.41 0.00 0.30 0.00 1.73

⁄
40 10 RC101_50 661.29 0.00 13.06 0.00 7.80
RC102_50⁄ 661.29 0.00 12.03 0.00 8.00
RC103_50⁄ 661.28 0.00 7.02 0.00 7.63
RC104_50⁄ 661.28 0.00 8.83 0.00 6.80
RC105_50⁄ 661.29 0.00 15.33 0.00 7.81
RC106_50⁄ 661.29 0.00 14.32 0.00 6.47
RC107_50⁄ 661.28 0.00 12.62 0.00 6.58
RC108_50⁄ 661.28 0.00 10.05 0.00 6.12
Average 661.29 0.00 11.66 0.00 7.15

90 10 RC101_100 1133.52 5.20 3600.00 10.20 25.24

RC102_100 1130.40 5.80 3600.00 10.50 27.53
RC103_100 1124.06 6.00 3600.00 11.10 29.46
RC104_100 1128.76 6.00 3600.00 10.60 28.62
RC105_100 1130.99 5.80 3600.00 10.40 29.27
RC106_100 1132.18 5.70 3600.00 10.70 25.93
RC107_100 1122.93 6.50 3600.00 11.60 28.24
RC108_100 1131.38 5.40 3600.00 10.80 24.82
Average 1129.28 5.80 3600.00 10.74 27.39

Overall average 719.32 1.93 1203.98 3.58 12.09

Table 4
Computational results for C type test problems with wide time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 C201_25⁄ 292.17 0.00 0.25 0.00 1.26
C202_25⁄ 292.09 0.00 0.34 0.00 1.23
C203_25⁄ 291.99 0.00 0.24 0.00 1.22
C204_25⁄ 291.91 0.00 0.33 0.00 1.43
C205_25⁄ 292.16 0.00 0.28 0.00 1.27
C206_25⁄ 292.15 0.00 0.25 0.00 1.27
C207_25⁄ 292.11 0.00 0.28 0.00 1.31
C208_25⁄ 292.13 0.00 0.36 0.00 1.51
Average 292.09 0.00 0.29 0.00 1.31

40 10 C201_50⁄ 511.85 0.00 2.79 0.00 6.66

C202_50⁄ 511.63 0.00 10.16 0.00 5.53
C203_50⁄ 511.51 0.00 3.63 0.00 5.56
C204_50⁄ 511.41 0.00 3.52 0.00 5.53
C205_50⁄ 511.82 0.00 2.85 0.00 6.62
C206_50⁄ 511.79 0.00 1.97 0.00 5.62
C207_50⁄ 511.73 0.00 2.12 0.00 5.34
C208_50⁄ 511.76 0.00 2.06 0.00 5.66
Average 511.69 0.00 3.64 0.00 5.81

90 10 C201_100 912.35 2.10 3600.00 1.90 23.69

C202_100 909.05 1.80 3600.00 1.90 25.34
C203_100 911.26 1.50 3600.00 1.70 22.83
C204_100 907.82 1.80 3600.00 2.00 22.67
C205_100 910.95 2.20 3600.00 2.20 23.96
C206_100 901.53 3.10 3600.00 2.90 23.66
C207_100 896.25 3.90 3600.00 3.80 22.22
C208_100 909.62 2.10 3600.00 2.00 23.75
Average 907.35 2.31 3600.00 2.30 23.52

Overall average 570.38 0.77 1201.31 0.77 10.21

 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550 547

Table 5
Computational results for R type test problems with wide time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 R201_25⁄ 476.67 0.00 0.19 0.00 1.30
R202_25⁄ 476.66 0.00 0.19 0.00 1.60
R203_25⁄ 476.61 0.00 0.19 0.00 1.99
R204_25⁄ 476.60 0.00 0.14 0.00 1.56
R205_25⁄ 476.66 0.00 0.17 0.00 1.36
R206_25⁄ 476.65 0.00 0.17 0.00 1.37
R207_25⁄ 476.61 0.00 0.19 0.00 1.80
R208_25⁄ 476.60 0.00 0.14 0.00 1.24
R209_25⁄ 476.64 0.00 0.17 0.00 1.80
R210_25⁄ 476.65 0.00 0.16 0.00 1.25
R211_25⁄ 476.64 0.00 0.16 0.00 2.18
Average 476.64 0.00 0.17 0.00 1.59

40 10 R201_50⁄ 728.82 0.00 1.50 1.40 6.18

R202_50⁄ 728.78 0.00 0.61 1.50 6.51
R203_50⁄ 728.76 0.00 0.75 1.50 5.23
R204_50⁄ 728.76 0.00 1.37 0.00 6.14
R205_50⁄ 728.80 0.00 1.04 0.00 6.37
R206_50⁄ 728.77 0.00 1.39 0.00 5.76
R207_50⁄ 728.76 0.00 1.03 0.00 5.23
R208_50⁄ 728.75 0.00 1.00 0.00 6.39
R209_50⁄ 728.79 0.00 1.14 0.00 6.49
R210_50⁄ 728.79 0.00 1.15 0.00 5.87
R211_50⁄ 728.78 0.00 1.23 0.00 6.14
Average 728.78 0.00 1.11 0.40 6.03

90 10 R201_100⁄ 967.03 0.00 715.64 0.60 28.75

R202_100⁄ 966.29 0.00 301.91 0.00 29.55
R203_100⁄ 966.26 0.00 177.62 0.10 26.94
R204_100⁄ 966.21 0.00 241.22 0.10 26.18
R205_100⁄ 966.35 0.00 273.37 0.50 27.23
R206_100⁄ 966.25 0.00 202.41 0.10 28.20
R207_100⁄ 966.23 0.00 182.72 0.10 27.28
R208_100⁄ 966.19 0.00 170.52 0.10 26.64
R209_100⁄ 966.28 0.00 365.95 0.00 29.29
R210_100⁄ 966.30 0.00 180.07 0.10 27.50
R211_100⁄ 966.28 0.00 330.75 0.10 27.11
Average 966.33 0.00 285.65 0.16 27.70

Overall average 723.92 0.00 95.64 0.19 11.77

procedure proposed by Prins (2004) for the VRP. Let Xk = {c1, c2, - graph is constructed for each group (Fig. 1c). In this network, the
. . . , cs} be a node set of corresponding group where s is the number first node is the facility/depot, from which the route will begin,
of customers assigned to depot k, and G(Nk, Ak) be a directed acyclic and the others are the customers that will be routed. Each arc in
auxiliary graph where Nk = {k [ Xk} is the set of vertices and the graph represents a possible route and the cost of each arc gives
Ak = {(ci, cj)jci, cj 2 Nk and i < j} is the set of directed arcs. Each arc the total cost of corresponding route. For example, arc (5–3) repre-
(ci, cj) 2 Ak represents a feasible route where the vehicle departs sents the feasible route departed from depot 1 and serving only
from depot (node k) and visits nodes ci+1, ci+2, . . . , cj1 and cj, con- customer 3 (i.e. 1–3-1), and arc (3–4) represents the feasible route
secutively, and turns back to the depot, such that load and time departed from depot 1 and serving customers 7 and 4 (i.e. 1–7-4–
windows restrictions on any arc are satisfied. After the construc- 1). Similarly, for each possible route, these arcs are constructed. It
tion of auxiliary graph, a minimum cost-path on G from node k is important to note that infeasible arcs should not be included to
to last node (cs) defines the optimal partition of the sequence the graph, due to capacity or time window restrictions. For exam-
which leads to a feasible solution originating from depot k. ple, arc (8–10) which represents the route 2-14-9-10-2 is not in-
An illustrative example for a chromosome is given in Fig. 1. In cluded to the graph since the total demand of customer (70
this example, TS_VRP_ATW consist of a facility (numbered as 0), units) is greater than the vehicle capacity (40 units). After con-
2 depots (numbered as 1 and 2) and 12 customers (numbered be- struction of auxiliary graph, the shortest path from first node to
tween 3 and 14). Customers with number between 3–7 and 8–14 last node is found and optimal solution of corresponding sequence
are served from depot 1 and 2, respectively. Besides, the capacity is obtained by the routes that are represented by arcs used in that
of vehicle which is available between the facility and the depots path (Fig. 1d).
is 200 units, and the one which is available between the depots
and the customers is 40 units. Numbers in each gene determine 5.2. Crossover operator
the priorities of nodes relative to each other (Fig. 1a). Smaller num-
bers indicate that it will be routed earlier than the nodes having The crossover operator is used to explore new solutions in the
greater numbers. By considering relative priority values, initially search space. In general, crossover operator is applied to obtain
the depots are assigned to the facility, and then, the customers to new solutions by exchanging some information between selected
the depots (Fig. 1b). After this separation process, an auxiliary parents. Since our encoding scheme uses permutation encoding,
548 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

Table 6
Computational results for RC type test problems with wide time windows.

jKj jJj Problem name LB MIP model Memetic algorithm

Gap CPU Gap CPU
20 5 RC201_25⁄ 303.91 0.00 0.31 0.00 2.34
RC202_25⁄ 303.87 0.00 0.28 0.00 1.65
RC203_25⁄ 303.85 0.00 0.30 0.00 1.51
RC204_25⁄ 303.85 0.00 0.36 0.00 1.70
RC205_25⁄ 303.89 0.00 0.30 0.00 1.36
RC206_25⁄ 303.89 0.00 0.42 0.00 1.34
RC207_25⁄ 303.88 0.00 0.55 0.00 1.34
RC208_25⁄ 303.87 0.00 0.30 0.00 2.27
Average 303.88 0.00 0.35 0.00 1.69

⁄
40 10 RC201_50 511.12 0.00 9.86 0.00 7.76
RC202_50⁄ 511.09 0.00 5.40 0.00 6.61
RC203_50⁄ 511.08 0.00 7.60 0.00 6.19
RC204_50⁄ 511.07 0.00 5.05 2.70 6.77
RC205_50⁄ 511.11 0.00 8.00 0.00 5.55
RC206_50⁄ 511.10 0.00 7.53 0.00 6.14
RC207_50⁄ 511.09 0.00 5.87 0.00 6.12
RC208_50⁄ 511.08 0.00 6.72 2.70 5.29
Average 511.09 0.00 7.00 0.68 6.30

90 10 RC201_100 1020.50 3.60 3600.00 3.40 30.56

RC202_100 1009.95 4.50 3600.00 4.50 27.50
RC203_100 1023.05 3.10 3600.00 3.10 28.35
RC204_100 1023.53 3.20 3600.00 3.10 27.62
RC205_100 1018.68 3.60 3600.00 3.60 28.98
RC206_100 1026.46 2.80 3600.00 2.80 29.79
RC207_100 1022.77 3.20 3600.00 3.10 27.20
RC208_100 1017.66 3.70 3600.00 3.70 27.29
Average 1020.33 3.46 3600.00 3.41 28.41

Overall average 611.76 1.15 1202.45 1.36 12.13

several crossover operators (i.e. Partially Mapped Crossover (PMX), parent (P2). We implement steady-state replacement strategy in
Order Crossover (OX), Cycle Crossover (CX)) can be used directly. which only one new solution is obtained by PMX operator, im-
After some preliminary experiments, we utilize PMX operator as proved by local search and inserted to the current population,
illustrated in Fig. 2. In this operator two cut points is selected ran- and then the worst chromosome is removed from population. It
domly and the substring between these points is directly inherited should be noted that local search algorithm is also implemented
from first parent (P1) to the same positions of the child. The sub- to the initial population to improve the solution quality.
string is also used to determine the mapping relation. Then, each According to these definitions, the overall pseudo-code of MA
gene between cut points in the second parent (P2) are swapped for solving TS_VRP_ATW is outlined in Fig. 3.
using mapping relation, and copied to the empty positions of the
child.
6. Computational results

5.3. Local search This section presents our computational results to investigate
the performance of the proposed MIP formulation and MA. The
In this paper, for the purpose of improvement of newly pro- state-of-the-art LP/MIP solver CPLEX (version 10.2) is used to solve
duced individuals, local search algorithm is developed. In this algo- the MIP formulation. We used the default CPLEX parameters to
rithm, an individual that were obtained by crossover is converted solve the formulation, and the total computation time for CPLEX
to the real solution as explained in Section 5.1 and successive swap has been limited to one hour (3600 s). The proposed MA has been
operations are applied to this solution. In swap operation, loca- coded using C++ programming language. After preliminary experi-
tions of two nodes which were randomly selected in the same ments, population size is accepted as jNj and total number of gen-
route are exchanged. In each iteration of local search, the best eration as 50 ⁄jNj. All experiments were performed on Intel Xeon
solution is selected amongst jNj new neighbor obtained from cur- 3.16 GHz equipped (with 1 GB RAM).
rent solution by swapping randomly selected nodes. If new solu- Since there are no benchmark instances in the literature for
tion is better than current solution, new solution is accepted as TS_VRP_ATW, we have derived our test problems considering the
current solution. These steps are repeated until no improvement VRP_TW from test set generated by Solomon (1987). The detailed
is obtained. description of the test problems and the conversion steps are given
in Appendix B.
5.4. Algorithm outline In comparison, following performance measures are used: (i)
Average solution time (CPU). (ii) Percentage Gap (Gap): calculated
In our implementation, we use tournament selection mecha- as 100⁄[(UB  LB)/LB] where LB is the best lower bound provided
nism. In this mechanism, two individual selected randomly from by CPLEX (MIP bound) in one hour computation time and UB is
the population and the best individual is accepted as first parent the upper bound found by CPLEX (best integer feasible solution
(P1). The same steps are also used for the selection of second of MIP) or MA.
 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550 549

Tables 1–6 present computational results for the proposed MIP derived from the literature. Experimental results indicate that the
formulation and MA on each TS_VRP_ATW test set, respectively. In proposed MIP formulation gives the optimal solutions for small-
these tables, while the first two columns represent the number of and medium-size test problems, and some large-size test prob-
depots and customers, respectively, the third column gives the lems. Moreover, the proposed MA achieves good quality solutions
problem name. The names are indicated with ‘‘⁄’’ for the test prob- for the test problems in a short computation time. According to
lem that optimal solution is found by CPLEX within one hour com- these results, we conclude that the problem with up to 50 nodes
putation time. Next column shows the best lower bound (LB) should be solved using proposed MIP formulation. Otherwise, we
provided by CPLEX. Finally, last two columns gives the percentage recommend that it should be solved using proposed MA.
gap value (Gap) and computation time (CPU) of the proposed MIP Further research can be performed on exact solution methods
formulation and MA, respectively. In these columns, better results (i.e. Branch and Cut) for the large-size test problems to obtain opti-
are shown in bold. mal solutions in a reasonable computation time. Moreover, the
As can be seen from Tables 1–3, optimal solutions are obtained proposed MIP formulation and MA can be adapted to different type
by not only the proposed MIP model but also the proposed MA for of this problem such as VRP with multiple arc time windows, mul-
all test problems with 25 and 50 nodes in a very short computing ti-period VRP_ATW, stochastic VRP_ATW, and VRP_ATW with het-
time. erogeneous vehicles. Complex distribution systems including arc
For the test problems with 100 nodes, good quality solutions time windows restrictions such as location routing problem with
have been obtained by the proposed MIP formulation within one arc time windows, multistage VRP in which vehicles can be used
hour CPU time. The gap values are obtained as 8.16, 0.83 and in each stage, can also be investigated. Finally, the performance
5.80, for C, R and RC-type of problems, respectively. According to of the proposed MA can be increased by using new neighborhood
results, it can be concluded that the test problems where the cus- mechanisms (i.e. insert and/or swap-range operators in addition
tomers are clustered are more difficult to solve than other test to swap).
problems. Meanwhile, good quality solutions are also obtained
by the proposed MA. For C, R and RC-type of problems, 10.44, Acknowledgements
4.57 and 10.74 gap values are obtained, respectively. Although bet-
ter feasible solutions are obtained with the proposed MIP formula- We thank to the Associate Editor and anonymous referees for
tion, the solution times of the proposed MA are very short (23.14, their helpful comments and suggestions, which greatly contributed
25.17 and 27.39 s, respectively). It is well known that the solution to the improvement of the paper.
time of any procedure should be as small as possible for the oper-
ational problems such as node routing problems. Therefore, the
Appendix A. Supplementary material
proposed MA is preferable to the proposed MIP formulation espe-
cially for the large-size problems.
Supplementary data associated with this article can be found, in
Similar results are also obtained for the test problems with wide
the online version, at http://dx.doi.org/10.1016/j.ejor.2013.05.001.
time windows. Tables 4–6 gives the computational results for C, R
and RC-type of test problems with wide time windows, respec-
References
tively. As can be seen from these tables, the best solutions are ob-
tained by both solution methods for all test problems with 25 Achuthan, N.R., Caccetta, L., Hill, S.P., 2003. An improved branch-and-cut algorithm
nodes in a very short computation time. for the capacitated vehicle routing problem. Transportation Science 37, 153–
As for the test problems with 50 nodes, the best solutions are 169.
Ambrosino, D., Graziascutella, M., 2005. Distribution network design: new
obtained by the proposed MIP formulation. Meanwhile, good qual- problems and related models. European Journal of Operational Research 165,
ity solutions are also obtained by the proposed MA in a short com- 610–624.
putation time. Only 5 out of 27 test problems (namely R201-50, Ambrosino, D., Sciomachen, A., Scutella, M., 2009. A heuristic based on multi-
exchange techniques for a regional fleet assignment location-routing problem?
R202-50, R203-50, RC204-50 and RC208-50) cannot optimally be Computers & Operations Research 36, 442–460.
solved with the proposed MA. For C, R and RC-type of problems, Araque, J., Kudva, G., Morin, T.L., Pekny, J.F., 1994. A branch-and-cut algorithm for
0.00, 0.40 and 0.68 gap values are obtained within 5.81, 6.03 and vehicle routing problems. Annals of Operations Research 50, 37–59.
Augerat, P., Belenguer, J.M., Benavent, E., Corberan, A., Naddef, D., 1998. Separating
6.30 s, respectively.
capacity constraints in the CVRP using tabu search. European Journal of
With regard to the test problems with 100 nodes, while the Operational Research 106, 546–557.
optimal solutions are obtained for R-type test problems within Baldacci, R., Bartolini, E., Laporte, G., 2010. Some applications of the generalized
285.65 s on average, no optimal solutions are obtained for C and vehicle routing problem. Journal of the Operational Research Society 61, 1072–
1077.
RC-type test problems (the gap values are obtained as 2.31 and Baldacci, R., Mingozzi, A., Roberti, R., 2012. Recent exact algorithms for solving the
3.46) within predetermined maximum solution time. Meantime, vehicle routing problem under capacity and time window constraints. European
good solutions results are obtained with the proposed MA in a Journal of Operational Research 218, 1–6.
Boudia, M., Prins, C., Reghioui, M. 2007. An Effective Memetic Algorithm with
short computation times. For C, R and RC-type of problems, the Population Management for the Split Delivery Vehicle Routing Problem, Hybrid
gap values are obtained as 2.30, 0.16 and 3.41 within 23.52, Metaheuristics. In: Bartz-Beielstein, T., Blesa Aguilera, M., Blum, C., Naujoks, B.,
27.70 and 28.41 s, respectively. Moreover, better results are ob- Roli, A., Rudolph, G., Sampels, M. (Eds.), vol. 4771. Springer Berlin/Heidelberg.,
pp. 16–30.
tained with the proposed MA for 4 of 8 in C-type and 3 of 8 in Corbeŕn, A., Prins, C., 2010. Recent results on arc routing problems: an annotated
RC-type test problems, respectively. bibliography. Networks 56, 50–69.
Dantzig, G.B., Ramser, J., 1959. The truck dispatching problem. Management Science
6, 80–88.
Desrochers, M., Laporte, G., 1991. Improvements and extensions to the Miller–
7. Conclusion and future directions Tucker–Zemlin subtour elimination constraints. Operations Research Letters 10,
27–36.
Doerner, K., Schmid, V., 2010. Survey: matheuristics for rich vehicle routing
In this study, a variant of the vehicle routing problem, called
problems. In: Blesa, M., Blum, C., Raidl, G., Roli, A., Sampels, M. (Eds.), Hybrid
Two-Stage Vehicle Routing Problem with Arc Time Windows Metaheuristics, vol. 6373. Springer, Berlin/Heidelberg, pp. 206–221.
(TS_VRP_ATW) has been dealt with. We have proposed a Mixed Eksioglu, B., Vural, A.V., Reisman, A., 2009. The vehicle routing problem: a
Integer Programming (MIP) formulation and a Memetic Algorithm taxonomic review. Computers & Industrial Engineering 57, 1472–1483.
Funke, B., Grünert, T., Irnich, S., 2005. Local search for vehicle routing and
(MA) for the problem. The performance of the proposed solution scheduling problems: review and conceptual integration. Journal of Heuristics
methods have been investigated on a wide range of test problems 11, 267–306.
550 C. Çetinkaya et al. / European Journal of Operational Research 230 (2013) 539–550

Hadjar, A., Soumis, F., 2009. Dynamic window reduction for the multiple depot Mullaseril, P.A., 1997. Capacitated Rural Postman Problem with Time Windows and
vehicle scheduling problem with time windows. Computers & Operations Split Delivery. Ph.D. thesis, University of Arizona, Tucson, Arizona.
Research 36, 2160–2172. Ngueveu, S.U., Prins, C., Wolfler Calvo, R., 2010. An effective memetic algorithm for
Kara, I., 2008. Two Indexed Polynomial Size Formulations for Vehicle Routing the cumulative capacitated vehicle routing problem. Computers & Operations
Problems. Basßkent University, Ankara/Turkey. Research 37, 1877–1885.
Kara, I., Laporte, G., Bektas, T., 2004. A note on the lifted Miller–Tucker–Zemlin Prins, C., 2004. A simple and effective evolutionary algorithm for the vehicle routing
subtour elimination constraints for the capacitated vehicle routing problem. problem. Computers and Operations Research 31, 1985–2002.
European Journal of Operational Research 158, 793–795. Prins, C., 2009. Two memetic algorithms for heterogeneous fleet vehicle routing
Kulkarni, R.V., Bhave, P.R., 1985. Integer programming formulations of vehicle problems. Engineering Applications of Artificial Intelligence 22, 916–928.
routing problems. European Journal of Operational Research 20, 58–67. Ralphs, T.K., 2003. Parallel branch and cut for capacitated vehicle routing. Parallel
Labadi, N., Prins, C., Reghioui, M., 2008. A memetic algorithm for the vehicle routing Computing 29, 607–629.
problem with time windows. RAIRO – Operations Research 42, 415–431. Reghioui, M., Prins, C., Labadi, N., 2007. GRASP with path relinking for the capacitated
Liu, R., Jiang, Z., 2012. The close–open mixed vehicle routing problem. European arc routing problem with time windows. In: vol. 4448. LNCS, pp. 722–731.
Journal of Operational Research 220, 349–360. Solomon, M.M., 1987. Algorithms for the vehicle routing and scheduling problems
Lysgaard, J., Letchford, A.N., Eglese, R.W., 2004. A new branch-and-cut algorithm for with time window constraints. Operations Research 35, 254–265.
the capacitated vehicle routing problem. Mathematical Programming 100, 423– Toth, P., Vigo, D., 2002. The Vehicle Routing Problem. Society for Industrial and
445. Applied Mathematics. SIAM, Philadelphia.
Miller, C., Tucker, A., Zemlin, R., 1960. Integer programming formulations and Vansteenwegen, P., Souffriau, W., Sörensen, K., 2010. Solving the mobile mapping
traveling salesman problems. Journal of the ACM 7, 326–329. van problem: a hybrid metaheuristic for capacitated arc routing with soft time
Moccia, L., Cordeau, J.F., Laporte, G., 2011. An incremental tabu search heuristic for windows. Computers & Operations Research 37, 1870–1876.
the generalized vehicle routing problem with time windows. Journal of the Wøhlk, S., 2005. Contributions to Arc Routing. Ph.D. thesis, University of Southern
Operational Research Society. Denmark.