Sie sind auf Seite 1von 15

JOURNAL OF ADVANCED TRANSPORTATION

J. Adv. Transp. 2013


Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/atr.1237

A differential evolution approach for the vehicle routing problem


with backhauls and time windows

İlker Küçükoğlu* and Nursel Öztürk


Uludag University, Industrial Engineering Department, Bursa, Turkey

SUMMARY
This paper presents a differential evolution algorithm (DEA) to solve a vehicle routing problem with
backhauls and time windows (VRPBTW) and applied for a catering firm. VRPBTW is an extension of the
vehicle routing problem, which includes capacity and time window constraints. In this problem, customers
are divided into two subsets: linehaul and backhaul. Each vehicle starts from a depot and goods are delivered
from the depot to the linehaul customers. Goods are subsequently brought back to the depot from the
backhaul customers. The objective is to minimize the total distance that satisfies all of the constraints. The
problem is formulated using mixed integer programming and solved using DEA. Proposed algorithm is
tested with several benchmark problems to demonstrate effectiveness and efficiency of the algorithm and
results show that our proposed algorithm can find superior solutions for most of the problems in comparison
with the best known solutions. Hence, DEA was carried out for catering firm to minimize total transportation
costs. Copyright © 2013 John Wiley & Sons, Ltd.

KEY WORDS: vehicle routing; transport planning; differential evolution algorithm

1. INTRODUCTION

Vehicle routing is an important logistics activity in both the public and private sectors. Therefore, even
small improvements in routing efficiency can result in large cost reductions for goverments and industries
[1]. Thus, improving routing efficiency will become even more important as a result of recent increases in
fuel prices. The vehicle routing problem (VRP) involves the design of a set of minimum-cost vehicle
routes, originating and terminating at a central depot, for a fleet of vehicles that services a set of customers
with known demands. Each customer is serviced exactly once, and furthermore, all customers must be
assigned to vehicles without exceeding vehicle capacities [2]. The route cost of a vehicle is the total dis-
tance it travels, and the objective is to minimize the total cost of all routes with the minimum number of
vehicles without violating any constraints. In literature, it is seen that one of the current issues in the op-
eration of transportation systems includes vehicle dispatch problems [3].
The VRP with backhauls and time windows (VRPBTW) is an extensive variant of the classical
VRP. The VRPBTW contains two different subsets of customers: linehauls and backhauls. A linehaul
customer requires a given quantity of goods from a central depot, whereas a given quantity of goods is
collected from a backhaul customer and brought back to the depot. Moreover, the backhauls must be
visited after the linehauls in each route. Fleet type can be homogeneous as well as heterogeneous

*Correspondence to: İlker Küçükoğlu, Uludag University, Industrial Engineering Department, Bursa, Turkey. E-mail:
ikucukoglu@uludag.edu.tr

Copyright © 2013 John Wiley & Sons, Ltd.


İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

fleet of vehicles with different capacities [4]. In addition, each customer must be serviced within
a specified time window, and a vehicle is not allowed to begin service at a customer location af-
ter its time window’s upper bound. Furthermore, a waiting time is incurred if a vehicle reaches a
customer before the lower bound. VRP with limited capacitated vehicles is demonstrated as a
NP-Hard problem type [5]. In addition to VRP, adding time windows and backhauls constraints
do not simplify the problem.
Much progress has been accomplished in the field of vehicle routing from the past to the
present. Laporte [6] presented the developments of the solution approaches on VRP in the last
50 years. Some of the main exact algorithms, heuristics, and meta-heuristics for the VRP were
reviewed. Although many different heuristic and exact algorithms are presented in the literature
for VRPs and their extensive types (capacitated VRP, VRP with backhauls, VRP with time
windows), only a few recent studies are devoted to the VRPBTW. Gelinas et al. [7] proposed
a new branching strategy for branch-and-bound approaches based on column generation. This al-
gorithm finds optimal solutions to different test problems with up to 100 customers. Thangiah
et al. [8] presented an insertion algorithm using Solomon’s test problems. Duhamel et al. [9]
designed a tabu search heuristic for the VRPBTW with customer precedence. Reimann et al.
[10] improved the insertion algorithm using an ant system. Cheung and Hang [11] developed la-
bel-matching algorithms for solving the VRPBTW. Their heuristic approach can handle the addi-
tion of complex real-world constraints, such as vehicles of different capacities and penalties for
vehicles that arrive early. Zhong and Cole [12] used an adapted sweep algorithm and guided lo-
cal search heuristic for VRPBTW. The guided local search heuristic was used for the initial so-
lution, and the adapted sweep algorithm was used to improve the solution. Cho and Wang [13]
adapted the nearest neighbor heuristic to the threshold-accepting algorithm. Ropke and Pisinger
[14] used a neighbor search algorithm by transforming the VRPBTW to VRPB and ordered
routes according to time windows constraints. Aghdaghi and Jolai [15] developed a two-phase
heuristic using a goal-programming approach. It can be clearly seen from the literature that heu-
ristic algorithms are performed for this specified problem when the problem size become large
scale. However, differential evolution algorithm (DEA), which is one of the recent population
based algorithms, is not applied for this problem to the best of our knowledge.
In this research, we present a DEA that was developed by adapting a distinct encoding scheme
to solve the VRPBTW. Section 2 defines the industrial problem and provides a mathematical
formulation of the VRPBTW. Section 3 describes the proposed DEA that was adapted for the
VRPBTW. Section 4 describes the parametric analysis of the proposed algorithm.Computational
results are reported in Section 5, and Section 6 concludes the paper.

2. PROBLEM DEFINITION AND MATHEMATICAL FORMULATION

2.1. Problem definition


Distribution procedures cause high costs for catering services. For this reason, distribution
operations must be well designed. In our problem, the catering firm under study has one
distribution center and eight cars which to service 148 customers placed in a wide area. Firm
must satisfy the customers’ daily demand amount. Each customer must be visited once a day
for linehaul and backhaul operations. First, vehicles visit the customers for food distributions
and then starts to collection operations or returns to depot directly. Linehaul and backhaul
customers can be visited in a same route. However, linehaul services must be performed before
backhaul services. Also, each customer has a specified time window to get service. In addition,
each vehicle has capacity that could not be exceeded. These restrictions are defined with three
different constraints that convert the VRP to a VRPBTW:

• time windows constraint;


• capacity constraint;
• service type (linehauls–backhauls) constraint.

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
VEHICLE ROUTING

2.2. Model formulation


Notation
M Number of linehaul customers
N Number of backhaul customers
K Number of cars
distanceij Distance from i to j (i, j = 1,. . .,M + N + 1; i 6¼ j)
ai The linehauling demand of customer i (i = 2,. . .,M + 1)
bi The backhauling demand of customer i (i = M + 2,. . .,M + N + 1)
capk The capacity of service vehicle k (k = 1,. . .,K)
pi Service time of customer i (i = 1,. . .,M + N + 1)
ei Earliest arrival time at customer i (i = 1,. . .,M + N + 1)
li Latest arrival time at customer i (i = 1,. . .,M + N + 1)
ti Service starting time at customer i (i = 1,. . .,M + N + 1)
cij Travel time from customer i to customer j (i, j = 1,. . .,M + N + 1; i 6¼ j)
T Maximum route time allowed for vehicles
MM Penalty number
S Solution sets which have no ti subsets for each vehicle

Decision variables
xijk 1 if vehicle k travels from customer i to j and 0 otherwise
uik 1 if vehicle k services to linehaul customer i and 0 otherwise
vik 1 if vehicle k services to backhaul customer i and 0 otherwise

Model building
The VRPBTW can be formulated into the following model:
X MþNþ1
MþNþ1 X XK
Min distanceij xijk (1)
i¼1 j¼1 k¼1

S.t.
X
M þ1
ai uik ≤capk k ¼ 1; . . . ; K (2)
i¼2

X
K
uik ¼ 1 ði ¼ 2; . . . ; M þ 1Þ (3)
k¼1

X
MþNþ1
bi vik ≤capk ðk ¼ 1; . . . ; K Þ (4)
i¼Mþ2

X
K
vik ¼ 1 ði ¼ M þ 2; . . . ; M þ N þ 1Þ (5)
k¼1

X
Mþ1
xijk ¼ ujk ðj ¼ 2; . . . ; M þ 1; k ¼ 1; . . . ; K Þ (6)
i¼1

X
MþNþ1
xijk ¼ vjk ðj ¼ M þ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (7)
i¼1

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

X
MþNþ1
xijk ¼ uik ði ¼ 2; . . . ; M þ 1; k ¼ 1; . . . ; K Þ (8)
j¼1

X
MþNþ1

xijk ¼ vik ði ¼ M þ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (9)


j¼1
j 6¼ ½2; M; þ; 1

X X
MþNþ1 K
xijk ¼ 1 ði ¼ 2; . . . ; M þ N þ 1Þ (10)
j¼1 k¼1

X X
MþNþ1 K
xijk ¼ 1 ð j ¼ 2; . . . ; M þ N þ 1Þ (11)
i¼1 k¼1

X X
MþNþ1 K
x1jk ≤K (12)
j¼1 k¼1

X X
MþNþ1 K
xi1k ≤K (13)
i¼1 k¼1

X
MþNþ1 X
MþNþ1
xijk ¼ xjik ði ¼ 1; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (14)
j¼1 j¼1

X
MþNþ1
x1jk ¼ 1 ðk ¼ 1; . . . ; K Þ (15)
j¼1

X
MþNþ1
xi1k ¼ 1 ðk ¼ 1; . . . ; K Þ (16)
i¼1

 
p1 þ c1j  t j ≤MM 1  x1jk ð j ¼ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (17)

t i þ pi þ ci1  T≤MM ð1  xi1k Þ ði ¼ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (18)

 
t i þ pi þ cij  t j ≤MM 1  xijk ði; j ¼ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (19)

ei ≤t i ≤li ði ¼ 1; . . . ; M þ N þ 1Þ (20)

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
VEHICLE ROUTING

xijk 2 f0; 1g ði; j ¼ 1; . . . ; M þ N þ 1; k ¼ 1; . . . ; K; i 6¼ jÞ (21)

uik 2 f0; 1g ði ¼ 2; . . . ; M þ 1; k ¼ 1; . . . ; K Þ (22)

vjk 2 f0; 1g ðj ¼ M þ 2; . . . ; M þ N þ 1; k ¼ 1; . . . ; K Þ (23)

According to the model, objective (1) is to minimize the sum of the route distances. Constraints (2) and
(4) are the capacity restrictions that ensure vehicle k does not exceed its capacity. Constraints (3) and (5)
assign linehaul and backhaul customers to a vehicle. Constraints (6)–(9) define the routes of the vehicles
that provide linehaul and backhaul customer restrictions. Constraint (6) provides that if customer j is
linehaul customer then customer i can be linehaul or backhaul customer. Constraint (7) provides that if
customer j is backhaul customer, then customer i must be backhaul customer. Constraints (8) and (9)
complement constraints (6) and (7): If customer i is backhaul customer, then customer j must be
backhaul customer, and also, if customer i is linehaul customer, then customer j can be linehaul or
backhaul customer. Constraints (10)–(16) define classic VRP constraints. Constraints (10) and (11) say
that each customer must be visited by only one vehicle, and constraint (14) is duration constraint.
Constraints (12), (13), (15), and (16) ensure the fleet size. Constraints (17)–(20) define time windows that
each customer must be serviced in time windows and also these constraints prevent subtours. For exam-
ple, a subtour consists of three customers i1, i2, and i3 where x12 = x23 = x31 = 1. After substituting the time
windows constraints, a feasible solution can be provided by only c12 + c23 + c31 ≤ 0, which is clearly
infeasible for nonnegative travel times [16]. Constraints (21)–(23) guarantee that all decision variables
must be either 1 or 0.
To find an optimal solution for the catering service problem, proposed mathematical model was solved
using the prominent commercial solver CPLEX 11.1 on a personal computer with a Pentium (Duo) 2.2
Ghz CPU. Although the original problem had 148 customers, an optimum solution could not be determined
for more than 32 customers. Table 1 shows the results of the problem restricted with 32 customers.
Optimum solution of the problem is 219.5 km could be found after CPLEX run 10 hours.

Table I. Optimum solution for VRPBTW with 32 customers.

Vehicle 1 Service type Vehicle 2 Service type Vehicle 3 Service type


11 Linehaul 8 Linehaul 3 Linehaul
12 Linehaul 24 Backhaul 6 Linehaul
7 Linehaul 9 Linehaul
13 Linehaul 2 Linehaul
14 Linehaul 10 Linehaul
16 Linehaul 1 Linehaul
15 Linehaul 5 Linehaul
31 Backhaul 4 Linehaul
32 Backhaul 20 Backhaul
30 Backhaul 21 Backhaul
29 Backhaul 17 Backhaul
23 Backhaul 26 Backhaul
28 Backhaul 18 Backhaul
27 Backhaul 25 Backhaul
22 Backhaul
19 Backhaul

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

3. DIFFERENTIAL EVOLUTION ALGORITHM

The DEA grew out of Price’s attempts to solve the Chebychev polynomial fitting problem [17]. The
DEA can be classified as an evolutionary optimization algorithm that is similar to a genetic algorithm
and uses the same operators: crossover, mutation, and selection. The main difference in constructing
better solutions is that genetic algorithms rely on crossover while differential evolution relies on muta-
tion. This main operation is based on the differences of randomly sampled pairs of solutions in the pop-
ulation [18]. In addition, DEA includes simple aritmetic operators to evolve population as well as
crossover, mutation, and selection [19]. The critical task of these parameters is to generate trial param-
eter vectors, which is replaced if it yields a better solution than existing vector solution [20].
As with all evolutionary optimization algorithms, DEA works with a population that is formed with
continuous variables [21]. However, VRP solutions consist of integer values. Thus, DEA needs to be
developed such that it includes an encoding procedure to represent integer solution vectors. Wang
et al. [22] proposed a modified DEA, which can solve the binary-coded optimization problems directly.
In addition, Erbao and Mingyong [20,23], Hou et al. [24], and Mingyong and Erbao [19] applied DEA
to solve VRP and its extensions. Wisittipanich and Kachitvichyanukul [25] proposed two enhanced
DEA for job shop scheduling problems to minimize makespan and total weighted tardiness. This paper
represents a different encoding scheme to construct integer solution vectors and improve the
performance of the algorithm.

3.1. Coding and initialization


As described earlier, DEA works with a population including multiple solutions, and each population
of generation G contain NP solution vectors composed of D parameters. The parameters of a solution
vector must be within the range of their lower and upper boundary values x(L) and x(U), respectively. To
generate initial population, parameter vector can be generated randomly by the following equation:
 
xi;j;0 ¼ xðLÞ þ rnd ½0; 1 xðU Þ  xðLÞ ; i ¼ 1; . . . ; D; j ¼ 1; . . . ; NP

where rnd[0,1] represents a uniformly distributed random value with the range of 0 and 1.
In this research, we used two blocks in each chromosome to denote the customer and vehicle
numbers. In these blocks, we used continuous variables for DEA procedures and labeled these
variables with integer variables. First-block labels contain the customer numbers without replication,
and second-block labels contain vehicle numbers. Figure 1 shows the proposed population structure.
Each chromosome of a population includes disordered routes. Thus, route construction procedure is
applied for each chromosome to generate a solution. This procedure consists of two section. In the first
section, customers are separated to vehicles with respect to their vehicle number label. After the
seperation section, customers are sorted according to customer specifications such that each linehaul
customer must be placed before backhaul customers. Figure 2 shows an illustrative example of route
construction procedure.

Figure 1. DEA population.

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
VEHICLE ROUTING

Figure 2. An illustrative example of route construction procedure.

3.2. Crossover
DEA typically uses a uniform crossover strategy [21]. For each variable in a chromosome, if the
randomly generated number is smaller than the crossover ratio, this variable is choosen from the mutand
vector and used in a trial vector. For any generation G, crossover operation is exacuted with the
following equation:

(
vi;j;G if rnd ½0; 1≤CR or i ¼ jrnd
ui;j;G ¼ ; i ¼ 1; . . . ; D; j ¼ 1; . . . ; NP
xi;j;G1 ; otherwise

where CR is the crossover ratio and jrnd 2 1, . . ., D is randomly choosen index for chromosome j.

3.3. Mutation
For each target vector Xj,G, a mutant vector Vj,G is generated by:

 
vi;j;G ¼ xi;r3;G1 þ F xi;r1;G1  xi;r2;G1 ; i ¼ 1; . . . ; D; j ¼ 1; . . . ; NP (26)

where r1, r2, r3 2 [1,NP], r1 6¼ r2 6¼ r3 6¼ j and F 2 [0,2] is the mutation factor. As distinct from the
classic DEA, mutation operation described with Equation (26) is applied to continious variables,
and then, integer values are shifted according to increasing order of the continious values.

3.4. Selection
After the crossover and mutation operations, the trial and target vectors are compared for the selection
process. If the trial vector has an equal or lower objective function value, this vector is replaced with
the target vector in the next generation. Otherwise, the target vector remains in the population for at
least one more generation [26]. This greedily selection operation is given by the following:

(    
uj;G if f uj;G ≤f X j;G1
X j;G ¼ ; j ¼ 1; . . . ; NP (27)
X j;G1 ; otherwise

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

3.5. Stopping criteria


In this research, maximum iteration is used as the stopping criterion. Mutation, crossover, and selection
operators repeat until generation number G reaches the maximum generation number. The pseudo code
of the proposed algorithm is shown in the Figure 3.

4. PARAMETRIC ANALYSIS

As with most of the heuristics, DEA works with a number of parameters. Furthermore, these parameters
affect performance of the algorithm. In this section we focused on to find intelligent parameter
combinations that improve the solution quality at proposed algorithm.
DEA has only three or four operational parameters: population size (NP), crossover ratio (CR),
mutation constant (F), and maximum generation 27,19. In this study, three parameters of the DEA
are analyzed with the specified ranges: NP = [5,15], CR = [0,1], F = [0,2]. For the experimental design,
we used three points for each parameter: maximum value of the parameter (denoted by 1), minimum
value of the parameter (denoted by 1), and the center points of the parameter ranges (denoted by
0). Experiments are carried out with the small-sized problem, which is mentioned in Section 2
consisting of 32 customers. Prepared experimental points and their results are shown in Table 2. Exper-
iments include totally nine different points, and each experiment has three replications. As a result of the
experiments, the best solution is 229.1 km, which is observed with CR = 0.1, F = 0.1, NP = 15.

Figure 3. Pseudo code of the proposed DEA.

Table II. Experimental design.

Number of experiments CR F NP Replication 1 Replication 2 Replication 3


1 1 1 1 244.90 235.90 243.30
2 1 1 1 229.10 244.50 233.90
3 1 1 1 288.00 353.00 244.90
4 1 1 1 229.50 235.20 234.50
5 1 1 1 815.23 813.66 653.47
6 1 1 1 532.27 786.35 714.75
7 1 1 1 562.99 664.64 707.54
8 1 1 1 472.69 474.90 379.80
9 0 0 0 331.15 374.45 430.74

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
VEHICLE ROUTING

Figure 4. Pareto chart.

The parameters and their interaction effects were examined using a Pareto chart generated by
STATISTICA 7. Figure 4 shows that CR, F, NP, and CR–F interactions affect the solution. Moreover,
CR parameter has a dominating effect on the solution with respect to the other parameters. Thus, in our
problem, the CR parameter range was created frequently.

5. COMPUTATIONAL RESULTS

The proposed DEA was coded by the Visual Basic 9.0 programming language in Visual Studio 2008.
For the initialization, customer and vehicle informations were taken from an Access database. Figure 5
shows the interface for the DEA formed in VS 2008. The numerical experiments were executed on a
personal computer with an Intel(R) Core(TM)2 Duo 2.20GHz CPU, 4 GB RAM, and Windows 7.
Computational analysis consist of two parts. First, we performed our algorithm on benchmark
problems. Then, we solved our catering firm problem.
To evaluate performance of the proposed DEA, we used a set of data for VRPBTW, namely R101 to
R105, which were developed by Gelinas et al. [7]. Gelinas’ problems were obtained from Solomon’s
R1 type VRPTW instances by randomly selecting 10%, 30%, and 50% of backhaul customers in each

Figure 5. Interface of the proposed DEA.

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
Table III. Computational result of benchmark problems for 25 customers.

Best known solution DEA


Difference (%)
Problem Backhaul Optimal Total Total
type (%) solution distance CPU Reference distance CPU Optimal S. Best known

Copyright © 2013 John Wiley & Sons, Ltd.


R101 10 643.4 643.4 5.90 Potvin et al. [29] 643.4 1.20 0.00 0.00
30 711.1 717.0 0.62 Thangiah et al. [8] 721.8 0.90 1.50 0.69
50 674.5 682.3 5.60 Potvin et al. [29] 676.8 1.90 0.34 0.81
R102 10 563.5 565.0 0.80 Thangiah et al. [8] 565.3 2.30 0.32 0.05
30 622.3 631.0 0.85 Thangiah et al. [8] 629.0 2.10 1.07 0.46
50 584.4 583.0* 0.76 Thangiah et al. [8] 585.3 1.80 0.15 0.15
R103 10 476.6 496.0 0.71 Thangiah et al. [8] 489.0 1.50 2.60 1.41
30 507.0 520.0 1.05 Thangiah et al. [8] 510.9 1.50 0.77 1.75
50 475.6 505.0 0.59 Thangiah et al. [8] 495.0 2.70 4.08 1.98
R104 10 452.5 461.0 0.67 Thangiah et al. [8] 459.1 2.20 1.46 0.41
30 467.6 470.0 1.01 Thangiah et al. [8] 469.6 2.60 0.43 0.08
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

50 446.8 448.0 0.90 Thangiah et al. [8] 458.7 1.90 2.66 2.39
R105 10 565.1 565.1 5.60 Potvin et al. [29] 565.1 3.10 0.00 0.00
30 623.5 630.2 6.80 Potvin et al. [29] 630.2 2.40 1.07 0.00
50 591.1 592.1 5.40 Potvin et al. [29] 598.5 2.90 1.25 1.08
Total 8405.0 8509.1 37.26 — 8497.7 31.00 17.70 2.54
*Heuristic solution is less than optimal solution.

DOI: 10.1002/atr
J. Adv. Transp. 2013
Table IV. Computational result of benchmark problems for 50 customers.

Best known solution DEA


Difference (%)
Problem
type Backhaul (%) Optimal solution Total distance CPU Reference Total distance CPU Optimal S. Best known

Copyright © 2013 John Wiley & Sons, Ltd.


R101 10 1122.3 1138.1 16.20 Potvin et al. [29] 1138.3 9.00 1.43 0.02
30 1191.5 1192.7 16.90 Potvin et al. [29] 1245.8 9.00 4.56 4.45
50 1168.6 1183.9 17.50 Potvin et al. [29] 1183.9 10.00 1.31 0.00
R102 10 974.7 979.0 3.23 Thangiah et al. [8] 978.7 7.00 0.41 0.03
30 1024.8 1039.0 3.54 Thangiah et al. [8] 1046.0 8.00 2.07 0.67
50 1057.2 1062.0 3.67 Thangiah et al. [8] 1115.3 8.00 5.50 5.02
R103 10 811.4 849.0 4.18 Thangiah et al. [8] 831.1 11.00 2.43 2.11
30 882.8 931.0 3.56 Thangiah et al. [8] 895.1 12.00 1.39 3.86
VEHICLE ROUTING

50 — 903.0 3.78 Thangiah et al. [8] 887.7 11.00 0.63 1.69


R104 10 — 691.0 3.88 Thangiah et al. [8] 688.7 18.00 — 0.33
30 — 744.0 2.56 Thangiah et al. [8] 737.7 13.00 — 0.85
50 733.6 766.0 2.34 Thangiah et al. [8] 742.2 13.00 1.17 3.11
R105 10 970.6 1002.5 14.90 Potvin et al. [29] 972.8 10.00 0.23 2.96
30 1007.5 1047.8 15.20 Potvin et al. [29] 1030.0 12.00 2.23 1.70
50 993.4 1018.0 15.40 Potvin et al. [29] 1022.2 15.00 2.90 0.41
Total 12 820.5 14 547.0 126.84 14 515.5 166.00 26.26 6.07

DOI: 10.1002/atr
J. Adv. Transp. 2013
Table V. Computational result of benchmark problems for 100 customers.

Best known solution DEA


Difference (%)
Problem type Backhaul (%) Optimal solution Total distance CPU Reference Total distance CPU Optimal S. Best known

Copyright © 2013 John Wiley & Sons, Ltd.


R101 10 1767.9 1815.0 169.60 Potvin et al. [29] 1811.6 142.00 2.47 0.19
30 1877.6 1896.6 163.30 Potvin et al. [29] 1925.9 153.00 2.57 1.54
50 1895.1 1905.9 211.60 Potvin et al. [29] 1930.2 124.00 1.85 1.27
R102 10 1600.5 1653.2 121.00 Ropke and Pisinger [14] 1649.8 179.00 3.08 0.21
30 1639.2 1750.7 114.00 Ropke and Pisinger [14] 1758.2 141.00 7.26 0.43
50 1721.3 1746.0 24.40 Thangiah et al. [8] 1777.1 150.00 3.24 1.78
R103 10 — 1348.4 — Reimann et al. [10] 1356.3 165.00 — 0.59
30 — 1390.3 115.00 Reimann et al. [10] 1389.2 108.00 — 0.08
50 — 1456.5 115.00 Ropke and Pisinger [14] 1465.0 117.00 — 0.58
R104 10 — 1084.2 132.00 Ropke and Pisinger [14] 1105.4 138.00 — 1.96
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

30 — 1154.8 122.00 Ropke and Pisinger [14] 1146.5 147.00 — 0.72


50 — 1191.4 119.00 Ropke and Pisinger [14] 1199.6 123.00 — 0.69
R105 10 — 1544.8 — Reimann et al. [10] 1527.7 95.00 — 1.11
30 — 1583.3 102.00 Potvin et al. [29] 1582.6 112.00 — 0.04
50 — 1633.0 — Reimann et al. [10] 1608.6 103.00 — 1.49
Total 10 501.6 23 154.1 1508.90 23 233.7 1997.00 20.47 5.00

DOI: 10.1002/atr
J. Adv. Transp. 2013
VEHICLE ROUTING

problem [28]. R1-type problem instances consist of totally 100 randomly located demand nodes and a
depot node that is located in the center. Also, smaller problems are obtained by just considering the
first 25 and 50 demand nodes. For each basic problem, the capacity of the vehicle is 200, and the
planning horizon of a vehicle is 230 time units.
Tables 3–5, which are classified according to problem sizes (25, 50, and 100 demand nodes,
respectively), compare the performance of the proposed DEA with the optimal and best known
solutions obtained from existing studies. It should be noted that the columns due to “Difference%” of
the tables indicate the percentage gap of the DEA solutions with the optimal and best known solutions.
The results for 25 customers, which have been reported in Table 3, shows that the proposed DEA
results are close to the optimal solutions and also better than the best known solutions in general.
The average percentage gap (%Gap) between optimal solutions and proposed DEA solutions is
1.18%, and the average gap between the best known solutions and the proposed DEA solutions is
0.17%. Furthermore, average computational time of this case (2.07 seconds) is less than the average
computational time of the best known solutions (2.48 seconds) even though this difference is about a
half second.
Table 4 presents the results of second case problems, which consider 50 demand nodes. Although
the average computational time is a little poor according to the best known solutions, DEA could
achieve new best solutions for many problems as in the first case results. Also, the average percentage
gap between best known solutions and proposed DEA is 0.41%, which demonstrates better
performance than the first case.
Finally, the last case of the DEA performance test results are shown in Table 5. For this case, average
results are slightly worse according to small-sized instances for both of the objective function values
and computational times. However, note that our algorithm could obtain seven new best solutions.
Summarily, for all of the 45 instance results, DEA outperforms the existing methods with respect to
average travel distances such that DEA exhibits 0.08% better performance on average and 27 solutions
of the instances indicated with bold characters in tables, are better than or equivalent to the best known
solutions. Additionally, our algorithm has found close results to optimal solutions especially for the
small-sized instances. On the other hand, computational times are compared with the existing methods,
and results show that DEA finds solutions in earlier time for some of the instances.
In the second part of the computational analysis, we performed our algorithm to catering firm
problem described earlier. The firm under this study constructs its vehicle routes manually on a
daily basis on the basis of the experience of staff and transportation reviews, and no advanced
computer programming has been used to organize this activity. Therefore, any feasible distribu-
tion plan has been applied for daily distributions. For this reason, DEA is implemented to
increase the effectiveness and efficiency of the firm’s distribution activities and also reduce the
total transportation costs. For the computational applications, we could compare our algorithm
for only one case because the catering firm has a single distribution center and fixed truck routes.
Nevertheless, we run our algorithm with various parameter combinations described in Section 4
to reach best solution for the catering firm problem. A total of 81 experiments with different
parametric settings were performed, and according to these experiments, the best solution found
by DEA is 817.96 km and provides a 22.32% better solution than the existing route distance
(1052.52 km), which is used by the catering firm. Moreover, the computational times took less
than 5 minutes on average, and it is reasonable to conclude that DEA can be implemented for
the catering firm to support decisions on vehicle routing.

6. CONCLUSION

In this study, we proposed differential evolution meta-heuristic algorithm to solve VRPBTW. We


discussed the case of a catering firm that uses a meal delivery system to serve 148 customers.
The mathematical formulation of this problem was defined as the VRPBTW, and CPLEX was
used to solve the problem. An optimum solution could be found for up to 32 customers, and
the modified DEA with different encoding operator was used to solve the actual problem. To
evaluate the efficiency of our algorithm, we tested DEA with benchmark problem data sets and
compared the results with the optimal and best known solutions. It was also shown that the

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
İ. KÜÇÜKOĞLU AND N. ÖZTÜRK

proposed algorithm could provide acceptable solutions with respect to the optimum solutions and
also mostly better than the best known solutions. Finally, proposed algorithm was applied to the
original problem, and the solutions showed that our algorithm provided 22.32% better route
distance than the existing route distance.

7. LIST OF ABBREVIATIONS

DEA Differential Evolution Algorithm


VRP Vehicle Routing Problem
VRPB Vehicle Routing Problem with Backhauls
VRPBTW Vehicle Routing Problem with Backhauls and Time Windows

REFERENCES
1. Tan KC, Lee LH, Ou K. Artificial intelligence heuristics in solving vehicle routing problems with time window
constraints. Engineering Applications of Artificial Intelligence 2001; 14:825–837.
2. Canen AG, Pizzolato ND. The vehicle routing problem. Logistics Information Management 1994; 7(1):11-13.
3. Bandara JMSJ, Wong SC. Special issue: some recent advances in transportation modeling, Editorial. Journal of
Advanced Transportation 2010 published online. DOI: 10.1002/atr.5670390302.
4. Repoussis PP, Tarantilis CD. Solving the fleet size and mix vehicle routing problem with time windows via adaptive
memory programming. Transportation Research Part C 2010; 18:695–712.
5. Archetti C, Feillet D, Gendreau M, Speranza MG. Complexity of the VRP and SDVRP. Transportation Research
Part C 2011; 19:741–750.
6. Laporte G. Fifty years of vehicle routing. Transportation Science 2009; 43(4):408–416.
7. Gelinas S, Desrochers M, Desrosiers J, Solomon M. A new branching strategy for time constrained routing problems
with application to backhauling. Annals of Operations Research 1995; 61:91–109.
8. Thangiah SR, Potvin JY, Sun T. Heuristic approaches to vehicle routing with backhauls and time windows.
Computers and Operations Research 1996; 23(11):1043-1057.
9. Duhamel C, Potvin J, Rousseau J. A tabu search heuristic for the vehicle routing problem with backhauls and time
windows. Transportation Science 1997; 31(1):49–59.
10. Reimann M, Doerner K, Hartl RF. Insertion based ants for vehicle routing problems with backhauls and time
windows. In Ant Algorithms, Dorigo M, et al. (eds). Springer: Berlin, 2002; 135–147.
11. Cheung RK, Hang DD. Multi-attribute label matching algorithms for vehicle routing problems with time windows
and backhauls. IIE Transactions 2003; 35(3):191–205.
12. Zhong Y, Cole MH. A vehicle routing problem with backhauls and time windows: a guided local search solution.
Transportation Research Part E 2005; 41:131–144.
13. Cho Y, Wang S. A threshold accepting meta-heuristic for the vehicle routing problem with backhauls and time
windows. Journal of the Eastern Asia Society for Transportation Studies 2005; 6:3022–3037.
14. Ropke S, Pisinger D. A unified heuristic for a large class of vehicle routing problems with backhauls. European
Journal of Operational Research 2006; 171:750–775.
15. Aghdaghi M, Jolai F. A goal programming model for vehicle routing problem with backhauls and soft time
windows. Journal of Industrial Engineering International 2008; 4(6):7–18.
16. Bard JF, Kontoravadis G, Yu G. A branch-and-cut procedure for the vehicle routing problem with time windows.
Transportation Science 2002; 36(2):250–269.
17. Storn R, Price K. Differential evolution – a simple and efficient heuristic for global optimization over continuous
spaces. Journal of Global Optimization 1997; 11:341–359.
18. Karaboğa D, Ökdem S. A simple and global optimization algorithm for engineering problems: differential evolution
algorithm. Turkish Journal of Electrical Engineering 2004; 12(1):53-60.
19. Mingyong L, Erbao C. An improved differential evolution algorithm for vehicle routing problem with simultaneous
pickups and deliveries and time windows. Engineering Applications of Artificial Intelligence 2010; 23:188–195.
20. Erbao C, Mingyong L. A hybrid differential evolution algorithm to vehicle routing problem with fuzzy demands.
Journal of Computational and Applied Mathematics 2009; 231:302–310.
21. Onwubolu G, Davendra D. Scheduling flow shops using differential evolution algorithm. European Journal of
Operational Research 2006; 171:674–692.
22. Wang L, Fu X, Mao Y, Menhas MI, Fei M. A novel modified binary differential evolution algorithm and its
applications. Neurocomputing 2012; 98:55–75.
23. Erbao C, Mingyong L. The open vehicle routing problem with fuzzy demands. Expert Systems with Applications
2010; 37:2405–2411.
24. Hou L, Zhou H, Zhao J. A novel discrete differential evolution algorithm for stochastic VRPSPD. Journal of
Computational Information Systems 2010; 6(8):2483–2491.

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr
VEHICLE ROUTING

25. Wisittipanich W, Kachitvichyanukul V. Two enhanced differential evolution algorithms for job shop scheduling
problems. International Journal of Production Research 2012; 50(10):2757–2773.
26. Onwubolu GC. Design of hybrid differential evolution and group method of data handling networks for modeling
and prediction. Information Sciences 2008; 178:3616–3634.
27. Mayer DG, Kinghorn BP, Archer AA. Differential evolution – an easy and efficient evolutionary algorithm for
model optimisation. Agricultural Systems 2005; 83:315–328.
28. Solomon MM. Algorithms for vehicle routing and scheduling problems with time windows constraints. Operations
Research 1987; 35(2):254–265.
29. Potvin J, Duhamel C, Guertin F. A genetic algorithm for vehicle routing with backhauling, Applied Intelligence
1996; 6(4):345–355.

Copyright © 2013 John Wiley & Sons, Ltd. J. Adv. Transp. 2013
DOI: 10.1002/atr

Das könnte Ihnen auch gefallen