DISCRETE AND CONTINUOUS

DYNAMICAL SYSTEMS
SUPPLEMENT 2007

Website: www.AIMSciences.org
pp. 240249

SOLVE THE VEHICLE ROUTING PROBLEM WITH TIME

WINDOWS VIA A GENETIC ALGORITHM

Yaw Chang and Lin Chen

Mathematics and Statistics
University of North Carolina at Wilmington
601 S. College Rd
Wilmington, NC 28403, USA
Abstract. The objective of vehicle routing problem (VRP) is to deliver a set
of customers with known demands on minimum-cost routes originating and
terminating at the same depot. A vehicle routing problem with time windows
(VRPTW) requires the delivery made in a specific time window given by the
customers. Prins (2004) proposed a simple and effective genetic algorithm
(GA) for VRP. In terms of average solution, it outperforms most published
tabu search results. We implement this hybrid GA to handle VRPTW. Both
the implementation and computation results will be discussed.

1. Introduction. Vehicle routing problem (VRP) is defined on a undirected network G = (V, E) with a node set V = {0, 1, . . . , n} and an edge set E. Node 0 is a
depot. Each other node i > 0 represents a customer with a known service time di
and each edge has a non-negative travel cost cij = cji . The VRP consists of determining a set of m vehicle trips of minimum total cost, such that each vehicle starts
and ends at the depot, each client is visited exactly once, and the total demand
handled by any vehicle does not exceed the vehicles capacity W .
The vehicle routing problem is a well-known integer programming problem which
falls into the category of NP-Hard problems. For such problems it is often desirable
to obtain approximated solutions, so they can be found quickly enough and are
sufficiently accurate for the purpose. Usually this task is accomplished by using
various heuristic methods, which rely on some insight into the nature of the problem.
The VRP arises naturally as a central problem in the fields of transportation,
distribution, and logistics [4]. In some market sectors, transportation means a high
percentage of the value is added to goods. Therefore, the utilization of computerized
methods for transportation often results in significant savings ranging from 5% to
20% in the total costs, as reported in [11].
The vehicle routing problem with time windows (VRPTW) is the generalization
of the vehicle routing problem (VRP) with the additional time constraints. In these
problems, the service of a customer can begin within a time window [ei , li ] defined
by customer i. The vehicle can not arrive earlier than time ei and no later than
time li . The regular VRP can be viewed as the situation ei = 0 and li = for all
1 i n. When 0 < ei < li < , it is also known as a double-sided time window
while a single-sided time window is refer to either ei = 0 or li = , but not both. A
2000 Mathematics Subject Classification. Primary: 58F15, 58F17; Secondary: 53C35.
Key words and phrases. vehicle routing problem, vehicle routing problem with time window,
genetic algorithm.

240

SOLVE THE VRPTW VIA A GENETIC ALGORITHM

241

vehicle arriving earily than the earliest service time of a customer will incur waiting
time. This penalizes the scheduling either in the direct waiting cost or the increase
number of vehicles.
Due to its complexities and usefulness in real life, the VRPTW continue to
draw attention from researchers and has been a well-known problem in network
optimization. Early surveys of solution methods for VRPTW can be founded in
Golden and Assad [6, 7], Desrochers et al.[5], and Solomon and Desrosiers [10]. The
recent developments can be found in Ioannou, Kritikos, and Prastacos [8], Braysy
and Gendreau [2, 3], and Braysy, Dullaert, and Genderau [1]. The goal of this paper
is to adapt and implement Prins algorithm to handle VRP with time windows. For
simplicity, we will only consider a single-sided time window with ei > 0 and li =
for all i.
2. Problem Formulation. Before we formally give the mathematical formulation
of the vehicle routing problem with time windows, we need to introduce some basic
terminologies from graph theory.
Definition 1. An undirected graph G is an ordered pair G = (V, E) where V =
{0, 1, 2, ..., n} is the set of vertices and E = {(i, j)|i, j V } is the set of unordered
pairs of distinct vertices, called edges.
Definition 2. Let G = (V, E) be a graph, a path P is a sequence of vertices
i1 , i2 , ..., ik , such that (ij , ij+1 ) E, for all 1 j k 1.
Definition 3. Let G = (V, E) be a graph. G is called a connected graph if there is
a path between every pair of vertices .
Now let G be a given undirected graph corresponding to a vehicle routing problem
where vertex 0 corresponds to the depot and vertex i corresponds to customer i,
for 1 i n. For convenience, let us define the following known constants and
variables:
cij : the cost associated with edge (i, j), the cost could be the travel distance, the travel time, or the travel cost from customer i to customer j. For
simplicity, we assume cij = cji . If there is no edge between vertex i and j,
cij = .
bi : the vehicle arrival time at customer i.
di : the service time for customer i.
qi : the demand for customer i.
ei : the earliest arrival time for customer i.
L: maximal operation time for each vehicle.
Q: the vehicle capacity.
Ri : a sequence (i1 , i2 , ...ik ) corresponding to a route where ij V , such that
a particular vehicle will serve customers (i1 , i2 , ...ik ) according to the order of
the sequence.
Wij : the waiting time for vehicle i at customer j.
Consider a given route Ri = (i1 , i2 , ..., ik ), the actual arriving time bij can be
defined by
bij = bij1 + dj1 + cij1 ,ij ,
i.e., the arrival time at customer ij1 , plus the service time at customer ij1 and
the travel time from customer ij1 to customer ij . To ensure the earliest arrival

242

YAW CHANG AND LIN CHEN

time constraint, it can be modified as

bij = max{eij , bij1 + dij1 + cij1 ,ij }.
The waiting time for vehicle i at customer j can be defined as
Wij = max{0, eij bij1 dij1 cij1,j }.
An additional constraint
bik + dik + cik ,0 L,
will ensure a vehicle will not operate over the operation time allowed, ie. Ri is a
feasible route. The total travel time for route Ri can be defined as
C(Ri ) = c0,i1 +

k
X

(cij1 ,ij ) +

j=2

k
X

W ij +

j=1

k
X

dij + cik ,0 .

j=1

A solution S is a collection of routes R1 , R2 , ..., Rl , such that each customer will

be covered by exactly one route Ri . The Vehicle Routing with time windows can
be formulated as the following optimization problem
min C(S) =
S

l
X

C(Ri ).

i=1

3. Chromosomes and Evaluation. Similar to most GA that a chromosome S

is a permutation of n positive integers, such that each integer is corresponding
to a customer without trip delimiters. A chromosome may be broken into several
different routes. For example, a chromosome with 5 customers, S = (1, 2, 3, 4, 5)
may be broken into R1 = (1, 2) and R2 = (3, 4, 5), or R1 = (1, 2, 3) and R2 = (4, 5);
etc.. Christian Prins [9] proposed an optimal splitting procedure to get the best
solution (among all the possible routes) respecting to a given chromosome. We
will review the main idea of this splitting procedure in Section 3.1 and discuss our
implementation in Section 3.2.
3.1. A Splitting Algorithm by Prins. Because a chromosome can be split into
many different routes, Prins [9] proposed a splitting procedure which can find an
optimal splitting i.e a trip delimiter so that the total cost is minimized. The main
idea can be described as follows. Without loss of generality, let S = (1, 2, 3, ..., n)
where V =
be a given chromosome. Consider an auxiliary graph H = (V , E)

{0, 1, 2, ..., n}. An arc (i, j) E if

Eij = c0,i+1 +

j1
X
k=i+1

(dk + Ck,k+1 ) + dj + cj,0 L;

j
X

qk Q.

(1)

k=i+1

Then Eij is the total travel cost(time) for the route (i + 1, i + 2, ..., j). An optimal
split for S corresponds to shortest path P from vertex 0 to vertex n in H.
The following example (Prins, [9]) demonstrates the construction. The first graph
of Fig. 1 shows a sequence S=(a, b, c, d, e). The number associates with each edge is
the travel time. Let us assume the service time (delivery time) is 0 and the demands
are 5, 4, 4, 2 and 7, respectively. The capacity of the vehicle is 10 and the maximum
operation time is 90. Then the auxiliary graph H is the second graph in Fig. 1.
The edge in H with weight 55 corresponds to the travel time of the trip (0, a, b, 0).
The weight associated with the other edges are similarly defined. The shortest path

SOLVE THE VRPTW VIA A GENETIC ALGORITHM

243

from 0 to e is (0, b, c, e) with minimal total travel time 205. The last graph gives
the optimal result with three trips.
The auxiliary graph helps us understand the idea how to split a given chromosome
S into optimal trips. But in practice, we do not have to construct such graph
H. It can be done by a labeling algorithm and a splitting procedure [9]. Let
S = (1, 2, ..., n) be a given chromosome. Two labels Vj and Pj for each vertex j in
S are computed. Vj is the cost of the shortest path from node 0 to node j in H,
and Pj is the predecessor of j on this path. The minimal cost is given at the end by
Vn . For any given i, note that the increment of j stops when L or Q are exceeded.
The labeling algorithm can be described as follow:

V0 := 0
for i := 1 to n do Vi := + endfor
for i := 1 to n do
cost := 0; load := 0; j := i
repeat
load := load + qSj
if i = j then
cost := c0,Sj , +dSj + cSj,0
else
cost := cost cSj1 ,0 + cSj1 ,Sj + dSj + cSj,0
endif
if (cost L) and (load < Q) then
if Vi1 + cost < Vj then
Vj := Vi1 + cost
Pj := i 1
endif
j := j + 1
endif
until (j > n) or (cost > L) or (load > Q)
endfor

3.2. Implementation for Time Windows. Recall Eij in (1) represents the total
travel time for route (i + 1, i + 2, ..., j). To incorporate the earliest arrival time ei ,
the total travel time Eij for the same route can be calculated by
Pj1
Eij = c0,i+1 + k=i+1 (Wi + dk + ck,k+1 ) + Wj + dj + cj,0 ,
where Wi is the waiting time defined earlier. If Eij L, then the arc (i, j) exists
in the auxiliary graph H, and the shortest path from 0 to n in H corresponding to
an optimal split for S.
To accommodate this modification, the splitting procedure can be modified as
follows:

244

YAW CHANG AND LIN CHEN

V0 := 0
for i := 1 to n do Vi := + endfor
for i := 1 to n do
cost := 0; load := 0; j := i
repeat
load := load + qSj
if i = j then
cost := max[c0,Sj , eSj ] + dSj + cSj,0
else
cost := max[cost cSj1 ,0 + cSj1 ,Sj , eSj ] + dSj + cSj,0
endif
if (cost L) and (load < Q) then
if Vi1 + cost < Vj then
Vj := Vi1 + cost
Pj := i 1
endif
j := j + 1
endif
until (j > n) or (cost > L) or (load > Q)
endfor
Here c0,Sj is the regular arriving time at customer Sj , and eSj is the earliest arrival
time at customer Sj . If eSj < c0,Sj , we will pick eSj . Otherwise, we will pick the
regular arrival time c0,Sj .
4. CROSSOVER. As mentioned earlier, the reproductive process in GA can be
done either by crossover or by mutation procedures. A crossover (or recombination)
operation is performed upon the selected chromosomes. We will use the order
crossover which can be explained as follow: First, two chromosomes are randomly
selected from the initial population and the least-cost one becomes the first parent
P1 . Two cutting sites i and j are randomly selected in P1 , here i=4 and j=6. Then,
the substring P1 (i)...P1 (j) is copied into C(i)...C(j). Finally, P2 is swept circularly
from j + 1 onward to complete C with the missing nodes. C is also filled circularly
from j + 1. The other child maybe obtained by exchanging the roles of P1 and P2 .
Table 1 demonstrates the process. Let i = 4 and j = 6, so C(4) = P1 (4),
C(5) = P1 (5), and C(6) = P1 (6). Now C(7) should equal to P2 (7). Because
C(6) = 3, we shift to P2 (8), so C(7) = P2 (8) = 2. Now C(8) should equal to P2 (9).
Again C(5) = 10 = P2 (9), we will skip P2 (9) and let C(8) = P2 (10) = 1. Repeat
this process until C is constructed.
5. MUTATION. The classical genetic algorithm framework must be hybridized
with some kind of mutation procedure. In this paper, we simply adopt the local
search procedure in [9].
A child C produced by crossover could be improved by local search with a fixed
mutation rate pm which stands for the rate or the probability of mutation and is
a fundamental parameter in genetics and evolution. Here pm is the measure of the
impact of mutations on the child C. The mutation procedure can be described
as follows. A chromosome C is converted into a VRPTW solution. Then for all
possible pairs of distinct vertexes (u, v), the following simple moves are tested. Let

SOLVE THE VRPTW VIA A GENETIC ALGORITHM

245

c
30

25

d
30

10
25

40

15

a
35
20

55

95

40
0

40

55
50

115
60

150
80

70

e 205

85
90

120

Trip 2
60
Trip 1

55

Trip 3
90

Figure 1. An example of chromosome evaluation

x and y be the successors of u and v in their respective trips. The following rules
are used to create new chromosomes.
M1. Remove u from C and insert it after v, ie, the new sequence is (...v, u, ...).
M2. If x is a client, remove (u, x) from C then insert (u, x) after v, ie, the
new sequence is (...v, u, x, ...).
M3. If x is a client, remove (u, x) from C then insert (x, u) after v, for example,
the new sequence is (...v, x, u, ...).
M4. Swap u and v, for example, the new sequence is (...v, ..., u, ...).
M5. If x is a client, swap (u, x) and v, ie, the new sequence is (...v, ...u, x, ...).
M6. If x and y are clients, swap (u, x) and (v, y), ie, the new sequence is
(...v, y, ...u, x, ...).

246

YAW CHANG AND LIN CHEN

Rank 1 2

i=4

j=6

10

P1

9 8

10

P2

9 8

10

7 6 4
5
10
3
2 1 9
Table 1. Example of crossover

M7. If u and v are in the same trip, replace (u, x) and (v, y) by (u, v) and
(x, y), ie, the new sequence is (...u, v, ...x, y, ...).
M8. If u and v are not in the same trip, replace (u, x) and (v, y) by (u, v) and
(x, y), ie, the new sequence is (...u, v, ...x, y, ...).
M9. If u and v are not in the same trip, replace (u, x) and (v, y) by (u, y) and
(x, v), ie, the new sequence is (...u, y, ...x, v, ...).
Let C be a new chromosome produced by the mutation procedure. If C is
better than the worst in the current population, C will be added into the current
population, and the worst one will be moved. If C is better than the current best
solution, we consider the current mutation process as a productive mutation.
6. THE IMPLEMENTED ALGORITHM FOR VRPTW. The population
is implemented as an array of ( is the population size) chromosomes, always
sorted in increasing order of cost to ease the basic genetic algorithm iteration. Thus,
the best solution is 1 .
Clones (identical solutions) are forbidden in to ensure a better dispersal of
solutions. This also allows a higher mutation rate pm by local search, giving a
more aggressive genetic algorithm. To avoid comparing chromosomes in details and
to speed-up clone detection, we impose a stricter condition: the costs of any two
solutions generated by crossover or mutation must be spaced at least by a constant
4 > 0.
A population satisfying the following condition will be said to be well-spaced.
The simplest form with 4 = 1 (solutions with distinct integer costs) was already
used in an efficient Genetic Algorithm. That is,
P1 , P2 : P1 6= P2 |F (P1 ) F (P2 )| 4
where P1 and P2 are the parents we select from the initial population, and F (P1 )
and F (P2 ) are the costs of P1 and P2
The stopping criterion for GA is either after reaching a maximum number of
general iterations, max (number of crossovers that do not yield a clone) or a maximal number of iterations without improving the best solution, max . Let be
the population size, pm be the mutation rate, and space = 5. The GA can be
described as follows:
Step 1: (Initial population)

SOLVE THE VRPTW VIA A GENETIC ALGORITHM

247

Randomly generate chromosomes.

Step 2: (Sort Initial population)
Sort the initial population according to the total travel time of each chromosome.
Step 3: (setup counter)
Set , =0
Step 4: (Crossover)
Let = + 1. Select two parents (P1 ) and (P2 ) by binary tournament. Apply
Crossover to (P1 , P2 ) and select one child C at random.
Step 5: (Mutation)
Compare r (randomly generated) and pm . If r > pm then go to Step 6. Otherwise,
mutation is applied to the child and C 0 is generated. The split procedure is then
used to get the solution of C 0 .
Step 6: (Productive iteration)
Since 1 is the best solution of the population, if (F (C 0 ) F (k )) , 1 k ,
then let = 0 and shift C 0 to re-sort population. Otherwise, = + 1.
Step 7: (Check stopping criterion)
If = max or = max , stop. Otherwise, go to Step 4.
7. COMPUTATIONAL RESULTS. In general, some good solutions must be
computed using heuristics to build initial populations. Because the testing problems
size are relative small, we just build the initial populations randomly without using
any heuristics. This will be implement when we test our algorithm on standard
VRPTW instances.
The Genetic Algorithm was implemented in the Visual Basic script, on a Pentium4 PC clocked at 2.7 GHz under the operating system Windows XP. We ran two different randomly generated data sets with n=10. For each data set we ran 3 different
population sizes with 3 different mutation rates. For all the cases, we set max =1000
and max =100. The best solution was calculated by complete enumeration for both
date sets.
In the following tables, the first column is the population size, the second column is the mutation rate, and the third column is the best solution in the initial
population with the number of trips in the parentheses. Columns 4 and 5 give the
final values of and , respectively. Column 6 gives the final solution produced
by the GA with number of trips in the parentheses. Column 7 is the best solution
with the number of trips in the bracket. The last column is the relative error.
Table 2, sorted in increasing order of population size, shows very encouraging
results. When the population size equals 30 and 100, the mutation rate equals to
0.3, and our solutions reach the best solution with reaching the maximal number
of iterations without improving the best solution, which means that the mutation
procedure was applied to the child C at least 100 times.
In general, when the mutation rate increases, the relative errors are decreasing,
and the number of trips is decreasing. Only two relative errors are greater than 5%.
One exception that we notice is that when the population size equals 50 and the
mutation rate equals 0.1, our solution also reaches the best solution that has two
trips instead of one.
Table 3 is based on a second data set, sorted in increasing order of population
size. It shows results similar to those in table 1. When population size equals 50
and 100, mutation rate equals to 0.3. Our solutions also reach the best solution with
reaching the maximal number of iterations without improving the best solution.

248

YAW CHANG AND LIN CHEN

Population
size n

Mutation
Rate r

Original
Solution

Final
Solution

Runtime
t(seconds)

Best
Solution

Relative
Error

30

0.1

1380(3)

928

100

520(1)

90

490(1)

6.12%

0.2

1550(2)

990

100

510(1)

91

490(1)

4.08%

0.3

1280(3)

551

100

490(1)

98

490(1)

0%

0.1

1050(2)

1000

87

490(2)

90

490(1)

0%

0.2

990(2)

1000

98

510(1)

90

490(1)

4.08%

0.3

1360(1)

996

100

500(1)

98

490(1)

2.04%

0.1

1270(3)

1000

98

520(2)

90

490(1)

6.12%

0.2

1823(4)

864

100

500(1)

93

490(1)

4.08%

490(1)

0%

50

100

0.3
1900(4) 980 100 490(1)
106
Table 2. Computational results for date set 1
Population
size n

Mutation
Rate r

Original
Solution

Final
Solution

Runtime
t(seconds)

Best
Solution

Relative
Error

30

0.1

1787(4)

506

100

1022(2)

85

962(2)

6.3%

0.2

2629(5)

1000

75

997(2)

88

962(2)

3.6%

0.3

1860(4)

579

100

987(2)

96

962(2)

2.6%

0.1

1640(4)

583

100

987(2)

90

962(2)

2.6%

0.2

1744(4)

573

100

987(2)

89

962(2)

2.6%

0.3

1702(4)

516

100

962(2)

97

962(2)

0%

0.1

1787(4)

506

100

1022(2)

89

962(2)

6.3%

0.2

1441(3)

1000

83

970(2)

93

962(2)

0.83%

0.3

1518(3)

1000

71

962(2)

97

962(2)

0%

50

100

Table 3. Computational results for date set 2

SOLVE THE VRPTW VIA A GENETIC ALGORITHM

249

Also, as population size and mutation rate increase, the solutions are getting better,
the relative errors are decreasing, and the number of trips is decreasing. Only two
relative errors are greater than 5%.
8. CONCLUSION. We implement a hybrid GA Vehicle Routing Problem to
solve The Vehicle Routing Problem with Time Windows. The final results are
encouraging when the customer size is small. We will have more comparable experiments with large number of customers in the future.
For simplicity, we only consider the earliest arrival time and assume the capacity
of the vehicle is unlimited. We will re-examine the module to handle the capacity
cases in the future. We would also like to implement the current algorithm to handle
a real time window.
Acknowledgements. The authors are grateful to editor and referees for suggestion
to improve the paper.
REFERENCES
[1] O. Br
aysa, W. Dullaert and M. Gendreau, Evolutionary Algorithms for the Vehicle Routing
Problem with Time Windows, J. of Heuristics, 10 (2004), 587611.
[2] O. Br
aysa and M. Gendreau, Vehicle Routing Problem with Time Windows, Part I: Route
Construction and Local Search Algorithms, Transportation Science, 39 (2005), 104118.
[3] O. Br
aysa and M. Gendreau, Vehicle Routing Problem with Time Windows, Part I: Metaheuristics, Transportation Science, 39 (2005), 119139.
[4] G. B. Dantzig and R. H. Ramser, The Truck Dispatching Problem, Management Science, 6
(1959), 8081.
[5] M. Desrochers, J. K. Lenstra, M. W. S. Savelsbergh and F. Soumis, Vehicle Routing with Time
Windows: Optimization and Approxinmation, In B. Golden and A. Assad (eds.) Vehicle
Routing: Methods and Studies, Amsterdam: Elsevier Science Publishers, 1988.
[6] B. L. Golden and A. A. Assad, Perspectives on Vehicle Routing: Exciting New Developments,
Operations Research, 34 (1986), 803809.
[7] B. L. Golden and A. A. Assad, Vehicle Routing: Methods and Studies, Amsterdam: Elsevier
Science Publishers, 1988.
[8] G. Ioannou, M. Kritikos and G. Prastacos, A Greedy Look-Ahead Heuristic for the Vehicle
Routing Problem with Time Windows, J. of Operational Research Society, 52 (2001), 523537.
[9] Christian Prins, A simple and effective evolutionary algorithm for the vehicle routing problem,
Computer and Operations Research, 31 (2004), 19852002.
[10] M. M. Soloman and J. Desrosiers, Time Window Constrained Routing and Scheduling Problems, Transportation Science, 22 (1988), 113.
[11] P. Toth and D. Vigo, The Vehicle Routing Problem, Monographs on Discrete Mathematics
and Applications, SIAM, Philadelphia, 2001.

Received September 2006; revised June 2007.

E-mail address: [email protected]