You are on page 1of 14

254 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO.

2, MARCH 2012

Optimizing the Vehicle Routing Problem With


Time Windows: A Discrete Particle Swarm
Optimization Approach
Yue-Jiao Gong, Student Member, IEEE, Jun Zhang, Senior Member, IEEE, Ou Liu, Rui-Zhang Huang,
Henry Shu-Hung Chung, Senior Member, IEEE, and Yu-Hui Shi, Senior Member, IEEE

Abstract—Vehicle routing problem with time windows routes for a fleet of vehicles to serve a set of customers. Each
(VRPTW) is a well-known NP-hard combinatorial optimization route starts from and terminates at the depot. Each customer is
problem that is crucial for transportation and logistics systems. visited once and only once by exactly one vehicle. The routes
Even though the particle swarm optimization (PSO) algorithm is
originally designed to solve continuous optimization problems, in should satisfy the vehicle capacity and the time window con-
this paper, we propose a set-based PSO to solve the discrete com- straints. Besides, the VRPTW has two objectives. The primary
binatorial optimization problem VRPTW (S-PSO-VRPTW). The objective is to minimize the number of vehicle (NV) routes. The
general method of the S-PSO-VRPTW is to select an optimal sub- secondary objective is to minimize the total travel distance (TD)
set out of the universal set by the use of the PSO framework. As with the same number of routes.
the VRPTW can be defined as selecting an optimal subgraph out
of the complete graph, the problem can be naturally solved by In the literature, many researchers have made great effort to
the proposed algorithm. The proposed S-PSO-VRPTW treats the solve the VRPTW and proposed both exact and approximate
discrete search space as an arc set of the complete graph that is approaches. Desrosiers et al. [3] developed an exact approach
defined by the nodes in the VRPTW and regards the candidate by the use of column generation. Kolen et al. [4] described a
solution as a subset of arcs. Accordingly, the operators in the al- branch-and-bound method to minimize the total TD. Madsen [5]
gorithm are defined on the set instead of the arithmetic operators
in the original PSO algorithm. Besides, the process of position applied the Lagrangian relaxation to compute the lower bounds.
updating in the algorithm is constructive, during which the con- Kohl et al. [6] introduced a Dantzig–Wolfe decomposition ap-
straints of the VRPTW are considered and a time-oriented, nearest proach to divide paths and applied a branch-and-bound method
neighbor heuristic is used. A normalization method is introduced to minimize the distance for each path. These exact approaches
to handle the primary and secondary objectives of the VRPTW. dealt with only the small-scale VRPTW (except the work in [6],
The proposed S-PSO-VRPTW is tested on Solomon’s benchmarks.
Simulation results and comparisons illustrate the effectiveness and which has succeeded in to solve a few problems with up to 100
efficiency of the algorithm. customers) because the required computational time accumu-
lates exponentially as the number of customers increases [2].
Index Terms—Combinatorial optimization problems (COPs),
set-based particle swarm optimization (S-PSO), vehicle routing Therefore, researchers started to pay more attention to ap-
problem with time windows (VRPTW). proximate approaches that are capable to find near-optimal so-
lutions within a shorter computational time. In 1995, Rochat and
I. INTRODUCTION Taillard [7] applied a probabilistic diversification and intensifi-
cation technique to a local search algorithm for the VRPTW.
HE VEHICLE routing problem with time windows
T (VRPTW) is a crucial issue in transportation and logis-
tics systems [1], [2]. The problem can be defined as designing
In 1997, a tabu search heuristic was developed by Taillard
et al. [8], which further improved the best known solutions
reported in [7] and other previous works. Later, the local search
and the tabu search have been widely used to solve the VRPTW,
as in the works of Nanry and Barnes [9], Cordeau et al. [10],
Manuscript received January 19, 2010; revised November 15, 2010; accepted
March 24, 2011. Date of publication May 27, 2011; date of current version Bräysy and Gendreau [11], Ho and Haugland [12], Qi et al. [13],
February 17, 2012. This work was supported in part by the Project U0835002 Shaw [14], Li et al. [15], Ibaraki et al. [16], Lee et al. [17], Bent
of National Natural Science Foundation of China Joint Fund, and in part by Na- and Hentenryck [18], and Danna and Pape [19]. Besides, Chiang
tional Natural Science Foundation of China under Grant 61070004. This paper
was recommended by Associate Editor J. Lazansky. and Russell [20] proposed simulated annealing, while Czech
Y.-J. Gong and J. Zhang (corresponding author) are with the Department of and Czarnas [21] and Debudaj-Grabysz and Czech [22] ap-
Computer Science, Sun Yat-sen University, Guangzhou, P.R. China, the Key plied parallel simulated annealing. The aforementioned heuris-
Laboratory of Digital Life, (Sun Yat-sen University), Ministry of Education,
P.R. China, and the Key Laboratory of Software Technology, (Sun Yat-sen Uni- tic/metaheuristic approaches have succeeded in dealing with the
versity), Education Department of Guangdong Province, P.R. China (e-mail: relatively large-scale VRPTW.
junzhang@ieee.org). Because of the recent advancements in evolutionary compu-
O. Liu and R.-Z. Huang are with the School of Accounting and Finance, The
Hong Kong Polytechnic University, Kowloon, Hong Kong. tation (EC), much research effort has been paid on to develop
H. S.-H. Chung is with the Department of Electronic Engineering, City Uni- evolutionary algorithms (EAs) to solve the VRPTW. As a fa-
versity of Hong Kong, Kowloon, Hong Kong. mous population-based metaheuristic algorithm, genetic algo-
Y.-H. Shi is with the Department of Electrical and Electronic Engineering,
Xi’an Jiaotong-Liverpool University, Suzhou, Jiangsu, P.R. China. rithm (GA) has been widely and successfully used [23]–[30].
Digital Object Identifier 10.1109/TSMCC.2011.2148712 Tan et al. [25], [26] authored two papers on the usage of the GA

1094-6977/$26.00 © 2011 IEEE


GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 255

for the VRPTW. In [25], the GA incorporated an integer string characterize the problem space of discrete combinatorial opti-
representation, several novel crossovers, and an adaptive muta- mization problems (COPs) very well. The VRPTW belongs to
tion scheme. In [26], the chromosomes in the GA are finetuned COPs. That is, the problem space of the VRPTW is a directed
by a local search approach. Results showed the effectiveness complete graph, and each candidate solution is a spanning sub-
of the hybrid GA for the VRPTW. Recently, Alvarenga [27] graph of the complete graph. Thus, the VRPTW can be naturally
combined the GA with a set partition formulation, Ghoseiri and solved by PSO by the use of a set-based methodology. In this pa-
Ghannadpour [28] used a push-forward insertion heuristic in per, a set-based PSO to optimize the VRPTW (S-PSO-VRPTW)
the initialization of the GA, and Nazif and Lee [29] designed is proposed. Different from previous PSO algorithms to solve
a specific crossover operator that is based on graph theory for the VRPTW [52]–[60], in the S-PSO-VRPTW, operations de-
the VRPTW. Besides the GA, other evolutionary techniques, fined on set and possibilities are applied. We define a universal
such as ant colony optimization (ACO), are also employed to set as the arc set of the complete graph in the VRPTW. Each
solve the VRPTW. Gambardella et al. [31] developed a MACS- candidate solution is a crisp subset of the arc set. Initialized by a
VRPTW algorithm in which the two objectives of the VRPTW nearest neighbor heuristic (NNH) with a predefined possibility
are optimized simultaneously by the search of two ant colonies. (otherwise, the particle is randomly initialized), the particle flies
Tan et al. [32] enhanced the MACS-VRPTW by the modifi- through the search space by dynamically adjusting its velocity
cation of the heuristic that is used in the algorithm. Chen and according to historical search experience. What should be em-
Ting [33] combined ACO with simulated annealing, and Yu et al. phasized is that the S-PSO-VRPTW generates new positions in
[34] combined ACO with tabu search. Repoussis et al. [35] pro- a constructive way. The constraints of the VRPTW are satis-
posed a novel arc-guided EA. There were also parallel struc- fied during the construction. The algorithm combines the two
tures to solve the VRPTW, such as the works by Berger and objectives of the VRPTW into one objective by the use of a
Barkaoui [36], [37] and Bouthillier and Crainic [38]. Moreover, novel decision-making method. An arc tangent function is used
some multiobjective algorithms were proposed to deal with the to normalize the TD within the range of [0, 1], which makes the
problem in recent years [39], [40]. objective to minimize the NV in preference to the objective to
Particle swarm optimization (PSO) that was first proposed by minimize the TD. This normalization method to handle the two
Kennedy and Eberhart [41] in 1995 is a relatively new technique. objectives of the VRPTW is novel, simple, and useful. It is not
In recent years, PSO has been extensively studied [42]–[46] and only appropriate for the S-PSO-VRPTW, but also potentially
widely used [47]–[51]. In PSO, each particle “flies” through useful in other approaches to solve the VRPTW.
the problem space by the adjustment of its “flying” trajectory The main contribution of this paper is the proposed PSO,
according to its previous best performance (self-cognitive) and i.e., S-PSO-VRPTW, to solve the VRPTW. Experimental re-
the historical best performance in its neighborhood (social in- sults show the advantages of the proposed algorithm, which
fluence). This search mechanism enables PSO to have high op- refreshes the best known results for Solomon’s VRPTW bench-
erating efficiency, fast convergence speed, and implementation marks [2]. More importantly, the S-PSO-VRPTW is the first
simplicity. These advantages of PSO and its great success to PSO algorithm that has been tested on all Solomon’s 56 bench-
solve various NP-hard problems encourage researchers to de- marks. The promising results will encourage other researchers
velop PSO algorithms for the VRPTW. However, most existing to develop PSO algorithms for the VRPTW. On the other hand,
PSO algorithms are operated in continuous space, where the the success of our work extends the application field of the PSO
trajectories of particles are defined as the changes of their phys- algorithm to discrete COPs.
ical positions. In order to apply the PSO algorithm to solve the The rest of this paper is organized as follows. Section II
VRPTW, in [52]–[60], some variants of PSO algorithms have presents a formal definition of the VRPTW. Section III gives
been proposed. However, most of these approaches have simi- a brief description on continuous and set-based PSO. In
larities of applying the PSO as if the particles were searching in Section IV, the S-PSO-VRPTW is proposed and described. Nu-
continuous space and, then, converted the real-value positions merical results are presented in Section V. Conclusion is given
into integer values to represent routes. These PSO algorithms in Section VI.
cannot fully embody the characteristics of the discrete prob-
lem space of the VRPTW. The operating efficiency of these
approaches is unsatisfactory. The approaches that are proposed II. VEHICLE ROUTING PROBLEM
in [52]–[58] only solved the VRPTW with fewer than ten cus- Capacitated vehicle routing problem (CVRP), which is the
tomers and could not deal with the large-scale VRPTW. The most elementary version of the VRP, can be defined as fol-
work in [59] was just a test to find out whether the swarm lows [31]: n customers are waiting to be served, each of which
moved toward the feasible region of the problem. The PSO al- requires a quantity qi of goods (i = 1, 2,. . ., n). A depot has a
gorithm in [60] was tested on a few Solomon’s 100-customer fleet to deliver the goods. Each vehicle has a capacity of Q units,
benchmarks, but performed unsatisfactorily in datasets R. i.e., the total demand of customers served on each route could
Recently, Chen et al. [61] proposed a new methodology of not exceed Q. Therefore, the vehicle has to periodically return
PSO that is based on set and possibility theory. The search space to the depot for reloading, which is equivalent to dispatching a
is treated as a universal set, and each candidate solution that is new vehicle for delivery. Besides, each customer must be vis-
represented by each particle’s position is regarded as a crisp ited once and only once by exactly one vehicle. A solution of
subset of the universal set. This novel methodology is found to the CVRP is a collection of routes that a vehicle starts from
256 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

the depot, serves some specific customers, and terminates at the 


v

depot. y0k = v (7)


Define G = (V, A) as a directed complete graph, k =1

where V = (c0 , c1 , . . . ,cn ) is the node set, and A = ti + wi + si + tij = tj ∀i, j = 0, 1, . . . , n, i = j (8)
{ci , cj  |ci , cj ∈ V, ci = cj } is the arc set. c0 represents the
ej ≤ tj ≤ lj ∀j = 0, 1, . . . , n (9)
depot, and ci represents the customer i (i = 1, 2,. . ., n). Each
node is associated with a demand quantity qi , among which wi = max{ei − ti , 0} ∀i = 0, 1, . . . , n. (10)
node c0 is associated with q0 = 0. Each arc ci , cj  is associated
with a travel time tij , which is represented by the Euclidean dis- Constraint (5) denotes that the quantity of goods that each ve-
tance dij between nodes ci and cj (dij = dj i ). The objective of hicle carries could not exceed the capacity Q. Constraint (6)
the CVRP is to minimize the total travel time of all the vehicles. stands for that each customer can only be served by one vehicle.
The VRPTW is an extension of the CVRP. In the VRPTW, Constraint (7) represents that all the routes start from the depot.
nodes are associated with more properties, and the solution has Formulas (8)–(10) define the time window constraint, where ti
to satisfy more constraints. A service time si is considered; is the time when the vehicle arrives at node i; wi is the waiting
therefore, the vehicle has to stay at the location of customer ci time of a vehicle at the customer location until the time ei ; si
for a time interval at least si (s0 = 0 is associated with the depot is the service time; and tij is the travel time between nodes i
c0 ) for service. A time window [ei , li ] during which the service and j.
has to start is considered. Therefore, when a vehicle arrives at The CVRP is NP-hard [1], and in addition, the VRPTW has
customer ci earlier than ei , it has to wait until the beginning of more constraints and is, therefore, harder to solve [2]. In [62],
the time window to serve the customer. On the other hand, if a Savelsberg proved that to find a feasible solution for the VRPTW
vehicle cannot arrive at ci before li , the vehicle cannot serve ci . is NP-hard. Researches on the VRPTW have great theoretical
At this time, customer ci should be served by another vehicle. and practical values.
For depot c0 , the time window is defined as that e0 is the earliest
start time, and l0 is the latest return time of all the vehicles. III. PARTICLE SWARM OPTIMIZATION
The VRPTW has two objectives. The primary objective is to Since 1995, PSO has undergone extensive improvements
minimize the number of the vehicle routes v. The secondary and various new versions have been derived [43]–[46]. As a
objective is to minimize the total TD with the same number of population-based search algorithm, in PSO, a group of ran-
routes. domly initialized individuals explore the search space simul-
The VRPTW can be stated mathematically as follows. taneously and, finally, converge to the optimal solution of the
Define variable problem. Different from other EC techniques, PSO is inspired
 by the behavior of bird flocking (or fish schooling). Individuals
1, if vehicle k travels directly from i to j in the population that are called particles are associated with ve-
xkij =
0, otherwise locities. Particles fly through the search space by dynamically
 adjusting their velocities according to historical search experi-
1, if customer i is served by vehicle k
yik = ence. The design method and the algorithm process of PSO are
0, otherwise .
very easy to understand and simple to implement. Moreover,
The goal of the VRPTW is to minimize PSO algorithms have been found to have a global search ability,
fast convergence speed, strong robustness, etc. The algorithm is
min Z1 = v (1) a very powerful optimization technique. In recent years, PSO
and algorithms have been widely used in a great many applications
[49]–[52].

n 
n 
v
min Z2 = tij × xkij (2)
i=0 j =0 k =1 A. Particle Swarm Optimization in Continuous Space
s.t. The original PSO is used for continuous optimizations. It

n proceeds as follows. Each particle i in the population consists
xkij = yjk ∀k = 1, . . . , v, ∀j = 1, . . . , n (3) of three D-dimensional vectors including the velocity vector,
i=0 i.e., Vi = [Vi1 , Vi2 , . . . , ViD ], the current position vector, i.e.,

n Xi = [Xi1 , Xi2 , . . . , XiD ], and the previous best position vector,
xkij = yik ∀k = 1, . . . , v, ∀i = 1, . . . , n (4) i.e., pBesti = [pBest1i , pBest2i , . . . , pBestD i ], where D is the
j =0 dimensionality of the search space. Moreover, the whole pop-
ulation maintains a global best-so-far position vector gBest =

n
yik × qi ≤ Q ∀k = 1, . . . , v (5) [gBest1 , gBest2 , . . . , gBestD ]. An overall flowchart of PSO is
i=0 shown in Fig. 1. In each iteration of the evolutionary process,
particle i learns from its own search experience pBesti and

v
yik = 1 ∀i = 1, . . . , n (6) the swarm’s search experience gBest to update its velocity Vi
k =1 and position Xi . The update rules of a particle’s velocity and
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 257

B. Set-Based Particle Swarm Optimization


Original PSO algorithms are commonly used to solve contin-
uous optimization problems, whereas S-PSO [61] is proposed
to solve discrete COPs. COPs can be defined as (S, f, Ω), where
S is the set of candidate solutions that is called the problem
space, f is the objective function that calculates the cost value
for each candidate solution, and Ω is the set of constraints. A
set-based representation method is proposed in S-PSO to char-
acterize the discrete search space of COPs [61]. The search
space is represented by a universal set S. The elements in S can
be divided into D dimensions. Each dimension stands for a sub-
set Si (i = 1, . . . ,D) of S, i.e., S = S 1 ∪ S 2 ∪ · · · ∪ S D . Each
candidate solution X is a subset of S, which satisfies X ⊆ S
via X = X 1 ∪ X 2 ∪ · · · ∪ X D , X i ⊆ S i (i = 1, . . . , D). X is
considered to be a feasible solution if and only if it satisfies all
the constraints in Ω. The objective of S-PSO is to find a feasible
solution X∗ (to choose a subset out of S) that minimizes the cost
function f.
The framework of S-PSO is identical with original PSO al-
gorithms, which is shown in Fig. 1. The velocity and position
update rules in S-PSO also have the same format as that in origi-
nal PSO algorithms that are shown in formulas (11) and (12) (or
formulas (13) and (12) if it employs the comprehensive learn-
ing strategy of CLPSO), respectively. However, the particle’s
position in S-PSO is represented by a crisp subset, i.e., X ⊆ S,
instead of the numerical digits in original PSO algorithms, and
the velocity is represented by a set with possibilities. There-
Fig. 1. Flowchart of PSO algorithms.
fore, the operators in formulas (11) and (12) are defined on set
and possibilities instead of the arithmetic operators in original
position are very simple, which are defined as PSO algorithms. The definitions of these operators when us-
ing S-PSO to solve the VRPTW will be described in the next
Vid = ω × Vid + c1 × randd1 × (pBestdi − Xid ) + c2 × randd2 section.
× (gBestd − Xid ) (11)
IV. SET-BASED PARTICLE SWARM OPTIMIZATION FOR VEHICLE
Xid = Xid + Vid (12)
ROUTING PROBLEMS WITH TIME WINDOWS
where ω is the inertia weight; c1 and c2 are the acceleration In this section, the S-PSO-VRPTW is proposed and described.
coefficients that determine the relative weight of self-cognitive As we know, the problem space of the VRPTW is a directed
and social influence; randd1 and randd2 are two random numbers complete graph, i.e., G = (V, A), and each candidate solution is
that are uniformly distributed over [0, 1]; i represents the current a spanning subgraph of G. The candidate solutions are judged
particle; and d represents the current dimension. to be feasible solutions if they satisfy the constraints that are
Liang et al. [45] proposed a representative PSO variant that is defined by formulas (3)–(10). The goal of the VRPTW is to
called comprehensive learning PSO (CLPSO). The CLPSO uses find a feasible solution that minimizes the NV and the total TD.
a novel velocity update rule to prevent premature convergence Because of these features of the VRPTW, the problem can be
as naturally solved by S-PSO (as described in Section III that the
  basic methodology of S-PSO is to select an optimal subset out
Vid = ω × Vid + c × randd × pBestdfi (d) − Xid (13)
of the universal set).
where fi (d) ∈ {1, 2, . . . , M } (M is the population size) defines In the S-PSO-VRPTW, the search space of the particle swarm
that the dth dimension of which particle’s pBest particle i should is the arc set A of the complete graph G. Each particle’s position
follow. The decision of fi (d) depends on a probability Pc. That is represented by a crisp set of arcs, which is a subset of A,
is, for each dimension of particle i, a random number “rand” in to denote a feasible solution of the problem. The velocity is
a range [0, 1] is generated. If rand is larger than Pc, the corre- represented by a set of arcs with possibilities. The arcs may
sponding dimension of particle i will learn from its own pBest, be selected to construct the new position of the particle, while
i.e., fi (d) = i. Otherwise, particle i will employ a tournament an arc associated with a larger possibility is more likely to be
selection to choose another particle’s pBest to learn from. Exper- selected in the construction. As the position and velocity in the
imental results in [45] demonstrated that CLPSO performs very S-PSO-VRPTW are represented as arc sets and possibilities, the
well when solving complex multimodal optimization problems. corresponding velocity and position update rules are defined as
258 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

2) Velocity: The velocity of a particle is defined as


Vi = [Vi0 , Vi1 , . . . , Vin ] (16)
Vid = {u, v/p(u, v)|u, v ∈ A } d
(17)
where Ad is the set of arcs that are adjacent to node d in the com-
plete graph, and p(u, v) ∈ [0, 1] is the corresponding possibility
of each arc u, v. p(u, v) denotes the possibility of u, v to be
selected when updating the position of a particle, which will be
illustrated later. If p(u, v) = 0, arc u, v is omitted in Vid .

B. Velocity Updating
As the problem space of the VRPTW is always unknown
and complex, the S-PSO-VRPTW employs the comprehensive
learning strategy of CLPSO to prevent premature convergence,
i.e., the velocity updating is according to formula (13). However,
the S-PSO-VRPTW employs novel operators that are defined on
set and possibilities instead of the arithmetic operators in orig-
inal PSO algorithms. Inspired by the work in [61], the defini-
tions of these operators are shown in (18)–(21), respectively.
Fig. 2. Illustrations of the constraints-based decoder in the S-PSO-VRPTW. The coefficient×velocity operator and the velocity+velocity
(a) Decode an edge. (b) Decode a solution. operator are defined as changing the value of p(u, v), the
position−position operator is defined as the subtraction of the
two arc sets, and the coefficient×(position−position) operator
operators and procedures that are based on these arc sets and
is defined as converting a crisp set into a set with possibilities:
possibilities. 
c × Vid = {u, v/p (u, v) u, v ∈ Ad },
A. Particle Representation 
1, if c × p(u, v) > 1
p (u, v) = (18)
1) Position: In the S-PSO-VRPTW, the position of a particle c × p(u, v), otherwise
is represented by 
Vid + Vjd = {u, v/ max(pi (u, v), pj (u, v)) u, v ∈ Ad }
Xi = [Xi0 , Xi1 , . . . , Xin ] (14)
(19)
Xid = [nb1 , d, d, nb2 ], nb1 , nb2 ∈ {0, 1, . . . , d − 1, d + 1, n}, 
Xid − Xjd d 
= U = {u, v u, v ∈ Xi and u, v ∈
d
/ Xj }
d

nb1 = nb2 (15)


(20)

where Xi is an arc set with each dimension Xid that is composed c × U d = {u, v/p (u, v) u, v ∈ Ad },
of two arcs adjacent to node d. In (15), n is the number of nodes; ⎧
d is the index of current dimension; nb1 is the previous node of ⎨ 1, if u, v ∈ U d and c > 1
node d, and nb2 is the subsequent node of node d. The position p (u, v) = c, if u, v ∈ U d and 0 ≤ c ≤ 1 (21)

of each particle represents a directed Hamilton circle in the 0, if u, v ∈ / Ud.
complete graph.
For example, suppose that Vi1 = {1, 3/0.5, 1, 4/0.2,
The Hamilton circle can be converted into a collection of ve-
4, 1/0.3}, Xi1 = {5, 1, 1, 3}, pBest1f i (1) = {4, 1,
hicle routes, which stands for a feasible solution of the problem
by the use of a constraint-based decoder. The illustration of the 1, 3}, ω = 0.7, c = 2.0, rand1 = 0.3. Then, we have ω ×
constraint-based decoder is shown in Fig. 2. Starting from the Vi1 = {1, 3/0.35, 1, 4/0.14, 4, 1/0.21}, pBest1f i (1) −
outarc of the depot, for each arc in the Hamilton circle, if the Xi1 = {4, 1}, and c × rand1 × (pBest1f i (1) −Xi1 ) = {4, 1/
arc satisfies all the constraints in the VRPTW, it would be pre- 0.6}. Finally, it can be deduced that the new velocity Vi1 = ω ×
served; otherwise, the depot would be inserted between the two Vi1 + c × rand1 × (pBest1f i (1) − Xi1 ) = {1, 3/0.35, 1, 4
nodes, and the arc is replaced by two new arcs. This process /0.14, 4, 1/0.6}.
is equivalent to sending a vehicle to deliver the goods along By learning from pBest1f i (1) , the corresponding possibility of
the Hamilton circle at the beginning, and when the vehicle is arc 4, 1 in the velocity set is improved. The example clearly
unable to serve the next customer on the circle because of the shows that the significance of updating of the arcs with possibil-
capacity or the time window constraints, one more new vehicle ities in the velocity set of the S-PSO-VRPTW is identical with
route is created. By the definitions of the particle position and the significance of the velocity updating in original PSO algo-
the constraint-based decoder, each particle represents a feasible rithms, i.e., adjusting the velocity by learning from the historical
solution of the VRPTW in the S-PSO-VRPTW. best information.
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 259

C. Position Updating
The position of particle i is updated by (12), where the opera-
tor “+” is also redefined in the S-PSO-VRPTW. In the position
updating process, we first convert the velocity Vi into a crisp set
by

Cut(Vid ) = {u, v u, v/p(u, v) ∈ Vid and p(u, v) ≥ rand}
(22)
where rand ∈ [0, 1] is a random number. For an arc u, v, it is
kept in the velocity set if and only if its corresponding possibility
p(u, v) is not smaller than rand. This way, the arc with a larger
p(u, v) is more likely to be preserved in the crisp set after the
conversion.
Then, the new position (Xi ) is built in a constructive way,
during which the capacity and time window constraints are
taken into account. (Xi ) is set as an empty set at first. In
accordance with that each vehicle starts from the depot, the
construction starts at (Xi0 ) of (Xi ) , selects an arc adjacent
to node 0, and chooses the next customer iteratively. Sup-
pose the current customer that the vehicle is serving is k,
the next customer m that the vehicle will visit comes from
three crisp sets: the SV = {m |k, m ∈ Vi , k, m satisfies Ω},
SX = {m |k, m ∈ Xi , k, m satisfies Ω}, and SA =
{m |k, m ∈ A , k, m satisfies Ω}. If there are nodes avail-
able in SV , m is selected from SV . Otherwise, if there are nodes
available in SX , m is selected from SX . Otherwise, m is selected
from SA . After selecting m, arc k, m is added to (Xi ) . When
there is no available node in all the SV , SX , and SA , the con-
straints of the VRPTW cannot be satisfied unless a new route Fig. 3. Pseudocode for the position update procedure in the S-PSO-VRPTW.
is created. At this point, we suppose a depot node is inserted
after k and, then, reselect m. This way, the feasibility of (Xi ) is
guaranteed. After constructing (Xi ) , the current position Xi of
this possibility, the particle is initialized by the NNH using the
particle i is replaced by (Xi ) . The pseudocode of the position
heuristic defined by (23). Otherwise, the particle is randomly
update rule is shown in Fig. 3.
initialized.
Moreover, we employ a time-oriented NNH to select better
elements. In other words, the selection of m in the position
updating is heuristic based. In every available set of nodes, the D. Objective Function
customer that is “closest” to the current customer k will be
selected. We use a matrix to record the heuristic information It has been noted in Section II that the VRPTW has two
between every two nodes, in which the closeness is defined objectives. The S-PSO-VRPTW incorporates a novel decision
according to making method to deal with the primary and secondary objec-
tives of the VRPTW. It is easy to implement and is demon-
heuristick m = bk m + rk m = (tk m + wm ) + (lm − tm ) strated to be very effective by the experimental results in
(23) Section V. The fitness of each particle depends both on the NV
where the definitions of tk m , wm , lm , and tm have been given in and the total TD. Notice that the NV, i.e., the primary objective
Section II; bk m = tk m + wm represents the time span between of the VRPTW, is an integer number, which indicates that every
the time when the vehicle completes the service for customer k two candidate solutions with different NV are certain to satisfy
and the time when the vehicle begin to serve customer m; rk m = |NV(Xi ) − NV(Xj )| ≥ 1. Therefore, if we normalize the TD
lm − tm defines the remaining time of customer m allowed to be within the range of [0, 1], the objective to minimize the NV is
served. in preference to the objective to minimize the TD. The objec-
At the beginning of the S-PSO-VRPTW, the position of each tive function of the S-PSO-VRPTW is defined by (25) and (26).
particle is initialized by Here, an arc tangent function is introduced for normalization,
 and the graph of which is shown in Fig. 4:
NNH(), if rand < ϕ
Xi = (24)
Random(), otherwise
fitness (Xi ) = NV(Xi ) + normalize (TD(Xi )) (25)
where “rand” is a random number that is uniformly distributed
over [0, 1], and ϕ is the possibility of greedy initialization. Under normalize (x) = arctan(x)/(π/2). (26)
260 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

V. NUMERICAL RESULTS
This section presents the numerical experimental results for
the S-PSO-VRPTW. The experiments are done on Solomon’s
benchmarks that consist of six different datasets [2]. Datasets
C have clustered customers, and the time windows are gener-
ated by the known vehicle routes for these clustered customers.
Datasets R have randomly generated customers that are dis-
tributed uniformly over a square. Datasets RC have mixed cus-
tomers of datasets C and R. Datasets C1, R1, and RC1 have
narrow time windows and small delivery capacity, which allow
few customers to be served by one vehicle. In contrast, datasets
C2, R2, and RC2 have large time windows and capacity, which
allow more customers to be served by one vehicle.
Fig. 4. Arc tangent function for normalization.

A. Experimental Settings
As mentioned earlier, the S-PSO-VRPTW employs the
CLPSO framework. In the experiments, the inertia weight ω
in (13) is initialized as 0.9 and linearly decreased from 0.9 to
0.4 during the training process, and the acceleration coefficient
c is set as 2.0. Besides, the refreshing gap “rg” and learning
probability Pc of CLPSO are set according to what reported
in the paper which proposed CLPSO [45]; the population size
is set as M = 20; and the possibility of greedy initialization is
set as ϕ = 0.3. We employ a flexible terminal condition that
if the gBest vector ceases the improvement for a certain num-
ber of generations (called stopping gap “sg”), the algorithm is
terminated. Solomon’s 25-customer/50-customer/100-customer
benchmarks are tested in this paper. For the 25-customer bench-
marks, sg is set as 1000, whereas for the 50-customer and 100-
customer benchmarks, sg is set as 10 000. Our algorithm is
coded in C and runs on a machine with Intel Pentium 4 CPU,
Fig. 5. Illustrations of the local search operator. (a) Before local search. 2.79 GHz/512 MB of RAM. Each test is conducted 30 times
(b) After local search. independently.
In the remaining of this section, the experimental results are
presented. The results obtained by the S-PSO-VRPTW are com-
E. Additional Local Search pared with the best-so-far results that are reported at Solomon’s
After each particle finishes updating its position, a simple lo- web page: http://web.cba.neu.edu/∼msolomon/problems.htm.
cal search scheme is applied. The scheme has potential to further These best known results are derived from 23 different algo-
reduce the required NV, which can accelerate the convergence rithms that include exact, heuristic, and metaheuristic algo-
of the particle swarm. The procedure of the local search for rithms. The algorithms consist of KDMSS [6], RT [7], TBGGP
particle i in the S-PSO-VRPTW is described as follows. [8], CLM [10], S97 [14], LLH [15], IKMUY [16], BVH [18],
Step 1: Choose the route r that has the fewest customers. DLP [19], CC [21], GTA [31], BB [37], S98 [63], CR [64],
Step 2: For each customer in r, try to insert it into the other L [65], HG [66], SSSD [67], H [68], KLM [69], RGP [70],
vehicle routes on the premise such that the insertion M [71], IV [72], and C [73].
does not change the start and the end time of the ser- In addition, we have also implemented a previous PSO
vice for all the customers in which the route is being algorithm for the VRPTW termed I-PSO [53]. In this paper, the
inserted and satisfies the constraints of the VRPTW. I-PSO algorithm is tested on Solomon’s 25-customer instances,
Step 3: If all the customers in r can be inserted into other with parameter settings and stopping criterion the same as in the
routes, route r is deleted, and the position of particle i S-PSO-VRPTW. Furthermore, the proposed S-PSO-VRPTW
is updated. Otherwise, the position of particle i stays is compared with some other state-of-the-art population-based
the same. techniques that include the GAs that are proposed by Tan et al.
As Fig. 5 shows, route 3 has the fewest customers; thus, the (HGA) [26] and Nazif and Lee (OCGA) [29], and a hybrid
local search intends to insert the three customers of route 3 into algorithm combining ACO with tabu search (ACO-TS) [34].
route 1 and 2. If the insertions succeed, route 3 is deleted, and The detailed results of the three algorithms to optimize
the position of the particle is updated. Otherwise, the position Solomon’s 100-customer instances are reported in [26], [29],
is kept as that before the local search. and [34], respectively.
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 261

TABLE I
RESULT COMPARISONS FOR SOLOMON’S 25-CUSTOMER INSTANCES

B. Results and Comparisons routes that we obtained is equal to that of the best known, which
means that the S-PSO-VRPTW produces the best or near-best
Tables I–III show the result comparisons, in which “NV”
result, the problem is marked with (∗), and the best result is
represents the NV routes and “TD” represents the total TD.
“Authors” denotes who have proposed the best results. “CPU” presented in italic.
1) 25-Customer Instances: As shown in Table I, for 25-
stands for the computational time (in minutes). The best and
mean results are presented. A problem marked with (∗∗) and customer instances, the S-PSO-VRPTW produces 29 out of
the best result in bold indicate that, for the instance, the S-PSO- 56 new best results, among which four come from the R1-
type datasets, ten come from the R2-type datasets, seven come
VRPTW provides new best result. In other cases, if the NV
from the C2-type datasets, and eight come from the RC2-type
262 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

TABLE II TABLE III


RESULT COMPARISONS FOR SOLOMON’S 50-CUSTOMER INSTANCES RESULT COMPARISONS FOR SOLOMON’S 100-CUSTOMER INSTANCES

In order to make further comparisons, we modify the stopping


criterion of I-PSO as that when its running time reaches the
computational time of the S-PSO-VRPTW for each instance, it
will be terminated. This way, results that are obtained by the
I-PSO and the S-PSO-VRPTW by the use of the same computa-
tional time are compared. However, we find that the new results
of I-PSO are almost the same as those reported in Table I. It
means that I-PSO converges fast to the local optimum and stops
evolving in an early generation. When using the flexible stop-
ping criterion as described earlier, the required computational
time of I-PSO is, thus, less than that of the S-PSO-VRPTW.
datasets. Compared with the previous algorithms that provided The comparisons in Table I and the aforementioned investi-
the best known, the S-PSO-VRPTW has better performance in gation show that the proposed S-PSO-VRPTW presents a much
exploring irregular search space (datasets R and RC). Moreover, better performance than I-PSO. This is because the I-PSO that
the proposed algorithm is very good at reducing the required uses the methodology of original continuous PSO is not well fit
NV. For the instances with large time windows and capacities, for the discrete VRPTW, whereas by the use of the methodology
the required NV is very small. Particularly, in some cases, all of S-PSO, the S-PSO-VRPTW is a natural technique to solve
the customers can be served by only one vehicle. Besides the 29 the VRPTW.
new best known, the S-PSO-VRPTW obtains another 26 solu- 2) 50-Customer Instances: As shown in Table II, for 50-
tions that are equal or close to the current best known solutions. customer instances, the S-PSO-VRPTW produces 32 out of 56
Therefore, the proposed algorithm performs well in optimizing new best results, among which seven come from the R1-type
55 out of the 56 25-customer instances. datasets, 11 come from the R2-type datasets, six come from the
In Table I, it can be observed that the S-PSO-VRPTW com- C2-type datasets, and eight come from the RC2-type datasets. It
prehensively outperforms I-PSO. However, the computational can be observed that the algorithm is very good at dealing with
time of I-PSO is less than that of the proposed S-PSO-VRPTW. instances with large time windows and vehicle capacities. For
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 263

Fig. 6. Schematic illustrations of the search process for R101, R201, and C101.

datasets R2 and RC2, the S-PSO-VRPTW refreshes all the best still competitive, especially, to deal with datasets with large time
known results that are published so far, while for datasets C2, windows including R2, C2, and RC2.
the S-PSO-VRPTW obtains six out of eight new best results. Furthermore, the S-PSO-VRPTW is compared with HGA
Besides, for 23 of the remaining 24 instances, the S-PSO- [26], OCGA [29], and ACO-TS [34]. In Table III, it can be
VRPTW produces best or near-best results compared with the observed that the S-PSO-VRPTW provides results with fewer
algorithms in the literature. Therefore, the S-PSO-VRPTW is a NV than HGA, OCGA, and ACO-TS for 31, 33, and 24 in-
very powerful technique to solve almost all the Solomon’s 50- stances, respectively. Only when optimizing R103 and RC108
customer instances. The only expectance is that, for instance, out of the 56 instances, HGA can achieve fewer NV than the
R106, the S-PSO-VRPTW requires a larger NV than KMDSS S-PSO-VRPTW. None of the NV that are produced by OCGA
[6]. is fewer than those obtained by the S-PSO-VRPTW. ACO-TS
3) 100-Customer Instances: In Table III, results of the provides fewer NV than the S-PSO-VRPTW only to solve R108
S-PSO-VRPTW that optimizes Solomon’s 100-customer in- and RC101. Moreover, taking the TD into consideration, we can
stances are presented. Compared with the best known results find that the S-PSO-VRPTW obtains the shorter TD than HGA,
that are summarized on Solomon’s website, which is gained OCGA, and ACO-TS in optimizing a lot of instances, such as
by 23 different algorithms, the S-PSO-VRPTW obtains 39 R101, R103, etc. These comparisons show the effectiveness and
best or near-best solutions, among which four come from R1- competitiveness of the usage of the S-PSO-VRPTW to solve
type datasets, nine come from R2-type datasets, eight come the VRPTW, for the proposed algorithm is good at reducing the
from C1-type datasets, eight come from C2-type datasets, three required NV routes as well as the total TD.
come from RC1-type datasets, and seven come from RC2-type We investigate the convergence behavior of the S-PSO-
datasets. Although no new best results are reported, the per- VRPTW. Figs. 6 and 7 show the convergence curves of the
formance of the S-PSO-VRPTW for 100-customer instances is best runs to optimize R101, R201, C101, C201, RC101, and
264 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

Fig. 7. Schematic illustrations of the search process for C201, RC101, and RC201.

RC201, respectively. It can be observed that the NV decreases dle the primary and secondary objectives of the VRPTW. This
as the iteration increases. The TD oscillates somewhat, but it method is not only appropriate for the S-PSO-VRPTW, but also
decreases within the same NV. The fitness primarily depends on potentially useful in other approaches to solve the VRPTW.
the vehicle number and, then, depends on the TD, and decreases The proposed S-PSO-VRPTW was the first PSO algorithm
as the iteration increases. The convergence behavior of the S- that was tested on all Solomon’s benchmarks. Experimental re-
PSO-VRPTW indicates the usefulness of our decision making sults illustrate the effectiveness and efficiency of the proposed
method for the two objectives of the VRPTW. algorithm, for we provide better results than the existing best
known results for Solomon’s instances. Moreover, the S-PSO-
VRPTW obtains promising results on almost all the tested in-
VI. CONCLUSION
stances, which shows the robustness of the algorithm. Compared
In this paper, we have proposed a set-based PSO to solve the with another PSO algorithm, two GAs, and a hybrid ACO algo-
VRPTW. In the S-PSO-VRPTW, the search space is a universal rithm, the S-PSO-VRPTW has been further demonstrated to be
set of arcs in the complete graph of the VRPTW. Each particle’s competitive to solve the VRPTW.
position that represents a set of delivery routes for a fleet of
vehicles has been defined as a subset of arcs and built construc-
tively. In this representation, the characteristics of the VRPTW
have been fully embodied. During the construction of particle’s ACKNOWLEDGMENT
position, the constraints of the VRPTW have been taken into ac- The authors would like to thank the Editor-in Chief, Asso-
count, and a time-oriented NNH has bee applied. Furthermore, ciate Editor, and Reviewers for their valuable comments and
a novel decision making method has been proposed to han- suggestions that improved the quality of the paper.
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 265

REFERENCES [25] K. C. Tan, L. H. Lee, Q. L. Zhu, and K. Ou, “Heuristic methods for vehicle
routing problem with time windows,” Artif. Intell. Eng., vol. 15, no. 3,
[1] P. Toth and D. Vigo, The Vehicle Routing Problem. Philadelphia, PA: pp. 281–295, 2001.
SIAM, 2001. [26] K. C. Tan, L. H. Lee, and K. Ou, “Artificial intelligence heuristics in
[2] M. M. Solomon, “Algorithms for the vehicle routing and scheduling prob- solving vehicle routing problems with time window constraints,” Eng.
lems with time window constraints,” Oper. Res., vol. 35, no. 2, pp. 254– Appl. Artif. Intell., vol. 14, no. 6, pp. 825–837, 2001.
265, 1987. [27] G. B. Alvarenga, G. R. Mateus, and G. D. Tomi, “A genetic and set
[3] J. Desrosiers, F. Soumis, and M. Desrochers, “Routing with time win- partitioning two-phase approach for the vehicle routing problem with
dows by column generation,” Networks, vol. 14, no. 4, pp. 545–565, time windows,” Comput. Oper. Res., vol. 34, no. 6, pp. 1561–1584,
1984. 2007.
[4] A. W. J. Kolen, A. H. G. R. Kan, and H. W. J. M. Trienekens, “Vehicle [28] K. Ghoseiri and S. F. Ghannadpour, “Hybrid genetic algorithm for vehicle
routing with time windows,” Oper. Res., vol. 35, no. 2, pp. 266–273, routing and scheduling problem,” J. Appl. Sci., vol. 9, no. 1, pp. 79–87,
1987. 2009.
[5] O. Madsen, “Lagrangian relaxation and vehicle routing,” Working Paper, [29] H. Nazif and L. S. Lee, “Optimized crossover genetic algorithm for vehicle
IMSOR, Tech. Univ., Denmark, 1990. routing problem with time windows,” Amer. J. Appl. Sci., vol. 7, no. 1,
[6] N. Kohl, J. Desrosiers, O. B. G. Madsen, M. M. Solomon, and F. Soumis, pp. 95–101, 2010.
“2-path cuts for the vehicle routing problem with time windows,” Transp. [30] W.-J. Wang, “Improved genetic algorithm for vehicle routing problem
Sci., vol. 33, no. 1, pp. 101–116, 1999. with time windows,” in Proc. Int. Conf. Intell. Comput. Cognitive Inf.,
[7] Y. Rochat and E. D. Taillard, “Probabilistic diversification and intensifi- 2010, pp. 203–206.
cation in local search for vehicle routing,” J. Heuristics, vol. 1, no. 1, [31] L. M. Gambardella, É. Taillard, and G. Agazzi, “MACS-VRPTW: A mul-
pp. 147–167, 1995. tiple ant colony system for vehicle routing problems with time windows,”
[8] É. Taillard, P. Badeau, M. Gendreau, F Guertin, and J. Y. Potvin, “A tabu in New Ideas in Optimization, New York: McGraw-Hill, pp. 63–76, 1999.
search heuristic for the vehicle routing problem with soft time windows,” [32] X. Tan, X. Zhuo, and J. Zhang, “Ant colony system for optimizing vehicle
Transp. Sci., vol. 31, no. 2, pp. 170–186, 1997. routing problem with time windows (VRPTW),” Lect. Notes Comput.
[9] W. P. Nanry and J. W. Barnes, “Solving the pickup and delivery problem Sci., vol. 4115, pp. 33–38, 2006.
with time windows using reactive tabu search,” Transp. Res. B, Method., [33] C.-H. Chen and C.-J. Ting, “A hybrid ant colony system for vehicle routing
vol. 34, no. 2, pp. 107–121, 2000. problem with time windows,” J. Eastern Asia Soc. Transp. Stud., vol. 6,
[10] J. F. Cordeau, G. Laporte, and A. Mercier, “A unified tabu search heuristic pp. 2822–2836, 2005.
for vehicle routing problems with time windows,” J. Oper. Res. Soc., [34] B. Yu, Z. Z. Yang, and B. Z. Yao, “A hybrid algorithm for vehicle routing
vol. 52, no. 8, pp. 928–936, 2001. problem with time windows,” Expert Syst. Appl., vol. 38, no. 1, pp. 435–
[11] O. Bräysy and M. Gendreau, “Tabu search heuristics for the vehicle routing 441, 2010.
problem with time windows,” TOP: Official J. Spanish Soc. Stat. Oper. [35] P. P. Repoussis, C. D. Tarantilis, and G. Ioannou, “Arc-guided evolutionary
Res., vol. 10, no. 2, pp. 211–237, 2002. algorithm for the vehicle routing problem with time windows,” IEEE
[12] S. C. Ho and D. Haugland, “A tabu search heuristic for the vehicle routing Trans. Evol. Comput., vol. 13, no. 3, pp. 624–647, Jun. 2009.
problem with time windows and split deliveries,” Comput. Oper. Res., [36] J. Berger and M. Barkaoui, “A memetic algorithm for the vehicle routing
vol. 31, no. 12, pp. 1947–1964, 2004. problem with time windows,” in Proc. 7th Int. Command Control Res.
[13] M.-Y. Qi, L.-X. Miao, L. Zhang, and H.-Y. Xu, “A new tabu search Technol. Symp., 2002.
heuristic algorithm for the vehicle routing problem with time win- [37] J. Berger and M. Barkaoui, “A parallel hybrid genetic algorithm for the
dows,” 15th Int. Proc. Conf. Manage. Sci. Eng., pp. 1648–1653, vehicle routing problem with time windows,” Comput. Oper. Res., vol. 31,
2008. no. 12, pp. 2037–2053, 2004.
[14] P. Shaw, “A new local search algorithm providing high quality solutions to [38] A. L. Bouthillier and T. G. Crainic, “A cooperative parallel meta-heuristic
vehicle routing problems,” Working Paper, Dept. of Comput. Sci., Univ. for the vehicle routing problem with time windows,” Comput. Oper. Res.,
Strathclyde, Glasgow, Scotland, 1997. vol. 32, no. 7, pp. 1685–1708, 2005.
[15] H. Li, A. Lim, and J. Huang, “Local search with annealing-like restarts [39] K. C. Tan, Y. H. Chew, and L. H. Lee, “A hybrid multi-objective evolu-
to solve the VRPTW,” Working Paper, Dept. Comput. Sci., Nat. Univ., tionary algorithm for solving truck and trailer vehicle routing problems,”
Singapore, 2001. Eur. J. Oper. Res., vol. 172, no. 3, pp. 855–885, 2006.
[16] T. Ibaraki, M. Kubo, T. Masuda, T. Uno, and M. Yagiura, “Effective [40] B. Ombuki, B. J. Ross, and F. Hanshar, “Multi-objective genetic algo-
local search algorithms for the vehicle routing problem with general time rithms for vehicle routing problem with time windows,” Appl. Intell.,
windows,” Working Paper, Dept. Appl. Math. Phys., Kyoto Univ., Japan, vol. 24, no. 1, pp. 17–30, 2006.
2001. [41] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc. IEEE
[17] L. H. Lee, K. C. Tan, K. Ou, and Y. H. Chew, “Vehicle capacity planning Int. Conf. Neural Netw., 1995, vol. 4, pp. 1942–1948.
system: A case study on vehicle routing problem with time windows,” [42] R. Poli, J. Kennedy, and T. Blackwell, “Particle swarm optimization: An
IEEE Trans. Syst., Man Cybern. A, Syst., Humans, vol. 33, no. 2, pp. 169– overview,” Swarm Intell., vol. 1, no. 1, pp. 33–57, 2007.
178, Mar. 2003. [43] J. Kennedy and R. Mendes, “Neighborhood topologies in fully informed
[18] R. Bent and P. V. Hentenryck, “A two-stage hybrid local search for the and best-of-neighborhood particle swarms,” IEEE Trans. Syst., Man,
vehicle routing problem with time windows,” Transp. Sci, vol. 38, no. 4, Cybern. C, Appl. Rev., vol. 36, no. 4, pp. 515–519, Jul. 2006.
pp. 515–530, 2004. [44] Z.-H. Zhan, J. Zhang, Y. Li, and H. S.-H. Chung, “Adaptive particle swarm
[19] E. Danna and C. Le Pape, “Accelerating branch-and-price with local optimization,” IEEE Trans. Syst., Man, and Cybern. B, Cybern., vol. 39,
search: A case study on the vehicle routing problem with time windows,” no. 6, pp. 1362–1381, Dec. 2009.
ILOG, Cédex, Tech. Rep. 03-006, 2003. [45] J. J. Liang, A. K. Qin, P. N. Suganthan, and S. Baskar, “Comprehensive
[20] W. C. Chiang and R. A. Russell, “Simulated annealing metaheuristics for learning particle swarm optimizer for global optimization of multimodal
the vehicle routing problem with time windows,” Ann. Oper. Res., vol. 63, functions,” IEEE Trans. Evol. Comput., vol. 10, no. 3, pp. 281–295, Jun.
no. 1, pp. 3–27, 1996. 2006.
[21] Z. J. Czech and P. Czarnas, “A parallel simulated annealing for the vehicle [46] Z. H. Zhan, J. Zhang, Y. Li, and Y. H. Shi, “Orthogonal learning particle
routing problem with time windows,” 10th Euromicro Workshop Parallel, swarm optimization,” IEEE Trans. Evol. Comput., to be published.
Distrib. Network-Based Process., 2002, pp. 376–383. [47] K. Veeramachaneni, L. A. Osadciw, and P. K. Varshney, “An adaptive
[22] A. Debudaj-Grabysz and Z. Czech, “Theoretical and practical issues of multimodal biometric management algorithm,” IEEE Trans. Syst., Man,
parallel simulated annealing,” Parallel Process. Appl. Math., vol. 4967, and Cybern.C, Appl. Rev., vol. 35, no. 3, pp. 344–356, Aug. 2005.
pp. 189–198, 2008. [48] C.-J. Lin, C.-H. Chen, and C.-T. Lin, “A hybrid of cooperative particle
[23] S. R. Thangiah, “Vehicle routing with time windows using genetic algo- swarm optimization and cultural algorithm for neural fuzzy networks and
rithms,” in Application Handbook of Genetic Algorithms: New Frontiers, its prediction applications,” IEEE Trans. Syst., Man, Cybern. C, Appl.
vol. 2. CRC Press, Boca Raton, FL, 1995, pp. 253–277. Rev., vol. 39, no. 1, pp. 55–68, Jan. 2009.
[24] K. Zhu, “A new genetic algorithm for VRPTW (2000),” in Proc. Int. Conf. [49] R. V. Kulkarni and G. K. Venayagamoorthy, “Bio-inspired algorithms for
Artif. Intell., 2000. autonomous deployment and localization of sensor nodes,” IEEE Trans.
266 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 42, NO. 2, MARCH 2012

Syst., Man, Cybern. C, Appl. Rev., vol. 40, no. 6, pp. 663–675, Nov. Yue-Jiao Gong (S’10) received the B.S. degree
2010. in computer science from Sun Yat-Sen University,
[50] P. Kanakasabapathy and K. S. Swarup, “Evolutionary tristate PSO for Guangzhou, China, in 2010, where she is currently
strategic bidding of pumped-storage hydroelectric plant,” IEEE Trans. working toward the Ph.D. degree.
Syst., Man, Cybern. C, Appl. Rev., vol. 40, no. 4, pp. 460–471, Jul. 2010. Her current research interests include artificial in-
[51] R. V. Kulkarni and G. K. Venayagamoorthy, “Particle swarm optimization telligence, evolutionary computation, swarm intelli-
in wireless-sensor networks: A brief survey,” IEEE Trans. Syst., Man, gence, and their applications in design and optimiza-
Cybern. C, Appl. Rev., vol. 41, no. 2, pp. 262–267, 2011. tion of intelligent transportation systems.
[52] F. Wang and Q.-Z. Wu, “Particle swarm optimization for vehicle routing
problem with time windows,” in Proc. Int. Conf. Risk Rel. Manage., 2008,
pp. 962–966.
[53] Q. Zhu, L. M. Qian, Y. C. Li, and S. J. Zhu, “An improved particle swarm
optimization algorithm for vehicle routing problem with time windows,”
in IEEE Congr. Evol. Comput., 2006, pp. 1386–1390.
[54] T. Zhen, H. Ge, Y. Che, and Y Jiang, “Research on particle swarm opti-
mization for grain logistics vehicle routing problem,” in Proc. IITA Int.
Conf. Serv. Sci., Manage. Eng., 2009, pp. 274–277.
[55] T. J. Ai and V. Kachitvichyanukul, “A particle swarm optimization for Jun Zhang (M’02–SM’08) received the Ph.D. degree
vehicle routing problem with time windows,” Int. J. Oper. Res., vol. 6, in electrical engineering from the City University of
no. 4, pp. 519–537, 2009. Hong Kong, Kowloon, Hong Kong, in 2002.
[56] X. X. Liu, W. G. Jiang, and J. W. Xie, “Vehicle routing problem with time He was a Brain Korean 21 Postdoctoral Fellow
windows: a hybrid particle swarm optimization approach,” in Proc. 5th with the Department of Electrical Engineering and
Int. Conf. Natural Comput., vol. 4, 2009, pp. 502–506. Computer Science, Korea Advanced Institute of Sci-
[57] W. G. Jiang, Y. B. Zhang, and J. W. Xie, “A particle swarm optimiza- ence and Technology, Daejeon, Korea, from 2003
tion algorithm with crossover for vehicle routing problem with time win- to 2004. Since 2004, he has been with Sun Yat-Sen
dows,” in Proc. IEEE Symp. Comput. Intell. Scheduling. 2009, pp. 103– University, Guangzhon, China, where he is currently
106. a Professor and the Ph.D. Supervisor with the Depart-
[58] S. Amini, H. Javanshir, and R. Tavakkoli-Moghaddam, “A PSO approach ment of Computer Science. He is the author of seven
for solving VRPTW with real case study,” Int. J. Res. Rev. Appl. Sci., research books and book chapters, and more than 80 technical papers in his
vol. 4, no. 3, pp. 118–126, 2010. research areas. His research interests include computational intelligence, data
[59] J. P. Castro, D. Landa-Silva, and J. A. M. Pérez, “Exploring feasible and mining, wireless sensor networks, operations research, and power electronic
infeasible regions in the vehicle routing problem with time windows using circuits.
a multi-objective particle swarm optimization approach,” Nat. Inspired Dr. Zhang is an Associate Editor of the IEEE TRANSACTIONS ON INDUS-
Cooperative Strategies Optim., vol. 236, pp. 103–114, 2009. TRIAL ELECTRONICS. He is also the Chair of the IEEE Beijing (Guangzhou)
[60] C. T. Lin, “Using predicting particle swarm optimization to solve the Section Computational Intelligence Society Chapter and the Chair of the IEEE
vehicle routing problem with time windows,” in Proc. IEEE Int. Conf. Guangzhou Subsection.
Ind. Eng. Eng. Manage., 2008, pp. 810–814.
[61] W. N. Chen, J. Zhang, H. Chung, W. L. Zhong, and Y. H. Shi, “A novel
set-based particle swarm optimization method for discrete optimization
problems,” IEEE Trans. Evol. Comput., vol. 14, no. 2, pp. 278–300, Apr.
2010.
[62] M. Savelsbergh, private communication, 1984.
[63] P. Shaw, “Using constraint programming and local search methods to
solve vehicle routing problems,” in Proc. Principles Pract. Constraint Ou Liu received the Ph.D. degree in information
Program., LNCS 1520, 1998, pp. 417–431. systems from the City University of Hong Kong,
[64] W. Cook and J. L. Rich, “A parallel cutting plane algorithm for the vehicle Kowloon, Hong Kong, in 2006.
routing problem with time windows,” Working paper, Dept. Comput. He has been with The Hong Kong Polytechnic
Appl. Math., Rice Univ., Houston, TX, 1999. University, Kowloon, Hong Kong, since 2006, where
[65] J. Larsen, “Parallelization of the vehicle routing problem with time win- he is currently an Assistant Professor with the School
dows,” Ph.D. dissertation, IMM-PHD-1999-62, Dept. Math. Modelling, of Accounting and Finance. His current research in-
Tech. Univ. Denmark, Lyngby, Denmark, 1999. terests include business intelligence, decision support
[66] J. Homberger and H. Gehring, “Two evolutionary metaheuristics for the systems, ontology engineering, genetic algorithms,
vehicle routing problem with time windows,” Inf. Syst. Oper. Res., vol. 37, ant colony systems, fuzzy logic, and neural networks.
pp. 297–318, 1999.
[67] G. Schrimpf, J. Schneider, H. Stamm-Wilbrandt, and G. Dueck, “Record
breaking optimization results using the ruin and recreate principle,” J.
Comput. Phys., vol. 159, no. 2, pp. 139–171, 2000.
[68] J. Homberger, Verteilt-parallele Metaheuristiken zur Tourenplanung.
Wiesbaden, Germany: Gaber, 2000.
[69] B. Kallehauge, J. Larsen, O. B. G. Madsen, “Lagrangian duality and non-
differentiable optimization applied on routing with time windows: Exper- Rui-Zhang Huang received the M.Phil. and Ph.D.
imental results,” Dept. Math. Modelling, Tech. Univ. Denmark, Lyngby, degrees from the Department of Systems Engineering
Denmark, Internal Report IMM-REP-2000-8, 2000. and Engineering Management, Chinese University of
[70] L. M. Rousseau, M. Gendreau, and G. Pesant, “Using constraint-based Hong Kong, Shatin, Hong Kong, in 2003 and 2008,
operators to solve the vehicle routing problem with time windows,” J. respectively.
Heuristics, vol. 8, no. 1, pp. 43–58, 2002. She has been with The Hong Kong Polytechnic
[71] D. Mester, “An evolutionary strategies algorithm for large scale vehicle University, Kowloon, Hong Kong, since 2007. She is
routing problem with capacitate and time windows restrictions,” Working the author or coauthor of several papers of some pres-
Paper, Inst. Evol., Univ. Haifa, Israel, 2002. tigious journals and conferences. She was involved in
[72] S. Irnich and D. Villeneuve, “The shortest path problem with k-cycle elim- many research works in various fields, such as min-
ination (k ≥ 3): Improving a branch-and-price algorithm for the VRPTW,” ing unseen named entity translations, semisupervised
INFORMS J. Comput., vol. 18, no. 3, pp. 391–496, 2006. document clustering, hierarchical incremental document clustering, active learn-
[73] A. Chabrier, “Vehicle routing problem with elementary shortest path based ing for clustering, language modeling, etc. Her current research interests include
column generation,” Comput. Oper. Res., vol. 33, no. 10, pp. 2972–2990, data mining, text mining, machine learning, information retrieval, artificial in-
2006. telligence, and knowledge systems.
GONG et al.: OPTIMIZING THE VEHICLE ROUTING PROBLEM WITH TIME WINDOWS 267

Henry Shu-Hung Chung (M’95–SM’03) received Yu-Hui Shi (SM’98) received the Ph.D. degree in
the B.Eng. and Ph.D. degrees in electrical engineer- electronic engineering from South-East University,
ing from The Hong Kong Polytechnic University, Nanjing, China, in 1992.
Kowloon, Hong Kong, in 1991 and 1994, respec- He is currently a Professor with the Depart-
tively. ment of Electrical and Electronic Engineering, Xi’an
He has been with the City University of Hong Jiaotong-Liverpool University, Suzhou, China, where
Kong, Kowloon, Hong Kong, since 1995, where he he is also the Director of the Research and Postgrad-
is currently an Associate Dean with the College of uate Office. Before joining Xi’an Jiaotong-Liverpool
Science and Engineering and a Professor with the University, he was with Electronic Data Systems Cor-
Department of Electronic Engineering. He is the au- poration, Indianapolis, IN. His main research inter-
thor of seven research book chapters and more than ests include the areas of computational intelligence
280 technical papers including 130 refereed journal papers in his research ar- techniques (including swarm intelligence) and their applications.
eas and holds 16 patents. His research interests include time- and frequency- Dr. Shi is the Editor-in-Chief of the International Journal of Swarm In-
domain analysis of power electronic circuits, switched-capacitor-based convert- telligence Research and an Associate Editor of the IEEE TRANSACTIONS ON
ers, random-switching techniques, control methods, digital audio amplifiers, EVOLUTIONARY COMPUTATION. He is the Chair of the IEEE Chief Information
soft-switching converters, electronic ballast design, and circuit design with evo- Officer Task Force on Swarm Intelligence.
lutionary computation techniques.
Dr. Chung is an Associate Editor of the IEEE TRANSACTIONS ON
POWER ELECTRONICS and the IEEE TRANSACTIONS ON CIRCUITS AND
SYSTEMS—PART II.

You might also like