2009 Fifth International Conference on Natural Computation

Vehicle Routing Problem with Time Windows: A Hybrid Particle Swarm

Optimization Approach

Xiaoxiang Liu, Weigang Jiang, Jianwen Xie

Department of Computer Science of Zhuhai College,
Jinan University, Zhuhai, Guangdong, 519070, China
[email protected], [email protected], [email protected]

Abstract involves a fleet of vehicles set-off from a depot to

serve a number of customers, at diverse geographic
Vehicle routing problem (VRP) is a well-known locations, with various demands and within specific
combinatorial optimization and nonlinear time windows before returning to the depot eventually.
programming problem seeking to service a number of Because a great deal of problem can be transformed
customers with a fleet of vehicles. This paper proposes into the VRPTW problem to deal with (e.g. post
a hybrid particle swarm optimization (HPSO) mailing, scheduling of trains and buses), and the
algorithm for VRP. The proposed algorithm utilizes the qualities of their solutions directly affect the qualities
crossover operation that originally appears in genetic of service, the research of VRPTW bas been paid more
algorithm (GA) to make its manipulation more readily and more attention.
and avoid being trapped in local optimum, and The method of solving VRPTW can be broadly
simultaneously for improving the convergence speed of classified into three generations [3]: the first generation
the algorithm, level set theory is also added to it. We was simple heuristic method developed in the 60’s and
employ the HPSO algorithm to an example of VRP,
70’s, which were mainly based on local search; the
and compare its result with those generated by PSO,
second generation was mathematical programming
GA, and parallel PSO algorithms. The experimental
based heuristic, which were near-optimization
comparison results indicate that the performance of
HPSO algorithm is superior to others, and it will algorithms that are very different from normal
become an effective approach for solving discrete heuristics; the third generation is optimization
combinatory problems. algorithm and artificial intelligence methods, such as
genetic algorithm, Tabu search, simulated annealing,
1. Introduction improved PSO algorithm, etc.
PSO (Particle Swarm Optimization) [4] is a bionic
Vehicle Routing Problems (VRP) was proposed for algorithm which simulates the flight of flocks of birds,
the first time by Dantzig and Ramser [1] in 1959. This making good effect for the various multidimensional
problem is a complex and challenging combinatorial continuous space optimization problems, but less in
optimization task, which consists in designing the research and application of discrete field [5]. In PSO
optimal set of routes for fleet of vehicles in order to algorithm, very few parameters are needed to be
serve a given set of customers. The interest in VRP is adjusted, which makes it particularly easy to
motivated by its practical relevance as well as by its implement. At present, there are lots of best-known
considerable difficulty. Theoretically, VRP is a NP- improved PSO algorithms, such as PSOBC [6]
hard problem, which means that it is believed that one algorithm, DRPSO algorithm [7] and so on. This paper
may never find a computational technique that will proposes a HPSO, which is based on genetic algorithm
guarantee optimal solutions to larger instances for such and level set, to solve VRPTW. The results of the
problems [2]. experiment show desirable effect.
Vehicle Routing Problem with Time Window
(VRPTW) is an extension of normal VRP, encountered 2. Description and mathematical model for
very frequently in making decision about the VRPTW
distribution of goods and services. The problem

978-0-7695-3736-8/09 $25.00 © 2009 IEEE 498

502
DOI 10.1109/ICNC.2009.353
 VRPTW is generally described as: given a central L

warehouse, owned K vehicles, the capacity of each min z = ¦¦¦ cij xijk + pE ¦ max (ET − s , 0)
i i

vehicle is denoted qk (k=1,2,…,K); now, transport tasks i j k i =1

of L delivery points should be accomplished, denoted L

by 1,2,…,L. The cargo of the i th delivery point is gi

+ pL ¦ max (s − ET , 0)
i i
(1)
i =1
(i=1,2,…,L). The time needed to complete the tasks
(loading or unloading) of the i th delivery point is ti. ­s.t. ¦ gi yki ≤ qk ∀k (2)
And the task i must be completed in time window [ETi, ° i

LTi], in which ETi denotes the earliest start time for the ° ¦ yki = 1 i = 1, 2,",L (3)
task i and LTi denotes the latest start time for the task i.
° k
°
If vehicles arrive delivery point i earlier than ETi, ® ¦ xijk = ykj j = 0,1," ,L; ∀k (4)
waiting occurs; if later than LTi, the task i will be ° i

delayed. We should get a solution for vehicle running ° ¦ xijk = yki i = 0,1,",L; ∀k (5)
route with transport demands and minimal travel cost. ° j

VRPTW can be divided into two categories: hard ° x , y = 0or 1 i, j = 0,1," ,L; ∀k
time widow VRPTW and soft time window VRPTW. ¯ ijk ki
(6)
Hard time window VRP is that each task must be
completed within the specific time. However, soft time This model is widely used. It can be converted into a
window VRP is that if a task can not be completed mathematical model of different issues by different
within the specific time, given a punishment. setting of the parameters. When Eq. (1) PE=PLĺĞ,
In this paper, we use the mathematical model the above model is the hard time window VRP. If Eq.
presented by literature [8], giving 0 to the central (1) ETi=0, LTiĺĞ, VRPTW model become a common
warehouse and giving 1,2,…,L to delivery points. Both VRP. If only one car is used, the problem will become
tasks and central warehouse are denoted by node i (i=0, TSP problem. If removing constraint (2), it becomes
1,…, L). Variables are defined as follows: TSPTW problem.

­1 the task of delivery point i is 3. Basic particle swarm algorithm

°
yki = ® completed by vehicle k , PSO was presented by Kennedy and Eberhart [6] in
°0 1995. The algorithm simulates flight behavior of flocks
¯ otherwise
of birds. Birds aim to achieve optimization results
­1 vehicle k travels from point i through a collective collaboration between the groups.
° In the PSO system, each candidate solution is called a
xijk = ® to point j .
"particle". A number of particles coexist and cooperate
°0
¯ otherwise to find optimization. Each particle "flies" to a better
position in problem space in accordance with its own
cij is the cost of transportation from the point i to the "experience" and the best "experience" of the adjacent
point j, it could be the meanings of distance, cost, time particle swarm, searching the optimal solution.
and so on. si is the time when vehicles arrive at the task Mathematical notation of PSO algorithm is defined
i. PE is the unit time of waiting cost when vehicle as follows: Assume searching space is D-dimensional
arrive at task i before ETi. PL is the unit cost of delay and the total number of particles is n. The ith particle
fines when vehicle arrive at task i after LTi. If vehicles location is denoted by the vector: Xi=(xi1, xi2,…, xiD);
arrive at task point i before ETi, opportunity cost PE(si- The past optimal location of the ith particle in the
ETi) will be increased; otherwise, if arriving after LTi, "flight" history (that is, the location corresponds
punishment cost PL(LTi-si) will be increased. optimal solution) is Pi=(pi1, pi2,…, piD). The past
Mathematical model of vehicle optimizing optimal location Pg of the gth particle is optimal in all
scheduling is as follows: of Pi(i=1,2,…,n); The location changing rate (speed) of
the ith particle is denoted by the vector
Vi=(vi1,vi2,…,viD). The location of each particle changes
by the following formula:

503
499
 vid (t + 1) = w × vid (t ) + c1 × rand () × [ pid (t ) − xid (t )] Vehicle 3: 0ė2ė6ė8ė0.
+ c2 × rand () × [ p gd (t ) − xid (t )] (7) 4.2. Crossover operation
xid (t + 1) = xid (t ) + vid (t + 1) (1 ≤ i ≤ n; 1 ≤ d ≤ D ) (8) This paper presents a crossover manner, which is
improved on the CX manner [12], in order to retain the
where c1, c2 are positive constants called accelerating
good gene fragments of parents. Suppose two parent
factor; rand() is a random number between 0 and 1; w
individuals are f1 and f2, through crossover operation
is called inertia factor; w, set a litter greater, is suited to
to be the offspring individuals new1 and new2.
a wide range of exploration to solution space while
Specific operation such as˖
smaller is suited to a small range. [Xmin, Xmax] is the
Step 1: Randomly select one cross region (a, b) in
changing range of particle location. [vmin, vmax] is the
the second string;
changing range of speed. If the location and speed
Step 2: Put the number in the cross region of f2 to
exceed boundary range in iteration, given boundary
the front of f1, delete the number of the cross region of
value. Based on analysis of the above parameters,
f2 in f1, get a sub-string new1. Put the number in the
Maurice Clerc provides the parameter conditioning of
cross region of f1 to the front of f2, delete the number
PSO algorithm convergence [9].
of the cross region of f1 in f2, get a sub-string new2;
Step 3: Judge whether the sub-string is a feasible
4. A hybrid particle swarm optimization solution, if not, redo Step 2, otherwise go to Step 4;
algorithm to solve VRPTW Step 4: Calculate the fitness by Eq. (1), retain the
offspring of the lower fitness recorded as new.
In the basic PSO algorithm, the speed of particles v For example, given two parent strings:
is restricted by a maximum speed vmax. It decides f1=1 2 3 4 5 6 7 8 9, f2=9 8 7 6 5 4 3 2 1,
searching precision in the solution space. If vmax is too and the cross region (3, 6), after crossing, two new
big, particles may miss the optimal solution; otherwise, strings are:
the particles may get into a local search space and can new1=7 6 5 4 1 2 3 8 9, new2=3 4 5 6 9 8 7 2 1.
not carry through a globe search [10]. Therefore, in this Calculate the fitness for new1 and new2 by Eq. (1).
paper, the idea of crossover operation of genetic Lastly, crossover operation of two parent strings is
algorithm is brought into PSO algorithm. Meanwhile, completed. Obviously, owing to crossover operation,
because PSO has the problem of low-speed the number of particles become twice as many as that
convergence, level set theory is also employed for of the original one, namely 2N new particles are
solving this deficiency. In a word, this paper applies obtained.
both technique of GA and level set theory to improve
standard PSO algorithm, thus proposing a combined 4.3. Selection operation
PSO algorithm that is suitable to VRPTW. We know that in PSO algorithm, each particle
moves towards the global optimal location and the
4.1. Particle coding
optimal location of individual history as a criteria to
This paper makes reference to the encoding manner find a better location for survival. However, this
of literature [11]. For the VRPTW problem with L movement in PSO makes some inferior particles
customer points and K vehicles, each particle has continue to reproduce to become inferior communities
L+K+1 dimensional vector, in which the order of each unable to be eliminated which affects the algorithm
element value denotes the delivery order of the convergence rate. At the same time the resource of
customer point in the total path. particle swam can not be fully utilized. So in selection
For example: Assume the number of customers is 8 operation, we need to judge which particles will be
and the number of vehicles is 3. If the location vector eliminated and which will be chosed to next generation.
of the particle is: This requires that all particles should be divided into
Warehouse No: 1 2 3 4 5 6 7 8 0 0 two parts: the superior ones and the inferior ones. Level
X=[6 8 4 1 2 9 5 10 3 7] set theory is introduced here.
Then the total path of corresponding solution of the For the tth-generation P(t) = (P1, P2, ..., Pn), n denots
particle is: 0ĺ4ĺ5ĺ0ĺ3ĺ7ĺ1ĺ0ĺ2ĺ6ĺ8ĺ0, the number of particles, the fitness function of particles
corresponding separate path is: n
f (Xi )
Vehicle 1: 0ė4ė5ė0; is set to fi(x), order ft = ¦ , where t denotes
Vehicle 2: 0ė3ė7ė1ė0; i =1 n
tth-generation.

504
500
 We have H f = { xi ∈ P (t ) | f ( xi ) ≤ ft ,1, 2, " , n}
t
qi 2 1.5 4.5 3 1.5 4 2.5 3

H f is called the level set about f relative to P(t). After ti 1 2 1 3 2 2.5 3 0.8
t

that the population of each generation can be divided [ETi, LTi] [1,4] [4,6] [1,2] [4,7] [3,5.5] [2,5] [5,8] [1.5,4]

two parts: H f and P (t ) − H f . [8] In the process of

t t TABLE II. DISTANCE FROM POINT TO POINT
Crossover operation, 2N new particles were obtained, dij 0 1 2 3 4 5 6 7 8
while only N of them can be allowed to enter next 0 0 40 60 75 90 200 100 160 80
generation. Thus, we sort the 2N particles on their
1 40 0 65 40 100 50 75 110 100
fitness measured by Eq. (1) in descending order, and
the top N particles would be selected to enter the next 2 60 65 0 75 100 100 75 75 75
generation. 3 75 40 75 0 100 50 90 90 150
4 90 100 100 100 0 100 75 75 100
4.4. Algorithm implementation 5 200 50 100 50 100 0 70 90 75
In this paper, we use the hybrid PSO to solve 6 100 75 75 90 75 70 0 70 100
VRPTW, whose pseudocode of algorithm is described
7 160 110 75 90 75 90 70 0 100
in Fig 1:
8 80 100 75 150 100 75 100 100 0

Set iteration times for MaxN, randomly generated

initial particles N; The best known result is: vehicle 1:
Calculate the fitness of initial particle to be l0. 0ĺ3ĺ1ĺ2ĺ0; vehicle 2: 0ĺ8ĺ5ĺ7ĺ0; vehicle 3:
According to the initial fitness of each particle, initialize 0ĺ6ĺ4ĺ0. The total travel cost is Z=910.
plbesti, pxbesti, glbest and gxbest; Set parameters of the Hybrid-PSO as follows: the
WHILE (iteration times< MaxN) DO number of particles N=40, the iteration time
FOR i=1: MaxN
The ith particle Xi crosses with gxbest and
MaxN=200.Solving this problem uses GA in literature
selection to be Xi’; Xi’ crosses with pxbesti and [8] (N=40, MaxN=200, Pc=0.6, Pe=0.2 roulette
selection to be Xi’’, ordering Xi=Xi’’; manner choosing offspring); uses PSO in literature
Calculate fitness li according to current location; [13] (N=40, divided into two subgroups, subgroup
IF (li< plbesti) size=22, overlapping particle swarm between
pxbesti=Xi, plbesti=li; subgroups is 2, w=0.729, c1=c2=1.49445, MaxN=200);
END IF uses parallel PSO in literature [11] (divided into two
END FOR
subgroups, subgroup size is N1=N2=20, w=0.729,
Update plbesti, pxbesti, glbest, gxbest;
END WHILE c1=c2=1.49445, MaxN=200). These above three
Finally, export glbest and gxbest. methods and Hybrid-PSO algorithm use Matlab7.0 to
program respectively, each with the operator of 50
Figure. 1. Pseudocode of hybrid PSO algorithm to solve VRPTW times. The comparison results are shown in Tab III:

TABLE III COMPARED RESULTS IN 4 ALGORITHMS FOR SOLVING 50

5. Calculation example TIMES

Searching Average
In this paper, we use the example in literature [8]. Average
successful rate searching time
There are eight cargo transportation tasks (No. 1, travel cost
(%) (s)
2,…,8). The cargo qi of each task, loading or unloading
GA[8] 993.6 24 11
time ti, executing time range [ETi, LTi] are showed in
Tab I, distance dij from yard 0 to each task point in Tab PSO [13]
940.5 46 6
II. These tasks are completed by eight tons of vehicles,
Parallel PSO[11] 923.8 72 4
in which the speed is 50, unit transportation cost is 1,
over time window unit cost for punishment is: Hybrid-PSO 914 97 4
PE=PL=50.
Experimental results show that the searching
TABLE I. CHARACTERS AND DEMAND OF TASK
successful rate 97% of the proposed Hybrid-PSO in
Task point 1 2 3 4 5 6 7 8 solving VRPTW problem is far greater than the basic

505
501
PSO 46%, the basic GA 24% and 72% of parallel PSO. [5] DING Hai-Li, WANG Fang, GAO Cheng-xiu.
At the same time in the aspect of the average result and “Crossover particle swarm optimization for traveling
evaluating successful searching time, the proposed salesman problem”. Journal of Mathematics, 2008(8): pp.
85-89.
method compared with other three methods is optimal.
[6] Niu, B., Zhu, Y.L. and He, X.X., et al. An Improved
6. Conclusion Particle Swarm Optimization Based on Bacterial Chemotaxis,
In Proc. IEEE Congress on Intelligent Control, Dalian,
This paper presents a HPSO algorithm, which CHINA, 2006, pp.3193-3197.
combines the PSO algorithm with genetic algorithm
[7] Jiang Wei-gang, Zhang Yuan-biao, Xie Jian-wen. “A
and level set. Through an example about window time Particle Swarm Optimization Algorithm Based on Diffusion-
VRP validated, the algorithm which has the characters Repulsion and Application to Portfolio Selection,” 2008
of less number of parameters, simple implementation, International Symposium on Information Science and
and relatively shorter time of average search, is proved Engineering, 2008(2): pp. 498-501
as an effective algorithm for solving discrete
[8] LI Jun,GUO Yao-huang. “Dispatch optimizes theory and
optimization problems.
methods for logistic distribution vehicle”. Beijing: Chinese
Commodity Publish House, 2001.
7. Acknowledgment
[9] Maurice Clerc, James Kennedy. “The particle swarm-
The authors thank the anonymous reviewers for explosion, stability, and convergence in a multidimensional
their valuable comments and suggestions that helped complex space”. IEEE Transaction on Evolutionary
us for improving our work. Computation, 2002, 6(1), pp. 58-73.

[10] ZHANG Li-yan, PANG Xiao-hong, XIA Wei-jun, et al.

8. References “A hybrid particle swarm optimization algorithm for vehicle
routing problem with time windows”. Journal of Shanghai
[1] Dantzig G, Ramser J. “The truck dispatching problem”. Jiaotong University, 2006(11), pp. 1890-1895.
Management Science, 1959(6), pp. 80-91.
[11] WU Yong, YE Chun-ming, MA Hui-min, et al. “Parallel
[2] Olli Braysy. “Genetic algorithms for the vehicle routing particle swarm optimization algorithm for vehicle routing
problem with time windows”. Special issue on problems with time windows”. Computer Engineering and
Bioinformatics and Genetic Algorithms, 2001(1), pp. 33–38. Applications, 2007, 43(14), pp. 223-226.

[3] Hoong C L, Melvyn S, Kwong M T. “Vehicle routing [12] ZHANG Li-ping, CHAI Yue-ting, CAO Rui. “Improved
problem with time windows and a limited number of genetic algorithm for vehicle routing problem with time
vehicles”. European Journal of Operational Research, 2003, windows”. Computer Integrated Manufacturing Systems,
148: pp.559-569. 2002, 8(6), pp. 451-454.

[4] R C Eberhart, J Kennedy. “A New Optimizer Using [13] LI Ning, ZOU Tong, SUN De-bao. “Particle swarm
Particles Swarm Theory”. In: Proc Sixth International optimization for vehicle routing problem with time windows”.
Symposium on Micro Machine and Human Science, Nagoya, Systems Engineering-Theory & Practices, 2004(4), pp. 130-
Japan, 1995. 135.

506
502