WCCI 2012 IEEE World Congress on Computational Intelligence

June, 10-15, 2012 - Brisbane, Australia IEEE CEC

A Multiobjective Evolutionary Algorithm with

Enhanced Reproduction Operators for the Vehicle
Routing Problem with Time Windows

Wei-Huai Hsu Tsung-Che Chiang

Department of Computer Science and Information Department of Computer Science and Information
Engineering, National Taiwan Normal University Engineering, National Taiwan Normal University
Taipei, Taiwan, R.O.C. Taipei, Taiwan, R.O.C.
[email protected] [email protected]

Abstract—This paper addresses the vehicle routing problem with In the VRPTW, we are given N customers and each
time windows (VRPTW). The task is to assign customers to customer i is associated with demand qi, service time si, and
multiple vehicles and determine the visiting sequences of service time window [ei, li], for i = 1 … N. There is a central
customers for the vehicles without violating the vehicle capacity depot, denoted by location 0. Each vehicle starts from the depot,
constraint and customer service time window constraints. Two serves customers, and finally comes back to the depot. The
common objectives of VRPTW are to minimize the number of total demand of customers assigned to each vehicle cannot
vehicles and the total traveling distance. Most of previous studies exceed the maximum capacity, Q. Each customer must be
assumed that the number of vehicles is more important than the served exactly once by one vehicle. The distance between
total distance. Hence, they solved the VRPTW by minimizing the
number of vehicles first and then minimizing the total distance
locations i and j (0 ≤ i, j ≤ N) is dij, and the travel time is tij. The
under the minimal number of vehicles. Recently, researchers service time window constraint says that the vehicle must
started to solve the VRPTW without this assumption and tried to arrive at the location of customer i no later than the latest
minimize both objectives simultaneously through searching for service time, li. The vehicle is allowed to arrive before the
the Pareto optimal set of solutions. Following this perspective, we earliest service time ei, in which case it must wait until ei and
use a multiobjective evolutionary algorithm to solve the VRPTW. then start to serve the customer.
We propose enhanced crossover and mutation operators by
Two common objectives of the VRPTW are to minimize
incorporating the domain knowledge. Performance of the
the number of required vehicles and the total distance. While
proposed algorithm is verified on a widely used benchmark
problem set. Comparing with seven existing algorithms, our
some studies focused on minimizing only the total distance
algorithm shows competitive performance and contributes many [6]−[8], most literature considered both objectives. A classical
new best known Pareto optimal solutions. way to address these two objectives is to minimize the number
of vehicles first and then minimize the total distance [9]−[18].
Keywords- vehicle routing problem; multiobjective optimization; Several researchers [2][13][15] have noticed that sometimes
evolutionary algorithm; crossover; mutation; time windows there is conflict between the two objectives. In other words, the
reduction of the number of vehicles could increase the total
I. INTRODUCTION distance traveled, which means that there does not exist a
unique optimal solution that minimizes both objectives
The vehicle routing problem with time windows (VRPTW) simultaneously. When the number of vehicles is not absolutely
is an important research topic in the fields of transportation more important than the total distance, the decision makers
science and operations research. It has a lot of applications in need to find a set of optimal solutions by considering the trade-
the real world. Typical examples include bank deliveries, postal off between the two objectives. In this paper, we follow the
deliveries, school bus routing, repairmen scheduling, and so on. concept of Pareto dominance to define the set of optimal
The typical vehicle routing problem (VRP) is a generalization solutions. A solution x is said to dominate a solution y iff x is
of the traveling salesman problem (TSP) and belongs to the not worse than y in all objectives and x is better than y in at
class of NP-hard combinatorial optimization problems [1]. By least one objective. A solution is called Pareto optimal iff it is
adding the constraints of customer service time windows, the not dominated by any other solution. The set of Pareto optimal
VRPTW becomes even harder. The task is to assign customers solutions is called the Pareto set, and the set of their objective
to multiple vehicles and determine the visiting sequences of vectors is called the Pareto front. Although there are more and
customers for the vehicles without violating the capacity more studies on solving the VRPTW in the multiobjective or
constraint and time window constraints so that certain Pareto way, this domain is still young [19]. In this paper, we
objective(s) is/are optimized. Due to the high practical value will propose a multiobjective evolutionary algorithm (MOEA)
and challenging problem complexity, the VRPTW has been with enhanced crossover and mutation operators to tackle the
intensively investigated in the last two decades [2]−[5]. multiobjective VRPTW (MOVRPTW).

U.S. Government work not protected by U.S. copyright

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
 The rest of this paper is organized as follows. In Section II, through linear weighted summation. The weight values of
we review the literature on classical VRPTW and MOVRPTW. constraints were adjusted adaptively. If the best non-tabu
The proposed algorithm is detailed in Section III. Section IV solution is feasible, the weight of constraint violation is divided
presents the experiments and results. Conclusions and future by a constant (greater than one); otherwise, the weight is
work are given in Section V. multiplied by the constant. In other words, the weight of
constraint violation decreases when the search process finds
II. LITERATURE REVIEW feasible solutions easily. Berger et al. [17] proposed a GA with
two populations. One population was responsible for
The study of Solomon [9] is a seminal work on the minimizing the total distance, and the other focused on
VRPTW. He proposed several heuristic methods, including the minimizing the number of vehicles through minimization of the
savings heuristic, time-oriented nearest-neighbor heuristic, and constraint violation. The constraint violation was measured by
insertion heuristics. These heuristics select unrouted customers the linear weighted summation of the amount of time window
and insert them into proper positions in the currently building violation and the number of unsatisfied customers. Nagata et al.
route according to some criteria regarding the travel distance, [18] proposed a memetic algorithm (MA) with a new crossover
time delay, and so on. He also developed a set of benchmark operator. The offspring produced by the proposed crossover
instances, which consists of six problem categories and 56 can be infeasible, and they developed a repair procedure based
problem instances. The type-1 insertion heuristic (I1) and the on the hill climbing algorithm. The repair procedure tried to
set of benchmark instances were widely adopted in later minimize the linear weighted summation of violation of
research works. The research group of Potvin demonstrated the capacity and time window constraints.
possibility of solving the VRPTW by metaheuristics. Potvin et
al. [10] proposed a tabu search (TS), whose features include the In contrast to the classical VRPTW, the MOVRPTW
reduction of neighborhood size, the alternation between the 2- considers the trade-off between the objectives and aims at
opt* and Or-opt neighborhoods, and the introduction of a route finding the set of Pareto optimal solutions, instead of a single
saving phase. Potvin and Bengio [11] used a genetic algorithm (lexicographically) optimal solution. Hong and Park [20]
(GA) and proposed two crossover and three mutation operators. addressed the MOVRPTW with total travel time and total
waiting time as the objective functions. They did clustering of
Since the goal of the classical VRPTW is to minimize the customers first based on the linear weighted summation of the
number of vehicles and then the total distance, several two objective values. Then, they did routing by the goal
researchers developed their approaches also in a hierarchical programming method. They concluded that the proposed
manner. Bräysy [12] proposed a four-phase approach. An approach could generate non-dominated solutions by changing
initial solution was created in the first phase, followed by a the weight values. Geiger [21] proposed a MOEA, which
route-elimination procedure. The third and four phases assigned the fitness values to individuals by the number of
improved the solution in terms of the total distance using four dominating individuals. Barán and Schaerer [22] developed a
local search procedures. Bent and Hentenryck [13] presented a multiobjective ant colony system, regarding the number of
two-stage hybrid local search algorithm, in which the number vehicles, the total traveling time, and the total delivery time.
of vehicles was minimized by simulated annealing (SA) in the They maintained an archive of found non-dominated solutions,
first phase, and the total distance was minimized by a large and only these solutions could deposit pheromone. Jozefowiez
neighborhood search in the second phase. They emphasized et al. [23][24] solved the MOVRP by parallel NSGA [25] and
that minimization of the total distance may not be effective in NSGA-II [26]. These approaches are featured by the new
minimization of the number of vehicles. Therefore, in the first dominance operator and additional archives for diversification.
phase they compared solutions by a lexicographical ordering of
the number of vehicles, the sum of square of number of Tan et al. [27][28] proposed a hybrid MOEA to solve the
customers in the routes, and the minimal delay [14]. MOVRPTW. The fitness of an individual was calculated by the
Homberger and Gehring [15] also proposed a two-phase hybrid number of dominating individuals. They proposed a route-
metaheuristic. The structure of their algorithm is similar to that exchange crossover and used three mutation operators. One of
of Bent and Hentenryck’s one. They used an evolution strategy three local search heuristics was chosen randomly to improve
(ES) to minimize the number of vehicles in the first phase and a all individuals every 50 generations. The experimental results
TS to minimize the total distance in the second one. To showed that all problem categories except C1 and C2 in
minimize the number of vehicles, the ES was driven by the Solomon’s data set contain several problem instances with
lexicographical evaluation of solutions based on four criteria: conflicting objectives. Ombuki et al. [29], Ghoseiri and
the number of vehicles, the number of customers in the Ghannadpour [30], and Garcia-Najera and Bullinaria [31]
smallest route, the minimal delay, and the total distance. solved the same problem by the MOEA. They all adopted the
Pareto ranking scheme [32] as the fitness assignment
Most of the metaheuristic approaches in the literature mechanism. Garcia-Najera and Bullinaria proposed a similarity
maintained the feasibility of solutions during the search process. measure. During mating selection, one parent was chosen by
In other words, only feasible solutions (satisfying the capacity the rank, and the other was chosen by the similarity. After the
and time window constraints) are produced by the offspring were produced by crossover and mutation, the best
neighborhood functions. Some researchers allowed infeasible individuals with lower ranks survived. Ties were broken by
solutions to join the search process and introduced the keeping the individuals with smaller similarity survive. Their
constraint violation in the evaluation function. Cordeau et al. experimental results showed that 34 of the 56 Solomon’s
[16] proposed a TS to minimize the total travel time. They problem instances have conflicting objectives.
aggregated the original objective and constraint violation

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
 III. PROPOSED MULTIOBJECTIVE EVOLUTIONARY parents. The best route refers to the route with the smallest ratio
ALGORITHM WITH ENHANCED OPERATORS (MOEA-EO) of the total distance to the number of customers. After
exchanging the best routes, the duplicate customers in the
In this section we detail the proposed MOEA-EO. The flow
original routes are removed. The advantage of this crossover is
of MOEA-EO is given here, and the details of each step are
that it maintains the feasibility of solutions naturally since the
described in the following subsections.
removal of customers will not violate the capacity and time
1. Initialization: Generate the initial population of size NPOP. window constraints. Fig. 2 shows an example of how the route
Decode the solutions, calculate the objective values, and exchange crossover works. Offspring 1 receives the best route
assign the fitness values. Set the generation number t = 1. (0, 2, 9, 6, 0) from parent 2 as the fourth route. Then, the
customers 2, 9, and 6 are removed from the first, third, and
2. Reproduction: Repeat the following sub-steps for NPOP/2
second routes, respectively. Offspring 2 is generated in a
times to generate NPOP offspring. similar way.
2.1 Select two parents through mating selection.
2.2 Produce two offspring by the enhanced crossover.
Parent 1 Parent 2
2.3 Apply the enhanced mutation in probability pm.
3. Environmental selection: Select the best NPOP individuals
1 5 7 1 3 2
among the 2⋅NPOP individuals to survive to generation t+1.
4. Termination: If t > TGEN, stop; otherwise, t = t +1, go to
2 6 8 5 8 9
Step 2.
3 9 4 6
A. Chromosome Representation
Our chromosome representation follows that proposed by
Tan et al. [28]. It is a list of routes, where each route is a 4 7
sequence of customer indices. Fig. 1 shows a chromosome that
encodes three routes. Since the routes are directly encoded, we
can calculate the number of vehicles and total distance Offspring 1 Offspring 2
efficiently.
9 1 5 7 2 1
8 Chromosome

8 9 5 8 9 2
7
1 1 5 7
2 3 6 6 3
0
2 6 8
6
3 4 7 4
4 3 9
Figure 2. Original route exchange crossover
5
4 Although the route exchange crossover needs no repair
procedure and helps to achieve short total distance, we notice
Figure 1. Chromosome representation that the offspring usually needs one more vehicle than the
parents do. Since minimization of the number of vehicles is one
B. Fitness Assignment and Selection of our objectives, we add a route elimination procedure in the
crossover. We try to eliminate the worst route, i.e. the route
We assign the fitness values by the Pareto ranking scheme. with the highest ratio of the total distance to the number of
The non-dominated solutions in the population are assigned customers. Customers on this route are re-inserted into other
rank one. By temporarily disregarding the solutions with rank routes one by one in the order they appear. The best position to
values smaller than or equal to r, the non-dominated solutions insert a customer k is determined based on the two criteria in
in the remaining population are assigned rank (r+1). Mating [14]. Let bk denote the service start time of customer k, the cost
selection uses 2-tournament, where two individuals are to insert k between i and j is defined by
randomly picked up and the one with a smaller rank is selected
as a parent. If two individuals have the same rank, one is Cost(i, k, j) = PTk(i, j) + PDk(i, j),
selected randomly. Environmental selection adopts the (μ+λ)- where
strategy, where μ = λ.
⎧ max{bk + s k + d kj , e j } − b j ,j≠0
PTk (i, j ) = ⎨
C. Enhanced Route Exchange Crossover b
⎩ k + s k + d k0 − (bi + s i + d i0 ) j=0
After two parents are selected through mating selection, we and
produce two offspring by crossover. Tan et al. [28] proposed a
route-exchange crossover by exchanging the best routes of the PD k (i, j ) = d ik + d kj − d ij . (1)

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
PTk(i, j) calculates the delay of the service start time of the heuristic [9]. It starts a new route by selecting the customer
customer j after insertion of customer k, and PDk(i, j) calculates with the closest location to the depot. Then, from the unrouted
the extra travel distance caused by the insertion of k. Only customers, the customer with the smallest cij is added at the end
feasible positions are considered, and the customer is re- of the route. Let i denote the last customer in the currently
inserted into the position with the least cost. Fig. 3 is an constructing route, the cost to add j after i is defined by
illustration of re-inserting the customer on the second route to
eliminate one route for the offspring. cij = δ1dij + δ2(ej – (bi + si)) + δ3(lj – (bi + si + tij)), (2)
which is a linear weighted summation of time delay and
Offspring 1 Offspring 1 distance. (Recall that bi denotes the service start time of
customer i.) Here, we set equal weights by δ1 = δ2 = δ3 = 1/3. If
no unrouted customer can be added without violating the
1 5 7 2 1 5 2
constraints, we start a new route.

8 9
The other heuristic solution is generated by the I1 insertion
3 3 7 9
[9]. The unrouted customer k with the largest c2(i, k, j) is
inserted between (i, j) iteratively. Again, a new route starts
4 6 4 8 6 when no unrouted customer can be added without violating the
constraints.
Figure 3. Re-insertion in the enhanced route exchange crossover
c2(i, k, j) = λd0k − α1(dik + dkj − μdij) − α2(bj′ − bj) (3)
Another enhancement we do is about the selection of the
The symbol bj′ denotes the new service start time of customer j
best route for exchange. Toward the end of the evolutionary
process, the best routes of the individuals in the population may after customer k is added between i and j. Here, we set λ = μ =
be identical. In this case, exchange of this route is useless. Thus, 1 and α1 = α2 = 0.5
in our enhanced version, we select one of the best three routes As for the remaining (NPOP – 2) initial individuals, we add
randomly for exchange. the customers one by one at the end of the current route in a
random order. If the next customer cannot be added, start a new
D. Enhanced Mutation route.
The proposed mutation is based on the re-insertion of
customers. First we randomly select Nm customers and IV. EXPERIMENTS AND RESULTS
temporarily remove them from the routes. Since it should be
more difficult to find a feasible position to insert a customer A. Benchmark Instances
with smaller time window, we order the removed customers in In the experiments we tested the proposed MOEA-EO on
the non-decreasing order of the length of time window (li – ei). Solomon’s data set [9]. The data set contains 56 problem
Then, we re-insert the customers in this order with the same instances, which fall into six categories, R1, R2, C1, C1, RC1,
method we mentioned in the previous sub-section. Fig. 4 gives and RC2. In the R categories the locations of customers are
an illustration. We randomly remove three customers, 2, 5, and generated uniformly, and in the C categories the locations of
8. Based on their time window length, we re-insert customers 5, customers are clustered. The RC categories are a mixture. The
8, and 2 one after another. R1, C1, and RC1 categories have a shorter scheduling horizon
and require more vehicles, whereas the other categories have a
Chromosome Chromosome longer scheduling horizon and need fewer vehicles. The
l i – ei
number of customers is 100. This data set is widely used in the
literature on VRPTW.
Insertion order

1 5 7 5 50 1 2 5

8 60
B. Benchmark Algorithms
2 6 8 3 6 7
We compare the results of our algorithm with seven
2 85 benchmark algorithms. Lim and Zhang’s approach [33] and
3 9 8 9 Nagata et al.’ approach [18] belong to the lexicographical
approach, which minimizes the number of vehicles and then
the total distance. Ting and Huang’s approach [7] and
4 4
Alvarenga et al.’ approach [8] focused only on minimizing the
total distance. Labadi et al.’ approach [34] solved the VRPTW
Figure 4. Illustration of the enhanced mutation by the lexicographical way and by minimizing the total
distance as well. Ombuki et al.’ approach [29] and Garcia-
Najera and Bullinaria’s approach [31] are MOEAs, which
E. Initial Population aimed at finding the Pareto set of solutions.
We generate two initial solutions heuristically, and the
remaining population is generated randomly. The first heuristic
solution is generated by the time-oriented, nearest-neighbor

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. NET SET OF NON-DOMINATED SOLUTIONS FOUND BY SEVEN BENCHMARK ALGORITHMS AND PROPOSED MOEA-EO

Problem Number of Total Algorithms that found the Problem Number of Total Algorithms that found the
Instance routes Distance solution Instance routes Distance solution
R101 19 1650.80 [N][L][G] R204 2 825.52 [N]
. 20 1642.87 [A] . 3 749.417 [M]
R102 17 1486.12 [N][L] . 4 743.851 [M]
. 18 1472.62 [A] 5 735.861 [B]
R103 13 1292.68 [N][L] R205 3 994.43 [N][L]
. 14 1213.62 [A] . 4 959.74 [G]
R104 9 1007.31 [N][L] 5 954.16 [O]
. 10 992.933 [M] R206 3 906.14 [N][L]
. 11 986.096 [A] . 4 889.39 [O]
R105 14 1377.11 [N][L][G] 5 879.893 [B]
. 15 1360.78 [A][B] R207 2 890.61 [N]
R106 12 1252.03 [N][L] . 3 817.854 [M]
. 13 1240.41 [G] . 4 797.994 [M]
R107 10 1104.66 [N][L] R208 2 726.82 [N][L]
. 11 1074.24 [B] . 3 706.855 [B]
R108 9 960.88 [N][L] R209 3 909.16 [N][L]
. 10 946.422 [M] . 4 863.813 [M]
R109 11 1194.73 [N][L] 5 859.39 [B]
. 12 1154.31 [M] R210 3 939.37 [N][L]
. 13 1151.84 [A] . 4 929.023 [M]
R110 10 1118.84 [N][L] . 5 918.778 [M]
. 11 1097.75 [M] 6 912.533 [B]
. 12 1072.42 [B] R211 2 885.71 [N]
R111 10 1096.73 [N][L] . 3 784.56 [M]
. 11 1062.2 [M] . 4 755.949 [B]
. 12 1053.5 [A] RC201 4 1406.94 [N][L]
R112 9 982.14 [N] . 5 1311.52 [M]
. 10 960.675 [A] . 6 1298.78 [M]
RC101 14 1696.95 [N][L] . 7 1282.94 [M]
. 15 1636.92 [O] . 9 1274.54 [A]
RC102 12 1554.75 [N][L] 10 1273.04 [B]
. 13 1492.70 [G] RC202 3 1365.65 [N][L]
. 14 1466.44 [M] . 4 1161.29 [M]
RC103 11 1261.67 [N][L] . 5 1118.66 [M]
RC104 10 1135.48 [N][L][M] . 7 1102.06 [M]
RC105 13 1629.44 [N][L] 8 1099.54 [B]
. 14 1540.18 [G] RC203 3 1049.62 [N]
. 15 1519.44 [G] . 4 945.083 [M]
. 16 1518.60 [A] . 5 940.851 [M]
RC106 11 1424.73 [N][L] 6 937.449 [B]
. 12 1394.43 [G] RC204 3 798.46 [N][L]
. 13 1377.35 [A] . 4 791.399 [B]
RC107 11 1230.48 [N][L] RC205 4 1297.65 [N][L]
. 12 1212.83 [A] . 5 1238.7 [M]
RC108 10 1139.82 [N][L] . 6 1185.89 [M]
. 11 1117.53 [A] . 7 1161.81 [A]
R201 4 1252.37 [N][L] RC206 3 1146.32 [N][L]
. 5 1190.02 [M] . 4 1085 [M]
. 6 1171.83 [M] . 5 1080.50 [O]
. 7 1158.48 [M] . 6 1067.81 [M]
8 1150.92 [B] . 7 1054.61 [B]
. 9 1148.48 [A] RC207 3 1061.14 [N][L]
R202 3 1191.70 [N][L] . 4 1001.85 [G]
. 4 1081.60 [G] . 5 970.78 [M]
. 5 1041.10 [M] 6 966.372 [B]
7 1037.50 [B] RC208 3 828.14 [N]
R203 3 939.50 [N][L] . 4 785.93 [O]
. 4 900.794 [M] 5 783.932 [B]
. 5 890.50 [O] [L]: Lim and Zhang [33]: Table 3 [N]: Nagata et al. [18]: Table 7
6 874.869 [B] [A]: Alvarenga et al. [8]: Table 5 [O]: Ombuki et al. [29]: Table 1 and 2
[G]: Garcia-Najera and Bullinaria [31]: Table 7
[B]: Labadi et al. [34] Table 3 and 4
[M]: MOEA-EO proposed in this study

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
 Table I can serve as a benchmark for interested researchers
C. Parameter Setting to do performance comparison conveniently.
We set the population size and generation number by 100
and 500, respectively, using the same setting as that of the GA
in [31]. The number of evaluations (100×500) is smaller than E. Single-Objective Performance Comparison
that of the MA in [18] (200×20×100), that of the GA in [7] In Table II, we summarize the average number of routes
(100×1200), and that of the GA in [29] (300×350). We tested and average total distance over the best solutions of instances
nine combinations of the other two parameters by three values in each of the R1, R2, RC1, and RC2 categories for the eight
of the mutation rate rm, {0.4, 0.6, 0.8} and three values of the tested algorithms. The algorithms in [29], [31], and our
number of mutated customers Nm, {20, 30, 40}. We run the MOEA-EO solved the VRPTW in the Pareto manner and
nine variants on R101, R201, RC101, and RC201 for ten times. usually generated multiple non-dominated solutions for one
Based on their best-case solution quality and the consumed problem instance. We considered the solution with the minimal
computation time, we chose the setting of 0.6 and 30. Since the number of routes and the one with the minimal total distance.
instances in C1 and C2 categories are simple and all algorithms The algorithm in [34] also generated two solutions for each
perform well on them, these instances were not considered in problem instance. The statistics calculated based on the
the experiments. We solved the remaining 39 problem solutions with the minimal number of routes are denoted by (R),
instances for ten times and collected the non-dominated and the statistics calculated based on the solutions with the
solutions to compare with the benchmark algorithms. minimal distance are denoted by (D). Due to the lack of data in
[31], we only provide the statistics based on the solutions with
the minimal distance.
D. Multiobjective Performance Comparison
We collected the best solutions for all tested problem TABLE II. AVERAGE NUMBER OF VEHICLES AND AVERAGE TOTAL
instances obtained by all seven benchmark algorithms from DISTANCE OF SEVEN BENCHMARK ALGORITHMS AND PROPOSED MOEA-EO
their original papers. From all these solutions (including our
best solutions), we generated the net set of non-dominated R1 R2 RC1 RC2
solutions for each instance and listed the results in Table I. 11.92 2.73 11.50 3.25
Lim and Zhang [33]
From Table I, the first observation is that almost all 1210.76 953.94 1384.17 1120.40
instances have conflicting objectives. The only two exceptions 11.92 2.73 11.50 3.25
are instances RC103 and RC104. Comparing with the results in Nagata et al. [18]
1210.34 951.03 1384.17 1119.24
Fig. 10 and Table 7 in [31], two more instances, R108 and
Labadi et al. 12.75 3.09 12.37 3.62
R207 are shown to have conflicting objectives by the solutions
obtained by our MOEA-EO. [34] (R) 1188.01 920.86 1351.27 1087.18
13.58 4.91 13.50 5.38
There are 118 non-dominated solutions in total. The Ting and Huang [7]
benchmark algorithms in [18], [33], [34], [8], [31], and [29] 1222.65 928.32 1415.16 1089.78
found 39, 33, 19, 15, 10, and 6 solutions, respectively. The 13.25 5.55 12.88 6.50
algorithms in [18] and [33] took minimization of the number of Alvarenga et al. [8]
1183.4 899.9 1341.7 1015.90
vehicles as the primary objective and found the solution with
Labadi et al. 13.42 5.36 13.13 6.75
the minimal number of vehicles for almost all instances. The
algorithms in [34] and [8] minimized the total distance, and [34] (D) 1184.16 879.51 1352.02 1009.37
each of them found the solution with the minimal total distance Ombuki et al. 12.50 3.18 12.13 3.38
for about half of the instances. Our MOEA-EO found 33 non- [29] (R) 1220.91 933.14 1386.35 1125.44
dominated solutions. This is the best performance among the
three algorithms that considered both objectives simultaneously. Ombuki et al. 13.17 4.45 12.75 5.63
In addition, 32 of the 33 solutions were found only by MOEA- [29] (D) 1203.22 892.89 1370.84 1025.31
EO. We mark these solutions in bold in Table I. Garcia-Najera and 13.08 4.00 12.63 5.38
Our algorithm found the minimal total distance exclusively Bullinaria [31] (D) 1187.32 897.95 1348.22 1036.65
for 3 instances. In other words, it updates the best known 12.67 3 12.38 3.38
minimal total distance for these instances. It also provides extra MOEA-EO (R)
1193.57 944.25 1355.54 1113.11
non-dominated solutions for 20 instances. The ability of
MOEA-EO for generating compromise solutions between the 13.17 4.45 12.63 5.63
MOEA-EO (D)
extreme solutions (with the minimal number of vehicles or with 1186.67 889.07 1351.11 1015.89
the minimal total distance) provides the information on the
trade-off relationship between the two objectives to make a
good decision. In the instance R203, for example, the two In Table II, we mark the best average objective values in
extreme solutions obtained by [18] and [34] say that the total bold. For minimizing the number of routes, the two algorithms
distance can be reduced by 10.7% (1049.62 to 937.449) by in [18] and [33] achieve the best average-case performance in
doubling the number of vehicles (3 to 6). Our MOEA-EO gives all four categories. For minimizing the total distance, the
an alternative that reduces the total distance by 10% (1049.62 algorithm in [8] and [34] are the best performers in two
to 945.093) by just one more vehicle. categories, respectively. Our MOEA-EO has the performance

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
between the extreme points. Considering both objectives, V. CONCLUSION AND FUTURE WORK
MOEA-EO dominates 2, 3, 3, and 2 algorithm variants in the The VRPTW is a problem of high complexity and practical
four categories, respectively. In category R1, for example, the value. Two common objectives are to minimize the number of
average objective values of MOEA-EO (D) are 13.17 and vehicles and total distance. In this study, we proposed an
1186.67. The values of [7] are 13.58 and 1222.65. Our MOEA to minimize both objectives simultaneously by
algorithm improves the solution quality in both objectives searching for the Pareto optimal set of solutions. We did
simultaneously. improvements to existing crossover and mutation operators.
The new crossover operator incorporates a route-elimination
F. Effect of Enhanced Crossover and Mutation post-processing and adds more randomness to increase
offspring diversity. The new mutation operator uses domain
To examine the effect of proposed crossover and mutation knowledge to guide the re-insertion order of customers.
operators, we conducted an experiment with four variants of Through comparing with seven benchmark algorithms using 39
our algorithm using different combinations of crossover and problem instances, our algorithm shows competitive
mutation. The average number of routes and average total performance and provides 32 new best known non-dominated
distance of the solutions with the minimal distance in the solutions. It also provides the minimal distance for 3 instances
obtained set of solutions are calculated and summarized in exclusively.
Table III. The proposed crossover and mutation are denoted by
E-RCX and E-IM. The crossover RCX does not have the route- We will continue this research in three directions. First, our
elimination post-process and selects the best route algorithm is not good enough at minimizing the number of
deterministically. The mutation IM removes and re-inserts one vehicles. A possible improvement is to consider auxiliary
customer at a time in a random order. In Table III, we can see objectives like the minimal delay in [13] and [15]. Second,
that the combination of proposed crossover and mutation most studies tested the proposed algorithms by Solomon’s
provides the best performance in both objectives in R2 and instances with 100 customers. We plan to test our algorithms
RC1 categories. For another two categories, it has the minimal by larger-scale instances in [35] (200 to 1000 customers). The
number of routes. The combination of original crossover and huge solution space and long computation time will be
mutation has the best performance only in total distance in challenging. Last, we want to extend our approach in more
category RC2. The results show that the proposed reproduction realistic VRP with time windows, such as those considering
operators are beneficial in improving solution quality. heterogeneous fleet [36] and pickup and delivery [37].

TABLE III. AVERAGE NUMBER OF VEHICLES AND AVERAGE TOTAL ACKNOWLEDGMENT

DISTANCE OF MOEA-EO VARIANTS WITH FOUR COMBINATIONS OF
CROSSOVER AND MUTATION OPERATORS The authors want to thank Mr. Wei-Sin Wang for collecting
the results of benchmark algorithms. This research was
R1 R2 RC1 RC2 supported by the National Science Council of Republic of
13.17 4.45 12.63 5.63 China under research grant No. NSC100-2221-E-003-004.
E-RCX, E-IM
1186.67 889.07 1351.11 1015.89
13.25 5.55 12.88 5.63
REFERENCES
RCX, E-IM [1] J. Lenstra and K. A. Rinnooy, “Complexity of vehicle routing and
1182.32 890.01 1354.2 1016.40
scheduling problem,” Networks, vol. 11, pp. 221–227, 1981.
13.17 4.64 13 5.63 [2] O. Bräysy, M. Gendreau, “Vehicle routing problem with time windows,
E-RCX, IM
1188.26 890.81 1358.97 1019.98 Part I: Route construction and local search algorithm,” Transportation
Science, vol. 39, no. 1, pp. 104–118, 2005.
13.25 5.27 13.25 6.13
RCX, IM [3] O. Bräysy, M. Gendreau, “Vehicle routing problem with time windows,
1185.55 893.21 1363.58 1014.83 Part II: Metaheuristics,” Transportation Science, vol. 39, no. 1, pp. 119–
139, 2005.
[4] J. Y. Potvin, “A review of bio-inspired algorithms for vehicle routing,”
Comparing our crossover operator with those in the other In: F. B. Pereira and J. Tavares (Eds.) Bio-inspired Algorithm for the
two tested MOEAs in [29] and [31], the main difference is that Vehicle Routing Problem, pp. 1–34, 2009.
our operator exchanges the best routes between solutions [5] J. Y. Potvin, “Evolutionary algorithms for vehicle routing,” INFORMS
whereas their operators exchanged randomly selected routes. Journal on Computing, vol. 21, no. 4, pp. 518–548, 2009.
Sharing good components between solutions should help to [6] K. C. Tan, L. H. Lee, Q. L. Zhu, and K. Ou, “Heuristic methods for
vehicle routing problem with time windows,” Artificial Intelligence in
improve the solution more efficiently. Our crossover operator Engineering, vol. 15, no. 3, pp. 281–295, 2001.
also tries to refine the worst route in the solution, but theirs did
[7] C. J. Ting, C. H. Huang, “An improved genetic algorithm for vehicle
not. As for the mutation, our operator removes a large number routing problem with time windows,” International Journal of Industrial
of customers at a time, but theirs adjusted one customer at a Engineering, vol. 12, no. 3, pp.216－226, 2005.
time. The large-scale modification in our operator could be [8] G. B. Alvarenga, G. R. Mateus, G. de Tomi, “A genetic and set
more helpful to increase the population diversity. More partitioning two-phase approach for the vehicle routing problem with
experiments will be done in the near future to compare the time windows,” Computers & Operations Research, vol. 34, no. 6, pp.
performance of these operators. 1561–1584, 2007.
[9] M. M. Solomon, “Algorithms for the vehicle routing and scheduling
problem,” Operations Research, vol. 35, no. 2, pp. 254－265, 1987.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.
[10] J. Y. Potvin, T. Kervahut, B. L. Garcia and J. M. Rousseau, “The vehicle [25] N. Srinivas and K. Deb, “Multiobjective optimization using
routing problem with time windows, Part I: Tabu search,” Journal on nondominated sorting in genetic algorithms,” Evolutionary Computation,
Computing, vol. 8, no. 2, 158–164, 1996. vol. 2, no. 3, pp. 221 – 248, 1994.
[11] J. Y. Potvin, S. Bengio, “The vehicle routing problem with time [26] K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, “A fast and elitist
windows, Part II: Genetic search,” Journal on Computing, vol. 8, no. 2, multiobjective genetic algorithm：NSGA-II, ” IEEE Transactions on
pp. 165–172, 1996. Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.
[12] O. Bräysy, “A reactive variable neighborhood search for the vehicle [27] K. C. Tan, T. H. Lee, Y. H. Chew, and L. H. Lee, “A multiobjective
routing problem with time windows,” Journal on Computing, vol. 15, evolutionary algorithm for solving vehicle routing problem with time
no.4, pp. 347–368, 2003. windows,” IEEE International Conference on Systems, Man, and
[13] R. Bent, P. V. Hentenryck, “A two-stage hybrid local search for the Cybernetics, vol. 1, pp. 361–366, 2003.
vehicle routing problem with time windows,” Transportation Science, [28] K. C. Tan, Y. H. Chew, L. H. Lee, “A hybrid multiobjective
vol. 38, no. 4, pp. 515–530, 2004. evolutionary algorithm for solving vehicle routing problem with time
[14] J. Homberger and H. Gehring, “Two evolutionary metaheuristics for the windows,” Computational Optimization and Applications, vol. 34, pp.
vehicle routing problem with time windows,” Information Systems and 115–151, 2006.
Operational Research, vol. 37, pp. 297–318, 1999. [29] B. Ombuki, B. J. Ross, and F. Hanshar, “Multi-objective genetic
[15] J. Homberger and H. Gehring, “A two-phase hybrid metaheuristic for algorithms for vehicle routing problem with time windows,” Applied
the vehicle routing problem with time windows,” European Journal of Intelligence, vol. 24, no. 1, pp. 17–30, 2006.
Operational Research, vol. 162, no. 1, pp. 220–238, 2005. [30] K. Ghoseiri and S. F. Ghannadpour, “Multi-objective vehicle routing
[16] J. F. Cordeau, G. Laporte, and A. Mercier, “A unified tabu search problem with time windows using goal programming and genetic
heuristic for vehicle routing problems with time windows,” Journal of algorithm,” Applied Soft Computing, vol. 10, no. 4, pp. 1096–1107,
the Operational Research Society, vol. 52, no. 8, pp. 928–936, 2001. 2011.
[17] J. Berger, M. Barkaoui, and O. Bräysy, “A route-directed hybrid genetic [31] A. Garica-Najera, J. A. Bullinaria, “An improved multi-objective
approach for the vehicle routing problem with time windows,” evolutionary algorithm for the vehicle routing problem with time
Information Systems and Operational Research, vol. 41, pp. 179–194, windows,” Computer & Operations Research, vol. 38, no. 1, pp. 287–
2003. 300, 2011.
[18] Y. Nagata, O. Bräysy, and W. Dullaert, “A penalty-based edge assembly [32] D. E. Goldberg, Genetic algorithms in search, optimization, and machine
memetic algorithm for the vehicle routing problem with time windows,” learning, Addison-Wesley, 1989.
Computers & Operations Research, vol. 37, no. 4, 724–737, 2010. [33] A. Lim and X. Zhang, “A two-stage heuristic with ejection pools and
[19] N. Jozefowiez, F. Semet, and E. G. Talbi, “Multi-objective vehicle generalized ejection chains for the vehicle routing problem withtime
routing problems,” European Journal of Operational Research, vol. 189, windows,” Journal on Computing, vol. 19, no. 3, pp. 443–457, 2007.
pp. 293–309, 2008. [34] N. Labadi, C Prins, and M. Reghioui, “A memetic algorithm for the
[20] S. C. Hong and Y. B. Park, “A heuristic for a bi-objective vehicle vehicle routing problem with time windows,” RAIRO-Operations
routing with time window constraints,” International Journal of Research, vol. 42, no. 3, pp. 415–431, 2008.
Production Economics, vol. 62, ppl. 249–258, 1999. [35] H. Gehring and J. Homberger, “A parallel hybrid evolutionary
[21] M. J. Geiger, “Genetic algorithms for multiple objective vehicle metaheuristics for the vehicle routing problem with time windows,”
routing,” Metaheuristic International Conference, pp. 348–353, 2001. EUROGEN, pp. 57–64, 1999.
[22] B. Barán and M. Schaerer, “A multiobjective ant colony system for [36] F. Belmecheri, C. Prins, F. Yalaoui, and L. Amodeo, “Particle swarm
vehicle routing problem with time windows,” IASTED International optimization to solve the vehicle routing problem with heterogeneous
Conference on Applied Informatics, pp. 97–102, 2003. fleet, mixed backhauls, and time windows,” The 24th IEEE International
Parallel and Distributed Processing Symposium, 2010.
[23] N. Jozefowiez, F. Semet, and E. G. Talbi, “Parallel and hybrid models
for multi-objective optimization: Aplication to the vehicle routing [37] Y. H. Huang and C. K. Ting, “Ant colony optimization for the single
problem,” Lecture Notes in Computert Science, vol. 2439, pp. 271–280, vehicle pickup and delivery problem with time windows,” The 2010
2002. International Conference on Technologies and Applications of Artificial
Intelligence, pp. 537–543, 2010.
[24] N. Jozefowiez, F. Semet, and E. G. Talbi, “Enhacnements of NSGA-II
and its application to the vehicle routing problem with route balancing,”
Lecture Notes in Computer Science, vol. 3871, pp. 131–142, 2006.

Authorized licensed use limited to: Hong Kong Polytechnic University. Downloaded on May 04,2024 at 08:09:29 UTC from IEEE Xplore. Restrictions apply.