Computers & Industrial Engineering 169 (2022) 108221

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

A cluster-based optimization framework for vehicle routing problem with

workload balance
Jingjing LI , Yaohuiqiong Fang , Na Tang *
School of Computer Science, South China Normal University, Guangzhou 510631, China

A R T I C L E I N F O A B S T R A C T

Keywords: Reducing transport costs and guaranteeing the fairness of the workload are major concerns for logistic com­
Multi-objective optimization panies. This paper addresses a vehicle routing problem with workload balance (VRPWB) and a microcluster-
Vehicle routing problem based VRPWB(MVRPWB) to minimize the total traveling costs and balance the workload. Different from exist­
Workload balance
ing works, we propose an optimization framework that includes three components: clustering, microcluster
fusion, and route search. Various proposed algorithms, such as a clustering algorithm, a cluster fusion scheme, an
enhanced ant colony algorithm, and noise processing algorithms, are embedded in these components. We can
flexibly combine different components to solve VRPWB and MVRPWB. The experiment results on the traditional
and revised instances show that the framework can get satisfactory solutions in most instances. Furthermore, a
real case study illustrates that the framework can solve the first mile and last mile problems in practice.

1. Introduction packages. Under the vehicle capacity constraint, we try to minimize the
ratio of maximum workload and minimum workload.
Vehicle routing problem (VRP) is one of the most critical fields in The transportation process includes two main phases: first mile de­
transportation systems (Lu, Zhang, & Yang, 2019). It is defined as the livery and last mile delivery. The first mile delivery means shipping
process of determining routes for vehicles to distribute the products to goods from a warehouse to the next hub from where the goods are
customers (Bertazzi, Golden, & Wang, 2015). Many algorithms proposed forwarded further. In this scenario, the hub can be regarded as the
to reduce cost and improve customer satisfaction in the literature customer of the warehouse. A set of the customers can be randomly
(Marinaki & Marinakis, 2016; Nguyen, Jiang, Ng, & Teo, 2016; Ticha, assigned to the vehicles. The vehicles depart from the warehouse,
Absi, Feillet, & Quilliot, 2019) to solve the problem. deliver packages to the customers, and return to the warehouse. The last
Workload balance is a critical issue for VRP. The benefits of mile delivery involves the transportation of goods from the hub to the
achieving workload balance are twofold: (1) as to drivers, reducing final destination, such as customer’s doorstep (Perboli, Brotcorne, Bruni,
overtime and ensuring each driver is treated equally and fairly; (2) as to & Rosano, 2021). In this scenario, some problems, i.e., locations of some
logistics company, saving the shipping cost and improving the service customers may be challenging to be found, some customers need
quality. Different papers may have different definitions of workload. For particular services, are needed to solve. Thus, the drivers should be
example, Zhang, Sun, Xie, Teng, and Wang (2019b) considered route familiar with geographical distribution regions or satisfy the customer’s
balance, while Fallah, Didehvar, and Rahmati (2020) proposed load requirements. A popular method to deal with the problems is to partition
balance. the customers serviced into clusters according to specific criteria in
Logistics companies measure the driver’s working time by the advance, usually one per vehicle. The route of each vehicle is obtained
driving time and service time (Sivaramkumar, Thansekhar, Saravanan, based on the given clusters.
& Amali, 2018). The driving time is usually calculated by the length of In this paper, according to assumptions and constraints in the above
the route, while the service time is mainly related to unloading time that two scenarios, we first propose two problems: vehicle routing problem
is calculated by the number of units of delivery packages. To ensure with workload balance (VRPWB) and microcluster-based VRPWB
delivery efficiency and achieve fairness, we study the workload, which (MVRPWB). The objectives of these problems are to minimize the total
includes the traveling distance and the number of units of delivery traveling costs and balance the workload. Next, we design an optimi

* Corresponding author.
E-mail addresses: [email protected] (J. LI), [email protected] (Y. Fang), [email protected] (N. Tang).

https://doi.org/10.1016/j.cie.2022.108221
Received 20 June 2021; Received in revised form 28 March 2022; Accepted 1 May 2022
Available online 9 May 2022
0360-8352/© 2022 Elsevier Ltd. All rights reserved.
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

zation framework, which includes three components: (1) clustering: in Table 1

this component, we propose a new clustering scheme, named Symbol table.
DBSCANþ, which combines the improved density-based spatial clus­ Symbol Description
tering of applications with noise (DBSCAN) (Li, Zhang, & Zhu, 2019)
N Set of all nodes (including a depot)
and modified maximum-minimum distance clustering algorithm (i, j) Edge (i, j) represents movement from i to j; i, j ∈ N
(MMDC) (Min, Jin, & Lu, 2019). We use DBSCAN+ to group the cus­ di Demand at node i; i ∈ N
tomers into clusters. Here, the obtained cluster is called a microcluster. V Set of vehicle
(2) Microcluster fusion: in this component, different microclusters are n Number of customer nodes
m Number of vehicle
fused to form clusters for balancing the workload. We design a cij The travel cost from i to j; i, j ∈ N
Microcluster fusion scheme to execute fusion. (3) Route search: the xvij Binary variable equal to 1 if vehicle v can travel from
modified ACO, named ACOþ, is applied in the clusters to search, then i to j directly, and 0, otherwise; i, j ∈ N; v ∈ V
each vehicle gets the final route. The framework can solve VRPWB and bv Workload of vehicle v; v ∈ V
MVRPWB through different combinations. Extensive numerical studies yvi Binary variable which takes value 1 if vehicle v
serves customer i, and 0, otherwise; i ∈ N⧹{0}; v ∈ V
on traditional and generated test instances demonstrate that the pro­
α Weight of workload
posed algorithms are very effective. In addition, we apply the framework β Weight of total travel cost
for first mile and last mile delivery. The results show that our framework Cv Capacity of vehicle v; v ∈ V
can also obtain high-quality solutions. uvi The remaining capacity of vehicle v before visiting
The remaining sections are organized as follows. Section 2 presents customer i; i ∈ N⧹{0}; v ∈ V
related works. Section 3 presents problem formulations. Section 4 pro­ MicroCluster The set of microclusters
σ + (Z) The set of outgoing edges (i, j) ∈ Z × N⧹Z, Z is any subset
poses the framework and the algorithms for the problems. Sections 5 and of N that is not equal to N
6 present experimental results. Finally, Section 7 presents the σ − (Z) The set of incoming edges (i, j) ∈ N⧹Z × Z, Z is any subset
conclusion. of N that is not equal to N
mCrk The microcluster k; k = 1, 2, ⋯, m
2. Related work NmCrk Set of all customers in mCrk
ZvmCrk Binary variable which takes value 1 if vehicle v serves

Most VRPs focus on minimizing transport costs. In practice, the microcluster mCrk , and 0, otherwise

equitable workload is essential for the drivers and ensuring service

quality. The workload has different definitions in the historical litera­ 3. Problem formulation
ture. Initially, the papers only considered the route per vehicle traveled
as the workload. VRP with route balance (VRPRB) is proposed by We consider a transportation service of packages from a depot to
Jozefowiez, Semet, and Talbi (2009), in which route balance is to customers under the following scenarios: (1) the vehicles have no prior
minimize the difference between the longest and the shortest route knowledge, and the customers can be assigned to them randomly; (2) the
lengths among vehicles. For solving this problem, Lacomme, Prins, vehicles need to deliver based on the given customer clusters. In this
Prodhon, and Ren (2015) proposed a multi-start method based on two section, we present the problem formulations. The symbols used for the
search spaces. Mandal, Pacciarelli, Løkketangen, and Hasle (2015) problems are provided in Table 1.
proposed an NSGA-II extension algorithm based on Pareto’s local search.
Lacomme, Prins, and Ramdane-Cherif (2004) minimized the longest
route length to balance the route. In Sun, Liang, Zhang, and Wang 3.1. Vehicle Routing Problem with Workload Balance (VRPWB)
(2017), a memetic algorithm based on NSGA-II for this type of VRPRB is
developed. The workload can also be defined as load. Mancini, Gans­ VRPWB is illustrated by a distribution center that employs a fleet of
terer, and Hartl (2021) introduced the collaborative VRP with time capacitated vehicles delivering the packages to the corresponding cus­
service consistency and load balance. Londono, Rendon, and Ocampo tomers. All routes traveled by the vehicles should start and end at the
(2021) considered a green VRP and proposed a multi-objective method same distribution center. Each customer must be served by exactly one
considering load equity. In Chiang and Cheng (2017), the workload is vehicle, and each vehicle visits at least one customer. The goals of
related to profit. The proposed model included the performance bonus VRPWB are finding a set of routes with minimum total travel costs, and
into VRP and realized the load balance among employees. Zhang, Qin, the workloads of vehicles should be balanced.
and Li (2019a) used two multi-objective local search (MOLS) algorithms We define VRPWB on a complete undirected graph G = (N, E), where
to maximize the minimum unit profit of all transport companies when vertex set N = {0, 1, 2, ⋯, n} consists of depot 0 and n customers, E = {(i,
outsourcing transportation tasks. j)|i, j ∈ N } is the edge set. The travel cost of each edge (i, j) is represented
To improve the satisfaction of customers and delivery reliability, the by cij , and each customer i ∈ N⧹0 is associated with a demand di . The
customers are grouped into clusters based on the territory or delivery demand of depot is assumed as d0 = 0. The objectives of VRPWB are as
characteristics in advance in many practical applications. Expósito- follows.
Izquierdo, Rossi, and Sevaux (2016) proposed an approximate two-level
approach to find the routes for capacitated vehicles to serve the cus­ 1. The total travel cost,
tomers organized into clusters. Janssens, Van den Bergh, Sörensen, and ∑∑∑
Cattrysse (2015) modeled a microzone-based vehicle routing problem f1 = cij xvij (1)
aiming at workload balance while achieving the shortest journey path i∈N j∈N v∈V

length. In Mokhtarinejad, Ahmadi, Karimi, and Rahmati (2015),a novel

machine-learning-based heuristic method (MLBM) was proposed to
2. The max-min vehicle workload,
solve VRP. MLBM includes three stages that are clustering, sub-problems
modeling, and solving the whole model. max{bv }
f2 = (2)
In this paper, we focus on two VRPs with workload balance. Different min{bv }
from the methods mentioned above, the workload is the combination of
route length and load. The workload balance is to minimize the ratio of
maximum workload and minimum workload. We proposed a framework
to solve the two problems, which the previous works had not considered. In f1 , the cost of cij is defined as the distance of the edge (i, j).

2
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

∑( )
yvi di ⩽Cv , ∀v ∈ V (9)
i∈N

uvi − uvj + Cv xvij ⩽Cv − dj ,

(10)
∀i, j ∈ N⧹{0}, i ∕ = j, v ∈ V, di + dj ⩽Cv

di ⩽uvi ⩽Cv , ∀i ∈ N \ {0}, v ∈ V (11)

Fig. 1. An example of a computational workload. The constraints (4) along with (5), indicate that each node obeys
flow conservation constraints. If vehicle v enters node i, then it must
Minimization of f1 is to save the travel time and the energy consumption depart from node i. Constraint (6) ensures that each vehicle has at least
for vehicles, therefore the total transport costs are reduced. one customer. Constraint (7) ensures that each customer only served by
The second objective f2 is to promote equity among the vehicles. one vehicle. Constraint (8) ensures that all vehicles must depart from the
Firstly, we use bv to represent the workload of the vehicle v, as shown in same depot. Constraint (9) ensures that the total delivery demand of the
Eq. (3), which is the sum of the demands of the nodes visited by the nodes in each route should not exceed the vehicle capacity. Subtour
vehicle and the travel distance of the vehicle. The demand of nodes is elimination is given in (10) and (11), which are proposed by Miller,
abstracted by the number of the basic transportation units. Tucker, and Zemlin (1960).
[ { ( )}] Therefore, VRPWB in this paper can be summarized as
∑ ∑ ∑
bv = yvi αdi + β cij xvij (3) (minf 1 , minf 2 ) (12)
i∈N j∈N v∈V

subjects to (4)–(11).
According to the common method in VRP, such as paper Jozefowiez
et al. (2009), the difference between the maximum and minimum values
of bv is minimized. However, this method is sensitive to the orders of 3.2. Microcluster-based VRPWB(MVRPWB)
magnitude for fairy allocating workload. For example, as shown in
Fig. 1, square and circle indicate a distribution center and customers, In this problem, the customers are grouped into microclusters in
respectively. The number in a circle represents the node’s demand, and advance, that is MicroCluster = {mCr1 , mCr2 , ⋯, mCrk }, k = m. The
the number on a line represents the travel cost between nodes. Vehicle v1 number of microclusters is equal to the number of vehicles. Each
travels along the path marked by the solid line, and vehicle v2 travels microcluster is assigned to a vehicle and each vehicle servers a micro­
along the other path marked by the dotted line. Assume α and β are equal cluster. In addition to the above mentioned, the assumptions and ob­
to 1 and 0.1, according to Eq. (3), we get bv1 = 21 and bv2 = 20. The jectives of MVRPWB and VRPWB are indentical.
difference between the workloads of the two vehicles is 1. We can MVRPWB can be written as
conclude that the two vehicles along the two routes reach workload (minf 1 , minf 2 ), subjects to (4)–(11) and
balance. However, if all the values in Fig. 1 are multiplied by 10, the ∑ ∑ v ∑ ∑
workload difference of the two vehicles is increased 10 times. The xij = xvij = 1,
workloads of the vehicles are not balanced even they travel along the (i,j)∈δ+(NmCrk )v∈V (i,j)∈δ− (NmCrk ) v∈V (13)
same routes. ∀mCrk ∈ MicroCluster
In order to eliminate the orders of magnitude sensitivity, we intro­
⋂
duce a new optimization function f2 that is the ratio between the NmCri NmCrj = ∅, ∀mCri , mCrj ∈ MicroCluster, i ∕
= j (14)
maximum and minimum workload of the vehicles. Minimization of f2
keeps workload fair and avoids the influence of the orders of magnitude. ∑ ∑
xvij = xvij = zvmCrk ,
In Fig. 1, the value of f2 is 1.05. Even all the values are multiplied by 10,
the value of f2 is not changed. Hence, the workload balance can be
(i,j)∈δ+(NmCrk ) (i,j)∈δ− (NmCrk ) (15)
intuitively reflected by using f2 without errors caused by the instances
k
∀v ∈ V, ∀mCr ∈ MicroCluster
with the same topology and different orders of magnitude. ∑
In addition, the constraints for VRPWB are defined as follows: xvi0 = 1,
∑ (i,0)∈δ+(NmCrk )
(16)
xvij = yvj , ∀v ∈ V, ∀j ∈ N, j ∕
= i (4) ∀v ∈ V, ∀mCrk ∈ MicroCluster
i∈N

∑ Constraint (13) determines that each vehicle only serves one

xvij = yvi , ∀v ∈ V, ∀i ∈ N, i ∕
= j (5) microcluster. Constraint (14) indicates the microclusters are disjoint.
j∈N
Constraint (15) ensures that if vehicle v enters a microcluster mCrv , it
must depart from the microcluster mCrv . Constraint (16) guarantees that
xvij = 0, ∀v ∈ V, i = j = 0 (6)
each vehicle must return to the depot after departing from a
∑ microcluster.
yvi = 1, ∀i = 1, ⋯, n (7) However, to save costs and balancing workload, it could be better to
v∈V
relax the constraints that allow vehicles to enter and leave microclusters
∑
yv0 = m (8) multiple times, and the customers in a microcluster can be reassigned to
v∈V different available vehicles. We introduce this problem as MVRPWB
with soft microcluster constraints. Its mathematical formulation is

3
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

described as follows: 4.1. Clustering

(minf 1 , minf 2 ), subjects to (4)–(11), (14) and
∑ ∑ v ∑ ∑ In the traditional DBSCAN, there are two key parameters: the radius
xij = xvij ⩾1, of neighborhood Eps and density MinPts. DBSCAN determines the value
(i,j)∈δ+(NmCrk )v∈V
(i,j)∈δ− (NmCrk ) v∈V
(13’) of Eps and MinPts based on the location of the mutation point. The
k
∀mCr ∈ MicroCluster characteristic of the mutation point is that the i nearest neighbor dis­
tance Disti changes smaller before the point of mutation, and the i
∑
∑
xvij = xvij ⩾1, nearest neighbor distance Dist i of each sample point is larger after the
(i,j)∈δ+(NmCrk ) (i,j)∈δ− (NmCrk )
mutation point k (Li et al., 2019). However, the method for setting Eps
(15’)
and MinPts is not suitable for solving VRP. The reasons are listed as
k
∀v ∈ V, ∀mCr ∈ MicroCluster follows.
For the multi-objective problems, there exists a set of solutions that
1. In VRP, each node has demand, and each cluster has an upper bound
are superior to others when considering all objectives, that is called
of load. The traditional DBSCAN does not consider the load.
Pareto set or non dominated solutions (Doerner, Gutjahr, Hartl, Strauss,
2. There are no mutation points in some instances.
& Stummer, 2004). We need to design the optimization algorithms to
find a set of solutions that are not only close to the Pareto front
Therefore, in DBSCAN+, we modify the calculation methods of Eps
(convergence) but also well distributed along it (diversity) (Wang,
and MinPts, which can ensure the stability of results for the same
Weng, & Zhang, 2018).
instance and handle the situation that there is no mutation point in the
instance.
4. Solution
In addition, the traditional DBSCAN generates the initial core points
randomly for simplifying computation. However, two initial core points
In the paper, we propose an optimization framework for solving the
may be very close so that the effect of the clustering is influenced. We
two problems. Specially, we present the novel algorithms embedded into
use improved MMDC to solve the problem. Although traditional
the framework. The framework consists of three components: clustering,
DBSCAN recognizes the noise nodes, it does not process them. We pro­
microcluster fusion and route search. In the first component, for fully
pose a noise processing algorithm so that the outliers can be absorbed
utilizing the characteristics of the VRPWB, we propose DBSCAN+ to find
into the microclusters. The main procedure of DBSCAN+ is given in
microclusters by combining DBSCAN with the MMDC. We also carry out
Algorithm 1.
noise processing. In the second component, the microclusters obtained/
Algorithm 1 DBSCAN+
given will be fused according to specific rules. Then clusters with
workload balance are formed. Besides, the noise nodes will be treated Input: an instance to be solved
{ }
for the second time. The operations in the two components can effec­ Output: MicroCluster= mCr1 , mCr2 , ⋯, mCrk , k = m; a set of noise nodes

tively avoid the generation of unnecessary paths and prevent trapping 1: initialize MinPts and Eps, the node farthest from the warehouse as X1 , all nodes for
unvisited state;
into the local optimal. In the third component, a heuristic algorithm 2: Cnum←1;
called ACO+ is developed for carrying out route search in the clusters. 3: repeat
By combining different components in the framework, we can solve 4: let Ndirect be the set of nodes in Eps range of Xk ;
different problems. Fig. 2 illustrates the optimization framework and 5: for all nodes X in Ndirect do
combinations of the components for different problems. Next, we 6: if X is unvisited then
7: if X blongs to the core point and DemandmCrk +DemandX′ eps < Cv then
describe the algorithms in detail. ⋃
8: Ndirect ←Ndirect NX′ eps ;
9: end if
10: if DemandmCrk +dX < Cv then
⋃
11: NmCrk ←NmCrk X;
12: DemandmCrk + = dX ;
13: mark X as visited;
14: end if
15: end if
16: end for
17: Cnum + + ;
18: if Cnum > 1 and Cnum⩽m then
19: FlagCnum ←0;
20: calculate the distance arrays D from all unvisited nodes to X1 , X2 , ⋯, XCnum− 1 ;
21: if initial core point-compliant nodes exist among the unvisited nodes then
22: XCnum ←comply with initial core point and is an unvisited node for
max{min(D) };
23: else
24: XCnum ←comply with core point and is an unvisited node for max{min(D) },
or just is an unvisited node for max{min(D) };
25: end if
26: end if
27: until (Cnum > m)

Fig. 2. The optimization framework.

First, the algorithm calculates MinPts and Eps, while initializes the
initial core points and the state of the nodes (line 1). Next, it performs a
loop to form microclusters (line 3–27). In the loop, the improved
DBSCAN considering the load is performed (line 4–16). Then, the initial

4
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

core points are generated in improved MMDC (line 17–26). The loop
stops iterating when the number of microclusters is equal to the number Algorithm 2 ProcenoiseP1

of vehicles.
{ }
Require: MicroCluster= mCr1 , mCr2 , ⋯, mCrk , k = m ; a set of noise nodes
In the following paragraphs, we present the calculation of the pa­
{ 1 2 k
}
Ensure: MicroCluster= mCr , mCr , ⋯, mCr , k = m ; a small number of noise nodes
rameters and the treating process for noise. 1: for all unvisited nodes X do
2: initialize DCX , Min, Minclu, Flag;
3: IfMax←min(DCX );
1. MinPts is the minimum number of data points to define a cluster
4: for k←1 to m do
(Ester, Kriegel, Sander, & Xu, 1996). Here, MinPts is calculated as 5:
( )〈
if DistCkX < Min and DistCkX < γ2 IfMax and DemandmCrk + dX γ1 Ck then
√̅̅̅ 6: Min←DistCkX , Minclu←k, Flag←1;
n
MinPts = ⌊ ⌋, MinPts⩾4 (17) 7: end if
2
8: end for
9: if Flag == 1 then
For each customer node i, a distance array Dist i that stores the dis­ 10:
⋃
NmCrMinclu ←NmCrMinclu X;
tances between node i and other nodes in ascending order. The Total 11: DemandmCrMinclu + = dX ;
is the sum of values of Dist i[MinPts] in each array Dist i. Eps is the 12: mark X as visited;
distance that specifies the neighborhoods (Ester et al., 1996). Eps is 13: end if
calculated as 14: end for

Total
Eps = (18)
n
4.2. Mircrocluster fusion scheme
Note that if the number of initial core points is less than a certain
value LimitE (preferably between 20% ∼ 40% of the total number of There are two methods to get microclusters. Firstly, DBSCAN+ can
nodes), then the Total is the sum of values of Dist i[2⋅MinPts]. Thus, get the microclusters, and each vehicle only belongs to a microcluster.
Eps becomes greater. Secondly, according to the definition of MVRPWB, the microclusters can
A customer node i is an initial core point if at least MinPts unvisited be given. No matter which method is used, there is no guarantee that
customer nodes within distance Eps of it (excluding depot). A each microcluster will be workload balanced. Thus, the second objective
customer node j is a core point if at least MinPts customer nodes of the problems cannot be reached. To improve the performance of
within distance Eps of it (excluding depot). clustering and balance the workload of vehicles, we propose a micro­
2. The noise nodes are the nodes that do not belong to any microcluster. cluster fusion scheme.
They will be treated in algorithm ProcenoiseP1 for the first time. Above all, we provide definitions related to the proposed scheme.
As shown in Algorithm 2, ProcenoiseP1 computes the average Definition 1. Cluster: Cluster Crk is a set of microclusters after fusion.
distances DistCkX between the noise node X and the nodes in each It includes one or more microclusters.
microcluster mCrk in turn and records them in an array DCX . The
Definition 2. directly reachable node: A node Q is directly reachable
minimum value of DCX is selected as IfMax. ProcenoiseP1 then puts
from P if node P is within distance Eps of it.
the noise node X into a microcluster mCrk with the minimum DistCkX
that satisfies a specific condition. Here, the condition is: Definition 3. adjacent clusters/microclusters: if two nodes belonging to
( )〈
dX + DemandmCrk γ1 Ck && DistCkX < γ 2 IfMax where DemandmCrk two different clusters/microclusters are directly reachable, we call the
two clusters/microclusters are adjacent clusters/microclusters. The
is the current load of microcluster mCrk :
adjacent value between the adjacent clusters/microclusters is calculated
∑
Demand mCrk = di (19) as
i∈NmCrk
Aij = Ki *Kj (22)
k k
γ 1 C is the upper bound of the load of mCr . γ1 and γ2 are two pa­ where Ki is the number of nodes belonging to Cri /mCri that are directly
rameters. If γ 1 or γ 2 are too small, many noise nodes will not be
reachable from Crj /mCrj , Kj is the number of nodes belonging to Crj /
absorbed by the microclusters. If γ1 is too great, the differences of the
loads among the microclusters will become larger thereby affecting mCrj that are directly reachable from Cri /mCri .
the effect of following fusion. Moreover, if γ2 is too great, the Definition 4. nearby cluster/microcluster: if there are nodes in a circle
microcluster will absorb the noise node far from it, leading to poor with Icore as the center and 3L as the radius, we call that the cluster/
processing results. microcluster containing these nodes is nearby cluster/microcluster of
Therefore, we set parameters γ1 and γ 2 adaptively based on loads the microcluster mCrk . Herein, Icore and L is defined as
of microclusters and the distribution of noise nodes. γ 1 is defined as (∑ ∑ )
∑ i∈NmCrk xi i∈NmCrk yi
di Icore = , (23)
(20)
i∈N
γ 1 = ∑ visited nk nk
i∈N di
∑
where Nvisited represents nodes that have been clustered (marked as MCDist[i]
(24)
i∈NmCrk
L=
visited status), meanwhile, γ2 is defined as nk
∑ secondmin
i∈Nnoise DistiC i where iis the node in the current microcluster mCrk , nk is the size of the
γ2 = ∑ (21)
min
i∈Nnoise DistiC i
microcluster mCrk , and array MCDist records the distance between the
nodes in the microcluster and the coordinate center of the microcluster
where DistiCmin
i represents the smallest average distance between Icore.
noise node i and the nodes in the microcluster. DistiCsecondmin
i repre­ Definition 5. irrelevant clusters/microclusters: if two clusters/micro­
sents the smallest average distance between noise node i and the clusters are not adjacent or nearby clusters/microclusters, we call them
nodes in the microcluster except DistiCmin
i . Nnoise represents the set of irrelevant clusters/microclusters.
noise nodes.
The procedure of the microcluster fusion scheme is described as

5
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

follows: (continued )
Step1 : Initialization. Initializing all the microclusters as available that 26: Flag + + ;
means they can be fused into clusters. 27: end if
28: until (Flag = 2)
Step2 : Clusterconstruction. As to a microcluster mCrk , its different
states will fire different actions. When mCrk is available and has not been
fused, the scheme checks whether it has adjacent clusters/microclusters. Algorithm 4 ProcenoiseP2
If there are adjacent clusters/microclusters, mCrk is fused with the one { }
Input: Cluster= Cr1 , Cr2 , ⋯, Crl , l⩽m; a small number of noise nodes
with the largest adjacent value. Otherwise, if the load of mCrk is less than { 1 }
Output: Cluster= Cr , Cr , ⋯, Crl , l⩽m
2

50% of the average microcluster load (DemandAver ), the scheme searches 1: for all unvisited nodes X do
for its nearby clusters/microclusters. Then, it fuses mCrk with a nearby 2: initialize DCX , Min, Minclu, Flag;
3: for k←1 to l do
cluster/microcluster nearest to mCrk . When mCrk is available and fused ( )〈
4: if DemandCrk + dX Vnumk Ck andDistCkX < Min then
in a cluster Crk , the scheme checks whether the cluster Crk has adjacent 5: Min←DistCkX , Minclu←k, Flag←1;
clusters/microclusters. If there are adjacent clusters/microclusters, Crk 6: end if
is fused with the one with the largest adjacent value. After fusion, when 7: end for
8: if Flag == 1 then
the number of microclusters in the fused cluster is equal to the upper ⋃
9: NCrMinclu ←NCrMinclu X;
bound MCr , the contained microclusters become unavailable, where 10: DemandCrMinclu + = dX ;
√̅̅̅̅
MCr = ⌊ m ⌋. The process of cluster construction is repeated until all the 11: mark X as visited;
microclusters have been checked. The microcluster mCrk becomes a 12: end if
13: end for
cluster directly if it has not been fused at the end.
We find that there are still some noise nodes needed to be processed.
Therefore, ProcenoiseP2 is designed to treat the noise nodes for the 4.3. Route search
second time.
ProcenoiseP2 computes the average distances DistCkX between the Route search is performed in the clusters to find an optimal route
noise nodes and the nodes in each cluster Crk and records them in an meeting two objectives. ACO is an efficient heuristic algorithm for
finding optimal paths in multi-objective problems (Barán & Schaerer,
array DCX . The cluster Crk with the nearest distance from the noise node
2003). In the general ACO for VRP, each ant represents a vehicle. It visits
X and its DemandIf k smaller than the upper bound is selected to absorb
some customers until its capacity is full and the solution is constructed
the noise node X. DemandIf k is the sum of the demand of node X and the with the paths of ants (Mazzeo & Loiseau, 2004). For reaching the sec­
load of the cluster Crk . The upper bound is Vnumk ⋅Cv , where Vnumk is the ond objective of the problems, we propose ACO+ to improve ACO by
number of microclusters contained in Crk . Since a total load of vehicles is modifying the meaning of an ant and introducing the return mechanism.
more than the total demand of the customer nodes, the noise node can Herein, the path of each ant represents a feasible solution that avoids the
always be absorbed in a cluster. solution construction. Moreover the return mechanism enforces the ant
Algorithm 3 and Algorithm 4 are the pseudocode of the microcluster to return the depot when its workload reaches the upper bound,
fusion scheme and ProcenoiseP2, respectively, where function Fuse (A, balancing the vehicles’ workload.
B) fuses cluster/microcluster A with cluster/microcluster B to a cluster. The main procedure of the proposed ACO+ algorithm is given in
Algorithm 3 microcluster fusion scheme Algorithm 5, in which a set of Pareto optimal solutions is updated by
{ } modifying the concentration of pheromone in the clusters. The entire
Input: MicroCluster= mCr1 , mCr2 , ⋯, mCrk , k = m
{ 1 } inner loop (line 5–19) shows the process of solution construction. There
Output: Cluster= Cr , Cr2 , ⋯, Crl , l⩽m
are na ants at the depot. Each ant can leave and return the depot some
1: initialize all the microclusters in available state;
2: Flag←0; rounds. In a round, an ant departs from the depot, visits some customer
3: repeat nodes, and returns to the depot. The path visited by the ant in a round ri
4: for k←m to 1 do corresponds to the path traveled by a vehicle v. The number of rounds an
5: if mCrk is available and has been fused then ant needs to perform is equal to the number of vehicles. Therefore, each
6: if there are clusters/microclusters adjacent to mCrk ’s Cluster Cri and # of ant represents a solution, it finally visits all the customers. The ant uses
Cri after fusing ⩽MCr then
( ) the pheromone updating rule and return mechanism in the function
7: Cri ←Fuse Cri , Cr/mCrmaxAij
ChooseNN(Point i ) to select the next customer.
8: if # of Cri == MCr then
{ ⃒ } Algorithm 5 ACO+
9: mark mCrk ⃒mCrk ∈ Cri as unavailable;
{ }
10: end if Input: Cluster= Cr1 , Cr2 , ⋯, Crl , l⩽m
11: end if Output: The Pareto solution set P
12: else if mCrk is available and unfused then 1: P←∅;
13: if there are clusters/microclusters adjacent to mCrk and # of Cri after 2: initialize the pheromone trails on each arc;
fusing ⩽MCr then 3: for r←0 to R do
14:
(
Cri ←Fuse mCrk , Cr/mCrmaxAij
) 4: SP ←greedy algorithm gets the initial solution;
5: for a←0 to na (number of ants) do
15: else if there are clusters/microclusters nearby mCrk and DemandmCrk <
6: Sa ←depot 0, ri←0, Demandari ←0, and initialize tabu list;
0.5DemandAver and # of Cri after fusing ⩽MCr then
( ) 7: for k←1 to l do
16: Cri ←Fuse mCrk , Cr/mCrnearest 8: repeat
17: end if 9: Pointj ←ChooseNN(Pointi );
if # of Cri == MCr then
⋃
18: 10: Sa ←Sa Pointj , Pointi ←Pointj ;
{ ⃒ }
19: mark mCrk ⃒mCrk ∈ Cri as unavailable; 11: Demandari + = dPointj ;
20: end if 12: update the local pheromone and tabu list;
21: end if 13: if Pointj == depot 0 then
22: end for 14: ri + + , Demandari ←0;
23: if the Fuse(A,B) was executed in this loop then 15: end if
24: Flag←0; 16: until (NCrk ⊆ NSa )
25: else 17: end for
(continued on next column) (continued on next page)

6
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

(continued ) ⎧
18: SP ←update(Sa , SP ); ⎪
⎪ ( )
⎪
⎪ Kp
19: end for ⎪
⎪
⎪
⎪ Vp pheij + ,
20: P←update(SP , P); ⎪
⎪
⎪ lengthP
21: if Pdoesn’t change after R0 iterations then
⎪
⎪ { }
⎪
⎪
22: Break; ⎪
⎪ if (i, j) ∈ Smin{f1 (s) } , Smin{f2 (s) }
⎨
23: end if pheij = Kp (28)
24: delete the cross and repeated paths in the extreme solution on the second ⎪
⎪ Vp⋅pheij + (1 − Vp) ,
⎪
⎪ lengthP
objective; ⎪
⎪
⎪ { }
25: update the global pheromone according to equation (28); ⎪
⎪
⎪ if (i, j) ∈ P⧹ Smin{f1 (s) } , Smin{f2 (s) }
26: end for
⎪
⎪
⎪
⎪
⎪
⎪ Vp⋅pheij , otherwise
⎪
⎩
4.3.1. Return mechanism
The return mechanism defines the workload indicator doubari for each
where Vp and Kp are parameters, and lengthP is the average total path
vehicle v when an ant a is in a cluster Crk as follows: length in the Pareto optimal solution set. Additionally, we set a upper
nowBkari bound Maxpha and a lower bound Minpha for the pheromone pheij on
doubari = k (25) each edge (i, j).
avBkCr
Kp
where nowBkari is the workload for an ant a in round ri. avBkCr is the
k
Maxphe = (29)
(1 − Vp)lengthP
i
average workload of all the vehicles in the cluster Cr .
nowBkari is: Minphe =
Maxphe
(30)
2n
nowBkari = bv (26)

were the round ri performed by the ant a corresponds to the vehicle v. Algorithm 5.1 ChooseNN(Pointi )
Crk Input: the previous node Pointi
avBk is: Output: the next node Pointj
∑ k
1: initialize doubari , nowBkari and avBkCr according to equation (25)–(27);
(αdi ) + βClulenk
(27) 2: if Demandari == Cv then
Crk i∈NCrk
avBk =
Vnumk 3: return depot 0;
4: end if
where Clulenk represents the total length of paths that all the vehicles 5: if doubari > ε and doubari ⩽1 then
if Random1 < ε1 ⋅doubari then
k
have visited in the cluster Crk . And the value of avBkCr is from the 6:
7: return depot 0;
current Pareto optimal solution set P.
8: end if
The return mechanism is executed base on the workload indicator. 9: else if doubari > 1 then
When the doubari is greater than ε, the mechanism enforces the ant to go 10: if Random2 < ε2 ⋅doubari then
back to the depot with a probability (line 5–13 of Algorithm 5.1). If 11: return depot 0;
12: end if
doubari is smaller than 1 that means the workload of the ant is smaller
k
13: end if
than the average workload avBkCr of the cluster, then the ant returns 14: if doubari > ζ then
with a probability ε1 . Otherwise, the ant returns with a larger proba­ 15: if Random3 ⩽ζ1 then
bility ε2 . 16: return the node with the highest visibility Tpij in Crk or depot 0 according to
the tabu list;
(0, ζ] is the range of doubari that an ant cannot return to the depot,
k
17: end if
where ζ⋅avBkCr is the lower bound of workload that an ant should have 18: if Random3 > ζ1 then
in a round. 19: return the node randomly selected by roulette in Crk or depot 0 according to
the tabu list;
20: end if
4.3.2. Pheromone updating rule
21: else
There are two ways: (1) reducing the maximum workload of the 22: return the node randomly selected by roulette in Crk according to the tabu list;
vehicle; (2) increasing the minimum workload of the vehicle to achieve 23: end if
the second objective of the problems. However, in the algorithm, the
vehicle may take a detour to increase its workload. In this way, although
the solution is Pareto-optimal, some routes are not optimal in terms of The ant can choose the next node to visit when it is not forced to go
the total traveling costs (Wan, Wang, & Yu, 2014). Therefore, we back by the mechanism (line 14–20 of Algorithm 5.1). There are two
remove the cross paths and repeated paths before pheromone updating methods for finding the next node. First, the ant at the node i selects a
in the solution with the best workload balance. next node j according to Tpij , that is:
In ACO+, the local pheromone updating according to Mutar, Bur­ ( )K2
hanuddin, Hameed, Yusof, and Mutashar (2020), the global pheromone Tpij = pheK1
K3
(31)
ij ⋅
updating rule is shown in (28). cij

where K1, K2 and K3 are constant parameters, pheij is calculated by Eq.

(28). Second, the ant selects the next node randomly by roulette. The
selection probability of each method is ζ1 .

7
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

4.4. Time complexity 5.1.1. Comparison results

Table 2 shows a comparison between the proposed framework and
In this subsection, we mainly discuss the time complexity of the Jozefowiez’s method in aspects of the total traveling costs and the dif­
proposed algorithms in the three components. The time complexity of ference. The format of the instance name is “X − nnum1 − knum2”, where
Algorithm 1 is determined by the number of customers, which takes the number of customers num1, the number of vehicles num2, if X is
O(n2 ). The time complexity of Algorithm 2 and Algorithm 3 is deter­ equal to “E”, the instances are from Christofides and Eilon (1969),
mined by the number of customers and vehicles, which are both O(nm). otherwise, the instances are from Christofides et al. (1979). In this table,
The numbers of customers and fused clusters determine the time although the two approaches can obtain non-dominated solutions, we
complexity of Algorithm 4, that is O(nl). The time complexity of Algo­ only consider extreme solutions on two objectives separately.
rithm 5 is O(Rna nnl ) in the worst case, which is associated with the In terms of the total traveling costs, our framework obtains values
number of iterations, the number of ants, the total number of customers close to those found by Jozefowiez’s method in most instances. Columns
and the largest number of customers in the clusters nl . 8 and 10 show that the proposed framework produces better than
Jozefowiez’s method in the first instance. The average difference pro­
5. Experimental results vided by the framework is about 5.535 versus 0.614 for Jozefowiez’s
method when solely minimizing the difference. However, the results in
In order to validate the proposed algorithms, we implement the ex­ column 4 and 7 show that in terms of running time, our framework
periments in C++ on a PC (INTEL i7-8750H CPU 2.20 GHz with 16 GB significantly outperforms Jozefowiez’s method on all the instances. The
RAM). We execute the Algorithms 10 times per instance and compute values are less than 60% of Jozefowiez’s. It means that the framework
the average runtime. Unless mentioned otherwise, in the problem defi­ has better convergence than Jozefowiez’s method. Additionally, the
nitions, the weight of workload α and total travel cost β equal 1. For the advantage is more evident on larger scale instances. The framework is
basic parameters of ACO+, the best results achieved with Vp = 0.8, Kp = able to process the customer nodes in the first and second components,
K3 = 1 or 100, K1 = 2, K2 = 5. Through observation on the test in­ thus ACO+ computes the distance between nodes in the cluster instead
stances, we assign ε = 0.7, ε1 = 0.3, ε2 = 0.7 and ζ = 0.85 to the return of the distance between all the nodes which can accelerate iteration
mechanism. When selecting the next nodes in a path, we assume that the solution compared with Jozefowize’s method.
node is selected randomly by roulette at a ratio of ζ1 = 0.5. Since the An example of the balanced routes obtained by the two methods on
algorithms for the multi-objective problems should satisfy both instance E-n51-k5 is given in Fig. 3. This instance contains 5 vehicles,
convergence and diversity, we use two metrics: Inverted Generational and the routes in the same color belong to the exact vehicle. From Fig. 3
Distance (IGD) and Hypervolume (HV), to evaluate the performance of (b), we can find that many routes take detours, and thereby there are
the algorithms. IGD (Wang et al., 2015) is an indicator to quantify the many crossing points on the routes. Fig. 3(a) shows no cross path for
convergence and diversity of the nondominated solutions obtained. The each vehicle, indicating that the balance is not achieved by detouring.
smaller value of the IGD means the solution set obtained is denser and Therefore, we can conclude that although the solutions obtained by the
closer to the true Pareto front (Wang et al., 2018). HV (Wang et al., proposed framework are slightly worse than the solutions generated by
2015) considers the degree of accuracy and diversity of the solutions Jozefowiez’s method, the framework uses a much shorter running time
produced by the algorithm. The greater value of the HV means the more to generate the routes without crossing points.
closeness of the solutions to the Pareto-optimal front.

5.1. Solving VRPWB

In this section, we focus on solving VRPWB based on instances ob­

tained from Christofides and Eilon (1969),Christofides, Mingozzi, and
Toth (1979), and Augerat, Belenguer, Benavent, Corberan, and Rinaldi
(1995). Since no problem is identical with VRPWB, we compare our
proposed framework with Jozefowiez et al. (2009) which introduced a
similar problem, named the VRP with routing balance (VRPRB). The
VRPRB aims to minimize the total traveling distance and the difference
between the longest and the shortest path lengths. For better compari­
son, in our framework, we modified the second objective as minimizing
the difference between the longest and shortest path lengths and
changed the weight of the workload α to 0. Then, we design validation
Fig. 3. The solution for instance E-n51-k5 with the smallest difference between
experiments to demonstrate the effectiveness of the proposed algo­
the longest and shortest route lengths. (a) Generated by the proposed frame­
rithms. We also discuss which solution should enhance the phenomenon
work. (b) Generated by Jozefowiez’s method.
on paths.

Table 2
Performance comparison between Jozefowiez’s method (Jozefowiez et al., 2009) and the proposed framework on VRPRB instances.
Instance Solely minimizing the total traveling costs Soly minimizing the difference

Jozefowiez’s method the proposed framework Jozefowiez’s method the proposed framework

Best Found Balance Time Best Found Balance Time Best Found Length Best Found Length

E-n51-k5 524.61 20.07 511.00 550.96 5.46 174.80 0.24 618.22 0.22 682.68
E-n76-k10 835.32 78.10 1269.00 982.73 96.83 529.23 0.59 1203.98 3.93 1299.59
E-n101-k8 826.39 67.55 2014.75 1018.89 92.92 762.97 0.29 1871.06 1.10 1447.72
M-n101-k10 827.39 67.55 1771.50 828.94 76.49 406.63 1.15 1429.90 20.47 1122.84
M-n151-k12 1047.35 74.78 3280.00 1155.68 86.33 1300.09 0.80 1484.48 1.95 1520.91

Average 812.21 61.61 1769.25 907.44 71.61 634.74 0.61 1321.53 5.54 1214.75

8
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

Table 3
The performance comparison of three algorithms on 11 instances.
Instance MMDC+(ACO+) (DBSCAN+)+(ACO+) (DBSCAN+)+(Microcluster Fusion Scheme)+(ACO+)

IGD HV time IGD HV time IGD HV time

E-n51-k5 0.1241 0.3032 127.8800 0.2181 0.5872 119.5680 0.0018 0.9927 265.3320
E-n76-k10 0.2449 0.4089 108.9960 0.1838 0.4803 169.8790 0.0617 0.6598 365.7100
M-n101-k10 0.0063 0.7750 167.6530 0.3886 0.3947 192.1350 0.0030 0.8043 526.1840
M-n151-k12 0.7831 0.1094 255.0890 0.6708 0.1718 289.3970 0.0464 0.9021 1045.6200
A-n37-k5 0.3161 0.3985 72.0520 0.2265 0.4760 61.3340 0.0565 0.7401 131.1160
A-n46-k7 0.6290 0.1991 74.5500 0.1435 0.6756 72.3520 0.0099 0.8925 108.7350
A-n48-k7 0.2509 0.4225 45.4930 0.5359 0.1294 22.6240 0.0115 0.7525 141.1460
A-n60-k9 0.6030 0.1894 40.7010 0.3073 0.4440 140.1860 0.0236 0.8409 160.3110
A-n62-k8 0.3739 0.3909 87.5610 0.3105 0.2711 102.2760 0.0132 0.4742 231.6450
A-n63-k10 0.5379 0.2091 38.6830 0.2513 0.4782 139.1200 0.0041 0.8002 190.1330
A-n80-k10 0.2483 0.5005 116.7490 0.3911 0.3630 56.2600 0.0347 0.7584 228.8510

Average 0.3743 0.3551 103.2188 0.3298 0.4065 124.1028 0.0242 0.7834 308.6166

Table 4
The performance comparison of three algorithms on 11 revised instances.
Instance MMDC+(ACO+) (DBSCAN+)+(ACO+) (DBSCAN+)+(Microcluster Fusion Scheme)+(ACO+)

IGD HV time IGD HV time IGD HV time

E-n51-k5-240H 0.1917 0.4272 142.2050 0.2662 0.5096 181.6380 0.0196 0.7258 228.7460
E-n76-k10-210H 0.7421 0.1302 172.2320 0.8765 0.0549 191.8310 0.0369 0.8986 395.5400
M-n101-k10-300H 0.0032 0.7456 283.4900 0.0035 0.7389 522.1450 0.0032 0.7496 322.3930
M-n151-k12-300H 0.5251 0.2650 106.7770 1.0325 0.0008 115.0570 0.0250 0.9510 948.8510
A-n37-k5-150H 0.3997 0.3097 36.7440 0.3327 0.3442 83.9810 0.1077 0.6448 119.0910
A-n46-k7-150H 0.3229 0.4001 20.2650 0.3763 0.3050 25.4840 0.0034 0.8412 101.1160
A-n48-k7-150H 0.9328 0.1008 18.3670 0.4156 0.2379 20.6790 0.1151 0.4615 103.4460
A-n60-k9-150H 0.5818 0.2907 61.9960 0.3989 0.5463 28.0040 0.0309 0.9148 256.7880
A-n62-k8-150H 0.5950 0.5003 59.7150 0.3701 0.2537 79.8130 0.0118 0.5034 560.5720
A-n63-k10-150H 0.9329 0.0000 17.5030 0.2419 0.6568 26.3730 0.0172 0.9491 198.9410
A-n80-k10-150H 0.4790 0.3156 65.9410 0.5563 0.3053 76.6250 0.0400 0.8693 313.4630

Average 0.5188 0.3168 89.5668 0.4428 0.3594 122.8755 0.0374 0.7735 322.63155

5.1.2. Effects of algorithms in the components num3H, where num3 indicates the number of expanded vehicle capacity
In our framework, DBSCAN+ is used to group the microclusters and H indicates the previous data is hypothetical. Table 4 provides the
effectively. Microcluster fusion scheme forms the clusters for workload IGD and HV values of the three algorithms on the revised instances. In
balance, which is able to escape from local optimal. For validating the terms of HV and IGD, the third algorithm obtains the best values on all
effects of the algorithms, we compare the performances of algorithm the instances, while the second algorithm obtains better values than the
combinations as follows: MMDC+(ACO+), (DBSCAN+)+(ACO+), first algorithm in 4 instances, and the first algorithm obtains better
(DBSCAN+)+(Microcluster Fusion Scheme)+(ACO+). values than the second algorithm in 5 instances, respectively. We find
As shown in Table 3, the three algorithms are tested by executing 11 that the first and second algorithms perform worse on the revised in­
instances from Christofides and Eilon (1969),hristofides et al. (1979), stances than on the original ones. Further, the superiority of the third
and Augerat et al. (1995). As to the IGD value, the third algorithm is algorithm is more obvious on these revised instances, which indicates it
93.5% and 86% lower than the first and second algorithms, respectively. is more stable in dealing with different instances.
In terms of HV, the corresponding values are 120.6% and 92.7%
improved on the first and second algorithms, respectively. The third 5.1.3. Discussion on pheromone enhancement
algorithm significantly outperforms other algorithms in the instances For ACO+, we try to significantly increase pheromone on the paths of
while its runtime is about 2 and 3 times longer than other algorithms. the following four solutions: (1) extreme solutions on f1 ; (2) extreme
The first algorithm obtains the worst performance in most cases, indi­ solutions on f2 ; (3) the best solutions for optimizing both f1 and f2 ; (4)
cating that DBSCAN+ and microcluster fusion scheme can speed up the Pareto-optimal solutions, to find an optimal way for pheromone
convergence and improve diversity. enhancement. The performance comparison is obtained with the
It is to be noted that, the capacity of all vehicles is always very close different number of iterations on instance E-n101-k8 from Christofides
to the total demand of the customers in the instances proposed by and Eilon (1969), as shown in Fig. 4. We quickly find that the perfor­
Christofides and Eilon (1969),hristofides et al. (1979), and Augerat et al. mance of the third way outperforms other ways in terms of IGD and HV.
(1995). Therefore, the load of each vehicle is almost full before deliv­ Compared with the traditional method of increasing pheromone on
ering. However, in the real world, there is a usual situation where loads paths of all the Pareto-optimal solutions, increasing pheromone on the
of most vehicles are less than their capacities, only a few vehicles are paths of the best solutions for optimizing both f1 and f2 is an effective
fully loaded. According to this situation, we modified the instances method to improve both convergence and diversity.
above to expand the vehicle capacity to 1.5 times the original capacity.
The format of the revised instance is the original name appended by

9
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

Fig. 4. Comparison of IGD and HV with iterations from 10000 to 90000 produced by different ways of pheromone enhancement on instance E-n101-k8.

5.2. Solving MVRPWB

Table 6
The average percent and standard deviation of the proposed framework on 5
MVRPWB assumes microclusters are given, and each microcluster is revised instances.
known to be allocated to a vehicle. We use the cluster fusion component
Instance The optimal solution of the The optimal solution of the
and route search component to solve MVRPWB. In order to evaluate the first objective second objective
algorithms, we revised a set of instances from Augerat et al. (1995). We
Avg. percent Std Dev Avg. percent Std Dev
artificially group the customer nodes into microclusters according to a
rule and the number of vehicles. In this experiment, the rule is the B56-k7-(142) 92.727 8.644 90.909 11.333
location of the nodes, which is realistic in the city. Since the total de­ B57-k9-(173) 96.429 10.475 94.048 12.026
B64-k9-(137) 88.360 16.986 80.952 20.199
mand of each microcluster is not equal after clustering, we regard the B67-k10-(142) 88.887 13.200 84.849 15.410
maximum total demand of the microclusters as the capacity of each B78-k10-(126) 90.909 15.275 80.952 22.013
vehicle. The revised instance is named Xnum4-knum5-(num6), where
Average 91.462 12.916 86.342 16.196
num4 indicates the number of nodes, num5 represents the number of
vehicles, num6 indicates the vehicle capacity, if X is equal to “A” or “B”,
the instances are from Augerat et al. (1995). Based on the microclusters, microclusters to vehicles is fixed. However, the workload of micro­
we apply the proposed framework to find extreme solutions on two clusters may be imbalanced because of the stochastic nature of the
objectives. These results are compared with the results obtained by using problem. To balance the workload of each vehicle, we proposed
the traditional ACO. microcluster fusion, in which only the microclusters with close locations
Solely considering the first objective, from columns 2 and 4 in can be fused. Therefore, the driver does not need to take much time to
Table 5, we observe that our framework outperforms better than tradi­ familiarize himself with customers’ locations that are not in his micro­
tional ACO in 4 instances except for the instance B67-k10-(142) with an cluster. The cost of exchanging the customers is negligible. We analyzed
equal value. Furthermore, the proposed framework is also competitive the number of customer nodes inside the given microcluster and the
in terms of the associated balance (column 5) for all instances, especially number of customer nodes finally served for each vehicle. In Table 6, we
the results in bold. For solely considering the second objective, the calculated the average percentage of the amount of inside customers and
framework could always find better results for all instances. On average, standard deviation from the average percent. When solely considering
the improvement of optimizing workload balance when using the the first objective, column 2 shows that, on average, more than 90
framework amounts to about 13.5% compared to ACO. Besides, the percent of the customer nodes a vehicle served are inside its micro­
computation time of the proposed framework is almost as same as ACO. cluster. In column 3, we also find the highest standard deviation is
Those observations illustrate that our framework provides more satis­ 16.98%. It indicates that, even in the worst case (B64-K9-(137)), more
factory solutions in acceptable computational time. than 60 percent of the nodes served by a vehicle are inside its micro­
In the assumption for MVRPWB, the preferred assignment of cluster. Although a deviation of the solution from the initial assumption

Table 5
The performance comparison between ACO and the proposed framework on 5 revised instances.
Instance Solely minimizing the total traveling costs Solely minimizing the Workload balance Running time

Traditional ACO the proposed framework Traditional ACO the proposed framework ACO framework

Best Found Balance Best Found Balance Best Found length Best Found length

B56-k7-(116) 650.748 2.800 646.382 2.841 2.520 672.850 2.194 706.012 105.451 136.308
B57-k9-(173) 1550.250 2.169 1540.690 2.169 1.919 1584.930 1.890 1611.270 95.519 91.492
B64-k9-(136) 805.773 1.641 804.744 1.633 1.541 821.561 1.191 929.908 68.438 153.799
B67-k10-(142) 1047.610 2.678 1047.610 2.678 2.421 1068.300 2.025 1135.100 104.680 173.832
B78-k10-(121) 1211.090 2.169 1199.890 2.307 1.852 1277.020 1.570 1395.450 162.042 223.518

Average 1053.094 2.291 1047.863 2.326 2.050 1084.932 1.774 1155.548 107.226 155.790

10
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

is inevitable, when solely minimizing the workload balance, the average Table 7
percent and standard deviation are 86.34% and 16.19%, respectively, The details of the most balanced solution.
which are acceptable. The results show that the microcluster fusion Vehicle Total traveling distance # of completed demands Total workload
scheme respects the initial assumption and does not add the familiarity
v1 30212.400 5976.000 51128.400
cost of the vehicle. There are two reasons behind the results. The first v2 29707.300 6048.000 50875.300
one is that the upper bound of the number of microclusters to be fused is v3 28003.700 6480.000 50683.700
adaptively defined, avoiding too many microclusters are grouped v4 32197.900 5400.000 51097.900
together. The other reason is that different microclusters are fused ac­ v5 26994.600 6876.000 51060.600

cording to the location and the number of needed customer nodes.

Thereby the customer nodes far away from each other or leading to depart from the entrance and return to the entrance.
imbalance are ruled out. In this case, we set the weight parameters α and β as 1 and 3.5. By
applying our framework, we obtain a set of 43 non-dominated solutions.
6. Practical case studies The two extreme solutions are (114647,1.13145) and
(147116,1.00877). It can be observed that even in the solutions for
We study a real case from a large logistics company in Guangzhou, solely minimizing total traveling costs, the value of workload balance
China. The company daily ships the goods from a distribution center to objective f2 is 1.13145. As to the most balanced solutions, the vehicles
regional service points that can be regarded as the “first mile” of the only travel 147116 meters and the value of f2 reaches 1.00877 indi­
transportation chain. Then, the service points deliver the goods directly cating the workload of each vehicle is almost equal. In addition, the
to the customers, which is the “last mile” of the transportation chain. running time is 287.011s which demonstrates the framework is superior
The two shipping cases can correspond to VRPWB and MVRPWB in our in terms of time efficiency. Table 7 illustrates the details of the most
model respectively. In this section, we apply the proposed framework to balanced solution. The workload of the vehicles is within a small range
the cases faced by the logistics company. which is [50683.7, 51128.4]. It further shows that the solution gua­
rantees the fairness of the vehicles.
6.1. First mile delivery
6.2. Last mile delivery
We have collected a large number of delivery tasks in a month from a
central urban district (called Tianhe) of Guangzhou. There are 57 In the last mile delivery, the service points located at the commu­
regional service points in the district, and all packages to be distributed nities employ the vehicles to send the packages to the customers’ homes.
to these locations are from the distribution center named “Xintang We study a real case of a service point in South China Normal University
Distribution Center”. We randomly selected the delivery tasks of one day (SCNU), Tianhe District, Guangzhou. SCNU has five major regions,
to study. Each package can be described simply by its address and de­ namely, Central Area, Western Area, South Area, North Area, and Res­
mand. The packages are to 57 service points, and the average demand idential Area.
for these locations is 540 units. There are 5 large trucks, each one with a We collected a large number of historical delivery tasks in one month
capacity of 7200 units, depart from the Xintang distribution center. In and selected one day’s tasks at random. Fig. 6 shows the customer lo­
Fig. 5, the locations needed to be delivered in the Tianhe District are cations of this day, where each colored area corresponds to a region of
colored in blue. The Tianhe District is circled by a gray line, while the SCNU. Since the customer’s address may be incomplete or some cus­
red house icon marks Xintang center. Xintang center is in a suburb, tomers have special requirements, the drivers are usually allocated to
which is far from Tianhe District. Moreover, from the Xintang center to different areas in advance according to their familiarity with the area or
Tianhe District, only a path (marked by a red line) allows the truck to service requirements. Herein, there are 5 vehicles and each one is
pass. Hence, trucks from the Xintang center must pass through a unique responsible for a region, as well as 67 customers are grouped into the 5
entrance when entering Tianhe District. For improving the calculation different regions. The red house icon indicates the location of the service
efficiency and accuracy, we regard the entrance of the district (marked point. The average demand for these customers is 10 units, and the ca­
as an asterisk) as the location of the distribution center. All the trucks pacity of each vehicle is 200 units.

Fig. 5. Geographical location of Xintang distribution center and regional service points in Tianhe District.

11
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

Fig. 6. The distribution of the customers in SCNU.

by ACO. Moreover, the distribution of solutions generated by the

framework spreads wider. The running time of our framework is
373.634s which is similar to that of ACO (292.765s).
We further conduct the balanced solution. Table 8 shows that the
workload of vehicles v2 and v4 is improved in our framework compared
with ACO. The workload of vehicles v1 and v3 is reduced accordingly.
Hence, the framework reaches the workload balance of the vehicles. Due
to the microcluster fusion scheme, the vehicle may be required to serve
the customers outside its initial assigned area resulting in a higher cost
of the driver, e.g., the driver needs to become familiar with a new
location. The last column in Table 8 investigates the number of those
additional customers each driver needs to serve. It can be seen that the
total number of additional customers is only 14. On average, each
vehicle spends more than 80% of its total time inside the given area.

7. Conclusion

This paper proposed an optimization framework for solving vehicle

problems based upon two scenarios. It is very promising for logistics
Fig. 7. The approximate Pareto-optimal set containing 114 solutions produced companies in practice. The reasons are twofold: first, the framework can
by the proposed framework and 250 solutions produced by ACO for last mile help the company manager assign the delivery tasks and paths to drivers
delivery case. quickly for two stages of the logistics chain. Secondly, the company can
utilize the framework to reduce transportation costs and keep fairness
We apply our framework, which combines the cluster fusion among drivers.
component with the route search component, to solve this case. We also The proposed algorithms are to find the paths that can minimize total
use ACO to find the optimal route for each vehicle in its given area for traveling cost and balance the workload. Unlike the existing works, we
comparison. In Fig. 7, it is clearly illustrated that all the solutions pro­ characterized the traveling distance and the load for each vehicle, which
duced by the proposed framework are much better than those produced is a more realistic measurement of workload. In particular, to avoid the

Table 8
The comparison between the most balanced solution obtained by the proposed framework and the solution generated by ACO.
Vehicle Solution from ACO Balanced solution produced by the proposed framework

Total traveling # of completed Total Total traveling # of completed Total # of customers outside the given
distance demand workload distance demand workload area

v1 2065.820 156.000 4405.820 2620.140 114.000 4330.140 0.000

v2 2324.050 66.000 3314.050 2674.930 108.000 4294.930 5.000
v3 2110.210 182.000 4840.210 2059.570 150.000 4309.570 2.000
v4 1921.410 92.000 3301.410 2526.850 124.000 4386.850 7.000
v5 2170.320 174.000 4780.320 2170.320 174.000 4780.320 0.000

Total 10591.810 670.000 20641.810 12051.810 670.000 22101.810 14.000

12
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

influence of the orders of magnitude sensitivity, we introduced the ratio References

between maximum workload and minimum workload. To achieve
workload equity, we try to minimize the ratio. Many papers only focus Augerat, P., Belenguer, J. M., Benavent, E., Corberan, A., & Rinaldi, G. (1995).
Computational results with a branch and cut code for the capacitated vehicle routing
on solving a vehicle routing problem. However, in this paper, we problem. Rapport de recherche - IMAG, 34, 949–979.
formulated the two problems: VRPWB and MVRPWB, which correspond Barán, B., & Schaerer, M. (2003). A multiobjective ant colony system for vehicle routing
to two stages of the logistics chain. Our optimization framework in­ problem with time windows. In The 21st IASTED International Conference. Anaheim,
CA, USA.
cludes three components embedding different algorithms. It can solve Bertazzi, L., Golden, B., & Wang, X. (2015). Min–max vs. min–sum vehicle routing: A
the two problems by combining different components. In the clustering worst-case analysis. European Journal of Operational Research, 240, 372–381.
component, we developed DBSCAN+ and a noise processing method, Chiang, W. C., & Cheng, C. Y. (2017). Considering the performance bonus balance in the
vehicle routing problem with soft time windows. Procedia Manufacturing, 11,
while in the microcluster fusion component, we designed a microcluster 2156–2163.
fusion scheme and another noise processing method. We adapted ACO+ Christofides, N., & Eilon, S. (1969). An algorithm for the vehicle-dispatching problem.
by constructing the return mechanism and novel pheromones updating Journal of the Operational Research Society, 20, 309–318.
Christofides, N., Mingozzi, A., & Toth, P. (1979). The vehicle routing problem.
rules in the route search component.
Combinatorial Optimization, 40, 315–338.
We demonstrated the effectiveness of our algorithms through Doerner, K., Gutjahr, W. J., Hartl, R. F., Strauss, C., & Stummer, C. (2004). Pareto ant
extensive experiments and two case studies. In the first experiment, we colony optimization: A metaheuristic approach to multiobjective portfolio selection.
revised the proposed framework to solve VRPRB. Compared with Annals of Operations Research, 131, 79–99.
Ester, M., Kriegel, H.-P., Sander, J., Xu, X., et al. (1996). A density-based algorithm for
Jozefowiez et al. (2009), the results showed that the framework could discovering clusters in large spatial databases with noise. Proceedings of the Second
take less time to produce good quality solutions. This is because that International Conference on Knowledge Discovery and Data Mining, 96, 226–231.
DBSCAN+ preprocessed the customer nodes, and the microcluster Expósito-Izquierdo, C., Rossi, A., & Sevaux, M. (2016). A two-level solution approach to
solve the clustered capacitated vehicle routing problem. Computers & Industrial
fusion scheme avoided falling into a local optimal solution. In the second Engineering, 91, 274–289.
and third experiments, we compared the performances of different al­ Fallah, H., Didehvar, F., & Rahmati, F. (2020). Approximation algorithms for the load-
gorithm combinations for solving VRPWB on traditional and revised balanced capacitated vehicle routing problem. Bulletin of the Iranian Mathematical
Society, 1–28.
instances. The comparison indicated that the proposed components Janssens, J., Van den Bergh, J., Sörensen, K., & Cattrysse, D. (2015). Multi-objective
could enhance the performance and perform better than other algo­ microzone-based vehicle routing for courier companies: From tactical to operational
rithms in terms of IGD and HV. The microcluster fusion scheme planning. European Journal of Operational Research, 242, 222–231.
Jozefowiez, N., Semet, F., & Talbi, E.-G. (2009). An evolutionary algorithm for the
increased the diversity of solutions, and ACO+ intelligently balanced the vehicle routing problem with route balancing. European Journal of Operational
workload of each vehicle in the cluster. We used the fourth experiment Research, 195, 761–769.
to compare the performance of the proposed framework with ACO on Lacomme, P., Prins, C., Prodhon, C., & Ren, L. (2015). A multi-start split based path
relinking (msspr) approach for the vehicle routing problem with route balancing.
MVRPWB. The results showed that our framework is superior to ACO
Engineering Applications of Artificial Intelligence, 38, 237–251.
due to the cooperation between the microcluster fusion scheme and Lacomme, P., Prins, C., & Ramdane-Cherif, W. (2004). Competitive memetic algorithms
ACO+. Furthermore, the fifth experiment demonstrated that the for arc routing problems. Annals of Operations Research, 131, 159–185.
microcluster fusion scheme would not increase the familiarity cost of Li, X., Zhang, P., & Zhu, G. (2019). Dbscan clustering algorithms for non-uniform density
data and its application in urban rail passenger aggregation distribution. Energies, 12.
vehicles. We applied the framework for solving two practical logistics https://doi.org/10.3390/en12193722
cases. The observations indicated that compared with ACO, the total Londono, J., Rendon, R., & Ocampo, E. (2021). Iterated local search multi-objective
traveling cost of our solution is smaller, and the workload differences of methodology for the green vehicle routing problem considering workload equity
with a private fleet and a common carrier. International Journal of Industrial
all vehicles are within a narrow range. Engineering Computations, 12, 115–130.
Finally, we highlight several potential extensions of the proposed Lu, H., Zhang, X., & Yang, S. (2019). A learning-based iterative method for solving
algorithms and framework. First, the return mechanism in ACO+ could vehicle routing problems. In International Conference on Learning Representations. New
Orleans, United States.
be revised to achieve equity in other issues, for example, performance, Mancini, S., Gansterer, M., & Hartl, R. F. (2021). The collaborative consistent vehicle
load, and fuel consumption. Moreover, for solving a large-scale instance, routing problem with workload balance. European Journal of Operational Research,
such as the number of nodes is more than 1000 and the number of ve­ 293, 955–965.
Mandal, S. K., Pacciarelli, D., Løkketangen, A., & Hasle, G. (2015). A memetic nsga-ii for
hicles is more than 50, the microcluster fusion scheme could be changed the bi-objective mixed capacitated general routing problem. Journal of Heuristics, 21,
to evaluate each microcluster instead of each node. In doing so, the 359–390.
effectiveness of clustering might be reduced, but the computation Marinaki, M., & Marinakis, Y. (2016). A glowworm swarm optimization algorithm for the
vehicle routing problem with stochastic demands. Expert Systems with Applications,
complexity is significantly reduced. In addition, the proposed frame­
46, 145–163.
work can be focused on becoming a general framework for solving VRPs. Mazzeo, S., & Loiseau, I. (2004). An ant colony algorithm for the capacitated vehicle
Our framework can be extended to other VRPs with various constraints routing. Electronic Notes in Discrete Mathematics, 18, 181–186.
and objectives, such as time windows, different types of vehicles, and Miller, C. E., Tucker, A. W., & Zemlin, R. A. (1960). Integer programming formulation of
traveling salesman problems. Journal of the ACM (JACM), 7, 326–329.
multiple depots by adding and combining different components. Min, J., Jin, C., & Lu, L. (2019). Maximum-minimum distance clustering method for split-
delivery vehicle-routing problem: Case studies and performance comparisons.
CRediT authorship contribution statement Advances in Production Engineering & Management, 14, 125–135.
Mokhtarinejad, M., Ahmadi, A., Karimi, B., & Rahmati, S. H. A. (2015). A novel learning
based approach for a new integrated location-routing and scheduling problem within
Jingjing LI: Conceptualization, Methodology, Validation. Yaohui­ cross-docking considering direct shipment. Applied Soft Computing, 34, 274–285.
qiong Fang: Software, Writing - original draft, Writing - review & Mutar, M., Burhanuddin, M., Hameed, A., Yusof, N., & Mutashar, H. (2020). An efficient
improvement of ant colony system algorithm for handling capacity vehicle routing
editing. Na Tang: Writing - review & editing. problem. International Journal of Industrial Engineering Computations, 11, 549–564.
Nguyen, V. A., Jiang, J., Ng, K. M., & Teo, K. M. (2016). Satisficing measure approach for
Acknowledgments vehicle routing problem with time windows under uncertainty. European Journal of
Operational Research, 248, 404–414.
Perboli, G., Brotcorne, L., Bruni, M. E., & Rosano, M. (2021). A new model for last-mile
This work was supported by the Key Project of Science and Tech­ delivery and satellite depots management: The impact of the on-demand economy.
nology Innovation 2030 supported by the Ministry of Science and Transportation Research Part E: Logistics and Transportation Review, 145. https://doi.
org/10.1016/j.tre.2020.102184
Technology of China (Grant No. 2018AAA0101300).
Sivaramkumar, V., Thansekhar, M., Saravanan, R., & Amali, S. M. J. (2018).
Demonstrating the importance of using total time balance instead of route balance
on a multi-objective vehicle routing problem with time windows. The International
Journal of Advanced Manufacturing Technology, 98, 1287–1306.
Sun, Y., Liang, Y., Zhang, Z., & Wang, J. (2017). M-nsga-ii: A memetic algorithm for
vehicle routing problem with route balancing. In International Conference on

13
J. LI et al. Computers & Industrial Engineering 169 (2022) 108221

Industrial Engineering and Other Applications of Applied Intelligent Systems. Arras, Wang, J., Zhou, Y., Wang, Y., Zhang, J., Chen, C. P., & Zheng, Z. (2015). Multiobjective
France. Springer. vehicle routing problems with simultaneous delivery and pickup and time windows:
Ticha, H. B., Absi, N., Feillet, D., & Quilliot, A. (2019). Multigraph modeling and formulation, instances, and algorithms. IEEE Transactions on Cybernetics, 46,
adaptive large neighborhood search for the vehicle routing problem with time 582–594.
windows. Computers & Operations Research, 104, 113–126. Zhang, Z., Qin, H., & Li, Y. (2019a). Multi-objective optimization for the vehicle routing
Wan, W., Wang, C., & Yu, B. (2014). Improved ant colony algorithm for period vehicle problem with outsourcing and profit balancing. IEEE Transactions on Intelligent
routing problem. Operations Research and Management Science, 70–77. Transportation Systems, 21, 1987–2001.
Wang, J., Weng, T., & Zhang, Q. (2018). A two-stage multiobjective evolutionary Zhang, Z., Sun, Y., Xie, H., Teng, Y., & Wang, J. (2019b). Gmma: Gpu-based
algorithm for multiobjective multidepot vehicle routing problem with time multiobjective memetic algorithms for vehicle routing problem with route
windows. IEEE Transactions on Cybernetics, 49, 2467–2478. balancing. Applied Intelligence, 49, 63–78.

14