International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Level-based multi-objective particle swarm

optimizer for integrated production scheduling
and vehicle routing decision with inventory
holding, delivery, and tardiness costs

Jianyu Long, Panos M. Pardalos & Chuan Li

To cite this article: Jianyu Long, Panos M. Pardalos & Chuan Li (2022) Level-based multi-
objective particle swarm optimizer for integrated production scheduling and vehicle routing
decision with inventory holding, delivery, and tardiness costs, International Journal of
Production Research, 60:11, 3319-3338, DOI: 10.1080/00207543.2021.1919780

To link to this article: https://doi.org/10.1080/00207543.2021.1919780

Published online: 28 Apr 2021.

Submit your article to this journal

Article views: 395

View related articles

View Crossmark data

Citing articles: 7 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
2022, VOL. 60, NO. 11, 3319–3338
https://doi.org/10.1080/00207543.2021.1919780

Level-based multi-objective particle swarm optimizer for integrated production

scheduling and vehicle routing decision with inventory holding, delivery, and
tardiness costs
Jianyu Longa , Panos M. Pardalosb and Chuan Lia
a School of Mechanical Engineering, Dongguan University of Technology, Dongguan, People’s Republic of China; b Center for Applied
Optimization, Department of Industrial and Systems Engineering, University of Florida, Gainesville, FL, USA

ABSTRACT ARTICLE HISTORY

Integrated optimisation of production scheduling and distribution decision is necessary for reducing Received 16 September 2020
the whole cost of the supply chain in the make-to-order business environment. This paper studies a Accepted 8 April 2021
new integrated production scheduling and vehicle routing problem (IPSVRP) with inventory holding, KEYWORDS
delivery, and tardiness costs. The considered IPSVRP is modelled as a triple-objective optimisation Metaheuristics; integrated
problem, where the first objective aims to obtain the minimal total holding cost in the inventory, production-distribution;
the second one attempts to achieve the minimal total travelling cost, and the third one tries to multi-objective optimisation;
acquire the minimal total tardiness cost. To obtain a set of diverse non-dominated solutions in the particle swarm optimisation;
Pareto-optimal front of the problem, we first derive several key structural properties used to pro- level-based learning
vide necessary conditions for any solution to be Pareto-optimal through theoretical investigation.
Based on the derived structural properties, a level-based multi-objective particle swarm optimizer
(LMPSO) is subsequently designed. The performance of LMPSO is analysed by conducting a set of
experiments, and its superiority is verified through comparing with other optimisation algorithms.
Moreover, the convergence behaviour of LMPSO is also investigated, and the experimental results
prove that it has the ability to achieve a set of non-dominated solutions proximity to the true Pareto
front.

1. Introduction
selection (Hassanzadeh, Rasti-Barzoki, and Khosroshahi
Production scheduling and distribution decisions are two 2016), and pre-designed customer visiting sequence
essential supply chain functions for a manufacturing (Armstrong, Gao, and Lei 2008). In practice, the dis-
enterprise. Since the distribution of jobs (or goods) can tribution decision can usually be regarded as a classical
only start after their production has been completed, tra- vehicle routing problem (VRP), where jobs require to be
ditional optimisation of production scheduling and dis- transported to different locations (or customers) using
tribution decision is performed separately and sequen- vehicles so as to optimise one or more objectives. There-
tially. This uncoordinated optimisation approach is not fore, this study focuses on the integrated production
conducive to reduce the whole cost of the supply chain, scheduling and vehicle routing problem (IPSVRP).
especially in the make-to-order (MTO) business environ- Recently, growing attention has been attracted by both
ment. To deal with this problem, it is necessary to make researchers and engineers to develop approaches for solv-
an integrated optimisation of production scheduling and ing IPSVRP. In 2017, Moons et al. (2017) reviewed and
distribution decision from the global perspective (Chen classified the existing studies on IPSVRP according to
2010). their relevant production, inventory, and distribution
Although the integrated decision-making can achieve characteristics. A remarkable observation of this review
global optimisation, it also brings great challenges to is that very few studies explicitly consider inventories and
problem solving. To make the problem easy to solve, a inventory holding costs. Therefore, the authors appealed
large part of studies in this field focus on several simple that taking the inventory characteristics into account is a
delivery strategies, such as direct shipment without dis- promising research direction as it can influence the total
tribution decision (Soukhal, Oulamara, and Martineau incurred costs. After 2017, many state-of-the-art IPSVRP
2005), pre-specified route (Gupta et al. 2012), batch studies shown in Table 1 have been conducted. As we can

CONTACT Chuan Li [email protected] School of Mechanical Engineering, Dongguan University of Technology, Dongguan 523808, People’s
Republic of China

© 2021 Informa UK Limited, trading as Taylor & Francis Group

3320 J. LONG ET AL.

see, Miranda et al. (2018) and Tavares-Neto and Nagano weight combinations are tried (Deb 2001). To overcome
(2019) are the only ones who considered the inventory these problems, the Pareto-based evolutionary strategy
characteristics in their models. However, both in their is recommended to be used because it aims to obtain
studies, the distribution of jobs were made using a sin- a set of diverse non-dominated solutions rather than a
gle vehicle rather than a fleet of vehicles. Therefore, to single solution in the Pareto-optimal front (Jozefowiez,
fill this research gap, our study focuses on an important Semet, and Talbi 2008). The knowledge of such multi-
variation of IPSVRP, where inventory holding are con- ple Pareto-optimal solutions may help a decision-maker
sidered between the production performed on a single to compare and choose a compromised Pareto-optimal
machine and the distribution conducted by a fleet of vehi- solution for implementation. This study is devoted to
cles. Moreover, delivery time windows of jobs are also develop an efficient Pareto-based evolutionary algorithm
taken into account in this study to improve customer sat- to deal with the considered triple-objective (namely the
isfaction as much as possible. The considered IPSVRP minimisations of the total inventory holding cost, the
is a kind of common problem arising from those com- total travelling cost, and the total tardiness cost) IPSVRP.
panies whose products are perishable or time-sensitive Particle swarm optimization (PSO), as an outstand-
goods, such as fresh foods, blood products, ready-mixed ing evolutionary algorithm introduced by Eberhart and
concrete, etc. Kennedy in 1995 (1995; Kennedy and Eberhart 1995),
In general, IPSVRP is a multi-objective optimisation has been widely researched and applied both in the con-
problem because the costs of production and distribu- tinuous and discrete optimisation fields. To avoid stag-
tion need to be optimised simultaneously. As shown in nation or premature convergence of the classical PSO in
Table 1, most studies that considered two or three opti- dealing with complicated problems, various PSO vari-
misation objectives used the weighted-sum strategy to ants have been proposed by adopting new learning or
aggregate them into a single objective. The advantage of updating strategies, such as, the incremental social learn-
using weighted-sum strategy is that it can directly employ ing strategy (Montes de Oca et al. 2011), the inter-
the excellent existing single-objective optimisation algo- swarm interactive learning strategy (Qin et al. 2016),
rithms for problem-solving. However, the outcome of and the orthogonal learning strategy (Zhan et al. 2011).
such a multi-objective strategy greatly depends on the Although these algorithms can achieve better results
user-defined weight of each objective. Sometimes, it is than the original PSO, their performance still deteriorate
difficult to obtain sufficient prior knowledge to determine drastically when encountering high-dimensional prob-
the optimal weight set in practice. Moreover, when deal- lems (Mandavi, Shiri, and Rahnamayan 2015). In general,
ing with a multi-objective optimisation problem with a researchers try to design effective PSOs to overcome ‘the
non-convex objective space, some Pareto-optimal solu- curse of dimensionality’ from two different perspectives:
tions on special positions cannot be obtained even if all (1) combining the cooperative coevolution framework,

Table 1. Overview of the related IPSVRP studies after 2017.

Production Fleet of Delivery time Number of
Author(s) environment Inventory vehicles restrictions objectives Multi-objective strategy
√ √
Lacomme et al. (2018) Single machine — One −
√ √
Miranda, Morabito, and Ferreira One production line — Three Weighted-sum strategy
(2018) √
Zou et al. (2018) Single machine — — One −
√ √
Gharaei and Jolai (2018) Parallel factories — Two Pareto-based evolutionary
√ √ strategy
Dayarian and Desaulniers Single machine — Two Weighted-sum strategy
(2019) √
Feng and Xu (2019) Parallel machines — √ —
√ Two Weighted-sum strategy
Marandi and Ghomi (2019) Parallel factories — √ √ Two Weighted-sum strategy
Tamannaei and Rasti-Barzoki Single machine — Three Weighted-sum strategy
(2019) √
Tavares-Neto and Nagano Parallel machines — — One −
(2019) √ √
Wang et al. (2019) Parallel machines — √ √ Two Weighted-sum strategy
Wang et al. (2019) Single machine — √ Two Weighted-sum strategy
Wang et al. (2020) Hybrid flow shop — — One −
√
Liu et al. (2020) Single machine — — One −
√ √
Mohammadi, Al-e-Hashem, Flexible job shop — Two Pareto-based evolutionary
and Rekik (2020) √ √ √ strategy
Our study Single machine Three Pareto-based evolutionary
strategy
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3321

whose core idea is to divide all the decision variables into In Section 4, the details of the proposed LMPSO are pre-
several groups and evolve them separately (Li and Yao sented. Next, the experimental results are discussed in
2012); (2) proposing novel updating strategies, whose Section 5. In the end, Section 6 gives the conclusions of
core idea is to evolve all variables in a swarm and main- this study as well as future works.
tain high diversity in the swarm (Cheng and Jin 2015;
Cheng and Jin 2015). Recently, inspired from the mixed-
2. Problem identification and mathematical
level learning methodology in pedagogy, a level-based
modelling
learning swarm optimizer (LLSO) is proposed to deal
with large-scale optimisation (Yang et al. 2018). LLSO Given one machine and n jobs, the goal of the IPSVRP
considers particles in the swarm as students with different considered in this paper is to scheduling the produc-
levels and teach them differently according to their cog- tion process of the n jobs on the machine and deter-
nitive and learning abilities. LLSO is designed to solve the mine transportation routes and visiting time decisions
single-objective optimisation problems, and its competi- for delivering them to their individual location. Note
tive performance has been demonstrated in both solution that some jobs may belong to the same customer in real
quality and computational efficiency. world. Our model can still handle this case by placing
Unlike the single-objective optimisation that only these jobs in the same location. In order to model this
needs to find a global optimal solution, multi-objective problem, we define a complete graph G = (V,E), where
optimisation aims to search a set of solutions as close V = { 0,1,2, . . . ,n} represents the node set and E repre-
as possible to the Pareto-optimal front, while ensur- sents the arc set connecting any two different nodes in V.
ing that the non-dominated solutions in the set are as Node 0 denotes the machine and the other nodes denote
diverse as possible (Long et al. 2019). Based on the the jobs. Each arc in E is linked with a travelling time tij ,
level-based learning idea of LLSO, a level-based multi- which is proportional to the Euclidean distance between
objective particle swarm optimizer (LMPSO) is designed node (job) i and node (job) j. Therefore, for any two jobs
to deal with the considered multi-objective IPSVRP in in the same location, tij = 0. With respect to the produc-
this paper. There are three main motivations for adopt- tion process, at most one job can be processed on the
ing this idea: (1) due to the domination relation between machine at any point of time and preemption is forbid-
solutions in multi-objective optimisation, solutions in a den. Moreover, there are no processing setup times and
swarm can naturally form different non-domination lev- job release dates. The processing time of job i is denoted
els. Therefore, different from LLSO, no additional param- as pi . After a job is completed on the machine, it will
eters are needed to divide a swarm into multiple levels; be kept in the inventory nearby the machine until it is
(2) in the evolution process, particles belonging to dif- carried away by a vehicle. Each job i in the inventory
ferent non-domination levels have different exploration will incur holding cost and its unit holding cost is αi .
and exploitation abilities, which should be treated dif- At the depot (machine), at most m homogeneous vehi-
ferently as well; (3) the IPSVRP is a complicated opti- cles with the same capacity Q can be hired for delivering
misation problem with high-dimensional decision space the completed jobs. The cost for hiring a vehicle is δ.
because production scheduling, vehicle assignment, vis- Since different jobs will occupy different capacity sizes
iting sequencing and visiting time decision require to be in a vehicle (oi is used to represent the size of occupied
optimised simultaneously. Moreover, to further improve capacity for delivering job i), it should be guaranteed that
the performance of the LMPSO, we derive several key the total amount of capacity of jobs transported by each
structural properties used to provide necessary condi- hired vehicle cannot exceed its capacity Q. A due win-
tions for any solution to be Pareto-optimal through theo- dow [die ,dil ] is associated with each job i. The vehicle has
retical investigation, and incorporate them in the search- to wait if arriving at the location of job i before its earliest
ing process. due date die . However, a tardiness cost will be incurred if
The main contributions of this study are summarised the vehicle arrives after the latest due date dil and its unit
as: (1) a novel triple-objective IPSVRP variant is stud- tardiness cost is βi .
ied to derive a trade-off among production, inventory, Therefore, in this problem, production scheduling,
and distribution; and (2) an effective Pareto-based evo- vehicle assignment, visiting sequencing and visiting time
lutionary algorithm (LMPSO) is proposed to deal with decision need to be determined and optimised simulta-
the triple-objective IPSVRP. The remainder of this paper neously. Three optimisation objectives are considered as
is as below. The problem identification and mathemati- follows: (1) the total holding cost of all the jobs in the
cal modelling are described in Section 2. The derivation inventory should be minimised; (2) the total transporta-
of some structural properties is introduced in Section 3. tion cost including the sum of transportation time of all
3322 J. LONG ET AL.

hired vehicles and the sum of vehicle hiring cost should 

n 
m
be minimised; (3) the sum of tardiness costs should be z0j = hv (10)
minimised. j=1 v=1
As we analysed in the literature review of introduc- 
n 
m
tion, the triple-objective IPSVRP we considered here is zi0 = hv (11)
different from those in other studies. Therefore, a math- i=1 v=1
ematical model is built in this section to further describe si + tij − sj ≤ M(1 − zij ), i ∈ {1, 2, . . . , n},
this problem more comprehensively. To formulate the
j ∈ {1, 2, . . . , n}, i = j (12)
mathematical model, another parameter M is defined to
represent a sufficiently large positive integer. In addition, 
n

some decision variables are given as follows: yiv oi − (1 − hv )M ≤ Q, v ∈ {1, 2, . . . , m} (13)

i=1
ci : completion time of job i;
si : the time to visit the location of job i(it is also the 
n

delivery date of job i); yiv − hv M ≤ 0, v ∈ {1, 2, . . . , m} (14)

i=1
bv : the departure time of vehicle v;
xij : 1, if job i is processed before job j on the machine; 
m
yiv = 1, i ∈ {1, 2, . . . , n} (15)
0, otherwise;
v=1
yiv : 1, if job i is transported by vehicle v; 0, otherwise;
zij : 1, if there is a visiting path goes from node i to node si ≥ die , i ∈ {1, 2, . . . , n} (16)
j; 0, otherwise; ci ≥ pi , i ∈ {1, 2, . . . , n} (17)
hv : 1, if vehicle v is used in a solution; 0, otherwise. +
ci , si , bv ∈ R , i ∈ {1, 2, . . . , n}, v ∈ {1, 2, . . . , m} (18)
On the basis of the above-mentioned parameters and
decision variables, the mathematical model is formulated xij , yiv , hv ∈ {0, 1}, i ∈ {1, 2, . . . , n}, j ∈ {1, 2, . . . , n},
as below: i = j, v ∈ {1, 2, . . . , m} (19)

n 
m zij ∈ {0, 1}, i ∈ {0, 1, 2, . . . , n}, j ∈ {0, 1, 2, . . . , n}, i = j
min αi yiv (bv − ci ) (1) (20)
i=1 v=1
  

m 
n
The objective function (1) is used to minimise the
min hv yiv zi0 (si + ti0 ) − bv + δ (2)
total
n holding cost of all the jobs in the inventory. Since
v=1 i=1
i=1 yiv zi0 (s
i + ti0 ) is the time of vehicle v to return to

n
the depot, ni=1 yiv zi0 (si + ti0 ) − bv is the total travel-
min βi max{0, si − dil } (3)
i=1
ling time of vehicle v. Therefore, the objective function
(2) minimises the total travelling cost which includes,
subject to respectively, the sum of travelling time of all hired vehi-
cles and the sum of vehicle hiring cost. The objective
ci − cj ≥ pi − Mxij , i ∈ {1, 2, . . . , n}, function (3) minimises the total tardiness cost. Con-
straints (4) and (5) ensure that for two jobs processed
j ∈ {1, 2, . . . , n}, i = j (4)
on the machine, only when the processing of the preced-
cj − ci ≥ pj − M(1 − xij ), i ∈ {1, 2, . . . , n}, ing job has been completed can the next one be started.
j ∈ {1, 2, . . . , n}, i = j (5) Constraints (6) guarantee that only when all the process-
ing of the jobs assigned to a vehicle have been completed
bv ≥ ci − M(1 − yiv ), i ∈ {1, 2, . . . , n}, can the vehicle departure. Constraints (7) make sure the
v ∈ {1, 2, . . . , m} (6) feasibility of the time connectivity between the vehicle
bv ≤ si − t0i + M(1 − yiv ), i ∈ {1, 2, . . . , n}, departure time and the visiting time of each job assigned
to this vehicle. Constraints (8)-(9) guarantee that each
v ∈ {1, 2, . . . , m} (7) job can be visited at most once. Constraints (10)–(11)

n require that each hired vehicle must start from the depot
zij = 1, j ∈ {1, 2, . . . , n} (8) and return to the depot. Constraints (12) make sure the
i=0 & i=j feasibility of the time connectivity between jobs visited

n by each vehicle, and eliminate possible sub-tours. Con-
zij = 1, i ∈ {1, 2, . . . , n} (9) straints (13) guarantee that the total amount of capacity
j=0& i=j of jobs transported by a vehicle cannot exceed the vehicle
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3323

capacity. Constraints (14) require that no job is assigned decrease the tardiness costs, it may increase the travelling
to a non-hired vehicle. Constraints (15) guarantee that times of vehicles because the sign of the bv in objective
each job must be transported by exactly one vehicle. Con- (2) is negative. In other words, postponing the vehicle
straints (16) ensure that each job must be delivered after departure times may improve the quality of the solution
its earliest due date. Constraints (17) guarantee that the in terms of minimising the total travelling cost. There-
completion time of each job is greater than or equal to fore, in a Pareto-optimal solution, the completed jobs
its processing time. Constraints (18)–(20) determine the belonging to the same vehicle may be held in the inven-
value range of decision variables. tory for a period of time before being transported to their
destinations. The proof is complete. 
3. Structural properties Proposition 4: Given a solution where the machine is idle
Obtaining a set of diverse non-dominated solutions in the for a time period π after the processing of a job i, if the job
Pareto-optimal front, as we know, is the goal of multi- i is transported by vehicle v and its departure time bv >
objective optimisation. Some key structural properties ci , the processing of job i should be postponed to complete
used to provide necessary conditions for any given solu- at ci + π  , where π  is the maximum value in the interval
tion to be Pareto-optimal for the above model are derived (0, π ] which guarantees the validity of inequality bv ≥ ci +
in this section. These properties are the basis of the π .
proposed LMPSO designed for problem solving.
Proof: Assume that we have two solutions ξ1 and ξ2 . In
Proposition 1: In a Pareto-optimal solution, unforced solution ξ1 , the machine is idle for a time period π after
idleness is not necessarily prohibited on the machine. the processing of a job i and the departure time of its
assigned vehicle bv > ci . Let π  is the maximum value
Proof: From the model formulation, we can find that in the interval (0, π ] which guarantees the validity of
advancing the job completion times does not necessarily inequality bv ≥ ci + π  . In solution ξ2 , the processing of
reduce the objective function values. An obvious proof job i is postponed to complete at ci + π  and the others
is that the sign of the ci in the objective (1) is nega- are the same with solution ξ1 . Obviously, the departure
tive. Therefore, it means that increasing the job com- time of the vehicle that carries job i is not changed. The
pletion times may improve the quality of the solution. only difference between the two solutions is the holding
Since inserting any idle time before processing a job will cost for job i. The cost is αi (bv − ci ) in solution ξ1 , while it
increase the job completion time, a Pareto-optimal solu- is αi (bv − ci − π  ) in solution ξ2 . Therefore, ξ2 dominate
tion is not necessarily a non-delay schedule. The proof is ξ1 . The proof is complete. 
complete. 
Proposition 5: In any Pareto-optimal solution, the
machine is not allowed to be idle after the processing of a
Proposition 2: In a Pareto-optimal solution, the jobs
job unless the job is the last job in a vehicle.
assigned to the same vehicle are not necessarily processed
continuously on the machine. Proof: Based on the Proposition 4, the proof of this
proposition is obtained immediately because, except for
Proof: Assume that we have a solution ξ , where the the last job belonging to a vehicle v, any other job i in the
jobs assigned to the same vehicle are processed continu- vehicle v we necessarily have bv > ci (Note that we may
ously. Since postponing the processing of a job belonging have bv = ci for the last job i). The proof is complete. 
to a vehicle in solution ξ will increase the job comple-
tion time, based on the Proposition 1, the quality of the Proposition 6: Given a solution where job i is processed
solution ξ may be further improved. Therefore, discon- immediately before job j and they are assigned to the same
tinuous processing among the jobs assigned to the same vehicle. Their processing sequences should be interchanged,
vehicle may occur in a Pareto-optimal solution. The proof if αi pj > αj pi .
is complete. 
Proof: Assume that we have two solutions ξ1 and ξ2 . In
Proposition 3: In a Pareto-optimal solution, one vehicle solution ξ1 , job i is processed immediately before job j
does not necessarily have to departure immediately after and they are assigned to the same vehicle v. In solution
the last job assigned to this vehicle is completed. ξ2 , the processing sequence of job i and j is interchanged
and the others are the same with solution ξ1 . According to
Proof: Although advancing the vehicle departure times the Proposition 5, there is no idle time between the pro-
will reduce the inventory costs of jobs and may also cessing of these two jobs. Therefore, the departure time
3324 J. LONG ET AL.

of their assigned vehicle is not changed. The only differ- (i.e. for each job i, si > die ). Suppose that the departure
ence between the two solutions is the holding cost for job of vehicle v does not occur at the completion time of its
i and j. We use BT to represent the start processing time last job (i.e. bv > maxi∈v {ci }, where v is the set of all
of job i in solution ξ1 and the start processing time of job jobs in v). In solution ξ2 , the departure time of vehicle v
j in solution ξ2 . Moreover, HC(ξ1 ) and HC(ξ2 ) are used is advanced to bv − min{bv − maxi∈v {ci }, mini∈v {si −
to represent the holding costs for jobs i and j in ξ1 and ξ2 , die }}, and the others are the same with solution ξ1 . Since
respectively. Then, we have the departure time of vehicle v is advanced in solution
ξ2 , the holding costs for the jobs belonging to vehicle v
HC(ξ1 ) = αi (bv − BT − pi ) + αj (bv − BT − pi − pj ) must be reduced. In addition, the tardiness costs for these
(21) jobs in solution ξ2 must be less than or equal to that of
HC(ξ2 ) = αj (bv − BT − pj ) + αi (bv − BT − pj − pi ) solution ξ1 , and the total travelling cost are the same in
(22) the two solutions. Therefore, ξ2 dominate ξ1 . The proof
is complete. 
Now, the difference of holding costs can be calculated as
HC(ξ1 ) − HC(ξ2 ) = αi pj − αj pi . Therefore, ξ2 dominate
4. Level-based multi-objective particle swarm
ξ1 , if αi pj > αj pi . The proof is complete. 
optimizer
Proposition 7: Given a solution where every job in a vehi- Based on the derived structural properties and the level-
cle v is on time or tardy (i.e. for each job i, si > die ), if based learning idea of LLSO, LMPSO is designed for
the departure of vehicle v does not occur at the completion solving the multi-objective IPSVRP. The entire LMPSO
time of its last job (i.e. bv > maxi∈v {ci }, where v is the framework is described in Algorithm 1, and its key com-
set of all jobs in v), the departure time of vehicle v should ponents are introduced below in detail.
be advanced to bv − min{bv − maxi∈v {ci }, mini∈v {si −
die }}. 4.1. Encoding and decoding of particles

Proof: Assume that we have two solutions ξ1 and ξ2 . In For each particle p, Xp = {xp1 , xp2 , . . . , xpD } and Vp =
solution ξ1 , every job in a vehicle v is on time or tardy {vp1 , vp2 , . . . , vpD } are defined to represent its position

Algorithm 1 Framework of LMPSO

Input: swarm size Ns, control parameter φ used for particle update
Output: archive Ar
1: Initialise a swarm S with Ns particles, where each particle has its position Xp and velocity
Vp ;
2: Decode a complete solution and calculate objective values for each particle in S;
3: Enhance the quality of particles based on the derived structural properties;
4: While the stop criterion is not satisfied do
5: Divide S into various non-dominated levels (denoted as{L1 , L2 , . . . ,LR }) using the fast
non-dominated sorting approach;
6: Calculate particle’s crowding distance (CD) in each non-dominated level;
7: Ar = non-dominated solutions of {L1 ∪Ar };
8: For r = {R, R-1, . . . , 3} do
9: For each particle Xp,r in Lr do
10: Randomly select two levels from the top (r-1) levels, ensuring r1 < r2 < r;
11: Use CD-based roulette wheel selection method to select two particles Xp1,r1 and
Xp2,r2 from Lr1 and Lr2 , respectively;
12: Use the approach proposed in LLSO to update the particle Xp,r based on Xp1,r1
and Xp2,r2 ;
13: Decode a complete solution and calculate objective values for Xp,r ;
14: Enhance the quality of Xp,r based on the derived structural properties;
15: End For
16: End For
17: For each particle Xp,2 in L2 do
18: Use CD-based roulette wheel selection method to select two particles Xp1,1 and Xp2,1
from L1 , ensuring the crowding distance of Xp1,1 is larger than that of Xp2,1 ;
19: Use the approach proposed in LLSO to update the particle Xp,2 based on Xp1,1
and Xp2,1 ;
20: Decode a complete solution and calculate objective values for Xp,2 ;
21: Enhance the quality of Xp,2 based on the derived structural properties;
22: End For
23: Improve the quality of particles in L1 using the local search approach;
24: End While
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3325

vector and velocity vector, respectively. Based on the deci- value (SPV) rule (Tasgetiren et al. 2007) is used to obtain
sion items of the problem and its derived structural prop- the processing sequence and job visiting sequence by
erties, the position vector Xp is designed to include the converting the continuous position values to a discrete
following five subvectors: (1) subvector sv1 of size n (n job permutation. The nearest discrete value rounding
is the number of jobs) is used to determine the process- (NDVR) rule, which rounds the particle position to the
ing sequence of jobs; (2) subvector sv2 of size n is used to nearest discrete value in set {1, 2, . . . , m} , is employed to
allocate the vehicle for each job; (3) subvector sv3 of size achieve the vehicle allocation scheme.
n is used to determine the visiting sequence of jobs; (4) After obtaining a complete solution through decod-
subvector sv4 of size m (m is the number of vehicles) is ing, the objective values of this solution need to be calcu-
used to determine the idle time intervals on the machine. lated. Due to the particle update mechanism and the code
Based on the Proposition 5, after the machine starts to structure, the feasibility of constraint (13) may be violated
process, at most m − 1 idle time intervals may occur on in the searching process. However, allowing exploring the
the machine before the processing processes of all the infeasible solution space is beneficial to the search of evo-
jobs are completed. Therefore, vector sv4 is designed to lutionary algorithms (Vidal et al. 2012). Thus, in addition
contain m values, where the first value refers to the idle to the objective function values calculated according to
time between the decision-making start time and the pro- Equations (1)–(3) for a solution ξ , the constraint viola-
cessing start time of the first job on the machine, and tion penalty can be considered. The three objective values
the other values refer to the following m − 1 possible idle of solution ξ are calculated as below:
time intervals; (5) subvector sv5 of size m is used to deter-
mine the idle time between the departure time of each 
m
Obj1 (ξ ) = f1 (ξ ) + η Pv (ξ ) (23)
vehicle and the completion time of its last job. Therefore,
v=1
the dimension size D is equal to 3n + 2m.
Since all the key information for decision-making 
m
Obj2 (ξ ) = f2 (ξ ) + η Pv (ξ ) (24)
has been determined in the code, decoding a particle
v=1
becomes very easy. Figure 1 shows a simple example for
the encoding and decoding of a particle. In this example, 
m
Obj3 (ξ ) = f3 (ξ ) + η Pv (ξ ) (25)
we have five jobs (jobs 1-5) and two vehicles (vehicles 1-
v=1
2). The processing times of the five jobs are 4, 3, 2, 3 and
2, respectively. The travelling time between any two jobs where f1 (ξ ), f2 (ξ ) and f3 (ξ ) are the objective function
is assumed to 3. The due windows associated with jobs values calculated according to Equations (1)–(3), respec-
1–5 are [22, 23], [25, 26], [23, 24], [16, 17] and [21, 22], tively, η is the penalty coefficient for constraint violation,
respectively. As shown in Figure 1, the smallest position and Pv (ξ ) denotes the overload value of vehicle v. If the

Figure 1. An example for the encoding and decoding of a particle.

3326 J. LONG ET AL.

total amount of capacity of jobs transported by vehicle v is equal to TUB1,1 + 20. Therefore, the value of TUB1,1
to its vehicle capacity Q, Pv (ξ ) = 0,
is less than or equal can be subsequently calculated as 3 by setting TUB1,1 +
otherwise, Pv (ξ ) = ni=1 yiv oi − Q. 20 = 23. In addition, when calculating TUB3,2 of job 3
for the second item of sv5 , g(TUB3,2 ) is first calculated
as 2.5 + 2 + 4 + 2 + 3 + 2 + 3 + TUB3,2 + 3, which is
4.2. Initialisation of the swarm equal to TUB3,2 + 21.5. Therefore, the value of TUB3,2
The LMPSO algorithm begins with a swarm of Ns ini- can be subsequently calculated as 2.5 by setting TUB3,2 +
tial particles, which is randomly generated for searching 21.5 = 24.
more potential outstanding solutions in different parts of
the solution space. Following many existing PSO algo-
4.3. Particle quality enhancement
rithms (Tseng and Liao 2008; Tripathi, Bandyopadhyay,
and Pal 2007), the velocity vector of each particle is ini- In addition to using the derived structural properties
tialised as a zero vector. Therefore, only the position in the encoding process, they can also be employed to
vector needs to be generated for each particle. Which enhance the quality of particles. More specifically, if the
intervals used to randomly generate the position values solution of a particle is tested to satisfy the condition
is important in the initialisation process. Since the SPV defined by a proposition, it should be properly modi-
rule uses the relative size relationship between values to fied. Proposition 4, Proposition 6 and Proposition 7 are
determine the job permutation, we can randomly gener- implemented one by one in this procedure. Note that
ate position values of sv1 and sv3 directly in [−1, 1] using Proposition 4 is better to be applied in backward direc-
uniform distribution. For each item in sv2 , it is uniformly tion for saving computational time, which means that the
randomly generated in [0.5, m+0.5]. Note that the vehi- process starts from the last job in the schedule and goes
cle capacity constraint (13) should be guaranteed in the backward to the first one. As for the implementation of
generation process of sv2 . As for the generation of sub- Proposition 6, if two adjacent jobs assigned to the same
vectors sv4 and sv5 , the most important thing is how to vehicle are interchanged, the new solution also needs to
determine a tight upper bound (TUB) of the value range be checked with Proposition 6.
for each item. Considering that the latest due date infor-
mation of jobs and the information determined in sv1 , sv2
4.4. Non-dominated sorting and crowding distance
and sv3 can be used to determine the TUB, Algorithm 2 is
calculation
proposed to randomly generate sv4 and sv5 .
The calculation method for TUBi,u in line 4 consists As shown in the line 5 of Algorithm 1, the swarm requires
of two steps: (1) calculate the time to visit the location to be divided into different non-dominated levels. A
of job i by supposing the item u in v4 or v5 is TUBi,u . particle is not non-dominated by any other particle in
For ease of description, the calculated time is denoted as the same level, but all the particles in a level are dom-
g(TUBi,u ). Since we want to obtain a tight upper bound inated by each particle in a lower level. As a neces-
for item u, any idle time or waiting time is not allowed on sary step for a population-based multi-objective opti-
the machine or in the visiting journey of the vehicle when misation algorithm, non-dominated sorting is always
calculating g(TUBi,u ); (2) calculate the TUBi,u by setting used to distinguish the quality of solutions in a popu-
g(TUBi,u ) equal to the latest due date dil of job i. For lation according to the concept of domination. Besides
the example shown in Figure 1, when calculating TUB1,1 that, non-dominated sorting also makes it possible to
of job 1 for the first item of sv4 , g(TUB1,1 ) is first cal- apply the level-based learning idea of LLSO for multi-
culated as TUB1,1 + 2 + 4 + 2 + 3 + 3 + 3 + 3, which objective optimisation. In LMPSO, we employ the fast

Algorithm 2 Randomly generation for subvector sv4 (or sv5 )

Input: vector length m
Output: permutation 
1: For u = {1,2, . . . m} do
2: Obtain the job set u that contains all the jobs processed after the uth idle time (or all
the jobs allocated to the vehicle u);
3: For each job i in u do
4: Calculate a candidate solution TUBi,u of the tight upper bound TUBu for the value
of item u in ;
5: End For
6: TUBu = maxi∈ u {TUBi,u }
7: Randomly generate a value from [0, TUBu ] for the item u in ;
8: End For
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3327

Algorithm 3 CD calculation procedure

Input: solution set LS without CD values
Output: solution set LS with CD values
1: Set n = |LS|, where |LS| denotes the number of solutions in set LS;
2: For each solution i in LS do
3: LS[i].CD = 0
4: End For
5: For each objective k do
6: Sort LS according to the kth objective value of each solution;
7: LS[1].CD = LS[1].CD + 1, and LS[n].CD = LS[n].CD + 1;
8: For i = {2, 3, . . . , n − 1} do
9: LS[i].CD = LS[i].CD + (LS[i + 1]. k− value − LS[i − 1]. k− value)/(fkmax − fkmin ), where fkmax
and fkmin are the maximum and minimum values of the kth objective in set LS,
respectively;
10: End For
11: End For


non-dominated sorting approach proposed by Deb et al. its probability of being selected is CDp / ni=1 CDi , where
(2002) to divide the swarm into various non-dominated n is the number of particles in the non-dominated level
levels, which are denoted as L1 , L2 , . . . , LR . particle p belongs to. The purpose of using the CD-based
After the execution of the fast non-dominated sort- roulette wheel selection method is to diversity preserva-
ing, as shown in the line 6 of Algorithm 1, the crowding tion, because maintaining a good spread of solutions in
distance (CD) should be calculated for each solution in the final non-dominated solution set is one of the objec-
each non-dominated level. A solution with a smaller CD tives of multi-objective optimisation. After the two exem-
denotes that it is more crowded by other solutions. There- plars Xp1,r1 and Xp2,r2 are selected, the particle update
fore, for a particular solution, its CD can be viewed as an approach proposed in LLSO is employed here to update
estimate of the density of solutions surrounding it. CD the particle Xp,r as follows:
will be used in the particle update process of LMPSO for
d d d d d d
diversity preservation, which is described in detail in the vp,r = λ1 vp,r + λ2 (xp1,r1 − xp,r ) + φλ3 (xp2,r2 − xp,r )
following subsection. The CD calculation approach used (26)
in LMPSO is almost the same with the method intro- d d d
xp,r = xp,r + vp,r (27)
duced by Deb et al (2002). The only difference between
them is that we set the distance of a boundary solu- where Xp,r = {xp,r 1 , . . . , xd , . . . , xD } and V
p,r p,r p,r = {vp,r ,
1
tion, i.e. the solution with the smallest or largest func- d . . . , vD } are the position vector and velocity vec-
. . . , vp,r p,r
tion values, to 1 rather than a sufficiently large number. tor of the particle p from the level Lr , respectively. λ1 , λ2
The revised procedure of CD calculation is shown in and λ3 are three random numbers, generated uniformly
Algorithm 3. in the range [0, 1]. φ is the control parameter within [0,
1], which is used to control the influence of the second
4.5. Particle update process exemplar.
For the update of a particle Xp,2 in the second level,
In Algorithm 1, the pseudo codes shown in lines 8–22 its two exemplars are both selected in the first level using
describe the particle update process of LMPSO, where the CD-based roulette wheel selection method. Then, the
lines 8–16 are used for updating the particles in levels abovementioned particle update approach is employed to
LR to L3 , and lines 17–22 are employed for updating the update the particle Xp,2 . In Equation (26), the selected
particles in level L2 . particle with larger CD value is the first exemplar and the
For a particle Xp,r in Lr (3 ≤ r ≤ R), based on the other selected particle is the second exemplar.
update strategy of LLSO, two levels (r1 and r2 ) are ran-
domly selected first from the top (r-1) levels. For ease of
4.6. Local search
description, we assume that r1 < r2 < r. Subsequently,
different from LLSO, we use CD-based roulette wheel Local search algorithms are always used in the single-
selection method, rather than random selection method, objective optimisation to generate an improved solution
to select two particles Xp1,r1 and Xp2,r2 from Lr1 and Lr2 by searching the neighbourhoods around an incumbent
as two exemplars, respectively. CD-based roulette wheel solution (Yan et al. 2021). However, instead of improving
selection method is realised by using the CD value to an incumbent solution, searching a set of enhanced solu-
associate a probability of selection with each particle. tions to replace a non-dominated solution set is the aim
Therefore, if CDp represents the CD value of particle p, of local search for multi-objective optimisation. In this
3328 J. LONG ET AL.

paper, the local search used for the non-dominated parti- in the following experiments, some other information
cles in the first level L1 is described in Algorithm 4, which for the test instances were generated by a uniform dis-
is proposed on the basis of the variable neighbourhood tribution in the following range: job processing time
descent (Ilic et al. 2010) procedure. pi ∈ [10, 30], unit holding cost αi ∈ [0, 1], unit tardiness
By considering the performance and efficiency of the cost βi ∈ [0, 1]. The vehicle hiring cost δ was set to 20.
LMPSO, the seven neighborhood structures used in the For each test instance, the information of due window
local search are defined as follows: [die ,dil ] was determined using the following procedure:
(1) determine the matching relationship between jobs
(1). Neighbourhood N1 : exchange the processing and vehicles and the visiting sequence of jobs in their
sequence of two different jobs in the machine. assigned vehicle based on the optimal solution of the
(2). Neighbourhood N2 : a job from a vehicle route is corresponding benchmark problem; (2) arrange the pro-
removed from its original position and reinserted in cessing sequence of jobs in each vehicle according to the
all the remaining positions. value of αi /pi from small to large; (3) calculate the com-
(3). Neighbourhood N3 : exchange the positions of two pletion time of each job and the departure time of each
different jobs in a vehicle route. vehicle without considering any idle time on the machine;
(4). Neighbourhood N4 : transfer a job from a vehicle (4) calculate the delivery time si of each job based on the
route to another vehicle route. vehicle departure time and the visiting sequence; (5) gen-
(5). Neighbourhood N5 : exchange a job i from a vehicle erate a slack value εi uniformly from [20, 30], and set die =
route with a job j from another vehicle route. si − εi ; (6) generate another slack value εi uniformly from
(6). Neighbourhood N6 : reduce the value of an idle time [20, 30], and set dil = si + εi . Using this special procedure
on the machine by half if the value is greater than 0. to generate the due window information is beneficial to
(7). Neighbourhood N7 : reduce the value of an idle time the convergence analysis of the algorithm, which will be
between the departure time of a vehicle and the detailed in the following subsection. In our experiments,
completion time of its last job by half if the value we took 4 groups (i.e. group A, B, E and M) including
is greater than 0. 22 problems from the benchmark problems, and con-
structed 5 test instances for each benchmark problem.
5. Experimental results Thus, a total of 110 test instances were obtained. These
constructed test instances can be downloaded from
5.1. Instance generation
https://github.com/longjianyuGH/Instance_of_IPSVRP.
Due to the considered IPSVRP is a new multi-objective
optimisation problem and there are no existing benchmarks,
5.2. Comparison with the state-of-the-art
we generated a number of test instances by alter-
ing benchmark problems of capacitated vehicle routing In order to make quantitative comparisons, the first step
problem. The benchmark problem, which can be found is to determine the evaluation metrics. As stated before,
on http://www.coin-or.org/SYMPHONY/branchandcut/ the aim of multi-objective optimisation is to obtain a
VRP/data/index.htm.old, contains the data of node num- set of diverse non-dominated solutions in the Pareto-
ber (job number), node position (job delivery position), optimal front. Therefore, for a set of non-dominated
node demand (job occupied capacity), vehicle number, solutions obtained by algorithms, to measure their prox-
and vehicle capacity. In order to be able to use them imity to the Pareto-optimal front and their diversity in

Algorithm 4 Local search

Input: the first level L1
Output: set of improved non-dominated solutions L̃1
1: For each particle Xp,l in L1 do
2: initialise a neighbourhood list N = {N1 , N2 , . . . , N7 }, a empty set ;
3: While N = ∅ do
4: select randomly a neighbourhood Ni from N;
5: remove Ni from N, i.e. N = N − {Ni };
6: generate the neighbourhood solutions of Xp,l using Ni , denoted as Ni (Xp,l );
7: = non-dominated solutions of ( ∪ Ni (Xp,l ));
8: If there is a solution X∗p,l that dominates Xp,l in set do
9: Xp,l = X∗p,l ;
10: N = {N1 , N2 , . . . , N7 };
11: End if
12: End while
13: L1 = non-dominated solutions of (L1 ∪ );
14: End For
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3329

the solution set, four evaluation metrics are employed as The proposed LMPSO was then compared with
below: some state-of-the-art multi-objective optimisation algo-
rithms, which includes the nondominated sorting genetic
(1). D(Φ) (Pan, Wang, and Qian 2009), which is the algorithm II (NSGA-II) (Deb et al. 2002), the time vari-
distance between the non-dominated solution set ant multi-objective particle swarm optimization (TV-
Φ obtained by algorithms and the reference Pareto MOPSO) (Tripathi, Bandyopadhyay, and Pal 2007),
front Φ∗. The smaller D(Φ), the better Φ. D(Φ) is the competitive and cooperative co-evolutionary multi-
calculated as follows: objective particle swarm optimization algorithm
1  (CCPSO) (Goh et al. 2010), and the adaptive multi-
D(Φ) = dξ ∗ (Φ) (28) objective particle swarm optimization (AMOPSO) (Han,
|Φ ∗ |
ξ ∗∈Φ∗ Lu, and Qiao 2017). To configure these algorithms for

 3   solving our test instances, the encoding procedure of
 fk (ξ ) − fk (ξ ∗) 2
dξ ∗ (Φ) = min  (29) LMPSO was also used in TV-MOPSO, CCPSO, and
ξ ∈Φ fkmax − fkmin AMOPSO, and the discrete encoding method (i.e. the dis-
k=1
crete representation after performing the SPV and NDVR
where fk (ξ ) (or fk (ξ ∗)) represents the kth objec- rules on the real number representation of LMPSO) was
tive value of solution ξ (or ξ ∗), and fkmax (or fkmin ) employed in NSGA-II. Moreover, all the algorithms use
denotes the maximum (or minimum) value of the the same decoding method as the algorithm LMPSO. We
kth objective in the reference Pareto front Φ∗. In the programmed these algorithms using C++ language, and
following experiments, Φ∗ is constructed from the employed them to run the 110 test instances in the same
union of all the Φs obtained by the compared algo- operating environment.
rithms because the true Pareto front is unavailable. For a fair comparison, the same maximum elapsed
(2). N(Φ) (Pan, Wang, and Qian 2009), which is the num- CPU time of 800 × n × m milliseconds, where n is the
ber of non-dominated solutions in Φ that are not number of jobs and m is the number of vehicles, was used
dominated by any solution in Φ∗. The bigger N(Φ), as the termination criterion for each algorithm. Accord-
the better Φ. N(Φ) is calculated as follows: ing to the preliminary experimental results, the maxi-
mum elapsed CPU time set for each instance was shown
N(Φ) = |Φ − {ξ ∈ Φ|∃ξ ∗ ∈ Φ : to be sufficient for the convergence of each algorithm.
ξ ∗ DOMINATES ξ }| (30) Besides setting the same termination criterion, parameter
tuning was also conducted to obtain the best parameter
(3). M(Φ) (Deb 2001), which is the maximum spread setting for each algorithm because it always has a great
defined to measure the length of the diagonal of influence on the performance of a multi-objective evo-
a hyperbox formed by the extreme function val- lutionary algorithm. With respect to the setting of the
ues in Φ. The bigger M(Φ), the better Φ. M(Φ) is penalty coefficient η for constraint violation, we have
calculated as follows: tested eleven values (0.1, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5,
 4.0, 4.5, and 5.0) from the interval [0.1, 5.0]. The results

 3 2
showed that η is preferably set to any value larger than or

M(Φ) =  max fk (ξ ) − min fk (ξ ) (31) equal to 2.0. In this range, the value of η has little influ-
ξ ∈Φ ξ ∈Φ
k=1 ence on the algorithm performance. Therefore, η is set
to 2.0 for all the algorithms in the following experiments.
(4). R(Φ) (Deb 2001), which is the relative distance As for the other parameters for each algorithm, following
defined to measure the relative distance among solu- Long et al. (2018), the parameter tuning was performed
tions in Φ. The smaller R(Φ), the better Φ. R(Φ) is by designing different levels for each parameter first, and
calculated as follows: the experimental results of the parameter configurations

 generated by the orthogonal experimental design method
 1 
R(Φ) =  (uξ − ū)2 (32) were subsequently analysed using the analysis of variance
|Φ|
ξ ∈Φ (ANOVA). The best parameter settings of each algorithm
are reported in Table 2.

3
Based on the above parameter settings, all the algo-
uξ = min  |fk (ξ ) − fk (ξ  )| (33)
ξ  ∈Φ ∧ ξ =ξ rithms were employed to run the 110 test instances
k=1
 independently 30 times. Tables 3–6 list the experimental
ū = uξ /|Φ| (34) results with respect to the metrics D(Φ), N(Φ), M(Φ)
ξ ∈Φ and R(Φ) for the 4 groups of 22 benchmark problems,
3330 J. LONG ET AL.

Table 2. Parameter settings. focuses on measuring the spread of extreme solutions and
Algorithm Parameter settings R(Φ) reflects the relative distance among solutions. As
LMPSO swarm size 800, control parameter 0.4 shown in Tables 5 and 6, we can find that the LMPSO
NSGA-II population size 800, crossover probability 0.9, mutation performs better than the compared algorithms on most
probability 0.1
TV-MOPSO swarm size 600, archive size 500, initial value of the
of problems regarding the results of M(Φ) and R(Φ). In
cognitive acceleration coefficient 2.5, terminal value of details, the LMPSO obtains the largest M(Φ) in 19 prob-
the cognitive acceleration coefficient 0.5, initial value of lems and the smallest R(Φ) in 18 problems. However,
the social acceleration coefficient 0.5, terminal value of
the social acceleration coefficient 2.5, initial value of the we can find that the results of LMPSO are still not good
inertia weight 0.7, terminal value of the inertia weight enough because some M(Φ) and R(Φ) values of LMPSO
0.4, the degree parameter of dependence of mutation 5
CCPSO subswarm size 100, archive size 500, inertial weight 0.4 give no better results than other methods. The reason
AMOPSO swarm size 500, maximum number of nondominated may be that the local search designed in LMPSO pays too
solutions 500
much attention to finding solutions close to the Pareto
front. Therefore, how to design an effective diversity
maintenance strategy in the local search without affect-
where the metric values for each benchmark problem are ing its exploration performance is the focus in our future
the average metric values of the 5 test instances gener- studies.
ated for this benchmark problem. The ‘avg’ in these tables
refers to the average value of those metrics, and the ‘std’
refers to the standard deviations. As for the results of
5.3. Convergence analysis
D(Φ) and N(Φ) shown in Tables 3 and 4, one can find
that the LMPSO outperforms the compared algorithms Evolutionary algorithm usually faces such a difficult
on all the problems. Moreover, the superiority of LMPSO problem in solving optimisation problems, that is, it is
is more obvious in solving large-scale problems. From the difficult to know the gap between the obtained solution
definition of D(Φ) and N(Φ), we can see that they are (or non-dominated solution set) and the optimal solu-
used for evaluating the proximity degree of the obtained tion (or Pareto front). In this Section, the convergence
non-dominated solution set to the reference Pareto front. behaviour of the proposed LMPSO is investigated. As
Therefore, through comparing with these advanced described before, the true Pareto front is unknown for
multi-objective optimisation algorithms, the proposed each test instance. However, based on the concept of
LMPSO can obtain competitive and better performance domination, if we can construct a solution so that it can
in proximity to the Pareto-optimal front respect. obtain the optimal value of the corresponding objective
Different from D(Φ) and N(Φ), M(Φ) and R(Φ) are function according to a certain optimisation order, this
used to evaluate the diversity of solutions, where M(Φ) solution must be a Pareto-optimal solution. The quality

Table 3. The results obtained by different algorithms in terms of D(Φ).

LMPSO NSGA-II TV-MOPSO CCPSO AMOPSO
Problem avg std avg std avg std avg std avg std
A-n32-k5 0.0216 0.0119 0.0248 0.0217 0.0239 0.0228 0.0223 0.0157 0.0226 0.0125
A-n37-k6 0.0284 0.0134 0.0298 0.0153 0.0304 0.0157 0.0309 0.0155 0.0307 0.0182
A-n44-k6 0.0194 0.0402 0.0235 0.0277 0.0213 0.0257 0.0209 0.0283 0.0216 0.0208
A-n48-k7 0.0231 0.0147 0.0265 0.0153 0.0251 0.0198 0.0221 0.0141 0.0253 0.0162
A-n53-k7 0.0203 0.0500 0.0248 0.0340 0.0241 0.0166 0.0239 0.0340 0.0249 0.0181
A-n60-k9 0.0217 0.0167 0.0280 0.0110 0.0353 0.0152 0.0230 0.0109 0.0271 0.0149
A-n65-k9 0.0189 0.0163 0.0276 0.0263 0.0269 0.0278 0.0244 0.0273 0.0282 0.0215
A-n69-k9 0.0183 0.0312 0.0307 0.0110 0.0272 0.0152 0.0238 0.0109 0.0270 0.0149
A-n80-k10 0.0171 0.0194 0.0329 0.0263 0.0263 0.0278 0.0251 0.0273 0.0294 0.0225
B-n39-k5 0.0277 0.0202 0.0291 0.0280 0.0298 0.0651 0.0305 0.0298 0.0304 0.0368
B-n41-k6 0.0219 0.0130 0.0274 0.0166 0.0283 0.0319 0.0261 0.0168 0.0278 0.0297
B-n50-k7 0.0193 0.0532 0.0311 0.0127 0.0283 0.0337 0.0249 0.0253 0.0299 0.0367
B-n56-k7 0.0198 0.0498 0.0304 0.0333 0.0308 0.0351 0.0238 0.0317 0.0284 0.0417
B-n63-k10 0.0185 0.0583 0.0315 0.0154 0.0277 0.0204 0.0259 0.0189 0.0304 0.0271
B-n78-k10 0.0186 0.0551 0.0322 0.0316 0.0293 0.0327 0.0247 0.0319 0.0285 0.0407
E-n23-k3 0.0235 0.0142 0.0243 0.0132 0.0249 0.0178 0.0253 0.0132 0.0250 0.0214
E-n33-k4 0.0230 0.0327 0.0257 0.0162 0.0254 0.0193 0.0271 0.0363 0.0248 0.0187
E-n51-k5 0.0203 0.0550 0.0266 0.0208 0.0277 0.0328 0.0269 0.0196 0.0287 0.0422
E-n76-k10 0.0180 0.0551 0.0297 0.0243 0.0286 0.0487 0.0273 0.0241 0.0301 0.0319
E-n76-k14 0.0175 0.0811 0.0340 0.0407 0.0291 0.0751 0.0248 0.0506 0.0323 0.0528
M-n101-k10 0.0176 0.0861 0.0327 0.0302 0.0304 0.0594 0.0258 0.0444 0.0389 0.0517
M-n121-k7 0.0168 0.0693 0.0352 0.0535 0.0313 0.0780 0.0289 0.0368 0.0344 0.0465
Note: (The best average values are shown in boldface).
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3331

Table 4. The results obtained by different algorithms in terms of N(Φ).

LMPSO NSGA-II TV-MOPSO CCPSO AMOPSO
Problem avg std avg std avg std avg std avg std
A-n32-k5 83.58 31.80 78.19 26.04 77.58 25.81 82.20 25.59 76.16 26.18
A-n37-k6 78.31 29.79 76.75 28.40 74.62 26.43 75.83 27.77 74.41 28.25
A-n44-k6 93.23 27.29 81.18 26.03 86.56 19.58 85.24 24.43 84.45 22.03
A-n48-k7 82.45 33.48 77.01 28.34 78.83 21.35 80.19 26.33 79.03 21.26
A-n53-k7 98.08 29.57 86.15 22.23 83.02 14.94 84.25 20.14 85.17 19.98
A-n60-k9 91.67 29.50 84.49 24.07 85.28 17.41 87.24 22.1 82.09 18.91
A-n65-k9 89.58 37.58 81.06 35.31 82.47 23.9 84.28 34.26 83.43 22.18
A-n69-k9 93.92 32.16 78.78 23.30 81.50 13.68 85.19 21.31 83.32 16.98
A-n80-k10 117.23 19.26 95.81 18.80 100.85 18.47 104.49 18.91 97.23 19.58
B-n39-k5 86.21 39.16 83.55 37.55 83.24 32.34 81.87 34.96 84.05 37.14
B-n41-k6 90.17 35.76 82.09 33.13 81.16 24.19 83.28 30.84 85.45 27.49
B-n50-k7 93.24 36.96 78.94 32.95 81.68 25.21 86.24 31.54 83.57 23.96
B-n56-k7 91.79 29.70 82.11 28.73 82.40 20.61 85.89 27.4 83.74 23.39
B-n63-k10 104.26 41.44 84.88 40.71 90.24 23.1 93.37 37.75 85.56 23.46
B-n78-k10 116.29 44.78 99.93 35.18 104.52 36.96 109.74 36.4 106.19 34.63
E-n23-k3 89.86 27.66 87.60 27.18 86.87 25.6 85.59 27.62 88.42 25.07
E-n33-k4 92.18 38.09 86.72 35.32 87.03 27.78 84.24 34.94 88.16 27.12
E-n51-k5 97.24 32.27 86.25 35.21 82.04 33.71 87.11 41.18 83.72 30.11
E-n76-k10 126.36 46.54 108.01 42.87 108.26 20.95 110.27 41.4 105.40 21.89
E-n76-k14 122.20 42.40 100.57 46.62 107.29 21.15 111.89 41.75 106.31 28.12
M-n101-k10 129.67 36.81 107.29 35.44 109.26 27.92 114.25 36.11 102.17 27.17
M-n121-k7 135.26 43.83 109.27 29.33 110.24 20.35 115.37 34.52 120.89 22.84
Note: The best average values are shown in boldface.

Table 5. The results obtained by different algorithms in terms of M(Φ).

LMPSO NSGA-II TV-MOPSO CCPSO AMOPSO
Problem avg std avg std avg std avg std avg std
A-n32-k5 4253.53 332.79 3846.43 325.49 4192.22 331.29 4214.86 326.86 4268.26 318.84
A-n37-k6 4724.18 313.42 4125.63 306.85 4536.79 325.16 4688.28 318.42 4678.92 367.16
A-n44-k6 5635.24 259.15 4526.38 262.19 5029.25 269.98 5697.28 260.89 5429.25 272.18
A-n48-k7 5935.36 374.82 5028.16 370.43 5419.27 356.12 5835.62 368.11 5357.04 352.29
A-n53-k7 6825.18 362.12 5746.80 350.22 6325.87 351.94 6529.11 348.78 6120.93 357.51
A-n60-k9 8247.53 418.56 6326.11 407.31 7012.79 391.27 7824.17 403.02 7824.24 391.37
A-n65-k9 9021.42 541.85 6927.83 661.73 7347.89 692.3 7960.03 667.84 7925.27 624.51
A-n69-k9 9772.53 628.17 7457.32 710.01 7825.85 700.65 8457.28 709.74 8324.78 631.15
A-n80-k10 13467.68 786.07 8924.80 790.49 9543.28 698.5 10478.55 792.27 9626.89 719.12
B-n39-k5 5356.26 295.79 4525.18 298.64 5028.63 307.78 5259.27 304.26 5125.36 316.49
B-n41-k6 5625.99 413.53 5068.25 412.32 5293.58 407.23 5533.26 414.43 5225.76 414.18
B-n50-k7 6435.27 344.93 5348.02 344.53 5746.78 341.42 6025.67 349.47 625 l.26 368.15
B-n56-k7 7318.48 389.18 5490.13 403.19 6346.28 377.1 6850.26 394.22 6467.28 391.36
B-n63-k10 8626.79 321.98 6869.75 324.25 7579.02 329.75 8024.63 325.53 7836.77 341.55
B-n78-k10 10439.87 620.75 7168.05 528.74 7962.46 636.63 8687.20 731.61 9625.52 614.19
E-n23-k3 3189.48 338.37 3163.68 331.27 3257.48 317.65 3057.11 330.39 3321.15 310.71
E-n33-k4 4619.38 366.71 4328.18 376.82 4410.02 382.62 4027.41 377.28 4491.27 381.68
E-n51-k5 6529.69 494.57 5892.48 497.98 5436.28 516.68 6028.68 497.18 5956.29 471.09
E-n76-k10 12495.28 701.53 8629.69 685.46 9848.45 645.67 9825.64 689.7 10185.90 650.34
E-n76-k14 11984.29 865.59 8252.70 822.40 9257.28 749.65 9638.22 829.36 10639.25 739.14
M-n101-k10 23572.58 812.58 13578.28 734.36 15492.78 696.79 18682.79 742.71 18249.93 678.41
M-n121-k7 28983.96 703.02 16825.90 728.29 19682.68 714.64 22582.89 731.09 21683.65 719.53
Note: The best average values are shown in boldface.

of the convergent solution set obtained by LMPSO can be To construct the solution ξ1∗ , the first step is to obtain
subsequently evaluated using these constructed Pareto- the minimal total holding cost. In order to build a com-
optimal solutions. In this Section, we have constructed pact model to minimise this objective, the following key
three Pareto-optimal solutions for each test instance as structural properties used to provide necessary condi-
below: tions for any given solution to be optimal are derived.
Since only the total holding cost minimisation need to
(1). Pareto-optimal solution ξ1∗ , which is determined by be considered in this stage, Proposition 8 and Proposition
successively minimising the total holding cost and 9 are obtained by revising Proposition 2 and Proposi-
the total travelling cost. tion 3, respectively. Proposition 10 can be derived easily
3332 J. LONG ET AL.

Table 6. The results obtained by different algorithms in terms of R(Φ).

LMPSO NSGA-II TV-MOPSO CCPSO AMOPSO
Problem avg std avg std avg std avg std avg std
A-n32-k5 241.58 42.48 231.98 37.76 246.53 20.34 298.57 40.59 262.95 37.14
A-n37-k6 283.48 27.20 287.68 27.84 308.28 26.55 315.37 27.72 292.74 28.60
A-n44-k6 342.19 17.52 338.22 18.96 351.49 22.95 359.59 19.81 348.31 19.46
A-n48-k7 346.25 24.32 351.59 22.88 366.28 16.83 355.29 25.83 353.17 25.23
A-n53-k7 372.16 28.8 379.28 23.68 383.47 43.74 389.77 23.85 381.28 27.06
A-n60-k9 386.01 27.2 394.77 30.83 398.26 20.25 411.53 30.65 392.28 19.72
A-n65-k9 402.28 29.6 429.17 18.81 432.49 28.53 426.03 20.97 419.28 25.64
A-n69-k9 409.62 39.92 438.99 41.12 439.28 47.43 435.28 46.17 446.86 43.52
A-n80-k10 462.19 16.24 489.21 17.76 473.19 22.53 476.70 20.88 478.16 18.04
B-n39-k5 296.67 19.12 303.27 18.08 318.86 21.06 326.87 20.61 319.67 16.28
B-n41-k6 310.67 14.88 315.95 14.46 330.78 20.16 329.01 16.92 322.89 19.46
B-n50-k7 369.28 35.04 378.98 33.28 380.28 36.56 387.99 34.83 373.46 35.72
B-n56-k7 377.02 20.56 392.07 21.92 395.87 32.49 401.78 26.55 393.21 25.96
B-n63-k10 392.21 18.56 421.64 21.76 418.48 19.62 423.82 25.2 419.68 21.68
B-n78-k10 463.28 20.08 496.26 21.04 478.26 25.38 473.94 25.11 479.19 18.73
E-n23-k3 196.76 25.52 187.48 25.68 205.16 25.92 213.47 28.17 192.47 25.18
E-n33-k4 232.87 15.76 239.64 15.22 248.17 16.02 284.53 19.89 245.32 15.61
E-n51-k5 358.28 34.64 366.28 34.84 382.59 34.65 381.50 39.51 373.97 30.16
E-n76-k10 472.94 12.24 491.59 15.92 476.28 40.23 469.28 18.27 478.29 18.96
E-n76-k14 457.92 25.44 503.48 18.56 482.47 25.56 478.28 20.61 486.18 19.64
M-n101-k10 515.98 14.72 552.87 33.28 543.58 21.33 537.11 32.22 561.15 21.27
M-n121-k7 547.01 20.64 582.58 21.52 574.26 22.23 567.28 28.35 574.33 19.43
Note: The best average values are shown in boldface.

from Proposition 6. Finally, Proposition 11 is obviously 

m
true based on the calculation formula of job holding cost yiv = 1, i ∈ {1, 2, . . . , n} (37)
shown in Equation (1). v=1

yiv ∈ {0, 1}, i ∈ {1, 2, . . . , n}, v ∈ {1, 2, . . . , m} (38)

Proposition 8: In a solution with minimised total holding
cost, the jobs assigned to the same vehicle are necessarily
processed continuously on the machine. Compared with the model in Section 2, we can find
that only the decision variable yiv is remained in this
Proposition 9: In a solution with minimised total hold- model. The decision variables si and zij do not need to be
ing cost, a vehicle necessarily has to departure immediately considered because the decision-making of visiting time
after the last job assigned to this vehicle is completed. and visiting sequencing is not necessary to be considered
when minimising the total holding cost. According to the
Proposition 10: In a solution with minimised total hold- Proposition 9, Proposition 10 and Proposition 11, the deci-
ing cost, the processing sequence of jobs in each vehicle must sion variables bv , xij , and hv are not need any more. Based
be arranged according to the value of αi /pi from small to on the Proposition 8, we know that idle time is prohibited
large. among the jobs belong to the same vehicle, but not nec-
essarily prohibited between the jobs belong to different
Proposition 11: In a solution with minimised total hold- vehicles. However, according to the calculation formula
ing cost, all the vehicles must be used. of job holding cost, the idle time between the jobs belong
to different vehicles on the machine has no effect on the
According to the Proposition 10, all the jobs should total holding cost. For ease of calculation, we assume
be sorted first so that α1 /p1 ≤ α2 /p2 ≤ . . . ≤ αn /pn . The that there is no idle time on the machine. Therefore, in
compact model used to obtain the minimal total holding the above compact model, we only need to determine
cost is subsequently built as follows: the matching relationship between the jobs and vehicles
(i.e. the decision variable yiv ). The completion time ci of

m 
n−1 
n
min αi pj yiv yjv (35) each job can be subsequently calculated easily based on
v=1 i=1 j=i+1 the Proposition 8 and Proposition 10. The above objec-
tive function (35) is used to minimise the total holding
subject to cost. Constraints (36) are used to guarantee that the total
amount of capacity of jobs transported by a vehicle must

n
yiv oi ≤ Q, v ∈ {1, 2, . . . , m} (36) not exceed the vehicle capacity. Constraints (37) are used
i=1 to ensure that each job must be transported by exactly one
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3333

vehicle. The compact model is an integer quadratic pro- integer programming model. Due to the decomposition
gramming problem, which can be solved efficiently by the strategy, the scale of each TSPTW model will not be too
software package GUROBI. large in our experiments. Therefore, we also used the
The second step for constructing the solution ξ1∗ is software package GUROBI to solve them directly.
to determine the values of the remaining decision vari- After all the TSPTW models have been solved, all the
ables si and zij by minimising the total travelling cost. decision variables for constructing the solution ξ1∗ have
Since the matching relationship between the jobs and been determined. Finally, the three objective values of ξ1∗
vehicles has been already determined in the first step, the can be calculated according to the Eq.s (1), (2) and (3).
decision-making of visiting time and visiting sequenc-
ing of each vehicle is independent. Therefore, to solve (2). Pareto-optimal solution ξ2∗ , which is determined by
efficiently, the optimisation problem in this step can be successively minimising the total holding cost and
decomposed into multiple travelling salesman problems the total tardiness cost.
with time windows (TSPTW), whose number equals to
the number of vehicles. The model used to solve each The first step of constructing the solution ξ2∗ is to
TSPTW is built as follows: minimise the total holding cost. Therefore, the compact
model built for constructing the solution ξ1∗ is also used
min s0 (39) here. After the compact model is solved, the second step
is to determine the values of the remaining decision vari-
subject to ables si and zij by minimising the total tardiness cost.
Similarly, the decomposition strategy is also employed
s0 ≥ si + ti0 , i ∈ {1, 2, . . . , nv } (40)
to decompose the optimisation problem in this step into
si ≥ bv + t0i , i ∈ {1, 2, . . . , nv } (41) multiple TSPTWs. The only difference is that the objec-

nv tive is to minimise the total tardiness cost rather than the
zij = 1, j ∈ {0, 1, 2, . . . , nv } (42) total travelling cost.
 v Therefore, the objective (39) should
i=0 & i=j be replaced by ni=1 βi max{0, si − dil }. Since this non-
linear objective is not conducive to be solved using the

nv
zij = 1, i ∈ {0, 1, 2, . . . , nv } (43) software package GUROBI, we define a new decision
j=0 & j=i variable τi to represent the tardiness time of each job,
namely max{0, si − dil }. Now, the revised TSPTW model
si + tij − sj ≤ M(1 − zij ), i ∈ {1, 2, . . . , nv }, is subsequently built as follows:
j ∈ {1, 2, . . . , nv }, i = j (44)

nv
si ≥ die , i ∈ {1, 2, . . . , nv } (45) min βi τi (48)
+ i=1
si ∈ R , i ∈ {0, 1, 2, . . . , nv } (46)
zij ∈ {0, 1}, i ∈ {0, 1, 2, . . . , nv }, j ∈ {0, 1, 2, . . . , nv } subject to (40)–(47) and
(47)
τi ≥ si − dil , i ∈ {1, 2, . . . , nv } (49)
For each vehicle, s0 is defined as the time the vehi- τi ∈ R+ , i ∈ {1, 2, . . . , nv } (50)
cle returns to the depot. Therefore, the objective (39) can
be used to minimise the total travelling time because the After all the revised TSPTW models have been solved,
vehicle departure time has been determined. Constraints all the decision variables for constructing the solution ξ2∗
(40) ensure that the vehicle can return to the depot after have been determined. Finally, the three objective values
all the jobs contained in this vehicle have been delivered, of ξ2∗ can be calculated according to the formulas (1), (2)
where nv denotes the number of jobs contained in the and (3).
vehicle. Constraints (41) make sure the feasibility of the
time connectivity between the vehicle departure time and (3). Pareto-optimal solution ξ3∗ , which is determined by
the visiting time of each job assigned to this vehicle. Con- successively minimising the total travelling cost (or
straints (42)-(43) ensure that each job and the depot can the total tardiness cost) and the total holding cost.
be visited at most once. Constraints (44) make sure the
feasibility of the time connectivity between jobs visited by Since the test instances were constructed on the basis
the vehicle, and eliminate possible subtours. Constraints of the benchmark problems of CVRP, the minimal total
(45) guarantee that each job must be delivered after its travelling cost of ξ3∗ is equal to the sum of the optimal
earliest due date. The above TSPTW model is a mixed travelling distances of CVRP and the hiring costs of all
3334 J. LONG ET AL.

the vehicles. In addition, due to the special procedure minimal total holding cost of ξ3∗ can be calculated eas-
used to generate the due window information for each ily by arranging the processing sequence of jobs in each
test instance, the minimal total tardiness cost of ξ3∗ is vehicle according to the value of αi /pi from small to large.
equal to 0. After the matching relationship between jobs Now, based on the above constructed Pareto-optimal
and vehicles and the visiting sequence of jobs in their solutions for each instance, the metric D(Φ) shown in
assigned vehicle were determined based on the optimal Equation (28) can be recalculated by taking them as the
solution of the corresponding benchmark problem, the Pareto front Φ∗. The closer the revised D(Φ) value is to

Figure 2. The box plot of the revised D(Φ) values of LMPSO.

Figure 3. The distribution of the obtained solutions (expressed by blue circle ‘◦’) and the constructed Pareto-optimal solutions
(expressed by red asterisk ‘*’) of the first instance of A-n32-k5.
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3335

Figure 4. The distribution of the obtained solutions (expressed by blue circle ‘◦’) and the constructed Pareto-optimal solutions
(expressed by red asterisk ‘*’) of the first instance of B-n78-k10.

Figure 5. The distribution of the obtained solutions (expressed by blue circle ‘◦’) and the constructed Pareto-optimal solutions
(expressed by red asterisk ‘*’) of the first instance of M-n121-k7.
3336 J. LONG ET AL.

0, the closer the obtained non-dominated solution set is than 86% and 82% instances. In addition, the conver-
to the true Pareto front to some extent. Note that 5 test gence behaviour of LMPSO was also analysed, and the
instances were constructed for each benchmark problem, experimental results proved that the proposed LMPSO
and each instance was independently solved 30 times by has the ability to achieve a set of non-dominated solutions
the proposed LMPSO. The box plot, as shown in Figure 2, proximity to the true Pareto front on most of instances.
is used to describe the distribution of the revised D(Φ) From the experimental results, one can find that the
values obtained by LMPSO for each benchmark prob- performance of LMPSO in diversity maintenance is not
lem. In the box plot, the lower quartile, the median, and good enough. Designing an effective diversity mainte-
the upper quartile values are shown by each box. The nance strategy without deteriorating the exploration per-
lines extending from each end of the box are called as formance has to be considered for future work. More-
whiskers that used to show the extent of the remain- over, extending the proposed LMPSO to other realis-
ing data. Moreover, Figures 3–5 describe the distribution tic IPSVRPs with different production environment and
of the obtained solutions in set Φ and the constructed delivery characteristics is also the focus of future study,
Pareto-optimal solutions to graphically see their prox- such as flow shop, job shop, prize-collecting vehicle rout-
imity. We only give the results of the first instance of ing problem, multi-depot vehicle routing problem, etc.
A-n32-k5, B-n78-k10, and M-n121-k7 due to the space
constraint. From these figures, one can see that the pro- Disclosure statement
posed LMPSO can basically achieve very small D(Φ) val-
ues for small and medium-sized problems, which means No potential conflict of interest was reported by the author(s).
that the solution set Φ is close to the true Pareto front
to some extent. However, for large-scale problems, due Funding
to their very high-dimensional decision space, the per- This research is partially supported by the National Natu-
formance of the LMPSO needs to be further improved, ral Science Foundation of China (71801046 and 51775112),
which is our next research direction. Guangdong Basic and Applied Basic Research Foundation
(2019B1515120095), and the Intelligent Manufacturing PHM
Innovation Team Program (2018KCXTD029, TDYB2019010).
P.M. Pardalos is partially supported by the Paul and Heidi
Brown Preeminent Professorship at ISE (University of Florida,
6. Conclusions USA) and a Humboldt Research Award (Germany). The valu-
able comments and suggestions from the editors and the four
Integrated optimisation of production scheduling and anonymous reviewers are very much appreciated.
distribution decision has a strong application back-
ground in the MTO business environment. This paper
introduces a triple-objective optimisation model for Notes on contributor
managing a novel integrated production scheduling and Jianyu Long received the B.E. and Ph.D.
vehicle routing problem with inventory holding, deliv- degrees in metallurgical engineering from
ery, and tardiness costs. The model aims to optimise the Chongqing University, Chongqing, China,
trade-off among the total holding cost, the total travel- in 2012 and 2017, respectively. He is cur-
rently a Teacher of Industrial Engineering
ling cost, and the total tardiness cost. To this end, sev- with the School of Mechanical Engineer-
eral key structural properties used to provide necessary ing, Dongguan University of Technology,
conditions for any given solution to be Pareto-optimal Dongguan, China. From 2015 to 2016, he
are derived, and a level-based multi-objective particle was a Visiting Scholar with the Center for Applied Optimiza-
swarm optimizer (LMPSO) is subsequently designed tion, Department of Industrial and Systems Engineering, Uni-
versity of Florida, Gainesville, FL, USA. His research interests
for problem solving by incorporating these structural
include production scheduling, and prognostics and system
properties in the searching process. A set of experi- health management.
ments were conducted to test its performance on 110
Panos M. Pardalos received the B.E.
instances constructed based on the benchmark problems degree in Mathematics from Athens Uni-
of capacitated vehicle routing problem. Through com- versity, Greece, in 1977, the M.E. degree
paring with state-of-the-art multi-objective optimisation Mathematics and Computer Science from
algorithms, the proposed LMPSO obtains the best D(Φ) Clarkson University, Potsdam, USA, in
and N(Φ) values on all the instances, which demon- 1978, and the Ph.D. degree in Computer
and Information Sciences from University
strates that it can obtain competitive and better perfor-
of Minnesota, Minneapolis, USA, in 1985.
mance in proximity to the Pareto-optimal front respect. He is a Distinguished Professor of Industrial and Systems Engi-
As for the metrics M(Φ) and R(Φ), the LMPSO per- neering at the University of Florida, Gainesville. He is the
forms better than the compared algorithms in no less director of the Center for Applied Optimization. His current
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3337

research interests include global optimisation and applications, Han, H., W. Lu, and J. Qiao. 2017. “An Adaptive Multiobjective
networks, Parallel Computing in Mathematical Programming Particle Swarm Optimization Based on Multiple Adaptive
and machine learning. Methods.” IEEE Transactions on Cybernetics 47: 2754–2767.
Hassanzadeh, A., M. Rasti-Barzoki, and H. Khosroshahi. 2016.
Chuan Li received the Ph.D. degree in
“Two new Meta-Heuristics for a bi-Objective Supply Chain
mechatronics engineering from Chong-
Scheduling Problem in Flow-Shop Environment.” Applied
qing University, Chongqing, China, in
Soft Computing 49: 335–351.
2007. He was a Postdoctoral Fellow with
Ilic, A., D. Urosevic, J. Brimberg, and N. Mladenovic.
the University of Ottawa, Ottawa, ON,
2010. “A General Variable Neighborhood Search for Solv-
Canada, a Research Professor with Korea
ing the Uncapacitated Single Allocation p-hub Median
University, Seoul, South Korea, and a
Problem.” European Journal of Operational Research 206:
Senior Research Associate with the City
289–300.
University of Hong Kong, Hong Kong. He is currently a Profes-
Jozefowiez, N., F. Semet, and E. G. Talbi. 2008. “Multi-objective
sor of Mechanical Engineering with the School of Mechanical
Vehicle Routing Problems.” European Journal of Operational
Engineering, Dongguan University of Technology, Dongguan,
Research 189: 293–309.
China. His research interests include combinatorial optimi-
Kennedy, J., and R. C. Eberhart. “Particle Swarm Optimiza-
sation, prognostics and health management, and intelligent
tion.” International Conference on Neural Networks, Perth,
systems.
WA, Australia, Australia, 1995.
Lacomme, P., A. Moukrim, A. Quilliot, and M. Vinot. 2018.
“Supply Chain Optimisation with Both Production and
References Transportation Integration: Multiple Vehicles for a Sin-
Armstrong, R., S. Gao, and L. Lei. 2008. “A Zero-Inventory Pro- gle Perishable Product.” International Journal of Production
duction and Distribution Problem with a Fixed Customer Research 56: 4313–4336.
Sequence.” Annals of Operations Research 159: 395–414. Li, X. D., and X. Yao. 2012. “Cooperatively Coevolving Particle
Chen, Z. L. 2010. “Integrated Production and Outbound Dis- Swarms for Large Scale Optimization.” IEEE Transactions on
tribution Scheduling: Review and Extensions.” Operations Evolutionary Computation 16: 210–224.
Research 58: 130–148. Liu, L., W. Li, K. Li, and X. Zou. 2020. “A Coordinated Pro-
Cheng, R., and Y. C. Jin. 2015a. “A Competitive Swarm Opti- duction and Transportation Scheduling Problem with Min-
mizer for Large Scale Optimization.” IEEE Transactions on imum sum of Order Delivery Times.” Journal of Heuristics
Cybernetics 45: 191–204. 26: 33–58.
Cheng, R., and Y. C. Jin. 2015b. “A Social Learning Particle Long, J., Z. Sun, P. M. Pardalos, Y. Hong, S. Zhang, and C.
Swarm Optimization Algorithm for Scalable Optimization.” Li. 2019. “A Hybrid Multi-Objective Genetic Local Search
Information Sciences 291: 43–60. Algorithm for the Prize-Collecting Vehicle Routing Prob-
Dayarian, I., and G. Desaulniers. 2019. “A Branch-Price- lem.” Information Sciences 478: 40–61.
and-Cut Algorithm for a Production-Routing Problem Long, J. Y., Z. Zheng, X. Q. Gao, and P. M. Pardalos. 2018.
with Short-Life-Span Products.” Transportation Science 53: “Scheduling a Realistic Hybrid Flow Shop with Stage Skip-
829–849. ping and Adjustable Processing Time in Steel Plants.”
Deb, K. 2001. Multi-Objective Optimization Using Evolutionary Applied Soft Computing 64: 536–549.
Algorithms. Chichester: Wiley. Mandavi, S., M. E. Shiri, and S. Rahnamayan. 2015. “Meta-
Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan. 2002. “A Fast heuristics in Large-Scale Global Continues Optimization: A
and Elitist Multiobjective Genetic Algorithm: NSGA-II.” Survey.” Information Sciences 295: 407–428.
IEEE Transactions on Evolutionary Computation 6: 182–197. Marandi, F., and S. M. T. F. Ghomi. 2019. “Integrated Multi-
Eberhart, R. C., and J. Kennedy. “A New Optimizer Using Factory Production and Distribution Scheduling Applying
Particle Swarm Theory.” 6th International Symposium on Vehicle Routing Approach.” International Journal of Produc-
Micromachine and Human Science, Nagoya, Japan, 1995. tion Research 57: 722–748.
Feng, X., and Z. Xu. 2019. “Integrated Production and Miranda, P. L., R. Morabito, and D. Ferreira. 2018. “Optimiza-
Transportation Scheduling on Parallel Batch-Processing tion Model for a Production, Inventory, Distribution and
Machines.” IEEE Access 7: 148393–148400. Routing Problem in Small Furniture Companies.” Top 26:
Gharaei, A., and F. Jolai. 2018. “A Multi-Agent Approach to the 30–67.
Integrated Production Scheduling and Distribution Problem Mohammadi, S., S. M. J. M. Al-e-Hashem, and Y. Rekik. 2020.
in Multi-Factory Supply Chain.” Applied Soft Computing 65: “An Integrated Production Scheduling and Delivery Route
577–589. Planning with Multi-Purpose Machines: A Case Study from
Goh, C. K., K. C. Tan, D. S. Liu, and S. C. Chiam. 2010. a Furniture Manufacturing Company.” International Journal
“A Competitive and Cooperative co-Evolutionary Approach of Production Economics 219: 347–359.
to Multi-Objective Particle Swarm Optimization Algorithm Montes de Oca, M. A., T. Stutzle, K. Van den Enden, and
Design.” European Journal of Operational Research 202: M. Dorigo. 2011. “Incremental Social Learning in Particle
42–54. Swarms.” IEEE Transactions on Systems Man and Cybernetics
Gupta, V., I. E. Grossmann, S. Pathak, and J. André. “Assess- Part B-Cybernetics 41: 368–384.
ing the Benefits of Production-Distribution Coordination in Moons, S., K. Ramaekers, A. Caris, and Y. Arda. 2017. “Integrat-
an Industrial Gases Supply Chain.” Proceedings of Founda- ing Production Scheduling and Vehicle Routing Decisions at
tions of Computer-Aided Process Operations (FOCAPO), the Operational Decision Level: A Review and Discussion.”
Savannah, Georgia, 2012. Computers & Industrial Engineering 104: 224–245.
3338 J. LONG ET AL.

Pan, Q. K., L. Wang, and B. Qian. 2009. “A Novel Differen- Vidal, T., T. G. Crainic, M. Gendreau, N. Lahrichi, and W.
tial Evolution Algorithm for Bi-Criteria No-Wait Flow Shop Rei. 2012. “A Hybrid Genetic Algorithm for Multidepot and
Scheduling Problems.” Computers & Operations Research 36: Periodic Vehicle Routing Problems.” Operations Research 60:
2498–2511. 611–624.
Qin, Q., S. Cheng, Q. Zhang, L. Li, and Y. Shi. 2016. “Particle Wang, S., R. Wu, F. Chu, and J. Yu. 2020. “Variable Neigh-
Swarm Optimization With Interswarm Interactive Learning borhood Search-Based Methods for Integrated Hybrid Flow
Strategy.” IEEE Transactions on Cybernetics 46: 2238–2251. Shop Scheduling with Distribution.” Soft Computing 24:
Soukhal, A., A. Oulamara, and P. Martineau. 2005. “Complex- 8917–8936.
ity of Flow Shop Scheduling Problems with Transportation Wang, J., S. Yao, J. Sheng, and H. Yang. 2019a. “Minimizing
Constraints.” European Journal of Operational Research 161: Total Carbon Emissions in an Integrated Machine Schedul-
32–41. ing and Vehicle Routing Problem.” Journal of Cleaner Pro-
Tamannaei, M., and M. Rasti-Barzoki. 2019. “Mathematical duction 229: 1004–1017.
Programming and Solution Approaches for Minimizing Wang, D., J. Zhu, X. Wei, T. C. E. Cheng, Y. Yin, and Y. Wang.
Tardiness and Transportation Costs in the Supply Chain 2019b. “Integrated Production and Multiple Trips Vehicle
Scheduling Problem.” Computers & Industrial Engineering Routing with Time Windows and Uncertain Travel Times.”
127: 643–656. Computers & Operations Research 103: 1–12.
Tasgetiren, M. F., Y.-C. Liang, M. Sevkli, and G. Gencyil- Yan, Y., C. D. Li, L. J. Song, J. Yang, and J. F. Su. 2021. “A Multi-
maz. 2007. “A Particle Swarm Optimization Algorithm for dimensional Information Fusion-Based Matching Decision
Makespan and Total Flowtime Minimization in the Permu- Method for Manufacturing Service Resource.” IEEE Access
tation Flowshop Sequencing Problem.” European Journal of 9: 39839–39851.
Operational Research 177: 1930–1947. Yang, Q., W. N. Chen, J. Da Deng, Y. Li, T. L. Gu, and J.
Tavares-Neto, R. F., and M. S. Nagano. 2019. “An Iterated Zhang. 2018. “A Level-Based Learning Swarm Optimizer for
Greedy Approach to Integrate Production by Multiple Par- Large-Scale Optimization.” IEEE Transactions on Evolution-
allel Machines and Distribution by a Single Capacitated ary Computation 22: 578–594.
Vehicle.” Swarm and Evolutionary Computation 44: 612–621. Zhan, Z. H., J. Zhang, Y. Li, and Y. H. Shi. 2011. “Orthogonal
Tripathi, P. K., S. Bandyopadhyay, and S. K. Pal. 2007. “Multi- Learning Particle Swarm Optimization.” IEEE Transactions
Objective Particle Swarm Optimization with Time Variant on Evolutionary Computation 15: 832–847.
Inertia and Acceleration Coefficients.” Information Sciences Zou, X., L. Liu, K. Li, and W. Li. 2018. “A Coordinated
177: 5033–5049. Algorithm for Integrated Production Scheduling and Vehi-
Tseng, C. T., and C. J. Liao. 2008. “A Particle Swarm Opti- cle Routing Problem.” International Journal of Production
mization Algorithm for Hybrid Flow-Shop Scheduling with Research 56: 5005–5024.
Multiprocessor Tasks.” International Journal of Production
Research 46: 4655–4670.