INTERNATIONAL

TRANSACTIONS
IN OPERATIONAL
Intl. Trans. in Op. Res. 32 (2025) 863–887 RESEARCH
DOI: 10.1111/itor.13282

Clustered vehicle routing problem for waste collection with

smart operational management approaches
Jungmin Kima , Apurba Mannab , Arindam Roya,c and Ilkyeong Moona,d,∗
a
Department of Industrial Engineering, Seoul National University, Seoul 08826, South Korea
b
Research Centre in Natural Sciences, Raja N. L. Khan Women’s College, Medinipur 721102, India
c
Department of Computer Science & Application, Prabhat Kumar College, Contai 721404, India
d
Institute of Engineering Research, Seoul National University, Seoul 08826, South Korea
E-mail: andykim14@snu.ac.kr [Kim]; apurba_sukumar@rnlkwc.ac.in [Manna]; arcontai@gmail.com [Roy];
ikmoon@snu.ac.kr [Moon]
Received 29 July 2022; received in revised form 18 November 2022; accepted 27 February 2023

Abstract
Waste collection is one of the essential tasks in a smart city. The Internet of Things (IoT) is a promising
technology that offers potential solutions for transforming traditional systems. An IoT-based smart bin is a
modern technology that offers real-time fill level information to a cleaning authority. However, high uncer-
tainty associated with the smart bin’s fill levels and improper operation hinder efficient waste collection. In
order to tackle the uncertainty in a smart bin and improve the waste collection operation, the IoT sensor’s
usage must be combined with optimization procedures. The present work introduced two operational man-
agement approaches to define dynamic optimal routes and combined ant colony optimization with a k-means
clustering algorithm to solve the clustered vehicle routing problem for waste collection on a large scale. Op-
erational management approaches reflect practical constraints when using IoT-based smart bins. A hybrid
metaheuristic is proposed and performed with these approaches thereby showing the potential of building a
smart waste collection system.

Keywords: clustering; ant colony optimization; smart bin; waste collection; vehicle routing

1. Introduction

The Internet of Things (IoT) is gaining attention in numerous research areas because it has a good
effect on the way operations are managed (Atzori et al., 2017). To take advantage of this technol-
ogy, governments and companies are developing capabilities to maximize their utilities to work in
tandem with the IoT. Doing so will potentially boost the efficiency and better effectiveness of their

∗
Corresponding author.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
864 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
operations while streamlining decision-making within the organizations (Fayoumi and Loucopou-
los, 2016).
Waste has increased drastically with rapid urbanization, and the outbreak of COVID-19 has ac-
celerated mainly plastic waste production (Peng et al., 2021). According to recent findings, more
than eight million tons of pandemic-associated plastic waste have been generated globally, with
more than 25,000 tons entering the global ocean. While dealing with this unexpected surge in waste,
conventional incineration facilities and landfills have reached their limit. Hence, the waste manage-
ment industry faces intense pressure over handling hazardous waste generated from COVID-19.
The utilization of IoT technologies can pave the way for better monitoring and control of waste.
Some researchers have studied different areas within waste management. Del Pia and Filippi
(2006) considered moving depots for waste collection and proposed a heuristic to reduce the total
route duration. Zhao and Zhu (2016) considered planning tours and vehicle acquisitions for waste
collection. Sahebjamnia et al. (2018) proposed a multiobjective model with environmental consid-
erations in the context of a closed-loop supply chain. Coban et al. (2018) addressed the multicriteria
decision-making methods to dispose of municipal waste and indicated the prominence of recycling
and landfill technologies in developing countries. López-Sánchez et al. (2018) addressed the multi-
objective routing problem for waste collection and proposed a hybrid heuristic to solve a real-world
problem. Ramos et al. (2018) addressed the recyclable waste collection problem using smart bins
and developed optimization procedures. Lin et al. (2020) addressed a robust facility location model
for waste transfer stations under uncertain environments. Eryganov et al. (2020) presented possible
approaches to modeling cooperation between the waste producers in a certain location with limited
or banned landfilling using the cooperative game theory. Adeleke and Ali (2021) suggested point-
of-collection sorting for efficient collection. In this paper, we focus on the vehicle routing problem
for waste collection using a smart bin.
The remainder of this paper is organized as follows. Section 2 reviews relevant literature. Section 3
provides a detailed problem description and two operational management approaches with mathe-
matical models. In Section 4, a hybrid metaheuristic is presented. Computational experiments are
given in Section 5. Section 6 presents the contributions and summary of this paper.

2. Literature review

Increasing waste has become one of the more concerning issues in a smart city because of rapid
urbanization. To tackle this problem, many authorities spotlight the importance of smart bins. The
smart bin equipped with an IoT sensor provides real-time fill level information to the authorities
to aid them in making proper decisions to collect waste. Huang and Lin (2015) made decisions
on optimal route planning and scheduling for collecting waste. However, they did not consider
IoT-based smart bins. Ramos et al. (2018) presented operational management approaches using
a smart bin. They introduced some parameters to estimate the fill level of bins. López-Sánchez
et al. (2018) proposed a multiobjective problem for waste collection with a hybrid metaheuristic
and a case study in Spain is presented. Ferrer and Alba (2019) elaborated on the mechanism of
the IoT-based smart bin but failed to deliver a mathematical model. Ríos-Mercado et al. (2023)
presented a dispersion territory design problem considering legal constraints and proposed an ad-
vanced metaheuristic algorithm to solve this problem. Delgado-Antequera et al. (2020) proposed

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 865
a multiobjective approach for waste collection and proposed a hybrid heuristic, while Haque et al.
(2020) presented an IoT-based efficient waste collection system with smart bins. Nonetheless, their
works are confined to a systematic approach to collecting waste. Salamirad et al. (2023) performed
a multicriteria decision-making analysis to treat industrial wastewater. They also proposed a new
hybrid heuristic to suggest managerial insights in a case study. Sarkar et al. (2022a) focused on
food waste with a circular economy concept. Exhaustive experimental results are supported with
extensive supply chain analysis. Roy et al. (2022) considered IoT-based smart bins with multiple
waste types and utilized vehicle compartments for each waste type. Sarkar et al. (2022b) proposed
a green supply-chain management setup for biodegradable products with an outsourcing strategy
and transportation modes. To the best of our knowledge, this paper is the first study that considers
the usage of IoT-based smart bin in a mathematical model and suggests operational management
approaches for waste collection.
The clustered vehicle routing problem (CluVRP) generalizes the capacitated vehicle routing prob-
lem (CVRP). Sevaux et al. (2008) introduced CluVRP in the context of a real-world application
where containers are employed to carry goods. To the best of our knowledge, the literature describes
only two exact approaches for the CluVRP. Pop et al. (2012) presented two compact formulations
but did not show computational results. Battarra et al. (2014) developed two exact algorithms and
provided results for a set of benchmark instances.
In the CluVRP, numerous researchers suggested different criteria for generating clusters. Ghiani
et al. (2005) allocated arcs and edges to vehicles for satisfying service deadlines and container-to-
vehicle compatibility constraints. Defryn et al. (2016) investigated the effect of the cost allocation
of a partner’s strategy on nondelivery penalties and the properties of its customer locations (dis-
tance to the depot, degree of clustering). Fernández et al. (2018) introduced the shared customer
collaboration in a vehicle routing problem (VRP). They assumed that some customers are shared
by different vehicles, in the sense that they have demand from more than one vehicle. Defryn and
Sörensen (2017) presented an improved two-level heuristic to solve the CluVRP. And they intro-
duced a new variant of the CluVRP, the CluVRP with weak cluster constraints. Hintsch and Irnich
(2018) decomposed CluVRP into three subproblems (i.e., the assignment of clusters to routes, the
routing inside each cluster, and the sequencing of the clusters in the routes). In this paper, we follow
the decomposing method presented by Hintsch and Irnich (2018). For the assignment of clusters,
the k-means clustering algorithm is used. Furthermore, an ant colony optimization (ACO) is ap-
plied for calculating routing inside each cluster. Finally, for the sequencing of the clusters in the
route, smart bins equipped with IoT sensors are used so that vehicles start their journey when they
receive signals from smart bins. Al-Refaie et al. (2021) considered clusters for efficient waste col-
lection. Hubs are allocated in each cluster, and a vehicle visits these hubs instead of visiting whole
bins. In this paper, bins are segmented into each cluster, and a vehicle can enter and leave a cluster.
Some previous research in the waste collection are compared with the present study in Table 1.
The CluVRP-WC is a generalization of the CVRP in which smart bins are segmented by clusters.
Furthermore, because determining the optimal solution to VRP is an N P-hard, as per Toth and
Vigo (2002), and the CluVRP-WC is a special case of the VRP, the CluVRP-WC is also an N P-
hard. To tackle this problem, this paper suggests two operational management approaches. In the
first experiment, this paper extends the experiment presented by Kim et al. (2020) to compare the
operational management approaches. In the next experiment, the proposed hybrid metaheuristic
was implemented to solve this problem on a large scale.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
866 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
Table 1
Contribution of different authors

Smart
Authors (year) Model bin Clustering Waste type Solution methodology
Zhao and Zhu (2016) MDVRP Recyclable Lexicographic weighted Tchebycheff
method
Defryn and Sörensen (2017) CluVRP  Variable neighborhood search
Ramos et al. (2018) DVRP  Recyclable Heuristic
López-Sánchez et al. (2018) WCP Solid Variable neighborhood descent
Hintsch and Irnich (2018) CluVRP  Neighborhood search, Variable
neighborhood descent
Ríos-Mercado et al. (2023) MDTDP Electronics Metaheuristic
Delgado-Antequera et al. MCDM General Heuristic
(2020)
Salamirad et al. (2023) CVRP Water Variable neighborhood search
Jorge et al. (2022) VRPP  Recyclable Simulated annealing, Neighborhood
search
Sarkar et al. (2022a) CEM Food Stochastic dual coordinate ascent
Roy et al. (2022) CVRP  Food, Variable neighborhood search
Recyclable,
General
Sarkar et al. (2022b) GSC Biodegradable Operational management approaches
This paper CluVRP   General Operational management approaches
Abbreviations: CEM, circular economic model; DVRP dynamic VRP; GSC, green supply chain; MCDM, multicriteria decision-
making; MDTDP, maximum dispersion territory design problem; MDVRP, multi depot VRP; VRP, vehicle routing problem;
VRPP, VRP profit; WCP, waste collection problem.

3. Mathematical model

3.1. Motivation

In this section, we present the motivation for the smart waste collection problem occurring in a
smart city where IoT-based smart bins are placed around the city. Because of irregular cleaning
and improper waste management, most waste bins get overfilled frequently and consequently cre-
ate unhealthy and nonhygienic situations. In the proposed model, these problems can be solved by
setting up IoT-based smart bins, which send alerts to the cleaning authority when they are filled
up to a predefined level, say 80%. During routing, a driver cannot check all bins fill level infor-
mation status because of the limitation of data transmission. According to Kim (2015), the IoT
sensor range is classified by its length (i.e., short-range for Bluetooth mesh networking, Wi-Fi, ra-
dio frequency identification (RFID); medium-range for LTE and 5G; and long-range for low-power
wide-area networking and very small aperture terminals. In this paper, it is assumed that smart bins
are equipped with a short-range (maximum 10 m) IoT sensor. In addition, each vehicle is equipped
with an IoT sensor receiver so that neighboring bins locations and fill level information are sent to
the vehicle during routing. This paper introduces a neighboring bound to represent the radius of
the IoT sensor. To the best of our knowledge, this paper is the first study that suggests a set concept
for certainly filled bins, penalty bins, and neighboring bins in the mathematical model.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 867
3.2. Problem description

The problem addressed in this paper is clustered vehicle routing problem for waste collection
(CluVRP-WC), considering the use of real-time information about bin fill levels to define dynamic
routes. This problem can be defined as follows: A cluster set is defined as M, based on the k-means
clustering algorithm and elbow method. For simplicity, the k-means algorithm is used to generate
clusters. A set of the depot and all bins is defined as B. A set of bins in each cluster is defined as
(Bm , m ∈ M). Each bin i’s fill level is stated with Li . A complete undirected graph is considered to
connect all bins and a depot, with the distance, di j , for each edge (i, j) in the graph. K is assumed
to be a set of homogeneous vehicles with a maximum capacity of L each. All vehicles depart and
return to the depot. The maximum number of vehicles is less than or equal to the number of clus-
ters to utilize the vehicle fully. Following Expósito-Izquierdo et al. (2016), the CluVRP-WC can be
defined by the mathematical model below, which requires the definition of some additional vari-
ables. Suppose Z to be any subset of B that is different from B. Then, let δ + (Z) be the set of edges
(i, j) ∈ Z × B\Z, and δ − (Z) be the set of edges (i, j) ∈ B\Z × Z. This paper considers several sets
to represent the CluVRP-WC model. These are cluster sets, filled bin sets, penalty bin sets, and
neighboring bin sets. If the fill levels of bins are above the threshold fill level (T F L), these bins are
defined as filled bin set, B f . The penalty is imposed on overflowing bins (i.e., a penalty bin set is de-
fined when the fill level is more than 100% ) and is defined as B p. Sets of the filled bins and penalty
bins in each cluster are defined as Bm f and Bmp. Following the definition by Jorge et al. (2022),
these filled bins are included in “must-go bins.” Considering the IoT sensor range, which is defined
as aneighboring bound, R, a neighboring bin set is introduced. When a bin is given, neighboring
bins are within the neighboring bound from the bin. These bins are included in the neighboring bin
set. The total time horizon is considered as T . At the beginning of each day, t(∈ T ), in the morning,
the sensors located inside the waste bins transmit information on the bins’ fill levels. The problem
in hand selects the waste bins to be visited (must-go bins) and the optimal visiting sequence in each
day t for each vehicle k, which will minimize the summation of routing cost throughout the clusters
while satisfying the vehicles’ fixed capacity. Routing cost and penalty cost are included in the total
cost. The assumptions of the mathematical model are as follows.

Assumptions:
1. Smart bins are equipped with IoT sensors and identical with the same capacity. The working
principle is taken from Roy et al. (2022).
2. Vehicles are equipped with IoT sensor receivers and homogeneous with the same capacity. Dur-
ing routing, a vehicle can receive fill level information within a neighboring bound from its
current location.
3. A general type of waste is considered.
4. If the fill level in a bin reaches the T F L, then an alert message is transmitted to the cleaning
authority so that the bin can be visited.
5. Every bin in each cluster will be visited by the same vehicle during collection time and will be
visited if the fill level is above T F L or the bin is within the neighboring bound from the filled
bin.
6. All filled bins (above T F L) must be visited. If not, a penalty is imposed in proportion to the fill
level of the filled bin.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
868 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

Fig. 1. Operational management approaches.

3.3. Operational management approaches

Considering the problem description above, two operational management approaches to define
dynamic collection routes are studied. Figure 1 shows the outline of operational management ap-
proaches. On the first day of the planning period, a cleaning authority checks the fill level infor-
mation from all bins. If the bins’ fill levels are above T F L, an alert message is sent to the cleaning
authority so that a vehicle can depart from a depot. In the selection process for must-go bins, two
operational management approaches are introduced. After one of the approaches is decided, then
solve the CluVRP-WC model and waste collection proceeds.

(1) estimation-based collection approach, based on Phase 1 heuristic by Jorge et al. (2022) in which
fill level estimation is considered to select the must-go bins at each day. This is coupled with a

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 869
CluVRP-WC model that optimizes each vehicle’s route, defining the best sequence to visit the
selected bins.
(2) Neighborhood-based collection approach, in which a neighboring bin is proposed to select the
must-go bins to perform the collection operation, and the revised mathematical model is pre-
sented.

The first approach is based on the method followed by Jorge et al. (2022), where the selection of
the bins to be visited in each day is based on currently filled bins (above T F L) and future filled bins
considering a fill level estimation. Once the set of bins to be visited is defined, each vehicle route
is optimized through a CluVRP-WC model. In the neighborhood-based collection approach, a
neighboring bin is introduced. A feasible set that a vehicle visit is reduced to the set of filled bins and
their neighboring bins. While visiting filled bins, a vehicle will visit neighboring bins from the filled
bin if the vehicle has enough space to collect waste. Between these approaches, the only difference
is the must-go bin selection. The notation used to formulate the CluVRP-WC is as follows:

Sets
M Set of m clusters, M = {1, 2, . . . , m}
Bm Set of bins in cluster m, bmi ∈ Bm , bmi = bin i in cluster m, ∀m ∈ M
{0} Depot
B Set of the depot and all bins, B = {0} ∪ B1 ∪ B2 ∪ · · · ∪ Bm
Bm f Set of the filled bins in cluster m, fill level≥ T F L, Bm f ⊆ Bm , ∀m ∈ M
Bf Set of all filled bins, B f = B1 f ∪ B2 f ∪ · · · ∪ Bm f
Bmp Set of penalty bins in cluster m, fill level≥ 100%, Bmp ⊆ Bm f , ∀m ∈ M
Bp Set of all penalty bins, B p = B1p ∪ B2p ∪ · · · ∪ Bmp
K Set of vehicles, |K| ≤ m
δ + (Z) Set of edges (i, j) ∈ Z × B\Z
δ − (Z) Set of edges (i, j) ∈ B\Z × Z
N[bmi ] Neighbor of bmi , N[bmi ] = {bmi } ∪ {bmx : bmx ∈ Bm ; i = x, ∀ m ∈ M, ∀ i, x = 1, 2, . . . , |Bm | and
d (bmx , bmi ) ≤ R, d denotes the euclidean distance}
Tm Bm f ∪ N[bml ], bml ∈ Bm f

Parameters
T F L (%) Threshold fill level
di j (km) Travel cost required to move a vehicle from node i to node j
Li (kg) Amount of waste at bin i
Pi ($) Penalty if bin i’s fill level is more than 100%
L (kg) Maximum capacity of a vehicle
T Total time horizon
R (m) Neighboring bound

Variables

1 if a vehicle k traverses from node i to node j in the undirected graph
xi jk =
0 otherwise
i, j ∈B, k ∈ K
1 if a biniis visited in the solution
yi =
0 otherwise
i ∈ B\{0}

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
870 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
To build the CluVRP-WC model, routing cost for collecting waste is considered in the objective
function and penalty cost is added to the total cost throughout all clusters. The total cost is given
by
 
xi jk di j + Pi . (1)
k∈K i∈B j∈B i∈B p

The followings are constraints. Constraint (2) enforces that all vehicles start from a depot and
return to the depot.
 
x0ik = x j0k = 1, ∀k ∈ K. (2)
i∈B\{0} j∈B\{0}

Constraint (3) establishes that only one vehicle can enter and leave a cluster.
   
xi jk = xi jk = 1, ∀m ∈ M (3)
k∈K (i, j)∈δ + (Bm ) k∈K (i, j)∈δ − (Bm )

Constraint (4) verifies the same vehicle can enter and leave a cluster.
 
xi jk = xi jk , ∀m ∈ M, ∀k ∈ K. (4)
(i, j)∈δ + (Bm ) (i, j)∈δ − (Bm )

Constraint (5) indicates that each vehicle cannot exceed its maximum capacity.
 
xi jk Li ≤ L, ∀k ∈ K, i = j. (5)
i∈B\{0} j∈B

Constraint (6) guarantees that every bin is visited at most once in the solution.
 
xi jk = y j , ∀ j ∈ B\{0}, i = j. (6)
k∈K i∈B\{0}

Constraint (7) ensures the connectivity of the path of a single vehicle.

 
xipk = x p jk , ∀k ∈ K, ∀p ∈ B\{0}. (7)
i∈B\{0} j∈B\{0}

Constraint (8) represents the classic subtour elimination constraints.


xi jk ≤ |S| − 1, ∀k ∈ K, ∀S ⊆ B\{0}, |S| ≥ 2. (8)
i∈S j∈S

Constraint (9) represents all filled bins (above T F L) that must be visited.

yi = |B f |. (9)
i∈B f

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 871

Fig. 2. Must-go bins selection algorithm.

Constraints (10) and (11) present binary variables.

xi jk ∈ {0, 1}, ∀i ∈ B, ∀ j ∈ B, ∀k ∈ K, (10)
yi ∈ {0, 1}, ∀i ∈ B. (11)

3.3.1. Estimation-based collection approach

The fill level estimation and “must-go bins” selection algorithm are the ones proposed by Jorge
et al. (2022). We followed the notations, parameters, and the outline of the must-go bins selection
algorithm by Jorge et al. (2022), and the differences are described in Fig. 2. While they considered
full or overflowing bins for selecting must-go bins, we considered a certain threshold fill level for
selecting bins to avoid overflowing bins. The bins to be visited are decided in the morning of each

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
872 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
day t, after receiving the real-time information on the bins’ fill level (Li ). “Must-go bins” are defined
as the combination of bins in which the fill level is more than T F L at day t and the fill level is
estimated to be above T F L from t + 1 to t (t ∈ (t + 1, T )). T F L indicates a certain threshold fill
level of a bin. The first collection day is the day on which the bins’ fill level is above T F L at day
t. These bins are included in the must-go bins (B f + i). The following estimated collection day is
the day in which the bins that were above T F L on the first collection day are predicted to become
above T F L again (Ŵit ≥ T F L). Let us assume the following estimated collection day is t . Between
day t and day t , check bins’ fill level status which is above T F L and is not included in the must-go
bins on the first collection day. If the fill level is estimated to be above T F L during this time period,
these bins are included on the first collection day to avoid potential overflowing. The next step is
to solve the CluVRP-WC model. To build the cost minimization model for waste collection, the
following sets, parameters, and variables are defined. The number of clusters, m, is decided using a
k-means clustering algorithm and the elbow method. The following parameters are adopted from
Jorge et al. (2022).

Parameters
Wit (kg) Amount of waste at bin i at day t
Ŵit (%) Estimated waste fill level in bin i at day t
ntt Number of days between collection day t and day t
āi (%) Average daily accumulation rate of bin i
α Value for a probability retrieved from the normal distribution depending on the desired level of confidence
for Ŵit
σai Standard deviation of the daily accumulation rates of bin i

3.3.2. Neighborhood-based collection approach

In this approach, this paper introduced that a vehicle visits neighboring bins from the filled bins
instead of using fill level estimation. All the filled bins are included in the must-go bins as in the
estimation-based collection approach. Furthermore, the must-go bins selection algorithm is revised
in Fig. 3. A vehicle will visit all the filled bins (above T F L) and their
neighboring bins. If the vehicle
has enough space to hold waste after visiting a filled bin (L − i∈B f Wit > 0), then the vehicle
searches neighboring bins from the filled bin (B f + N[bmi ]). For the mathematical modeling part,
only the differences are presented below. A subset of filled bins and their neighboring bins in each
cluster is defined as Tm . Parameter R represents the radius of the IoT sensor on bins. A feasible area
that a vehicle can visit is reduced to the set of the filled bins and their neighboring bins. Therefore,
Bm in constraints (3) and (4) is replaced with Tm in constraints (12) and (13). Other constraints are
the same as the estimation-based collection approach.
   
xi jk = xi jk = 1, ∀m ∈ M, (12)
k∈K (i, j)∈δ + (Tm ) k∈K (i, j)∈δ − (Tm )

 
xi jk = xi jk , ∀m ∈ M, ∀k ∈ K. (13)
(i, j)∈δ + (Tm ) (i, j)∈δ − (Tm )

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 873

Fig. 3. Revised must-go bins selection algorithm.

The CluVRP-WC is a generalization of CVRP. For this reason, this problem can be solved in a
limited instance size. To tackle this problem, this paper presented a hybrid metaheuristic combining
the k-means clustering algorithm and ACO algorithm.

4. Hybrid metaheuristic

In this section, this paper proposed an algorithm that combines the k-means clustering algorithm
(Algorithm 1) and ACO. The waste bins are segmented based on the k-means clustering algo-
rithm, and paths in each cluster are optimized by using ACO. The methodology is described as
follows.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
874 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
4.1. Cluster computation and optimal path design

Numerous methods for generating clusters (e.g., distribution-based, density-based, graph-based,

prototype-based, etc.) are proposed by Tan et al. (2016). In this problem, waste bins are distributed
in distinct geographical locations. For simplicity, the k-means clustering algorithm is used, one of
the prototype-based clustering methods, to group the bins. This section uses the clustering tech-
nique to improve the performance of the evolutionary computation algorithm.
The k-means clustering algorithm is a distance-based clustering technique from which centroids
of each cluster are calculated. This algorithm performs an iterative alternating fitting process to
form a specific number of clusters. Initially, c points are randomly selected to be a first guess of
the centroids of the clusters. Each node is assigned to the nearest centroid to form temporary clus-
ters. In the next step, the centroids are replaced by means of clusters, and the nodes and centroids
are reassigned. This process continues until no further changes occur in the clusters. The k-means
clustering algorithm is one of the partition-based clustering techniques proposed by Hartigan and
Wong (1979). The classical k-means clustering algorithm is described in Algorithm 1. In this pro-
posed CluVRP model, all the waste bins are grouped using the k-means clustering algorithm. The
number of clusters is selected by the elbow method, the distance is based on Euclidean types, and
mean square errors examine its consistency. The k-means clustering algorithm is given below:

Algorithm 1. k-means clustering algorithm

Require: Set of bins and clusters

Ensure: Unique set of bins in each cluster; centroid of each cluster
Step 1: Initialize a positive integer c (c = |M|) using the elbow method
Step 2: Generate bins in each cluster (an initial set) randomly
Step 3: Determine the centroid of each cluster
Step 4: Allocate each bin to the cluster where it is nearest to the centroid
Step 5: Update the centroid in each cluster
Step 6: Repeat Steps 5 and 6 until the centroids no longer move

After allocating the bins to each cluster, the next aim is to find the optimum path in each cluster.
In the exact approach, it is solved in a limited instance size and difficult to determine the best
path that considers the routing cost in a feasible time. Thus, this paper presents the evolutionary
computation method. With this heuristic, all nodes are clustered into |M| groups, and an optimal
path is calculated in each cluster.

4.2. k-Means-ant colony optimization

The proposed ACO and its procedures in each cluster are given as follows.

4.2.1. Representation
In each cluster, a complete cycle of vehicle depot, filled bins, and neighboring bins represent a solu-
tion of ants. Therefore, an N(p + 1)-dimensional integer vector, Xi = ({0}, bm1 , bm2 , . . . , bmp ), m =
1, 2, . . . , |M|, is used to represent a vehicle depot and p the number of must-go bins. To be specific,

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 875

Algorithm 2. Path construction

Require: A depot and the set of bins of a cluster

Ensure: The optimum tour for the rth ant of the cluster;
Step 1: Set E = {1, 2, . . . ., N}, l = 1 and E = {2, 3, . . . ., N}. Node 1 to N in E represent a depot and must-go
bins, respectively
Step 2: xrl = a random element from the set E
Step 3: i = xrl
Step 4: Set W L=0; W L means waste load
Step 5: while {((L − W L) >= Li ) and (l < N)}
{W L = W L + Li
Set E =E -{i}
Let bin i be the present position of an ant. Then the next bin j ∈ E is selected by the ant, with
δ
τi j1
probability pi j given by the formula, pi j =  δ1 , where δ1 is a user-defined parameter that controls
τi j
j∈E
the relative importance of pheromone concentration. A roulette-wheel selection process is used for
this parameter.
l = l + 1, i = j}
Step 6: Print the optimal tour of the cluster

{0} represents a vehicle depot and bm1 , bm2 , . . . , bmp represent a set of must-go bins in cluster m in a
complete cycle. In the proposed ACO, τi j represents the amount of pheromone that lies on the path
between bin i and bin j.

4.2.2. Pheromone initialization

As the aim of the CluVRP-WC is to minimize the total routing cost on the proposed model, it is
assumed that in each cluster, an initial value of pheromone between bin i and bin j is τi j = √1 .
di j

4.2.3. Path construction

In each cluster, the following are required to construct a path Xr for the rth ant and presented in
Algorithm 2. n paths are constructed for n different ants and all vehicles, ∀k ∈ K.

4.2.4. Pheromone evaporation

In each cluster, the following formula is used between bin i and bin j for the evaporation of
pheromone:

τi j = (1 − ρ )τi j , (14)

where ρ lies between [0, 1]. The constant ρ specifies the pheromone evaporation rate, causing ants
to forget previous decisions.

4.2.5. Pheromone updating

After the completion by all ants, a pheromone is increased on the paths through which the ants have
traveled. Depending upon the nature of the present problem, a pheromone is updated using the

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
876 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

Algorithm 3. k-means-ACO

Require: Problem data, ACO parameters, number of clusters (|M|)

Ensure: The optimum tour in each cluster
Step 1: Initialize a set of bins and the number of clusters (|M|)
Step 2: Execute a k-means clustering algorithm (subsection 4.1)
Step 3: Start ACO
Step 4: Identify the fill level information above T F L
Step 5: Distribute the total number of ants
Step 6: Determine pheromone initialization (subsection 4.2.2)
Step 7: Execute path construction (subsection 4.2.3)
Step 8: Perform pheromone evaporation (subsection 4.2.4)
Step 9: Execute pheromone updating (subsection 4.2.5)
Step 10: If pheromone updating occurred, go to Step 7
Step 11: Find the optimal tour in each cluster

Table 2
Parameters of the k-means-ACO algorithm

Parameter α β ρ n δ1
Value 1 2 0.5 300 1

following rules, where ρ represents the rate of evaporation and n is the number of ants. Algorithm 3
shows the outline of the proposed hybrid heuristic.

ρ  best
n
τi j = (1 − ρ )τi j + τi j . (15)
n
i=1

5. Computational experiments

In this section, this paper presents the comparison results of two operational management ap-
proaches. At first, experiments were carried out in a small size with an exact method and the pro-
posed heuristic. These experiments extend the study presented by Kim et al. (2020). Their study
showed the practical usage of a smart bin and presented the outline of optimal waste collection.
However, the number of bins in their experiments is restricted to only five and they considered one
vehicle. In the small-size experiments, we extended this experiment with 15 bins and four vehicles. In
addition, 20 different instances are used to represent the diverse random fill level information. This
paper also considered penalty costs for overflowing bins and applied two operational management
approaches. The limitation of the estimation-based collection approach is presented, and large-size
experiments were conducted to support this result. To judge the effectiveness and feasibility of the
proposed hybrid algorithm, a dispersion test was conducted based on the traveling salesman prob-
lem library (TSPLIB) presented by Reinelt (1991). In the large-size experiment, this problem is
hard to solve with the exact method because of exponentially increased computation time. In this
regard, this paper applied a hybrid metaheuristic to solve this problem. The parameter setting for
the hybrid metaheuristic is described in Table 2. In Tables 4–6, a deviation is calculated between

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 877

Fig. 4. Gwanak district in Seoul, Korea.

operational management approaches. EBCA stands for estimated-based collection approach, and
NBCA stands for neighborhood-based collection approach. In Tables 7 and 8, an error is calculated
between the average solution and the best solution.
NBCA − EBCA
Deviation(%) = × 100,
EBCA
average solution − best solution
Error(%) = × 100.
best solution

5.1. Computational results on the small-size problem

An exact method was coded in Python 3.7.9 with CPLEX Optimizer 20.1, and the hybrid meta-
heuristic was coded in Python 3.6.8 on an Intel Core i7 Macintosh 2.6 GHz. In the small-size
problem, the Gwanak district (Fig. 4) in Seoul, Korea, is chosen for the location of IoT-based
smart bins. Kim et al. (2020) randomly chose five bins in Seoul. In this experiment, this paper also
selected 15 places randomly in Seoul. Each location corresponds to a bin. A depot is set in the mid-
dle of these places and denoted as 0. All these places have been taken from Google Maps, and their
distances are written in Table 3. For the waste generation, 20 instances are used and each instance
corresponds to a different random seed in the Python code. The computation times of the k-means-
ACO were calculated as the average values from 10 experiments. In these experiments, the unit for

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
878 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
Table 3
Distance matrix of bins and a depot: Gwanak district, Seoul, Korea

(i, j) 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 0 2.41 2.16 1.13 1.34 0.69 1.38 0.36 0.58 0.74 1.17 1.40 1.46 1.89 2.31 2.63
1 2.41 0 1.03 1.28 1.53 2.05 2.47 2.77 3.02 2.99 3.58 3.66 2.36 3.42 4.10 4.71
2 2.16 1.03 0 1.13 1.85 2.07 2.77 2.50 2.72 2.48 3.14 3.10 1.50 2.57 3.31 4.01
3 1.13 1.28 1.13 0 0.84 0.92 1.63 1.49 1.75 1.71 2.29 2.40 1.44 2.37 2.98 3.52
4 1.34 1.53 1.85 0.84 0 0.67 9.94 1.56 1.79 2.08 2.49 2.72 2.24 3.02 3.57 3.97
5 0.69 2.05 2.07 0.92 0.67 0 0.78 0.86 1.08 1.45 1.79 2.04 1.97 2.54 3.01 3.31
6 1.38 2.47 2.77 1.63 9.94 0.78 0 1.35 1.50 2.04 2.21 2.54 2.74 3.26 3.65 3.83
7 0.36 2.77 2.50 1.49 1.56 0.86 1.35 0 0.21 0.72 0.93 1.20 1.80 1.99 2.29 2.49
8 0.58 3.02 2.72 1.75 1.79 1.08 1.50 0.21 0 0.65 0.70 1.03 1.85 1.95 2.16 2.32
9 0.74 2.99 2.48 1.71 2.08 1.45 2.04 0.72 0.65 0 0.63 0.68 1.33 1.28 1.58 1.90
10 1.17 3.58 3.14 2.29 2.49 1.79 2.21 0.93 0.70 0.63 0 0.46 2.01 1.66 1.67 1.64
11 1.40 3.66 3.10 2.40 2.72 2.04 2.54 1.20 1.03 0.68 0.46 0 1.80 1.28 1.2 1.25
12 1.46 2.36 1.50 1.44 2.24 1.97 2.74 1.80 1.85 1.33 2.01 1.80 0 1.04 1.78 2.50
13 1.89 3.42 2.57 2.37 3.02 2.54 3.26 1.99 1.95 1.28 1.66 1.28 1.04 0 0.72 1.51
14 2.31 4.10 3.31 2.98 3.57 3.01 3.65 2.29 2.16 1.58 1.67 1.20 1.78 0.72 0 0.81
15 2.63 4.71 4.01 3.52 3.97 3.31 3.83 2.49 2.32 1.90 1.64 1.25 2.50 1.51 0.81 0

cost is a dollar, and the unit for computation time is a second. The distance unit is a kilometer,
and the maximum bin capacity is 100 kg. For the penalty cost, this paper imposed penalty cost in
proportion to a bin’s fill level so as to represent the relative amount of penalty.
In the small-size problem, the following parameters are used. |K| = 4, L = 625 kg, T = 7,
Pi = Li × 15%, R = 10 m. We assumed the unit traveling cost of a vehicle as $1 so that the distance
matrix in Fig. 4 corresponds to the traveling cost in the distance. Table 4 shows the exact method
results for operational management approaches. In both collection approaches, total routing costs
in the neighborhood-based collection approach were lower than those of the estimation-based
collection approach with an average 3.87%. However, there were significant differences in the
computation times. In the neighborhood-based collection approach, computation times were
almost half on average compared to those in the estimation-based collection approach.
In the estimation-based collection approach, this paper used the must-go bins selection algorithm
inspired by Jorge et al. (2022) and it had some limitations. First, because each bin has a different
fill rate, there is a possibility of unnecessarily including most of the bins in the must-go bins so that
a vehicle visits all bins, thereby increasing computation times. To be specific, suppose a bin is filled
up above T F L at the first time period but the fill rate is slow. Then the time when this bin is filled
up above T F L again is set far from the first time period so that most of the bins are included in
the must-go bins in the first time period. Second, they assumed normal distribution for the fill rate,
which is a too naive approach for tackling the uncertainty of fill level information. Therefore, the
must-go bins selection algorithm is revised and combined with a hybrid metaheuristic to solve the
problem on a large scale.
In the small-size problem, this paper also applied a proposed heuristic to compare the perfor-
mance with the exact algorithm. Table 5 presents the comparison of operational management
approaches. The parameter setting is the same as in the exact algorithm. In both collection
approaches, total routing costs in the neighborhood-based collection approach were lower than

© 2023 International Federation of Operational Research Societies.

 Table 4
Comparison of operational management approaches with an exact algorithm in a small size

Estimation-based collection approach Neighborhood-based collection approach Deviation (%)

Routing Penalty Total Computation Routing Penalty Total Computation Routing Penalty Total Computation
Instance cost ($) cost ($) cost ($) times (s) cost ($) cost ($) cost ($) times (s) cost cost cost times
Instance 1 947.87 70.20 1018.07 222.97 984.60 70.20 1054.80 174.01 3.87 0.00 3.61 −21.96
Instance 2 1263.83 0.00 1263.83 332.88 1159.37 0.00 1159.37 148.27 −8.27 0.00 −8.27 −55.46
Instance 3 704.33 15.00 719.33 213.47 684.07 15.00 699.07 38.75 −2.88 0.00 −2.82 −81.85
Instance 4 686.41 0.00 686.41 202.68 624.73 0.00 624.73 73.87 −8.99 0.00 −8.99 −63.55
Instance 5 1068.38 45.00 1113.38 249.08 1083.28 45.00 1128.28 90.72 1.39 0.00 1.34 −63.58
Instance 6 1197.18 45.00 1242.18 320.06 1091.43 45.00 1136.43 207.16 −8.83 0.00 −8.51 −35.28
Instance 7 907.70 33.30 941.00 247.92 892.84 32.10 924.94 170.17 −1.64 −3.60 −1.71 −31.36
Instance 8 1190.91 0.00 1190.91 303.34 1100.72 0.00 1100.72 88.53 −7.57 0.00 −7.57 −70.82
Instance 9 943.39 98.55 1041.94 274.62 1036.75 98.85 1135.60 132.05 9.90 0.30 8.99 −51.92
Instance 10 1201.21 0.00 1201.21 302.41 1008.78 0.00 1008.78 116.30 −16.02 0.00 −16.02 −61.54
Instance 11 899.16 112.20 1011.36 243.75 946.94 127.20 1074.14 49.27 5.31 13.37 6.21 −79.79
Instance 12 907.41 45.60 953.01 322.86 846.60 45.60 892.20 37.41 −6.70 0.00 −6.38 −88.41
Instance 13 912.62 101.70 1014.32 277.03 947.87 101.70 1049.57 192.14 3.86 0.00 3.48 −30.64
Instance 14 889.90 32.40 922.30 237.25 782.64 32.40 815.04 38.47 −12.05 0.00 −11.63 −83.79
Instance 15 863.65 34.20 897.85 235.48 846.60 34.20 880.45 110.57 −2.02 0.00 −1.94 −53.05
Instance 16 911.09 31.50 942.59 319.49 932.45 31.50 963.95 218.05 2.34 0.00 2.27 −31.75
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

Instance 17 1164.76 0.00 1164.76 304.74 1032.78 0.00 1032.78 106.67 −11.33 0.00 −11.33 −65.00
Instance 18 906.17 99.90 1006.07 295.73 785.33 99.90 885.23 70.70 −13.34 0.00 −12.01 −76.09
Instance 19 912.93 116.70 1029.63 262.64 947.87 116.70 1064.57 169.15 3.83 0.00 3.39 −35.60
Instance 20 918.15 15.00 933.15 299.10 828.91 15.00 843.91 79.51 −9.72 0.00 −9.56 −73.42
Average −3.87 −57.74

© 2023 International Federation of Operational Research Societies.

879

14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 880

Table 5
Comparison of operational management approaches with the proposed heuristic in a small size

Estimation-based collection approach Neighborhood-based collection approach Deviation (%)

Routing Penalty Total Computation Routing Penalty Total Computation Routing Penalty Total Computation
Instance cost ($) cost ($) cost ($) times (s) cost ($) cost ($) cost ($) times (s) cost cost cost times
Instance 1 951.60 70.20 1021.80 49.49 988.06 70.20 1058.26 37.32 3.83 0.00 3.57 −24.59
Instance 2 1265.71 0.00 1265.71 64.59 1253.91 0.00 1253.91 61.47 −4.88 0.00 −4.88 −4.83
Instance 3 771.60 15.00 786.60 29.63 689.57 15.00 704.57 29.88 −4.44 0.00 −4.35 0.84
Instance 4 702.75 0.00 702.75 29.05 639.68 0.00 639.68 29.93 −8.98 0.00 −8.98 3.03
Instance 5 1160.65 45.00 1205.65 50.72 1100.28 45.00 1145.28 42.16 −0.93 0.00 −0.90 −16.88
Instance 6 1198.01 45.00 1243.01 54.63 1092.73 45.00 1137.73 52.97 −8.79 0.00 −8.47 −3.04
Instance 7 912.27 33.30 945.57 48.76 935.06 15.00 950.06 52.71 2.50 −54.95 0.47 8.10
Instance 8 1196.28 0.00 1196.28 63.88 1132.27 0.00 1132.27 53.60 −5.35 0.00 −5.35 −16.09
Instance 9 943.96 98.55 1042.51 52.57 1058.85 114.15 1173.00 54.64 12.17 15.83 12.52 3.94

© 2023 International Federation of Operational Research Societies.

Instance 10 1208.05 0.00 1208.05 61.43 1039.86 0.00 1039.86 43.66 −13.92 0.00 −13.92 −28.93
Instance 11 908.21 112.20 1020.41 45.50 987.20 88.25 1075.45 43.02 8.70 −21.35 5.39 −5.45
Instance 12 911.39 45.60 956.99 43.98 855.39 49.50 904.89 43.88 −6.14 8.55 −5.44 −0.23
Instance 13 918.23 101.70 1019.93 50.49 955.84 94.85 1050.69 41.19 4.10 −6.74 3.02 −18.42
Instance 14 890.23 32.40 922.63 46.14 790.82 48.00 838.82 43.39 −11.17 48.15 −9.08 −5.96
Instance 15 865.25 34.20 899.45 47.21 855.25 30.60 885.85 45.20 −1.16 −10.53 −1.51 −4.26
Instance 16 913.53 31.50 945.03 49.67 948.11 47.10 995.21 44.98 3.78 49.52 5.31 −9.44
Instance 17 1167.25 0.00 1167.25 59.76 1040.13 0.00 1040.13 41.43 −10.89 0.00 −10.89 −30.67
Instance 18 909.06 99.90 1008.96 49.50 822.08 99.90 921.98 45.53 −9.57 0.00 −8.62 −8.02
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

Instance 19 917.84 116.70 1034.54 48.45 962.68 102.50 1065.18 49.62 4.89 −12.17 −2.96 2.41
Instance 20 920.59 15.00 935.59 47.22 855.10 15.00 870.10 45.82 −7.11 0.00 −7.00 −2.96
Average −2.81 −8.07

14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 881

Fig. 5. Performance gaps with respect to operational management approaches.

those of the estimation-based collection approach with an average 2.81%. In addition, computation
time is reduced to 8.07% on average. To judge the effectiveness of the heuristic, we calculated the to-
tal cost gap between the exact method and the heuristic with operational management approaches
in a small size, and the results are described in Fig. 5. BFS stands for the total cost in Table 4,
while HEU stands for the total cost in Table 5. Through all instances, these gaps are calculated
within 5%. As for computation time, the neighborhood-based collection approach outperforms the
estimation-based collection approach. However, as for total cost, the estimation-based collection
approach shows stable performances when using the exact method and heuristic.

HEU − BFS
Total cost gap (%) = × 100.
BFS

5.2. Computational results on the large-size problem

In the large-size problem, the eil51 instance is used from the TSPLIB for general location infor-
mation. In this instance, 51 nodes correspond to 51 locations of smart bins, and a depot is set in
the middle of the nodes. The following parameters are used for the experiment. |K| = 5, L = 1000
kg, T = 7, Pi = Li ∗ 15%, R = 10 m. This paper assumed the unit traveling cost of a vehicle as $1
so that the distance matrix in the TSPLIB corresponds to the traveling cost in the distance. We
consider eil51, a TSPLIB instance of 51 nodes corresponding to 51 bins. A total of 51 bins were in-
dexed from 1 to 51 and were segmented into five clusters. The number of clusters was selected using
the elbow method. In this experiment, total routing costs in the neighborhood-based collection ap-
proach were lower than those of the estimation-based collection approach with an average 2.97%.
However, there were also significant differences in the computation times. In the neighborhood-
based collection approach, computation times were reduced 10.61% on average compared to those

© 2023 International Federation of Operational Research Societies.

 882

Table 6
Comparison of operational management approaches with the proposed heuristic in a large size

Estimation-based collection approach Neighborhood-based collection approach Deviation (%)

Routing Penalty Total Computation Routing Penalty Total Computation Routing Penalty Total Computation
Instance cost ($) cost ($) cost ($) times (s) cost ($) cost ($) cost ($) times (s) cost cost cost times
Instance 1 2429.97 15.00 2444.97 228.29 2459.76 15.00 2474.76 169.67 1.23 0.00 1.22 −25.68
Instance 2 2455.41 45.00 2500.41 213.81 2282.37 60.60 2342.97 172.60 −7.05 34.67 −6.30 −19.27
Instance 3 1102.06 15.00 1117.06 124.71 1121.72 15.00 1136.72 103.60 1.78 0.00 1.76 −16.93
Instance 4 1113.96 15.00 1128.96 117.87 998.19 15.00 1013.19 112.60 −10.39 0.00 −10.25 −4.47
Instance 5 2268.96 60.00 2328.96 188.40 2245.60 47.55 2293.15 171.85 −1.03 −20.75 −1.54 −8.78
Instance 6 2176.97 135.00 2311.97 186.76 1805.62 120.60 1926.22 185.18 −17.06 −16.68 −16.68 −0.85
Instance 7 2291.84 15.00 2306.84 191.46 2102.51 46.95 2149.46 159.05 −8.26 213.00 −6.82 −16.93
Instance 8 2402.72 45.00 2447.72 202.83 2201.15 45.00 2246.15 180.09 −8.39 0.00 −8.24 −11.21
Instance 9 1843.00 48.00 1891.00 182.62 1788.20 30.00 1818.20 178.98

© 2023 International Federation of Operational Research Societies.

−2.97 −37.50 −3.85 −1.99
Instance 10 2357.46 15.00 2372.46 196.63 2316.25 32.55 2348.80 169.77 −1.75 117.00 −1.00 −13.66
Instance 11 2344.09 15.00 2359.09 204.28 2494.84 15.00 2509.84 172.50 6.43 0.00 6.39 −15.56
Instance 12 1754.29 45.00 1799.29 154.49 1743.31 45.00 1788.31 171.86 −0.63 0.00 −0.61 11.24
Instance 13 2365.67 31.95 2397.62 196.48 2242.91 30.60 2273.51 179.35 −5.19 −4.23 −5.18 −8.72
Instance 14 2360.13 15.00 2375.13 197.99 2335.97 30.60 2366.57 172.48 −1.02 104.00 −0.36 −12.88
Instance 15 1620.14 15.00 1635.14 148.42 1642.46 15.00 1657.46 151.77 1.38 0.00 1.36 2.26
Instance 16 2299.61 15.00 2314.61 189.05 2202.78 15.00 2217.78 164.29 −4.21 0.00 −4.18 −13.10
Instance 17 2272.27 0.00 2272.27 191.19 2165.50 0.00 2165.50 160.18 −4.70 0.00 −4.70 −16.22
Instance 18 2255.77 15.00 2270.77 197.90 2267.08 15.00 2282.08 164.59 0.50 0.00 0.50 −16.83
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

Instance 19 2297.73 15.00 2312.73 194.31 2358.98 15.00 2373.98 164.96 2.67 0.00 2.65 −15.10
Instance 20 2296.59 15.00 2311.59 199.18 2215.36 15.00 2230.36 184.06 −3.54 0.00 −3.51 −7.59
Average −2.97 −10.61

14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 883
Table 7
Dispersion results of k-means-ACO for different TSPLIB instances

Instance Clusters Best Worst Average SD Error

bayg29 2 2022 2188 2086.7 56.90 3.20
3 2049 2275 2189 80.16 6.83
4 2060 2196 2131.5 44.60 3.47
5 2135 2329 2239.6 55.61 4.90
dantzig42 2 714 851 792.6 40.58 11.01
3 711 847 793.3 41.69 11.58
4 741 817 773.8 20.34 4.43
5 841 898 861.1 16.37 2.39
eil51 2 615 694 649.5 27.90 5.61
3 782 846 809.4 23.14 3.50
4 881 909 896.6 7.88 1.77
5 1019 1070 1046.2 17.38 2.67
berlin52 2 7846 8416 8197.5 212.92 4.48
3 8150 8668 8464.3 166.44 3.86
4 8401 8963 8766.7 156.12 4.35
5 8377 9570 9024.7 497.06 7.73
st70 2 1058 1181 1143.7 39.40 8.10
3 1252 1439 1368.4 57.69 9.30
4 1597 1702 1653.2 40.30 3.52
5 1719 1780 1756.4 19.20 2.18
rat99 2 2068 2353 2172.3 96.77 5.04
3 2077 2392 2285.7 111.26 10.05
4 2192 2456 2315.7 91.62 5.64
5 2345 2648 2526.2 93.69 7.73
eil101 2 976 1088 1035.9 38.39 6.14
3 1158 1342 1234.5 47.31 6.61
4 1382 1491 1439.3 31.20 4.15
5 1519 1659 1572.8 47.56 3.54

in the estimation-based collection approach. Table 6 shows results for operational management
approaches with the proposed hybrid metaheuristic in a large size.

5.3. Performance test of k-means-ACO

The performance of the proposed hybrid algorithm was statistically tested by running it 10 times.
An average value, standard deviation (SD), and error are calculated according to the optimal solu-
tion against seven TSPLIB instances. To judge the effectiveness of the clustering algorithm, different
numbers of clusters are provided. The results are given in Table 7. To compare the dependency of
the clustering solution with other existing algorithms, an advanced GA is selected. The parameter
setting for GA is exactly the same from the paper. The modified GA is the combination of roulette
wheel selection, comparison crossover, and random mutation proposed by Maity et al. (2016). This
GA is combined with the k-means algorithm to compare with the k-means-ACO. The results are
presented in Table 8.

© 2023 International Federation of Operational Research Societies.

 884

Table 8
Performance test of k-means-ACO and k-means-GA on standard TSPLIB instances
Total cost($) Computation time(s)
Average Best SD Error(%) Average Best SD Error(%)
k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means- k-means-
Instance Nodes Clusters ACO GA ACO GA ACO GA ACO GA ACO GA ACO GA ACO GA ACO GA

bayg29 29 3 1846.5 1676.6 1646 1628 96.81 53.57 12.18 2.99 9.66 16.74 8.98 15.67 0.77 0.79 7.53 6.83
bays29 29 3 2252.7 2167.7 2033 2022 121.68 201.71 10.81 7.21 9.75 17.77 9.17 15.63 0.77 3.35 6.25 13.68
dantzig42 42 2 818.6 795.4 765 763 35.31 32.75 7.01 4.25 13.46 37.23 12.81 35.63 0.39 1.25 5.06 4.48
eil51 51 5 1027.8 956.8 982 894 35.93 54.93 4.66 7.02 15.16 44.01 14.87 42.02 0.23 1.93 1.97 4.75
berlin52 52 5 8597.8 9061.9 8267 8306 162.24 544.83 4.00 9.10 15.54 45.50 14.77 43.27 0.89 2.66 5.20 5.16
st70 70 4 1659.1 1544.1 1619 1389 29.49 96.85 2.48 11.17 20.82 60.76 19.64 57.60 0.97 2.87 6.00 5.50
eil76 76 4 1201.6 1134.6 1146 1081 39.35 46.11 4.85 4.96 22.33 67.14 21.52 64.47 1.03 2.01 3.77 4.14

© 2023 International Federation of Operational Research Societies.

pr76 76 5 138434.7 138493.7 123620 131538 8535.55 5005.62 11.98 5.29 21.72 67.85 21.06 64.31 0.52 1.76 3.15 5.50
rat99 99 3 2229.8 2584.7 2160 2369 40.74 162.24 3.23 9.11 45.46 113.59 44.63 92.11 1.03 26.76 1.87 23.32
eil101 101 4 1332.6 1415.1 1282 1325 33.38 61.91 3.95 6.80 49.58 101.30 46.04 93.76 3.80 8.34 7.68 8.04
pr107 107 4 47638.6 53725.9 45243 47540 1656.90 5115.19 5.29 13.01 65.95 113.65 64.59 98.51 1.41 12.26 2.11 15.37
pr124 124 4 70681.7 85259 67418 77200 2136.93 5414.06 4.84 10.44 76.06 135.20 75.28 122.38 0.54 10.82 1.03 10.47
pr136 136 4 120233.3 144288.5 115226 128099 3332.05 8293.75 4.35 12.64 84.35 159.94 82.70 140.00 1.81 19.12 2.00 14.24
pr144 144 4 80285.5 101878.6 75160 89002 4370.23 8710.08 6.82 14.47 90.98 185.87 86.99 162.14 2.59 23.93 4.59 14.64
kroA150 150 3 63218.7 83864.8 59283 78461 2896.72 2691.36 6.64 6.89 181.06 229.83 167.37 212.42 36.04 17.38 8.18 8.20
kroB150 150 3 60198.4 81029.8 54562 78160 3095.18 2463.38 10.33 3.67 160.77 249.30 157.41 214.25 3.52 33.72 2.14 16.36
pr152 152 4 87998.8 110102.3 81851 88351 3280.86 15575.98 7.51 24.62 98.20 184.31 97.13 168.95 0.77 17.32 1.10 9.09
u159 159 5 54682.9 68031.9 51632 62686 2094.34 3064.84 5.91 8.53 166.74 186.80 160.36 162.26 7.27 17.14 3.97 15.13
rat195 195 4 4448.3 5979.9 4233 5284 126.72 342.42 5.09 13.17 204.25 311.82 175.01 275.05 14.28 29.03 16.71 13.37
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887

tsp225 225 5 5045.3 6816.8 4649 6417 196.00 275.41 8.52 6.23 234.74 332.42 232.73 291.29 16.92 37.60 7.74 14.12

14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 885
6. Conclusions

6.1. Managerial insights

This work consists of three research questions: Why are clusters needed for waste collection?, Why
is optimization difficult, and how might we best suggest a heuristic?, and What are the contributions
of this research?’ First and foremost, vehicle routing can be carried out on a cluster basis instead
of on a bin basis by grouping bins into clusters, which reduces the complexity of the VRP by or-
ders of magnitude. However, applying only the k-means clustering algorithm has some limitations
in that this algorithm is based on only the location between nodes. Therefore, this paper intro-
duced sequential clustering using filled bin sets and their neighboring sets to reduce the subset
that a vehicle will visit. Second, it seemed possible to get an optimal solution to this problem.
According to Ramos et al. (2018) and Jorge et al. (2022), it took a long computation time to get
an optimal solution, which is far from a practical point of view. However, the proposed hybrid
heuristic suggests a good solution in a short time. Third, this paper first proposed the neighboring
bin concept considering the radius limitation of IoT sensors and this concept is implemented in a
mathematical model. In addition, this paper presented two operational management approaches
for smart waste collection. One of them is based on fill level estimation and the other one is based
on neighborhood. The first one showed good performance as for dynamic fill level information,
while the second one presented fast solutions as for computation times and considered practical
constraints.

6.2. Future study

This paper represents a vehicle routing problem for smart waste management. The above system is
configured with IoT-based smart bins. However, this paper has some limitations. First, we consider
only a single type of waste for simplicity. This may be extended to several types of waste within the
municipal waste system (e.g., food waste, recyclable waste, biodegradable waste). Second, multiple
compartments in a vehicle will facilitate collection of several types of waste. Furthermore, special
types of waste could be considered, such as medical waste, chemical waste, and industrial waste, to
broaden .CluVRP-WC model. The best waste management system has two aspects. One is efficient
waste collection and savings in transportation costs. The other is maximizing utilities for reusing
waste. Proper waste management is necessary to build a pollution-free smart city.

Acknowledgments

This research was supported by the National Research Foundation of Korea (NRF) funded by
the Ministry of Science, ICT, and Future Planning under Grant No. NRF-2019R1A2C2084616;
and the Brain Pool Program of the National Research Foundation of Korea (NRF) funded by the
Ministry of Science, ICT, and Future Planning under Grant No. NRF-2020H1D3A2A01085443.
We thank the associate editor and the two anonymous referees for their helpful comments.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
886 J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887
References

Adeleke, O.J., Ali, M.M., 2021. An efficient model for locating solid waste collection sites in urban residential areas.
International Journal of Production Research 59, 3, 798–812.
Al-Refaie, A., Al-Hawadi, A., Fraij, S., 2021. Optimization models for clustering of solid waste collection process. Engi-
neering Optimization 53, 12, 2056–2069.
Atzori, L., Iera, A., Morabito, G., 2017. Understanding the internet of things: definition, potentials, and societal role of
a fast evolving paradigm. Ad Hoc Networks 56, 122–140.
Battarra, M., Erdoğan, G., Vigo, D., 2014. Exact algorithms for the clustered vehicle routing problem. Operations Re-
search 62, 1, 58–71.
Coban, A., Ertis, I.F., Cavdaroglu, N.A., 2018. Municipal solid waste management via multi-criteria decision making
methods: a case study in Istanbul, Turkey. Journal of Cleaner Production 180, 159–167.
Defryn, C., Sörensen, K., 2017. A fast two-level variable neighborhood search for the clustered vehicle routing problem.
Computers & Operations Research 83, 78–94.
Defryn, C., Sörensen, K., Cornelissens, T., 2016. The selective vehicle routing problem in a collaborative environment.
European Journal of Operational Research 250, 2, 400–411.
Del Pia, A., Filippi, C., 2006. A variable neighborhood descent algorithm for a real waste collection problem with mobile
depots. International Transactions in Operational Research 13, 2, 125–141.
Delgado-Antequera, L., Caballero, R., Sánchez-Oro, J., Colmenar, J.M., Martí, R., 2020. Iterated greedy with variable
neighborhood search for a multiobjective waste collection problem. Expert Systems with Applications 145, 113101.
Eryganov, I., Šomplák, R., Nevrlỳ, V., Smejkalová, V., Hrabec, D., Haugen, K.K., 2020. Application of cooperative game
theory in waste management. Chemical Engineering Transactions, 81, 877–882.
Expósito-Izquierdo, C., Rossi, A., Sevaux, M., 2016. A two-level solution approach to solve the clustered capacitated
vehicle routing problem. Computers & Industrial Engineering 91, 274–289.
Fayoumi, A., Loucopoulos, P., 2016. Conceptual modeling for the design of intelligent and emergent information systems.
Expert Systems with Applications 59, 174–194.
Fernández, E., Roca-Riu, M., Speranza, M.G., 2018. The shared customer collaboration vehicle routing problem. Euro-
pean Journal of Operational Research 265, 3, 1078–1093.
Ferrer, J., Alba, E., 2019. Bin-ct: urban waste collection based on predicting the container fill level. Biosystems 186,
103962.
Ghiani, G., Guerriero, F., Improta, G., Musmanno, R., 2005. Waste collection in southern Italy: solution of a real-life
arc routing problem. International Transactions in Operational Research 12, 2, 135–144.
Haque, K.F., Zabin, R., Yelamarthi, K., Yanambaka, P., Abdelgawad, A., 2020. An IoT based efficient waste collection
system with smart bins. In 2020 IEEE 6th World Forum on Internet of Things (WF-IoT). IEEE, Piscataway, NJ, pp.
1–5.
Hartigan, J.A., Wong, M.A., 1979. Algorithm AS 136: a k-means clustering algorithm. Journal of the Royal Statistical
Society: Series C (Applied Statistics) 28, 1, 100–108.
Hintsch, T., Irnich, S., 2018. Large multiple neighborhood search for the clustered vehicle-routing problem. European
Journal of Operational Research 270, 1, 118–131.
Huang, S.H., Lin, P.C., 2015. Vehicle routing–scheduling for municipal waste collection system under the keep trash off
the ground policy. Omega 55, 24–37.
Jorge, D., Antunes, A.P., Ramos, T.R.P., Barbosa-Póvoa, A.P., 2022. A hybrid metaheuristic for smart waste collection
problems with workload concerns. Computers & Operations Research 137, 105518.
Kim, C., Yang, R., Lee, J., Jang, H., Ha, S., Kim, W., 2020. A study to find the optimal waste collection path. In Pro-
ceedings of the Korea Information Processing Society Conference, Korea Information Processing Society, Seoul, pp.
216–219.
Kim, J.S., 2015. The introduction of IoT (internet of things) technologies and policy directions. Review of Korea Contents
Association 13, 1, 18–24.
Lin, Z., Xie, Q., Feng, Y., Zhang, P., Yao, P., 2020. Towards a robust facility location model for construction and demo-
lition waste transfer stations under uncertain environment: The case of Chongqing. Waste Management 105, 73–83.

© 2023 International Federation of Operational Research Societies.

 14753995, 2025, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/itor.13282, Wiley Online Library on [20/01/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
J. Kim et al. / Intl. Trans. in Op. Res. 32 (2025) 863–887 887
López-Sánchez, A., Hernández-Díaz, A.G., Gortázar, F., Hinojosa, M.A., 2018. A multiobjective GRASP–VND algo-
rithm to solve the waste collection problem. International Transactions in Operational Research 25, 2, 545–567.
Maity, S., Roy, A., Maiti, M., 2016. An imprecise multi-objective genetic algorithm for uncertain constrained multi-
objective solid travelling salesman problem. Expert Systems with Applications 46, 196–223.
Peng, Y., Wu, P., Schartup, A.T., Zhang, Y., 2021. Plastic waste release caused by covid-19 and its fate in the global ocean.
Proceedings of the National Academy of Sciences 118, 47.
Pop, P.C., Kara, I., Marc, A.H., 2012. New mathematical models of the generalized vehicle routing problem and exten-
sions. Applied Mathematical Modelling 36, 1, 97–107.
Ramos, T.R.P., de Morais, C.S., Barbosa-Póvoa, A.P., 2018. The smart waste collection routing problem: Alternative
operational management approaches. Expert Systems with Applications 103, 146–158.
Reinelt, G., 1991. Tspliba traveling salesman problem library. ORSA Journal on Computing 3, 4, 376–384.
Ríos-Mercado, R.Z., González-Velarde, J.L., Maldonado-Flores, J.R., 2023. Tabu search with strategic oscillation for
improving collection assignment plans of waste electric and electronic equipment. International Transactions in Op-
erational Research 30, 1002–1030.
Roy, A., Manna, A., Kim, J., Moon, I., 2022. IoT-based smart bin allocation and vehicle routing in solid waste manage-
ment: a case study in South Korea. Computers & Industrial Engineering 171, 108457.
Sahebjamnia, N., Fathollahi-Fard, A.M., Hajiaghaei-Keshteli, M., 2018. Sustainable tire closed-loop supply chain net-
work design: hybrid metaheuristic algorithms for large-scale networks. Journal of Cleaner Production 196, 273–296.
Salamirad, A., Kheybari, S., Ishizaka, A., Farazmand, H., 2023. Wastewater treatment technology selection using a
hybrid multicriteria decision-making method. International Transactions in Operational Research 30, 1479–1504.
Sarkar, B., Debnath, A., Chiu, A.S., Ahmed, W., 2022a. Circular economy-driven two-stage supply chain management
for nullifying waste. Journal of Cleaner Production 339, 130513.
Sarkar, B., Ganguly, B., Pareek, S., Cárdenas-Barrón, L.E., 2022b. A three-echelon green supply chain management for
biodegradable products with three transportation modes. Computers & Industrial Engineering 174, 108727.
Sevaux, M., Sörensen, K., et al.2008. Hamiltonian paths in large clustered routing problems. In Proceedings of the
EU/MEeting 2008 Workshop on Metaheuristics for Logistics and Vehicle Routing, EU/ME, Vol. 8, pp. 411–417.
Tan, P.N., Steinbach, M., Kumar, V., 2016. Introduction to Data Mining. Pearson Education India, New Delhi.
Toth, P., Vigo, D., 2002. The Vehicle Routing Problem. Monographs on Discrete Mathematics and Applications. SIAM,
Philadelphia, PA.
Zhao, J., Zhu, F., 2016. A multi-depot vehicle-routing model for the explosive waste recycling. International Journal of
Production Research 54, 2, 550–563.

© 2023 International Federation of Operational Research Societies.