Cleaner Logistics and Supply Chain 12 (2024) 100164

Contents lists available at ScienceDirect

Cleaner Logistics and Supply Chain

journal homepage: www.journals.elsevier.com/cleaner-logistics-and-supply-chain

A practical vehicle routing problem in small and medium cities for fuel
consumption minimization
Yong-Jin Pak * , Kyong-Ho Mun
Faculty of Electronics and Automation, Kim Il Sung University, Ryongnam-Dong, Taesong-District, Pyongyang, Democratic People’s Republic of Korea

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

Keywords: Today, reducing fossil fuel consumption in the transport sector is one of the important requirements to preserve
Fuel consumption the ecological environment. In this paper, we investigated a logistics problem in small and medium cities. We
Vehicle routing problem proposes the time-dependent vehicle routing problem model in small and medium cities considering heteroge-
Variable neighborhood search
neous fleets, time windows, multi-trip, time dependency, and road network. We conducted a numerical analysis
Time-dependent speed
Road network
on fuel consumption minimizing path selection between two customer nodes and found that the path selection is
independent of the payload in most cases. The model is formulated as a mathematical integer linear program-
ming model, and small problems are solved using a commercial optimization solver. For large-scale instances, we
proposed a heuristic algorithm based on the variable neighborhood search and evaluated its performance. The
proposed algorithm can decrease fuel consumption by 25 % compared to manual routing and scheduling.
Furthermore, for provide useful managerial insights, sensitivity analyses were performed considering different
vehicle capacities.

1. Introduction several variants, more accurately reflecting the characteristics of the

real-life problem. VRP has many applications in reality. Parvasi et al.
Today, reducing fossil fuel consumption in the transport sector is one (2017, 2019) solved a school bus routing problem using VRP. Vidal et al.
of the important requirements to preserve the ecological environment. (2020) provided a comprehensive survey on VRP variants.
This work originated from the transport problem in small and me- Efforts for environmental protection in transport fields, such as
dium cities(SMCs). Cities are often classified by population, and SMCs reduction of greenhouse gas emissions, reduction of fossil fuel con-
refer to cities with populations smaller than 2.5 million (Bofeng et al., sumption, and use of electric vehicles, have generated research streams
2019). In the case of China, 88 cities of 146 cities are SMCs (Jia et al., such as GVRP (Green VRP), EVRP(Electric VRP) and PRP (Pollution
2019). Here, we consider the diameter of SMC to be smaller than 30 km, Routing Problem), new fields of VRP (Xiao et al., 2012; Koc et al., 2016;
which means in SMCs, it does not take 1 h to travel between any two Demir et al., 2014; Niu et al., 2017; Cinar et al., 2016; Bektas and
customer sites in the city. For SMCs, we can assume some characteristics Laporte, 2011; Bruglieri et al., 2023; Ferreira et al., 2024). A review of
related to the city’s scale and can reduce the complexity of the transport GVRP literature can be found in Garside et al. (2024) and Asghari et al.
problem. Those assumptions are explained in the Section 3.3. (2021).
Our research aims to solve the practical transport problem in small As environmental pollution problems caused by automotive exhaust
and medium cities. For this purpose, we should consider heterogeneous emissions in urban areas become serious, extensive studies have been
fleets, time windows, multi-trip, time dependency, and road networks. conducted to optimize logistics problems in urban areas. Cattaruzza
However, there are relatively few studies dealing with these character- et al. (2017) reviewed the literature on VRPs to solve city logistics
istics jointly. problems and concluded that the main challenges with VRP for city
In this study, an operational research(OR) approach to optimize logistics are time-dependency, multi-level and multi-trip organization of
transport in urban areas was proposed. Traditional OR problems the distribution, dynamic information, and surveyed the research results
modeling transport management problems are Vehicle Routing Prob- for each issue.
lems(VRP) (Dantzig and Ramser, 1959). VRP has been specialized into The contributions of this paper are as follows. (1) For small and

* Corresponding author.
E-mail address: [email protected] (Y.-J. Pak).

https://doi.org/10.1016/j.clscn.2024.100164
Received 15 April 2024; Received in revised form 6 July 2024; Accepted 13 July 2024
Available online 19 July 2024
2772-3909/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

medium cities, we conducted a numerical analysis on the optimal path (Figliozzi, 2011; Yao et al., 2015; Xiao and Konak, 2015; Franceschetti
selection in the meaning of fuel consumption minimization. We found et al., 2016; Hatami et al., 2020; Liu et al., 2020; Pan et al., 2021; Guo
that the optimal path selection is independent of the payload in most et al., 2021; Yeh and Tan, 2021; Vincent et al., 2021).
cases. (2) We formulated the MILP model for the practical vehicle In all the above studies, the optimal path between two nodes was
routing and scheduling problem in small and medium cities, which assumed to be given. In urban road networks, however, there is a need
considers all important characteristics arising in urban logistics, such as for path selection between two nodes considering the state of the urban
heterogeneous fleet, time window, multi-trip, time dependency, and road network over time, especially for the travel time minimization
road network. The above-mentioned path selection independence and problem or fuel consumption minimization problem rather than the
the condition on the size of the city reduce the problem dimension distance minimization problem. In this perspective, studies considering
significantly. (3) We designed the variable neighborhood search algo- urban road networks were also conducted (Garaix et al., 2010; Lai et al.,
rithm for our problem and examined its performance using randomly 2016; Hosseini-Nasab and Lotfalian, 2016; Ticha et al., 2017; Behnke
generated instances. For search efficiency, we designed a new adaptive and Kirschstein, 2017).
neighborhood structure. Our result will provide logistics designers and Studies that considered road network and time dependency simul-
executors in small and medium cities with useful insight and an actual taneously have also been conducted.
methodological approach. Qian and Eglese (2016) formulated fuel emissions optimization VRP
The remainder of this paper is organized as follows. Section 2 pre- on a graph in which a set of arcs connecting customer nodes exists for
sents a brief review of related works. Section 3 gives some preliminaries each pair of customers. They divided the time horizon for the delivery
for further discussion, and Section 4 provides an MILP model for the operation into time slots. They assumed that for each road segment, the
problem. The meta heuristic algorithm to solve a proposed problem is maximum speed allowed by the traffic conditions is given for each time
explained in Section 5, followed by computational results in Section 6. slot. In their model, the speed of the vehicle on each road segment and
Finally, the conclusion is discussed in Section 7. the waiting time at each customer are considered as decision variables.
They proposed a column generation based Tabu search algorithm and
2. Related works concluded that about 3 % of GHG emissions can be reduced compared to
an approach where the objective is to minimize total time. In their work,
This work examines a variant of the pollution routing problem that however, a fleet of homogeneous vehicles is considered, and multi–trip
addresses both routing and scheduling decisions while taking fuel con- and load dependent fuel emissions are not taken into account.
sumption into account. Androutsopoulos and Zografos (2017) proposed a Bi-objective Time
For the estimation of fuel consumption, a Comprehensive Modal and Load Dependent Vehicle Routing Problem with Time Windows
Emission Model (CMEM) Barth et al. (2004) is widely used in research (BTL-VRPTW). They used two objective functions that present the
that addresses green, pollution, and sustainable routing-based problems. minimization of the total duration of the routes and total fuel con-
Koc et al. (2014) examined the fleet size and mix pollution-routing sumption. They proposed a travel time model in which the average
problem. They extended the pollution-routing problem by considering travel speed on any link of the road network is modeled as a continuous
a heterogeneous vehicle fleet. They have developed a evolutionary piecewise linear function of the time of the day. They showed that in
metaheuristic for realistic benchmark instances. Cheng et al. (2017) their problem, the scheduled road paths between consecutive customer
introduced a green inventory routing problem considering environ- nodes included in a non-dominated solution could not be determined in
mental impacts and heterogeneous vehicles. They proposed a compre- advance. To cope with this property, they proposed a new network
hensive objective function that considered inventory costs and routing reduction approach that defines eligible road paths for each pair of
costs. The routing cost contains fuel and emission costs. customer nodes and examined the effectiveness of network reduction.
Karakostas et al. (2020) proposed the Fleet-size and Mix Pollution However, they used the path length as a criterion for selection, and
Location-Inventory-Routing Problem with Just-in-Time replenishment solution quality and computational effort depend on the selection of the
policy and Capacity Planning and proposed a General Variable Neigh- value of K. Even though they considered load-dependent fuel con-
borhood Search (GVNS)-based framework to solve that problem effi- sumption explicitly, they did not consider a heterogeneous fleet and
ciently. They introduced a novel adaptive shaking method and the multi-trip.
efficiency was proved by intensive numerical experiments. Karakostas Huang et al. (2017) proposed the deterministic/stochastic Time-
et al. (2024) introduced a multi-period, multi-commodity, inventory- Dependent Vehicle Routing Problem with Path Flexibility(TDVRP-PF)
routing problem with strategic fleet scheduling decisions. This prob- and derived some conclusions. They presented multiple paths between
lem considers speed limits as well as strict European Union regulations two customer nodes as path flexibility and flexible departure time at the
on truck drivers’ working and driving time. They proposed a MILP depot, and post-service waiting time at customer sites as time flexibility.
model and conducted several sensitivity analyses on key model pa- Using randomly generated benchmark problems on the real road
rameters. Karakostas et al. (2024) presented the Pollution Traveling network of Beijing, China, they confirmed that TDVRP-PF results in
Salesman Problem with Refueling. This integrates the Pollution Trav- significant savings in terms of cost and fuel consumption compared to
eling Salesman Problem and the Traveling Salesman Problem with TDVRP, and the role of path flexibility is more significant than the role
Refueling which are recently proposed variants of the Traveling of time flexibility. For numerical experiments, they divided the time
Salesman Problem. They also proposed a two-stage double adaptive horizon into 3 time periods and sorted road segments into 9 types by
general variable neighborhood search method. They verified that their speed limits in each time period. They used a modified Dijikstra’s al-
methodology consistently yields high performance. Experiments were gorithm to find the time-dependent shortest path between customer
conducted on TSPLib instances against the methodologies in the open nodes efficiently. They acknowledged the need for efficient algorithms
literature. to solve large-scale instances. Their approach was very realistic and
Lai et al. (2024) considered a joint pollution-routing problem with comprehensive. However, they did not consider the heterogeneous fleet
time windows and speed optimization (PRP-SO) and proposed a tailored and multi-trip.
branch-and-price (BP) algorithm. They found that fuel savings of up to Ehmke et al. (2016) proposed a load-dependent vehicle routing and
53 % can be achieved when considering payloads and road states. scheduling model that explicitly considers the path selection between
One of the important features of urban VRP is the time dependency customer nodes. They argued that the emission optimal path between
due to congestion in urban traffic networks. Malandraki and Daskin customer nodes cannot be pre-calculated because the vehicle load on the
(1992) first formulated the Time-Dependent Vehicle Routing Problem path between two customer nodes is not known until the routing deci-
(TDVRP), and the work to solve several TDVRP variants is ongoing sion is made. They found that if any path is optimal for both empty truck

2
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

and full loaded truck, then the path is optimal for all payloads. Using this Table 2
finding, they simplified the pre-computation phase and proposed a Tabu Model coefficients.
search algorithm to solve large-scale problems. However, they did not light-duty medium-duty
consider time windows, heterogeneous fleet, and multi-trip in their
wv1 0.001214 0.001404
model. wv2 0.0000261 0.0000411
The comprehensive model that considers the road network, time- wv3 7.47E-09 7.47E-09
dependent speed, load-dependent fuel consumption, heterogeneous wv4 1.93E-07 2.57E-07
fleet, and multi-trip simultaneously has been formulated by Raeesi and
Zografos (2019). They introduced their model as the Steiner Pollution-
several road segments, and each road segment belongs to one road class.
Routing Problem (SPRP). They analyzed the above last 4 papers and
There are many paths between any two customers. Fig. 1 shows only 3
pointed out the need for the exact identification of the full set of the
paths among many possible paths.
eligible emissions minimizing road-paths between customer nodes on a
The road segments have a time-dependent speed limit due to traffic
time-dependent graph. To solve that problem, they proposed an efficient
congestion and the safety rule. In Fig. 2, the speed limit in the k-th period
exact Path Elimination Procedure (PEP) and verified PEP’s effectiveness
is sk.
over other methods. SPRP fits well to our purpose except for the
For the sake of simplicity, we classify the functions of speed limit
following issues. First, there is no efficient algorithm to solve realistic
over time into several types. We call this a speed limit profile.
large-scale problems. They examined the problems with 10 customers.
Let us denote the set of road segments in path pij between two cus-
Second, their model does not allow waiting after customer service at the
customer site. tomers i, j which belong to m-th speed limit profile as pm ij and the total

The conclusion is that there is no practical research for logistics in length of a set of road segments x as d(x). Then the following relation
small and medium cities and that SPRP is too complicated to find and holds naturally.
evaluate a good solution. ( ) ∑ ( )
d pij = d pm
The following table summarizes the comparison of the existing ij
m
models and our model (Table 1).
The fuel consumption consumed when vehicle v carries the load of q
3. Problem definition through a path pij in time period k can be calculated as follows.
( ) ∑ ( ) ( )
C pij , k, v, q = d pmij FC v, sk , q
m
(2)
In this section, we provide a fuel consumption model and a speed m
limit profile that we have adopted is adopt in our work. And based on
some assumptions, we analyze the optimal path selection under different The optimal path p*ij to be selected when vehicle v carries the load of q
payloads. between customer i, j in time period k is as follows.
( )
p*ij (k, v, q) = argminC pij , k, v, q (3)
3.1. Fuel consumption model pij

We used the Comprehensive Modal Emission Model (CMEM), Barth 3.3. Assumptions
et al. (2004) and Barth and Boriboonsomsin (2008) to estimate fuel
consumption. Using the CMEM, when vehicle v runs in speed s on In this paper, we discuss the vehicle routing problem for small and
average with payload q, the fuel consumption per kilometer is calculated medium cities, based on the following assumptions.
as follows.
/ / • A fleet consists of light and medium-duty trucks, and each truck has
FC(v, s, q) = wv1 s + wv2 + wv3 • q + wv4 s2 (1)
different fuel consumption characteristics.
Here, wv1 , wv2 , wv3 , wv4 are the coefficients related to the vehicle type. Using • The transport demand of each consumer is given in advance, and
the same parameters from Niu et al. (2017), these coefficients is calcu- each consumer has a constraint on a time window in which they can
lated as Table 2. be served.
• Vehicles can perform multiple trips and serve several customers in
one trip.
3.2. Road network and speed limit profile • Vehicles are allowed to wait at the depot and customer sites.

As shown in Fig. 1, a path from customer i to customer j consists of

Table 1
Model comparison.
Qian and Eglese Androutsopoulos and Zografos Huang et al. (2017) Ehmke et al. Raeesi and Zografos Our model
(2016) (2017) (2016) (2019)

Model Name BTL-VRPTW TDVRP–PF SPRP FCVRSP-

SMC
Time-dependency ○ ○ ○ ○ ○ ○

Load-dependency ○ ○ ○ ○ ○

Time-windows ○ ○ ○ ○ ○

Heterogeneous fleet ○ ○

Road network ○ ○ ○ ○ ○ ○

multi-trip ○ ○

Waiting at customer site ○ ○ ○ ○

Dimension of decision variable 5 4 5 4

in MILP
Solution Approach for realistic Tabu Search Ant Colony System Route-Path Tabu Search − VNS
problem approximation

3
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

Fig. 1. Road network.

• Travel time between two customers is relatively short compared to

the length of the period, and travel between any pair of customers is
finished within one period.

The last assumption is introduced to simplify the problem in small

and medium cities. Considering road congestion dynamics in small and
medium cities, a time period can be assumed to be longer than one hour.
Considering the city dimension, the travel time between any pair of
customers can be assumed to be shorter than one hour.

3.4. Optimal path selection between two customers

Fig. 2. Speed limit on a road segment.
Assuming that the two paths rij , sij between two customers i, j consist
of road segments belong to three speed limit profiles (Fig. 3), we con-
ducted the computational experiments for the optimal path selection
according to the payload.
Given the vehicle type v and period (period is equivalent to speed
limits), we analyzed the ratio of the cases that optimal path (in the
meaning of fuel consumption minimization) is dependent on the
payload i.e. p*ij (k, v, q1 ) ∕
= p*ij (k, v, q2 ) to all possible cases. Here the case is
( ) ( ) ( ) ( ) ( ) ( )
specified by path configuration d rij1 , d rij2 , d rij3 , d s1ij , d s2ij , d s3ij
and payload q1 , q2 .
∑ ( ) ( ) ( ) ( ) ( ) ( )
1
d rij1 , d rij2 , d rij3 , d s1ij , d s2ij , d s3ij , q1 , q2
p* (k, v, q1 ) ∕
= p* (k, v, q2 )
Ndiff = ∑ ( ) (ij ) ( ) ( )ij ( ) ( )
1
d rij1 ,d rij2 ,d rij3 ,d s1ij ,d s2ij ,d s3ij ,q1 ,q2

We also calculated the average differences in fuel consumption be-

Fig. 3. Two paths consisting of 3 road segments.
tween two paths as follows.

∑( ) ( ) ( ) ( ) ( ) ( ) ⃒ ( ) ( )⃒
⃒ ⃒
⃒C p*ij (k, v, q1 ), k, v, q1 − C p*ij (k, v, q2 ), k, v, q1 ⃒
d rij1 ,d rij2 ,d rij3 ,d s1ij ,d s2ij ,d s3ij ,q1 ,q2
Fdiff = ∑( ) ( ) ( ) ( ) ( ) ( ) ( )
C p*ij (k, v, q1 ), k, v, q1
d rij1 ,d rij2 ,d rij3 ,d s1ij ,d s2ij ,d s3ij ,q1 ,q2

4
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

Table 3 SMC) is formulated as follows.

Ratio of cases that optimal path depends on payload. Indexes
k: time period
Speed limit (m/s) light-duty medium-duty
v: vehicle
s1k s2k s3k Ndiff (%) Fdiff (%) Ndiff (%) Fdiff (%) i, j, m: customer, 0 means the depot.
Parameters
5 7 10 0.3601 0.443029 0.53232 0.690998 V: Number of vehicles
5 7 13 0.4804 0.497681 0.650493 0.773212 F: Set of vehicles, F = {1, 2, …, V }
5 7 15 0.4139 0.549837 0.650493 0.772787 pi: process time at customer i, p0: process time at the depot
5 7 20 0.3524 0.465253 0.496307 0.648178 Di: demand of customer i
5 10 13 0.4693 0.714367 0.785547 0.960274 Qv: capacity of vehicle v
5 10 15 0.5356 0.696476 0.786987 0.958956 [bk, ek]: k-th time period
5 10 20 0.2427 0.856221 0.779187 0.790676 [gi, hi]: time window of customer i
5 13 15 0.574 0.867806 0.483013 1.338884 Sijkv: load independent fuel consumption when vehicle v leaves customer i in period k
5 13 20 0.4971 0.705612 0.707227 0.977642 to visit customer j
5 15 20 0.4862 0.740568 0.707227 0.977798 Lijkv: load independent fuel consumption per unit load when vehicle v leaves customer
7 10 13 0.273 0.335526 0.361867 0.545192 i in period k to visit customer j
7 10 15 0.2685 0.352115 0.361867 0.545922 tijk: time to travel from customer i to customer j in period k
7 10 20 0.2085 0.279266 0.215947 0.444849 Decision variables
7 13 15 0.4867 0.37373 0.4128 0.591924 xijkv: 1 if vehicle v leaves customer i in period k to visit customer j, else 0
7 13 20 0.2474 0.359179 0.318307 0.515854 xijk: 1 if some vehicle leaves customer i in period k to visit customer j, else 0
7 15 20 0.2802 0.370575 0.318307 0.515575 xij: 1 if some vehicle leaves customer i to visit customer j, else 0
10 13 15 0.0000 0.0000 0.0000 0.0000 li: leaving time from customer i
10 13 20 0.0000 0.0000 0.0000 0.0000 ai: arriving time at customer i
10 15 20 0.0000 0.0000 0.0000 0.0000 qijkv: vehicle load when vehicle v leaves customer i in period k to visit customer j
13 15 20 0.0000 0.0000 0.0000 0.0000 zij: 1 if a trip starting with customer j started just after a trip ending with customer i,
else 0
uv: 1 if vehicle v is used for transport, else 0
Given the vehicle type v and period, six road segment lengths are
changed from 0 km to 10 km and two payloads are changed from 0 to
4,000 kg for the light-duty truck and from 0 to 10,000 kg for the
medium-duty truck. Objective function
In the case of light-duty truck, the ratio(Ndiff ) is the highest with ∑∑ ∑ ∑
0.5356 % when speed limits are 5, 10, and 15 m/s, and the relative Minimize Sijkv • xijkv + Lijkv • qijkv (5)
i∈N j∈N 1≤k≤K v∈F
difference in fuel consumption in the two paths(Fdiff ) is the highest with
0.86 %. In the case of medium-duty truck, the maximum ratio is 0.78 % Subject to
and the relative difference is 1.33 %. Furthermore, last four rows of ∑
Table3 we can see that when speed limits are in the range from 10 m/s to xij = 1, ∀i ∈ N\{0} (6)
j∈N
20 m/s, optimal path selection is identical for all payloads. In other
words, when a truck runs faster than 36kph (kilometers per hour) and ∑
xij = 1, ∀j ∈ N\{0} (7)
slower than 72kph, the fuel consumption minimization path between i∈N
any two points is independent of the payload.
∑
From this observation, we can assume that the following relation xij = xijk , ∀i, j ∈ N (8)
holds. 1≤k≤K

∀i, j, k, v, q1 , q2 : p*ij (k, v, q1 ) ≅ p*ij (k, v, q2 ) (4) ∑

xijk = xijkv , ∀i, j ∈ N, 1 ≤ k ≤ K (9)
v∈F
Compared to the previous work related to optimal route selection
(Ehmke et al., 2016; Raeesi and Zografos, 2019), our result is for the ∑
xijkv ≤ 1, ∀i, j ∈ N, ∀v ∈ F (10)
special case that each journey between any two customers completes 1≤k≤K
within one period. In that case, we can say that the optimal path for an
∑∑ ∑∑
empty truck (or full-loaded truck) is optimal for all payloads. This re- xijkv = xjmkv , ∀j ∈ N\{0}, ∀v ∈ F (11)
flects the characteristics of logistics in small and medium city. i∈N 1≤k≤K m∈N 1≤k≤K

∑ ∑∑ ∑ ∑∑
4. Problem formulation qijkv − qjmkv = Dj , ∀j ∈ N\{0} (12)
1≤k≤K v∈F i∈N 1≤k≤K v∈F m∈N

From the analysis in Section 3, it can be seen that using only the
optimal path between two points determined for an empty vehicle, we xijkv • Dj ≤ qijkv ≤ xijkv •(Qv − Di ), ∀i, j ∈ N, ∀v ∈ F, 1 ≤ k ≤ K (13)
can find a solution that is almost optimal to the problem considering all
paths. xijk = 1→li , aj ∈ [bk , ek ], ∀i, j ∈ N\{0}, 1 ≤ k ≤ K (14)
Since the fuel consumption minimizing path between two customer
xi0k = 1→li , li + ti0k ∈ [bk , ek ], ∀i ∈ N\{0}, 1 ≤ k ≤ K (15)
nodes in the road network graph can be easily obtained using the well-
known Dijikstra’s algorithm, we can now consider that the amount of
x0jk = 1→aj , aj − t0jk ∈ [bk , ek ], ∀i ∈ N\{0}, 1 ≤ k ≤ K (16)
fuel consumption per unit distance per unit load between two customers
according to time period and vehicle type is given. ∑
xij = 1→aj = li + tijk • xijk , ∀i ∈ N\{0}, 1 ≤ k ≤ K (17)
Under the assumptions in Section 3.3, the fuel considering vehicle 1≤k≤K
routing and scheduling problems in small and medium cities(FCVRSP-

5
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

∑
li + ti0k • xi0k ≤ eK , ∀i ∈ N\{0} (18) 5. Variable neighborhood search algorithm
1≤k≤K

∑ Variable neighborhood search (VNS) was originally proposed by

t0jk • x0jk + p0 ≤ aj , ∀j ∈ N\{0} (19) Mladenovic and Hansen (1997) and has been used to solve many
1≤k≤K combinational optimization problems. The main feature of VNS is to
change the neighborhood systematically in the search process, and many
ai + pi ≤ li ∀i ∈ N\{0} (20) researchers adopt this approach to solve difficult problems in the field of
VRP (Todosijevic, 2016; Wei et al., 2015; Wassan et al., 2015).
gi ≤ ai ≤ hi ∀i ∈ N\{0} (21)
Key components of VNS are as follows:
∑ Initialization: The algorithm begins with an initial solution, which
zij ≤ 1, ∀i ∈ N (22)
can be generated randomly or using a heuristic method. This initial
j∈N
solution serves as the starting point for the search process.
∑ Neighborhood Definition: In VNS, a neighborhood is defined as a
zij ≤ 1, ∀i ∈ N (23)
i∈N set of solutions that can be obtained by applying a specific perturbation
rule to the current solution. For example, in the context of the traveling
2zij ≤ xi0 + x0j , ∀i, j ∈ N\{0} (24) salesman problem, a neighborhood might consist of all solutions ob-
∑∑ tained by swapping two cities in the current tour.
uv ≤ xi0kv ≤ |N| • uv , ∀v ∈ F (25) Local Search: The local search phase involves finding the best so-
i∈N 1≤k≤K lution within a given neighborhood. This is typically done using a
∑ ∑∑ ∑ greedy algorithm or another optimization technique. The goal is to
x0j − zij = uv , ∀j ∈ N\{0} (26) improve the current solution by exploring its immediate vicinity.
i∈N i∈N j∈N v∈F
Variable Neighborhood Change: This is the distinguishing feature
∑ ∑ of VNS. After a certain number of iterations or when a predefined con-
zij − 1 ≤ xi0kv − x0jkv ≤ 1 − zij , i, j ∈ N\{0}, ∀v ∈ F (27) dition is met, the algorithm changes the neighborhood structure by
1≤k≤K 1≤k≤K
applying a different perturbation rule. This allows the search to explore
∑ ∑ new regions of the solution space and potentially find better solutions.
zij = 1→li + ti0k • xi0k + t0jk • x0jk + p0 ≤ aj , ∀i, j ∈ N (28)
Acceptance Criteria: To ensure that the algorithm does not get stuck
1≤k≤K 1≤k≤K
in local optima, VNS uses acceptance criteria to decide whether to
xijkv , zij , uv ∈ {0, 1}, ai , li , qijkv ≥ 0, ∀i, j ∈ N, ∀v ∈ F, 1 ≤ k ≤ K (29) accept a new solution. Common criteria include the Metropolis criterion,
which accepts worse solutions with a certain probability, or the first
Objective function (5) represents fuel consumption minimization. improvement strategy, which accepts the first solution that is better than
This is calculated by summing load − independant and dependent fuel the current one.
consumptions for all road segments. Termination Criterion: The algorithm terminates when a stopping
Constraints (6)− (11) are typical vehicle flow conservation con- condition is met, such as reaching a maximum number of iterations,
straints. Constraints (12) are the commodity flow conservation con- achieving a satisfactory solution quality, or exhausting the available
straints. The first item represents the amount of incoming goods into a computational resources.
customer node, and the second represents the amount of outgoing goods The steps of Basic VNS (BVNS) method are given in Algorithm 1
from that customer node. Therefore, the difference between these two (Gendreau et al., 2010).
items means the amount served at that customer node. Algorithm 1: Basic VNS
Constraints (13) represent vehicle capacity. Constraints (14)− (16)
Function VNS (x, kmax, tmax)
are constraints related to time periods. Constraints (14) guarantee both 1: repeat
the leaving time and arriving time of the travel from customer i to 2: k ← 1
customer j are in time period k. Constraints (15)− (16) are conditions for 3: repeat
the trips from the depot or to the depot. Here the condition x = 1→a ∈ 4: x′ ← Shake(x, k) // Shaking
5: x″ ← BestImprovement(x′) // Local search
[b, e] can be transformed to the following widely used formation in
6: x ← NeighbourhoodChange(x, x″, k) // Change neighbourhood
MILP. 7: ′until k = kmax
8: ′t ← CpuTime()
b − (1 − x) • M ≤ a ≤ e + (1 − x) • M 9: until t > tmax
M is a sufficiently large number.
Constraints (17)− (20) reflect arriving and leaving time. Constraints
(17) forbidden waiting on the road. Constraints (18)− (19) are timing
5.1. Overview of VNS algorithm
conditions on arrival and departure to and from the depot. Constraints
(20) allow waiting at customer sites.
Here we describe details of VNS algorithm that we use. The main
Constraints (21) reflect the time window for customer service.
flow of the VNS algorithm consists of initial solution generation,
Constraints (22)− (28) are related to multi-trips. Constraints (28)
neighborhood change, and local search. After the initial solution is
represent a time limit for two subsequent trips to pass through the depot.
generated, VNS repeats the improvement of the solution by neighbor-
Constraints (29) are variable specification constraints.
hood change and local search while the stopping condition is not
One superiority of FCVRSP-SMC in comparison to SPRP is the small
satisfied (Algorithm 2). It randomly generates the neighborhood solu-
dimension. The decision variables of SPRP is 5 dimensional, whereas
tion of the current solution by changing the neighborhood one by one
ours is 4 dimensional. It has a great effeteness on the solution efficiency.
(line 7) and finds a local optimal solution starting from that solution
using local search (line 8). In line 6, K is the number of neighborhoods. It
checks whether this local optimal solution is better than the best solu-
tion found so far and updates the best solution (lines 9 to 12).

6
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

A slightly different point from the traditional VNS algorithm is that it

expands neighborhood structure using the solution update counter
Nnonimp in the search process. The main idea is to expand the neigh-
borhood structure to escape from local optima if the solution is not
updated for a long time (See Section 5.3).
Algorithm 2. VNS

1: X←InitSolution()
2: X* ← X, Nnonimp ← 0
3: while the stop condition is not satisfied
Fig. 4. Solution representation.
4: Nnonimp ← Nnonimp + 1
5: k ← 1
the depot. Several zeros are added for multiple trips that may be
6: while k ≤ K
7: X1 ← Neighbour(X, k, Nnonimp)
included in the solution in the future. The solution also includes waiting
8: X2 ← LocalSearch(X1) times at each customer site and depot.
9: if X2 is better than X*
10: k ← 0, Nnonimp ← 0
11: X*←X2, X←X2 5.3. Adaptive neighborhood structure
12: end if
13: k←k+1 Adaptive Neighborhood structure is a novel methodological feature
14: end while
within VNS. Adaptive Variable Neighborhood Search (AVNS) ap-
15: end while
16: return X* proaches have been widely used in solving routing and logistics opti-
mization problems due to their effectiveness and flexibility. These
approaches are designed to dynamically adapt to the problem at hand,
making them particularly suitable for complex and dynamic environ-
5.2. Initial solution generation
ments. One notable example is the application of AVNS to the Dynamic
Vehicle Routing Problem (DVRP). As described in Jeeu Fong Sze et al.
The initial solution is generated based on the nearest neighbor-first
(2024), an effective adaptive VNS (AVNS) was proposed to address the
heuristic (Algorithm where NS represents a set of customers that are
DVRP. This approach demonstrated the flexibility of AVNS as a powerful
not scheduled, V represents a set of vehicles, and t represents the current
re-optimization tool, capable of handling dynamic changes in real-time.
time. After the vehicle is selected (line 2, line 12), the process of
The study introduced critical nodes as potential diversion points and
customer selection is repeated to serve as many customers as possible.
presented new rules to construct dynamic information, reporting useful
Among the customers satisfying the constraints on the vehicle capacity
results.
and time window, the customer with the least fuel consumption from the
AVNS has been applied to the Vehicle Routing Problem with Inter-
current location is selected (line 4 ~ line7).
mediate Stops (VRPIS). Schneider et al. (2014) developed an adaptive
In some cases, the generation of the initial solution fails because the
variable neighborhood search (AVNS) to solve the VRPIS, considering
time window limitation and vehicle capacity limitations of customers
stopping requirements at intermediate facilities. This approach was
must be considered simultaneously. In order to handle such a case, we
shown to be highly effective in numerous practical vehicle routing
give a little opportunity the non-nearest customer is selected (line 6) and
applications.
repeat the initial solution generation until a feasible solution is gener-
Overall, adaptive VNS approaches have proven to be valuable tools
ated. The results of the computational experiments showed that the
in solving routing and logistics optimization problems, offering robust
number of repetitions of initial solution generation is usually between 1
and adaptable solutions to complex real-world challenges.
and 7 when the random selection probability Ps is equal to 0.9.
In this study, we used two intra-vehicle, two inter-vehicle, and one
Algorithm 3. InitSolution
waiting time adjustment neighborhood generation operations.
1: while a feasible solution is not found R1: Inter_move
2: t ← 0, NS←V, v ← 1
3: while NS ∕ =∅
This operation moves one customer to a location in the other vehicle.
4: Find customers CS in NS which vehicle v can serve from time t. Depending on the customer’s new location, a new trip can be generated.
5: if CS∕=∅ R2: Inter_exchange
6: Find a customer c in CS that fuel consumption is minimal when vehicle v This operation swaps the positions of two customers from two
serves c from time t with probability Ps.
different vehicles.
7: Schedule c with vehicle v, increase t. NS←NS \ {c}
8: else R3: Intra_move
9: if vehicle v is full This operation moves one customer to a new location in the same
10: Schedule vehicle v to return to the depot and start a new trip. vehicle. Depending on the customer’s new location, a new trip can be
11: if t = TL generated.
12: v ← v + 1, t ← 0
13: end if
R4: Intra_exchange
14: Continue loop. Go to line 3 This operation swaps the positions of two customers in one vehicle.
15: else if t < TL R5: Delay
16: t ← t + 0.1
17: Continue loop. Go to line 3
18: else Table 4
19: Retry. Go to line 1 Neighborhood expanding scheme.
20: end if Neighborhood Step Step 2 Step 3 Step 4 Step 5
21: end if Structure 1
22: end while
23: return with the generated schedule. N1 R1 R1 ⊕ R1 ⊕ R 2 R1 ⊕ R2 ⊕ R1 ⊕ R 2 ⊕ R 3
24: end while R2 ⊕ R3 R3 ⊕ R4 ⊕ R 4 ⊕ R5
N2 R2 R2 R2 ⊕ R 3 R2 ⊕ R3 ⊕ R2 ⊕ R 3 ⊕ R 4
R4 ⊕ R5
An example of the generated initial solution for the case that 3 vehicles N3 R3 R3 R3 R3 ⊕ R4 R3 ⊕ R 4 ⊕ R 5
serve eight customers is shown in Fig. 4. This shows the customer N4 R4 R4 R4 R4 R4 ⊕ R 5
N5 R5 R5 R5 R5 R5
numbers visited by each vehicle in the order of visit, and zeros represent

7
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

This operation changes the waiting time at depot or customer sites. It Table 5
increases the delay time by a random value from − 1.5 to 1.5. Parameters of test instance sets.
In order to increase the efficiency of the search, we proposed an Parameter TS1 TS2 TS3
adaptive neighborhood structure that enlarges the range of the neigh-
Number of customers 10 20 100
borhood in the case of a long time stay at the local optimum solution. Number of light-duty 1, 4000 2, 4000 5, 4000
Nnonimp is the number of repetitions without improvement. As Nnonimp vehicles, Capacity
increases, the neighborhood is expanded step by step (Table 4). A (kg)
neighborhood structure is first generated by one corresponding neigh- Number of medium- 1, 10,000 2, 10,000 8, 10,000
duty vehicles,
borhood generation operation and then is expanded as follows when Capacity(kg)
Nnonimp reaches certain rates (e.g., 0.2, 0.4, 0.6, 0.8) of MAXnonimp. Number of speed limit 3 3 3
Here ⊕ is a superposition operator of the neighborhood generation profiles
operations. R1 ⊕ R2 is the operation that generates Neighbors by Time periods [0,2],[2,6],[6,8] [0,2],[2,6],[6,8] [0,2],[2,6],[6,8]
Speed limits(m/s) (5,8,10), (5,8,10), (5,8,10),
applying the R2 operation to each neighbor generated by R1. As it can be
(5,12,20), (5,12,20), (5,12,20),
seen from the table, the neighborhood N5 by the time delay is not (5,10,15) (5,10,15) (5,10,15)
expanded. Customer demand Random number Random number Random number
Algorithm 4. Neighbor (X, k, Nnonimp) in (500, 2300) in (500, 2000) in (500, 1700)
Processing time at 0.3 0.4 0.2
1: while k ≤ K customers(h)
2: Find a random neighbor X1 of X using Rk Processing time at 0.5 0.5 0.5
3: X←X1 depot(h)
4: if Nnonimp < k / K * MAXnonimp
5: break
6: end if
6.1. Test instances
7: k ← k + 1
8: end while
9: return X Three problem sets TS1, TS2 and TS3 were generated as follows
(Table 5).
Parameters related to vehicles were set by following existing litera-
ture (Qian and Eglese, 2016; Androutsopoulos and Zografos, 2017;
5.4. Local search
Raeesi and Zografos, 2019) and other parameters were artificially set
considering real application scenarios. Considering the small and me-
In the local search phase, we used the method that updates the so-
dium cities, the maximum number of customers was set to be 100.
lution to the best solution in the neighborhood in turn using the five
Time windows were randomly generated as [0,4] or [4,8], reflecting
previously defined neighborhood generation operations (Algorithm 5).
that customer services are required in the morning or afternoon. The
If a solution is updated in any neighborhood structure, we restart the
depot was positioned in the city center, and customers’ locations were
updating process from the first neighborhood structure (line 7). This
randomly generated. Assuming that there are three paths between any
process is completed if no better solution is obtained in searching five
two customers, the path length was randomly generated based on the
neighborhood structures in turn.
geometric distance between the two customers. In addition, road seg-
Similar to Wassan et al. (2015), we used multilayer local search
ments for each path with three speed limit profiles were included in
optimization. In the second layer, we used the R5: Delay, considering
random proportions. We generated 20 problems for each problem set.
that scheduling and routing are important. If the solution is updated in
each neighborhood, we move to the best solution in the neighborhood
using the R5: Delay operation before returning to the first neighborhood
(line 6).
Algorithm 5. Local Search ( X )

1: k ← 1
2: while k ≤ K
3: Find the best neighbor X1 of X using Rk
4: if X1 is better than X
5: X←X1
6: Find a best neighbor X1 of X using R5
7: X ← X1, go to line 1
8: end if
9: k ← k + 1
10: end while
11: return X

5.5. Stopping criteria

If Nnonimp reaches MAXnonimp, VNS completes the search,

concluding that it is almost impossible to update the solution anymore.

6. Computational experiments

In this section, computational experiments are conducted to examine

the performance of our algorithm. VNS procedures are coded in C++,
and the CPLEX model is implemented using IBM ILOG CPLEX 12.6.10.
Experiments are performed on a personal computer with a 2.6 GHz
processor and 4GByte RAM. Fig. 5. Results of VNS according to MAXnonimp.

8
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

average objective function values and computation times were

compared (Fig. 5).
As MAXnonimp increases, the computation time increases, and the
solution quality improves, but after 1000, the improvement of the so-
lution quality is not significant. Especially for VNS-ADAPTIVE, when
MAXnonimp is greater than 1000, the quality of the solution is slightly
worse. This seems to be a negative effect due to the slow expansion of the
neighborhood structure. From these results, we conclude that it is
reasonable to set MAXnonimp as 1000, considering the solution quality
and computational time simultaneously.
For TS1 and TS2, VNS-ADAPTIVE using the adaptive neighborhood
structure increases the computation time, but the solution quality is
improved. In order to validate the effect of the adaptive neighborhood
structure, the updating process of the solution was analyzed in detail.
Fig. 6 shows the solution updates of VNS_FIX and VNS_ADAPTIVE for
Problem 3 in TS1(TS1-03). In the case of VNS_FIX (Fig. 6, a), updates

Table 7
Computational result for TS3.
Instance Obj_Cplex Vns_Init Obj_Vns CPU(s) Gap1 Gap2

TS3-01 17,319.00 19,386.12 16,028.39 1,184.05 8.05 20.95

TS3-02 15,789.05 17,772.97 14,210.15 1,108.62 11.11 25.07
TS3-03 15,560.36 19,697.38 15,047.85 1,161.25 3.41 30.90
TS3-04 14,029.62 16,651.01 13,187.84 680.10 6.38 26.26
TS3-05 21,044.02 20,593.07 18,729.18 533.45 12.36 9.95
TS3-06 18,145.51 18,791.67 13,790.59 1,586.49 31.58 36.26
TS3-07 13,357.48 15,059.49 13,183.81 1,130.64 1.32 14.23
TS3-08 21,061.36 19,638.12 15,796.02 1,050.29 33.33 24.32
Fig. 6. Updating process of solution (a-VNS_FIX, b-VNS_ADAPTIVE). TS3-09 17,546.21 19,197.57 13,686.05 963.53 28.21 40.27
TS3-10 18,362.98 20,326.26 15,037.72 1,347.41 22.11 35.17
TS3-11 13,811.23 18,201.52 12,496.54 1,163.24 10.52 45.65
6.2. Parameter calibration and effectiveness of the proposed adaptive TS3-12 16,455.14 18,017.08 15,860.39 425.29 3.75 13.60
neighborhood TS3-13 22,247.14 27,539.48 19,355.01 1,094.79 14.94 42.29
TS3-14 16,138.11 16,602.10 13,878.78 1,214.73 16.28 19.62
TS3-15 14,993.14 17,757.69 12,343.90 536.52 21.46 43.86
In the VNS algorithm, the main parameter affecting the computation TS3-16 20,997.99 20,164.30 16,378.43 1,257.16 28.21 23.11
time and solution quality is MAXnonimp, which specifies the stopping TS3-17 31,500.82 26,658.19 25,830.67 605.85 21.95 3.20
condition. TS3-18 26,516.78 21,461.75 21,130.27 982.95 25.49 1.57
To calibrate the parameter MAXnonimp and to compare two algo- TS3-19 17,019.90 20,525.41 15,658.31 625.69 8.70 31.08
TS3-20 20,661.93 24,182.20 18,595.74 1,231.73 11.11 30.04
rithms, computational experiments have been conducted as follows.
Average 994.19 16.01 25.87
Using VNS without adaptive neighborhood (VNS_FIX) and with
adaptive neighborhood (VNS_ADAPTIVE), each problem in TS1 and TS2 Gap1 = Obj_Cplex/Obj_Vns − 1.
was solved 10 times, changing MAXnonimp from 500 to 1500. The Gap2 = Vns_Init/Obj_Vns − 1.

Table 6
Computational results for TS1 and TS2.
Instance CPLEX VNS Gap Instance CPLEX VNS Gap

Obj1 CPU1 Obj2 CPU2 Obj1 CPU1 Obj2 CPU2

TS1-01 2,394.47 1.67 2,457.35 18.21 2.63 TS2-01 3,692.08 172.24 3,724.75 233.81 0.88
TS1-02 2,487.41 2.17 2,528.09 19.52 1.64 TS2-02 4,011.22 2,077.22 4,145.97 240.65 3.36
TS1-03 2,520.57 2.23 2,603.91 10.79 3.31 TS2-03 3,283.06 156.03 3,726.66 159.72 13.51
TS1-04 2,304.35 1.09 2,498.37 16.35 8.42 TS2-04 3,445.59 618.25 3,691.37 176.34 7.13
TS1-05 2,517.23 1.79 2,660.88 21.51 5.71 TS2-05 3,632.39 553.16 4,181.42 177.72 15.11
TS1-06 2,050.24 1.87 2,096.11 14.35 2.24 TS2-06 3,205.65 1,182.77 3,301.71 269.42 3.00
TS1-07 2,245.82 2.71 2,276.29 14.93 1.36 TS2-07* 4,043.96 3,601.02 4,118.79 141.62 1.85
TS1-08 2,831.16 2.45 2,881.63 12.59 1.78 TS2-08 3,928.34 1,121.51 3,940.92 338.25 0.32
TS1-09 3,223.55 2.26 3,223.56 10.95 0.00 TS2-09* 3,563.81 3,600.25 3,759.88 279.38 5.50
TS1-10 2,298.91 0.45 2,298.91 10.81 0.00 TS2-10 3,832.47 2,435.72 4,158.97 488.94 8.52
TS1-11 2,683.99 6.47 2,684.25 13.69 0.01 TS2-11* 3,843.65 3,600.00 3,988.69 274.59 3.77
TS1-12 2,327.62 0.59 2,374.90 13.54 2.03 TS2-12 3,408.51 2,737.05 3,519.41 367.52 3.25
TS1-13 2,377.72 1.23 2,378.20 9.13 0.02 TS2-13* 4,025.78 3,600.00 4,607.21 311.89 14.44
TS1-14 2,682.70 2.09 2,714.13 7.71 1.17 TS2-14* 4,091.74 3,600.00 4,340.17 195.83 6.07
TS1-15 2,581.31 4.60 2,799.96 14.37 8.47 TS2-15* 3,885.78 3,600.00 4,502.79 100.05 15.88
TS1-16 1,846.23 3.65 1,987.64 19.46 7.66 TS2-16 4,111.50 2,163.52 4,566.40 151.30 11.06
TS1-17 2,650.42 4.88 2,763.00 8.95 4.25 TS2-17 3,837.80 552.59 4,265.02 167.06 11.13
TS1-18 2,529.06 2.59 2,682.26 13.53 6.06 TS2-18 3,127.88 340.49 3,197.57 190.13 2.23
TS1-19 2,235.47 1.20 2,318.30 9.14 3.71 TS2-19* 4,525.52 3,600.38 4,755.61 209.71 5.08
TS1-20 1,832.22 1.51 1,877.46 12.10 2.47 TS2-20 3,296.65 100.70 3,474.67 257.36 5.40
Average 2.38 13.58 3.15 Average 1,970.64 236.56 6.88

Gap = Obj_VNS/Obj_CPLEX − 1.

9
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

were made within 3 s. And until the end of the calculation, the update of Table 8
the solution was not made. Comparation of VNS-Fix and VNS-Adaptive.
In the case of VNS_ADAPTIVE (Fig. 6, b), the search was completed at Instance Objective value (VNS- Objective value (VNS- Gap
about 43 s, and the updates of the solution were made after 30 s. In more ADAPTIVE) FIX)
detail, updates were made when Nnonimp was 587 at about 21 s, and TS3-01 16,028.39 16,725.62 4.35
Nnonimp was 458 at about 29 s, in which the neighborhood structure TS3-02 14,210.15 15,287.27 7.58
was expanded to step 3. TS3-03 15,047.85 15,278.09 1.53
TS3-04 13,187.84 14,072.74 6.71
TS3-05 18,729.18 19,448.38 3.84
6.3. Result and analysis TS3-06 14,237.34 14,237.34 0.00
TS3-07 13,183.81 13,183.81 0.00
We compared our VNS algorithm which has the adaptive neighbor- TS3-08 15,796.02 17,017.05 7.73
hood structure with CPLEX model. The computational results for TS1 TS3-09 13,686.05 14,434.67 5.47
TS3-10 15,115.57 16,161.57 6.92
and TS2 are given in Table 6. For TS1, the proposed VNS algorithm’s
TS3-11 12,496.54 12,765.21 2.15
average gap from the optimal solution by CPLEX is 3.15 %. In this case, TS3-12 15,860.39 16,548.73 4.34
the computation time of CPLEX (2.38 s) is much smaller than VNS TS3-13 19,355.01 19,804.05 2.32
(13.58 s). However, some problems in TS2 (TS2-07, TS2-09, TS2-11, TS3-14 13,878.78 14,692.07 5.86
TS2-13, TS2-14, TS2-15, TS2-19) have not been solved optimally within TS3-15 12,343.90 12,343.90 0.00
TS3-16 16,378.43 17,429.93 6.42
one hour using CPLEX. Therefore, we compared the results of CPLEX TS3-17 25,830.67 27,272.02 5.58
found within 1 h and the solution of the VNS algorithm. As a result, the TS3-18 21,130.27 22,347.37 5.76
VNS algorithm has an average gap of 6.88 % compared to CPLEX solu- TS3-19 15,658.31 15,844.64 1.19
tions. For the above seven problems without optimal solution, the gap TS3-20 18,674.12 20,035.46 7.29
Average 4.25
calculated using the lower bound is 8.39 %.
For the problems in TS3, the optimal solution is not obtained within Gap = VNS-FIX/VNS-ADAPTIVE − 1.
24 h when CPLEX is used, so for CPLEX, the search time limit is set as 1 h
and the results are compared (Table 7). The average computation time of
VNS is 994.19 s (about 16 min), and solution quality is improved by Table 9
16.01 % (Gap1) on average compared with CPLEX. Comparation of VNS-Fix and VNS-Adaptive.
In order to evaluate the effectiveness of vehicle routing and sched- Average fuel consumption Ratio (%)
uling using VNS compared to manual routing, a comparison is made 3800, 8000 18,342.62 114.44
with the VNS initial solution yielded by the nearest neighborhood 4000, 10,000 16,028.39 100.00
heuristic. In fact, the nearest neighborhood heuristic can be considered 4200, 12,000 15,814.74 98.67
4400, 14,000 15,917.51 99.31
as a reflection of the manual routing method in practice. It can be seen
that the fuel consumption decreases by 25.87 %(Gap2) on average. This
demonstrates the economic effectiveness of applying FCVRSP-SMC in consumption minimizing path selection between two customer nodes is
urban logistics. almost independent of the vehicle payload. Based on this finding, the
We also examined the performance of VNS-Adaptive compared to optimal path between two customer nodes is determined according to
VNS-Fix (Table 8). The computation time limit is set to be 1200 s (20 the time period and the vehicle type. Then, based on them, the FCVRSP-
min). The results show that the proposed adaptive neighborhood SMC is modeled to reflect multi-trips, heterogeneous vehicles, time-
structure can reduce the objective value by 4.25 percent on average. dependent travel speed, and time window. Since this problem is a
very complex optimization problem, we need a heuristic algorithm that
6.4. Sensitivity analysis can solve large-scale problems. We have designed a solution algorithm
for FCVRSP-SMC using VNS, widely used as an efficient approximate
In reality, logistics companies should choose vehicles with proper solution approach in previous studies. We designed an adaptive neigh-
capacities for efficient operations. Therefore, we conducted a sensitivity borhood structure to enhance the global search performance of VNS.
analysis to investigate how the capacities of vehicles have an impact on Simulation results validate the effectiveness of VNS with the proposed
fuel consumption. For problem set TS3, we changed the capacity of adaptive neighborhood structure. This method makes it possible to
light-duty vehicles to be 3800, 4000, 4200, 4400 and the capacity of reduce fuel consumption by about 25 %. This indicates the economic
medium-duty vehicles to be 8000, 10000, 12000, 14000, respectively. and environmental benefits of applying FCVRSP-SMC to urban logistics.
Table 9 shows the average fuel consumption for 4 cases. The ratio shows Through the sensitivity analysis, it is also confirmed that the fleet mix of
the ratio to Case 2 (4000, 10000). Case 2 is for the original capacities. logistics companies can have an impact on operation efficiency.
For the smaller capacities, fuel consumption increased by 14.44 %. The
reason is the increase in the number of trips due to small capacities. For Funding
the bigger capacities, fuel consumption dicreases a little. This sensitivity
analysis shows that the correct selection of vehicle capacities is impor- The authors declare that no funds, grants, or other support were
tant for logistics companies. received during the preparation of this manuscript.

7. Conclusions CRediT authorship contribution statement

In this paper, we studied a vehicle routing and scheduling problem in Yong-Jin Pak: Writing – review & editing, Writing – original draft,
small and medium cities with environmental considerations. The Software, Project administration, Methodology, Formal analysis, Data
important characteristics of urban logistics are time-dependent speed curation. Kyong-Ho Mun: Validation, Supervision, Resources, Formal
limitations and multipaths between two customer sites according to the analysis, Conceptualization.
urban road network. This paper provides a relatively low dimensional
MILP model that considers all those characteristics. Through computa-
tional experiments, we found that if the CMEM model is used as a fuel
consumption calculation model for small and medium cities, the fuel

10
Y.-J. Pak and K.-H. Mun Cleaner Logistics and Supply Chain 12 (2024) 100164

Declaration of competing interest Hosseini-Nasab, H., Lotfalian, P., 2017. Green routing for trucking systems with
classification of path types. J. Clean. Prod. 146, 228–233.
Huang, Y., Zhao, L., Woensel, T.V., Gross, J.-P., 2017. Time-dependent vehicle routing
The authors declare that they have no known competing financial problem with path flexibility. Transp. Res. B 95, 169–195.
interests or personal relationships that could have appeared to influence Jia, Y., Tang, L., Min, Xu., Yang, X., 2019. Landscape pattern indices for evaluating urban
the work reported in this paper. spatial morphology – a case study of Chinese cities. Ecol. Ind. 99, 27–37. https://doi.
org/10.1016/j.ecolind.2018.12.007.
Karakostas, P., Sifaleras, A., 2024a. Modeling of sustainable integrated supply chains
Data availability under the consideration of European Union regulations. Cent. Eur. J. Oper. Res.
https://doi.org/10.1007/s10100-024-00910-7.
Karakostas, P., Sifaleras, A., 2024b. The pollution traveling salesman problem with
Data will be made available on request. refueling. Comput. Oper. Res. 167 https://doi.org/10.1016/j.cor.2024.106661.
Karakostas, P., Sifaleras, A., Georgiadis, M.C., 2020. Adaptive variable neighborhood
Acknowledgements search solution methods for the fleet size and mix pollution location-inventory-
routing problem. Expert Syst. Appl. 153 https://doi.org/10.1016/j.
eswa.2020.113444.
This research did not receive any specific grant from funding Koc, C., Bektas, T., Jabali, O., Laporte, G., 2014. The fleet size and mix pollution-routing
agencies in the public, commercial, or not-for-profit sectors. problem. Transp. Res. B Methodol. 70, 239–254. https://doi.org/10.1016/j.
trb.2014.09.008.
Koc, C., Bektas, T., Jabali, O., Laporte, G., 2016. The impact of depot location, fleet
References composition and routing on emissions in city logistics. Transp. Res. B Methodol. 84,
81–102.
Androutsopoulos, K.N., Zografos, K.G., 2017. An integrated modelling approach for the Lai, D.S., Demirag, O.C., Leung, J.M., 2016. A tabu search heuristic for the heterogeneous
bicriterion vehicle routing and scheduling problem with environmental vehicle routing problem on a multigraph. Transp. Res. E 86, 32–52.
considerations. Transp. Res. C 82, 180–209. Lai, David, Costa, Yasel, Demir, Emrah, Florio, Alexandre M., Van Woensel, Tom, 2024.
Asghari, M., e hashem, M.A., J., S. M., 2021. Green vehicle routing problem: a state-of- The pollution-routing problem with speed optimization and uneven topography.
the-art review. Int. J. Prod. Econ. 231. Comput. Oper. Res. 164, 106557, ISSN 0305-0548. https://doi.org/10.1016/j.
Barth, M., Boriboonsomsin, K., 2008. Real-world co2 impacts of traffic congestion. cor.2024.106557.
Transp. Res. Record 2058, 163–171. Liu, C., Kou, G., Zhou, X., Peng, Y., Sheng, H., Alsaadi, F.E., 2020. Time-dependent
Barth, M., Scora, G., Younglove, T., 2004. Modal emissions model for heavy-duty diesel vehicle routing problem with time windows of city logistics with a congestion
vehicles. Transp. Res. Record 1880, 10–20. avoidance approach. Knowl.-Based Syst. 188, 104813.
Behnke, M., Kirschstein, T., 2017. The impact of path selection on GHG emissions in city Malandraki, C., Daskin, M., 1992. Time-dependent vehicle-routing problems
logistics. Transp. Res. E 106, 320–336. formulations, properties and heuristic algorithms. Transp. Sci. 26, 185–200.
Bektas, T., Laporte, G., 2011. The pollution-routing problem. Transp. Res. B 45, Niu, Y., Yang, Z., Chen, P., Xiao, J., 2017. Optimizing the green open vehicle routing
1232–1250. problem with time windows by minimizing comprehensive routing cost. J. Clean.
Bofeng, C., et al., 2019. A benchmark city-level carbon dioxide emission inventory for Prod.
China in 2005. Appl. Energy 233, 659–673. https://doi.org/10.1016/j. Pan, B., Zhang, Z., Lim, A., 2021. Multi-trip time-dependent vehicle routing problem
apenergy.2018.10.016. with time windows. Eur. J. Oper. Res.
Bruglieri, M., Paolucci, M., Pisacane, O., 2023. A matheuristic for the electric vehicle Parvasi, S.P., Mahmoodjanloo, M., Setak, M., 2017. A bi-level school bus routing problem
routing problem with time windows and a realistic energy consumption model. with bus stops selection and possibility of demand outsourcing. Appl. Soft Comput.
Comput. Oper. Res. 157, 106261 https://doi.org/10.1016/j.cor.2023.106261. 61, 222–238.
Cattaruzza, D., Absi, N., Feillet, D., Gonzalez-Feliu, J., 2017. Vehicle routing problems Parvasi, S.P., Tavakkoli-Moghaddam, R., Taleizadeh, A.A., Soveizy, M., 2019. A bi-level
for city logistics. EURO J. Transp. Logist. 6, 51–79. bi-objective mathematical model for stop location in a school bus routing problem.
Cheng, C., Yang, P., Qi, M., Rousseau, L.-M., 2017. Modeling a green inventory routing IFAC-PapersOnLine 52 (13), 1120–1125.
problem with a heterogeneous fleet. Transp. Res. Part E 97, 97–112. https://doi.org/ Qian, J., Eglese, R., 2016. Fuel emissions optimization in vehicle routing problems with
10.1016/j.tre.2016.11.001. time-varying speeds. Eur. J. Oper. Res. 248, 840–848.
Cinar, D., Gakis, K., Pardalos, P.M., 2016. A 2-phase constructive algorithm for Raeesi, R., Zografos, K.G., 2019. The multi-objective steiner pollution routing problem
cumulative vehicle routing problems with limited duration. Expert Syst. Appl. 56, on congested urban road networks. Transp. Res. B 122, 457–485.
48–58. Schneider, M., Stenger, A., Hof, J., 2014. An adaptive VNS algorithm for vehicle routing
Dantzig, G., Ramser, J., 1959. Time constrained routing and scheduling, in: M.o. ball, t.l. problems with intermediate stops. OR Spectr. 37, 353–387.
magnanti, c.l. monma,e.l. nemhauser (Eds.). Manag. Sci., 6, 80–91. Sze, J.F., Salhi, S., Wassan, N., 2024. An adaptive variable neighbourhood search
Demir, E., Bektas, T., Laporte, G., 2014. A review of recent research on green road freight approach for the dynamic vehicle routing problem. Comput. Oper. Res. 164 https://
transportation. Eur. J. Oper. Res. 231. doi.org/10.1016/j.cor.2024.106531.
Ehmke, J., Campbell, A., Thomas, B., 2016. Vehicle routing to minimize time-dependent Ticha, H.B., Absi, N., Feillet, D., Quilliot, A., 2017. Empirical analysis for the VRPTW
emissions in urban areas. Eur. J. Oper. Res. 251, 478–494. with a multigraph representation for the road network. Comput. Oper. Res. 88,
Ferreira, K.M., de Queiroz, T.A., Munari, P., Toledo, F.M.B., 2024. A variable 103–116.
neighborhood search for the green vehicle routing problem with two-dimensional Todosijevic, R., Hanafi, S., Urosevic, D., Jarboui, B., Gendrond, B., 2016. A general
loading constraints and split delivery. Eur. J. Oper. Res. variable neighborhood search for the swap-body vehicle routing problem. Comput.
Figliozzi, M.A., 2011. The time dependent vehicle routing problem with time windows: Oper. Res. 78, 468–479.
benchmark problems, an efficient solution algorithm, and solution characteristics. Vidal, T., Laporte, G., Matl, P., 2020. A concise guide to existing and emerging vehicle
Transp. Res. E 48, 616–636. routing problem variants. Eur. J. Oper. Res. 286, 401–796.
Franceschetti, A., Demir, E., Honhon, D., Woensel, T.V., Laporte, G., Stobbe, M., 2016. Vincent, F.Y., Normasari, N.M.E., Chen, W.H., 2021. Location-routing problem with
A meta heuristic for the time-dependent pollution-routing problem. Eur. J. Oper. time-dependent demands. Comput. Ind. Eng. 151, 106936.
Res. 259, 972–991. Wassan, N., Wassan, N., Nagy, G., Salhi, S., 2015. The multiple trip vehicle routing
Garaix, T., Artigues, C., Feillet, D., Josselin, D., 2010. Vehicle routing problems with problem with backhauls: formulation and a two-level variable neighborhood search.
alternative paths: an application to on-demand transportation. Eur. J. Oper. Res. Comput. Oper. Res. 78, 454–467.
204, 62–75. Wei, L., Zhang, Z., Zhang, D., Lim, A., 2015. A variable neighborhood search for the
Garside, A.K., Ahmad, R., Muhtazaruddin, M.N.B., 2024. A recent review of solution capacitated vehicle routing problem with two-dimensional loading constraints. Eur.
approaches for green vehicle routing problem and its variants. Oper. Res. Perspect. J. Oper. Res. 243, 798–814.
12, 100303. Xiao, Y., Konak, A., 2015. A simulating annealing algorithm to solve the green vehicle
Gendreau, M., Potvin, J.-Y., 2010. Handbook of Metaheuristics. Springer. routing and scheduling problem with hierarchical objectives and weighted tardiness.
Guo, R., Zhang, W., Guan, W., Ran, B., 2021. Time-dependent urban customized bus Appl. Soft Comput.
routing with path flexibility. IEEE Trans. Intell. Transp. Syst. 22 (4), 2381–2390. Xiao, Y., Zhao, Q., Kaku, I., Xu, Y., 2012. Development of a fuel consumption
https://doi.org/10.1109/TITS.2020.3019373. optimization model for the capacitated vehicle routing problem. Comput. Oper. Res.
Hatami, Sara, Eskandarpour, Majid, Chica, Manuel, Juan, Angel A., Ouelhadj, Djamila, 39, 1419–1431.
2020. Green hybrid fleets using electric vehicles: solving the heterogeneous vehicle Yao, E., Lang, Z., Yang, Y., Zhang, Y., 2015. Vehicle routing problem solution considering
routing problem with multiple driving ranges and loading capacities. SORT-Statistics minimizing fuel consumption. IET Intel. Transport Syst. 9, 523–529.
and Operations Research Transactions, [online], 44, 1, 141–70, https://doi.org/ Yeh, W.-C., Tan, S.-Y., 2021. Simplified swarm optimization for the heterogeneous fleet
10.2436/20.8080.02.98 [View: 30-09-2021]. vehicle routing problem with time-varying continuous speed function. Electronics
10, 1775. https://doi.org/10.3390/electronics10151775.

11