Received 22 September 2023, accepted 9 October 2023, date of publication 18 October 2023, date of current version 14 November 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3325787

Hybrid Algorithms for Energy Minimizing Vehicle

Routing Problem: Integrating Clusterization and
Ant Colony Optimization
NICOLÁS FRÍAS , FRANKLIN JOHNSON , AND CARLOS VALLE
Departamento de Ciencia de Datos e Informática, Facultad de Ingeniería, Universidad de Playa Ancha, Valparaíso 2360002, Chile

Corresponding author: Nicolás Frías ([email protected])

This work was supported by Universidad de Playa Ancha under Grant DGI APFON 03-22.

ABSTRACT In the field of engineering, complex problems often arise that require solutions. The
implementation of these algorithms plays a crucial role in achieving favorable outcomes with the available
resources. The Vehicle Routing Problem (VRP) has been a central topic in distribution and logistics
for decades. New VRP models and tools are developed to address the challenges of modern logistics.
The Energy Minimizing Vehicle Routing Problem (EMVRP) is a ‘‘green’’-oriented variant of the VRP
where the objective is to minimize the total amount of energy consumed by a fleet of vehicles. The VRP
literature has focused on solving the problem using a variety of approaches and techniques, including exact
methods, heuristics, metaheuristics, and hybrid algorithms. Hybrid algorithms combine different techniques
to obtain more effective and better solutions. This work presents four innovative hybrid algorithms to
address the EMVRP problem. These algorithms combine Machine Learning (ML) clustering techniques
with metaheuristic approaches inspired by an Ant Colony Optimization (ACO). The proposed algorithms
are: Free Ant + K-Means, Free Ant + K-Medoids, Restricted Ant + K-Means, and Restricted Ant + K-
Medoids. Each of them combines the benefits of clustering with the optimization capacity of ACO. Proposed
algorithms were subjected to testing using instances from CVRPLIB. Both Free Ant and Restricted Ant
efficiently solved EMVRP problems. The results obtained were analyzed and compared with the proposals
of other authors in the literature. Overall, the results are promising, but they also indicate a significant scope
for experimentation and parameter tuning of the proposed algorithms.

INDEX TERMS Complex problems, free ant, green-vrp, heuristics, k-means, k-medoids, machine learning,
metaheuristics, restricted ant, vehicle routing problem (VRP).

I. INTRODUCTION evolution of the TSP. Since its inception, this problem has
One of the most widely known and oldest combinatorial generated significant interest among the Operations Research
optimization problem is the Traveling Salesman Problem and mathematical communities. These problems are known
(TSP), which involves determining the shortest possible route as Vehicle Routing Problems (VRP) and belong to the field
that visits all cities (from a city list) exactly once and returns of combinatorial optimization.
to the starting city. Over time, the idea of optimizing routes Introduced in the early 1960s, the VRP has been exten-
for a finite number of cities has been extended to find optimal sively studied by the scientific community, primarily because
routes for vehicles from one or more depots to a set of of its practical applications in the field of distribution and
locations or customers, thus being understood as the natural logistics. Additionally, the VRP presents inherent compli-
cations that exist in this type of combinatorial optimization
The associate editor coordinating the review of this manuscript and problem, similar to the TSP, making it part of the NP-Hard
approving it for publication was Wentao Fan . problem set. Whereas the TSP seeks the shortest route to visit

2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
125800 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

a certain number of cities and return to the origin, the VRP Cartesian plane. A = (i, j) : i, j ∈ V , i ̸= j is the set of arcs or
aims to determine the optimal set of routes for vehicles, visit edges connecting each node in set V , where (i, j) represents
all customers once, and serve their associated demands per the arc connecting node vi to node vj .
vehicle. In the VRP, there is a finite number of routes, one Additionally, we have the following characteristics that
for each available vehicle, and the goal is to determine the complete the EMVRP model:
optimal set of routes. This characteristic makes the VRP a • dij is the distance between node i and node j,
generalization of TSP. • qi is the positive integer load of node vi , i.e., its demand
The formal definition of a VRP states that m vehicles or supply,
initially located at a depot must serve positive amounts of • m is the number of heterogeneous vehicles,
goods to n customers. The objective is to determine the • Q0 is the tare weight (weight of the vehicle when empty)
K optimal routes used by the fleet of vehicles to fulfill of these vehicles,
the demands of the n customers. VRP have many practical • Q is the maximum capacity of a vehicle.
applications, especially in transportation and distribution Kara et al. [1] defined the Energy Minimizing Vehicle
logistics. Consequently, various VRP variants have been Routing Problem (EMVRP) as a vehicle routing problem,
developed to better represent our reality and involve variables such that:
that approximate the actual costs associated with traversing • Each node vi is served by exactly one vehicle,
different vehicle routes. Models such as the Capacitated • Each route starts and ends at the depot v0 ,
Vehicle Routing Problem (CVRP), Dynamic Vehicle Routing • The load on the arcs accumulates as the sum of
Problem (DVRP), and Vehicle Routing Problem with Time quantities qi supplied by the preceding nodes (in the
Windows (VRPTW) are only a few current frameworks case of pickup) and decreases as the sum of quantities
within this problem class, and they can differ significantly qi demanded by the preceding nodes (in the case of
from one another or involve constraints from other types of delivery),
problems. • The load of a vehicle must not exceed the capacity Q,
This article presents the results of a research project • The objective is to find a set of K vehicle routes with the
focused on the Energy Minimizing Vehicle Routing Problem minimum total cost, that is, the minimum total energy.
(EMVRP), which is one of the most interesting models Furthermore, the following decision variables are defined
in the field of vehicle routing optimization. This model for the formulation of this problem:
stands out for its ability to minimize both the financial 1) Xij = 1 if the arc (i, j) is part of a route, and 0 otherwise,
costs and environmental impact of a fleet of vehicles 2) Yij is the weight of the vehicle when traveling from
while traversing different routes. Therefore, the EMVRP has node vi to node vj , and 0 otherwise.
emerged as a valuable tool for addressing logistical and As can be inferred, Yij is the component that gives meaning
environmental challenges in route planning and distribution. to this new energy-oriented approach. The weight of the
Originally published by Kara et al. [1], the EMVRP is vehicle on the first arc of any route must have a predetermined
a key problem within the Green VRP domain. In this value; for example, tare weight Q0 , and it should always
study, we propose an algorithm to solve EMVRP problems increase (or decrease) by units of qi immediately after visiting
using constructive metaheuristic techniques and clustering node vi . In the case of pickup, the flow of the variable
algorithms. The study is divided into three parts: the EMVRP exhibits an incremental step function, whereas in the delivery
problem, clustering algorithms, and the use of the Ant Colony case, it exhibits a decremental step function. Therefore,
Optimization (ACO) metaheuristic technique. Our hypothesis a model constructed for one case may not be suitable for
in this work is that we believe that mixed strategies combining its counterpart. However, Kara et al. demonstrated that when
metaheuristics and machine learning propose better solutions distances are symmetric, the optimal route for the delivery
than these techniques separately. We present two innovative (or pickup) case is equal to the optimal route for the pickup
metaheuristic algorithms for problem resolution based on the (or delivery) case, but traversed in the reverse order. Thus,
Best-Worst Ant System algorithm proposed by Cordón et al. if the distances are symmetric, there is no need to differentiate
between delivery and pickup, because the solution for one
[2]. These algorithms, named Free Ant and Restricted
case is the solution for the other, but in the reverse order.
Ant, are combined with clustering techniques, specifically
Therefore, the objective function to be minimized is as
the K-Means and K-Medoids algorithms with capacity
follows:
restriction, inspired by the work of Geetha et al. [3].
n X
X n

II. ENERGY MINIMIZING VEHICLE ROUTING PROBLEM min dij Yij . (1)
Let us consider an arbitrary Capacitated Vehicle Routing i=0 j=0

Problem (CVRP). This problem is defined by a graph G = This objective function (1) is subject to the following
(V , A), where V = v1 , v2 , . . . , vi is the set of nodes constraints:
representing the customers in the problem, with v0 is the n
X
depot, and vi is the i-th node or vertex. These nodes have X0i = m. (2)
positions with coordinates represented as ci = (xi , yi ) in a i=1

VOLUME 11, 2023 125801

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

n
X Before delving into the development of the algorithms
Xi0 = m. (3)
proposed in this study, it is essential to introduce the
i=1
n terms that will be frequently used throughout the following
X
Xij = 1, j = 1, 2, 3, . . . , n. (4) chapters. Therefore, this section presents some fundamental
i=0 concepts necessary to comprehend the processes involved in
n
X the algorithms proposed in this study. By doing so, readers
Xij = 1, i = 1, 2, 3, . . . , n. (5) will be able to gain a clear understanding of the terms
j=0 employed in the analysis of Free Ant and Restricted Ant.
Xn n
X
Yij − Yji = qi , i = 1, 2, 3, . . . , n. (6) A. ANT COLONY OPTIMIZATION
j=0 j=0
j̸=i j̸=i Ant Colony Optimization (ACO) acts as a tool for solving
Y0i = Q0 X0i , i = 1, 2, 3, . . . , n. (7) complex combinatorial optimization problems and was
originally designed to find the optimal path within a graph,
Yij ≤ (Q + Q0 − qj )Xij , (i, j) ∈ A. (8) specifically a Traveling Salesman Problem (TSP) graph.
Yij ≥ (Q0 + qi )Xij , ∀(i, j) ∈ A. (9) ACO can be understood as a bio-inspired metaheuristic
Xij = 0 o 1, (i, j) ∈ A. (10) with a general-purpose application, arising from the genuine
observation that a colony of ants is capable of finding an
The cost of traversing an arc (i, j) is the product of optimal path between their nest and a food source. Initially,
the distance dij (the distance between nodes vi and vj ) real ants, explore the area around their nest in a random
and the weight Yij carried by the vehicle on this arc, manner when searching for food, leaving behind a trail of
which is satisfied by the objective function given in (1). pheromones. This pheromone trail affects subsequent ants,
Furthermore, Constraints (2) and (3) ensure that all m vehicles gradually causing them to converge towards an optimal path.
are used. Constraints (4) and (5) represent the degrees The foundation of the Ant Colony Optimization theory was
of restriction for each node. Constraint (6) represents the introduced by Marco Dorigo in his doctoral thesis in 1992.
classical conservation of the flow equation, which balances However, it was not until the mid-1990s that Dorigo, M.,
the inflow and outflow of each node (in terms of demand Maniezzo, V., and Colorni, A. [4] published it as a novel
or supply) and prohibits any illegal sub-trips. Constraint (7) metaheuristic called Ant System (AS). AS employs a fixed
initializes the flow in the first arc of each route. Constraint (8) number of artificial ants (w) where each ant a (considered
handles the capacity constraints and forces the value of Yij to a computational agent) constructs a solution to the problem,
be zero when arc (i, j) is not a part of any route. Constraint (9) previously represented by a graph. In this graph, each arc (i, j)
sets the lower bounds for flow on any arc. The integrity corresponds to one of the possible choices available to the ant
constraint is given in (10), and corresponds to the binary to transition from its current state to a new state.
variable for each arc. Ant a is only aware of the information associated with these
arcs (i, j), which consists of the following:
III. BACKGROUND 1) Heuristic information η, which determines the heuristic
The state-of-the-art analysis in this work focused on address- preference of an ant to move from one state to
ing contemporary challenges in logistics and distribution, another. This information typically refers to the cost of
using the Energy Minimizing Vehicle Routing Problem transitioning from node vi to node vj , and in practice,
(EMVRP) as a case study. These logistical problems encom- it remains unchanged throughout the execution of the
pass a wide array of challenges and complexities. We believe algorithm.
that the application of heuristics, metaheuristics and hybrid 2) Pheromone trail information τ , which measures the
algorithms, such as ant colony optimization and clustering, learned preference for state transitions among the
has the potential to provide optimal or near-optimal solutions w ants. This information aims to mimic the real
for logistics and distribution-related issues. The combination pheromone deposited by natural ants and is a data value
of these techniques can be beneficial in approaching these that will be modified during the course of the execution
problems from various perspectives: of the algorithm, with the purpose of aiding future ants
1) Optimization: Metaheuristics can find optimal or in selecting better routes.
near-optimal solutions for logistics and distribution The development of a generic ACO algorithm requires
problems. artificial ant a to have memory Ma , where it stores each
2) Adaptability: Metaheuristics can adapt to changing visited node vi . This list is commonly known as the taboo
situations and cater to different types of problems, list because it contains all nodes that cannot be revisited (as
depending on current needs. they are part of a route). The list of the remaining nodes that
3) Complex Problems: Hybrid algorithms combine differ- the ant can visit is denoted as Ja . To determine the next node
ent techniques, such as machine learning and optimiza- vj that the ant will visit, the transition rule (11), also known
tion, to effectively address challenging problems. as the standard probability distribution pa of AS, is used to

125802 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

determine the next node vj that the ant will visit. achieved by setting a minimum value τmin and a maximum

α β value τmax of pheromone for each arc and updating the
 P [τ (r, s)] [η(r, s) ]

, ifs ∈ Ja (r), pheromone value only if it falls within this range. This
α β .
pa (r, s) = u∈Ja (r) [τ (r, u)] [η(r, u)] restriction ensures that the algorithm does not converge
rapidly to suboptimal solutions. The determination of τmin

 0, otherwise,
(11) and τmax in MMAS are given by:

where r corresponds to node vi and s corresponds to node 1 1

τmax = . (14)
vj . τ represents the pheromone trail on arc (i, j), and η 1 − ρ f (S gb )
√
denotes the heuristic information on the same arc, typically τmax (1 − n Pbest )
1 τmin = √ . (15)
the inverse of the distance d(r, s) ( d(r,s) ). Ja (r) represents (avg − 1) n Pbest
the set of nodes that are yet to be visited by ant a starting
from node r. Furthermore, α and β are adjustable parameters Here, f (S gb ) represents the value of the best global
that determine the relative importance of the pheromone trail solution of the problem, Pbest is the probability of conver-
versus the relative importance of the heuristic information, gence to the best-found solution, which is typically assumed
respectively. to be 0.05. Finally, avg denotes the average choice among the
In ACO, it is necessary to modify the pheromone trails different components of the solution, i.e., avg = n/2.
associated with arcs in each iteration. For example, in AS, In general, the MAX-MIN Ant System is an effective and
at the end of each iteration, there is an evaporation of the widely used optimization algorithm that has been success-
pheromone trails, followed by the deposition of a determined fully applied to a broad range of problems, including the
number of pheromones by the w ants. The update of the Traveling Salesman Problem (TSP), Quadratic Assignment
pheromone τrs , associated with the arc (i, j) connecting nodes Problem (QAP), and Vehicle Routing Problem (VRP).
vi and vj , occurs as follows:
w C. BEST-WORST ANT SYSTEM
X
τrs ← (1 − ρ) · τrs + 1τrsa . (12) The Best-Worst Ant System (BWAS) algorithm is an ACO
a=1 type algorithm developed by Cordón et al. [2], which
where ρ is the evaporation coefficient (a predefined value incorporates concepts from evolutionary computation. This
within the range [0, 1]), and 1τrsa is the amount of pheromone algorithm utilizes the AS transition rule (11) and does
deposited on arc (i, j) by ant a. The amount of pheromone not require online pheromone updates (12), thus making
1τrsa that ant a can deposit can be calculated as follows: it an optional step. Furthermore, this algorithm relies on
 three essential actions performed by an external program
 1 if ant a uses arc (i, j) in its solution, separate from the ant colony. This program is a special
1τrsa = La (13) type of non-interactive computational process known as a
0, otherwise,

‘‘daemon.’’ We briefly outline these actions as follows:
where La is the length of the solution constructed by ant a. 1) Only the best global solution, referred to as ‘‘best-
Since the introduction of AS, various variants of ACO global,’’ and the current worst solution, referred to
algorithms have been proposed to solve combinatorial as ‘‘worst-current,’’ are considered for performing
optimization problems. Some are more closely inspired by positive and negative pheromone updates, respectively
AS, whereas others diverge further. However, their main ideas (offline update). This update is applied once the w ants
are similar: have constructed their solutions and before the iteration
1) The use of a colony of ‘‘ants’’ that cooperate with each end.
other, a) The offline update is applied as follows:
2) Artificial ‘‘pheromone’’ trails for local communication, i
τrs ← (1 − ρ) · τrs + τrs . (16)
3) A sequence of local movements, and i
4) A stochastic decision policy that utilizes local  τrs = f (C(Sbest−global )),

information. if (r, s) ∈ Sbest−global , (17)
Some well-known ACO algorithms include the Ant Colony 
0, otherwise.

System (ACS) [5], ASrank [6], and MAX-MIN Ant System
(MMAS) [7], among others. For more details on Ant Colony where f (C(·)) corresponds to a function that
Optimization, please refer to [8]. evaluates the quality of any given solution.
b) On the other hand, the penalty of arcs in
B. MAX-MIN ANT SYSTEM the current worst solution, ‘‘worst-current,’’ is
As presented by Stützle, T. & Hoos, H. [7], the MAX-MIN performed as follows:
Ant System (MMAS) differs from standard Ant Colony
Optimization (ACO) algorithms in that it limits the amount / Sbest−global ,
∀(r, s) ∈ Sworst−current and (r, s) ∈
of pheromone that can be deposited on an arc (i, j). This is τrs ← (1 − ρ) · τrs . (18)

VOLUME 11, 2023 125803

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

2) The daemon process also incorporates a pheromone where n corresponds to the number of elements in group Gi ,
matrix reset procedure. This occurs whenever a defined cj represents the Cartesian coordinates of the j-th element
stagnation condition is satisfied. (i.e., the values xi ), and i is the average of the Cartesian
3) The pheromone matrix underwent a series of stochastic coordinates among the nodes in group Gi , representing the
mutations that were loosely inspired by evolutionary coordinates of the i-th centroid belonging to that group. The K
computations. The purpose of these mutations was to centroids (1, 2, 3, . . . , K ) are then recalculated as the average
introduce greater diversity into the search process: of the coordinates of all nodes within each respective group
a) Each row of the pheromone matrix is mutated in G. Because it is an iterative algorithm, the groups are
(with a probability of Pm ) as follows: redefined as the K centroids are recalculated, and the nodes
( are reassigned to the groups organically after each iteration
τrs + mut(it, τthreshold ), if e = 1
τrs =
′
. until convergence is reached.
τrs − mut(it, τthreshold ), if e = 0
(19) E. K-MEDOIDS
Another algorithm in the field of unsupervised machine
(r,s)∈Sbest−global τrs
P
τthreshold = . (20) learning is the Partition Around Medoids (PAM), originally
|Sbest−global | proposed by Kaufman, L., & Rousseeuw, P. J. [11], and is
where e is a random value in {0, 1}, it represents commonly referred to as ‘‘K-Medoids.’’ Similar to K-Means,
the current iteration, and τthreshold is the average this algorithm works by dividing the dataset into K groups.
pheromone value of the best global solution. The idea behind K-Medoids arises, in part, from the fact
b) Furthermore, the function mut() is described as that K-Means is often sensitive to outliers. This sensitivity
follows: occurs because the centroids, denoted as i, can be strongly
it − itr influenced by extreme values (elements whose coordinates
mut(it, τthreshold ) = · σ · τthreshold . are significantly distant from the rest), because the centroids
Nit − itr
are calculated as the average coordinates of the elements.
(21) Thus, K-Medoids is a classification technique that segments
where Nit is the maximum number of iterations, the elements of a dataset; however, unlike K-Means, where
itr is the last iteration where a pheromone matrix the centroids are not necessarily part of the groups, in K-
reset is performed, and σ is a constant that Medoids, an object from the dataset is selected as the centroid
specifies the intensity of the mutation relative to of its group. This central object is referred to as a medoid.
the number of iterations conducted so far. The working mechanism of K-Medoids follows an iterative
process similar to that of K-Means. In the initial step, the
D. K-MEANS medoids of each group are initialized randomly. Each element
One of the most commonly used algorithms for clustering is then associated with one of the groups in G based on
datasets is K-Means, which is known for its simplicity the closest proximity to its respective medoid Mi . In each
and effectiveness. This algorithm (also known as Lloyd’s subsequent iteration, the K medoids (M1 , M2 , M3 , . . . , MK )
Algorithm) was originally designed by Lloyd, S. P. [9], are recalculated in terms of the similarity among the elements
although the term ‘‘K-Means’’ was coined in 1967 by within the group. In other words, the new medoid Mi of the
MacQueen, J. B. [10]. K-Means belongs to the field of i-th group is determined as the object that, on average, has
unsupervised machine learning and takes a set of elements the lowest distance to the rest of the objects within the same
and assigns each of them to one of the K groups. This is based group. The cost of group Gi in K-Medoids is given by:
on the inherent features of each element, that is, their values
X
costGi = |cj − Mi |. (23)
xi . The goal of the algorithm is to find the optimal K centroids cj ∈Gi
that minimize the sum of the squared distances between all
the objects in a group and their centroid. where G represents the set of groups, n is the total number of
In the classical sense of this algorithm, the centroids non-medoid elements in a group, Mi denotes the coordinates
are initially placed randomly in space, and the algorithm of the medoid belonging to group Gi , and cj represents the
iteratively assigns each element vi to one of the K groups coordinates of the j-th non-medoid element within group Gi .
based on its similarity or proximity to the centroids. Each
F. VARIABLE NEIGHBORHOOD SEARCH
element is assigned to group G with the smallest distance
to its respective centroid, where G = G1 , G2 , G3 , . . . , GK Variable Neighborhood Search (VNS) is commonly
represents the set of groups. The equation for minimizing the described as a framework rather than a specific local search
sum of squared distances across all groups Gi , is given by: algorithm, as it provides guidelines for implementing more
effective local searches. The VNS was originally introduced
K X
n
X 2 by Hansen, P. & Mladenovic, N. [12] in 1997 as a method for
argmin ||cj − µi || . (22) constructing local search heuristics. It is based on systematic
i=1 cj ∈Gi changes of neighborhoods, involving a descent phase to reach

125804 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

a local minimum and a shaking phase to escape the current state transitions. These clustering algorithms are used to
valley, that is, to search for a better local minimum. generate groups of nodes that can potentially form a single
The VNS is considered a metaheuristic that exploits the route. Therefore, the clustering algorithms are executed at the
idea of changing neighborhoods to find improved solutions. beginning of the main program execution, once the data for
To achieve this, the algorithm relies on the following three an instance has been loaded, various parameters that control
observations: the algorithms have been set, and the distances between all
1) A local minimum with respect to one neighborhood nodes, as well as the energy cost Cij between each node, have
structure may not be the same for another. been calculated.
2) A global minimum is a local minimum with respect to
all possible neighborhood structures. 1) IMPROVED K-MEANS ALGORITHM FOR CAPACITATED
3) For many problems, the local minimum with respect to CLUSTERING PROBLEM
one or more neighborhoods is relatively close to each The K-Means algorithm typically does not consider any
other. constraints when generating clusters because it focuses solely
This empirical observation implies that a local optimum on identifying similar elements. To maintain the vehicle
often provides some information about the global optimum. capacity constraint within each route, an improved version
of K-Means called the Improved K-Means Algorithm for
G. CLARKE AND WRIGHT’S ALGORITHM Capacitated Clustering Problem, has been utilized. This
The Clarke & Wright algorithm (also known as Saving’s algorithm was proposed by Geetha et al. [3] and its main
Algorithm) was introduced by Clarke, G. & Wright, J.W. [13] advantage lies in considering the maximum capacity Q of
in 1964. It is one of the most popular heuristic techniques for the m vehicles in a Capacitated Vehicle Routing Problem
solving the Vehicle Routing Problem (VRP). The central idea (CVRP). To avoid breaking this capacity constraint, nodes are
of this algorithm is to combine two routes into a single new grouped using a prioritization system.
route, ensuring that ‘‘savings’’ are achieved. The potential The Improved K-Means Algorithm incorporates several
savings from combining two routes is given by: modifications to better handle situations where there is an
gij = di0 + d0j − dij . (24) intra-cluster maximum capacity constraint. The customer
nodes are assigned to the nearest groups in a similar
Here, gij represents the savings value when two nodes are manner to any K-Means algorithm, but the assignment occurs
integrated into the same route (greater values indicate higher in an ordered manner based on a characteristic of each
savings); di0 is the distance between node vi and the depot v0 , customer: their demand. Customers with higher demands qi
d0j is the distance between depot v0 and node vj , and dij is the and minimum distances dikj with respect to the centroids
distance between node vi and node vj . are given higher priority when being assigned to a group.
The algorithm operates as follows: first, each customer Therefore, it is necessary to sort the customers (based on
is assigned to a vehicle. Then, it attempts to combine their respective demands) at the beginning of the algorithm
two routes, provided that the vehicle’s capacity allows it. in descending order.
The order in which the routes are combined is determined Once the nodes are sorted, they are assigned to the
using (24). These savings values are sorted in descending groups whose centroid are closest to them. To determine
order, and the route that generates the highest savings is the initial centroid locations, the coordinates of the nodes
selected and added to the merged route. The Clarke & Wright with higher demand that are farthest from the depot and the
algorithm has two approaches: Parallel and Sequential. The already selected centroids are considered. This approach is
Parallel approach creates routes simultaneously, whereas the based on the work of Arthur, D. and Vassilvitskii, S. [14],
Sequential approach creates one route at a time. who demonstrated that the initial position of the centroids
directly affects the final result of the clustering process. They
IV. METHODOLOGY refer to this improved version of the K-Means algorithm as
This section presents the working methodology and the K-Means++.
processes involved in each of the proposed algorithms: Free To determine the optimal number of routes, that is, the
Ant and Restricted Ant. Furthermore, the notation k is used optimal K , the following equation is used:
interchangeably to refer to both centroids and medoids M n
from this point onward. X qi
K= . (25)
The overall behavior of the proposed algorithms in this Q
i=1
work can be observed in Fig. 1, which presents, in a flowchart,
the clustering stage and then the ant colony optimization On the other hand, to calculate the priority of each node vi ,
stage. the next equation is used:
Priority Pi = dikj /qi . (26)
A. CLUSTERING ALGORITHMS
The purpose of these algorithms is to obtain information Here, dikj corresponds to the distance between node vi and
that can assist the ants in making better decisions during the k-th centroid of each j-th group. Once all nodes have been

VOLUME 11, 2023 125805

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

FIGURE 1. Generalized flowchart of the proposed hybrid algorithms.

assigned to a group, the coordinates of each centroid kj is Each node is assigned to its nearest group, that is, the group
recalculated. If convergence is not achieved, the priority Pi whose medoid is closest in terms of distance. Similar to the
needs to be recalculated for a new iteration. Capacitated K-Means approach, a priority system based on
In the context of VRP problems, the optimal number distance and demand was used for node assignment. Once all
of clusters that clustering algorithms should generate will nodes have been assigned to a group, the K medoids must be
always be equal to the number of routes, that is, the optimal recalculated. To choose the new medoid for each group in Gj ,
K . Therefore, the number of clusters is determined by (25). it is necessary to determine which node in the group has the
lowest average cost compared with the other members of the
same group, according to (23).
2) IMPROVED K-MEDOIDS ALGORITHM FOR CAPACITATED
CLUSTERING PROBLEM
The K-Medoids algorithm for Capacitated Clustering follows B. ACO ALGORITHMS
the same principles as the algorithm proposed by Geetha et al. The ACO Algorithms developed in this work are Free Ant
[3], but it is governed by the criteria of K-Medoids. Once and Restricted Ant, and both are based on the BWAS system.
again, we employ the idea of K-Means++ to select the most The execution of these algorithms begins after the clustering
distant nodes as the initial medoids. process.

125806 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

Algorithm 1 Clustering Algorithms Algorithm 2 ACO Algorithms

1: Initialization of variables 1: Initialization of variables
2: K 2: unvisitedNodes
3: totalCost 3: routes
4: function GetFirstCentroids(K ) 4: while length of unvisitedNodes > 0 do
5: for idxk in K do 5: newRoute
6: if centroidsList is empty then 6: actualNode ←depot
7: Find the farthest node from the depot 7: maxCapacity
8: centroidsList ←node 8: vehicleWeight ←tare
9: else 9: Calculate all the validNodes from unvisitedNodes
10: while length of validNodes > 0 do
10: Find the farthest node from the rest of
11: Calculate the heuristics from actualNode to
centroids
validNodes
11: centroidsList ←node 12: Calculate the pheromones from actualNode to
12: end if validNodes
13: end for 13: pseudoRandomChoice
14: return centroidsList 14: if pseudoRandomChoice ≤ l0 then
15: end function 15: Calculate bestNode
16: while actualIteration < maxIterations do 16: nextNode ←bestNode
17: unassignedNodes 17: else
18: Calculate distances from every node to every 18: Calculate all the probabilities with heuristics
centroid and pheromones
19: for node in nodesList do 19: Choose a randomNode based on these
20: if noder exists in unassignedNodes then probabilities
21: Gets closest clusters to noder 20: nextNode ←randomNode
22: for cluster in closestClusters do 21: end if
23: Calculate distances from every node to 22: Append nextNode to newRoute
cluster 23: maxCapacity subtract demand of nextNode
24: Calculate priority for every node to 24: vehicleWeight sum demand of nextNode
25: unvisitedNodes remove nextNode
cluster
26: actualNode ←nextNode
25: Assign to this cluster the node with
27: Recalculate all the validNodes from
highest priority unvisitedNodes
26: Remove this node from unassignedNodes 28: end while
27: if noder not exists in unassignedNodes 29: Append newRoute to routes
then 30: end while
28: Exit this FOR loop 31: return routes
29: end if
30: end for
31: end if
32: end for The algorithm begins by initializing the variables and
33: function CalculateTotalCost setting up the necessary data structures. The main loop
34: Sum all distances from every node to its centroid iterates until all the nodes have been visited. In each
35: return newCost iteration, a new route is constructed. The algorithm selects
36: end function the next node to visit based on the heuristics and pheromone
37: if newCost < totalCost then information. The selection is influenced by a pseudo-random
38: bestClusters ←actualClusters choice using parameter l0 . If the pseudo-random choice is less
39: bestCentroids ←actualCentroids than or equal to l0 , then the best node is chosen. Otherwise,
40: totalCost ←newCost the selection is based on the probabilities calculated from the
41: end if heuristics and pheromones. The selected node is added to the
42: Recalculate centroids new route and the algorithm updates the remaining capacity
43: if actualCentroids = bestCentroids then and weight of the vehicle. The selected node is removed from
44: Exit this WHILE loop the list of unvisited nodes, and the process continues until all
45: end if the nodes have been visited. The resulting routes are returned
46: end while as the outputs of the algorithm.

1) FREE ANT
The pseudo-code for the ACO Algorithms is outlined in The Free Ant algorithm utilizes information obtained from
Algorithm 2. the best K clusters to determine the nodes belonging to each

VOLUME 11, 2023 125807

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

group. In addition, this information is used to construct the where γ and δ are constants that determine the relative
initial pheromone matrix. For this purpose, an amount of importance of Cij and gij , respectively. Thus, the standard
pheromone τmax is assigned to each arc (i, j) connecting probability distribution used by Free Ant and Restricted Ant
nodes within the same group, whereas arcs (i, j) connecting is given by:
nodes from different groups are assigned an initial amount of 
[τ (r, s)]α · ([dij ]β [Cij ]γ [gij ]δ )
pheromone τ0 . The calculation of τ0 is based on the equation

 P α β γ δ
,
 u∈Ja (r) [τ (r, u)] · ([dij ] [Cij ] [gij ] )


developed by M. Dorigo and L.M. Gambardella in their work

on the Ant Colony System [5]: pa (r, s) = if s ∈ Ja (r), .


 0,
1 
τ0 = .

(27)

otherwise,
Lc
(29)
Here, Lc corresponds to the best total energy obtained by a
greedy algorithm. The details of this strategy, along with others (such as
To prevent possible stagnation, the amount of pheromone the use of a candidate node list), can be found in the
associated with each arc (i, j) is limited to the ranges application recommendations of ACO for VRP problems by
{τmin , τmax }. The specific values for the minimum pheromone Bullnheimer, B., Hartl, R., and Strauss, C. [15] and Rizzoli,
amount τmin and maximum pheromone amount τmax are A.E., Montemanni, R., Lucibello, E. et al. [16].
determined according to (15) and (14), proposed in the
Max-Min Ant System (MMAS). 2) RESTRICTED ANT
Once the pheromone matrix is created and initialized, In Restricted Ant, the global problem is decomposed into
iterative execution of the Free Ant algorithm begins. In each smaller sub-problems: groups. Each group corresponds to one
iteration, a limited number of ants are deployed, and to of the K routes. In this manner, we transform the original VRP
maximize exploration, each ant is placed in a different node problem into K TSP problems of lower complexity.
to start its construction. However, to minimize computation In this version of the BWAS algorithm, all elements τij are
time, a candidate node list is used. For the design of the ACO initialized with a value of τmax .
algorithms in this work, it was decided that the composition The iterative process of this algorithm operates differently
of this list would be a homogeneous mixture of the closest from the Free Ant algorithm, as the iterations are performed
nodes to depot v0 in each group, the first selected nodes in for each group. For each group Gi , a determined number of
each route from the solutions with the best results obtained ant deployments are repeated, and each ant a creates only a
in the previous iteration, and a small number of randomly partial solution to the problem, specifically the solution for
selected nodes. that group Gi .
Once ant a is placed in its initial node, it starts constructing Once again, we work with a candidate node list, one list
one of the K routes. The ant continues to select nodes to visit for each group, but this time it will consist of a homogeneous
as long as the accumulated total demand in route r is less distribution of nodes that are closest to the depot, the initial
than the maximum capacity Q of the vehicle. When it is no nodes of the best solutions obtained from the previous
longer possible to add nodes without violating this constraint, iteration, and a small number of randomly selected nodes.
or when all nodes have been visited, the ant returns to the Each ant a starts its tour at one of these candidate nodes and
depot. This means that although there are nodes that can be probabilistically creates a solution based on (29). Naturally,
added to the list representing the current route, the size of ant a can only visit the nodes belonging to the same group.
this list continues to increase by adding more nodes. When Once all iterations in each group are completed, the best
no more nodes are available to visit, a new route is created solution found in each group is combined into a single
for the remaining available nodes. Once all nodes have been final solution, which represents the solution to the original
visited, the quality of the generated solution is calculated. problem.
The quality of the solution is determined by the sum of
all energies required for each route. This objective function C. VNS ALGORITHMS
is given by (1). Thus, we can observe that each ant a in In this study, the VNS was used as a tool to guide the design
the iterative process of the Free Ant algorithm has the total and development of local search algorithms. In our case, these
freedom (probabilistically) to choose which node in the graph types of algorithms require an initial solution, provided by
to visit. Hence its name. the ACO algorithms. Based on these solutions, the algorithm
To improve the performance and decision-making capa- starts exploring the neighborhood space to find (or not) a
bility of each ant a, this work decided to complement the better local optimum.
heuristic information η by considering not only the distance Two VNS-inspired algorithms were used in the program:
dij but also the energy cost Cij and the savings value gij , such one focused on improving the individual solution of a
that: group, that is, the solution obtained by the Restricted Ant,
and another to improve the solution of the entire problem
η = [dij ]β · [Cij ]γ · [gij ]δ . (28) provided by Free Ant.

125808 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

1) 2-OPT VNS TABLE 1. Specifications of the testing machine.

This first algorithm is a simple variant of VNS and was

implemented to improve the solutions delivered by the ants
of the Restricted Ant. As the solutions provided by each
of these ants are solutions to specific groups, the design of
this algorithm is based solely on two intra-cluster operators:
‘‘single-route-relocate’’ and ‘‘single-route-swap’’.
The single-route-relocate operator relocates a node to a
new position within the same route, whereas the single-route-
swap operator exchanges the positions of two nodes, within advantages of K-Means and ACO algorithms. The work of
the same route. Kusumahardhini, N., Hertono, G. F. & Handari, B. D. [18]
was also reviewed, in which they combined K-Means with
ACO to solve the Multiple Traveling Salesman Problem
2) GENERAL VNS (MTSP).
This algorithm corresponds to a more comprehensive design
of the VNS, and its implementation is aimed at improving the
V. EXPERIMENTAL RESULTS
solutions obtained in the Free Ant.
In this section, we will present the results of experiments
In this version of the algorithm, the operators are
obtained with the Free Ant and Restricted Ant algorithms,
divided into two groups: intra-cluster and inter-cluster. The
which were conducted on a series of instances from the
intra-cluster operators are the same as in the 2-OPT VNS,
CVRPLIB library. The libraries used for experimentation
while the inter-cluster operators are as follows: ‘‘two-routes-
correspond to ‘‘Set A (Augerat, 1995),’’ ‘‘Christofides,
relocate,’’ which relocates a node from one route to a new
Mingozzi and Toth (1979),’’ and ‘‘Golden et al. (1998).’’
position in another route; ‘‘two-routes-swap,’’ that exchanges
The choice of these libraries was made because during the
two nodes between different routes; and finally, ‘‘two-routes-
literature review stage of the EMVRP, two articles related to
exchange,’’ which exchanges two consecutive nodes from one
EMVRP were identified. The first corresponds to the authors
route with two consecutive nodes from another route.
Rathnakumara and Rupasinghe [19], who developed a Tabu
The inter-cluster operators are only viable and applicable
Search (TS) algorithm guided by a Machine Learning (ML)
when the newly generated solution does not violate the
model. This algorithm is called ML-TS and was tested by
capacity constraint Q for any of the K routes.
the authors on the ‘‘Set A (Augerat, 1995)’’ instance as an
EMVRP problem.
D. HYBRID ALGORITHMS The second article is by the authors Kramer et al. [20],
Due to the NP-Hard nature of many problems in the field who developed a metaheuristic that combines Local Search
of engineering, especially those of a combinatorial nature, (LS) with an integer programming approach and a speed
various authors have proposed algorithms that successfully optimization algorithm. This metaheuristic is called ILS-
combine different techniques, such as heuristics, metaheuris- SP-SOA and was tested by the authors on the instances
tics, machine learning, etc., in an attempt to solve these ‘‘Christofides, Mingozzi and Toth (1979)’’ and ‘‘Golden et al.
problems. A hybrid algorithm is one that combines two (1998)’’ as an EMVRP problem.
or more algorithms that solve the same problem, either by The specifications of the testing machine used for the
choosing one (based on the data) or switching between them benchmarks are listed in Table 1. The parameter settings
during the course of the algorithm. This is generally done are presented in Table 2. The choice of the values for these
to combine desired features of each algorithm so that the parameters is the result of an experimentation phase where
final algorithm is better than its individual components. Some empirically, values were sought that would generally yield
examples of hybrid algorithms include Machine Learning the best results for both algorithms. The configuration used
(ML) with Tabu Search (TS), iterative algorithms with integer by Stützle, T., Hoos, H. [7] and that of Cordón et al. [2] served
programming, Genetic Algorithm (GA) with Local Search as a starting point, and the search for an optimal configuration
(LS), Ant Colony Optimization (ACO) with Local Search began from there.
(LS), and so on. For this study, each instance of a CVRPLIB library
Since the algorithms proposed in this work are hybrid algo- was executed 10 times for each variant of the proposed
rithms, as they use a combination of clustering techniques algorithms. For each instance of each CVRPLIB library,
with the ACO metaheuristic, a review of the state of the ‘‘Set A (Augerat, 1995),’’ ‘‘Christofides, Mingozzi and Toth
art related to these areas was conducted. We highlight the (1979),’’ and ‘‘Golden et al. (1998),’’ two tables are presented.
recent proposals of two studies focused on using K-Means The first table corresponds to the best result of each program
to subdivide the search space for ACO. execution, that is, the best total energy cost of all routes
One of these studies is by S.A. El-Khatib, Y.A. Skobtsov (objective function (1)) and the total computation time in
& S.I. Rodzin [17], who present an efficient algorithm for seconds to obtain this solution. The calculation of time
image segmentation using a method that leverages all the considers the time required by the clustering algorithm, added

VOLUME 11, 2023 125809

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 2. Parameter configuration. Ant + K-Means, and Restricted Ant + K-Medoids. The best
energy value found is the lowest value in terms of kilojoules
among all the applied algorithms for a particular instance.
The ant-based algorithms are divided into two types: ‘‘free’’
and ‘‘restricted.’’ The ‘‘free’’ algorithms use groups generated
by the clustering-based algorithms to initially populate the
pheromone matrix. In contrast, clustering-based algorithms
are used to establish the search space in the ‘‘restricted’’
algorithms.
During the development of the experiments, we noticed
that in some instances of CVRPLIB, the maximum capacity
constraint per route was very tight compared to the total node
demand. This caused the clustering algorithm (K-Means or
K-Medoids) to be unable to find optimal solutions with each
to the time of the ACO algorithm, and the time of the node assigned to a group. At the end of an iteration of this
VNS local searches. Additionally, a second table presents the algorithm, some nodes remained unassigned. For this reason,
averaged results of the 10 executions and their average Gap. the results tables do not show data for Restricted Ant in
The average Gap is calculated by: certain instances. However, in the case of Free Ant, this issue
did not hinder the generation of the necessary pheromone
Savg − Sbest
Gap = 100 · . (30) matrix. The arcs related to the unassigned nodes still had an
Sbest initial pheromone trail.
where Savg corresponds to the average value of the 10 solu-
tions found for an instance, and Sbest is the value of the best A. SET A (AUGERAT, 1995)
solution found among the 10 solutions. This was the first set of instances considered in this study.
The parameter configurations presented in Table 2 This set consists of 22 groups of instances ranging from 30 to
were used during the execution of the experiments. For 78 nodes (excluding the depot). A vehicle capacity of zero
instances from the libraries ‘‘Set A (Augerat, 1995)’’ and has been considered for this set. The results obtained for this
‘‘Christofides, Mingozzi and Toth (1979),’’ 200 iterations set are presented in Table 3.
were performed for each execution of the Free Ant algorithm,
and 100 iterations were performed per group in the Restricted B. CHRISTOFIDES, MINGOZZI AND TOTH (1979)
Ant algorithm. For instances from the ‘‘Golden et al. This set consists of six groups of instances ranging from 50 to
(1998)’’ library, the number of iterations were 300 and 200, 199 nodes (excluding the depot). A vehicle tare equivalent
respectively. The number of ants per iteration for the proposed to 15% of the maximum capacity was used for this set. For
algorithms was as follows: for Free Ant, it was determined by example, if the vehicle capacity Q is 200 units, the tare would
considering half of the total number of nodes in the problem, be 30 units. 200 iterations were run per execution for the Free
meaning that the number of ants was equal to 50% of the total Ant algorithm, and 100 iterations were run for the Restricted
number of nodes. In the case of Restricted Ant, the number Ant algorithm. The results obtained for this set are presented
of nodes varies depending on the group being solved. The in Table 4.
number of ants was calculated by taking half of the node count
of the current group, resulting in the number of ants being C. GOLDEN ET AL. (1998)
50% of the number of nodes per group. This set consists of 15 groups of instances ranging
In Tables 3-5, we have six columns: Instance, Algorithm, from 200 to 483 nodes (excluding the depot). The same tare
Best Energy (Kj), Avg. Energy (Kj), Gap (%) and Time used for the Christofides, Mingozzi, and Toth set, i.e., 15% of
(Sec). The ‘‘Instance’’ column indicates the name of the the maximum capacity Q, was used for this set as well. The
instance being solved. The ‘‘Best Energy (Kj)’’ column results obtained for this set are presented in Table 5.
provides the best energy value found for that instance in terms
of kilojoules across the 10 executions. The ‘‘Avg. Energy VI. RESULTS ANALYSIS
(Kj)’’ column provides the average energy value across In this section, we analyze the obtained results. To compare
all executions for each algorithm. The ‘‘Gap (%)’’ column the results of the proposed algorithms, we referred to
represents the averaged gap of all executions, as presented two different works. In the first one, Rathnakumara and
in (30). The ‘‘Time (Sec)’’ column indicates the time taken by Rupasinghe [19] presented the ML-TS algorithm, which is
the algorithm to find the best energy value in seconds. Finally, based on clustering and tabu search, to solve the EMVRP.
the ‘‘Algorithm’’ column indicates the algorithm applied to In the second work, Kramer et al. [20] proposed the ILS-SP-
solve the instance and to which others columns values corre- SOA algorithm, which combines exact techniques and local
spond. Four different algorithms are compared in the table: searches to solve the Pollution-Routing Problem and other
Free Ant + K-Means, Free Ant + K-Medoids, Restricted Green-VRP problems.

125810 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 3. Results of Set A (Augerat, 1995).

VOLUME 11, 2023 125811

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 3. (Continued.) Results of Set A (Augerat, 1995).

Both authors achieved favorable results in their respective in most problems. This is evident in the lower values in the
benchmarks using the CVRPLIB library. Time (Sec) column.
Compared to ML-TS, the Free Ant algorithms achieved
A. SET A (AUGERAT, 1995) better results for 20 out of 28 instances in Set A (Augerat,
Table 6 present the best results obtained and compare them 1995), with an average Gap of -7.08% below the results
with the best results of the ML-TS algorithm. Additionally, reported by G. W. H. H. P. Rathnakumara and T. D.
the difference (Gap) between the best energy obtained by Free Rupasinghe. Restricted Ant yielded an average Gap of
Ant + K-Means, Free Ant + K-Medoids, Restricted Ant + K- 12.21% above the results.
Means, and Restricted Ant + K-Medoids is shown compared
to the results provided by Rathnakumara and Rupasinghe. B. CHRISTOFIDES, MINGOZZI AND TOTH (1979)
The calculation of this Gap is done using the next equation: Table 7 presents data corresponding to the ‘‘Christofides,
SML−TS − Sbest Mingozzi and Toth (1979)’’ dataset. In this case, the results
Gap Best Energy = 100 · . (31) obtained in the experimentation phase are compared to
Sbest
those presented by Kramer et al., and unlike Table 6, two
Here, SML−TS corresponds to the best energy obtained by
additional columns are added: ‘‘Avg. Energy (Kj)’’ and ‘‘Gap
ML-TS for the instances in Set A (Augerat, 1995).
Avg. Energy (%)’’. The first column shows the average
The data presented in Table 3 indicate that, in general,
energy obtained in all runs for each instance, whereas the
in terms of energy minimization, the Free Ant algorithms
second column shows a new Gap (33), which evaluates the
achieve better results than the Restricted Ant algorithms.
average differences between each ACO algorithm and that of
Furthermore, Table 3 reveals that Free Ant + K-Medoids
Kramer et al.
achieves a higher solution quality, where lower values
indicate better solutions. On the other hand, the Restricted
SILS−SP−SOAbest − Sbest
Ant algorithms tend to generate lower-quality solutions Gap Best Energy = 100 · . (32)
compared to Free Ant algorithm. However, Restricted Ant Sbest
algorithms are much more efficient, with Restricted Ant + K- SILS−SP−SOAavg − Sbest
Gap Avg. Energy = 100 · . (33)
Means being the fastest algorithm in completing its iterations Sbest

125812 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 4. Results of Christofides, Mingozzi and Toth (1979).

Here, SILS−SP−SOAbest corresponds to the best solution pre- assisted by clusters, it tended to converge towards a set of
sented by Kramer et al. for each instance, and SILS−SP−SOAavg very similar local optima. This behavior did not manifest
is the average of Kramer et al.’s solutions for each instance. in the same way when using Free Ant without clustering;
During the experimentation phase, we decided to test in this case, the convergence occurred in a much more
the effectiveness of using clusters with the ant colony staggered manner throughout the iterations. This discrepancy
optimization metaheuristic. As a result, we conducted tests is attributed to the bias introduced by the information of
with Free Ant, both with and without the use of K-Means cluster arcs in the pheromone matrix. The convergence graph
or K-Medoids. We observed that when the algorithm was without the use of clustering can be seen in Fig. 2, while

VOLUME 11, 2023 125813

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 5. Results of Golden et al. (1998).

125814 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 5. (Continued.) Results of Golden et al. (1998).

FIGURE 2. Convergence chart of the free ant algorithm without the use of FIGURE 3. Convergence chart of the free ant algorithm with the use of
a clustering algorithm in the CMT2 instance. K-Means in the CMT2 instance.

the convergence graph with the use of K-Means is depicted Gap of 30.20% for the Restricted Ant in the same column.
in Fig. 3. In the ‘‘Gap Avg. Energy (%)’’ column, Free Ant had an
Now, according to Table 4, we can see that once again, average gap of 12.90% above ILS-SP-SOA, and 29.88% for
the Free Ant algorithms achieve better solutions in terms Restricted Ant.
of energy consumption than the Restricted Ant algorithms. This time, there is an increase in the average Gap value for
However, it is also evident that the Restricted Ant has both Free Ant and Restricted Ant. This increase in the average
significantly shorter execution times than the Free Ant. This Gap is approximately double that of the Set A (Augerat, 1995)
suggests that allowing ants to explore the search space more dataset.
freely and using clustering techniques such as K-Means and
K-Medoids to guide pheromone formation produces superior C. GOLDEN ET AL. (1998)
results. However, the idea that restricting ant exploration Table 5 presents the results obtained in Golden et al. (1998).
helps reduce the time required to find solutions is also In general, the Free Ant algorithms achieved the best results
confirmed. It is important to note that this restriction limits in terms of the energy consumption. The K-Means and
the ability of the algorithm to find optimal solutions. K-Medoids algorithms showed similar results in terms of the
Table 7 indicate that the Free Ant algorithms achieved number of optimal solutions found. On this occasion, the
better results in 1 out of the 14 instances of Christofides, Restricted Ant algorithms proved to be much more efficient
Mingozzi and Toth (1979), specifically CMT13 using Free and faster in completing the executions, suggesting that
Ant + K-Means. However, an average Gap of 10.33% above for medium to large-sized problems, this type of restricted
the results reported by Kramer et al. was observed in the ‘‘Gap algorithm is a better choice than Free Ant when the runtime
Best Energy (%)’’ column for the Free Ant algorithms, and a is crucial.

VOLUME 11, 2023 125815

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 6. Comparison results of Set A (Augerat, 1995).

125816 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 6. (Continued.) Comparison results of Set A (Augerat, 1995).

Furthermore, the results presented in Table 5 were 24.28% above the results reported by Kramer et al. was
compared once again with those obtained by Kramer et al. observed for the Free Ant algorithms in the ‘‘Gap Best
A comparison with the results of Kramer et al. is presented Energy (%)’’ column, and a 32.87% Gap for the Restricted
in Table 8. The Free Ant algorithms achieved better Ant algorithms in the same column. Regarding the ‘‘Gap
results in one out of the 20 instances of Golden et al. Avg. Energy (%)’’ column, Free Ant had an average of
(1998), specifically in Golden_17 using the Free Ant + K- 27.45% above ILS-SP-SOA, and a 35.00% difference for the
Means algorithm. On the other hand, an average Gap of Restricted Ant.

VOLUME 11, 2023 125817

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 7. Comparison results of Christofides, Mingozzi and Toth (1979).

125818 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 8. Comparison results of Golden et al. (1998).

VOLUME 11, 2023 125819

 N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

TABLE 8. (Continued.) Comparison results of Golden et al. (1998).

Again, we observed an increase in the average Gap value The results obtained in an extensive experimentation
for both algorithms. However, it is important to note that the phase indicate that the algorithms proposed in this work
increase in Restricted Ant is considerably smaller than the are capable of consistently generating high-quality solutions.
increase experienced by Free Ant in relation to the previous In particular, the results show that the Free Ant algorithm
set of instances. This difference suggests that the Restricted outperformed its counterpart, Restricted Ant, in all scenarios,
Ant can operate more efficiently with a smaller number while the latter demonstrated greater robustness and balance
of iterations, while Free Ant requires a higher number of in medium and large-scale instances, primarily due to its
iterations as the problem size increases. superior runtime efficiency. This may be attributed to the
Finally, we encountered a limitation regarding the scalabil- advantages of the clustering algorithms used, which allowed
ity of Free Ant in comparison to Restricted Ant. In the case Restricted Ant to find higher-quality routes in a much more
of the former, both the Gap and the total algorithm execution confined search space. It is important to note that the
time, begin to increase as we attempt to solve problems with performance of the algorithms should be tested on a larger
a greater number of inputs (i.e., a higher number of nodes). number of instances and iterations to draw more conclusive
These issues are not reflected in the same manner in the conclusions.
Restricted Ant algorithms, which prove to be more efficient While the results are promising, it is important to highlight
with larger-scale instances. that there is still room for improvement and further research
Another limitation of the proposed algorithms (Free Ant is needed to achieve more efficient solutions. Green VRP
and Restricted Ant) is that, in order to operate efficiently, problems are still relatively unexplored, leaving an ample
a proper parameter configuration must be established. This scope for further investigations.
entails a preliminary experimentation process to determine Nonetheless, both algorithms are innovative and utilize
the configuration that yields the best results for a specific techniques that have not been extensively explored together
instance. in the resolution of combinatorial optimization problems. The
proposed algorithms are novel in how they use clustering
VII. CONCLUSION techniques to guide the search of the ACO metaheuris-
Combinatorial optimization is a complex and challenging tic. One of the novelties is the combined use of two
field in which significant advances have been achieved clustering algorithms (K-Means with capacity-constrained
through operations research, applied mathematics, and and K-Medoids with capacity-constrained) with the ant
computer science research. In this work, two hybrid ant colony system. Additionally, the Free Ant algorithm is
colony-based algorithms with clustering techniques, namely a completely new proposal to the best of the authors’
Free Ant and Restricted Ant, were presented for solving the knowledge. Future work includes considering the reuse of
Energy Minimizing Vehicle Routing Problem (EMVRP), a clustering algorithms in later stages after initialization, using
‘‘green’’-oriented variant of the Vehicle Routing Problem active re-clustering techniques to modify the ants’ search
(VRP). These algorithms has applications in green logistics space. It is believed that this new approach will help the
and distribution operations. Restricted Ant algorithm improve its results, as it tends to

125820 VOLUME 11, 2023

N. Frías et al.: Hybrid Algorithms for EMVRP: Integrating Clusterization and ACO

converge to worse solutions than Free Ant due to the cluster [12] P. Hansen and N. Mladenovic, ‘‘Variable neighborhood search,’’ in Search
shapes. Methodologies, E. K. Burke and G. Kendall, Eds. Boston, MA, USA:
Springer, 2005, doi: 10.1007/0-387-28356-0_8.
Furthermore, it would be interesting to investigate the [13] G. Clarke and J. W. Wright, ‘‘Scheduling of vehicles from a central depot
combined use of these algorithms with other machine to a number of delivery points,’’ Oper. Res., vol. 12, no. 4, pp. 568–581,
learning techniques to improve the self-regulation of the Aug. 1964. [Online]. Available: http://www.jstor.org/stable/167703
[14] D. Arthur and S. Vassilvitskii, ‘‘K-means++: The advantages of careful
ACO system’s hyperparameters, which are highly variable- seeding,’’ in Proc. 18th Annu. ACM-SIAM Symp. Discrete Algorithms
dependent. Additionally, the implementation of a paralleliza- (SODA), vol. 8, Philadelphia, PA, USA. New Orleans, LA, USA:
tion system for deploying ants in the iterations of Free Ant Society for Industrial and Applied Mathematics, 2007, pp. 1027–1035.
[Online]. Available: https://www.bibsonomy.org/bibtex/2553bbfa74b13c
and Restricted Ant should be considered. 47b4e9c7c0034a8406e/sdo, doi: 10.1145/1283383.1283494.
One of the main challenges researchers in this field [15] B. Bullnheimer, R. F. Hartl, and C. Strauss, ‘‘Applying the ANT system
face is the lack of documented instances for benchmarking. to the vehicle routing problem,’’ in Meta-Heuristics, S. Voß, S. Martello,
I. H. Osman, and C. Roucairol, Eds. Boston, MA, USA: Springer, 1999
Currently, there are only a limited number of standardized doi: 10.1007/978-1-4615-5775-3_20.
and widely used instance sets, such as those available in [16] A. E. Rizzoli, R. Montemanni, E. Lucibello, and L. M. Gambardella,
CVRPLIB. However, these solutions do not always expose ‘‘Ant colony optimization for real-world vehicle routing problems,’’ Swarm
Intell., vol. 1, pp. 135–151, Sep. 2007, doi: 10.1007/s11721-007-0005-x.
the optimal solution for the EMVRP model, where the goal [17] S. S. El-Khatib, Y. A. Skobtsov, and S. I. Rodzin, ‘‘Theoretical and
is to minimize the total energy consumed. Therefore, more experimental evaluation of hybrid ACO-k-means image segmentation
research and collaboration are needed to develop additional algorithm for MRI images using drift-analysis,’’ Proc. Comput. Sci.,
vol. 150, pp. 324–332, Jan. 2019, doi: 10.1016/j.procs.2019.02.059.
instances and optimal solutions for such problems. [18] N. Kusumahardhini, G. F. Hertono, and B. D. Handari, ‘‘Implementation
In summary, the proposed algorithms were able to solve of K-means and crossover ant colony optimization algorithm on multiple
EMVRP problems and achieved good results in the instances traveling salesman problem,’’ J. Phys., Conf. Ser., vol. 1442, no. 1,
Jan. 2020, Art. no. 012035, doi: 10.1088/1742-6596/1442/1/012035.
from CVRPLIB where they were tested. However, although [19] G. W. H. H. P. Rathnakumara and T. D. Rupasinghe, ‘‘Clustering-based
the algorithms did not reach the best results in all instances, augmented Tabu search for energy minimizing vehicle routing problem
(EMVRP),’’ in Proc. 12th Annu. Sessions. Nugegoda, Sri Lanka: Sri Lanka
this work lays the foundation for future research and improve- Association for Artificial Intelligence, 2017, p. 77.
ments in the use of these metaheuristic techniques. The [20] R. Kramer, A. Subramanian, T. Vidal, and L. D. A. F. Cabral, ‘‘A matheuris-
proposed algorithms are innovative in the efficient integration tic approach for the pollution-routing problem,’’ Eur. J. Oper. Res.,
vol. 243, no. 2, pp. 523–539, Jun. 2015, doi: 10.1016/j.ejor.2014.12.009.
of clustering techniques and the ACO metaheuristic.
REFERENCES NICOLÁS FRÍAS received the Graduate degree
[1] I. Kara, B. Y. Kara, and M. K. Yetis, ‘‘Energy minimizing vehicle routing from Universidad de Playa Ancha, Valparaíso,
problem,’’ in Combinatorial Optimization and Applications (Lecture Notes Chile. He is currently a computer engineer
in Computer Science), vol. 4616, A. Dress, Y. Xu, and B. Zhu, Eds. Berlin, in Chile. With a specialization in information
Germany: Springer, 2007, doi: 10.1007/978-3-540-73556-4_9. management, he has built a successful career
[2] O. Cordon, I. F. D. Viana, F. Herrera, and L. Moreno, ‘‘A new ACO as a software developer. He stands out for his
model integrating evolutionary computation concepts: The best-worst ant
ability to merge theoretical concepts with practical
system,’’ in Proc. Ants, 2000, pp. 22–29.
applications.
[3] S. Geetha, G. Poonthalir, and P. T. Vanathi, ‘‘Improved k-means
algorithm for capacitated clustering problem,’’ INFOCOMP J.
Comput. Sci., vol. 8, no. 4, pp. 52–59, 2009. [Online]. Available:
https://infocomp.dcc.ufla.br/index.php/infocomp/article/view/282
FRANKLIN JOHNSON received the B.E. degree,
[4] M. Dorigo, V. Maniezzo, and A. Colorni, ‘‘Ant system: Optimization by
the master’s degree in informatics, and the Ph.D.
a colony of cooperating agents,’’ IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. 26, no. 1, pp. 29–41, Feb. 1996, doi: 10.1109/3477.484436.
degree from Pontificia Universidad Católica de
[5] M. Dorigo and L. M. Gambardella, ‘‘Ant colony system: A cooperative Valparaíso, in 2006, 2010, and 2017, respectively.
learning approach to the traveling salesman problem,’’ IEEE Trans. Evol. He is currently the Head of the Department
Comput., vol. 1, no. 1, pp. 53–66, Apr. 1997, doi: 10.1109/4235.585892. of Computer Sciences, Universidad de Playa
[6] B. Bullnheimer, R. Hartl, and C. Strauß, ‘‘Working papers SFB adaptive Ancha, Valparaíso, Chile. His research interests
information systems and modelling in economics and management include metaheuristics, autonomous search, and
science,’’ WU Vienna Univ. Economics Bus., Working Paper, Apr. 1997 combinatorial optimization. His research interests
[7] T. Stützle and H. Hoos, ‘‘Improvements on the ant-system: Introducing the include metaheuristics, autonomous search, and
MAX-MIN ant system,’’ in Artificial Neural Nets and Genetic Algorithms. combinatorial optimization.
Vienna, Austria: Springer, 1998, doi: 10.1007/978-3-7091-6492-1_54.
[8] F.-Y. Wang and D. Liu, Advances in Computational Intelligence. CARLOS VALLE received the Ph.D. degree in
World Scientific, 2006. [Online]. Available: https://www.worldscientific. computer science from Universidad Técnica Fed-
com/doi/abs/10.1142/6072 erico Santa María, Chile, in 2014. He is cur-
[9] S. Lloyd, ‘‘Least squares quantization in PCM,’’ IEEE Trans. Inf. Theory, rently with the Department of Computer Science
vol. IT-28, no. 2, pp. 129–137, Mar. 1982, doi: 10.1109/TIT.1982.1056489. and Informatics, Universidad de Playa Ancha,
[10] J. B. MacQueen, ‘‘Some methods for classification and analysis
Valparaíso, Chile. From 2014 to 2018, he was a
of multivariate observations,’’ in Proc. 5th Berkeley Symp. Math.
Researcher with the Computer Science Depart-
Statist. Probab., Statist., vol. 1. Berkeley, CA, USA: University of
California Press, Berkeley, 1967, pp. 281–297. [Online]. Available: ment, Universidad Técnica Federico Santa María.
http://projecteuclid.org/euclid.bsmsp/1200512992 Since 2018, he has been an Associate Professor
[11] L. Kaufman and P. J. Rousseeuw, ‘‘Partitioning around medoids with the Department of Computer Science and
(program PAM),’’ in Finding Groups in Data, L. Kaufman and Informatics, Universidad de Playa Ancha, Chile. His research interests
P. J. Rousseeuw, Eds. Wiley, 1990, ch. 2, pp. 68–125. [Online]. Available: include machine learning, ensemble learning, deep neural networks, and
https://onlinelibrary.wiley.com/doi/pdf/10.1002/9780470316801.ch2, doi: pattern recognition with applications to time series.
10.1002/9780470316801.ch2.

VOLUME 11, 2023 125821