Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Applied Soft Computing Journal 108 (2021) 107403

Contents lists available at ScienceDirect

Applied Soft Computing Journal


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

Solving capacitated vehicle routing problem using cooperative firefly


algorithm

Asma M. Altabeeb a , Abdulqader M. Mohsen a,b , , Laith Abualigah c , Abdullatif Ghallab a
a
Faculty of Computing and Information Technology, University of Science and Technology, Sana’a, Yemen
b
Faculty of Computer and Information Technology, University of Aden, Aden, Yemen
c
Faculty of Computer Sciences and Informatics, Amman Arab University, Amman, Jordan

article info a b s t r a c t

Article history: The capacitated vehicle routing problem (CVRP) is a classical combinatorial optimization problem,
Received 28 May 2020 which has received much attention due to its main challenges as distribution, logistics, and transporta-
Received in revised form 20 March 2021 tion. This proposed attempts to find the vehicle routes with minimizing traveling distance, in which the
Accepted 6 April 2021
excellent solution delivers a set of customers in one visit by capacitated vehicle. For solving the CVRP
Available online 18 April 2021
problem, a cooperative hybrid firefly algorithm (CVRP-CHFA) is proposed in this paper with multiple
Keywords: firefly algorithm (FA) populations. Each FA is hybridized with two types of local search (i.e., Improved
Discrete firefly algorithm 2-opt as a local search and 2-h-opt as a mutation operator) and genetic operators. The proposed
Parallel firefly algorithm algorithms (FAs) communicate from time to time for exchanging some solutions (fireflies). The main
Parallel metaheuristics aim of the hybridization and communication strategies is to maintain the diversity of populations to
Local search prevent the proposed algorithm from falling into local optima and overcome the drawbacks of a single
Hybrid firefly algorithm swarm FA. The experiments are conducted on 108 instances from eight standard benchmarks. The
results revealed that the proposed CVRP-CHFA got promising results compared to other well-known
methods. Moreover, the proposed CVRP-CHFA significantly outperformed the recent three hybrid firefly
algorithms.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction One of the most popular VRPs is the symmetric homogeneous


capacitated vehicle routing problem (CVRP). It differs from the
Dantzig and Ramser introduced the vehicle routing problem conventional VRP with capacity limitations, where the carrying
(VRP) in [1]. The difficulties in solving the VRP and its practical capacity limits vehicles. CVRP belongs to the class of NP-hard
importance are the main reasons for the researchers to pay great complete problems [8]. CVRP aims to determine the smallest
attention. Because of VRP’s wide applicability and its importance routing cost of similar vehicles that serve customers with known
in determining efficient strategies for reducing operational costs demands.
in distribution networks, it is so widely studied [2]. In practice, Recently, many optimization algorithms have been proposed
many real-world transportation applications indicate the signif- in the literature to solve CVRP [9]. These algorithms are di-
icance of this problem, particularly when suppliers are to deal vided into three main classes; exact, classical heuristic, and meta-
with many large products, and the freight component is not heuristic algorithms [4]. Metaheuristic algorithms are one of the
trivial [3]. The critical issue in this problem is how to allocate and
best alternatives that have been used to solve the large-scale
order a distributed network geographically by the commitment to
CVRP [10]. Natural phenomena usually inspire this class of algo-
vital roles in reducing the cost and improving the service. Based
rithms; it proved its ability to solve various complex problems in
on the constraints added to VRP, the problem can be classified
many fields. Several algorithms have been suggested for solving
into different VRPs, such as capacitated VRP, VRP pickup and
CVRP [4], such as tabu search, genetic algorithm, ant colony
delivery, VRP with time windows, multiple depot VRP, open VRP,
optimization, cuckoo search, and firefly algorithm. Despite the
and rich VRP. These constraints make the VRP more complex and
difficult to solve [4–7]. efficiency of existing metaheuristic algorithms, they suffer from
several drawbacks that affect their performance, such as slow
∗ Corresponding author. convergence rate, premature convergence, or stagnation. Other
studies are proposed to overcome these drawbacks either by
E-mail addresses: asmam.altabeeb@gmail.com (A.M. Altabeeb),
a.alabadi.fcit@aden-univ.net (A.M. Mohsen), aligah.2020@gmail.com improving existing metaheuristic algorithms or by introducing
(L. Abualigah), ghallab@gmail.com (A. Ghallab). new metaheuristic algorithms [8,11,12].

https://doi.org/10.1016/j.asoc.2021.107403
1568-4946/© 2021 Elsevier B.V. All rights reserved.
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Firefly algorithm (FA) is a metaheuristic optimization method to customer j directly, Yijk = 1 otherwise Yijk = 0. The objective
that has been used in many domains to solve versions optimiza- function of CVRP (Eq. (1)), which minimizes the total distance
tion problems [13]. The primary motivations to investigate the traveled by the vehicles, is subject to some constraints such as
FA in solving the CVRP are due to its advantages, such as its (i) customers can only be serviced by one vehicle (Eqs. (2) and
simplicity, flexibility, and versatility, etc. [14] besides the promis- (3)), (ii) the total demand of all customers on any route must
ing results of the previous version [15]. It is worth to mention not exceed the vehicle capacity Q (Eq. (4)), (iii) all routes must
that the firefly algorithm was used to address CVRP only in three start and finish at the same depot after servicing the customers
studies [15–17]. Like other metaheuristics, the basic FA suffers (Eq. (5)), and (iv) some CVRP instances may have an additional
from different problems leading to the low quality of results. route duration limit, which restricts the duration (or length) of
In our recently published research [15], CVRP-FA revealed any route to a predefined bound D (Eq. (6)) [22,23].
promising results compared to other CVRP algorithms. However,
N N K
one drawback of it is the possibility of getting stuck in local ∑ ∑ ∑
optima. These results can be attributed to the mechanism of Minimize cij Yijk (1)
i=0 j=0 k=1
applying the second local search procedure, i.e., 2-h-opt, which
applied for each solution in every iteration. In CVRP-FA, the solu- Subject to
tion representation is converted from CVRP to Traveling Salesman K N
Problem (TSP). Then the 2-h-opt is used to seek for an existing ∑ ∑
Yijk = 1 j ∈ {1, . . . , N } : i ̸ = j (2)
twisted connection and open it to reduce the cost of the route.
k=1 i=0
This mechanism does not guarantee to produce better solutions
K N
all the time; however, CVRP-FA only accepts the improved so- ∑ ∑
lutions. Hence, the algorithm loses diversity and is likely to fail Yijk = 1 i ∈ {1, . . . , N } : i ̸ = j (3)
in the local optima or converges fast to the nonpromising area. k=1 j=0
N N
These reasons are the main motivation to propose a new, im- ∑ ∑
proved method to solve the mentioned problems and obtain new Yijk qi ≤ Q k ∈ {1, . . . , K } (4)
best solutions for different optimization problems. i=0 j=0
In this paper, an enhanced variant of CVRP-FA called CVRP- N
∑ N

CHFA is introduced to solve the CVRP with two main contri- Yijk = Yjik ≤ 1 for i = 0 and k ∈ {1, . . . , K } (5)
butions. First, the applying of the 2-h-opt [18] as a mutation j=1 j=1
procedure instead of a local search. The key idea of this change N
∑ N

is that the mutation is occasionally applied according to a pre- Yijk (dij + Si ) ≤ D k ∈ {1, . . . , K } (6)
defined probability (mutation rate). On the contrary to CVRP-FA, i=0 j=0
the proposed algorithm accepts all solutions produced by this
change. If the solution is improved, the algorithm intensifies A solution of CVRP is feasible if all routes satisfy the capacity
its search in the promising areas; otherwise, it diversifies the constraint and no customer is visited more than once [24]. For
search space and comes out of the local minima. Second, the the number of vehicles (routes) K that serve the customers, some
concepts of the cooperative island model are utilized to maintain researches such as [25] considered it as a constraint variable
diversity further, avoid trapped in local minima, and speed up while other researches considered it as a decision variable such
the convergence. To the best of our knowledge, there is no study as [26].
apply such a model for FA to solve CVRP. The effectiveness of
the proposed algorithm is tested on 102 instances from seven 3. Related work
standard benchmarks, and the results are compared with recent
optimization algorithms that have been used to solve CVRP such Various optimization algorithms have been used in several
as hybrid firefly [17], CFAGA [16], CVRP-FA [15], DIWO [12], studies to solve CVRP [4]. Kir et al. [27] proposed a heuristic
LNS-ACO [19], ISOS [8], CVRP_GELS [20] and GVNS [21]. The algorithm (HA) based on tabu search (TS) and adaptive massive
obtained results showed that the proposed CVRP-CHFA got better neighborhood search. Dam et al. in [28] introduced a chemical
results in most cases compared to other well-known methods. reactor optimization (CRO) algorithm combined with an adaptive
Furthermore, the Shapiro–Wilk statistical test is employed to an- unified tabu search (UTS) algorithm. Mouaouia et al. in [25]
alyze the obtained results; the proposed CVRP-CHFA significantly developed a new tour split strategy, called Lagrangian split (Ls),
outperformed different optimization algorithms. and integrated it with variable neighborhood search.
The rest of the paper is organized as follows. The CVRP is Besides, ant colony optimization (ACO) was proposed for solv-
formulated in Section 2. Section 3 briefly reviews the meta- ing CVRP. A novel membrane algorithm (MA_ACO) is presented
heuristics and their application to CVRP. Section 4 presents the to solve CVRP [29]. MA_ACO combined membrane systems with
cooperative FA model CVRP-CHFA. The computational results and two versions of ACO. Then Akpinar in [19] used ACO as a neigh-
discussion are introduced in Section 5. Finally, Section 6 presents borhood search to provide a better diversification level. The ef-
the conclusion and highlights future work. fect of the elite pool was studied in the three hybrid meta-
heuristic algorithms, i.e., elitist-ant system, big bang-big crunch,
2. Capacitated vehicle routing problem formulation and scatter search [30]. Wang et al. developed ACO to solve
CVRP [31]. Their algorithm allowed every ant to represent all
CVRP is defined as a graph G = (V , E) consists of a set of nod- paths of a feasible solution instead of a single path. Goel and
es V = {0, 1, . . . , N } and an edge set E = {(i, j) : i, j ∈ V }. Maini enhanced the diversity of ACO by using the firefly algo-
Node 0 represents the depot and the other nodes [1, . . . , N ] rithm [32]. According to their computational results, the authors
represent the customers who have specific demands qi , where claimed that the computation time of the proposed algorithm
i = {1, 2, . . . , N }, to be delivered or picked up. The travel cost was far less compared to other comparative methods in [20,
between node i and j is defined by cij > 0. For a single depot 25,33]. However, this comparison is not entirely fair because
0, a set of homogeneous vehicles K vehicles with capacity limit the computation time is affected by many factors such as the
Q depart from and return to. If vehicle k travels from customer i machine(s) used for conducting the experiments, the operating
2
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

system, the programming language used, the compiler, and fi- sentation to represent the solutions of CVRP [16]. In addition, the
nally, the skills of the programmer. To make a fair comparison, the authors borrow the mutation operator from genetic algorithms
authors should re-implement all the compared algorithms using to reflect the fireflies’ movement. The third is the representation
the same computation environment [34]. of the firefly movement by the Insertion function and computed
Genetic algorithms (GA) was developed for solving CVRP [35– as the hamming distance (HD) between two fireflies [45]. Such
37]. Mazidi et al. in [38] hybridized GA with the ACO algorithm. movement concept has shown promising results. It is worth
The initial population was constructed by the ACO algorithm and noting that FA has also applied the same movement concept for
improved by GA. A new multiple population-based metaheuris- solving TSP and CVRP [15,46,47]. Accordingly, this concept of
tic called golden ball (GB) was developed in [11]. GB outper- movement will be used in this study.
formed GA and distributed genetic algorithms in terms of quality, The FA has been used in few studies to solve VRPs. A discrete
convergence behavior, and robustness.
FA variant was proposed for the time window VRP (VRPTW)
Akhand et al. in [39] applied adaptive sweep and velocity ten-
and another for solving the rich VRPTW in [45,48]. However,
tative particle swarm optimization (VTPSO) algorithms to solve
the results obtained were not promising compared to those in
CVRP. The authors claimed that the use of an adaptive sweep
the literature. To address the CVRP, FA was targeted to optimize
algorithm to construct the initial population of VTPSO is better
than the standard sweep algorithm. Four different chaos maps heterogeneous fixed fleet VRP (HFFVRP) with local search (2-
were embedded with a discrete artificial bee colony algorithm opt) [17]. The authors were forced to use the CVRP dataset for
that was proposed in [40]. The computation results showed that a lack of the HFFVRP dataset in the literature. The experimental
the use of chaos maps with a discrete artificial bee algorithm was results reflect the shortcomings of the proposed algorithm. For
significantly better than stock pseudo-random number genera- example, it was able to solve the small-scale instances only.
tors. A water flow-like algorithm was developed to solve CVRP Furthermore, most of the time, it was stuck in local optima. The
by Zainudin et al. in [41]. The authors claimed that the proposed FA with adaptive mutations and GA (CFAGA) combinations have
algorithm obtained promising results. However, the quality of the been introduced [16]. This research has three drawbacks. First
obtained results was not as good as those presented in the liter- is the use of only five small-scale instances. Second, CFAGA has
ature. Ehsan et al. in [42] proposed an enhanced cuckoo search been unable to meet the BKS of the five cases. Third, comparing
(CS), called ACS, via fragmentation of cuckoos; furthermore, they the proposed algorithm to the simple FA and GA, rather than the
improved intelligent water drops (IWD) via hybridizing the IWD recent existing algorithms in the literature.
with 2-opt heuristic. In conclusion, metaheuristic algorithms that have been pro-
Ruskartina et al. in [43] adapted symbiotic organism search posed to solve CVRP still have drawbacks, which led to low
(SOS) to solve CVRP. To make SOS suitable to CVRP, they added a quality of solutions. However, the appropriate hybridization tech-
sophisticated decoding method in the basic SOS framework. The niques and/or cooperative models improve their performance
authors claimed that SOS could obtain reasonable and promising dramatically. The performance of a metaheuristic algorithm can
results. Then, Vincent et al. in [8] improved SOS by added two be improved by integrating a component of a particular meta-
new interaction strategies, namely amensalism and competition.
heuristic algorithm instead of the entire algorithm. Another di-
The computational results of [8] were significantly better than
rection of improvement is the use of multiple populations (coop-
other methods in [43]. Yanwei et al. in [12] developed a dis-
erative model) of a metaheuristic algorithm. It has been proven
crete hybrid invasive weed optimization algorithm (DIWO). The
from the literature and practical point of view that the hybridiza-
experimental results showed that DIWO could obtain three new
tion and the cooperative model improve the performance of the
solutions better than the best-known solution (BKS). However,
two of these solutions are unreliable for the following reasons: algorithm significantly [21,49–51].
(i) the first solution violated vehicles’ capacity constraint. Thus, it
was an infeasible solution for CVRP; and (ii) there was an error 4. The proposed algorithm (CVRP-CHFA)
when the authors calculated the overall distance of the second
solution. Hosseinabadi et al. in [20] enhanced GELS to solve large
In this section, the proposed CVRP-CHFA algorithm is intro-
size CVRP, called CVRP_GELS. The computational results showed
duced for solving CVRP. CVRP-CHFA is a continuation of our previ-
that CVRP_GELS still consumed more computational time.
ously published work [15]. The improvement involves two folds.
Jianyong et al. in [22] solved CVRP using a cooperative parallel
First, modifying the local search and mutation of the published
metaheuristic based on TS. They used a shared solution pool as
a communication strategy. Ammi and Chikhi, in [44], developed CVRP-FA to produce a new variant called Hybrid FA (HFA). Then,
cooperative parallel metaheuristics using a generalized island building a parallel cooperative model of multi HFA, called CVRP-
model, which consists of ACO and GA. The authors used the ex- CHFA. The purpose of CVRP-CHFA is to improve the problem-
change strategy for global communication between archipelagoes solving process by maintaining the population diversity and ac-
and local communication between islands. Kalatzantonakis et al. celerating the convergence speed simultaneously. In Section 4.1,
in [21] proposed a cooperative general variable neighborhood we illustrate the main steps of the HFA. In Section 4.2, we present
search method, and the results proved that the cooperative model the cooperative HFA model, CVRP-CHFA. Finally, we introduce the
provided better quality solutions for CVRP. main phases of the CVRP-CHFA algorithm in Section 4.3.
FA is a nature-inspired metaheuristic that simulates the pat-
terns and behavior of the fireflies in nature. Yang first developed
4.1. Hybrid FA for CVRP (HFA)
it in 2007 [13]. FA has initially been designed to tackle the con-
tinuous optimization problems and cannot be used to address the
discrete problems directly. Therefore, FA should be modified to In HFA, two types of hybridization are involved. First, FA is
address this kind of problem, such as VRPs [45]. Three strategies hybridized with local search techniques to speed up the con-
have been suggested in the literature to make FA efficient in vergence rate. Second, FA is hybridized with genetic operators
solving VRPs. The first is the adaptation of the CVRP with the real- (i.e., crossover and mutation) to equalizer the diversification and
coded scheme [17]. The results obtained by such adaptation were intensification search strategies. Algorithm 1 displays the general
not good. The second is the use of FA with permutation repre- structure of HFA, and its main steps are outlined as follows.
3
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Algorithm 1: Pseudo-code of hybrid FA list and deleted from the same list to ensure that the customer is
1 begin
not selected again (Eqs. (2) and (3)); (iii) the selected customer
2 Initialize algorithm parameters (population size (P-S), maximum has added to the route accordingly; (iv) If the total demand of all
number of iterations (MI), crossover rate (C-R), and mutation rate customers in this route <= vehicle’s capacity (Eq. (4)) and the
(M-R)); route’s duration limit <= the duration (Eq. (6)); steps ii and iii
3 Initialize firefly population with their positions (solutions)
X = (x1 , x2 , ...., xP −S );
are repeated; otherwise the endpoint (the depot) is added to the
4 Evaluate the objective function of each position F (xi ) ; route (Eq. (5)); (v) the number of routes is increased by 1; and
5 Calculate the light_intensity at each position (xi ) according to Eq. (7); (vi) steps i to v are repeated until all costumers are served.
6 Sort the population and determine the best position x∗ ; After that, for each generated solution (xi ), its objective func-
7 t=0 ; /* t is the iteration number */
8 while stopping criteria are not met do
tion is calculated. The objective function is the total distance of
9 for i= 1 to P-S do all routes in the corresponding solution. The distance of each
10 for j = 1 to i do route is calculated by the sum of euclidean distances between
11 if Light_intensityi > Light_intensityj then all its costumers including the depot in the start and the end
12 if rand < C − R then
13 Apply PMX procedure between (xi ) and x∗
points. Here, the more attractive solution has the lowest objective
according to C-R; function value since CVRP is a minimization problem.
14 end
15 Apply improved 2-Opt local search for (xi );
4.1.3. Calculating the fireflies’ light intensity
16 Apply one mutation_procedure for (xi ) according to
M − R; HFA is determining the light intensity (Ii ) of a solution i start-
17 Evaluate the objective function of each new ing by calculating the Hamming Distance (HD) as the number of
position F (xi ) ; non-matching items in position xi of solution i and xbest . After
18 end calculating HD, Eq. (7) is used to assigning the light intensity (Ii )
19 end
of the solution i.
20 end
21 Assign the light_intensity at each new position (xi ) using Eq. (7); Ii = Random(1, HDi,best ) (7)
22 Sort the population and determine the best position x∗ ;
23 t=t+1;
24 end 4.1.4. Crossover
25 Return global best solution x∗ ; The purpose of applying crossover is to increase the exploita-
26 end
tion ability of HFA. The crossover is performed between the
solution position xi and xbest position to produce two new po-
sitions H1 and H2 . Consequently, the new solution position x′i
is the best position among xi , H1 and H2 . Since the solution is
represented by permutations, the partially matched crossover
(PMX) is applied.

4.1.5. Local search


Instead of using two types of local search as in the CVRP-
FA, only improved 2-opt local search [12] is applied in HFA.
Improved 2-opt will move solution i to its new position x′i , which
Fig. 1. Two solutions representations.
is supposed to be the best position in the neighborhood. This step
aims to accelerate convergence and improves the quality of the
solutions.
4.1.1. Solution representation
HFA has two representations to represent the CVRP solution.
Firstly, an array list is used for CVRP solution representation 4.1.6. Mutation
(CVRP-SR), as shown in Fig. 1(a). Secondly, one dimension array HFA combines three separate mutation procedures to preserve
(N) is used for the TSP solution representation (TSP-SR), where the diversity of the solutions populations and to avoid them
N represents the number of customers in addition to the depot , from being stuck in local minima. These mutation procedures
as shown in Fig. 1(b). These representations help HFA to convert are applied to adjust the position xi of solution i to produce
from CVRP to TSP in an easy way and vice versa during the run. a new position x′i . The first is the 2-h-opt mutation, which is
The key idea of this exchange is explained in Section 4.1.5. one of the most effective mutations for complex problems [36].
This mutation starts by converting solution representation from
4.1.2. Initialization multiple routes to a single rout representation according to a
Firstly, HFA’s parameters, including probabilities of mutation predefined probability M − R. In other words, representation is
(M-R) and crossover (C-R), population size (P-S), the maximum converted from CVRP-SR to TSP-SR. Then, the 2-h-opt is used to
number of iterations (MI), are initialized. Then, the initial popu- seek for an existing twisted connection and open it to reduce the
lation is generated. cost of the route. After that, the CVRP-SR is rebuilt as a multi-
In the literature, two methods were used to initialize the route representation. This process leads to exploring a new area
population (solutions) either by applying heuristics [8,22,39] or of the search space. With the probability of 0.5, HFA applies the
randomly [11,26,27]. According to Osaba et al. [52], the most ap- second mutation; otherwise, it applies the third one. The second
propriate way to prove the optimization quality of meta-heuristic mutation selects two customers randomly for swapping from
algorithms is the random initialization process. Therefore, the the same route in the CVRP-SR solution. In contrast, the third
initial population of the proposed algorithm is randomly gener- mutation selects two customers randomly for swapping from
ated. Each solution (firefly) (xi ) is constructed according to the two different random routes in the CVRP-SR solution. Algorithm
following steps: (i) a route is created with a start point (the 2 illustrates the three different types of mutation operation in
depot); (ii) a random customer is selected from the costumers’ pseudo-code.
4
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Fig. 2. The cooperative CVRP-CHFA model with a communication strategy.

Algorithm 2: Pseudo-code of mutation procedure


1 begin
2 Input: The current solution position xi and M − R;
3 Output: The solution position x′i ;
4 Generate rand = random(0,1);
5 if rand <= M − R then
6 Transform xi to TSP-SR y;
7 Enhance y by 2-h-opt;
8 Transform y to CVRP-SR x′i ;
9 end
10 Generate rand = random(0,1);
11 if rand <= M − R then
12 Generate R = random(0,1);
13 if R <= 0.5 then
14 Step1: for solution position xi , select one route r1 , randomly;
15 Step2: select two customers c1 and c2 ,in the route r1 at
random;
16 Step3: Swap xi (r1 (c1 )) and xi (r1 (c2 )); Fig. 3. A full connected, cooperative HFA model.
17 end
18 else
19 Step1: for solution position xi , select two routes r1 and r2 ,
run independently. At every communication interval, HFAs com-
at random;
20 Step2: select two customer c1 from route r1 and c2 from municate with each other by exchanging some solutions accord-
route r2 randomly; ing to a predefined communication strategy. The communication
21 Step3: swap xi (r1 (c1 )) and xi (r2 (c2 )); topology is fully connected (each HFA can communicate with any
22 end other HFA in the model).
23 end
The following parameters are used to control the communica-
24 Return the modified solution position x′i ;
25 end
tion between the HFAs:
1. Number of HFAs(n): determines semi-isolated HFAs’ num-
ber in the cooperative HFA model i.e. HFAs = {P1 , P2 , . . . , Pn }.
Thus, the total number of solutions in the model (µ) is defined
4.1.7. Termination criterion by Eq. (8).
Steps of HFA (i.e., crossover, local search, and mutation) are n

repeated until one of the two termination criteria is met; either µ= Pi (8)
reach the maximum number of iterations or find the best-known i=1
solution (BKS). Then, the best solution x∗ is returned.
Generally, each HFA has the same size of the population, Pi . It
is worthy to note that the selection of n and Pi is restricted by the
4.2. Cooperative HFA model (CVRP-CHFA)
available computational resources.
The cooperative HFA model (CVRP-CHFA) has multiple pop- 2. The Communication Interval (CI): determines the number
ulations that exchange solutions (fireflies) through populations of iterations to start communication between HFAs. It determines
occasionally. As shown in Fig. 2, several hybrid HFAs are ran- when solutions should be transferred from one HFA to another.
domly initialized. Each of HFA creates and evolves its population One critical issue in the design of the cooperative HFA model is
independently. They only communicate and exchange solutions the chosen length of the communication interval. The suitable
between them when the communication interval is triggered. interval length allows each HFA to reach a local optimum. That is,
In this model, CVRP-CHFA is multi populations of n different every HFA lost the diversity in each interval. This diversity may
HFAs [P1 , P2 , P3 , . . . , Pn ] where n is the number of HFAs in the regain again after the communication process, where migration
model. HFAs are exploring the research space independently. occurs between HFAs. If the interval length decreases, the co-
Each of these is randomly initialized and distributed through the operative model works more closely, like one large HFA. That is
available processors in the multi-thread architecture. The HFAs mean a large HFA has a fast convergence rate, which may affect
communicate from time to time to exchange some solutions for the solution quality. On the other hand, if the interval length
sharing the learning experience. More precisely, some solutions increases, the cooperative model works more closely like several
from HFA Pi , where i is the HFA index, are sent to other HFA independent sequential HFA. Therefore, it may be trapped in local
P(Rand(1,n)) . optima. Accordingly, the medium length of the communication
Fig. 3 shows a virtual representation of a cooperative HFA interval is expected to provide an appropriate balance between
model with three HFAs. Each HFA has its parameters, and they diversification and intensification.
5
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

4.3. CVRP-CHFA algorithm steps

CVRP-CHFA is divided into four main steps. The first step is


the initialization step where CVRP-CHFA parameters, including
the maximum number of iterations (MI), number of HFAs(n),
communication interval (CI), and communication rate (CR), are
initialized.
The second step is the creation of new HFA populations, where
number of new HFAs = {P1 , P2 ,..., Pn } are generated indepen-
dently. Each HFA’s population Pi is initialized as an array-list of
size P − S i.e., population size. The HFA parameter values, i.e., P-S,
Fig. 4. A full connected graph.
C-R, and M-R, are assigned to each HFA. Each of HFA is assigned
to a thread.
The third step is the HFA populations update. Each population,
3. The Communication Rate (CR): defines as the percentage Pi in the model, is updated twice. First, as in HFA, each solution i
of solutions selected for exchange among the HFAs. The main moves from its current position xi to an enhanced position x′i by
component of the cooperative HFA model is the communication applying crossover, local search, and mutation. Second, in case
between the solutions to address the local optima in the source the communication interval is reached, communication occurs
and destination HFAs. Since different HFAs use different initial among the HFAs. The source HFA Pi , where i is the HFA index,
solutions, they may converge to different local solutions. To avoid tries to influence the destination HFA P(Rand(1,n)) by sending some
local optima, some solutions are selected to migrate from one of its solutions. After the source and destination HFAs have been
HFA to another at predefined communication intervals. In gen- determined, some solutions equal to communication rate are
eral, a high communication rate will rapidly exchange solution transferred according to the communication policy. After each
materials among the HFAs. This rate may lead to the global mix- stage of communication, each HFA will absorb the transferred
ing of solutions where the HFAs virtually are reduced to a single solutions by reordering its solutions.
large population, which may lead to losing the diversity. Loss of Finally, the fourth step is the termination criterion check. In
diversity may imply the premature convergence of the HFAs to this step, the algorithm stops the search process if the maximum
local optima. On the other hand, the low communication rate may number of iterations is reached or one of HFAs found the optimal
lead to isolated HFAs; thus, the algorithm may suffer from slow solution; otherwise, it repeats step three. Algorithm 4 shows the
convergence. Consequently, the average communication rate may pseudo-code of the proposed CVRP-CHFA algorithm.
lead to maintain the balance between diversification and intensi-
fication. The pseudo-code of the communication model structure
Algorithm 4: CVRP-CFA algorithm
is listed in Algorithm 3.
1 begin
2 Initial FA parameters for each HFA;
Algorithm 3: Pseudo-code of exchange solutions procedure 3 for i = 1 to n do
4 Initialize random population (Pi ); /* for HFA[i] */
1 begin
5 for j = 1 to P − S do
2 Input: Source_HFA, Destination_HFA, C-R;
6 objectiv efunction[j] = E v aluate(Pi [j]);
3 for i=1 to Communication_Rate do
7 end
4 X= select a random solution from Source_HFA;
5 Y= find the worst solution in Destination_HFA; 8 end
6 Let Y=X; 9 t=0 ; /* t is the iteration number */
10 while termination criterion not reached do
7 end
11 for i = 1 to n do
8 end
12 HFA (Pi ) ; /* execute the HFA independently */
13 end
14 if t MOD communication_inter v al == 0 then
4. The Communication Topology: the HFAs interconnection is 15 for m = 1 to n do
typically referred to as the communication topology. The cooper- 16 Select some solutions to transfer based on
ative model can be viewed as a connected graph. Fig. 4 shows a communication rate from HFA[m];
topology of a fully connected graph. Given n HFAs, each HFA is a 17 Send selected solutions to another population HFA[rand]
based on communication topology;
vertex in the graph. These vertices are connected by edges spec-
18 Replace a number of solutions in HFA[rand] with the
ifying communication links between the existing vertices. Thus, incoming solutions based on communication policy;
each HFA can communicate with any other HFA in the model 19 end
during the communication process to send or receive solutions. 20 Sort each HFA and determine its best position x∗ ;
At the other extreme, as the connectivity of the topology de- 21 end
creases, the cooperative model approximates more closely several 22 t=t+1;
independent runs of a sequential HFA. 23 end
24 Return the best solution among all HFAs;
5. The Communication Policy: specifies the manner of select-
25 end
ing solutions, for exchange, from source HFA, and substituting
them in the destination HFA. Standard communication policies
send a percentage of random solutions or the best solutions. A
potential drawback of selecting the best solutions is that some- 5. Computational results and discussion
times the majority of the best solutions may approach each other,
especially at the end of a run when the HFA has converged. This The experimental results of the CVRP-CHFA are discussed in
leads to the sending of very similar solutions that cover a minimal this section to analyze its efficiency. CVRP-CHFA is encoded in
area of the search space. In the event of a local minimum being java and runs code on 12 PC with Intel i-5 at 3.2 GHz CPU and
found, these similar solutions may not contain enough solution 4 GB RAM on 32-bit Windows 7.
materials for helping HFA to jump out from the local minima and CVRP-CHFA is investigated on 102 instances from seven stan-
find the overall global maxima. dard benchmarks (Augerat et al. A, B, and P [53], Christofides
6
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 1 Table 3
Benchmarks with their instances and range. S/N ratios obtained from the Taguchi experimental design.
Benchmarks #instances Range Level Number of HFAs Communication interval Communication rate
A 27 32–80 1 −62.13 −62.12 −62.11
B 22 31–78 2 −62.12 −62.12 −62.11
E 9 22–101 3 −62.11 −62.11 −62.11
P 24 16–101 4 −62.11 −62.11 −62.11
F 2 72–135 5 −62.10 −62.11 −62.11
M 4 101–200 Delta 0.03 0.01 0.00
CMT 14 51–200 Rank 1 2 3

Table 2
The orthogonal array L25 .
Experiment Number of HFAs Communication Communication
interval rate
1 7 25 10
2 7 50 20
3 7 70 30
. . . .
. . . .
. . . .
25 15 125 40

and Eilon E [54], Fisher F [55] and Christofides M and CMT [56])
as illustrated in Table 1. Each instance is executed ten times to
obtain results, including the best and average found solutions,
and tie together numerous text scripts and EXCEL sheets. The
CVRP-CHFA stopping criterion is met either when it reaches the
maximum number of iteration, i.e., 2000 iterations, or obtains the
BKS.
Fig. 5. Main effects plot for S/N ratio (higher value is better).
5.1. CVRP-CHFA’s parameters tuning

In this section, the values of CVRP-CHFA’s parameters will be It is clear from Fig. 5 that the better performance and ro-
specified. The parameters of each HFA in the cooperative model, bustness of CVRP-CHFA is when its parameters: number of HFAs,
i.e., population size (P-S), crossover rate (C-R), mutation rate communication interval, and communication rate are set to the
(M-R) are determined by [15]. levels 5, 3, and 4, respectively. It is worthy to note that CVRP-
To tune and choose the proper combination of the cooperative CHFA performance increase as the number of HFAs increase.
model’s parameters, the statistical Taguchi experimental design The results of the Taguchi experiment indicate that the more
was applied. These parameters are the number of HFAs(n), com- the number of HFAs, the better the results which the algorithm
munication interval (CI), and communication rate (CR). Taguchi produced. Due to the limitation of computational resources such
classifies parameters into controllable and noise parameters. This as the number of processors and memory space, no further exper-
method is used to minimize the effect of noisy parameters and iments will carry out. Furthermore. The communication interval
determine optimal levels of important controllable parameters must be long enough to allow for the exploitation of the search
based on the concept of robustness [57]. space. Therefore, it is important to keep it big enough. Too fre-
Minitab software is used to generate and analyze the Taguchi quent communication should be avoided because it makes HFAs
results. CVRP-CHFA’s parameters and their levels are as follows: exchange solutions very fast, such that all the HFAs begin to
(i) number of HFAs (7, 9, 11, 13, 15); (ii) communication interval share the same solutions. This loss of diversity negatively affects
(25, 50, 70, 100, 125); and (iii) communication rate (10, 20, 30, the performance of CVRP-CHFA. Statistically, the results indicated
40, 50). In an experiment where classic full factorials are used, that interval length at lower or equal to 100 iterations yielded
35 = 243 experiments would be needed for each observation good results. The communication rate must be optimum and not
disruptive. Statistically, the results indicate that the communica-
value. However, the Taguchi method generates an appropriate
tion rate of 20% of HFAs is equal roughly the communication rate
orthogonal array with 25 experiments (L25 ), which provides a
of 40%. Moreover, the results demonstrate that a communication
balance among the orthogonal index, parameters, and levels to
rate over 40% is disruptive, and when it is less than 20%, it
identify the effective parameters and levels as shown in Table 2.
is not enough (isolated). The results show a negative influence
To conduct the 25 experiments, four instances, i.e., A-n80-k10,
of big communication (close to HFAs sizes). This influence may
P-n101-k4, E-n101-k14, and M-n200-k17, are used. Each instance contribute to the loss of diversity.
is executed ten times for each combination to increase the re- According to the discussion above and the Taguchi method,
liability of the experiments. The average results of all different with very little change, the better performance and robustness of
observation values are evaluated by transforming the results into CVRP-CHFA can be obtained when its parameters are adjusted to
the signal-to-noise ratio (S/N ratio) in the Taguchi experimental the values shown in Table 4.
design. The S/N ratios are used to identify the control factor
settings that minimize the variability caused by the noise factors. 5.2. Comparison to the state-of-the-art algorithms
From Table 3, it is clear that the effect of parameters on the S/N
ratio had the following order: number of HFAs (Rank = 1), fol- CVRP-CHFA performance obtained is compared with the cor-
lowed by communication interval (Rank = 2) and communication responding performance obtained by the previous version CVRP-
rate (Rank = 3). FA [15]. Furthermore, The results obtained using CVRP-CHFA are
7
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 4 5.2.1. Comparison to the previous version CVRP-FA


Parameter settings for CVRP-CHFA. To evaluate CVRP-CHFA against the previous version CVRP-FA,
Parameters Value four different comparisons are conducted. The first comparison is
Number of HFAs (n) 15 between our three firefly variants with different strategies. The
Communication interval (CI) 70
first one is the previous version CVRP-FA [15] that applies 2-h-opt
Communication rate (CR) 20%
Population size (P-S) 110 by transform CVRP to TSP in the local search phase. This variant
Crossover rate (C-R) 0.95 outperformed all other FA algorithms for solving CVRP. The sec-
Mutation rate (M-R) 0.1 ond one is HFA that applies 2-h-opt by transforming CVRP to TSP
in the mutation phase, and the third one is the cooperative HFA
Table 5
model (CVRP-CHFA). These variants are evaluated and compared
Algorithms-specific parameter settings. using ten CVRP instances representing different benchmarks.
Algorithm Parameter settings As shown in Table 6, with respect to the best solution, all
CVRP-CHFA MI = 2000, P-S = 110, C-R = 0.95, M-R = 0.1, n = 15, CI variants attained similar results to BKS in the following instances
= 70, CR = 20% (A-n36-k5, A-n37-k6, B-n35-k5, E-n33-k4, E-n51-k5, P-n20-k2, P-
CVRP_FA MI = 1000, P-S = 110, C-R = 0.95, M-R = 0.1 n50-k7, P-n60-k15). However, HFA and CVRP-CHFA outperformed
DIWO NA CVRP-FA in the B-n45-k6 instance. Moreover, CVRP-CHFA sig-
ISOS MI = 500, 1000, 1500, eco_size = 25, 50, 75, T0 = 15, 25,
35, pf = 0.7, 0.8, 0.9
nificantly achieved a better solution than HFA and CVRP-FA for
HA NA the A-n32-k5 instance. In terms of the average solution (Avg.),
LNS-ACO MI = 5000, r1 = 0.5, r2 = 0.3, r3 = 0.2, α = 1, β = 1, q HFA and CVRP-CHFA significantly outperformed CVRP-FA for all
= 4 ≤ randi() ≤ min(100, 0.4n), p = 6 instances with superiority for CVRP-CHFA. The optimum solu-
CVRP_GELS NA
tions reached are 2,3 and 9 for CVRP-FA, HFA, and CVRP-CHFA,
GVNS MI = 300, 500, 800, thread = 28
respectively. The number of iterations (Iter.) shows how HFA
MI is the maximum number of iterations; P-S is population size; C-R is crossover was improved over CVRP-FA and how CVRP-CHFA was improved
rate; M-R is mutation rate.
n is the number of HFAs; CI is communication interval; CR is the communication
over HFA. The average number of iterations needed by CVRP-
rate. FA, HFA, and CVRP-CHFA was 290, 69.4, and 9, respectively. The
eco_size is population size; T0 is initial temperature; pf is parasite_force. superiority of HFA over CVRP-FA is referred to using 2-h-opt as
r1, r2, r3 are a random number ∈ [0, 1]; q defines the number of nodes removed a mutation instead of a local search technique. In addition to
from the solution at each iteration.
this, the superiority of CVRP-CHFA is referred to the utilization of
p avoids removing the same customers again and again.
the parallelized multi-population model in the search process of
the algorithm, which speeds up the convergence and fosters the
diversity in case the solutions are trapped into the local optima.
compared to those obtained using five other metaheuristic algo- In the second comparison, the performance of the CVRP-CHFA
rithms, including DIWO [12], ISOS [8], HA [27], LNS-ACO [19], algorithm is verified with the CVRP-FA algorithm using 80 in-
and CVRP_GELS [20]. Finally, CVRP-CHFA is tested against the stances from four standard benchmarks covering Augerat et al.
recently developed parallel GVNS algorithm [21]. The values of A, B, P [53], Christofides and Eilon E [54].
the parameters of each algorithm are shown in Table 5. Tables 7–10 show the comparison of CVRP-CHFA results with
In the following tables of the results, the best-obtained so- CVRP-FA results. In these tables, CVRP instances and their cor-
lution is highlighted in bold. The new optimal solution better responding BKS are placed in columns 1 and 2, respectively.
than the BKS with the same number of routes is highlighted in The other columns represent the Best, B_Gap, Avg., A_Gap, and
bold, followed by ‘‘*’’. The newly obtained optimal solution better iteration (Iter.), respectively. Observations of Tables 7–10 indicate
than the BKS with an increasing number of routes by one is that the CVRP-CHFA algorithm outperforms, in terms of solution
highlighted in bold, followed by ‘‘**’’. The new obtained optimal quality, CVRP-FA.
In detail, the CVRP-CHFA algorithm can produce significantly
solution better than the BKS with a decreasing number of routes
better results in comparison to the CVRP-FA algorithm, according
by one is highlighted in bold, followed by ‘‘***’’. By running the
to the reported results in the literature. CVRP-CHFA algorithm has
proposed algorithm independently ten times, we record the best
less average of A_Gap values for the average of best results for all
(Best), and average (Avg.) The Euclidean distances are calculated benchmarks. Similar to CVRP-FA, it can obtain the best solution
as arc costs for all the test problems. Comparisons are made better than the recorded BKS in the literature for three instances
taking into account the relative best percentage gaps, B_Gap = from the benchmark P.
((Best − BKS)/BKS) ∗ 100, from BKS values and the relative average In terms of the number of iterations (Iter.), CVRP-CHFA out-
percentage gaps, A_Gap = ((Av g . − BKS)/BKS) ∗ 100, from BKS performs CVRP-FA in all instances. The number of iterations for
values. CVRP-CHFA that needs to reach optimal or near-optimal solutions

Table 6
Computational results of CVRP-CHFA, HFA and CVRP-FA.
Instance BKS CVRP-FA HFA CVRP-CHFA
Best Avg. Iter. Best Avg. Iter. Best Avg. Iter.
A-n32-k5 784 796 798.1 830 796 797.3 420 786 792.2 54
A-n36-k5 799 799 804.2 49 799 803.3 1 799 799 1
A-n37-k6 949 949 955.5 399 949 952 24 949 949 1
B-n35-k5 955 955 955.1 406 955 955 49 955 955 1
B-n45-k6 678 686 692.8 560 678 681.8 30 678 678 1
E-n33-k4 835 835 845.2 168 835 838.7 4 835 835 1
E-n51-k5 521 521 521 2 521 521 1 521 521 1
P-n20-k2 216 216 216 3 216 216 1 216 216 1
P-n50-k7 554 554 557.3 131 554 557.1 115 554 554 4
P-n60-k15 968 968 983.2 352 968 977.3 49 968 968 15

8
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 7
Computational results of CVRP-CHFA and CVRP-FA for benchmark A.
Instance BKS CVRP-FA CVRP-CHFA
Best B_Gap Avg. A_Gap Iter. Best B_Gap Avg. A_Gap Iter.
A-n32-k5 784 796 1.5 798.1 1.8 830 786 0.26 792.2 1.05 54
A-n33-k5 661 661 0 661 0 2 661 0 661 0 1
A-n33-k6 742 742 0 742.7 0.09 4 742 0 742 0 1
A-n34-k5 778 778 0 778 0 5 778 0 778 0 1
A-n36-k5 799 799 0 804.2 0.65 49 799 0 799 0 1
A-n37-k5 669 669 0 669 0 27 669 0 669 0 1
A-n37-k6 949 949 0 955.5 0.69 399 949 0 949 0 1
A-n38-k5 730 730 0 730.7 0.1 26 730 0 730 0 1
A-n39-k5 822 822 0 822 0 1 822 0 822 0 1
A-n39-k6 831 831 0 834.6 0.43 25 831 0 831 0 1
A-n44-k6 937 937 0 937 0 5 937 0 937 0 1
A-n45-k6 944 953 0.95 959.4 1.63 474 944 0 944 0 2
A-n45-k7 1146 1147 0.09 1153.3 0.64 180 1146 0 1147.3 0.11 1
A-n46-k7 914 914 0 914 0 5 914 0 914 0 1
A-n48-k7 1073 1073 0 1073 0 11 1073 0 1073 0 1
A-n53-k7 1010 1011 0.099 1014.8 0.48 5 1010 0 1010.5 0.05 1
A-n54-k7 1167 1172 0.43 1172 0.43 22 1167 0 1167 0 2
A-n55-k9 1073 1074 0.093 1078.6 0.52 320 1073 0 1073.3 0.03 4
A-n60-k9 1354 1355 0.07 1364.7 0.79 533 1354 0 1355.6 0.12 7
A-n61-k9 1034 1039 0.48 1048.4 1.39 532 1035 0.1 1036.7 0.26 12
A-n62-k8 1288 1298 0.78 1312.2 1.88 418 1294 0.47 1297.6 0.75 1219
A-n63-k10 1314 1314 0 1333.4 1.48 320 1315 0.08 1315 0.08 8
A-n63-k9 1616 1630 0.87 1644.8 96 1.78 1616 0 1616 0 6
A-n64-k9 1401 1420 1.36 1424.9 1.71 6 1411 0.71 1411.4 0.74 75
A-n65-k9 1174 1178 0.34 1180.7 0.57 8 1177 0.26 1177 0.26 2
A-n69-k9 1159 1162 0.26 1174 1.29 835 1159 0 1160.2 0.10 6
A-n80-k10 1763 1773 0.57 1787.2 1.37 65 1763 0 1768.3 0.30 103

Table 8
Computational results of CVRP-CHFA and CVRP-FA for benchmark B.
Instance BKS CVRP-FA CVRP-CHFA
Best B_Gap Avg. A_Gap Iter. Best B_Gap Avg. A_Gap Iter.
B-n31-k5 672 672 0 672 0 7 672 0 672 0 1
B-n34-k5 788 788 0 789.7 0.22 83 788 0 788 0 1
B-n35-k5 955 955 0 955.1 0.01 406 955 0 955 0 1
B-n38-k6 805 806 0.12 806.2 0.15 134 805 0 805 0 1
B-n39-k5 549 550 0.18 553.9 0.89 195 549 0 549 0 1
B-n41-k6 829 829 0 829.8 0.1 20 829 0 829 0 1
B-n43-k6 742 742 0 742 0 1 742 0 742 0 1
B-n44-k7 909 909 0 913.3 0.47 207 909 0 909 0 1
B-n45-k5 751 751 0 754.2 0.43 149 751 0 751 0 1
B-n45-k6 678 686 1.18 692.8 2.18 560 678 0 678 0 2
B-n50-k7 741 741 0 744.8 0.51 70 741 0 741 0 1
B-n50-k8 1312 1318 0.46 1329.6 1.34 586 1312 0 1314.2 0.17 427
B-n51-k7 1032 1032 0 1021 −1.07 289 1032 0 1024.3 −0.75 1
1016a −1.6 1017.4 −1.41 93 1016a −1.6 1016 −1.6 3
B-n52-k7 747 747 0 747.6 0.08 84 747 0 747 0 1
B-n56-k7 707 709 0.28 713.9 0.98 119 707 0 707.7 0.10 10
B-n57-k9 1598 1610 0.75 1615.2 1.08 112 1603 0.31 1605.2 0.45 165
B-n63-k10 1496 1503 0.47 1540.6 2.98 938 1496 0 1512.4 1.10 148
B-n64-k9 861 862 0.116 887.8 3.11 806 861 0 861 0 25
B-n66-k9 1316 1319 0.23 1324.8 0.67 265 1316 0 1317.3 0.10 85
B-n67-k10 1032 1042 0.97 1067.6 3.45 654 1033 0.1 1034.6 0.25 353
B-n68-k9 1272 1278 0.5 1288.1 1.27 701 1273 0.08 1274.8 0.22 61
B-n78-k10 1221 1224 0.3 1248 2.21 703 1221 0 1222.4 0.11 190
a
Denotes a new solution better than BKS with the increasing number of routes by 1.

is less than CVRP-FA. In most instances, CVRP-CHFA reaches BKS 71 out of 80 instances, leading to an average success rate of 0.89.
before starting the communication phase; This means that using In addition, Table 12 summarizes the results in Tables 7–10 with
2-h-opt as a mutation operator is better than using it in local respect to the average B_Gap and A_Gap values of each instance
search phase as in CVRP-FA. and their overall average values. In terms of the average of B_Gap,
By taking into account the best obtained results (Best) for the proposed algorithm can solve the benchmarks A, B, P, and
each benchmark in Tables 7–10, we have summarized these re- E with an average of B_Gap rates of 0.07, 0.02, 0.00, and −0.18,
sults in Table 11 to provide better understanding. This summary respectively. The overall B_Gap average is −0.02. In terms of the
reveals more clearly the efficiency of the CVRP-CHFA algorithm average of A_Gap, CVRP-CHFA is able to solve the benchmarks A,
than CVRP-FA. Moreover, we can infer that CVRP-CHFA performs B, P, and E on an average of A_Gap rates at 0.14, 0.08, 0.10, and
satisfactorily in solving CVRP instances concerning the quality of −0.14, respectively. The overall A_Gap average is 0.04.
the solution. With success rates of 0.78, 1.86, 1,00, 1.00, and 0.89, Over the four benchmarks, the average B_Gap and A_Gap
the CVRP-CHFA addresses benchmarks A, B, P, and E, respectively. values show the ability of the CVRP-CHFA algorithm to obtain
It is noted from the table that the CVRP-CHFA is capable of solving nearly optimal (or optimal) solutions for all 80 instances. The
9
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 9
Computational results of CVRP-CHFA and CVRP-FA for benchmark E.
Instance BKS CVRP-FA CVRP-CHFA
Best B_Gap Avg. A_Gap Iter. Best B_Gap Avg. A_Gap Iter.
E-n22-k4 375 375 0 375 0 1 375 0 375 0 1
E-n23-k3 569 569 0 569 0 1 569 0 569 0 1
E-n30-k3 534 534 0 534.6 0.11 22 534 0 534 0 1
50a −5.24 512.5 −4.03 4 503a −5.81 504.3 −5.56 1
E-n33-k4 835 835 0 845.2 1.22 168 835 0 835 0 1
E-n51-k5 521 521 0 521 0 3 521 0 521 0 1
E-n76-k7 682 683 0.2 688.9 1.01 225 682 0 682.7 0.10 24
E-n76-k8 735 738 0.41 743.3 1.13 564 736 0.14 736.8 0.25 38
E-n101-k14 1067 1082 1.41 1101.9 3.27 151 1071 0.38 1078.2 1.05 262
a
Denotes a new solution better than BKS with an increasing number of routes by 1.

Table 10
Computational results of CVRP-CHFA and CVRP-FA for benchmark P.
Instance BKS CVRP-FA CVRP-CHFA
Best B_Gap Avg. A_Gap Iter. Best B_Gap Avg. A_Gap Iter.
P-n16-k8 450 450 0 450 0 1 450 0 450 0 1
P-n19-k2 212 212 0 212 0 1 212 0 212 0 1
P-n20-k2 216 216 0 216 0 3 216 0 216 0 1
P-n21-k2 211 211 0 211 0 1 211 0 211 0 1
P-n22-k2 216 216 0 216 0 1 216 0 216 0 1
P-n22-k8 603 590a −2.16 590 −2.16 4 590a −2.16 590 −2.16 1
P-n23-k8 529 529 0 529 0 3 529 0 529 0 1
P-n40-k5 458 458 0 459.9 0.42 218 458 0 458 0 1
P-n45-k5 510 510 0 510 0 13 510 0 510 0 1
P-n50-k7 554 554 0 557.3 0.6 131 554 0 554 0 4
P-n50-k8 631 631 0 633.4 0.38 167 631 0 630.2 −0.13 2
629a −0.32 632.9 0.3 308 629a −0.32 629 −0.32 1
P-n50-k10 696 697 0.2 702.9 0.99 71 696 0 696.9 0.13 180
P-n51-k10 741 742 0.2 750.3 1.26 220 741 0 742.1 0.15 12
P-n55-k7 568 568 0 573.5 0.97 21 568 0 568 0 2
P-n55-k8 588 576b −2.04 577 −1.87 47 576b −2.04 576 −2.04 1
P-n55-k10 694 698 0.6 699.4 0.78 111 694 0 697.4 0.49 63
P-n60-k10 744 749 0.7 752.6 1.16 28 744 0 744.6 0.08 100
P-n60-k15 968 968 0 983.2 1.57 352 968 0 968 0 15
P-n65-k10 792 792 0 799.3 0.92 293 792 0 792 0 3
P-n70-k10 827 827 0 827.9 0.11 29 827 0 827 0 2
P-n76-k4 593 593 0 598.8 0.98 200 593 0 593.1 0.02 5
P-n76-k5 627 628 0.2 630.7 0.59 47 627 0 627.3 0.05 61
P-n101-k4 681 681 0 684.7 0.54 7 681 0 682.2 0.18 6
a
Denotes a new solution better than BKS with an increasing number of routes by 1.
b
Denotes a new better solution than BKS with decreasing number of routes by 1.

Table 11
Computational analysis summary of CVRP-CHFA against CVRP-FA in terms of the success rate over the four benchmarks.
Algorithm A B E P Avg.
OA NP SR OA NP SR OA NP SR OA NP SR SR
CVRP-FA 13 27 0.48 10 22 0.45 5 8 0.63 18 23 0.78 0.58
CVRP-CHFA 21 27 0.78 19 22 0.86 8 8 1 23 23 1 0.89

OA is the number of optimal results obtained.


NP is the number of problems.
SR is the success rate (OA/NP).

Table 12
Computational analysis summary of CVRP-CHFA against CVRP-FA in terms of B_Gap and A_Gap over the four benchmarks.
Algorithm A B E P Avg. Avg.
B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap
CVRP-FA 0.29 0.73 0.25 0.96 0.19 0.73 −0.11 0.31 0.16 0.67
CVRP-CHFA 0.07 0.14 0.02 0.08 0 0.1 −0.18 −0.14 −0.02 0.04

negative values indicate that the proposed algorithm can reach The third comparison shows further evidence for the superior-
optimal solutions better than the existing BKS. Besides, the overall ity of CVRP-CHFA over CVRP-FA. This comparison is between the
average of the values B_Gap and A_Gap reveal the consistency diversity of CVRP-CHFA and CVRP-CHFA in solving two selected
of the CVRP-CHFA algorithm compared to CVRP-FA. As a result, CVRP instances. The conclusion can, therefore, be generalized for
this summary highlights the CVRP-CHFA algorithm’s adequate other instances. The diversity values using P-n101-k4 and M-
performance in solving CVRP instances. n200-k17 from an independent run are plotted in Fig. 6. Overall,
10
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Fig. 6. The population diversity monitor of CVRP-CHFA and CVRP-FA with P-n101-k4 and M-n200-k17 in a single run.

Fig. 7. The convergence behavior of CVRP-CHFA and CVRP-FA with P-n101-k4 and M-n200-k17 in a single run.

with the same configurations of evolution operators, the pro-


posed CVRP-CHFA model outperforms CVRP-FA in maintaining
consistency and adequate diversity and striking balance between
explorative and exploitative capabilities. In terms of diversity
shown in Fig. 6a, CVRP-CHFA can maintain diversity at a relatively
high level compared to CVRP-FA. Moreover, further evidence can
be concluded from the line graph of the diversity in Fig. 6b, which
shows that CVRP-FA maintained lower diversity as compared to
CVRP-CHFA. With CVRP-CHFA, diversity is also stable as iterations
proceeds. The population diversity is enhanced when reaching
the communication interval after certain iterations, as can be
observed from Figs. 6a and 6b. The value of population diversity
is kept at an adequate number during the search, which could
help the solutions ‘‘jump out’’ local optima. There are reasons:
(a) well-distributed population’s solutions resulted in consistent
Fig. 8. P-n16-k8 route with distance 422 and violated capacity constraint that
diversity in all populations, (b) the interaction between its sub- is proposed by DIWO.
populations after each communication interval, and (c) CVRP-FA
tends to converge into its only population while CVRP-CHFA can
maintain diversity by forcing its sub-populations to be different.
benchmarks namely Augerat et al. A, B, and P [53], Christofides
In the fourth comparison, to fairly compare the convergence
and Eilon E [54], Fisher F [55] and Christofides M and CMT [56].
behavior of CVRP-CHFA and CVRP-FA algorithms, Fig. 7 plots the
The proposed CVRP-CHFA algorithm is compared against the
convergence behavior achieved by each algorithm during the first
algorithms in Tables 13–18. According to the provided results
100 iterations. It is clear from the figure that CVRP-CHFA con-
verged faster than CVRP-FA with the two instances. CVRP-CHFA in the literature for the compared algorithms, the results of the
quickly moved from the current best solution to the new best CVRP-CHFA algorithm given in the tables are for an indepen-
solution. That means CVRP-CHFA has a high capability to escape dent run and the average of ten independent runs. In these
from local optima than CVRP-FA. Moreover, with the P-n101- tables, CVRP instances and their corresponding BKS are placed
k4 instance, CVRP-CHFA converged to BKS faster than CVRP-FA in columns 1 and 2, respectively. The other columns represent
(i.e., in fewer than 60 iterations). the Best, B_Gap, Avg., A_Gap, respectively. The best results of the
algorithms are highlighted in bold.
5.2.2. Comparison of CVRP-CHFA with other CVRP algorithms Tables 13–18 shows a performance evaluation test that shows
The objective of this section is to demonstrate the effective- the effectiveness of the CVRP-CHFA algorithm as it has reached
ness of the proposed CVRP-CHFA algorithm over five previously BKS or less in 82 out of 102 instances. According to the com-
proposed CVRP metaheuristic algorithms including DIWO [12], parison study, we can firmly state that CVRP-CHFA performs
ISOS [8], HA [27], LNS-ACO [19], and CVRP_GELS [20]. For the effectively in addressing the CVRP instances compared with the
comparison, we use 102 instances selected from seven popular other CVRP metaheuristic algorithms. In detail, for benchmark
11
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 13
Computational results of CVRP-CHFA and other CVRP algorithms for benchmark A.
Instance BKS DIWO ISOS HA LNS-ACO CVRP-CHFA
Best B_Gap Best Avg. A_Gap Best B_Gap B_Gap Best B_Gap Avg. A_Gap
A-n32-k5 784 784 0 – – – 784 0 0 786 0.26 792.2 1.05
A-n33-k5 661 661 0 661 663.65 0.40 661 0 0 661 0 661 0
A-n33-k6 742 – – – – – 742 0 0 742 0 742 0
A-n34-k5 778 778 0 – – – 778 0 0 778 0 778 0
A-n36-k5 799 – – – – – 799 0 0 799 0 799 0
A-n37-k5 669 – – – – – 669 0 0 669 0 669 0
A-n37-k6 949 – – – – – 949 0 0 949 0 949 0
A-n38-k5 730 – – – – – 730 0 0 730 0 730 0
A-n39-k5 822 – – – – – 822 0 0 822 0 822 0
A-n39-k6 831 – – – – – 831 0 0 831 0 831 0
A-n44-k6 937 – – – – – 939 0.21 0 937 0 937 0
A-n45-k6 944 – – – – – 955 1.17 1.48 944 0 944 0
A-n45-k7 1146 – – – – – 1153 0.61 0 1146 0 1147.3 0.11
A-n46-k7 914 914 0 914 939.33 2.77 915 0.11 0 914 0 914 0
A-n48-k7 1073 – – – – – 1073 0 1.03 1073 0 1073 0
A-n53-k7 1010 – – – – – 1026 1.58 0 1010 0 1010.5 0.05
A-n54-k7 1167 – – – – – 1169 0.17 0 1167 0 1167 0
A-n55-k9 1073 – – – – – 1074 0.09 0 1073 0 1073.3 0.03
A-n60-k9 1354 1354 0 1354 1392.71 2.86 1366 0.89 0 1354 0 1355.6 0.12
A-n61-k9 1034 – – – – – 1045 1.06 3.19 1035 0.1 1036.7 0.26
A-n62-k8 1288 – – – – – 1302 1.09 1.55 1294 0.47 1297.6 0.75
A-n63-k10 1314 – – – – – 1325 0.84 1.14 1315 0.08 1315 0.08
A-n63-k9 1616 – – – – – 1644 1.73 2.04 1616 0 1616 0
A-n64-k9 1401 – – – – – 1442 2.93 1 1411 0.71 1411.4 0.74
A-n65-k9 1174 – – – – – 1189 1.28 0.94 1177 0.26 1177 0.26
A-n69-k9 1159 – – – – – 1169 0.86 0.95 1159 0 1160.2 0.10
A-n80-k10 1763 1764 0.06 – – – 1790 1.53 2.95 1763 0 1768.3 0.30

Table 14
Computational results of CVRP-CHFA and other CVRP algorithms for benchmark B.
Instance BKS DIWO ISOS LNS-ACO CVRP-CHFA
Best B_Gap Best Avg. A_Gap PD_B Best B_Gap Avg. A_Gap
B-n31-k5 672 672 0 – – – 0 672 0 672 0
B-n34-k5 788 788 0 – – – 0 788 0 788 0
B-n35-k5 955 – – 955 955.22 0.02 0 955 0 955 0
B-n38-k6 805 – – – – – 0 805 0 805 0
B-n39-k5 549 – – – – – 0 549 0 549 0
B-n41-k6 829 829 0 – – – 0 829 0 829 0
B-n43-k6 742 – – – – – 0 742 0 742 0
B-n44-k7 909 – – – – – 0 909 0 909 0
B-n45-k5 751 751 0 751 754.76 0.50 0 751 0 751 0
B-n45-k6 678 – – – – – 0 678 0 678 0
B-n50-k7 741 741 0 – – – 0 741 0 741 0
B-n50-k8 1312 – – – – – 0 .53 1312 0 1314.2 0.17
B-n51-k7 1032 – – – – – – 1032 0 1024.3 −0.75
– – – – – −1.6 1016a −1.6 1016 −1.55
B-n52-k7 747 – – – – – 0 747 0 747 0
B-n56-k7 707 – – – – – 0 707 0 707.7 0.10
B-n57-k9 1598 – – – – – 0 1603 0.31 1605.2 0.45
B-n63-k10 1496 1497 0.07 – – – 1.2 1496 0 1512.4 1.10
B-n64-k9 861 – – – – – 1.51 861 0 861 0
B-n66-k9 1316 – – – – – 1.06 1316 0 1317.3 0.10
B-n67-k10 1032 1035 0.29 – – – 1.74 1033 0.1 1034.6 0.25
B-n68-k9 1272 – – 1274 1282.72 0.84 1.42 1273 0.08 1274.8 0.22
B-n78-k10 1221 1223 0.16 1221 1256.89 2.94 0.57 1221 0 1222.4 0.11
a
Denotes a new solution better than BKS with an increasing number of routes by 1.

A, the CVRP-CHFA algorithm can produce better results in com- For benchmark P, the CVRP-CHFA algorithm is able to produce
parison to ISOS, HA, and LNS-ACO algorithms according to their the BKS for 21 out of 25. Furthermore, it can find new solutions
reported results in the literature. CVRP-CHFA algorithm has less better than the BKS for the remaining four instances, including
average of B_Gap values for all 27 instances, i.e., 0.07. The slight P-n22-k8, P-n50-k8, P-n55-k8, and P-n55-k15. As a result, the
improvement provides by DIWO over CVRP-CHFA is referring to average of B_Gap values is less than 0, i.e., −0.36. It is essential to
the use of only six instances out of 27.
highlight that the results of DIWO is unreliable and not valid for
For benchmark B, CVRP-CHFA outperforms DIWO, ISOS, HA,
the P-n16-k8 instance because there is a violation of the capacity
and LNS-ACO with less average of B_Gap values at 0.02%. Also,
it can obtain an optimal solution for B-n51-k7 better than BKS. constraint in route 1,4,9,1, as shown in Fig. 8. Moreover, from
Similarly, the CVRP-CHFA algorithm has less average of B_Gap Fig. 9, the two solutions of P-n22-k8 that are found by CVRP-CHFA
values (0.09) comparing to ISOS and LNS-ACO in the benchmark and DIWO has the same routes and the same distance; therefore,
E. Here, the optimal solution obtained for E-n30-k3 is better than recomputing the distance of P-n22-k8 which reported by DIWO
BKS. finds that the total distance is 590, not 589 as listed in [12].
12
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 15
Computational results of CVRP-CHFA and other CVRP algorithms for benchmark E.
Instance BKS ISOS LNS-ACO CVRP-CHFA
Best Avg. A_Gap B_Gap Best B_Gap Avg. A_Gap
E-n22-k4 375 – – – 0 375 0 375 0
E-n23-k3 569 – – – 0 569 0 569 0
E-n30-k3 534 534 551.93 0.18 – 534 0 534 0
– – – −5.81 503a −5.81 504.3 −5.56
E-n33-k4 835 – – – 0 835 0 835 0
E-n51-k5 521 521 546.13 0.25 – 521 0 521 0
E-n76-k7 682 682 708.22 0.26 1.91 682 0 682.7 0.10
E-n76-k8 735 – – – 1.22 736 0.14 736.8 0.24
E-n76-k14 1021 – – – 0.88 1021 0 1025.5 0.44
E-n101-k14 1067 – – – 1.41 1071 0.37 1078.2 1.05
a
Denotes a new solution better than BKS with the increasing number of routes by 1.

Table 16
Computational results of CVRP-CHFA and other CVRP algorithms for benchmark P.
Instance BKS DIWO ISOS LNS-ACO CVRP-CHFA
Best B_Gap Best Avg. A_Gap PD_B Best B_Gap Avg. A_Gap
P-n16-k8 450 422 – – – – 0 450 0 450 0
P-n19-k2 212 – – – – – 0 212 0 212 0
P-n20-k2 216 216 0 – – – 0 216 0 216 0
P-n21-k2 211 – – – – – 0 211 0 211 0
P-n22-k2 216 216 0 – – – 0 216 0 216 0
P-n22-k8 603 589 – – – – −2.16 590a −2.16 590 −2.16
P-n23-k8 529 – – – – – 0 529 0 529 0
P-n40-k5 458 – – – – – 0 458 0 458 0
P-n45-k5 510 – – – – – 0 510 0 510 0
P-n50-k7 554 – – – – – 0 554 0 554 0
P-n50-k8 631 – – – – – 1.9 631 0 630.2 −0.13
– – – – – – 629a −0.32 629 −0.32
P-n50-k10 696 696 0 – – – 0 696 0 696.9 0.13
P-n51-k10 741 – – – – – 0.81 741 0 742.1 0.15
P-n55-k7 568 – – – – – 0 568 0 568 0
P-n55-k8 588 – – – – – 0 576b −2.04 576 −2.04
P-n55-k10 694 – – – – – 0 694 0 697.4 0.49
P-n55-k15 989 955a −3.44 – – – 0 945a −4.45 945 −4.45
P-n60-k10 744 – – – – – 1.48 744 0 744.6 0.08
P-n60-k15 968 – – – – – 0.93 968 0 968 0
P-n65-k10 792 – – – – – 1.01 792 0 792 0
P-n70-k10 827 – – – – – 1.21 827 0 827 0
P-n76-k4 593 595 0.34 593 633.8 6.88 0.84 593 0 593.1 0.02
P-n76-k5 627 – – – – – 2.87 627 0 627.3 0.05
P-n101-k4 681 681 0 681 696.43 2.27 – 681 0 682.2 0.18
a
Denotes a new solution better than BKS with an increasing number of routes by 1.
b
Denotes a new better solution than BKS with decreasing number of routes by 1.

Fig. 9. The obtained P-n22-k8 route by (a) CVRP-CHFA and (b) DIWO.

The CVRP-CHFA algorithm has less average of B_Gap values The optimal solution obtained for M-n200-k17 is better than BKS
(−0.74) comparing to ISOS and HA in the benchmarks F and M. with the same number of routes. Finally, for benchmark CMT,
13
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 17
Computational results of CVRP-CHFA and other CVRP algorithms for benchmarks F and M.
Instance BKS ISOS HA CVRP-CHFA
Best Avg. A_Gap Best B_Gap Best B_Gap Avg. A_Gap
F-n72-k4 237 237 241.1 1.73 – – 237 0 237.1 0.04
F-n135-k7 1162 1162 1227.76 5.66 – – 1163 0.09 1164.4 0.21
M-n101-k10 820 820 835.74 1.92 – – 829 0.98 830.6 1.29
M-n121-k7 1034 1037 1045.97 1.16 – – 1034 0 1034.1 0.01
M-n151-k12 1015 – – – 1048 3.25 1021 0.59 1025.8 1.06
M-n200-k17 1373 – – – 1331a −3.06 1289a −6.12 1294.4 −5.72
a
Denotes a new solution better than BKS with an equal number of routes.

Table 18
Computational results of CVRP-CHFA and other CVRP algorithms for benchmark CMT.
Instance BKS LNS-ACO ISOS CVRP_GELS CVRP-CHFA
Best Best Avg A_Gap Best Avg A_Gap Best Avg A_Gap
CMT1 524.611 524.611 524.611 530.93 1.20 524.611 528 0.65 524.611 524.611 0
CMT2 835.26 835.26 835.74 857.68 2.68 835.26 838 0.33 835.26 835.853 0.07
CMT3 826.14 826.14 826.14 845.9 2.39 826.14 830 0.47 827.844 828.422 0.28
CMT4 1028.42 1046.9 1028.42 1090.86 6.07 1028.42 1035 0.64 1028.42 1033.33 0.48
CMT5 1291.29 1341.4 1305.49 1370.64 6.15 1294.21 1302 0.83 1300.306 1312.47 1.64
CMT6 555.43 555.43 555.43 563.26 1.41 555.43 559 0.64 555.43 555.43 0
CMT7 909.68 909.68 914.68 928.93 2.12 914.13 920 1.13 909.68 913.917 0.47
CMT8 865.94 865.94 865.94 878.8 1.49 869.34 876 1.16 865.94 867.107 0.13
CMT9 1162.55 1164.93 1162.55 1248.36 7.38 1162.55 1174 0.98 1163.185 1172.061 0.82
CMT10 1395.85 1419.7 1417.9 1523.94 9.18 1395.85 1405 0.66 1414.328 1420.732 1.78
CMT11 1042.11 1042.11 1042.11 1049.95 0.75 1042.11 1058 1.52 1042.11 1042.327 0.02
CMT12 819.56 819.56 819.56 834.5 1.82 819.56 829 1.15 830.162 837.236 2.16
CMT13 1541.14 1547.1 1541.14 1574.67 2.18 1541.14 1564 1.48 1542.863 1545.216 0.26
CMT14 866.365 866.37 866.53 868.61 0.26 866.365 873 0.77 866.365 866.55 0.02

Table 19
Computational analysis summary of CVRP-CHFA against other CVRP algorithms in terms of the success rate over the seven benchmarks.
Algorithm A B E P F M CMT Avg.
OA NP SR OA NP SR OA NP SR OA NP SR OA NP SR OA NP SR OA NP SR SR
DIWO 5 6 0.83 5 8 0.63 – – – 6 7 0.86 – – – – – – – – – 0.76
ISOS 3 3 1 3 4 0.75 3 3 1 2 2 1 2 2 1 1 2 0.5 10 14 0.71 0.85
HS 11 27 0.41 – – – – – – – – – – – – 1 2 0.5 – – – 0.45
LNS-ACO 17 27 0.63 14 22 0.64 3 7 0.43 15 23 0.65 – – – – – – 8 14 0.57 0.61
CVRP_GELS – – – – – – – – – – – – – – – – – – 11 14 0.79 0.79
CVRP-CHFA 21 27 0.78 19 22 0.86 7 9 0.78 24 24 1 1 2 0.5 2 4 0.5 8 14 0.57 0.80

OA is the number of optimal results obtained.


NP is the number of problems.
SR is the success rate (OA/NP).

Table 20
Computational analysis summary of CVRP-CHFA against other CVRP algorithms in terms of B_Gap and A_Gap over the seven benchmarks.
Algorithm A B E P F M CMT Avg. Avg.
B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap B_Gap A_Gap
DIWO 0.01 NA 0.07 NA – – −0.77 NA – – – – – – −0.23 NA
ISOS NA 2.01 NA 1.08 NA 0.26 NA 4.575 NA 3.7 NA 1.54 NA 3.22 NA 2.34
HS 0.6 NA – – – – – – – – 0.1 NA – – 0.35 NA
LNS-ACO 0.6 NA 0.29 NA 1.084 NA 0.387 NA – – – – 0.57 NA 0.59 NA
CVRP_GELS – – – – – – – – – – – – 0.08 0.89 0.08 0.89
CVRP-CHFA 0.07 0.14 −0.05 0.01 0.09 0.31 −0.36 −0.32 0.05 0.13 −1.14 −0.84 0.26 0.58 −0.15 0.001

the CVRP-CHFA algorithm can achieve better results compared to rate of 0.78, 0.86, 0.78, 1,00, 0.50, 0.50, and 0.57 for benchmarks
LNS-ACO and ISOS in terms of the average of B_Gap and average A, B, P, E, F, M and CMT respectively. CVRP-CHFA is typically
of A_Gap. Despite the slight improvement of CVRP_GELS over capable of offering the optimal solution for 82 of 102 instances
CVRP-CHFA in terms of the average of B_Gap, CVRP-CHFA has less (21 of 27 A, 19 of 22 B, 7 of 9 E, 24 of 24 E, 1 of 2 F and 2 of 4 of
average of A_Gap values with 0.58. M and 8 of 14 CMT). The average success rate here is 0.80.
To better understanding, the results in Tables 13–18 is sum- Additionally, by taking into account overall average of both
marized in Table 19. In this regard, we consider the best results B_Gap and A_Gap, the results in Tables 13–18 are summarized
for every benchmark achieved by the compared algorithms. From in Table 20. In terms of B_Gap, the CVRP-CHFA algorithm can
our standpoint, this summary more specifically highlights the solve the instances of CVRP with an overall B_Gap average of
effectiveness of the proposed CVRP-CHFA algorithm relative to −0.15. In details, the averages of B_Gap are 0.07, −0.05, 0.09,
other CVRP algorithms. −0.36, 0.05, −1.14, and 0.26 achieved for instances A, B, P, E,
Based on Table 19, it can be noted that CVRP-CHFA is working F, M and CMT respectively. In terms of A_Gap, CVRP-CHFA is
efficiently for CVRP benchmarks. CVRP-CHFA achieve a success able to solve the benchmarks A, B, P, E, F, M, and CMT with an
14
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 21
Computational results of CVRP-CHFA and parallel GVNS for benchmark A.
Instance BKS GVNS CVRP-CHFA
Best B_Gap Best B_Gap Avg.
A-n32-k5 787.0819 787.0819 0 797.451 1.32 797.451
A-n33-k5 662.1101 662.1101 0 662.1101 0 662.1101
A-n33-k6 742.6932 742.6932 0 742.6932 0 742.6932
A-n34-k5 780.9361 780.9361 0 780.9361 0 780.9361
A-n36-k5 802.1318 802.1318 0 802.1318 0 802.1318
A-n37-k5 672.4652 672.4652 0 672.4652 0 672.4652
A-n37-k6 950.8522 950.8522 0 950.8522 0 950.8522
A-n38-k5 734.1846 734.1846 0 733.945a −0.03 733.945
A-n39-k5 828.9891 828.9891 0 828.9891 0 828.9891
A-n39-k6 833.2046 833.2046 0 833.2046 0 833.4092
A-n44-k6 939.3346 939.3346 0 938.181a −0.12 938.181
A-n45-k6 944.8763 944.8763 0 944.8763 0 945.3445
A-n45-k7 1146.909 1146.909 0 1146.771a −0.01 1149.607
A-n46-k7 917.9073 917.9073 0 917.724a −0.02 917.724
A-n48-k7 1074.338 1074.338 0 1074.338 0 1019.818
A-n53-k7 1012.326 1012.326 0 1012.249a −0.01 1016.184
A-n54-k7 1171.784 1183.575 1.006 1171.682a −0.01 1172.908
A-n55-k9 1074.464 1074.464 0 1074.464 0 1074.81
A-n60-k9 1355.799 1355.799 0 1355.799 0 1357.716
A-n61-k9 1039.078 1040.309 0.12 1040.309 0.12 1042.975
A-n62-k8 1294.282 1313.048 1.45 1302.346 0.62 1302.349
A-n63-k9 1622.145 1633.936 0.73 1622.145 0 1625.864
A-n63-k10 1313.729 1316.852 0.24 1316.852 0.24 1317.982
A-n64-k9 1400.832 1403.278 0.17 1410.829 0.71 1413.648
A-n65-k9 1181.687 1181.687 0 1184.661 0.25 1184.718
A-n69-k9 1165.995 1170.538 0.39 1165.995 0 1169.218
A-n80-k10 1766.5 1790.451 1.36 1768.737 0.13 1777.565
Average 0.20 0.12
a
Denotes a new solution better than BKS with an equal number of routes.

Table 22
Computational results of CVRP-CHFA and parallel GVNS for benchmark B.
Instance BKS GVNS CVRP-CHFA
Best B_Gap Best B_Gap Avg.
B-n31-k5 676.0884 676.0884 0 676.0884 0 676.0884
B-n34-k5 790.1838 790.1838 0 789.84a −0.04 790.0834
B-n35-k5 956.2941 956.2941 0 956.2941 0 956.2941
B-n38-k6 807.8788 807.8788 0 807.8788 0 808.29
B-n39-k5 553.1564 553.1564 0 553.1564 0 553.1564
B-n41-k6 833.8063 833.8063 0 833.663a −0.02 833.7059
B-n43-k6 746.9838 746.9838 0 746.694a −0.04 746.8963
B-n44-k7 914.9648 914.9648 0 914.9648 0 914.9648
B-n45-k5 754.4388 754.4388 0 753.96a −0.06 754.0669
B-n45-k6 680.4379 680.4379 0 680.4379 0 680.6794
B-n50-k7 744.228 744.228 0 744.228 0 744.228
B-n50-k8 1321.524 1321.524 0 1315.473a −0.46 1319.614
B-n51-k7 1019.325 1019.325 0 1019.325 0 1020.156
B-n52-k7 749.9697 749.9697 0 749.9697 0 749.9904
B-n56-k7 712.9161 712.9161 0 712.9161 0 716.6029
B-n57-k7 1143.946 1143.946 0 1150.278 0.55 1155.215
B-n57-k9 1603.371 1603.371 0 1613.11 0.61 1614.265
B-n63-k10 1501.271 1506.883 0.37 1499.095a −0.14 1507.463
B-n64-k9 869.3157 869.3157 0 868.193a −0.13 868.2048
B-n66-k9 1325.355 1327.441 0.16 1326.496 0.09 1326.844
B-n67-k10 1039.359 1039.359 0 1040.523 0.11 1054.149
B-n68-k9 1278.211 1283.816 0.44 1277.688a −0.04 1280.809
B-n78-k10 1229.273 1232.601 0.27 1227.898a −0.11 1229.409
Average 0.05 0.01
a
Denotes a new solution better than BKS with an equal number of routes.

average of A_Gap values of 0.14, 0.01, 0.31, −0.32, 0.13, −0.84, B_Gap and A_Gap show that the CVRP-CHFA algorithm is con-
and 0.58, respectively with an overall A_Gap average of 0.001. The sistent compared with the other CVRP algorithms. The summary
overall average of B_Gap and A_Gap values show the capability thus stresses that the CVRP-CHFA has performed satisfactorily in
of the CVRP-CHFA to obtaining optimal or near-optimal solutions addressing CVRP.
over the seven benchmarks in all 102 instances. Moreover, the
negative values indicate that the ability of the proposed algorithm 5.2.3. Comparison of CVRP-CHFA with parallel GVNS algorithm
to find a new optimum solution better than the current BKS In this section, the performance of the CVRP-CHFA algorithm is
for seven instances. Besides, the overall average values of the validated with a parallel GVNS algorithm using 65 instances from
15
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Table 23
Computational results of CVRP-CHFA and parallel GVNS for benchmarks E and M.
Instance BKS GVNS CVRP-CHFA
Best B_Gap Best B_Gap Avg.
E-n13-k4 247 247 0 247 0 247
E-n22-k4 375.2798 375.2798 0 375.2798 0 375.2798
E-n23-k3 568.5625 568.5625 0 568.5625 0 568.5625
E-n30-k3 538.7947 538.7947 0 505.011a −6.27 506.5078
E-n31-k7 379 379 0 379 0 379
E-n33-k4 837.6716 837.6716 0 837.6716 0 837.6716
E-n51-k5 524.9442 524.9442 0 524.611a −0.06 524.7621
E-n76-k7 666.8325 697.7697 4.64 687.602 3.11 689.9682
E-n76-k8 740.6554 740.6554 0 740.6554 0 743.3249
E-n76-k14 1026.706 1026.706 0 1024.605b −0.205 1031.117
E-76-10 837.3556 838.9785 0.19 835.321a −0.24 837.9059
E-101-14 1082.65 1094.638 1.11 1087.406 0.44 1094.324
E-101-8 826.908 832.7041 0.701 829.13 0.27 831.6122
M-n101-k10 819.5575 819.5575 0 840.591 2.57 846.7091
M-n200-k16 1274 1331.238 4.49 1319.225 3.55 1326.323
Average 0.74 0.21
a
Denotes a new solution better than BKS with an equal number of routes.
b
Denotes a new solution better than BKS with an increasing number of routes by 1.

Table 24
Computational results of CVRP-CHFA on large-scale CVRP problems.
Instance BKS ILS-RVND-SP PMNTS LsxVNS HAFA HA CVRP-CHFA
Best Avg. A_Gap Best Avg. A_Gap Best B_Gap Best B_Gap Best B_Gap Best B_Gap Avg. A_Gap
G-n262-k25 6119 – – – – – – – – – – 5875 −3.99 5644a −7.76 5704.8 −6.77
G9(225) 579.71 583.24 585.21 0.95 581.73 583.15 0.59 586.79 1.22 593.71 2.42 – – 585.5 1 588.54 1.52
G10(323) 736.26 741.96 744.17 1.07 738.5 739.97 0.50 749.92 1.86 758.66 3.04 – – 749.7 1.83 753.55 2.35
G13(252) 857.19 857.19 860 0.33 861.92 864.45 0.85 861.27 0.48 871.19 1.63 – – 871.7 1.69 875.83 2.17
G17(240) 707.76 707.76 707.81 0.01 707.83 708.46 0.1 708.53 0.11 718.76 1.55 – – 705.8a −0.28 706.96 −0.11
G18(300) 995.13 995.65 997.85 0.27 1000.27 1002.53 0.74 1012.79 1.77 1017.13 2.21 – – 995a −0.01 1002.5 0.74
Average 0.53 0.56 1.09 2.17 −3.99 −0.59 −0.02
a
Denotes a new solution better than BKS with an equal number of routes.

the benchmarks A and B [53], benchmark E [54] and a subset of 5.2.4. Comparison of CVRP-CHFA on large-scale CVRP problems
benchmark M(M-n101-k10, M-n200-k16) [56]. The performance of the CVRP-CHFA algorithm is validated on
The CVRP-CHFA algorithm is compared against parallel GVNS large-scale CVRP problems in this section.
algorithm in Tables 21–23. In these tables, CVRP instances and The CVRP-CHFA algorithm is compared against ILS-RVND-
their corresponding BKS are placed in columns 1 and 2, respec- SP [58], PMNTS [59], LsxVNS [25], HAFA [32] and HA [27] al-
tively. The other columns represent the Best, B_Gap, Avg., A_Gap, gorithms in Table 24. In this table, CVRP instances and their
corresponding BKS are placed in columns 1 and 2, respectively.
respectively. The best-obtained results are highlighted in bold.
The other columns represent the Best, B_Gap, Avg., A_Gap, re-
The results show that the CVRP-CHFA algorithm can pro-
spectively. The best-obtained results are highlighted in bold.
duce significantly better results than the parallel GVNS algorithm,
The results show that the CVRP-CHFA algorithm can produce
according to its reported results in the literature. CVRP-CHFA significantly better results than HA, LsxVNS and HAFA algorithms,
algorithm has less average of B_Gap values for the average of according to their reported results in the literature for G-n262-
best results for all benchmarks. CVRP-CHFA algorithm can find k25, G9, G10, G17 and G18. Moreover, CVRP-CHFA algorithm can
the best solution better than the recorded BKS in the literature achieve better results compared to ILS-RVND-SP and PMNTS in
for 19 instances from benchmarks A, B, and E. terms of best solution and B_Gap for G17 and G18. Despite the
Moreover, for benchmark A, the CVRP-CHFA algorithm can slight improvement of ILS-RVND-SP and PMNTS over CVRP-CHFA
produce better results in comparison to the parallel GVNS algo- in terms of the best and average solution for G9, G10 and G13.
rithm, according to the reported results in the literature. CVRP- Furthermore, CVRP-CHFA algorithm can find the best solution
CHFA algorithm has less average of B_Gap values for all 27 in- better than the recorded BKS in the literature for G-n262-k25,
stances, i.e., 0.12. Furthermore, it can obtain an optimal solu- G17 and G18.
tion for A-n38-k5, A-n44-k6, A-n45-k7, A-n46-k7, A-n53-k7, and
A-n54-k7 better than BKS. 5.3. CVRP-CHFA convergence behavior
For benchmark B, CVRP-CHFA outperforms parallel GVNS with
The convergence behavior of CVRP-CHFA in solving two se-
less average of B_Gap values at 0.01. In addition, it can obtain
lected CVRP instances is described in this section. The conclu-
an optimal solution for B-n34-k5, B-n41-k6, B-n43-k6, B-n45-k5,
sion can, therefore, be generalized for other instances. The con-
B-n50-k8, B-n63-k10, B-n64-k9, B-n68-k9, and B-n78-k10 better vergence behavior of CVRP-CHFA with P-n101-k4 and M-n200-
than BKS. Similarly, the CVRP-CHFA algorithm has less average of k17 from an independent run are shown in Fig. 10. This figure
B_Gap values (0.21) comparing to GVNS in the benchmark E, and plots the best-obtained solution and the average solution overall
a subset of benchmark M. Here, the optimal solutions obtained solutions in each iteration during the independent run.
for E-n30-k3, E-n51-k5, E-n76-k14, and E-76-10 are better than It can be noted that CVRP-CHFA quickly converged. Further-
BKS. more, CVRP-CHFA has a high capability to escape from local
16
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Fig. 10. The convergence behavior of CVRP-CHFA with P-n101-k4 and M-n200-k17 in a single run.

optima, regained its diversity, and converged to global optima, Table 25


as shown with the P-n101-k4 instance. Moreover, Fig. 10 shows Mann–Whitney U test statistics for CVRP-CHFA, CVRP-FA, LNS-ACO, HA, ISOS,
CVRP_GELS and parallel GVNS.
how quickly the CVRP-CHFA founds BKS after about 320 and 500
Algorithm 1 Algorithm 2 Mann–Whitney Z Asymp. Sig.
iterations for P-n101-k4 and M-n200-k17, respectively. Notice U (2-tailed)
that CVRP-CHFA has a high speed of convergence with different
CVRP-FA 1293.000 −6.122 0.000
solutions during the initial stage. After that, diversity decreases LNS-ACO 3595.000 −2.936 0.003
gradually with the progress of the search process. CVRP-CHFA CVRP-CHFA HA 220.000 −3.452 0.001
can regain some diversity when diversity is lost immediately, as ISOS 118.000 −4.914 0.000
shown at iterations 67 in Fig. 10a. CVRP-CHFA strikes a good CVRP_GELS 53.000 −2.068 0.039
parallel 1259.000 −0.334 0.738
balance between diversification and intensification throughout
GVNS
the search process. This balance allows CVRP-CHFA to efficiently
exploring search space and converges on the optimal solution to
the promising areas. Furthermore, during the search process, no
stagnation occurs. to the introduction of genetic operators that help in retaining
a balance between intensification and diversification. Moreover,
5.4. Statistical analysis it can be attributed to the utilization of the parallelized multi-
population, which fosters diversity in case the solutions are
The statistical analysis tests are conducted for rigorous and fair trapped into the local optima and speeds up the convergence.
comparisons. Besides, the tests determine if there are statistically Owing to the hardware and software variations employ to exe-
significant differences between B_Gaps or A_Gaps among the vari- cute the compared algorithms, we do not evaluate the running
ous algorithms. Based on the Shapiro–Wilk test, the distribution is time of the related algorithms.
not normally distributed for the tested variables, i.e., algorithms. According to the Mann–Whitney U test, CVRP-CHFA signifi-
Therefore, a non-parametric test, such as the Mann–Whitney cantly outperforms CVRP-FA, LNS-ACO, HA, ISOS, and CVRP_GELS.
U test, should be used. DIWO is excluded from the statistical Furthermore, it is worth pointing out that CVRP-CHFA’s compu-
analysis test because it used a few instances. Furthermore, it has tational capability can obtain new solutions better than BKS for
errors in its results. seven instances. As shown in Fig. 11, the cost of the new solutions
The mean rank of each pair of groups is the same under the are 1016, 503, 590, 629, 576, 945 and 1289 for B-n51-k7, E-n30-
null hypothesis. With the results depicted in Table 25, the p- k3, P-n22-k8, P-n50-k8, P-n55-k8, P-n55-k15 and M-n200-k17
value p is 0.000, 0.000, 0.003, 0.001, 0.000, and 0.039 and is less respectively. The current BKS of these instances are 1032, 534,
than 0.05 for groups CVRP-CHFA with hybrid firefly, CVRP-FA, 603, 631, 588, 989, and 1373.
LNS-ACO, HA, ISOS and CVRP_GELS respectively. Based on these
results, the Mann–Whitney U test’s null hypothesis is rejected. It 6. Conclusions
can, therefore, be inferred that the results from CVRP-CHFA vary
significantly from those of hybrid fireflies, LNS-ACO, HA, ISOS, In this paper, a new, improved version of the FA was proposed
and CVRP_GELS. Although there are no significant differences to solve CVRP, called CVRP-CHFA. To speed up the convergence,
between CVRP-CHFA and GVNS, CVRP-CHFA has better results FA was hybridized with an improved 2-opt local search. PMX
in terms of the average B_Gap for the three used benchmarks, crossover and three mutations were integrated to enhance the
as shown in Tables 21–23. Moreover, GVNS reports only the local exploitation capability in the promising area, preserve solu-
best-obtained results and delegates the average results. tions diversity, and avoid converging solutions into local minima.
Moreover, FA was implemented as multiple populations that
5.5. Discussion exchange solutions occasionally. By evaluating several solutions
simultaneously, this version preserved diversity and accelerated
Generally, the overall results show that CVRP-CHFA gets BKS the convergence.
for the majority of the small and medium instances in ten runs. The effectiveness of the proposed algorithm was tested on 102
In terms of the best-found solution, for the all tested instances, instances from seven standard benchmarks, and the results were
CVRP-CHFA is considerably superior to CVRP-FA, HA, DIWO, LNS- compared with recent optimization algorithms that have been
ACO, and parallel GVNS and comparable to ISOS and CVRP_GELS. used to solve CVRP. The results demonstrated the significance of
In terms of the average solution, CVRP-CHFA outperforms most CVRP-CHFA as follows: (a) it outperformed our previously pub-
of the compared algorithms. It is comparable to other algorithms lished work (CVRP-FA), which already outperformed all other FA
in large-scale instances. CVRP-CHFA significantly outperforms algorithms for solving CVRP problem in routing cost, the number
CVRP-FA, and ISOS in all tested instances. While it is comparable of iterations, the diversity, and convergence rate; (b) CVRP-CHFA
to CVRP_GELS, the superiority of CVRP-CHFA can be attributed outperformed other metaheuristic algorithms including CVRP-FA,
17
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

Fig. 11. The new solutions obtained by CVRP-CHFA.

and ISOS significantly and was comparable to CVRP GELS; and (c) can be applied to address other VRPs problems such as VRP
CVRP-CHFA’s computational capability can obtain new solutions pickup and delivery, VRP with time windows, multiple depot
better than BKS for 10 instances inducing B-n51-k7, E-n30-k3,
P-n22-k8, P-n50-k8, P-n55-k8, P-n55-k15, M-n200-k17, G-n262- VRP, open VRP, rich VRP, etc. and various discrete optimization
k25, G17 and G18. In future research, the CVRP-CHFA algorithm problems.
18
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

CRediT authorship contribution statement [22] J. Jianyong, G.C. Teodor, L. Arne, A cooperative parallel metaheuristic for
the capacitated vehicle routing problem, Comput. Oper. Res. 44 (2014)
33–41.
Asma M. Altabeeb: Conceptualization, Methodology, Software,
[23] L. Shih-Wei, L. Zne-Jung, Y. Kuo-Ching, L. Chou-Yuan, Applying hybrid
Validation, Formal analysis, Investigation, Data curation, Visual- meta-heuristics for capacitated vehicle routing problem, Expert Syst. Appl.
ization, Writing - original draft. Abdulqader M. Mohsen: Concep- 36 (2, Part 1) (2009) 1505–1512.
tualization, Methodology, Supervision, Validation, Formal analy- [24] S. Irnich, P. Toth, D. Vigo, Chapter 1: The family of vehicle routing prob-
sis, Investigation, Data curation, Visualization, Writing - original lems, in: Vehicle Routing: Problems, Methods, and Applications, second
ed., SIAM, 2014, pp. 1–33.
draft, Writing - review & editing. Laith Abualigah: Validation,
[25] C.B. Mouaouia, A.t.H. Hacene, S. Said, An integration of Lagrangian split
Investigation, Writing - review & editing. Abdullatif Ghallab: and VNS: The case of the capacitated vehicle routing problem, Comput.
Validation, Formal analysis, Investigation, Writing - review & Oper. Res. 78 (2017) 513–525.
editing. [26] M. Amous, S. Toumi, B. Jarboui, M. Eddaly, A variable neighborhood search
algorithm for the capacitated vehicle routing problem, Electron. Notes
Discrete Math. 58 (2017) 231–238.
Declaration of competing interest [27] S. Kir, H.R. Yazgan, E. Tüncel, A novel heuristic algorithm for capacitated
vehicle routing problem, J. Ind. Eng. Int. 13 (3) (2017) 323.
The authors declare that they have no known competing finan- [28] T.-L. Dam, K. Li, P. Fournier-Viger, Chemical reaction optimization with
cial interests or personal relationships that could have appeared unified tabu search for the vehicle routing problem, Soft Comput. (2016)
1–13.
to influence the work reported in this paper.
[29] Y. Niu, S. Wang, J. He, J. Xiao, A novel membrane algorithm for capacitated
vehicle routing problem, Soft Comput. 19 (2) (2015) 471–482.
References [30] G. Jaradat, M. Ayob, I. Almarashdeh, The effect of elite pool in hybrid
population-based meta-heuristics for solving combinatorial optimization
[1] G.B. Dantzig, J.H. Ramser, The truck dispatching problem, Manage. Sci. 6 problems, Appl. Soft Comput. 44 (2016) 45–56.
(1) (1959) 80–91. [31] X. Wang, T.-M. Choi, H. Liu, X. Yue, Novel ant colony optimization methods
[2] S.N. Kumar, R. Panneerselvam, A Survey on the Vehicle Routing Problem for simplifying solution construction in vehicle routing problems, IEEE
and its Variants, Scientific Research Publishing, 2012. Trans. Intell. Transp. Syst. 17 (11) (2016) 3132–3141.
[3] Y. Tao, F. Wang, An effective tabu search approach with improved loading [32] R. Goel, R. Maini, A hybrid of ant colony and firefly algorithms (HAFA) for
algorithms for the 3l-cvrp, Comput. Oper. Res. 55 (2015) 127–140. solving vehicle routing problems, J. Comput. Sci. (2018).
[4] B. Kris, R. Katrien, V.N. Inneke, The vehicle routing problem: State of the [33] W. Liu, X. Li, A problem-reduction evolutionary algorithm for solving the
art classification and review, Comput. Ind. Eng. 99 (2016) 300–313. capacitated vehicle routing problem, Math. Probl. Eng. 2015 (2015).
[5] J. Caceres-Cruz, P. Arias, D. Guimarans, D. Riera, A.A. Juan, Rich vehicle [34] K. Socha, M. Dorigo, Ant colony optimization for continuous domains,
routing problem: Survey, ACM Comput. Surv. 47 (2) (2015) 32. European J. Oper. Res. 185 (3) (2008) 1155–1173.
[6] J. Li, T. Li, Y. Yu, Z. Zhang, P.M. Pardalos, Y. Zhang, Y. Ma, Discrete [35] M.L. Shahab, D.B. Utomo, M.I. Irawan, Decomposing and solving capacitated
firefly algorithm with compound neighborhoods for asymmetric multi- vehicle routing problem (cvrp) using two-step genetic algorithm (tsga), J.
depot vehicle routing problem in the maintenance of farm machinery, Theor. Appl. Inf. Technol. 87 (3) (2016).
Appl. Soft Comput. 81 (2019) 105460. [36] E. Osaba, F. Diaz, E. Onieva, R. Carballedo, A. Perallos, Amcpa: A population
[7] A.Z. Şevkli, B. Güler, A multi-phase oscillated variable neighbourhood metaheuristic with adaptive crossover probability and multi-crossover
search algorithm for a real-world open vehicle routing problem, Appl. Soft mechanism for solving combinatorial optimization problems, Int. J. Artif.
Comput. 58 (2017) 128–144. Intell. 12 (2) (2014) 1–23.
[8] F.Y. Vincent, A.A.N. Perwira Redi, Y. Chao-Lung, R. Eki, S. Budi, Symbi- [37] H. Daglayan, M. Karakaya, The impact of crossover and mutation operators
otic organisms search and two solution representations for solving the on a GA solution for the capacitated vehicle routing problem, Univ. J. Eng.
capacitated vehicle routing problem, Appl. Soft Comput. (2016) –. Sci. (2016).
[9] A. Fenton, The bees algorithm for the vehicle routing problem, 2016, arXiv [38] A. Mazidi, M. Fakhrahmad, M.H. SADREDDINI, A meta-heuristic approach
preprint arXiv:1605.05448. to cvrp problem: local search optimization based on ga and ant colony, J.
[10] M. Manfrin, Ant colony optimization for the vehicle routing problem, 2004, Adv. Comput. Res. (2016).
DEA defense at ULB. [39] A. Akhand, Z.J. Peya, K. Murase, Capacitated vehicle routing problem
[11] E. Osaba, F. Diaz, E. Onieva, Golden ball: a novel meta-heuristic to solve solving using adaptive sweep and velocity tentative PSO, Int. J. Adv.
combinatorial optimization problems based on soccer concepts, Appl. Comput. Sci. Appl. (IJACSA) (2017).
Intell. 41 (1) (2014) 145–166. [40] M. Metlicka, D. Davendra, Chaos driven discrete artificial bee algorithm for
[12] Z. Yanwei, L. Longlong, Q. Zhenyu, W. Wanliang, A discrete hybrid invasive location and assignment optimisation problems, Swarm Evol. Comput. 25
weed optimization algorithm for the capacitated vehicle routing problem, (2015) 15–28.
Procedia Comput. Sci. 91 (2016) 978–987, Promoting Business Analytics [41] S. Zainudin, M. KERWAD, Z.A. Othman, A water flow-like algorithm for
and Quantitative Management of Technology: 4th International Conference capacitated vehicle routing problem., J. Theor. Appl. Inf. Technol. 77 (1)
on Information Technology and Quantitative Management (ITQM 2016). (2015).
[13] X.-S. Yang, Firefly algorithms for multimodal optimization, in: International [42] T. Ehsan, K. Vahid, K. GH.M., Z. M., Enhanced intelligent water drops
Symposium on Stochastic Algorithms, Springer, 2009, pp. 169–178. and cuckoo search algorithms for solving the capacitated vehicle routing
[14] I. Fister, I. Fister Jr, X.-S. Yang, J. Brest, A comprehensive review of firefly problem, Inform. Sci. 334–335 (2016) 354–378.
algorithms, Swarm Evol. Comput. 13 (2013) 34–46. [43] E. Ruskartina, F.Y. Vincent, B. Santosa, A.P. Redi, Symbiotic organism search
[15] A.M. Altabeeb, A.M. Mohsen, A. Ghallab, An improved hybrid firefly (sos) for solving the capacitated vehicle routing problem, World Acad. Sci.
algorithm for capacitated vehicle routing problem, Appl. Soft Comput. 84 Eng. Technol. Int. J. Ind. Manuf. Eng. 2 (5) (2015).
(2019) 105728. [44] M. Ammi, S. Chikhi, A generalized island model based on parallel and
[16] M. Saraei, S.A. Ghaheri, An effective hybrid algorithm for vehicle routing cooperating metaheuristics for effective large capacitated vehicle routing
problem by indicating capacity using genetic and firefly algorithms, Int. J. problem solving, J. Comput. Inf. Technol. 23 (2) (2015) 141–155.
Eng. Educ. (IJEE) (2017). [45] E. Osaba, X.-S. Yang, F. Diaz, E. Onieva, A.D. Masegosa, A. Perallos, A
[17] P.-P. Matthopoulos, S. Sofianopoulou, A firefly algorithm for the discrete firefly algorithm to solve a rich vehicle routing problem modelling
heterogeneous fixed fleet VRP, Int. J. Ind. Syst. Eng. (2018). a newspaper distribution system with recycling policy, Soft Comput. (2016)
[18] J.J. Bentley, Fast algorithms for geometric traveling salesman problems, 1–14.
ORSA J. Comput. 4 (4) (1992) 387–411. [46] L. Zhou, L. Ding, X. Qiang, Y. Luo, An improved discrete firefly algorithm for
[19] S. Akpinar, Hybrid large neighbourhood search algorithm for capacitated the traveling salesman problem, J. Comput. Theor. Nanosci. 12 (7) (2015)
vehicle routing problem, Expert Syst. Appl. 61 (2016) 28–38. 1184–1189.
[20] A.A.R. Hosseinabadi, N.S.H. Rostami, M. Kardgar, S. Mirkamali, A. Abraham, [47] A.M. Mohsen, W. Al-Sorori, A new hybrid discrete firefly algorithm
A new efficient approach for solving the capacitated vehicle routing for solving the traveling salesman problem, in: Applied Computing and
problem using the gravitational emulation local search algorithm, Appl. Information Technology, Springer, 2017, pp. 169–180.
Math. Model. 49 (2017) 663–679. [48] E. Osaba, R. Carballedo, X.-S. Yang, F. Diaz, An evolutionary discrete firefly
[21] P. Kalatzantonakis, A. Sifaleras, N. Samaras, Cooperative versus non- algorithm with novel operators for solving the vehicle routing problem
cooperative parallel variable neighborhood search strategies: a case study with time windows, in: Nature-Inspired Computation in Engineering,
on the capacitated vehicle routing problem, J. Global Optim. (2019) 1–22. Springer International Publishing, Cham, 2016, pp. 21–41.

19
A.M. Altabeeb, A.M. Mohsen, L. Abualigah et al. Applied Soft Computing Journal 108 (2021) 107403

[49] W. Al-Sorori, A.M. Mohsen, Multi-population discrete bat algorithm with [54] N. Christofides, S. Eilon, An algorithm for the vehicle-dispatching problem,
crossover to solve TSP, in: International Conference on Hybrid Intelligent J. Oper. Res. Soc. 20 (3) (1969) 309–318.
Systems, Springer, 2016, pp. 466–478. [55] M.L. Fisher, Optimal solution of vehicle routing problems using minimum
[50] W. Al-Sorori, A. Mohsen, W. Aljoby, An improved hybrid bat algorithm for k-trees, Oper. Res. 42 (4) (1994) 626–642.
traveling salesman problem, in: International Conference on Bio-Inspired [56] N. Christofides, The vehicle routing problem, Combin. Optim. (1979).
Computing: Theories and Applications, Springer, 2016, pp. 504–511. [57] M. Alssager, Z.A. Othman, Taguchi-based parameter setting of cuckoo
[51] A.M. Mohsen, Annealing ant colony optimization with mutation operator search algorithm for capacitated vehicle routing problem, in: P.J. Soh, W.L.
for solving tsp, Comput. Intell. Neurosci. 2016 (2016). Woo, H.A. Sulaiman, M.A. Othman, M.S. Saat (Eds.), Advances in Machine
[52] E. Osaba, R. Carballedo, F. Diaz, E. Onieva, A.D. Masegosa, A. Perallos, Good Learning and Signal Processing: Proceedings of MALSIP 2015, Springer
practice proposal for the implementation, presentation, and comparison of International Publishing, Cham, 2016, pp. 71–79.
metaheuristics for solving routing problems, Neurocomputing 271 (2018) [58] A. Subramanian, E. Uchoa, L.S. Ochi, A hybrid algorithm for a class of
2–8. vehicle routing problems, Comput. Oper. Res. 40 (10) (2013) 2519–2531.
[53] P. Augerat, J.M. Belenguer, E. Benavent, A. Corberán, D. Naddef, G. Rinaldi, [59] J. Jianyong, G.C. Teodor, L. Arne, A parallel multi-neighborhood cooperative
Computational Results with a Branch-And-Cut Code for the Capacitated tabu search for capacitated vehicle routing problems, European J. Oper. Res.
Vehicle Routing Problem, Citeseer, 1998. 222 (3) (2012) 441–451.

20

You might also like