A Hybrid Genetic Algorithm For The Traveling Salesman Problem With Drone
A Hybrid Genetic Algorithm For The Traveling Salesman Problem With Drone
A Hybrid Genetic Algorithm For The Traveling Salesman Problem With Drone
https://doi.org/10.1007/s10732-019-09431-y
Quang Minh Ha1 · Yves Deville1 · Quang Dung Pham2 · Minh Hoàng Hà3
Abstract
This paper addresses the traveling salesman problem with drone (TSP-D), in which a
truck and drone are used to deliver parcels to customers. The objective of this prob-
lem is to either minimize the total operational cost (min-cost TSP-D) or minimize the
completion time for the truck and drone (min-time TSP-D). This problem has gained
a lot of attention in the last few years reflecting the recent trends in a new delivery
method among logistics companies. To solve the TSP-D, we propose a hybrid genetic
search with dynamic population management and adaptive diversity control based
on a split algorithm, problem-tailored crossover and local search operators, a new
restore method to advance the convergence and an adaptive penalization mechanism
to dynamically balance the search between feasible/infeasible solutions. The compu-
tational results show that the proposed algorithm outperforms two existing methods
in terms of solution quality and improves many best known solutions found in the lit-
erature. Moreover, various analyses on the impacts of crossover choice and heuristic
components have been conducted to investigate their sensitivity to the performance of
our method.
B Minh Hoàng Hà
minhhoang.ha@vnu.edu.vn
Quang Minh Ha
quang.ha@uclouvain.be
Yves Deville
yves.deville@uclouvain.be
Quang Dung Pham
dungpq@soict.hust.edu.vn
123
Q. M. Ha et al.
1 Introduction
The past few years have witnessed a rapid growth of interest in research on utilizing
drones with trucks for delivering parcels to customers. This new method deploys drones
with trucks to not only reduce delivery time and operational cost but also improve
service quality. A problem related to this new delivery method is called the routing
problem with drones, which is a generalization of the well-known traveling salesman
problem (in the case of one truck and one drone) and vehicle routing problem (in the
case of a fleet of trucks and drones); they are denoted TSP-D and VRP-D, respectively,
and their objective is to minimize either the total operational cost (min-cost) or the
completion time for a truck and drone (min-time).
In the literature, the very first work on this class of problems is the work of Murray
and Chu (2015), in which the authors proposed two subproblems. The first is a TSP-D
problem, called the flying sidekick traveling salesman problem (FSTSP) in which a
truck and drone cooperate with each other to deliver parcels. The authors introduced
a mixed integer programming formulation and a simple and fast heuristic with the
objective of minimizing the completion time for two vehicles. In the second problem,
the parallel drone scheduling TSP (PDSTSP), a single truck and a fleet of drones
are in charge of delivering parcels. The truck is responsible for parcels far from the
distribution centre (DC), and the drones are responsible for serving customers in its
flight range around the DC. Again, the objective is to minimize the latest time that a
vehicle returns to the depot. The problem description and hypothesis used in FSTSP
has been adapted in numerous subsequent studies such as in Ha et al. (2018), Ponza
(2016) and Freitas and Penna (2018) as well as in this paper.
Agatz et al. (2018) also introduced a TSP-D problem with assumptions differing
from those of the FSTSP. The most notable is that the drone may be launched and
returned to the same location (whereas this is forbidden in FSTSP). Additionally,
the two vehicles (truck and drone) share the same road network, hence the same
distance matrix (they are in different networks in FSTSP). The authors proposed a
mathematical model for this problem and developed several route-first, cluster-second
heuristics based on local search and dynamic programming to solve it with instances
with up to 10 customers. The above work has been extended further by Bouman et al.
(2018), who presented exact solution approaches, proving that the problem with larger
instances can be solved.
In a recent work, Freitas and Penna (2018) proposed a hybrid heuristic named
HGVNS to solve two TSP-D variants by Murray and Chu (2015) and Agatz et al.
(2018) with the min-time objective. In detail, HGVNS first obtains the initial solu-
tion by using a mixed-integer program (MIP) solver to solve the TSP optimally and
then applies a heuristic in which some trucks’ customers are removed and reinserted
as drone customers. Next, the initial solution is used as the input for a general vari-
able neighbourhood search in which eight neighbourhoods are shuffled and chosen
randomly. The authors conducted the experiments on three instance sets from Ponza
(2016) and Agatz et al. (2018) and TSPLIB. The computational results show that the
proposed approach can decrease delivery time by up to 67.79%.
A generalization of the TSP-D called the vehicle routing problem with drones
(VRPD or VRP-D) was first studied by Wang et al. (2017) where a fleet of trucks
123
A hybrid genetic algorithm for the traveling salesman…
and drones is responsible for delivering parcels. Several theoretical aspects have been
studied in terms of bounds and worst cases. An extension of that work was studied
in Poikonen et al. (2017), in which the author considered more practical aspects such as
drone endurance and cost. In addition, connections between VRPD and other classes
of VRPs have been made in the form of bounds and asymptotic results. Other works
related to drone applications are also presented in a survey conducted by Otto et al.
(2018).
In this paper, we introduce a new hybrid genetic algorithm (HGA) with adaptive
diversity control to effectively solve the TSP-D under both min-cost and min-time
objectives. HGA is a combination of the genetic algorithm and local search technique
together with a population management, diversity control and penalization mechanism
to balance the search between feasible and infeasible search spaces. This method was
initially proposed by Vidal et al. (2012) and has been used to solve many variants of
VRP efficiently, as in Vidal et al. (2012, 2013, 2014) and Bulhões et al. (2018). We
also present problem-tailored components to significantly improve the performance of
the algorithm. Different computational experiments show the improvements in terms
of solution quality under both objectives and different instance sets and the importance
of the new proposed elements.
The main contributions of this paper are as follows.
– We propose an efficient hybrid genetic algorithm that includes a new crossover, a
set of 16 local search operators, and a penalization and restore mechanism to solve
the TSP-D under both min-cost and min-time objectives.
– We conduct extensive computational experiments to evaluate the performance of
HGA under instance sets from Murray and Chu (2015), Ha et al. (2018) and Freitas
and Penna (2018). The proposed method outperforms two existing approaches
(GRASP and Murray et al.’s best) in terms of solution quality and can improve
upon a number of best known solutions among three instance sets.
– We analyze the efficiency and importance of the new components to the perfor-
mance of the overall algorithm.
The remaining parts of the paper are organized as follows. Section 2 introduces
the TSP-D and related assumptions considered in the problem. Section 3 discusses
the proposed hybrid genetic algorithm (HGA). Section 4 presents the computational
results, and Sect. 5 concludes the paper.
2 Problem description
In this section, we briefly discuss the description of the TSP-D, which was first pro-
posed in Murray and Chu (2015) and then developed further in Ha et al. (2018) to solve
the min-cost objective. In this problem, given a graph G = (V , A), V = {0 . . . n + 1}
is a set of depot and customer locations and A is a set of arcs that link two pair of nodes
in V . We need to deliver parcels to a set N = {1, . . . n} customers using a truck and a
drone (an unmanned aerial vehicle used for delivery). In this graph, 0 is the depot, and
n + 1 is its duplication. We denote di j and τi j (di j and τij ) as the distance and time
traveled from node i to node j by truck (drone), respectively. The effective arrival
123
Q. M. Ha et al.
times of the truck and drone are denoted by ti and ti . We have t0 = t0 = 0. Different
from actual arrival time, the effective arrival time of a vehicle (drone or truck) takes
into account both the actual arrival time and the time required to retrieve and (possi-
bly) prepare the drone for the next launch. This definition was initially described in
the work of Murray and Chu (2015). The drone is managed by the truck driver and
is carried in the truck while not in service. To make a delivery, the drone is launched
from either the truck at a customer location or the starting depot and later returns to the
truck at a different customer location or the return depot. The delivery plan of these
two vehicles (truck and drone) is subjected to the following requirements.
– Both vehicles (truck and drone) must start from and return to the depot.
– Each customer can only be serviced once by either a truck or drone. If a customer
is served by a truck (a drone), we call it a truck delivery (drone delivery).
– A drone delivery is represented as a 3-tuple i, j, k, where i, j, k are customer
locations that are described as follows:
– i is the node where the truck launches the drone, which we call the launch
node. We also denote s L as the time required for the truck driver to prepare the
drone for launch.
– j is the node the drone will fly to and make the delivery. We call it the drone
node. Most importantly, node j must be eligible for the drone to visit, as not
all parcels can be delivered by the drone due to capacity limitation (i.e., the
parcel is too heavy). We denote the set of nodes that can be served by drone
as VD ⊆ N .
– k is the rendezvous node, where the drone—after making a delivery—rejoins
the truck to have its battery replaced and to be made ready for the next launches.
The time required for the truck driver to replace drone’s battery and load new
parcel for possible next launch is called recovery (or retrieval) time and is
denoted as s R . In addition, two vehicles are required to wait for each other at
the rendezvous point, and while waiting for the truck, the drone is assumed to
be in constant flight.
– In a drone delivery, both truck and drone are required to satisfy the endurance
constraint, which is, in detail:
– Truck travel time constraint the truck travel time from the launch node to the
rendezvous node plus its recovery time cannot exceed the drone endurance
(the maximum operational time of a drone without recharging),
where τi→k is the truck travel time from i to k, and Γtr uck is the time taken for
the truck to recover the drone and possibly prepare it for the next launch. More
specifically, if the truck just recovers the drone without relaunching it at the
same location, then Γtr uck = s R . Otherwise, if the truck relaunches the drone
at the same location, then Γtr uck = s R + s L . This constraint is not imposed
when the launch node is the depot (node 0).
123
A hybrid genetic algorithm for the traveling salesman…
– Drone travel time constraint the drone travel time plus its recovery time
cannot exceed the drone endurance:
wT = α × ΨT , and (2)
wD = β × ΨD , (3)
where ΨT and Ψ D are the waiting times, and α and β are the per-unit-time waiting
fees of the truck and drone, respectively.
The objective of the TSP-D is either to minimize the total operational cost of the
system or to minimize the completion time of two vehicles. We denote the problem
with the first objective as “min-cost TSP-D” and with the latter as “min-time TSP-D”.
In this section, we describe a hybrid genetic algorithm with adaptive diversity control
method for solving TSP-D. The framework, as proposed in Vidal et al. (2012), is a
hybrid metaheuristic that combines the exploration capability of genetic algorithms
with efficient local search and diversity control. We adapt this general framework
with modifications to match the characteristics of the TSP-D. They include new local
search operators, crossovers, a penalized mechanism and a restoration method to con-
vert from a TSP-D solution to a giant-tour chromosome. We describe the approach
in Algorithm 1. In detail, starting from an initial population (Line 1), for each iter-
ation, two parents are selected to generate an offspring individual using a crossover
operator (Line 4). This offspring then goes through a split procedure [proposed in Ha
et al. (2018)] to obtain the drone delivery and truck delivery chromosome (Line 5).
Subsequently, the offspring is “educated” by a local search method—which contains
multiple operators—to improve its quality. The educated offspring then employs a
restore method to update its giant tour chromosome (Line 7). The offspring is then
checked for feasibility and is added to the appropriate subpopulation (feasible or
infeasible). It also has a probability of being repaired of Pr ep = 50% and is added
123
Q. M. Ha et al.
to a feasible subpopulation if the repair succeeds (Lines 8–14). In the next step, if
a subpopulation reaches its maximum size, a survivor selection method is called to
eliminate a number of individuals in that subpopulation, keeping only the best ones
(Lines 15–17). The method then adjusts the penalty parameters (Line 18) and calls
the diversification procedure if the search is not improved after a certain number of
iterations (Line 19). Finally, we return the best feasible solution found (Line 22).
The rest of this section is arranged as follows. We first define the search space in
Sect. 3.1. Section 3.2 describes the solution representation. Section 3.3 presents the
evaluation of individuals. Parent selection and crossover are described in Sect. 3.4.
Section 3.5 discusses the local search procedure, and various operators are presented.
The restore method is introduced in Sect. 3.6. Finally, Sect. 3.7 regards the population
management with the population initialization, adjustment of penalty coefficients,
survivor selection and diversity control.
It has been well studied that by exploiting infeasible solutions, we can significantly
improve the performance of a heuristic (Glover and Hao 2011). In this section, we
define the search space S, which includes the feasible and infeasible solutions s ∈
S. Infeasible solutions comprise drone deliveries that violate the drone endurance
constraint. More specifically, a drone delivery in a TSP-D solution is not valid in the
following two scenarios.
123
A hybrid genetic algorithm for the traveling salesman…
– The truck travel time constraint is violated (except for the case where the launch
node is the depot, as described in the problem description above):
In these two cases, the drone cannot feasibly be flown because its battery will be
depleted before the retrieval operation undertaken by the truck driver is completed.
Let sol(s) represent a TSP-D solution in the search space. We have sol(s) =
(T D, D D), where T D = e0 , . . . ek , ei ∈ V is the truck tour, and D D ⊆ P is the
set of drone deliveries in solution s.
We now define the fitness evaluation function for min-time and min-cost TSP-D
separately.
For min-cost TSP-D, the operational cost of solution s, denoted cost(T D, D D), is
calculated as
where
k−1
– cost(T D) = C1 di,i+1 is the cost of the truck tour;
i=0
– cost(D D) = C2 (di j + d jk ) is the cost of drone deliveries; and
i, j,k∈D D
– costW (D D) = costW T (i, j, k) + cost D (i, j, k is the wait cost of the
W
i, j,k∈D D
T = α × max(0, τ
truck and drone. We have costW i→k − τi jk ), where τi→k is the
truck travel time from i to k (in the truck tour), and τijk is the drone travel time
D = β × max(0, τ − τ
from i to j to k. In addition, costW i jk i→k ).
Let ω represent the penalty for violating the drone endurance constraint. We define
the penalized cost of a solution s as the sum of the operational cost and the weighted
sum of the truck’s or drone’s excess travel time during drone deliveries. This penalized
cost is computed as
φ(s) = cost(T D, D D) + ω max(0, τi→k + Γtr uck − ) × ΥT × C1
i, j,k∈D D (5)
+ max(0, τij + τ jk + Γdr one − ) × Υ D × C2 ,
123
Q. M. Ha et al.
where ω is the penalty for violating the constraint, and ΥT and Υ D are the speeds of the
truck and drone, respectively. This penalized cost is then used as the fitness function
to compute the fitness of the individuals.
Similar to min-cost TSP-D, we also have the penalized cost of a solution s in the min-
time objective as the sum of the completion time of two vehicles and the penalties for
violating the constraint. It is computed as follows:
φ(s) = time(s) + ω max(0, max(τi→k + Γtr uck ,
i, j,k∈D D (7)
τij + τ jk + Γdr one ) − ).
A solution in HGA is represented as a giant TSP tour (giant tour) with the two depots
removed. We also denote this as a (giant-tour) chromosome. When a TSP-D solution
is needed for a local search method, it can be obtained using the split procedure, which
runs in polynomial time (Ha et al. 2018). Conversely, we can retrieve a giant tour from
a TSP-D solution by using a restore method, which will be discussed in the coming
section. To conclude, by having a transformation between a giant tour and TSP-D
solution using the split and restore method, we can use the fast and efficient operators
in both the crossover and local search step. A demonstration of this transformation is
described in Fig. 1.
This section discusses the individual assessment method which has been proposed
in Vidal et al. (2012). To evaluate any individual P1 in the population, we consider
two factors: its penalized cost φ(P1 ) (described in Sect. 3.1) and its contribution to the
diversity of the population, denoted Δ(P1 ) and calculated as the average “distance”
from P1 to its closest neighbours in the population. By taking into account these two
factors, we can obtain a balance between intensification and diversification. Other-
wise, the heuristic might either converge too soon and too quickly (focusing only on
improving the penalized cost) or will always explore completely different giant tours,
leading to a random search. In detail, the diversity contribution described above is
presented in Eq. 8:
123
A hybrid genetic algorithm for the traveling salesman…
1
Δ(P1 ) = δ(P1 , P2 ), (8)
n close
P2 ∈Nclose
where n close is the number of considered closest neighbours, and Nclose is the set of
closest neighbours of P1 (i.e., the set of elements sorted using Eq. 9). The distance
between two individuals P1 and P2 , denoted δ(P1 , P2 ), is a normalized Hamming
distance based on the differences between the nodes in the same positions of the
giant-tour chromosome. This distance is shown in Eq. 9, where 1(condition) is a
valuation function that returns 1 if the condition is true and 0 otherwise.
1 gt gt
δ(P1 , P2 ) = [1(P1 (i) = P2 (i))], (9)
n
i=1,...,n
gt gt
where 1(P1 (i) = P2 (i)) returns 1 if the node in position i of the giant-tour chro-
mosome in P1 is different than the node in position i of the giant-tour chromosome of
P2 , and 0 otherwise.
The evaluation of an individual P, or as we call it, the biased fitness, denoted
B F(P), is then computed as in Eq. 10, where f it(P) is the rank of P in the subpopu-
lation of size nbI ndiv with respect to its penalized cost φ(P), and dc(P) is the rank
of P in the subpopulation in terms of diversity contribution. The parameter nbElite
ensures that a certain number of elite individuals will survive to the next generation
during the survival selection process [proven in Vidal et al. (2012)].
nbElite
B F(P) = f it(P) + 1 − dc(P) (10)
nbI ndiv
123
Q. M. Ha et al.
In detail, Algorithm 2 first takes the two parents P1 , P2 as one of its inputs.
Moreover, in line 1, it also takes into account the corresponding TSP-D solution
(T D1 , D D1 ) of P1 , which was obtained during the “education” process (Line 6 of
Algorithm 1). Subsequently, it defines two TSP tours, T S P1 , T S P2 , in Line 2 by tak-
ing two parents and adding two depots to them. An empty offspring with two depots
is also initialized in Line 3. In Line 4, a random number is generated to decide from
which component—T D1 or D D1 —the algorithm will inherit. In either case, it will
choose a random segment of the chosen component by generating two random cut
points, a, b, with a < b, and copy the nodes between those cut points to C while
keeping their original positions in T S P1 (Lines 5–9). Finally, the remaining nodes of
C are filled one by one, starting from position 1, by taking the remaining nodes of
T S P1 and copying to C while keeping their relative orders in T S P2 (Line 10). The
offspring is returned by removing two depots of C (Line 11).
123
A hybrid genetic algorithm for the traveling salesman…
The main role of the education step is to improve the quality of solutions by means
of the local search procedure. We design a hill-climbing and first-improvement local
search for both min-cost and min-time objectives. Similar to Vidal et al. (2012), we
also apply the technique proposed in Toth and Vigo (2003) to restrict the search to
the h × n closest vertices, where h = 0.1 is the granular threshold. This technique
significantly reduces the computation time consumed by the education process. We
define neighborhoods for the TSP-D based on a set of 16 move operators in which
3 operators—N1 , N13 and N14 —are inherited from the work of Ha et al. (2018). In
each operator, the evaluation separately evaluates the move costs for the min-cost and
min-time objectives. For min-cost, it is the total truck and drone costs of the affected
arcs, while the total truck and drone travel times of the affected arcs are calculated
in the min-time problem. Moreover, the truck and drone cumulative time and cost as
well as the cost and time of all drone tuples in set P are pre-computed at the beginning
of the HGA to effectively accelerate the algorithm.
123
Q. M. Ha et al.
Fig. 3 Illustrations of N1 to N6
123
A hybrid genetic algorithm for the traveling salesman…
– N11 (Intradrone delivery drone rdv swap 1–1): This is similar to the above move
operator, except that we swap j and k.
– N12 (Intradrone delivery launch rdv swap 1–1): Again, it is similar to the above
move operator, but instead, we swap i and k.
– N13 (Drone insertion): Select a random node j such that j is either a truck-only
node or the node in between a drone delivery, choose two other nodes i and k in
the truck tour—i is before k—and create a new drone delivery i, j, k. This move
is only valid when no drone delivery interference exists between i and k or when
we can say that there is no drone launch or retrieval between i and k.
– N14 (Drone remove): We select a random drone node j, remove the associated
drone delivery, and reinsert j between two consecutive nodes i and k in the truck
tour.
– N15 (Drone swap 1–1): Select two random drone deliveries i 1 , j1 , k1 and
i 2 , j2 , k2 , and swap j1 and j2 . We will therefore have two new drone deliveries:
i 1 , j2 , k1 and i 2 , j1 , k2 .
– N16 (Drone relocation 1–1): Select a random drone delivery i, j, k, and choose a
new launch i and rendezvous node k for j to have a new drone delivery i , j, k .
To more efficiently guide the search for good solutions, a restoration method is devel-
oped in which we use the educated TSP-D solution (the one that has been improved
123
Q. M. Ha et al.
using local search) to update the existing giant tour individual. In detail, the new giant
tour is constructed by reinserting drone nodes in the drone deliveries of the educated
TSP-D solution to a random position between their launch node and rendezvous node
on the truck delivery tour of that solution. After the insertion operation is finished,
two depots are removed to obtain a valid giant tour individual. As a result, we have a
new giant tour individual that is formed by an “educated” truck tour with drone nodes
being reinserted. An illustration of this process is shown in Fig. 6.
It is possible that our randomized restoring can generate a random solution based
on the provided solution but the new solution is at least as good as the old one due to
the optimality of our split procedure. This ensures the convergence of our algorithm.
In the restore method, with each drone delivery i, j, k, we have tried three ways to
choose the position for reinserting drone node j: (1) the position right after node i;
(2) the position right before node k; and (3) a random position between i and k. The
experimental results show that the third option performs the best. Our guess is that
random restore method could help to increase the diversification of the algorithm, thus
avoid the algorithm converging too fast.
123
A hybrid genetic algorithm for the traveling salesman…
Fig. 6 Reinsertion in restore method. Truck travels the solid lines and drone travels the dashed lines
and the tested result in Ha et al. (2018), which suggests the use of high-quality TSP
tours instead of completely random ones. We obtain the giant tour chromosomes after
these generation steps. These tours then pass through the split method to obtain the
corresponding TSP-D solutions of each individual. In the next step, individuals’ TSP-
D solutions are processed using the education process to improve their qualities, and
when an infeasibility occurs, they are probablistically repaired. After that, the restore
method is called to update individuals’ giant tour chromosomes. The individuals are
then added to the appropriate subpopulations based on their feasibilities.
Any subpopulation that exceeds the size of μ+λ is passed through a select survivors
method in which λ individuals are discarded. The discarded ones are ones defined as
“clones” or the worst individuals with respect to their biased fitnesses. Solutions are
defined as clones if and only if they have the same giant tour (possibly in reversed
order).
Furthermore, the penalty coefficient ω is dynamically adjusted during the search for
each 100 iterations. This mechanism is necessary to guide the algorithm in two search
spaces. More specifically, the penalty coefficient is increased when the search produces
too many infeasible solutions (meaning that it falls too deeply into the infeasible
search space) and is decreased in the opposite case. In detail, let E R E F be the targeted
proportion of the feasible solution, and we then adjust the parameter ω as follows: if
the naturally feasible proportion is below E R E F − ζ % (is higher than E R E F + ζ %),
then the penalty coefficient is increased by η I (decreased by η D ). This means that when
the feasible proportion is in the range E R E F ± ζ %, the coefficient remains unchanged
to avoid the search jumping too quickly between regions in the search space.
When the search is not improved after I ter D I V iterations, the diversification method
is called, in which we retain the best n best individuals with respect to their biased fitness
123
Q. M. Ha et al.
and generate n init Pop new individuals as in the initialization phase. This technique is
important because it creates new genetic materials for the search when the population
has lost its diversification characteristic.
4 Computational results
This section presents the computational results of the HGA, which has been imple-
mented in C++ and compiled with the “-O3” flag. The experiments are run on a desktop
computer with an Intel Core i7-6700, 3.4 GHz processor.
Because the parameters proposed in Vidal et al. (2012) have been proven to work
well on many variants of VRP, we retained most of them. In detail, the default param-
eters of HGA are μ = 15, λ = 25, n init Pop = 4μ, ζ = 5%, η I = 1.2, η D =
0.85, n best = μ/3, nbElite = 6, E R E F = 0.3, n close = 0.2, ω = 1.0, I ter N I =
2500, and I ter D I V = 0.3 × I ter N I .
For the TSP-D parameters, we used the parameters proposed in Murray and Chu
(2015): the truck speed and drone speed were set to 40 km/h, and the drone endurance
was 20 min. The time required to launch and retrieve the drone (s L and s R ) were
both set to 1 min.
As described in Sect. 3.1, there are two types of infeasibilities in a TSP-D: truck
travel time and drone travel time constraint violations. From those constraint violations,
we define three levels of relaxations.
– RelaxAll We accept both types of infeasibilities.
– RelaxTruck We only accept the truck travel time constraint violation in infeasible
solutions.
– RelaxDrone We only accept the drone travel time constraint violation in infeasible
solutions.
The impacts of these different types of relaxations are investigated in Sect. 4.3. By
default, RelaxAll is used. The default selection for the crossover is DX, which is the
best performing crossover as tested in Sect. 4.2.
The following sections are organized as follows. We first evaluate the performance
of HGA with different instance sets and compare with the existing methods. Next, an
analysis of the impacts of different crossovers is presented. Finally, we evaluate the
sensitivity of each component in HGA.
In this section, we test HGA on three sets of instances: (1) 72 min-time instances of
10 customers from Murray and Chu (2015); (2) 60 instances of 50 and 100 customers
from Ha et al. (2018) under both min-time and min-cost objective functions; and (3) 24
instances with various size introduced in the recent work of Freitas and Penna (2018).
For the HGA, we collected its best found solutions and computed the objective func-
tion’s value of solutions on average over 10 runs. Current best methods—GRASP
in Ha et al. (2018), HGVNS in Freitas and Penna (2018) and different approaches
proposed in Murray and Chu (2015)—were selected to compare with HGA. As men-
123
A hybrid genetic algorithm for the traveling salesman…
tioned before, the standard version of HGA with DX and RelaxAll was used in this
experiment. The results for Instance Sets (1), (2) and (3) are presented in Tables 1, 2,
3 and 4.
In Table 1, we compare HGA with the best results found by Murray and Chu (2015)
and GRASP Ha et al. (2018) among 36 instances of Set 1 with two settings of drone
endurance (20 and 40 min). The column shows the drone endurance in minutes.
Column H G A represents the best found solutions while column H G A reports the
average values among 10 runs of our new algorithm. The values in bold text imply the
best result found among the three approaches. Overall, HGA was able to improve the
existing best found solutions in 9 tests and obtained results as good as the best ones in
60 tests. Column H G A shows the stability of HGA in this context when the solutions
over 10 runs generally reach the best ones in all instances but two. The results also
demonstrate a dominance of our HGA over GRASP in terms of solutions’ quality.
However, HGA is in general slower than GRASP.
Tables 2 and 3 report the comparisons of objective value and average run time (in min-
utes) between HGA and GRASP in Ha et al. (2018) on Instance Set 2. We collected
the average value (Column “H G A”) and best solution of HGA found among repeated
runs (Column “H G A”) and its average run time in minutes (Column “TH G A ”). The
corresponding values of GRASP are reported in Column “G R AS P”, “G R AS P”,
and “TG R AS P ”. Column “Change (%)” calculates the percentage change between best
found objective values of HGA and GRASP. A negative value indicates an improve-
ment of our new method in comparison to GRASP. With respect to this comparison,
HGA shows improvements in terms of solutions‘ quality in both min-cost and min-time
objectives.
In detail, for min-cost TSP-D (Table 2), the average objective values of solutions
of HGA are even better than those of the best found solutions of GRASP on most
instances (see Columns “H G A”and “G R AS P”). The proposed algorithm can signif-
icantly improve existing best known solutions by 6.40% and 15.10% on average (up to
nearly 15% and 20%) for 50- and 100-customer instances, respectively. We can observe
that the algorithm performs better in large instances (i.e., 100-customer instances).
However, it is worth mentioning that GRASP performs better on two instances D5
and D6. Regarding run time, HGA is 1.5 to 2 times slower than GRASP due to its
more complex design. This result is acceptable since it still can deliver significantly
better results in less than 1 min for 50-customer instances and less than 5 min for
100-customer instances.
For min-time TSP-D (Table 3), HGA can also improve the existing best known
solutions found by GRASP on all instances but not as significantly as in min-cost
TSP-D. In detail, the improvements are 2.39% and 4.05% on average (and up to nearly
6% and 8%) for 50- and 100-customer instances, respectively. Again, HGA performs
123
Q. M. Ha et al.
437v1 20 56.468 57.446 56.468 56.468 440v7 20 49.996 49.776 49.422 49.422
437v1 40 50.573 50.573 50.573 50.573 440v7 40 49.204 49.204 49.204 49.204
437v2 20 53.207 53.207 53.207 53.207 440v8 20 62.796 62.700 62.576 62.576
437v2 40 47.311 47.311 47.311 47.311 440v8 40 62.270 62.004 62.004 62.004
437v3 20 53.687 54.664 53.687 53.687 440v9 20 42.799 42.566 42.533 42.533
437v3 40 53.687 53.687 53.687 53.687 440v9 40 42.799 42.566 42.533 42.533
437v4 20 67.464 67.464 67.464 67.464 440v10 20 43.076 43.076 43.076 43.076
437v4 40 66.487 66.487 66.487 66.487 440v10 40 43.076 43.076 43.076 43.076
437v5 20 50.551 50.551 50.551 50.551 440v11 20 49.204 49.204 49.204 49.204
437v5 40 45.835 44.835 44.835 44.835 440v11 40 49.204 49.204 49.204 49.204
437v6 20 45.176 47.601 47.311 47.311 440v12 20 62.004 62.004 62.004 62.004
437v6 40 45.863 43.602 43.602 43.602 440v12 40 62.004 62.004 62.004 62.004
437v7 20 49.581 49.581 49.581 49.581 443v1 20 69.586 69.586 69.586 69.586
437v7 40 46.621 46.621 46.621 46.621 443v1 40 55.493 55.493 55.493 55.493
437v8 20 62.381 62.381 62.381 62.381 443v2 20 72.146 72.146 72.146 72.146
437v8 40 59.776 59.416 59.416 59.416 443v2 40 58.053 58.053 58.053 58.053
437v9 20 45.985 42.945 42.416 42.416 443v3 20 77.344 77.344 77.344 77.344
437v9 40 42.416 42.416 42.416 42.416 443v3 40 69.175 68.431 68.431 68.431
437v10 20 42.416 41.729 41.729 41.729 443v4 20 90.144 90.144 90.144 90.144
437v10 40 41.729 41.729 41.729 41.729 443v4 40 82.700 83.700 82.700 82.700
437v11 20 42.896 42.896 42.896 42.896 443v5 20 55.493 58.210 54.973 55.077
437v11 40 42.896 42.896 42.896 42.896 443v5 40 53.447 51.929 51.929 51.929
437v12 20 56.696 56.425 56.273 56.273 443v6 20 58.053 58.053 55.209 55.209
437v12 40 55.696 55.696 55.696 55.696 443v6 40 52.329 52.329 52.329 52.329
440v1 20 49.430 50.164 49.430 49.430 443v7 20 64.409 65.523 65.523 65.523
440v1 40 46.886 46.886 46.886 46.886 443v7 40 60.743 60.743 60.743 60.743
440v2 20 50.708 51.828 50.708 50.708 443v8 20 77.209 78.323 78.323 78.323
440v2 40 46.423 46.423 46.423 46.423 443v8 40 73.967 72.967 72.967 72.967
440v3 20 56.102 58.502 56.102 56.102 443v9 20 49.049 45.931 45.931 45.931
440v3 40 53.933 53.933 53.933 53.933 443v9 40 47.250 45.931 45.931 45.931
440v4 20 69.902 73.091 69.902 69.902 443v10 20 47.935 46.935 46.935 46.935
440v4 40 68.397 68.397 68.397 68.397 443v10 40 47.935 46.935 46.935 46.935
440v5 20 43.533 44.624 43.533 43.533 443v11 20 57.382 56.395 56.395 56.395
440v5 40 43.533 43.533 43.533 43.533 443v11 40 56.395 56.395 56.395 56.395
440v6 20 44.076 44.122 43.949 43.949 443v12 20 69.195 69.195 69.195 69.195
440v6 40 44.076 43.944 43.810 43.853 443v12 40 69.195 69.195 69.195 69.195
123
Table 2 Comparison with GRASP under min-cost objective—instance set 2
B1 1372.82 1413.24 0.27 1225.78 1239.85 − 10.71 0.5 E1 2206.53 2255.99 2.28 1775.1 1802.47 − 19.55 3.47
B2 1491.3 1513.98 0.26 1381.89 1402.98 − 7.34 0.38 E2 2210.61 2273.09 2.28 1795.03 1830.95 − 18.80 3.66
B3 1503.78 1521.67 0.28 1357.17 1370.82 − 9.75 0.45 E3 2248.16 2312.76 2.48 1818.16 1861.59 − 19.13 3.5
B4 1396.17 1426.2 0.27 1282.16 1292.87 − 8.17 0.49 E4 2179.06 2223.97 2.97 1776.58 1822.36 − 18.47 3.68
B5 1457.91 1500.9 0.31 1351.37 1357.61 − 7.31 0.4 E5 2286.16 2360.3 2.87 1866.22 1899.25 − 18.37 3.55
B6 1316.08 1353.76 0.27 1159.79 1174.30 − 11.88 0.44 E6 2244.62 2313.86 3.26 1795.17 1831.23 − 20.02 4.12
B7 1370.05 1399.71 0.24 1308.25 1322.70 − 4.51 0.42 E7 2249.09 2313.67 3.18 1892.41 1923.18 − 15.86 3.46
B8 1484.93 1517.23 0.25 1255.61 1275.61 − 15.44 0.62 E8 2220.88 2272.55 3.15 1813.73 1831.46 − 18.33 4.3
B9 1442.09 1468.86 0.28 1355.32 1363.53 − 6.02 0.52 E9 2279.91 2326.29 2.87 1882.71 1901.76 − 17.42 5.01
B10 1392.54 1429.57 0.25 1252.9 1257.48 − 10.03 0.47 E10 2324.74 2384.52 3.41 1870.55 1932.75 − 19.54 4.07
A hybrid genetic algorithm for the traveling salesman…
C1 2870.41 2935.87 0.21 2679.1 2703.14 − 6.66 0.29 F1 4569.83 4648.2 1.85 3766.63 3854.56 − 17.58 3.24
C2 2804.47 2868.67 0.26 2750.74 2755.04 − 1.92 0.36 F2 4186.76 4318.78 2.38 3469.54 3575.90 − 17.13 3.95
C3 3087.55 3185.09 0.16 2932.78 2952.48 − 5.01 0.32 F3 4414.38 4563.64 2.45 3751.07 3891.27 − 15.03 3.02
C4 2844.1 2916.86 0.20 2655.25 2676.82 − 6.64 0.67 F4 4499.09 4600.27 2.14 3818.62 3862.57 − 15.12 3.27
C5 3323.92 3367.34 0.19 3133.69 3156.88 − 5.72 0.36 F5 4381.37 4597.32 2.66 3756.78 3807.86 − 14.26 3.28
C6 3433.99 3472.39 0.19 3238.92 3268.49 − 5.68 0.4 F6 4032.9 4171.8 2.63 3465.56 3560.66 − 14.07 2.54
C7 3001.13 3047.71 0.21 2681.01 2738.71 − 10.67 0.91 F7 4076.31 4213.52 2.84 3601.78 3660.90 − 11.64 4.83
123
Table 2 continued
123
C8 3481.17 3557.99 0.22 3250.19 3259.22 − 6.64 0.91 F8 4491.2 4597.9 2.77 3803.14 3930.37 − 15.32 3.91
C9 3267.23 3306.38 0.19 3032.73 3056.08 − 7.18 0.58 F9 4388.91 4463.39 2.55 3873.51 3904.18 − 11.74 4.07
C10 3291.2 3356.29 0.23 3082.07 3117.05 − 6.35 0.61 F10 4173.64 4567.84 2.57 3837.47 3895.06 − 8.05 3.1
D1 4159.39 4389.24 0.21 3927.97 3928.12 − 5.56 0.43 G1 5947.97 6148.5 1.94 5084.56 5312.36 − 14.52 3.08
D2 4275.46 4334.4 0.19 4097 4113.47 − 4.17 0.58 G2 5882.97 5987.64 2.63 5198.89 5234.67 − 11.63 3.16
D3 4085.71 4191.08 0.18 3846.84 3861.61 − 5.85 0.39 G3 6074.57 6138.94 2.82 5063.87 5116.38 − 16.64 4.08
D4 4612.46 4714.62 0.21 4334.32 4334.32 − 6.03 0.31 G4 6458.96 6632.14 2.39 5542.19 5703.23 − 14.19 4.1
D5 4717.67 4793.39 0.20 4569.16 4584.22 − 3.15 0.65 G5 6198.95 6329.25 2.59 5496.77 5566.09 − 11.33 2.94
D6 4405.02 4485.87 0.20 4384.00 4385.85 − 0.48 0.77 G6 6049.34 6343.26 2.95 5377.51 5463.62 − 11.11 2.67
D7 4749.57 4796.23 0.25 4634.53 4657.77 − 2.42 0.73 G7 5889.08 6023.11 2.85 5318.17 5396.12 − 9.69 4.98
D8 4143.03 4287.87 0.20 3911.94 3963.02 − 5.58 0.64 G8 5599.55 5871.96 2.62 5112.58 5246.50 − 8.70 5.03
D9 4653.73 4688.16 0.22 4469.78 4490.67 − 3.95 0.66 G9 6050.8 6254.5 3.08 4996.42 5187.45 − 17.43 4.3
D10 4260.6 4301.83 0.20 4208.93 4232.38 − 1.21 0.95 G10 6249.69 6534.13 2.70 5473.91 5598.25 − 12.41 3.77
Mean 0.22 − 6.40 0.51 2.65 − 15.10 3.68
Q. M. Ha et al.
Table 3 Comparison with GRASP under min-time objective—instance set 2
B1 120.68 121.69 0.45 115.65 116.43 − 4.17 0.76 E1 188.58 192.08 5.45 187.67 188.32 − 0.48 3.6
B2 118.53 119.46 0.48 118.39 118.39 − 0.12 0.33 E2 190.55 192.88 5.71 187.21 188.01 − 1.75 5.6
B3 119.7 120.25 0.52 116.21 116.39 − 2.92 0.57 E3 189.05 192.83 5.65 188.09 188.89 − 0.51 4.58
B4 123.02 124.7 0.36 118.71 119.26 − 3.50 0.47 E4 188.61 191.27 4.54 186.23 186.99 − 1.26 4.69
B5 119.46 120.77 0.48 115.78 115.91 − 3.08 0.58 E5 190.47 193.61 4.34 187.71 188.26 − 1.45 4.06
B6 119.54 121.46 0.39 114.31 115.46 − 4.38 0.88 E6 190.32 193.86 4.10 189.16 189.44 − 0.61 4.84
B7 118.54 121.02 0.33 115.52 115.63 − 2.55 0.62 E7 191.51 194.41 4.33 190.39 190.89 − 0.58 3.84
B8 119.36 119.99 0.35 117.9 118.04 − 1.22 0.78 E8 190.47 193.74 3.86 189.02 189.54 − 0.76 4.22
B9 118.26 119.86 0.42 117.64 117.72 − 0.52 0.39 E9 191.12 193.7 4.31 189.76 189.94 − 0.71 4
B10 119.8 121.27 0.37 117.38 117.70 − 2.02 0.6 E10 189.71 193.28 4.17 189.45 189.91 − 0.14 3.4
A hybrid genetic algorithm for the traveling salesman…
C1 220.63 222.6 0.27 215.07 215.37 − 2.52 0.6 F1 341.68 344.98 2.65 322.94 326.10 − 5.48 5.73
C2 210.39 211.14 0.41 209.23 210.11 − 0.55 0.53 F2 325.7 330.63 2.87 308.74 310.89 − 5.21 5.24
C3 214.61 215.31 0.28 212.02 212.22 − 1.21 0.38 F3 336.35 340.92 3.88 309.67 313.55 − 7.93 5.61
C4 225.15 225.47 0.25 212.08 213.27 − 5.81 0.6 F4 326.79 334.69 1.98 311.37 314.96 − 4.72 6.06
C5 226.36 233.97 0.32 223.06 224.57 − 1.46 0.48 F5 335.88 344.61 2.04 314.82 317.83 − 6.27 6.57
C6 240.37 242.22 0.27 234.01 235.56 − 2.65 0.31 F6 309.71 319.22 2.23 294.38 297.47 − 4.95 4.7
C7 227.73 229.56 0.20 222.27 223.40 − 2.40 0.51 F7 317.86 330.83 1.67 311.41 316.15 − 2.03 4.92
123
Table 3 continued
123
C8 242.19 245.12 0.37 234.26 237.53 − 3.27 0.46 F8 345.44 350.75 1.96 323.74 326.40 − 6.28 5.21
C9 237.98 241.07 0.26 226.01 227.43 − 5.03 0.68 F9 339.53 342.41 1.83 315.56 318.47 − 7.06 4.66
C10 230.03 235.47 0.33 226.17 226.17 − 1.68 0.48 F10 332.05 340.35 1.87 312.7 315.13 − 5.83 3.94
D1 315.8 318.9 0.31 306.39 307.09 − 2.98 0.61 G1 437.48 450.23 1.87 417.92 425.19 − 4.47 4.45
D2 317.15 322.85 0.29 313.93 315.64 − 1.02 0.57 G2 415.32 424.88 2.67 389.64 390.14 − 6.18 2.4
D3 300.4 303.26 0.29 295.86 297.54 − 1.51 0.6 G3 446.6 454.98 2.45 411.47 415.14 − 7.87 4.9
D4 333.47 336.7 0.31 323.72 324.60 − 2.92 0.56 G4 449.68 465.83 1.83 433.09 435.56 − 3.69 4.67
D5 324.68 326.46 0.25 321.46 321.83 − 0.99 0.4 G5 434.6 446.73 1.67 421.05 422.49 − 3.12 4.48
D6 315.16 317.73 0.28 313.21 313.65 − 0.62 0.49 G6 450.28 462 1.73 415.46 420.84 − 7.73 5.51
D7 329.31 330.24 0.33 316.65 317.83 − 3.84 0.32 G7 420 439.62 1.42 409.31 412.14 − 2.55 5.21
D8 306.28 312.12 0.31 293.76 296.51 − 4.09 0.58 G8 442.67 453.19 1.71 406.51 407.89 − 8.17 5.08
D9 326.09 331.31 0.27 317.85 318.31 − 2.53 0.41 G9 456.78 469.49 1.32 428.16 435.75 − 6.27 5.91
D10 306.1 309.54 0.29 305.51 305.54 − 0.19 0.41 G10 460.89 470.44 2.15 426.82 430.94 − 7.39 5.4
Mean 0.33 − 2.39 0.52 2.66 − 4.05 4.51
Q. M. Ha et al.
Table 4 Comparison with HGVNS in Freitas and Penna (2018) under min-time objective
Instance T S P∗ HGA Gap H G A (%) HGA Gap H G A (%) TH G A (s) HGVNS Gap H GV N S (%) H GV N S Gap H GV N S (%) TH GV N S (s)
berlin52 239.75 198.00 − 17.41 199.80 − 16.66 14.15 210.03 − 12.39 220.23 − 8.14 6.50
bier127 3665.60 3499.11 − 4.54 3506.42 − 4.34 64.06 3456.80 − 5.70 3587.88 − 2.12 53.69
ch130 187.83 182.86 − 2.64 182.86 − 2.64 76.12 178.16 − 5.15 180.40 − 3.95 44.13
d198 463.45 460.16 − 0.71 461.16 − 0.50 114.00 461.83 − 0.35 461.83 − 0.35 67.69
eil51 13.45 13.45 0.00 13.45 0.00 10.51 13.45 0.00 13.68 1.71 11.57
eil76 16.90 16.90 0.00 16.90 0.00 26.71 16.35 − 3.25 16.68 − 1.30 27.14
kroA100 661.30 539.91 − 18.36 541.36 − 18.14 97.87 587.80 − 11.11 609.71 − 7.80 30.95
kroA150 822.60 688.35 − 16.32 693.612 − 15.68 145.16 764.42 − 7.07 780.93 − 5.07 40.95
kroA200 922.05 806.87 − 12.49 820.86 − 10.97 169.53 870.65 − 5.57 873.99 − 5.21 46.53
kroB150 823.65 656.36 − 20.31 676.11 − 17.91 146.29 763.15 − 7.35 773.72 − 6.06 50.20
kroB200 917.95 799.64 − 12.89 801.36 − 12.70 152.21 835.43 − 8.99 838.40 − 8.67 31.94
A hybrid genetic algorithm for the traveling salesman…
kroC100 662.30 544.68 − 17.76 547.38 − 17.35 79.22 658.38 − 0.59 660.93 − 0.21 36.63
kroD100 661.00 544.88 − 17.57 547.22 − 17.21 65.43 606.45 − 8.25 652.34 − 1.31 40.15
kroE100 690.35 576.97 − 16.42 581.86 − 15.71 69.41 651.31 − 5.65 659.48 − 4.47 48.57
lin105 420.60 377.95 − 10.14 381.69 − 9.25 90.89 378.25 − 10.07 380.43 − 9.55 40.27
pr107 1222.50 1032.64 − 15.53 1038.11 − 15.08 79.05 1204.42 − 1.48 1224.35 0.15 32.54
pr124 1687.25 1615.88 − 4.23 1618.10 − 4.10 46.67 1653.80 − 1.98 1996.62 18.34 25.45
pr136 2762.00 2397.25 − 13.21 2474.30 − 10.42 141.84 2642.00 − 4.34 2789.00 0.98 44.50
pr144 1688.75 1675.75 − 0.77 1675.75 − 0.77 175.92 1666.25 − 1.33 1675.75 − 0.77 43.33
pr152 2123.95 1969.80 − 7.26 1973.67 − 7.08 119.13 2114.04 − 0.47 2128.53 0.22 61.29
rat99 37.45 37.45 0.00 37.45 0.00 54.71 37.15 − 0.80 37.33 − 0.32 35.37
rat195 71.50 71.50 0.00 71.50 0.00 168.83 71.40 − 0.14 71.93 0.60 44.89
rd100 246.70 217.00 − 12.04 219.42 − 11.06 85.14 240.46 − 2.53 243.84 − 1.16 33.87
st70 21.00 21.00 0.00 21.00 0.00 21.91 20.50 − 2.38 21.00 0.00 3.85
Average − 9.19 − 8.65 92.28 − 4.46 − 3.69 37.58
123
Q. M. Ha et al.
approximately 1.5 times slower than GRASP but can still deliver better solutions in
less than 1 min and 5 min for 50- and 100-customer instances, respectively.
We report the comparison between HGA and HGVNS proposed by Freitas and Penna
(2018) on Instance Set 3 under Min-time objective in Table 4. In this table, Column
“T S P ∗ ” is the optimal TSP value obtained by Concorde (Applegate et al. 2006),
Columns “HGA” and “H G A”, respectively, are the best and average results among 10
repeated runs. Columns “Gap H G A ” and “Gap H G A ” represent the best and average
gaps between HGA and optimal TSP while Column “TH G A ” reports the average
running time of HGA. Similarly, Columns “HGVNS”, “H GV N S”, “Gap H GV N S ”,
“Gap H GV N S ” and TH GV N S represent the best and average value and gap as well as
the running time of HGVNS.
In overall, among 24 instances, HGA can improve existing best known solution
found by HGVNS in 16 instances, performs worse than HGVNS in 7 instances and
a draw in one instance (eil51). In average, HGA performs approximately 5% better
than HGVNS and up to 17.17% in kroC100. Regarding the computational cost, HGA
is about 2 times slower than HGVNS but the running time between two algorithms
is not fairly compared since they are run on two different machine configurations
(HGVNS runs on a faster machine Intel Core i7 Processor 3.6 GHz). In detail, HGA
mostly performs better than HGVNS on instances where solutions contain many drone
deliveries (such as the kro instances) while HGVNS, on the other hand, can find better
solutions among instances where solutions contain very few drone deliveries. This can
be explained by the fact that HGVNS explores the search space from the optimal TSP
solutions, which are very closed to the final TSP-D solutions.
We evaluate the performance of HGA when using our proposed crossover versus
four classical crossovers (Potvin 1996) (OX, PMX, OBX, and PBX) in Table 5 under
two objectives with Instance Set 2 mentioned in the above section. Again, HGA was
repeatedly run 10 times for each choice of crossover, and we have conducted 6000 tests
in total. For each crossover, we report the average percentage gap with the best found
solution (regardless of crossover), the run time in minutes (Column “T ”), the standard
deviation (Column “sd”) and the geometric mean value (row “Mean”). Furthermore,
a comparison of the convergence of these crossovers in both objectives is presented in
Figs. 7 and 8, where the Y-axis shows the averaged percentage gap with the best found
solutions, and the X-axis contains the maximum number of iterations over which an
improvement could be made.
Overall, DX delivers the best value among other crossovers in terms of percentage
gap. For min-cost, DX is approximately 18%, 5.7%, 283%, and 16.5% better than OX,
PMX, OBX, and PBX, respectively. For min-time, that superiority is approximately
26.5%, 10.2%, 283%, and 46.9%. As can be seen, OBX performs worst among the
crossovers, possibly due to its design, for which only a random number of separated
123
A hybrid genetic algorithm for the traveling salesman…
Gap T (min) sd Gap T (min) sd Gap T (min) sd Gap T (min) sd Gap T (min) sd
Min-cost 1.39 1.37 0.86 1.64 1.44 0.87 1.47 1.31 0.92 5.33 0.87 1.28 1.62 1.53 0.95
Min-time 0.49 1.55 0.33 0.62 1.50 0.41 0.54 1.48 0.30 1.88 0.96 0.51 0.72 1.73 0.40
nodes is copied from the parent. This causes the OBX to have a smaller chance of
transmitting “good” materials from its parent such as good drone deliveries or good,
complete truck deliveries. The performances of OX and PMX, on the other hand, were
much closer to those of DX, especially for PMX in the min-cost problem, being only
5.7% inferior. This result is because OX and PMX are both designed to copy a random
subsequence of the parent to the children, thus having a high chance of transmitting
“good” materials such as complete drone or truck deliveries from parent to offspring.
With respect to run time, OBX performs nearly 1.5–2 times faster than other
crossovers. However, due to its poor performance, this fast run time is not valuable.
Other crossovers deliver similar run times—less than 2 min on average—which is an
acceptable value.
When considering standard deviation, DX, OX, PMX and PBX perform stably, the
values of which are mostly less than 0.5% and no more than 1%, while OBX shows
its instability in delivering values that are more than 0.5% and up to nearly 1.3%.
Finally, from Figs. 7 and 8, we can see a similar pattern in the convergences of all
the crossovers. They all converge quickly in the first 5000 iterations.
123
Q. M. Ha et al.
This section provides analyses, as shown in Table 6, of the impact of the key compo-
nents of HGA as based on the measurement of percentage gap on average of solutions
over 10 runs to the best known solutions (BKS). The investigated components are
the restore method, repair mechanism, relaxation choice, infeasibility of solutions,
and diversity contribution. We adapted the standard setting (crossover DX is used
with parameters mentioned at the beginning of Sect. 4) and modified each of the key
components to test their impact. In detail, we have the following.
– No INF Instead of relaxing the endurance constraint on truck and drone travel
times, we insist that it hold. Therefore, no infeasible solution is allowed.
– No DIV We do not count the diversity contribution (setting it to 0) during the
calculation of biased fitness.
– No REPAIR We do not use a repair method in HGA.
– No RESTORE We do not use a restore method in HGA.
– RelaxTruck We only allow for infeasible solutions in which the endurance con-
straint is violated by truck travel times but not the drone’s time.
– RelaxDrone We only allow for infeasible solutions in which the endurance con-
straint is violated by the drone’s travel times but not the truck’s time.
The experiment results show that HGA is indeed sensitive to its parameters (infeasi-
bility, diversity contribution, repair, and restore method) in such a way that any change
to those values negatively impact the algorithm’s performance. However, those nega-
tive changes do not share the same impact. In detail, eliminating the role of the restore
method (No RESTORE) strongly reduces the performance of HGA, which proves
the necessity of this problem-specific component to the general framework in order to
efficiently solve the TSP-D problem.
123
A hybrid genetic algorithm for the traveling salesman…
123
Q. M. Ha et al.
5 Conclusion
Acknowledgements This research is funded by Vietnam National Foundation for Science and Technology
Development (NAFOSTED) under Grant No. 102.99-2016.21. The authors would like to thank the anony-
mous reviewers for the valuable comments that helped to considerably improve the quality of this work.
We also express our thanks to Júlia Cária de Freitas and Professor Puca Huachi Vaz Penna for sending us
the instance files so that we could conduct the comparison with the HGVNS algorithm.
References
Agatz, N., Bouman, P., Schmidt, M.: Optimization approaches for the traveling salesman problem with
drone. Trans. Sci. 52(4), 965–981 (2018)
Applegate, D., Bixby, R., Chvatal, V., Cook, W.: Concorde TSP solver (2006)
Bouman, P., Agatz, N., Schmidt, M.: Dynamic programming approaches for the traveling salesman problem
with drone. Networks 72(4), 528–542 (2018)
Bulhões, T., Hà, M.H., Martinelli, R., Vidal, T.: The vehicle routing problem with service level constraints.
Eur. J. Oper. Res. 265(2), 544–558 (2018)
de Freitas, J.C., Penna, P.H.V.: A variable neighborhood search for flying sidekick traveling salesman
problem. Int.Trans. Oper. Res.(2018)
Glover, F., Hao, J.K.: The case for strategic oscillation. Ann. Oper. Res. 183(1), 163–173 (2011)
Ha, Q.M., Deville, Y., Pham, Q.D., Hà, M.H.: On the min-cost traveling salesman problem with drone.
Trans. Res. Part C Emerg. Technol. 86, 597–621 (2018)
Murray, C.C., Chu, A.G.: The flying sidekick traveling salesman problem: optimization of drone-assisted
parcel delivery. Trans. Res. Part C Emer. Technol. 54, 86–109 (2015)
Otto, A., Agatz, N., Campbell, J., Golden, B., Pesch, E.: Optimization approaches for civil applications of
unmanned aerial vehicles (UAVs) or aerial drones: a survey. Networks 72(4), 411–458 (2018)
Poikonen, S., Wang, X., Golden, B.: The vehicle routing problem with drones: extended models and con-
nections. Networks 70(1), 34–43 (2017)
Ponza, A.: Optimization of drone-assisted parcel delivery. Master’s thesis, University of Padova, Italy (2016)
Potvin, J.Y.: Genetic algorithms for the traveling salesman problem. Ann. Oper. Res. 63(3), 337–370 (1996)
Toth, P., Vigo, D.: The granular tabu search and its application to the vehicle-routing problem. Inf. J. Comput.
15(4), 333–346 (2003)
Vidal, T., Crainic, T.G., Gendreau, M., Lahrichi, N., Rei, W.: A hybrid genetic algorithm for multidepot
and periodic vehicle routing problems. Oper. Res. 60(3), 611–624 (2012)
Vidal, T., Crainic, T.G., Gendreau, M., Prins, C.: A hybrid genetic algorithm with adaptive diversity man-
agement for a large class of vehicle routing problems with time-windows. Comput. Oper. Res. 40(1),
475–489 (2013)
Vidal, T., Crainic, T.G., Gendreau, M., Prins, C.: A unified solution framework for multi-attribute vehicle
routing problems. Eur. J. Oper. Res. 234(3), 658–673 (2014)
123
A hybrid genetic algorithm for the traveling salesman…
Wang, X., Poikonen, S., Golden, B.: The vehicle routing problem with drones: several worst-case results.
Optim. Lett. 11(4), 679–697 (2017)
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.
123