Traveling Salesman Problem With A Drone Station

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

42 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO.

1, JANUARY 2019

Traveling Salesman Problem With a Drone Station


Sungwoo Kim and Ilkyeong Moon

Abstract—The importance of drone delivery services is TABLE I


increasing. However, the operational aspects of drone delivery C OMPARISON OF T RUCKS AND D RONES
services have not been studied extensively. Specifically, with
respect to truck-drone systems, researchers have not given suf-
ficient attention to drone facilities because of the limited drone
flight range around a distribution center. In this paper, we pro-
pose a truck-drone system to overcome the flight-range limitation.
We define a drone station as the facility where drones and
charging devices are stored, usually far away from the pack-
age distribution center. The traveling salesman problem with
a drone station (TSP-DS) is developed based on mixed integer
programming. Fundamental features of the TSP-DS are ana- range. To overcome these limitations, drone and truck delivery
lyzed and route distortion is defined. We show that the model services can be used such that the characteristics of one com-
can be divided into independent traveling salesman and parallel plement the other. Table I summarizes comparison of trucks
identical machine scheduling problems for which we derive two and drones [3], [4]. To demonstrate the combined means of
solution approaches. Computational experiments with randomly
generated instances show the characteristics of the TSP-DS and delivery, the HorseFly team at the University of Cincinnati
suggest that our decomposition approaches effectively deal with developed a system in which a drone can attach to and launch
TSP-DS complexity problems. from a truck [4].
Index Terms—Drone delivery, drone station, mixed integer Drones seem a good logistic alternative for industries, but
programming, truck-drone service. the technology needs further development to overcome some
realistic problems. Aside from controlling a certain type of
drones [5] or motion sensing issues [6], battery capacity is a
I. I NTRODUCTION main concern for drone utilization. As many distribution cen-
ROWING e-commerce and m-commerce increases the ters with drone facilities are located far from central cities,
G importance of efficient logistics. In 2013, Amazon
announced drone technology as a future logistic innovation,
relatively few customers are serviceable by drones. For this
reason, large retail companies such as Amazon strive to build
and many companies have invested into drone research. For more distribution centers near major cities, but the expenses
example, Amazon unveiled Amazon Prime Air, and Google of constructing distribution centers are still a huge obstacle to
announced Project Wing [1], [2]. Drones have many advan- completion. To deal with this logistical problem, a different
tages over the typical truck delivery system [3], [4]. As drones concept of drone facilities is proposed. Roblin [7] introduced
operate independently, they are free from operating labor costs Pylons Dronairports, which contain drone recharge and shelter
and have relatively unlimited working time. Further, they move devices. Designed by Bruni and Sardo, these compact devices
through the air and thus avoid the traffic congestion prob- can be easily installed any place. In addition, Amazon plans
lems of ground transportation. These advantages lead to the to use street lights and church steeples as drone docking sta-
highly energy-efficient use of drones. Moreover, the trans- tions [8]. Another problem is that the weight and volume
portation cost per kilometer is much lower than that of other capacities of drones are not enough to accommodate com-
means. However, because of technological limitations, a drone mercial delivery services [9]. Drone security is also affected
can carry only one parcel of limited weight and volume, by issues with GPS and sensor accuracies [10].
and it can deliver to a single customer within a short flight Because many researchers and companies have tried to over-
come these problems, some companies have been able to
Manuscript received May 17, 2018; accepted August 19, 2018. Date of utilize drones for commercial purposes. For instance, DHL
publication October 10, 2018; date of current version December 14, 2018.
This work was supported by the National Research Foundation of Korea grant express launched the first commercial delivery drone, called
funded by the Korea government (MSIT) under Grant 2017R1A2B2007812. Parcelcopter, in 2014 [11], [12], and the plan for building the
This paper was recommended by Associate Editor X. Xu. (Corresponding first airport for drones is ongoing in Rwanda [13]. In con-
author: Ilkyeong Moon.)
S. Kim is with the H. Milton Stewart School of Industrial and Systems trast, research on the operational aspects of drone delivery has
Engineering, Georgia Institute of Technology College of Engineering, Atlanta, been neglected, and only a handful of papers in drone-truck
GA 30332 USA. systems have been presented. One of the initial papers about
I. Moon is with the Department of Industrial Engineering, Seoul National
University, Seoul 08826, South Korea, and also with the Institute for Industrial the traveling salesman problem (TSP) in tandem with drones
Systems Innovation, Seoul National University, Seoul 08826, South Korea was conducted by Murray and Chu [14], who described two
(e-mail: ikmoon@snu.ac.kr). different models. The flying sidekick TSP (FSTSP) describes
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. the way a single drone is used with a truck. A drone is attached
Digital Object Identifier 10.1109/TSMC.2018.2867496 onto the truck, and a truck driver launches the drone and also
2168-2216 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
KIM AND MOON: TSP-DS 43

Applegate et al. [15] and a review of multiple TSP (MTSP)


problems was written by Bektas [16]. Other excellent
overviews of the vehicle routing problem were provided by
Golden et al. [17] and Toth and Vigo [18]. The proposed
model is also related to the PMS. Allahverdi et al. [19],
Ruiz and Vázquez-Rodríguez [20], and Baker and
Trietsch [21] summarized studies of the PMS. As tradi-
tional studies do not exploit drones, we concentrate on the
drone-truck models to which they are directly related.
A drone station can operate drones after a truck arrives
(a) (b)
and supplies parcels. This characteristic is closely related to
Fig. 1. Comparison of the (a) PDSTSP and (b) TSP-DS (red circle: drone- the PMS with precedence constraints. Tanaka and Sato [22]
serviceable customer, white circle: truck-only customer, and blue circle: a studied a single machine scheduling problem with precedence
drone station). constraints. The objective was to minimize total job comple-
tion time, and job idle time was not permitted. A successive
retrieves it. The other model is the parallel drone scheduling sublimation dynamic programming method was applied to find
TSP (PDSTSP) and is the key reference for this paper. Unlike the exact solution. Bilyk et al. [23] defined a batch schedul-
the FSTSP, the PDSTSP can utilize a sufficiently large num- ing problem with precedence constraints. Identical machines
ber of drones. However, drones deliver parcels only within the were assumed, and ready time for each job was considered.
flight range of the distribution center such that problems arise A variable neighborhood search and a greedy randomized
when the distribution center is far away from a majority of adaptive search procedure were applied to solve the problem.
customers. Davari et al. [24] solved a single machine scheduling problem
To overcome the limitations of the PDSTSP, we developed with time windows and precedence constraints. A branch-
the TSP with a drone station (TSP-DS), through which we and-bound algorithm was proposed to solve the problem.
exploit a drone station, defined as a facility that stores drone Hassan et al. [25] studied a PMS with precedence con-
and charging devices. The station is ready to launch drones straints to minimize the makespan. Three valid inequalities
that is, it is “activated” after a truck supplies parcels for drone were proposed, and their strengths were checked by compu-
delivery. We assume that the station can furnish a sufficiently tational experiments. Nicosia and Pacifici [26] addressed a
large number of drones and that the location of the station does multiple machine scheduling problem with precedence con-
not depend on that of the distribution center. Specifically, the straints. A heuristic method related to the bin packing problem
drone station is located near customer areas and away from was developed, and a lower bound was proposed. Because tra-
the distribution center. The facility can deliver parcels using ditional studies did not exploit drones, we concentrate on the
drones after a truck supplies the deliverables to the drone sta- drone-truck models in this paper.
tion, and a truck and a drone station operate independent of Murray and Chu [14] offered one of the earliest studies of
the distribution center after the truck supplies parcels for drone truck-drone delivery problems and introduced two fundamen-
delivery. Fig. 1 depicts the difference between the PDSTSP tal models. First, the PDSTSP describes a drone facility within
and the TSP-DS. a distribution center. To our knowledge, it is the only model in
We first analyze the fundamental features of the TSP-DS. which a drone facility is considered in truck-drone problems.
We define route distortion, and the lower bound of the num- A sufficiently large number of drones can be utilized at the dis-
ber of drones to eliminate route distortion was presented. By tribution center, but the limited flight range creates practical
applying the assumptions of the sufficient number of drones issues. To alleviate this problem, the FSTSP was developed
and by considering the distance between the distribution center to describe a truck driver launching and retrieving a drone.
and a drone station is far enough, we show that the TSP-DS This model overcomes the flight range limitation from the
can be divided into the TSP and the parallel identical machine distribution center of the PDSTSP, but it only applies to a
scheduling problem (PMS). Through this approach, we suc- single drone. Our research is directly related to the PDSTSP
cessfully reduce the complexity of the problem, and obtain the and serves as a complementary model applicable to a drone
exact solution. In addition, we explain the tradeoff between the facility separated from the distribution center. To solve the
TSP-DS and PDSTSP. PDSTSP, Murray and Chu [14] developed a heuristic method
The remainder of this paper is composed as follows. based on decomposition of the model into the TSP and PMS.
Section II introduces previous research related to truck-drone We also used the similar decomposition approach; however,
systems. Section III describes the TSP-DS. Fundamental fea- our approach focused on the conditions on the decomposition
tures of the TSP-DS is presented in Section IV. Section V which guarantees the optimal solution.
shows the analyses of computational results and discussion of Although we take into account a drone facility problem with
several issues, and Section VI presents conclusions. a truck TSP, a majority of research has concentrated on truck-
launch delivery problems, which are intricately related to the
II. L ITERATURE R EVIEW FSTSP. In related studies, Agatz et al. [3] assumed that drones
The TSP-DS is one variation of the TSP and the vehicle and a truck share the same road network, which allowed them
routing problem. A recent review of the TSP was offered by to find the worst-case approximation ratios for the heuristics.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
44 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 1, JANUARY 2019

However, the assumption fails to take advantage of the drones a drone station which is relatively cheap and easy to install
capacity to freely move off truck paths and remain unaffected was defined. We need to emphasize that our study is focused
by road conditions. on a more general work of the PDSTSP model which over-
Ha et al. [27] introduced the TSP with a drone. They comes the limited usage of drones; TSP-DS can be reduced
assumed that launching and retrieving a drone is impossible at to the PDSTSP when the locations of the distribution center
the same customer node. The mathematical formulation and and drone station are the same.
two heuristic algorithms were developed. Mathew et al. [28] A drone station can store and utilize a sufficiently large
described the heterogeneous delivery problem by considering number of drones that deliver drone-fitting parcels with a lim-
a team using a truck and drones with complementary capabili- ited flight range. A large number of drones seems to be vague,
ties based on the assumption that drone-serviceable customers therefore, we present the lower bound of the number of drones
can only receive deliveries by drones. The problem can be which guarantees the minimum makespan of the total deliv-
reduced to the generalized traveling problem, which can be ery time in a latter section. After a truck arrives at the station,
solved with many heuristics methods. In addition, they defined drone-fitting parcels are processed for drone delivery and the
the multiple warehouse delivery problem by showing a special station is said to be activated. We assume that the location
case of the heterogeneous delivery problem and developing of the station is relatively far from the distribution center; a
two heuristic approaches. Ferrandez et al. [29] compared the drone station is farther than the maximum flight distance of a
overall travel times and energy consumption of truck-only and drone launched from the distribution center. If a drone station
truck-drone tandem deliveries. They proposed a clustering-first is located to near a distribution center, there are not mean-
and routing-second approach. K-means algorithm, used to find ingful differences between the TSP-DS and PDSTSP. In this
an efficient launch location of drones, and genetic algorithms case, a decision maker does not operate a drone station in that
were applied to solve a truck-routing problem. operating an independent facility needs additional costs. If a
We introduce several studies not directly related to a truck- drone station is far away from a city center, the decision maker
drone delivery service; however, these works show solutions also does not operate the station because of the same reason.
to drone problems. Boone et al. [30] introduced the MTSP That is why we assumed that the location of a drone station is
which can be applied to the drone swarm route plan. They far away from the distribution center but near the city center.
divided the MTSP into two components: 1) clustering and Although the decision where to build a drone station can be an
2) TSP problems. The K-means clustering method was applied important issue, the location of the drone station is assumed
to divide cities into multiple clusters, and each drone was to be given. The reason for this assumption is that the location
allocated to each cluster. A constructive heuristic approach, problem should be solved based on the long-term perspective
called 2-Opt, was applied to solve the TSP in each cluster. while our topic mainly focuses on the daily delivery service.
This approach helps to reduce significant computation time. Because of safety and weight issues, a single drone can-
Dorling et al. [31] developed the vehicle routing problem not carry multiple parcels. Therefore, a drone visits only one
for drone delivery services by deriving an approximated customer per sortie while a truck can visit multiple customers
and linearized cost function that accounts for the energy in one trip. In addition, some customers order products that
consumption model of multiple drones and by developing exceed the volume and weight capacities of drones. The lim-
mixed integer-based programming for the problem. Further, ited flight range is due to the capacity of drone batteries.
Dorling et al. [31] built a string-based simulated annealing We assume that the travel time of vehicles are proportional
heuristic. A drone system in an indoor environment was intro- to distances and drones are faster than a truck because the
duced by Khosiawan and Nielsen [32]. The system focused on drones cross air space and the truck must follow ground routes.
a scheduling issue, and a system architecture for drone appli- Because charged batteries are supplied from a drone station,
cations in an indoor environment was developed. Furthermore, battery charging times for returned drones are not considered.
a framework of scheduler component was presented. A truck or a drone delivers an order only once to a customer.
Travel times between nodes were assumed to be symmetric.
The truck departs from the distribution center and returns to it
III. T RAVELING S ALESMAN P ROBLEM W ITH after packages are delivered. Likewise, drones return to the sta-
D RONE S TATION tion after delivering parcels. The delivery service is considered
The TSP-DS is an extension of the PDSTSP, with the major ended when a truck returns to the distribution center and all
difference in the location of off-duty drones. In the TSP-DS, drones return to their drone station. We defined the last deliv-
drones are stored in and launched from a drone station, not ery time as the time to finish the total delivery service. The
the package distribution center. In the previous research con- objective of the TSP-DS is to minimize the last delivery time.
ducted by Murray and Chu [14], it was assumed drones can
only be used at a distribution center and the number of drones
are infinite. However, in many cases, distribution centers are A. Notation
located far from a city center where a majority of customers We regard each customer as a single node and make a
are located. Constructing additional distribution centers might network with N = {1, . . . , c} as a node set of customers and
be a solution to handle this problem, but it is hard to be s ∈ N as a drone station node index. In a customer network,
realized because of enormous costs. The motivation of this we add the distribution center node. We define 0 as the index
paper was to relax the first assumption, and the concept of of the distribution center, and to avoid symmetric problems,

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
KIM AND MOON: TSP-DS 45

we define c + 1 as the index of the distribution center node Constraints (7)–(9) specify the flow of the truck.
for returns. We also define origin set N0 = {0, 1, . . . , c} and Constraint (7) means that a single truck leaves the distribution
destination set N1 = {1, 2, . . . , c + 1}. Multiple drones are center and Constraint (8) means that the truck must return to
located in a drone station, and a set of drones is defined as the distribution center. Constraint (9) ensures that the truck
V = {1, 2, . . . , v}. leaves customer j ∈ N to deliver parcels after it arrives to
Customers are sorted by their package information. Weights, customer node j ∈ N from customer node i ∈ N0
volumes, and distances from the drone station are considered 
ui − uj + 1 ≤ (c + 2) · 1 − xi,j ∀i ∈ N, j ∈ {N1 : j = i}
to distinguish drone-serviceable customers. We define D as a
set of drone-serviceable customers, which is a subset of N. (10)
The travel time of a truck between a pair of nodes (i, j) (∀i ∈ 1 ≤ ui ≤ c + 2 ∀i ∈ N1 . (11)
N0 , j ∈ N1 ) is defined as τi,j and that of the drones is defined as
Subtours of the truck are eliminated by
d (∀i ∈ N , j ∈ N ). The binary decision variable x equals 1
τi,j 0 1 i,j
Constraints (10) and (11)
if the truck travels from node i ∈ N0 to node j ∈ {N1 : j = i};
it is 0 otherwise. Similarly, the decision variable xsi,j is defined xi,j ∈ {0, 1} ∀i ∈ N0 , j ∈ {N1 : j = i} (12)
for the route of a truck until it arrives at a drone station. The
binary decision variable yi,v is 1 if customer i ∈ D is served
Ê
xsi,j ∈ + ∀i ∈ N0 , j ∈ {N1 : j = i} (13)
yi,v ∈ {0, 1} ∀i ∈ D, v ∈ V (14)
by drone v ∈ V launched from a drone station. Variable z
refers to the last possible delivery time of a truck and drones. ui ∈ Ê+ ∀i ∈ N1 . (15)
ui indicates the position of node i ∈ N1 in the truck’s path. Constraints (12)–(15) define the decision variables.
Although the assumptions on the number of drones and
B. Mathematical Formulation location of the station are relaxed, the proposed mathemati-
We can formulate the TSP-DS as follows: cal formulation still works to the relaxed problem. In a latter
section, we address that this formulation is decomposed to
Minimize z (1)
  reduce the complexity by using the assumptions.
subject to xi,j + yj,v = 1 ∀j ∈ N (2)
i∈N0 v∈V IV. F UNDAMENTAL F EATURES OF TSP-DS
i =j j∈D
xi,j ≤ xi,j ∀i ∈
s
N0 , j ∈ N1 . (3) In the TSP-DS, a loaded truck reaches a drone station, and
activates the drone delivery process. This station activation
The objective function (1) minimizes the delivery time of a condition has important features, and we demonstrate the main
truck and drones. Constraint (2) suggests that neither a truck characteristics of the TSP-DS in this section.
nor a drone can deliver the parcel to a customer more than Proposition 1: The activation time of a drone station is
once. Constraint (3) ensures that xsi,j follows the path of xi,j : always less than or equal to z/2.
⎧ Proof: The travel time matrix of a truck is symmetric, and
  ⎨ 1, if i = 0 the total distance of a truck does not change when the travel
xsi,j − xsj,i = −1, if i = s ∀i ∈ N0 ∪ {c + 1}. direction of the truck is reversed on the route. For this reason,

j∈N2 j∈N2 0, otherwise. when the activation time of a drone station is greater than z/2,
(4) a truck can be chosen to the same travel route with the reverse
Constraint (4) restricts the route of a truck until it arrives direction which activates the drone station before z/2.
at a drone station
   A. Routes Distortion
z≥ τi,j · xsi,j + τs,i
d
+ τi,s
d
· yi,v ∀v ∈ V (5) Generally, a drone station is used to maximize the use of
i∈N0 j∈N1 i∈D drones, and a truck is used on the shortest routes. However,

z≥ τi,j · xi,j . (6) in some cases, a truck driver takes a longer route to activate a
i∈N0 j∈N1 drone station earlier which results in the overall reduction in
i =j the objective function. Because the proposed model searches
Constraint (5) imposes the criterion that z is greater than or the optimal schedule of the global truck-drone system, in
equal to the last delivery time of drone v ∈ V launched from which a drone station and a truck interact, we define this
a drone station. Constraint (6) restricts that z should not less case as route distortion. In analyzing the fundamental features
than the last delivery time of a truck of the drone-truck system, we do not take into consideration
 two assumptions: the sufficiently large number of drones in
x0,j = 1 (7) a drone station and the minimum distance between the distri-
j∈N1 bution center and the drone station. However, we assume that

xi,c+1 = 1 (8) the number of drones are less than that of customers.
i∈N0 There are two types of the route distortions. In one, a
  truck delivers parcels to customers who can be serviced
xi,j = xj,k ∀j ∈ N. (9)
by drones. This happens when the last delivery time of a
i∈N0 k∈N1
i =j k =j drone is later than that of a truck. In this case, use of

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
46 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 1, JANUARY 2019

(a) (b) (a) (b)

(c) (c)

Fig. 2. Example of disadvantage of using drones to deliver parcels to Fig. 3. Routes of a truck can be influenced by a drone station.
(a) Optimal schedule of a truck. (b) Optimal schedule of a truck and a drone.
all drone-available customers. (a) Drone station delivering parcels to all
drone-serviceable customers. (b) Optimal schedule of a truck and a drone. (c) Comparison of schedules between (a) and (b).
(c) Comparison of schedules between (a) and (b).

a truck to deliver to drone-serviceable customers is more


efficient (Fig. 2). Fig. 2(a) shows a drone delivering parcels to
serviceable Customers 3–5. However, the last delivery time of
a drone delivery is later than that of the truck delivery. In con-
trast, Fig. 2(b) shows that a truck delivers parcels to Customer
3 and alleviates the burden on the drone station which results (a) (b)
in an earlier delivery service time.
In the other route distortion case, a truck uses a long deliv-
ery route to arrive at a drone station early. A driver would
make this decision because drones are only able to deliver
to customers after a truck supplies parcels to the station, and
thus, an early activation time means an early delivery time
by drones (Fig. 3). For example, in the case of Fig. 3(a), the
shortest route of a truck is 0-1-6-2-0. However, as shown in
Fig. 3(b), a truck using longer route 0-6-1-2-0 can activate
a drone station so that the drones can start delivering earlier. (c)
The duration of delivery times from a drone station is indepen-
dent of activation time, but the sooner the station is activated
by the truck, the earlier the drones can finish deliveries. The
last delivery shown in Fig. 3(b) is earlier than that shown in
Fig. 3(a).
When N0 is given, three factors affect route distortions. The
main factors correspond to the number of drones in a station as
well as the velocities and flight ranges of drones (the number
of drone-serviceable customers) (Fig. 4). When the number of (d)
drones increases, a truck takes the shorter routes. In the case of
Fig. 4(a), the last delivery time from a drone station is earlier Fig. 4. Three factors can affect the routes. (a) Optimal schedule of a truck
and two drones. (b) Optimal schedule of a truck and a high-speed drone.
than that of the shortest truck routes; further, early activation (c) Optimal schedule of a truck and a drone with a smaller flight range.
of a drone station is unnecessary when many drones are avail- (d) Comparison of schedules between (a), (b), and (c).
able. Likewise, faster drones affect the best route choices for a
truck, as shown in Fig. 4(b). Decreasing the flight range or the
number of potential drone-serviceable customers also offers a drone-serviceable customer. As a result, the last delivery time
the same result that a drone station needs not to be activated from a drone station is earlier than that of a truck from a dis-
in the early stage of delivery service. In the case of Fig. 4(c), tribution center. These examples show that the factors related
the flight range is decreased and Customer 5 is not considered to the drone station workload affect the truck route.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
KIM AND MOON: TSP-DS 47

B. Conditions for Elimination of Route Distortion (R is the diameter of the flight range). As cmax does not
Based on the assumption that a sufficient number of drones exceed the radius of the flight range, the following inequal-
is available in a station, we can draw the inequality that ity holds: τ0,s ≥ R ≥ 2 · τs,cmax . As the drone velocity
eliminates the route distortion. is the same or exceeds the speed of a truck, our problem
Proposition 2: Let cmax be the farthest drone-serviceable always satisfies Proposition 2. we also assume that a drone
customer from the drone station s and α be the travel rate of station can utilize a sufficiently large number of drones,
the drone speed to the truck speed. If the number of drones is and this assumption satisfies Proposition 3. Therefore, our
sufficient and the problem satisfies τs,0 ≥ 2 · τs,cmax /α, drones problem fulfils the elimination condition of route distortion
can finish parcel deliveries to all drone-serviceable customers (Corollary 1).
before the truck returns to the distribution center. When the problem satisfies conditions for Corollary 1, a
Proof: When the number of drones in a station is sufficient, drone station can successfully initiate delivery of all drone-
each drone can deliver a parcel to a single customer. In this compatible parcels, and a truck does not need to deliver parcels
case, the upper bound for the flight time of a drone from the to any customer serviceable by drones. Because the route dis-
station (UBd ) is the delivery time of a drone to cmax . As the tortion was eliminated, the model can be divided into two
travel time matrix of a truck is symmetric, τcmax ,s = τs,cmax and independent problems. The first problem is the TSP through
UBd = {τcmax ,s + τs,cmax }/α = 2 · τs,cmax /α. The lower bound which one finds the shortest truck routes by considering only
of the truck travel time LBt to return to the distribution center customers who cannot be serviced by drones. The second
after leaving a drone station s is τs,0 . Therefore, if UBd is less problem finds the drone station schedule that minimizes the
than LBt , the last delivery time from a drone station s can be last delivery time using drones. Because the objective value of
earlier than or the same as the delivery time of the truck. the second problem is always less than or equal to the objective
In the real world, a sufficiently large number of drones value of the first problem (Corollary 1), these two independent
is not needed, and the number of customers is the logical problems successfully solve the TSP-DS. We define these two
upper bound for drone inventory. However, when many drones problems as an independent traveling salesman and parallel
are needed, and although we cannot find the minimum num- machine scheduling problem (TSPMS).
ber before solving the problem, we can find the bound that However, in terms of a drone station schedule, the PMS
likely allows for a sufficient number of drones for delivery model can suggest an overuse of drones because the model
services. is not designed to minimize them. Furthermore, it does not
exploit the information from the solution of the TSP which

Proposition
3: If the number of drones is
provides the arrival time of a truck at the drone station. For this
|N|/ [(α · τ0,s )/(2 · τcmax ,s )] , additional drones are
not necessary to shorten the schedule. reason, a two-stage traveling salesman and modified parallel
Proof: In Proposition 2, UBd = 2 · τs,cmax /α, and LBt is machine scheduling problem (TSMPMS) is developed to find
τs,0 . Therefore, the lower bound of the maximum number a schedule that minimizes the number of drones used at a
of customers to which a drone can deliver before a truck station by exploiting the solution of the TSP to set the drone
station schedule. The first stage is the same as the ordinary
returns to the distribution center is [(α · τ0,s )/(2 · τcmax ,s )] .
The number of is |N|, and thus, TSP. After the TSP is solved, the activation time of a drone

customers
the required number
station as and the last delivery time of a truck zt are known. As
of drones is |N|/ [(α · τ0,s )/(2 · τcmax ,s )] .
Combining Propositions 2 and 3, we can define the follow- the problem satisfies Corollary 1, z = zt and the last delivery
ing general condition. time of a drone station can be earlier or the same as z. This
finding means the upper bound of the drone flight time UBf is

Corollary
1: If the number of drones is more than z − as . Reflecting this information, a modified PMS problem
|N|/ [(α · τ0,s )/(2 · τcmax ,s )] and the problem satisfies
τs,0 ≥ 2 · τs,cmax /α, then the route distortion is eliminated. is solved to minimize the number of drones used under the
upper bound of the flight time. The process to calculate UBf
is described in Algorithm 1.
C. Decomposition of TSP-DS The start node (start_node) is initialized as 0 node. The
The TSP-DS is an NP-hard problem, and the typical math- activation time of a drone at a station (active_time) and the
ematical formulation can be solved very limited size of upper bound of the flight time (UBf ) are set as 0. The algo-
instances; it was hard to solve problems with more than rithm finds the next node from the start node. When the next
11 nodes of instances. One of our motivation is to reduce node j is found, the activation time and the new start node
the complexity. By analyzing the mathematical structure of is updated. The algorithm repeats until the new start node is
the TSP-DS, we found that there are special characteris- c. After the activation time is fully updated, the upper bound
tics of the mathematical formulation and exploited them of the flight time is calculated. Because the problem satis-
to derive decomposition methods which guarantee optimal fies Proposition 1, UBf can be always greater than or equal
solutions. to z/2.
For our problem, we address the situation in which the After UBf is calculated, we can find the schedule of a drone
majority of customers are located far from the distribution station that utilizes the minimum number of drones without
center and the maximum flight distance of a drone from changing the last delivery time. We define a new binary vari-
the distribution center is less than the distance between the able dv ; it is 1 if drone v ∈ V is used for the delivery and 0
drone station and the distribution center. It means τ0,s ≥ R otherwise. The mathematical formulation of the modified PMS

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
48 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 1, JANUARY 2019

Algorithm 1 Algorithm for UBf


Initialization : start_node, arrival_station, UBf = 0;
While{i ∈ N0 − Nd }
{
While{j ∈ N0 − Nd }
{
if(xstart_node,j = 1) then
arrival_station + = τstart_node,j
start_node = j
break (a) (b)
end-if
}
if(start_node = c) then break;
}
if(arrival_station > z/2) then
UBf = arrival_station
else
UBf = z - arrival_station
end-if (c)
Output(UBf )
Fig. 5. Difference between the proposed model and PDSTSP. (a) Optimal
schedule of the TSP-DS. (b) Optimal schedule of the PDSTSP. (c) Comparison
of schedules between (a) and (b).
is as follows:

Minimize dv (16) Proposition 4: If the problem satisfies Corollary 1, the min-
v∈V
  imum number of drones in a drone station is, at most, three
subject to τs,i
d
+ τi,s
d
· yi,v ≤ UBf · dv ∀v ∈ V. (17) times larger than the number needed to satisfy the PDSTSP
i∈D when the sum of flight distances from a drone station and the
The objective function (16) minimizes the required number of last delivery time of the TSP-DS are the same as those of the
drones for delivery service. Constraint (17) suggests that dv is PDSTSP.
1 when drone v is used and the flight time of a drone so it Proof: Let zs be the last delivery time of the TSP-DS and
does not exceed the upper bound of the flight time zd be that of the PDSTSP. The lower bound of the last deliv-
 ery time of the TSP-DS is z. When the problem satisfies
yj,v = 1 ∀j ∈ D (18) Corollary 1, the route distortion does not exist. In addition,
v∈V we assume that the sum of flight distances from a drone sta-
yj,v ∈ {0, 1} ∀j ∈ D, ∀v ∈ V (19) tion and the last delivery time are the same as those from the
dv ∈ {0, 1} ∀v ∈ V. (20) distribution center. Therefore, zs = zd = z and z is the same
as the last delivery time of a truck. Because a drone station
Constraint (18) shows that each customer serviceable by needs to finish the delivery service before a truck returns to
drones should receive deliveries by a drone. Constraints (19) the distribution center, the lower bound of the delivery time
and (20) define decision variables. from a drone station is z/2. In other words, the truck can
arrive the drone station before or equal to z as mentioned in
D. Tradeoff Between TSP-DS and PDSTSP Proposition 1.
In the proposed model, the activation time of a drone station Let A be a drone in the PDSTSP, and list the delivery jobs
depends on the arrival time of a parcel-laden truck. However, of Drone A with the shortest process time order. Consider two
in the case of the PDSTSP, a drone station is located at the drones in the TSP-DS. Assign the odd index jobs of Drone A
distribution center, so drones can be used at time 0. In this to a drone in the TSP-DS and the even index jobs to the
regard, even when the last delivery time and the sum of flight other drone in the TSP-DS with the same order. The sum
times of the TSP-DS is the same as those of the PDSTSP, of flight distances of Drone A cannot exceed z; therefore, at
the required number of drones for the delivery service might most one drone’s last delivery time in the TSP-DS, defined as
differ (Fig. 5). For instance, according to Fig. 5(a), a drone Drone B, can exceed over z/2. Let c max is the last delivery job
starts the delivery service from a drone station after a truck of Drone B, then z/2 < the last delivery time of Drone B <
arrives. Therefore, an additional drone is needed in the TSP- z/2 + 2 · τs,c max . If the last delivery job of Drone B is allocated
DS so the last delivery time is no later than the time when a to a new drone, then the last delivery time of Drone B − 2 ·
truck returns to the distribution center. In contrast, Fig. 5(b) τs,c max < z/2 holds. It shows that jobs in Drone A can be
shows that when all drones are activated at time 0, they can divided into at most three segments in which each of them is
deliver the parcels to drone-serviceable customers before a less than or equal to z/2. Therefore, three drones are needed
truck finishes its delivery route. at most.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
KIM AND MOON: TSP-DS 49

TABLE II
AVERAGE C OMPUTATION T IMES (S ECONDS ) OF THE TSP-DS, TSPMS,
AND TSMPMS W ITH R ESPECT TO THE N UMBER OF C USTOMERS
IN THE E XPERIMENT R EGION

Fig. 6. Experimental design.

V. C OMPUTATIONAL E XPERIMENTS
Results of computational experiments and the insight of the
developed model are presented in this section. The models
were built in XPRESS-IVE 7.9 with the XPRESS-MP math-
ematical programming solver. Experiments were conducted
with an Intel Core i5-3570 CPU 3.4 GHz with 8.00 GB of
RAM in Windows 10.
According to Murray and Chu [14], the flight range of
a commercial drone is approximately 16 km (≈10 miles).
generated to evaluate the performances of the other models,
Therefore, we assumed that a circle with 16 km radius is a fea-
and the number of customers is increased from 20 to 80. In
sible flight region. To compare the PDSTSP and the TSP-DS,
each customer set, ten random instances were generated. The
we set two different flight areas. The feasible flight area from
travel rate α was fixed at 2. We stopped the experiment of each
drone station s is defined as Region A while that from the
model when it took over 18 000 s. The detailed information
distribution center is defined Region B. To avoid overlapping
of the experiments and results are shown in Table II. Fig. 7
feasible flight regions, we made a gap between them. As a
illustrates computation times for each data set with respect to
result, the experiments were conducted in a square region of
different models.
32 km · 65 km (Fig. 6).
Although three models give the same objective value, the
Due to the probabilistic nature of parcel ordering, customers
computation times are distinct between them. With fewer cus-
were assigned randomly to specific locations. Furthermore, to
tomers, the gap is small. However, the computation times of
concentrate on the effect of a drone station on delivery, we
the TSP-DS are much greater for the data set with more cus-
only consider small and light parcels that can be delivered
tomers. Although the computation time increases are relatively
by drones. For this reason, if customers are located in the
small, the TSPMS and TSMPMS models are also not free
flight-feasible region, they are assumed to be drone-serviceable
from increased computation times. The computation time dif-
customers. When we solve the TSP-DS, customers located
ference between the TSPMS and the TSMPMS was negligible
in Region A are classified as drone-serviceable customers
in the small problems. However, in the large problems, the
but others are considered truck-only customers. However,
TSMPMS was much faster than the TSPMS. The gap between
in the PDSTSP, customers in Region B can be serviced
the computations are increased according to the size of the
by drones while those in Region A cannot be serviced by
problem because the second-stage problem of the TSMPMS
drones. In addition to this, customer locations are restricted to
uses bounds derived from the first-stage model while the
Regions A and B.
TSPMS solves two problems independently.
The number of drones in a station was calculated using
the bound of drones needed to satisfy Proposition 3. In
detail, the radius of Region A is 16 km and the distance B. Comparison Between PDSTSP and TSP-DS
between the drone station and the distribution center is 33 km. As we assume that the TSP-DS is justifiable when customers
When the travel rate α is set as 2, a drone can deliver parcels live in city centers far away from the distribution center, we
to at least two customers before a truck at the drone sta- consider the case in which more than one-half of customers
tion returns to the distribution center. Therefore, the minimum are near drone station s. To analyze the characteristics men-
number of drones to satisfy the condition for Proposition 3 is tioned in Section IV-A, we conduct experiments by varying
no more than |N|/2 . the number of customers in Region A, the number of drones,
and travel rate α. The total number of customers is set at 10.
A. Computation Times In each case, ten experiments are conducted, and the savings
Two data sets are generated to evaluate the computation between the objective values of the TSP-DS (PDSTSP) and
times of the models. A small data set is used to compare the ordinary TSP is found. Each saving is calculated as [the
the performance between the TSP-DS and other models. Due objective value of the TSP—that of the TSP-DS (PDSTSP)]/
to the complexity characteristic of the TSP-DS, the number the objective value of the TSP. The detailed environment set-
of customers is increased from 7 to 11. A large data set is ting and results are shown in Table III. The results show that

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
50 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 1, JANUARY 2019

TABLE IV
AVERAGE G APS (%) OF THE O PTIMAL VALUE B ETWEEN THE L AST
D ELIVERY T IME OF THE TSP-T AND T HAT OF THE TSP-TC W ITH
R ESPECT TO THE N UMBER OF C USTOMERS IN R EGION A,
THE N UMBER OF D RONES , AND T RAVEL R ATES

(a)

The objective value of the TSP was much later than those
of the TSP-DS and PSDTSP, which justifies use of the truck-
drone system. Moreover, the objective value of the TSP-DS
(b) is lower than that of the PDSTSP, and the maximum saving
of the optimal value is increased according to the number of
Fig. 7. Comparison of the computation times of three models. customers in Region A. It strengthens our argument that uti-
(a) Computation times of a small set. (b) Computation times of a large set. lizing a drone station helps to make the last delivery time
earlier when the distribution center is far away from a major-
TABLE III ity of customers. Likewise, the increasing number of drones
AVERAGE S AVINGS (%) OF THE O PTIMAL VALUE B ETWEEN THE TSP-DS
(PDSTSP) AND THE TSP W ITH R ESPECT TO THE N UMBER OF or increased travel rate α enlarge the saving because releas-
C USTOMERS IN R EGION A, THE N UMBER OF D RONES , ing the burden of the drone alleviates the burden created by
AND T RAVEL R ATES
inefficient truck routes.
To analyze the impacts of the main factors, such as the
number of drones, customers serviceable by drones, and
travel rate α, on route distortion, the differences between the
last delivery time of the truck side in the TSP-DS (TSP-T)
and the last delivery time of the TSP with truck only cus-
tomers (TSP-TC) are considered. The gap is calculated as
(the objective value of the TSP-T—that of the TSP-TC)/(the
objective value of the TSP-T), and detailed information is
given in Table IV. Obviously, the last delivery time of the
TSP-T is larger than or equal to that of the TSP-TC when
the number of drones was not sufficient because of route
distortion.
If the same number of drones is used, the gap is smaller
when few customers are in Region A. When 6 customers are
in Region A and travel rate α is 1.5, two drones are sufficient
to avoid route distortion. However, more than three drones
are needed at the same travel rate to serve 9 customers to
avoid route distortion. It can be observed that the number of
drones in a station has significant impact on the truck route.
when the number of drones is increased from 1 to 3, the deliv- Because the generated examples satisfy the distance condition
ery rates are not appreciably changed and the route distortions of Proposition 3, increasing the number of drones corresponds
do not happen. to the shortened truck route.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
KIM AND MOON: TSP-DS 51

TABLE V
AVERAGE AND M AXIMUM N UMBER OF D RONES U SED IN A
Algorithm 2 Mosel Code of Algorithm 1
S TATION , AND R ATIOS β AND γ st_node := 0
arrival_station := 0
flight_time := 0
forall(i in N_tsp) do
forall(j in N_tsp|getsol(x(st_node,j))>0.99) do
arrival_station := arrival_station
+ tau_original(st_node,j)
st_node := j
break
end-do
if(st_node = s) then
break
C. Minimum Number of Drones Used in Drone Station end-if
To analyze the relationship between the number of cus- end-do
tomers and number of drones used in the station without route if(arrival_station > getobjval/2) then
distortion, we calculate the minimum number of drones in the flight_limit := arrival_station
TSMPMS. We vary the number of customers from 10 to 50 else
to check the trend. Because the ratio of the number of cus- flight_limit := getobjval − arrival_station
tomers serviceable by drones/the total number of customers end-if
can affect the required number of drones, we vary this ratio
from 0.6 to 0.9, and 10 experiments are performed for each
ratio. Therefore, 40 experiments are conducted for each num- Another interesting point is that the required number of
ber of customer group. The average and maximum number of drones to eliminate route distortion is relatively small. The
drones for each customer group are derived. Moreover, ratio β required number of drones are less than one-third of cus-
shows the average number of drones used / total number tomers. Moreover, ratio β is negatively affected by the number
of customers. Similarly, γ (the maximum number of drones of customers because increasing the number of customers
used/total number of customers) is defined to check the upper leads drones to offer more options to deliver parcels to cus-
bound. The details of the experiments and results are shown tomers in the drone-service area. Therefore, drones deliver
in Table V. more parcels in a given time period if the number of customers
The average number of drones used is much less than the increases.
upper bound derived from Proposition 3. The maximum num-
ber of drones used is also smaller than the upper bound.
VI. C ONCLUSION
The maximum number of drones is approximately twice the
average. Ratios β and γ decreased for increasing number of We define a new drone and truck-drone TSP by exploring
customers. The decreasing rate of β is higher than γ . use of a drone station with three features: 1) it can utilize many
drones; 2) it is located far away from the distribution center;
and 3) it is activated for delivery after a truck arrives with
D. Discussion parcels. The TSP-DS was formulated based on mixed integer
We analyzed the flight range (the number of customers ser- programming and we analyzed characteristics of the TSP-DS.
viceable by drones), the velocity of drones, and the number of We proved that the mathematical model can be divided into
drones as main factors affecting the route distortion. However, two different mathematical models, and derived the TSPMS
in a realistic-world problem, the drone range and velocity and the TSMPMS to give the exact solution of the TSP-DS.
are difficult to control because of safety issues and limited Computational experiments showed that the fundamental char-
technologies. Fortunately, increasing the number of drones is acteristics of the TSP-DS and the TSMPMS could effectively
relatively easy because the sufficient number of drones can be reduce the complexity problem. Another experiments revealed
utilized at a drone station which leads to elimination of route that the TSP-DS is more effective than the PDSTSP when
distortion. a majority of customers are located far from the distribution
The tradeoff between the number of customers and the center. We also showed that route distortion can be eliminated
required number of drones for the delivery service provides with relatively small number of drones. We expect our model
insights into practical application. If the distribution center is can be used as a means to overcome the limits of drone facility
far away from a majority of drone-serviceable customers, con- problems, and it can be used to establish drone-truck delivery
structing a drone station will lead to better delivery outcomes. systems in the near future.
However, when the customers are distributed more evenly, the In this problem, we assumed that the locations of customers,
tradeoff for building a drone center or using trucks out of a dis- a drone station, and the distribution center are given, and the
tribution center should be considered. Although at most three results show that the distance between a drone station and
times of drones compared to that for the PDSTSP is required, the distribution center is an important factor. Therefore, the
more customers could be efficiently serviced by drones. location problem of a drone station is an extended topic of

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.
52 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 49, NO. 1, JANUARY 2019

our problem. Consideration of multiple drone stations may [13] L. Kuo. (Oct. 2015). The World’s First Airport for Drones Will Be Built
also inform future research. When some of the flight ranges in Rwanda. [Online]. Available: http://qz.com/519849/the-worlds-first-
airport-for-drones-will-be-built-in-rwanda/
of each drone station overlap, drones could freely move to each [14] C. C. Murray and A. G. Chu, “The flying sidekick traveling salesman
station, which would improve the utilization rates of drones. problem: Optimization of drone-assisted parcel delivery,” Transport. Res.
C Emerg. Technol., vol. 54, pp. 86–109, May 2015.
[15] D. L. Applegate, R. E. Bixby, V. Chvatal, and W. J. Cook, The
A PPENDIX Traveling Salesman Problem: A Computational Study. Princeton, NJ,
USA: Princeton Univ. Press, 2011.
See Algorithm 2. [16] T. Bektas, “The multiple traveling salesman problem: An overview
of formulations and solution procedures,” Omega, vol. 34, no. 3,
ACKNOWLEDGMENT pp. 209–219, Jun. 2006.
[17] B. L. Golden, S. Raghavan, and E. A. Wasil, The Vehicle
The authors would like to thank the anonymous reviewers Routing Problem: Latest Advances and New Challenges, vol. 43.
and the Associate Editor for their valuable and constructive New York, NY, USA: Springer, 2008. [Online]. Available:
https://www.springer.com/us/book/9780387777771
comments to improve earlier version of this paper. [18] P. Toth and D. Vigo, Vehicle Routing: Problems, Methods, and
Applications. Philadelphia, PA, USA: Soc. Ind. Appl. Math., 2014.
[19] A. Allahverdi, C. T. Ng, T. C. E. Cheng, and M. Y. Kovalyov, “A survey
R EFERENCES of scheduling problems with setup times or costs,” Eur. J. Oper. Res.,
[1] M. Grothaus. (Jan. 2016). This is How Google’s Project vol. 187, no. 3, pp. 985–1032, Jun. 2008.
Wing Drone Delivery Service Could Work. [Online]. Available: [20] R. Ruiz and J. A. Vázquez-Rodríguez, “The hybrid flow shop
https://www.fastcompany.com/3055961/fast-feed/this-is-how-googles- scheduling problem,” Eur. J. Oper. Res., vol. 205, no. 1, pp. 1–18,
project-wing-drone-delivery-service-could-work/ Aug. 2010.
[2] D. Muoio. (Aug. 2016). Google’s Secretive Drone Delivery Project [21] K. R. Baker and D. Trietsch, Principles of Sequencing and Scheduling.
Just Got Cleared for Testing—Here’S Everything We Know About the New York, NY, USA: Wiley, 2013.
Program. [Online]. Available: http://www.techinsider.io/google-project- [22] S. Tanaka and S. Sato, “An exact algorithm for the precedence-
wing-drone-service-2016-8/ constrained single-machine scheduling problem,” Eur. J. Oper. Res.,
[3] N. Agatz, P. Bouman, and M. Schmidt, “Optimization approaches vol. 229, no. 2, pp. 345–352, Sep. 2013.
for the traveling salesman problem with drone,” Transport. Sci., [23] A. Bilyk, L. Mönch, and C. Almeder, “Scheduling jobs with
vol. 52, no. 4, pp. 965–981, Apr. 2018. [Online]. Available: ready times and precedence constraints on parallel batch machines
https://doi.org/10.1287/trsc.2017.0791 using metaheuristics,” Comput. Ind. Eng., vol. 78, pp. 175–185,
[4] M. Wohlsen. (Jun. 2014). The Next Big Thing You Missed: Amazon’s Dec. 2014.
Delivery Drones Could Work—They Just Need Trucks. [Online]. [24] M. Davari, E. Demeulemeester, R. Leus, and F. Nobibon, “Exact
Available: https://www.wired.com/2014/06/the-next-big-thing-you-miss algorithms for single-machine scheduling with time windows and
ed-delivery-drones-launched-from-trucks-are-the-future-of-shipping/ precedence constraints,” J. Scheduling, vol. 19, no. 3, pp. 309–334,
[5] M. Tanaka, K. Tanaka, and H. O. Wang, “Practical model construction Jun. 2016.
and stable control of an unmanned aerial vehicle with a parafoil- [25] M. A. Hassan, I. Kacem, S. Martin, and I. M. Osman, “Valid inequalities
type wing,” IEEE Trans. Syst., Man, Cybern., Syst., to be published, for unrelated parallel machines scheduling with precedence constraints,”
doi: 10.1109/TSMC.2017.2707393. in Proc. Int. Conf. Control Decis. Inf. Technol. (CoDIT), Apr. 2016,
[6] S. Cai, Y. Huang, B. Ye, and C. Xu, “Dynamic illumination opti- pp. 677–682.
cal flow computing for sensing multiple mobile robots from a drone,” [26] G. Nicosia and A. Pacifici, “Scheduling assembly tasks with caterpil-
IEEE Trans. Syst., Man, Cybern., Syst., vol. 48, no. 8, pp. 1370–1382, lar precedence constraints on dedicated machines,” Int. J. Prod. Res.,
Aug. 2017. vol. 55, no. 6, pp. 1680–1691, Mar. 2017.
[7] A. Roblin. (Jan. 2015). Pylons Dronairports Would Serve [27] Q. M. Ha, Y. Deville, Q. D. Pham, and M. H. Ha, Heuristic Methods for
As Charging Stations and Safe Storage. [Online]. Available: the Traveling Salesman Problem With Drone, FPT Lecturers, Sep. 2015.
http://www.trendhunter.com/trends/drone-dock/ [Online]. Available: http://ds.libol.fpt.edu.vn/handle/123456789/1994
[8] T. Mogg. (Jul. 2016). Amazon Wants to Turn Street Lights and [28] N. Mathew, S. L. Smith, and S. L. Waslander, “Planning paths for pack-
Church Steeples Into Drone Docking Stations. [Online]. Available: http:// age delivery in heterogeneous multirobot teams,” IEEE Trans. Autom.
www.digitaltrends.com/cool-tech/amazon-prime-air-docking-stations/ Sci. Eng., vol. 12, no. 4, pp. 1298–1308, Oct. 2015.
[9] D. Gross. (Dec. 2013). Amazon’s Drone Delivery: How Would [29] S. M. Ferrandez, T. Harbison, T. Weber, R. Sturges, and R. Rich,
It Work? [Online]. Available: http://edition.cnn.com/2013/12/02/tech/ “Optimization of a truck-drone in tandem delivery network using
innovation/amazon-drones-questions/ k-means and genetic algorithm,” J. Ind. Eng. Manag., vol. 9, no. 2,
[10] L. L. Arnold and P. A. Zandbergen, “Positional accuracy of the wide area pp. 374–388, Mar. 2016.
augmentation system in consumer-grade GPS units,” Comput. Geosci., [30] N. Boone, A. Sathyan, and K. Cohen, “Enhanced approaches to solving
vol. 37, no. 7, pp. 883–892, Jul. 2011. the multiple traveling salesman problem,” in Proc. AIAA Infotech Aerosp.
[11] E. Adams. (May 2016). DHL’s Tilt-Rotor ‘Parcelcopter’ is Conf. AIAA SciTech Forum (AIAA), Kissimmee, FL, USA, Jan. 2015.
Both Awesome and Actually Useful. [Online]. Available: https:// [Online]. Available: https://doi.org/10.2514/6.2015-0889
www.wired.com/2016/05/dhls-new-drone-can-ship-packages-around- [31] K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski, “Vehicle
alps/ routing problems for drone delivery,” IEEE Trans. Syst., Man, Cybern.,
[12] A. Hern. (Sep. 2014). DHL Launches First Commercial Drone Syst., vol. 47, no. 1, pp. 70–85, Jan. 2017.
‘Parcelcopter’ Delivery Service. [Online]. Available: https:// [32] Y. Khosiawan and I. Nielsen, “A system of UAV application in
www.theguardian.com/technology/2014/sep/25/german-dhl-launches- indoor environment,” Prod. Manuf. Res., vol. 4, no. 1, pp. 2–22,
first-commercial-drone-delivery-service/ Jan. 2016.

Authorized licensed use limited to: UNIVERSIDAD SANTO TOMAS. Downloaded on November 19,2022 at 17:48:48 UTC from IEEE Xplore. Restrictions apply.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy