Applied Soft Computing 126 (2022) 109225

Contents lists available at ScienceDirect

Applied Soft Computing

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

Routing and scheduling optimization for UAV assisted delivery system:

A hybrid approach
∗
Mohammad Sajid a , , Himanshu Mittal b , Shreya Pare c , Mukesh Prasad c
a
Department of Computer Science, Aligarh Muslim University, Aligarh, India
b
Department of Computer Science Engineering and IT, Jaypee Institute of Information technology, Noida, India
c
School of Computer Science, FEIT, University of Technology Sydney, Sydney, Australia

article info a b s t r a c t

Article history: This paper proposes a joint-optimization framework for UAV-routing and UAV-route scheduling
Received 24 May 2021 problems associated with the UAV-assisted delivery system. The mixed-integer linear programming
Received in revised form 11 June 2022 (MILP) models for UAV-routing and UAV-route scheduling problems are proposed considering the
Accepted 20 June 2022
effect of incidental processes and the varying payload on travel time. A hybrid genetic and simulated
Available online 1 July 2022
annealing (HGSA) algorithm is proposed for the UAV-routing problem to minimize travel time. In
Keywords: HGSA, genetic algorithm (GA) employs a novel stochastic crossover operator to search for the optimal
Smart logistics global position of customers, whereas simulated annealing (SA) utilizes local search operators to avoid
Unmanned aerial vehicle the local optima. A UAV-Oriented MinMin (UO-MinMin) algorithm is also proposed to minimize the
Routing makespan of the UAV-route scheduling problem. It employs a UAV-oriented view to generate the route-
Route-scheduling scheduling order with minimal computational efforts without affecting the quality of the makespan.
Genetic algorithm A Monte Carlo simulation-based sensitivity analysis is conducted to evaluate the impact of the
Simulated annealing
hybridization probability of GA and SA in the proposed HGSA algorithm. To assess the performance of
the HGSA algorithm, a set P of 24 benchmark instances is adopted and adjusted to meet the constraints
of the UAV-Assisted delivery system. The proposed HGSA outperforms the state-of-the-art algorithms
such as genetic algorithm (GA), Particle Swarm Optimization & Simulated Annealing algorithm (PSO-
SA), Differential Evolution & Simulated Annealing (DE-SA), and Harris-hawks optimization (HHO). For
all 24 instances, the aerial routes generated by HGSA have been used to evaluate the effectiveness
of the UO-MinMin algorithm for different numbers of UAVs. The proposed UO-MinMin algorithm
outperforms the base algorithms such as minimum completion time (MCT) and opportunistic load
balancing (OLB).
© 2022 Elsevier B.V. All rights reserved.

1. Introduction have gained extensive attention in recent years. Apart from the
traditional traffic, E-commerce companies contribute majorly to
The global population will be approximately 8.5 billion by urban traffic to supply goods and services to the customers,
2030, leading to many unavoidable challenges, including traffic which is a variant of the vehicle routing problem (VRP). VRP is a
management, environmental degradation, economic conditions, well-known optimization problem arising in transportation and
and others [1,2]. It has been reported that approximately 1.3 logistics distributions [5–7]. In VRP, there is a requirement to
million people die each year worldwide due to poor traffic man- transport the goods or services from a supply point (Depot) to
agement and road traffic crashes [3]. It also costs 3% of gross various geographically dispersed points (Customers) to optimize
domestic product (GDP) to developing countries [3]. In the Indian the cost. The efficient management of logistic distribution can
city of Mumbai, people waste almost 11 days a year being stuck
indirectly help to save time and money, improve overall customer
in traffic, and this poor traffic management also costs Mumbai
satisfaction, and reduce traffic congestion and carbon emissions.
city nearly 17% of its GDP [4]. This traffic congestion becomes
Consequently, a small saving in VRP’s cost has a significant im-
more exacerbated because of stop-and-go traffic patterns. Various
pact due to its applications in various fields, and thus VRP is a
e-commerce companies like Amazon, Alibaba, eBay, Flipkart, etc.,
strategically and economically important problem. To improve
∗ Corresponding author. the existing delivery system, the unmanned aerial vehicle (UAV)
(a.k.a. drones) can be used to deliver the demands/services, and
E-mail addresses: [email protected] (M. Sajid),
[email protected] (H. Mittal), [email protected] (S. Pare), this delivery system is defined as a UAV-assisted delivery sys-
[email protected] (M. Prasad). tem [8–11]. Commercial companies like Amazon Prime Air, DHL

https://doi.org/10.1016/j.asoc.2022.109225
1568-4946/© 2022 Elsevier B.V. All rights reserved.
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Parcelcopter, UPS Flight Forward, Flytrex, Wingcopter, Zipline, chromosomes to explore the search space. In contrast, the SA
and others develop the UAV-assisted delivery system and ser- operators are used for the remaining chromosomes to exploit the
vices [12]. To effectively realize the UAV-assisted delivery system, neighborhood of the already explored chromosomes. The second
many more issues and challenges exist like UAV-Routing, UAV- phase takes aerial routes to generate the appropriate schedule
route scheduling, joint optimization, cooperated truck and UAV- for available UAVs. The second phase employs the proposed UO-
routing problems, the communication and coordination, fog node MinMin algorithm to schedule the aerial routes on a fleet of UAVs
selection for UAV, situational and context-aware UAV control, to minimize the makespan. The UO-MinMin algorithm provides
UAV-Human interaction, safety and security of UAVs, real-time an order of execution of all aerial routes using given UAVs in the
target tracking, attitude and altitude control of UAVs, route track- minimum possible time [35,36]. The UO-MinMin algorithm works
ing of UAVs, etc. [13–24]. Some other issues are UAV cluster task based on the UAV-oriented view rather than the aerial route-
scheduling, routing & scheduling for multi-UAV inspection, task oriented view, i.e., the UAV-oriented view provides a schedule
allocation for multi-UAV systems, multi-target tracking, location, of aerial routes with minimal computational efforts and without
and routing problems with or without land vehicles to name a affecting the quality of makespan [36].
few [15–20]. The UAVs’ routing and route-scheduling problems The followings are main contributions of this paper-
play a major role in UAV-assisted delivery system. The UAVs’
routing and route scheduling problems can also be extended to • The mixed-integer linear programming (MILP) models for
many real-life applications in various fields such as drug delivery, UAV-routing and UAV-route scheduling problems are pro-
food delivery, disaster management, finance, banking, defense, posed.
and others. • The effect of incidental processes (i.e., Loading, Take-off,
The joint optimization of routing and route scheduling prob- Landing, and delivery of parcels to customers) and varying
lems of UAVs must be addressed for the efficient realization of payload on travel time are considered.
the UAV-assisted delivery system. The UAV-routing problem is • A join-optimization framework consisting of UAV-routing
concerned with finding the aerial routes of geographically dis- and UAV-route scheduling algorithms is proposed for the
tributed customers to deliver their demands with minimum pos- UAV-assisted delivery system.
sible travel time. In contrast, the UAV-route scheduling problem • Due to the complex nature of the UAV-routing problem, a
involves scheduling aerial routes using a fixed number of UAVs hybrid algorithm, i.e., HGSA, is proposed to generate feasible
to minimize the makespan. The UAVs’ routing and route schedul- aerial routes and search for the near-optimal solution.
ing problems must be resolved jointly from the customers’ and • Due to the time constraints, many aerial routes, and limited
suppliers’ points of view. Due to UAVs’ battery and capacity con- UAVs, a UO-MinMin scheduling algorithm is proposed to
straints, the UAVs must execute many trips (routes) to fulfill the generate scheduling orders of aerial routes in a computa-
customers’ demands. The e-commerce organizations can afford tionally efficient way.
a limited number of UAVs due to budget constraints. Thus, it • The dataset P for capacitated vehicle routing problem (CVRP)
requires managing (scheduling) the several UAV trips every day benchmark instances has been adopted and adjusted to
to serve the customers. The inefficient management (scheduling) meet the distance, flow, and other constraints of the UAV-
of UAV trips leads to high waiting times for customers (customer Assisted delivery system [25].
dissatisfaction) and affects the organization’s overall business or • A Monte Carlo simulation-based sensitivity analysis is con-
process. Therefore, the UAV-route scheduling algorithms are also ducted to evaluate the effect of input probability for the
required in addition to UAV-routing algorithms. The aerial routes hybridization of GA and SA in the proposed HGSA algorithm.
with customers are decided first, followed by the scheduling • The simulation study on 24 instances of the dataset P has
of aerial routes using UAVs. In this scenario, aerial routes are been performed. Results show the effectiveness of our pro-
assumed tasks, and UAVs are considered processing elements; the posed HGSA algorithm against genetic algorithm (GA), Par-
problem becomes a variant of an independent task scheduling ticle Swarm Optimization & Simulated Annealing algorithm
problem [25–28]. Fig. 1 depicts a problem instance in which a (PSO-SA), Differential Evolution & Simulated Annealing (DE-
routing algorithm is executed in the first phase, followed by a SA), and Harris-hawks optimization (HHO) algorithms [33,
route scheduling algorithm in the second phase. In the first phase, 34,37–39].
the routing algorithm takes customers’ information and generates • For all 24 instances, the best solutions generated by HGSA
clusters of customers known as aerial routes. The aerial routes have also been utilized to evaluate the effectiveness of
are fed to a route scheduling algorithm that schedules the aerial the UO-MinMin algorithm against minimum completion
routes on appropriate UAVs. The aerial routes are scheduled on time (MCT) and opportunistic load balancing (OLB) algo-
two available UAVs using a route-scheduling algorithm. rithms [35,36].
The UAV-routing and UAV-route scheduling problems fall into
NP-Hard combinatorial optimization problems [28–30]. Gener- The organization of the remaining paper is as follows. Sec-
ally, the combinatorial optimization problems (i.e., routing and tion 2 reports the related work in the same domain. Section 3
scheduling problems, etc.) have been addressed using exact meth- presents the mixed-integer linear programming (MILP) mod-
ods (branch-&-bound, Google OR tools); meta-heuristics (GA, els for UAVs’ routing and route scheduling problems. Section 4
ACO, firefly algorithms); heuristics (sweep algorithm, Clarke-&- explains the proposed HGSA routing algorithm, while the UO-
Wright saving algorithm,); local search algorithms (swap, ex- MinMin algorithm is described in Section 5. Section 6 presents
change, 2-Opt*) and their combinations [28,29]. This paper devel- the simulation study, including systems parameters, sensitivity
ops meta-heuristic and heuristic algorithms to address the UAV- analysis, UAV-routing results, UAV-route scheduling results, and
routing and UAV-route scheduling problems. Fig. 2 depicts the observations. Finally, Section 7 puts the conclusion remark and
proposed two-phase joint optimization framework. In the first future research directions.
phase, the routing algorithm fetches customers’ information and
generates aerial routes of customers. The proposed hybrid genetic 2. Related work
and simulated annealing (HGSA) algorithm addresses the routing
problem. HGSA is an amalgamation of genetic algorithm (GA) [31, A significant amount of literature exists on UAV-based ser-
32] and simulated annealing (SA) algorithm [33,34]. In the pro- vices. Several routing and scheduling algorithms for the hybrid
posed HGSA, the GA operators are randomly applied to 50% of truck-UAV system or UAV-assisted delivery system are reported.
2
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 1. Routing and Scheduling in UAV-Assisted Delivery System.

is solved by constructing a complete tour for all targets, split-

ting it by GV routes, constructing the GV tour, and assigning
UAV flights to the route. Agatz et al. [44] introduced a different
version of the traveling salesman problem with the drone (TSP-
D). The truck collaborates with a drone to perform deliveries,
and its mixed-integer linear programming (MILP) formulation
is also presented. The heuristics based on dynamic program-
ming and local search techniques were proposed to optimize
TSP-D. The worst-case approximation ratio for heuristics was
also proved. Freitas and Penna [45] also proposed to solve TSP-
D [43] by creating an initial solution using a Concorde solver
and next applying the randomized variable neighborhood descent
(RVND) heuristic to obtain the problem solution. The proposed
heuristic performs much better on large-scale instances. Ham [46]
solved the integrated scheduling problem of m-truck, m-drone,
Fig. 2. Joint Optimization Framework for UAV-Assisted Delivery System.
and m-depot, which is constrained by time-window, drop-pickup,
and m-visit using an application of constraint programming (CP).
The study considered that the drone either flies back to the
Savuran and Karakaya [40] proposed a GA-based route op- depot to receive the parcel for the next delivery or can di-
timization for a carrier-launched drone routing problem. The rectly fly to another customer for pickup. The CP model has been
carrier keeps on moving on its own, and the UAV serves the tested for small instances, and it performs well in terms of the
target customers. The effect of UAV take-off and landing on quality of solutions and computational efforts. Yurek and Oz-
tour length was considered, and the performance of the route mutlu [47] proposed a decomposition-based iterative algorithm
optimization algorithm was compared with the Nearest-Neighbor for traveling salesman problems with the drone (TSP-D). The
(NN) heuristic. Murray and Chu [41] introduced the flying side- delivery truck carries a drone on its roof, and both truck and
kick traveling salesman problem (FSTSP), in which a truck acts drone serve customers. The algorithm starts with the shortest
like a mobile station for drones. Aimed to optimal routing and truck route, performs the assignment of customers to drone, and
scheduling of unmanned aircraft and delivery trucks, a heuristic improves routing decisions iteratively to optimize delivery com-
based on the truck first, drone second (TFDS) approach was pletion time. Kim and Moon [48] proposed a truck-drone system
proposed in which truck-only routes are constructed first and in which a fixed drone station far away from the distribution cen-
next, drones replace the truck nodes to optimize the objective ter is established to store drones and charging devices. The drones
value. Ferrandez et al. [42] investigated the time and energy are activated to perform delivery upon arrival of the truck from
required in a truck-drone delivery network compared to a stand- the distribution center. A mixed-integer programming model is
alone truck or drone delivery network. A K-means and genetic developed for traveling salesman problems with a drone station
algorithm are proposed to find out the optimal routes and launch (TSP-DS). The experiments show that TSP-DS is more effective
sites. The algorithm assumed that the number of drones per than parallel drone scheduling TSP, in which parcel delivery is
truck and delivery requirements was known in advance. The K- performed by the drones in the flight range of the distribution
means clustering was used to find the vehicle stops, whereas a center. Jeong et al. [49] studied the hybrid delivery system in
genetic algorithm was used to find the vehicle routes. The study which trucks and drones perform parcel deliveries. The study
recommended two or more drones per truck, and the speed of the considered the effect of parcel weight on drone energy con-
drone should be approximately three times (or more) that of the sumption, the federal government’s range, and restricted flying
truck for efficient deliveries in real-time savings. Luo et al. [43] areas. A mathematical model was developed and solved using
proposed solving a two-echelon cooperated routing problem for CPLEX OR-tool. Due to the high computational requirements of
the ground vehicle (GV) and UAVs. The ground vehicles serve CPLEX for large instances, a heuristic algorithm TPCSA was devel-
the customers along the road, and their associated UAVs serve oped that combines constructive heuristics and search heuristics.
the customers in the unreachable areas beyond the road. A 0– Liu et al. [50] proposed a cooperated truck and drone routing
1 integer programming model is formulated, and the problem problem in which the truck and drone are used cooperatively
3
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

to complete the deliveries of all parcels. In addition to parcel on flight time, and UAV behavior during the idle time were all
delivery of customers, the truck provides a moving launching considered. A receding horizon task assignment (RHTA) heuristic
pad for drones. The energy consumption model for the routing was proposed and tested for large scale numerical examples
problem and a two-stage route-based modeling approach is pro- in island-area delivery. Torabbeigi et al. [62] studied a parcel
posed to optimize the truck’s main route and the drone’s adjoint delivery system using drones, considering the payload effect on
flying routes. The hybrid algorithm consisting of a simulated the battery consumption rate, which may be problematic in over-
annealing algorithm and Tabu search is proposed to solve the estimation and under-estimation. It was shown that a linear
routing problem. The random instances and a road network in relationship exists between payload and battery consumption
Changsha, China, have been used to conduct the experiments. Luo rate. Linear regression-based models were proposed for strategic
et al. [51] proposed multi-objective genetic algorithm to solve the and operational planning for a given region. The proposed models
collaborative single truck-multiple drone problem to optimize the help determine the depot locations and delivery flight paths for
customer satisfaction and distribution cost. The multi-objective drones. Kim et al. [63] proposed a mixed-integer programming
GA adopted greedy initialization method, problem specific solu- model for parcel delivery using drones in dense urban areas. The
tion representation and adaptive strategy to balance exploration proposed model considers rooftops of city buildings as launching
and exploitation. The multi-objective GA was evaluated based pads for operation planning. A solution heuristic based on block
on benchmark instances and a real-world case in Changsha city, stacking is proposed to perform the maximum deliveries. The
China. Murray and Raj [52] proposed a three-phase heuristic for proposed heuristic is evaluated on three different problem sizes,
cooperative truck-drone delivery problem which is an extension i.e., small (10 destinations and two drones), medium (20 goals
of the flying sidekick TSP. The proposed MILP models of the and four drones), and large (50 destinations and ten drones), to
problem considers the retrieval activities and drone queuing. assess the performance.
First phase partitioned customers into two sets based on their Schermeret al. [19] proposed a matheuristic for vehicle rout-
serving vehicle. The second phase generates sorties to identify ing problems with drones (VRPD) such that drones are associ-
the routes of truck and UAVs followed by the third phase in ated with every truck to serve the customers to optimize the
which exact action timing is computed. The numerical analysis is makespan. In addition, an algorithm for drone assignment and
conducted to evaluate the performance of the proposed heuristic. scheduling problems (DASP) was proposed to search for the opti-
Peng et al. [53] proposed a hybrid genetic algorithm (HGA) for mal assignment and schedule of drones with minimum makespan.
cooperative truck-drones routing problem. The proposed HGA The numerical results produced by the proposed algorithm con-
consists of three phases populations’ management, initialization, firm that drones could significantly reduce the makespan.
and education to support the delivery of multiple parcels in Sawadsitang et al. [64] developed Bayesian Shipper Cooper-
different locations. As shown, HGA performs better compared to ation in Stochastic Drone Delivery (BCoSDD) framework which
Lin–Kernighan Heuristic (LKH), a greedy algorithm, and a variant employs the dynamic Bayesian coalition formation game and
of adaptive memory programming (AMP). Nguyen et al. [54] stochastic programming. The framework considers the uncer-
addressed min-cost Parallel Drone Scheduling Vehicle Routing tainty factors such misbehavior of shippers and drone breakdown
Problem using Ruin and Recreate heuristic (SISSRs) algorithm to to minimize the drone delivery cost. The framework was eval-
minimize the transportation cost. A classification and regression uated using Solomon benchmark instances and Singapore logis-
trees (CART) based sensitivity analysis is conducted to evaluate tics industry. The framework offers effective solution compared
the impact of different components, i.e., local search, threshold to deterministic drone delivery (DDD), stochastic drone deliv-
acceptance, perturbation, sweep removal, and random removal. ery (SDD), cooperative in deterministic drone delivery (CoDDD),
It is reported that SISSRs outperforms state-of-the-art algorithms and cooperative in stochastic drone delivery (CoSDD). Huang
and offers twenty six new best known results. et al. [65] proposed iterated heuristic framework, Stochastic Event
Moreover, an agent-based method for Truck-multi-Drone Team Scheduling, to address the dynamic task scheduling problem with
Logistics Problem [55], a matheuristic for Mothership vehicles stochastic task arrival times and due dates. The SES was proposed
and coordinating drones [56], an ACO based scheme [57] and for the intelligent on-demand meal delivery system (UIOMDS)
hybrid-sweep and genetic algorithm [58] for instant delivery by to optimize the total tardiness. The simulated annealing was
UAV in cooperation with land vehicles, a heuristic algorithm also integrated with SES to improve exploration capability. As
for cooperated refrigerated land vehicles and drones for food shown, the SES performs better compared to Earliest Due Date
industry [59], and others are also proposed. first (EDDF) and Guided-local search (GLS).
The routing and scheduling algorithms are also proposed for In addition to routing and scheduling problems, the UAVs
UAV-assisted delivery systems. Some of the recent works are are also considered for other issues such as joint optimization
reported here. San et al. [60] proposed to solve the parcel delivery of routing and scheduling for vehicle-assisted multi-UAV inspec-
problem using a swarm of unmanned aerial vehicles (UAVs). A tion [15], Decentralized control of UAVs for cluster task schedul-
feasibility study was performed to avoid deadlock situations. A ing [16], adaptive task allocation [17], surveillance mission prob-
genetic algorithm of chromosomes of multi-dimensional genes lems using strategic routing of a fleet of UAVs [18,20], control
was proposed to solve the parcel delivery problem to optimize and tracking of UAVs [21–24], simultaneous sensor selection and
two fitness functions, i.e., vertical in terms of payload and weight routing of UAVs for military reconnaissance missions [66], mini-
while horizontal in terms of completion time, using a concurrent mum time coverage of ground areas using a group of UAVs [67],
scheduler approach. Dorling et al. [11] proposed to solve two persistent intelligence, surveillance, and reconnaissance (PISR)
variants of the vehicle routing problem (VRP) for drone delivery, routing problem [68,69], to name a few.
considering the effect of payload and battery weight on energy In majority of literature, the effect of incidental processes
consumption. A string-based simulated annealing (SA) algorithm (i.e., Loading, Take-off, Landing, and delivery of parcels to cus-
is employed to optimize the delivery time and budget constraints tomers) and varying payload on travel time are neglected or
to solve the drone delivery problem. Song et al. [61] proposed simplified. The joint-optimization of routing and route scheduling
a scheduling model for the UAV logistics systems where UAVs problems are also neglected and both problems are addressed
can replenish their battery and deliver the product at different exclusively. Moreover, the scheduling of aerial routes on limited
service stations distributed in the delivery system. The flight time, UAVs and multiple parcel delivery by UAV was also not consid-
limited loadable products, the effect of loaded product weight ered. This paper considers the joint-optimization of routing and
4
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

route-scheduling algorithms in a UAV-assisted delivery system. is represented as J = {Az : 1 ≤ z ≤ M }. The aerial route
Multiple parcel deliveries are allowed, and the effects of the (cycle)Az begins and ends at the depot o0 and is represented by
incidental processes on travel times are also considered. The Az = (o0 , oj , ol , . . . . . . o0 ). The payload assigned to any UAV will
proposed framework can be implemented for delivery of food, be maximum at the depot and reduced at subsequent nodes. Let
i = (oj , ol , . . . . . . o0 ) be the remaining path of UAV u at node oi
Azk k
drug, goods, defense and other services.
z
from the aerial route A and node oj is the successor node of oi .
3. Problem formulation The aerial routes Azk z
0 and A are equivalent to each other at the
depot node o0 . Therefore, the remaining payload of the UAV uk at
This section presents the design principles, UAV routing, and node oi is given as follows:
UAV-route scheduling problems. ∑
Lki = rj (1)
3.1. UAV routing problem oj ∈Azk
i

Where node oj is the successor node of oi in the aerial route Az

The layout of the UAV Routing Problem can be described in
For aerial route Az from node oi to oj , the battery power of the
the form of a Euclidian graph G = (O+ , C ). The element O+
UAV uk can be represented as [10,11,50]:
denotes the set of n+1 nodes, i.e., O+ = {o0 } ∪ O, where node
o0 represents the depot, and the set O consists of n nodes from (W + Lki )vijk
o1 to on representing the customers and given by O = {oi : 1 ≤ pkij = + pe ∀oi , oj ∈ O+ , oi ̸= oj (2)
370ϕγ
i ≤ n}. Each node oi is represented by geometric coordinates
(xi , yi ) and its demand ri . The depot o0 has no demand while Where γ , pe and ϕ are lift-to-drag ratio, power consumption
other nodes o1 to on have some positive demands, i.e., ri > 0. The of electronics, and conversion efficiency of motor and propeller,
element respectively.
{( C is) the set of edges }joining the nodes represented by In aerial route Az from oi to oj , the energy consumed by the
C = oi , oj : oi , oj ∈ O, i ̸ = j and dij represents the Euclidian
UAV uk is given as-
distance between nodes oi and oj . A fleet U of m homogeneous
UAVs is represented by U = {uk : 1 ≤ k ≤ m} where all UAVs W + Lki vijk
(( ) )
∫ ∫ dij
1
have identical weight W, payload capacity L, and battery capacity Fij = pkij dt = + pe ∗ ds
370ϕγ vijk
B. The payload assigned to the UAV uk at node oi is represented 0

by Lki . When UAV uk flies from node oi and oj with velocity vijk . It ∀oi , oj ∈ O+ , oi ̸= oj (3)
takes time tijk and consumes power pkij . The incidental processes,
i.e., loading, take-off, landing, and parcel delivery to the node For efficient delivery of customer’s demand in aerial route Az , it
(customer), are assumed to take equal times and are represented is assumed that UAV uk consumes maximum power to fly with
by τ . In contrast, the incidental processes, i.e., take-off, landing, maximum speed:
and parcel delivery to the customer, consume almost equal en- pkij = PH ∀oi , oj ∈ O+ , oi ̸= oj (4)
ergies and are represented by ξ .The assumptions and constraints
k
are as follows: Therefore, the velocity of the UAV u to move from node oi to oj
in aerial route Az is given as:
• In an aerial route, a single UAV may fulfill the demands of 370ϕγ (PH − pe )
more than one customer. vijk = ( ) ∀oi , oj ∈ O+ , oi ̸= oj (5)
• The depot has a launching pad, and UAV can be launched. W + Lki
The battery of the UAV is fully recharged before leaving the In aerial route Az from node oi to oj , the time required by the UAV
depot. uk is given as:
• The customers are assigned to a UAV before it leaves the ∫ dij
depot. 1
tijk = ds ∀oi , oj ∈ O+ , oi ̸ = oj (6)
• All UAVs may perform deliveries using a tether to lower a 0 vijk
package to the ground.
Therefore, the time required by UAV uk to move from node oi to
• The payload capacity of every UAV must be observed, i.e., the
oj in aerial route Az is given as:
sum of demands of all customers assigned to a UAV in an
aerial route must not exceed the maximum payload capacity dij (W + Lki )
L. tijk = ∀oi , oj ∈ O+ , oi ̸= oj (7)
370ϕγ (PH − pe )
• The battery capacity of every UAV must be observed, i.e., the
power consumed in fulfilling the demands of all customers In aerial route Az from node oi to oj , the energy consumed by the
assigned to a UAV for a given aerial route must not exceed UAV uk is given as:
the maximum battery capacity B. PH dij (W + Lki )
• The demand of every customer is delivered once using one ekij = PH tijk = ∀oi , oj ∈ O+ , oi ̸= oj (8)
370ϕγ (PH − pe )
UAV only.
• Every customer is visited exactly once by one UAV only. So, the energy consumed by the UAV uk for the assigned aerial
• Every UAV route must start and end at the depot. route Az ∈ J is given as:

The problem is to fulfill the demands of all n customers using m

∑ PH dij (W + Lki )
Ezk = ∗ Xijk ∀uk ∈ U (9)
UAVs while meeting the payload and battery capacities, flow, and 370ϕγ (PH − pe )
(oi , oj )∈Az
integrity constraints. For the UAV routing problem, let Xijk be the
decision variable to decide whether the delivery is performed by Therefore, the total energy consumed to fulfill the demands of all
UAV or not. If UAV uk flies from node oi to next node oj , the value customers is given as:
of Xijk is set to 1 and 0 otherwise. ∑ ∑ ∑ PH dij (W + Lki )
It is assumed that the customer set is partitioned into the set E O+ = Ezk = ∗ Xijk + M ∗ ξ (10)
of M aerial routes, which is known as a batch-of-routes (BoR) and 370ϕγ (PH − pe )
Az ∈ J Az ∈J (oi , oj )∈Az

5
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

The time required by the UAV uk to fulfill the demands of all route Az is exclusively executed by one UAV only due to the
customers on the aerial route Az ∈ J ∈ is given as: independence between aerial routes. By atomic, it means every
∑ aerial route Az is an indivisible unit and cannot be partitioned
Tzk = tijk ∗ Xijk + (M + 1) ∗ τ ∀ uk ∈ U (11)
and distributed across multiple UAVs to perform deliveries. A
(oi , oj )∈Az
fleet U of m homogeneous UAVs is represented by U = {uk :
Therefore, the total time required to fulfill the demands of all 1 = k = m} and it is assumed that every UAV uk visits the
customers is given as: aerial routes in a non-preemptive manner. In non-preemptive
∑ route scheduling, the UAV is not interrupted until it delivers to all
T = Tzk customers in the assigned aerial route. After executing one route,
Az ∈J the UAV battery is recharged or replaced for processing another
⎛ ⎞
( ) aerial route. It is assumed that all UAVs take almost equal time to
dij W + Lki
charge the battery at level B fully, and it is represented as τ c . The
∑ ∑
= ⎝ ( ∗ Xijk + (M + 1) ∗ τ ⎠
370ϕγ (PH − pe ) time required for the aerial route Az by UAV uk is represented by
Az ∈J (oi , oj )∈Az
Tzk . Since, all UAVs are homogeneous; it is assumed all UAVs take
k
∀u ∈ U (12) almost equal time for execution of the aerial route Az i.e., Tzk =
The problem is to fulfill the demands of all n customers using Tz , ∀k.
available m homogeneous UAVs while meeting the payload and For each UAVuk , the depot o0 maintains scheduling infor-
battery capacities, flow, and integrity constraints to minimize the mation in a queue Qk which consists of a tuple (z, g, CTz ) for
total travel time. The objective function is given as follows: each scheduled aerial route Az on UAVuk . The g represents the
scheduling slot at the assigned UAV uk while the CTz represents
Minimize T the expected completion time of the aerial route Az at assigned
⎛ ⎞
( ) UAV uk . For the UAV route scheduling problem, let Yzk be the
∑ ∑ dij W + Lki
= ⎝ ( ∗ Xijk + (M + 1) ∗ τ ⎠ decision variable to decide whether the aerial route Az is assigned
370ϕγ (PH − pe ) to UAV uk or not. If Yzk equals to 1, it means UAV uk executes aerial
Az ∈J (oi , oj ) ∈Az
route Az for delivery and queue Qk consists of one entry (z, g, CTz )
With the following constraints for Az ; and otherwise Yzk will be equal to 0.
∑ ∑
Xijk = 1 (13)
• URT k : It represents the expected ready time of the UAVuk
uk ∈U oi , oj ∈O+
∑ offered to each unscheduled aerial route Az . The expected
rj ≤ L ∀oi ∈ O+ , ∀A ∈ J , ∀u ∈ U z k
(14) ready time of the UAVuk represents the time when all pre-
oj ∈Azk
viously scheduled aerial routes have been executed, and it
i
∑ is ready to execute the next aerial route Az . It is given as
pkij Xijk ≤ B ∀uk ∈ U , ∀Az ∈ J (15) follows:
(oi , oj )∈Az
if Q k is empty
{
0
∑ URT k = (20)
CT w + τ Other w ise
k k c
X0j = 1 ∀u ∈ U (16)
oj ϵ O
∑ ∑ where CT w represents the completion time of the last scheduled
Xihk − Xhjk = 0, ∀oh ∈ O, ∀uk ∈ U (17) aerial route Aw at UAVuk .
oi ϵ O+ oj ϵ O+
∑ • RT z : It represents the expected ready time of the aerial route
k
Xi0 = 1, ∀uk ∈ U (18) Az and it is computed as:
oi ϵ O
RT z = min{URTk : ∀uk ∈ U } (21)
Xijk ∈ {0, 1}, ∀oi , oj ∈ O+ , ∀uk ∈ U (19)
The constraint (13) ensures that every customer must be visited In conjunction with the expected ready time RTz of aerial route
once using a single UAV only, and constraint (14) provides that Az , the UAV ul that offers the minimum of expected ready time,
the quantity delivered by a UAV must not exceed the payload and the value of g is decided.
capacity L at any node in an aerial route. The constraint (15)
ensures that the total power consumed by the UAV in a route
• CT z : The expected completion time of the aerial route Az is
must not exceed the battery capacity B. The constraints (16), (17), the sum of the expected ready time of the assigned UAV ul
and (18) ensure the flow constraints at the depot as well as on and the expected time required to execute the aerial route
customers. Constraint (16) provides that each UAV departs from Az . It is computed as follows:
the depot; constraint (17) ensures that every UAV that visits the CTz = RTz + Tz (22)
customers must leave, and constraint (18) ensures that the UAV
must return to the depot. The constraint (19) ensures the integrity In conjunction with the expected completion time CTz of aerial
of the decision variable [29]. route Az scheduled on UAV ul , one entry (z, g+1, CTz ) is enqueued
in queue Ql and the g represents the scheduling slot of the last
3.2. UAV-route scheduling problem aerial route scheduled at UAVul . The value of Yzl is also set to 1.

In general, the depot has limited UAVs. It may be required • MKUJ : It represents the makespan of BoR J of M aerial routes
to execute multiple aerial routes by a single UAV, which require concerning fleet U of m homogeneous UAVs and is defined
efficiently scheduling aerial routes on available UAVs. The set of as the maximum completion times of all aerial routes from
M aerial routes can be considered as a batch-of-routes (BoR) and BoR J. It can be computed as:
represented as J = {Az : 1 ≤ z ≤ M }. Every aerial route
MK UJ = max CTz , ∀Az ∈ J
{ }
Az is atomic and independent of other aerial routes. Every aerial (23)
6
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 3. Proposed HGSA UAV-Routing Algorithm.

The problem is to schedule M aerial routes from BoR J on the the fittest chromosomes are selected, combined, and mutated
available fleet of m homogeneous UAVs while meeting the batch to produce offspring for the upcoming generation. The SA algo-
size, flow, and integrity constraints to minimize the makespan. It rithm is also a well-known probabilistic technique that works
can be written as: based on annealing in metallurgy, consisting of heating and con-
trolled cooling of a material to enlarge the size of its crystal
Minimize MK UJ = max CTz , ∀Az ∈ J
{ }
and then reduce its imperfections. The SA algorithm randomly
s. t. selects a neighbor chromosome close to the current chromosome,
∑ measures its quality, and then selects the better or worse chro-
Yzk = 1, ∀Az ∈ J (24) mosome based on the temperature-dependent probabilities. As
uk ∈U shown in Fig. 3, the proposed HGSA algorithm begins with a
∑ population of randomly generated individuals known as chro-
Yzk < M , uk ∈ U (25)
mosomes. The chromosomes consist of all customers’ random
Az ∈J
∑∑ permutations, as shown in Fig. 4(a). It then evolves a new pop-
Yzk = M (26) ulation until the stopping criterion is reached. To create the
Az ∈J uk ∈U
population for the next generation, the HGSA employs GA and
SA randomly and equally, i.e., half offspring are generated using
Yzk ∈ {0, 1}, ∀Az ∈ J , ∀uk ∈ U (27) genetic operators. The remaining 50% of offspring are generated
using SA. The GA works based on tournament selection, giant
The constraint (24) represents that each aerial route must be
tour best cost crossover (GTBCX) operator [59,60], and swap
executed once using a single UAV. The constraint (25) states that
mutation operators, while SA works based on inversion, 2-opt*,
the total aerial routes assigned to UAV uk must not exceed the
relocate, and exchange operators [61]. The proposed HGSA is an
total number of aerial routes M. The constraint (26) ensures that
elitist algorithm, i.e., it merges the old and the newly formed
the total number of aerial routes scheduled on all UAVs must be populations, and then the merged population is sorted based on
equal to the total number of aerial routes M. The constraint (27) the objective values; the best individuals are retained, and the
ensures the integrity of the decision variable [35,36]. worst individuals are discarded. The algorithm terminates after a
specified number of iterations. The following sub-sections explain
4. Proposed HGSA UAV-routing algorithm the chromosome’s structure and evaluation strategy, GA, and SA
operators.
This section explains the proposed routing algorithm, i.e., hy- It is possible to hybridize the GA and SA using two methods. In
brid genetic and simulated algorithm (HGSA) for UAV-routing the first method, the SA algorithm is executed after each iteration
problem (Section 3.1), and all essential components. Fig. 3 depicts of the GA for any number of chromosomes. In this method, the
the proposed HGSA algorithm. time required to execute the GA and SA algorithms contribute
The proposed HGSA algorithm is an amalgamation of the ge- separately to the overall algorithm’s time complexity. In the
netic algorithm (GA) [31,32] and the simulated annealing (SA) second method (proposed HGSA algorithm), the GA operators are
algorithm [33,34]. It works based on the natural selection pro- randomly applied to 50% of chromosomes, and the SA operators
cess to solve search and optimization problems where the fittest are used for the remaining chromosomes. This method ensures
chromosomes are selected, combined, and mutated to produce a balance between exploratory and exploitative behavior. This
the offspring for the upcoming generation. The GA is a global method also provides half exploitative chromosomes in each gen-
evolutionary optimization algorithm that works based on the nat- eration. Moreover, the time required to execute the GA is reduced
ural selection process to solve the optimization problems where to half compared to the first method.
7
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

stochastically and accepts the offspring with the best available fit-
ness value in constant time. As shown in Fig. 5, two offspring are
selected for reproduction, two consecutive customers are chosen
from both, and subsequently, the selected customers are removed
from other offspring. The removed customers are reinserted at k
randomly chosen locations, and then the second customer is also
reinserted at k randomly selected locations. The best offspring
among the k offspring is determined based on the fitness value.
As shown in Fig. 5, two offspring are chosen, and two customers
are selected from both, whereas Fig. 6 shows the removal of
customers. Fig. 7 shows the stochastic insertion of the first cus-
tomer in both the offspring, whereas Fig. 8 depicts the stochastic
insertion of the second customer in the offspring.
Fig. 4. (a) Chromosome Structure (b) Chromosome divided into UAV’s Aerial
Routes.
4.3. Selection and mutation operators
Table 1
Customer demands. This work employs tournament selection with tournament
Customer 1 2 3 4 5 6 7 8 9 10 size 8 to select the offspring for reproduction. To maintain the
Demand 1 0.5 1 1 1 1 1.5 1.5 1.5 1
population’s diversity and avoid premature optimum solutions,
mutation operators introduce new offspring by inserting random
genes at random locations. This work employs swap mutation in
which two randomly selected customers are swapped.
4.1. Chromosome structure and its evaluation
4.4. Simulated annealing operators and parameters
The random permutation of all customers represents a chro-
mosome, as shown in Fig. 4(a), in which each customer ap- During the reproduction process for a new population in
pears precisely once. The initial population consists of random HGSA, the SA algorithm also searches the neighborhoods of a
permutations of customers representing the chromosomes. Each given offspring based on different neighborhood structures. The
chromosome is considered a giant aerial route without route SA algorithm employs three well-known neighborhood struc-
delimiters. A cluster-first route-second approach is used to covert tures, namely 2-Opt*, Relocate, and Exchange operators, as pre-
a giant aerial route into UAV’s aerial routes [70]. The giant aerial sented in Fig. 8(a), (b), and (c), respectively [72]. As shown in
route is decomposed into clusters of customers based on cus- Fig. 8(a), the 2-opt* operator reduces the overlapping of routes.
tomers’ demands and the payload of the UAV. The customers The customers served after the customer oj in the second route
in the giant aerial route are allocated to the first UAV until its are removed and reinserted to be served after the customer oi in
payload capacity is exhausted. The customers assigned to the the first route. Similarly, the customers served after the customer
first UAV are removed from the giant aerial route. The exact oi on the first route are removed and reinserted to be visited
process is recurring until all customers of the giant aerial route after the customer oj on the second (route. It)can be performed
are allocated to UAVs. Using the demands of customers given in
by removing the edges (oi , oi+1 )( and oj ,) oj+1 and replace them
Table 1, 3 kg as the capacity of each UAV and cluster-first route-
with the edges (oi , oj+1 ) and oj , oi+1 , respectively. Fig. 8(b)
second approach, the giant aerial route as shown in Fig. 4(b) is
shows that the relocate operator removes outlier customers and
decomposed into UAV’s aerial routes. Next, UAV visits customers
reinserts them into a route, forming a nearby cluster. The edges
as per the order given in the assigned aerial route. The aerial
(oi−1 , oi ) and (oi , oi+1 ) are removed from the first route and
routes can be used to compute the fitness value, i.e., total travel
are replaced by the edge (oi−1 , oi+1 ). The removed customer is)
time, using Eq. (12). Whenever the aerial routes are more than
relocated to the second route by replacing the edge oj , oj+1
(
the total available UAVs, the UAV-scheduling algorithm can be
with the edges (oj , oi ) and (oi , oj+1 ). Fig. 8(c) shows that the
employed to schedule the aerial routes.
exchange operator exchanges customers from different routes to
reduce overlapping routes. Two edges in both routes are removed
4.2. Best-aerial route with stochastic Insertion Crossover operator
and replaced by two other edges. In the first route, the edges
(oi−1 , oi ) and ((oi , oi+1)) are removed and replaced by edges
This work proposes a problem-specific crossover operator,
(oi−1 , oj ) and oj , oi+1 , respectively. In the second route, the
which searches neighborhoods to find the best offspring with
edges (oj−1 , oj ) and ((oj , oj+1)) are removed and replaced by the
constant computational time (i.e., O(1) time). The proposed oper-
edges (oj−1 , oi ) and oi , oj+1 , respectively.
ator always generates viable offspring without requiring an extra
The SA algorithm produces three neighborhood offspring, and
repair procedure. The basic idea is that every customer must
the best offspring is chosen and compared with the old off-
reach its optimal global position, resulting in better-fitted off-
spring. It is decided if the best offspring is better than the old
spring. The proposed operator is inspired by the best-cost route-
offspring; otherwise, the worse offspring are accepted based on
crossover (BCRC) proposed by B. Ombuki [71], in which two off-
the temperature and cost-dependent probability as per Eq. (28).
spring (giant routes) are selected. Then some customers are cho-
The temperature value is updated using Eq. (29) [33,34].
sen from one offspring and are removed from another offspring. {
The removed customers are then again re-inserted in search 1 ∆c ≤ 0
of the best possible neighborhood. The BCRC operator requires ProbAccept = − ∆θc
(28)
checking every possible neighborhood offspring that leads to e ∆c > 0
linear, i.e., O(N) computational efforts, making the reproduction
θ = θ ∗ (1 − α) (29)
process very slow.
This work proposes a best Aerial Route with a stochastic Where the ∆c is the change in cost, the θ is the current temper-
insertion crossover operator that looks at neighborhood offspring ature, and α is the rate of temperature change.
8
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 5. Two selected offspring and selected customers.

Fig. 6. Removal of Customers.

Fig. 7. Best-Aerial Route with stochastic Insertion Crossover Operator.

Fig. 8. Neighborhood Operators for SA Algorithm.

5. Proposed UAV-oriented MinMin route scheduling algorithm starts processing aerial routes as soon as possible without adding
any arbitrary delay [35]. The steps are given in Algorithm 1.
This section discusses the proposed route-scheduling algo- The algorithm creates an empty queue Q k , empty the priority
rithm, i.e., UAV-Oriented MinMin (UO-MinMin) for the aerial queue and computes the expected ready time URT k for every UAV
route-scheduling problem (Section 3.2). As discussed in Sec- uk ϵ U using Eq. (20). Next, the algorithm computes the expected
ready time RT z and the expected completion time CT z for each
tion 3.2, the set of M aerial routes is known as a batch-of-routes
aerial route Az concerning every UAV uk ϵ U. Since, all UAVs are
(BoR) and represented as J = {Az : 1 ≤ z ≤ M }. Every aerial
identical, the completion times remain similar concerning each
route Az is atomic in nature and independent. A fleet U of m
UAV initially; therefore, all completion times are added to all
homogeneous UAVs is represented by U = {uk : 1 = k = m} priority queues. The algorithm selects the aerial route with a
and it is assumed that every UAV uk visits the aerial routes in a minimum completion time for each UAV. Next, the pair of the
non-preemptive manner. aerial route Aw and UAV ul with a minimum value of the expected
The UO-MinMin route scheduling algorithm schedules the BoR ready time RT w is selected, and the tie is broken arbitrarily, if any.
J of M aerial routes on fleet U of m UAVs and maintains schedul- The aerial route Aw is then scheduled on UAV ul and the value of g
ing information on the Depot O0 . The proposed scheduling al- is fetched from the queue Ql and the scheduling information (w,
gorithm is an eager route scheduling algorithm, i.e., the UAV g+1, CTw ) is enqueued in queue Ql . The value of Yzl is also set to
9
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

1, and the scheduled aerial route Aw is removed from the set of and other constraints of the UAV-Assisted delivery system. The
unscheduled routes J. The scheduling of aerial route Aw on UAV characteristics of the dataset P are given in Table 2. i.e., it consists
ul only changes the expected ready time of only UAV ul without of 24 instances, the geometric coordinates are assumed to be in
affecting the expected ready times of other UAVs. Therefore, the the form of (x.10−1 km, y.10−1 km), and the distances of nearest
expected completion times for all unscheduled aerial routes are and farthest customers from the depot are 0.224 km and 4.992
computed with respect to the recently selected UAV ul . This km, respectively. The demands of customers are generated as
procedure is repeated until all aerial routes are scheduled. follows:
The proposed algorithm works based on the UAV-oriented σi %L σi %L ̸= 0
{
view rather than the aerial route-oriented view [36]. In an aerial ri = (30)
1.5 Other w ise
route-oriented view, the expected completion times for all un-
scheduled aerial routes are computed for all UAVs after the Where ri represents the demand of the customer oi for the UAV
scheduling of a single aerial route Aw and UAV ul . The ex- routing problem, L is the capacity of the UAV and σi is the demand
pected completion times for all unscheduled aerial routes will be of the customer oi in the problem instance of set P.
changed with respect to chosen UAV ul and it does not affect the Table 3 lists all parameters of UAVs, including battery capacity,
expected completion times of other UAVs. In the UAV-oriented recharging time, and time required for take-off, landing, loading,
view, the expected ready time for the UAV whose ready time and dropping a package at the customer location. The parameter
changes at scheduling decisions will be computed. Therefore, values have been selected from literature [10,43,63].
the UAV-oriented view removes the unnecessary computational The proposed HGSA routing algorithms, and the peer algo-
efforts of the algorithm without having any effect on the quality rithms, i.e., GA, PSO-SA, DE-SA, and HHO, are developed and
of the makespan. tested using Python 3.6.9 on Desktop Intel i7, core processor
at 2.30 GHz, and 8 GB RAM with Windows 10 OS. The perfor-
mance of HGSA is compared with GA, PSO-SA, DE-SA, and HHO
6. Simulation study
algorithms in terms of total travel time, total UAV trips, and
total energy consumed. For each instance out of 24 instances,
The simulation study is divided into five sub-sections. Sec-
the population size and the number of iterations are 100. The
tion 6.1 presents systems parameters and the peer algorithms.
developed HGSA UAV-routing algorithm, GA, PSO-SA, and DE-SA,
Section 6.2 conducts the sensitivity analysis of the hybridization
are executed ten times for each instance. The best travel time,
probability of GA and SA in the proposed HGSA algorithm. Sec- mean values, and standard deviations are recorded.
tion 6.3 discusses the results of UAV-routing algorithms, i.e., HGSA, The PSO-SA algorithm represents the amalgamation of particle
GA, PSO-SA, DE-SA, and HHO algorithms. Section 6.4 discusses swarm optimization (PSO) and simulated annealing (SA) algo-
the results offered by UAV-route scheduling algorithms, i.e., UAV- rithms [33,34,37]. In PSO-SA, the particle’s search space bounds
Oriented MinMin, MCT, and OLB heuristics algorithms. Section 6.5 are −10 to +10, and the inertia weight ψ and acceleration coef-
puts the observations of the study. ficients χ 1 and χ 2 are taken as 0.09, 0.01, and 0.01, respectively.
The DE-SA algorithm represents the hybrid version of differential
6.1. System parameters and peer algorithms evolution (DE) and simulated annealing (SA) algorithms [38].
In DE-SA, the bounds for agents are (−10, 10), while the mu-
This section tests the proposed algorithms on benchmark tation factor and recombination rate are taken as 0.3 and 0.9,
dataset P proposed by Augerat et al. [73]. The problem instance respectively. Harris-Hawks optimization (HHO) is a continuous-
P-n76-k4 consists of 76 nodes, including one depot and 75 cus- valued algorithm that finds approximate solutions to the prob-
tomers, whereas the depot is considered the UAV control center lem by repeatedly improving the candidate solution using ex-
(CC). The dataset P has been adjusted to meet the distance, flow, ploratory and exploitative phases, surprise pounce, and different
10
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Table 2
Characteristics of Problem Instances in Set P.
Set Total geometric coordinates Demand of Customer Distance of Nearest Node from Depot Distance of Farthest Node from Depot
Instances (10−1 km, 10−1 km) (Kg) (10−1 km) (10−1 km)
Set P 24 (xi , yi ) As per eq. (30) 2.24 49.92

Table 3
UAV’s Parameters.
Parameter Symbol Value
UAV Weight W 7.5 kg.
Payload Capacity of UAV L 2.5 kg.
conversion efficiency of motor and propeller ϕ 0.5
Lift-to-drag ratio γ 3
Power consumption of electronics/Incidental processes pe 0.1 kWh
Battery Capacity B 1.7 kW
Maximum rate of Power PH 0.6 kW
Recharging Time τc 1.25 h
Take-off/landing/loading/dropping packages τ 0.15 h
Fig. 9. Monte Carlo Procedure for Sensitivity Analysis of input probability.

Table 4
System parameters.
HGSA were conducted 100 times for each input probability and
Parameter Values
each problem instance from set P. For each problem instance, the
Population Size 200 value for the RandV variable in the HGSA was chosen from the
Generations 100
set SV, and ten simulations were conducted for input probability.
GA Crossover Operator Probability 0.9
GA Mutation Operator Probability 0.3 The mean values of travel time were recorded for the analysis.
Number of Experiments for every instance 10 As observed from the sensitivity analysis results, the mean
Objective Value Chosen (in Table) Mean, Standard Deviation travel times for all input probabilities remain almost close for
Aerial Route Chosen (Fig. 12) Best problem instances with few customers, i.e., for problem instances
P-n16-k8, P-n19-k2, P-n20-k2, and P-n21-k2. It is due to very
few customers in these problem instances. Fig. 10 shows the
attacking strategies [39]. In HHO, the exploratory and exploita- mean travel times offered by HGSA for each input probability
tive phases correspond to global search and local neighborhood corresponding to problem instances P-n50-k10 and P-65-k10. It
search, respectively. It consists of Lévy flights to realize the ex- can be observed that the HGSA offered a minimum mean travel
ploitative behavior of the HHO. The Lévy flights are realized time for 0.5 value of input probability for both instances. A more
using the same local search operators as in the HGSA algo- or less similar pattern was observed for all other remaining prob-
rithm, and this is why not hybridize HHO with SA. The three lem instances, and the graphical results for only two instances
algorithms, i.e., PSO-SA, DE-SA, and HHO, employ a random- are shown in Fig. 10 due to space constraints. It was noticed
key method to convert the real-valued solution into permutation that the proposed HGSA algorithm performs significantly better
due to the continuous-valued nature of the solutions [74]. Ta- for input probability 0.5 for all problem instances of set P. The
ble 4 presents the parameters required for the proposed HGSA sensitivity analysis ensures to use of 0.5 hybridization probability
algorithm. The parameter values are selected from the literature to combine GA and SA in the proposed HGSA algorithm.
[32–34].
6.3. UAV routing results
6.2. Sensitivity analysis
This section presents the results offered by HGSA, GA, PSO-
The proposed HGSA algorithm combines GA and SA to gener- SA, DE-SA, and HHO for 24 instances of dataset P. The graphical
ate a new population at each iteration. The GA and SA algorithms results for only three instances are shown in Figs. 11, 12, and
are global and local optimization algorithms, respectively, i.e., GA 13 due to space constraints, and the mean results for all 24
explores new areas of the solution domain, and SA exploits al- instances are provided in Table 6. Fig. 11 shows the five algo-
ready explored areas of the solution space. Thus, GA and SA rithm’s results for the problem instance P-n55-k15, consisting
algorithms are hybridized in HGSA to balance the exploration of 55 nodes (including 54 customers and depot (UAV CC)) and
and exploitation of solution space. The performance of HGSA is their geometric coordinates. The demand was generated using
dependent on the input probability (RandV) corresponding to GA Eq. (30) and demand given in problem instance for each customer.
and SA, as shown in Fig. 3. Therefore, it is necessary to conduct a Fig. 11(a) depicts the locations of UAV CC and customers in the
sensitivity analysis of the proposed HGSA concerning the input 2D geometric plane. Fig. 11(b)–(f) represent the total aerial routes
probability. To evaluate the proposed HGSA algorithm, Monte offered by HGSA, GA, PSO-SA, DE-SA, and HHO. As can be seen,
Carlo (MC) procedure is employed to conduct the sensitivity anal- the aerial routes offered by HGSA have minimum overlapping
ysis [75–77]. MC is a random sampling and statistical modeling of routes compared to others. It ensures that the HGSA offers
to simulate the operations of complex systems. MC procedure more efficient aerial routes, which leads to minimum possible
repeats the simulation of a complex system many times using travel time, aerial routes, and energy consumption. For ten runs,
inputs generated from suitable probability density functions and the best travel times provided by HGSA, GA, PSO-SA, DE-SA,
then analyzes the results. Fig. 9 represents the MC procedure for and HHO algorithms are 1222.37, 1258.39, 1269.60, 1270.23, and
the proposed HGSA algorithm over set P to conduct the sensitivity 1344.87 min, respectively, and the energy consumption by the
analysis of input probability. First of all, the input probability was five algorithms for the same solution are 4.49, 4.68, 4.61, 4.66,
sampled using a discrete uniform distribution in the interval [0, and 5.29 kW, respectively. The total number of aerial routes in the
1], and all sampled values are given as SV = [0, 0.1, 0.2, 0.3, 0.4, best solution offered by HGSA, GA, and DE-SA is 26, and provided
0.5, 0.6, 0.7, 0.8, 0.9, 1]. Next, the simulations of the proposed by PSO-SA and HHO is 27 and 28 aerial routes. Fig. 11(g) and (h)
11
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 10. Mean Travel time for 11 sampled Values in [0, 1] for problem instances P-n50-k10 and P-n65-k10.

present the mean travel time and mean energy consumption for proposed HGSA algorithm can easily be employed for real-world
100 generations. As it can be seen, the HGSA continuously offers UAV-assisted delivery systems.
the minimum travel times and minimum energy consumption Tables 5 and 6 present the results of all 24 problem instances
compared to GA, PSO-SA, DE-SA, and HHO algorithms. in terms of mean travel time, mean energy consumption, and
Fig. 12 shows the results offered by five algorithms for the standard deviations for ten runs of each algorithm. Out of 24
problem instance P-n76-k4. This problem instance consists of problem instances, the first problem instance consists of 16 nodes
76 nodes (including 75 customers and a depot (UAV CC)), their (including 15 customers and UAV CC), whereas the last problem
geometric coordinates, and the demands. The actual demand is instance consists of 101 nodes (including 100 customers and UAV
generated using Eq. (30) for each customer. Fig. 12(a) depicts the CC). To highlight the best results, the magenta color has been
locations of UAV CC and customers in the 2D geometric plane. used. As it can be seen, the proposed HGSA algorithm offers the
Fig. 12(b)–(f) show the total aerial routes offered by HGSA, GA, minimum mean travel time and minimum energy consumption
PSO-SA, DE-SA, and HHO algorithms. As it is visible, the aerial compared to all the four algorithms, i.e., HGSA, GA, DE-SA, PSO-
routes offered by HGSA have minimum overlapping of routes SA, and HHO for all 24 instances. The mean travel time and
compared to others. It ensures that the HGSA offers more efficient standard deviation offered by the HGSA algorithm are low com-
aerial routes, which leads to minimized travel time, aerial routes, pared to all peer algorithms. The low standard deviation ensures
and energy consumption. The best travel times provided by HGSA, the consistent performance of the algorithm, i.e., the proposed
GA, PSO-SA, DE-SA, and HHO algorithms are 1762.17, 1805.37, HGSA is more consistent in comparison to GA, DE-SA, PSO-SA,
1801.60, 1793.43, and 1933.23 min, respectively, and the energy and HHO algorithms. Table 5 also shows the energy consumption
consumption by the five algorithms correspond to the same so- corresponding to five algorithms and 24 problem instances. As it
lution are 6.503, 6.89, 6.69, 6.68, and 7.87 kW, respectively. The can be observed that the mean energy consumption and standard
total number of UAV aerial routes offered by HGSA, GA, and PSO- deviation offered by the HGSA algorithm are low. It ensures
SA is 37, whereas the DE-SA and HHO offered 38 and 39 UAV the superiority of the proposed HGSA algorithm compared to all
aerial routes. Fig. 12(g) and (h) present the mean travel time and peer algorithms. The performance of five algorithms, i.e., HGSA,
energy consumption over 100 generations. As it can be seen, the GA, PSO-SA, DE-SA, and HHO, is almost the same for problem
HGSA unceasingly offers the minimum travel times and minimum instances P-n16-k8, P-n19-k2, P-n20-k2, and P-n21-k2. It is due
energy consumption compared to the peer algorithms. to very few customers in these problem instances. For the re-
Fig. 13 presents the results produced by five algorithms, i.e., maining problem instances, the performance of HGSA is better
HGSA, GA, PSO-SA, DE-SA, and HHO, for the problem instance in comparison to the remaining algorithms. The gap between
P-n101-k4. This problem instance consists of 101 nodes (in- the travel times and energy consumption offered by the five
cluding 100 customers and one depot (UAV CC)), geometric co- algorithms increases with the increasing number of customers.
ordinates, and demands. Fig. 13(a) illustrates the problem in- The performance order is either ‘‘HGSA, DE-SA, GA, PSO-SA and
stance. Fig. 13(b)–(f) depict the total aerial routes offered by HHO’’ or ‘‘HGSA, GA, DE-SA, PSO-SA, and HHO’’. The HHO is the
HGSA, GA, PSO-SA, DE-SA, and HHO algorithms. The overlapping worst performer, whereas the performance of GA and DE-SA is
in the aerial routes offered by HGSA is minimum compared to conflicting, i.e., GA performs better than DE-SA for some problems
others, ensuring that the HGSA offers more efficient aerial routes. instances and vice-versa. In both cases, the HGSA is the best
The best travel times provided by HGSA, GA, PSO-SA, DE-SA, performer among the group of five algorithms as it offers the
and HHO algorithms are 2373.18, 2467.17, 2431.36, 2426.02, minimum mean travel time, minimum mean energy consump-
and 2616.10 min, respectively, and the energy consumption by tion, and low standard deviations for all 24 problem instances.
the four algorithms correspond to the same solution are 8.52, It can be observed that the proposed HGSA algorithm performs
9.31, 8.90, 8.93, and 10.42 kW, respectively. The total number almost equivalent or better compared to the peer algorithms
of UAV aerial routes offered by HGSA, GA, and DE-SA is 52, for problem instances with very few customers. On the other
whereas the PSO-SA and HHO offered 53 and 57 UAV aerial hand, the proposed HGSA algorithm performs significantly better
routes. Fig. 13(g) and (h) present the mean travel time and mean than peer the algorithms for the problem instances with a higher
energy consumption for 100 generations over ten runs of each number of customers.
algorithm. The HGSA incessantly offers the minimum mean travel Table 7 presents aerial routes offered by the five algorithms,
time and minimum mean energy consumption compared to GA, i.e., HGSA, GA, DE-SA, PSO-SA, and HHO, for 24 problem instances.
PSO-SA, DE-SA, and HHO algorithms. The HGSA algorithm offers minimum aerial routes compared to
It can be observed from Figs. 11(g), 12(g), and 13(g), the GA, PSO-SA, DE-SA, and HHO algorithms for all problem instances.
proposed HGSA algorithm offered the best solution in very few It can be observed from Table 7 that all algorithms offer an equal
generations in comparison to the peer algorithms. Hence, the number of routes for problem instances with a smaller number
12
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 11. For problem instance P-n55-k15 (a) Problem Instance (b)–(f) Routes offered (g) Travel Time (h) Energy Consumption.

of customers. The aerial routes provided by the HGSA, GA, DE- the neighbor customers are clustered and served by the minimum
SA, PSO-SA, and HHO are 52, 52, 52, 53, and 57, respectively, for number of UAVs. The reason behind this performance of HGSA is
the problem instance P-n101-k4. It is also visible from Figs. 11, the balance between exploration and exploitation in two phases
12, and 13 that the aerial routes offered by HGSA have minimum of the HGSA algorithm as it employs the hybridization of GA
overlapping. The minimum overlap of aerial routes ensures that and SA algorithms. GA algorithm searches for the globally best
13
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 12. For problem instance P-n76-k4 (a) Problem Instance (b)–(f) Routes offered (g) Travel Time (h) Energy Consumption.

solution, and SA looks for the best neighborhood solution and employ SA in the second phase to perform neighborhood search;
reduces overlapping aerial routes using 2-Opt*, Exchange, and however, both of them cannot beat the performance of HGSA for
relocate operators. The peer algorithms PSO-SA and DE-SA also all instances of the set P.
14
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 13. For problem instance P-n101-k4 (a) Problem Instance (b)–(f) Routes offered (g) Travel Time (h) Energy Consumption.

On the other hand, the HHO algorithm employs the SA opera- a permutation. The issue with the decoding methods is that an
tors during Lévy flights in the exploitation phase, even though it infinite number of continuous-valued vectors can be decoded to
cannot beat the performance of HGSA for any problem instance. a single discrete solution because of cardinality; thus, continuous-
The continuous algorithms, i.e., PSO-SA, DE-SA, and HHO, employ valued algorithms may need to explore bulky plateaus in the
the random-key method [74] to convert a real-valued vector into fitness space. This leads to a wastage of computational efforts.
15
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Table 5
Travel Time (Minutes): Set P (µ-Mean, ψ -Standard Deviation).
Problem Instances HGSA GA PSO-SA DE-SA HHO
µ ψ µ ψ µ ψ µ ψ µ ψ
P-n16-k8 348.66 0.0 348.66 0.0 348.66 0.0 348.66 0.0 348.66 0.0
P-n19-k2 442.96 0.0 442.96 5.6e−14 443.42 0.39 443.58 0.11 451.73 11.05
P-n20-k2 474.45 0.0 474.46 0.0 475.06 0.86 474.99 0.911 477.92 1.83
P-n21-k2 484.05 6.56e−14 484.05 0.0 484.29 0.399 484.36 0.354 490.22 5.37
P-n22-k2 494.23 0.0 494.43 0.229 494.31 0.135 494.38 0.26 497.59 2.04
P-n22-k8 517.84 0.0 518.89 1.28 529.26 9.89 519.33 1.29 531.52 8.26
P-n23-k8 524.91 0.0 525.06 0.135 524.98 0.135 524.98 0.135 528.27 0.933
P-n40-k5 947.07 0.0 950.54 0.642 954.48 10.52 953.58 10.38 994.27 2.84
P-n45-k5 1060.67 0.398 1069.84 1.36 1067.63 9.48 1066.0559 2.33 1128.82 11.26
P-n50-k7 1165.644 0.662 1174.964 2.074 1170.998 3.282 1173.024 1.389 1246.784 16.095
P-n50-k8 1157.65 7.17 1173.164 4.610 1169.76 1.25 1168.76 2.792 1241.93 24.038
P-n50-k10 1155.19 4.26 1174.003 7.596 1165.008 8.971 1173.14 2.293 1232.91 12.88
P-n51-k10 1176.98 7.20 1192.95 10.529 1195.522 1.209 1193.775 5.662 1279.59 19.88
P-n55-k7 1247.30 1.145 1280.33 6.953 1267.213 4.114 1270.029 4.667 1342.89 19.428
P-n55-k8 1246.318 0.786 1273.226 3.425 1270.473 2.967 1272.593 5.985 1343.672 13.450
P-n55-k10 1248.52 1.56 1278.363 7.338 1278.089 7.338 1270.995 2.2449 1342.273 10.210
P-n55-k15 1246.38 0.472 1279.19 4.091 1273.888 8.653 1275.77 9.090 1352.743 18.972
P-n60-k10 1389.371 3.0412 1421.147 4.359 1409.646 23.259 1405.922 10.349 1510.673 22.1005
P-n60-k15 1388.101 1.147 1419.9114 5.246 1410.896 3.314 1412.225 5.2133 1502.118 14.367
P-n65-k10 1512.465 9.242 1549.114 3.956 1543.493 18.947 1538.222 2.129 1662.754 1.794
P-n70-k10 1626.879 1.833 1668.384 2.963 1667.807 4.899 1657.904 13.695 1777.3598 12.394
P-n76-k4 1765.75 3.1362 1810.574 10.789 1805.664 6.164 1798.900 18.394 1950.257 4.070
P-n76-k5 1770.282 4.0699 1820.528 8.653 1805.3515 3.585 1796.525 10.458 1961.773 7.209
P-n101-k4 2383.7206 0.8134 2482.156 13.3972 2441.730 12.675 2437.0725 8.2964 2666.101 18.953

Table 6
Energy Consumed (Kw) Set P, (µ-Mean, ψ -Standard Deviation).
Problem Instances HGSA GA PSO-SA DE-SA HHO
µ ψ µ ψ µ ψ µ ψ µ ψ
P-n16-k8 1.16 1.16e−16 1.16 1.16e−16 1.16 1.16e−16 1.16 1.16e−16 1.16 1.16e−16
P-n19-k2 1.57 1.28e−16 1.57 1.28e−16 1.58 0.003 1.58 0.001 1.61 0.024
P-n20-k2 1.66 1.28e−16 1.67 1.28e−16 1.68 0.009 1.67 0.009 1.70 0.018
P-n21-k2 1.69 1.28e−16 1.69 2.22e−16 1.69 0.003 1.69 0.003 1.75 0.053
P-n22-k2 1.72 1.28e−16 1.72 0.002 1.72 0.001 1.72 0.002 1.75 0.020
P-n22-k8 2.10 2.56e−16 2.11 0.012 2.12 0.012 2.12 0.0128 2.24 0.084
P-n23-k8 1.80 1.28e−16 1.80 0.001 1.80 0.001 1.80 0.001 1.83 0.009
P-n40-k5 3.40 2.56e−16 3.43 0.006 3.42 0.022 3.41 0.017 3.77 0.067
P-n45-k5 3.86 0.004 3.95 0.0135 3.88 0.012 3.91 0.023 4.438 0.102
P-n50-k7 4.23 0.007 4.324 0.020 4.29 0.032 4.305 0.014 4.94 0.104
P-n50-k8 4.25 0.022 4.36 0.042 4.27 0.013 4.26 0.028 4.94 0.154
P-n50-k10 4.27 0.042 4.365 0.040 4.28 0.0147 4.31 0.0229 4.804 0.0423
P-n51-k10 4.21 0.146 4.329 0.022 4.305 0.012 4.287 0.056 4.895 0.122
P-n55-k7 4.52 0.011 4.753 0.063 4.57 0.041 4.60 0.046 5.228 0.1857
P-n55-k8 4.513 0.0078 4.68 0.0595 4.60 0.029 4.63 0.0598 5.236 0.0877
P-n55-k10 4.535 0.015 4.68 0.035 4.630 0.0133 4.609 0.022 5.172 0.102
P-n55-k15 4.513 0.0047 4.742 0.0459 4.618 0.861 4.677 0.0254 5.327 0.1822
P-n60-k10 5.118 0.0304 5.386 0.0435 5.221 0.060 5.234 0.061 6.031 0.170
P-n60-k15 5.106 0.0114 5.374 0.0466 5.1839 0.0331 5.197 0.0521 5.996 0.067
P-n65-k10 5.624 0.006 5.941 0.0647 5.784 0.1039 5.7822 0.0213 6.777 0.0898
P-n70-k10 5.993 0.0183 6.408 0.0296 6.253 0.04899 6.204 0.0621 7.298 0.1758
P-n76-k4 6.5173 0.0361 6.930 0.1078 6.7316 0.0616 6.764 0.098 7.977 0.04766
P-n76-k5 6.527 0.0406 7.030 0.0865 6.778 0.0875 6.740 0.0181 7.992 0.072
P-n101-k4 8.5372 0.0081 9.33715 0.1426 8.917 0.0401 8.9707 0.0332 10.51 0.0484

6.4. UAV route-scheduling results The proposed UO-MinMin, MCT, and OLB heuristic algorithms
are also developed and tested using Python 3.6.9 programming
Table 5 presents the mean travel times and standard devia- languages on Desktop Intel i7, core processor at 2.30 GHz, and
tions, while the total number of aerial routes is given in Table 7. 8 GB RAM with Windows 10 OS. For all 24 instances, the set of
As observed from Tables 5 and 7, the mean travel time and all aerial routes corresponding to the best solution generated by
total aerial routes offered by the proposed HGSA for the prob- HGSA is considered batch-of-aerial routes, i.e., all aerial routes are
lem instance P-n101-k4 consisting of 101 nodes (including 100 considered tasks that need to be scheduled on available UAVs. The
customers and one UAV CC) are 39.73 h (2383.7206 min) and travel times offered by HGSA for all aerial routes are regarded
52, respectively. If one UAV serves all 100 customers, it needs to as the time required to execute the aerial routes. The times
visit 52 aerial routes and takes approximately 39.20 h. Therefore, provided by HGSA for aerial routes are based on the problem
it demands multiple UAVs to serve all customers in a feasible instance, which is static and based on the travel time between
time per day, which ultimately requires an aerial route schedul- two points without considering real-time effects. Therefore, the
ing algorithm. The aerial-route scheduling problem formulation time required for each aerial route is augmented by a random
and UO-MinMin algorithm are explained in Sections 3.2 and 5, time factor from 1 to 15 min to simulate the real conditions.
respectively. To evaluate the performance of the UO-MinMin algorithm MCT
16
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Table 7 are approximately equal to the UO-MinMin algorithm. It can be

Aerial Routes for Set P.
established that UO-MinMin consistently performs better than
Instances HGSA GA PSO-GA DE-SA HHO
MCT and OLB algorithms. The reason behind the excellent per-
P-n16-k8 8 8 8 8 8 formance of the UO-MinMin is that it considers all unscheduled
P-n19-k2 10 10 10 10 10
P-n20-k2 11 11 11 11 11
aerial routes and UAVs at every scheduling decision and sched-
P-n21-k2 11 11 11 11 11 ules the aerial route, which will finish its execution as early
P-n22-k2 11 11 11 11 11 as possible. The MCT algorithm considers only one unscheduled
P-n22-k8 10 10 10 10 10 aerial route and all UAVs at every scheduling decision. It ignores
P-n23-k8 12 12 12 12 12
P-n40-k5 21 21 21 21 21
different aerial routes from consideration, which leads to its
P-n45-k5 23 23 23 23 23 poor performance compared to the UO-MinMin algorithm. The
P-n50-k7 25 25 25 25 25 OLB algorithm considers only one aerial route and one UAV at
P-n50-k8 24 24 25 25 25 every scheduling decision, which leads to its poor performance
P-n50-k10 24 24 24 25 25
P-n51-k10 25 25 26 26 27
compared to UO-MinMin and MCT algorithms.
P-n55-k7 26 26 27 27 28 It can be observed from Table 7 the time required to serve
P-n55-k8 26 26 27 27 27 all 100 customers corresponding to problem instance P-n101-
P-n55-k10 26 27 27 27 28 k4 are 960 min (16 h), 572 min (9.53 h), 384 min (6.4 h), and
P-n55-k15 26 26 27 26 28
P-n60-k10 29 29 29 29 31
314 min (5.2 h) using 3, 5, 8 and 10 UAVs, respectively. For
P-n60-k15 29 29 30 30 30 the problem instance P-n76-k5, the time required to serve all
P-n65-k10 31 31 32 32 34 75 customers is 729 min (12.15 h), 448 min (7.47 h), 295 min
P-n70-k10 34 34 35 34 35 (4.92 h), and 238 min (3.97 h) using 3, 5, 8, and 10 UAVs, respec-
P-n76-k4 37 37 38 37 39
P-n76-k5 37 37 38 37 40
tively. A similar pattern can also be observed for the remaining
P-n101-k4 52 52 53 52 57 problem instances. It can be established that it is required to em-
ploy the routing & route-scheduling algorithms to handle UAV-
assisted delivery systems. Also, a minimum of 5 and 15 UAVs
are required for medium-scale and large UAV-assisted delivery
and OLB, the numbers of UAVs are considered 10, 8, 5, and 3.
systems, respectively.
The UO-MinMin, MCT, and OLB algorithms are static algorithms,
and therefore, these algorithms are executed one time for each
instance, and the solution is recorded. Due to space constraints, 6.5. Observations and research directions
the results for three problem instances are given using stacked
lines in Figs. 14, 15, and 16, and the results for the remaining • It was observed from the simulation study that the pro-
instances are provided in Table 8. posed HGSA algorithm unceasingly performs better than the
Fig. 13 presents the makespan offered by the UO-MinMin, peer algorithms. The HGSA algorithm offers minimum mean
MCT, and OLB algorithms for the problem instance P-n55-k8. This travel time and minimum energy consumption compared to
problem instance consists of 55 nodes (including 54 customers all the four algorithms, i.e., GA, DE-SA, PSO-SA, and HHO,
and a depot (UAV CC)), their geometric coordinates, and the for all 24 instances of set P. The low mean and standard
demands. The actual demand is generated using Eq. (30) for deviation provided by HGSA ensure that the proposed algo-
each customer. As given in Table 6, the number of aerial routes rithm is more consistent than GA, DE-SA, PSO-SA, and HHO
generated by HGSA is 26 for problem instance P-n55-k8. These algorithms. The performance order from best to worst is
26 aerial routes are scheduled using UO-MinMin, MCT, and OLB HGSA, GA, DE-SA, PSO-SA, and HHO.
algorithms on 3, 5, 8, and 10 UAVs. It is visible that UO-MinMin • It was observed that the aerial routes offered by HGSA
performs better than the MCT and OLB algorithms as it offers a have minimum overlapping of routes compared to others.
lower makespan compared to the peer algorithms. For the prob- The minimum overlap of aerial routes ensures that HGSA
lem instance P-n76-k4, Fig. 15 depicts the makespan offered by generates aerial routes by combining the nearby customers.
the UO-MinMin, MCT, and OLB algorithms. This problem instance
Thus, HGSA offers more efficient aerial routes to minimize
consists of 76 nodes (including 75 customers and a depot (UAV
travel time, aerial routes, and energy consumption.
CC)), their geometric coordinates, and the demands. The number
• The proposed HGSA algorithm performs better due to the
of aerial routes generated by HGSA is 37 for problem instance P-
balance between the exploration and exploitation of the two
n76-k4. These 37 aerial routes are scheduled using UO-MinMin,
phases. The HGSA employs the hybridization of GA and SA
MCT, and OLB algorithms on 3, 5, 8, and 10 UAVs. As shown in
algorithms. In the first phase, the GA algorithm searches
Fig. 13, the UO-MinMin performs better than the MCT and OLB
for the global best solution using a stochastic crossover
heuristic algorithms. For the problem instances P-n01-k4, Fig. 16
depicts the makespan offered by the UO-MinMin, MCT, and OLB operator. In the second phase, SA removes overlapping of
algorithms. This instance consists of 101 nodes (including 100 aerial routes using 2-Opt*, Exchange, and relocate operators.
customers and a depot (UAV CC)), their geometric coordinates, • It was perceived from the mean travel times and total aerial
and the demands. The HGSA algorithm generates 52 aerial routes routes for many problem instances that it is not possible
scheduled using UO-MinMin, MCT, and OLB algorithms on 3, 5, 8, to serve 40 or more customers using one UAV in a single
and 10 UAVs. As shown in Fig. 16, the UO-MinMin again performs day. Therefore, it establishes the need for route-scheduling
better than the MCT and OLB algorithms. algorithms for UAV-assisted delivery systems.
Table 8 puts the makespan offered by the UO-MinMin, MCT, • It was observed from the simulation study that the proposed
and OLB algorithms for all 24 instances. To highlight the best UO-MinMin algorithm performs better compared to MCT
results, the magenta color has been used. Table 7 shows that the and OLB algorithms for all problem instances of set P. The
UO-MinMin offers the best makespan compared to the MCT and reason behind the excellent performance of the UO-MinMin
OLB for all 24 problem instances. For problem instances P-n16-k8, is that it considers all unscheduled aerial routes and UAVs
P-n19-k2, P-n20-k2, P-n21-k2, P-n50-k8, P-n55-k7, P-n55-k15, at every scheduling decision and schedules the aerial route,
and P-n60-k15, the performance of MCT and OLB algorithms which will finish its execution as early as possible.
17
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Fig. 14. Makespan offered by UO-MinMin, MCT, and OLB for problem instance P-n55-k8.

Fig. 15. Makespan offered by UO-MinMin, MCT, and OLB for problem instance P-n76-k4.

Fig. 16. Makespan offered by UO-MinMin, MCT, and OLB for problem instance P-n101-k4.

• The proposed HGSA algorithm can be investigated in many • The reinforcement learning algorithms (SARSA, Q-Learning,
directions, including specialized initialization methods, du- etc.) can also be used to improve the performance of the
plicate solutions handling, local search operators, crossover proposed algorithm, operators, hybridization, and other com-
operators, etc. It was observed that the quality of aerial ponents of the algorithm.
routes is highly dependent on the quality of initial solu- • The UAV routing problem could also be investigated in dif-
tions. Therefore, intelligent initialization heuristics, i.e., dy- ferent directions, such as time windows, multiple UAV con-
namic, delayed, complementary initialization, could be used trol, recharge centers, multiple echelons, pick-up, multi-
to improve convergence speed and solution accuracy. Clever mode transportation mode, multi-objectives, large-scale op-
heuristics can be developed to handle duplicate solutions timization, traffic prediction, etc. [29,78]. The proposed al-
during algorithm evolution. Moreover, intelligent static and gorithms can also be extended for context-aware routing
stochastic local search, crossover, and mutation operators and scheduling, routing with fog node selection for UAV
could also be designed. communication, and others.
18
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Table 8
Makespan Results.
INSTANCES UO-MinMin MCT OLB
#UAV #UAV #UAV
3 5 8 10 3 5 8 10 3 5 8 10
P-n16-k8 154 100 58 52 157 102 61 58 158 101 58 61
P-n19-k2 206 108 103 59 211 118 105 59 209 120 103 61
P-n20-k2 205 112 99 93 219 125 99 97 219 128 99 93
P-n21-k2 203 147 100 97 211 157 105 97 217 151 117 99
P-n22-k2 206 148 101 93 209 151 106 97 216 150 111 97
P-n22-k8 224 134 117 64 234 138 120 67 234 139 121 68
P-n23-k8 215 151 100 91 220 155 104 91 222 157 107 102
P-n40-k5 363 251 167 143 403 263 176 157 409 261 176 156
P-n45-k5 443 275 177 162 457 284 183 169 462 291 190 175
P-n50-k7 460 268 199 156 467 271 206 168 478 272 207 176
P-n50-k8 440 286 168 162 474 299 177 162 449 293 168 172
P-n50-k10 457 269 174 168 467 288 193 170 474 302 184 179
P-n51-k10 456 270 190 160 469 283 197 175 469 293 210 176
P-n55-k7 474 311 204 161 497 333 215 161 498 326 212 172
P-n55-k8 511 325 213 173 521 333 220 179 524 329 226 189
P-n55-k10 503 338 231 182 534 353 246 190 536 363 245 206
P-n55-k15 496 318 209 172 524 328 209 173 517 332 215 181
P-n60-k10 541 329 216 166 573 342 218 175 566 344 226 189
P-n60-k15 564 349 237 189 581 355 245 195 576 355 249 189
P-n65-k10 556 358 220 206 590 372 233 212 586 380 239 209
P-n70-k10 635 366 257 210 645 385 265 212 637 409 262 218
P-n76-k4 706 439 276 233 718 441 280 244 733 445 282 246
P-n76-k5 729 448 295 238 734 464 303 256 744 467 306 245
P-n101-k4 960 572 384 314 1001 605 390 319 1009 619 396 324

• The UAV routing and route-scheduling problems are than GA, PSO-SA, DE-SA, and HHO algorithms for all 24 instances.
permutations-based problems. The continuous-valued al- The performance order from best to worst is HGSA, GA, DE-SA,
gorithms like PSO and HHO require a decoding method PSO-SA, and HHO. The aerial routes offered by HGSA have min-
that enhances the overall time complexity. Therefore, it imum routes overlapping compared to others. Therefore, HGSA
is necessary to evolve new meta-heuristics which work offers more efficient aerial routes, minimizing travel time, aerial
exclusively for permutation-based problems. Similarly, the routes, and energy consumption. The UO-MinMin algorithm is
operator-specific to permutation-based problems is an ex- compared with MCT and OLB algorithms. It is also observed that
cellent avenue to investigate. the proposed UO-MinMin consistently performs better than MCT
• The novel routing and scheduling model based on fuzzy- and OLB algorithms for all 24 instances. The UO-MinMin consid-
system, i.e., fuzzy-evolutionary algorithms, can also be de- ers all unscheduled aerial routes and UAVs at every scheduling
veloped, which could be realized in the fog-enabled cloud
decision and schedules the aerial routes to finish their execution
environment. A different research line is to build energy-
as early as possible.
efficient and quantum-aware routing and scheduling al-
gorithms for UAV-assisted delivery systems. The energy-
efficient routing and scheduling algorithms can be designed, 7.2. Future research directions
requiring minimum energy to produce aerial routes and
schedules for UAV-assisted delivery systems handling many It is anticipated that the proposed models, algorithms, and
customers. simulation setup will be helpful for future research on the UAV-
assisted delivery system. The proposed HGSA algorithm can be
7. Conclusion and future research directions
investigated in many research directions, such as employing spe-
cialized initialization methods, handling duplicate solutions, local
7.1. Conclusion
search operators, crossover operators, etc. The variants of the UAV
This paper proposes a joint-optimization framework to ad- routing problem could also be investigated, i.e., UAV routing with
dress UAV-routing and UAV-route scheduling problems associ- time windows, multiple UAV control and recharge centers, mul-
ated with a UAV-assisted delivery system. The proposed mixed- tiple echelons, pick-up, multi-mode transportation mode, multi-
integer linear programming (MILP) models for both problems objectives, large-scale optimization, etc. The proposed algorithms
consider the effect of incidental processes and the varying pay- can also be extended for context-aware routing and scheduling,
load. A hybrid genetic and simulated annealing (HGSA) is pro- routing with fog node selection for UAV communication, and
posed to address the UAV-routing problem. HGSA is the hy- others.
bridization of GA and SA to balance the exploration and ex-
ploitation behavior. GA explores the search space using stochastic
CRediT authorship contribution statement
crossover and mutation operators, whereas SA exploits the al-
ready searched space using 2-Opt*, Exchange, and relocate op-
erators. The proposed UO-MinMin algorithm uses a computa- Mohammad Sajid: Investigation, Methodology, Writing – orig-
tionally efficient UAV-oriented approach to handle the UAV-route inal draft, Writing – review & editing. Himanshu Mittal: Writing
scheduling problem to minimize the makespan. The performance – original draft, Writing – review & editing. Shreya Pare: Writing
of HGSA is compared with GA, PSO-SA, DE-SA, and HHO algo- – original draft, Writing – review & editing. Mukesh Prasad:
rithms using 24 well-known benchmark instances of set P. The Investigation, Methodology, Writing – original draft, Writing –
simulation study confirms that HGSA consistently performs better review & editing.
19
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

Declaration of competing interest [22] A. Altan, O. Aslan, R. Hacıoglu, Real-time control based on NARX neural
network of hexarotor UAV with load transporting system for path tracking,
in: IEEE 2018 6th International Conference on Control Engineering &
The authors declare that they have no known competing finan-
Information Technology (CEIT), Istanbul, Turkey, 25-27 Oct, 2018.
cial interests or personal relationships that could have appeared [23] A. Sunay, A. Altan, E. Belge, R. Hacioğlu, Investigation of route tracking
to influence the work reported in this paper. performance with adaptive PID controller in quadrotor, Eur. J. Tech. (EJT)
10 (1) (2020).
Data availability [24] A. Altan, O. Aslan, R. Hacıoglu, Model predictive control of load trans-
porting system on unmanned aerial vehicle (UAV), in: Proc. of the
Fifth International Conference on Advances in Mechanical and Robotics
Data will be made available on request. Engineering (AMRE), Rome, Italy, 2017.
[25] M. Sajid, Z. Raza, Energy-efficient quantum-inspired stochastic Q-hype
Acknowledgment algorithm for batch-of-stochastic-tasks on heterogeneous DVFS-enabled
processors, Concurr. Comput.: Pract. Exp. (Wiley) 31 (20) (2019) e5327.
[26] R.A. Haidri, M. Alam, M. Shahid, S. Prakash, M. Sajid, A deadline aware
All authors have read and agreed to the published version of
load balancing strategy for cloud computing, Concurr. Comput.: Pract. Exp.
the manuscript. (Wiley) (2021).
[27] M. Sajid, Z. Raza, Energy-aware stochastic scheduler for batch of
References precedence-constrained jobs on heterogeneous computing system, Energy
125 (2017) 258–274.
[1] United Nations Population Fund (UNFPA), State of World Population 2007 – [28] J.Y.-T. Leung, Handbook of Scheduling: Algorithms, Models, and Perfor-
Unleashing the potential of urban growth, 2007, United Nations Population mance Analysis, in: Computer and Information Science Series, CRC Press,
Fund, New York, USA. Boca Raton, FL, 2004.
[2] M. Grote, I. Williams, J. Preston, S. Kemp, Including congestion effects [29] P. Toth, D. Vigo, Vehicle Routing: Problems, Methods, and Applications,
in urban road traffic CO2 emissions modeling: Do local government second ed., Society for Industrial and Applied Mathematics, Philadelphia,
authorities have the right options? Transp. Res. D 43 (2016) 95–106. PA, 2015.
[3] Global status report on road safety, 2018, World Health Organization [30] M.R. Garey, D.S. Johnson, Computers and Intractability: A Guide to the
(WHO). Theory of NP-Completeness, W. H. Freeman & Co., New York, NY, USA,
[4] Mumbai’s congested roads are costing It – Both Time & Money, The Quint, 1990.
https://www.youtube.com/watch?v=02xFx7Z_wQ4. [31] J.-Y. Potvin, State-of-the-art review evolutionary algorithms for vehicle
[5] S. Juyal, A. Saxena, S. Sharma, A. Srivastava, NITI aayog & the Boston routing’, INFORMS J. Comput. 21 (4) (2009) 518–548.
consulting group, 2018, Transforming India’s Mobility: A perspective. [32] C.K.H. Lee, A review of applications of genetic algorithms in operations
[6] L. Wen, J. Kenworthy, X. Guo, D. Marinova, Solving traffic congestion management, Eng. Appl. Artif. Intell. 76 (2018) 1–12.
through street renaissance: A perspective from dense Asian cities, Urban
[33] S. Kirkpatrick, C.D. Gelatt Jr., M.P. Vecchi, Optimization by simulated
Sci. 3 (1) (2019).
annealing, Science 220 (4598) (1983) 671–680.
[7] A. Mor, M.G. Speranza, Vehicle routing problems over time: A survey.
[34] B. Rabbouch, F. Saâdaoui, R. Mraihi, Empirical mode simulated annealing
4OR-Q, J. Oper. Res. (2020).
for solving the capacitated vehicle routing problem, J. Exp. Theor. Artif.
[8] W. Shi, H. Zhou, J. Li, W. Xu, N. Zhang, X. Shen, Drone assisted vehicular
Intell. 32 (3) (2020) 437–452.
networks: Architecture, challenges, and opportunities, IEEE Netw. 32 (3)
(2018) 130–137. [35] M. Sajid, Z. Raza, M. Shahid, Energy efficient scheduling algorithms for
[9] S.C. Liu, Y.F. Zhang, Y. Liu, L.H. Wang, X.V. Wang, An internet of things batch-of-tasks (BoT) applications on heterogeneous computing systems,
enabled dynamic optimization method for smart vehicles and logistics Concurr. Comput.: Pract. Exper. 28 (9) (2015) 2664–2669.
tasks, J. Clean Prod. 215 (2019) 806–0820. [36] E.K. Tabak, B.B. Cambazoglu, C. Aykanat, Improving the performance of
[10] R. D’Andrea, Guest editorial can drones deliver? IEEE Trans. Autom. Sci. independent task assignment heuristics MinMin, MaxMin and sufferage,
Eng. 11 (3) (2014) 647–648. IEEE Trans. Parallel Distrib. Syst. 25 (5) (2014) 1244–1256.
[11] K. Dorling, J. Heinrichs, G.G. Messier, S. Magierowski, Vehicle routing [37] T.J. Ai, V. Kachitvichyanukul, A particle swarm optimization for the vehicle
problems for drone delivery, IEEE Trans. Syst. Man Cybern.: Syst. 47 (1) routing problem with simultaneouspick-up and delivery, Comput. Oper.
(2017) 70–85. Res. 36 (5) (2009) 1693–1702.
[12] S. UELAND, 8 commercial drone delivery companies, practical ecom- [38] K. Sethanan, T. Jamrus, Hybrid differential evolution algorithm and genetic
merce, 2021, https://www.practicalecommerce.com/8-commercial-drone- operator for multi-trip vehicle routing problem with backhauls and het-
delivery-companies. erogeneous fleet in the beverage industry, Comput. Ind. Eng. 1436 (2020)
[13] M. Alwateer, S.W. Loke, A.M. Zuchowicz, Drone services: issues in drones 106571.
for location-based services from human-drone interaction to information [39] A.A. Heidari, S. Mirjalili, H. Faris, I. Aljarah, M. Mafarja, H. Chen, Harris
processing, J. Location Based Serv. 13 (2) (2019) 94–127. hawks optimization: Algorithm and applications, Future Gener. Comput.
[14] S. Poikonen, J.F. Campbell, Future directions in drone routing research, Syst. 97 (2019) 849–872.
wiley, Networks 77 (2020) 116–126. [40] H. Savuran, M. Karakaya, Route optimization method for unmanned air
[15] M. Hu, W. Liu, J. Lu, R. Fu, K. Peng, X. Ma, J. Liu, On the joint design of vehicle launched from a carrier, Lecture Notes Softw. Eng. 3 (4) (2015)
routing and scheduling for vehicle-assisted multi-UAV inspection, Future 279–284.
Gener. Comput. Syst. 94 (2019) 214–223. [41] C.C. Murray, A.G. Chu, The flying sidekick traveling salesman problem:
[16] J. Yang, X. You, G. Wu, M.M. Hassan, A. Almogren, J. Guna, Application Optimization of drone-assisted parcel delivery, Transp. Res. C, Emerg.
of reinforcement learning in UAV cluster task scheduling, Future Gener. Technol. 54 (2015) 86–109.
Comput. Syst. 95 (2019) 140–148.
[42] S.M. Ferrandez, T. Harbison, T. Weber, R. Sturges, R. Rich, Optimization
[17] H. Kurdi, M.F. AlDaood, S. Al-Megren, E. Aloboud, A.S. Aldawood, K. Youcef-
of a truck-drone in tandem delivery network using k-means and genetic
Toumi, Adaptive task allocation for multi-UAV systems based on bacteria
algorithm, Int. J. Ind. 9 (2016) 374–388.
foraging behavior, Appl. Soft Comput. 83 (2019) 105643.
[43] Z. Luo, Z. Liu, J. Shi, A two-echelon cooperated routing problemfor a ground
[18] X. Wang, H. Zhao, T. Han, H. Zhou, C. Li, A grey wolf optimizer using
vehicle and its carried unmanned aerial vehicle, Sensors 17 (5) (2017)
Gaussian estimation of distribution and its application in the multi-
1144.
UAV multi-target urban tracking problem, Appl. Soft Comput. 78 (2019)
240–260. [44] N. Agatz, P. Bouman, M. Schmidt, Optimization approaches for the traveling
[19] D. Schermer, M. Moeini, O. Wendt, A matheuristic for the vehicle routing salesman problem with drone, Transp. Sci. 54 (4) (2018) 739–1034.
problem with drones and its variants, Transp. Res. C 106 (2019) 166–204. [45] J.C. de Freitas, P.H.V. Penna, A randomized variable neighborhood descent
[20] A. Altan, R. Hacıoglu, Model predictive control of three-axis gimbal system heuristic to solve the flying sidekick traveling salesman problem, Electron.
mounted on UAV for real-time target tracking under external disturbances, Notes Discr. Math. 66 (2018) 95–102.
Mech. Syst. Signal Process. 138 (2020) 106548. [46] A.M. Ham, Integrated scheduling of m-truck, m-drone, and m-
[21] A. Altan, Performance of metaheuristic optimization algorithms based on depotconstrained by time-window, drop-pickup, and m-visit using
swarm intelligence in attitude and altitude control of unmanned aerial constraint programming, Transp. Res. C, Emerg. Technol. 91 (2018) 1–14.
vehicle for path following, in: 2020 4th International Symposium on [47] E.E. Yurek, H.C. Ozmutlu, A decomposition-based iterative optimization
Multidisciplinary Studies and Innovative Technologies (ISMSIT), Istanbul, algorithm for traveling salesman problem with drone, Transp. Res. C,
Turkey 22-24 Oct, 2020. Emerg. Technol. 91 (2018) 249–262.

20
M. Sajid, H. Mittal, S. Pare et al. Applied Soft Computing 126 (2022) 109225

[48] S. Kim, I. Moon, Traveling salesman problem with a drone station, IEEE [64] S. Sawadsitang, D. Niyato, P.S. Tan, P. Wang, S. Nutanong, Shipper coop-
Trans. Syst. Man, Cybern. Syst. 49 (1) (2019) 42–52. eration in stochastic drone delivery: A dynamic Bayesian game approach,
[49] H.Y. Jeong, B.D. Song, S. Lee, Truck-drone hybrid delivery routing: Payload- IEEE Trans. Veh. Technol. 70 (8) (2021) 7437–7452.
energy dependency and no-fly zones, Int. J. Prod. Econ. 214 (2019) [65] H. Huang, C. Hu, J. Zhu, M. Wu, R. Malekian, Stochastic task scheduling in
220–233. UAV-based intelligent on-demand meal delivery system, IEEE Trans. Intell.
[50] Y. Liu, Z. Liu, J. Shi, G. Wu, W. Pedrycz, Two-echelon routing problem Transp. Syst. (2021).
for parcel delivery by cooperated truck and drone, IEEE Trans. Syst. Man [66] F. Mufalli, R. Batta, R. Nagi, Simultaneous sensor selection and routing of
Cybern.: Syst. 51 (12) (2021) 7450–7465. unmanned aerial vehicles for complex mission plans, Comput. Oper. Res.
[51] Q. Luo, G. Wu, B. Ji, L. Wang, P.N. Suganthan, Hybrid multi-objective 39 (11) (2012) 2787–2799.
optimization approach with Pareto local search for collaborative truck- [67] G.S.C. Avellar, G.A.S. Pereira, L.C.A. Pimenta, P. Iscold, Multi-UAV routing
drone routing problems considering flexible time windows, IEEE Trans. for area coverage and remote sensing with minimum time, Sensors 15 (11)
Intell. Transp. Syst. (2021). (2015) 27783–27803.
[52] C.C. Murray, R. Raj, The multiple flying sidekicks traveling salesman [68] V.K. Shetty, M. Sudit, R. Nagi, Priority-based assignment androuting of a
problem: Parcel delivery with multiple drones, Transp. Res. C 110 (2020) fleet of unmanned combat aerial vehicles, Comput. Oper. Res. 35 (6) (2008)
368–398. 1813–1828.
[53] K. Peng, J. Du, F. Lu, Q. Sun, Y. Dong, P. Zhou, M. Hu, A hybrid genetic [69] S.G. Manyam, S. Rasmussen, D.W. Casbeer, K. Kalyanam, S. Manickam,
algorithm on routing and scheduling for vehicle-assisted multi-drone Multi-UAV routing for persistent intelligence surveillance & reconnaissance
parcel delivery, IEEE Access 7 (2019) 49191–49200. missions, in: Proc. Int. Conf. UnmannedAircraft Syst. (ICUAS), Miami, FL,
[54] M.A. Nguyen, G.T.H. Dang, M.H. Hàa, M.T. Pham, The min-cost parallel USA, 2017, pp. 573–580.
drone scheduling vehicle routing problem, European J. Oper. Res. 299 (3) [70] M. Sajid, J. Singh, R.A. Haidri, M. Prasad, K. Kotecha, V. Vijayakumar, D.
(2022) 910–930. Garg, A novel algorithm for capacitated vehicle routing problem for smart
[55] J.M.L. Blanco, R. P.L. Gonzalez, J.L.A. Pineda, D. Canca, M. Calle, A multi- cities, Symmetry 13 (10) (2021) 1923.
agent approach to the truck multi-drone routing problem, Expert Syst. [71] B. Ombuki, B.J. Ross, F. Hanshar, Multi-objective genetic algorithms for
Appl. 195 (2022) 116604. vehicle routing problem with time windows, Appl. Intell. 24 (1) (2006)
[56] L. Amorosi, J. Puerto, C. Valverde, Coordinating drones with mothership 17–30.
vehicles: The mothership and drone routing problem with graphs, Comput. [72] O. Bräysy, M. Gendreau, Vehicle routing problem with time windows, Part
Oper. Res. 136 (2021) 105445. I: Route construction and local search algorithms, Transp. Sci. 39 (1) (2005)
[57] Q. Gu, T. Fan, F. Pan, C. Zhang, A vehicle-UAV operation scheme for instant 104–118.
delivery, Comput. Ind. Eng. 149 (2020) 106809. [73] Augerat, et al., The VRP Web, 2013, [Online]. Available: https://neo.lcc.uma.
[58] S.Y. Lee, S.R. Han, B.D. Song, Simultaneous cooperation of refrigerated es/vrp/vrp-instances/capacitated-vrp-instances/.
ground vehicle (RGV) and unmanned aerial vehicle (UAV) for rapid delivery [74] V. Santucci, M. Baioletti, A. Milani, Algebraic differential evolution al-
with perishable food, Appl. Math. Model. 106 (2022) 844–866. gorithm for the permutation flowshop scheduling problem with total
[59] J. Euchi, A. Sadoka, Hybrid genetic-sweep algorithm to solve the vehicle flowtime criterion, IEEE Trans. Evol. Comput. 20 (5) (2016) 682–694.
routing problem with drones, Phys. Commun. 44 (2021) 101236. [75] S.S. Turgut, A.H. Feyissa, E. Küçükoner, E. Karacabey, Uncertainty and sen-
[60] K.T. San, E.Y. Lee, Y.S. Chang, The delivery assignment solution for swarms sitivity analysis by Monte Carlo simulation: Recovery of trans-resveratrol
of UAVs dealing with multi-dimensional chromosome representation of from grape cane by pressurised low polarity water system, J. Food Eng.
genetic algorithm, in: Proc. IEEE 7th Annu. Ubiquitous Comput. Electron. 292 (2021) 110366.
Mobile Commun. Conf. (UEMCON), New York, NY, USA, 2016, pp. 1–7. [76] R.L. Harrison, Introduction to Monte Carlo simulation, AIP Conf. Proc. 1204
[61] B.D. Song, K. Park, J. Kim, Persistent UAV delivery logistics:MILP (2010) 17–21.
formulation and efficient heuristic, Comput. Ind. Eng. 120 (2018) 418–428. [77] E. Borgonovo, E. Plischke, Sensitivity analysis: A review of recent advances,
[62] M. Torabbeigi, G.J. Lim, S.J. Kim, Drone delivery scheduling optimization European J. Oper. Res. 248 (3) (2016) 869–887.
considering payload-induced battery consumption rates, J. Intell. Robot. [78] H. El-Sayed, S. Sankar, Y.A. Daraghmi, P. Tiwari, E. Rattagan, M. Mohanty,
Syst. 97 (2020) 471–487. D. Puthal, M. Prasad, Accurate traffic flow prediction in heterogeneous
[63] J. Kim, H. Moon, H. Jung, Drone-based parcel delivery using the rooftops vehicular networks in an intelligent transport system using a supervised
of city buildings: Model and solution, Appl. Sci. 10 (2020) 4362. non-parametric classifier, Sensors 18 (6) (2018) 1696.

21