Metaheuristic-based optimal 3D positioning of UAVs forming aerial mesh network to provide emergency communication services

The optimal placement of unmanned aerial vehicles (UAVs) to facilitate post-disaster emergency communication services is a paramount research domain. The novelty of the authors’ work is to optimally place available UAVs in 3D space to meet the objectives prominent during such situations. The objectives considered here are target coverage, QoS, energy consumption and two newly characterized objectives, i.e. equal load distribution over UAVs and fault tolerance for improving network connectivity and lifetime. To improve these conﬂicting objectives altogether, the authors proposed two metaheuristic-based hybrid optimization algorithms, namely, HWWO-HSA, a hybrid of Water Wave Optimization (WWO), Harmony Search (HS) and Simulated Annealing (SA); and HGA-SA, a hybrid of Genetic Algorithm (GA) and SA. Furthermore, to improve network performance, the authors performed their parameter tuning using well-known Taguchi’s design of experiment. For maintaining network connectivity, the authors used the concept of graphs for connected components. The proposed hybrids are compared with their originals, i.e. GA, HS, SA and WWO over different scaled test scenarios. The non-parametric Kruskal–Wallis test and Dunn’s post hoc test suggest that hybrids perform better than the originals. HGA-SA signiﬁcantly outperforms in small-scale scenarios while HWWO-HSA signiﬁcantly out-performs in medium-scale and large-scale scenarios. However, being hybrid the computational time of the hybrids is greater than the originals.


FIGURE 1
A deployed Aerial Mesh Network (AMN) in disaster-struck area to facilitate communication services service (QoS), energy consumption, connectivity, deployment cost (or number of UAVs), network lifetime, fault tolerance and robustness, security, etc. However, it has been observed that most of these issues are greatly affected by the positioning of UAVs [8][9][10]. Furthermore, the optimal positioning of UAVs to improve these issues is mostly observed as NP-hard problems [9][10][11]. For such problems, well-known metaheuristic-based optimization algorithms like genetic algorithms (GA), swarmbased algorithms, bio-inspired algorithms, etc. have proved to be a good solution. The NP-hard problems whose exact optimal solutions does not exist, metaheuristic-based optimization algorithms give near to optimal (or sub-optimal) solutions to such problems within a reasonable time. These optimization algorithms can search for solutions within a large continuous search space where UAVs can be positioned at any point in 3D space of the given target area. Also, these algorithms require less mathematical modelling for the behaviour of each and every UAV of the network, in contrast to the traditional gradient-based techniques. Please note that finding an exact optimal solution for an NP-hard problem is not possible, therefore, we called suboptimal solutions as optimal solutions in the rest of the paper to keep the things simple.
On this concept, Sabino et al. [8] aimed at facilitating differentiated data requirements for the end users with minimum deployment cost through optimal placement of UAVs using multi-objective evolutionary algorithms. Caillouet and Razafindralambo [9] centred on objectives, i.e. QoS, full target coverage and minimum deployment cost along with maintaining network connectivity through optimal deployment of UAVs using εconstraint method. Reina et al. [10] used a multi sub-populationbased GA for optimal deployment of UAVs in order to meet target coverage, fault tolerance and data redundancy. To deal with fault tolerance, Ref. [10] considered to maximize the size of minimum vertex cuts, whose exclusion can disconnect the whole network. In addition to this, we contributed to improve fault tolerance of the network by minimizing the number of such minimum vertex cuts and maximizing the number of UAVs connected to the control station. This keeps the network to remain connected for longer time. Furthermore, we focused on equal distribution of load over UAVs so that the UAVs get less unequally drained out of energy, and thus, keep the network connected for longer time. In summary, while considering the familiar network issues, i.e. target coverage, QoS and energy consumption of the network, we also focused on equal load distribution and fault tolerance of the network that are important when providing emergency communications in postdisaster regions.
Thus, the novelty of our work is to optimally place the available number of UAVs in 3D space to improve the above-mentioned network issues/objectives while keeping the network connected. In one of our prior work [12], we have considered objectives like target coverage, QoS and energy consumption by optimizing the altitude of UAVs arranged in 2D fixed grid structure using some optimization algorithms. The reason to consider the grid structure is to deal with the issue of connectivity. This paper used the concepts of graph theory for connected components in place of the grid structure to address the issue of connectivity. The concept of connected components helps to optimize the positions of UAVs in 3D space, which adds more flexibility to the system in comparison to the system proposed in [12], that only optimizes the altitude of the grid formed by UAVs having same altitude. Furthermore, this paper proposed two novel network objectives, i.e. equal load distribution and fault tolerance. The equal load distribution helps to maintain a balance of resource utilization between homogenous UAVs. It assures that each UAV should cover an equal number of targets so that the load gets equally distributed between all homogenous UAVs. This improves network connectivity as the number of UAVs died due to unequal resource drainage will decrease, which decreases the probability of network for getting disconnected due to dead UAVs, and thus, improving the overall network lifetime. Another network objective, i.e. fault tolerance keeps a check on the weaker connections present in the network in order to maintain the network connectivity for longer time. To meet the above-mentioned network objectives altogether, we proposed new hybrid optimization algorithms. We have further improved the network's performance by tuning the parameters of these hybrid optimization algorithms using Taguchi's Design of Experiment (DoE) [13], which is not incorporated in many previous works in the past.
The target coverage objective focuses on improving the number of targets covered by the network. The QoS objective focuses on improving QoS provided by the network while the energy consumption objective focuses on reducing the total energy consumed by the network during UAVs' flight and data transmission. However, the objectives, i.e. target coverage, QoS and energy consumption get inversely affected with respect to the altitude of UAVs [9,14]. The equal load distribution focuses on the objective that each UAV should cover an equal number of targets, which may reduce the target coverage when the targets are densely distributed. The reason is when the targets are densely distributed then a UAV cannot cover more than a certain number of targets to keep the equal load distribution, and thus, its coverage decreases. Furthermore, fault tolerance keeps the UAVs closer when the targets are sparsely distributed in order to keep the network connected, and thus, may reduce both target coverage and equal load distributions. As we can see that these objectives, which have different intended purposes, may have an inverse effect on each other while positioning UAVs. Therefore, optimizing the placement of UAVs in a given 3D space, which is a large continuous search space, is quite challenging. Furthermore, most of the optimal placement problem of UAVs is observed as NP-hard [9][10][11]. To trade-off these conflicting objectives simultaneously, we used a metaheuristic-based optimization approach to get the optimum positions of UAVs. The metaheuristic-based optimization algorithms apply mathematical-based search models over data collected while network simulation to find the optimal solutions.
The optimization algorithms use both exploitation (i.e. local search) and exploration (i.e. global search) to find the optimal results of a particular problem. However, it is not possible for any optimization algorithm to solve each and every optimization problem through proper balancing of both exploitation and exploration (i.e. No Free Lunch [NFL] theorem). In the past, various hybrid optimization algorithms are designed to solve different optimization problems [15][16][17]. This inspired us to develop new hybrid optimization algorithms which properly balance exploitation and exploration search ability to solve our optimal placement problem of UAVs. This study presents the first attempt to hybridize Water Wave Optimization (WWO) with Harmony Search (HS) and Simulated Annealing (SA). The strong exploration capability of WWO in both small and large search spaces is integrated with the exploitation capability of HS and Metropolis Monte Carlo criteria of SA for removing random unfit solutions at later execution phase of the hybrid, which causes its faster convergence. Moreover, we proposed a hybrid of GA and SA that balances both exploration and exploitation and has lesser computational time (CT) in comparison to the prior hybrids of GA and SA [18][19][20]. By performing experiments, it is observed that the proposed hybrid optimization algorithms perform quite efficiently than their original ones, i.e. GA, HS, SA and WWO while finding optimal locations of UAVs in 3D space.
On the basis of the proposed research work, a brief literature review is presented in Section 2. Section 3 elaborates the system model of our work. Section 4 describes the proposed hybrid optimization algorithms for optimal positioning of UAVs in our system model. Furthermore, Section 5 designs the optimal solution framework for our system model. Section 6 presents the results of each tuned optimization algorithm and compared them in different test scenarios to find the best one in each scenario. Furthermore, Kruskal-Wallis test (KW test) and Dunn's post-hoc test are used to validate the results. Section 7 presents the software and hardware challenges in our system model. Section 8 concludes the whole research work of the paper with future scopes.

LITERATURE REVIEW
The problem of optimized positioning of UAVs to perform a particular task mostly covers under NP-hard problems [21]. This is because searching an optimal solution in huge multidimensional and multi-modal solution space is not feasible in polynomial time. The optimized positioning of UAVs could be optimizing UAV altitude at fixed position [21] or UAV position at fixed altitude [22] or both [9]. The positioning of UAVs directly affects its coverage area [10]. Furthermore, coverage has been categorized as static and dynamic based on the movement of UAVs. If UAVs hover over a particular point during their flight time, it is called static coverage while if they keep moving from one point to another, it is called dynamic coverage [23]. In this work, we have preferred static coverage due to the interest of convenience. Apart from coverage, there are other issues which get affected by the positioning of UAVs. The issues include service quality, deployment cost, energy consumption, connectivity, mission completion time, network lifetime, interference, etc. These issues could be considered either as objectives or constraints based on the system requirements which cause change in the definition of the problem as well as its respective solution. Mozaffari et al. [14] centred over target coverage and network lifetime by applying circle packing theory to optimally place UAVs in 3D space. Reina et al. [10] aimed to meet target coverage, fault tolerance and data redundancies by applying multi subpopulation-based GA for optimal positioning of UAVs. Furthermore, Caillouet et al. [9] centred on QoS and full target coverage while maintaining network connectivity, by optimally deploying minimum number of UAVs using epsilon-constraint method. Hu et al. [24] incorporated smallest enclosing circle concept with evolutionary algorithm for UAV's optimal placement to ensure reliable communication in target area using minimum number of UAVs. However, there could be a condition when we cannot accommodate minimum number of UAVs to meet the defined goals. Furthermore, there is a need to focus on equal load distribution and the weak connections present in the network to keep the network connected for longer time. Apart from familiar network objectives like target coverage, energy consumption, QoS, this study also focused on equal load distribution and fault tolerance to improve network performance in critical situations.
The metaheuristic-based optimization algorithms, capable of solving any complex optimization problem, search the solution space using both exploitation and exploration property. The exploitation property makes the algorithm to rigorously search the local solution space while the exploration property makes the algorithm to fully explore the global space. However, according to No Free Lunch (NFL) theorem, it is not possible for any optimization problem to solve each and every optimization problem. In the past, various hybrid optimization algorithms are developed to solve different optimization problems. Žulj et al. [15] designed a hybrid of adaptive large neighbourhood search (ALNS) and tabu search (TS) to solve the orderbatching problem which is observed as NP-hard problem.Şenel et al. [16] proposed a hybrid of particle swarm optimization (PSO) and grey wolf optimizer (GWO) to solve one of the toughest industrial problem, i.e. leather nesting. Murthy et al. [17] developed a hybrid of PSO, cuckoo and chemical reaction optimization algorithms to solve the problem of hiding association rules in database transactions.
On the basis of NFL, we proposed HWWO-HSA which is a hybrid of WWO, HS and SA. The WWO algorithm has strong exploration but weak exploitation property due to its random function used in its propagation phase. Earlier, Singh et al. [25] developed a hybrid of WWO and sequential quadratic programming (SQP) where SQP is used for exploitation while WWO for exploration. Furthermore, Sahebjamnia et al. [26] developed hybrid of WWO and GA where the crossover operation of GA is used in the propagation phase of WWO for its better exploitation. Rong et al. [27] designed a hybrid of WWO and PSO where PSO uses its property for faster convergence and exploitation along with exploration property of WWO. This study proposed HWWO-HSA which used the strong exploitation property of HS and Metropolis Monte Carlo criteria of SA, for discarding random unfit solutions at the later execution phase of the hybrid, along with the exploitation property of WWO. It is observed through experiments that the new hybrid balances between both exploitation and exploration with faster convergence to find optimal locations of UAVs.
Also, we proposed a hybrid of GA and SA which improves the exploration ability of GA using SA's random search for solutions in large search spaces. Earlier, Mamaghani et al. [18], Huang et al. [19], Anto et al. [20], etc. have designed hybrid of GA and SA. However, these earlier designs have high time complexity which is not suitable for our problem. The proposed HGA-SA is designed in such a way that it has comparatively much lesser time complexity as compared to earlier ones.

SYSTEM MODEL
Here, a system model is described to provide a better insight to the situation where there is need to optimize the performance of AMNs. In this model, we have M number of targets that are present in a disaster struck area A. The base stations that used to provide communication facility in such an area are collapsed. Therefore, to facilitate communication in the given area, we deploy an AMN using UAVs. The wireless communication between UAVs and targets is supported by IEEE 802.11 g (Wi-Fi) technology. We assume N available number of UAVs, forming an AMN, to be deployed in the given area A. The UAVs should be placed in 3D space in such a way that they remain connected with each other and with the control station. Therefore, consider the control station while finding whether the network is connected or not. To check network connectivity, we used the concept of graph theory dealing with connected components in undirected graphs. The reason to consider undirected graphs is that the transmission range of each UAV is assumed to be equal. In this concept, we count the number of connected components

METHODOLOGIES USED
In order to deal with optimal placement problem of available number of UAVs for increasing coverage of given targets, QoS and equal load distribution with reduced energy consumption by the network, we use popular metaheuristic optimization approaches. This is because these kinds of problems are usually multi-modal and multi-dimensional and mostly come under NP-hard problems, as discussed earlier, which has no exact solution and usually converge near to optimal (or suboptimal solutions) [28]. To keep things simple sub-optimal and optimal solutions are considered same in all over the paper. The metaheuristic-based optimization approach uses mathematical models over collected data to find optimum solutions. In this section, we presented proposed metaheuristic-based hybrid optimization algorithms, i.e. HWWO-HSA and HGA-SA along with original ones, i.e. GA [29], HS [30] and SA [31], WWO [32]. Each of these algorithms is discussed one by one.

Genetic algorithm
GA [29] is influenced by nature's behaviour of evolution. It evolves the generations using biological operator such as selection, crossover and mutation. These operators are applied over each individual/chromosome of the population of size P. At each generation, two parents are selected based on their fitness values using selection operator like Roulette wheel, tournament selection, etc. These parents perform crossover operation with probability Pc to produce two offspring. Each offspring further undergo through mutation with probability Pm to get out from local maxima/minima. This algorithm will stop either at convergence or maximum number of iterations MaxIt. According to our system model, we assume a given area where we deploy N available UAVs in 3D space. Now create population Pop of size P which contains connected AMNs. To find whether the network is connected or not, the concept of connected components is used (Section 3). Each of these connected AMNs act as a chromosome with their corresponding fitness values. These chromosomes undergo through selection, crossover and mutation operations. If an individual (or AMN) generated after such operations is not a connected network, discard it and again perform the operations. After maximum number of iterations, we find the optimal chromosome or connected AMN with best fitness value.
Individual structure: An individual (i.e. either chromosome or particle), used by an optimization algorithm, is defined as an AMN formed using N UAVs that are placed in 3D space. Therefore, an individual, denoted as p, is structured as: there is a need to be assured that each individual remains connected throughout the execution of an algorithm. To check whether an individual is connected or not, find out its number of connected components. If the number of connected components is one, only then the network is said to be connected otherwise not (Section 3). If an individual is not connected, then discard it and generate new one.

1.
For a given area A, place available N UAVs in 3D space with their heights within minimum and maximum values. These UAVs will form an AMN.

2.
Set transmission range of each UAV as R which makes the network to be treated as an undirected graph.

4.
Input: Initialize population Pop of size P which contains randomly created connected AMNs represented as chromosomes. To check whether network is connected or not, count its number of components. If the count is one, then it is connected otherwise not.

5.
Calculate fitness value f(p) of each chromosome p.

Harmony search
The HS optimization [30] uses the principle used by a musician to compose a new harmony. The musician tries different music pitches stored in his/her mind to compose a new harmony. This analogy is used by HS to solve an optimization problem using all solutions stored in a memory to get the optimal solution. For a d-dimensional problem, HS initializes a memory, called as HS memory or HSM, of size P with random solutions where each solution has d dimensions/components. At each iteration, it either randomly selects any solution in HSM using Harmony Memory Considering Rate (HMCR) or generates a random value lying within lower bound (LB) and upper bound (UP) with probability (1-HMCR

Simulated annealing
SA [31] is based on the relationship between temperature and motion of particles. If the temperature is high the randomness in motion is high and if it is low the randomness in motion is low. Therefore, at initial stage we search for all possible positions with high randomness and at later stage we try to converge to the optimal solutions while keeping less randomness to select the positions using Metropolis Monte Carlo criteria. In order to have lower temperature at later stages, we decrease the temperature at the rate of alpha, α, where 0 ≤ α ≤ 1, after each iteration. However, each temperature update is maintained for a number of sub-iterations SubIt. After maximum number of iterations, the optimal solution is obtained.

Water wave optimization
The WWO [32] is based on the concepts of water waves that lie in different regions of the sea. The waves lying in shallow water are more still in comparison to the waves lying in deep water. The purpose of WWO is to make those waves that lie in deep water (having low fitness) to come up in the shallow water (having high fitness). It initializes a population Pop of P random waves having height h max and their fitness value is inversely proportional to their height from the water surface. At each iteration, a wave undergoes through propagation by randomly choosing a position in each dimension. Also, the wavelength of the wave is updated using wavelength reduction coefficient μ.
If this new position is better than current one, then update it with new one. Furthermore, if it is the new best position till now then break it upto k max number of solitary waves. Select the best solitary wave and if it the best position till now then make it the new best position. Then, break operation is performed with breaking coefficient β, to search around the new best position for finding the best one in that region. However, if the new position generated through propagation is not better than current one, then decrease its wave height by one. We perform refraction of the wave when its height becomes zero that prevents search operation from getting stuck into local optima. After maximum iterations, the algorithm produces an optimal solution.
x ←x_new  [18][19][20]. First, it applies GA operators like parent selection, crossover and mutation to get the new population. If the new population's best chromosome has not improved from the previous one, SA is applied to find the better solution. If SA finds better solution, it will be merged to the new population while discarding the chromosome having least fitness value; otherwise the new population remains unchanged. After maximum number of iterations, the optimal solution will be produced.

Hybrid Water Wave optimization with Harmony Search and Simulated Annealing
The hybrid of WWO, HS and SA applies useful features of HS and SA over WWO. The random function of WWO, used for propagation, cannot fully exploit the search space over a given region and may get out of that region after some time using its quick exploration capability. To improve the exploitation and faster convergence of WWO, we incorporated the concept of harmonics of HS. The HS makes use of every member of HS memory to generate a component of new solution with probability of HMCR and randomly generate a component of a solution with probability of 1-HMCR. This exploitation property of HS replaces the random propagation of WWO to improve its exploitation property. Furthermore, we observed that during exploration phase of WWO, its refraction operation accepts new solution even when the new one is not better than the current one. Here, we used the Metropolis Monte Carlo criteria of SA to accept the new unfit solution with certain probability. This probability is higher during initial phase when the temperature is high in order to have more exploration of the search space. At later phase, the probability of exploration decreases with decrease in temperature in order to improve convergence to the global optimum. After maximum iterations, the optimal solution is achieved.

OPTIMIZATION FRAMEWORK
The UAVs' optimum positioning in 3D space optimizes the following network objectives, i.e. improved target coverage, QoS, equal load distribution and fault tolerance with reduced energy consumption. The optimization framework, as shown in   17). Based on these fitness values, the optimization algorithm uses its optimization technique to select a better solution among all solutions. After running maximum iterations, the algorithm produces an optimal solution/AMN having optimal positioning of UAVs as shown in Figure 3.
At a given disaster-struck area A having randomly scattered set of targets/end devices E (discussed in Section 3), a set of UAVs U is positioned in 3D space to facilitate communication services. Furthermore, height of a UAV u ∈ U , i.e. h u should be bound to a minimum and maximum value, defined by users, as follows: Now, we define each objective function one by one.

Target coverage
In order to define TC, we first define the coverage radius r u of a UAV u having transmission range R as follows [14]: where Ѳ is UAV's visibility angle. Now, target eεE located at (x e ,y e ,0) is within coverage radius of UAV u located at (x u ,y u ,h) if the distance between target e and projection of UAV u, i.e. s(e, u) is less than the coverage radius r u , i.e.
where s (e, u) = √ (x e − x u ) 2 + (y e − y u ) 2 . Now, the objective, i.e. TC, which is defined as the total number of targets covered by the network [14], is formulated as: where c eu = { 1 if for target e, min(s(e, u))∀u ∫U and s (e, u) ≤ r u 0 else.
The variable c eu is set to 1 only if target e is closest to any UAV uϵ U, i.e. the distance between target e and projection of UAV u at the ground is smallest as compared to the distance between target e and projection of other UAVs at the ground, and target e is within coverage radius r u of UAV u (Equation 3).

Quality of service
We define another objective function, i.e. QoS provided by the network using metrics like packet delivery ratio (PDR) and endto-end delay (E2ED). Thus, QoS is formulated as a linear sum of PDR and E2ED as follows: The weight of PDR, i.e. v1 is kept positive as we have to improve PDR in order to improve QoS while, the weight of E2ED, i.e. v2 is kept negative as we have to reduce E2ED to order to improve QoS. Setting v1 = v2 = 0.5, we get,

Energy consumption (E)
The objective function, i.e. energy consumption by the network is evaluated using both energy consumed during UAVs' flight as well as during data transmission. The energy consumed by a UAV during flight of T seconds is defined as [33]: where γ is minimum power required by a UAV to fly in air, ρ is motor speed multiplier. P max , s and T are the maximum power of the motor, speed and flight time of UAV, respectively. Thus, the total energy consumed by N UAVs during their flight time of T seconds is: The energy consumed during transmission is evaluated using simple model of wireless communication for 1-hop distance [12]. According to this model, the amount of energy spent by a source to send m bits of data is defined as: where E elect is defined as electronics energy and E fs .d 2 or E amp .d 4 are defined as amplifier energies which depend on the distance (d) between source and destination. The threshold distance, i.e. d o is calculated as: The amount of energy dissipated by a destination node to receive m bits of data is formulated as: On the basis of this wireless communication model, the total energy consumed by a network while transmitting data for a period of T seconds is calculated as E2. Now, the total energy consumed by an AMN for a period of T seconds is defined as:

Equal load distribution
In order to achieve novel network objective, i.e. equal load distribution, we try to assure that each homogenous UAV should cover equal number of targets. The equal distribution maintains a balance of resource utilization between UAVs' so that each and every UAV gets equally utilized. To achieve equal load distribution of the network, we define load deviation of the network (LDN) as the average of the square difference of the number of targets covered by each UAV u ∈ U from the average number of targets that should be assigned to each UAV. It is formulated as: where CO u is the number of targets covered by a UAV u ∈ U , M is the total number of targets and N is the total number of UAVs and ANT is the average number of targets that should be assigned to each UAV of an AMN for equal load distribution. The minimization of LDN ensures more equal load distribution over all UAVs.

Network fault tolerance (NFT)
The above-defined objectives primarily deal with the connection between UAVs and targets. However, there is also a need to focus on the connection between UAVs and the connection between UAVs and the control station to improve the network performance. Reina et al. [10] focused on the connection between UAVs to improve the fault tolerance of the network. They focused on maximizing the size of minimum vertex cuts in the network to keep the UAVs connected for longer time. The reason to consider minimum vertex cut is that it is the smallest vertex set (or set of UAV nodes) whose removal can disconnect the whole network. We further contributed to this direction by minimizing the number of minimum vertex cuts in the network. This in turn will reduce the number of such vertex sets in the network whose removal can disconnect the whole network, and thus, will keep the network connected for longer time. Moreover, we focused on increasing the number of UAVs connected to the control station so that the control station remains connected with the network of UAVs for longer time. Now, we combined all these factors into the objective function, i.e. NFT, which considers: (i) Maximizing the size of minimum vertex cuts in the network, denoted as m. (ii) Minimizing the number of minimum vertex cuts in the network, denoted as k.
(iii) Maximizing the number of UAVs connected with the control station, denoted as dd.
Thus, the NFT improves the connection between UAVs as well as the connection between UAVs and the control station to keep the UAVs and the control station connected with each other for longer time. We defined it as a linear sum of the size of minimum vertex cuts (m), number of minimum vertex cuts (k) and the number of UAVs connected with the control station (dd) as follows: The u1, u2 and u3 are the weights of m, k and dd, respectively, which represents their respective contribution to the NFT. The sign with m and dd is positive as we have to increase m and dd to improve NFT, while the sign with k is negative as we have to decrease k to improve NFT.
Here, we assume that increasing size of minimum vertex cuts is more important than minimizing number of minimum vertex cuts and maximizing number of UAVs connected to the control station. This is because the minimum vertex cut can break the whole network, therefore, its size need to be increased to maintain the network connectivity. Therefore, we kept u1 = 0.5 while u2 = u3 = 0.25, such that u1 + u2 + u3 = 1. Thus, Equation (14) becomes:

Fitness function
Now, we converted the above five network objectives, i.e. TC, QoS, E, LDN and NFT into a single objective as the optimization algorithms can optimize only a single objective. Therefore, a single objective is created by the linear combination of TC, QoS, E, LDN and NFT as follows: The significance of Utility is to measure the quality of UAVs' deployment while forming an AMN. Here, we have to maximize the fitness function Utility, therefore, the sign of TC, QoS and NFT is positive while the sign of E and LDN is negative. Setting w1 = 0.4 (as target coverage is comparatively more significant than other objective) and w2 = w3 = w4 = w5 = 0.15, Equation (16) becomes: Based on Equation (17), the network's Utility value is maximized by an optimization algorithm to find the optimal UAVs' positions. However, to make all objectives to contribute equally, we normalized each objective function within the range of [0, 1] using the following formula [34]:

EXPERIMENTAL ANALYSIS
In this section, the rigorous experiments are performed by the optimization algorithms and their results are analysed to find the best one each scenario.

Experimental requirements
The system requirements to run the experiments are described as follows:

Test scenarios
In the test scenario, we assumed a disaster struck area A of size 1000 × 1000 m 2 where N = 30 UAVs have been deployed to facilitate communication to M number of disconnected targets. To simulate this scenario, we used ns-2 simulator where we have set the values of other simulation parameters according to Table 1. The wireless signals between UAVs and targets suffer from fading effect of the objects that lie between them, therefore, shadowing propagation model is used in this scenario. Furthermore, to allow data transmission, physical layer uses IEEE 802.11s between UAVs in order to support mesh topology between them and IEEE 802.11 g between UAVs and targets in order to support Wi-Fi communication between them. The routing layer uses Geographic Greedy Forwarding (GGF) and transport layer uses User Datagram Protocol (UDP). The flow of data generated by targets and control station is assumed to be constant with packet size of 1500 bytes and CBR rate of 10 Mbps for each of the 10 sessions. This whole scenario is simulated for a period of 60 seconds. In addition to this, the values of different parameters of UAVs are set as follows: transmission range is set to 250 m, visibility angle is set to 60 o and initial energy of each UAV is set to 100 kJ. The parameters that effect energy consumed during UAV's flight, i.e. ρ, γ, P max and s are set to 1, 30, 85 Watt and 2 m/s, respectively. Furthermore, the parameters that effect energy consumed during data transmission, i.e. E elect , E fs and E amp are set to 50 nJ/bit, 10 pJ/bit/m 2 and 0.0013 pJ/bit/m 4 , respectively. Now, we create three different scenarios based on number of targets/end users in a given area. Here, if number of targets is between 10 and 40, then it is called as small-scale scenario. If number of targets is between 50 and 80, then it is called as a medium scale and if number of targets is between 90 and 120, then it is called as a large-scale scenario. These different scenarios are presented in Table 2.

Parameter tuning
The parameters of the optimization algorithms (Section 4) need to be set initially before running it. The proper selection of these parameters enables the optimization algorithms to adapt in accordance with the problem, and thus, considerably affects convergence rate of the algorithms. In order to tune the parameters, we used famous Taguchi's DoE method [13]. According to this method, we select a set of values for each parameter where each value signifies a level. Here, we selected five levels for the parameters of GA, HS, SA, WWO and HGA-SA as shown in column 3-7 of Table A1 (see Appendix). However, HWWO-HSA has large number of parameters, therefore, it can have only three levels. Then, we created an orthogonal array (i.e. L 25 , L 27 , etc.) which is a subset of all the combinations of different parameter levels and is sufficient enough to get the best level for each parameter. The subscript of L Using different parameter combinations of an orthogonal array, the fitness function, i.e. Utility (Equation 17) is obtained for each optimization algorithm which is further used to calculate its corresponding relative percentage deviation (RPD) using the formula: The RPD represents the percentage deviation of a solution from the best one. According to Taguchi's DoE, the RPD values are used to calculate signal-to-noise ratio (SNR) for each parameter using formula for maximization problem as: where n is the number of observations and y i is the ith observation value. The higher value of SNR infers higher amount of information and lesser noise, and thus, its higher value is preferred. Therefore, a parameter value having higher SNR will be more informative than the one having smaller SNR. This will ensure that a metaheuristic-based optimization algorithm is well adopted according to the given problem and thus has been tuned quite well. The SNR for different parameters of each algorithm is presented in Figure 4. On the basis of the SNR value for different parameter levels, the best level which has highest SNR value is obtained in column 8 of Table A1 (see Appendix).

Comparison of optimization algorithms
The tuned optimization algorithms run each test problem for 20 times. These 20 values of Utility generated during each run of a test problem by an optimization algorithm are averaged and presented in Table A2 (see Appendix). On the basis of these Utility values, the corresponding RPD values are calculated using Equation (19). This is so because RPD provide more significance to the results while comparison by scaling the Utility val-ues in the range of 0-100. Furthermore, it directly shows the percentage deviation (either increase or decrease) of an optimization algorithm in comparison to the best one. Figure 5 presents RPD values by the different optimization algorithms for each test problem. These RPD values suggest that hybrid algorithms are performing quite better than originals, i.e. GA, HS, SA and WWO. Furthermore, the RPD values of HGA-SA are lower as compared to that of others in small-scale test problems, suggesting that HGA-SA is performing better than others The RPD values of the optimization algorithms in different test problems

FIGURE 6
The computational time of the optimization algorithms in different test problems in small scale. The RPD values of HWWO-HSA are lower as compared to that of other in medium and large-scale test problems, suggesting that HWWO-HSA performs better than others in medium and large-scale scenarios. In order to validate the conclusions drawn from RPD values, we have performed validation test in Section 6.5. The CT is defined as the time required for executing an algorithm for maximum number of iterations. The CT of the different optimization algorithms for each test problem is presented in Figure 6. The average CT for each algorithm is presented in Figure 7, which shows that being hybrid, the CT of hybrids is comparatively higher than that of the originals. The WWO has lowest CT while HWWO-HSA has highest CT. Furthermore, GA has second lowest CT, then HS, then SA and then HGA-SA. The CT of HWWO-HSA is 18.4% more than that of HGA-SA. Figure 8 shows the iteration at which the optimization algorithms converge to an optimal solution. The iteration at which we decide that the convergence is achieved is the one after which there is either no change or negligible change in the optimal solution. In this figure, we can observe that the proposed HWWO-HSA converges in comparatively lower number of iterations than that of WWO. This is due to Metropolis Monte Carlo selection criteria of SA which discard random unfit solutions at later execution phase of the hybrid algorithm.

Validation test
In order to validate the conclusions drawn from Figure 5 (RPD plot), we performed a validation test. Before applying validation test, we performed normality and equal variance test. The Shapiro-Wilk's normality test, as shown in Figure 9(a), shows that all the residuals are not align to the straight diagonal line. This suggests that RPD values are not normally distributed. Furthermore, Levene's equal variance test, as shown in Figure 9(b), shows that confidence interval of standard deviation of all the algorithm overlap with each other. This suggests that RPD values generated by each optimization algorithm have equal variance. In such a situation where data set is not normally distributed (i.e. non-parametric data) but have equal variance, KW test can perform well. The KW test compares the median values of different samples. If the p-value of KW test is less than the significance level, α = 0.05, then with 95% confidence it suggests that there is a significant difference in median values between at least two algorithms. In all scenarios, it is observed that the p-value < 0.05, as shown in column 6 of Table A3 (see  Appendix), which suggests that the results are significant. Furthermore, the confidence interval plots for RPD having median RPD value (blue circle in Figure 10) for each optimization algorithm in different scenarios are presented in Figure 10. On the basis of the confidence interval plots, it is observed that the proposed HGA-SA has the lowest median RPD value than others in small-scale scenarios while in medium and large-scale scenarios, the proposed HWWO-SA has the lowest median RPD value than others. The median RPD values of each algorithm in different test problems are presented in column 3 of Table A3  (see Appendix).
On the basis of KW test, it has been suggested that the results are significant and at least two algorithms have different median RPDs in all the test scenarios. Now, to find which pairs of algorithms are different we have performed Dunn's post-hoc test. This post-hoc test finds out the significance difference in median RPD between each pair of algorithms as shown in column 5 of Table A3 (see Appendix). On the basis of this, we can say that median RPD of HGA-SA has significant difference with that of GA and HS in small-scale scenarios. This suggests that in small scale, the average utility of HGA-SA is 11.01% and 12% more than that of GA and HS, respectively. Further in medium scale, HWWO-HSA has significant difference with GA, HS, SA and WWO, while HGA-SA has significant difference with WWO. This infers that in medium scale, the average utility of HWWO-HSA is 12.7%, 6.5%, 15.3% and 23.2% more than that of GA, HS, SA and WWO, respectively, while HGA- SA is 17.2% more than that of WWO. In large scale, HWWO-HSA has significant difference with GA, HS and SA. This suggests that in large scale, the average utility of HWWO-HSA is 7.08%, 11.18% and 9.86% more than that of GA, HS and SA, respectively.

HARDWARE AND SOFTWARE CHALLENGES
According to Figure 1, the end devices communicate with the control station with the help of an AMN. An AMN is formed using UAVs that mount mesh routers over them to provide data transmission between end devices and the control station.
Furthermore, VoIP server is used as a framework which enables voice over IP services between communicating devices. The hardware and software requirements for our system model are presented below:

End devices
• The end devices can be Android Phone, iPhone, wired or wireless phones, PC (Windows, Linux or Mac) • The software installed over end device is Zoiper (https:// www.zoiper.com/) which provides session initiated protocol (SIP)-based calling between the control station and end devices. The Zoiper associates an extension number to each end device present in the asterisk network handled through VoIP server.

VoIP server
The VoIP server facilitates communication between end devices and control station by providing a communication framework between them. The server enables voice communication through packet transmission within the network framework.
• The hardware for VoIP server can be Linux

Control station
The control station should also be configured with the VoIP server so that it can connect to the AMN and able to communicate to the end devices.

Unmanned aerial vehicles
The major concern while selecting UAVs according to the application requirements are as follows: • On the basis of the KW test, it is observed that the proposed hybrids are performing better than the original ones. Moreover, HGA-SA has the lowest median RPD value as compared to that of others in small-scale scenarios while HWWO-HSA has the lowest median RPD value as compared to that of others in both medium and large-scale scenarios. Also, the p-values of the KW test are less than the significance level in each of the test scenarios, suggesting that the results are significant. Furthermore, Dunn's post hoc test is used to find which pairs of optimization algorithms differ significantly. The test results suggest that in small scale the proposed hybrid HGA-SA significantly performs 11.01% and 12% better than GA and HS, respectively. In medium scale, the proposed hybrid HWWO-HSA significantly performs 12.7%, 6.5%, 15.3% and 23.2% better than GA, HS, SA and WWO, respectively, while hybrid HGA-SA performs 17.2% better than WWO. Moreover, in large scale, HWWO-HSA significantly performs 7.08%, 11.18% and 9.86% better than GA, HS and SA, respectively. However, the CT of hybrids is comparatively higher than that of originals. The CT of HWWO-HSA is 18.4% more than that of HGA-SA. In future, we may consider implementing this model in a disaster-struck region having dynamic targets. Moreover, other objectives like interference, replacement of dead UAVs, battery charging, etc. can be considered for improving the performance of AMNs. Also, we can apply tuned metaheuristic-based multiobjective optimization algorithms in our designed problem to analyse their results.