Northern Goshawk Optimization: A New Swarm-Based Algorithm for Solving Optimization Problems

Optimization algorithms are one of the effective stochastic methods in solving optimization problems. In this paper, a new swarm-based algorithm called Northern Goshawk Optimization (NGO) algorithm is presented that simulates the behavior of northern goshawk during prey hunting. This hunting strategy includes two phases of prey identification and the tail and chase process. The various steps of the proposed NGO algorithm are described and then its mathematical modeling is presented for use in solving optimization problems. The ability of NGO to solve optimization problems is evaluated on sixty-eight different objective functions. To analyze the quality of the results, the proposed NGO algorithm is compared with eight well-known algorithms, particle swarm optimization, genetic algorithm, teaching-learning based optimization, gravitational search algorithm, grey wolf optimizer, whale optimization algorithm, tunicate swarm algorithm, and marine predators algorithm. In addition, for further analysis, the proposed algorithm is also employed to solve four engineering design problems. The results of simulations and experiments show that the proposed NGO algorithm, by creating a proper balance between exploration and exploitation, has an effective performance in solving optimization problems and is much more competitive than similar algorithms.


I. INTRODUCTION
Optimization means choosing the best solution out of all available candidate solutions for an optimization problem. An optimization problem consists of three main parts: decision variables, constraints (equality and inequality), and objective functions [1]. From the general point of view, optimization problem solving methods can be grouped into deterministic methods and stochastic methods. Deterministic methods implement the optimization problem-solving process based on the use of information about the derivatives of objective functions or based on information in the form of the first-order and the second-order derivatives. This information enables deterministic methods to effectively find the exact optimal for linear or convex nonlinear problems. However, these methods fail to solve more complex The associate editor coordinating the review of this manuscript and approving it for publication was R. K. Tripathy .
problems, especially those with many local optimizations. The time-consuming process of solving complex problems, high-dimensional problems, non-convex problems, problems for non-differentiable objective functions, problems with random or unknown search space are other issues that challenge deterministic methods [2]. Challenges and inability of deterministic methods led to the introduction of stochastic methods and optimization algorithms. Stochastic-based optimization algorithms are efficient tools in solving optimization problems that are able to provide suitable solutions to optimization problems without using information about the derivatives of the objective function and relying only on random scanning of the search space and random operators [3]. The process of solving the optimization problem in optimization algorithms is such that at first, a certain number of solvable solutions are generated randomly as candidate solutions. Then in an iteration-based process and based on the steps of the algorithm, these candidate solutions are improved. After the full implementation of the algorithm, the best candidate solution is selected as the solution to the problem. The solution obtained from the optimization algorithm is at best equal to the global optimal, otherwise it must be very close to it. For this reason, the solutions obtained from the optimization algorithms are called quasi-optimal [4]. The desire to achieve better quasi-optimal solutions and closer to the global optimal has led to the design of numerous optimization algorithms by researchers.
Optimization algorithms can be divided according to the type of their inspiration in nature or society into four groups: evolutionary-based, swarm-based, physics-based, and gamebased optimization algorithms.
Evolutionary-based optimization algorithms rely on the simulation of biological sciences, genetics, and the use of evolutionary operators such as natural selection. Genetic Algorithm (GA) is one of the oldest evolutionary algorithms developed based on the modeling of the reproductive process and the use of selection, crossover, and mutation sequence operators [5]. Differential Evolution (DE) algorithm is another popular evolutionary optimization algorithm that has a good ability to optimize non-differentiable nonlinear functions, which has been introduced as a powerful and fast way to optimize problems in continuous spaces [6].
Swarm-based optimization algorithms are introduced based on modeling the natural behaviors of animals, insects, aquatic animals, plants, and other living things. Particle Swarm Optimization (PSO) is one of the most widely used swarm-based algorithms, which is inspired by the intelligent behavior of birds and fish [7]. Modeling ant swarm behavior in finding the shortest path between the food source and the nest has inspired the design of the Ant Colony Optimization (ACO) [8]. Hierarchical leadership behavior modeling as well as the strategy of gray wolves during hunting have been used in the design of the Grey Wolf Optimization (GWO) [9]. In the design of the Whale Optimization Algorithm (WOA) is inspired by the bubble net hunting method performed by humpback whales [10]. Some other swarm-based algorithms are Raccoon Optimization Algorithm (ROA) [11], Teaching-Learning Based Optimization (TLBO) [12], Crow Search Algorithm (CSA) [13], Grasshopper Optimization Algorithm (GOA) [14], Tunicate Swarm Algorithm (TSA) [15], and Marine Predators Algorithm (MPA) [16].
Physics-based optimization algorithms have been developed based on the simulation of various laws and phenomena in physics. One of the oldest algorithms in this group is Simulated Annealing (SA), which is inspired by the simulation of the annealing process by melting and cooling operations in metallurgy [17], [18]. Simulation of the gravitational force that objects exert on each other at different distances has led to the design of a Gravitational Search Algorithm (GSA) [19]. Water Cycle Algorithm (WCA) is inspired by the water cycle in nature by modeling the evaporation of water from the ocean, cloud formation, rainfall, and river formation, as well as modeling the overflow of water from pits [20]. Some other physics-based algorithms are Artificial Chemical Reaction Optimization Algorithm (ACROA) [21], Multi-Verse Optimizer (MVO) [22], Electromagnetic Field Optimization (EFO) [23], Nuclear Reaction Optimization (NRO) [24], Optics Inspired Optimization (OIO) [25], Atom Search Optimization (ASO) [26], and Equilibrium Optimizer (EO) [27].
Game-based optimization algorithms are based on modeling the behavior of players in different games and the rules of these games. Simulation of competition and interactions between teams in the game of volleyball, the coaching process during the game, is employed in the design of the Volleyball Premier League (VPL) algorithm [28]. Mathematical modeling of players' behavior in tug-of-war game led to the Tug of War algorithm Optimization (TWO) [29].
With the advancement of science and technology, engineering problems become more complex, which require effective and efficient optimization methods. Therefore, this issue is resolved by improving existing methods or introducing newer optimization algorithms. An important issue in improving the capability of optimization algorithms is to increase the exploration power to global search the problem-solving space and to increase the exploitation power to local search the optimal area discovered, while a proper balance must be struck between these two indicators [30].
A major question that arises in the study of optimization algorithms is that given the existing optimization algorithms, is there still a need to design new optimization algorithms?
The answer to this question lies in the No Free Lunch (NFL) Theorem [31]. The NFL states that an algorithm that provides effective performance in solving one or more optimization problems has no guarantee that it will perform effectively in solving other optimization problems and may even fail. This means it cannot be claimed that a particular optimization algorithm is the best optimizer for all problems. It is always possible to design new algorithms that solve optimization problems better than existing algorithms. The NFL encourages researchers to be motivated to design newer optimization algorithms that can solve optimization problems more effectively. The concepts expressed in the NFL theorem have also motivated the authors of this paper to develop a new optimizer.
Northern goshawk is a bird of prey whose hunting strategy represents an optimization process. In this strategy, the northern goshawk first selects the prey and attacks it, then hunts the selected prey in a chase process. However, to the best of our knowledge of the literature, no optimization algorithm has been developed based on northern goshawk behavior. This research gap motivated the authors to develop a new optimization algorithm by mathematically modeling the northern goshawk strategy while hunting.
The novelty of this paper is in designing a new swarmbased optimization algorithm called Northern Goshawk Optimization (NGO) that mimics the behavior of northern goshawks while hunting. The various steps of the proposed NGO algorithm are expressed and then mathematically modeled. Sixty-eight objective functions are employed to evaluate the capability of NGO. The performance of the proposed NGO algorithm in optimization is compared with the performance of eight well-known algorithms. In order to analyze the NGO for solving real-world problems, this algorithm has also been implemented on four design optimization problems.
The structure of the paper is created in such a way that the proposed NGO algorithm is introduced and modeled in Section II. Simulation studies are presented in Section III. The performance of NGO in solving engineering design problems is evaluated in Section IV. Conclusions and suggestions for further study of this paper are provided in Section V.

II. NORTHERN GOSHAWK OPTIMIZATION
In this section, the proposed Northern Goshawk Optimization (NGO) algorithm is introduced and then its mathematical modeling is presented.

A. INPIRATION AND BEHAVIOR OF NORTHERN GOSHAWK
The northern goshawk is a medium-large hunter in the family Accipitridae, which was first described by the current scientific name, i.e., Accipiter gentilis by Linnaeus in his Systema naturae in 1758 [32]. Northern goshawk is a member of the Accipiter genus that hunts on a variety of prey, including small and large birds and possibly other birds of prey, small mammals such as mice, rabbits, squirrels, and even animals such as foxes and raccoons. Northern goshawk is the only member of this genus which is distributed in Eurasia and North America [33]. The male is slightly larger than the female. The male body length is 46 to 61cm, the distance between the two wings is 89 to 105 cm and weighs about 780 grams. However, the female species is 58 to 69 cm long with a weight of 1220 grams and the distance between the two wings is estimated at 108 to 127 cm [34], [35]. A photo of the northern goshawk is shown in Figure 1. The northern goshawk hunting strategy consists of two stages, so that in the first stage, after identifying the prey, it moves towards it at a high speed, and in the second stage, it hunts the prey in a short tail-chase process [36]. Northern goshawk behavior when hunting and catching prey is an intelligent process. Mathematical modeling of the mentioned strategy is the main inspiration in designing the proposed NGO algorithm.

B. ALGORITHM INITIALIZATION PROCESS
The proposed NGO is a population-based algorithm that northern goshawks are searcher members of this algorithm. In NGO, each population member means a proposed solution to the problem that determines the values of the variables. From a mathematical point of view, each population member is a vector, and these vectors together form the population of the algorithm as a matrix. At the beginning of the algorithm, population members are randomly initialized in the search space. The population matrix in the proposed NGO algorithm is determined using (1).
The proposed NGO is a population-based algorithm that northern goshawks are searcher members of this algorithm. In NGO, each population member means a proposed solution to the problem that determines the values of the variables. In fact, from a mathematical point of view, each population member is a vector, and these vectors together form the population of the algorithm as a matrix. At the beginning of the algorithm, population members are randomly initialized in the search space. The population matrix in the proposed NGO algorithm is determined using (1).
where X is the population of northern goshawks, X i is the ith proposed solution, x i,j x i,j is the value of the jth variable specified by the ith proposed solution, N is the number of population members, and m is the number of problem variables. As stated, each population member is a proposed solution to the problem. Therefore, the objective function of the problem can be evaluated based on each population member. These values obtained for the objective function can be represented as a vector using (2).
where F is the vector of obtained objective function values and F i is the objective function value obtained by ith proposed solution.
The criterion for deciding which solution is best is the value of the objective function. In minimization problems, the smaller the value of the objective function, and in maximization problems, the larger the value of the objective function, the better the proposed solution. Given that in each iteration new values are obtained for the objective function, the best proposed solution should be updated in each iteration. VOLUME 9, 2021

C. MATHEMATICAL MODELLING OF PROPOSED NGO
In designing the proposed NGO algorithm to update the population members, the simulation of northern goshawk strategy during hunting has been employed. The two main behaviors of northern goshawk in this strategy, including (i) prey identification and attack and (ii) chase and escape operation are simulated in two phases.

1) PHASE 1: PREY IDENTIFICATION (EXPLORATION)
Northern goshawk in the first phase of hunting, randomly selects a prey and then quickly attacks it. This phase increases the exploration power of the NGO due to the random selection of prey in the search space. This phase leads to a global search of the search space with the aim of identifying the optimal area. A schematic of northern goshawk behavior in this phase involving prey selection and attack is shown in Figure 2. The concepts expressed in the first phase are mathematically modeled using (3) to (5).
where P i is the position of prey for the ith northern goshawk, i is its objective function value based on first phase of NGO, r is a random number in interval [0, 1], and I is a random number that can be 1 or 2. Parameters r and I are random numbers used to generate random NGO behavior in search and update.

2) PHASE 2: CHASE AND ESCAPE OPERATION (EXPLOITATION)
After the northern goshawk attacks the prey, the prey tries to escape. Therefore, in a tail and chase process, the northern goshawk continues to chase prey. Due to the high speed of the northern goshawks, they can chase their prey in almost any situation and eventually hunt. Simulation of this behavior increases the exploitation power of the algorithm to local search of the search space. In the proposed NGO algorithm, it is assumed that this hunting is closed to an attack position with radius R. The chase process between the northern goshawk and prey is shown in Figure 3. The concepts expressed in the second phase are mathematically modeled using (6) to (8).
where t is the iteration counter, T is the maximum number of iterations, X new,P2 i is the new status for ith proposed solution, i is its objective function value based on second phase of NGO.  Select the prey at random using (3). 9.
Calculate new status of jth dimension using (7). 17. end for j = 1: m 18. Update ith population member using (8). 19. end for i = 1: N 20. Save best proposed solution so far. 21. end for t = 1: T 22. Output best quasi-optimal solution obtained by NGO for given optimization problem. End NGO.

complexity of the update process is equal to O(2T · N · m)
where T is the maximum number of iterations, and m is the number of problem variables. Therefore, the computational complexity of the proposed NGO algorithm is equal to O(N · (1+2T · m)).

III. SIMULATION STUDIES AND DISCUSSION
In this section, the performance of the proposed NGO algorithm in solving optimization problems is tested. For this purpose, NGO is implemented on sixty-eight different objective functions including unimodal, high-dimensional multimodal, fixed-dimensional multimodal [37], CEC2015 [38], and CEC2017 [39]. The performance of the proposed NGO algorithm is compared with eight well-known algorithms PSO, GA, GSA, TLBO, GWO, WOA, MPA, and TSA. The values set for the control parameters of these algorithms are specified in Table 1. The proposed NGO algorithm and each of the competing algorithms are implemented in twenty independent executions on every objective function, while each execution contains 1000 iterations. The optimization results are reported using two indicators (i) the average of the best proposed solutions and (ii) the standard deviation of the best proposed solutions. The experimentation has been done on Matlab R2020a version using 64 bit Core i7 processor with 3.20 GHz and 16 GB main memory.

A. EVALUATION OF UNIMODAL OBJECTIVE FUNCTION (F1-F7)
The optimization results of F1 to F7 functions using the proposed NGO algorithm and eight competitor algorithms are reported in Table 2. The simulation results show that NGO has been able to provide the optimal global for F6. The NGO algorithm is the first best optimizer in solving F1, F2, F3, F4, F5, and F7 functions. What can be deduced from the analysis of the simulation results is that the proposed NGO algorithm has a superior and much more competitive performance than the eight compared algorithms.

B. EVALUATION OF HIGH-DIMENSIONAL MULTIMODAL OBJECTIVE FUNCTION (F8-F13)
The implementation results of the proposed NGO algorithm and eight compared algorithms on the objective functions of F8 to F13 are presented in Table 3. The NGO with its high exploration power has been able to achieve the optimal global value for F9 and F11. In the F8 function optimizer, GA is the first best optimizer while NGO is the second best optimizer for this function. GSA is the first best optimizer and NGO is the second best optimizer for the F13 function. The proposed NGO algorithm is the first best optimizer for solving F10 and F12 functions. The simulation results show that the proposed NGO algorithm has an acceptable ability to solve high-dimensional multimodal optimization problems.

C. EVALUATION OF FIXED-DIMENSIONAL MULTIMODAL OBJECTIVE FUNCTION (F14-F23)
The solving results of the objective functions F14 to F23 using the NGO and eight competitor algorithms are presented in Table 4. The proposed NGO algorithm has been able to converge to the global optimum for F14 and F17. The NGO is the first best optimizer in solving F15 and F20 functions. In optimizing the functions of F16, F18, F19, F21, F22, and F23, the proposed NGO algorithm has the same performance in the avg index as some competing algorithms. However, in these functions, the proposed NGO algorithm has better conditions in the std index. Analysis of the simulation results shows that the proposed NGO algorithm has a high capability in solving F14 to F23 functions and is much more competitive than the eight compared algorithms.
The performance of NGO and eight competitor algorithms in optimizing F1 to F23 functions is shown in the form of a boxplot in Figure 5. The analysis of this boxplot shows that the NGO has less width and a more efficient center than competitor algorithms in optimizing most F1 to F23 functions. This means that the NGO has offered close and almost similar solutions in different implementations. Therefore, NGO is able to provide more efficient solutions to optimal problems.

D. STATISTICAL ANALYSIS
Comparison of optimization algorithms based on avg and std criteria provides valuable information about their capabilities. However, it may be a chance that one algorithm is superior to another, even after twenty independent executions with the least probability. Therefore, in this subsection, a statistical analysis is presented to further analyze the performance of the proposed algorithm in effectively solving optimization problems than the eight competitor algorithms. For this purpose, Wilcoxon rank sum test is used to show whether the superiority of the proposed algorithm over the competing algorithms is significant or not. In this test, a p-value is used to show the superiority of one algorithm over another algorithm.    The results of statistical analysis of the proposed NGO algorithm against eight competitor algorithms are presented in Table 5. According to the results of the Wilcoxon rank sum test, in cases where a p-value is less than 0.05, the proposed NGO algorithm is significantly better than all competitor algorithms. According to Table 5, the NGO has a significantly      superiority over each of the competitor algorithms in optimizing unimodal and fixed-dimensional multimodal functions. Also, NGO has a significant superiority in optimizing    high-dimensional multimodal functions compared to MPA, TSA, WOA, GWO, TLBO, and PSO.

E. SENSITIVITY ANALYSIS
The proposed NGO algorithm is a population-based algorithm that solves optimization problems in a repetition-based process. Therefore, the two parameters of the population, VOLUME 9, 2021  number of northern goshawks (N ) and the maximum number of iterations (T ) affect the performance of the proposed NGO algorithm. Therefore, in this subsection, the sensitivity analysis of the NGO to the two parameters N and T is presented.
To evaluate the sensitivity analysis to parameter N , the proposed NGO algorithm for different values of population members equal to 20, 30, 50, and 80 has been implemented on the functions F1 to F23. The results of the sensitivity analysis of the NGO with respect to parameter N are reported in Table 6. The simulation results show that increasing the number of population members has improved the performance of the NGO and the values of the objective functions have decreased. The behavior of the convergence curves of the NGO in the study of this analysis is shown in Figure 6. These convergence curves show that increasing the number of population members leads to an increase in the exploratory power of NGO in identifying the optimal area more quickly and thus converging to more appropriate solutions.
In order to evaluate the sensitivity analysis for the T parameter, the NGO for different values of the maximum  number of iterations equal to 100, 500, 800, and 1000 is employed to solve the functions of F1 to F23. The results of this analysis are presented in Table 7. Sensitivity analysis of the NGO to parameter T shows that as the maximum number of iterations increases, the value of all objective functions decreases. The behavior of the convergence curves of the proposed NGO algorithm under the influence of different values of the T parameter is presented in Figure 7. These convergence curves show that increasing the value of T gives the NGO more opportunity to converge towards better solutions.

F. EVALUATION OF IEEE CEC2015 (CEC1-CEC15)
The results of optimization of CEC2015 functions using the NGO and eight competitor algorithms are presented in Table 8. The simulation results show that the NGO has better results than the eight competitor algorithms in CEC1, CEC3, CEC7, CEC8, CEC9, CEC10, CEC11, CEC12, CEC13, CEC14, and CEC15 functions. In optimizing CEC5 and CEC6 the WOA performed better. However, the proposed NGO is the second best optimizer to solve these functions.

G. EVALUATION OF IEEE CEC2017 (C1-C30)
The performance results of the proposed NGO and eight competitor algorithms on the CEC2017 objective functions VOLUME 9, 2021    are presented in Table 9. What is clear from the analysis of the results is that the proposed NGO algorithm offers better quasi-optimal solution for objective functions C4, C5, C8, C10, C11, C12, C13, C20, C22, C25, C26, C29, and C30.

IV. NGO APPLICATION FOR ENGINEERING DESIGN PROBLEMS
In this section, the performance of the NGO in solving problems in real-world applications is evaluated. For this purpose, the NGO is implemented on four optimization problems, namely pressure vessel design, welded beam design, tension/compression spring, and speed reducer design.

A. PRESSURE VESSEL DESIGN OPTIMIZATION PROBLEM
The mathematical model used was adapted from [40]. Figure 8 shows the schematic view of the pressure vessel problem. In this design, T s is the thickness of the shell, T h is the thickness of the head, R is the inner radius, and L is the length of the cylindrical section without considering the head. Tables 10 and 11 report the performance of the NGO and other algorithms. The NGO provides an optimal solution at (0.7781779, 0.3846819, 40.31963, 200.00000) with a corresponding fitness value of 5885.4958. Figure 9 presents the convergence analysis of the NGO for the pressure vessel design optimization problem.

B. WELDED BEEM DESIGN OPTIMIZATION PROBLEM
The mathematical model of a welded beam design was adapted from [10]. Figure 10 displays the schematic view of the welded beam problem. In this design, h is the thickness of weld, l is the length of the clamped bar, t is the height of the bar, and b is the thickness of the bar. The results to this optimization problem are presented in Tables 12 and 13.    The NGO provides an optimal solution at (0.20576, 3.471, 9.0361, 0.20577) with a corresponding fitness value equal: 1.725202. Figure 11 presents the convergence analysis of the NGO for the welded beam design optimization problem.

C. TENSION/COMPRESSION SPRING DESIGN OPTIMIZATION PROBLEM
The mathematical model of this problem was adapted from [10]. Figure 12 displays the schematic view of the tension/compression spring problem. in this design, d is the wire diameter, D is the mean coil diameter, and P is the number of active coils. The results to this optimization problem are displayed in Tables 14 and 15. The AMBOA provides the optimal solution at (0.0523593, 0.372854, 10.4093) with a corresponding fitness value of 0.012672. Figure 13 shows the convergence analysis of the NGO for the tension/compression spring optimization problem.

D. SPEED REDUCER DESIGN OPTIMIZATION PROBLEM
This problem is modeled mathematically in [41], [42]. Figure 14 displays the schematic view of the speed reducer design problem. In this design, b is the face width, m is the module of teeth, p is the number of teeth in the pinion, l 1 is the length of the first shaft between bearings, l 2 is the length of the second shaft between bearings, d 1 is the diameter of first shafts, and d 2 is the diameter of second shafts. The results of the optimization problem are presented in Table 16 and 17. The optimal solution was provided by the    Figure 15 presents the convergence analysis of the NGO for the speed reducer design optimization problem.

V. CONCLUSION AND FUTURE WORKS
In this paper, a new intelligence swarm-based algorithm called Northern Goshawk Optimization (NGO) was designed, which its main inspiration is to simulate the VOLUME 9, 2021 behavior and strategy of northern goshawk while hunting. Mathematical modeling of the proposed NGO algorithm was presented and then its performance in optimization was tested on sixty-eight objective functions. The optimization results indicate the ability of NGO to provide desired quasi-optimal solutions for optimization problems. The performance of NGO in optimization was compared with eight well-known algorithms including PSO, GA, GSA, TLBO, GWO, WOA, MPA, and TSA. The analysis of the simulation results showed the obvious superiority of the proposed NGO algorithm over the eight competitor algorithms. In addition, the implementation of NGO on four design problems showed that the proposed algorithm was highly capable of solving real-world problems.
The authors make several suggestions for future studies of this paper. Attempts to design the binary as well as the multi-objective version of the proposed NGO algorithm are among the main study potentials for the future. In addition, the application of NGO in solving optimization problems in different sciences and comparing it with other existing algorithms are other suggestions for further studies in line with this paper.  Minimize f (x) = 0.6224x 1 x 3 x 4 + 1.778x 2 x 2 3 + 3.1661x 2 1 x 4 + 19.84x 2 1 x 3 . Subject to: g 1 (x) = − x 1 + 0.0193x 3 ≤ 0, g 2 (x) = −x 2 + 0.00954x 3 ≤ 0, Minimize f (x) = 1.10471x 2 1 x 2 + 0.04811x 3 x 4 (14.0 + x 2 ). Subject to:   Subject to: