Structured Clanning-Based Ensemble Optimization Algorithm: A Novel Approach for Solving Complex Numerical Problems

. In this paper, a novel swarm intelligence-based ensemble metaheuristic optimization algorithm, called Structured Clanning-based Ensemble Optimization, is proposed for solving complex numerical optimization problems. The proposed algorithm is inspired by the complex and diversiﬁed behaviour present within the ﬁssion-fusion-based social structure of the elephant society. The population of elephants can consist of various groups with relationship between individuals ranging from mother-child bond, bond groups, independent males, and strangers. The algorithm tries to model this individualistic behaviour to formulate an ensemble-based optimization algorithm. To test the eﬃciency and utility of the proposed algorithm, various benchmark functions of diﬀerent geometric properties are used. The algorithm performance on these test benchmarks is compared to various state-of-the-art optimization algorithms. Experiments clearly showcase the success of the proposed algorithm in optimizing the benchmark functions to better values.


Introduction
With the ever increasing demand for new and better utilitydriven technologies, the challenges surrounding them are becoming more and more complex. ese optimization problems were traditionally solved using classical deterministic methods [1], which were often quite efficient in finding the solutions. But the ever increasing complexity of these problems has made these classical techniques quite unreliable for various real-world engineering problems. Here, various metaheuristic optimization [1] algorithms prove to be quite successful. Often inspired by nature, these algorithms makes use of various stochastic [2] techniques for finding the solution. ese metaheuristics are higherlevel heuristics that try to find a partial solution called heuristic representing a sufficiently good solution to the optimization problem [3][4][5][6]. e presence of randomness drives the algorithms towards promising regions of the search space. Simplicity of use, high reliability, and high flexibility has made these algorithms quite popular in last few years [7].
Among these algorithms, many are based on multiagent paradigms. ese include various evolutionary algorithms [8] inspired by the evolutionary mechanisms found in nature. ese algorithms [9][10][11][12][13][14][15] use various mechanisms like crossover, mutation, and selection for solving the optimization problems. Among the most popular metaheuristic, GA [9] was proposed by Holland in 1970s. It represents the information in the form of genetic representation as bits and performs evolutionary processes like selection, mutation, and crossover on it. Other populationbased algorithms utilize swarm intelligence-based mechanisms [16][17][18][19][20] to solve the optimization problem. ese algorithms make use of information distributed in the individual participants of the swarm. e lack of any central control structure makes it vital for the distributed information to be shared amongst the individuals of the swarm. e interaction between these individuals leads to the global intelligent behaviour which plays an important role in the performance of these algorithms. us swarm intelligence can be seen as the accumulation of various agents generally comprising social creatures like birds, ants, bees, fish, and so on and their intelligent behaviour at the individual level to replicate a complicated and collaborative behaviour at the social level. Social researchers around the world have analyzed such behaviours and developed algorithms replicating these behaviours that can be used to solve optimization problems in scientific and engineering domains.
Most of these algorithms focus on one specific strategy to evaluate the search space for the solution. is makes them vulnerable to solve problems that they are not designed to solve. Due to the presence of multiple solution update strategies, ensemble-based approaches can generally be more robust to find the solution in a more complex and unknown search space. e proposed algorithm incorporates an ensemble of solution update strategies to efficiently evaluate the problem search space.
Formulating any swarm intelligence-based optimization algorithm can be a tricky task. For any stochastic optimization method to be successful, it requires to maintain a proper balance between its exploratory and exploitative nature. It requires the understanding of whether a group replicates features of self-organization or not. Self Organization is quite common in various decentralized societies like fish schools and insect colonies. ese systems utilize the local information individually to achieve a common complex task. Here, the self-organization process helps in the form of local interactions between the lower system of entities and enables some coordination in the working of the group. is process occurs with some sort of initial randomness and is not controlled by any external/central authority. is overall process takes place in a decentralized manner such that individual entities work on the basis of local information.
is paper proposes a novel swarm intelligence-based metaheuristic optimization algorithm "Structured Clanning-based Ensemble Optimization (SCEO)" inspired by the social organization of the elephant clan. e algorithm utilizes the individual diversity found within the complex social structure of the elephant population to formulate an ensemble-based optimization algorithm. So far, only a small amount of work has been done on developing elephant-social-structure-inspired algorithms. In 2016, Wang et al. [21] proposed the elephant herding optimization (EHO) algorithm inspired by herding behaviour in the elephant groups. In elephant herding optimization (EHO) algorithm, only one type of male elephants (apart from the bounded clan elephants) are simulated that although live in isolation with the parent clan, stay in touch with the clan by simulating lowfrequency vibrations. Compared to that, the proposed SCEO algorithm simulates a clan with elephants possessing multiple types of behaviour. e algorithm simulates elephants with varying relational strength to the clan. Apart from the various types of bounded elephants (child, female, bounded, and semibounded), some male elephants are unbounded and can move independently to any particular clan. is enables the proposed algorithm to use an ensemble of solution search strategies within it making it more robust to different types of optimization problem. e rest of the paper is organized as follows: Section 2 gives an overview of the social structure of the elephant population followed by the explanation of the proposed algorithm in Section 3. Section 4 contains analysis of impact of various parameters on the performance of the algorithm. Section 5 compares the proposed algorithm with various state-of-the-art optimization algorithms. Finally, conclusion of this study is given in Section 6.

Social Structure of the Elephants
From the Elephantidae family and the Proboscidea order, elephants are one of the largest land animals on Earth. Studies had shown them to be among the most intelligent organisms. Elephants are well known to have a very complex social structure. eir social order is highly fission-fusion based with relationship between individuals ranging from mother-child bond, bond groups, independent males, and strangers [22]. Apart from this, the degree of bonding can vary quite a lot with time. e behaviour of individuals and their relationship with others varies widely over time. As with many sexually dimorphic mammals, the roles and behaviour characteristics of male and female elephant are quite different.
An elephant society is organized in the form of groups of families bound together in the form of a clan. Each clan is composed of those families who share the same foraging area. A clan is usually made up of several bond groups and numerous families such that several hundred elephants may make up a clan. Each family is led by a female elephant called Matriarch. ese female leaders are generally the largest and oldest female in the group and largely affect the structure and fortune of their group. e family unit can vary widely in size with anywhere from 2 to 50 elephants. Apart from the female leader "Matriarch," each family consists of a mix of other individuals including child elephants, younger female elephants, and younger dependent male elephants.
In contrast to female elephants, the male elephants can show different kinds of behaviour. Young male elephants are bounded with their native family and at 9 to 18 years of age, they leave the group as independent elephants. ese independent male elephants can form a small group with other male elephants. But these groups are generally loosely bound. During the mating period, these independent elephants join different family groups moving from one group to other. ey do not have preferences for specific family units and will randomly move to different groupings for reproductively receptive females.
is diversity in the individual behaviour and fissionfusion-based social structure is the main inspiration for the proposed algorithm Structured Clanning-based Ensemble Optimization (SCEO).

Elephant Social Order-Based Ensemble Optimization Algorithm
To properly mimic the fission-fusion-based social structure of the elephant clan, the population is divided into 2 or more groups. Further, the population as a whole consists of different types of agents with different individualistic property associated with each. ese agents are based on behaviour of female leaders, child elephants, other female elephants, bounded male elephants (young males), semibounded male elephants, and unbounded male elephants (fully independent males) (Figure 1). e agents have different behaviours associated with them modelled in the form of the properties and strategies associated with them.

Global Female Leader (GFL).
e global female leader is a local female leader that occupies the best position found by the population so far. is leader retains its position in the next iteration and is used to guide other female local leaders to possibly better position in the search space.

Local Female Leader (LFL).
e local female leader occupies the best position found by her family so far. All the other agents of the family move in a position relative to her; thus, this female acts as a guide to all other members in the family. e position of the local leader is updated relative to the global female leader (GFL). is helps in maintaining an interaction between different families of the swarm. e position of the local female leader is updated as per the following equation: (1) e local female leader LFL[f] belongs to the fth family unit. j represents the particular dimension of the position of the agent to be updated. r 1 is a randomly generated number between 0 and 1. w 1 is the weight coefficient affecting the influence of global female leader on the local female leader. Its value is taken between 0 and 2. By taking w 1 to be 2, the r 1 * w 1 value allows the local female leader (LFL) to explore the hypervolume covered by maximum 2 * Δd distance in each dimension around the global female leader (GFL) (where Δd is distance between LFL and GFL). e manuscript is now improved to include the above explanation.
is limits the net movement of LFL to be within 0 to 2 times the distance between GFL in each direction. is can also be seen in Figure 2. In vector form, the position update equation for local female leader can be formulated as given in Equation (2). Here, P old is the old position and P new is the new position of the agent. e matrix operation (r ≤ pr 1 ) is a Boolean operation and compares the two vectors resulting in 1 at indexes where the condition is met and 0 where the condition is not met: e numerical example shown in Equation (3) Not all the dimensions of the position is changed by GFL. is is decided on the basis of perturbation rate (pr 1 ). e value of perturbation rate is taken between 0 and 1 with larger value causing more dimensions to be affected. Perturbation rate (pr 1 ) in the SCEO algorithm is taken to introduce mutation factor into the position update of the LFL agents. In many ways, this is similar in principle to the perturbation factors used in various evolutionary algorithms like differential evolution. By generating a random number (r 5 ) between 0 and 1 and comparing it to the pr 1 value, some of the dimensions of the LFL can be restricted from change (if r 5 > pr 1 ). us, the perturbation rate can be defined as the average ratio of dimensions that are mutated in the direction of GFL.

Other Permanently Bounded Agents.
ese include agents inspired by the behaviour of child elephants, other female elephants, and young male elephants which are permanently bounded to their native family unit. ese agents move in a fixed proximity of their group's local female leader. e child agents move relatively near to the local female leader forming the innermost family circle. Child agents are followed by the female agents that are still in close Modelling and Simulation in Engineering proximity to the local leader as compared to the young male agents. ese agents change their position as per the following equation: In vector form, the position update equation for bounded agents can be formulated as given in Equation (5). Equation (6) shows the numerical calculation example for solution update of a bounded agent. e example uses child agents with w 2 � 0.02. e α value in the example is taken to be 1 representing very first iteration in the solution search space: Here, Swarm[i] represents the ith agent of the swarm bounded to the fth group. Ub and Lb are the upper and lower bounds of the search space. r 2 is a randomly generated number between −1 and 1. w 2 is the weight coefficient defining the local leader proximity region in which the ith agent can move. Its value is taken to be the lowest for child agents and highest for the young male agents. e low value of w 2 helps in better exploitation of the search region. α is used to control the balance between exploration and exploitation of the algorithm. Its value is linearly decreased from high value (near 1) at the start to a low value (near 0) at the end of the algorithm. e high value at the start helps in better exploration as the agents are able to move larger area around the local female leader. e low value of α at the end reduces the movement region for the agent, thus helping in the exploitation of the search space near the female leaders at the end of Algorithm 1.

Semibounded Male Agents.
ese male agents are not bounded to a single group and instead move from one group to other randomly. e movement of these agents in the search space is implemented using the following equation: e position update equation in the vector form for semibounded agents can be formulated as given in Equation (8). Equation (9)   for solution update of a semibounded agent. In the example, the weights w 3 and w 4 are taken to be 1.33: Here, Swarm[i] represents the ith agent of the swarm. r f is the randomly selected family and bestpos is the best position that has been found by the ith so far. r 3 and r 4 are random numbers between 0 and 1. w 3 and w 4 are weight coefficients affecting the influence of the female leader and the agent's best position on its movement.

Unbounded Male Agents.
ese are independent agents that are not bound to any particular family and move randomly in the search space. is is formulated by random selection of three agents in the search space and using their positions to change the position of this unbounded agent. e position update equation is as follows:.
e position update equation in the vector form for unbounded agents can be formulated as given in Equation (11). Equation (12) shows the numerical calculation example for solution update of an unbounded agent. In the example the weight w 5 is taken as 0.5 and perturbation rate is taken 0.6: re 1 , re 2 , and re 3 are 3 randomly selected agents from the swarm. e update probability of the position is decided on j < D and r < pr 2 j < D j < D and r < pr 1 j < D Semibounded Local leader the basis of the perturbation rate (pr 2 ), with higher value causing more dimensions to get modified. e overall SCEO algorithm can be implemented as in Algorithm 19. Figure 3 shows the overall flowchart of the proposed SCEO algorithm. To find out the proper configuration of the hyperparameters, a tuning strategy called F-Race [23] is used. F-Race is type of racing algorithm that utilizes Friedman test [24] to evaluate a set of configurations and eliminate the poor ones based on their performance. is leads to efficient allocation of the computational resources towards testing the promising configurations more. e proposed algorithm was tested with 4374 different configurations with different possible values of parameters as shown in Table 1.

Impact of Parameters
For F-Race, the number of initial races r was taken as 5 and significance level α was taken as 0.05. e maximum executions for the algorithm was set to 60000. For evaluation, 8 different functions (as mentioned in Table 2) were taken. e configurations were evaluated using vertical approach based on the mean fitness with a precision up to 8 decimals.
e maximum function evaluation was set to 300000 (� 30 * 10000). For first 40000 executions, the functions g 1 , g 3 , g 4 , and g 5 were used for evaluation; after that, functions g 2 , g 6 , g 7 , and g 8 were used. In this way, both the learning and testing sets have functions with similar characteristics with regards to modality. e resulting tuned values are shown in Table 1.

Experimental Setup
is section benchmarks the performance of the proposed SCEO algorithm compared to other state-of-the-art algorithms on various benchmark functions taken from [25][26][27][28]. e benchmark functions in Tables 3-5 provide a wide range of optimization benchmarks with varying degree of complexity in terms of shift, rotation, and modality. e benchmark function tables show the number of dimensions, range, optimal value, and modality of the functions along with their mathematical expressions. Further, to crank up the benchmark difficulty, the proposed algorithm is tested on full set of CEC 2014 [28] test set and compared against some relatively new algorithms.
(1) Procedure Structured Clanning-based Ensemble Optimization (2) Initialize Swarm of N g groups of agents with each group containing 1 female leader, N c child agents, N f female agents and N b bounded male agents. D is dimension of the search space. (3) Initialize N sb semi bounded agents and assign a random group to it. (4) Initialize N ub unbounded agents. (6) Find Fitness of all the agents. (7) for Stopping criteria not met do (8) Assign best fitness position in each fth group to the Local Leader (LFL f ).

Evaluation on Benchmark Functions.
For each function, the optimization algorithm is used for 10000 function evaluations per dimension of the benchmark function. Each function is tested 51 times and mean, standard deviation, and best and worst of these runs are used to evaluate the performance of the compared algorithms. An error of 10 −8 has been taken as error tolerance and is considered to be zero error. e proposed algorithm (SCEO) has been compared with other state-of-the-art optimization algorithms: FEP [15], ABC [29], GWO [7], and GSA [30]. e comparison results are tabulated in Tables 6-9. e parameters of these algorithms were tuned by using sensitivity analysis [31]. e hyperparameter setting for the compared algorithms is shown in Table 10. Further comparison of these algorithms has been done using parametric and nonparametric tests. t-value and h-value of the algorithm pairs is calculated using two-tailed t-test (parametric test with 5% significance level and 100 degrees of freedom) and tabulated in Tables 11 and 12. In Tables 11 and 12, the negative t-value [λ 1 , λ 2 , λ 3 , . . . , λ 10 ] � [1/5, 1/5, 10, 10, 5/100, 5/100, 5/32, 5/32, 5/100, 5/100]  indicates that SCEO algorithm is better than the compared algorithm with higher value indicating more statistical difference between from compared algorithms. Further, Shapiro-wilk test [32,33] was conducted and tabulated in Table 13. is tests the null hypothesis that if the result data are coming from a normal distribution or not. h-value of 1 indicates that the benchmark performance data is not coming from normal distribution and the parametric test is not sufficient to compare the performance of the algorithms. us, to properly compare the algorithms, a nonparametric test in the form of Wilcoxon ranked sum test [34,35] is conducted and tabulated in Table 14.
It is clearly seen from Tables 6-9, 11, and 12 that SCEO easily outperformed other state-of-the-art algorithms for most of the tested benchmark functions.
is is clearly

Evaluation on CEC 2014 Benchmark Functions.
In order to test the efficiency and reliability of the algorithm, it has been tested on CEC 2014 [28] test set against some of the newly proposed algorithms. ese include L-SHADE [36], MVMO [37,38], jDE [39], CETMS [40], FWA-DM [41], TSC-PSO [42], and HCA-SA [43] algorithms. L-Shade [36] proposed by Tanabe et al. is a modification of the SHADE [44] algorithm incorporating linear population size reduction (LPSR) strategy within it. MVMO [37] proposed by Erlich et al. makes use of a mechanism that adopts a single parent-offspring pair approach along a normalized search space. jDE [39] algorithm introduces self-adaptivity to the vanilla DE [14] algorithm. CETMS [40] algorithm uses various features of tissue membrane systems and FWA-DM [41] uses differential mutation strategy on the vanilla fireworks algorithm [45]. TSC-PSO [42] applies spatial correlation-based movement strategy on the PSO [17] algorithm. HCA-SA [43] introduces self-adaptivity in the vanilla cuckoo search algorithm (CSA) [46]. Tables 15 and 16 compares the proposed algorithm with L-SHADE [36], MVMO [37,38], jDE [39], CETMS [40], FWA-DM [41], TSC-PSO [42], and HCA-SA [43] algorithms on CEC 2014 [28] test set. Comparison is done in terms of mean and standard deviation of the 51 runs of each algorithm as per the guidelines of the CEC 2014 [28]. Clearly, the SCEO algorithm offers a comparable performance against all the newly proposed algorithms. To analyze the results, Shapiro-Wilk and Wilcoxon tests were performed as shown in Tables 17 and 18, respectively. As most of the h-values returned by Shapiro-Wilk test are 1, it is possible that the samples are not representing normal distribution, making parametric tests like t-test inefficient for comparing the algorithms. us, Wilcoxon test was performed and tabulated in Table 18. e Wilcoxon test Table 9: Comparison with FEP [15], ABC [29], GWO [7], and GSA [30].    Modelling and Simulation in Engineering 13   [15], ABC [29], GWO [7], and GSA [30] samples.    [15], ABC [29], GWO [7], and GSA [30].      [37], jDE [39], CETMS [40], FWA-DM [41], TSC-PSO [42], and HCA-SA   is clearly proves the reliability of the proposed algorithm is complex optimization scenarios. is was even more evident in case of composition functions (f 23 −f 30 ) where the SCEO algorithm performed much better than all the compared algorithms. Further, the high convergence rate (Figures 4-5) clearly shows the applicability of the proposed algorithm in real-world optimization problems.

Conclusion
Elephant population in nature is generally composed of highly complex fission-fusion-based social structure. e individuals in the elephant society can exhibit diverse types of behaviour. e population of elephants can consist of various groups with different types of individuals exhibiting different types of behaviours with member of the society. e relationship between individuals can range from mother-child bond, bond groups, independent males, and strangers. e paper proposes a novel swarm intelligence-based metaheuristic optimization algorithm  (1) [37], jDE [39], CETMS [40], FWA-DM [41], TSC-PSO [42], and HCA-SA [43] algorithms. e proposed algorithm is inspired by this complex social order of the elephant society and tries to model the individualistic behaviour. e algorithm is tested against other algorithms (FEP, ABC, GWO, and GSA) on 60 different benchmark functions. e experiments clearly show the superiority of the SCEO algorithm against the other tested algorithms proving it to be reliable and efficient for solving complex real-world numerical optimization problems. e proposed algorithm still has a scope for improvement in its performance. Future research can be directed towards studying the various hyperparameters of the algorithm by incorporating adaptability within them. Further, constrained and multiobjective models of the proposed algorithm can be formulated to extend the utility of the proposed algorithm in the real-world optimization applications.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.