Received February 27, 2019, accepted March 15, 2019, date of publication March 22, 2019, date of current

version April 13, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2907012

Fitness Dependent Optimizer: Inspired by the

Bee Swarming Reproductive Process
JAZA MAHMOOD ABDULLAH 1,2 AND TARIK AHMED RASHID 3, (Member, IEEE)
1 Information
Technology Department, College of Commerce, University of Sulaimani, Sulaymaniyah, Iraq
2 Computer Engineering Department, College of Engineering, Komar University of Science and Technology, Sulaymaniyah 65-1154, Iraq
3 Computer Science and Engineering, University of Kurdistan Hewler (UKH), Erbil, KRG, Iraq

Corresponding author: Jaza Mahmood Abdullah ([email protected])

This work was supported by the University of Sulaimani and the Komar University of Science and Technology.

ABSTRACT In this paper, a novel swarm intelligent algorithm is proposed, known as the fitness dependent
optimizer (FDO). The bee swarming the reproductive process and their collective decision-making have
inspired this algorithm; it has no algorithmic connection with the honey bee algorithm or the artificial
bee colony algorithm. It is worth mentioning that the FDO is considered a particle swarm optimization
(PSO)-based algorithm that updates the search agent position by adding velocity (pace). However, the FDO
calculates velocity differently; it uses the problem fitness function value to produce weights, and these
weights guide the search agents during both the exploration and exploitation phases. Throughout this paper,
the FDO algorithm is presented, and the motivation behind the idea is explained. Moreover, the FDO
is tested on a group of 19 classical benchmark test functions, and the results are compared with three
well-known algorithms: PSO, the genetic algorithm (GA), and the dragonfly algorithm (DA); in addition,
the FDO is tested on the IEEE Congress of Evolutionary Computation Benchmark Test Functions (CEC-
C06, 2019 Competition) [1]. The results are compared with three modern algorithms: (DA), the whale
optimization algorithm (WOA), and the salp swarm algorithm (SSA). The FDO results show better per-
formance in most cases and comparative results in other cases. Furthermore, the results are statistically
tested with the Wilcoxon rank-sum test to show the significance of the results. Likewise, the FDO stability
in both the exploration and exploitation phases is verified and performance-proofed using different standard
measurements. Finally, the FDO is applied to real-world applications as evidence of its feasibility.

INDEX TERMS Optimization, swarm intelligence, evolutionary computation, metaheuristic algorithms,

fitness dependent optimizer, FDO.

I. INTRODUCTION programming. Evolutionary algorithms include heuristic or

From the time when computers were invented, searching for metaheuristic algorithms and many hybrid techniques.
the unknown and looking for the best solution were points Traditional algorithms are efficient in their work; however,
of focus. As early as 1945, Alan Turing used a type of search several facts can be discussed about them. They are mostly
algorithm for breaking German Enigma ciphers during World deterministic; for example, a given input will always obtain
War II [2]. To date, hundreds of types of algorithms have been the same output (except hill climbing when using random
developed for various purposes, including optimization prob- restart). Moreover, they perform local searches, which is why
lems. Optimization algorithms are used to find suitable solu- there is no guarantee that global optimality will be reached
tions for a problem. There might be many different solutions for most of the optimization problems. Consequently, they
for a single problem, but the optimum solution is preferable. have limited diversity in the obtained solutions. Additionally,
Usually, optimization problems are nonlinear with a complex they use some information about the problems, and therefore,
landscape. Generally, optimization algorithms can be classi- they tend to be problem-specific. Furthermore, these tradi-
fied into traditional and evolutionary algorithms. Traditional tional algorithms cannot effectively solve multimodal prob-
algorithms include gradient-based algorithms and quadratic lems because they do not work on highly nonlinear problems.
Evolutionary algorithms could be the correct answer to
The associate editor coordinating the review of this manuscript and previous limitations as they have stochastic behaviors. They
approving it for publication was Danilo Pelusi. come in two forms: heuristic and meta-heuristic algorithms.

2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 7, 2019 Personal use is also permitted, but republication/redistribution requires IEEE permission. 43473
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

Heuristic algorithms search for a solution by trial and error; optimization algorithm (WOA) in 2016 [17], and the salp
they hope that a quality solution will be found in a reasonable swarm algorithm (SSA) in 2017 were proposed by the same
amount of time. Similarly, they tend to use specific random- author [18]. Two new variants of the ABC are proposed by
ization mechanisms and local searches in various ways. More Laizhong et al, the authors showed that they managed to
studies and developments have been conducted on heuristic enhance the exploitations of the novel ABC algorithm, as it is
algorithms to make what is known as metaheuristic algo- well known that the novel ABC has a good exploration ability,
rithms. Metaheuristic algorithms have better performance however, it suffers from slow exploitations. In their first work,
than heuristics algorithms, which is why the ‘‘meta’’ prefix they employed an adaptive method for the population size
was added, which means ‘‘higher’’ or ‘‘beyond’’. However, (AMPS) [19]. In the second paper, they proposed a ranking-
researchers currently use these two terms (heuristic and meta- based adaptive ABC algorithm (ARABC) [20], the atten-
heuristic) interchangeably, as there is little difference in their tion on both works was to improve exploitations ability of
definitions [3]–[5]. the novel ABC. Nonetheless, two more improvements were
The complexity of real-world problems that exist around us suggested on the novel ABC in 2018, firstly, by propos-
makes it impossible to search every possible solution simply ing the distance-fitness-based neighbor search mechanism
because of time, space, and cost considerations. As a result, (DFnABC), which is a new variant of the ABC [21], and sec-
low cost, fast, and more intelligent mechanisms are required. ondly, by proposing the dual-population framework (DPF),
Therefore, researchers have studied the behaviors of animals again to enhance ABC convergence speed [22]. Additionally,
and natural phenomena to understand how they solve their in 2018, a new algorithm which inspired by vapour-liquid
problems. For example, how ants find their path, how a group equilibrium (VLE) was proposed by Enrique M. Cortés-Toro
of fish, birds or flies avoid the enemy or hunt their prey, and his colleagues, the authors claim that their algorithm can
and how gravity works. Thus, these algorithms, which are solve highly nonlinear optimization problems in continuous
inspired by nature, are known as nature-inspired algorithms. domains [23].
Development in nature-inspired metaheuristic algorithms Various research has been conducted in the field of
began in the 1960s at the University of Michigan. John nature-inspired metaheuristic algorithms; additionally, many
Holland and his colleagues published their genetic algo- efficient algorithms have been proposed in the literature.
rithm (GA) book in 1960 and republished it in 1970 and Alternatively, there is always room for new algorithms,
1983 [6]. An algorithm that is inspired by the annealing as long as the proposed algorithm provides better or com-
process of metal, known as simulated annealing (SA), was parative performances, as explained by David H. Wolpert
developed by Kirkpatrick et al. [7]. Nevertheless, in the past and William G. Macready in their work titled ‘‘No Free
two decades, this field has witnessed many major signs of Lunch Theorems for Optimization’’ in 1997. Thus, there is no
progress. For instance, particle swarm optimization, which single global algorithm that can provide the optimum solution
was proposed by James Kennedy and Russel C. Eberhart, for every optimization problem. For example, if algorithm
has been used for many real-world applications [8]. PSO was ‘‘A’’ works better than algorithm ‘‘B’’ on optimization prob-
inspired by the swarm intelligence of fish and birds while the lem X, then there is a high chance that there is an optimization
authors were studying a flock of birds. They found that they problem Y, that works better on algorithm ‘‘B’’ than on algo-
could apply these behaviors to optimization problems; later, rithm ‘‘A’’ [24]. For these reasons, a new algorithm (FDO)
PSO became a base algorithm for other algorithms, including is proposed in this paper. This algorithm is inspired by the
our algorithm. R. Storn and K. Price developed differential swarming behavior of bees during the reproductive process
evolution (DE) in 1997. It is a vector-based algorithm that when they search for new hives. This algorithm has nothing in
outperforms GA in many applications [9]. After that, in 2001, common with the ABC algorithm (except both algorithms are
Zong WooGeem et al. developed the harmony search (HS), inspired by bee behavior, and both are nature-inspired meta-
which was applied in many optimization problems such as heuristic algorithms).
transport modeling and water distribution [10]. Then, in 2004, The major contributions of this paper are summarized as
C. Tovey and S. Nakrani developed the honey bee algorithm. follows:
They used it for Internet hosting center optimization [11]. 1- A new novel swarm intelligent algorithm is pro-
This was followed by the development of a novel bee algo- posed, which is using certain characteristics of the bee
rithm proposed by Pham et al. [12], and one year later, swarms. For example, it uses a fitness function for
D. Karaboga et al. created the artificial bee colony (ABC) generating suitable weights that help the algorithm in
algorithm in 2005. In 2009, Xin-She Yang developed the both exploration and exploitation phases, as it provides
firefly algorithm (FA) [13]; and then, the cuckoo search fast convergence towards global optimality with respect
(CS) algorithm was proposed by the same author [14]. Addi- to fair coverage of the search landscape.
tionally, Xin-She Yang proposed a bat-inspired algorithm 2- One more unique feature of FDO is that it stores the
in 2010 [15]. Then, in 2015, Mirjalili A. S. proposed the past search agent pace (velocity) for potential reuse in
dragonfly algorithm (DA) [16], which is a PSO-based algo- future steps (more on this is discussed in section IV).
rithm inspired by the dragonfly swarm behavior of attrac- 3- FDO can be considered a PSO-based algorithm since
tion to food and distraction by the enemy, then the whale it uses a similar mechanism for updating agents’

43474 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

FIGURE 1. Honey bee anatomy [27]. FIGURE 2. Bee swarming process cycle [28].

positions; however, FDO does it in a very different way,

and it is statistically proven in this paper that FDO work under the command of queen bees; they also create
outperforms PSO, DA, GA, WOA, and SSA in many queen cells throughout the year. Finally, scout bees explore
benchmark test functions and has comparative results the environment and exploit the preferable targets, which is
on others. the most important feature of this work. Usually, when the
The remainder of the paper is organized as follows. number of bees in the hive increases and the inside colony
It begins by explaining the motivation behind the FDO conditions and outside weather conditions are suitable, then
algorithm and then debates the unique aspects that show the queen lays eggs into the queen cells, and the bee colony
the novelty of FDO algorithm. Then, the bee swarm fea- starts the reproductive processes by swarming [25], [26].
tures that inspired the FDO algorithm are presented. After Swarming is mostly a late spring phenomenon; it is the
that, the FDO algorithm is introduced by showing the pseu- process by which a new honeybee colony is formed. The
docode, equations, and rules. Furthermore, FDO for the sin- queen bee leaves the old colony with a group of worker bees
gle objective problem is described. Moreover, in the results and some scout bees; Figure (2) shows the swarming cycle.
and discussion section, detailed information is provided about A swarm typically consists of thousands to tens of thousands
FDO performance against other algorithms. In addition, of bees. They settle 20–30 meters away from the natal hive
FDO is applied to two real-world case scenarios. Finally, temporarily for a few hours to a few days. They may gather
the main points about FDO, its limitation, and future works in a tree or on a branch where they cluster around their queen,
are described. and then, they send 20–50 scout bees out to find suitable new
hives, usually after several tries, which might take several
II. BEE SWARMING hours or up to three days. Eventually, with the guidance from
Since ancient times, this remarkable social insect has the scouts, the rest of the bees flying overhead in the proper
been one of the most famous creatures on the planet. direction.
Honeybees have been the subject of scientific observations. A swarm may fly a kilometer or more to the scouted
Likewise, considerable research and many books have been location. Through direct observation, it can be said that
published about them, for example, ‘‘Behavior and the Social the scout bee has several criteria for a suitable hive. For
Life of Honeybees’’ by Ribbands in 1953. Snodgrass wrote instance, a suitable hive has to be large enough to accom-
‘‘Anatomy of the Honey Bee in 1956, also ‘‘the wisdom modate the whole swarm (minimum of 15 liters, prefer-
of hive’’ by Thomas D. Seeley was written in 1995, and ably 40 liters in volume). It should have a small entrance
many other great works. Figure (1) shows the anatomy of (approximately 12.5 cm2), as well as being located at the
a bee. Presenting the colony structure of bees and their bio- lowest point of the hive, and obtain a certain amount of
logical details are beyond the scope of this paper. However, warmth from sunlight [26], [29].
the swarming behavior of the bee life cycle will be discussed What inspired us were the scouts’ collective decision-
shortly since it is related to FDO. making processes. When a number of scout bees discover
As widely known, bees live and work in groups inside a some suitable hives, they will choose the most suitable hive,
colony called a hive (nest site). In brief, there are several types and they keep the swarm intact. Typically, scout bees com-
of bees: queen bees, worker bees, and scout bees. As their municate through moving their legs and wings, which is
names suggest, the queen bee is responsible for making deci- known as a bee dance. Usually, a decision will be made when
sions and producing the next generation of bees. Worker bees approximately 80% of the scouts have agreed upon a certain

VOLUME 7, 2019 43475

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 1. FDO-related Bee biological entities. of artificial scout bees is expressed as follows:

Xi,t+1 = Xi,t +pace (1)

where i represents the current search agent, t represents

the current iteration, x represents an artificial scout bee
(search agent), and pace, is the movement rate and direction
of the artificial scout bee. pace is mostly dependent on the
fitness weight fw. However, the direction of pace is com-
hive location or when there is a quorum of 20-30 scout bees pletely dependent on a random mechanism. Thus, the fw for
present at a potential hive [26], [29]. minimization problems can be calculated as:
Algorithmically speaking, each hive that a scout bee
 ∗ 
x 
 i,t fitness 
exploits, represents a possible solution exploited by an arti- fw =   wf (2)
 xi,t fitnees 
ficial search agent, and the best hive represents the global
optimum solution, as shown in Table (1). The hive specifi- ∗
The xi,t fitness , is a fitness function value of the best global
cations, such as its volume, entrance size, entrance location, solution that has been discovered thus far. xi,t fitness is a value
and amount of sunlight, can also be considered as fitness of the fitness function of the current solution, wf is a weight
functions of the solution. The scout’s collective decision- factor, and its value is either 0 or 1, which is used for control-
making process, represented by fitness weight (fw) in the ling the fw. If it is equal to 1, then it represents a high level
algorithm, fw is discussed further in the next section. of convergence and a low chance of coverage. Nonetheless,
if wf = 0, then it is not affecting the Equation (2), thus it
III. FITNESS DEPENDENT OPTIMIZER ALGORITHM can be neglected, setting wf = 0 provides a more stable
This algorithm replicates what a swarm of bees is doing search. However, this is not always the case; sometimes,
during reproduction. The main part of this algorithm is taken the opposite occurs because the fitness function value is
from the process of scout bees searching for a new suit- completely optimization problem dependent. Nevertheless,
able hive among many potential hives. Every scout bee that the fw value should be in the [0, 1] range; however, there are
searches for new hives represents a potential solution in this some cases where fw = 1, for example, when the current
algorithm; furthermore, selecting the best hive among several solution is the global best, or when the current and global
good hives is considered as converging to the optimality. best solutions are identical or have the same fitness value.
The algorithm begins by randomly initializing an artifi- Additionally, there is a chance that fw = 0, which occurs
cial scout population in the search space Xi (i= 1, 2, . . .n); when xi,t∗
fitness = 0. Finally, division by zero should be
each scout bee position represents a newly discovered hive avoided when xi,t fitness = 0. Therefore, the following rules
(solution). Scout bees try to find better hives by randomly should be used (3)–(6), shown at the bottom of the next page.
searching more positions; each time a better hive is found, Here, r is a random number in the [−1, 1] range. There
the previously discovered hive is ignored; thus, each time the are different implementations of the random walk; however,
algorithm identifies a new, better solution, then the previously Levy flight has been chosen because it provides more stable
discovered solution will be ignored. In addition, if the current movements because of its good distribution curve [3].
move is not leading the artificial scout bee to a better solution Regarding FDO mathematical complexity: For each itera-
(hive), it will continue in its previous direction, hoping that tion, it has an O (p∗n + p∗CF) time complexity, where p is
the previous direction takes the scout to a better solution. the population size, n is the dimension of the problem, and
However, if the previous direction does not lead to a better CF is the cost of the objective function. Whereas, for all
solution, it will then continue to the current solution, which iterations, it has an O (p∗CF +p∗ pace) space complexity,
is the best solution that has been found to that point. where the pace is the best previous paces stored. From here,
In nature, scout bees search for hives randomly. In this FDO time complexity is proportional to the number of itera-
algorithm, artificial scouts initially search the landscape ran- tions. However, its space complexity will be the same during
domly using a combination of a random walk and fitness the course of iterations.
weight mechanism. Accordingly, every time an artificial FDO has a simple calculation mechanism in terms of
scout bee moves by adding pace to the current position, objective value calculations, it has only (fitness weight and
the scout hopes to explore a better solution. The movement one random number) to be calculated for each agent, whereas,

 

 fw = 1 or fw = 0 orxi,t fitness = 0, pace
 = xi,t ∗ r (3)  

 r < 0, pace = xi,t − x ∗ ∗ fw∗ − 1
 
i,t (4) 
 fw > 0 and fw < 1  

  r ≥ 0, pace = xi,t − x ∗ ∗ fw (5)  

i,t

43476 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

in PSO for calculating each solution, there are global best,

agent best, search factors C1 and C2, and random numbers
(R1 and R2 parameters) to be calculated [8]. Also, in the DA,
there are five different parameter weights to be calculated
(separation, alignment, cohesion, attraction, distraction, and
some random values), and most of these parameters have
accumulative nature (summation and multiplication), and
their values depend on all other agents’ value, resulting in
even more complex calculations [16].

IV. FDO WITH SINGLE OBJECTIVE OPTIMIZATION

PROBLEMS
The FDO with single objective optimization problems
(FDOSOOP) begins by initializing artificial scouts at random
locations on the search landscape, using upper and lower
boundaries. For every iteration, the global best solution is
selected; then, for every artificial scout bee, the fw is cal-
culated according to Equation (2). After that, the fw value
is checked to determine if fw = 1 or 0, also whether
xi,t fitness = 0. Then Equation (3) is used for generating
the pace. However, if fw > 0 and fw < 1, then a random
number r will be generated in the [−1, 1] range. If r < 0
then Equation (4) is used to calculate the pace, in this case,
fw gets a negative sign, but if r ≥ 0 then Equation (5) is
used to calculate the pace, accordingly, fw gets a positive sign.
Randomly selecting negative or positive sign for a a fw will
guarantee that the artificial bee will search randomly in every
direction.
In FDO randomization mechanism controls the pace size
and direction, whereas in most cases, the randomization
mechanism only controls the pace direction; in these cases,
the pace size depends on the fw. Moreover, each time the
artificial scout bee finds a new solution, it checks whether
the new solution is better than the current solution, depending FIGURE 3. Pseudocode of FDOSOOP.
on the fitness function as shown in the pseudocode of the
single objective FDO (see Figure (3)). If the new solution
be replaced with the line: ‘‘if (Xt+1,i fitness > Xt,i fitness)’’
is better, then it is accepted, and the old solution is ignored.
in both occurrences in the pseudocode shown in Figure (3).
Additionally, one of the special features of FDO is that if
the new solution is not better, then the artificial scout bee V. RESULTS AND DISCUSSION
continues using the previous direction (using the previous To test the performance of this algorithm, a number of stan-
pace value if available), but only if it takes the scout bee to a dard benchmark test functions exist in the literature is used.
better solution. In addition, if using the previous pace is not Additionally, our results are compared to five other well-
leading the scout bee to a better solution, then FDO maintains known algorithms in the literature: PSO, GA, DA, WOA,
the current solution until the next iteration. In this algorithm, and SSA. It is worth mentioning that results of (19 classical
every time the solution is accepted, its pace value is saved for benchmark test functions) PSO, GA, and DA is taken from
potential reuse in the next iteration. this work [16], however, we conducted the CEC-C06 tests.
When implementing FDO for maximization problems, Also, all test results were compared using the Wilcoxon rank-
two minor changes are needed. First, Equation (2) must be sum test to prove their statistical significance. Moreover,
replaced by Equation (6), as Equation (6) is simply an inverse four measurement metrics were used for further observation.
version of Equation (2). Finally, the FDO was used for optimizing two real-world
  applications; thus, the section consists of five parts as follows.
x 
 i,t fitness 
fw =  ∗  − wf (6) A. CLASSICAL BENCHMARK TEST FUNCTIONS
 xi,t fitness 
Three sets of test functions are selected to test the perfor-
Second, the condition for selecting a better solution should mance of the FDO algorithm [16]. The test functions have dif-
be changed. The line: ‘‘if (Xt+1,i fitness < Xt,i fitness)00 must ferent characteristics, for instance, unimodal test functions,

VOLUME 7, 2019 43477

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 2. Classical benchmark results of selected algorithms with FDO [16].

multimodal test functions, and composite test functions. for all test functions wf was equal to 0 except test function
Each set of these test functions is used to benchmark cer- (2 and 6) where wf equal to 1. Every test function was
tain perspectives of the algorithm. Unimodal benchmark minimized towards 0.0 except TF8, which was minimized
functions, for example, are used for testing the exploitation towards -418.9829 (see Appendix Tables 6, 7 and 8 for more
level and convergence of the algorithm, as their name might details about the test function conditions). For example, some
imply that they have a single optimum. However, multimodal test functions were shifted by some degrees from the origin
benchmark functions have multi optimal solutions, and they point to prove that the algorithms were not biased towards the
are used for testing the local optima avoidance and explo- origin.
ration levels. As in multimodal algorithms, there are many In Table (2), the results of FDO, DA, PSO, and GA are
optimum solutions; one of them is a global optimum solution presented. The TF1 to TF6 results showed that FDO generally
and most local optimum solutions. An algorithm must avoid provided better results than the other algorithms; however,
local optimum solutions and converge to a global optimum the TF7 results showed the other algorithms were better.
solution. Furthermore, the composite benchmark functions FDO in TF8 showed poor performance even though it had
are mostly combined, shifted, rotated, and biased versions of better results than PSO. In contrast, TF9 FDO provided a
other test functions. Composite benchmark functions provide better result than both GA and DA, and comparative results
diverse shapes for different regions of the search landscape; were produced by PSO. In TF10 to TF13 and TF18, FDO pro-
they also have a very large number of local optima. This type vided relatively comparative results to the other algorithms.
of benchmark function demonstrates that complications exist However, the results of TF14 to TF17 and TF19 confirm that
in real-world search spaces (see Table (6, 7 and 8) in the the FDO algorithm outperformed DA, PSO, and GA in all
appendix) [16]. cases.
Each algorithm in Table (2) has been tested 30 times by
using 30 search agents each with 10 dimensions; in each B. CEC-C06 2019 BENCHMARK TEST FUNCTIONS
test, the algorithm was allowed to look for the best optimum A group of 10 modern CEC benchmark test functions is used
solution in 500 iterations, and then, the average and standard as an extra evaluation on FDO, these test functions were
deviation were calculated. Regarding parameter sets, GA, improved by professor Suganthan and his colleges for a single
PSO, and DA parameter sets described in this paper [16]. But objective optimization problem [1], the test functions are
for FDO parameters, there is only wf to be tuned. In Table (2), known as ‘‘The 100-Digit Challenge’’, which are intended to

43478 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 3. IEEE ECE 2019 benchmark results.

be used in annual optimization competition. See Table (9) in TABLE 4. The wilcoxon rank-sum test for classical benchmarks.
the appendix.
Functions CEC04 to CEC10 are shifted and rotated,
whereas functions CEC01 to CEC03 are not. However,
all test functions are scalable. The parameter set where
defined by the CEC benchmark developer, as functions
CEC04 to CEC10 where set as 10-dimensional minimiza-
tion problem in [−100, 100] boundary range, however,
CEC01 to CEC03 have different dimensions as shown in the
Appendix in Table 9. For more convenient, all CEC global
optimum where unified toward point 1. FDO is competed
with three modern optimization algorithms: DA, WOA, and
SSA. The reasons behind selecting these algorithms are:
1) They are all PSO-based algorithms same as FDO. 2) All
of them are well cited in the literature. 3) They are Proven
to have an outstanding performance both on benchmark test
functions and real-world problems. 4) These algorithm imple-
mentations are publicly provided by their authors. Regard-
ing algorithms parameter settings, their default parameter
settings were not modified during the tests, all competi-
tors are set the same as the settings used in their original
papers [16]–[18]. Interested readers can find these algorithms
MATLAB implementations and their parameter setting spec-
ification here [30]. Additionally, FDO default parameter set
wf = 0 is used for all test functions. a statistical comparison are shown in Table (4) and Table (5).
Each algorithm where allowed to search the landscape In Table (4), the comparison is conducted only between the
for 500 iterations using 30 agents. As shown in Table (3), FDO and DA algorithms because the DA algorithm was
FDO outperforms other algorithms except in CEC06. Even already tested against both PSO and GA in this paper [16].
though other algorithms have a comparative result in CEC03, According to the mentioned work, it has been proven that
CEC05, and CEC09 benchmarks, for example, WOA has the the DA results are statistically significant compared with PSO
same result as FDO in CEC03, but the WOA standard devi- and GA.
ation is equal to 0, this shows that WOA has the same result Again, as shown in Table (4), the FDO results are con-
every time it uses with no chance for further improvements. sidered significant in all statistical tests (unimodal, multi-
modal and composite test functions), except in TF2, that is
C. STATISTICAL TESTS because the results are more than 0.05. There are two unusual
To show that the results presented in Table (2) and Table (3) results in the composite test functions in both TF16 and TF18
are statistically significant, the p values of the Wilcoxon rank- because the DA algorithm provided the same fitness function
sum test are found for all test functions, and the results of value for each of the 30 different individual tests.

VOLUME 7, 2019 43479

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

FIGURE 4. Search history of the FDO algorithms on unimodal, multimodal, and composite test functions.

FIGURE 5. The trajectory of FDO’s search agents on unimodal, multimodal, and composite test functions.

FIGURE 6. Average fitness of FDO’s search agents on unimodal, multimodal, and composite test functions.

Table (5) shows the Wilcoxon rank-sum test of FDO D. QUANTITATIVE MEASUREMENT METRICS
against DA, WOA, and SSA for 10 CEC benchmark test func- For more detailed analyses and in-depth observation of
tions, the results show that FDO performances are statistically the FDO algorithm, four more quantitative metrics were
significant in all cases, except in test function CEC03 for used, as shown in Figures (4, 5, 6 and 7). In each exper-
DA and WOA algorithms, and test function CEC04 and iment, the first test function is selected from the unimodal
CEC08 for WOA algorithm. The results of Table (4) and benchmark functions (FT1 to FT7), the second test function
Table (5) prove that FDO results are statistically significant, selected is from the multimodal test functions (TF8 to TF13),
consequently, the existence of the FDO algorithm is statisti- and the last test function is selected from the composite
cally feasible. benchmark functions (TF14 to TF19). The experiment was

43480 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

FIGURE 7. Convergence curve of the FDO’s algorithms on unimodal, multimodal, and composite test functions.

TABLE 5. The wilcoxon rank-sum test (p-value) for CEC 2019.

FIGURE 9. Array configurations for 10-elements [32].

FIGURE 10. Global best with average fitness results for 200 Iteration with
20 artificial scout bees on aperiodic antenna array designs.

the scout quickly explores the overall area first and then
gradually moves towards optimality.
The second metric measures the value of the search agent
(fitness function value), as shown in Figure (5). The values
start with large values and then steadily decrease. This behav-
ior guarantees that FDO will eventually reach optimality [31].
FIGURE 8. Nonuniform antenna array and a thinned antenna array [22]. The third test metric is shown in Figure (6) and shows
that the average fitness value of all FDO agents decreased
dramatically over the course of the iterations, which verifies
conducted using 10 search agents, each allowed to search the that the algorithm not only improves the global best agent (xi∗ )
two-dimensional landscape through 150 iterations. but also improves the overall agent fitness values.
The first metric measures the convergence and illustrates The fourth metric measures the convergence of the global
how well the artificial scout covers the search landscape. best agent through the course of the iteration. This proves
This is merely a search history of artificial scout movements that xi∗ becomes more accurate as the number of iterations
because the position of the artificial scouts is recorded from increases again, clear abrupt changes can be seen due to the
the beginning to the end of the test. As presented in Figure (4), emphasis on the local search and exploitation, see Figure (10).

VOLUME 7, 2019 43481

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 6. Unimodal benchmark functions [16].

1) FDO USAGE ON APERIODIC ANTENNA ARRAY DESIGNS.

Since the 1960s, with the advances in both radar techniques
and radio astronomy, aperiodic antenna arrays have received
considerable attention, as shown in Figure (8); there are two
types of aperiodic antenna arrays: nonuniform antenna arrays
and thin antenna arrays.
In particular, to obtain the peak sidelobe level (SLL) in
nonuniform arrays, the element position should be optimized
in terms of a real number vector, as shown in Figure (9).
Moreover, to avoid grating lobes, a certain element spacing
limit exists for conventional periodic arrays (see constraints
in Equation (7)). Interested readers can review [32] for more
details on this problem.
FIGURE 11. Global best with average fitness results for 200 Iteration with Again, as shown in Figure (9), there are 10 elements of
30 artificial scout bees on the FM synthesis problem.
a nonuniform isotropic array, and only four element loca-
tions need to be optimized on each side. Since the outer-
most element is fixed at location 2.25λ0 with an average
Overall, in this section, measurement metrics showed that element spacing of davg = 0.5λ0 , this is a four-dimensional
FDO is capable of effectively exploring the search space, optimization problem with the following constraints:
improving the overall solution, avoiding local optimum and
xi ∈ (0, 2.25) xi − xj  > 0.25λ0
 
fairly converging towards optimality.
min {xi } > 0.125λ0 .i= 1, 2, 3, 4.i 6 = j. (7)
E. FDO REAL WORLD APPLICATION
Similar to any other metaheuristic algorithm, FDO can be The constraints show that there is a boundary between
used to solve real-world application problems. In this section, 0 and 2.25 for every element. However, each element cannot
FDO is applied to two different applications: be smaller than 0.125λ0 or larger than 2.0λ0 ; that is because

43482 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 7. Multimodal benchmark functions (10 dimensional) [16].

of 2.25λ0 is a fixed element and two adjacent elements can- 2) FDO ON FREQUENCY MODULATED SOUND WAVES
not get closer than 0.25λ0 . The fitness function problem is FDO is used on frequency-modulated sound waves (FM)
described as: to optimize the parameter of an FM synthesizer, which
has an essential role in several modern music systems; this
f = max {20 log |AF(θ )|} (8) problem has six parameters to be optimized as indicated
in Equation (10).
where
4 X = {a1 ,w1 , a2 , w2 , a3 , w3 } (10)
X
AF (θ ) = cos [2π xi (cos θ − cos θs )]
The objective of this problem is to generate a sound,
i=1
as in Equation (11), that is similar to the target sound, as
+ cos [2.25 × 2π (cos θ − cos θs )] (9)
in Equation (12).
Consider that θs = 90◦ in this work is defined in
y(t) = a1 . sin (w1 .t. + a2 . sin (w2 .t.θ + a3 . sin (w3 .t.θ )))
Figure (9) [32].
The DFO algorithm is used to optimize this problem, con- (11)
sidering the constraints mentioned in Equation (7). Twenty yo (t) = (1.0). sin ((5.0).t. + (1.5). sin ((4.8).t.θ
artificial scout search agents are used for 200 iterations, +(2.0). sin ((4.9).t.θ ))) (12)
and the presented result in Figure (10) includes the global
best fitness in each iteration and the average fitness value where the parameters should be in the range [−6.4, 6.35] and
according to Equation (8). The result shows that the global θ = 2π/100, the fitness function can be calculated using
best solution reached its optimum solution in iteration 78 with Equation (13), which is simply the summation of the square
element positions = {0.713, 1.595, 0.433, 0.130}. root between the result of Equation (11) and Equation (12),

VOLUME 7, 2019 43483

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 8. Composite benchmark functions [16].

while t = 100 turns. w2 = −0.0193, a3 = −0.5701, w3 = 4.937} were also

100 generated at iteration 200. The global best value converges
X
f (Ex ) = (y (t) − yo (t))2 (13) to the near-global optimal value from iteration 64.
t=0

Interested readers can find more details on this problem VI. CONCLUSION
in [33]. A new swarm intelligent algorithm was proposed called the
FDO is applied to the problem with 30 agents for 200 fitness dependent optimizer; it is inspired by the bee repro-
iterations, and records of the global best solutions and aver- ductive swarming process, where scout bees search for a
age fineness values can be seen in Figure (11). Parameter- new nest site. Additionally, the algorithm is inspired by their
set X = {a1 = 0.974, w1 = −0.241, a2 = −4.3160, collective decision-making. It has no algorithmic connection

43484 VOLUME 7, 2019

J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

TABLE 9. CEC-C06 2019 Benchmarks ‘‘The 100-Digit Challenge:’’ [1].

with the ABC algorithm. FDO employs fitness function val- REFERENCES
ues to generate weights that drive the search agents towards [1] K. V. Price, N. H. Awad, M. Z. Ali, and P. N. Suganthan, ‘‘The 100-digit
optimality. Additionally, FDO depends on the randomization challenge: Problem definitions and evaluation criteria for the 100-digit
challenge special session and competition on single objective numerical
mechanism in the initialization, exploration and exploitation optimization,’’ School Elect. Electron. Eng., Nanyang Technol. Univ.,
phases. A group of 19 single objective benchmark testing Singapore, Tech. Rep., Nov. 2018.
functions was used to test the performance of the FDO. [2] B. J. Copeland, Alan Turing’s Automatic Computing Engine. Oxford, U.K.:
Oxford Univ., 2005.
The benchmark testing functions were divided into three
[3] X.-S. Yang, Nature-Inspired Metaheuristic Algorithms. London, U.K.:
subgroups (unimodal, multimodal and composite test func- Luniver Press, 2010.
tions). Additionally, FDO tested on 10 modern CEC-C06 [4] I. Fister, Jr., X.-S. Yang, I. Fister, J. Brest, and D. Fister, ‘‘A brief review
benchmarks. The FDO results compared to two well-known of nature-inspired algorithms for optimization,’’ Elektrotehniški Vestnik,
vol. 80, no. 3, pp. 116–122, 2013.
algorithms (PSO and GA) and three modern algorithms [5] L. Bianchi, M. Dorigo, L. M. Gambardella, and W. J. Gutjahr, A Sur-
(DA, WOA, and SSA), FDO outperformed the competing vey on Metaheuristics for Stochastic Combinatorial Optimization, vol. 8.
algorithms in the majority of cases and produced a compara- Amsterdam, The Netherlandsl: Springer, 2008, pp. 239–287.
[6] M. Melanie, An Introduction to Genetic Algorithms. Cambridge, MA,
tive result on the others. The test results were compared using USA: MIT Press, 1999.
the Wilcoxon rank-sum test to prove their statistical signif- [7] S. Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi, ‘‘Optimization by
icance. Four additional experiments were conducted on the simulated annealing,’’ Science, vol. 220, no. 4598, pp. 671–680, May 1983.
FDO algorithm to measure, prove and verify the performance [8] J. Kennedy and R. Eberhart, ‘‘Particle swarm optimization,’’ in Proc. IEEE
Int. Conf. Neural Netw., Nov. 1995, pp. 1942–1948.
and credibility. Furthermore, FDO was practically applied to [9] R. Storn and K. Price, ‘‘Differential evolution—A simple and efficient
two real-world examples as evidence that the algorithm can heuristic for global optimization over continuous spaces,’’ J. Global
address real-life applications. Optim., vol. 11, pp. 341–359, Dec. 1997.
[10] Z. W. Geem, J. H. Kim, and G. V. Loganathan, ‘‘A new heuristic optimiza-
Generally, we found that the number of search agents was tion algorithm: Harmony search,’’ Simulation, vol. 76, no. 2, pp. 60–68,
related somehow to FDO performance after testing on many Feb. 2001.
standard test functions and real-world applications. Thus, [11] S. Nakrani and C. Tovey, ‘‘On honey bees and dynamic server allocation
using a small number of agents (below five) would notably in internet hosting centers,’’ Adapt. Behav., vol. 12, nos. 3–4, pp. 223–240,
Dec. 2044.
decrease the accuracy of the algorithm, and a large number of [12] D. T. Pham, M. Castellani, and M. Sholedolu, ‘‘The bees algorithm,’’
search agents would improve the accuracy and cost more time Manuf. Eng. Centre, Cardiff Univ., Cardiff, U.K., Tech. Rep., 2005.
and space, partially because the algorithm depends on the [13] X.-S. Yang and X. He, ‘‘Firefly algorithm: Recent advances and applica-
tions,’’ Int. J. Swarm Intell., vol. 1, no. 1, pp. 36–50, 2013.
fitness weight on the significant part of its searching mech- [14] X. S. Yang and S. Deb, ‘‘Cuckoo search via Lévy flights,’’ in Proc.
anism; in view of this, it is known as the fitness dependent World Congr. Nature Biologically Inspired Comput. (NaBIC), Dec. 2009,
optimizer. pp. 210–214.
Future works will adapt, implement and test both multi- [15] X.-S. Yang, ‘‘A new metaheuristic bat-inspired algorithm,’’ in Nature
Inspired Cooperative Strategies for Optimization, vol. 284. Berlin,
objective and binary objective optimization problems on Germany: Springer, 2010, pp. 65–74.
FDO. Finally, integrating evolutionary operators into FDO [16] S. Mirjalili, ‘‘Dragonfly algorithm: A new meta-heuristic optimization
and hybridizing it with other algorithms can be considered technique for solving single-objective, discrete, and multi-objective prob-
lems,’’ Neural Comput. Appl., vol. 27, no. 4, pp. 1053–1073, May 2015.
as potential future research. [17] S. Mirjaliliab and A. Lewisa, ‘‘The whale optimization algorithm,’’ Adv.
Eng. Softw., vol. 95, pp. 51–67, May 2016.
[18] S. Mirjalilia, A. H. Gandomibf, S. Z. Mirjalili, C. Saremia, H. Farisd,
and S. M. Mirjalilie, ‘‘Salp swarm algorithm: A bio-inspired optimizer for
VII. APPENDIX engineering design problems,’’ Adv. Eng. Softw., vol. 114, pp. 163–191,
See Tables 6–8. Dec. 2017.

VOLUME 7, 2019 43485

 J. M. Abdullah, T. A. Rashid: FDO: Inspired by the Bee Swarming Reproductive Process

[19] L. Cui et al., ‘‘A novel artificial bee colony algorithm with an adaptive JAZA MAHMOOD ABDULLAH was born in
population size for numerical function optimization,’’ Inf. Sci., vol. 414, Sulaimaniyah, Iraq, in 1985. He received the B.Sc.
pp. 53–67, Nov. 2017. degree in the field of computer science from the
[20] L. Cui, G. Li, X. Wang, Q. Lin, J. Chen, N. Lu , J. Lu, ‘‘A ranking-based University of Sulaimani, Iraq, in 2008, and the
adaptive artificial bee colony algorithm for global numerical optimiza- master’s degree (Hons.) in the field of software
tion,’’ Inf. Sci., vol. 417, pp. 169–185, Nov. 2017. systems and internet technology from the Univer-
[21] L. Cui et al., ‘‘A smart artificial bee colony algorithm with distance-fitness- sity of Sheffield, U.K., in 2012. He is currently
based neighbor search and its application,’’ Future Gener. Comput. Syst.,
pursuing the Ph.D. degree with the University of
vol. 89, pp. 478–493, Dec. 2018.
Sulaimani, working in the field of Artificial Intel-
[22] L. Cui et al., ‘‘An enhanced artificial bee colony algorithm with dual-
population framework,’’ Swarm Evol. Comput., vol. 43, pp. 184–206 , ligence in the subject of swarm-based algorithms
Dec. 2018. He worked as an Assistant Programmer for two years, and then continued
[23] E. M. Cortés-Toro, B. Crawford, J. A. Gómez-Pulido, R. Soto, and his study in the University of Sheffield. After that, he joined the Department
J. M. Lanza-Gutiérrez, ‘‘A new metaheuristic inspired by the vapour- of Informations Technology, University of Sulaimani, where he is currently
liquid equilibrium for continuous optimization,’’ Appl. Sci., vol. 8, no. 11, affiliated with.
p. 2080, Oct. 2018.
[24] D. H. Wolpert and W. G. Macready, ‘‘No free lunch theorems for optimiza-
tion,’’ IEEE Trans. Evol. Comput., vol. 1, no. 1, pp. 67–82, Apr. 1997.
[25] J. D. Villa, ‘‘Swarming behavior of honey bees (Hymenoptera: Apidae)
in southeastern louisiana,’’ Ann. Entomological Soc. Amer., vol. 97, no. 1,
pp. 111–116, Jan. 2004.
[26] A. Avitabile, R. A. Morse, and R. Boch, ‘‘Swarming honey bees
guided by pheromones,’’ Ann. Entomological Soc. Amer., vol. 68, no. 6,
pp. 1079–1082, Nov. 1975.
[27] H. Blackiston, Beekeeping For Dummies, 2nd ed. Medina, OH, USA:
A.I. Root Company, 2009.
[28] T. D. Seeley, Honeybee Democracy, Princeton, NJ. USA: Princeton Univ.
Press, 2010.
[29] K. M. Schultz, K. M. Passino, and T. D. Seeley, ‘‘The mechanism of flight
guidance in honeybee swarms: subtle guides or streaker bees?’’ J. Exp. TARIK AHMED RASHID received the Ph.D.
Biol., vol. 211, pp. 3287–3295, Oct. 2008. degree in computer science and informatics from
[30] A. Mirjalili and S. Mirjalili. (2015). Seyedali Mirjalili. Accessed: the College of Engineering, Mathematical and
Jan. 01, 2019.[Online]. Available: http://www.alimirjalili.com/Projects. Physical Sciences, University College Dublin
html (UCD), in 2006, where he was a Postdoctoral
[31] F. van den Bergh and A. P. Engelbrecht, ‘‘A study of particle swarm Fellow of the Computer Science and Informatics
optimization particle trajectories,’’ Inf. Sci., vol. 176, no. 8, pp. 937–971,
School, from 2006 to 2007. He joined the Uni-
Apr. 2006.
versity of Kurdistan Hewlêr, in 2017. His research
[32] N. Jin and Y. Rahmat-Samii, ‘‘Advances in particle swarm optimization
for antenna designs: Real-number, binary, single-objective and multiob- interests include three fields. The first field is the
jective implementations,’’ IEEE Trans. Antennas Propag., vol. 55, no. 3, expansion of machine learning and data mining
pp. 560–562, Mar. 2007. to deal with time series applications. The second field is the development
[33] S. Das and P. N. Suganthan, ‘‘Problem definitions and evaluation criteria of DNA computing, optimization, swarm intelligence, and nature inspired
for CEC 2011 competition on testing evolutionary algorithms on real world algorithms and their applications. The third field is networking, telecommu-
optimization problems,’’ Jadavpur Univ., Kolkata, India, Nanyang Technol. nication, and telemedicine applications.
Univ., Singapore, Tech. Rep., Dec. 2010.

43486 VOLUME 7, 2019