Heliyon 9 (2023) e15378

Contents lists available at ScienceDirect

Heliyon
journal homepage: www.cell.com/heliyon

Research article

Quasi-reflection learning arithmetic optimization algorithm firefly

search for feature selection
Nebojsa Bacanin a , Nebojsa Budimirovic a , K. Venkatachalam b,∗ ,
Hothefa Shaker Jassim c , Miodrag Zivkovic a , S.S. Askar d ,
Mohamed Abouhawwash e,f
a
Singidunum University, Danijelova 32, 11000 Belgrade, Serbia
b
Department of Applied Cybernetics, Faculty of Science, University of Hradec Králové, 50003 Hradec Králové, Czech Republic
c
Modern College of Business and Science, AL-Khuwair 133, Muscat, Oman
d
Department of Statistics and Operations Research, College of Science, King Saud University, P.O. Box 2455, Riyadh 11451, Saudi Arabia
e
Department of Mathematics, Faculty of Science, Mansoura University, Mansoura 35516, Egypt
f
Department of Computational Mathematics, Science and Engineering (CMSE), Michigan State University, East Lansing, MI 48824, USA

A R T I C L E I N F O A B S T R A C T

Keywords: With the whirlwind evolution of technology, the quantity of stored data within datasets is rapidly
Feature selection expanding. As a result, extracting crucial and relevant information from said datasets is a gruelling
Metaheuristics task. Feature selection is a critical preprocessing task for machine learning to reduce the excess
Aritmetic optimisation algorithm
data in a set. This research presents a novel quasi-reflection learning arithmetic optimization
Quasi-reflection-based learning
Firefly algorithm
algorithm - firefly search, an enhanced version of the original arithmetic optimization algorithm.
Quasi-reflection learning mechanism was implemented for enhancement of population diversity,
while firefly algorithm metaheuristics were used to improve the exploitation abilities of the
original arithmetic optimization algorithm. The aim of this wrapper-based method is to tackle
a specific classification problem by selecting an optimal feature subset. The proposed algorithm
is tested and compared with various well-known methods on ten unconstrained benchmark
functions, then on twenty-one standard datasets gathered from the University of California,
Irvine Repository and Arizona State University. Additionally, the proposed approach is applied
to the Corona disease dataset. The experimental results verify the improvements of the presented
method and their statistical significance.

1. Introduction

Considering that the complexities of the machine learning models are increasing, more and more datasets are becoming high-
dimensional. The analysis of datasets alike is notably strenuous since irrelevant, surplus, and raw data are profuse within the sets.
A copious amount of inessential data impedes the machine learning procedure and makes it computationally costly with inferior
performance and accuracy. Thus, obtaining relevant information is pivotal for addressing this problem. The feature selection (FS)
mechanism detects an optimal, informative feature subset from high-dimensional datasets, intending to mitigate the redundant
variables and minimize the complexity while increasing the prediction accuracy of the machine learning model.

* Corresponding author.
E-mail address: [email protected] (K. Venkatachalam).

https://doi.org/10.1016/j.heliyon.2023.e15378
Received 15 April 2022; Received in revised form 3 April 2023; Accepted 4 April 2023
Available online 6 April 2023
2405-8440/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

In literature, feature selection discerns three methods: filter, wrapper, and embedded methods. Filter-based methods use statistical
specifications to select and assess feature subsets. However, the wrapper-based method utilizes a specific classifier to evaluate and
determine essential features. Filter-based methods are less computationally demanding than wrapper-based, but the latter often yield
the best-performing feature subsets. With the embedded-based methods, algorithms perform feature selection throughout the model
training. Regarding computational complexity, embedded methods are intermediate to wrappers and filters. Metaheuristic algorithms
establish as the most productive optimization approaches when solving demanding, high-dimensional, real-world problems, such as
the early detection of serious diseases [1], the COVID-19 patient health predictions [2–4], traffic air pollution [5], landslide surveys
[6], radio-frequency identification readers collision management [7] or resolving optimization problems of artificial neural network
parameters [8] and expensive 200-dimensional problems [9]. Their technique consists of two search strategies: exploration and
exploitation. The exploration phase involves surveying the search space extensively, an aptitude that enables avoiding local optima.
Exploitation is the capacity to examine nearby auspicious solutions to advance their accuracy locally. A total balance between these
phases enables the exceptional performance of the algorithm.
No algorithm can resolve all optimization problems, per the No Free Lunch (NFL) theorem [10]. This theorem continually inspires
researchers to present new methods and hybridize or modify the existing algorithms to resolve numerous issues. Precisely this is be-
hind the proposal of a hybridized arithmetic optimization algorithm (AOA), adjusted for resolving FS problems in a wrapper-based
method. The objective is to improve the original AOA incorporating firefly search and quasi-reflection learning mechanism, thus
achieving high classification accuracy and efficient dimensionality reduction. The proposed algorithm is evaluated on 10 uncon-
strained benchmark functions exhibited on the Congress on Evolutionary Computation 2019 (CEC2019), and after on twenty-one
benchmark datasets from the University of California, Irvine (UCI) Repository and Arizona State University (ASU) and the COVID-19
dataset. The presented approach is tested against various established FS algorithms on datasets mentioned above for an extensive
evaluation. The results exhibit the superiority of the presented algorithm over other compared algorithms in selecting the most
significant features in most datasets.
Motivation for this research arises from the following question: Is it achievable to further enhance resolving FS problems with the
introduced algorithm by improving the accuracy and obtaining features with a significant effect on the target variable?
The summary of contributions of this study is as follows: an improved version of AOA metaheuristics that boosts exploration and
exploitation abilities of original AOA is proposed; the FS challenge solving is further improved in terms of classification metrics and
the number of utilized features and novel approach can be successfully applied to COVID-19 health prognosis, which is a considerate
benefit to health care.
The rest of this paper is organized as follows: Section 2 proposes an improved, hybridized version of the original AOA. Section 3
provides the results of the introduced approach and the comparative analysis with several well-known algorithms for standard
CEC2019 benchmarks, followed by 21 standard datasets. Moreover, this section also shows the application of the proposed method
on the COVID-19 dataset. Section 4 outlines the conclusions.

2. Proposed method

The optimization process of population-based algorithms starts with a randomly generated candidate solutions set. These solutions
are enhanced with optimization rules and evaluated by a determined fitness function. As a consequence of the stochastic nature of
population-based algorithms, finding the optimal solution demands an adequate number of random solutions and optimization
iterations.
In this study, a hybridized AOA metaheuristic that targets deficiencies of the original approach was developed. The AOA [11] em-
ploys arithmetic operators division (D), multiplication (M), subtraction (S) and addition (A) in the search process. In the initialization
phase of AOA, a randomly generated set of candidate solutions (X) is given in matrix, shown in Eq. (1).

𝑋 = [𝑥𝑖,𝑗 ]𝑁×𝑛 (1)

Prior to the initialization phase, the algorithm selects the search phase (i.e. exploration or exploitation). Therefore, Math Optimizer
Accelerated (MOA) function is used, as shown in Eq. (2):
( )
𝑀𝑎𝑥 − 𝑀𝑖𝑛
𝑀𝑂𝐴(𝐶_𝐼𝑡𝑒𝑟) = 𝑀𝑖𝑛 + 𝐶_𝐼𝑡𝑒𝑟 ⋅ (2)
𝑚𝑎𝑥𝐼𝑡𝑒𝑟

where 𝑀𝑂𝐴(𝐶_𝐼𝑡𝑒𝑟) marks the function value at the 𝑡th iteration. 𝐶_𝐼𝑡𝑒𝑟 marks the current iteration, and 𝑚𝑎𝑥𝐼𝑡𝑒𝑟 denotes the
maximum number of iterations. 𝑀𝑖𝑛 designates the minimum, and 𝑀𝑎𝑥 - the maximum value of the 𝑀𝑂𝐴 function.
The exploration phase utilizes arithmetic operators, division and multiplication because their high-distributed values make them
suitable for the exploration search mechanism. Updating positions for these two exploration strategies (division and multiplication
search strategies) are described in Eq. (3). This exploration phase is executed for the condition of 𝑟1 > 𝑀𝑂𝐴(𝐶_𝐼𝑡𝑒𝑟), where 𝑟1 is a
random number.
⎧
⎪ 𝑏𝑒𝑠𝑡(𝑥𝑗 ) ÷ (𝑀𝑂𝑃 + 𝜖) ⋅ ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 ) ⋅ 𝜇 + 𝑙𝑏𝑗 ),
⎪ 𝑟2 < 0.5
𝑥𝑖,𝑗 (𝐶_𝐼𝑡𝑒𝑟 + 1) = ⎨ (3)
⎪ 𝑏𝑒𝑠𝑡(𝑥𝑗 ) ⋅ 𝑀𝑂𝑃 ⋅ ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 ) ⋅ 𝜇 + 𝑙𝑏𝑗 ),
⎪ 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
⎩

2
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

𝑥𝑖 (𝐶_𝐼𝑡𝑒𝑟 + 1) marks the 𝑖th solution in the next iteration, 𝑥𝑖,𝑗 (𝐶_𝐼𝑡𝑒𝑟) designates the 𝑗th position of the 𝑖th solution at the current
iteration, while 𝑏𝑒𝑠𝑡(𝑥𝑗 ) is the 𝑗th position in the best-gained solution thus far. 𝜖 is a small integer number, 𝑢𝑏_𝑗 and 𝑙𝑏_𝑗 denote upper
and lower bound value of the 𝑗th position, respectively. 𝜇 is a control parameter of the search process, set to 0.5, according to the
original paper, and 𝑟2 is a random number. Math optimizer probability (𝑀𝑂𝑃 ) coefficient is calculated as shown in Eq. (4):
𝐶_𝐼𝑡𝑒𝑟1∕𝛼1
𝑀𝑂𝑃 (𝐶_𝐼𝑡𝑒𝑟) = 1 − (4)
𝑚𝑎𝑥𝐼𝑡𝑒𝑟1∕𝛼1
where 𝑀𝑂𝑃 (𝐶_𝐼𝑡𝑒𝑟) denotes the function value at the 𝑡th iteration. 𝛼1 is a parameter that defines the exploitation accuracy through
iterations and is set to 5, according to the original paper. The exploitation phase utilizes arithmetic operators S and A because
they provide high-dense results suitable for exploitation search mechanisms. This exploitation phase is executed for the condition
𝑟1 ≤ 𝑀𝑂𝐴(𝐶_𝐼𝑡𝑒𝑟). Updating positions in the exploitation phase are modelled in the following Eq. (5):

⎧
⎪ 𝑏𝑒𝑠𝑡(𝑥𝑗 ) − 𝑀𝑂𝑃 ⋅ ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 ) ⋅ 𝜇 + 𝑙𝑏𝑗 ),
⎪ 𝑟3 < 0.5
𝑥𝑖,𝑗 (𝐶_𝐼𝑡𝑒𝑟 + 1) = ⎨ (5)
⎪ 𝑏𝑒𝑠𝑡(𝑥𝑗 ) + 𝑀𝑂𝑃 ⋅ ((𝑢𝑏𝑗 − 𝑙𝑏𝑗 ) ⋅ 𝜇 + 𝑙𝑏𝑗 ),
⎪ 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
⎩
where 𝑟3 is a random number. The exploitation search strategies generally aim to prevent from getting trapped in the local search
space. This process aids the exploration search strategies to obtain the optimal solution while maintaining the diversity of candidate
solutions.
During extensive empirical simulations on constrained and bound-constrained CEC benchmarks, it was observed the AOA ex-
ploration is not enough in the early phases of some runs, and the search process gets trapped in sub-optimal domains. Moreover,
notwithstanding the exploitation abilities of the original AOA, it was noticed that in later iterations, the exploitation process could
be improved to facilitate search around the global optimum region of the search space.
These deficiencies were addressed by incorporating two changes in the original approach: the quasi-reflection learning (QRL)
mechanism, which showed potential in enhancing population diversity [12], and the search strategy from the prominent firefly
algorithm (FA) metaheuristics, that exhibits strong intensification abilities [13].
The QRL is first applied in the population initialization in the following way: when a standard random population 𝑃 of size 𝑁 is
generated, for each solution 𝑋 ∈ 𝑃 , its quasi-reflective individual 𝑋 𝑞𝑟 is generated by applying Eq. (6) and population 𝑃 𝑞𝑟 is formed;
the fitness is calculated for all solutions from 𝑃 ∪ 𝑃 𝑞𝑟 and best 𝑁 individuals are selected. In this way, initial pseudo-random solutions
are closer to optimum.
The quasi-reflected component 𝑗 of the solution 𝑋 (𝑋𝑗𝑞𝑟 ) is calculated as:
( )
𝑙𝑏𝑗 + 𝑢𝑏𝑗
𝑋𝑗𝑞𝑟 = rnd , 𝑥𝑗 (6)
2
( )
𝑙𝑏𝑗 + 𝑢𝑏𝑗 𝑙𝑏𝑗 + 𝑢𝑏𝑗
where is the arithmetic mean (center) of the interval [𝑙𝑏𝑗 , 𝑢𝑏𝑗 ], while rnd , 𝑥𝑗 generates uniformly distributed
2 2
[ ]
𝑙𝑏𝑗 + 𝑢𝑏𝑗
pseudo-random from the interval , 𝑥𝑗 .
2
Also, in early 𝜓 iterations, with the goal of enhancing AOA’s exploration abilities, quasi-reflective solution 𝑋 ∗,𝑞𝑟 of current best
individual 𝑋 ∗ is created and evaluated and better solution is retained in population. However, in later phases (after 𝜓 iteration), this
mechanism is no longer needed and the search process of the basic AOA is replaced with the FA’s search equation Eq. (7), which is
employed for every parameter 𝑗 of every solution 𝑖 [13]:
−𝛾𝑟2𝑖,𝑧
𝑋𝑖𝑡+1 = 𝑥𝑡𝑖 + 𝛽0 ⋅ 𝑒 (𝑥𝑡𝑧 − 𝑥𝑡𝑖 ) + 𝛼2𝑡 (𝜅 − 0.5) (7)
where parameters 𝛽0 and 𝛼2 are standard FA’s search parameters and 𝜅 denotes a random number derived from either Gaussian or
uniform distribution. Solution 𝑧 is randomly selected solutions from population, while parameter 𝛼2 is dynamic and its value adjusts
during a run as suggested in [13].
Inspired by the introduced changes, presented method’s name is quasi-reflection learning AOA - firefly search (QRLAOA-FS). Its
high-level pseudo-code, with emphasis on the incorporated modifications, is given in Algorithm 1.

3. Experimental results and comparative analysis

Firstly, the proposed approach is substantiated on unconstrained benchmark functions, then, tested on twenty-one standard FS
datasets and finally, applied on COVID-19 dataset for health prediction. In all experiments, control parameters for QRLAOA-FS were
set as follows: basic AOA parameters 𝛼1 = 5 and 𝜇 = 0.5, and initial values for FA parameters 𝛼2 = 0.5 and 𝛽0 = 1. All observed
approaches in comparative analysis were applied for the purpose of this study and validated with control parameters’ values set
according to the original papers.
All methods were implemented in Python 3.9 programming language, while for machine learning algorithms, such as KNN, the
scikitlearn library was used. For data preprocessing, pandas library was utilized. Also, for the purpose of visualization Python library
pyplot from matplotlib package and seaborn were used. Python was chosen due to its relatively fast execution speed (very important

3
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

Algorithm 1 QRLAOA-FS algorithm pseudo-code.

Initialize control parameters of AOA and FA algorithms
Generate initial population 𝑃 of 𝑁 individuals
Generate quasi-reflection population 𝑃 𝑞𝑟
Calculate fitness of individuals from 𝑃 ∪ 𝑃 𝑞𝑟 and select 𝑁 best solutions
while 𝑡 ≤ 𝑚𝑎𝑥𝑖𝑡𝑒𝑟 do
for (𝑖 = 1 to 𝑁 do
if 𝑡 < 𝜓 then
Perform basic AOA exploitation and exploration
else
Perform firefly search according to expression Eq. (7)
end if
end for
Determine 𝑋 ∗ solution
if 𝑡 < 𝜓 then
Generate 𝑋 ∗,𝑞𝑟 by applying QRL mechanism according to Eq. (6) and evaluate its fitness
Perform greedy selection between 𝑋 ∗ and 𝑋 ∗,𝑞𝑟
end if
end while
Return 𝑋 ∗
Post-process results and visualization

Table 1
Mean fitness function and standard deviation of fitness functions results comparison, and average ranking for Friedman test of different approaches on CEC2019
benchmark functions.
Function Stats EHOI EHO SCA SSA GOA WOA BBO MFO PSO FFO AOA QRLAOA-FS

mean 4.76E+04 1.35E+07 9.82E+09 3.21E+09 1.61E+10 1.03E+10 3.52E+10 7.17E+09 6.74E+11 1.46E+03 2.38E+03 7.28E+02
CEC01
std 2.69E+03 7.74E+06 5.47E+08 2.06E+09 1.14E+10 8.81E+09 2.55E+10 7.58E+09 6.53E+11 4.46E+02 3.85E+02 2.45E+01
mean 1.70E+01 1.72E+01 1.76E+01 1.73E+01 1.74E+01 1.74E+01 8.86E+01 1.74E+01 8.56E+02 2.80E+00 4.89E+00 2.54E+00
CEC02
std 1.07E-15 4.82E-15 4.00E-02 8.07-05 1.40E-02 2.78E-03 2.50E+01 3.83E-15 3.97E+01 1.38E+02 2.91E-04 7.96E-05
mean 1.26E+01 1.26E+01 1.26E+01 1.26E+01 1.26E+01 1.26E+01 1.26E+01 1.26E+01 1.26E+01 9.72E+00 9.56E+00 5.83E+00
CEC03
std 1.90E-15 1.90E-15 1.09E-04 2.34E-15 1.21E-04 1.39E-07 2.58E-07 3.39E-05 6.61E-04 4.27E-01 7.16E-05 3.48E-05
mean 1.28E+01 1.55E+01 8.33E+02 3.25E+01 1.51E+02 2.66E+02 6.95E+01 1.39E+02 6.92E+01 3.99E+00 4.78E+00 1.16E+00
CEC04
std 3.85E+00 6.38E+00 2.92E+02 1.33E+01 1.49E+02 1.28E+02 2.35E+01 1.57E+02 8.02E+00 4.21E-01 5.84E-01 2.65E-01
mean 1.06E+00 1.07E+00 2.24E+00 1.35E+00 1.33E+00 1.67E+00 1.31E+00 1.14E+00 1.55E+00 1.03E+00 1.07E+00 1.00E+00
CEC05
std 2.13E-02 2.20E-02 7.80E-02 1.12E-01 1.41E-01 4.19E-01 9.68E-02 8.23E-02 1.17E-01 2.34E-02 1.98E-02 2.04E-02
mean 8.334E+00 9.46E+00 1.05E+01 3.79E+00 6.20E+00 9.13E+00 5.80E+00 4.93E+00 1.04E+01 1.61E+00 2.39E+00 1.54E+00
CEC06
std 8.57E-01 1.25E+00 7.58E-01 1.27E+00 1.37E+00 1.04E+00 6.44E-01 2.21E+00 6.78E-01 6.17E-02 1.76E-01 5.92E-02
mean 1.43E+02 1.80E+02 6.39E+02 2.89E+02 2.87E+02 4.54E+02 4.92E+00 3.19E+02 6.97E+02 5.91E+00 4.83E+00 5.07E+00
CEC07
std 3.97E+02 1.43E+02 1.39E+02 2.36E+02 1.74E+02 2.17E+02 1.26E+02 2.10E+02 1.63E+02 1.08E+02 6.93E+01 7.18E+01
mean 2.70E+00 3.16E+00 5.77E+00 5.08E+00 5.49E+00 5.75E+00 4.80E+00 5.46E+00 5.11E+00 1.15E+00 1.34E+00 1.03E+00
CEC08
std 8.64E-01 1.18E+00 5.49E-01 6.42E-01 8.13E-01 7.77E-01 1.13E+00 5.79E-01 7.38E-01 7.99E-01 1.05E+00 5.37E-01
mean 2.29E+00 2.41E+00 9.74E+01 2.38E+00 2.45E+00 5.16E+00 3.75E+00 2.46E+00 2.65E+00 3.91E+00 2.58E+00 2.05E+00
CEC09
std 6.35E-03 1.39E-02 9.22E+01 4.52E-02 7.29E-02 7.60E-01 2.51E-01 6.24E-02 9.41E-02 2.85E-01 5.23E-02 5.49E-03
mean 1.92E+01 2.11E+01 2.08E+01 2.03E+01 2.01E+01 2.05E+01 2.08E+01 2.02E+01 2.06E+01 2.11E+01 2.06E+01 1.88E+01
CEC10
std 1.52E+00 1.04E-01 8.27E-02 8.29E-02 9.24E-02 4.92E-02 2.29E-02 1.49E-01 1.08E-01 1.14E-05 2.45E-03 7.16E-04

average rank 4.4 6.45 10.08 6.25 7.65 9.25 7.7 7.15 9.45 4 3.7 1.2

for conducted simulations) and because of high availability of machine learning and data pre-processing libraries. All simulations
were executed on Intel ® I9 11900K platform with 64GB of RAM. Computing platform also has NVIDIA GPU 1080 11G, however,
the GPU was not used because used libraries do not support executing on GPU by using CUDA ® technology.
As it is explained in Subsection 3.1, instead of maximum number of iterations (𝑚𝑎𝑥𝐼𝑡𝑒𝑟), number of fitness function evalua-
tions (𝐹 𝐹 𝐸𝑠) is utilized as the termination condition and QRLAOA-FS 𝜓 control parameter value was determined empirically and
calculated according to expression 𝐹 𝐹 𝐸𝑠∕3.

3.1. Simulation for unconstrained functions

The original AOA and ten well-known metaheuristic methods: salp swarm algorithm (SSA) [14], particle swarm optimization
(PSO) [15], biogeography-based optimization (BBO) [16], elephant herding optimization (EHO) [17], EHO improved (EHOI) [18],
fruit fly optimisation algorithm (FFO) [19], whale optimization algorithm (WOA) [20], moth-flame optimization (MFO) [21], sine
cosine algorithm (SCA) [22] and grasshopper optimization algorithm (GOA) [23] are evaluated on 10 recent benchmark functions
set (CEC2019) [24], under similar conditions. Obtained results are put into comparison with the ones gathered using the QRLAOA-FS
algorithm. Characteristics of the CEC2019 bound-constrained benchmark function are stated in [24].
The simulation results of the aforementioned algorithms are obtained for the same benchmarks in research [18]. The same simu-
lation is rerun to provide an unbiased comparison and to confirm the findings from [18]. Control parameters for tested approaches in
[18], population size 𝑁 = 50 and a maximum number of iterations 𝑚𝑎𝑥𝐼𝑡𝑒𝑟 = 500, tend to be biased since algorithms do not employ
an equal amount of fitness function evaluations (𝐹 𝐹 𝐸𝑠) in a single iteration. Most of these methods require 𝑁 evaluations during
the initialization stage and then complete one more 𝐹 𝐹 𝐸 for each member of the population in each subsequent iteration. Therefore,
the termination condition is set to 25050 for all methods, establishing an unbiased comparative analysis. This experiment is repeated
in 30 separate runs.
The experimental results are shown in Table 1, correlating mean values and standard deviations of proposed and compared
approaches.

4
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

For each benchmark instance, the best mean fitness value and the best standard deviation value are marked with bold and bold
italic style, respectively. The stochastic nature of the observed algorithms is why the results obtained deviate slightly from those in
the study [18]. QRLAOA-FS gave the optimal mean value regarding nine of ten functions. AOA performed best on function CEC07,
marginally better than the QRLAOA-FS. The proposed algorithm provided the lowest standard deviation on five benchmark functions,
followed by AOA and EHOI.
The Friedman test [25], a two-way variance analysis by ranks, is carried out in an effort to validate the importance of improve-
ments statistically. Additionally, the ranking of 12 algorithms used on 10 functions is shown in Table 1. The Friedman test average
ranking for QRLAOA-FS is 1.2, demonstrating its superiority to the other eleven algorithms. The Friedman statistics (65.073) is greater
than the 𝜒 2 critical value (𝜒 2 = 19.673) at the significance level 𝛼 = 0.005, which leads to the conclusion that QRLAOA-FS is signifi-
cantly different from the compared algorithms. The second test also rejects 𝐻0 since the 𝐹 -distribution critical value (1.887) is smaller
than the realized Iman-Davenport statistic (13.036). In both tests, the significance level is higher than the 𝑝 − 𝑣𝑎𝑙𝑢𝑒 (1.11𝐸 − 16).
Given that both tests rejected the null hypothesis, Holm’s step-down approach is applied as a post hoc procedure. The provided
approach significantly outperformed all compared methods with a significance level 𝛼 set to 0.1 and ten out of eleven (all methods
omitting HLBDA) with a significance level 𝛼 set to 0.05.
All this leads to the conclusion that QRLAOA-FS outperformed the original AOA and other tested metaheuristics.

3.2. Feature selection simulation results

The proposed algorithm, original AOA and ten popular algorithms: FFO, binary particle swarm optimization (BPSO) [26], binary
coyote optimization algorithm (BCOA) [27], binary dragonfly algorithm (BDA) [28], hyper learning binary dragonfly algorithm
(HLBDA) [29], chaotic crow search algorithm (CCSA) [30], binary multiverse optimizer (BMVO) [31], binary artificial bee colony
(BABC) [32], success-history based adaptive differential evolution with linear population size reduction (LSHADE) [33] and evolution
strategy with covariance matrix adaptation (CMAES) [34] are tested on twenty-one standard datasets, taken from the UCI repository
[35] and Arizona State University [36]. The parameters for these algorithms are given in [29]. A comprehensive description of
employed datasets containing a various number of dimensions and features is presented in [29]. Thus, the proposed algorithm’s per-
formance is estimated on differing models, demonstrating its effectiveness in varying dimensions. The performance of the algorithm
is evaluated using four metrics: classification accuracy, feature selection ratio, mean fitness value, and standard deviation of fitness
value.
This study chooses the fitness function to increase classification accuracy while minimizing the number of selected features. The
|𝑆|
fitness function is as follows 𝐹 𝑖𝑡 = 𝛼3 𝐸𝑅 + 𝛽 |𝑂| , where 𝐸𝑅 is the classification’s error, |𝑆| is the length of the subset of selected
features, while |𝑂| represents the length of the original features. To illustrate how classification error and feature size affect the
fitness function, two weight infectors 𝛼3 ∈ [0, 1] and 𝛽 = (1 − 𝛼3 ), are employed.
Multiple optimization problems are modelled in binary space, such as feature selection in data classification. Abundant binary
metaheuristic methods [37,29] are generated, primarily for wrapper-based FS, intending to solve these issues. Transfer functions (TF)
are utilized to convert continuous values to binary. QRLAOA-FS algorithm is tested using S-shaped and V-shaped transfer functions.
V-shaped TF [37] yielded excellent results and is consequently applied in the introduced approach.
The dataset is split into a training set, and an evaluation set using the stratified 10-fold cross-validation approach described in
the paper [29]. The K nearest neighbour (KNN), where 𝑘 is set to 5, is applied to determine classification error in wrapper-based
feature selection. Simplicity and low computational cost are the assets of utilizing KNN as a learning technique. Owing to the non-
deterministic nature of optimization algorithms, each approach is applied 20 times. The outcomes’ averages are compiled. As not all
algorithms employ the equal amount of 𝐹 𝐹 𝐸𝑠 per iteration, the maximum number of iterations, 𝑚𝑎𝑥𝐼𝑡𝑒𝑟 = 100, can result in a biased
comparative study. As a result, 1010 is set as the termination condition 𝑚𝑎𝑥𝐹 𝐹 𝑒s. Population size is set to 𝑁 = 10.
Table 2 displays testing results of mean fitness, a standard deviation of fitness function and feature selection ratio for tested
algorithms. QRLAOA-FS detected the optimal mean fitness value on thirteen datasets, followed by HLBDA, which obtained optimal
values on four datasets. QRLAOA-FS identified the lowest standard deviation in nine datasets, followed by BABC with five. The
results show that QRLAOA-FS yielded the smallest feature size in eleven datasets, accompanied by HLBDA with six. In comparison
with AOA and other approaches, QRLAOA-FS can repeatedly select a small-scale, most relevant subset of features.
To demonstrate the stability of the proposed method, Fig. 1 provides boxplot analysis of classification accuracy in Subfig-
ures 1b, 1d, 1f, 1h, and average error rate convergence graphs in Subfigures 1a, 1c, 1e, 1g, of QRLAOA-FS, AOA and four more
methods on four datasets.
The proposed QRLAOA-FS is relatively stable, upholding the best classification accuracy performance on three of four datasets
(Glass, Colon, Horse Colic and Libras Movement), followed by HLBDO. QRLAOA-FS surpasses the original AOA in all four cases.
Fig. 1 provides average error rate convergence graphs for 1000 𝐹 𝐹 𝐸s, proving that QRLAOA-FS generated a drastically better initial
population than AOA on two datasets (Horse Colic and Glass) and significantly better initial population than the other four methods
on tree datasets (Horse Colic, Glass and Colon).

3.3. COVID-19 dataset and results

Since its inception, the coronavirus disease (COVID-19) has been responsible for millions of fatalities worldwide. Artificial intel-
ligence can assist in COVID-19 diagnosis, prevention and detection [38]. The use of the suggested approach in predicting COVID-19

5
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

Table 2
The result of the mean fitness value, standard deviation of fitness value and feature selection ratio of algorithms.

Dataset Stats HLBDA BDA BABC BMVO BPSO CCSA BCOA CMAES LSHADE FFO AOA QRLAOA-FS

mean 0.0112 0.0112 0.0111 0.0116 0.0111 0.0116 0.0111 0.0118 0.0116 0.0114 0.0116 0.0112
Glass std 0.0033 0.0033 0.0033 0.0033 0.0033 0.0032 0.0033 0.0036 0.0033 0.0033 0.0032 0.0032
fs ratio 0.2900 0.2900 0.3050 0.3250 0.3050 0.3300 0.3050 0.2900 0.3050 0.2900 0.3050 0.2900
mean 0.1312 0.1369 0.1384 0.1452 0.1335 0.1455 0.1424 0.1428 0.1400 0.1402 0.1338 0.1304
Hepatitis std 0.0094 0.0063 0.0059 0.0100 0.0079 0.0065 0.0138 0.0133 0.0092 0.0068 0.0076 0.0061
fs ratio 0.3185 0.3660 0.3159 0.3604 0.3261 0.3527 0.3053 0.3501 0.3370 0.3248 0.3327 0.3145
mean 0.1312 0.1360 0.1351 0.1525 0.1340 0.1513 0.1418 0.1394 0.1475 0.1353 0.1382 0.1334
Lymphography std 0.0129 0.0120 0.0126 0.0105 0.0140 0.0115 0.0133 0.0151 0.0148 0.0122 0.0120 0.0103
fs ratio 0.4499 0.4806 0.5082 0.4890 0.5002 0.4973 0.4471 0.5085 0.4334 0.4652 0.4561 0.4328
mean 0.5849 0.5933 0.5845 0.6086 0.5849 0.5998 0.5943 0.5935 0.5990 0.5898 0.5896 0.5819
Primary Tumor std 0.0103 0.0100 0.0086 0.0081 0.0114 0.0118 0.0136 0.0136 0.0135 0.0102 0.0111 0.0092
fs ratio 0.6678 0.6120 0.6707 0.5943 0.6675 0.6439 0.6060 0.6264 0.6414 0.6071 0.6107 0.6012
mean 0.2125 0.2214 0.2256 0.2593 0.2246 0.2481 0.2170 0.2179 0.2209 0.2206 0.2215 0.2113
Soybean std 0.0088 0.0095 0.0088 0.0120 0.0043 0.0110 0.0102 0.0114 0.0111 0.0099 0.0091 0.0084
fs ratio 0.6530 0.6230 0.6429 0.5745 0.6415 0.5972 0.6373 0.6284 0.6485 0.6091 0.6343 0.5819
mean 0.1360 0.1429 0.1481 0.1673 0.1410 0.1701 0.1432 0.1420 0.1481 0.1413 0.1431 0.1358
Horse Colic std 0.0040 0.0090 0.0075 0.0163 0.0071 0.0141 0.0109 0.0099 0.0156 0.1008 0.0079 0.0053
fs ratio 0.0871 0.1368 0.2425 0.2576 0.1964 0.2926 0.1055 0.1129 0.1501 0.1218 0.1312 0.0967
mean 0.0843 0.0930 0.1014 0.1107 0.0960 0.1112 0.0871 0.0883 0.0911 0.0976 0.0931 0.0869
Ionosphere std 0.0086 0.0104 0.0102 0.0054 0.0056 0.0091 0.0105 0.0071 0.0100 0.0092 0.0089 0.0050
fs ratio 0.2191 0.2678 0.2899 0.2881 0.2439 0.3425 0.2266 0.2455 0.2822 0.2476 0.2436 0.2201
mean 0.0400 0.0408 0.0400 0.0480 0.0368 0.0494 0.0439 0.0471 0.0501 0.0434 0.0427 0.0395
Zoo std 0.0080 0.0081 0.0072 0.0101 0.0067 0.0098 0.0093 0.0100 0.0114 0.0082 0.0078 0.0069
fs ratio 0.4562 0.4469 0.4969 0.5189 0.4251 0.4937 0.4532 0.4499 0.4938 0.4463 0.4472 0.4240
mean 0.0674 0.0832 0.0959 0.1081 0.0930 0.1018 0.0791 0.0844 0.0795 0.0819 0.0829 0.0751
Musk 1 std 0.0063 0.0100 0.0064 0.0075 0.0079 0.0079 0.0078 0.0075 0.0098 0.0081 0.0079 0.0075
fs ratio 0.4686 0.4782 0.4946 0.4604 0.4963 0.5033 0.4460 0.4947 0.4848 0.4607 0.4596 0.4441
mean 0.3159 0.3340 0.3451 0.3540 0.3415 0.3528 0.3292 0.3351 0.3272 0.3286 0.3328 0.3150
Arrhythmia std 0.0094 0.0089 0.0044 0.0075 0.0073 0.0066 0.0099 0.0059 0.0131 0.0090 0.0069 0.0052
fs ratio 0.4051 0.4699 0.4805 0.4302 0.4706 0.4789 0.4050 0.4628 0.4496 0.4381 0.4417 0.4095
mean 0.0173 0.0192 0.0202 0.0254 0.0195 0.0238 0.0209 0.0181 0.0198 0.0197 0.0234 0.0186
Dermatology std 0.0020 0.0034 0.0023 0.0025 0.0020 0.0039 0.0025 0.0023 0.0035 0.0027 0.0025 0.0023
fs ratio 0.4575 0.4826 0.5572 0.5425 0.5341 0.5425 0.4370 0.4926 0.5074 0.4819 0.4803 0.4491
mean 0.1507 0.1543 0.1573 0.1705 0.1541 0.1632 0.1612 0.1646 0.1600 0.1545 0.1554 0.1483
SPECT Heart std 0.0068 0.0083 0.0082 0.0096 0.0078 0.0097 0.0203 0.0185 0.0169 0.0090 0.0080 0.0066
fs ratio 0.4477 0.4499 0.4866 0.5274 0.5044 0.5249 0.4161 0.4751 0.4274 0.4614 0.4691 0.4463
mean 0.1814 0.1938 0.2026 0.2094 0.2005 0.2074 0.1858 0.1889 0.1900 0.1927 0.1941 0.1803
Libras Movement std 0.0084 0.0103 0.0086 0.0051 0.0075 0.0091 0.0101 0.0089 0.0122 0.0078 0.0084 0.0049
fs ratio 0.4161 0.4518 0.4598 0.4313 0.4488 0.4645 0.4061 0.4302 0.4323 0.4256 0.4329 0.4046
mean 0.2789 0.2788 0.2803 0.2815 0.2792 0.2801 0.2815 0.2844 0.2835 0.2801 0.2799 0.2746
ILPD std 0.0050 0.0049 0.0049 0.0054 0.0047 0.0051 0.0063 0.0081 0.0065 0.0050 0.0050 0.0046
fs ratio 0.2950 0.3150 0.3350 0.2800 0.3250 0.3200 0.3050 0.3450 0.2950 0.2950 0.3250 0.2800
mean 0.0557 0.0557 0.0529 0.0533 0.0529 0.0529 0.0532 0.0538 0.0559 0.0537 0.0533 0.0529
Seeds std 0.0152 0.0152 0.0057 0.0060 0.0057 0.0057 0.0061 0.0066 0.0151 0.0061 0.0061 0.0060
fs ratio 0.3143 0.3143 0.3143 0.3214 0.3143 0.3143 0.3214 0.3430 0.3214 0.3214 0.3143 0.3143
mean 0.3008 0.3165 0.3173 0.3169 0.3174 0.3294 0.3280 0.3204 0.3337 0.3201 0.3176 0.3017
LSVT std 0.0312 0.0273 0.0190 0.0209 0.0243 0.0221 0.0227 0.0201 0.0217 0.0203 0.0188 0.0190
fs ratio 0.2845 0.3483 0.4501 0.4007 0.4426 0.4495 0.2986 0.4199 0.3166 0.3781 0.3624 0.3002
mean 0.1260 0.1311 0.1344 0.1415 0.1331 0.1412 0.1292 0.1285 0.1264 0.1321 0.1314 0.1249
SCADI std 0.0081 0.0091 0.0058 0.0064 0.0063 0.0066 0.0117 0.0074 0.0080 0.0076 0.0069 0.0056
fs ratio 0.2888 0.3728 0.4372 0.4154 0.4250 0.4523 0.3182 0.3980 0.3487 0.3972 0.3628 0.2879
mean 0.1577 0.1836 0.2138 0.2371 0.2072 0.2299 0.1816 0.2197 0.1778 0.1829 0.1842 0.1558
TOX 171 std 0.0213 0.0273 0.0179 0.0156 0.0177 0.0200 0.0181 0.0141 0.0213 0.0182 0.0145 0.0139
fs ratio 0.4796 0.4831 0.5001 0.4453 0.4974 0.4984 0.4592 0.4981 0.4960 0.4597 0.4726 0.4438
mean 0.0509 0.0625 0.0696 0.0754 0.0609 0.0741 0.0553 0.0681 0.0628 0.0616 0.0630 0.0502
Leukemia std 0.0134 0.0153 0.0137 0.0156 0.0114 0.0142 0.0134 0.0113 0.0199 0.0136 0.0131 0.0119
fs ratio 0.4573 0.4696 0.4910 0.4177 0.4945 0.4934 0.4146 0.4913 0.4776 0.4268 0.4307 0.4170
mean 0.0774 0.0834 0.0940 0.1041 0.0908 0.0981 0.0832 0.0892 0.0811 0.0848 0.0832 0.0751
Lung discrete std 0.0095 0.0105 0.0073 0.0092 0.0090 0.0085 0.0110 0.0078 0.0102 0.094 0.0097 0.0083
fs ratio 0.3713 0.4391 0.4862 0.4482 0.4809 0.4855 0.3808 0.4704 0.4378 0.4467 0.4328 0.3922
mean 0.1331 0.1500 0.1628 0.1701 0.1506 0.1684 0.1434 0.1603 0.1436 0.1548 0.1590 0.1345
Colon std 0.0335 0.0350 0.0257 0.0305 0.0274 0.0264 0.0296 0.0335 0.0269 0.0308 0.0289 0.0264
fs ratio 0.4380 0.4630 0.4865 0.4442 0.4909 0.4886 0.4179 0.4884 0.4667 0.4522 0.4425 0.4145

patient health is demonstrated in this section. The COVID-19 case dataset is compiled from the [39]. Fifteen features incorporated
in the mentioned dataset are described in [29]. The purpose is to predetermine the conditions for recovery and death as determined
by the listed factors. The data is split into two distinct sets for training and testing, aiming to validate. Each feature is assigned a
numeric form. Classification error is calculated using the K nearest neighbour (KNN), where 𝑘 is set to 5. KNN is often used in the
literature due to its simplicity and low computational cost. Additionally, it proved to be a very suitable classifier for wrapper-based
methods.
As Table 3 exhibits, QRLAOA-FS exceeded the original AOA and other algorithms by yielding optimal mean fitness value, followed
by HLBDA. BMVO achieved the lowest standard deviation result. The proposed approach attained the smallest feature selection ratio
of 0.1533, outperforming the original AOA (0.1680) The best average classification accuracy is obtained by QRLAOA-FS, followed by
HLBDA and AOA, achieving results of 92.88%, 92.31% and 92.24%, respectively. Collecting extra clinical features can prompt a more
precise accuracy of health prediction in the future.

6
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

Fig. 1. Average error rate convergence graphs and accuracy boxplot diagrams of six algorithms on four datasets.

7
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

Table 3
The result of the mean fitness value, standard deviation of fitness value, feature selection ratio and accuracy of algorithms on the COVID-19 dataset.

HLBDA BDA BABC BMVO BPSO CCSA BCOA CMAES LSHADE FFO AOA QRLAOA-FS

Mean fitness 0.0778 0.0793 0.0804 0.0825 0.0824 0.0839 0.0786 0.0809 0.0812 0.0800 0.0785 0.0720
Standard deviation 0.0082 0.0087 0.0075 0.0073 0.0091 0.0094 0.0118 0.0121 0.0108 0.0089 0.0091 0.0081
Feature selection ratio 0.1627 0.1740 0.2101 0.2427 0.1540 0.2313 0.1593 0.1987 0.1807 0.1753 0.1680 0.1533
Accuracy 0.9231 0.9217 0.9209 0.9191 0.9183 0.9176 0.9222 0.9203 0.9198 0.9210 0.9224 0.9288

Table 4
Shapiro-Wilk test for multi-problem analysis.

HLBDA BDA BABC BMVO BPSO CCSA BCOA CMAES LSHADE FFO AOA QRLAOA-FS
0.015682 0.016325 0.012733 0.025307 0.030842 0.013288 0.029549 0.035672 0.016235 0.036582 0.024853 0.016446

3.4. Validation of improvements of the proposed method

The NFL theorem claims that there is no universal method in the study of metaheuristic optimizations. When balancing between
local intensive exploitation and global search, one improvement is at the expense of the other. The aim is to achieve as much overall
improvement as possible. The suggested QRLAOA-FS was validated against twenty-one problem instances (twenty-one standard
datasets) and compared with eleven other metaheuristic methods, including the original AOA. On average, it demonstrated the best
performance. In addition, each simulation is executed in fifty independent runs due to the stochastic nature of tested methods. In
line with the most recent computer science research, it is essential to determine whether or not generated improvements of one
method against other methods are statistically significant because comparisons based on obtained empirical results are insufficient.
As a consequence, performing additional results analysis through statistical tests is advised [25].
The testing of twelve methods against twenty-one instances belongs to the field of multi-problem, multiple-methods analysis [40].
It is advised to use the average objective function value for each issue as a benchmark when making these comparisons. This method
has limitations if the results obtained from multiple reruns of the same problem don’t come from normal distribution [25,40,41].
Therefore, first, the Shapiro-Wilk test for single-problem analysis [42] is performed in the following manner: for each algorithm and
every problem, a data sample is generated for every single algorithm and each problem by using the results gained in every run and
corresponding 𝑝 − 𝑣𝑎𝑙𝑢𝑒 are determined for each method-problem pair. All obtained 𝑝 − 𝑣𝑎𝑙𝑢𝑒𝑠 are higher than 𝛼 = 0.05. Therefore the
conclusion is that the independent runs generate results from a normal distribution, and in this case, the averages are representative.
Next, the criteria needed for the secure application of parametric tests are then examined, including independence, homoscedas-
ticity of variances in the data, and normality [43]. Every run begins with a unique pseudo-random number seed, completely fulfilling
the criteria of independence. When Levene’s test [44] was used to examine homoscedasticity, it produced a 𝑝 − 𝑣𝑎𝑙𝑢𝑒 above the
𝛼 = 0.05 threshold. As a result, this requirement is met as well.
Consequently, the normality for each method is assessed separately using the Shapiro-Wilk test for multi-problem, multiple-
methods analysis [42]. To accomplish this, an average objective function is used for each instance and each approach. The findings
of the Shapiro-Wilk test, which are displayed in Table 4, demonstrate that the derived 𝑝 − 𝑣𝑎𝑙𝑢𝑒𝑠 are always lower than the threshold
value for a significance level of 0.05. It was decided to proceed forward with non-parametric testing because the assumption that the
data came from a normal distribution is disputed, which leads to the conclusion that the normality condition for justified employment
of parametric tests is not achieved.
Thus, the Friedman aligned test [45,46] is performed, together with the Holm post-hoc procedure in conjunction with the two-
way variance analysis by ranks. These results are shown in Table 5. This strategy is advised for simulations of multi-problem,
multiple-methods comparisons [25].
The proposed QRLAOA-FS metaheuristic outperformed all other approaches, as can be seen from the Friedman aligned test results,
achieving an average rank of 39.5238. The HLBDA, which has an average rank of 46.2619, is the second-best approach, followed
by the BDA, which has an average rank of 106.4286. Original AOA has an average rank of 120.4524. Additionally, the calculated
Friedman 𝑝 − 𝑣𝑎𝑙𝑢𝑒 is 1.11𝐸 − 16, and the Friedman statistics 𝜒𝑟2 is 116.90, which is greater than the 𝜒 2 = 19.68 with 5 degrees of
freedom critical value at 𝛼 = 0.05. From all of these statistical indicators, it can be concluded that there is a statistically significant
difference between the suggested QRLAOA-FS and other approaches and that the 𝐻0 , which maintains that there is no significant
performance difference between methods, can be rejected.
The Iman and Davenport’s test [47] was carried out in accordance with the [48] because it may produce more trustworthy results
than the 𝜒 2 test. The test’s result, 21.37, is higher than the critical value for the 𝐹 -distribution (1.83). The 𝑝 − 𝑣𝑎𝑙𝑢𝑒 for the Iman
and Devenport’s analysis is 1.11𝐸 − 16, which is smaller than 𝛼 = 0.05. Therefore, the conclusion is that the Iman and Davenport’s
analysis is rejecting a null hypothesis as well.
In the end, the non-parametric post-hoc Holm’s step-down procedure is then employed, with significant values 𝛼 = 0.05 and
𝛼 = 0.1. The results are shown in Table 6 (values 0 and 1 indicate whether the null hypothesis is rejected). These results also show
that the QRLAOA-FS clearly outperformed all other metaheuristics at the critical level of 0.1, while for the critical level of 0.05, the
null hypothesis was rejected against all methods except for HLBDA.

8
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

Table 5
Friedman aligned tes results.

Functions HLBDA BDA BABC BMVO BPSO CCSA BCOA CMAES LSHADE FFO AOA QRLAOA-FS

Glass 148.5 148.5 137 151 137 148.5 137 163 157.5 154 157.5 148.5
Hepatitis 39 97 146 202 56 208 197 195 168 174 120 34
Lymphography 32 73 59 229 50 227 181 129 216 65 107 53
Primary Tumor 37 171 35 239 38 215 184 172 212 98 95 27
Soybean 13 75 160 250 126 248 33 41 63 64 74 11
Horse Colic 26 78 165 244 49 246 87 66 175 60 84 23
Ionosphere 25 101 204 236 170 235 40 48 70 188 103 19
Zoo 79 89 71 196 47 201 162 192 203 153 128 77
Musk 1 7 86 220 247 210 233 45 100 46 68 85 22
Arrhythmia 8 132 222 245 213 242 62 164 43 54 104 6
Dermatology 88 110 118 198 106 187 167 92 122 127 189 105
SPECT Heart 44 81 124 228 67 199 194 209 186 82 91 28
Libras Movement 15 139 217 234 206 232 36 58 69 111 156 12
ILPD 108.5 102 143 173 108.5 142 169 193 191 152 135 55
Seeds 177.5 177.5 115.5 130 115.5 115.5 125 140 179 144 133 115.5
LSVT 10 123 112 119 121 225 226 182 238 185 155 3
SCADI 61 131 180 219 166 218 99 80 57 159 145 52
TOX _171 2 21 243 252 230 251 14 249 9 20 24 1
Leukemia 17 134 205 231 93 224 42 200 141 113 161 16
Lung discrete 29 76 207 240 190 223 83 176 51 94 72 18
Colon 5 90 221 241 96 237 31 214 30 183 211 4

Average Ranking 46.2619 106.4286 157.3571 214.4762 123.5714 212.2381 110.1905 145.8571 125.0714 116.5714 120.4524 39.5238
Rank 2 3 10 12 7 11 4 9 8 5 6 1

Table 6
The Holm’s step-down procedure findings.

Comparison p_values Ranking alpha=0.05 alpha=0.1 H1 H2

QRLAOA-FS vs BMVO 0 0 0.004545 0.009091 1 1

QRLAOA-FS vs CCSA 1.80E-14 1 0.005000 0.010000 1 1
QRLAOA-FS vs CMAES 3.48E-08 2 0.005556 0.011111 1 1
QRLAOA-FS vs BABC 3.85E-07 3 0.006250 0.012500 1 1
QRLAOA-FS vs LSHADE 3.85E-07 4 0.007143 0.014286 1 1
QRLAOA-FS vs AOA 1.13E-06 5 0.008333 0.016667 1 1
QRLAOA-FS vs FFO 2.86E-06 6 0.010000 0.020000 1 1
QRLAOA-FS vs BCOA 2.88E-05 7 0.012500 0.025000 1 1
QRLAOA-FS vs BDA 5.87E-05 8 0.016667 0.033333 1 1
QRLAOA-FS vs BPSO 3.61E-04 9 0.025000 0.050000 1 1
QRLAOA-FS vs HLBDA 7.80E-02 10 0.050000 0.100000 0 1

The proposed algorithm uses a greater number of FFEs due to initialization with QRL. The standard number of FFEs is N (number
of solutions). The proposed algorithm uses 2N FFEs. For an unbiased comparison, the number of FFEs is used as a termination
condition.

4. Conclusion

This research paper proposes a quasi-reflection learning AOA - firefly search to tackle feature selection problems - an approach that
is adapted for binary search space and utilizes arithmetic operators for positions update. The proposed and other well-known methods
are subjected to extensive testing on various datasets. The following evaluation parameters are used to compare the testing results:
mean fitness function, average classification accuracy, feature selection ratio, standard deviation and average error rate. Aimed at a
profound analysis, several statistical tests, such as Friedman, Iman-Davenport and Holm’s step-down procedure, are conducted. These
comparative analyses show that the proposed algorithm outperformed other tested algorithms in most cases, significantly upgrading
the original AOA performance. The emphasis should be on QRLAOA-FS’s efficient performance on high-dimensional datasets (TOX
171, Leukemia), successfully selecting small-size feature subsets that optimize the classification accuracy.
Compared to the original AOA, QRLAOA-FS has an initial advantage on both simulations, for unconstrained functions and feature
selection, due to the population diversity provided by the incorporated QRL mechanism. Convergence speed is enhanced by com-
bining this with the solid exploitative capabilities of the FA. These strategies collaborate to avoid the algorithm from falling into the
current local optimum and improve AOA in accelerating overall performance.
On the downside, QRLAOA-FS uses twice as many evaluations per iteration precisely because of the QRL mechanism. There are
a large number of datasets that are unbalanced. This study was done on partially balanced datasets. As part of future research, the
proposed QRLAOA-FS can be tested and improved on imbalanced data sets.

Funding

This project is funded by King Saud University, Riyadh, Saudi Arabia.

9
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

CRediT authorship contribution statement

Nebojsa Bacanin: Conceived and designed the experiments; Wrote the paper. Nebojsa Budimirovic: Conceived and designed the
experiments; Analyzed and interpreted the data; Wrote the paper. Venkatachalam K: Analyzed and interpreted the data; Contributed
reagents, materials, analysis tools or data. Hothefa Shaker Jassim; Miodrag Zivkovic: Performed the experiments; Contributed
reagents, materials, analysis tools or data; Wrote the paper. S.S. Askar; Mohamed Abouhawwash: Performed the experiments;
Contributed reagents, materials, analysis tools or data.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgments

Researchers Supporting Project number (RSP2023R167), King Saud University, Riyadh, Saudi Arabia.

References

[1] T. Bezdan, M. Zivkovic, E. Tuba, I. Strumberger, N. Bacanin, M. Tuba, Glioma brain tumor grade classification from MRI using convolutional neural networks
designed by modified FA, in: International Conference on Intelligent and Fuzzy Systems, Springer, 2020, pp. 955–963.
[2] M. Zivkovic, N. Bacanin, K. Venkatachalam, A. Nayyar, A. Djordjevic, I. Strumberger, F. Al-Turjman, COVID-19 cases prediction by using hybrid machine
learning and beetle antennae search approach, Sustain. Cities Soc. 66 (2021) 102669.
[3] N. Budimirovic, E. Prabhu, M. Antonijevic, M. Zivkovic, N. Bacanin, I. Strumberger, K. Venkatachalam, COVID-19 severity prediction using enhanced whale
with salp swarm feature classification, Comput. Mater. Continua 72 (1) (2022) 1685–1698, https://doi.org/10.32604/cmc.2022.023418.
[4] N. Bacanin, N. Budimirovic, K. Venkatachalam, I. Strumberger, A.F. Alrasheedi, M. Abouhawwash, Novel chaotic oppositional fruit fly optimization algorithm
for feature selection applied on COVID 19 patients’ health prediction, PLoS ONE 17 (10) (2022) e0275727, https://doi.org/10.1371/journal.pone.0275727.
[5] B. Cao, W. Zhang, X. Wang, J. Zhao, Y. Gu, . Y. Zhang, A memetic algorithm based on two Arch2 for multi-depot heterogeneous-vehicle capacitated arc routing
problem, Swarm Evol. Comput. 63 (2021) 100864, https://doi.org/10.1016/j.swevo.2021.100864.
[6] R. Li, X. Wu, H. Tian, N. Yu, C. Wang, Hybrid memetic pretrained factor analysis-based deep belief networks for transient electromagnetic inversion, IEEE Trans.
Geosci. Remote Sens. 60 (2022) 5920120, https://doi.org/10.1109/TGRS.2022.3208465.
[7] B. Cao, Y. Gu, Z. Lv, S. Yang, J. Zhao, . Y. Li, RFID reader anticollision based on distributed parallel particle swarm optimization, IEEE Int. Things J. 8 (5) (2021)
3099–3107, https://doi.org/10.1109/JIOT.2020.3033473.
[8] N. Bacanin, T. Bezdan, E. Tuba, I. Strumberger, M. Tuba, Monarch butterfly optimization based convolutional neural network design, Mathematics 8 (6) (2020)
936.
[9] J. Tian, M. Hou, H. Bian, J. Li, Variable surrogate model-based particle swarm optimization for high-dimensional expensive problems, Complex Intell. Syst.
(2022), https://doi.org/10.1007/s40747-022-00910-7.
[10] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE Trans. Evol. Comput. 1 (1997) 67–82.
[11] L. Abualigah, A. Diabat, S. Mirjalili, M.A. Elaziz, A.H. Gandomi, The arithmetic optimization algorithm, Comput. Methods Appl. Mech. Eng. 376 (2021) 113609,
https://doi.org/10.1016/j.cma.2020.113609.
[12] Q. Fan, Z. Chen, Z. Xia, A novel quasi-reflected Harris hawks optimization algorithm for global optimization problems, Soft Comput. (2020) 1–19.
[13] X. Yang, Firefly Algorithms for Multimodal Optimization. Stochastic Algorithms: Foundations and Applications, Springer Berlin Heidelberg, 2009, pp. 168–178.
[14] S. Mirjalili, A.H. Gandomi, S.Z. Mirjalili, S. Saremi, H. Faris, S.M. Mirjalili, Salp swarm algorithm: a bio-inspired optimizer for engineering design problems,
Adv. Eng. Softw. 114 (2017) 163–191.
[15] J. Kennedy, R. Eberhart, Particle swarm optimization, Proc. IEEE Int. Conf. Neural Netw. 4 (2002) 1942–1948.
[16] D. Simon, Biogeography-based optimization, IEEE Trans. Evol. Comput. 12 (2008) 702–713.
[17] G.G. Wang, S. Deb, X.Z. Gao, L.D.S. Coelho, A new metaheuristic optimisation algorithm motivated by elephant herding behaviour, Int. J. Bio-Inspir. Comput. 8
(2016) 394–409.
[18] H. Muthusamy, S. Ravindran, S. Yaacob, K. Polat, An improved elephant herding optimization using sine–cosine mechanism and opposition based learning for
global optimization problems, Expert Syst. Appl. 172 (2021) 114607.
[19] W.T. Pan, A new evolutionary computation approach: fruit fly optimization algorithm, in: Conference of Digital Technology and Innovation Management, 2011,
pp. 382–391.
[20] S. Mirjalili, A. Lewis, The whale optimization algorithm, Adv. Eng. Softw. 95 (2016) 51–67.
[21] S. Mirjalili, Moth-flame optimization algorithm: a novel nature-inspired heuristic paradigm, Knowl.-Based Syst. 89 (2015) 228–249.
[22] S. Mirjalili, SCA: a sine cosine algorithm for solving optimization problems, Knowl.-Based Syst. 96 (2016) 120–133.
[23] S. Saremi, S. Mirjalili, A. Lewis, Grasshopper optimization algorithm: theory and application, Adv. Eng. Softw. 105 (2017) 30–47.
[24] K. Price, N. Awad, M. Ali, P. Suganthan, Problem definitions and evaluation criteria for the 100-digit challenge special session and competition on single
objective numerical optimization, Technical Report, Nanyang Technological University, 2018.
[25] J. Derrac, S. García, D. Molina, F. Herrera, A practical tutorial on the use of nonparametric statistical tests as a methodology for comparing evolutionary and
swarm intelligence algorithms, Swarm Evol. Comput. 1 (1) (2011) 3–18, https://doi.org/10.1016/j.swevo.2011.02.002.
[26] J. Kennedy, R.C. Eberhart, A discrete binary version of the particle swarm algorithm, in: Computational Cybernetics and Simulation IEEE International Conference
on Systems, Man, and Cybernetics, vol. 5, 1997, pp. 4104–4108.
[27] R.C. Thom de Souza, C.A. de Macedo, L. dos Santos Coelho, et al., Binary coyote optimization algorithm for feature selection, Pattern Recognit. 107 (2020)
107470, https://doi.org/10.1016/j.patcog.2020.107470.
[28] S. Mirjalili, Dragonfly algorithm: a new meta-heuristic optimization technique for solving single-objective, discrete, and multi-objective problems, Neural
Comput. Appl. 27 (2016) 1053–1073, https://doi.org/10.1007/s00521-015-1920-1.

10
N. Bacanin, N. Budimirovic, K. Venkatachalam et al. Heliyon 9 (2023) e15378

[29] T. Jingwei, S. Mirjalili, Hyper learning binary dragonfly algorithm for feature selection: a COVID-19 case study, Knowl.-Based Syst. 215 (2021) 106553, https://
doi.org/10.1016/j.knosys.2020.106553.
[30] G.I. Sayed, A.E. Hassanien, A.T. Azar, Feature selection via a novel chaotic crow search algorithm, Neural Comput. Appl. (2017) 1–18.
[31] N. Al-Madi, H. Faris, S. Mirjalili, Binary multi-verse optimization algorithm for global optimization and discrete problems, Int. J. Mach. Learn. Cybern. (2019).
[32] Y. He, H. Xie, T.L. Wong, X. Wang, A novel binary artificial bee colony algorithm for the set-union knapsack problem, Future Gener. Comput. Syst. 78 (2018)
77–86, https://doi.org/10.1016/j.future.2017.05.044.
[33] R. Tanabe, A.S. Fukunaga, Improving the search performance of SHADE using linear population size reduction, in: 2014 IEEE Congress on Evolutionary Compu-
tation, CEC, 2014, pp. 1658–1665.
[34] N. Hansen, S. Kern, Evaluating the CMA evolution strategy on multimodal test functions, in: Parallel Problem Solving from Nature - PPSN VIII, Springer, Berlin,
Heidelberg, 2004, pp. 282–291.
[35] UCI Machine Learning Repository, https://archive.ics.uci.edu/ml/index.php. (Accessed 14 April 2019).
[36] Datasets - Feature Selection @ ASU, http://featureselection.asu.edu/datasets.php. (Accessed 9 November 2019).
[37] S. Mirjalili, A. Lewis, S-shaped versus V-shaped transfer functions for binary particle swarm optimization, Swarm Evol. Comput. 9 (2013) 1–14, https://
doi.org/10.1016/j.swevo.2012.09.002.
[38] A.T. Sahlol, D. Yousri, A.A. Ewees, M.A.A. AlQaness, Ro. Damasevicius, M. Abd Elaziz, COVID-19 image classification using deep features and fractional-order
marine predators algorithm, Sci. Rep. (2020).
[39] C. Iwendi, A.K. Bashir, A. Peshkar, et al., COVID-19 patient health prediction using boosted random forest algorithm, Front. Public Health 8 (2020), https://
doi.org/10.3389/fpubh.2020.00357.
[40] T. Eftimov, P. Korošec, B.K. Seljak, Disadvantages of statistical comparison of stochastic optimization algorithms, in: Proceedings of the Bioinspired Optimization
Methods and their Applications, BIOMA 2016, pp. 105–118.
[41] S. García, D. Molina, M. Lozano, F. Herrera, A study on the use of non-parametric tests for analyzing the evolutionary algorithms’ behaviour: a case study on the
CEC’2005 special session on real parameter optimization, J. Heuristics 15 (2009) 617–644.
[42] S.S. Shapiro, R. Francia, An approximate analysis of variance test for normality, J. Am. Stat. Assoc. 67 (1972) 215–216.
[43] A. LaTorre, D. Molina, E. Osaba, J. Poyatos, J. Del Ser, F. Herrera, A prescription of methodological guidelines for comparing bio-inspired optimization algo-
rithms, Swarm Evol. Comput. 67 (2021) 100973.
[44] G.V. Glass, Testing homogeneity of variances, Am. Educ. Res. J. 3 (1966) 187–190.
[45] M. Friedman, The use of ranks to avoid the assumption of normality implicit in the analysis of variance, J. Am. Stat. Assoc. 32 (1937) 675–701.
[46] M. Friedman, A comparison of alternative tests of significance for the problem of m rankings, Ann. Math. Stat. 11 (1940) 86–92.
[47] R.L. Iman, J.M. Davenport, Approximations of the critical region of the fbietkan statistic, Commun. Stat., Theory Methods 9 (1980) 571–595.
[48] D.J. Sheskin, Handbook of Parametric and Nonparametric Statistical Procedures, Chapman and Hall/CRC, 2020.

11