Egyptian Informatics Journal xxx (xxxx) xxx

Contents lists available at ScienceDirect

Egyptian Informatics Journal

journal homepage: www.sciencedirect.com

A new evolutionary algorithm: Learner performance based behavior

algorithm
Chnoor M. Rahman a,b,⇑, Tarik A. Rashid c
a
Applied Computer Department, College of Medicals and Applied Sciences, Charmo University, Sulaimany, Iraq
b
Technical College of Informatics, Sulaimany Polytechnic University, Sulaimany, Iraq
c
Computer Science and Engineering Department, University of Kurdistan Hewler, Erbil, Iraq

a r t i c l e i n f o a b s t r a c t

Article history: A novel evolutionary algorithm called learner performance based behavior algorithm (LPB) is proposed in
Received 1 October 2019 this article. The basic inspiration of LPB originates from the process of accepting graduated learners from
Revised 10 January 2020 high school in different departments at university. In addition, the changes those learners should do in
Accepted 12 August 2020
their studying behaviors to improve their study level at university. The most important stages of opti-
Available online xxxx
mization; exploitation and exploration are outlined by designing the process of accepting graduated
learners from high school to university and the procedure of improving the learner’s studying behavior
Keywords:
at university to improve the level of their study, respectively. To show the accuracy of the proposed algo-
Evolutionary algorithms
Genetic algorithm
rithm, it is evaluated against a number of test functions, such as traditional benchmark functions, CEC-
LPB C06 2019 test functions, and a real-world case study problem. The results of the proposed algorithm
Learner performance based behavior are then compared to the DA, GA, and PSO. The proposed algorithm produced superior results in most
algorithm of the cases and comparative in some others. It is proved that the algorithm has a great ability to deal
Optimization with the large optimization problems comparing to the DA, GA, and PSO. The overall results proved
Metaheuristic optimization algorithm the ability of LPB in improving the initial population and converging towards the global optima.
Moreover, the results of the proposed work are proved statistically.
Ó 2020 Production and hosting by Elsevier B.V. on behalf of Faculty of Computers and Information, Cairo
University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction sion making, either in economic or engineering fields. Decision-

making tasks involve making the best decision to choose between
The computational intelligence (CI) term as a branch of artificial different alternatives. Numerous optimization algorithms exist;
intelligence (AI) was first invented by Bezdek in the early 1990s however, no single algorithm fits all the different problems. It is
[1], which motivated a new field in computer-based intelligence. crucial for the appropriate optimizer to guarantee that the optimal
CI in principle consists of any technologies and science- solution is always reachable. NP-hard problems, for example, are
supported approaches for creating, analyzing, and developing usually not easy to be solved. However, most combinatorial opti-
intelligent systems [2]. It mainly depends on a set of nature- mization problems, for example, N-Queens, traveling salesperson,
inspired computational patterns and a numerical collection of data and 0/1 Knapsack are NP-hard. To solve this type of problem and
[3]. The study of optimization techniques is one of the main sub- depending on the size of the problem, two approaches exist,
jects of CI. Optimization is part of any problem that requires deci- namely; exact methods and metaheuristic methods [4]. Exact
methods are useful when the number of decision variables is small.
⇑ Corresponding author. These methods find the optimal solution for the problem. Exam-
E-mail addresses: [email protected] (C.M. Rahman), tarik. ples for exact methods are branch and bound algorithm [5],
[email protected] (T.A. Rashid). dynamic and linear programming, and so on. The problem with
Peer review under responsibility of Faculty of Computers and Information, Cairo these methods is that they are known as time expensive methods,
University. so that it is not recommended to use them for solving difficult or
NP-hard problems. Likewise, where the decision space is discrete
or when a large number of decision variables exist, which occurs
in most if not in all practical problems of optimization, exact meth-
Production and hosting by Elsevier

https://doi.org/10.1016/j.eij.2020.08.003
1110-8665/Ó 2020 Production and hosting by Elsevier B.V. on behalf of Faculty of Computers and Information, Cairo University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
2 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx

ods cannot show good performance, instead, metaheuristics can be cation behavior of micro bats [17]. In 2014, Mirjalili on the base
used [4]. of hunting behavior, and social hierarchy of grey wolf proposed a
Depending on the characteristics, metaheuristic optimization new optimization algorithm named as grey wolf optimization
algorithms can be classified in various ways. They can be classified (GWO) algorithm [18]. In 2015, the same author proposed the
into population-based algorithms and trajectory-based or single- dragonfly algorithm (DA). DA was mainly inspired by the hunting
point search algorithms. In the latter case, the algorithm uses a sin- and migrating behaviors of a dragonfly. The latter is called a
gle solution, which means in each iteration only a single solution dynamic (migratory) swarm, and the former is called static (feed-
will manipulate. Hill climbing, tabu search, and simulated anneal- ing) swarm [19]. Finally, in 2019 fitness dependent optimizer
ing are examples of this class of algorithms. On the other hand, (FDO) developed. It is inspired by the bee swarming reproductive
population-based algorithms use a population of agents and the process. FDO mimics the PSO in utilizing velocity to update search
whole population is modified in each iteration. Examples for agent’spositions. However, FDO uses the fitness function of the
population-based algorithms are genetic algorithm, particle swarm problem to produce weights, and these weights are then used to
optimization, ant colony optimization, and so on [4]. guide the agents in the exploration and exploitation phases [20].
Since introducing these algorithms for optimization, many
1.1. Related works researchers utilized them to optimize problems in various fields.
However, some other researchers aimed at improving those algo-
During the 1960s and 1970s, the metaheuristic optimization rithms. The satisfactory results produced by these algorithms for
algorithms were bloomed. At the beginning of the 1960s, the different optimization problems proved the importance and neces-
genetic algorithm (GA) [6] was developed by John Holland and sity of them [21–26]. Consequently, researchers continue to pro-
his collaborators. GA is a search technique; it is based on Darwin’s pose new algorithms in the field. Many of these algorithms do
theory of evolution and selection of biological systems. The ability not have a good balance between exploitation and exploration.
of the GA for optimization makes the researchers use it in optimiz- Having high exploitation traps the algorithm in the local optimum.
ing a wide range of problems. Since then, it has been modified and Moreover, a high degree of exploration raises the probability of
hybridized with other techniques to solve various problems. In [7] finding global optima but decreases the efficiency. Therefore, hav-
GA was combined with an active set technique (AST). The hybrid ing a good balance between exploration and exploitation can make
technique was used for optimizing the unsupervised artificial neu- an algorithm perform better compared to the other algorithms
ral network. The aim of this work was to accurately estimate the [27].
temperature profiles of the heat conduction model in the head of
humans. The results revealed that the hybrid technique produced
better and accurate results comparing to standalone approaches, 1.2. Innovative contribution
such as GA and AST. Additionally, in [8], GA combined with an inte-
rior point technique to optimize a new approach. The approach In this paper, a new optimization method, learner performance
was solving the initial value of the equation of a Painlev́e II, and behavior based algorithm is proposed. The LPB method mimics the
its variants, utilizing the feed-forward artificial network. Moreover, process of accepting graduated learners from high school in differ-
in [9], GA combined with IPT to optimize a feed-forward artificial ent colleges and the behaviors of learners that affect their perfor-
neural network for solving porous fin equation. Better accuracy mance during the college study, and the factors that may help
achieved comparing to other numerical techniques. Similarly, ref- the learners to change their high-school study behaviors that are
erence [10] designed a neuro-heuristic schema for non-linear sec- not effective anymore for studying in the college. To implement
ond order Thomas-Fermi system. To optimize the schema, GA and this, multi-populations can be utilized to demonstrate the learners
sequential quadratic programming was utilized. It was discovered that have a GPA in different ranges. Consequently, this causes a
that the examined schema was feasible, precise, and effective. Hol- good balance between exploration and exploitation [27]. The most
land’s work encouraged many to adopt and develop identical tech- important features of the proposed work are:
niques in their research works. Later, in 1966, Fogel et al.
developed an evolutionary programming technique [11]. In this
It is a population based algorithm.
work, finite state machines were used to represent the solution,
The initial population is created randomly.
and stochastically one of the machines was mutated. Afterward,
A percentage of the population is separated.
in 1983, Kirkpatrick et al. developed simulated annealing (SA)
The population is divided into a number of sub-populations.
[12]. SA mimics the process of annealing that utilized for crystal-
The highest fitness in the separated group is used to divide the
lization, which is a physical process in metals and glasses to harden population into sub-populations.
the material. Furthermore, at the beginning of the 1990s, Marco
The sub-populations that contain the best individuals have pri-
Dorigo completed his Ph.D. thesis on optimization and nature- ority to go through the optimization process first.
inspired algorithms. In his thesis, he examined a novel idea known
Mutation and crossover operators are used to make changes in
as an ant colony optimization algorithm (ACO) [13]. ACO was the structure of new individuals.
inspired by the swarming behavior of social ants utilizing the pher-
omone to find the source of food and bring the food back to their
nest. Later, in 1995, James Kennedy proposed particle swarm opti- 1.3. Organization
mization (PSO) [14]. PSO can be counted as another significant
improvement in the field. It mimics the behaviors of the school The rest of the paper is organized as follows: Section 2 shows
of birds or fish. A particle represents a single solution that has a the inspiration of the proposed algorithm. Section 3 presents the
position in the search space. In 2005, Karaboga introduced an arti- features (operators) of GA that are utilized in the proposed tech-
ficial bee colony (ABC) [15]. ABC mimics the behaviors of honey- nique. The LPB operators and techniques along with the pseudo
bees. It provides well-balanced exploitation and exploration code are presented in section 4. Furthermore, the results of the
ability. Thereafter, in 2007, Chu and Tsai proposed a new swarm- algorithm and an inclusive and comparative study on some bench-
based optimization algorithm named cat swarm optimization mark test functions along with a real-world problem are presented
(CSO) algorithm. CSO mimics the behaviors of cats [16]. Yang in in section 5. Finally, the conclusion of the work and directions for
2010 introduced the bat algorithm, which is based on the echolo- future researches are shown in section 6.

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx 3

2. Inspiration level of learning. Most fresh learners at colleges have learned to

study skills in high school that are no longer effective. They might
Every year groups (population) of learners finish their high have a proper sense of metacognition, which accurately informed
school and apply to the universities. The applications for some of them when they studied sufficiently during high school, but it is
these learners are accepted and the rest are rejected. Depending not accurate anymore. This means that entering college requires
on their GPA, the learners are divided into different groups. The overcoming the old study strategies with new ones [29,32].
process of transferring learners from high school to university Besides, having an adequate level of metacognition can cause a
starts with a group of graduated learners M from high school. good improvement in the learner’s level of study and it may have
Departments from universities specify the number of learners that an effect on all the strategies used by the learner. The main inspi-
they want to accept to study in their department. Furthermore, ration of this algorithm originates from the following steps that a
each department specifies the minimum GPA that the learners learner goes through:
should have in order to study in that department. This is like
grouping the graduated learners from high school M to a number 1) The strategies used to group the learners according to their
of different departments (groups) according to their GPA. The GPA, and almost all the learners that are accepted in a
department accepts the applications of the learners if the learner’s department have a GPA in a specific range.
GPA is in the range of the required GPAs by that department. 2) After accepting the learners in the departments, their study-
Among the learners who apply to a specific department, there ing behaviors can be improved to make them good college
are a number of learners, which their GPA is under the required learners. The learner’s behaviors influenced by each other
GPA. The application for those learners will be rejected. Further- while they study together.
more, there are learners that their GPA is much higher than the 3) The level of metacognition for learners has a big impact on
required GPA, thus, these applications have a priority and they will all the studying behaviors.
be accepted first, and then the lower ranges, and so on, until the
number of accepted applications is equal to the number of learners In this algorithm, the first step is used to choose individuals
specified by the department. Furthermore, sometimes it happens from the population. The importance of this step is dividing the
that in general, the GPA of learners is low. Thus, some of the main population to some sub-populations and then the individuals
departments cannot have a specified number of learners with the will be selected from the sub-populations depending on their fit-
required GPA. At these situations before finalizing the list of ness. This prevents converge to local optima because selecting
accepted learners, the department, and the university should individuals will start from the perfect sub-population. The latter
decide whether they want to accept learners with less GPA or not. two steps are used to improve the individuals by letting the learn-
After accepting the graduated learners from high school in dif- ers work in groups and ask for help from each other. Furthermore,
ferent colleges and departments, the learners go through a number having a good level of metacognition will influence the overall
of difficulties. Because the environment they came from was differ- studying behaviours of a learner in a stochastic way (mutation).
ent from the environment they are in now. In addition, the study- On the other hand, learners affect the studying behaviours of each
ing behaviors that they had as high school learners may not be other when they study together (crossover).
effective anymore. It is normal that many fresh learners are not
prepared either academically nor in terms of study skills for
3. Genetic algorithm operators
college-level study. Working on the learner’s studying behaviors,
such as seeking help and group working will help them to study
The genetic operators imitate the procedure of the heredity of
more effectively and will result in improving their score during
genes to produce new individuals at each generation. The opera-
their study in the college [28,29]. Studying behaviors of a learner
tors are utilized to make changes in the structure of individuals
can be affected by studying behaviors of learners in the same
during the representation. The common genetic operators are
department or learners in any other department.
crossover, mutation, and selection. Here, we only define crossover
The level of learning of transition learners from high school to
and mutation operators.
college can be improved by adopting some effective strategies,
which are quite different from those in high school. A number of
behaviors have been considered to judge between the strong and 3.1. Crossover
weak learners, they include (level of interest, deep processing,
effective note taking, problem-solving, group working, seeking Crossover is the most fundamental genetic operator. It works on
help, and self-study). Additionally, according to [30] the learners two individuals at the same time and produces offspring by inte-
with a high level of creativity are always strong learners. Depend- grating features of both individuals. Various crossover techniques
ing on the previously mentioned resources, it can be concluded are available; however, the most used one is choosing a stochastic
that learners who have a good level of the aforementioned behav- cut-point to produce the offspring by integrating the part of one
iors are good learners. parent to the right of the cut-point with the part of the second par-
Moreover, it has been noticed that the quality of metacognition ent to the left of the cut-point. For example, one-cut point cross-
is another key-difference between strong and weak learners. over, two-cut point crossover, multi-cut point crossover, etc. [33].
Metacognition refers to the learner’s awareness level of under-
standing of a topic. Those who have poor metacognition are confi- 3.2. Mutation
dent and believe that they have done well on exams while it is not
and their low score shocks them. When learners get low marks on Mutation creates random changes in different individuals. The
exams, they often believe that they should spend more time study- simplest form of mutation is altering one or more genes. Mutation
ing subjects. In addition to studying more (although that often in the genetic algorithm has a great role of either a) restoring the
helps), however, learners with poor metacognition should change lost genes during the selection process, hence, they can be used
the way they study [31]. Learners with poor metacognition levels in another context or b) serving the genes that were not available
usually have poor study strategies, which rise false confidence that in the initial population. Various ways of mutation are available for
they have studied the material well without increasing their actual different representations of individuals. For example, uniform

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
4 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx

mutation, replacement mutation, dynamic mutation, boundary Table 1

mutation, and so on [33]. Parameter settings for lpb.

Parameters Parameter Value

4. Learner performance based behavior algorithm Crossover rate 2*round (0.7*population size)
Mutation rate round (0.2*population size)
As the first step in the algorithm, randomly create a population Population Size 80
dp 0.5
of graduated learners M who want to apply for different depart-
ments in different universities. Furthermore, we have an operator
and we call it division probability dp. As discussed, every depart-
ment accepts learners that have a GPA greater than or equal to dp: the percentage of individuals chosen from M
the minimum required GPA. To show this in the algorithm, at first, O: the sub-population chosen from M according to the dp
we use the dp parameter to randomly choose a percentage of ele- operator.
ments from M. Afterwards, we calculate the fitness of each of the BP: bad population
chosen individuals and sort them. Then we divide them into two GP: good population
groups, good and bad, depending on their fitness. The former con- PF: perfect population
tains the individuals that have a higher GPA and the bad group k: is a counter utilized to count the number of newly created
contains the rest. After this, the fitness of the individuals in the individuals
main population M is calculated and then filtered. Those individu-
als that have fitness smaller or equal to the highest fitness (best fit- 5. Results and discussion
ness) in the bad population will be moved to the bad population.
The rest of the individuals will be divided into two groups. Those In this section, a number of standard benchmark functions in
who have fitness smaller or equal to the highest fitness (best fit- the literature are used to examine LPB algorithm. The results are
ness) in the good population will be moved to a good population, then evaluated against three popular algorithms in the literature:
and those who have fitness higher than the highest fitness in the DA, PSO, and GA. The results for 19 classical benchmark functions
good population will be moved to the perfect population. Then for PSO, DA, and GA are taken from [19]. Nevertheless, we exam-
the number of learners specified by the department will be chosen ined the CEC-C06 2019 test functions to show the ability of the
from the perfect population and good population. If the number of algorithm in solving large scale optimization problems [37]. Addi-
individuals in these two populations was smaller than the number tionally, the processing time (PT) in seconds of the algorithm for
of specified learners by the department this is when the number of both groups of the test functions is calculated to show the ability
learners that have got the required GPA is small and the depart- of the algorithm compared to the others in quickly finding the opti-
ment should decide whether they should accept other learners mal results. Furthermore, to prove the significance of the results,
with less GPA or not. If they decided to accept other learners, the the Wilcoxon rank-sum test [35] is used. Then the algorithm is
rest of the individuals will come from the bad population. used to optimize a real-world problem. The parameter settings
After accepting the graduated learners from high school in the for LPB algorithm are shown in Table 1.
departments, as discussed, they may not have effective studying
behaviours [29,32]. However, improving behaviours like help- 5.1. Classical benchmark test functions
seeking, group working can have a positive impact on them. In
addition, as mentioned in [28,29] learners can influence each To test the performance of the LPB algorithm a group of bench-
other’s behaviour. For example, when they work in groups or when mark functions is used. These benchmark functions are divided
they ask help from each other their studying behaviours will be into three groups: unimodal, multi-modal, and composite test
affected. To show this in the algorithm crossover operator from functions [36–39]. Each group has different properties. Unimodal
genetic algorithm is used. Utilizing a crossover operator will let test functions, for example, benchmark the convergence and the
the individuals exchange some studying behaviours. Consequently, exploitation of the algorithm. This group of test functions has a sin-
the learner has a set of studying behaviours, which is different gle optimum. However, multi-modal test functions, as their name
from the original studying behaviours owned by the learners. implies, have multi optimum. They have one global optimum and
Hence, the overall, behaviours of both individuals will be affected multi-local optima. To approach the global optimum an algorithm
and the produced individuals have different behaviours. should avoid the entire local optimal solutions. Hence, this group
In addition, the level of metacognition has an impact on the of test functions can benchmark exploration and avoid local
overall studying behaviours of a learner. Whenever the metacogni- optima.
tion level of a learner is affected, stochastically, the overall study-
ing behaviours of the learner will be affected too [31,34]. The level
of metacognition according to [34] is affected by training the lear- 1. [Initialization]
ner using a number of strategies. Using these strategies is excluded Randomly create a population M
from this work. Consequently, the level of metacognition of learn- 2. [Specify parameters]
ers can be affected using a rate that can be specified in the algo- Specify the number of required learners N for a department,
rithm. As mentioned the metacognition level may affect the crossover rate and mutation rate
overall behaviours of the learners in a stochastic way. So that, ran- 3. [Create Sub-Populations]
domly changing positions of the behaviours of that individual Use dp parameter to randomly choose a percentage of
according to a specific rate or randomly updating the values of individuals O from M
studying behaviours of that learner can do this. This is presented Evaluate the fitness of individuals in O
in the algorithm by using the mutation operator from the genetic Depending on their fitness, sort the individuals in O
algorithm. Visual 1 shows the pseudo code for LPB algorithm. (descending order), use one of the sorting methods
Definition of symbols: Divide O to two populations, good (individuals with high
M: the initial random population fitness) and bad (individuals with low fitness)
N: the number of individuals in the new population

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx 5

a (continued) ing to the GA are evident that LPB algorithm has a greater exploita-
tion rate and convergence speed. In addition, LPB algorithm
1. [Initialization]
outperforms both PSO and DA in the last unimodal test function
While termination condition is not met (TF7) and PSO in TF5 as well. Nevertheless, the LPB algorithm pro-
Use dp parameter to randomly choose a percentage of vides better results than the other algorithms in all the other test
individuals O from M functions. PSO, however, provided a better result in TF12. These
Evaluate the fitness of individuals in O results show the ability of the proposed algorithm in avoiding local
Depending on their fitness, sort the individuals in O optima, exploring the search space, and balancing exploration and
(descending order), use one of the sorting methods exploitation. Results of the test functions TF7-TF19 proved that
Divide O to two populations, good (individuals with high LPB has a superior exploration, and a perfect ability in avoiding local
fitness) and bad (individuals with low fitness) optima, and also it has a superior balance between exploration and
Find fitness for all individuals in the population M exploitation phases comparing to the DA, PSO, and GA. As shown in
Find the highest fitness in good and bad populations Table 2, it can be concluded that the LPB algorithm has the first rank
if an individual from M has fitness <= highest fitness in the among the other algorithms because it outperformed the other
bad population algorithms in 12 functions out of 19 functions. Fig. 1 shows the con-
Move it to the bad population BP vergence curve for the proposed algorithm. In Fig. 1, for each group
else if an individual from M has fitness <= highest fitness in of the test functions, one function is selected (F2 for unimodal, F9
the good population for multi-modal, and F17 for composite test functions), and cost
Move it to the good population GP refers to the fitness value for the global best.
Else For the traditional benchmark functions, the PT of the LPB is
Move it to the perfect population PF much smaller comparing to the DA. The reason for this is that in
end if the first stage of the LPB, a subset of the population is chosen based
while k <= N on this smaller group other subpopulations are built. The perfect
if PF is not empty subpopulation has priority to be optimized first, then the good sub-
Select an individual from PF population and so on. Since the subpopulations are much smaller
else if GP is not empty compared to the main population, searching for the solutions in
Select an individual from GP these subpopulations is speeder. This improves the randomness
Else and saves the optimization time simultaneously. However, com-
Select an individual from BP pared to the PSO and GA, the PT of the LPB is higher.
end if
k = k + 1; 5.2. CEC-C06 2019 benchmark test functions
end while
4. Crossover Many real-world problems exist in which time is not as impor-
5. Mutation tant as getting an accurate answer. In addition, practically people
6. [Termination] tune an algorithm and execute it more than one trail if they
Repeat the procedure from step 3 until termination condition wanted. This means users try to find the most successful algorithm
is met. for their scenario regardless of time. It is this feature of numerical
end while optimization, which the CEC-C06 benchmark test functions also
7. [Optimal Solution] known as ‘‘The 100-digit challenge” examine. They calculate the
Select the best solution from the perfect population values of functions at ‘‘horizontal” slices of the convergence plot
VISUAL 1: PSEUDO CODE FOR LPB ALGORITHM [39]. These test functions are considered for use in an annual com-
petition of optimization. They are used to evaluate the algorithm
for large scale optimization problems. The first three functions,
Finally, the composite test functions are mostly combined, biased, CEC01 to CEC03, have various dimensions as shown in Appendix
rotated, and shifted versions of the aforementioned groups [39]. B Table 9. On the other hand, the CEC04 to CEC10 functions set
They demonstrate the difficulties exist in the real search spaces as 10-dimensional minimization problems in the range [100,
by providing a huge number of local optima and diverse shapes 100], and they are shifted and rotated. All the CEC functions are
for various regions. This type of benchmark functions can bench- scalable and all global optimum of these functions were united
mark the combined exploitation and exploration of an algorithm. towards point 1. The results of the CEC-C06 2019 test functions
See Appendix A, Tables 6, 7, and 8 for more information about the for the LPB, DA, and PSO are shown in Table 3. For each test func-
test functions and their conditions [19]. Ultimately, for each algo- tion in Table 3, superior results are shown in bold. The test func-
rithm in Table 2, the test functions are solved 30 times, 80 search tions are solved 30 times utilizing 80 search agents over 500
agents are utilized over 500 iterations. The average and standard iterations. The average, standard deviation, and processing time
deviation are then calculated. Parameters for GA, PSO, and DA are are then calculated. The results of the CEC-C06 2019 benchmark
discussed in reference [19]. For all test functions in Table 1, dp is functions for DA and PSO are taken from [40]. As shown in Table 3,
set to 0.5. The average and standard deviation of the optimal solu- the value of metrics, average, and standard deviation for the LPB
tion is calculated in the last iteration. These two metrics are used to algorithm in almost all the CEC-C06 2019 test functions are smaller
evaluate the overall performance of the algorithms, and to show the than DA, and PSO. However, PSO showed its superiority in CEC04.
stability degree of the algorithms to solve the test functions. Additionally, the results of the LPB and PSO for optimizing CEC05,
For each test function in Table 2, superior results are shown in and CEC09 are comparative. The results of the CEC-C06 2019
bold. As shown in Table 2, for the first six unimodal test functions benchmark functions revealed that for large scale optimization
(TF1-TF6), the DA algorithm outperforms the LPB algorithm, and problems LPB provides better results compared to the DA, and PSO.
also PSO performs better in the (TF1-TF4, and TF6). This proves that The processing time for the LPB and DA for the CEC-C06 2019 is
the exploitation and the convergence speed of the algorithm are not also shown in Table 3. As clear, the PT for the LPB for optimizing all
better than the algorithms used in the comparison. However, the the functions is much smaller. The reason for this, as mentioned
results of the unimodal test functions of the LPB algorithm compar- earlier, is that in the first stage of the LPB, a subset of the popula-

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
6 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx

Table 2
Comparison of results of the classical benchmark function between LPB, DA, PSO, and GA.

Test Function LPB DA PSO GA

TF1 Ave. 0.001877545 2.85E-18 4.2E-18 748.5972
Std. 0.002093616 7.16E-18 1.31E-18 324.9262
PT (Seconds) 160.840946 1445.243327 249.665030 65.422226
TF2 Ave. 0.005238111 1.49E-05 0.003154 5.971358
Std. 0.003652512 3.76E-05 0.009811 1.533102
PT (Seconds) 169.076368 1259.496468 3.826913 55.040008
TF3 Ave. 36.4748883 1.29E-06 0.001891 1949.003
Std. 29.22415523 2.1E-06 0.003311 994.2733
PT (Seconds) 202.408611 1216.762524 12.702411 80.126424
TF4 Ave. 0.393866 0.000988 0.001748 21.16304
Std. 0.135818 0.002776 0.002515 2.605406
PT (Seconds) 191.301934 1399.014810 2.877756 63.099468
TF5 Ave. 16.76919 7.600558 63.45331 133307.1
Std. 22.19251 6.786473 80.12726 85007.62
PT (Seconds) 130.846636 1707.285731 5.224432 55.818782
TF6 Ave. 0.00203173 4.17E-16 4.36E-17 563.8889
Std. 0.0027832 1.32E-15 1.38E-16 229.6997
PT (Seconds) 157.547318 1550.130722 2.795879 51.284046
TF7 Ave. 0.004975 0.010293 0.005973 0.166872
Std. 0.002965 0.004691 0.003583 0.072571
PT (Seconds) 158.642028 1593.877054 8.982616 56.555067
TF8 Ave. 3747.65 2857.58 7.1E + 11 3407.25
Std. 189.0206 383.6466 1.2E + 12 164.4776
PT (Seconds) 162.354305 1738.794894 8.266467 55.234252
TF9 Ave. 0.001567 16.01883 10.44724 25.51886
Std. 0.001842 9.479113 7.879807 6.66936
PT (Seconds) 159.074029 1638.957037 4.816792 84.833759
TF10 Ave. 0.017933 0.23103 0.280137 9.498785
Std. 0.013532 0.487053 0.601817 1.271393
PT (Seconds) 128.431567 1297.325669 8.013542 84.666823
TF11 Ave. 0.066355 0.193354 0.083463 7.719959
Std. 0.030973 0.073495 0.035067 3.62607
PT (Seconds) 130.664299 1210.086084 9.429028 56.656545
TF12 Ave. 2.78659E-05 0.031101 8.57E-11 1858.502
Std. 3.83626E-05 0.098349 2.71E-10 5820.215
PT (Seconds) 140.837076 1464.060419 22.898798 102.745164
TF13 Ave. 0.000309 0.002197 0.002197 68047.23
Std. 0.000512 0.004633 0.004633 87736.76
PT (Seconds) 139.449467 1339.438272 16.752814 103.377836
TF14 Ave. 0.998004 103.742 150 130.0991
Std. 1.26E-11 91.24364 135.4006 21.32037
PT (Seconds) 170.207352 1034.450489 86.298548 152.142368
TF15 Ave. 0.002358 193.0171 188.1951 116.0554
Std. 0.003757 80.6332 157.2834 19.19351
PT (Seconds) 247.224271 1659.652400 8.250347 54.974533
TF16 Ave. 1.03163 458.2962 263.0948 383.9184
Std. 2.46E-06 165.3724 187.1352 36.60532
PT (Seconds) 181.858429 969.827007 4.247415 80.998874
TF17 Ave. 0.397888 596.6629 466.5429 503.0485
Std. 3.16E-06 171.0631 180.9493 35.79406
PT (Seconds) 141.213291 1018.757437 2.607163 50.990811
TF18 Ave. 3.000142 229.9515 136.1759 118.438
Std. 0.000283 184.6095 160.0187 51.00183
PT (Seconds) 180.663489 1001.716543 2.718852 80.273981
TF19 Ave. 3.86278 679.588 741.6341 544.1018
Std. 9.61E-07 199.4014 206.7296 13.30161
PT (Seconds) 169.415055 1312.805448 8.952319 77.905123

tion is chosen based on this smaller group other subpopulations of the test functions the LPB showed significantly better results
are built. The perfect subpopulation has priority to be optimized compared to the DA. Again, in reference [19] it was proved that
first, then the good subpopulation and so on. Since the subpopula- the results of the DA are statistically significant comparing to the
tions are much smaller compared to the main population, search- PSO and GA. This means that there is no need to compare the pro-
ing for the solutions in these subpopulations is speeder. posed algorithm with PSO and GA statistically since it has proved
Consequently, this improves the randomness and saves the opti- its superiority against DA. As shown in Table 4, all the results
mization time simultaneously. However, compared to the PSO except (TF6, TF11, TF12, and TF19) were smaller than 0.05, which
and GA, the PT of the LPB is higher. proves the importance of the results of the proposed algorithm.

5.3. Statistical tests 5.4. Real world application

The Wilcoxon rank-sum test function [35] is used to verify the In this section, the proposed algorithm is used to optimize a
importance of the results statistically. The p values reported in generalized assignment problem. The problem and its representa-
Table 4 for classical benchmark test functions prove that for most tion are discussed in the following two sections.

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx 7

Fig. 1. Convergence curve for LPB on unimodal, multi-modal, and composite benchmark function.

Table 3
IEEE CEC 2019 benchmark test results.

CEC Function LPB DA PSO

CEC01 Ave. 7494381363.65768 543  108 1.47127E + 12
Std. 8138223463.28023 669  108 1.32362E + 12
PT (Seconds) 377.373846 2034.958870 382.330436
CEC02 Ave. 17.63898 78.0368 15183.91348
Std. 0.31898 87.7888 3729.553229
PT (Seconds) 140.912536 2122.108475 6.064791
CEC03 Ave. 12.7024 13.7026 12.70240422
Std. 0 0.0007 9.03E-15
PT (Seconds) 144.194876 2223.799974 8.901970
CEC04 Ave. 77.90824 344.3561 16.80077558
Std. 29.88519 414.0982 8.199076134
PT (Seconds) 137.305797 1720.974833 5.179151
CEC05 Ave. 1.18822 2.5572 1.138264955
Std. 0.10945 0.3245 0.089389848
PT (Seconds) 138.406681 1722.243949 5.370252
CEC06 Ave. 3.73895 9.8955 9.305312443
Std. 0.82305 1.6404 1.69E + 00
PT (Seconds) 142.041586 1401.682147 131.167162
CEC07 Ave. 145.28775 578.9531 160.6863065
Std. 177.8949 329.3983 104.2035197
PT (Seconds) 122.135692 1376.289834 5.436392
CEC08 Ave. 4.88769 6.8734 5.224137165
Std. 0.67942 0.5015 0.786760649
PT (Seconds) 138.207450 1802.883649 5.527832
CEC09 Ave. 2.89429 6.0467 2.373279266
Std. 0.23138 2.871 0.018437068
PT (Seconds) 141.699472 1365.799778 4.446880
CEC10 Ave. 20.00179 21.2604 20.28063455
Std. 0.00233 0.1715 0.128530895
PT (Seconds) 147.995515 1699.088096 9.462923

5.4.1. Problem definition way that the total hours of assigning cases to the justice teams
A generalized assignment problem known as (GAP) is a popular are minimized. To form the problem mathematically, first we
NP-hard combinatorial optimization problem [41]. The main goal define the following symbols:
in the GAP is assigning a set of tasks to a set of workers with min- i ? row number indicating ith case i e [1, N]
imum cost. In this work, we assign cases in the court to justice j ? column number indicating jth justice team j e[1, N]
teams in a way that the cases could be finished within a minimum C[i][j] ? cost of allocating ith case to the jth team
number of working hours. Assigning cases and justice administra- X[i][j] = 1 if jth justice team is assigned to ith case
tion in the judicial system is routine works, however, they are very X[i][j] = 0 otherwise.
time-consuming. Increasing caseloads at any time will make the The problem can be formulated mathematically as:
problem more series. In this work, we use the proposed algorithm X
N X
N
to assign the right case to the right justice team and to assign a Min C ½i½jX ½i½j ð3Þ
proper time to deliver the decision of the court. The cases should i¼1 j¼1

be assigned to the teams in the base of the number of hours Subject to:
required by that team to deal with that case. So that, it can be con-
sidered that N cases and N justice teams are available where we X
N
X ½i½j ¼ 1; 8i 2 N ¼ f1; 2;    Ng
have to assign each case to one and only one justice team in a
i¼1

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
8 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx

Table 4 two points between 1 and N were generated and values of those
The wilcoxon rank-sum test overall two positions were swapped. The proposed algorithm was applied
runs for classical benchmark test
functions.
to the problem using 80 individuals, for 200 iterations. To verify the
ability of the algorithm to solve the problem different size of the
Test Function LPB Vs. DA matrix was given to the algorithm, as shown in Table 5. To run
TF1 7.72E-06 the program a standard laptop with processor Intel Core i7,
TF2 1.07E-10 16 GHz was used. The results for different matrix sizes are shown
TF3 5.52E-09
in Table 5.
TF4 3.42E-06
TF5 0.006739 In all the cases the population size was kept to 80, and the val-
TF6 0.75328294 ues for the matrix were generated in the range [10 100]. Fig. 2
TF7 7.77E-13 shows the convergence of the algorithm towards the global mini-
TF8 4.23E-27
mum for solving the aforementioned problem using different size
TF9 1.91E-05
TF10 1.08E-09
for the matrix. The Figures show that the size of the matrix will
TF11 5.96E-17 not affect the accuracy of the proposed algorithm and its conver-
TF12 0.138213 gence towards the global minimum.
TF13 0.185156
TF14 0.04631
TF15 0.025386 6. Conclusions
TF16 0.033765
TF17 0.089253
TF18 0.007899
This paper proposed another metaheuristic algorithm based on
TF19 0.35758 the process of transferring graduated learners from high school to
university and improving the studying behaviors of the learners at
colleges. The genetic algorithm inspired this algorithm. The two
X
N most important phases of metaheuristic algorithms (exploitation
X ½i½j ¼ 1; 8j 2 N ¼ f1; 2;    Ng and exploration) were outlined. Mimicking the process of transfer-
j¼1 ring graduated learners from high school to college and dividing
them into different groups according to their GPA outlined the for-
X ½i½j 2 f0; 1g mer phase. The exploration phase, however, was designed by mim-
icking the process of improving the level of learners by utilizing
5.4.2. Problem representation various affective study skills. The parameters used in the LPB were
Representing the problem will be a row from 1 to N examining dp, crossover, mutation. The dp parameter is used in the first steps
the square cost matrix. Every individual in the population is a per- of the algorithm to divide the population into different groups. The
mutation from 1 to N. If the jth element in the row is i, thus, the ith latter two parameters were utilized in the process of improving
case will be given to the jth justice team. For instance, let’s consider learners studying skills.
the following matrix: The ability of the proposed work was benchmarked using tradi-
tional test function and the CEC-C06 2019 functions. The results
were compared to PSO, GA and one of the most recently developed
Team1 Team2 Team3 Team4 Team5
algorithms, which is DA. It was proven that the LPB performed bet-
Case1 23 21 12 30 19 ter in most of the cases. Moreover, The processing time of the algo-
Case2 30 25 13 22 21 rithm was compared to the GA, PSO, and DA. the PT of the proposed
Case3 21 23 32 40 15 work was much smaller compared to the DA. However, it was
Case4 12 32 40 32 29 found that the processing time of the PSO, and GA is smaller than
Case5 20 15 21 27 22 the LPB. Additionally, the results of the Wilcoxon rank-sum test
function proved the significance of the produced results by the
If the solution is [4 5 2 3 1] that means case 4 in column 1 with cost proposed technique. Furthermore, the ability of the algorithm
12 will be given to the first justice team, case 5 in column 2 with was tested using a real-world NP-hard problem. Again, the results
cost 15 will be given to the second team, case 2 with cost 13 in col- proved the effectiveness of the proposed algorithm in solving a
umn 3 will be given to the third team, and so on. Because of the real-world problem. As per finding of the examined work, it can
constraint that says every case should be assigned to one and only be concluded that the proposed work is able to outperform most
one team and according to the encoding used, elements in each of the algorithms in the literature. However, bigger problem sizes
tuple should be unique. Thus, partially mapped crossover [42] for combinatorial optimization could be a challenge for LPB. There-
was used where the individuals are permutations of numbers fore, it is recommended for researchers in different fields to use it
between 1 and N. For mutation, swap mutation was used, randomly as an optimization technique.
For future works, a number of research directions can be rec-
ommended. First of all, the authors will focus on reducing the
processing time of the algorithm. Moreover, implementing the
Table 5 multi objective version of the algorithm is another research direc-
Result of the court case assignment problem with varying size. tion. Modifying the algorithm to improve the exploitation phase
Size of matrix Optimal No. Of Time Required of LPB is another area that the authors are planning to implement
Solution Generations (Sec.) in the future. Besides, another future work is finding new
10  10 218 17 0.14 parameters to replace the parameters from the genetic algorithm.
15  15 350 15 0.17 In addition, utilizing the proposed technique to optimize different
20  20 425 34 0.33 problems and compare the results with other heuristic
30  30 676 57 0.53
techniques.

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx 9

Fig. 2. Convergence to the global minimum using a different number of cases and justice teams, A) 10x10, B) 15x15, C) 20x20, D) 30x30.

Acknowledgment Appendix A

The authors would like to send special thanks to Mr. Ahmed Single-objective test problems are used in this work. See Tables
Saadaldin Qosaeri from the University of Kurdistan Hewler, for 6, 7, and 8 for the mathematical representation of traditional
his thoughtful ideas and discussion. benchmark functions used in this work.

Table 6
Unimodal benchmark functions.

Function Dimension Range Shift position fmin

Pn
TF1ðxÞ ¼ x2 10 [100, 100] [30, 30, . . . 30] 0
Pi¼1 i Q
TF2ðxÞ ¼ ni¼1 jxi j þ ni¼1 jxi j 10 [10,10] [3, 3, . . . 3] 0
Pn Pi 2 10 [100, 100] [30, 30, . . . 30] 0
TF3ðxÞ ¼ i¼1 j¼1 xj

TF4ðxÞ ¼ maxfjxj; 16i6ng 10 [100, 100] [30, 30, . . . 30] 0

i
Pn1 h 2
i
10 [30, 30] [15, 15, . . . 15] 0
TF5ðxÞ ¼ i¼1 100ðxiþ1  x21 Þ þ ðxi  1Þ2
Pn 2 10 [100, 100] [750, . . . 750] 0
TF6ðxÞ ¼ i¼1 ð½xi þ 0:5Þ
Pn 4 10 [1.28, 1.28] [0.25, . . . 0.25] 0
TF7ðxÞ ¼ i¼1 ixi þ random½0; 1

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
10 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx

Table 7
Multi-modal benchmark functions.

Function Range Shift Position fmin

Pn pffiffiffiffiffiffiffi
TF8ðxÞ ¼ i¼1 x sin
2
jxi j [500, 500] [300, . . . 300] 418.9829 X 5
Pn  2i 
TF9ðxÞ ¼ i¼1 i  10 cos ð2
x pxi Þ þ 10 [5.12, 5.12] [2, 2, . . . 2] 0
 qP
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  Pn 
TF10ðxÞ ¼ 20exp 0:2 n 2  exp 1 [–32, 32] 0
i¼1 xi n i¼1 cosð2pxi Þ þ 20 þ e
Pn 2 Qn  
TF11ðxÞ ¼ 4000 xiffi [600, 600] [400, . . . 400] 0
i¼1 xi  þ1
1 p
i¼1 cos i
n P h i o P
TF12ðxÞ ¼ pn 10sinðpy1 Þ þ i¼1 ðyi  1Þ2 1 þ 10sin2 ðpyi þ 1Þ þ ðyn  1Þ2 þ ni¼1 uðxi ; 10; 100; 4Þ
n1 [50, 50] [30, 30, . . . 30] 0
n P h i h io P
TF13ðxÞ ¼ 0:1 sin2 ð3px1Þ þ ni¼1 ðxi  1Þ2 1 þ sin2 ð3pxi þ 1Þ þ ðxn  1Þ2 1 þ sin2 ð2pxn Þ þ ni¼1 uðxi ; 5; 100; 4Þ [50, 50] [100, . . . 100] 0

Table 8
Composite benchmark functions.

Function Dimension Range fmin

TF14ðCF1Þf 1; f 2; f 3    f 10 ¼ Spherefunctiond1; d2; d3    d10 ¼ ½1; 1; 1;    1k1; k2; k3;    k10 ¼ ½100
5
; 100
5
; 100
5
;    100
5


10 [5, 5] 0
TF15ðCF2Þf 1; f 2; f 3    f 10 ¼ Grienwank0 sfunctiond1; d2; d3    d10 ¼ ½1; 1; 1;    1k1; k2; k3;    k10 ¼ ½100
5
; 100
5
; 100
5
;    100
5


10 [5, 5] 0
TF16ðCF3Þf 1; f 2; f 3    f 10 ¼ Grienwank0 sfunctiond1; d2; d3    d10 ¼ ½1; 1; 1;    1k1; k2; k3;    k10 ¼ ½1; 1; 1;    1 10 [5, 5] 0
0 0
TF17ðCF4Þf 1; f 2 ¼ Ackley sfunctionf 3; f 4 ¼ Rastrigin sfunctionf 5; f 6 ¼ Weierstrass0 sfunctionf 7; f 8 ¼ Griewank0 sfunctionf 9;
f 10 ¼ Spherefunctiond1; d2; d3    d10 ¼ ½1; 1; 1;    1k1; k2; k3;    k10 ¼ ½32
5 5
; 32 ; 1; 1; 0:5
5
; 0:5
5
; 100
5
; 100
5
; 100
5
; 100
5


10 [5, 5] 0
0 0
TF18ðCF5Þf 1f 2 ¼ Rastrigin sfunctionf 3; f 4 ¼ Weierstrass0 sfunctionf 5; f 6 ¼ Griewank0 sfunctionf 7; f 8 ¼ Ackley sfunctionf 9;
f 10 ¼ Spherefunctiond1; d2; d3    d10 ¼ ½1; 1; 1;    1k1; k2; k3;    k10 ¼ ½15 ; 15 ; 0:5
5
; 0:5
5
; 100
5
; 100
5
; 32 ; 32 ; 100
5 5 5
; 100
5


10 [5, 5] 0
0 0
TF19ðCF6Þf 1f 2 ¼ Rastrigin sfunctionf 3; f 4 ¼ Weierstrass0 sfunctionf 5; f 6 ¼ Griewank0 sfunctionf 7; f 8 ¼ Ackley sfunctionf 9;
f 10 ¼ Spherefunctiond1; d2; d3    d10 ¼ ½0:1; 0:2; 0:3; 0:4; 0:5; 0:6; 0:7; 0:8; 0:9; 1k1; k2; k3;    k10
¼ ½0:1  15 ; 0:2  15 ; 0:3  0:5
5
; 0:4  0:5
5
; 0:5  100
5
; 0:6  100
5
; 0:7  32
5
; 0:8  32
5
; 0:9  100
5
; 1  100
5


10 [5, 5] 0

Table 9
CEC-C06 2019 benchmark functions [37].

Function Functions Dimension Range fmin

CEC01 STORN’S CHEBYSHEV POLYNOMIAL FITTING PROBLEM 9 [8192, 8192] 1
CEC02 INVERSE HILBERT MATRIX PROBLEM 16 [16384, 16384] 1
CEC03 LENNARD-JONES MINIMUM ENERGY CLUSTER 18 [4, 4] 1
CEC04 RASTRIGIN’S FUNCTION 10 [100, 100] 1
CEC05 GRIENWANK’S FUNCTION 10 [100, 100] 1
CEC06 WEIERSRASS FUNCTION 10 [100, 100] 1
CEC07 MODIFIED SCHWEFEL’S FUNCTION 10 [100, 100] 1
CEC08 EXPANDED SCHAFFER’S F6 FUNCTION 10 [100, 100] 1
CEC09 HAPPY CAT FUNCTION 10 [100, 100] 1
CEC10 ACKLEY FUNCTION 10 [100, 100] 1

Appendix B 629. Available at: https://www.sciencedirect.com/science/article/abs/pii/

S0377221708003597 [Accessed 26 Jun. 2019].
[5] Lawler, E. and Wood, D. Branch-and-Bound Methods: A Survey. Operat Res
The CEC-C06 2019 benchmark functions are shown in the fol- 1966; [online] 14(4), pp.699-719. Available at: https://pubsonline.informs.
lowing table: org/doi/abs/10.1287/opre.14.4.699 [Accessed 18 May 2019].
[6] Goldberg DE, Holland JH. Genetic algorithms and Machine Learning. Mach
Learn 1988;3(2–3):95–9.
[7] Raja, M., Umar, M., Sabir, Z., Khan, J. and Baleanu, D. A new stochastic
computing paradigm for the dynamics of nonlinear singular heat conduction
References model of the human head. Eur Phys J Plus, [online]; 2018: 133(9). Available at:
https://link.springer.com/article/10.1140/epjp/i2018-12153-4 [Accessed 10
Dec. 2019].
[1] Bezdek JC. On the relationship between neural networks, pattern recognition,
[8] Zahoor Raja, M., Shah, Z., Anwaar Manzar, M., Ahmad, I., Awais, M. and
and intelligence. Int J Approximate Reasoning 1992;6(2):85–107.
Baleanu, D. A new stochastic computing paradigm for nonlinear Painlevé II
[2] Xing B, Gao WJ. Innovative computational intelligence: A rough guide to 134
systems in applications of random matrix theory. Eur Phys J Plus, [online];
clever algorithms. Cham, Switzerland: Springer Publication; 2014.
2018: 133(7). Available at: https://link.springer.com/article/10.1140/epjp/
[3] Du KL, Swamy MNS. Search and optimization by
i2018-12080-4 [Accessed 10 Dec. 2019].
metaheuristics. Switzerland: Springer Publication; 2016.
[9] Ahmad, I., Zahid, H., Ahmad, F., Raja, M. and Baleanu, D. Design of
[4] Jourdan, L., Basseur, M. and Talbi, E. Hybridizing exact methods and
computational intelligent procedure for thermal analysis of porous fin
metaheuristics: A taxonomy. Eur J Operat Res 2009; [online] 199(3), pp.620-

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003
 C.M. Rahman, T.A. Rashid / Egyptian Informatics Journal xxx (xxxx) xxx 11

model. Chin J Phys, [online]; 2019: 59, pp.641-655. Available at: https:// [27] Črepinšek, M., Liu, S. and Mernik, M. Exploration and exploitation in
www.sciencedirect.com/science/article/abs/pii/S0577907318308396?via% evolutionary algorithms-A survey. ACM Computing Surveys 2019; [online]
3Dihub [Accessed 10 Dec. 2019]. Available at: https://romisatriawahono.net/lecture/rm/survey/softcomputing/
[10] Sabir, Z., Manzar, M., Raja, M., Sheraz, M. and Wazwaz, A. Neuro-heuristics for Crepinsek%20-%20Exploration%20and%20Exploitation%20in%20Evolutionary%
nonlinear singular Thomas-Fermi systems. Appl Soft Comput, [online]; 2018: 20Algorithms%20-%202013.pdf [Accessed 27 Jun. 2019].
65, pp.152-169. Available at: https://www.sciencedirect.com/science/article/ [28] Qayyum, A. Learner help-seeking attitudes and behaviors in a digital era. Int J
abs/pii/S1568494618300152 [Accessed 10 Dec. 2019]. Educat Technol Higher Education 2018; [online] 15(1). Available at: https://
[11] Fogel, L. J., Owens, A. J., & Walsh, M. J. Intelligent decision making through a educationaltechnologyjournal.springeropen.com/articles/10.1186/s41239-
simulation of evolution. Behav Sci 1966; 11(4), 253–272. Available at: 018-0100-7 [Accessed 31 May 2019].
https://doi.org/10.1002/bs.3830110403 [29] CHEW, S. Improving Classroom Performance by Challenging Learner
[12] Kirkpatrick S, Gelatt DC, Vechhi MP. Optimization by simulated annealing. Misconceptions About Learning online 2010 Association for Psychological
Science 1983;1966(220):671–80. Available at: https://pdfs.semanticscholar. Science Available at: https://www.psychologicalscience.org/observer/
org/e893/4a942f06ee91940ab57732953ec6a24b3f00.pdf. improving-classroom-performance-by-challenging-learner-misconceptions-
[13] Dorigo M, Blum C. Ant colony optimization theory: A survey. Theoret Comput about-learning [Accessed 31 May 2019].
Sci 2005;344(2–3):243–78. [30] Karwowski, M. Are creative learners really welcome in the classrooms?
[14] Kennedy, J. Particle Swarm Optimization. Encyclopedia of Machine Learning Implicit theories of ‘‘good” and ‘‘creative” learner’ personality among polish
and Data Mining 2017; [online] pp.967-972. Available at: https://www. teachers. Procedia - Soc Behav Sci 2010; [online] 2(2), pp.1233-1237. Available
semanticscholar.org/paper/Particle-Swarm-Optimization-Kennedy/ at: https://www.sciencedirect.com/science/article/pii/S1877042810002193
20a5cda34f158ace5881d3f6635c3eb5b531c199 [Accessed 17 May 2019]. [Accessed 31 May 2019].
[15] Basturk, B., Karaboga, D. An artificial bee colony (ABC) algorithm for numeric [31] Roediger, H.L., III, & Karpicke, J.D. Test-enhanced learning: Taking memory
function optimi- zation. In: Proceedings of the IEEE Swarm Intelligence tests improves long-term retention. Associat Psychol Sci 2006; Available at:
Symposium 2006; Indianapolis, Indiana, USA, 12–14 May; 2006. http://learninglab.psych.purdue.edu/downloads/
[16] Chu, S., Tsai, P. and Pan, J. Cat Swarm Optimization. Lecture Notes in Computer 2006_Roediger_Karpicke_PsychSci.pdf , 17, 249-255.
Science 2006; [online] pp.854-858. Available at: https://link.springer.com/ [32] Cao L, Nietfeld J. College Learners’ Metacognitive Awareness of Difficulties in
chapter/10.1007/978-3-540-36668-3_94 [Accessed 18 May 2019]. Learning the Class Content Does Not Automatically Lead to Adjustment of
[17] Yang XS. A New Metaheuristic Bat-Inspired Algorithm. Nat Insp Cooperat Study Strategies. Austral J Educat Dev Psychol 2007;7:31–46.
Strateg Optimiz 2010;84:65–74. Available at: https://link.springer.com/ [33] Gen, M. Lin, L. Genetic Algorithms. Wiley Encyclopedia of Computer Science
chapter/10.1007/978-3-642-12538-6_6. and Engineering 2008; [online] Available at: https://doi.org/10.1002/
[18] Mirjalili, S., Mirjalili, S. and Lewis, A. Grey Wolf Optimizer. Adv Eng Softw 9780470050118.ecse169 [Accessed 8 Jun. 2019].
2014; [online] 69, pp.46-61. Available at: https://www.sciencedirect. [34] Millis, B. Using Metacognition to Promote Learning. IDEA Paper #63 2016;
com/science/article/pii/S0965997813001853 [Accessed 18 May 2019]. IDEA, no. 63. Available at: https://www.ideaedu.org/Portals/0/Uploads/
[19] Mirjalili, S. Dragonfly algorithm: a new meta-heuristic optimization technique Documents/IDEA%20Papers/IDEA%20Papers/PaperIDEA_63.pdf
for solving single-objective, discrete, and multi-objective problems. Neur [35] Derrac, J., García, S., Molina, D. and Herrera, F. A practical tutorial on the use of
Comput Appl 2015; [online] 27(4), pp.1053-1073. Available at: https://link. nonparametric statistical tests as a methodology for comparing evolutionary
springer.com/article/10.1007/s00521-015-1920-1 [Accessed 18 May 2019]. and swarm intelligence algorithms. Swarm Evolution Comput 2011; [online] 1
[20] Abdullah, J., Ahmed, T. Fitness Dependent Optimizer: Inspired by the Bee (1), pp.3-18. Available at: https://www.sciencedirect.com/science/article/abs/
Swarming Reproductive Process. IEEE Access 2019; [online] 7, pp.43473- pii/S2210650211000034 [Accessed 2 Jun. 2019].
43486. Available at: https://ieeexplore.ieee.org/ielx7/6287639/8600701/ [36] Yao, X., Liu, Y. and Lin, G. Evolutionary programming made faster. IEEE Trans
08672851.pdf?tp=&arnumber=8672851&isnumber=8600701&ref= [Accessed Evolut Comput 1999; [online] 3(2), pp.82-102. Available at: https://ieeexplore.
31 May 2019]. ieee.org/document/771163 [Accessed 1 Jun. 2019].
[21] Chandra Mohan, B. and Baskaran, R. A survey: Ant Colony Optimization based [37] Price KV, Awad NH, Ali MZ, Suganthan PN. The 100-Digit Challenge: Problem
recent research and implementation on several engineering domain. Exp Syst Definitions and Evaluation Criteria for the 100-Digit Challenge Special Session
Appl 2012; [online] 39(4), pp.4618-4627. Available at: https:// and Competition on Single Objective Numerical
www.sciencedirect.com/science/article/pii/S0957417411013996 [Accessed Optimization. Singapore: Nanyang Technological University; 2018.
18 May 2019]. [38] Molga, M., Smutnicki, C. Test functions for optimization needs. Test functions
[22] Karaboga, D., Gorkemli, B., Ozturk, C. and Karaboga, N. A comprehensive for optimization needs 2005; Available at: http://www.robertmarks.org/
survey: artificial bee colony (ABC) algorithm and applications. Artif Intellig Rev Classes/ENGR5358/Papers/functions.pdf [Accessed 1 Jun. 2019].
2012; [online] 42(1), pp.21-57. Available at: https://link.springer.com/article/ [39] Liang, J., Suganthan, P., Deb, K. Novel composition test functions for numerical
10.1007/s10462-012-9328-0 [Accessed 18 May 2019]. global optimization. Proceedings 2005 IEEE Swarm Intelligence Symposium,
[23] Zhang, Y., Wang, S. and Ji, G. A Comprehensive Survey on Particle Swarm 2005. SIS 2005, [online] (Proceedings of 2005 IEEE), pp.68-75. Available at:
Optimization Algorithm and Its Applications. Mathemat Probl Eng 2015; https://ieeexplore.ieee.org/document/1501604 [Accessed 2 Jun. 2019].
[online] Available at: https://www.hindawi.com/journals/mpe/2015/931256/ [40] Rahman, C. Rashid, T. Dragonfly Algorithm and Its Applications in Applied
[Accessed 18 May 2019]. Science Survey. Computat Intellig Neurosci, [online] 2019, pp.1-21. Available
[24] Chawla, M. Duhan, M. Bat Algorithm: A Survey of the State-of-the-Art. Appl at: https://new.hindawi.com/journals/cin/2019/9293617/ [Accessed 28 Dec.
Artif Intellig 2015; [online] 29(6), pp.617-634. Available at: https:// 2019].
www.tandfonline.com/doi/abs/10.1080/08839514.2015.1038434? [41] Sahu, A. Tapadar, R. Solving the Assignment problem using Genetic Algorithm
journalCode=uaai20 [Accessed 18 May 2019]. and Simulated Annealing. IAENG Int J Appl Mathemat 2007; [online] Available
[25] Lee, N., Li, X. Wang, D. A comprehensive survey on genetic algorithms for DNA at: http://www.iaeng.org/IJAM/issues_v36/issue_1/IJAM_36_1_7.pdf
motif prediction. Informat Sci 2018; [online] 466, pp.25-43. Available at: [Accessed 24 May 2019].
https://www.sciencedirect.com/science/article/pii/S0020025518305206 [42] Goldberg, D. Lingle, R. Alleles, Loci, and the Traveling Salesman Problem. In:
[Accessed 18 May 2019]. Proceedings of the First International Conference on Genetic Algorithms and
[26] Mohammed, H., Umar, S. Rashid, T. A Systematic and Meta-Analysis Survey of Their Application 1985, L. Erlbaum Associates Inc. Hillsdale, NJ, USA Ó; 1985,
Whale Optimization Algorithm. Computat Intellig Neurosci 2019; [online] pp.154-159.
2019, pp.1-25. Available at: https://www.hindawi.com/journals/cin/2019/
8718571/cta/ [Accessed 18 May 2019].

Please cite this article as: C. M. Rahman and T. A. Rashid, A new evolutionary algorithm: Learner performance based behavior algorithm, Egyptian Infor-
matics Journal, https://doi.org/10.1016/j.eij.2020.08.003