See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/336445054

Improved Swarm Intelligence Optimization using Crossover and Mutation for

Medical Classification

Conference Paper · October 2019

DOI: 10.1109/ICTCS.2019.8923027

CITATIONS READS
0 112

2 authors:

Mais Yasen Nailah Al-Madi

Princess Sumaya University for Technology Princess Sumaya University for Technology
6 PUBLICATIONS   16 CITATIONS    21 PUBLICATIONS   211 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

EvoloPy-NN: An open source nature-inspired optimization Framework for Training Multilayer Perceptron Neural Network in Python View project

IMPROVED GENETIC PROGRAMMING TECHNIQUES FOR DATA CLASSIFICATION View project

All content following this page was uploaded by Mais Yasen on 11 October 2019.

The user has requested enhancement of the downloaded file.

 Improved Swarm Intelligence Optimization using
Crossover and Mutation for Medical Classification
Mais Yasen, Nailah Al-Madi
Department of Computer Science
Princess Sumaya University for Technology
Amman, Jordan
[email protected], [email protected]
Abstract – Early diagnoses helps in curing most diseases or in and adjusts the weights without returning back to the input layer,
making them more bearable, it is vital to enhance the accuracy of and avoids getting stuck in local optima. This can explain why
predicting chronic diseases. Extreme Learning Machine (ELM) is a ELM has good generalization performance without using cycles,
classifier which can be efficiently used to predict diseases. Artificial thus learning faster than other training methods such as
Bee Colony algorithm (ABC) and Dragonfly Algorithm (DA) have
backpropagation [5].
been efficiently used in several optimization problems, including the
optimization of ELM settings. Evolutionary Computation is a type To increase the prediction accuracy of ELM, it can be
of optimization algorithm, which has biological operators to find implemented in conjunction with optimization algorithms to
desired solutions. Two of these operators are crossover and mutation efficiently choose the number of its hidden layer nodes and values
(CM) that are used to generate new solutions from old ones, and can of weights throughout the learning process [27]. Swarm
be integrated with swarm intelligence algorithms to enhance their Intelligence (SI) is a type of population based and nature inspired
results. In this paper, models that make use of ABC and DA to metaheuristic optimization algorithms that reflects the natural
optimize the number of hidden neurons and weights of ELM are behavior of biological swarm groups [6]. Artificial Bee Colony
presented. Moreover, crossover and mutation are combined with the algorithm (ABC) and Dragonfly Algorithm (DA) are SI algorithms
swarm search of ABC and DA for chronic diseases forecasting, in
that can be applied in the optimization of the number of hidden
models called ELM-ABC-CM and ELM-DA-CM. Using 4 real
datasets to evaluate the proposed models, and compare their results nodes and weights from an ELM. The reason why ABC and DA
with the results of standard ABC and DA, and other well-known are chosen is that ABC has a feature of grouping the solution, and
classifiers, including regular ELM, using different evaluation DA has a feature of distracting from enemies, these features and
metrics. The results show that crossover and mutation improved the their phases will enable the employment of natural operations.
outcome of ABC and DA. Moreover, ELM-DA-CM proved its Also, ABC and DA proved their efficiency before in previous
efficiency over ELM-ABC-CM. works [7, 8].
Keywords—Machine Learning; Swarm Intelligence; Evolutionary Evolutionary Computation (EC) is another type of population
Computation; Extreme Learning Machine; Dragonfly Algorithm; based and nature inspired metaheuristic optimization algorithms.
Artificial Bee Colony; Crossover; Mutation; Medical Prediction. EC iteratively applies biological evolution to generate solutions
I. INTRODUCTION [9]. Crossover and mutation are two vital biological operators in
Early diagnoses is important to cure most diseases or to manage EC that are used to generate new populations from an existing one
them by preventing their consequences and making them more and enhance the results by having more exploration and
bearable [1]. Therefore, it is an essential requirement to increase exploitation [9]. These operators can be applied with SI
the accuracy of predicting diseases such as heart disease, hepatitis, optimization algorithms to enhance the accuracy of prediction of
diabetes, and diabetic retinopathy. The symptoms of these diseases an optimized classification algorithm. The contribution of this
need to be taken into consideration when forecasting them using paper is summarized as the following:
machine learning [2]. 1. Using crossover and mutation on ABC and DA.
Machine learning (ML) in artificial intelligence enables the 2. Optimizing ELM using ABC-CM and DA-CM and improving
computers to learn without being explicitly programmed [2]. It the tuning of ELM.
finds patterns by searching through the data and uses the detected 3. Using 4 real datasets for training and testing our models.
patterns to alter program actions accordingly [2]. The process 4. Evaluating the proposed models and comparing them with other
when algorithms reflect what has been learned in the past from classifiers.
training data to predict new data is called supervised ML [3]. This paper is structured as follows: Section II includes the
Classification is one of the main supervised ML tasks that aims related literature in the area of work. Section III describes the
to build a model based on previous data to classify new data. background of methods used in this work. Section IV includes the
Extreme Learning Machine (ELM) is a neural network that is proposed methodology used in the development. Section V
inspired by the biological brain, it consists of a computational presents the experiments and the results, and Section VI concludes
model that contains a number of processing nodes called neurons the research and discusses future work.
[4]. Neurons send signals to one another over large number of
weighted connections that link input, hidden and output layer
together for communication purposes. ELM training method is
feedforward, which travels from the input layer to the output layer
 II. RELATED WORK III. BACKGROUND
Crossover and mutation operations of EC can be used to This section discusses the methods used in this work, starting with
enhance the search process of SI algorithms, and they were ELM, then ABC and DA, and lastly crossover and mutation.
previously applied on particle swarm optimization (PSO).
As mentioned in [10], the PSO adjusts itself based on the A. Extreme Learning Machine
previous information about particles and performance of Extreme learning machine consists of a single hidden layer
neighbors. The work presented a PSO with discrete binary feedforward Neural Network (NN) used for classification. A
variables, where the authors tested 5 different De Jong evaluation neural network could be defined as “a computing system made up
functions, where evolutionary algorithms with crossover combine of a number of simple, highly interconnected processing elements,
information from the parents to allow leaps. When there are many which process information by their dynamic state response to
global solutions, crossover could be harmful because two solutions external inputs” [15]. Fig. 1 shows the structure of NN, where NNs
may give better results than any crossover of them. The results are distributed in layers that consist of a group of interconnected
showed that PSO was able to solve various problems which means nodes, also known as neurons, which have an activation function
it is extremely flexible and robust, but it had a problem in getting inside and are inspired by the biological brain [4].
out of a good local optimum.
The work presented in [11], used PSO for solutions updates and
combined it with Gaussian Mutation. The results were compared
with the original PSO and Genetic Algorithm (GA) using DeJong's
functions. PSO with Gaussian Mutation was able to be better than
GA. Furthermore, PSO with mutation was applied on a gene
network where it got better results than standard GA and PSO.
The authors in [12], proposed a theta-PSO with crossover and
mutation to enhance PSO. Their proposed algorithm was capable
Fig. 1 ELM NN Structure [4]
of getting out of local minima by adjusting the parameters
properly. The results were tested on 4 multi-modal functions, their In ELM the training method is feedforward which has no
algorithm reached the global optimal in a limited number of cycles, taking into consideration that ELM is feedforward using an
iterations achieving a high success rate. On the other hand, the appropriate number of hidden nodes and weights will enable the
value of fitness increased in the beginning of crossover and algorithm to learn fast without propagating back. Choosing ELM
mutation, resulting in a long iteration time. settings unwisely can slow down ELM or result in a low accuracy,
In [13], the authors addressed that PSO is widely used as a therefore it is vital to select the number of hidden neurons and
stochastic technique in global optimization. As PSO includes the weights wisely, instead of just generating them randomly.
variables: local and global best position, and because of its early In the process of training, sigmoid function is used in hidden
convergence it can be easily stuck in a local optimal solution. neurons, which is used in the hidden layer to transfer between
Keeping a big search space and ensuring the population diversity different neurons as shown in Equation (1) [16]. Where x is the
can help in preventing that problem by balancing the exploration value of input for each node.
and exploitation. The authors introduced using crossover and / …
mutation with PSO where it is performed on all of the particles in The execution steps of ELM are:
the current iteration if the diversity of the particles reaches a value 1. Extract features for the input layer nodes.
less than a predefined threshold. The results were applied on a 12 2. Randomly specify the number of hidden layer nodes.
widely used nonlinear functions and it showed that the proposed 3. Randomly generate the weights and biases connecting the input
approach had better performance than standard PSO. layer with the hidden layer.
As mentioned in [14], the authors presented a new PSO 4. For each hidden node compute the sigmoid value.
algorithm for solving global optimization problems called QPSO. 5. Calculate the inverse Moore-Penrose generalized hidden layer
QPSO is a combination of Quadratic crossover and the Basic PSO output matrix.
algorithm, where they maintained diversity by preventing the
6. For each output compute the sigmoid value.
search space from shrinking and accepting any new solution even
From the steps above it can be seen that ELM could be altered
if it is worse than the best solution found so far. The results were
to work with an optimization method to solve the problem of
tested on 12 benchmark functions and showed that QPSO
initializing the number of hidden nodes and weights randomly.
performed better than BPSO algorithm in dimensions up to 50.
From the related work and to the best of our knowledge, it can B. Artificial Bee Colony
be seen that applying crossover and mutation with SI enhanced the Artificial Bee Colony algorithm (ABC) has been efficiently
results, and was never studied on DA. Moreover, optimizing ELM implemented in many optimization problems. The optimization of
using EC combined with ABC and DA was never studied before. hidden nodes and weights of an ELM could be one of these
problems. The ABC algorithm was first proposed by Karaboga in survive, thus all dragonflies move towards the food sources in the
2005 [17]. And it is a meta-heuristic SI optimization algorithm attraction to food principle [19]. Fifth, to survive all dragonflies
inspired by the foraging behavior of honeybees in nature. The move away as far as possible from the enemy sources in the
solution of ABC is represented in a multi-dimensional search distraction from enemies principle [19]. To calculate the values of
space as food sources and a population of three different types of the different principles the following equations are used [19]:
bees (employed, onlooker, and scout). Let xi be the food source set ∑ … (5) ∑ / … (6)
found by the employed bees for each iteration of the ABC, xi = ∑ / … (7) … (8)
{xi1, xi2, …, xin} where n is the number of solutions needed. … (9)
Equation (2) is used to calculate a new derived solution [18]. The separation is calculated using Equation (5), where Xi is the
, , , ∗ , , … (2) position of the current dragonfly (i), Xj is the position of the jth
Where ϕ is a random number between 0 and 1, y is a random dragonfly close to the current, and n is the total of dragonflies. The
number between 0 and the maximum number of food sources, y alignment is found using Equation (6), where Vj is the velocity
should not equal the current food source (i), and j is a random value of the jth dragonfly close to the current (i). The cohesion is
number generated between 0 and maximum number of solutions. calculated as shown in Equation (7). The attraction to food is
Equation (3) is to calculate the probability of each solution calculated by using Equation (8), where Xf is the position of the
suggested by the employed bees, it is also known as the roulette solution. The distraction from enemy is calculated as shown in
wheel equation that evaluates the solutions based on the fitness Equation (9), where Xe is the position of the enemy. Values of ∆X
values achieved, this phase is called the onlooker bee phase [18]. and X are calculated using Equations (10) and (11) [19]. Where s,
/ ∑ … (3) a, c, f, e, and w are the weights of their correspondent principle (S,
A, C, F, E, and ∆X). e is calculated using Equation (12) where i is
Where i is the current solution, pi is the probability of solution
the current iteration and I is the maximum iterations. s, a, and c are
i, fiti is the fitness value of solution i, sn is the solutions total, j is
three different random numbers between 0 and 2e, f is a random
the solutions counter, and fitj is the fitness of each solution j.
number between 0 and 2, and w is calculated using Equation (13).
The scout bee phase is the final stage, it is responsible of
∆ , ∆ , … (10)
checking the epoch reached so far, which is the number of times
, , ∆ , … (11)
the solution is allowed to get worse than the solution produced
. ∗ . / / … (12)
before. The scout bee abandons the old solution and discovers a
. ∗ / … (13)
new solution for the employed bees to work on it in the following
iterations. Equation (4) is used to calculate the new solution [18]. D. Crossover and Mutation
, ∗ … (4) Evolutionary Computation (EC) is another type of population
Where ub, lb are the vectors that contain the upper bounds and based and nature inspired metaheuristic optimization algorithms.
lower bounds allowed for the solution, ϕ is a random number What distinguishes EC is the use of biological evolution on
between 0 and 1, i is the current food source, and j is a random candidate solutions to remove the worst, and to change solutions
number between 0 and maximum number of solutions. iteratively [9]. Crossover and mutation are popular examples on
the operators used in EC. These operators are applied to generate
C. Dragonfly Algorithm new solutions from existing ones [9].
Dragonfly Algorithm (DA) was first proposed by Seyedali Crossover usually occurs in every iteration to combine the
Mirjalili in 2016 [19]. And it is an algorithm that can be used in genetic information of two parents and generate new children [23].
ELM number of hidden nodes and weights optimization. DA is a Crossover has many types, uniform crossover is illustrated in Fig.
meta-heuristic SI optimization algorithm inspired by the static and 2, where two parents integrate in a uniform pattern to generate a
dynamic behaviors of dragonflies in nature [20]. In the static new child [24]. There are two reasons why uniform crossover was
behavior a large number of dragonflies migrate in a certain chosen to be implemented; first is that using a uniform pattern will
direction travelling for long distances [21]. On the other hand, in guarantee having stable and proportional new derived solutions.
the dynamic behavior dragonflies get into groups and fly over Second is that uniform crossover will help in reaching the best
different areas to find food resources [22]. solution faster, because the amount of solution change is high. On
DA has five principles that are important in finding the the other hand, mutation usually happens less frequently to find
solutions required. First, the separation principle implies the static better solutions by altering the genetic information of one or more
collision avoidance of a dragonfly from other dragonflies that are genes of members of a solution [23]. Fig. 3 shows bit inversion
close to its position [19]. Second, the alignment principle reflects mutation where a single gene is altered [24].
the process of velocity matching of a dragonfly to other
dragonflies that are close to its position [19]. Third, the cohesion
principle is the tendency of a dragonfly towards the center of the
space that contains other dragonflies close to its position [19].
Fourth, the main aim of dragonfly swarms is to stay alive and Fig. 2 Crossover [24] Fig. 3 Mutation [24]
 IV. PROPOSED APPROACHES A. Data
The following are the execution steps of the proposed models, The performance evaluation was done on four Medical datasets
where the fitness calculation is done by sending the proposed [25]. Where first, we applied feature selection on the datasets using
number of hidden nodes and their weights to ELM. ABC-CM and gain ratio to consider only the most relevant features to the class
DA-CM steps are: attribute using WEKA [26]. Then we split the data into two sets,
1. Calculate the solution probability using roulette wheel. Check if 66% for training and 34% for testing, as shown in Table 1.
the solution probability is lower than probability of mutation. Table 1 Number of Records and Features in the data files
2. In the employed bee or dynamic phase: if the probability is Dataset Training Testing Features (Selected)
lower, alter the current solution using mutation, where a new Heart disease 177 93 14 (10)
solution is derived using the equations explored in section III. Hepatitis 102 53 20 (11)
3. Choose the two parent solutions that got the highest fitness. Diabetes 506 262 9 (5)
Retinopathy 760 391 19 (16)
4. Calculate the solution probability using roulette wheel. Check if
the solution probability is lower than probability of crossover.
5. In scout bee or static phase: if probability is lower, reset solution
that reached the epoch and generate a new solution by
combining the two parents selected in a uniform pattern.
6. Repeat steps 1 to 5 in each iteration of ABC or DA.
The execution steps of ELM-ABC, as shown in Fig. 4, are:
1. Initialize all food sources randomly.
2. Employed bees find all the possible solutions.
3. Find the fitness value for each proposed solution using ELM,
and retrieving the resulting accuracy.
4. Onlooker bee phase calculates the probability of each solution.
Then decides greedily based on a random number whether to Fig. 4 ELM-ABC Process Fig. 5 ELM-DA Process
follow the solution or not. B. Experiments settings
5. Scout bee phase checks if each solution reached the epoch time. For the evaluation of our model the fitness function was
6. Store the best solution based on a greedy selection. accuracy. The settings of ABC and DA used are; Iterations: 100,
7. Repeat from step 2 to 6 until maximum iterations is reached. Swarm size: 20, Seed: Random, Number of Sources: 50, Upper
The execution steps of the ELM-DA, as shown in Fig. 5, are: bound: 1, Lower bound: 0, Epoch: 50, Crossover probability: 0.8,
1. Initialize the dragonfly positions and positions difference (∆X) Mutation probability: 0.2. ELM settings are; Output Neurons: 2,
randomly. Seed: Random, Hidden Layers: 1, Hidden Layer Nodes: Random.
2. Calculate the fitness values for the proposed solutions. After preparing the datasets and building our proposed models,
3. Start the static phase by updating the best fitness value. ELM-ABC-CM and ELM-DA-CM need to be run 30 times to
4. If the fitness value was better than the best fitness found so far, cover the randomness of ABC and DA solutions. To evaluate our
then update the best food source with the solution. models, they are compared with seven classifiers implemented in
5. If the fitness value was worse than the worst fitness found so far, WEKA with their default settings: Bayes Network (BN), Naïve
then update the worst enemy source with the solution. Bayes (NB), Decision Tree (J48), K-Nearest Neighbors (IBK), K-
6. Start the dynamic phase by calculating s, a, c, f, e, w, and ∆X. star (K*), Repeated Incremental Pruning (J-Rip), Artificial Neural
7. Calculate the separation, alignment, cohesion, attraction to food, Network (ANN).
and distraction from enemy values. To evaluate the efficiency of the classifiers we can use the
8. Update dragonfly positions difference (∆X) and dragonfly following metrics; Accuracy, recall, precision, Fmeasure, and
positions (X). AUC, using the Equations (14-18). Where TN is true negative, TP
9. Repeat from step 3 to 8 until maximum iterations is reached. is true positive, FN is false negative and FP is false positive.
/ … (14)
V. EXPERIMENTS AND RESULTS / … (15)
The performance of our approaches was evaluated by conducting / … (16)
a number of experiments that are explained in this section. / … (17)
… (18)
 Table 2 Results (*1 Accuracy, *2 Precision, *3 Recall, *4 F-measure, *5 AUC)
Classifier Heart Disease Hepatitis
*1 *2 *3 *4 *5 *1 *2 *3 *4 *5
BN 81.52 0.88 0.80 0.84 0.82 81.13 0.45 0.56 0.50 0.71
NB 82.61 0.88 0.82 0.85 0.83 86.79 0.60 0.67 0.63 0.79
J48 67.39 0.79 0.62 0.69 0.69 86.79 0.67 0.44 0.53 0.70
IBK 78.26 0.89 0.73 0.80 0.80 81.13 0.45 0.56 0.50 0.71
K* 71.74 0.78 0.73 0.75 0.71 90.57 0.75 0.67 0.71 0.81
J-Rip 72.83 0.83 0.69 0.75 0.74 83.02 0.50 0.67 0.57 0.77
ANN 76.09 0.87 0.71 0.78 0.77 88.68 0.67 0.67 0.67 0.80
ELM 75.00 0.90 0.65 0.76 0.77 84.91 0.60 0.33 0.43 0.64
ELM-ABC 83.70 0.83 0.91 0.87 0.82 84.91 0.56 0.56 0.56 0.73
STDEV 1.35 0.03 0.02 0.01 0.02 5.05 0.12 0.13 0.12 0.07
best runs 85.87 0.88 0.93 0.88 0.85 88.68 0.67 0.67 0.67 0.80
ELM-DA 83.70 0.83 0.91 0.87 0.82 88.68 0.80 0.44 0.57 0.71
STDEV 0.00 0.01 0.01 0.00 0.00 0.82 0.04 0.15 0.09 0.06
best runs 83.70 0.84 0.93 0.87 0.83 88.68 0.75 0.67 0.67 0.80
ELM-ABC-CM 84.78 0.86 0.89 0.88 0.84 88.68 0.80 0.44 0.57 0.71
STDEV 1.33 0.03 0.01 0.01 0.02 0.96 0.02 0.06 0.05 0.03
best runs 85.87 0.88 0.93 0.88 0.85 90.57 0.83 0.56 0.67 0.77
ELM-DA-CM 84.62 0.84 0.93 0.88 0.82 90.57 0.83 0.56 0.67 0.77
STDEV 0.54 0.01 0.01 0.00 0.01 0.96 0.05 0.08 0.04 0.03
best runs 84.78 0.84 0.95 0.88 0.83 90.57 0.83 0.89 0.70 0.88

Classifier Diabetes Diabetic Retinopathy

*1 *2 *3 *4 *5 *1 *2 *3 *4 *5
BN 74.05 0.83 0.76 0.80 0.73 77.39 0.79 0.88 0.84 0.73
NB 77.48 0.80 0.87 0.84 0.73 74.33 0.80 0.80 0.80 0.72
J48 76.34 0.83 0.80 0.82 0.74 83.14 0.80 0.98 0.88 0.77
IBK 68.32 0.75 0.79 0.77 0.63 91.57 0.89 1.00 0.94 0.88
K* 69.47 0.74 0.83 0.78 0.63 88.51 0.85 1.00 0.92 0.84
J-Rip 77.48 0.81 0.86 0.83 0.73 81.61 0.82 0.91 0.87 0.77
ANN 76.34 0.83 0.80 0.82 0.74 84.67 0.84 0.95 0.89 0.80
ELM 76.72 0.79 0.89 0.84 0.71 72.80 0.75 0.88 0.81 0.66
ELM-ABC 77.86 0.84 0.82 0.83 0.76 75.10 0.78 0.85 0.82 0.71
STDEV 2.22 0.03 0.03 0.02 0.03 3.82 0.04 0.02 0.02 0.06
best runs 77.86 0.84 0.83 0.83 0.76 77.39 0.82 0.89 0.83 0.75
ELM-DA 78.63 0.85 0.82 0.84 0.77 75.86 0.77 0.91 0.83 0.69
STDEV 0.18 0.00 0.01 0.00 0.00 0.58 0.02 0.05 0.01 0.02
best runs 82.44 0.88 0.85 0.87 0.81 76.63 0.80 0.96 0.84 0.72
ELM-ABC-CM 78.63 0.86 0.80 0.83 0.78 75.38 0.79 0.85 0.82 0.71
STDEV 8.15 0.07 0.12 0.08 0.09 0.00 0.00 0.00 0.00 0.00
best runs 88.55 0.96 0.86 0.91 0.90 75.10 0.79 0.85 0.82 0.71
ELM-DA-CM 79.77 0.83 0.87 0.85 0.76 76.25 0.78 0.89 0.83 0.71
STDEV 2.31 0.03 0.05 0.02 0.03 0.71 0.01 0.02 0.00 0.02
best runs 82.06 0.86 0.91 0.87 0.78 76.63 0.79 0.91 0.83 0.72
C. Results ELM-ABC-CM and ELM-DA-CM improved the results of
It can be concluded from Table 2 that DA is performing more ELM-ABC and ELM-DA, crossover and mutation decreased the
productively with ELM (ELM-DA) than ABC (ELM-ABC) in 3 STDEV values of the 30 runs, which means that the randomness
datasets. This is explained by comparing the Standard Deviation of the solutions is decreased. The diabetes, heart and hepatitis
(STDEV) values of ABC with DA, where ABC has higher STDEV datasets give a high indication that the ELM-DA-CM is capable of
due to the huge effectiveness of the randomness of its operation on reaching the best solution, in comparison with all of the classifiers
the solutions it produces. However, best runs of ABC are very mentioned in the table. Moreover, ELM-ABC-CM has the best
competitive in comparison with DA. Furthermore, ELM-ABC and accuracy in the heart disease dataset. On the other hand, the results
ELM-DA enhanced the results of ELM in all datasets.
of both approaches were very competitive in the other datasets, and [10] J. Kennedy and R. C. Eberhart, "A discrete binary version of the
their best runs were better than most classifiers in the table. particle swarm algorithm", (1997), IEEE International Conference
on Systems, Man, and Cybernetics, Vol. 5, PP. 4104-4108.
VI. CONCLUSION AND FUTURE WORK [11] N. Higashi and H. Iba, "Particle swarm optimization with Gaussian
This work goal was to construct models that can predict mutation", (2003), IEEE Swarm Intelligence Symposium, PP. 72-79.
chronic diseases and evaluate their performance. The proposed [12] Weimin Zhong, Jianliang Xing and Feng Qian, "An improved theta-
models are swarm-based which integrate crossover and mutation PSO algorithm with crossover and mutation", (2008), 7th World
with the search of ABC and DA (called ABC-CM, DA-CM). The Congress on Intelligent Control and Automation, PP. 5308-5312.
enhanced ABC and DA models were used to improve the results [13] Dong G, Cooper J., “Particle Swarm Optimization with Crossover
of ELM classifier. The datasets used in this research were real and Mutation Operators Using the Diversity Criteria”, (2013),
patients’ records of four different medical cases. Results were ASME International Design Engineering Technical Conferences and
compared with other well-known classifiers including ELM using Computers and Information in Engineering Conference, Vol. 3A, PP.
different evaluation metrics. The results showed that ELM-ABC- V03AT03A010.
CM and ELM-DA-CM improved the efficiency of ELM-ABC and [14] Pant M., Thangaraj R., Abraham A., (2007), “A New PSO Algorithm
ELM-DA, and crossover and mutation decreased the randomness with Crossover Operator for Global Optimization Problems”,
of the solutions produced. Moreover, ELM-DA-CM reached the Innovations in Hybrid Intelligent Systems, Advances in Soft
best prediction in three datasets, and ELM-ABC-CM got the best Computing, Vol. 44, PP. 215-222.
accuracy in one dataset. [15] Maureen Caudill, (1989), "Neural Network Primer", San Francisco:
Based on the results, as a future work it is necessary to enlarge Miller Freeman Inc., PP 321.
the search space of ABC and DA to increase their accuracy. [16] A. C. C. Coolen, (1998), “A Beginner’s Guide to the Mathematics of
Moreover, the running time was long, thus it is important to find a Neural Networks”, Springer, Chapter 2, PP 13-70.
way of parallelizing these models to achieve good results in a [17] Dervis Karaboga, (2005), “An Idea Based on Honey Bee Swarm for
Numerical Optimization”, Technical Report-TR06, PP 1-10.
meaningful time.
[18] Yunfeng Xu, Ping Fan, Ling Yuan, (2013), “A Simple and Efficient
REFERENCES Artificial Bee Colony Algorithm”, Mathematical Problems in
[1] WEBMD, “Health Screening: Finding Health Problems Early”, Engineering (MPE), Volume 2013, PP 1-9.
Retrieved on: February 11, 2019. From: www.webmd.com. [19] Seyedali Mirjalili, (2016), “Dragonfly algorithm: a new meta-
[2] Margaret Rouse, (2016), “Analytics tools help make sense of big heuristic optimization technique for solving single-objective,
data”, AWS, Retrieved on: December 6, 2018, From: discrete, and multi-objective problems”, Springer, PP 1053–1073.
searchbusinessanalytics.techtarget.com. [20] M. A. Salam, H. M. Zawbaa, E. Emary, K. K. A. Ghany and B. Parv,
[3] Jerome H. Friedman, (1997), “Data mining and statistics: What’s the (2016), “A hybrid dragonfly algorithm with extreme learning
connection”, Proceedings of the 29th Symposium on the Interface machine for prediction”, INnovations in Intelligent SysTems and
Between Computer Science and Statistics, PP 1-7. Applications (INISTA), PP. 1-6.
[4] Jun-Shien Lin, and Shi-Shang Jang, (1998), “Nonlinear Dynamic [21] Robert W. Russell, Michael L. May, Kenneth L. Soltesz, John W.
Artificial Neural Network Modeling Using an Information Theory Fitzpatrick, (1998), “Massive Swarm Migrations of Dragonflies in
Based Experimental Design Approach”, American Chemical Eastern North America”, University of Notre Dame, PP 325-342.
Society, Vol. 37, PP 3640–3651. [22] Martin Wikelski, David Moskowitz, James S Adelman, Jim
[5] Guang-Bin Huang, Qin-Yu Zhu, Chee-Kheong Siew, (2006), Cochran, David S Wilcove, Michael L May, (2006), “Simple rules
“Extreme learning machine: Theory and applications”, guide dragonfly migration”, PMC, PP 325-329.
Neurocomputing, Vol. 70, PP 489-501. [23] Zakir H. Ahmed, (2010), “Genetic Algorithm for the Traveling
[6] Beni G., Wang J., (1993), “Swarm Intelligence in Cellular Robotic Salesman Problem using Sequential Constructive Crossover
Systems”, Robots and Biological Systems: Towards a New Bionics?, Operator”, International Journal of Biometrics and Bioinformatics
Vol. 102, PP 703-712. (IJBB), Vol. 3, PP 96-105.
[7] M. Z. Yasen, R. A. Al-Jundi and N. S. Al-Madi, (2017), “Optimized [24] Marek Obitko, (1998), “Introduction to Genetic Algorithms”,
ANN-ABC for Thunderstorms Prediction”, 2017 International Obitko, Retrieved on: February 14, 2019, From: obitko.com.
Conference on New Trends in Computing Sciences (ICTCS), PP 98- [25] David Aha, (2013), “UCI Machine Learning Repository”, University
103. of California Irvine.
[8] M. Yasen, N. Al-Madi and N. Obeid, (2018), “Optimizing Neural [26] WEKA, Version: 3.8, Retrieved on: September 5, 2016, From:
Networks using Dragonfly Algorithm for Medical Prediction”, 2018 www.cs.waikato.ac.nz.
8th International Conference on Computer Science and Information [27] Faris, H., Ala’M, A. Z., Heidari, A. A., Aljarah, I., Mafarja, M.,
Technology (CSIT), PP 71-76. Hassonah, M. A., & Fujita, H. (2019). “An intelligent system for
[9] Al-Jundi, Ruba, Mais Yasen, and Nailah Al-Madi, (2017), spam detection and identification of the most relevant features based
“Thunderstorms Prediction using Genetic Programming”, on evolutionary random weight networks”. Information Fusion, 48,
International Journal of Information Systems and Computer 67-83.
Sciences, Vol. 7, PP 1-7.

View publication stats