Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Jul 23, 2019

Content may be subject to copyright.

Congestion Control in Wireless Sensor Networks

based on Support Vector Machine, Grey Wolf

Optimization and Differential Evolution

Haﬁza Syeda Zainab Kazmi1, Nadeem Javaid1∗

1COMSATS University Islamabad

Islamabad, Pakistan

{zainab.kazmi13, nadeemjavaidqau}@gmail.com

∗Correspondence: nadeemjavaidqau@gmail.com; www.njavaid.com

Muhammad Imran

Kind Saud University

Saudi Arabia

dr.m.imran@ieee.org

Fatma Outay

College of Technological Innovation, Zayed University

United Arab Emirates

fatma.outay@zu.ac.ae

Abstract—Transmission rate is one of the contributing factors

in the performance of Wireless Sensor Networks (WSNs). Con-

gested network causes reduced network response time, queuing

delay and more packet loss. To address this issue, we have

proposed a transmission rate control method. The current node

in a WSN adjusts its transmission rate based on the trafﬁc

loading information gained from the downstream node. Multi

classiﬁcation is used to control the congestion using Support

Vector Machine (SVM). In order to get less miss classiﬁcation

error, Differential Evolution (DE) and Grey Wolf Optimization

(GWO) algorithms are used to tune the SVM parameters. The

comparative analysis has shown that the proposed approaches

DE–SVM and GWO-SVM are more proﬁcient than the other

classiﬁcation techniques in terms of classiﬁcation error.

Index Terms—Wireless Sensor Networks, Support Vector Ma-

chine, Transmission Rate, Congestion Control, Differential Evo-

lution, Grey Wolf Optimization

I. INT ROD UC TI ON

Wireless Sensor Network (WSN) consists of large number

of scattered sensor nodes. The data sensed by the sensor nodes

is sent to the sink or a base station. Sensor nodes are being

used for:

•animal tracking,

•ﬂood detection,

•forecasting of the weather data,

•monitoring of patients, and

•vehicle monitoring.

Sensor nodes independently perform some sensing task and

carry out some processing. These sensor nodes communicate

with each other in order to forward the collected data to the

sink node. Some nodes act as relay nodes. Relay nodes are

used to collect the sensed data and route the data to the sink.

WSNs are prone to communication failures. Sensor nodes have

the ability to work in harsh environments. However, sensor

nodes have the following constraints [1]:

•limited battery time,

•less memory, and

•fast energy depletion.

Large number of sensors in a wide geographical area

provides better accuracy. Congestion at a node occurs if the

arrival of data packets at a particular node are greater than

the number of outgoing data packets. Congestion can cause

packet loss and reduced response time. Response time of the

network is described as the amount of time needed for a

packet transmission from a sender to a receiver. Response

time decelerates with the reduced network throughput in a

congested network. Special considerations are required to deal

with congestion in a WSN. Several transmission rate control

mechanisms have been proposed in the past years. Congestion

control has been tackled by the adjustment of transmission

rate at each node [2]. The incoming and outgoing rate of data

packets can be handled in order to avoid retransmission and

packet loss. Congestion should be controlled at each hop to

avoid the problem of packet loss [2]. Mechanisms to detect

and avoid congestion can serve the purpose. Data mining al-

gorithms have been used to recognize such complex problems

and make smart decisions. Learning techniques are categorized

as supervised learning and unsupervised learning that works

on labeled and non-labeled data, respectively. For the problem

mentioned earlier, several classiﬁcation methods can be used to

classify the data and predict the right amount of transmission

rate of sensor nodes in a WSN. Classiﬁcation methods like

Support Vector machine (SVM), k-Nearest Neighbor (k-NN),

Naive Bayes, Neural Networks (NN) and Decision Tree can

be used to classify such data. The optimization algorithms like

2019 Wireless Days (WD)

978-1-7281-0117-0/19/$31.00 ©2019 IEEE

Genetic Algorithm (GA), Harmony Search Algorithm (HSA),

Grey Wolf Optimization (GWO), Differential Evolution (DE),

Fireﬂy algorithm (FA) and Particle Swarm Optimization (PSO)

can be used to optimize the classiﬁer parameters in order

to accurately classify the data. To avoid the problem of

congestion, a huge amount of packet re-transmissions, fast

energy depletion and reduced throughput in WSN, a reliable

transmission rate adjustment methodology is required. In [2],

the authors have shown high classiﬁcation error using GA

and their proposed technique does not sufﬁciently reduce the

amount of re-transmitted packets. Better accuracy has been

achieved using SVM in the presented scenario in [2]. We have

used the same classiﬁcation method SVM to avoid congestion.

However, we have tuned the SVM parameters using DE and

GWO algorithms. This is because DE-SVM and GWO-SVM

are capable of yielding better results than other classiﬁcation

methods. In this paper, we present two novel methods to

avoid the congestion at every hop in a WSN. The congestion

problem is tackled by adjusting the transmission rate using

Support Vector Machine (SVM). The parameters of SVM

are tuned using Differential Evolution (DE) and Grey Wolf

Optimization (GWO) algorithms. We have used these classi-

ﬁers because they have the ability to produce better results

than other classiﬁcation methods. The paper is organized as

follows: Section 2 consists of the related work. Section 3

describes the strategy for transmission rate adjustment. Section

4 provides the proposed method which consists of a brief

background of SVM and the transmission rate adjustment

based by SVM tuned by Differential Evolution (DE) and Grey

wolf Optimization (GWO-SVM). The performance evaluation

and comparison of the presented work with other classiﬁers

is presented in Section 5. Finally, the paper is concluded in

Section 6.

II. RE LATE D WOR K

The problem of congestion in WSNs is tacked in [2] by

adjusting the transmission rate at current node. The node

adjusts its transmission rate by taking buffer occupancy ratio

and congestion degree estimate from the upstream node to

avoid congestion and improve the network throughput. Mul-

ticlassiﬁcation is done by Support Vector Machine (SVM).

The authors have used Genetic Algorithm (GA) for parameter

tuning. The parameters adjusted for all SVMs are acceptable

error, penalty ratio and deviation of Gaussian kernel function.

Authors of [3] have proposed a clustering routing protocol

in WSNs. The method used to enhance the performance

and network lifetime is a three-level hybrid algorithm. The

Multilevel Hybrid Protocol (MLHP) combined tree-based

techniques. At level one, cluster heads are selected, whereas

Grey Wolf Optimization is used for data transfer. To save

energy, nodes select the best route using GWO. At level tree,

distributed clustering is proposed. MLHP gives comparatively

more residual energy, more stability and improved network

lifetime in WSNs. Finding location of unknown nodes is an

important issue to be tackled. GWO [4] can be used for

localization problems. Node localization problem articulates

using range-based technique to calculate the coordinates of

unknown nodes using the positions of the known nodes. The

known nodes are called the anchor nodes which have a GPS

device. Using the GPS device, the anchor nodes determine

their positions. GWO gives better performance in terms of less

computation time and success rate of localized nodes. It can

be combined with other heuristic algorithms for ﬁnding the

location of nodes. Deployment of sensor nodes in unreceptive

environments causes the unreliable data collection. To gain

the accurate information, anomaly detection mechanisms have

been proposed earlier [5]. In order to make decisions from

the gathered data, it is noteworthy to detect anomalies in a

sensor network. Discovering anomaly is an extensive process

to determine its variance in behavior than the expected perfor-

mance. Authors of [5] took the initiative to solve the one-class

classiﬁcation issue. The issue of anomaly detection is resolved

by One Class SVM (OCSVM). Support vector machine has

been proven to be the efﬁcient classiﬁcation method. Radial

base function can be used as kernel in OCSVM. Optimization

of hyperparameters is done using OCSVM for the core purpose

of anomaly detection. Authors of [6] have catered the fault

identiﬁcation by initially classifying the sensor data using

SVM. The sensor faults are detected using the proposed

Online Sparse Least Squares SVM (OS-LSSVM). The features

of faults are extracted using Error Correcting Code SVM

(ECOC-SVM). The initial characteristics are separated, and

the fault states are classiﬁed. ECOC-SVM and OS-LSSVM

are considered to be highly efﬁcient for real-time requirements

of fault identiﬁcation and prediction. Sensor location is a key

element that contributes in the performance of WSN because

most of the applications in wireless sensor network domain

need the known location of sensor nodes. Several optimization

algorithms have been used to reduce the localization error

of sensor nodes. Authors of [7] have used metaheuristics

to solve this optimization problem. Optimization algorithms

like, Particle Swarm Optimization (PSO), Fireﬂy Algorithm

(FA), Grey Wolf Optimization (GWO) algorithm are used to

estimate the position of sensor nodes. The localization problem

is resolved by minimizing the localization error using efﬁcient

optimization algorithms. GWO comparatively worked better

and reduced more error than other algorithms. As anchor

(nodes with known position) nodes are used to estimate the

location of other sensor nodes, transmission range should be

increased to localize more targets. However, it takes more

computation time. As sensors have less energy and their

energy depletes faster, providing a better network lifetime is

challenging in WSNs. According to [8], GWO outperforms

other optimization algorithms. It gives more accuracy and

most importantly, GWO takes less execution time in an energy

constrained environment.

III. TRANSMISSION RATE ADJUSTMENT

Congestion occurs in wireless sensor networks when the

number of incoming packets at a sensor node are more than

the number of outgoing packets. 100 nodes were randomly

deployed [2] in an area of 100m*100m with a sink and the

congestion information was stored as shown in Fig. 1. The

authors of [2] controlled the congestion on each hop. Trans-

mission rate is increased or decreased based on the channel

information of the downstream node. An awareness packet is

sent from each node to the upstream node regarding the trafﬁc

information. Here, upstream node is the one from which the

data is being received, whereas, downstream node refers to

the node that will receive the data. Normalized queue length

(4B) and congestion degree (4C) are referred as trafﬁc

loading. Based upon these two, trafﬁc loading information

is estimated at each sensor node. Normalized buffer size at

any node v is the ratio of number of packets in queue and

buffer size. Congestion degree is calculated as the ration of

average processing time of packs and the interval of arrival

time of two adjacent packets. If the current node has more

trafﬁc loading than the downstream node, then there is a need

of increased transmission rate. Clearly, if the buffer occupancy

of the current node is greater than the buffer occupancy of the

downstream node, the current node is more congested and

it should increase the data transmission rate. 4Band 4C

determine the change in buffer and congestion whereas, 4R

represents the increased or decreased data transmission rate.

The amount of data transmission is determined on the basis

of trafﬁc loading information. The packet loss or number of

retransmission of packets is determined using the values of

4B,4Cand 4R. The data transmission rate which gives

the less amount of packet loss is selected.

IV. PROPOSED TECHNIQU E

A. Transmission rate adjustment based on DE-SVM and

GWO-SVM

Support Vector Machine: Suppose we have a data space

X and we have to classify the data in two classes. We have

d1, d2 . . . , dk data points or the training points with labels

y1, y2 . . . , yk. we need to classify them in classes C or C1.

The prediction is made whether the data point d belongs to

a particular class or not. SVM can work efﬁciently on this

problem. SVM [1] is used to separate the hyperplane optimally

to classify the input data into positive or negative class. It

produces the supreme distance between the data and the plane.

A kernel function is used in non-linear classiﬁcation to map the

low-dimensional feature space classiﬁcation data into a high-

dimensional feature space. SVM [16] is a supervised learning

machine that classiﬁes the objects by ﬁnding a hyperplane. The

hyperplane segments or divides the objects and determines in

which category the object lies. Non-linear classiﬁcation is done

by changing the kernel function and generating hyperplane

lines using Gaussian Radial Basis (RBF). SVM [17] uses

different parameters like Penalty, Loss (loss function i.e. hinge

and squared hinge), Dual (for optimization problem), Tol

(for stopping criteria) and Random state (to generate random

number). We have used the data set provided by [2] of

100 randomly deployed sensor nodes and used GWO and

DE algorithms for SVM parameter tuning. The steps of the

proposed work are taken as follows:

1) For sensor nodes, retransmission values are determined

using the provided values of 4B,4Cand 4R

2) The data is divided into independent variables and

response variable that are 4B,4C,4Rand the num-

ber of retransmission packets, respectively. The (4B),

(4C) and (∆R) are used to interpret the retransmission

values.

3) 80% and 20% data is used as training and testing data.

SVM is designed for each retransmission value. Zero

retransmission data values and other data values are

labeled with 1 and -1, respectively. Five SVMs are

designed for ﬁve retransmission values.

4) Grey Wolf Optimizer is used to tune SVM parameters.

The adjusted parameters are penalty ratio (C), acceptable

error and the deviation of the gaussian kernel function.

Maximum iterations and number of search agents taken

are 50 and 5, respectively. GWO depicts the same mech-

anism as grey wolves hunting. Grey wolves always hunt

in a pack. Each pack consists of four types of wolves

that are alpha, beta, delta, and omega. Alpha wolves are

known to be the leaders, the dominant members or more

accurately the decision makers. Beta wolves support the

alpha wolves and help them in decision making. Delta

wolves follow the commands of alpha and beta. Omega

are not considered an important entity. With a good

hierarchy, each pack successfully hunts the prey. They

track the prey, encircle and then harass it and attacks

the prey when it attempts for self-defense. The pseudo

code of GWO is given below in algorithm 1.

•Social Hierarchy: Social hierarchy of grey wolves

is distinguished into alpha, beta and delta which are

considered as the best or optimum solution, second

best and third best solution, respectively. Here the

goal is to get a required solution or prey.

•Encircling Prey: It includes the encircling of a prey

for an optimal solution. The values of A and C

coefﬁcient vectors can be adjusted in order to reach

near the best agent.

•Hunting: The core of GWO algorithm is hunting.

It means to move towards the solution and up-

dating the alpha solution. With the alpha score,

beta and delta can calculate their positions. The

omega wolves are the remaining solutions and up-

date themselves in reference with alpha, beta and

delta solutions.

•Attacking Prey (exploitation): When the prey stops

moving, the wolves attack the prey to ﬁnish the

hunt. The ﬂuctuation of the coefﬁcient vector A is

decreased by a. The random value A [-2a, 2a] where

a is decreased from 2 to 0. With the operators, GWO

search agents can update their positions using alpha,

and delta positions.

•Search for prey (exploration): Random population is

generated, and the position of prey is estimated by

alpha, beta and delta wolves. The distance of solu-

tion from prey is updated. To highlight exploration

and exploitation, parameter a tends to decrease from

2 to 0.

5) Differential Evolution is also used to tune SVM pa-

rameters. The adjusted parameters are penalty ratio

(C), acceptable error and the deviation of the gaussian

kernel function. DE works the same way as GA. It

performs crossover, mutation and selection. It takes two

independent elements and accumulates the difference of

these two.The they are multiplied by the mutation factor

to generate a mutant element. The second step involves

making the trial elements same as the population rate to

perform crossover. The last step is known as selection

as it selects the elements estimated in the previous step

[18]. We have used DE for getting the suitable parameter

values of SVM. Algorithm 2 explains the DE algorithm.

Algorithm 1 Pseudo code of Grey Wolf Optimization

Require: Input:[parameters, minimum value, maximum

value]

1: Initialize the population X

2: Initialize a, A, C

3: Calculate the ﬁtness of each search agent

4: Update alpha=the best search agent

5: Update beta= the second search agent

6: Update delta=the third search agent

7: while itertaion < M axIter ation do

8: for edoach search agent

9: State Update the position of current search agent

10: end for

11: Update a, A and C

12: Calculate the ﬁtness of all search agents

13: Update alpha

14: Update beta

15: Update delta

16: itertaion + +

17: end while

18: Return the best search agent

Fig. 1 shows the system model of tuning of support vector

machine parameters using grey wolf optimization and differ-

ential evolution algorithms. The datasets [2] are divided into

train and test sets. In the train phase, GWO and DE are used to

obtain the SVM parameters. The ﬁtness value for each solution

is estimated. The optimized parameters from GWO and DE are

used to re-train the SVM. Then, the errors of classiﬁcation are

calculated which shows the amount of misclassiﬁcation made

by the proposed methods.

V. SIMULATION RESULTS

The results section consists of two subsections: At ﬁrst, the

proﬁciency of Support Vector Machine (SVM) is evaluated. In

the second section, the presented technique is compared with

other classiﬁcation techniques based on Mean Square Error

(MSE), Mean Absolute Error (MAE) and Root Mean Square

Error (RMSE). In order to evaluate the performance of the

Algorithm 2 Pseudo code of Differential Evolution

Require: Input:[parameters, No. of iterations, crossover, mu-

tation]

Initialize the population X

2: for each individual j in the population X do

Choose three members n1, n2 and n3 such that,1>

n1, n2, n3 6N

4: Create a random integer i (1,N)

while itertaion < M axIter ation do

6: for each parameter i do

Calculate the ﬁtness values of

8: all individuals

Create mutant vectors using mutation

10: strategy

Create trial vectors by recombining

12: noisy

vectors with parent vectors

14: Evaluate trial vectors with their ﬁtness

values

16: end for

Select winning vectors as individuals in the new

18: generation

itertaion + +

20: end while

end for

22: Return the best values

proposed technique,we have performed simulations in python

3.7. Speciﬁcations of the system used are: 1.61 GHz processor,

8.00 GB RAM, 2.66 GHz processor base frequency and 8 MB

cache. We have used the same data set of readings as provided

in [2].

A. Proﬁciency of GWO-SVM and DE-SVM

We have taken total data of 400 inputs for simulations. The

data used for training phase and test phase are 80% and 20%

respectively. The SVM parameters are tuned using the GWO

and DE techniques. Maximum iterations and number of search

agents taken are 50 and 5 respectively.

Fig. 2 and 4 show that the obtained results from GWO-SVM

and DE-SVM match the actual data. The training data and real

data are represented by blue and green lines, respectively. Fig.

3 and 5 display the compliance of test and real data by red and

blue lines respectively. In both of the above ﬁgures, x-axis or

horizontal axis and y-axis or vertical axis display the available

data and amount of packet loss, respectively. We can conclude

from these overlapping lines that, the presented classiﬁers

produce better results. A contingency table or confusion matrix

is also calculated in order to get a glance of predictions.

The error matrix gives the visualization of errors being made

during classiﬁcation by the classiﬁer. All correct and incorrect

predictions are speciﬁed in a matrix. The matrix consists of

rows and columns and presents the instances in predicted and

actual class. The proposed techniques are made using ﬁve

SVMs, so the proposed confusion matrix is a matrix of ﬁve

F

u

n

c

t

i

o

n

INPUT

SVM parameters

Apply GWO and DE for

parameter tuning

Train SVM

Evaluate fitness of parameters Is stopping criteria

met?

Optimized Parameters

Retrain SVM

OUTPUT

Classi ca on

Accuracy

MeanSquare

Error

YES

NO

Training

Set

All dataset

Packet loss of each

data

Training

Set

Fig. 1. System Model

Fig. 2. Compliance of data in GWO-SVM

Fig. 3. Compliance of data in GWO-SVM

rows and ﬁve columns. The advantage of using the confusion

matrix is to have a clear idea of what types of errors the

Fig. 4. Compliance of data in DE-SVM

Fig. 5. Compliance of data in DE-SVM

classiﬁcation model has made and how much data is predicted

accurately. The confusion matrices C1 and C2 of the applied

GWO-SVM and DE-SVM techniques are presented as follows:

C1 =

20 2 0 0 0

2 6 1 1 0

0 3 8 0 0

1 0 6 8 1

0 0 0 5 16

C2 =

20 2 0 0 0

3 6 1 0 0

0 3 8 0 0

1 0 6 8 1

0 0 0 6 15

The correctly predicted values are shown on the diagonals

of the matrices. The values that are predicted more than the

real data, and less than the real data are located as the upper

and lower triangular elements of the matrices respectively.

As shown in the error matrices, more than 70% data are

located on the diagonal which means more than 70% data

are accurately predicted. Upper triangular data shows higher

transmission rate in the node. To conclude, the techniques

correctly determined the amount of retransmission based on

inputs as the predicted values are mostly correct.

B. Comparison of GWO-SVM with other classiﬁers

The overall error is calculated to evaluate the quality of

the presented technique with other classiﬁcation methods like

GASVM, Naive Bayes (NB), Random forest (RF), and k-NN.

The data taken in training and testing phases are similar in all

methods.

1) Support Vector Machine using Genetic Algorithm (GA-

SVM): Genetic Algorithm (GA) [5] is a search heuristic

inspired by the biological evolution. Individuals of GA are

termed as chromosomes. Multi classiﬁcation is done using

SVM and the parameters are tuned using GA. The adjusted

parameters are penalty ratio, acceptable error in SVM and

the deviation of the Gaussian kernel function. Implementation

of GA is done using uniform crossover and mutation. The

values used for population size, crossover and mutation are

50, 0.7 and 0.3, respectively [2]. The error shows that genetic

algorithm efﬁciently adjusted the parameters and very well

classiﬁed the data. The confusion matrix C3 of GA-SVM is

given below:

C3 =

22 0 0 0 0

3 4 3 0 0

0 1 9 1 0

1 0 6 7 2

0 0 1 5 15

Fig. 6 shows that the obtained results from GA-SVM match

the actual data. The training data and real data are represented

by blue and green lines, respectively. Fig. 7 displays the com-

pliance of test and real data by red and blue lines respectively.

In both of the above ﬁgures, x-axis or horizontal axis and y-

axis or vertical axis display the available data and amount

of packet loss, respectively. The comparison of the presented

Fig. 6. Compliance of data in GA-SVM

Fig. 7. Compliance of data in GA-SVM

techniques with GA-SVM is shown in 10, 11, and 12 and from

the performance evaluation, we came to the conclusion that the

proposed DE-SVM and GWO-SVM produce less classiﬁcation

error as compared to GA-SVM.

2) Random Forest: A Random Forest (RF) [14] includes

numerous different decision trees. Each decision tree analyses

and votes on how the feature must be is classiﬁed. New items

are classiﬁed based on voting done by the trees in the forests.

Number of estimators and random state are taken as 9 and

42, respectively. MSE, MAE, and RMSE of random forest are

displayed in ﬁg. 10, 11, and 12. The bar plots display that

random forest works quite well on this data set. However,

random forest did not produce as much accurate results as

the proposed techniques. Fig. 8 displays the tree generated

using Weka tool. M5P is a well known binary regression model

in which the the last nodes produce continous attributes. A

standard deviation reduction or divergence metric is used to

construct a tree.

3) Naive Bayes: Naive Bayes [14] is an efﬁcient supervised

learning algorithm that uses conditional probabilities to predict

an outcome. It works accurate in real world scenarios. Nave

Bayes is based on statistics and assesses each feature indepen-

dently in the data set. It deals with two features independently.

In this way, a ﬁrm correlation between the factors is made.

However, we have used Gaussian Nave Bayes (GaussianNB)

and checked the performance of this classiﬁer on the given

problem. This class (GaussianNB) assumes the features to be

normally distributed. At ﬁrst, we have scaled the features and

then classiﬁed them using gaussian nave bayes. MSE, MAE

and RMSE are displayed in ﬁg. 10, 11, and 12, respectively.

Fig. 8. The Graph of M5P tree results

This classiﬁer handles features independently and assumes that

the presence of a feature is unassociated to the presence of

other features. The results proved that naive bayes does not

work well on the given dataset because provided features are

related in our scenario i.e., transmission rate is dependent upon

the the features of trafﬁc loading information.

4) k-Nearest Neighbor (k-NN): This classiﬁer is used to

measure the difference based upon a distance function. It ﬁnds

the closest neighbors of an instance and assigns a class to the

instance based on voting [6]. The results depend upon the

number of neighbors selected. We have randomly selected the

number of neighbors and number of jobs 5 and 2, respectively.

The K value impacts the accuracy of the predictions. Fig. 9

shows the error for the predicted values of test set for all the

K values between 1 and 25. The reason behind low testing

performance of k-NN is that it overﬁts the data and produce

unreliable training predictions of observations if the data is

ﬁnite.

Fig. 9. Miss classiﬁcation in k-NN

The amount of retransmission packets is obtained by using

the values of buffer occupancy ratio, congestion degree and

transmission rate. Before providing SVM with the available

data for predictive analytics, we should split the dataset into

training and testing sets. In classiﬁcation, ﬁrst the model is

Fig. 10. Classiﬁcation errors comparison

Fig. 11. Classiﬁcation errors comparison

constructed and then this model is tested. SVM works well

on small datasets, therefore, we have used it to predict the

amount of packet loss based on the available data i.e. 4B,

4Cand ∆R. Simulation results show that the proposed

technique GWO-SVM and DE-SVM outperform the other

classiﬁers. Comparative analysis of all classiﬁers is performed

taking into consideration on the same dataset, however, DE-

SVM has done classiﬁcation more accurately with less errors

than all other methods. Proposed GWO-SVM presented the

Fig. 12. Classiﬁcation errors comparison

second best results and proved to be a good classiﬁcation

method in our scenario. As compared to other classiﬁcation

methods like GA-SVM, random forest, naive bayes and k-NN,

DE is more robust, and it performs computations efﬁciently.

DE better classiﬁes the continuous data and provides results

faster whereas GE provides good enough results on discrete

problems. The comparison results given in table I concluded

that the proposed techniques GWO-SVM and DE-SVM solve

the congestion problem and adjust the rate of transmission in

a better way.

TABLE I

ERRO RS OF A LL C LAS SI FIER S

Errors GA-

SVM

Random

Forest

Nave

Bayes

k-NN GWO-

SVM

DE-

SVM

MSE 0.425 0.463 0.664 0.458 0.412 0.387

MAE 0.325 0.400 0.58 0.333 0.312 0.312

RMSE 0.425 0.46 0.96 0.458 0.412 0.387

VI. CONCLUSION AND FUTURE WO RK

Many researches have alluded the efﬁciency of SVM clas-

siﬁcation method. This study aims to control congestion in

WSNs by adjusting the transmission rate. Congestion degree

and buffer occupancy ratio for different values of transmission

rate are used to obtain the amount of retransmission pack-

ets. The congestion problem is solved using DE and GWO

algorithms. As, it is difﬁcult to classify the complex data,

SVM parameters are tuned using GWO and DE to reduce

the classiﬁcation error. The simulation results show that the

proposed approaches efﬁciently deal with the complex data

and outperforms the GA-SVM, K-NN, Naive Bayes, and

Random Forest in terms of classiﬁcation error. We aim to

apply other optimization algorithms for parameter tuning in

the future in order to get more accurate classiﬁcation. We will

also introduce and classify different type of faults to decrease

sensor failures and handle network trafﬁc appropriately and

we aim to consider more dynamic and practical test scenarios

in WSNs.

REFERENCES

[1] Shah, S.A., Nazir, B. and Khan, I.A., 2017. Congestion control algorithms

in wireless sensor networks: Trends and opportunities. Journal of King

Saud University-Computer and Information Sciences, 29(3), pp.236-245.

[2] Gholipour, M., Haghighat, A.T. and Meybodi, M.R., 2017. HopbyHop

Congestion Avoidance in wireless sensor networks based on genetic

support vector machine. Neurocomputing, 223, pp.63-76.

[3] Al-Aboody, N.A. and Al-Raweshidy, H.S., 2016, September. Grey wolf

optimization-based energy-efﬁcient routing protocol for heterogeneous

wireless sensor networks. In Computational and Business Intelligence

(ISCBI), 2016 4th International Symposium on (pp. 101-107). IEEE.

[4] Rajakumar, R., Amudhavel, J., Dhavachelvan, P. and Vengattaraman,

T., 2017. GWO-LPWSN: Grey wolf optimization algorithm for node

localization problem in wireless sensor networks. Journal of Computer

Networks and Communications, 2017.

[5] Tran, K.P. and Huong, T.T., 2017, October. Data driven hyperparameter

optimization of one-class support vector machines for anomaly detection

in wireless sensor networks. In Advanced Technologies for Communica-

tions (ATC), 2017 International Conference on (pp. 6-10). IEEE.

[6] Deng, F., Guo, S., Zhou, R. and Chen, J., 2017. Sensor multifault

diagnosis with improved support vector machines. IEEE Transactions on

Automation Science and Engineering, 14(2), pp.1053-1063.

[7] Kaur, R. and Arora, S., 2017. Nature inspired range based wireless

sensor node localization algorithms. International Journal of Interactive

Multimedia and Artiﬁcial Intelligence, 4(6), pp.7-17.

[8] Shieh, C.S., Sai, V.O., Lee, T.F., Le, Q.D. and Lin, Y.C., 2017. Node

Localization in WSN using Heuristic Optimization Approaches.

[9] Zhang, Y., Liu, Y., Chao, H.C., Zhang, Z. and Zhang, Z., 2018. Classi-

ﬁcation of Incomplete Data Based on Evidence Theory and an Extreme

Learning Machine in Wireless Sensor Networks. Sensors, 18(4), p.1046.

[10] Goeschel, K., 2016, March. Reducing false positives in intrusion de-

tection systems using data-mining techniques utilizing support vector

machines, decision trees, and naive Bayes for off-line analysis. In

SoutheastCon, 2016 (pp. 1-6). IEEE

[11] Gupta, S., Mittal, M. and Padha, A., 2017, December. Predictive Analyt-

ics of Sensor Data Based on Supervised Machine Learning Algorithms.

In 2017 International Conference on Next Generation Computing and

Information Systems (ICNGCIS) (pp. 171-176). IEEE.

[12] Yun, S., Lee, J., Chung, W., Kim, E. and Kim, S., 2009. A soft

computing approach to localization in wireless sensor networks. Expert

Systems with Applications, 36(4), pp.7552-7561.

[13] Alam, M.N., Das, B. and Pant, V., 2015. A comparative study of

metaheuristic optimization approaches for directional overcurrent relays

coordination. Electric Power Systems Research, 128, pp.39-52.

[14] Jiang, C., Zhang, H., Ren, Y., Han, Z., Chen, K.C. and Hanzo, L., 2017.

Machine learning paradigms for next-generation wireless networks. IEEE

Wireless Communications, 24(2), pp.98-105.

[15] Swain, R.R. and Khilar, P.M., 2017. Composite fault diagnosis in

wireless sensor networks using neural networks. Wireless Personal Com-

munications, 95(3), pp.2507-2548.

[16] Abuassba, A.O., Zhang, D., Luo, X., Shaheryar, A. and Ali, H., 2017.

Improving Classiﬁcation Performance through an Advanced Ensemble

Based Heterogeneous Extreme Learning Machines. Computational intel-

ligence and neuroscience, 2017.

[17] Shakeel, U., Jan, N., Qasim, U., Khan, Z.A. and Javaid, N., 2016, July.

DRADS: Depth and reliability aware delay sensitive routing protocol

for underwater WSNs. In Innovative Mobile and Internet Services in

Ubiquitous Computing (IMIS), 2016 10th International Conference on

(pp. 78-83). IEEE.

[18] Javaid, N., Majid, A., Sher, A., Khan, W. and Aalsalem, M., 2018.

Avoiding Void Holes and Collisions with Reliable and Interference-Aware

Routing in Underwater WSNs. Sensors, 18(9), p.3038.

[19] Ali, B., Sher, A., Javaid, N., Aurangzeb, K. and Haider, S.I., 2018.

Retransmission avoidance for reliable data delivery in underwater WSNs.

Sensors, 18(1), p.149.

[20] Ahmed, F., Wadud, Z., Javaid, N., Alrajeh, N., Alabed, M.S. and Qasim,

U., 2018. Mobile Sinks Assisted Geographic and Opportunistic Routing

Based Interference Avoidance for Underwater Wireless Sensor Network.

Sensors, 18(4), p.1062.

[21] Javaid, N., Hussain, S., Ahmad, A., Imran, M., Khan, A. and Guizani,

M., 2017. Region based cooperative routing in underwater wireless sensor

networks. Journal of Network and Computer Applications, 92, pp.31-41.