CHAPTER 5

TWOFOLD GRAVITATIONAL SEARCH OPTIMIZATION

BASED CLUSTERING PROTOCOL FOR WIRELESS
SENSOR NETWORKS

5.1 Introduction

Wireless Sensor Networks (WSNs) have always attracted the interests of the researchers

because of numerous applications they have in store with them. These networks are formed

with hundreds of small sensors which sense data from the environment and forward it to

centralized base station either directly or via relay nodes [148-152]. These networks are

provide support in various areas such as healthcare, agriculture, military applications, home

applications etc. However, the sensor nodes are powered by smaller batteries which become a

major constraint for these networks [153-155]. WSNs also provide backbone support to

Internet of Things (IoT) applications these days. If the constraint of the sensor nodes, i.e.

their limited battery is not taken care of, the entire structure of IoT becomes of no use as well.

In order to increase the lifetime of the sensor nodes, clustering is one of the many famous

techniques which has been practiced by the researchers in the past. These clustering

techniques have been successful in increasing the lifetime of the sensor nodes [156-160]. In

such techniques, the entire network is divided into clusters with each cluster having its own

cluster head and cluster members [161].

The concept of clustering was originated from Low Energy Adaptive Clustering Hierarchy

(LEACH) protocol and ever since researchers have come forward with numerous variations

to this protocol [162-165]. In these variations, the focus has been given on optimal selection

of cluster head for the network. While some of the approaches go ahead with selection of
cluster head according to residual energy of the node or some other parameter of the node

[166-168], the others make use of multi hop data transmission approach [169-170] among

cluster heads to prolong the lifetime of WSN. Apart from this, many bio inspired or swarm

intelligence approaches such as Ant Colony Optimization [171-173], krill herd algorithm

[174-175], whale and grey wolf optimization algorithm [176-178], gravitational search

algorithm (GSA) [179-182] have been used to optimize the performance of WSN. Apart from

this concept of Mobile Agent (MA) has also attracted interest of the researchers. MA is used

to collect data from the cluster heads to improve the network lifetime. Such approaches have

been described in [183-186] while the work related to optimizing the itinerary of the MA can

also be seen in [187-190].

Although a lot of work has been done in the recent past related to optimizing the cluster head

selection process or optimizing the data transmission process, still work involving emergency

applications has not been focused upon lately. This chapter presents two fold gravitational

search algorithm based clustering protocol for IoT application involving emergency data. The

proposed cluster protocol optimizes the cluster head selection process using GSA and data

transmission has been done using MA. Furthermore, the itinerary of the MA has also been

optimized using GSA.

5.2 Related work

The authors in [191] presented concept of threshold value of energy which has been used to

avoid rotation of cluster heads in every round. Usually, the cluster heads are rotated every

round so as to balance the load among them. However, in the presented protocol, the cluster

heads having energy more than threshold value are not rotated and kept same for the next

round. It saves the energy consumed in broadcasting of advertisement packets and thereby

increases the network lifetime. The same concept of saving the energy by avoiding re-

broadcast of advertisement packets in each round was presented in [192]. In the presented
protocol, nodes exchange certain list of information in pilot rounds initially which is used for

formation of clusters in the subsequent rounds thus avoiding broadcasting of advertisement

packets. Apart from saving the energy by avoiding the rotation of cluster heads, optimization

strategies for cluster head selection are also considered in the past research works. The

authors in [193] have presented hybrid cluster head selection method using harmony search

and firefly optimization. While energy efficient cluster heads were selected at primary stage

using harmony search algorithm, the clusters formed by them are refined using firefly

optimization. The cluster formation was optimized considering the density of the nodes,

compactness of the cluster formed and energy required in cluster formation. On the other

hand, the authors in [194] have made use of gravitational search algorithm for optimal

selection of cluster heads. The parameters considered for optimal cluster head selection were

inter cluster distance, residual energy of the node and distance from base station. Node

spacing is one of the optimization parameter explored by the authors in [195] apart from

remaining energy and total energy of the network to optimize the cluster head selection

process. Another use of algorithms namely dolphin echolocation and crow search

optimization in a hybrid way was made by the authors in [196] for selecting the cluster heads.

Since the network lifetime can also be improved by optimizing the data transmission process

from cluster heads to the base station, this can be seen in [197] where the mobile agents were

used to collect data from the cluster heads. The empower Hamilton loop based approach was

used for itinerary planning of the MA. An approach of data collection using Mobile data

collectors has been presented in [198]. The itinerary of mobile data collector has been

optimized using Particle Swarm Optimization (PSO) algorithm [199] and cluster heads

forward the data to the mobile data collector, when it reaches the anchor points, using the

space division multiple access technique. The optimization of itinerary of MA was also
explored by the authors in [200] using the fuzzy based approach. The mobile agent would

choose next node to visit based on the outcome of the fuzzy rules.

5.3 Motivation

Wireless sensor networks that provide support to Internet of Things applications which is

very prominent field of research these days. WSNs have numerous sensors powered by

smaller batteries and researchers have used clustering approaches to increase their lifetime.

Increasing the lifetime reduces the re-deployment cost of the sensors and makes IoT

applications more successful. Clustering originated from LEACH which defines threshold

value and a probability for each node to become cluster head. Ever since research has been

conducted to optimize the selection of cluster head for better network performance. In some

of the previous researches [200-205], threshold values for every node has been modified to

have better cluster head or probability of the node has been modified to optimize the cluster

head selection process. Apart from cluster head selection process, data transmission phase

(sending data from cluster heads to the base station) has also seen modifications in the recent

past. Authors in [206] proposed a clustering technique where the probability of the node to

become cluster head was modified using remaining energy of the node. Furthermore, the

authors used the Hamilton loop concept for MA to collect data from the cluster heads.

However, the authors have not focused on including other parameters, such as distance of the

node from the base station etc., to optimize the cluster head selection process. Authors in

[207] proposed the collection of data from the nodes using the MA and optimized the

itinerary for it. However, the author has not worked on allocating the source nodes to the

respective MA in a distance-efficient way. It was also observed that researchers have not

focused on specific applications of WSN or IoT that take into account emergency data sensed

by the nodes and most of the clustering protocols designed had only considered normal data

sensed by the nodes. Therefore, taking into consideration these shortcomings, the proposed
clustering protocol has been described in this chapter. The proposed protocol uses GSA to

optimize the cluster head selection process and to optimize the itinerary for the MA to collect

data from the nodes.

5.4 System Model

 In this chapter, clustering protocol for WSN has been designed which is expected to

support IoT application involving sensing of emergency data by the sensor nodes. For

example, consider an IOT application backed by the WSN deployed for forest

monitoring. The sensor network consists of sensors such as temperature sensor,

moisture sensor etc. (to monitor the environmental conditions) and sound sensors (to

monitor if someone is illegally cutting down the trees). Now sound sensors have

emergency data as compared to other sensors, and upon sensing the sound of higher

frequency (related to the tree cutting) can forward the data to the monitoring station

wherein smart alarm system (IOT device) can receive data from the sound sensors and

can trigger the alarm.

 This chapter considers the homogenous environment for the nodes in context of

energy given to them and transmission range of the nodes.

 All the nodes are randomly deployed in the network and these nodes are not mobile.

 Out of ‘N’ total nodes, ‘em’ percentage of total nodes are assumed to have emergency

data.

 The nodes in the network consume energy according to first order radio energy

dissipation model [26]. The energy consumed in transmitting and receiving a packet

of size ‘L’ bits is given by:

𝐸𝑒𝑙𝑒𝑐 ∗ 𝐿 + 𝐸𝑎𝑚𝑝 ∗ 𝐿 ∗ 𝑑 4 𝑖𝑓 𝑑 > 𝑑0

𝐸 (𝑡𝑥 ) = { (5.1)
𝐸𝑒𝑙𝑒𝑐 ∗ 𝐿 + 𝐸𝑎𝑚𝑝 ∗ 𝐿 ∗ 𝑑 4 𝑖𝑓 𝑑 ≤ 𝑑0
 𝐸 (𝑟𝑥 ) = 𝐸𝑒𝑙𝑒𝑐 ∗ 𝐿 (5.2)

Where E(tx) is the energy consumed by the node which is sending L bit packet, E(rx)

is the energy consumed by the node which is receiving it.Eelec is the energy consumed

per bit by the transmitter or the receiver,Eamp and Efsare the amplifier parameters of

transmission corresponding to the multi-path fading model and the free-space model,

respectively.

d is the Euclidean distance between two communicating nodes and d0 is the threshold

𝐸
distance which is computed by 𝑑0 = √ 𝑓𝑠⁄𝐸 .
𝑎𝑚𝑝

 The various notations and symbols used in this chapter have been shown in the table

5.1.

5.5 Two Fold Gravitation Search Optimization Based Clustering Protocol

The proposed clustering protocol exploits Gravitational Search Algorithm (GSA) twice,

hence called as Two Fold GSA), to optimize the performance of the wireless sensor network

(WSN) for Internet of Things (IOT) application involving emergency data.

The proposed protocol groups the nodes into clusters, selects an optimal cluster head from

them which then aggregates the data from the sensor nodes and forwards it to server/base

station with the help of Mobile Agents (MA). In the Two Fold GSA, the first use of GSA was

made in optimizing the cluster head selection process while the second one was made in

optimizing the itinerary of MA for data aggregation process. The protocol has two phases

namely set up phase and steady phase. While the setup phase deals with selection of cluster

heads and formation of clusters, the steady phase deals with the data aggregation from cluster

members to cluster heads and data transmission from cluster heads to base station using MAs.
 Table 5.1: Symbols and notations used

p Probability of node to become Cluster head

N Total number of nodes in the network

M*M Network size

𝑋𝑖𝑑 Random location of ith node in ‘d’ dimension

𝐸в𝑖 Bit value representing emergency/normal data

𝑁𝑒𝑖𝑖 Neighbor set for ith node

k Number of neighbors for a node

𝐸𝑖 Energy of the ith node

𝑑 Gravitational Force between the nodes iand j

𝐹𝑖,𝑗

𝐺 Gravitational constant

𝑀𝑝𝑖 Passive gravitational mass of node ‘Ni’

𝑀𝑎𝑗 Active gravitational mass of node ‘Nj’

𝐷𝑖,𝑗 Euclidean distance between the nodes iand j

d0 Threshold Distance

em Percentage of nodes with emergency data

𝐶𝐻𝑖 Cluster head set for ith node

L Packet Size

FMMA Free memory of mobile agent

5.5.1 Setup Phase

Initially every node is eligible to become cluster head such that their probability ‘p’ of

becoming cluster head is equal. The equal probability is defined in traditional clustering

protocol LEACH. However, two nodes having equal probability may have different

characteristics. For example, a node may be located nearer to the base station while the other

one may be located far away, a node may have dense neighborhood while the other one may

have sparse neighborhood (this factor influences the size of cluster formed by the nodes), a

node may have some emergency data to forward to server/base station while the other one

may have sensed only normal data which is not as important as emergency data. Taking into

consideration such concerns, it is unfair to assign equal probability to the nodes to become

cluster head. Therefore, the probability of the node must be adjusted in a way that the node

having better characteristics has higher chance of becoming cluster head. Consequently, in

the proposed protocol, the probability ‘p’ has been adjusted according to the acceleration

parameter defined by GSA.

5.5.1.1 Use of GSA in selection of optimal cluster head

According to GSA, every eligible node which can become cluster head acts as an agent.

Therefore, in network having set of ‘N’ randomly distributed nodes in the area of ‘M*M’ sq.

units, we have ‘N’ agents in the initialization stage of GSA. These agents have different set of

characteristics which are defined below:

 Let 𝑋𝑑𝑖 = {𝑥𝑑1 , 𝑥𝑑2 , … … . , 𝑥𝑑𝑁 } represent their random locations in ′𝑑′ dimension search

space. Since this chapter focuses on 2-D network, therefore we have set of X and Y

coordinates for each node which will represent their location.

 𝐸в𝑖 = {𝐸в1 , 𝐸в2 , … … . 𝐸в𝑁 } represents the bit value for the emergency data sensed by

𝑖𝑡ℎ node such that

 0 𝑖𝑓 𝑛𝑜𝑑𝑒 𝑁𝑖 ℎ𝑎𝑠 𝑛𝑜𝑟𝑚𝑎𝑙 𝑑𝑎𝑡𝑎
𝐸в𝑖 = { (5.3)
1 𝑖𝑓 𝑛𝑜𝑑𝑒 𝑁𝑖 ℎ𝑎𝑠 𝑒𝑚𝑒𝑟𝑔𝑒𝑛𝑐𝑦 𝑑𝑎𝑡𝑎

 𝑁𝑒𝑖𝑖 = {𝑁𝑒𝑖1 , 𝑁𝑒𝑖2 , … … . 𝑁𝑒𝑖𝑘 } represents the set of ‘k’neighbors for node ‘Ni’ such at

any point of time Euclidean distance between the node and its neighbor is subjected to

condition:

𝐷𝑖,𝑛𝑒𝑖 (𝑡) < 𝑑0 (5.4)

where d0 is the threshold distance and 𝐸𝑖 = {𝐸1 , 𝐸2 , … … . 𝐸𝑁 } represents the energy

levels of the nodes.

GSA is based on Newton’s first and second Law of Gravity. While the first law states that

each particle has certain amount of attraction (known as Gravitational Force) with another

particle, the second law defines the acceleration produced in the particle which depends on

the force exerted on it by another particle and its own mass.

Extending the law to the sensor network, each node ‘Ni’ must attract another node ‘Nj’ in the

neighborhood. The force exerted between the nodes is directly proportional to the product of

their Gravitational Mass and inversely proportional to square of distance between them. The

force 𝐹𝑑𝑖,𝑗 (𝑡) acting in the ‘dth’ dimension with which node ‘Nj’ pulls or pushes the node ‘Ni’

at any point of time ‘t’ is given as:

𝑀𝑝𝑖 (𝑡)∗𝑀𝑎𝑗 (𝑡)

𝐹𝑑𝑖,𝑗 (𝑡) = 𝐺 𝐷𝑖,𝑗 (𝑡)
(𝑥𝑑𝑗 (𝑡) − 𝑥𝑑𝑖 (𝑡))(5.5)

Where 𝐺 is the gravitational constant, 𝑀𝑝𝑖 is the passive gravitational mass of node ‘Ni’, 𝑀𝑎𝑗

is the active gravitational mass of node ‘Nj’, 𝑥𝑑𝑗 (𝑡) and 𝑥𝑑𝑖 (𝑡) represents the position of the

node ‘Nj’ and ‘Ni’ and 𝐷𝑖,𝑗 (𝑡) is the Euclidean distance between the nodes. Since the network
is two dimensional, therefore the nodes will exert force over the neighbors in two dimensions,

thus we have:

𝑀𝑝𝑖 (𝑡)∗𝑀𝑎𝑗 (𝑡)

𝐹𝑥𝑖,𝑗 (𝑡) = 𝐺 𝐷𝑖,𝑗 (𝑡)
(𝑋𝑗 (𝑡) − 𝑋𝑖 (𝑡)) (5.6)

𝑦 𝑀𝑝𝑖 (𝑡)∗𝑀𝑎𝑗 (𝑡)

𝐹𝑖,𝑗 (𝑡) = 𝐺 𝐷𝑖,𝑗 (𝑡)
(𝑌𝑗 (𝑡) − 𝑌𝑖 (𝑡)) (5.7)

𝑦
Where 𝐹𝑥𝑖,𝑗 (𝑡), 𝐹𝑖,𝑗 (𝑡) represents the force acting in the X and Y dimension over the node.

This can be seen in the figure below:

Figure 5.1: Force acting between two nodes

Since a node has ‘k’ number of neighbors, therefore all these neighbors will exert a force over

the node ‘Ni’ such that the total force exerted over the node ‘Ni’ is randomly weighted sum of

all the forces and is given by:

𝐹𝑑𝑖 (𝑡) = ∑𝑘𝑗=1,𝑗≠𝑖 𝑟𝑎𝑛𝑑𝑗 ∗ 𝐹𝑑𝑖,𝑗 (𝑡) (5.8)

Where 𝑟𝑎𝑛𝑑𝑗 is the random number in the range of (0, 1).

The gravitational masses for the nodes are computed as:

𝑚𝑖 (𝑡)
𝑀𝑝𝑖 = 𝑀𝑎𝑗 = ∑𝑘 (5.9)
𝑗=1 𝑚𝑗 (𝑡)

Such that

𝑓𝑖𝑡 −𝑤𝑜𝑟𝑠𝑡(𝑡)
𝑖
𝑚𝑖 (𝑡) = 𝐵𝑒𝑠𝑡(𝑡)−𝑤𝑜𝑟𝑠𝑡(𝑡) (5.10)

Where 𝑓𝑖𝑡𝑖 is the fitness function of the 𝑖𝑡ℎ node, 𝑤𝑜𝑟𝑠𝑡 is the maximum value of the fitness

function of some node in the neighborhood of ‘k’neighbors when the minimization problem

is taken into account and vice-versa for the maximization problem. Similarly, 𝐵𝑒𝑠𝑡 is the

minimum value of the fitness function of some node in the neighborhood of ‘k’neighbors in

case of minimization problem and vice-versa. Thus for maximization problem we have:

𝐵𝑒𝑠𝑡(𝑡) = 𝑚𝑎𝑥{𝑓𝑖𝑡𝑗 (𝑡)} 𝑗 ∈ {1,2, … … . , 𝑘} (5.11)

𝑊𝑜𝑟𝑠𝑡(𝑡) = 𝑚𝑖𝑛{𝑓𝑖𝑡𝑗 (𝑡)} 𝑗 ∈ {1,2, … … . , 𝑘} (5.12)

The fitness function of the nodes is dependent upon following three sub-routines or three sub-

fitness functions:
 Number of emergency nodes in the neighborhood: The first sub-fitness function is

dependent on the number of neighbors having the emergency data to forward to base

station. If a node, having more number of neighbors that has emergency data, is

elected as cluster head then probability of emergency data getting lost from the

network will be reduced significantly. This is because in such a cluster formed, more

nodes with emergency data will forward data to cluster head over shorter distance as

compared to the scenario when they are in direct communication with base station and

their distance is significantly more. As the energy consumed is directly proportional

to square of distance between two communicating entities, therefore the energy

consumed by the nodes possessing emergency data gets reduced considerably.

Consequently, the fitness function is computed as:

∑ 𝑁𝑒𝑖𝑖 |𝐸в𝑁𝑒𝑖 =1
𝑓𝑖𝑡𝑒𝑚𝑒𝑟𝑔𝑒𝑛𝑐𝑦
𝑖
= 𝑒𝑚∗𝑁
𝑖
(5.13)

 Remaining energy of the node: This is another factor that needs to be maximized to

have optimal network performance. If a node (fulfilling condition of the first sub-

fitness function) has formed cluster such the majority of the cluster members have

emergency data, then its remaining energy needs to be on the higher side too for

reliable data transfer to base station. Otherwise, in the contrary scenario, all the data

accumulated by the low-energy cluster head will be lost. Therefore, the fitness

function is computed as:

𝑒𝑛𝑒𝑟𝑔𝑦 𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙 𝐸𝑛𝑒𝑟𝑔𝑦

𝑓𝑖𝑡𝑖 = (5.14)
𝐸𝑖

 Communication cost: This is another imperative parameter which needs to be

considered for optimal cluster head election. This parameter defines how much energy

cluster members have to spend to forward data to cluster head, i.e. intra cluster
 communication cost, and to forward data to the base station, i.e. direct communication

cost. Even if the node fulfills the first two sub-fitness functions, it cannot afford to

have more communication cost which is again deadly for the network’s performance.

The communication cost needs to minimized. Therefore, the fitness function is

computed as:
2 2
∑𝑘𝑗=1(𝐸𝑒𝑙𝑒𝑐 ∗ 𝐿 + 𝐸𝑓𝑠 ∗ 𝐿 ∗ 𝐷 𝑖,𝑗 ) (𝐸𝑒𝑙𝑒𝑐 ∗ 𝐿 + 𝐸𝑓𝑠 ∗ 𝐿 ∗ 𝐷 𝑖,𝐵𝑆 )
𝑓𝑖𝑡𝑐𝑜𝑠𝑡
𝑖
= +
∑𝑘𝑗=1(𝑁𝑒𝑖𝑗 ∗ 𝐸𝑗 ) 𝐸𝑖

(5.15)

Thus, final fitness function of the node is:

𝑓𝑖𝑡𝑖 = 𝛼 ∗ 𝑓𝑖𝑡𝑒𝑚𝑒𝑟𝑔𝑒𝑛𝑐𝑦
𝑖
+ 𝛽 ∗ 𝑓𝑖𝑡𝑒𝑛𝑒𝑟𝑔𝑦
𝑖
+ 𝛾 ∗ (1 − 𝑓𝑖𝑡𝑐𝑜𝑠𝑡
𝑖
) (5.16)

Where 𝛼, 𝛽, 𝛾 are constants having the sum equal to 1. The optimal cluster head would be the

node that has formed cluster with more number of neighbors having emergency data, the

node must have more remaining energy with it and the energy cost of communication should

be less too. Once the masses have been computed, the acceleration can be computed as:

𝐹𝑑 (𝑡)
𝑎𝑐𝑐𝑑𝑖 = 𝑀𝑖 (𝑡) (5.17)
𝑖

The nodes having higher value of the mass tend to experience least acceleration and are

considered fit. Therefore, the probability ‘p’ of the node to become cluster head is adjusted

as:

𝑝
𝑝𝑎𝑑𝑗 = 𝑎𝑐𝑐𝑖 𝑑 (5.18)
𝑖

After adjusting for the probabilities of the nodes, each node generates a random number and

compares it with threshold value. If the random number is less than threshold value, the node

becomes cluster head for the current round provided that it has not been cluster head in last

‘1/p’ rounds.
 𝑝𝑎𝑑𝑗 (𝑟)
1 ; if node(i) ∈ G(r)
𝑇ℎ(𝑖 ) = {1−𝑝𝑎𝑑𝑗 (𝑟)(𝑟 𝑚𝑜𝑑 𝑝𝑎𝑑𝑗(𝑟)
) (5.19)
0; 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Where ‘r’ is the current round, G(r) is the set of the nodes which have not become as cluster

head in last ‘1/p’rounds. All the elected cluster heads broadcast the advertisement packet in

their communication range to their neighbors. All the neighboring nodes that receive the

packet decide to join the cluster head and form cluster with them. However, a node can

receive the advertisement packet from more than one cluster head, in such a case the nodes

join the cluster head for which the variance of the distance (between the node and cluster

head) is least. Let us assume that 𝐶𝐻𝑖 = {𝐶𝐻1 , 𝐶𝐻2 , … … . 𝐶𝐻𝑞 } represents set of ‘q’ cluster

heads from which the node ‘i’ has received the advertisement packet such that 𝐷𝑖,𝐶𝐻 =

{𝐷𝑖,1 , 𝐷𝑖,2 , … … . 𝐷𝑖,𝑞 } is the set of distance between them. The node makes the cluster head as

parent cluster head subjected to condition that:

2 2
(𝐷𝑖,𝐶𝐻 −𝐴)
𝐶𝐻𝑝𝑎𝑟𝑒𝑛𝑡 = 𝐶𝐻𝑖 |Var𝑖 = ( √∑𝑞𝑖=1 𝑞
) = 𝑚𝑖𝑛 (Var𝑖 ) (5.20)

𝑞 𝐷𝑖,𝐶𝐻
Where 𝐴 = ∑𝑖=1 is the average value of the distance of the node from all the cluster
𝑞

heads and Var𝑖 represents variance of the distance. Thus, the formation of the clusters marks

the end of setup phase.

5.5.2 Steady Phase:

In this, the data transmission is executed between the base station, cluster heads and cluster

members. After the formation of the clusters, the cluster heads broadcast Time Division

Multiple Access (TDMA) schedule to the member nodes for the data transmission. All the

cluster members aggregate the sensed data at their respective cluster head in the assigned

time slots. After aggregating the data from cluster members, the cluster heads need to
forward the data to the base station. For this purpose, the MAs have been used to collect data

from the cluster heads. The most imperative step is the planning of itinerary of the mobile

agent which will decide the order of visit of MA at the cluster heads to collect data from

them. The proposed protocol makes use of GSA second time here to decide optimal itinerary

for the MA.

5.5.2.1 Use of GSA for optimal itinerary planning of MA

In order to plan the itinerary for the MA in an optimal way, various steps such as deciding the

number of MAs for a particular number of source nodes (the nodes which have data to send

to base station, cluster heads in our case), allocation of the source nodes to the MA needs to

be done first.

 Deciding the number of MAs: In this step, the total data that needs to be collected and

free memory of the MA decides how many number of MAs are required for data

collection. Mathematically, it is computed as:

𝑝
∑𝑖=1 𝐿𝐶𝐻
𝑖
𝑅𝑀𝐴 = (5.21)
𝐹𝑀𝑀𝐴

Where RMA is the required number of mobile agents, FMMA is the free memory of

the MA and 𝐿𝐶𝐻𝑖 is the data packet size with the ithcluster head (source node). Let

𝑀𝐴 = {𝑀𝐴1 , 𝑀𝐴2 , … … . 𝑀𝐴𝑅𝑀𝐴 } be the set of MAs that will be collecting data from
𝑝
the p number of cluster heads. Therefore, 𝑅𝑀𝐴
number of cluster heads will be

allocated to every MA.

 Allocating the source nodes to MAs: In traditional approaches, the allocation has been

done using Greatest Information in Greatest Memory (GIGM) algorithm [25].

According to this, the source node, having greatest (highest amount) data to forward,

is allocated to the MA that has greatest memory free/available with it. Since, in this

approach the source nodes are allocated based on the amount of data available with
 them therefore there is always an uncertainty about the distance among the allocated

nodes. For instance, the source nodes allocated to a single MA may be located far

away from each other which increases the tour length of the MA and will

consequently lead to higher energy consumption.

In order to tackle this issue, the k-means clustering approach has been used with an

objective of reducing the distance among the assigned source nodes. This is done by

allocating source nodes nearer to each other to same MA instead of allocating the

nodes to the MA based on the greatest data.

In the proposed allocation strategy, a random source node is first given as input to the

k-means clustering algorithm with an intent of forming ‘RMA’ clusters. Then a cluster

is formed with nearest node to the randomly chosen source node. In the next iteration,

another node is added to the cluster such that it is closest to the centroid of the cluster

formed. The iteration continues till number of members in the cluster is not more than
𝑝
𝑅𝑀𝐴
. Thus, using k-means clustering algorithm, the source nodes can be allocated to

the MA in a distance-efficient way. At the end of this step, each MA will have to visit

set of cluster heads and its itinerary is represented by:

𝐼𝑀𝐴 = {𝐶𝐻1 , 𝐶𝐻2 , … … . 𝐶𝐻 𝑝 } (5.22)

𝑅𝑀𝐴

𝑝
 Optimal Itinerary planning for MA: For the set of 𝑅𝑀𝐴
source nodes, the itinerary

planning is done using GSA algorithm. As defined earlier, the agent having lesser

acceleration is considered more efficient. The fitness functions defined earlier for

optimal cluster head election are changed in this step and rest of the computation of

the masses and gravitational force remains same. This is again modelled as

maximization problem where the priority to visit the nodes by the MA is to be

maximized. The fitness function of the source nodes are dependent on three sub-

routines or sub-fitness functions again which are explained below:

o Type of data with node: Since a node may have some emergency data, it needs

to be sent to base station with more priority as compared to other nodes having

normal data. Therefore, the fitness function is evaluated according to the

𝐸в𝑖 value of the node. This needs to be maximized.

𝑓𝑖𝑡𝑑𝑎𝑡𝑎_𝑡𝑦𝑝𝑒
𝑖
= 𝐸в𝑖 (5.23)

o Energy of the node: The nodes having the emergency data have more priority

than the nodes having the normal data. However, among the nodes with

emergency data, the ones having the minimum remaining energy left with

them or the ones having higher energy cost of communication with the base

station have more priority. Thus, this fitness function encompasses the

remaining energy of the node as well as energy cost of communication with

the base station.

𝐸𝑒𝑙𝑒𝑐 ∗𝐿+ 𝐸𝑓𝑠 ∗𝐿∗𝐷 2𝑖,𝐵𝑆

𝑓𝑖𝑡𝑒𝑛𝑒𝑟𝑔𝑦
𝑖
= 𝐸𝑖
(5.24)

o Distance with the base station: A node, having the emergency data and having

minimum remaining energy, if located far away from the base station is again

a more priority for the MA to visit as early as possible. Therefore, this fitness

function is related to distance of the node from the base station.

𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 2 2
𝑓𝑖𝑡𝑖 = √(𝑋1 − 𝐵𝑋) + (𝑌1 − 𝐵𝑌) (5.25)
 Where X1 and Y1 are coordinates of the node and BX and BY are coordinates

of the base station. If the distance is more, then priority to visit the node

increases.

The final fitness function is again the weighted sum of the three sub-fitness functions

and can be computed as:

𝑓𝑖𝑡𝑖 = 𝛼 ∗ 𝑓𝑖𝑡𝑑𝑎𝑡𝑎_𝑡𝑦𝑝𝑒
𝑖
+ 𝛽 ∗ 𝑓𝑖𝑡𝑒𝑛𝑒𝑟𝑔𝑦
𝑖
+ 𝛾 ∗ 𝑓𝑖𝑡𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒
𝑖
(5.26)

Where 𝛼, 𝛽, 𝛾 are constants having the sum equal to 1. After computing the fitness

function for the nodes, the force exerted over the node, masses of the node as well as

acceleration for the node is computed. The MA visits the node having the minimum

acceleration first and the node having the maximum acceleration at last. Therefore, if

𝑎𝑐𝑐𝐶𝐻𝑖 represents the computed acceleration, then the optimal itinerary can be

represented as:

𝐼𝑜𝑝𝑡−𝑀𝐴 = {𝑎𝑐𝑐𝐶𝐻1 , 𝑎𝑐𝑐𝐶𝐻2 , … … . 𝑎𝑐𝑐𝐶𝐻 𝑝 } |𝑎𝑐𝑐𝐶𝐻1 > 𝑎𝑐𝑐𝐶𝐻2 > ⋯ … . >

𝑅𝑀𝐴

𝑎𝑐𝑐𝐶𝐻 𝑝 (5.27)
𝑅𝑀𝐴

Thus, when MA visits the cluster heads, they aggregate their data at the MA which

carries it towards the base station. This marks the end of steady phase as well as one

round. At the start of next round, same process is repeated and new cluster heads are

chosen to balance the load among them.

5.6 Results and Discussion

The proposed TF-GSA clustering protocol was implemented in MATLAB environment. A

network of 200 nodes, that were randomly deployed, was created in area of 200*200 sq.
units. The base station was considered to be located in the center of the network. Other

simulation parameters used for the simulation have been defined in table 6.1. The

performance of proposed clustering protocol was analyzed in terms of network lifetime and

energy consumption. The comparison was done with the protocol described in [147] to check

effectiveness of the TF-GSA.

Table 5.2: Simulation Parameters Set 1

Parameter Value

Number of nodes 200

Deployment Type Random

BS location 100,100

Network area 200 * 200 sq. units

Packet size 2000 bits

Initial Energy 0.5 Joules

Eelec 50 nJ/bit

Efs 10 pJ/bit/m2

Eamp 0.0013 pJ/bit/m4

Eda 5 nJ/bit

Threshold Distance, d0 87.7 m

The figure 5.1 shows the variation of number of alive nodes versus number of rounds for

existing and proposed clustering protocol. The number of alive nodes is an apt measurement

of the network lifetime. The round when the first node dies defines the network stability

period while the round at which the last node dies defines the network lifetime. It was

observed that the network stability period was least for LEACH-GA clustering protocol (400

rounds for LEACH-GA) followed by CBRP (600 rounds) and EEUC (720 rounds). The EHL

had second highest network stability period of 960 rounds whereas the proposed clustering

protocol had network stability period of 1100 rounds. After the first node dies, the network

experiences steep decline in number of alive nodes. However, the proposed protocol manages

to achieve somewhat gradual decline in the number of alive nodes which extended its

network lifetime. EHL optimized the cluster head selection process taking into account

remaining energy of the node only. It ignored the distance between the chosen cluster head

and the base station which can be important factor in deciding energy consumption of the

cluster head to forward data to the base station. While for the proposed clustering protocol,

GSA was used to optimally select the cluster heads taking into focus parameters such as

energy cost of communication, remaining energy of the node as well as number of neighbors

with emergency data. This enables to choose the cluster head that is having higher energy and

has least cost of communication. Also, the use of MA enables multi hop communication

between the nodes which is more energy efficient than single hop communication.

Furthermore, better network lifetime (which is because of more number of alive nodes) also

helps us to infer that more data can be sent to the base station. This includes the normal data

as well as emergency data. Consequently, we can infer from the better results that the

proposed clustering protocol provides more reliability in terms of sending more emergency
 Figure 5.2: Comparison of Network Lifetime
data to the base station. Thus, for IoT application involving emergency data, the proposed

clustering protocol promises to deliver better performance.

Table 5.3: Network Stability Period for Different Protocols

Protocol Network Stability Period

 LEACH-GA 400 rounds

CBRP 600 rounds

EEUC 720 rounds

EHL 960 rounds

TF-GSA 1100 rounds

Table 5.4: Network Lifetime for Different Protocols

Protocol Network Stability Period

LEACH-GA 790 rounds

CBRP 935 rounds

EEUC 970 rounds

EHL 1180 rounds

TF-GSA 2230 rounds

The figure 6.2 shows the energy consumed in the network against number of rounds. Initially

the network was supplied with 100 Joules of energy (0.5 Joules per node). As the number of

rounds progress, the energy gets consumed at uniform rate and when all the nodes die out, the

energy of the network gets fully consumed. The proposed protocol however showed gradual

increase in the rate of energy consumption as compared to other protocols that had steep

energy consumption rate. This is due to better optimized selection of cluster head using GSA

and multi hop data communication process between cluster heads using MA.

Furthermore, second set of simulation parameters was used to analyze and check the

effectiveness of the proposed optimal itinerary planning technique against other state-of-art

itinerary planning techniques for MA defined in [99]. These simulation parameters have been
 Figure 5.3: Comparison of Energy Consumption

shown in table 6.4. For this simulation, 800 nodes were randomly deployed in the network of

bigger size having dimensions 1000*500 sq. units. The number of source nodes which have

data to forward to base station were varied from 10-80 and the nodes were supplied with

energy of 2 Joules each. The performance of the network was analyzed in terms of success

rate of MA trip and task energy consumption.

 Table 5.5: Simulation Parameters Set 2

Parameter Value

Number of nodes 800

BS location 500,250

Network area 1000 * 500 sq. units

Packet size 2048 bits

Initial Energy 2.0 Joules

Number of source nodes 10-80

MA parameters Value

MA processing code 1024 bits

Raw data reduction ratio 0.8

Aggregation ratio 0.9

The figure 5.3 shows the variation of the success rate of the mobile agent’s data collection

trip. Success rate is defined as ratio of number of MAs received at the base station to the

number of MAs dispatched by the base station for data collection. MA is dispatched from the

base station to collect data from the source nodes. If the MA is not received back at the base

station, it is considered as failure. It happens when the node does not have enough energy to

send the MA to next node in the trip. It can be seen from the figure that when the number of
 Figure 5.4: Success rate of Mobile Agent Trip

nodes were less, almost all of the techniques have 100% success rate meaning that all the

mobile agents which were dispatched to gather data from the network came back

successfully. However, as the number of nodes increase, the success rate reduces. This

happens because more number of source nodes would mean more data to be collected from

the network. It eventually increases the tour length as well resulting in the increased energy

consumption and reduced success rate of MA trip. The proposed GSA based scheme however

showed higher success rate than the other schemes even when the number of source nodes

were more. The reason for better success rate of TF-GSA is attributed to two factors. The first

factor is related to the use of k-means clustering algorithm to assign the nearby source nodes

to particular MA instead of using GIGM algorithm (which is used by existing techniques).

This leads to allocation of closely packed group of nodes to MA which reduces the intra-

cluster/intra-group tour length for the MA. The second factor is the use of GSA for

optimizing the itinerary of MA which considers remaining energy of the nodes, their

communication cost and distance to the base station to check the priority of the nodes to be

visited by MA. The nodes with higher priority are visited first which also includes the priority

for the nodes having emergency data. The higher success rate of MA as compared to other

techniques justifies the proposed itinerary planning technique.

The figure 5.4 shows the energy consumed by the mobile agent to gather the data from the

source nodes against varying number of source nodes. The energy consumed by the mobile

agent was least for the proposed scheme as compared to other schemes which indicate better

optimized itinerary for the mobile agent. Energy consumption depends is directly

proportional to the square of distance between two communicating entities. The distance

between the source nodes is reduced by using better assignment strategy of allocating the

source nodes to the particular MA (k-means approach). Furthermore, the GSA has been used

to optimize the itinerary of the MA. It takes into account the distance of the nodes from the

base station and remaining energy plus energy cost of communication with the base station as

well. This leads to better itinerary for the MA thus leading to lesser task energy consumption.