A Biologically-Inspired Clustering Protocol for Wireless Sensor Networks

8/6/2019 A Biologically-Inspired Clustering Protocol for Wireless Sensor Networks

1/16

A biologically-inspired clustering protocol for wireless sensor networks

S. Selvakennedy a,*, S. Sinnappan b, Yi Shang c

a School of Information Technologies J12, University of Sydney, NSW 2006, Australiab School of Economics & Information Systems, University of Wollongong, Wollongong, NSW 2522, Australia

c Department of Computer Science, University of Missouri-Columbia, Columbia, MO 65211, USA

Available online 15 June 2007

Abstract

Lately, wireless sensor networks are garnering a lot of interests, as it is feasible to deploy them in many ad hoc scenarios such as forearthquake monitoring, tsunami monitoring and battlefield surveillance. As sensor nodes may be deployed in hostile areas, these battery-powered nodes are mostly expected to operate for a relatively long period. Clustering is an approach actively pursued by many groups inrealizing more scalable data gathering and routing. However, it is rather challenging to form an appropriate number of clusters with wellbalanced memberships. To this end, we propose a novel application of collective social agents to guide the formation of these clusters. Inorder to counter the usual problems of such meta-heuristics, we propose a novel atypical application that allows our protocol to convergefast with very limited overhead. An analysis is performed to determine the optimal number of clusters necessary to achieve the highestenergy efficiency. In order to allow for a realistic evaluation, a comprehensive simulator involving critical components of the communi-cation stack is used. Our protocol is found to ensure a good distribution of clusterheads through a totally distributed approach. To quan-tify certain clustering properties, we also introduced two fitness metrics that could be used to benchmark different clustering algorithms.Crown copyright 2007 Published by Elsevier B.V. All rights reserved.

Keywords: Clustering; Routing; Biologically-inspired; Meta-heuristic; Simulations

1. Introduction

Wireless sensor technology is evolving rather rapidly invarious aspects, as it draws significant interests from differ-ent research communities such as wireless networking,embedded system, database, data mining and distributedcomputing groups. For instance, in terms of the sensornode hardware, the Mica2 mote has roughly eight timesthe memory and communication bandwidth as its prede-

cessor, the Rene mote, developed in 1999 for the samepower budget [1]. These sensor nodes coupled with wirelesscommunication capability have found use in many applica-tions such as earthquake monitoring, target tracking andsurveillance, structural monitoring and precision agricul-ture. The nodes are typically stationary due to their uniqueapplication needs, substantially more resource constrainedand more densely deployed than mobile ad hoc networks

(MANETs). Even though, there have been significantadvances in recent years to improve wireless sensor nodescapabilities, more energy-efficient solutions are requiredwithin the communication stack and middleware for theconservation of the battery power. As the entire wirelesssensor network (WSN) utility relies on its useful lifetime,these solutions however might come at the tradeoff againstthe traditional performance measures such as packet deliv-ery ratio, mean latency and throughput. Within the com-

munication stack, an approach that is likely to succeed inthis regard is the use of a hierarchical structure for routing[2,3].

Clustering with data aggregation is an important tech-nique in this direction, and it makes the tradeoff betweenenergy efficiency and data resolution. Even though manyprotocols have been proposed in the literature to minimizeenergy dissipation on the forwarding paths, some nodesmay still be drained quickly. By employing a dynamic clus-tering technique, these nodes could lose their popularity ascertain roles are rotated dynamically [3]. Various clustering

0140-3664/$ - see front matter Crown copyright 2007 Published by Elsevier B.V. All rights reserved.

doi:10.1016/j.comcom.2007.05.010

* Corresponding author. Tel.: +61 2 9351 4494.E-mail address: [email protected] (S. Selvakennedy).

www.elsevier.com/locate/comcom

Computer Communications 30 (2007) 27862801
mailto:[email protected]:[email protected]


2/16

techniques in different context have been proposed. Mostalgorithms aim at generating the minimum number of clus-ters and transmission distance. These algorithms also dis-tinguish themselves by how the clusterheads (CHs) areelected.

Another important design issue to consider is network

reliability or robustness. Since the wireless medium andthe sensor nodes themselves are considered less reliable rel-ative to the wired technology and its tethered nodes, adopt-ing a deterministic approach in the solution design, andassuming only average case performance would be disas-trous in practice. To this end, social insect swarm behaviormay provide an ideal model for the design of such less con-trollable systems. To our knowledge, very few researchershave considered or adopted such biologically-inspiredapproaches for the WSN design. However, a number ofrecent works has been based on different swarm behaviorsin the design of routing protocols for MANETs. As thereare many important similarities between these two ad hoc

technologies, we believe building on these knowledge maybe useful for WSNs. Most of these swarm-based routingalgorithms are simple yet robust as well as adaptive totopological changes. However, such algorithms cannotestablish the shortest or appropriate paths before a suffi-cient number of agents is flooded [4].

In this paper, we propose to biologically inspire theclustering approach, whereby the network is clusteredaround certain nodes deemed biologically fit. A prelimin-ary version of this algorithm is introduced in [5]. Accord-ing to this approach, when a node has a special agent, itself-elects itself to become a CH. Such election obviates

the need to maintain many state variables. A fixed numberof such agents (i.e. in biological term, the swarm size) areused to ensure that a certain number of clusters areformed throughout the network useful life. This numberis derived analytically to minimize energy dissipationthrough data aggregation. We derive the optimal numberof clusters assuming a Voronoi tessellation incorporatingboth the influence of intra- and inter-traffic. Based on thisoptimal number and our approach of clustering, we couldguarantee that the network always operates close to opti-mal once system convergence is achieved. In addition, wecould ensure that the CHs are well distributed across thenetwork to promote load balancing. Thus, using thisapproach, we address a multi-objective optimization prob-lem of energy efficiency and load balancing. To quantifythe clustering properties, we also introduce two fitnessmetrics that will be useful for comparing any clusteringalgorithms.

To the best of our knowledge, this is the first attempt toadapt a biologically-inspired technique to the resource-starved device class in a unique approach. To counter thegeneral problems of such biologically-inspired meta-heuris-tics whereby they requires significant number of agents anda long time for convergence, we only adopt its certain use-ful features, and modify the mechanics of the agents. The

main features that we still adopt are:

The notion of biological fitness of a solution; The use of a collection of simple agents interacting

locally. Interaction is not achieved through explicit mes-sage exchanges, but implicitly by leaving some simplestate at the nodes.

In a typical swarm intelligence use, an agent has to bepresent at an object (here, a node) to affect its biological fit-ness. In our approach, an agent is allowed to influence evenfrom remote. We demonstrate later how this unique adap-tation has reduced the need for a large agent populationand also reduced the convergence time in our atypicalapplication. Also, our dynamic clustering protocol exhibitsfault-tolerance in the face of node failures comparedagainst fixed or deterministic clustering approaches in theliterature. Even when a node fails, this protocol could stillachieve the desired properties of good CH distribution andenergy efficiency.

The rest of the paper is organized as follows: Section 2

presents the perspective of this area of research. Variousclustering and biologically-inspired algorithms proposedin the literature are discussed there. The basic swarm prin-ciples are presented in Section 3. In Section 4, the details ofour protocol are described. The comprehensive simulatorused to experiment with this algorithm is described in Sec-tion 5. Many different experiments comparing our protocolagainst well-known proposals and their correspondingresults are presented and analyzed in Section 6. The mainfindings of this work are summarized in the final sectionwith some directions for further work.

2. Related work

Intense research in the field of sensor network technol-ogy in recent years has fueled further development inmicro-sensor technology and low-power analog/digitalelectronics. Some of the ongoing challenging issues facingresearchers are low-power signal processing architecture,routing protocol and data gathering strategy, data storageand indexing, data aggregation, low-power security proto-col, and localization. It is realized that the approach that islikely to succeed to provide a scalable and energy-efficientsolution is of a hierarchical structure. To this end, variousclustering algorithms have been proposed in different con-text. Initially, these algorithms focused on the connectivityproblem [68] but later energy-efficiency was more of inter-est in wireless ad hoc and sensor networks [915]. However,almost all focuses on reducing the number of clustersformed, which may not necessarily entail minimum energydissipation.

Generally, clustering algorithms segment a network intonon-overlapping clusters comprising a CH each. Non-CHnodes transmit sensed data to CHs, where the sensed datacould be aggregated as these signals could be sufficientlycorrelated due to the nodes spatial proximity, and trans-mitted to the base-station (aka the sink). Clustering algo-

rithms may be distinguished by the way the CHs are

S. Selvakennedy et al. / Computer Communications 30 (2007) 27862801 2787


3/16

elected. The Linked Cluster Algorithm (LCA) [8] selects aCH based on the highest id among all nodes within one-hop. This is enhanced by LCA2 [16] that selects the nodewith the lowest id among all nodes that is neither a CHnor is one-hop of the previously selected CHs. In [17],the authors developed a similar distributed algorithm to

LCA2, which identifies the CH by choosing the node withthe highest degree. Other algorithms such as the Distrib-uted Cluster Algorithm (DCA) [18] and Weighted Cluster-ing Algorithm (WCA) [19] rely on weights to select CHs.

Load balancing heuristics are proposed in [10] and [20].In [20], the proposed clustering algorithm is designed toensure each cluster has equal number of nodes while keep-ing minimal distance between nodes and their CH. Theseload-balanced algorithms focus mainly on balancingthe intra-cluster traffic load without due consideration tothe external traffic. The maxmin d-cluster algorithm isproposed to achieve better load balancing among CHs aswell as to reduce the number of CHs as compared to

LCA or LCA2 [11]. It generates d-hop clusters with arun-time of O(d) rounds. In [12], some clustering algo-rithms are proposed to maximize the network lifetime byvarying the cluster size and the duration of a node beingnominated as a CH based on the assumption that the loca-tions are known a priori. These algorithms ignore inter-cluster traffic, and also need the recognition of the wholenetwork topology, which may not be feasible in manycases. The proposal in [21] incorporates the effect ofinter-cluster traffic in the determination of the optimal sinkplacement that maximizes the network topological lifetime.In [22], a fixed clustering algorithm that performs energy

load balancing to improve network lifetime is proposed.It also takes into consideration the interaction betweenclustering and routing. Two schemes were introduced.The first scheme finds the optimal cluster size and location,whereas the second allows a CH to probabilistically chooseto either relay the traffic to the next-hop or to deliver itdirectly to the sink. It assumes a heterogeneous network,where the CH nodes have larger resources than the others.

The Low-Energy Adaptive Clustering Hierarchy(LEACH) algorithm [13] and its related extensions[14,15,23] use probabilistic self-election, where each sensornode has a probability p of becoming a CH in each round.It guarantees that every node becomes a CH only once in 1/p rounds. Such role rotation aims to distribute the energyusage for a more load-balanced operation. However,LEACH assumes all nodes are able to reach the sinkdirectly, and requires position knowledge to perform a pre-cise transceiver power control. Some of these algorithmswere designed to generate stable clusters in environmentswith mobile nodes. In a typical WSN, the sensor nodesare stationary, and the instability of clusters due to mobil-ity of these nodes may not be an issue. Drawing from thewealth of clustering proposals for both MANETs andWSNs, The Hybrid Energy-Efficient Distributed (HEED)clustering is proposed in [3] to adopt real-valued weight-

based clustering. It uses residual energy as the primary

clustering parameter to probabilistically elect a numberof tentative CHs, and then advertises to their neighborsof their intentions to become CHs. Such messages includea secondary cost measure that is a function of neighborproximity or cluster density. This secondary cost is mainlyused to avoid elected CHs to be in range of each other, and

to guide the regular nodes in choosing the best cluster to join. This proposal is generally able to achieve a goodCH distribution across the network.

Another crucial design aspect of WSNs to consider is thenetwork reliability and fault-tolerance. It has been demon-strated in different context that the collective behavior ofsocial insects has many attractive features, not the leastrobustness and reliability through redundancy. However,only a handful WSN proposals have been inspired by nat-ure. Due to some parallels to MANETs, we will considersome biologically-inspired algorithms proposed for thisdomain. The first MANET routing algorithm based onthe ant-colony principles is the Ant-colony-based Routing

Algorithm (ARA) [24] derived from AntNet [25], an algo-rithm for wired networks. It exploited the pheromone layingbehavior of ants.

Pheromone is a quality metric indicating the goodness ofa path. Although pheromone evaporates over time, subse-quent ants leave additional pheromone and thus reinforcethe path. Ants gradually establish the shortest pathbetween food and their nest in a fully distributed andautonomous manner. Ants are flooded towards destina-tions while establishing the reverse paths to the source.The fact of gradual decay of pheromone introduces a formof a negative feedback to prevent old routes from remain-

ing in the forwarding tables when routes fall out of favorwith ants. As the number of ants that completes journeysto food in a given time is larger on a shorter path thanon a longer path, a shorter path accumulates more phero-mone and attracts more ants. The shortest paths becomepreferable, and most ants use them. However, longer pathsare not entirely lost as some ants may still maintain suchroutes. Routing schemes based on such ant-colony behav-ior is both robust and adaptable. When the shortest routeis lost due to some event, the longer routes provide alterna-tive options.

Another MANET routing algorithm that inspired bytermite is Termite [26]. In this approach, no exclusiveagents are used though, as the agents are piggybacked ontodata packets. These packets follow the pheromone trail fortheir destinations, and leave new pheromone for theirsources. In [27], routing is established based on a swarmof different types of honeybees. This algorithm, termedBeeAdHoc, is a hybrid routing approach where it couldeither be proactive or reactive. It is shown that BeeAdHocachieves better energy savings while still maintaining per-formance in terms of the traditional metrics against DSR,AODV and DSDV.

As the number of nodes grows, the number of agentsrequired to establish the routing infrastructure may

explode. A way to overcome the overhead explosion and

2788 S. Selvakennedy et al. / Computer Communications 30 (2007) 27862801


4/16

attain scalability is by using the hierarchical routingapproach. For MANETs, an adaptive Swarm-based Dis-tributed Routing (adaptive-SDR) [28] and Mobile AntsBased Routing (MABR) [29] are proposed, whereby bothschemes consider dividing the network into small clusters/zones, and then perform intra- and inter-routing. To our

knowledge, the only work to have considered swarm intelli-gence in the wireless sensor context is [30]. They imple-mented a distributed network of mobile sensor nodes andcontrolled the nodes physical movements using swarm prin-ciples. It was demonstrated that the swarm behavior couldbe used to ensure safe separation between the agents andcoverage efficiency while enforcing a level of cohesion thatmaintains a level of connectivity between the mobile agents.These nodes however are less resource constrained than typ-ical mote-class devices, and the overhead due the swarmintelligence meta-heuristic was not a major concern there.

We propose the use of both a hierarchical structure andthe biological inspiration to realize a scalable and robust

data routing strategy. To appreciate this integration, thebasic principles of swarm intelligence are presented next,and then followed by our proposed protocol.

3. Swarm principles

Swarms are useful in many optimization problems. Aswarm of agents is used in a stochastic algorithm to obtainnear optimum solutions to complex, non-linear optimiza-tion problems [31,32]. In this work, we consider the useof a swarm of agents and its control aspects. Swarm con-trol issues are important as the swarm behavior could be

used to establish logical network topology. In the descrip-tion of his boids model, Reynolds [33] presents the classicswarm control theory. There are three basic controllingbehaviors that govern movements of agents within theswarm. Kadrovach and Lamont [30] have summarizedthese behaviors as shown in Table 1.

When agents migrate in a swarm, every agent mustensure that they do not collide with one another. Also, inorder to ensure the best survival of an agent, it should stayas close as possible to the others. This implies that an agentshould match its velocity with neighboring agents to keepabreast. Any swarm behavior is solely based on locally

observable phenomena, and is reflected in Table 1 by theadjective nearby. The integration of these behaviors resultsin a stable swarm formation, where every agent is at leastsome minimum distance from others. In social insects,many sophisticated community behaviors emerge fromthe interaction of individuals where each insect carriesout simple tasks. Some known collective behaviors are for-

aging, nest construction, thermoregulation, brood sortingand cemetery formation [34]. These collective behaviorsof social insects have inspired computer scientists to repli-cate them as they exhibits many attractive features, such asrobustness and reliability through redundancy [4,34].

In this paper, we propose a biologically-inspired cluster-

ing protocol that uses a collection of agents. We exploit onthe first two swarm principles, namely separation andalignment, through pheromone control to achieve a stableand near uniform distribution of the CH nodes. Moreover,it is observed in [35] that the algorithm converges faster toan optimal or near optimal solution when pheromone isalso reduced drastically from those elements that makeup the worst solution in each iteration. Thus, subsequentants are discouraged from returning to poorer solutionsseen in the past. This constitutes the simplest way to imple-ment anti-pheromone where the evaporation of pheromoneis simulated by a significant reduction in existing phero-mone levels. The integrated use of these separate compo-

nents as part of the proposal is described in-depth further.

4. The T-ANT protocol

Our main design goal is to consistently form an optimalnumber of clusters with a good CH distribution for a load-balanced and energy efficient network operation. Each reg-ular node is mapped to exactly one cluster. Accordingly, weenumerate the requirements for the proposed designstrategy:

1. The protocol should be completely distributed. Each

node only makes a local decision on its role.2. The CH election should be completed in constant time.3. The protocol should be efficient in terms of processing

complexity and message exchange.4. The elected CHs should be well distributed.

4.1. Overview of the T-ANT protocol

As T-ANT is a dynamic clustering protocol, its opera-tion is divided into rounds. Each round comprises a clustersetup phase and a steady state phase. A number of timersare utilized to control the operations for optimal perfor-mance. However, if time synchronization is absent, thisprotocol will still function albeit at suboptimal perfor-mance. During the cluster setup phase, CHs are electedand clusters are formed around them. To avoid the main-tenance of many state variables as in many current cluster-ing proposals, we utilize social agents in a unique way tocontrol CH election. (Just for simplicity of reference, weterm them ants; these ants however do not behave as antsin the ant-colony optimization technique.) A node with anant becomes a CH, whereas the others would choose to jointhe best cluster in range. As we reveal further, ourapproach represents an atypical use of the swarm intelli-

gence approach. During the steady state phase, the cyclic

Table 1Agent behaviors in a swarm

Behavior Description

Separation Avoid collisions with nearby agents.Alignment Attempt to match velocity with nearby agents.

Cohesion Attempt to stay close to nearby agents.



5/16

process of data collection from members, data aggregationand transfer to the sink occurs at fixed intervals. Since antsplay a critical role in this protocol, we need to determine anappropriate swarm size and the way they are introducedinto the network. This issue is addressed through the antrelease component used during network initialization.

The details of all aspects of the protocol are individuallydescribed further with appropriate analyses wherenecessary.

4.2. The ant release component

During network initialization, the sink releases a fixednumber of ants (i.e. control messages) into the network.When the sink releases an ant, it chooses one of its neigh-bors at random. The ant makes a random walk into thenetwork as deep as restricted by its time-to-live (TTL)value. Assuming the terrain is a square M M and nodeswith radio range r, TTL is initially fixed at

dM= ffiffiffi2p re. Thisrepresents half of the diagonal terrain length (in hops) to

facilitate the ants to wander away from each other. Toavoid attraction between ants, they do not leave any pher-omone trail at this stage. Before releasing the next ant, thesink waits for a time proportional to one-hop delay (s) anda random offset to ensure its subsequent transmission doesnot self-interfere. When an ant arrives at a node, the nextstop is again randomly chosen (excluding the sender) ifTTL has not expired. If TTL expires, the ant remains atthis node. However, if the final ant location overlaps withanother ant, the latter ant must find another location.Algorithm 1 summarizes this initialization phase. Beyond

this initialization phase, ants only migrate in single-hopsat the start of each round (i.e. TTL = 1).

Algorithm 1. The Ant Release Protocol

Sink:Repeat

Choose a random neighbor (node i) to release an ant

Send the ant to node i

Wait for s(1 + rand), where 0 < r and< 1Until all ants are released

Other nodes:If an ant wanders to node i

Decrement its TTL

If TTL > 0Choose a random neighbor, j

Send the ant to node j

ElseIf TTL == 0 and node i already has an antPick a random neighbor

Send the ant to it

Else

Store the ant

End

End

Deciding on the appropriate swarm size is crucial for T-ANTs optimal performance. The analysis goal here is toorganize nodes into clusters for optimal data aggregationto minimize overall energy dissipation. Towards this end,an analysis to determine the appropriate swarm size is per-formed. The following assumptions are made for this

analysis:

The nodes are randomly scattered in a two-dimensionalplane and follows a homogeneous spatial Poisson pro-cess with k intensity [15].

All nodes in the network are homogeneous with radiorange r.

The communication from each node follows isotropicpropagation model.

The energy needed for the transmission of one bit ofdata from node u to node v is the same as to transmitfrom v to u (i.e. a symmetric channel).

The sink is located the center of the terrain (an alternatederivation is indicated for corner locations).

The overall idea of the derivation of this optimal systemparameter value is to define a function for the energy usedin the network to disseminate information to the sink.

As per the assumptions, the nodes are distributedaccording to a homogeneous spatial Poisson process. Thenumber of nodes in a square area of side M is a Poissonrandom variable, Nwith mean kA where A = M M. Letsassume that for a particular realization of the process, thereare n nodes in this area. If a node requires an ant to becomea CH (the details of CH election is explained in the next

subsection), and there are p% ants as to the number ofnodes, there will be np nodes elected as CHs. Also, theCHs and regular nodes are distributed as per independenthomogeneous spatial Poisson processes P1 and P0 of inten-sity k1 = pk and k0 = (1 pk), respectively.

Using the ideas in stochastic geometry, each node joinsthe cluster of the closest CH to form a Voronoi tessellation[36]. The plane divides into zones called Voronoi cells, witheach cell corresponding to a P1 process point termed itsnucleus. IfNv is the random variable representing the num-ber of P0 process points in each Voronoi cell and Lv is thetotal length of all segments connecting the P0 process

points to the nucleus in a Voronoi cell, then based on theresults in [23]:

ENvjN n % ENv k0k1

1 pp

1

ELvjN n % ELv k02k

321

1 p2p

32

ffiffiffik

p 2

Now, to derive the energy dissipation, the free space (d2

path loss) channel model is used. [For other path loss expo-nents (i.e. 36), it follows a similar derivation, and has aconstant effect on the optimal number of CHs.] To transmitan l-bit packet a distance r (i.e. its radio range), the radio

expends:



6/16

ETxl; d lEelec lefsr2 3

where Eelec is the electronic energy that depends on factorslike digital coding, modulation, filtering and spreading ofthe signal, and efsr

2 is the amplifier energy that dependson the distance and the acceptable bit-error rate. As to re-

ceive a packet, the radio expends:

ERxl; d lEelec 4

The T-ANT protocol guarantees that the number of clus-ters per round is always np. The dissipated energy by thenodes can be analytically estimated using the computa-tion and communication energy models. Each CH dissi-pates energy receiving signals from its members,aggregating the signals and transmitting the aggregatesignal to the sink. Since a CH could be located at any(x,y) point on the terrain with uniform intensity, theprobability density function of its location is constant

(1/M2). The transmission to the sink may also be multi-hop. We could estimate the average distance to the sink(dToS) by integrating the distance function over the areaas follows:

EdToS Zy

Zx

ffiffiffiffiffiffiffiffiffiffiffiffiffiffix2 y2

p 1M2

dxdy

ffiffiffiffiffiffiffiffi

1=6p

M

5

Limits of the definite integral are [M/2, M/2] for both xand y. If we however assume the sink is located at acorner (say bottom left origin), range of this integral is

changed to [0,M] giving E[dToS] ffiffiffiffiffiffiffiffi2=3p M. Accordingly,the average hop to the sink is E[dToS]/r. Now, let C1 rep-resent the energy spent by a CH node during a singleround:

EC1jN n ENvlEelec ENv 1lEDA


1=6p

M

r2lEelec lefsr2 6

We assume lossy data aggregation is performed at the CHwith the energy for aggregation is EDA.

As for each non-CH node, it only needs to transmit its

data to the CH once during a data interval. Let C2 repre-sent the energy used by each non-CH node:

EC2jN n lEelec lefs ELvENv 2

7

This allows us to determine the energy spent in a cluster(C3) during each round as:

EC3jN n EC1jN n EC2jN n ENv 8

Since there are np clusters in the network, we can now de-rive the total energy usage. Let Crepresent the total energy

spent in the system, then:

ECjN n npEC3jN n nl 21 pEelec EDA


1=6p

Mp

r2Eelec efsr2 efs1 p

4pk

#9

Removing the conditioning on N yields:

EC EECjN n

ENl 21 pEelec EDA ffiffiffiffiffiffiffiffi

1=6p

Mp


4pk

" #

kM2l 21 pEelec EDA ffiffiffiffiffiffiffiffi

1=6p

Mp


4pk

" #

10E[C] is minimized by a value of p that is a solution of:

efs

4p2k 2Eelec

ffiffiffiffiffiffiffiffi1=6

pM

r2Eelec efsr2

0 11

Eq. (11) has two roots where only one is positive. The sec-ond derivative of the above equation is also positive forthis root, hence minimizing the total energy spent. Thisonly positive root of (11) is given by:

popt 1

2

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiefs

k

ffiffiffiffiffi1=6

pM

r2Eelec efsr2 2Eelec

!vuuut 12

The above expression gives the optimal percentage of antsrequired to achieve the optimal clustering performance interms of energy dissipation.

4.3. The clustering algorithm

Any clustering algorithm comprises two components,namely CH election and cluster formation. Both compo-nents are invoked during the cluster setup phase ofT-ANT, whereby the phase is activated through the CS_Timer.When this timer expires, a node checks to see whether itpossesses an ant. If the node has an ant, it is guaranteedto become a CH. As there are only a fixed number of antsin the network, a fixed number of CHs are formed. Sincethere is no looping statement as a function of the numberof nodes, it is trivially evident that the election terminates

in constant time. This CH election approach has a verysmall constant time complexity as opposed to other pro-posals in the literature.

Observation 1. The CH election process terminates in O(1)time (c.f. requirement 2).

The cluster formation is triggered by these CHs,whereby they advertise to the neighbors by broadcastingan advertisement (ADV) message with their node id. Uponreceiving such message, a regular node records the CH idand the number of ADV messages received thus far. At thisstage, the most critical difference between our biologically-

inspired approach against other typical swarm intelligence



7/16

applications occurs. When regular nodes receive an ADVmessage from a CH, we allow this message to affect thepheromone level at these nodes. Here, an ant does not needto be at a node to leave a pheromone trail, which is con-trary to its typical application. As such, the influence ofan ant is broader and quicker than their typical implemen-

tation. This atypical property of our approach has contrib-uted to a significant reduction in the required number ofants for the system behavioral convergence, and thus over-comes a common problem highlighted in many biologi-cally-inspired proposals.

The actual clustering process happens once anothertimer expires. A node decides to join a cluster when itsJOIN_Timer expires. The node computes its pheromonelevel based on the number of CHs (nc) in its neighborhoodand its normalized residual energy. The pheromone func-tion (pi) is based on the forwarding probability formulaused in the uniform ant routing algorithm [37], butexpanded as:

pi pi1 Dp

1 Dp 13

where Dp is given by:

Dp k EresiEmax

1n2c

14

Eresi is a nodes residual energy, Emax is the reference max-imum battery energy and k is the learning rate of the algo-rithm ( = 0.1). This pheromone function is adopted toachieve dual clustering objectives, namely load balancingas well as energy efficiency. The effect of quadratic drop

in pheromone with the number of CHs in range not onlypromotes a good distribution of CHs, but also indirectlyaffects the load balancing objective. A regular node choosesthe nearest cluster to join by sending a JOIN message withits id, the selected CH id and its pheromone level. When aCH receives JOIN messages, it finds the member with thehighest pheromone level to attract its ant for the nextround. It is obvious that the amount of state required forthis clustering is small compared against the stochastic orweight-based clustering algorithms proposed in theliterature.

Observation 2. T-ANT is completely distributed (c.f.

requirement 1). A node locally decides to become a CH ifan ant wanders to it or may join a cluster.

Lemma 1. T-ANT has a worst-case processing time com-

plexity of O(n) per node, where n is the number of nodes in

the network (c.f. requirement 3).

Proof. For a homogeneous spatial Poisson process, theprocess intensity is uniform. Thus, the average nodedegree is pr2n/M2, where r is the radio range and M isthe side of the square region. Generally r


8/16

If node i is not a CH and this CH k is nearer

Select CH k as the best CH

End

End

/* Only for nodes with ADV */When JOIN_Timer expires at node i

Compute the pheromone level, pCreate a JOIN message (CH_id, p)Send JOIN to the selected CH

End

Steady State:When DATA_Timer expires at node i

Capture sensory value, val

If node i is a CH

Wait for data messages from all members

Aggregate data signals

Create a data message (sink_id, val)Send to sink

Else

If node i belongs to a cluster

/* for covered nodes */Create a data message (CH_id, val)Send to CH

Else

/* for uncovered nodes */Create a data message (sink_id, val)Forward to sink

End

End

End/* only for CH nodes */When ANT_Timer expires

Pick the best neighbor, j

Send ant to node j

End

Lemma 3. The probability of two CH nodes in each others

radio range is at least inverse of n, implying the CHs are

mostly well distributed (c.f. requirement 4).

Proof. Without loss of generality, lets assume that thereare two ants in a network of n nodes. In the worst-case,two CHs may be formed in almost complete overlappingranges as shown in Fig. 1, especially during network initial-ization when the ants just follow a random walk. At thisstage, the normalized residual energy is fairly uniformacross the network (it does not need to be exactly same).Thus, the only factor that influences the pheromone levelof the cluster members is the number of CHs (nc) in range.The nodes in the shaded area will have a quadratic in nclower pheromone than the rest. This is due to the quadraticdrop introduced into the pheromone function [see Eq.

(14)]:

Dp$ 1n2c

pi $ pi11=n2

c11=n2c $ 1n2c

At the next cluster setup phase, the ants would only wanderto anyof their neighbors outside the shaded area (e.g. ant awanders to any node in A \ B 0), as their pheromone level ishigher. These ants would only return to any nodes in theshaded area only when all other nodes (in the order of n)are visited. This implies that in a cycle ofn visits, only oncean ant comes in range of another ant giving a probability1/n. By induction, a network with more ants will exhibita similar behavior. Moreover, when more than two antscome into range, the repulsion effect is even greater. Thus,CHs are mostly well distributed. h

Theorem 1. The T-ANT protocol promotes load balancing.

Proof. In this dynamic protocol, the role of CH is contin-uously shared among the nodes. With Lemma 3, the electedCHs are mostly well distributed throughout the networklifetime. As each regular node chooses to associate to itsnearest CH, this results in similar cluster sizes as well.Therefore, T-ANT achieves load balancing. h

It is possible that the wandering ants may die due to theenvironment uncertainty or any node failure. When thereare lesser ants than the optimal number, the network oper-ation would still proceed as before albeit suboptimally.Due to the robustness of the proposed protocol, theremaining CHs would strive to share the aggregation task.If there is any uncovered node, it would to resort to direct(multihop) transmissions towards the sink. The event ofmost ants dying simultaneously has a probability in theorder of all nodes dying at about the same time, as the antsare mostly well distributed. Thus, such an event is onlylikely when the network is nearing the end of its lifetime.To limit the impact of isolated node failures, we couldintroduce a timer on the ants. When this timer expires, cur-

rent ants are flushed out and the sink re-releases the same

abA

B

Fig. 1. An example to illustrate the worst-case scenario of closelyoverlapping CH nodes. Sets A and B represent nodes in range of CH aand b, respectively. The shaded area represents nodes with significantlylower pheromone level than the rest.



9/16

optimal number of ants to restart the above process. Fur-ther robustness and reliability study on this issue is leftfor future study.

4.4. The clustering properties

The use of a specific swarm size as derived earlier guar-antees the network mostly operates with the optimal num-ber of clusters compared to a stochastic approach that mayhave fewer or more than the optimal number at any time.The given pheromone expression guides the evolution ofthe swarm to achieve the separation behavior among antsin the swarm, which is one the swarm principles highlightedin Section 3. Separation may not be achieved immediatelyupon the random release of these ants by the sink. How-ever, it is found empirically that separation is attainedrather quickly within a small constant number of rounds.To quantify this phenomenon, the CH election fitness met-ric is introduced. It is defined as follows:

S 1nc

Xnci1

ni 15

where nc is the number of CH nodes and ni is the number ofADVs seen by CHi. By normalizing against numberof CHs, it captures the average number CHs in the rangeof another. This metric will have a small value for a net-work with well distributed CHs. For instance, a value oftwo represents on average two CHs in the range of anotherCH. Another desirable swarm behavior is alignment. In ourcontext, the number of members served by a CH is used torepresent alignment. When the swarm evolves to achieve

separation, alignment is also achieved as a side-benefit.The clustering fitness metric A is defined to represent align-ment as follows:

A ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1

nc

Xnci1

mi l2s

16

where mi is the number of members of CH i and l* is theideal average number of nodes served by a CH in a perfectuniform CH distribution. l* is computed as n/nc. Thus,this clustering fitness metric is the standard deviation ofthe clusters membership against the ideal average. A smal-ler metric value implies that the membership varianceamong the clusters is small thereby the clusters are wellbalanced.

5. The simulation framework

We evaluated the performance of the T-ANT clusteringusing a discrete-event simulator. To enable a comprehen-sive study, the effects of both routing and MAC protocolsare also integrated. In the description of the simulator, weassume that each sensor node is aware of:

1. Its neighbors (even if it changes) due to the occasional

beaconing by the sink and the cluster setup phase; and

2. The network is synchronized by means of any time syn-chronization protocols [42].

The simulator is developed in C++ and adopts theobject-oriented approach to allow natural mapping to areal sensor network. The radio model is currently assumed

to follow an isotropic propagation. As for the MACchoice, we adopted the CSMA protocol due to its simplic-ity and its promise of scalability [39,43]. However, astraightforward application of this protocol in a converge-cast scenario is a recipe for failure. In the periodic monitor-ing type application, when the sensor DATA_Timerexpires, all nodes capture their sensory value and convertto digital via an analog-to-digital converter (ADC) linkedto the sensing hardware. Assuming a time-synchronizedclustered network, all nodes simultaneously generate thesensory message for transmission towards their CHs. Ifno precautions were taken, their transmissions would inter-fere resulting in many collisions and retransmissions. This

many-to-one transmission is known as convergecast. Inorder to reduce collisions, Huang and Zhang [39] proposedthat a node should delay its transmission relative to its dis-tance to the hop destination (h) and the node density. Asthere is likely to be many nodes at each hop, an additionalrandom offset is also included to further reduce the colli-sion probability. We chose to adopt a similar temporalcoordination with respect to a nodes CH. However, itwas modified to be less conservative to reflect the smallerscope of a cluster rather than an entire network as in[39]. Accordingly, the random wait time function (T) isgiven as:

Th 1 rand sh 17

where randis a uniformly distributed random number from0 to 1 and s is the average one-hop delay.

At the network layer, we adopted a simple routingmechanism in Greedy Routing Scheme (GRS) [44] to con-trol the networks forwarding behavior. The forwardingobjective is to minimize the number of hops between thesink and the other nodes. To establish this minimum hoprouting tree, the sink occasionally broadcasts a beacon mes-sage with a hop count, which is initialized to zero. To over-come the problem of asymmetrical links that may beprevalent in WSNs [45], the sink could broadcast at apower level lower than the maximum level of a regularnode, a move similar to Hes proposal [46]. This is possiblefor a sensor node with tunable transmit power as in Mica2motes. Upon receiving the beacon, each node records thesender id, increments the hop count by one, and thenrebroadcasts at a power level below its maximum level. Anode only rebroadcasts if the new hop count is smaller thanits stored value. Since we are focusing on a stationaryscenario, the sink node only needs to perform occasionalbeaconing to avoid significant overhead. This forwardingrule establishes a minimum hop tree rooted at the sink.

Finally, the T-ANT protocol is implemented between the



10/16

application and the network layer, and thus, the overallsystem framework is as shown in Fig. 2.

Based on the given simulation framework, we investi-gate T-ANTs performance against LEACH, HEED anda flat strategy (i.e. the application sits directly on GRS).Since LEACH cannot be applied directly to a multihopnetwork, we modified this algorithm to use the GRS rout-ing protocol to forward messages whenever the destinationis not within the radio range. As for HEED, even thoughthis algorithm is able to achieve a good CH distributionin general compared to others in the literature, the use oftentative CHs that do not finally become CHs leaves someregular nodes uncovered (i.e. no CH in range to join).According to the proposed HEED implementation [3],such uncovered nodes are forced to become CHs. In our

view, this move goes against its main design goal of achiev-ing a good CH distribution. Such forced CHs may be inrange of other CHs, and they are mostly single-node clus-ters. The issue of having many single-node clusters wasaddressed by increasing the radio range of the nodes in[3]. We, however, feel that instead of forcing these nodesto become CHs, they should just be left uncovered. Onlythe nodes that did not receive any messages at all shouldbe made CHs. Such a move has a notional structurechange, but it is operationally the same. Thus, for ourcomparative study here, we adopt this suggestedimplementation.

The results from the comparison and other evaluationsare presented further.

6. Results and discussions

For these simulation experiments, we assumed that thereare 100 sensor nodes distributed randomly in a squareM M region with M= 100 m. The transceiver energyparameters are set as: Eelec = 50 nJ/bit and efs = 10 pJ/bit/m2. The energy for data aggregation is set toEDA = 5 nJ/bit per signal [13]. The control and data mes-sage sizes are fixed at 30 bytes, and sensory data is gener-

ated at 2-s interval. Each CH node retains its CH status

for 20 s. The anti-pheromone rate is fixed at 0.1. Unlessotherwise stated, all the following investigations adoptthese values as their system parameters as summarized inTable 2. For all simulation results in this paper, each exper-iment is repeated 10 times and a 95% confidence interval isobtained.

For the purpose of benchmarking our protocol againstothers in the literature, we chose to compare against twowell-known proposals, LEACH and a more recent one,HEED, as indicated in the previous section. These twoprotocols belong to two different approaches of cluster-ing; the former is a stochastic clustering scheme andmakes a clever use of dynamic thresholding, whereasthe latter adopts a widely used approach in the MANETenvironment, namely weight-based clustering for ensuringgood CH distribution. The key design goals of these pro-tocols are the same as our proposal. Both protocols aregood baseline for comparison due to the following

features:

Clustering is distributed and entirely based on localinformation.

Clusters are disjoint. They are designed to achieve energy efficiency through

load balancing. Each employs a different mode of CH election.

For the HEED protocol implementation, we used thevalues suggested in [3] for its parameters, and chose touse node degree as the secondary parameter for clustering,as it results in the best load-balanced performance. In thissection, a comprehensive evaluation of T-ANT is provided.The followings are the different aspects of study performed:

The number of clusters required for optimal T-ANTperformance.

The clustering properties of the clustered formations. The distribution of residual energy of nodes. Topology formation as the system evolves towards

convergence. Network lifetime.

These are the performance metrics utilized in our

investigations:

Application

T-ANT

GRS

CSMA

Radio

ADC

Sensor

Fig. 2. Simulation framework involving the core components of acommunication stack.

Table 2The system parameters

Parameter Value

Field dimension, M 100 mMessage size 30 bytesSensor DATA_Timer 2 sCS_Timer 20 s

Anti-pheromone rate, b 0.1Electronics energy, Eelec 50 nJ/bitAmplifier energy, efs 10 pJ/bit/m

2

Data aggregation energy, EDA 5 nJ/bit/signalNode radio range 30 m



11/16

Clustering fitness: This is based on the fitness metricgiven as Eq. (16). It represents the fitness of the clusterformation in terms of membership.

CH election fitness: This is based on the fitness metricgiven in Eq. (15). It represents the fitness of all theelected CH nodes in terms of their distribution.

Energy per round: This metric represents the energy dis-sipated by all nodes in a round of data collection. Network lifetime: This metric represents the time period

from the instant the network is deployed to the momentwhen the first node runs out of energy.

6.1. T-ANT characteristics

Since T-ANTs performance depends on the number ofants used to form the clusters, the analysis results are usedto determine this optimal number. These results are verifiedusing simulation. Fig. 3 depicts the comparison betweenthose results. Even though the absolute energy values aredifferent, it is obvious that both curves suggest a very sim-ilar pattern. The discrepancy is mainly due to the analysisconsidering the energy usage of T-ANT alone, whereasthe simulator considers the overhead of the entire stack.Both results concur that the optimal number of ants forthis scenario is nine. This optimal number of ants, whichdetermines the number of clusters, is attainable using Eq.(12). When there are fewer or more clusters in the network,the energy dissipation worsens. When there are fewer clus-ters, there will be many cases of uncovered nodes. Thesenodes will resort to sending their messages, possibly multi-

hop, to the sink. However, when there is more than theoptimal number, the overhead of managing these clusterswill outweighs their benefit. Thus, it is necessary to identifythe optimal number of ants to achieve T-ANTs best per-formance based on the deployment parameters.

In order to observe the generality of the above finding,we used Eq. (12) to determine the optimal number of clus-ters for different terrain areas as well as number of nodes.In Fig. 4, a wireframe 3D plot is given to depict how theoptimal number varies with the two factors. For a fixed

area, the optimal number of clusters varies in O ffiffiffinp

,

where n is the number of nodes. As for a fixed number ofnodes, the optimal number of clusters is linear of terrainarea.

6.2. Clustering properties

In this study, we characterize the clustering properties ofT-ANT against LEACH and HEED in terms of clusteringand CH election fitness. Fig. 5 depicts clustering fitness ofthese protocols at different simulation time. This fitnessmetric represents the alignment property, which indicateshow well aligned the cluster sizes are. Both T-ANT and

HEED mostly exhibit small standard deviation values.This implies each cluster has similar number of members.This behavior is only achievable if the CHs are always welldistributed. As claimed in [3], we verified that HEED isable to achieve this property with our proposed modifica-tion. As for LEACH, the fitness value varies quite unpre-dictably. This is mainly due to its stochastic electionnature, where the number of clusters formed at each roundchanges quite significantly. Moreover, it is also possiblethat the elected CHs may even be clumped. When the

36

36.5

37

37.5

38

38.5

39

39.5

40

4 8 10 12 14 16 18 20

Number of ants

EnergyperRound(mJ)

2.66

2.67

2.68

2.69

2.7

2.71

2.72

2.73

EnergyperRound(mJ)

Sim

Math

6

Fig. 3. Average energy dissipated against the number of ants for T-ANT.The simulation results (i.e. dots) are plotted against the left y-axis, and the

analysis results (i.e. line) are against right y-axis.

50

150

250

350

450

550

650

750

850

950

2500

10000

22500

40000

0

10

20

30

40

50

60

OptimalNumberofClusters

Number of nodes Terrain

Area(m2)

Fig. 4. The optimal number of clusters [using Eq. (12)] against the numberof nodes and the terrain area. The number of nodes is in range [50,1000],whereas area size is in [2500,40,000].

0

5

10

15

20

25

30

35

40

45

50

0 200 400 600 800 1000

Time (s)

ClusteringFitness

T-ANT

HEED

LEACH

Fig. 5. The clustering fitness at different simulation time for T-ANT,

LEACH and HEED.



12/16

CHs are clumped, the disparity among clusters is large interms of their membership. At other times, the number ofclusters may be substantially lesser than necessary toachieve complete network coverage. Many nodes will not

be covered and the few clusters may have too many mem-bers far from the ideal mean. The two outlying LEACHpoints on Fig. 5 represent such a scenario.

In Fig. 6, the CH election fitness is depicted for theseprotocols. This fitness metric captures how well the electedCHs are distributed. It represents the average number ofCHs in the range of another. A rather consistent behavioras above is obtained. Both T-ANT and HEED have smallervalues with HEED showing the least CH overlaps with theproposed modification. Due to its biological inspiration,T-ANT has a higher value initially before the systemachieves convergence. As the swarm evolves, the ants areable to move to better locations, and within five rounds,the swarm is able to achieve separation. Even after a goodCH distribution is achieved, the fitness value keepschanging as the ants are forced to wander to achieve loadbalancing. It is also found that HEED mostly forms lessernumber of CHs than T-ANT. The number of clustersformed using T-ANT always follows the suggested optimalnumber as per our analysis. Thus, in terms of energyefficiency, HEED will not be able to perform as well asT-ANT due to its clustering process. As for LEACH, itsprobabilistic CH election forces the value to vary signifi-cantly.At some rounds, there are no CHsin range of anotherbut at other rounds, there are more than a few CHs in range.

6.3. Load distribution

In order to observe how well T-ANT promotes load bal-ancing among the nodes, we ran a simulation on periodicdata collection at 2-s interval for 2000 s. At the outset, eachnode had 2-J battery energy. Figs. 7(a) and (b) shows resid-ual energy across the nodes at the end of simulation com-paring three protocols. The plots are separated to avoidclutter. It is obvious that T-ANT achieved the best perfor-mance by maintaining a near uniform battery dischargeamong the nodes. HEED did not perform as well mainly

due to its use of tentative CHs that do not ultimately

become CHs. This causes the affected nodes to become

uncovered, which in turn forces them to send their mes-sages directly, possibly multihop, to the sink. As expected,LEACH also did not perform as well. Since LEACH doesnot guarantee an optimal clustering throughout its opera-tion, the number of clusters formed and their sizes varygreatly. When a suboptimal clustered topology is formedand operated for certain interval, this causes some nodesto become more energy drained than the others.

6.4. Dynamics of clustered topology

The following figures visually demonstrate how T-ANTpromotes good distribution of CHs in thenetwork comparedto HEED and LEACH. Figs. 8 and 9 depict the clusteredtopology formation at different rounds. For T-ANT, wechose to use the optimal number of ants based on the analy-sis. In the following figures, a ring represents a CH, a filledcircle represents a regular node, and a line segment links amember to its CH. Any unlinked regular nodes are uncov-ered. For T-ANT, it is visually evident that when the sinkrandomly releases the ants, they could still be forced toneighboring nodes, which results in some regular nodes leftuncovered as shown in Fig. 8(a). However, as the swarmevolves, the ants are able to achieve some separation byround three [Fig. 8(b)]. When the system converges by round

five, the network achieves good CH distribution.

0

1

2

3

4

5

6

7

8

9

10

0 200 400 600 800 1000

Time (s)

CHFitness

T-ANT

HEED

LEACH

Fig. 6. The CH election fitness at different simulation time for T-ANT,LEACH and HEED.

1.65

1.7

1.75

1.8

1.85

1.9

1.95

2

0 20 40 60 80 100

Node ID

Resid

ualEnergy(J)

T-ANT

HEED

1.65

1.7

1.75

1.8

1.85

1.9

1.95

2

0 20 40 60 80 100

Node ID

ResidualEnergy(J)

T-ANT

LEACH

Fig. 7. Residual energy distribution of nodes after 2000-s simulation ofT-ANT, HEED and LEACH. The nodes are initially equipped with 2-Jbattery energy. (a) T-ANT vs. HEED; (b) T-ANT vs. LEACH.



13/16

HEED is also able to achieve a fair CH distribution. How-ever, it is not able to achieve complete network coverageresulting in many regular nodes left un-clustered. There aretwo issues that might have contributed to this shortcoming.Firstly, the issue of tentative CHs that do not finally becomeCHs, as highlighted earlier. The other issue is with the proto-cols Finalize phase, its last phase of cluster setup. Duringthis phase, any node that did not receive a CH message isforced to become a CH. This causes some neighboring nodesto become CHs, as depicted in Fig. 9(b). LEACH exhibitssignificant variations to its clustered topology, as expected.For certain rounds, there are very few clusters formed withmany nodes left uncovered, as shown in Fig. 9(c). At otherrounds, there are many more clusters than necessary, whichforce many of them to be in range of quite a few CHs[Fig. 9(d)]. Such a formation results in significant disparityin their cluster membership, leading to imbalance in energydissipation when the network has to operate under this sub-optimal structure for a certain interval.

6.5. Network lifetime

In Fig. 10, the improvement gained through T-ANT isfurther exemplified by the network lifetime graph. For this

investigation, we set initial battery energy at only 0.1 J to

let the nodes to die sooner. This however does not changethe pattern of behavior of these protocols. It is evident thatT-ANT exhibits the longest lifetime with all nodes remainingfully functional. It is found that T-ANT achieves more thantwice the lifetime of HEED and LEACH. It also achieves asmuch as five times the lifetime of the flat routing approach.There is a distinctive stair-case effect across all the curves.This is caused by the chosen routing protocol. Whenever cer-tain critical nodes on the routing tree die, a part of the net-work is partitioned resulting in a number of nodessimultaneously losing connectivity to the sink. Thus, thereare sudden plunges in the numberof active nodes throughoutthe curves. T-ANT avoids the use of such critical nodes aslong as possible with its ants. This obviously indicates thatT-ANT promotes good load balancing across the entire net-work to sustain the network for its longest possible use.Fig. 10 also indicates the utility of a clustering algorithmagainst a flat routing approach. All the clustering protocolswith data aggregation are able to sustain network lifetime atleast twice the lifetime of the minimum hoprouting protocol.

7. Conclusions

In this work, we proposed the first novel application of

the swarm intelligence technique to the resource-starved

Fig. 8. The dynamics of T-ANT clustered topology at different rounds as the swarm evolves. During earlier rounds, quite a few nodes are uncovered as theswarm behavior hasnot converged. However, thecoverageis almosttotalwhenconvergence is achieved.Topologyformed at round:(a) 1 (b)3 (c)10 and(d) 20.



14/16

wireless sensor nodes domain. Due to the unique challengesof this domain, we devised an approach that allows thealgorithm to converge very quickly with only limitedoverhead. Unlike the traditional applications of suchbiologically-inspired technique, we allow an agent to influ-ence all nodes within the radio range of the host node, andnot just this node alone.

The T-ANT clustering protocol utilizes these social

agents to guide the election of clusterheads in a totally

distributed manner. An analysis was performed to deter-mine the number of clusterheads necessary to achieve theoptimal performance in terms of energy dissipation.Accordingly, we employed this fixed number of agents toelect clusterheads. This unique approach allowed T-ANTto form and maintain the optimal number of clustersthroughout the network operation. Due to the robustnessof any biologically-inspired algorithm, this protocol couldhandle unforeseen circumstances in the environment andnode failures. We have also introduced two fitness metricsthat are useful for comparison of clustering propertiesbetween algorithms. It is demonstrated here that T-ANTis able to achieve two desirables swarm behaviors, namelyseparation and alignment, in a fully distributed way. Dueto the separation behavior, the elected clusterheads aremostly well distributed across the network. Moreover, asa side-benefit of this behavior and the alignment behavior,T-ANT also achieved an even distribution of membersamong the clusters. It is also found that T-ANT maintainsubstantially lesser state overhead in memory comparedto LEACH or HEED.

Such a biologically-inspired approach may also be use-ful in applications that require an in-network actuation,to assist in the sensoractuator coordination. The feasibil-

ity of this use is left for our future work.

Fig. 9. The dynamics of HEED and LEACH clustered topologies at different rounds. The coverage of both protocols is not total. HEED at round (a) 10(b) 20; LEACH at round (c) 10 and (d) 20.

0

20

40

60

80

100

0 1000 2000 3000 4000 5000

Time (s)

Numberofalivenodes

T-ANT

HEED

LEACH

Flat

Fig. 10. Network lifetime against simulation time of T-ANT, HEED andLEACH.



15/16

Acknowledgements

The authors thank the anonymous reviewers for theirvaluable comments and suggestions. This work wassupported in part by grants from ARC Discovery ProjectDP0664782 and USyd R&D L2844 U3230.

References

[1] J. Hill, M. Horton, R. Kling, L. Krishnamurthy, The platformsenabling wireless sensor networks, Communications of the ACM 47(6) (2004) 4146.

[2] K. Akkaya, M. Younis, A survey on routing protocols for wirelesssensor networks, Elsevier Ad Hoc Networks 3 (3) (2005) 325349.

[3] O. Younis, S. Fahmy, Heed: a hybrid, energy-efficient, distributedclustering approach for ad hoc sensor networks, IEEE Transactionson Mobile Computing 3 (4) (2004) 366379.

[4] Y. Ohtaki, N. Wakamiya, M. Murata, M. Imase, Scalable ant-basedrouting algorithm for ad-hoc networks, in: 3rd IASTED InternationalConference on Communications, Internet, and Information Techno-logy, 2004, pp. 5055.

[5] S. Selvakennedy, S. Sinnappan, Y. Shang, Data dissemination basedon ant swarms for wireless sensor networks, in: IEEE ConsumerCommunications and Networking Conference, 2006, pp. 132136.

[6] M. Gerla, J.T.-C. Tsai, Multicluster, mobile, multimedia radionetwork, Wireless Networks 1 (3) (1995) 255265.

[7] F. Kuhn, T. Moscibroda, R. Wattenhofer, Initializing newly deployedad hoc and sensor networks, in: 10th Annual International Confer-ence on Mobile Computing and Networking, 2004, pp. 260274.

[8] D. Baker, A. Ephremides, The architectural organization of a mobileradio network via a distributed algorithm, IEEE Transactions onCommunications 29 (11) (1981) 16941701.

[9] G. Gupta, M. Younis, Load-balanced clustering of wireless sensornetworks, in: IEEE International Conference on Communications,May 2003, pp. 18481852.

[10] A.D. Amis, R. Prakash, Load-balancing clusters in wireless ad hocnetworks, in: 3rd IEEE Symposium on Application-Specific Systemsand Software Engineering Technology, March 2000, pp. 2532.

[11] A.D. Amis, R. Prakash, T.H.P. Vuong, D.T. Huynh, Max-mind-cluster formation in wireless ad hoc networks, in: 19th JointConference of the IEEE Computer and Communications, March2000, pp. 3241.

[12] C.-F. Chiasserini, I. Chlamtac, P. Monti, A. Nucci, An energy-efficient method for nodes assignment in cluster-based ad hocnetworks, Wireless Networks 10 (3) (2004) 223231.

[13] W.B. Heinzelman, A.P. Chandrakasan, H. Balakrishnan, An appli-cation-specific protocol architecture for wireless microsensor net-works, IEEE Transactions on Wireless Communications 1 (4) (2002)660670.

[14] S. Selvakennedy, S. Sinnappan, A configurable time-controlled

clustering algorithm for wireless sensor networks, in: 1st InternationalWorkshop on Heterogenous Wireles Sensor Networks, IEEE Com-puter Society, 2005, pp. 368372.

[15] S. Bandyopadhyay, E.J. Coyle, An energy efficient hierarchicalclustering algorithm for wireless sensor networks, in: 22nd IEEEJoint Conference of the IEEE Computer and CommunicationsSocieties (INFOCOM), 2003, pp. 1713 1723.

[16] A. Ephremides, J. Wieselthier, D. Baker, A design concept for reliablemobile radio networks with frequency hopping signaling, Proceedingsof the IEEE 75 (1) (1987) 5673.

[17] C.R. Lin, M. Gerla, Adaptive clustering for mobile wireless networks,IEEE Journal on Selected Areas in Communications 15 (7) (1997)12651275.

[18] S. Basagni, Distributed clustering for ad hoc networks, in: 4thInternational Symposium on Parallel Architectures, Algorithms, and

Networks, June 1999, pp. 310315.

[19] M. Chatterjee, S.K. Sas, D. Turgut, An on-demand weightedclustering algorithm (wca) for ad hoc networks, in: IEEE GlobalTelecommunications Conference, December 2000, pp. 16971701.

[20] S. Ghiasi, A. Srivastava, X. Yang, M. Sarrafzadeh, Optimalenergy aware clustering in sensor networks, Sensors 2 (2002) 258269.

[21] J. Pan, Y.T. Hou, L. Cai et al., Topology control for wireless sensornetworks, in: 9th International Conference on Mobile Computing andNetworking, September 2003, pp. 286299.

[22] T. Shu, M. Krunz, S. Vrudhula, Power balanced coverage-timeoptimization for clustered wireless sensor networks, in: 6th ACMInternational Symposium on Mobile Ad Hoc Networking andComputing, May 2005, pp. 111120.

[23] M.J. Handy, M. Haase, D. Timmermann, Low energy adaptiveclustering hierarchy with deterministic cluster-head selection, in: 4thInternational Workshop on Mobile and Wireless CommunicationsNetwork, 2002, pp. 368372.

[24] M. Genes, U. Sorges, I. Bouazizi, Ara the ant-colony based routingalgorithm for manets, in: ICPP Workshop on Ad Hoc Networks,August 2002, pp. 7985.

[25] G.D. Caro, M. Dorigo, Antnet: distributed stigmergetic control forcommunication networks, Journal of Artificial Intelligence 9 (1998)317365.

[26] M. Roth, S. Wicker, Termite: Emergent ad-hoc networking, in:Second Mediterranean Workshop on Ad-Hoc Networks, 2003.

[27] H.F. Wedde, M. Farooq, T. Pannenbaecker et al., Beeadhoc: anenergy efficient routing algorithm for mobile ad hoc networks inspiredby bee behavior, in: Conference on Genetic and EvolutionaryComputation, June 2005, pp. 153160.

[28] I. Kassabalidis, M.A. El-Sharkawi, R.J. Marks et al., Adaptive-sdr:adaptive swarm-based distributed routing, in: International JointConference on Neural Networks, May 2002, pp. 351354.

[29] M. Heissenbuttel, T. Braun, Ants-based routing in large scale mobilead-hoc networks, in: Kommunikation in Verteilten Systemen, Feb-ruary 2003, pp. 9199.

[30] B.A. Kadrovach, G.B. Lamont, A particle swarm model for swarm-based networked sensor systems, in: ACM Symposium on AppliedComputing, March 2002, pp. 918924.

[31] M.J. Mataric, Issues and approaches in the design of collective autono-mousagents, Robotics and AutonomousSystems 16 (24)(1995) 321331.

[32] M. Dorigo, G.D. Caro, L.M. Gambardella, Ant algorithms fordiscrete optimization, Artificial Life 5 (2) (1999) 137172.

[33] C.W. Reynolds, Flocks, herds and schools: a distributed behavioralmodel, ACM SIGGRAPH Computer Graphics 21 (4) (1987) 2534.

[34] V. Hartmann, Evolving agent swarms for clustering and sorting, in:Conference on Genetic and Evolutionary Computation, June 2005,pp. 217224.

[35] R. Schoonderwoerd, O.E. Holland, J.L. Bruten, L.J.M. Rothkrantz,Ant-based load balancing in telecommunications networks, AdaptiveBehavior 2 (1996) 169207.

[36] S.G. Foss, S.A. Zuyez, On a voronoi aggregative process related to abivariate poisson process, Advances in Applied Probability 28 (4)(1996) 965981.

[37] D. Subramanian, P. Druschel, J. Chen, Ants and reinforcementlearning: a case study in routing in dynamic networks, in: Interna-tional Joint Conference on Artificial Intelligence, Morgan Kaufman,1997, pp. 832838.

[38] J. Montgomery, M. Randall, Anti-pheromone as a tool for betterexploration of search space, in: Third International Workshop on AntAlgorithms, Springer-Verlag, 2002, pp. 100110.

[39] Q. Huang, Y. Zhang, Radial coordination for convergecast in wirelesssensor networks, in: 29th IEEE International Conference on LocalComputer Networks, November 2004, pp. 542549.

[40] H. Gupta, V. Navda, S.R. Das, V. Chowdhary, Efficient gathering ofcorrelated data in sensor networks,in: MobiHoc05, ACM Press, 2005,pp. 402413.

[41] J. Chou, D. Petrovic, K. Ramachandran, A distributed and adaptivesignal processing approach to reducing energy consumption in sensor

networks, INFOCOM (2003) 10541062.



16/16

[42] B. Sundararaman, U. Buy, A.D. Kshemkalyani, Clock synchroniza-tion for wireless sensor networks: a survey, Elsevier Ad HocNetworks 3 (3) (2005) 281323.

[43] A. Woo, D. Culler, A transmission control scheme for media access insensor networks, in: 7th Annual International Conference on MobileComputing and Networking, July 2001, pp. 221235.

[44] G. Finn, Routing and addressing problems in large metropolitan-scale internetworks, ISI, March 1987.

[45] D. Ganesan, B. Krishnamachari, A. Woo et al., Complex behavior atscale: an experimental study of low-power wireless sensor networks,UCLA, LA, 2002, pp. 111.

[46] T. He, S. Krishnamurthy, J.A. Stankovic et al., Wide-area monitoringof mobile objects: energy-efficient surveillance system using wirelesssensor networks, in: 2nd ACM International Conference on MobileSystems, Applications, and Services, ACM, 2004, pp. 270283.

S. Selvakennedy obtained his Ph.D. in computercommunications in 1999 from the University ofPutra, Malaysia. He is currently serving as aSenior Lecturer at the University of Sydney,Australia. His current research interests lies indeveloping algorithms for media access, routing,localization and topology control issues in wire-less sensor networks, wireless ad hoc networksand wireless mesh networks. Dr. Selvakennedy isa member of ACM and IEEE.

Sukunesan Sinnappan received his Ph.D. inInformation Science from University of New-castle, Australia in 2004. He has been working asa Lecturer in the Department of InformationSystems, University of Wollongong, Australiasince 2004. He is a member of IEEE (Society onSocial Implications of Technology), Associationof Information Systems (SIG E-Business), andInternational Association for Development ofInformation Society (IADIS). His researchinterests include online profiling, location-basedservices and wireless sensor networks.

Yi Shang is an Associate Professor in theDepartment of Computer Science at the Univer-sity of Missouri at Columbia. He received Ph.D.degree from University of Illinois at Urbana-Champaign in 1997. His research interests includewireless sensor networks, bioinformatics, intelli-gent distributed systems and non-linear optimi-zation. His research has been supported by NSF,

NIH, DARPA, Microsoft, and Raytheon. He is amember of the IEEE and ACM.


Documents

A Biologically-Inspired Clustering Protocol for Wireless Sensor Networks