adaptive design optimization of wireless sensor networks...

Computer Networks 51 (2007) 1031–1051

www.elsevier.com/locate/comnet

Adaptive design optimization of wireless sensor networksusing genetic algorithms q

Konstantinos P. Ferentinos *, Theodore A. Tsiligiridis

Informatics Laboratory, Agricultural University of Athens, 75 Iera Odos street, Athens 11855, Greece

Received 31 July 2005; received in revised form 29 January 2006; accepted 28 June 2006Available online 2 August 2006

Responsible Editor: N.B. Shroff

Abstract

We present a multi-objective optimization methodology for self-organizing, adaptive wireless sensor network designand energy management, taking into consideration application-specific requirements, communication constraints andenergy-conservation characteristics. A precision agriculture application of sensor networks is used as an example. Weuse genetic algorithms as the optimization tool of the developed system and an appropriate fitness function is developedto incorporate many aspects of network performance. The design characteristics optimized by the genetic algorithm systeminclude the status of sensor nodes (whether they are active or inactive), network clustering with the choice of appropriateclusterheads and finally the choice between two signal ranges for the simple sensor nodes. We show that optimal sensornetwork designs constructed by the genetic algorithm system satisfy all application-specific requirements, fulfill the existentconnectivity constraints and incorporate energy-conservation characteristics. Energy management is optimized to guaran-tee maximum life span of the network without lack of the network characteristics that are required by the specificapplication.� 2006 Elsevier B.V. All rights reserved.

Keywords: Wireless sensor networks; Genetic algorithms; Adaptive network design; Energy conservation; Optimal design

1. Introduction

Wireless Sensor Networks (WSNs) generally con-sist of a large number of low-cost, low-power, mul-

1389-1286/$ - see front matter � 2006 Elsevier B.V. All rights reserved

doi:10.1016/j.comnet.2006.06.013

q Parts of this paper have been presented at the 2nd IEEEConference on Sensor and Ad Hoc Communications andNetworks (SECON 2005), Santa Clara, CA, USA, 26–29September 2005.

* Corresponding author. Tel.: +30 210 529 4203; fax: +30 210529 4199.

E-mail address: [email protected] (K.P. Ferentinos).

tifunctional sensor nodes that are small in size andcommunicate over short distances [1]. Their struc-ture and characteristics depend on their electronic,mechanical and communication limitations but alsoon application-specific requirements. In WSNs,sensors are generally deployed randomly in the fieldof interest; however, there are certain applicationswhich provide some guidelines and insights, leadingto the construction of an optimal architecture interms of network infrastructure limitations andapplication-specific requirements.

.

mailto:[email protected]

1032 K.P. Ferentinos, T.A. Tsiligiridis / Computer Networks 51 (2007) 1031–1051

One of the major and probably most importantchallenges in the design of WSNs is the fact thatenergy resources are significantly more limited thanin wired networks [1,2]. Recharging or replacing thebattery of the sensors in the network may be diffi-cult or impossible, causing severe limitations in thecommunication and processing time between allsensors in the network. Note that failure of regularsensors may not harm the overall functioning of aWSN, since neighboring sensors can take over,provided that their density is high. Therefore, thekey parameter to optimize for is network lifetime,or the time until the network gets partitioned.

Another issue in WSN design is the connectivityof the network according to the selected communi-cation protocol [2,3]. The most common protocolfollows the cluster-based architecture, wheresingle-hop communication occurs between sensorsof a cluster and a selected clusterhead sensor thatcollects all information gathered by the other sen-sors in its cluster. Usually, connectivity issuesinclude the number of sensors in each cluster,because a clusterhead can handle up to a specificnumber of connected sensors, as well as coverageissues related to the ability of each sensor to reachsome clusterhead.

Finally, design issues that have been ratherneglected in the research literature are those thatdepend on the particular application of WSNs.Energy and connectivity issues are certainly impor-tant in a WSN design, but one must not forget thepurpose of the sensor network, which is the collec-tion and possibly management of measured datafor some particular application. This collectionmust meet specific requirements, depending on thetype of data that are collected. These requirementsare turned into specific design properties of theWSN, which in this work are called ‘‘application-specific parameters’’ of the network.

Several analyses of energy efficiency of sensornetworks have been realized [2–5] and several algo-rithms that lead to optimal connectivity topologiesfor power conservation have been proposed [6–11].However, most of these approaches do not take intoaccount the principles, characteristics and require-ments of application-specific WSNs. When thesefactors are considered, then the problem of optimaldesign and management of WSNs becomes muchmore complex.

A WSN designer who takes into account all thedesign issues discussed above obviously deals withmore than one nonlinear objective functions or

design criteria which should be optimized simulta-neously (this problem is discussed in [12]). Thus,the focus of the problem is how to find manynear-optimal non-dominated solutions in a practi-cally acceptable computational time. There areseveral interesting approaches to tackling suchproblems, but one of the most powerful heuristics,which is also appropriate to apply in our multi-objective optimization problem, is based on GeneticAlgorithms (GAs) [13]. GAs try to imitate naturalevolution by assigning a fitness value to each candi-date solution of the problem and by applying theprinciple of survival of the fittest. Their basiccomponents are the representation of candidatesolutions to the problem in a ‘‘genetic’’ form (geno-type), the creation of an initial, usually random pop-ulation of solutions, the establishment of a fitnessfunction that rates each solution in the population,the application of genetic operators of crossoverand mutation to produce new individuals fromexisting ones and finally the tuning of the algorithmparameters like population size and probabilities ofperforming the pre-mentioned genetic operators.The successful application of GAs in a sensornetwork design in [14] led to the development ofseveral other GA-based application-specific app-roaches in WSN design, mostly by the constructionof a single fitness function [15–18], but also by con-sidering Pareto optimality in the evaluation of fit-ness values [19]. However, in most of theseapproaches, either very limited network characteris-tics are considered, or several requirements of theapplication cases are not incorporated into the per-formance measure of the algorithm.

The novelty of this work stands in the develop-ment of an integrated GA approach, both in thedirection of degrees of freedom of network charac-teristics and of application-specific requirementsrepresented in the performance metric of the GA.The primary goal is to find the optimal operationmode of each sensor so that application-specificrequirements are met and energy consumption ofthe network is minimized. More specifically,network design is investigated in terms of activesensors placement, clustering and signal range ofsensors, while performance estimation includes,together with connectivity and energy-related char-acteristics, some application-specific properties likeuniformity and spatial density of sensing points.Thus, the implementation of the proposed method-ology results in an optimal design scheme, whichspecifies the operation mode for each sensor. The

K.P. Ferentinos, T.A. Tsiligiridis / Computer Networks 51 (2007) 1031–1051 1033

ultimate objective of this research is to find adynamic sequence of operation modes for each sen-sor, i.e. a sequence of WSN designs, which will leadto maximization of network lifetime in terms ofnumber of data collection (measuring) cycles. Thisis achieved by implementing the algorithm repeat-edly in order to develop a dynamic network designthat adapts to new energy-related information con-cerning the status of the network after each measur-ing cycle or at predefined time intervals.

In the following section we describe the WSNmodeling approach and the problem statementand complexity. In Section 3 we describe the GAapproach that was used to develop the WSN designalgorithm by analyzing the representation schemethat was used, the development of the fitnessfunction that drives the evolution process of thealgorithm and finally, the steps of the algorithmtowards design optimization and further adaptationfor energy conservation. In Section 4 we present thenetwork design capabilities of the algorithm when itis applied on a set of sensors with full batterycapacities. The procedure leads to an optimal designof the WSN, which is further used as the initialnetwork in the sequence of runs in the dynamicalgorithm. Its capability of sensor usage rotationand avoidance of using sensors with low-batterylevels is shown in Section 5 where the algorithm isapplied on the re-design of battery-constrainedWSNs. Section 6 discusses the performance of thealgorithm in adaptive design of WSNs duringseveral consecutive measuring cycles, both at thelevels of network characteristics, such as communi-cation issues and application-specific requirements,as well as of energy-conservation characteristics,such as life-time maximization. Finally, in Section7, some overall conclusions are drawn and trendsof future work are stated.

2. Problem outline

The methodology of WSN design that wedevelop in this work, although general, takes intoaccount several application-specific characteristics,such as those posed in the framework of precisionagriculture, to show the performance of the devel-oped algorithm. Precision agriculture refers to theapproach of agricultural control and managementbased on direct chemical, biological and environ-mental sensing. Sensor networks play a vital rolein that approach by maximizing the quantity, diver-sity and accuracy of information extracted from a

WSN deployment. The parameters to be sensedinclude regular environmental parameters liketemperature, humidity and solar radiation, as wellas soil moisture, dissolved inorganics such as nitro-gen and phosphorous species, and finally herbicidesand pesticides. There are several sensing approachesthat contribute to data collection, including remotesensing via satellites and airborne sensors, autono-mous mobile systems and embedded, networkedsystems. WSNs belong to this last category.

2.1. WSN modeling

The salient features of the proposed WSN are thefollowing: A square grid of 30 by 30 length units isconstructed and sensors are placed in all 900 junc-tions of the grid, so that the entire area of interestis covered. The grid is applied to open field cultiva-tion, where a length unit is an abstract parameter sothat the developed system for optimal design isgeneral enough. The length unit is defined as thedistance between the positions of two neighboringsensor nodes in the horizontal or vertical dimension.Sensors are identical and may be either active orinactive. They are assumed to have power controlfeatures allowing manual or automatic adjustmentof their transmit power through the base station.In this way, they are capable of transmitting inone of three supported signal ranges. Providedthat a sensor is active, it may operate as a cluster-head transmitting at an appropriate signal range(CH sensor) that allows the communicationwith the remote base station (sink), or it mayoperate as a ‘‘regular sensor’’ transmitting at eitherhigh or low-signal range (HSR/LSR sensor,respectively).

We consider a cluster-based network architec-ture. There are several sophisticated clusteringmethodologies in the literature of WSNs towardsenergy saving [20–23]. However, our work tacklesthe energy saving issue through the optimizationof the operating modes of sensors, thus a simpleapproach of clustering sensors in regular operatingmodes with their closest CH sensor is adopted forthe formation of clusters in the network. Conse-quently, sensors are divided into clusters and in eachcluster a sensor is chosen to act as a clusterhead. Allsensors in regular operating modes in a clustercommunicate directly (one-hope) with the closestclusterhead and this is how clusters are formed.Clusterheads communicate directly with the remotebase station (single-hop transmission).

Table 1Correspondences between objectives and optimization parameters

Objectives Optimizationparameters

Parameter symbols inGA methodology

J1 Operational energy OE

J2 Communication energy CE

J3 Battery capacity penalty BCP

J4 Uniformity ofmeasurements

–

J 04 Mean relative deviation ofmeasurement points

MRD

J5 Sensors-per-CH error SCE

J6 Sensors out of range SORE

J7 Spatial density error SDE


It is assumed that communication between clus-terheads and the base station can always beachieved when required and that the base stationis able to communicate with every sensor in thefield, meaning that every sensor is capable ofbecoming a clusterhead at some point. In addition,it is assumed that traffic load is uniformly distrib-uted among sensors in regular operating modes.Since clusterheads have to handle all traffic gener-ated by and destined to the cluster, they have totransmit, receive and process a much larger amountof traffic than ‘‘regular sensors’’. Clusterheads needto perform long range transmissions to the basestation, data collection and aggregation at specificperiods including some computations, as well ascoordination of MAC within a cluster. The problembecomes more complex in the cases of multi-hoptransmissions, where clusterheads need to cover dis-tances that are usually much greater than the‘‘regular sensors’’ transmission range. Althoughthe analysis of this operation is out of the scope ofthis work, the clear result is that clusterheads expe-rience high energy consumption and exhaust theirenergy resources more quickly than ‘‘regularsensors’’ do.

2.2. Problem statement

We propose an algorithm to dynamically designWSN topologies by optimizing energy-relatedparameters that affect the battery consumption ofthe sensors and thus, the life span of the network.At the same time, the proposed algorithm tries tomeet some embedded connectivity constraints andoptimize some physical parameters of the WSNimplemented by the nature of the specific applica-tion. The multiple objectives of the optimizationproblem are blended into a single objective function,the parameters of which are combined to formulatea fitness function that gives a quality measure toeach WSN topology and it is optimized by the pro-posed algorithm, as it is shown in Section 3.

We identify three sets of parameters which dom-inate the design and the performance of a WSN forprecision agriculture. The first set is the application-

specific parameters which include two parametersregarding the deployment of sensors for the specificcase considered here. These are the highest possibleuniformity of sensing points and some desiredspatial density of measuring points. The second setis the connectivity parameters which include anupper bound on the number of sensors that each

clusterhead sensor can communicate with, and thefact that all sensors must have at least one cluster-head within their signal range. Finally, the thirdset refers to the energy-related parameters whichinclude the operational energy consumptiondepending on the types of active sensors, the com-munication energy consumption depending on thedistances between sensors that communicate withtheir corresponding clusterhead, and finally the bat-tery energy consumption.

The optimization problem is defined by the min-imization of the energy-related parameters (say,objectives J1, J2 and J3) and the maximization ofsensing points’ uniformity (objective J4), subject tothe connectivity constraints (say, constraints C1

and C2) and the spatial density requirement (con-straint C3) (see Table 1 for the exact correspon-dences). In order to combine all objectives into asingle objective function (weighted sum approach),the optimization parameters are formed in such away that all of them are minimized. Thus, objectiveJ4 is expressed by its dual objective, say J 04, whichhas to be minimized. Further, the penalization ofthe constraints C1, C2 and C3 allows their transfor-mation into objectives J5, J6, and J7, respectively,which have to be minimized. Thus, a single objectivefunction that blends all (obviously conflicting)objectives is of the form

f ¼ minX7

i¼1i6¼4

wiJ i þ w4J 04

8>><>>:

9>>=>>;: ð1Þ

This form of objective function is suitable for theformulation of a numeric evaluation function [24](namely a ‘‘fitness function’’ in the terminologyof GAs), which gives a quality measure to eachpossible solution of the optimization problem. The


details of that formulation are presented in Section3. What follows describes the mathematical repre-sentation of the optimization parameters in their‘‘minimization’’ form.

1. Application-specific parameters: The main goalof a WSN used in precision agriculture is to takeuniform measurements over the entire area of inter-est, so that an overall and uniform picture of theconditions of the area is realized. This has beenachieved using the following two parameters:

(a) First, the measure of uniformity of measure-ments. The metric of the uniformity of mea-surement points that was used here was theMean Relative Deviation (MRD). The entirearea of interest was divided into several over-lapping sub-areas. Sub-areas are defined byfour factors: two that define their size (lengthand width) and two that define their overlap-ping ratio (ratios in the two directions). Allthese factors are expressed in terms of theunit length of each direction. The larger theoverlapping ratio is, the higher precision isachieved in the evaluation of uniformity,but also, the slower the algorithm becomes.In order to define MRD, the notion of spatialdensity (q) of measurements was used. Morespecifically, qSi, the spatial density of mea-surements in sub-area Si, was defined as thenumber of measurements over the area ofthe ith sub-area, i = 1,2, . . . ,N, where N isthe number of overlapping sub-areas intowhich the entire area, say S, was divided. Inaddition, qS, the spatial density of the entirearea of interest, was defined as the totalnumber of measurements of the networkover the total area of interest. Thus, MRD

was defined as the relative measure of thedeviation of the spatial density of measure-ments in each sub-area from the totalspatial density of measurements in the entirearea

MRD ¼PN

i¼1 qSi� qS

�� N � qS

: ð2Þ

Low values of MRD mean high uniformity ofmeasurement points. Acceptable values for ourapplication example are of MRD below 0.15.

(b) The second application-specific parameter ofthe fitness function was the Spatial DensityError (SDE) that was used to penalize net-work designs that did not meet the minimum

required spatial density of measurementpoints that would suffice adequate monitoringof the measured variables (e.g., air or soiltemperature, air or soil relative humidity,solar radiation, etc.) in the area of interest.The desired spatial density qd was set equalto 0.2 measurement points per square lengthunit and the SDE factor was evaluated by

SDE ¼qd�qs

qdif qs < qd ;

0 otherwise:

(ð3Þ

2. Connectivity parameters: A crucial issue inWSNs is the assurance that network connectivityexists and all necessary constraints are satisfied.Here, these necessary characteristics of the sen-sor network were taken into account by the inclu-sion of the following parameters in the fitnessfunction:

(a) A Sensors-per-Clusterhead Error (SCE)parameter to ensure that each clusterheaddid not have more than a maximum prede-fined number of sensors in regular operatingmodes in its cluster. This number is definedby the physical communication capabilitiesof the sensors as well as their data manage-ment capabilities in terms of quantity of datathat can be processed by a clusterhead sensor.It was assumed to be equal to 15 for the appli-cation considered here. If nfull is the numberof clusterheads (or clusters) that have morethan 15 active sensors in their clusters and ni

is the number of sensors in the ith of thoseclusters, then

SCE ¼Pnfull

i¼1ni

nfull if nfull > 0;

0 otherwise:

(ð4Þ

(b) A Sensors-Out-of-Range Error (SORE)parameter to ensure that each sensor can com-municate with its clusterhead. This of coursedepends on the signal range capability of thesensor. It is assumed that HSR-sensors covera circular area with radius equal to 10 lengthunits, while LSR-sensors cover a circular areawith radius equal to 5 length units. If nout isthe number of active sensors that cannot com-municate with their clusterhead and n is thetotal number of active sensors in the network,then


SORE ¼ noutn

: ð5Þ

3. Energy-related parameters: Energy consump-tion in a wireless sensor network, as explainedearlier, is a crucial factor that affects the perfor-mance, reliability and life span of the network. Inthe optimization process during the evolutionarydesign of the sensor network, three differentenergy-related parameters were taken into account:

(a) Operational Energy (OE) consumptionparameter, which refers to the energy that asensor consumes during some specific time ofoperation. It basically depends on the opera-tion mode of the sensor, that is, whether itoperates as a CH, a HSR or a LSR sensor,or whether it is inactive. The correspondingrelevance factors for the energy consumptionof the three active operating modes of the sen-sors are taken proportional to 20:2:1, respec-tively and zero for inactive. The meaning isthat the energy consumption of a sensoroperating in CH mode is 10 times more thanthat of a sensor operating in HSR mode and20 times more than that of a sensor operatingin LSR mode. These relevant factors wereused to simplify the analysis and did notnecessarily represent accurately the real energyrelations between the available operationmodes of the sensors. Their exact valuesdepend on electromechanical characteristicsof the sensors and were not further consideredin the analysis presented here. The OE con-sumption parameter was then given by

OE ¼ 20 � nchnþ 2 � nhs

nþ nls

n; ð6Þ

where, nch, nhs and nls are the number ofCH, HSR and LSR sensors in the network,respectively.

(b) Communication Energy (CE), which refers tothe energy consumption due to communica-tion between sensors in regular operatingmodes and clusterheads. It mainly dependson the distances between these sensors andtheir corresponding clusterhead, as defined in[6]. It is depicted by

CE ¼Xc

i¼1

Xni

j¼1

l � dkji; ð7Þ

where c is the number of clusters in the net-work, ni is the number of sensors in the ithcluster, dji is the Euclidean distance from sen-sor j to its clusterhead (of cluster i) and land k are constants, characteristic of the topol-ogy and application site of the WSN. For thespecific precision agriculture application foropen field monitoring, the values of l = 1and k = 3 were chosen.

(c) Battery life. An important issue in WSNs isself-preservation of the network itself, that is,the maximization of the life span of the sensors.Each sensor consumes energy from somebattery source in order to perform its vitaloperations, like sensing, communication, dataaggregation if the sensor is a clusterhead, etc.Battery capacity of each sensor of the networkwas taken into account in the design optimiza-tion process by the introduction of a BatteryCapacity Penalty (BCP) parameter. Sincethe operation mode of each sensor is known,its Battery Capacity (BC) can be evaluatedat each time. Thus, when the design optimiza-tion algorithm is applied at a specific timet (measuring cycle), the BCP parameter isgiven by

BCP ½t� ¼Xngrid

i¼1

PF ½t�i �1

BC½t�i

� 1

!; t ¼ 1; 2; . . .

ð8Þ
Note that BCi is updated according to theoperation mode (CH, HSR or LSR) of eachsensor i, during the previous measuring cyclet � 1 of the network
BC½t�i ¼ BC½t�1�i � BRR½t�1�

i : ð9Þ
In the above:
• BCP[t] is the Battery Capacity Penalty of theWSN at measuring cycle t. It is used topenalize the use of sensors with low-batterycapacities, giving at the same time largerpenalty values to operating modes that con-sume more energy (especially CH mode).

• ngrid is the total number of available sensornodes.

• PF ½t�i is the Penalty Factor assigned to sensori. The values it takes are given according tothe operation mode of sensor i. The valuesused here are proportional to the relevant


battery consumptions of the sensor modes,namely, 20:2:1 for active sensor modes(CH, HSR and LSR, respectively) and 0for inactive. They provide different penaltiesaccording to the specific modes of thesensors in the WSN of the following measur-ing cycle. However, as it is explained in thenext section, further exploration of the opti-mal relevance values needs to be performed.

• BC½t�i and BC½t�1�i are the Battery Capacities

of sensor i at measuring cycles t and t � 1,respectively, taking values between 0 and1, with 1 corresponding to full batterycapacity and 0 to no capacity at all.

• BRR½t�1�i is the Battery Reduction Rate that

depends on the operation mode of sensor i

during the measuring cycle t � 1 andreduces its current battery capacity accord-ingly, using the percentage of the relevancefactors for the energy consumption of theoperating modes of the sensor as follows:0.2 for CH, 0.02 for HSR 0.01 for LSRoperation modes and 0 for inactive sensors.

2.3. Problem complexity

By considering the connectivity constraint of theoptimization problem which upper bounds thenumber of allowed sensors per cluster in the WSNtopology (15 sensors in our case), the problem isequivalent to finding the Minimum DegreeSpanning Tree (MDST) over the active sensors ofthe WSN, which is NP-hard [25]. In other words,deciding whether there exists a spanning tree whosedegree is upper-bounded by a number, say D, isequivalent to finding the MDST.

The information on the Euclidean distances ofthe active sensors reduces the problem to a Mini-mum Weight Spanning Tree (MWST). In the casewhere all nodes are placed on a two-dimensionalplane and the weights of the edges between twonodes correspond to the Euclidean distances, thedegree of a MWST is upper-bounded by 6 [26].However, the other constraints of our optimiza-tion problem (e.g., all active nodes other thanclusterheads have degree equal to 1, energyrequirements, etc.), might not allow the construc-tion of a connected MWST. Therefore, the prob-lem still needs to be solved in the context of theMDST, which as we mentioned above, is NP-hard.

3. Methodology of GA

The methodology and formulation of GAs forsome specific application incorporates three basicsteps: the problem representation, i.e. the encodingmechanism of the problem’s phenotypes into geno-types that GAs manipulate and evolve, the formula-tion of the fitness function that gives to eachindividual (i.e. possible network design) a measureof performance, and finally the choice of the geneticoperators and the selection mechanism used. Thesesteps are of major importance, as they drasticallyaffect the performance of the final results and theyare described in detail in the following Sections3.1–3.3, respectively. Section 3.4 presents the algo-rithm that is dynamically applied to achieve adap-tive design of the WSN towards continuous energyconservation.

3.1. WSN representation

The variables that are included in the WSN rep-resentation are those that give all the required infor-mation so that the performance of a specificnetwork design can be evaluated. These variablesare the placement of the active sensors of thenetwork, the operation mode of each active sensor,that is, whether it is a clusterhead or a ‘‘regular sen-sor’’, and in the case of a ‘‘regular sensor’’, the rangeof its signal (high or low).

Each individual in a GA population specifies thecomposition and arrangement of sensors encoded asa vector of genes. Fig. 1 shows an example individ-ual which represents a grid of sensors with r rowsand c columns. For a sensor placed at each of ther Æ c grid positions, there are four possibilities repre-sented by a two-bit encoding scheme: being an inac-tive sensor (00), being an active sensor operating ina low-signal range (10), being an active sensor oper-ating in a high-signal range (01) and being an activeclusterhead sensor (11). The grid junctions areencoded row by row in the bit string, as shown inFig. 1. Each position needs two bits for the encod-ing, thus, the length of an individual (GA string)is 2rc. In the specific design problem analyzed here,the sizes of r and c are both equal to 30, thus thelength of the individuals are equal to 1800.

3.2. Fitness function

In the case under investigation the fitness func-tion is a weighting function that measures the

c

r 1 2 3 . . . 2c . . . 2rc1 1 0 0 0 1 1 0 0 0 0 0 0 0

active sensor - clusterhead 11active sensor - high signal range 10active sensor - low signal range 01inactive sensor 00

. . .

bit number :

Fig. 1. Binary representation (on the right) of the location and state of sensors in a randomly generated WSN (on the left). Representationof the first row is shown.

Table 2Weighting coefficients of GA fitness function

Weighting coefficient ‘‘Equal importance’’ values Final values

a1 102 102

a2 104 104

a3 2 106

a4 103 105

a5 10 10a6 5 · 10�3 10�2


quality and the performance of a specific sensornetwork design. This function is maximized by theGA system in the process of evolutionary optimiza-tion. A fitness function must include and correctlyrepresent all or at least the most important param-eters that affect the performance of the WSN design.Having described these parameters (Section 2), thenext issue is the decision on the importance of eachparameter on the final quality and performancemeasure of the network design. The final form ofthe weighting linear fitness function f of a specificWSN design is given by

f ¼ 1=ða1 �MRDþ a2 � SDE þ a3 � SCE þ a4 � SORE

þ a5 � OE þ a6 � CE þ a7 � BCP Þ: ð10Þ

The significance of each parameter is defined bysetting appropriate weighting coefficients ai: i =1,2, . . . , 7 in the fitness function that will be maxi-mized by the GA. The values of these coefficientswere determined based on experience about theimportance of each parameter. First, weightingcoefficients that resulted, in average the same impor-tance of each parameter were determined (firstcolumn of Table 2) and after some rudimentalexperimentation, the final values that best repre-sented the intuition about relevant importance ofeach parameter were set (second column of Table2). As can be seen in Table 2, the final weights weresuch that network connectivity parameters (weightsa3, a4) were treated as constraints, in the sense thatall sensors should be in range with a clusterhead andno clusterhead should be connected to more thanthe predefined maximum number of sensors. Therewas no need for an increase of the SDE weight valuebecause all GA-generated designs seemed to meet

that specific constraint (i.e. the desired spatial den-sity of measurement points). Note that the coeffi-cients were determined based on normalizationwith respect to the value of a5 which is set equalto 10. It should be noted that the BCP parameterwas not taken into account in the optimization ofthe initial design of the WSN, as it was assumed thatall sensor nodes had full battery capacities at thebeginning. The final value of a7 was the result of atrade-off between energy management optimizationand network characteristics optimization, particu-larly of the characteristics concerning the applica-tion-specific properties of the WSN, as it is furtherexplained in Section 4.

3.3. Genetic operators and selection mechanism

The types of crossover and mutation are of majorimportance to the performance of the GA optimiza-tion. Two types of the classical crossover operatordefined in [27] were tested, the one-point and thetwo-point crossover. The mutation type that wasused was the classical one for binary representation,that is, the swapping of the bits of each string (0becomes 1 and vice versa) with some specific lowprobability. Crossover is also applied with some


specific probability. Both these probabilities aretuned after proper experimentation, as explainedin Section 4.

The adopted selection mechanism was the rou-lette wheel selection scheme. The probability ofselecting some individual to become a parent forthe production of the next generation was propor-tional to its fitness value. In addition, in order toassure that the best individual of each generationwas not destroyed by the crossover and mutationoperators during the evolution process, ‘‘elitism’’was included in the algorithm, meaning that thecurrent best individual at each generation of thealgorithm always survived to the next generation.

3.4. Dynamic optimal design algorithm

Having completed the development of a repre-sentation scheme and forming the fitness function,the dynamic genetic algorithm for optimal adaptivedesign of the WSN could be developed. The algo-

Set population size M; Set max # of gen

Generate random initial population of M

for t=1 to G

Evaluate parameters for each individ

Assign fitness value to each individ

for i=1 to M/2

Select 2 parent individuals (acco

Crossover the 2 individuals with

Store the 2 output offspring

end for i

for i=1 to M

Mutate offspring i with probabili

end for i

Replace old population with new offs

end for t

return best individual in current popul

Fig. 2. Pseudocode of the optimal

Apply “ODA”

while WSN is “alive”

Initiate new measuring cycle usin

Evaluate battery capacities at th

Update battery capacities using(9

Re-apply “ODA” to sensors with up

Wait until current measuring cycl

end while

Fig. 3. Pseudocode of the dynamic optim

rithm consisted of two parts: the Optimal DesignAlgorithm (ODA), which is applied to a set of sen-sors with specific battery capacities (Fig. 2), and theDynamic Optimal Design Algorithm (DODA),which updates the battery capacities of the sensorsand reapplies the optimal design algorithm accord-ingly (Fig. 3). Both algorithms as well as all simula-tions presented in the following sections wereimplemented in Matlab.

Some of the issues that have to be clarifiedfollow.

1. Optimal WSN design algorithm:• The size of the population is a parameter of

exploration that is further discussed in thenext section.

• In the assignment of a fitness value to eachindividual, specific weighting coefficients areused in (10) (Table 2).

• The probability of selection of parent individ-uals is proportional to their fitness value.

erations G;

WSN designs

ual in current popul. using (2)-(8)

ual using (10)

rding to fitness values)

probability pc

ty pm

pring to form current population

ation (Optimal_WSN_design)

WSN design algorithm (ODA).

g current Optimal_WSN_design

e end of current cycle, using(9)

)

dated battery capacities

e is completed

al WSN design algorithm (DODA).


• The genetic operators of crossover and muta-tion are applied with specific probabilities, asit is explained in the next section.

2. Dynamic optimal design algorithm:• The measuring cycle is defined as the period of

time during which a clusterhead sensor con-sumes 20% of its full battery capacity.

• The steps of ‘‘battery capacities update’’ and‘‘re-application of the optimal WSN designalgorithm’’ are performed during data collec-tion of the measuring cycle. This is becausebattery capacities at the end of the cycle canbe evaluated based on the developed model,without having to wait until the actual endof the measuring cycle. Thus, at the end ofeach measuring cycle, the next optimal WSNdesign has already been formed and it is thenused for the next data measuring cycle.

• The life span of the network, which is referredto as ‘‘WSN is alive’’ in the pseudocode,defines the application time of the dynamicalgorithm. The network, i.e. the set of sensorsin the field, is considered to be ‘‘alive’’ if theset of sensors with battery capacities abovezero is such that some operational WSN canbe designed and applied to the next measuringcycle.

The number of iterations performed by the algo-rithm in a single measuring cycle are in the order ofG Æ l Æ M2, where G is the number of generations ofthe GA, l is the bit-string length and M is thepopulation size. If n is the total number of availablesensors in the WSN design, then obviously the com-putational complexity of the algorithm is O(n), asonly the l parameter depends on n (l = 2 Æ n).

4. GA experimentation and initial WSN design

GAs have a number of parameters that are prob-lem specific and need to be explored and tuned sothat the best algorithm performance is achieved.These parameters are the population size, the prob-abilities of crossover and mutation and the type ofcrossover. In the beginning, a number of experi-ments were carried out to determine the most appro-priate population size. Sizes from 100 to 1000individuals in orders of hundreds were tested. Thebest performance, by means of maximizing the cor-responding fitness function, was achieved with apopulation size of 300 individuals. Then, severalexplorations were performed with probabilities of

crossover ranging from 0.3 to 0.9 for both one-pointand two-point crossover types and probabilities ofmutation ranging from 0.0001 to 0.01. The resultsled to the use of one-point crossover with probabilitypc = 0.8 and probability of mutation pm = 0.005.

GAs incorporate stochastic operations during theoptimization process while the quality of therandomly generated initial population drasticallyaffects the final performance. Thus, in any explora-tion and then further application of the algorithmthat are presented, several runs were tested with dif-ferent random initial populations. Average resultsover the several runs as well as the best solutionsachieved by each set of parameters were used todraw conclusions.

The developed algorithm was tested in three waysand the results are shown in the current and thefollowing two sections. First, the performance ofthe algorithm in designing initial optimal WSNtopologies and sensor operation modes was exam-ined. Thus, ‘‘ODA’’ was applied in a field of fullbattery capacity sensor nodes. Then, the batterycapacity update term was included and the inte-grated algorithm was tested off-line in some prede-termined WSN designs with limited batteryresources, that is, with specific limited or zero bat-tery capacities at some sensor nodes, so that itscapability of avoiding low-battery nodes would beshown. Finally, ‘‘DODA’’ was applied dynamicallyto examine its performance on adaptive optimaltopology and energy management that would leadto the maximization of the life span of the entireWSN.

The algorithm was started, having availableall sensor nodes of the grid at full battery capacities.The three initial populations that gave thebest results after 3000 iterations of the GA wererecorded (abbreviated as ‘‘GA1’’, ‘‘GA2’’ and‘‘GA3’’, starting from the fittest design). The evolu-tion progress of the best GA run is shown in Fig. 4,where both the fitness progress of the best individualfound by the algorithm as well as the average fitnessof the entire population at each generation are plot-ted. The optimization in the entire GA populationcan be seen from the general increase of the averagepopulation fitness, despite the numerous fluctua-tions caused by the search process through thegenetic operators of crossover and mutation.

The optimization performed by the GA evolutionprocess can also be seen by the progress of the val-ues of some of the parameters of the WSN designsfound during the evolution. These parameters are

Fig. 4. Evolution progress of the best individual (best fitnessvalue) and the entire population (average fitness value) of the GAduring the two best runs of the algorithm.


shown in Fig. 5 for the best run of the GA which ledto the ‘‘GA1’’ design. More specifically, plot (a)shows the evolution of MRD which represents uni-formity of measurement points (the lower the valueof MRD, the better the value of the achieved unifor-mity), plot (b) shows the evolution of the opera-tional energy consumption (OE), plot (c) showsthe evolution of the communication energyconsumption (CE), while plot (d) shows the numberof clusterheads (lower line), high-signal range (mid-dle line) and low-signal range sensors (upper line)in the sensor networks as they evolved duringoptimization.

Fig. 5. Evolution of WSN parameters during 3000 generations. The initvalues for estimation of uniformity of measurement point; (b) operatconsumption parameter; (d) number of active sensors for the three pos

The optimization process can easily be observedby the evolution of WSN characteristics as shownin these graphs. The conducted experiments showedthat in cases where the initial random designssuffered with communication limitation issues, thealgorithm at the beginning of the evolution wasalways trying to find designs that at least satisfiedthe communication and the application-specific con-straints. Afterwards, the other parameters likeenergy issues and clustering were optimized withthe best possible minimization of operation energyconsumption factor, the decrease of clusterheadsexistence, the increase of low-signal range sensorsexistence and so on. Details on all sensor networkcharacteristics for the three GA-generated designscan be seen in Table 3. A comparison with the per-formance and characteristics of some additionaldesigns can be found in [28]. Comparison resultsfavored the GA-generated designs in all aspects ofperformance evaluation, that is, energy consump-tion, connectivity and application-specific charac-teristics.

5. Performance on battery-constrained WSNs

The algorithm was applied on specific initialWSN designs with sensor nodes of various batterycapacities, in order to show the quality of decisionsthat the algorithm makes on the operation modes ofthe sensors for the next measuring cycle. Table 4

ial population has 3:1 ratio of active to inactive sensors. (a) MRD

ional energy consumption parameter; (c) communication energysible operation modes (CH, HSR, LSR).

Table 3WSN designs parameter values

‘‘GA1’’ ‘‘GA2’’ ‘‘GA3’’

MRD 0.0840 0.1018 0.1141SDE 0 0 0OE 5.0086 4.6827 4.9711CE · 103 1.4323 1.6422 1.4965OOR 0 0 0OCC 0 0 0Active 699 602 622CH 133 105 117HSR 275 222 247LSR 291 275 258CH/Active 0.19 0.17 0.19HSR/Active 0.39 0.37 0.40LSR/Active 0.42 0.46 0.41Fitness 0.0137 0.0136 0.0131

Parameter values for the three GA-generated wireless sensornetwork designs. OOR: Out-of-Range sensors (sensors that can-not communicate with some clusterhead); OCC: Over-ConnectedClusters (clusters with more than 15 sensors).

Table 4Initial WSN design scenarios

Batterycapacity (%)

Scenario I(%)

Scenario II(%)

Scenario III(%)

0 15 15 1510 0 30 3050 30 33 070 33 0 0100 22 22 55

Percentages of sensors’ battery levels over all available sensors,for the three examined scenarios.


shows the three scenarios that were used for the ini-tial designs as far as the battery capacities of thesensors are concerned. Battery capacities were givenas a percentage of the full battery capacity offered atthe beginning of the initial measuring cycle, whereasin each scenario, the number of sensors with the

Table 5Battery-capacity usage as active sensors and clusterheads

Batterycapacity (%)

Scenario I Scenario II

Active sensors (%) Clusterheads (%) Active sensors

0 0 (0) 0 (0) 0 (0.4)10 – – 70 (2.3)50 78 (1.1) 14 (2.5) 78 (1.4)70 78 (1.5) 17 (1.1) –100 78 (3.2) 16 (1.6) 77 (2.4)

Average percentages (std’s) of specific battery levels of sensors used asmeasuring cycle, for the three examined scenarios.

specified battery capacity was given as a percentageof the total number of sensors (900). In all three sce-narios, 15% of the sensors were considered havingzero battery capacities. The construction of thesescenarios was based on the percentages of operatingmodes of the sensors in the best GA-generated opti-mal design (namely, the ‘‘GA1’’ design). In ScenarioI, the values of 0%, 50%, 70% and 100% batterycapacity levels are taken equal to the percentagesof CH, HSR, LSR and inactive operating modesrespectively, in ‘‘GA1’’, over all 900 sensors. Therest two scenarios were obviously produced in asimilar way (see Table 4).

The algorithm was run several times for each sce-nario for 3000 generations in each run and the aver-age results are shown in Tables 5 and 6. Both tablesrepresent average rates of used (active) sensor nodesof the proposed by the algorithm WSN designs.Table 5 shows the average percentages and standarddeviations (values in the parentheses) of the sensorsof each initial battery capacity that were active orused as clusterheads in the proposed designs of thenext measuring cycle, for all three scenarios. Wedo not explicitly include HSR and LSR usage ofsensors in these results because these operatingmodes are quite similar. We investigate usage ofactive sensors in general, which is an importantparameter, and from these active sensors, we furtherpresent usage of the CH operating mode, becauseclusterheads drastically differ in energy consump-tion from sensors in regular operating modes, andtherefore their usage and re-usage are of majorimportance. Thus, for example, in ‘‘Scenario I’’,78% of the sensors with 50% battery capacity wereactive in the new WSN design of the next measuringcycle, while 14% of the 50% battery capacity sensorswere used as clusterheads in that new design. Simi-larly, in ‘‘Scenario III’’, only 3% of the sensors with10% battery capacity were used as clusterheads in

Scenario III

(%) Clusterheads (%) Active sensors (%) Clusterheads (%)

0 (0) 0 (0.3) 0 (0)5 (1.5) 67 (1.9) 3 (0.9)

19 (2.3) – –– – –21 (2.5) 78 (1.7) 22 (1.1)

active sensors in general or clusterheads in the WSN of the next

Table 6Active sensors and clusterheads battery-capacity distributions

Batterycapacity (%)

Scenario I Scenario II Scenario III

Active sensors (%) Clusterheads (%) Active sensors (%) Clusterheads (%) Active sensors (%) Clusterheads (%)

0 0 (0) 0 (0) 0 (0.1) 0 (0) 0 (0.1) 0 (0)10 – – 33 (0.8) 12 (3.5) 32 (0.5) 8 (1.9)50 36 (0.6) 32 (5.5) 39 (0.4) 50 (2.9) – –70 38 (0.6) 42 (3.1) – – – –100 26 (1.0) 26 (2.8) 27 (0.7) 38 (3.5) 68 (0.5) 92 (1.9)

Average distribution (percentages and std’s) of active sensors and clusterheads in the WSN of the next measuring cycle over existingbattery levels of sensors, for the three examined scenarios.


the new WSN, while 22% of the full battery capacitysensors were used as clusterheads in the same WSN.As can be seen, there was no case where some sensorwith no battery capacity was used in any of the pro-posed designs, in any scenario. The avoidance ofusing sensors with low-battery capacities is notevident in Scenario I (the battery level distributionof 0/50/70/100 did not help towards that), but itcan be seen in both Scenarios II and III, especiallyin the percentages that represent clusterhead usage.It is evident that sensors with high-battery capacitieswere preferred over low-battery ones, especially inthe case where these sensors served as clusterheadsin the new design.

A different approach of presenting the usage ofsensors in the WSN of the next measuring cycleaccording to their previous battery capacities is usedin Table 6. In that table, the average percentages(and standard deviations in the parentheses) ofactive nodes or clusterheads in each scenario’sdesign of the next measuring cycle that used specificinitial battery capacity sensors are presented. Forexample, in Scenario II, 33% of the active nodesof the new WSN design of the next measuring cyclehad 10% battery capacity, 39% had 50% batterycapacity and 27% had full battery capacity, or, inScenario III, 8% of the sensors chosen to serve asclusterheads in the WSN design of the next measur-ing cycle had 10% battery capacity while 92% of the

Table 7WSN design main characteristics

MRD O

Avg. Std. A

Initial optimal WSN 0.0840 – 5.Scenario I 0.1227 0.0088 5.Scenario II 0.1555 0.0116 5.Scenario III 0.1594 0.0115 5.

Design characteristics of initial optimal WSN design and designs of th

clusterheads had full capacity. The complete avoid-ance of using sensors with no battery is evident heretoo, while the preference in sensors with high-bat-tery capacities can be seen, mainly in Scenarios IIand III where the battery distributions were moreproblematic.

An important issue in the off-line testing ofthe developed system (as well as in the dynamicapplication of the algorithm examined later) is theconservation of the application-specific WSN char-acteristics, while the system tries to avoid the usageof sensors with no-battery or low-battery capacities.For this reason, direct comparison with the LEACHmodel [8] or other models appearing in the litera-ture, needs considerable attention so as to avoid fur-nishing misleading results. It should be noted thateven better energy-conservation usage could beachieved by the developed algorithm, but limita-tions of application-specific parameters and com-munication constraints, limit that ability. As it isshown in Table 7, the values of uniformity andoperational and communication energy consump-tions of the proposed designs were kept quite closeto the optimal values of the original WSN design.This becomes even more important, consideringthe fact that in all three scenarios, 15% of the avail-able sensors had no battery capacity and they werecompletely avoided by the design algorithm. Inaddition, in all three cases, all communication

E CE · 103

vg. Std. Avg. Std.

0086 – 1.4323 –1516 0.0950 1.6953 0.15910047 0.3136 2.0106 0.22023593 0.1729 1.8045 0.1031

e next measuring cycle, for the three examined scenarios.


constraints were met and spatial densities of mea-suring points were kept within the appropriaterange.

6. Adaptive design performance

The self-organizing (adaptation) capabilities ofthe algorithm towards energy conservation but alsotowards connectivity sustainability and nursing ofapplication-specific requirements were examinedby the dynamic application of the algorithm to asequence of measuring cycles. As described inSection 2, battery consumption during one measur-ing cycle was set to 20% of the total (full) batterycapacity for sensors operating as clusterheads, 2%for high-signal range sensors and 1% for low-signalrange sensors, while there was no battery consump-tion for sensors that were inactive during somemeasuring cycle. Therefore, if a static clusteringalgorithm was used, the life span of the WSN wouldhave been five measuring cycles. It should be notedhere that the duration of a measuring cycle was setlarge enough to better demonstrate the way theproposed algorithm operates in avoiding low-bat-tery sensors and maximizing life span of the entirenetwork. In addition, the necessary setup time fornetwork re-configuration and updating was nottaken into account. The performed simulations tryto give an approximation of lifetime duration ofthe WSN in terms of the number of measuringcycles.

The optimal design ‘‘GA1’’ was used as thestarting design in the dynamic application of thealgorithm, which was tested during 15 consecu-tive measuring cycles. A comparison of some preli-minary results with those of static clustering onthe initially optimal WSN (‘‘GA1’’) presented inprevious work [28] showed clear evidence of theenergy conservation that is performed by the adap-tive design of the algorithm. Here, we focus on theanalysis of the effect of the adaptation factor con-cerning energy conservation of the dynamicallyapplied algorithm. The variability of this effect isdetermined by the weighting factor of the BCPparameter in the fitness function of the GA (a7),which from now on we call Energy-ConservationFactor (ECF).

6.1. Adaptation analysis and performance

The dynamic adaptation of WSN design by thedeveloped algorithm during several measuring

cycles was based not only on the conservation ofenergy that would lead to the maximization of thelife span of the network, but also on the conserva-tion of the performance characteristics of theWSN, like measurement uniformity and spatialdensity, faultless connectivity, and minimization ofoperating and communication costs. The algorithmperformed a trade-off between the satisfaction ofthese performance measures and energy conserva-tion. The proper adjustment of the ECF parametercould give dynamic design capabilities that would‘‘prefer’’ either the energy-conservation part or thenetwork performance part. Because of the fact thatthis trade-off is not stable and depends on the user’spreference and the specific demands of the applica-tion that the sensor network is used to, only a suit-able range can be suggested for the specific WSNdesign. After some experimentation with severalvalues of ECF in orders of 10, it was found that areasonable trade-off is performed for ECF valuesbetween 0.01 and 10.

The final analysis of the energy-conservationcharacteristics of the adaptive design process thatis presented in Section 6.2 was based on an appli-cation of the algorithm with ECF parameter equalto 0.1, which kept a balance between energy con-servation and network performance. Here, in thepresentation of the network performance charac-teristics during 15 consecutive measuring cycles,three representative applications of the algorithmare shown, with ECF equal to 10, 0.1 and 0.01,that is, its ‘‘boundary values’’ and the value thatis considered the most appropriate, as explainedbefore. In Fig. 6 it can be seen that the uniformitylevel (MRD) and the communication energy con-sumption of the WSN are highly influenced bythe value of ECF. The adaptive WSN designs withECF equal to 0.1 and 0.01 (especially the latter)kept the MRD values quite low during all measur-ing cycles. There is a small general trend ofincrease in the value of MRD, but this is reason-able as more and more energy limitations are intro-duced into the network as time passes. Similarly, inthe case of communication energy consumption ofthe WSNs, the adaptive design with ECF = 0.01preserved the best values during the entire testingperiod, with values very close to the initial con-sumption of the network. It should be notedhere that spatial density of sensing points was notpresented in the graphs of Fig. 6 because allapproaches gave zero penalty values of SDE

during the entire testing period. In addition, no

Fig. 7. Percentages of sensors with battery capacities below 50%, 40%, 30% and 20% of full battery capacity at the end of each measuringcycle, for three different ECF values.

Fig. 6. MRD, OE and CE performance measures of the WSNs over the testing period of 15 measuring cycles for three different values ofthe ECF.


communication faults occurred throughout theadaptive design processes.

Fig. 7 shows the effect of ECF to the availableenergy of the sensors of the WSN during the period


of the dynamic application of the algorithm. It pre-sents the percentage of sensors that have batterycapacity below certain levels at the end of each mea-suring cycle, with the three ECF values discussedbefore. Except for the indication that appropriateenergy management of the WSN is achieved (as itis analyzed in the next section), these graphs alsoshow that the ECF parameter seems to play animportant role in the life span of the network too.Relevant analysis on remaining sensors with batterycapacities above certain percentage-levels indicatedsimilar effects on the conservation of energyresources.

6.2. Energy-conservation characteristics

As mentioned before, the analysis of the energy-conservation characteristics of the adaptive designprocess that is presented here was based on an appli-cation of the algorithm with ECF parameter equal to0.1, which kept a balance between energy conserva-tion and network performance. The graphs in Fig. 8show the frequencies of sensor usages over the

Fig. 8. Frequency of usage of each sensor of the network, over all measurange sensor, (c) usage as low-signal range sensor, (d) general usage (in

dynamic application of adaptive WSN design (15measuring cycles), i.e. the number of measuringcycles during which each sensor was used. The threepossible usages of clusterhead, high-signal range andlow-signal range are shown in graphs (a)–(c), respec-tively, while graph (d) shows the number of measur-ing cycles during which each sensor was active, ingeneral. For example, in the usages of just 10 sensorswhich are shown in Fig. 9 for convenience (sensornumbers 501–510), it is clear that, sensor number503 for example, was used once as a clusterheadnode during the entire period of 15 measuring cycles(graph (a)), five times as a high-signal range sensor(graph (b)), and seven times as a low-signal rangesensor (graph (c)). Thus, it has been used as an activesensor during 13 measuring cycles (graph (d)).

In average, all sensors were used for 1.6 measur-ing cycles as clusterheads (0.7 standard deviation),for 4.0 measuring cycles as HSR sensors (1.8 std),for 4.7 measuring cycles as LSR sensors (1.8 std)and in general, they were active for 10.3 measuringcycles in average (1.7 std). The average values showthe general tendency to avoid repeatedly using the

ring cycles. (a) Usage as clusterhead node, (b) usage as high-signaldependent of operating mode).

Fig. 9. Detail of Fig. 8, for sensor numbers 501 to 510. Explanation of graphs is the same as that of Fig. 8.


same sensors, especially as clusterheads. In addition,the algorithm manages to avoid the repetitive use ofthe same sensors in HSR mode in a larger degreethan in LSR mode, which is reasonable. The actualplots of course provide more information on theperformance of the dynamic application of the algo-rithm than the average values, especially the plotsconsidering clusterhead usage and usage in general(active nodes).

From these plots, the following remarks can bemade about the dynamic design performed by theproposed algorithm:

• All available sensors eventually become active atsome point, with the vast majority of them beingused more than 5 times during the 15 measuringcycles (in average, each sensor was used around10 times).

• No sensor was used more than three times as aclusterhead during the 15 measuring cycles, withthe vast majority of them being used just once ortwice in that operating mode.

A similar representation that includes the timefactor of the re-use of sensors at each operatingmode is shown in the graphs of Fig. 10. The threeavailable operating modes as well as the general

use of sensors are shown, while the number ofsensors that are used in each operating mode forspecific times during the dynamic application ofthe algorithm is shown for each measuring cycle.More specifically, graph (a) shows the number ofsensors in each cycle that had not yet been used asclusterheads, as well as those that had been usedonce, twice and three times. It can be seen thatthe third reuse in most of those few sensors thatwere used three times as clusterheads, was clearlydelayed. In a comparison of graphs (b) and (c),the slight preference in earlier re-use of sensors inlow-signal range than in high-signal range is shown.The general patterns of all these graphs give a clearindication that some energy-conservation optimiza-tion is performed in the adaptive design of theWSNs. Of course, it should not be forgotten thatthis optimization is restricted by the concurrentoptimization of the rest performance parametersof the WSN.

Table 8 shows the distribution of operatingmodes of the sensors at each of the 15 measuringcycles tested, as well as the average number ofsensors that each clusterhead coordinates respec-tively (standard deviations in the parentheses). Itcan be seen that the number of active sensorsremains constant after the first three measuring

Fig. 10. Number of sensors that were used for specific times (or not used) over the testing period of the adaptive design at each measuringcycle, as clusterhead nodes (a), in high-signal range operating mode (b), in low-signal range operating mode (c), and as active nodes ingeneral (d).


cycles, and the same holds for the allocation of theactive nodes into HSR and LSR operating modes,while there is a slight decrease in the number ofCH sensors, which leads to the general increaseof the average number of active sensors coordi-nated by each clusterhead. However, these averagevalues are much smaller than the actual capabilityof clusterhead sensors (15 sensors), which leadsto the conclusion that less clusterheads could beused, but the energy conservation of the operat-

Table 8Distribution of operating modes of sensors and average clustering

Measuring cycle CHs HSR LSR T

1 133 275 291 62 125 273 302 73 119 276 298 64 98 253 258 65 107 229 292 66 103 235 264 67 93 237 278 68 91 234 275 69 88 227 287 610 83 220 293 511 86 238 276 612 84 234 281 513 87 224 287 514 75 225 262 515 82 219 296 5

ing cost of such a design would have beenneglected by the increase in communication energyconsumption.

Fig. 11 shows the percentage of sensors (over theentire grid of 900 sensors) with battery capacitiesbelow certain percentage-levels after each measuringcycle, based on the assumption that all sensors had100% battery capacity at the beginning of the firstmeasuring cycle. It is clear that the percentage ofsensors with battery capacity below 40% is kept very

otal active Inactive Avg. sensors/CH (std’s)

99 201 4.26 (1.80)00 200 4.60 (2.08)93 207 4.82 (2.03)09 291 5.21 (2.23)28 272 4.87 (2.11)02 298 4.84 (2.26)08 292 5.54 (2.38)00 300 5.59 (2.14)02 298 5.84 (2.25)96 304 6.18 (2.44)00 300 5.98 (2.89)99 301 6.13 (2.53)98 302 5.87 (2.74)62 338 6.49 (2.78)97 303 6.28 (2.35)

Fig. 11. Percentages of sensors with battery capacities belowcertain values (as percentages of full battery capacity) at the endof each measuring cycle of the adaptive WSN design.


low during the 15 measuring cycles, even while atthe end of the 15th measuring cycle there is no sen-sor with battery capacity below 20%. Correspond-ing results on the analysis of remaining sensorswith battery capacities above certain percentage-levels also showed high conservation of energyresources.

7. Conclusions

In this paper, we presented an algorithm for theoptimal design and dynamic adaptation of applica-tion-specific WSNs, based on the evolutionary opti-mization properties of genetic algorithms. A fixedwireless network of sensors of different operatingmodes was considered on a grid deployment andthe GA system decided which sensors should beactive, which ones should operate as clusterheadsand whether each of the remaining active normalnodes should have high or low-signal range. Duringoptimization, parameters of network connectivity,energy conservation as well as application require-ments were taken into account so that an integratedoptimal WSN was designed. From the evolution ofnetwork characteristics during the optimizationprocess, we can conclude that it is preferable tooperate a relatively high number of sensors andachieve lower energy consumption for communica-tion purposes than having less active sensors withconsequently larger energy consumption for com-munication purposes. In addition, GA-generateddesigns compared favorably to random designs ofsensors. Uniformity of sensing points of optimal

designs was satisfactory, while connectivity con-straints were met and operational and communica-tion energy consumption was minimized.

We also showed that dynamic application of thealgorithm in adaptive WSN design can lead to theextension of the network’s life span, while keepingthe application-specific properties of the networkclose to optimal values. The algorithm showedsophisticated characteristics in the decision of sen-sors’ activity/inactivity schedule as well as the rota-tion of operating modes (clusterhead or ‘‘regularsensor’’ with either high or low-signal range), whichled to considerable energy conservation on availablebattery resources.

Future work will deal with the development ofheuristic methodologies for optimal routing ofdynamically selected clusterhead sensors, throughsome multi-hop communication protocol.

Acknowledgements

This work was supported in part by the‘‘PYTHAGORAS-II’’ research project which isco-funded by the European Social Fund and Greeknational resources (EPEAEK II).

References

[1] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci,Wireless sensor networks: a survey, Computer Networks 38(2002) 393–422.

[2] S. Slijepcevic, M. Potkonjak, Power efficient organization ofwireless sensor networks, in: Proc. IEEE Int. Conf. onCommunications, Helsinki, Finland, 2001, pp. 472–476.

[3] B. Krishnamachari, F. Ordonez, Analysis of energy-efficient,fair routing in wireless sensor networks through non-linearoptimization, in: Proc. IEEE Vehicular Technology Confer-ence – Fall, Orlando, FL, 2003, pp. 2844–2848.

[4] A. Trigoni, Y. Yao, A. Demers, J. Gehrke, R. Rajaraman,WaveScheduling: energy-efficient data dissemination forsensor networks, in: Proc. Int. Workshop on Data Manage-ment for Sensor Networks (DMSN), in conjunction withVLDB, 2004.

[5] V. Mhatre, C. Rosenberg, D. Kofman, R. Mazumdar, N.Shroff, A minimum cost heterogeneous sensor network witha lifetime constraint, IEEE Trans. Mobile Comput. 4 (1)(2005) 4–15.

[6] S. Ghiasi, A. Srivastava, X. Yang, M. Sarrafzadeh, Optimalenergy aware clustering in sensor networks, Sensors 2 (2002)258–269.

[7] V. Rodoplu, T.H. Meng, Minimum energy mobile wirelessnetworks, IEEE J. Select. Areas Commun. 17 (8) (1999)1333–1344.

[8] W.R. Heinzelman, A. Chandrakasan, H. Balakrishnan,Energy-efficient communication protocol for wireless


microsensor networks, in: Proc. 33rd Hawaii Int. Conf. onSystem Sciences, Maui, Hawaii, 2000.

[9] J.-H. Chang, L. Tassiulas, Energy conserving routing inwireless ad-hoc networks, in: Proc. IEEE INFOCOM’00, TelAviv, Israel, 2000, pp. 22–31.

[10] D.J. Chmielewski, T. Palmer, V. Manousiouthakis, On thetheory of optimal sensor placement, AlChE J. 48 (5) (2002)1001–1012.

[11] C. Zhou, B. Krishnamachari, Localized topology generationmechanisms for wireless sensor networks, in: IEEE GLO-BECOM’03, San Francisco, CA, December 2003.

[12] K.P. Ferentinos, T.A. Tsiligiridis, Heuristic dynamic clus-tering in wireless sensor networks towards uniform sensing,in: 15th IST Mobile & Wireless Communications Summit,Myconos, Greece, June 2006.

[13] J.H. Holland, Adaptation in natural and artificial systems,University of Michigan Press, Ann Arbor, 1975.

[14] S. Sen, S. Narasimhan, K. Deb, Sensor network design oflinear processes using genetic algorithms, Comput. Chem.Eng. 22 (3) (1998) 385–390.

[15] D. Turgut, S.K. Das, R. Elmasri, B. Turgut, Optimizingclustering algorithm in mobile ad hoc networks using geneticalgorithmic approach, in: IEEE GLOBECOM’02, Taipei,Taiwan, November 2002.

[16] G. Heyen, M.-N. Dumont, B. Kalitventzeff, Computer-aideddesign of redundant sensor networks, in: Escape 12, TheAague, The Netherlands, May 2002.

[17] S. Jin, M. Zhou, A.S. Wu, Sensor network optimizationusing a genetic algorithm, in: 7th World Multiconferenceon Systemics, Cybernetics and Informatics, Orlando, FL,2003.

[18] S.A. Aldosari, J.M.F. Moura, Fusion in sensor networkswith communication constraints, in: Information Process-ing in Sensor Networks (IPSN’04), Berkeley, CA, April2004.

[19] D.B. Jourdan, O.L. de Weck, Layout optimization for awireless sensor network using a multi-objective geneticalgorithm, in: IEEE Semiannual Vehicular TechnologyConference, Milan, Italy, May 2004.

[20] K. Sohrabi, J. Gao, V. Ailawadhi, G.J. Pottie, Protocols forself-organization of a wireless sensor network, IEEE Per-sonal Commun. Mag. 7 (5) (2000) 16–27.

[21] M. Younis, M. Youssef, K. Arisha, Energy-aware routing incluster-based sensor networks, in: 10th IEEE/ACM Inter-national Symposium on Modeling, Analysis and Simulationof Computer and Telecommunication Systems (MASCOTS2002), Fort Worth, TX, October 2002.

[22] S. Bandyopadhyay, E.J. Coyle, An energy efficient hierar-chical clustering algorithm for wireless sensor networks, in:IEEE INFOCOM 2003, San Francisco, CA, April 2003.

[23] O. Younis, S. Fahmy, Distributed clustering in ad-hoc sensornetworks: a hybrid, energy-efficient approach, in: INFO-COM 2004, Hong Kong, March, 2004.

[24] Z. Michalewicz, D.B. Fogel, How to Solve It: ModernHeuristics, Springer-Verlag, Berlin, Germany, 2002.

[25] M.R. Gary, D.S. Johnson, Computers and Intractability,Freeman, San Francisco, CA, 1979.

[26] C. Monma, S. Suri, Transitions in geometric minimumspanning trees, Discrete Comput. Geometry 8 (3) (1992).

[27] D.E. Goldberg, Genetic Algorithms in Search, Optimizationand Machine Learning, Addison-Wesley, Reading, MA,1989.

[28] K.P. Ferentinos, T.A. Tsiligiridis, Evolutionary energymanagement and design of wireless sensor networks, in:IEEE SECON 2005, Santa Clara, CA, September 2005.

Konstantinos P. Ferentinos received theB.Sc./M.Sc. degree in AgriculturalEngineering from the Agricultural Uni-versity of Athens, Greece, in 1997 andthe M.S. and Ph.D. degrees from CornellUniversity in 1999 and 2002 respectively,in Biological and Environmental Engi-neering, with a minor in ComputerScience. During the academic year 2003–2004 he was a postdoctoral researcher inthe Department of Biological and Envi-

ronmental Engineering at Cornell University, Ithaca, NY, USA.Currently he is a postdoctoral fellow in the Informatics Labo-
ratory at the Agricultural University of Athens and a visitinglecturer in the Department of Mathematics at the University ofAthens, Greece. His research interests include applications ofArtificial Intelligence in Controlled Environment Agriculture(CEA), wireless sensor networks, neural network modeling, faultdetection and evolutionary and biologically inspired optimizationalgorithms. He is a member of IEEE, INNS and ASAE.
Theodore A. Tsiligiridis received the B.Scin mathematics from the University ofAthens, Greece in 1976, his M.Sc inprobability and statistics from the Man-chester-Sheffield University, UK, and hisPh.D in telecommunications from theUniversity of Strathclyde, Glasgow,Scotland, UK in 1989. Shortly after hisgraduation he joined the Computer Sci-ence and Mathematics Division of theAgricultural University of Athens

(AUA), where he is currently a Professor. He has worked invarious public and academic posts and he coordinated many
research and development projects. He actively involved in theprojects RACE I/II/EC (Advanced Telecommunications),DELTA/EC (Distance Learning), and ORA/EC (Tele-services inRural Areas). He particularly worked in the areas of mobilecellular systems (R1044/RACE I/EC; 1987), performance evalu-ation on LANs (RSRE/MOD; 1988), and tele-services in ruralareas (R2022/RACE II/EC; 1992, DART/TAP/EC 1996). Healso participated in the EC project: GREEK PLAN forRestructuring Agricultural Surveys (Decisions: 85/360, 90/386,92/587 of the EU Council of Ministers), coordinating many of itsactivities. He is currently coordinating two major projects. Thefirst is under the auspice-supervision of Eurostat and theNational Statistical Service of Greece (EUROFARM/EU pro-ject, 1999/2000 Farm Structure Survey). The second is under theauspice-supervision of the European GoDigital/EU project, inwhich he is responsible for introducing internet services ande-commerce practices in around 3000 Small Medium Enterprises(SMEs).
He is a member of IEEE, ACM, Mathematical Society, theStatistical Institute, and the OR Society. His research inter-ests include traffic modeling and performance evaluation of


broadband, high-speed networks, wireless multimedia commu-nication. He is currently working in medium access control,routing, congestion and flow control, scheduling, optimal designmethods, security and multimedia quality of service applied onLANs, TCP/IP, ATM, including wired, wireless, mobile, cellular

and sensor networks, as well as new switching technologies. Healso works in the area of tele-services, interactive multimediaapplications, geographical information systems, location basedservices, and e-commerce, particularly applied in environmentalscience.

adaptive design optimization of wireless sensor networks...

Documents