Movement-Assisted Sensor Deployment

Transcription

1 640 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 Movement-Asssted Sensor Deployment Gulng Wang, Student Member, IEEE, Guohong Cao, Member, IEEE, and Thomas F. La Porta, Fellow, IEEE Abstract Adequate coverage s very mportant for sensor networks to fulfll the ssued sensng tasks. In many workng envronments, t s necessary to make use of moble sensors, whch can move to the correct places to provde the requred coverage. In ths paper, we study the problem of placng moble sensors to get hgh coverage. Based on Vorono dagrams, we desgn two sets of dstrbuted protocols for controllng the movement of sensors, one favorng communcaton and one favorng movement. In each set of protocols, we use Vorono dagrams to detect coverage holes and use one of three algorthms to calculate the target locatons of sensors f holes exst. Smulaton results show the effectveness of our protocols and gve nsght on choosng protocols and calculaton algorthms under dfferent applcaton requrements and workng condtons. Index Terms Moble sensor networks, sensor coverage, dstrbuted algorthm. æ 1 INTRODUCTION WIRELESS sensor networks can greatly enhance our capablty to montor and control the physcal envronment. Sensor networks are revolutonzng the tradtonal methods of data collecton, brdgng the gap between the physcal world and the vrtual nformaton world [11], [15], [24], [28]. Sensor nodes must be deployed approprately to reach an adequate coverage level for the successful completon of the ssued sensng tasks [5], [21]. In many potental workng envronments, such as remote harsh felds, dsaster areas, and toxc urban regons, sensor deployment cannot be performed manually. To scatter sensors by arcraft s one possble soluton. However, usng ths technque, the actual landng postons cannot be controlled because of the exstence of wnd and obstacles, such as trees and buldngs. Consequently, the coverage may be nferor to the applcaton requrements no matter how many sensors are dropped. Moreover, n many cases, such as durng n-buldng toxc leaks [12], [13], chemcal sensors must be placed nsde a buldng from the outsde. In these scenaros, t s necessary to make use of moble sensors, whch can move to the correct places to provde the requred coverage. One example of a moble sensor s the Robomote [26]. These sensors are smaller than 0: m 3 and cost less than $150. Most prevous research efforts on deployng moble sensors are based on centralzed approaches. For example, the work n [30] assumes that a powerful cluster head s avalable to collect the sensor locatons and determne the target locatons of the moble sensors. However, n many sensor deployment envronments such as dsaster areas and battlefelds, a central server may not be avalable. It may also be hard to organze sensors nto clusters due to network parttons. Further, centralzed approaches ntroduce a sngle pont of falure. Sensor deployment has also been addressed n the feld of robotcs [12], where sensors are deployed teratvely one by one, utlzng the locaton nformaton obtaned from the prevous deployment. Snce sensors are deployed one by one, the deployment tme s very long, whch can sgnfcantly ncrease the network ntalzaton tme. In ths paper, we propose two sets of dstrbuted protocols for controllng the movement of sensors to acheve target coverage: basc protocols and vrtual movement protocols. In the basc protocols, sensors move teratvely, eventually reachng the fnal destnaton. In each teraton, sensors detect coverage holes usng a Vorono dagram. If holes exst, they calculate the target locatons to heal the holes and move. In the vrtual movement protocols, sensors do not perform teratve physcal movement. Instead, after calculatng the target locatons, sensors move vrtually and exchange these new vrtual locatons wth the sensors whch would be ther neghbors f they had actually moved. The real movement only occurs when the communcaton cost to reach ther logcal neghbors s too hgh or when they determne ther fnal destnatons. In both the basc and vrtual movement protocols, three algorthms, VEC, VOR, and Mnmax, are proposed to calculate the target locatons f coverage holes exst. In VEC, sensors move away from a dense area; n VOR, sensors mgrate towards holes; n Mnmax, sensors also move towards holes, but more conservatvely wth the consderaton of not generatng new holes. Smulaton results show that our dstrbuted protocols are effectve n terms of coverage, deployment tme and movement. The rest of the paper s organzed as follows. Secton 2 ntroduces some prelmnares. In Secton 3, we present the basc self-deployment protocols and, n Secton 4, we present the vrtual movement protocols. Secton 5 evaluates the performance of the proposed protocols. Based on the smulaton results, we justfy our desgn and dscuss future work n Secton 6.. The authors are wth the Department of Computer Scence and Engneerng, The Pennsylvana State Unversty, Unversty Park, PA E-mal: {guwang, gcao, tlp}@cse.psu.edu. Manuscrpt receved 17 Jan. 2005; revsed 5 June 2005; accepted 19 Aug. 2005; publshed onlne 17 Apr For nformaton on obtanng reprnts of ths artcle, please send e-mal to: tmc@computer.org, and reference IEEECS Log Number TMC PRELIMINARIES 2.1 Localzaton Technques Locaton awareness s mportant for wreless sensor networks snce many applcatons such as envronment montorng and target trackng depend on knowng the locatons of sensor nodes. Due to the ad hoc nature of such /06/$20.00 ß 2006 IEEE Publshed by the IEEE CS, CASS, ComSoc, IES, & SPS

2 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 641 Fg. 1. (a) Vorono dagram. (b) Vorono polygon G 0 of s 0. networks, each node must determne ts locaton through a locaton dscovery process. For outdoor systems, the Global Postonng System (GPS) [3] s one method for ths purpose. GPS may not be cost effectve or work well ndoors. Many technques have been proposed to enable each node to determne ts locaton ndoors wth only lmted communcaton wth nearby nodes. Most of these methods explot receved sgnal strength [22], tme dfference of arrval of two dfferent sgnals [25], and angle of arrval [7]. Hu and Evans [14] have provded detaled dscusson of these technques. In subsequent dscusson n ths paper, we assume that sensor nodes know ther locatons. 2.2 Path Plannng In systems that explot moble sensors, fndng paths on whch these moble sensors can move to desrng destnatons, especally when there exst obstacles n the feld, s an mportant problem. The problem has been studed n the area of robotcs [6], [17]. Recently, L et al. [18] studed the problem n sensor networks. They combned the above methods to fnd the best moton path and modfed them to explot the dstrbuted nature of sensor networks. In ths paper, we do not study ths problem further; we assume that moble sensors can move to any locaton where they are asked to move based on the exstng technques. We comment more on the mpact of ths assumpton n Secton Sensng Model Each type of sensor has ts unque sensng model characterzed by ts sensng area, resoluton, and accuracy. The sensng area depends on multple factors such as the strength of the sgnals generated at the source, the dstance between the source and the sensor, the attenuaton rate n propagaton, and the desred confdence level of sensng. Let us consder an applcaton [8] n whch a network of acoustc sensors s deployed for detectng moble vehcles. Due to sgnal attenuaton, sensors closer to a vehcle can detect hgher strength of acoustc sgnals than sensors farther away from the vehcle and, thus, have hgher confdence for detectng the vehcle. Therefore, gven a confdence level, we can derve a sensng range surroundng each sensor. In ths paper, we only consder the sotropc sensng models. Each sensor node s assocated wth a sensng area whch s represented by a crcle wth the same radus. Ths s a common assumpton when comparng algorthms for sensng coverage [20], [21]. 2.4 Vorono Dagram The Vorono dagram [4], [9] s an mportant data structure n computatonal geometry. It represents the proxmty nformaton about a set of geometrc nodes. The Vorono dagram of a collecton of nodes parttons the space nto polygons. Every pont n a gven polygon s closer to the node n ths polygon than to any other node. Fg. 1a s an example of the Vorono dagram, and Fg. 1b s an example of a Vorono polygon. We defne the Vorono polygon of s 0 as G 0 ¼hV 0 ; E 0, where V 0 s the set of Vorono vertces of s 0, and E 0 s the set of Vorono edges. As shown n Fg. 1b, V 0 ¼fV 1 ;V 2 ;V 3 ;V 4 ;V 5 g, and E 0 ¼fV 1 V 2 ;V 2 V 3 ;V 3 V 4 ;V 4 V 5 ;V 5 V 1 g. We use N 0 to denote the set of Vorono neghbors of s 0. In Fg. 1b, N 0 ¼fs 1 ;s 2 ;s 3 ;s 4 ;s 5 g. The Vorono edges of s 0 are the vertcal bsectors of the lne passng s 0 and ts Vorono neghbors, e.g., V 1 V 5 s s 0 s 1 s bsector. Our sensor deployment protocols are based on Vorono dagrams. As shown n Fg. 1, each sensor, represented by a number, s enclosed by a Vorono polygon. These polygons together cover the target feld. The ponts nsde one polygon are closer to the sensor nsde ths polygon than the sensors postoned elsewhere. If ths sensor cannot detect the expected phenomenon n ts Vorono polygon, no other sensor can detect t. Therefore, to examne coverage holes, each sensor only needs to check ts own Vorono

3 642 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 Fg. 2. Snapshot of the executon of VEC. (a) Round 0. (b) Round 1. (c) Round 2. polygon. If ts sensng area cannot cover the polygon, there are some coverage holes. To construct the Vorono polygon, sensors frst calculate the bsectors of ther neghbors and themselves. These bsectors (and possbly the boundary of the target feld) form several polygons. The smallest polygon encrclng the sensor s the Vorono polygon of ths sensor. 2.5 Sensng Range versus Communcaton Range In a dstrbuted case, sensors can exchange the locaton nformaton by broadcastng. It s possble that some Vorono neghbors of a sensor are out of ts communcaton range, and, consequently, the calculated polygon of ths sensor s not accurate. If the sensng range s much shorter than the communcaton range, then the naccurate constructon of Vorono cells wll not affect the detecton of coverage holes. Ths s because f Vorono neghbors cannot reach each other by drect communcaton, ther dstance s large enough that there s a coverage hole. If communcaton range s smlar to the sensng range, sensors may msdetect coverage holes. We descrbe our heurstcs to deal wth the naccurate constructon of the Vorono polygons n Secton 3. 3 THE BASIC DEPLOYMENT PROTOCOLS Our deployment protocol runs teratvely. In each round, sensors frst broadcast ther locatons and construct ther Vorono polygons based on the receved neghbor nformaton. Sensors then determne the exstence of coverage holes by examnng ther Vorono polygons. If any hole exsts, sensors calculate where to move to elmnate or reduce the sze of the coverage hole. Three algorthms are proposed to calculate the target locatons: VEC pushes sensors away from a densely covered area, VOR pulls sensors to the sparsely covered area, and Mnmax moves sensors to the center of ther Vorono polygon. Termnaton condtons are defned for each algorthm. 3.1 The VECtor-Based Algorthm (VEC) VEC s motvated by the attrbutes of electromagnetc partcles: When two electromagnetc partcles are too close to each other, an expellng force pushes them apart. Assume dðs ;s j Þ s the dstance between sensor s and sensor s j. d ave s the average dstance between two sensors when the sensors are evenly dstrbuted n the target area, whch can be calculated beforehand snce the target area and the number of sensors to be deployed are known. The vrtual force between two sensors s and s j wll push them to move ðd ave dðs ;s j ÞÞ=2 away from each other. In case one sensor covers ts Vorono polygon completely and should not move, the other sensor wll be pushed d ave dðs ;s j Þ away. In summary, the vrtual force wll push the sensors d ave away from each other f a coverage hole exsts n ether of ther Vorono polygons. The vrtual force exerted by s j on s s denoted as ~F j, wth the drecton from s j to s. In addton to the vrtual forces generated by sensors, the feld boundary also exert forces, denoted as ~F b, to push sensors too close to the boundary nward. ~F b exerted on s wll push t to move d ave =2 d b ðs Þ, where d b ðs Þ s the dstance of s to the boundary. Snce d ave s the average dstance between sensors, d ave =2 s the dstance from the boundary to the sensors closest to t when sensors are evenly dstrbuted. The fnal overall force on sensors s the vector summaton of vrtual forces from the boundary and all Vorono neghbors. These vrtual forces wll push sensors from the densely covered area to the sparsely covered area. Thus, VEC s a proactve algorthm, whch tres to relocate sensors to be evenly dstrbuted. As an enhancement, we add a movement-adjustment scheme to reduce the error of vrtual-force. When a sensor determnes ts target locaton, t checks whether the local coverage wll be ncreased by ts movement. The local coverage s defned as the coverage of the local Vorono polygon and can be calculated by the ntersecton of the polygon and the sensng crcle. If the local coverage s not ncreased, the sensor should not move to the target locaton. Although the general drecton of the movement s correct, the local coverage may not be ncreased because the target locaton s too far away. To address ths problem, the sensor wll choose the mdpont or 3=4 pont between ts target locaton and ts current locaton as ts new target locaton. If the local coverage s ncreased at the new target locaton, the sensor wll move; otherwse, t wll stay. Fg. 2 shows an operatonal example of VEC. Round 0 s the ntal random deployment of 35 sensors n a 50 m 50 m flat space, wth the sensng range of sx meters

4 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 643 Fg. 5. Inaccurate Vorono polygon. Fg. 3. The VEC protocol at sensor s. and communcaton range of 20 meters. The ntal coverage s 75.7 percent. After Round 1 and Round 2, the coverage s mproved to 92.2 percent and 94.7 percent, respectvely. A formal descrpton of the VEC algorthm s shown n Fg The VORono-Based Algorthm (VOR) Contrary to the VEC algorthm, VOR s a pull algorthm whch pulls sensors to cover ther local maxmum coverage holes. In VOR, f a sensor detects the exstence of coverage holes, t wll move toward ts farthest Vorono vertex (denoted as V far ), and stop when the farthest Vorono vertex can be covered. As n VEC, n VOR, a sensor needs only to check ts own Vorono polygon. Fg. 4 llustrates VOR. Pont A s the farthest Vorono vertex of s 0, and dða; s 0 Þ s longer than the sensng range. To heal the hole, s 0 moves along lne s 0 A to Pont B, where dða; BÞ s equal to the sensng range. We lmt the maxmum movng dstance to be at most half of the communcaton range mnus the sensng range to avod the stuaton shown n Fg. 5, n whch s 0 s not aware of the exstence of s 1 because of communcaton lmtatons. When s 0 does not know s 1, t wll calculate ts local Vorono polygon as the dotted one and vew the area around A as a coverage hole. If s 0 moves toward pont A and stops at a dstance dða; BÞ (sensng range), apparently s 0 has moved more than needed and t tres to cover the area whch s already covered by s 1. It s qute possble t has to move back after t gets to know s 1. Therefore, we set the maxmum movng dstance such that a sensor moves towards the coverage hole step-by-step. After s 0 moves a certan dstance and gets closer to s 1, t can communcate wth s 1 and calculate the correct Vorono polygon. Then the rsk of movng oscllaton can be greatly reduced. VOR s a greedy algorthm whch tres to fx the largest hole. Movng oscllatons may occur f new holes are generated due to a sensor s leavng. To deal wth ths problem, we add oscllaton control whch does not allow sensors to move backward mmedately. Before a sensor moves, t frst checks whether ts movng drecton s opposte to that n the prevous round. If yes, t stops for one round. In addton, the movement adjustment mentoned n VEC s also appled here. The deployment protocol usng VOR s smlar to the VEC Protocol, except that n lne (2.2) VECðÞ s replaced by VORðÞ, whch s shown below. We run VOR on the same ntal deployment as shown n Fg. 2a. After round 1 and round 2, the coverage s mproved to 89.2 percent and 95.6 percent, respectvely. Fg. 4. VOR. 3.3 The Mnmax Algorthm Smlarly to VOR, Mnmax fxes holes by movng closer to the farthest Vorono vertex, but t does not move as far as VOR to avod stuatons n whch a vertex that was

5 644 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 orgnally close becomes a new farthest vertex. Mnmax chooses the target locaton as the pont nsde the Vorono polygon whose dstance to the farthest Vorono vertex (V far ) s mnmzed. We call ths pont the Mnmax pont, denoted as O m. Ths algorthm s based on the belef that a sensor should not be too far away from any of ts Vorono vertces when the sensors are evenly dstrbuted. Mnmax can reduce the varance of the dstances to the Vorono Vertces, resultng n a more regular shaped Vorono polygon, whch better utlzes the sensor s sensng crcle. Compared wth VOR, Mnmax consders more nformaton and t s more conservatve. Compared wth VEC, Mnmax s reactve ; t fxes the hole more drectly by movng toward the farthest Vorono vertex. The Mnmax pont s the center of the smallest enclosng crcle of the Vorono vertces and can be calculated by the algorthms descrbed n [19], [27], [29]. In the deployment protocol usng Mnmax, we also specfy the maxmum movng dstance and do oscllaton control as n VOR. We run Mnmax on the same ntal deployment as shown n Fg. 2a. After round 1 and round 2, the coverage s mproved to 92.7 percent and 96.5 percent, respectvely. 3.4 Termnaton The algorthm termnates naturally based on the movementadjustment heurstc (explaned n Secton 3.1), whch does not allow sensors to move unless the local coverage can be ncreased. The total coverage, bounded by 100 percent, ncreases as the local coverage ncreases. Based on the attrbutes of Vorono dagram, the local coverage ncrease of one sensor does not affect the local coverage of another sensor. Thus, sensors wll stop naturally when the coverage cannot be ncreased. The formal proof s shown n the Appendx. In some applcatons, the coverage requrement may be met wthout achevng maxmum coverage. In these cases, t may be prudent to termnate the deployment process before the maxmum coverage s reached to save power and reduce the deployment tme. To termnate the deployment procedure earler, we use a threshold, defned as the mnmum ncrease n coverage below whch a sensor wll not move. Wth a larger, the deployment wll fnsh earler. When ¼ 0, sensors stop when the best coverage s obtaned. 3.5 Optmzatons Dealng Wth Message Loss Hello messages may be lost due to collsons. Consequently, sensors may fal to know the exstence of some Vorono neghbors and mstakenly determne coverage holes. To address ths problem, we assocate each tem n a sensor s neghbor lst wth a number whch ndcates the freshness of ths tem, that s, for how many rounds ths neghbor has not been heard. When constructng the Vorono polygon, sensors only consder the sensors n ts neghbor lst wth certan freshness. For example, only sensors that have been heard wthn the last two rounds can be consdered when constructng the Vorono polygon. Supposng the probablty of message loss s 5 percent, the probablty that a message s lost two consecutve tmes s 0.25 percent. Therefore, f a sensor does not hear a Hello message from a neghbor for two consecutve cycles, t can assume that the neghbor has moved and be correct wth a percent chance. Fg. 6. Workng procedure (VOR). Ths soluton ntroduces a new problem. If a sensor actually moves to a new place, ts prevous neghbors cannot hear t. If these old neghbors stll consder ths sensor n ther formaton of the Vorono polygons untl the freshness threshold s volated, t wll prolong the deployment process. To address ths problem, we propose that a sensor broadcasts ts new locaton before t moves so that ts neghbors can react promptly f a new hole s generated by ts leavng Dealng wth Poston Clusterng In some cases, the ntal deployment of sensors may form clusters, as shown n Fg. 6, resultng n low ntal coverage. In ths case, sensors located nsde the clusters cannot move for several rounds, snce ther Vorono polygons are well covered ntally. Ths problem prolongs the deployment tme, as s shown n Fg. 6, n whch some sensors are stll clustered together after the sxth round. To reduce the deployment tme n ths stuaton, we propose an optmzaton whch detects whether too many sensors are clustered n a small area. The algorthm explodes the cluster to scatter the sensors apart. Each sensor compares ts current neghbor number to the neghbor number t wll have f sensors are evenly dstrbuted. If a sensor fnds the rato of these two numbers s larger than a threshold, t concludes that t s nsde a cluster and chooses a random poston wthn an area centered at tself whch wll contan the same number of sensors as ts current neghbors n the even dstrbuton. The exploson algorthm only runs n the frst

6 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 645 round. It scatters the clustered sensors and changes the deployment to be close to random. 4 DEPLOYMENT PROTOCOLS WITH VIRTUAL MOVEMENT The basc protocols requre sensors to move teratvely, eventually reachng the fnal destnaton. Other approaches can be envsoned n whch the sensors move only once to ther destnaton to mnmze the sensor movement. One such approach s to let sensors stay fxed and obtan ther fnal destnatons by smulated movement. Wth the same round-by-round procedure, sensors calculate ther target locatons, vrtually move there, and exchange these new vrtual locatons wth the sensors whch would be ther neghbors as f they had actually moved. The real movement only happens at the last round after fnal destnatons are determned. We dd not deploy ths alternatve method for two reasons. Frst, ths approach s susceptble to poor performance under network parttons, whch are lkely to occur n a sensor deployment. If a network partton occurs, each partton wll exercse the movement algorthms wthout knowledge of the others. Consequently, the obtaned fnal destnaton s not accurate and the requred coverage cannot be reached. Usng real movement, the network parttons wll be healed allowng all sensors to be eventually consdered n the algorthm. A second reason s the hgh communcaton overhead. To guarantee logcal neghbors are reached, a network-wde broadcast s needed when usng smulated moblty. If ths network-wde broadcast s mplemented by gosspng, the message complexty s at mnmum 2rn 2 (here, r s the number of rounds needed and n s the number of sensors n the network). Usng actual moblty as n the basc protocols, a much lower message complexty, 2rn, s suffcent. To get balance between movement and message complexty, we propose to let sensors do vrtual movement when the communcaton cost to reach the logcal Vorono neghbors s reasonable, and physcal movement otherwse. The challenge s to determne f a sensor can reach ts logcal neghbors wth reasonable communcaton cost. We propose the followng heurstcs. Frst, f a sensor s dstance to ts farthest Vorono vertex s shorter than half of the communcaton range, t must know all ts Vorono neghbors. In ths case, one hop broadcast (same n the basc protocols) s enough to exchange the locaton nformaton wth ts logcal neghbors and physcal movement s not necessary. Otherwse, t s possble that some Vorono neghbors are out of the communcaton range. To get the locatons of these Vorono neghbors, sensors request ther neghbors wthn the communcaton range to broadcast ther neghbor lsts, thus obtanng the logcal postons of sensors located wthn two broadcast hops. When the dstances between the physcal locatons of sensors and ther farthest Vorono Vertces are larger than two tmes the maxmum movng dstance, sensors should move physcally. In realzaton, we dvde the dscovery phase nto two subphases. In the frst subphase, sensors broadcast hello messages; n the second subphase, sensors broadcast the locatons of known neghbors. In one round, f a sensor s dstance to ts farthest Vorono vertex s larger than half of the communcaton range, t wll calculate the target locaton as n the basc schemes and do logcal movement. In the next round, t wll set a flag n the hello messages, Fg. 7. Vrtual movement protocols at s. ndcatng that t wants ts neghbors to broadcast ther neghbor lst. Any sensor that receves a hello message wth such a flag wll broadcast ts neghbor lst n the second subphase of the dscovery phase. In ths way, the message complexty s at most two tmes the basc scheme n one round. The flag wll not be reset untl the sensor moves physcally. Sensors move physcally under two condtons: One s that ts physcal poston s two tmes the maxmum movng dstance to ts farthest Vorono vertex, as dscussed above. The second condton s that a sensor s logcal poston has not changed for several rounds. Then, the sensor can determne that t has obtaned ts fnal locaton and t can move. The formal descrpton of the protocol wth the vrtual movement s shown n Fg. 7.

7 646 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 Fg. 8. Coverage. (a) Basc protocols. (b) Vrtual movement. 5 PERFORMANCE EVALUATIONS 5.1 Objectves, Metrcs, and Methodology We mplement our deployment protocols n the ns-2 (verson 2.1b9a), a standard network smulator. Our objectves n conductng ths evaluaton study are threefold: frst, testng the effectveness of our protocols n provdng hgh coverage; second, by comparng VEC, VOR and Mnmax, and comparng the basc protocols and the vrtual movement protocols, gvng some nsght on choosng protocols n dfferent stuatons; fnally, studyng the effectveness of controllng the trade-off among varous metrcs by adjustng parameters. We analyze the performance of our protocols from two aspects: deployment qualty and energy consumpton. Deployment qualty s measured by the sensor coverage and the tme (number of rounds) to reach ths coverage. Deployment tme s determned by the number of rounds needed and the tme of each round. The duraton of each round s prmarly determned by the movng speed of sensors, whch s the mechancal attrbute of sensors. Thus, we only use the number of rounds to measure the deployment tme. Energy consumpton ncludes two parts, mechancal movement and communcaton. Message complexty s used to measure the energy consumed n communcaton. As for movement, the energy consumed n movng a sensor n meters conssts of two parts: startng/brakng energy and movng energy. Therefore, we use movng dstance and the number of movement as the metrcs. We run smulatons under dfferent sensor denstes, whch determnes the sensor coverage that can be reached and the dffculty to reach t. In a 100 m 100 m target feld, we dstrbute four dfferent numbers of sensors, rangng from 120 to 180, n ncrements of 20 sensors. The ntal deployment follows the random dstrbuton. Most smulaton results are under the termnaton condton that s equal to 1 percent dvded by the number of sensors. To evaluate each metrc under dfferent parameter settngs, we run 10 experments based on dfferent ntal dstrbuton and calculate the average results. We choose as the MAC layer protocol and DSDV as the routng protocol. The physcal layer s modeled after the RF MOTE from Berkeley, wth MHz OOK 5 kbps as the bandwdth and 20 meters as the transmsson range. Based on the nformaton from [1], we set the sensng range to be 6 meters. Ths s consstent wth other current sensor prototypes, such as Smart 1 (Unversty of Calforna, Berkeley), CTOS dust, and Wns (Rockwell) [2]. 5.2 Smulaton Results Coverage Fg. 8 shows the coverage obtaned when the coverage ncrease threshold s equal to 1 dvded by the number of sensors. From the fgure, we can see that the coverage s greatly ncreased by all three algorthms compared to the ntal random dstrbuton. For example, when 140 sensors are deployed, Mnmax and VOR can ncrease the coverage to be more than 98 percent from 77.7 percent. In contrast, to obtan the same coverage under random deployment, on average, 340 sensors are requred. Also, by analyzng the trace of the basc protocols, we fnd the coverage ncreases very quckly durng the frst several rounds. For example, n most of the cases, the coverage can be ncreased to more than 85 percent after the frst round when 120 sensors are deployed and more than 90 percent n hgher sensor denstes. In the vrtual movement protocols, actual coverage ncrease happens after the real movement. Among VEC, VOR, and Mnmax, VEC performs the worst. The prmary reason s that VEC s senstve to the ntal deployment. Consder an extreme stuaton n whch sensors are located n the same lne wth equal spacng. In ths case, no sensor wll move, snce the vrtual forces offset each other, though there are large coverage holes. If the sensors are located n smlar relatve postons ntally, VEC does not perform well. In addton, VEC nether consders coverage holes nor utlzes any geometrc nformaton from the Vorono polygons when choosng the target locatons. It tres to reach relatvely balanced postons among the sensors, despte the dffculty of obtanng an exact, global, even dstrbuton from only local nformaton. VOR and Mnmax acheve qute smlar coverage. They both move to heal the holes drectly. VOR s more greedy and may move more than needed, thus generatng new coverage holes. But fnally, VOR wll move sensors back to the correct postons f coverage can be ncreased by dong so.

8 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 647 Fg. 9. Movng dstance. (a) Basc protocols. (b) Vrtual movement. Fg. 10. The number of movements. (a) Basc protocols. (b) Vrtual movement. Between vrtual movement protocols and basc protocols, vrtual movement protocols acheve almost the same coverage as the basc protocols, as expected. We can conclude that usng vrtual movement wll not affect the acheved coverage Energy Consumpton Fg. 9 and Fg. 10 show the movng dstance and the number of movements, respectvely. Fg. 11 shows the message complexty. Here, message complexty s defned as the number of messages exchanged when the protocol termnates. To evaluate the energy consumpton, we normalze the movng dstance and the number of movements nto message complexty. That s, wth the same amount of energy consumed n movement, how many messages can be transmtted. Calculated from Robomote [26], approxmately, to move a sensor one meter consumes a smlar amount of energy as transmttng 300 messages. The energy consumpton n startng/brakng s vared n dfferent systems. Fg. 12 shows the unfed energy consumpton when the startng/brakng to one meter movng energy consumpton rato s 1. (We also have plotted the case when the rato s 4. The results are smlar to Fg. 12, so we do not Fg. 11. Message complexty. (a) Basc protocols. (b) Vrtual movement.

9 648 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 Fg. 12. Unfed energy consumpton: Energy consumed n startng/brakng s equal to movng one meter. (a) Basc protocols. (b) Vrtual movement. show them here.) From the fgure, we can conclude that vrtual movement protocols are much more energy-effcent than the basc protocols. The mprovement s larger when the startng energy s hgh. Among the three algorthms to calculate the target locatons, Mnmax s always the most energy-effcent, except when the sensor densty s qute low. At low sensor densty, VEC consumes the least energy. VEC pushes sensors nto relatvely regular postons and does not perform the fne adjustment of ther locatons to reach hgh coverage. Therefore, VEC consumes the least energy and reaches the lowest coverage among these three algorthms under low densty. Between VOR and Mnmax, Mnmax moves less n most cases. Mnmax s proposed to address the aggressve feature of VOR and the smulaton results verfy ts effectveness. Because Mnmax s senstve to the constructon of Vorono polygon, t s senstve to message loss. In our prevous paper [10], we showed that f message loss s not accounted for, Mnmax has the largest movng dstance. Ths s because when message loss occurs, the calculated Vorono polygon s not correct. From Fg. 9, we can observe an nterestng phenomenon of VEC: The movng dstance s smlar under dfferent sensor denstes. Ths s because VEC fxes coverage holes by pushng sensors nto a relatvely even dstrbuton. In VEC, sensors are pushed by the vrtual forces, whch are determned by the dfference between the average dstance of sensors when they are evenly dstrbuted and the ndvdual nterdstances. Both values ncrease wth a low densty and decrease wth a hgher densty. Thus, VEC s not senstve to sensor densty. In contrast, Mnmax and VOR relocate sensors by measurng the coverage holes, whch are larger under lower densty and smaller under hgh densty. Therefore, the movng dstance n VOR and Mnmax s decreased wth a hgher sensor densty. Fg. 10 shows the number of movements usng basc protocols and vrtual movement protocols, respectvely. We can see that the number of movements s greatly reduced by usng vrtual movement. In partcular, sensors move about once when the number of nodes s more than 140, and less than 1.5 tmes when the number of nodes s 120. Ths shows the effectveness of the heurstc to determne when to move physcally. In the basc protocols, sensors move several tmes on average. They move less wth a hgher node densty and fewer coverage holes. As descrbed n Secton 2.2, path plannng s requred to overcome obstacles. We expect a greater negatve mpact on movng dstance when obstacles must be overcome on the protocols that requre the larger number of movements. Fg. 11 shows the message complexty. Message complexty s prmarly determned by the number of rounds to fnsh the deployment process and the number of messages n each round. Wthn one round, VEC transmts more messages than VOR and Mnmax snce sensors need to send one message to notfy neghbors that ther Vorono polygons are well covered. In addton, VEC needs more rounds to termnate (explaned n the next secton). Therefore, VEC has the hghest message complexty. Between Mnmax and VOR, Mnmax needs more rounds to termnate and has hgher message complexty Convergence Tme In ths secton, we evaluate the convergence tme of our protocols. We set to be 0. Fg. 13 shows the coverage n each round when the number of sensors s 140. From the fgure, we can see that the coverage ncreases very quckly durng the frst several rounds. Here for each algorthm, we run 50 experments. After 10 rounds, the algorthms acheve at least 98 percent of the best coverage they can reach n all these experments, and the algorthms acheve at least 99 percent of the best coverage, n about percent of Fg. 13. Convergence tme.

10 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 649 Fg. 14. Termnaton. (a) Basc protocols. (b) Vrtual movement. these experments. From the experments, we can conclude that our algorthm can quckly converge. Our algorthm resembles the steepest descent algorthm [16], n that both try to move along the drecton that s locally optmal. Unfortunately, the convergence theory of steepest descent s not satsfactory from a theoretcal pont of vew. It s shown n [16] that t converges f the objectve functon satsfes certan condtons and proper step lengths are taken. However, no result on the convergence rate of steepest descent for general objectve functons can be found n the lterature. For most applcatons, steepest descent s the best choce f no global nformaton of the objectve functon s avalable Termnaton Fg. 14 shows deployment tme under dfferent sensor densty. As expected, vrtual movement protocols requre more tme to termnate. In vrtual movement, each sensor wats several rounds before real movement. In general, the deployment procedure can be roughly dvded nto two parts: One s overall redstrbuton, whch moves a group of sensors from a dense area to a sparse area to acheve a relatvely even dstrbuton of sensors. Another s the mnor adjustment of postons n the local area to acheve better coverage. When the sensor densty s not hgh, tme consumed n the overall redstrbuton s the major factor of deployment. When the sensor densty s very hgh, no largescale movement s needed and the poston adjustment s the major factor. VOR s good at the overall redstrbuton because of ts aggressve feature. It termnates the quckest n low or medum sensor densty. Mnmax calculates the target locatons to heal coverage holes most accurately, and t fnshes the quckest n very hgh densty. VEC pushes sensors away from dense area by vrtual force. The sensors n a sparse area may not move for a long tme snce no sensor s present to push them. These sensors only move after the sensors from a dense area are propagated nto ther area. Ths propagaton process may take a long tme Impact of Coverage Increase Threshold In ths secton, we study the effectveness of controllng the trade-off between coverage and deployment tme by adjustng, the coverage ncrease threshold. Table 1 shows the termnaton round, coverage reached, and other metrcs for dfferent for the vrtual movement protocols. We can see that, wth a smaller, a hgher coverage can be reached, whle the deployment cost and the deployment tme are also ncreased. By properly settng ths threshold, we can save tme and energy by tradng off a small amount of coverage. For example, n Mnmax, when s ncreased from 0.5 percent to 1 percent, the deployment tme can be shortened by seven rounds (32 percent), whle the ultmate coverage acheved s only reduced by 0.4 percent. Among VEC, VOR, and Mnmax, the termnaton tme s qute dfferent when s small and smlar when s larger. For example, when s equal to 2 percent, the three algorthms termnate n a smlar tme. As for other metrcs, Mnmax performs the best. Therefore, Mnmax s the best choce f can be set to be a relatvely large number. 6 CONCLUSION AND DISCUSSIONS Ths paper addressed the problem of movng sensors n a target feld to get hgh coverage. Based on Vorono dagrams, we desgned two sets of dstrbuted protocols TABLE 1 Impact of (n ¼ 140) R s the round number when all sensors stop. E measures the effectveness of movng, whch s the rato of actual movng dstance to the dstance between the ntal poston and the fnal poston. M shows the average number of movements of sensors. C and D refer to the coverage and to the movng dstance, respectvely. These values are obtaned n the stoppng round R.

11 650 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 to teratvely move moble sensors from densely deployed areas to sparsely deployed areas. Smulaton results verfed the effectveness of our protocols and provded a baselne for performance under deal condtons. The vrtual movement protocols can sgnfcantly reduce mechancal movement wth a cost of less than two tmes an ncrease n the message complexty over the basc protocols. In each set of the protocols, three algorthms to calculate the target locaton were proposed: VEC, VOR, and Mnmax. Mnmax s the best choce n most cases. VEC moves least at a low sensor densty and can be deployed when the coverage requrement s not hgh. VOR termnates the earlest when the sensor densty s not very hgh, and t can be deployed when both the deployment tme requrement and coverage requrement are strct. Below, we dscuss some open ssues. 6.1 Dstrbuted Scheme versus Centralzed Scheme To address the problem of moble sensor deployment, we propose that sensors calculate ther target locatons n a dstrbuted fashon. We dd not deploy a centralzed approach for the followng reasons. Frst, a central server archtecture may not be feasble n some deployments. Further, the centralzed approach suffers from the problem of a sngle pont of falure. Although a centralzed approach s not feasble for many scenaros, t s nterestng to study t and compare t wth our dstrbuted algorthms. We consder a centralzed approach n an deal case, n whch the optmal postons to place sensors s decded a pror. For n sensors, there are n regular postons. It does not matter whch sensor s placed n whch poston. There exsts a central server, whch can collect the locatons of sensors and drect them to move, such that the average movng dstance s mnmzed. Bascally, ths s a bparty matchng problem and the classc Hungaran method [23] can be used to calculate how to allocate sensors from ther ntal postons to the target locatons such that the movng dstance s mnmzed. We have compared the average movng dstance of our dstrbuted protocols wth ths deal case. When the node densty s lower than about 110 sensors per 100 m 100 m, the centralzed scheme outperforms our scheme; otherwse, our dstrbuted algorthms are better. (Here, we choose Mnmax under vrtual movement for comparson.) Ths s because n our scheme, sensors only move when there are coverage holes, whle n the centralzed scheme, sensors always move to an optmal pont even f t does not mprove coverage. In our scheme, when the node densty s hgh, sensors need only move a short dstance to reach hgh coverage. Under low densty, when most sensors are requred to move, the centralzed scheme results n a lower average movng dstance. In terms of coverage, the centralzed approach can always guarantee the optmal coverage snce the optmal postons are decded a pror. 6.2 Sensng Area In ths paper, the sensng area of each sensor s assumed to be a dsk wth radus 6 m. Ths s the deal case, whch provdes us wth a baselne of the sensor placement problem. In future work, we wll address varyng sensng ranges. Here, we dscuss these ssues. Our protocols can deal well wth the case of a larger or smaller sensng radus f the sensng area s unformly a dsk. The performance of the protocols depends more on the rato of communcaton range to sensng range than the absolute sensng range. As the sensng range decreases wth regard to the communcaton range, our protocols wll perform very well because they can accurately construct the Vorono dagrams. As the sensng range ncreases, we need to enlarge the broadcast hops to better construct the Vorono polygons. If the sensng area s an rregular shape, nstead of a dsk, sensors can stll check ther Vorono polygons to determne the coverage holes. In ths case, we can decrease the sensng range used n our protocols to account for the reduced coverage. In future work, we wll study our protocol s senstvty to the sensng area. 6.3 Senstvty to Communcaton Range In our prevous paper [10], we evaluated the mpact of communcaton range on the basc protocols and found that when communcaton s more than two tmes of the sensng range, the performance s smlar to the results presented here (6 m sensng range and 20 m communcaton range). Ths requrement s reasonable and most hardware can satsfy t. In stuatons n whch ths requrement cannot be satsfed, we can ncrease the broadcast hop lmt. In the vrtual movement protocols, the broadcast hop can be ncreased accordngly f the communcaton range s short and the performance wll not be affected. APPENDIX A We denote the followng terms for the proof. The locaton of sensor s n the rth round s denoted as ½x ðrþ ;y ðrþ Š. The Vorono polygon of s, G, n the rth round s denoted as G ðrþ. G changes n dfferent rounds f s or ts neghbors move. The area of the covered part of G ðrþ n the rth round s denoted as, and that n the ðr þ 1Þth round s denoted ðrþ as ^A. G ðrþ s a fxed polygon n the target feld, but the covered porton of G ðrþ may be changed n dfferent rounds snce sensors may move and they cover dfferent areas f they move. Therefore, ^ s not equal to. Also, ^ s not equal to A ðrþ1þ. They refer to the covered area of dfferent polygons. The area of the covered porton of G ðrþ by a sensor located at ½x; yš s denoted as ð½x; yšþ. The area of the covered porton n the whole target feld n the rth round s total. Lemma 1. (a) total ¼ P n ¼1 AðrÞ. (b) A ðrþ1þ total ¼ P n ðrþ ¼1 ^A. Proof. The Vorono dagram s a partton of the target feld, so (a) s obvous. For the same reason, (b) s also correct. The summaton of the covered area of parttons s the whole covered area n the target feld, whatever a parttonng method s used. Therefore, the summaton of the covered area of the Vorono polygons n the prevous round s also the whole covered area n the current round. tu Lemma 2. ¼ ð½x ðrþ ;y ðrþ ŠÞ. Proof. Ths s the drect result of the attrbute of Vorono dagram. Every pont wthn G ðrþ s closer to ½x ðrþ ;y ðrþ Š than to any other sensor. Any pont not covered by s s also not be covered by any other sensor. tu

12 WANG ET AL.: MOVEMENT-ASSISTED SENSOR DEPLOYMENT 651 Theorem 1. A ðrþ1þ total > total before all sensors stop movng. Proof. At the ðr þ 1Þth round, there may be areas n G ðrþ whch are not covered by s but are covered by other sensors, because the current Vorono polygon of s s G ðrþ1þ but not G ðrþ. Therefore, ^ ð½x ðrþ1þ ;y ðrþ1þ ŠÞ: ð1þ By enforcng the movement adjustment heurstcs (descrbed n Secton 3.1), our algorthms guarantee that, f s moves, ð½x ðrþ1þ ;y ðrþ1þ ŠÞ > Certanly, f s does not move, ð½x ðrþ1þ ;y ðrþ1þ Š¼½x ðrþ because ½x ðrþ1þ ;y ðrþ1þ From (1), (2), and (3), X n ¼1 ^ > Xn ¼1 ŠÞ ¼ ;y ðrþ Š. ð½x ðrþ ð½x ðrþ f some sensor moves n the rth round. From Lemma 2, From (4) and (5), X n ¼1 ¼ ^ A ðrþ1þ total ;y ðrþ ŠÞ: ð2þ ;y ðrþ ŠÞ; ð3þ ð½x ðrþ ;y ðrþ ŠÞ ð4þ ð½x ðrþ ;y ðrþ ŠÞ: ð5þ > Xn ¼1 f some sensor moves n the rth round. By Lemma 1, ( total ¼ P n ¼1 AðrÞ ¼ P n ðrþ ¼1 ^A : From (6) and (7), A ðrþ1þ total > total f some sensor moves n the rth round. Corollary 1. Our dstrbuted algorthms are convergent, and thereby termnate naturally. Proof. Followng from Theorem 1 and the fact that total s upper bounded by the total area of the target feld, our dstrbuted algorthms converge and termnate naturally. All sensors stop movng when no coverage ncrease can happen. tu ACKNOWLEDGMENTS Ths work was supported n part by the US Natonal Scence Foundaton CNS and a grant from MARCO/ DARPA Ggascale Systems Research Center. The authors also want to thank Dr. Sergo Palazzo and the anonymous revewers for ther helpful comments on an early verson of ths paper. Prelmnary results of ths paper have been presented n part at IEEE INFOCOM 2004 n Hong Kong [10]. ð6þ ð7þ ð8þ tu REFERENCES [1] Berkeley Sensor and Actuator Center, berkeley.edu, [2] Wreless Sensng Networks, [3] US Naval Observatory (USNO) GPS Operatons, tycho.usno.navy.ml/gps.html, Apr [4] F. Aurenhammer, Vorono Dagrams A Survey of a Fundamental Geometrc Data Structure, ACM Computng Surveys, vol. 23, pp , [5] T. Clouqueur, V. Phpatanasuphorn, P. Ramanathan, and K.K. Saluja, Sensor Deployment Strategy for Target Detecton, Proc. Frst ACM Int l Workshop Wreless Sensor Networks and Applcatons, [6] D. Kodtschek, Plannng and Control va Potental Functons, Robotcs Rev. I, pp , [7] D. Nculescu and B. Nath, Ad Hoc Postonng Systems (APS) Usng AoA, Proc. IEEE Infocom, [8] F. Zhao and L. Gubas, Wreless Sensor Networks. Morgan Kaufmann, [9] S. Fortune, D. Du, and F. Hwang, Vorono Dagrams and Delaunay Trangulatons, Eucldean Geometry and Computers, [10] G. Wang, G. Cao, and T. La Porta, Movement-Asssted Sensor Deployment, Proc. IEEE Infocom, March [11] W.R. Henzelman, J. Kulk, and H. Balakrshnan, Adaptve Protocols for Informaton Dssemnaton n Wreless Sensor Network, Proc. ACM MobCom, [12] A. Howard, M.J. Matarc, and G.S. Sukhatme, An Incremental Self-Deployment Algorthm for Moble Sensor Networks, Autonomous Robots, specal ssue on ntellgent embedded systems, Sept [13] A. Howard, M.J. Matarc, and G.S. Sukhatme, Moble Sensor Networks Deployment Usng Potental Felds: A Dstrbuted, Scalable Soluton to the Area Coverage Problem, Proc. Sxth Int l Symp. Dstrbuted Autonomous Robotcs Systems, June [14] L. Hu and D. Evans, Localzaton for Moble Sensor Networks, Proc. ACM MobCom, [15] C. Intanagonwwat, R. Govndan, and D. Estrn, Drected Dffuson: A Scalable and Robust Communcaton, Proc. ACM MobCom, [16] J. Nocedal and S.J. Wrght, Numercal Optmzaton. New York: Sprnger, [17] J. Lengyel, M. Rechert, B. Donald, and D. Greenberg, Real-Tme Robot Moton Plannng Usng Rasterzng Computer Graphcs Hardware, Proc. SIGGRAPH, [18] Q. L, M. De Rosa, and D. Rus, Dstrbuted Algorthms for Gudng Navgaton across a Sensor Network, Proc. ACM MobCom, [19] N. Megddo, Lnear-Tme Algorthms for Lnear Programmng n R 3 and Related Problems, SIAM J. Computng, vol. 12, pp , [20] S. Meguerdchan, F. Koushanfar, G. Qu, and M. Potkonjak, Exposure In Wreless Ad-Hoc Sensor Networks, Proc. ACM MobCom, [21] S. Meguerdchan, F. Koushanfar, M. Potkonjak, and M.B. Srvastava, Coverage Problems n Wreless Ad-Hoc Sensor Network, Proc. IEEE Infocom, [22] N. Patwar and A. Hero III, Usng Proxmty and Quantzed RSS for Sensor Locaton n Wreless Locaton n Wreless Networks, Proc. Workshop Wreless Sensor Networks and Applcatons, [23] C.H. Papadmtrou and K. Stegltz, Combnatoral Optmzaton: Algorthms and Complexty. Dover, [24] G.J. Potte and W.J. Kaser, Wreless Integrated Network Sensors, Comm. ACM, May [25] A. Savvdes, C. Han, and M.B. Strvastava, Dynamc Fne- Graned Localzaton n Ad-Hoc Networks of Sensors, Proc. ACM MobCom, [26] G.T. Sbley, M.H. Rahm, and G.S. Sukhatme, Robomote: A Tny Moble Robot Platform for Large-Scale Sensor Networks, Proc. IEEE Int l Conf. Robotcs and Automaton, [27] S. Skyum, A Smple Algorthm for Computng the Smallest Enclosng Crcle, Informaton Processng Letters, vol. 37, pp , [28] K. Sohrab, J. Gao, V. Alawadh, and G.J. Potte, Protocols for Self-Organzaton of a Wreless Sensor Network, IEEE Personal Comm., vol. 7, no. 5, pp , Oct

13 652 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 6, JUNE 2006 [29] E. Welzl, Smallest Enclosng Dsks (Balls and Ellpsods), New Results and New Trends n Computer Scence, pp , [30] Y. Zou and K. Chakrabarty, Sensor Deployment and Target Localzaton Based on Vrtual Forces, Proc. IEEE Infocom, Gulng Wang receved the BS degree from Nanka Unversty, Chna. She s currently pursung the PhD degree n computer scence and engneerng at the Pennsylvana State Unversty. Her research nterests nclude dstrbuted systems, wreless networks, and moble computng, wth a focus on wreless sensor networks. She s a student member of the IEEE. Guohong Cao receved the BS degree from Xan Jaotong Unversty, Xan, Chna, and the MS and PhD degrees n computer scence from the Oho State Unversty n 1997 and 1999, respectvely. Snce then, he has been wth the Department of Computer Scence and Engneerng at the Pennsylvana State Unversty, where he s currently an assocate professor. Hs research nterests are wreless networks and moble computng. He has publshed 100 papers n the areas of sensor networks, cache management, data dssemnaton, resource management, wreless network securty, and dstrbuted fault-tolerant computng. He s an edtor of the IEEE Transactons on Moble Computng and the IEEE Transactons on Wreless Communcatons, a co-guest edtor of a specal ssue on heterogeneous wreless networks of ACM/Kluwer Moble Networkng and Applcatons, and has served on the program commttee of many conferences. He was a recpent of the Presdental Fellowshp at the Oho State Unversty n 1999 and a recpent of the NSF CAREER award n He s a member of the IEEE. Thomas F. La Porta receved the PhD degree n electrcal engneerng from Columba Unversty, New York, NY. He joned the Computer Scence and Engneerng Department at the Pennsylvana State Unversty (Penn State) n 2002 as a full professor. He s the drector of the Networkng and Securty Research Center at Penn State. Pror to jonng Penn State, he had been wth Bell Laboratores snce There, he was the drector of the Moble Networkng Research Department, where he led varous projects n wreless and moble networkng. He s a fellow of the IEEE, a Bell Labs fellow, and receved the Bell Labs Dstngushed Techncal Staff Award n He was the foundng edtor-n-chef of the IEEE Transactons on Moble Computng. He also served as edtor-n-chef of IEEE Personal Communcatons magazne for three years. He has publshed more than 50 techncal papers and holds 28 patents. He was an adjunct member of the faculty for seven years at Columba Unversty, where he taught courses on moble networkng and protocol desgn.. For more nformaton on ths or any other computng topc, please vst our Dgtal Lbrary at