Aggregation in Sensor Networks: An Energy- Accuracy Trade-off Athanassios Boulis, Saurabh Ganeriwal, and Mani B. Srivastava [Networked and Embedded Systems Lab, BE Departeat, Univesity of Califomia at Los Angeles ‘email: { boul, sauraba, mbs }@ee.uelaedu Abaract = Wirtss a hoe sensor networks (WASNS) are n need of the study of seal appicatons that wil belp the researchers ‘ow them as distributed phsialy coupled systems, coletve {hat estinaes the piysial environment, abd. no just energy limited adhoc networks. We develop this perspective ung large and. inering dass of WASN’ applications called errercion aplcatins. In particular, we consider the ‘challenging peiodle aggregation problem where the WASN provides the wer with periodic mats of the eavzoament, 5 Spposed to simpler and previowly studied suupshot aggregation problems. fn periodic aggregation our approach allows the Spatla-tempora correlation mong vals sensed atthe various ‘odes to be exploited towards energy-efficient estimation of the sggregated valve f interest Ove approach also ereates a sytem level energy vs acuracy knob wherey the more the estimation ‘ror thatthe wer ean tolerate, the let ithe energy consumed, We present a sihuted estimation agorchm that canbe applied to explore the energs-acearacysulspace fora suvlass ‘tperioie aggregation problems,” and present extensive ‘mulation resals tbat validate our approach. ‘The resullag iporith, apart from being more Nexibe inte energy-accurac} fulspace and mare robust, cam aso beng coaiderale eens) Savings fora tpleal accuracy requirement (fivefold decense In fenery) consumption Tor 59% esmatlon error) compared. 10 ‘repeated snapshot aggregations -Reyword: sensor neowors, aggregation epplctions, dtribated ‘timation energy, accuracy trade. Lintropucion ‘The technologial advances in embedded computers, sensor, and radios have Ted to the emergence of wireless ad-hoc fensor networks CWASNS) as 3 new class of system with ses in diverse and useful applications. Indeed te early papers in the area (6]7(13]{15} talk aboot the vision of cheap self organizing ad-hoe networks that are able to perform s higher evel sensing task through the collaboration of a large number of cheaper and resource constrained wireless sensor nodes. Leveraging numerous sensing devices placed close to the sctualpiysical phenomena, the information that_ such networks can provide is more accurate and richer than the information povided by a system of few, expensive, sat-of thear sensing devices, Since” WASNS operate largely unatended, ofen in environments where the access cost of eploying or muinaining nodes is high, a key problem in 93) respectively. ‘The global aggregated value should atleast be equal to the ‘current measurement. Hence we should assign zero probability to te points on the Left of x. Thus w(x can be defined a (oves 4-20, IWe> 4,36, Note that xy is equal vo 4-307, where sy and oy represent ‘the mean and variance off) respectively. Figure 4 represents ‘he case, when the mean of local measurement i smaller than ‘he mean of the current estimate. We assign a zero probability 1 points on the let of y. The sssumpaon is thatthe value won't fal ton drastically within a sampling period. Moreover as explained caries, we should asign zero probability to the points on te lef of x. Thus wi) ean be defined as (Ov < max(, ~30,, 44-304) Ix > max(gh, ~30) ls ~302) a= © v0 © | » ‘ounoeD BETRIRLTIONS.Note tht i ult = 30 where and eps themmean and variance off sespetey. ‘We normalize the resultant function appropriately so that it represents a valid probabily distribution. Although we are combining two bounded gaussian distribution, we make an assumption that the obained distribution is sil represented by ‘8 gaussian distribution, This is dane in order to maitan the consistency ofthe estimates, The mean and variance of the ‘ew gaussian distribution represents the new estimate and the confidence (or cenainty) associated with this new estimate respectively Making the decision ‘This subsection corresponds to modules C and D in Figure 1. These modules diced affect the teadeofs one can achieve in the energy-accuracy subspace. When a node calcultes 2 new estimate, it should not necessarily wansmat i. Infact should tansnt only if this new estimate can bring @ significant change in one ofits neighbor estimate ‘This rises a question, how does a node know whether its new estimate is ping t affect the estimates of its neighbors. For this purpose every node mainsin a table whee it stores the most recent estimates received from its neighbors. Let us this able the estimate_sable. A node, after calculating its ‘ew estimate (by using modules A or 2), combines this estimate wath every estimte inthe estimate table, by wsing the sume algorithm as in modole B. The node then calulstes the difference between the new generated estimate for every ‘neighbor and the old estimate, I ths diference is beyond & preset threshold for any ofthe neighboring nodes, the node broadcasts its new estimate. The tveshold is expressed ia terms of faction, fa mmber in [0.1] [Note thatthe threshold should be on both the mean and the variance, So even if the mean remains the same, but the variance is changed significantly, the node should stil broadcast its value. We represent the thresholds on mean and variance as mean; and variance, respectively. In the curtent implementation, we keep both meany and variancey tO the same valve Let us observe closely what does the threshold physically ‘mean. Suppose we set mean; to be 0.05. This means if the ode observes a change greater than 5%, it shall broadcast its estimate. This also means that all changes less than 5% are left unnoticed. Thus, the algorithm can converge to 2 global aggseyated value that is off by at most 9% from the actual vale, Thus, the thesbold diced relates to the accuracy of, ‘he global estimate in each node, and consequently the accuracy ofthe estimate thatthe user receives. The Ueshold also relates dreely to the energy expenditure of the sensor ‘network, as it decides the amount of wafic being generated in the network. Thus, the higher the threshold is, the lower the aceuracy iy and the higher the energy savings ae. In fct, he tiveshold gives the user contol knob i the energy accuracy subspace. This parameter is set depending on the usee equiement 136 We should also uy to contol the overhead of passing on fedundat information. Ths wil be cleat from the scenario shown in Figure 5. Suppose that nodes /, J and K are in each other's neighborhood. Suppose node | measures a local value, ‘which changes is estimate. After combining tis estimate with the values ints extmate table, ode I decides to broadcast its new estimate, Both node J & K get this new value and will ‘Update thee respective estimates accordingly. Now node K (or ‘is unaware atthe estimate of J (or K) has also changes. So wien node K (or J) run its decision making sep, it will Find out that i should broadcast this new estimate as it can change the estimate of J (or K) considerably. Thus the same estimate i propagated again, ‘This overhead can easly be avoided by simply maining {information about two-hop neighbors at every node. When 8 node gets a teading from another node (say X), it only checks for those nodes (in the estimate able) that ae in its Ov neighborbood but notin the neighborhood of X. Thus, in the shove example when K (or J) ges the realng from 1, ¢ does not broadcast as it realizes that J (or A) is also in the neighborhood of 1. The two-hop neighborhood information can easily be acquired with a protocol like the one in (1) a the network bootup phase. However when nodes die this Information might get outdated. Ia ode doesnot keop wack fof such dying. nodes, it might just potentially speed some unnecessary energy in making sure that information reaches dead nodes (which it believes are sll alive), Note that inespectve ofthe strategy that anode fellows, the accuracy of the algorithm remains unaffected, F Insights In this subsection, we give some insight on the current model, a wells motivate mare complex models F.1 Robustness of the algorthen In this subsection we will discuss the robusiness of our proposed scheme and compare it with that of snapshot aggregation, First lol us summarize some of the key properties for ou algorithm ‘©The algorithm i completely disuibuied and localized ‘nodes exchange information ony with immediate one- hop neighbor) Fes ‘© Thor is no need for tine synchronization between the nodes, Moreover, there is no requirement on the number ‘or order of messages tata node should receive. Ifa new ‘estimate i received the curren: estimate i updated“+ Alle nodes perform the same function, and compute te same results, Every node has an estimate of the lobal agrezated value. There are no epecal nodes, Contrast these properties 10 the properties of snapshot aggregation ‘+ Theresa hierarchical parent-child structure “The parent should recive an exact mumber of messages, equal tthe numberof it hildren, ‘As a result our algorithm performs exactly the same with mobile users and multiple users (te global estimate is available everywhere). Failing nodes, failing links, noa- reliable anspor protocols and MACS, deteriorate accuracy slowly, as only the information of th filing, nodesinks is lost and this is usualy redundant information. ‘This robust ‘operation ofthe algorithm isa great advantage inthe dyamic fnvironment of WASNs where node failures and mobile users face the norm ater than the exception. Snapshot aggregation on the other hand is very sensitive 10 the stability of the hierarchical structure. I a parent node dies, all its descendents are cutoff from the fuetonality of the algorithm and it results in generating poor estimates of the ‘lobe aggregated value. Puther more fa user moves, a ne dugrepaion tee noeds to be formed in onder to receive the aggrepned value. This is ansated in more enerpy consumption F.2 Choosing the sveshold ‘he energy ffciency and accuracy ofthe algorithm heavily depend on the chosen threshold. As the simulation results reveal in section V there isa very concrete dependency of threshold, enenuy spent, snd sccuracy achieved. If a user knows some of the micro characteristics of the physical process being monitored, then based on the graphs provided in ection V be can knowledgably chose a threshold that ft his needs. What the user needs to know is how dyamic the proces is. For example, how fast do the values fluctuate and {in what dynamic range? Inuiively a more dynamic physical process will puta grestr stain in our distributed estimation proces, requiring more enerey for the same accuracy. Indeed ‘his intuition ie validated from he simaation results, ‘The more ineresting problem rises though when the wer ‘as no information on the physical process being monitored Tn such a cae, sting a tveshold does not provide immediate information on the actual accuracy achieved, The only known fact is thatthe accuracy is Tess or equal than the chosen ttxeshold. The ony eer parameter than an provide the user with some indirect information about the accuracy 4s the Variance associated with the mean estimate, Recall that variance represents the degree of certainty that should be associated withthe estimate. Hence, sf user gets an esate With large variance, he should ty to decease the threshold and vice versa, Carenly we are in the process of developing an analytical model to provide guidelines to he user, s0 that 1s te can adjust the threshold on-line. Furthermore, the energy spent atthe user node can be usod as an estimate of the average energy spent in aodes (ecll Gt our algorithm is fairly homogenous in energy consumption). So even with no physical process information the user can acquire an estimate for the accuracy achieved, as well as fr the energy spent, on the My, Tn some rare cases with very dynamic processes and. Hie spaia-temporal comelation betwoen the nodes" valves the user could be advised to resort othe periodic execution of snapshot agregatons. F.3 Data Suaeness ‘The estimates generated by every node are based on information, the node has acquired until that time. These ‘estimates are going to be valid a that time but a time passes the data wil become stale. Thus, we need to diferemiate between the freshness ofthe estimacs. Ideally, we would like 1o give more weigh (0 an estimate that has Deen receatly generated. A very simple way of dong this i o timestamp the Prediction. Any old predictions (before a cei time) wil be neglected. However, this way we do not achieve a graceful degradation in estimate imporcance. A more proper way of schieving the gradual staleness ofthe estimate is to increase their variance over time. In our curentimplemeatation, We increase the variance associated with the estimate at a constant rac, ‘This approach motivated us to look for more gener ‘models, where mean and variance are represented time functions rather than simple scalars. This is a subject of on- sing work. V. SIMULATION RESULTS In this section we present the simulation platform, a general ‘model to describe diverse physical processes, and finally measurements on Several aspects of the energy and accuracy ‘quantities. A. The simulation payor ‘The algorithms described in the previous section were implemented. in our simulation platform [1B]. based on PARSEC (PARallel Simulation Environment for Complex systems) [12]. In our simulation platform a sensor network is ‘modeled as: i) collection of sensor nodes, (i) a channel and iy a supervising entity to creste the nodes, wigger the Sensing devices, and keep statistics of several quanties. Esch Sentor node consists of two enties. 1) A procesting entity, ‘where all node specific algorithms run, and 2) a radio entity in forder 1 interact with other sensor nodes, The radio eatty Implements diferent MAC protocols and measures the energy spent due t transmissions and receptions Note that enerzy is spent on receptions even if thee is collision of packets, For ‘our measurements & TDMA MAC was used. To transform luunsmnission and reception times to energy we used the data, from the hardvare specifications provided in (14),1B. Modeling a physical process In order to acquire concrete measurements, the underline physical process needs to be defined. Since we donot wish to bind our algorithm to specific physical process soenarios we have built a general model to define & large class of physical processes, We adopt difusion model of multiple, aduive, tmobile, and vacable sources, That is if there 8 a source at some point, its vale i diffied tothe environment according to some power law of distance. For a single pot in the WASN coverage are, the influences from mulliple sources ‘re added. The sources can move and change thet value in asbitary ways, An example ofa physical process described by ‘our model could be a terrain with a fre moving through i ‘The measured heat follows our model, More specifically our model sates: von Ledisi +0" Vip. +noise OD Here Vip, the value of the sensed quant at point p at time ¢, dist) is the distance between point p and source time ¢, p(s the positon of source fat ime f and k, a are parameters, The noise is gaussian with zero mean and unit ‘variance and it models the measurement nose. For our current experiments we used one source with a specific waveform (shown in Figure 6), with k=025 ant. We altered the period ofthe waveform and also made the source mobile in ord to test the response of our algorithm in increasingly dynamic processes, C. Mearuremenis All results are derived using networks of $0 nodes distribute in uniformly random fashion over a sensoe train of size 45m x 45m. Each node has a ansmission range of [5m and a sampling peviod of 05 seconds. Simulation is run fora time’ interval ‘of 200 seconds. All simulation results are averaged over 500 individual cases Figure 6 shows how the actual maximum value varies over time and the way itis tracked by our algorithm, for two ferent Uxesbolis. The waveform of the maximum value also reveals the shape of the source waveform used in all our measurements. AS stated earier the shape of the source's ‘waveform remains the same and only its period changes in different simulation scenarios. Figue 7 shows the cumulative energy consumption of all the nodes in the sensor network versus diferent threshold valves. The figure shows energy consumption using repeated Snapshot aggregation and our slgonhm in dee diferent fecenarios. As can be seen from Figure 7, the energy consumption in our algorithm varies fom 5% to 67% ofthe energy consumed in repeated snapshot aggregation according to the threshold and the physical process examined. AS expected with increasing thresholds, energy consumption decreases. Reducing the period ofthe waveform ofthe source or making the source imbile results in a more dynamic sting. The energy consumption naturally increases but in a ‘rice manner. The figure clearly highlights the stiffness Sesocinted with snapshot aggregition. ‘The energy 16 consumption for snapshot aggregation isthe same fr all the Scenarios and for all ueshold values, . Fe? Fes [ESTIMATION ERROR VS THRESHOLD USED.Figure 8 plots the estimation eror, versus diferent Utveshold values, The estimation enor basicaly ives us the accuracy ofthe algorithm. Inthe case of repented snapshot sggrezation, the user node gets the actual global aggregated value and hence there is no discrepancy between the estimated value and the actual aggregated value (Le. the eror is 0). ‘Again one can se thatthe snapshot aggregation case is rigid, ‘ot able to accommodate differen Uzeshold values, In our algo, a8 tbe theshold increases the acouracy decreases (estimation enor increases) ee Peto As expected, the performance is worse in a more dynamic setting. A key thing to notice from Figure 8 is thatthe estimation eror is much smaller than he threshold wed ‘Theoretically we Know that the threshold aes 88 an upper ound 0 the estimation ere. Th practice we discover that the fervor is considerably mullet. This is avery. positve ‘observation a one can get significant energy raving’ for typical accuracy levels. For instance, if 95% accuracy is needed (Le. 58% ero) forthe processes examined above, then Wwe can set the threshold to 0.125 and consume only 20% of, the energy thatthe repeated snapshot aggregation approach 17 C1 Spatial variation In this subsection, we examine the spatial variation of energy consumption and accuracy inthe sensor network, The simulations ate canied out fora fixed ttreshold of 0.1. The four is static and the period of the source is 200, The Source is located at (25.25) ie. approximately atthe center of the sensor terain We average out the results for 100 simulation runs Figure 9 plots the spatial variation of energy consumption in the senor tein. The nodes located near the center ofthe eran, closer to the source, are the nodes that ae going © siasure a higher reading as compared 10 the boundary nodes. ‘Thus these nodes generate new estimates of the global aggregated value and transmit thom with greater frequency (ince they are tacking the global mux), consuming larger mounts of energy. On the other hand, boundary nodes that are measuring lower values do not need to tansmit their tetimates since they notice that they cannot influence the fstimates of other odes. Altbough the plot for repeated Snapshot agazegation isnot shown, it ean be easily infeed that every node consumes the same amount of energy iespesive ofits location, Note that even for nodes located at. the center ofthe senior terrain, the absolute vale of energy ‘consumption is mich larger for repeated snapshot aggregation 38 compared to our algorithm. Figure 10 plots the spatial variation of estimation error in the Sensor tern. AS can be observed from the figure, almost every node inthe sensor terrain achieves the same accurcy. ‘This because of the bisie operation of our algonthm scoording to which every node inthe sensor train shuld be having the same global picture as any oer node in the ‘network. Only the node withthe actual maximum has an error ‘ery close to zero. Infact in evry simulation run atleast 95% ofthe nodes had similar eatimae asthe user node. This i ‘leary in contrast to snapshot aggreyation, ia which only few fof the nodes (the nodes in path of the aggregation tee from the node measuring the global maximum and the User node), Ihave the same accuracy (zero) 35 the wser node Sumsmarizing. the simulation resus show how one can achieve a desired energy consumption and accuracy by choosing the right threshold. Funhermore, the energy consumption is not homogeneous in the network, at east in Scenarios where the maximum remins in the same lotion, ‘The nodes, which possibly have the global maximum, end up spending more ence. Howeser, every node in the network has the same global picture and hence has the same accuracy level. Ina more dynamic soting (lor example where the Source is mobile), we expect both energy consumption and accuracy tobe homogeneous throughout the network, VI. Concusions In this paper we study a large clas of applications in |WASNS, rimely aggregation applications. We propose a Aistibed estimation algoritym that ean be applied 10 2 fubclas of periodic aggregation problems. Our algorithm exploits the energy-accurcy trade Of in WASNS to providethe user with a larger solution space than the conventional approsch of periodically ramning snapshot aggregations. We Validate our algorithm Uuough simulation, achieving promising results. ‘VIL ACKNOWLEDGMENTS ‘This paper is based in par on research funded through ice of Naval Research's AINS program, and DARPA's PACIC program, The views expressed inthis paper are those ofthe author's, and do not necessarily represent those ofthe shove funding agencies. VIIL REFERENCES ED $a M, Mein Bioday PrtonsfrLow Energy Deglyman an lene Spor Desay Ad Hoe Wit [evr Paces of Metso 201, Long Bech USA, 2 DDL A Baws and M, B. Sevan, “Agmesion spain a ‘ctcecomtran! dad sae” TM. UCLASESL: SOI Om, Meet oewl eT BL A. Boule “srming Debate Agus fc Somer ‘Reena hp ade ole (4) NH. Coen, A Pinata, J. Te. L. Weng. D. Yeh Chala ose Ageaon of Penaeus? IBM Teel Repo [5], Covance tescion Wong Coup (CTWG) "A Cling ‘Adan nthe Tey ad Ps of Da Fas, Fo an Deel Reet hte ats! (6). DEsen, coins, LHedenaon (Eos), Enbeing te Inner Canmanctn fte AEM. wo p3E-A May (7), Bs, R Govt, Henan, 8 Kms, “Next Cry Gate Sele Const Io Senor Nectrs™ ACM ‘Maca Contes, Seale WA, August 1398 (2), con Tmt eons Moanin See [ews Tang Lesou tea MPEG. ACM Compe: Comusiains Hei, ve et. Oe 20. (9). S.tttmann FSC intone Goda. DBs D. Ginsu fee Wels Ser Mtr wth Lowel Rusia Proceeds of Symes ed Open ‘Shan Pile Gt 300 uae, Siehporpa. aed C. Sen, “Quing and ‘Tatieg ef Ser Network, SPICY 1h Anal urinal Syapesim ce ese Seng Sule and Gti ote Raspace), Oran, ss: Apa 5 U.P Chere Gata, FZ.“ al Space Apes “Tocting nd Serr Managemot Wao Sener Reo, ‘do Ae Reach Comer fecha Rept F002 1007, Mach Er PARSEC. PARstMI Sinloion Enminemet for Comper sens pplesacani Gi. Pee and Wee, “Wines Iepaed Newo Senor Camanistins he ACM VOL, Sr My 00. [A Sarde Can MB Svat, “Dyan fo pnd Iatiaen in sive nar of sor Medien 20, oe, ay 9166.9. ly 30 D.Temesbowe, “Pose Cong’ Commasiains he [ACHE Vol 8h no 8 pas. My 200. 1 Phan, Re Gian, D. Ean esa Eery Sa fo Monteriog Soar eres ace ot WERE 02 9} an. 13. 03}. (141 us. 161 138 07. 18. Sl. Maen, Me 1 ak, ough M. Hele, tnd Wel Hong, TAG ATi AGgpion Sri fo A [NESS apr can soenbNESLed