A Bayesian Tree Learning Method for Low-Power Context-Aware System in Smartphone

Transcription

1 A Bayesia Tree Learig Method for Low-Power Cotext-Aware System i Smartphoe Kyo-Mo Yag, Sug-Bae Cho Dept. of Computer Sciece Yosei Uiversity Seoul, Korea kmyag@sclab.yosei.ac.kr, sbcho@yosei.ac.kr Abstract Cotext-aware services usig smartphoe have bee proliferated for ubiquitous computig. However, the capacity of smartphoe battery is extremely limited so that the services caot be effectively used. I this paper, we propose a lowpower cotext-aware system usig tree-structured Bayesia etwork. Bayesia etwork, oe of the probabilistic models, is kow to hadle the ucertaity flexibly. A well-kow problem of the probabilistic model, however, is high time complexity, which leads to sigificat cosumptio. To reduce the time complexity, we propose a tree-structure learig method. The key idea lies i how to cosider the relatio of each ode. For the reaso, we coduct the spaig tree based o the mutual iformatio amog odes. The data for experimet were collected from Adroid phoe for two weeks. The amout of the collected data is 7,464. The accuracy of proposed method achieves 94.3%. The eergy cosumptio is measured usig the power tutor applicatio. Keywords-Low-power cosumptio, cotext-awareess, treestructure Bayesia etwork, Structure learig. I. INTRDUCTIN Recet proliferatio of the smartphoes leads to developig a large variety of applicatios ad ivestigatig o the use of various sesors through cotext-awareess. Previous researchers focus o the accuracy of cotextawareess usig all possible sesors []. The battery capacity of smartphoe is well behid the developmet of service applicatio. I a typical case, the user has to carry a extra battery or charge it frequetly. There is the critical issue of how to reduce the battery cosumptio for the cotextawareess i smartphoe. I this paper, we propose a low-power cotext-aware system usig tree-structure Bayesia etwork. Bayesia etwork is oe of the powerful probabilistic methods for cotext-awareess []. It ca ifer cotext i ucertai situatio or with the icomplete data. However, the probability model geerally has high time complexity, because the model has to calculate the probability of each ode every time. It causes the sigificat cosumptio for cotext-awareess i smartphoe. We propose a treestructure learig method to reduce the time complexity. We compare the accuracy usig differet structure learig methods ad evaluate the time complexity of the proposed method. I additio, we verify the low- cosumptio feature of the proposed method i a real smartphoe eviromet. The paper is orgaized as follows. Sectio presets the related works for cotext-awareess, battery problem, ad Bayesia etworks. Sectio 3 describes i detail the proposed low-power cotext-aware system. Fially, sectio 4 reports the experimets coducted to compare the power cosumptio of the proposed method ad the moolithic BN. II. RELATED WRKS A. Cotext-aware service i smartphoe Cotext-aware services aim to provide the coveiet services to users who are i the cotexts recogized. Cotext is all the iformatio related to the iteractios betwee user ad applicatios [3]. Iteractios are becomig importat as research i pervasive computig progresses. Cotext-aware services i smartphoe recogize the situatio ad provide services. The applicatios are developed usig various sesors. Lee ad Cho proposed a method usig the KeyGraph ad Bayesia etwork to ifer mobile life log [4]. Ravi, et al. proposed battery maagemet service [5]. The service recommeds the battery chargig time depedig o curret state. Phithakkitukoo ad Datu proposed a three-step approach i desigig the model based o the embedded sesor data for cotrollig alert mode [6]. Lester, et al. preseted the approach to buildig a system that exhibits these properties ad provided evidece based o data for 8 differet activities; Sittig, stadig, walkig, walkig up stairs, walkig dow stairs, ridig elevator dow, ridig elevator up, ad brushig teeth [7]. Satos, et al. described the architecture, operatio ad potetial applicatios of a prototype system developed withi the User-Programmable Cotext-aware Services (UPCASE) project [8]. A lot of applicatios have applied to the cotext-awareess. However, these previous works lack of dealig with the power cosumptio of smartphoe. B. Bettery lifetime problem Fig. shows the key features of mobile devices. Accordig to the TIME Mobility Poll, coducted i Jue ad July 0, the mai issue for mobile users is battery life. 6 percet of America mobile users wish for improvemets i that area. Device size does ot seem to be a problem for most 6

2 of them, as oly 5 percet of America users wat a smaller device. Because the battery life is short, the user has to carry a extra battery or charge it frequetly. There is the critical issue of how to reduce power cosumptio for cotextawareess. the system. The method requires the sufficiet amout of data, but it is ot easy to obtai reliable data i may realworld problems. The secod oe ca be costruct it based o the domai kowledge. The experts idetify the structure ad set of parameters accordig to their kowledge, if we do ot have eough data i the domai. The time complexity of the Bayesia etwork is calculated usig the LS algorithm as follows []. Here, represets the umber of odes, k represets the maximum umber of parets for each ode, r deotes the umber of values for each ode, ad w represets the maximum umber of clique. 3 CMP = ( k + w +( wr + r ) ) k w w Figure. Key features of mobile devices by TIME Mobility Poll. May researchers have proposed low-power applicatio usig cotext-awareess for solvig the problems. Seo, et al. proposed a cotext-aware cofiguratio maager for smartphoes [9]. The system chages the cofiguratio of a smartphoe accordig to the user-defied policy rules. Bareth ad Kupper proposed a hierarchical positioig algorithm [0]. The algorithm dyamically deactivates differet positioig techologies ad oly activates the positioig method with the least eergy cosumptio. Miraoui, et al. proposed limited resources-aware service []. The system chages the services cosiderig the curret resources of mobile phoe. These researchers lack of reducig the power cosumptio of the cotext-aware module. I this paper, the learig method maitais the advatage of probabilistic model ad reduces the time complexity for the power cosumptio. C. Bayesia etworks Bayesia etwork is devised as a powerful techique for hadlig the ucertaity. Bayesia etwork has a structure of a directed acyclic graph which represets the lik relatios of the ode, ad has coditioal probability tables (CPT). Assume that a ode is idepedet of each other. U ) = A, A,, A ) = A ) A = i= A )... A A pa( A )). i i A, A,, A - ) The coditioal probability distributio of variable A ca be represeted as P ( A pa( A)), where pa (A) deotes the set of paret variables of variable A, where U is a set of ode, ad the joit probability distributio is computed by the chai rules as (). There are two approaches to idetify the structure ad parameter of Bayesia etwork. The first approach is to lear model from the data o problem domais. The structure learig is useful if we have a lack of uderstadig about D. Structure desig of Bayesia etwork TABLE I. Authors RELATED WRKS FR REDUCING INFERENCE TIME Necessity of kowledge Extra procedure Pearl [3] Heckerma ad Breese [4] Zhag ad Poole [5] Kjaerulff [6] Koller ad Pfeffer [7] ude, et al. [8] Descriptio Usig oly relevat CPT about curret iferece Removig the ucertai iteractio betwee causes ad effects Removig weak depedecies before iferece Removig weak depedecies before iferece Applyig object cocept to BN Hierarchical modular approach desiged for multiple agets Bayesia etwork is a robust tool for practical problems which ivolve high level of ucertaity. However, utilizig it i the large-scale domais is difficult because cosiderable effort is put o desigig ad maitaiig the etwork. Besides, it is uable to etirely apply o ubiquitous devices sice lots of computatio power ad resources are required i the iferece process. For these reasos, there have bee may studies o reducig the time complexity. Table I shows two types of the related works for reducig iferece time. First, the ecessity of kowledge implies that the etwork is ot desiged automatically. Secod, whe the system ifers, it eeds extra procedure for modifyig the etwork structure. The Noisy-R model was proposed by Pearl [3]. The model ca compute the distributios required for the CPT from a set of distributios, elicited from the expert, ad the magitude which grows liearly with the umber of parets. Heckerma ad Breese proposed a exteded versio of the method called Noisy-MA Gate [4]. This method showed a collectio of coditioal idepedece assertios ad fuctioal relatioships ad removed the represetatio of 63

3 the ucertai iteractios betwee cause ad effect. Zhag et al. removed weak depedecies before iferece [5]. The method evaluated the relatio of each ode with query ode ad modified the structure of etwork through removig the odes ad edges. Kjaerulff preseted a method for reducig the computatioal complexity through removal of weak depedeces [6]. Koller ad Pfeffer proposed a method where Bayesia etwork is applied to a object cocept called BN [7]. This method used a Bayesia etwork fragmet to describe the probabilistic relatios betwee the attributes of a object. ude divided BN model ito several smaller multi-level modules ad iferred each module sequetially from the low level to the high level [8]. Its compositio is similar to MBN, but it restricts the etworks i hierarchical structure. The previous works i the table have problems. If the desig method eeds domai kowledge, it requires the time to aalyze the domai. If the system eeds extra procedure, it cosumes extra battery. To solve these problems we propose a tree structure learig method. III. LW-PWER CNTET-AWARE SYSTEM I this paper, we propose the low-power cotext-aware system. The proposed method cosiders for the power cosumptio of iferece module. A. Sesor data preprocessig The proposed system focuses o how to reduce the power cosumptio for cotext-awareess usig sesor iformatio. Therefore, the other source that geerates the eergy cosumptio such as memory, sychroizatio, ad so o, is ot cosidered i the system. Abdesslem, et al. measured the eergy cosumptio of differet sesors as show i Table II [9]. Each sesor rus cotiuously o a Nokia N95 8GB smartphoe util the battery was depleted. I this research, the power cosumptio of GPS is 63mW, ad it is 6 times more power cosumptio tha the accelerometer sesor. TABLE II. Sesors PWER CNSUMPTIN IN SENSRS Battery life (hrs) Average power cosumptio (mw) Camera IEEE GPS Microphoe Bluetooth.6 Accelerometer We select the sesors which ca use durig half-day because of the battery life time. Figure 3. A example of the sesor data collected Figure. System architecture Fig. illustrates the system architecture. The system cosists of four modules: Sesor collectio, data preprocessig, etwork desig, ad cotext-awareess. I this system, we do ot use the sesors that require high power cosumptio. The sesor collectio module obtais the cotiuous sesor data i smartphoe. The data are set to the data preprocessig module that discretizes them usig decisio tree. The etwork learig module trais the structure ad parameter of the BN. The cotext-awareess module ifers user situatio usig the tree-structure Bayesia etwork. If the result of iferece is higher tha the threshold, it is the curret situatio. Fig. 3 shows a example of the sesor data collected. If these cotiuous data were used as the states of the iput values, the time complexity of etwork would be very high. It affects to icrease the size of CPT because it is related to the umber of the states of values []. I other word, if the cotiuous values map to the states of the value, the umber of state is almost ifiite. For this reaso, the states of sesor data are preprocessed. TABLE III. DEFINITIN F INPUT AND UTPUT Type Sesors Values Iput Sesor: Accelerometer: _axis Sesor: Accelerometer: Y_axis Sesor: Accelerometer: Z_axis Sesor: Accelerometer: rietatio Sesor: Accelerometer: Pitch Sesor: Accelerometer: Roll Sesor: Magetic: Sesor: Magetic: 64

4 utput Sesor: Magetic:3 Sesor: Proximity Sesor: Light Sesor: Gyro: Sesor: Gyro: Sesor: Gyro:3 User: Situatio {Sleepig, Exercisig, Movig street, Havig meal, Shoppig, Studyig, Viewig} The attributes N i ad N j are used for iput odes i the etwork. The class C represets the output ode i the etwork. There are two discretizatio techiques. First, the rage of iput values ca be divided ito a predefied umber of itervals of equal width. Secod, it ca be divided usig statistical methods. A decisio tree is oe of the powerful ad popular tools for makig rules. All the cotiuous iput such as accelerometer, gyroscope, ad so o make the rules with a rage of the divisio usig the decisio tree, because the iput data do ot eed to chage ito the sematic data. It just eeds to divide three rages: Low, middle, ad high. The Geeral Social Survey (GSS) collected the data o social treds i order to moitor chages i the livig coditios [0]. The survey defied the category of situatio. The output of etwork is defied with referrig to the survey. Table III represets the iput values ad output values. B. Tree-structure learig The purpose of tree-structure learig is to reduce the time complexity by cosiderig the relatio of each ode. The proposed method does ot eed the extra computatioal time for modifyig the structure. I additio, the method does ot eed expert s kowledge because of learig from the collected data. Fig. 4 shows the flowchart of the proposed method. The learig method cosists of six steps. Figure 5. Creatio result of maximum spaig tree Next, maximum spaig tree is created usig the calculated mutual iformatio. The mutual iformatio sets the weight of spaig tree. The tree radomly selects oe ode. The, it selects aother ode that has maximum weight from the ode. If the lik does ot create cycle, the ode is selected. This procedure is repeated util all odes are selected. Fig. 5 shows this step. I this figure, ode is root ode. That is selected radomly. Nodes 6, 3, ad 4 have the higher degree of associatio to the ode tha other odes. Nodes 7 ad 5 have the highest degree of associatio to ode 6. Iput: Node N, Mutual iformatio I ij utput: Maximum spaig tree Settig the lik weight to I ij ; Select ode N i ; While (Not all odes N are selected) Lik Ni to N j, I ij has the maximum weight; If (Cycle is created by the lik) Delete lik Ni to N j ; Retur to the first step; Figure 6. Creatio algorithm of maximum spaig tree Figure 4. Flowchart of tree-structure learig method First, we calculate mutual iformatio with the relatio of class. The mutual iformatio is to calculate the relatio betwee the attribute ad attribute Y from data as (3). I, / c) i j ij ( Ni; N j / C) = i, j, c)log p( N N c C i / c) p( j / c) i i, j j, Figure 7. Groupig the maximum spaig tree 65

5 Fig. 6 shows the creatio algorithm of maximum spaig tree. The etwork ca maitai the relatio of each ode though this algorithm. Third, the method costructs a sub-tree through groupig some odes by cosiderig the relatio of each ode. Fig. 7 shows how to make sub-trees whe the group level is. The group level meas how may depths ca be grouped. data: Accelerometer, magetic, gyroscope, light, orietatio, pitch, ad roll. Fig. 9 shows the tree-structured etwork usig the proposed structure learig method. The etwork cosists of fourtee iput odes, four itermediate odes, ad oe output ode. The moolithic BN is also traied usig EM algorithm []. B. Time complextity This experimet verifies that the proposed method has lower time complexity tha other methods. We calculate it usig LS algorithm. We compare moolithic BN (BN), Treeagumeted BN (TAN) ad the proposed method. It is assumed that the umber of clique w equals umbers of parets k. The maximum umber of states is 7. The maximum umber of paret ode of moolithic BN is 7 while the maximum umber of paret ode of the proposed BN is. The umber of state of the itermediate module chages to 5. Fig. 0 shows the time complexity of each method. Figure 8. Creatig tree-structured etwork The group liks to a itermediated ode, which is used for cosiderig relatio of the grouped odes. Fially, the class ode liks all itermediate odes. Fig. 8 shows the fial step. The parameters of the etwork are traied usig Maximum Likelihood Estimatio (MLE) []. IV. EPERIMENTAL RESULTS A. Experimetal settig The data were collected from three graduate studets for two weeks. We used the Samsug Galaxy S3. Adroid phoe collects sesor data twice per a secod, ad the amout of the collected data is 7,464. Figure 0. Time complexity usg EM algorithm Figure 9. Leared tree-structured etwork We collected o seve situatios: Sleepig, exercisig, movig street, havig meal, shoppig, studyig, ad viewig. Whe they collected the data, the smartphoe was put ito their pocket. The studets selected the situatio ad coducted it. We leared the etwork usig seve sesor Figure. Compariso of eergy cosumptio The experimetal result shows that the time complexity is slightly icreased accordig to the umber of states i the proposed etwork. However, the time complexity of the moolithic BN is dramatically icreased i compariso with the proposed etwork. To verify the relatio betwee 66

6 reducig the time complexity ad reducig the eergy cosumptio, we measured the eergy cosumptio usig power tutor applicatio []. The applicatio ifers 00 times per a secod. Fig. cofirms the differece of the time cosumptio of the moolithic BN (BN) ad the proposed BN (PBN). Although the moolithic BN cosumes the,954j for a hour, the proposed BN cosumes the 960J for a hour. C. Compariso of accuracy We coduct 0-fold cross validatio to calculate the accuracy of each etwork as show i Fig.. Figure. Accuracy of the etworks As a result, although the proposed method has slightly lower accuracy tha moolithic BN, there is the relatively small differece of the time complexity. If the system selects eight or ie as the umber of states, the accuracy of the etwork is 93.7%, although the system ca have better battery life tha moolithic BN. V. CNCLUDING REMARKS I this paper, we have proposed tree-structure learig method for a low-power cotext-awareess. The method does ot require the extra computatioal time ad the domai kowledge cosiderig the relatio of each ode. The system is aware of seve situatios. To verify the efficiecy of the proposed system, we compare the accuracy of the proposed method agaist the moolithic Bayesia etwork ad calculate the time complexity. I additio, we cofirm the power cosumptio usig power tutor applicatio ad verify that the system has lower cosumptio tha the moolithic BN. We will improve the method through modular approach by cosiderig the relatio of odes. The system will be applied to various cotext-aware service applicatios. ACKNWLEDGMENT This work was supported by Samsug Electroics, Ic. REFERENCES [] G. Che, D. Kotz, A survey of cotext-aware mobile computig research, Techical Report, Dept. of Computer Sciece, Dartmouth College, vol., o., pp. 6, 000. [] F. V. Jese, Bayesia Networks ad Decisio Graphs, Spriger, 007. [3] A. K. Dey, Uderstadig ad usig cotext, Persoal ad Ubiquitous Computig, vol. 5, o., pp. 4 7, 00. [4] Y. S. Lee ad S.-B. Cho, Extractig meaigful cotexts from mobile life log, Itelliget Data Egieerig ad Automated Learig, vol. 488, pp , 007. [5] N. Ravi, J. Scott, L. Ha, ad L. Iftode, Cotext-aware battery maagemet for mobile phoes, Pervasive Computig ad Commuicatios, pp. 4 33, March, 008. [6] S. Phithakkitukoo ad R. Datu, Cotext-aware alert mode for a mobile phoe, IEEE Pervasive Computig ad Commuicatios, vol. 6, o. 3, pp. 3, 00. [7] J. Lester, C. Tazeem, ad G. Borriello, A practical approach to recogizig physical activities, Pervasive Computig, vol. 3968, pp. 6, 006. [8] A. C. Satos, G. M. P. Cardos, D. R. Ferreira, P. C. Diiz, ad P. Chaiho, Providig use cotext for mobile ad social etworkig applicatios, Pervasive ad Mobile Computig, vol. 6, o, 3, pp , 00. [9] S.-S. Seo, A. Kwo, J. M. Kag, J. Strasser, ad J. W. Hog, PYP: Desig ad implemetatio of a cotext-aware cofiguratio maager for smartphoes, Proc. of It. Workshop o Smart Mobile Applicatios, pp. 5, Jue, 0. [0] U. Bareth ad A. Kupper, Eergy-efficiet positio trackig i proactive locatio-based services for smartphoe eviromets, Computer Software ad Applicatios, pp. 56 5, July, 0. [] M. Miraoui, C. Tadj, J. Fattahi ad C. B. Amar, Dyamic cotext-aware ad limited resources-aware service adaptatio for pervasive computig, Software Egieerig, vol. 0, pp. -, July, 0. [] V. K. Namasivayam ad V. L. Prasaa, Scalable parallel implemetatio of exact iferece i Bayesia etworks, Cof. Parallel ad Distributed Systems, vol., pp. 8 6, July, 006. [3] J. Pearl, Bayesia Networks, Uiversity of Califoria, Los Ageles, CA 90095, 0. [4] D. Heckerma ad J. S. Breese, Causal idepedece for probability assessmet ad iferece usig Bayesia etworks, IEEE Tras. o Systems, Ma ad Cyberetics, vol. 6, o. 6, pp , 996. [5] N. L. Zhag ad D. Poole, A simple approach to Bayesia etwork computatios, Proc. of Cof. o Artificial Itelligece, pp. 7-78, May, 994. [6] U. Kjaerulff, Reductio of computatioal complexity i Bayesia etworks through removal of weak depedecies, Proc. of Cof. o Ucertaity i Artificial Itelligece, pp , 994. [7] P. Weber, ad L. Jouffe, Complex system reliability modellig with dyamic object orieted Bayesia etworks (DBN), Reliability Egieerig ad System Safety, vol. 9, o., pp. 49 6, 006. [8] P, De ude, G. Pavli, ad T. Hood, A modular approach to adaptive Bayesia iformatio fusio, Iformatio Fusio, pp. 8, 007. [9] F. B. Abdesslem, A. Phillips, ad T. Hederso, Less is more: Eergy-efficiet mobile sesig with seseless, Proc. of Workshop o Networkig, Systems, ad Applicatios for Mobile Hadhelds, pp. 6 6, 009. [0] Geeral Social Survey, [] A power moitor for adroid-based mobile platforms, Available from: 67