An Activity Based Mobility Prediction Strategy Using Markov Modeling for Wireless Networks

Transcription

1 An Actvty Based Moblty Predcton Strategy Usng Markov Modelng for Wreless Networks R.V. Mathvarun and V.Vadeh Abstract: The foremost objectve of a wreless network s to facltate the communcaton of moble users regardless of ther pont of attachment to the network. The system must dscern the locaton of the moble termnal, to afford flawless servce to the moble termnal. Moblty predcton s wdely used to assst handoff management, resource reservaton and servce preconfguraton. Predcton technques that are currently used don t consder the motvaton behnd the movement of moble nodes and ncur huge overheads to manage and manpulate the nformaton requred to make predctons. Ths paper proposes an actvty based moblty predcton technque that uses actvty predcton and Markov modelng technques to devse a predcton methodology that could make accurate predctons than exstng technques. 1. INTRODUCTION The fundamental requrement for the wreless networks to provde flawless servce to the moble termnal s that t must know the pont of attachment of the moble termnal (MT) at any pont of tme. Moblty predcton s wdely employed to ad network management tasks. At network level, moblty predcton can be used to assst handoff management, resource allocaton, congeston control and qualty of servce provsonng [1]. At the applcaton level, t s exploted to provde Moble Locaton Servces. Moblty predcton s an mportant maneuver that determnes the locaton of the moble termnal by carefully manpulatng the avalable nformaton. The predcton accuracy depends on the user movement model and the predcton algorthm used. Two dfferent technques are already avalable for moblty predcton. Frst technque uses the hstorcal movement patterns of the user to predct the user s whereabouts n future. The second technque uses the contextual knowledge. Pattern based technque [2] s used based on the fact that often movng objects have some degree of regularty n ther moton. In regular pattern based predcton technques, the user moblty pattern s created usng the tme sequence n whch the locatons are vsted. Here, users are classfed nto dfferent categores. The man drawback wth pattern based technque s that the pattern perod and perodcty vares from one moble termnal to other and the locatons are ncluded n R.V. Mathvarun s wth Department of Electroncs Engneerng, Madras Insttute of Technology, Chenna , Inda (emal: math_varun@yahoo.co.n) V.Vadeh s wth Department of Electroncs Engneerng, Madras Insttute of Technology, Chenna , Inda (emal: vadeh@annaunv.edu) the profle, wthout consderng the possble reasons for the vst and hence the predcton could be msleadng [3]. Larger the sze and more up-to-date the data s, t s more lkely to make accurate predctons usng pattern based technques. But the restrcton on the sze of the user s hstory leads to loss of nformaton, whch results n msleadng predctons. As the user movement pattern s hghly uncertan, assumptons about the user s movement pattern should be made wth utmost care. Contextual knowledge based Predcton (CKbP) s proposed to overcome these problems [4]. Contextual knowledge based technque predcts the whereabouts of the MT usng knowledge of user s preferences, goals and analyzed spatal nformaton wthout mposng any assumptons about the avalablty of users movements hstory. User context and spatal conceptual maps are combned for the predcton process. But, the nformaton gatherng and evdence extracton are laborous n contextual knowledge based technque. The nformaton requred s huge and keepng nformaton up to date s a dffcult task. There s a clear need for a technque, whch could help to overcome the drawbacks wth current predcton technques. Ths paper proposes an Actvty based Moblty Predcton algorthm usng Markov modelng (AMPuMM) to predct the future locaton of the MTs. Normally n Pattern based and Contextual knowledge based technques, the volume of nformaton that needs to be gathered, stored and manpulated s huge. To overcome ths problem AMPuMM uses actvty patterns to make predcton n contrast to exstng technques that uses the tme sequence of cells vsted by a user. The transton between actvtes s modeled as a Markov chan [5] to predct (n+1) th day locaton usng n day s nformaton. In actvty based predcton technque proposed n [3] an attempt s made to dentfy the actvty performed by the user n each vsted locaton lke home, offce, shoppng etc usng real tme traces. Such an approach would gve rse to serous prvacy concerns. To overcome ths problem, the actvtes are frst classfed nto more abstract groups. The actvtes are classfed based on the regularty of ther occurrence as regular and rregular and further, t s classfed nto short, medum and long based on the length of vst. Through varous surveys t had been establshed, that majorty of user actvtes occur only at specfc locatons, for lmted tme perods n the case of most the users. The number and locaton of everyday actvtes that may be performed by a person s lmted by the amount of tme avalable and the space-tme constrants assocated wth varous oblgatory

2 actvtes (e.g. work) and jont actvtes wth others. The geographcal behavor of ndvdual are found to be stable over longer perods of tme whle the tme-sequence of vsts and repeatablty has a shorter lfe span. Hence patterns based on the tme sequence n whch varous locatons are vsted could have very short lfe span and accurate predctons may not be possble. Based on these facts ths thess proposes an actvty based locaton predcton scheme that could reflect changes n user s behavor earler than other pattern based technques and could provde accurate predctons on MT s future locatons [3]. The rest of the paper s organzed as follows. Secton 2 explans the actvty model and the profle generaton. Markov based predcton method and path fndng algorthms are descrbed n secton 3. The smulaton work and results are presented n secton 4 followed by concluson and future work n secton ACTIVITY BASED MOBILITY MODEL Wth actvty based moblty models [6], future locaton can be predcted wth better accuracy. Actvty models accumulate and store nformaton about the moblty behavor of the users n terms of the tme sequence n whch actvtes are performed. The actvtes performed by a user can be broadly classfed nto navgaton actvty and locaton actvty. Locaton actvtes are typcally performed whle stayng at a locaton. In ths paper a mnmum threshold t mn s set for an actvty to graduate as a locaton actvty. Navgaton actvtes are those assocated wth physcal movement between places where a locaton actvty occurs. Users everyday actvtes are observed and recorded nto sx groups based on the regularty wth whch t occurs and the duraton. Based on these crtera the actvtes are classfed nto the followng sx groups. n + 1 th day s used to develop a model that predcts the user s movements gven n days of nformaton. The Markov chan model uses the current actvty on the n + 1 th day to predct the next actvty for the day. A= {a1, a2.a7} s the set of all actvty states. The transton probablty from an actvty to actvty j s gven by the relatve frequency of the sequence aa, a A} where j { j a. Ths corresponds to the, j entry of the transton probablty matrx p.. p p71.. p p In the matrx P, 17 s the transton probablty from p actvty 1 to actvty 7 and 71 denotes the transton probablty from actvty 7 to actvty 1. The sum of all the transton probabltes n any row s equal to 1.e. j= 7 = 1, j= 1 p j = 1. p Type Duraton Symbol Regular Short RS Medum RM Long RL Irregular Short IS Medum IM Long IL p j a 1- p - p l Each day on whch the observatons are made s segmented nto 9 dfferent tme slots. For each slot the actvty data s gathered and stored separately to make accurate predctons even when the sequence n whch actvtes are carred out s varable. For each tme slot, the probablty of occurrence of the actvty and the probablty to move from one actvty to next can be extracted from users trace. p jj a j 1- p kk - p kl p kl 3. MARKOV MODEL FOR ACTIVITY TRANSITION User s movement patterns attan a steady state and are predctable after an ntal volatle perod. The actvty transton s modeled as a Markov chan usng the seven states consdered for the study. The user actvty pattern hstory s 1- p jj - p j 1- pll a l p ll a k p kk Fgure 1: An llustraton of Markov model for actvty transton wth four states

3 A sngle transton matrx for the whole day wll not gve correct predctons for the next probable actvty, gven the current actvty. For example a user mght move from actvty home to actvty workout every mornng, come back to home and agan move to actvty offce. So a sngle actvty transton matrx wll gve equal probablty of occurrence for both the actvtes at any tme of the day. In order to overcome ths problem and to gve a temporal dmenson and to ad tmed predctons, the day s dvded nto eght fxed tme slots (0-360, , , , , , and ). Indvdual actvty transton matrces are created and mantaned for each of the eght tme slots. Fgure 1 s llustraton of the Markov model for actvty transton wth 4 actvtes. State space A corresponds to dfferent A = { a, aj, ak, al} actvtes. Transton probablty matrx ' T(a, a j) = N(a aj ) / N(a aj ) where N (a a j) s the number of tmes actvty aj follows ' N(a a ) actvty a and j s the number of tmes actvty aj doesn t follow a. Intal state probablty vector s defned by Λ (a) = N (a) / Σ N (a ), next state probablty vector s defned by A (t) = A (t-1)* T. 4. MOBILITY PREDICTION ARCHITECTURE The User Moblty Hstory (UMH) s derved from UAPs observed over a perod of n days. The UMH contans two sets of data for locaton actvtes for each tme slot. The frst sute of fgures furnshes the probablty of transton from one actvty to another. The second set of data gves the set of all actvty locatons of nterest vsted durng the sad perods arranged n the order of tme-of-vst. It furnshes the nformaton about the cell n whch the actvty was carred out and other nformaton to dentfy the exact next actvty locaton n the cases where more than one locaton s avalable for a sngle actvty. When updates were made to the actvty transton matrx dependng upon the number of days and the days on whch an actvty occur weghts are assgned for obtanng the transton probablty. By ths way the system adapts to changes n pattern much faster whle the hstory s also gven due mportance. Snce weghts are assgned based on the occurrences, regular and rregular actvtes are clearly dentfed. Regular actvtes carry more weght when compared to rregular actvtes. Navgaton actvtes are recorded after proper scrutny to avod repettons. They are dentfed a seres of cells that occur n the trace between to end cells where an actvty has been dentfed. The schematc representaton of the proposed moblty predcton archtecture s shown n Fgure2. Trace Actvty Recognzer Transton Matrx Generator UMH LA/NA C A B B A Locaton Predctor Path Fnder Predcted Locaton & Path Comparator C GIS +RCM Fgure 2 Moblty predcton archtecture

4 4.1 Locaton predcton Future sgnfcant locatons of the moble termnal can be predcted usng the UMH when the current locaton of the moble termnal s known. Let be the current actvty, [tmn, tmax] be the current tme slot, t current be the current tme, t elapsed be the elapsed duraton n the current actvty n the tme slot, t mean, be the mean actvty duraton of I n the tme slot, and tpredcted be the tme untl whch the moblty of the user s to be determned. If tmn tpredcted tmax ( ( mean elapsed predcted ( && t current + t, t )) > t next actvty= current actvty else t t current t + t, ) = t whle ( current ( current mean elapsed t < predcted ) t mean na, Get next actvty and of next actvty from the UMH next actvty (na) = actvty wth maxmum transton probablty gven the current actvty ( t mean, na predcted ) f current + t t go to locaton predcton module wth seres of actvty predctons made else t current = tcurrent + tmean, na end whle By dong ths the seres of actvtes that are between current tme and the predcted tme can be determned whch can n turn be used to fnd the correspondng locatons and the path between them. If the t predcted les n a tme slot dfferent from the current tme slot, the seres of actvtes n the current tme slot, f any s frst predcted. Then the actvty the user s n at the tme t max of the current tme slot s taken as the current actvty for the tmn of the next tme slot and the actvty transton matrx of the next tme slot s used to predct the seres of actvtes a user wll be nvolved n untl tpredcted s reached. Once the actvty s predcted the locaton where the actvty s lkely to take place has to be determned. If there s only one locaton actvty wth act_d = a predcted wthn perod m then t automatcally qualfes as the locaton of the MT at tme tpredcted. If more than one locaton actvty qualfes wth act_d = a predcted then (fromx_act_d, fromx_cell_d, freqx) values are used to determne the locaton. For example the number of shoppng locatons of nterest for a user could be more than one. In case of ncdental shoppng people prefer ) places close to ther home or offce and the choce of shoppng locaton s generally dependant on current locaton from where the user ventures. 4.2 Path fnder Once the future locaton or set of all locatons that are lkely to be vsted by the MT s determned the path taken can be obtaned from the UMP. The navgaton actvty records of the UMP provde the path between locatons of nterest. Alternatvely t s possble to obtan the path whch s a sequence of rado cells lyng between these locatons of nterest usng the GIS+RCM system that was used n the actvty recognton phase. To determne the path when several alternatves exsts a shortest path algorthm was used. In ths study to obtan the shortest path when alternatves exst we have used Djktra s shortest path algorthm. 5 SIMULATION AND RESULTS Ths secton summarzes smulaton results of the proposed framework. Smulaton archtecture was constructed to model and predct MT s traveled trajectores. The predcton and path fndng algorthms explaned earler were mplemented usng C++. All actvtes on workng days alone were consdered. Rado coverage maps obtaned from a major cellular servce provder n Chenna, Inda were used to determne and represent user s locaton and navgaton actvtes as a sequence of rado cells wth ther geographc postons. User moblty traces collected n Chenna, Inda alone were used as rado coverage maps for the other set of data was not avalable. The locaton actvtes were dentfed and segregated and the UAPs were obtaned. The UMH summarzes the actvty traces and creates the database wth records and attrbutes based on whch predctons were made. The mnmum locaton actvty duraton du mn was taken as 30 mnutes. As far as vsted locatons were concerned the nformaton were mantaned for nne fxed tme slots (0-360, , , , , , & ). Usng n days old UMH the n+1 th day s locaton and navgaton actvtes were predcted and compared wth the orgnal values and error graphs were plotted. The performance of the algorthm, for users chosen randomly from predctable and quas-predctable predctablty groups s provded. Fgure 3 and Fgure 4 are snapshots of the smulator output showng predcted locaton for a gven tme n two consecutve tme slots. A weghted update system s followed that helps the system to react qucker to changes n moblty pattern. Ths helps the scheme to overcome some of the drawbacks assocated wth regular pattern based predcton technques. Fg 5 gves the performance mprovement n AMPuMM technque over the CKbP technque. Ths

5 mprovement s acheved as the motvaton for each move has also been consdered. Updatng new values and correctng the errors further tran the system that shows mprovement n system performance. Ths performance mprovement s shown n fg 6. Here, performance s valued by the ht rato. A ht refers to a correct predcton that matched wth the actual locaton and an ncorrect future locaton predcton s referred to as a mss. These results are observed after tranng the system for 25 days. Fgure 3 Snapshot of the smulator output showng predcted locaton for slot 4 Fgure 4 Snapshot of the smulator output showng predcted locaton for slot 5

6 Ht Rato Days AMPuMM CKbP Fg 5 Performance Improvement Performance Chart be converted nto GSM trace format. Solely usng GPS nformaton would also lead to a problem of shadowed regon [8]. Sensors could be fxed nsde the buldngs to overcome ths problem. But, ths needs to facltate each buldng wth sensors. To avod ths, for the shadowed regons, the GSM traces could be used. By combnng the GPS and GSM trace nformaton, the system performance could be mproved. REFERENCES [1]. Jngyuan zhang, Locaton Management n Wreless Cellular Networks, The Computer Journal, Page 20 of 22, January [2]. Alejandro Quntero, A User Pattern Learnng Strategy for Managng Users Moblty n UMTS Networks, IEEE Transactons On Moble Computng, vol. 4, no. 6, pp , [3]. M.S.Srcharan, V.Vadeh, P.P.Arun, An Actvty based Moblty Predcton Strategy for Next Generaton Wreless Networks, IEEE, IFIP, Wreless Optcal and communcaton Networks, Ht rato [4]. Nancy Samaan and Ahmed Kormouch, A Moblty Predcton Archtecture Based on Conceptual Knowledge and Spatal Conceptual Maps, IEEE Transactons on Moble Computng, vol 4, no 6, pp , [5]. Ramesh R. Sarukka Lnk Predcton and Path Analyss usng Markov Chans, 9th Internatonal World Wde Web Conference (WWW9) Refereed Paper, [6]. Kar Laasonen, Clusterng and Predcton of Moble User Routes from Cellular Data, Basc Research Unt, HelAsnk Insttute for Informaton Technology Department of Computer Scence, Unversty of Helsnk, Days Performance Fg 6, Performance Improvement 6 CONCLUSIONS AND FUTURE WORK [7]. L. Lao, D. Fox, and H. Kautz, Extractng Actvtes and Sgnfcant Places from GPS Traces, Proceedngs of the Internatonal Symposum of Robotcs Research (ISRR), [8]. Zanab R.Zad and Bran L.Mark, Real-tme Moblty Trackng Algorthm for Cellular Networks Based on Kalman Flterng, IEEE Transacton on Moble computng, Vol. 4, No. 2, March/Aprl 2005 The Actvty based Moblty Predcton usng Markov Modelng (AMPuMM) s free from mantanng large user hstory. By perodcally updatng the transton matrx and state probablty vector, the system performance could be mproved. Wth ths AMPuMM technque, 82.6% accuracy s acheved, as the motvaton for each move s also ncluded. Some tmes, the MT s present whereabouts, whch wll be used to generate the user profle, can t be found because of sgnal loss. To fx ths problem, n future, GPS traces could be used to fnd the MT s locaton [7]. Usng the longtude and lattude nformaton extracted from the GPS traces, the present cell d could be found. Thus, the GPS trace nformaton could