Sunteți pe pagina 1din 5

Mobility Prediction and Spatial-Temporal Trafc Estimation in Wireless Networks

Haitham Abu-Ghazaleh and Attahiru Sule Alfa


Department of Electrical and Computer Engineering University of Manitoba, Canada. Email: {haitham, alfa}@ee.umanitoba.ca location, as well as the expected trafc demands, both spatially and temporally. This paper is organized as follows. After briey mentioning some of the related previous work in Section 2, Section 3 presents our proposed mobility prediction model, using the theory of Markov Renewal processes. This model is extended to our next framework on Trafc load prediction to provide for a spatial-temporal trafc estimation, which is described in Section 4. Before concluding our work, a network resource reservation scheme which utilizes the developed frameworks is proposed in Section 5. II. P REVIOUS W ORK I. I NTRODUCTION The rapid growth of mobile networking and the diversity in network applications have prompted the need for future generation of wireless networks to support a range of QoS levels. Future wireless networks are expected to safeguard these QoS agreements despite the users movement and the networks trafc. Occasionally, mobile users may need to have their ongoing connections transferred between different base stations (or cells) in order to maintain their active sessions; a process known as handoff. Successful handoffs are only possible if sufcient resources can be granted to the ongoing session by the new network access point. Otherwise, the session will be prematurely terminated, or dropped, as a result of insufcient resources available at the new cell. Sufcient resources along the users known path could be reserved ahead in time if the users mobility behavior was known to the network. One way of minimizing the dropping of handoff requests is to reserve enough resources at all neighboring cells for the handoff request. However, such an approach leads to a wastage of resources as well as the possibility of signicantly increasing the blocking rate for new connection requests. A better alternative would be to limit the reservation of these resources to the regions of the network where a user is likely to visit. Therefore, an efcient prediction is important since misplaced reservation of network resources will not only fail to uphold the desired QoS, but also likely degrade the performance of the overall network. In this paper, we develop a mobility prediction model using Markov Renewal theory and show how to apply it for predicting a mobile users future Over the recent years, there has been many work done on developing mobility prediction and network trafc estimation techniques, and both areas have been addressed independently. Various mobility prediction algorithms have been proposed such as the ones given in [1], [2], and are based on historical data such as aggregate mobility and handoffs history, at each location. An example of prediction algorithms that requires no prior information on the users mobility patterns are given in [3], [4], which rely on monitoring the current users speed, location, or direction. Incorporating road topology information [5] and geographic maps with identiable landmark objects [6] into the user-mobility prediction schemes have been previously proposed, which may improve the predictions by eliminating or afrming certain paths. A more recent work on Markovbased mobility prediction models can be found in [7]. The work on spatial trafc estimation have been covered by many, including the authors in [8], who proposed to model the time-varying population distributions at distinct locations. The work in [9] focused on modeling the spatial network trafc demands at discrete locations for the purpose of optimizing the network coverage areas. The authors in [10] proposed a stochastic model for capturing the spatial trafc distribution of mobile users with multiple classes of services. III. M OBILITY P REDICTION A. User Mobility Patterns In this work, the entire network space is assumed to be divided into zones and the zone-ID is used to specify the users location, e.g. the cell-ID in a cellular network that is currently serving the user. An example of such a representation is given in Figure 1. Each cell can be further divided into a number

AbstractAn understanding of the network trafc behavior is essential in the evolution of todays wireless networks, and thus leads to a more efcient planning and management of the networks scarce bandwidth resources. Prior reservation of radio resources at the future locations of a users mobile trajectory can help with optimizing the allocation of the networks limited resources, as well as help with sustaining a desirable level of QoS. The objective of this study is to propose a framework for a mobility prediction model using Markov Renewal Processes, for computing the likelihoods of the next-cell transition, along with anticipating the duration between the transitions, for an arbitrary user in a wireless network. The proposed technique can also be used to estimate the expected trafc load and activity at each location in a networks coverage area.

978-1-4244-1645-5/08/$25.00 2008 IEEE

2203

of sectors with distinct sector-IDs, which is used to further describe the intra-cell movement patterns, as proposed in [11].

Gi,j (t) = P r{Tn+1 Tn t | Xn+1 = j , Xn = i} (3) Gi,j (t) represents the conditional probability that a transition will take place within an amount of time t, given that the process has just entered state i and will next transition to state j. The sojourn times in such a process are assumed to follow a general distribution. Dene the random variable i as the time spent in the current state i before making a transition, then the cumulative waiting time probability is given as follows. Wi ( t) = Pr{i t} =
j

7
1

6
5

Pi,j Gi,j (t)

(4)

Fig. 1. 7-Cell Network, with Cell-IDs used to describe the terminals mobility path

Other performance metrics from MRP can be found in [14]. C. Model Description

The users mobility history patterns can be periodically recorded using the cell-ID representations. For each cell that is accumulating a mobility pattern prole [12], the number of handoffs made to a neighboring cell, as well as the time spent in the current cell before the transition occurs, can be recorded. This allows for the computing of the cell-transition probabilities, Pi,j , and the mean time spent in the current location of cell i, i,j , before making a transition to the neighboring cell j. Alternatively, the distribution of the cell sojourn times at each location can be extrapolated from the set of recorded sojourn times. These stats can either be recorded for each user, or collectively for all users, into a single history prole per location. The latter method is more suitable for situations where all users will generally exhibit similar behavior at a given access point, and are also not signicantly inuenced by erratic behaviors from a single user. Moreover, it has been shown in [1] that the accuracy of the prediction for the next-cell transition can be improved by additionally considering the prior location of the user immediately before the transition into the current location. For this case, the cell-transition probability can be modied to Ph,i,j , which is a second-order Markov Chain, where the index h is the cell-ID of the previous location immediately before entering cell i. The same applies to h,i,j . B. A Markov Renewal Process A Markov Renewal Process (MRP) is a semi-Markov process where the successive state occupancies are governed by the transition probabilities Pi,j of a Markov process, and the sojourn time in any state depends on both the current state and the next state transition. A more detailed description of such processes can be found in [13]. The semi-Markov kernel for a time-homogeneous process is given by Qi,j (t), where Qi,j (t) = P r{Xn+1 = j Tn+1 Tn t | Xn = i} (1)

The mobility prediction process can be modeled as a Markov Renewal process, and can be used to estimate the likelihoods of an arbitrary user making a particular transition from its current location, within a time t (and not at time t, for reasons explained in [7]). The model will assume the availability of the information regarding the transition probabilities and the conditional distributions of the cell-sojourn times, using the aggregate mobility history that is collected in each cell of the network. The probability Qi,j (t) dened in equations (1) to (4) can be computed to evaluate the likelihoods of an arbitrary user making a transition to a neighboring location, and depends on the length of time spent in the current location. The majority of the mobility prediction models focus only on those users in a network with an active session. However, a more elaborate model should consider distinguishing the users that are mobile with active and idle sessions, as well as the changes in the session conditions (i.e. from active to idle and vice versa). Our proposed mobility prediction scheme is based on the assumption that a users session-activity patterns are correlated with the location and movement patterns in a network. In addition, the presence of users with idle sessions may exert some inuence on the density and behavior of users with active sessions, and vice versa. The transition probabilities for this model can be further described as follows.

Pi,j denotes the probability that a users ongoing and active session is transferred from cell i to cell j. Pi,i denotes the probability that a users ongoing session is terminated in the current cell i. Pi,i denotes the probability that a users idle connection becomes active in the current cell i. Pi,j denotes the probability that a users idle connection is transferred from cell i to cell j.

Qi,j (t) denotes the probability that immediately after making the transition into state i, the process makes a transition into state j in an amount of time less than or equal to t. Qi,j (t) = Pi,j Gi,j (t) , Pi,j = lim Qi,j (t)
t

(2)

The subscript i and i denotes a user in location i with an active and idle session, respectively. We further assume that Pi,i = 0 and Pi,i = 0 since this model describes a Renewal process for predicting the future transitions. Furthermore, we assume that a change in location and sessionactivity cannot occur simultaneously in a single transition. For

2204

W (<=t)

each of the probabilities given above, we can dene a cumulative distribution function in the form given by equation (4), and subsequently compute the semi-Markov kernel Qi,j (t). The Markov Renewal Process for mobility prediction can also be extended for the case where the users previous location is considered in the mobility pattern, i.e. extending Qi,j (t) to Qh,i,j (t) which has the following form. P r{Xn+1 = j Tn+1 Tn t | Xn = i , Xn1 = h} (5) D. Numerical Example Consider the simple case where an entire wireless network is composed of 7 cells only, as shown in Figure 1, such that mobile users can only roam between those 7 cells. In this example, users with active sessions will only be considered, i.e. the transition probabilities Pi,j where 1 i 7 and 0 j 7. The state j = 0 implies the completion of an active session in the current cell i, e.g. P3,0 is the likelihood of an arbitrary user terminating his active session in cell 3. The sojourn times associated with the transitions Pi,j > 0 are assumed to follow a Weibull distribution. Some ctitious values were used to generate the probability transition matrix P= {Pi,j } as well as the parameters for the Weibull distribution functions in the matrix G(t) = {Gi,j (t)}. Figure 2 shows
0.7

0.8

0.6

W1(<=t)
0.4

W (<=t)
2

W (<=t)
3

0.2

W5(<=t)

10

Time(t)

Fig. 3.

The Cumulative Waiting Time Probabilities Wi ( t).

current cell before making a transition to the other locations. Such information may aid with determining when a resource reservation should be made in the neighboring cells. IV. T RAFFIC L OAD P REDICTION An understanding of how the population of users in a network varies with time can further provide network operators with an insight of the trafc demands at each location in the network. Such information can assist with optimizing the allocation of the networks resources. In this section, an equal amount of resources is assumed to be demanded by each of the users in the network, with the difference being in the amount of time spent utilizing the resources. An example of such a network are those that support voice calls only, whereby each connection is given a distinct channel that is equal in bandwidth with all the other available channels. Hence, the network trafc load can be studied by estimating the number of users that have active sessions at each location. A. Model Description Our proposed mobility prediction model can be used to further predict the population of users at each of the locations within time t. Let Xi denote the current (i.e. at t = 0) population of users in location i, and Yi (t) be the estimated population of active and idle users in location i, and within time t. Given Qi,j (t), for all possible i and j pairs, the estimated population within time t is computed as follows, Yi (t) = Xi i (t) +
j (j=i)

Q5,1(t)
0.6

5,4

(t)

0.5

Q5,6(t) Q
5,0

(t)

Qi,j(t)

0.4

0.3

0.2

0.1

10

Time(t)

Fig. 2.

The Transition Probabilities Q5,j (t).

a plot of the probability functions Qi,j (t) for the case where the current cell that is serving an arbitrary user is cell 5. For this plot, only 1, 4, 6, and 0 are permitted for the next state j (refer to Figure 1). The graph shows how the likelihoods of making a particular transition varies with different time-period considerations. For example, during the initial time period of 2 units (from the time when an arbitrary user entered cell 5), the network would predict that a user in cell 5 will most likely complete its session in the current cell. Further ahead in time, the user is instead more likely to make the next transition to cell 4, which would prompt the network to modify its resource reservation decisions. A plot of the cumulative waiting time probabilities in cells 1 to 3 and 5, for the network in this example is shown in Figure 3. These set of graphs show the probability of how long an arbitrary user may spend in the

Xj Qj,i (t) Qi,j (t)


j (j=i)

(6)

where,

i (t) = 1

Equation (6) describes the proportion of the current population of users in location i that are estimated to remain in the same location, in addition to the estimated contribution to the population of users in location i from the surrounding locations, all within time t. Hence, Yi (t) denes the estimated population of users in location i that is expected within an

2205

amount of time t from the known current population Xi , and governed by the probabilities Qi,j (t). We can rewrite the previous denitions in matrix form, such that Y(t) and X are row vectors with components Yi (t) and Xi , respectively. Let Q(t) be the square matrix that incorporates the components Qi,j (t). Note that the diagonal elements in each row of the matrix Q(t) is i . B. 7-Cell Network Example Consider the same 7-cell network shown in Figure 1. Let us denote Xi and Xi as the current (i.e. at t = 0) population of active and idle users in location i, respectively. Hence, X = [X1 , . . . , X7 , X1 , . . . , X7 ]. Subsequently, let Yi (t) and Yi (t) be the estimated population of active and idle users in location i and within time t, respectively. Therefore, Y(t) = [Y1 (t) , . . . , Y7 (t) , Y1 (t) , . . . , Y7 (t)]. For the case of the 7-cell network, the matrix Q(t) that describes the allowable transitions is given in the following form. MA SI (7) Q(t) = S A MI The inner matrix MA , denoting the transitions sessions between the different locations, is of the form, where mi,j = Qi,j (t). 1 (t) m1,2 m1,3 m1,4 m1,5 m1,6 m2,1 2 (t) m2,3 m3,1 m3,2 3 (t) m3,4 m4,1 m4,3 4 (t) m4,5 m5,1 m5,4 5 (t) m5,6 m6,1 m6,5 6 (t) m7,1 m7,2 m7,6 of active following

the capacity of idle sessions can be assumed to be as large as the total population of users in the network, or innite. Let Ci be the maximum number of active connections that can be allowed by the network at each location i. Therefore, active connections that are transferred into location i which is serving at full capacity is terminated and included in the population of users with idle connections in the same location. This is equivalent to having a users active connection being dropped after attempting a handoff to the new location. Using equation (9), Y(t), the population distribution with capacity Ci in location i, is computed as follows, where Yi (t) = min {Yi (t), Ci }, and Yi (t) = max {Yi (t) , Yi (t) + Yi (t) Ci }. Y(t) = Yi (t) , Yi (t) : i , i (10)

C. Numerical Example The case of the 7-cell network will be used in this example, and some ctitious values were assigned to the transition probabilities as well as the parameters that describe the sojourn times associated with each transition. These sojourn times were again assumed to follow a Weibull distribution. The
22 20

m1,7 m2,7 m6,7 7 (t) (8) Note the the inner matrix MI is similar to the form given by MA , with the difference being that the elements are now of the form Qi,j (t), since this inner matrix describes the transition of idle sessions between the different locations. The inner matrix SI , with diagonal elements Qi,i (t), describes the transition of an active session to being idle within location i. The inner matrix SA is similar to the form given by SI , expect that the elements are now of the form Qi,i (t) rather than Qi,i (t). The inner matrix SA describes the session transition from being idle to active within the current location. Therefore, the estimated population of active and idle users within time t can be computed in matrix form as follows. Y(t) = X Q(t) (9)

total in location 2 idle sessions in location 2 active sessions in location 2

population of users

18 16

14

12 10

8 6

10

time t

Fig. 4.

Estimated Population of Active and Idle Sessions in Cell 2.

This model assumes that the network can serve an undened or innite number of users (with active sessions) at each location. In reality, this can hardly be true since the number of active connections at each time period is limited by the capacity of the network at each of the locations, due to the networks limited resources. Hence, the model could be further extended to consider a given maximum number of active connections that can be served by the network at each location. The maximum number of active connections is limited, whereas

estimated population of both active and idle sessions within a time period t was computed using equation (9). A set of random integers were used to assign the initial population X in the system. Figure 4 illustrates how the estimated population of active and idle sessions changes with the time range at which the estimation is made, for cell 2. The estimate for the number of active sessions in cell 2 within a time of 1 unit is much higher than the estimate made with a higher time range. Such changes are due to the population ows from the surrounding cells into the current cell, as well as those leaving it or changing session-state. Note that the changes in the population estimate for cell 2 is dependent on the changes in those cells that can be directly reached from cell 2. For a xed capacity of active sessions, the same initial data that was applied previously were used to analyze the same example with the capacity in each cell being 8 active sessions,

2206

22 20

VI. C ONCLUSIONS AND F UTURE W ORK


total in location 2 idle sessions in location 2 active sessions in location 2

population of users

18 16

14

12 10

8 6

10

time t

Fig. 5. Estimated Population of Active and Idle Sessions in Cell 2, with C2 = 8.

i.e. Ci = 8 for all i. Figure 5 shows the estimated population of active and idle sessions for cell 2. Note how the number of active sessions never exceeds the assigned capacity of 8, and the saturation of active sessions at Ci has a direct impact on the population of idle sessions. V. N ETWORK R ESOURCE R ESERVATION Apart from predicting the mobility paths and activities, and the expected network trafc load at each location, the proposed MRP model could further be used to estimate the amount of resources that need to be reserved (for maintaining ongoing connections) in the neighboring locations. If is the amount of network resources that are required for each mobile connection (e.g. = 1 channel for each connection in a voice cellular network), then the amount of network resources that are expected to be needed at location i during time t is given as Ri (t), such that Ri (t) =
j

This paper presents the frameworks for predicting the subsequent transitions in mobility and activity of network users, as well as estimating the trafc population of users with active and idle sessions within a time period t. The prediction model was developed as a Markov Renewal process. The proposed framework is limited to making the necessary predictions and estimations for single-step transitions. This work will be extended to deal with estimating future trafc loads in terms of trafc volume rather than users. The current network architectures caters for both voice and data connectivity, where the latter type of network service is likely to dominate future network use. The prediction for multi-step transitions will also be considered, for estimating the chances of end-to-end QoS up to a certain level. We are currently in the process of collecting some actual mobility trafc traces in order to evaluate the performance of our proposed predictor. VII. ACKNOWLEDGEMENTS This research work is supported in part by a Grant from NSERC (Natural Sciences and Engineering Research Council of Canada) to A. S. Alfa. R EFERENCES
[1] J. Chan and A. Seneviratne. A practical user mobility prediction algorithm for supporting adaptive QoS in wireless networks. In Proceedings IEEE International Conference on Networks, October 1999. [2] S. Choi and K. G. Shin. Adaptive bandwidth reservation and admission control in qos-sensitive cellular networks. IEEE Transactions on Parallel and Distributed Systems, 13(9), September 2002. [3] T. Liu and P. Bahl and I. Chlamtac. Mobility modeling, location tracking, and trajectory prediction in wireless atm networks. IEEE Journal on Selected Areas in Communications, 16(6), August 1998. [4] H.-W. Ferng et al. A dynamic resource reservation scheme with mobility prediction for wireless multimedia networks. In IEEE 60th Vehicular Technology Conference VTC2004-Fall, September 2004. [5] W.-S. Soh and Hyong S. Kim. Dynamic bandwidth reservation in cellular networks using road topology based mobility predictions. In INFOCOM, March 2004. [6] N. Samaan and A. Karmouch and H. Kheddouci. Mobility prediction based service location and delivery. In Canadian Conference on Electrical and Computer Engineering 2004, May 2004. [7] L. Song, U. Deshpande, et al. Predictability of WLAN Mobility and its Effects on Bandwidth Provisioning. In IEEE INFOCOM 2006 [8] V. Bermudez, D. M. Rodriguez, C. Molina, and K. Basu. Adaptability theory modeling of time variant subscriber distribution in cellular systems. In IEEE Vehicular Technology Conference VTC 1999, vol. 3. [9] K. Tutschku and P. Tran-Gia. Spatial trafc estimation and characterization for mobile communication network design. IEEE Journal on Selected Areas in Communications, 16(5), June 1998. [10] F. Ashtiani and J. A. Salehi. Mobility modeling and analystical solution for spatial trafc distribution in wireless multimedia networks. IEEE Journal on Selected Areas in Communications, 21(10), December 2003. [11] S. Kwon and H. Park and K. Lee. A novel mobility prediction algorithm based on user movement history in wireless networks. In Systems Modeling and Simulation: Theory and Applications: Third Asian Simulation Conference, AsianSim 2004, October 2004. [12] C. H. Choi and M. I. Kim and T. J. Kim and S. J. Kim. Adaptive bandwidth reservation mechanism using mobility probability in mobile multimedia computing environment. In Proceedings 25th Annual IEEE Conference on Local Computer Networks LCN 2000, November 2000. [13] E. Cinlar. Introduction to Stochastic Processes. Prentice Hall, 1975. [14] R. A. Howard. Dynamic Probabilistic Systems: Volume II - SemiMarkov and Decision Processes. Wiley, 1971.

Xj Qj,i (t)

(11)

where Xj is the number of active connections in location j. Hence, the network should, on average, reserve Ri (t) resources at location i within time t. However, this reservation scheme only caters for those users that have ongoing connections transferred from one location to another. To consider further reserving resources for new connections at location i, then the amount of network resources that are expected to be needed at location i within time t, for both new and transferred connections, is given as Ri (t), such that Ri (t) = Ri (t) + Xi Qi,i (t) (12)

The proposed scheme is an example of a fractional-based resource reservation scheme. An alternative approach would be to adapt a threshold-based resource reservation scheme, whereby the decision to reserve a certain amount of resources is based on some known threshold value Hx . For example, if Ri (t) or Ri (t) Hx , then x amount of resources will be reserved at location i within time t.

2207

S-ar putea să vă placă și