You are on page 1of 21

The Predictive User Mobility Prole Framework for Wireless Multimedia Networks

Broadband and Wireless Networking Laboratory Georgia Institute of Technology, Atlanta, GA 30332 Email: Abstract User Mobility Prole (UMP) is a combination of historic records and predictive patterns of mobile terminals, which serves as fundamental information for mobility management and enhancement of Quality of Service (QoS) in wireless multimedia networks. In this paper, a UMP framework is developed for estimating service patterns and tracking mobile users, including descriptions of location, mobility, and service requirements. For each mobile user, the service requirement is estimated using a mean square error method. Moreover, a new mobility model is designed to characterize not only stochastic behaviors, but historical records and predictive future locations of mobile users as well. Therefore, it incorporates aggregate history and current system parameters to acquire UMP. In particular, an adaptive algorithm is designed to predict the future positions of mobile terminals in terms of location probabilities based on moving directions and residence time in a cell. Simulation results are shown to indicate that the proposed schemes are effective on mobility and resource management by evaluating blocking/dropping probabilities and location tracking costs.
KeywordsWireless Network, User Mobility Prole, Quality of Service, Mobility and Resource Management.

Ian F. Akyildiz

Department of Electrical and Computer Engineering North Carolina State University, Raleigh, NC 27695 Email: Call admission control (CAC) and resource management. An efcient CAC algorithm demands the knowledge of UMP in order to accommodate the maximum number of users or to yield maximum system throughput [23]. Routing optimization. User mobility information can also be used to assist trafc routing in wireless networks to ease the bottleneck effect in overloaded base stations or access points [25]. Location update and paging. Many mobility management schemes utilize UMP to improve system performance with regards to reducing signaling costs and call loss rates [8], [14]. Since mobility and resource management are critical to supporting mobility and providing QoS in wireless networks, it is very important to describe movement patterns of mobile objects accurately. Moreover, the development of UMP frameworks will benet the evaluation of many design issues, such as routing protocols and handoff algorithms. Consequently, there have been some efforts in predicting future locations of mobile users. In [2], the location probability at the time of a call arrival is calculated by assuming that the MTs take the shortest paths when they move from one cell to another with four possible directions. Another prediction method is proposed in [7] in which the next probable cell is determined based on path information. In [25], a hierarchical location prediction algorithm is described in which a two-level user mobility model, global and local, is used to represent the movement behavior of a mobile user. The next cell is predicted by considering speed and direction of a users trajectory. In [13], a bandwidth reservation scheme for handoff in QoSsensitive cellular networks is introduced by using the estimation of user mobility. The authors focused their efforts on the premise that handoff behavior of an MT is expected to be probabilistically similar to the users coming from the same previous cell. Therefore, this estimation is based on statistical measurement. A movement-based location management mechanism based on historical records is proposed, assuming that each terminal is equipped with a path length counter storing the length of a users records and IDs of most recently visited cells [18]. In [6], the authors proposed a model of mobility proles that included the trajectory estimation of mobile users and arrival/departure times in each cell along a terminals path. These cells constitute the most likely cluster into which a terminal will move. As for the estimation of resource demands, two methods are proposed in [47]. One of them is based on the Wiener Process

Wenye Wang

I. I NTRODUCTION Diverse mobile services and development in wireless networks have stimulated an enormous number of people to employ mobile devices such as cellular phones and portable laptops as their communications means. The most salient feature of wireless networks is mobility support, which enables mobile users to communicate with others regardless of location. It is also the very source of many challenging issues, relating to the mobility and service patterns of mobile terminals (MTs), namely user mobility prole (UMP). For each mobile user, a UMP consists of detailed information of service requirements and mobility models that is essential to Quality of Service (QoS) and roaming support. In general, the applications of UMP can be categorized as follows: Development and analysis of handoff algorithms. One of the most important QoS issues is to design efcient hanfoff algorithms to reduce handoff dropping probability caused by bandwidth shortage and mobility when mobile users move from one cell to another [13], [29].

This work is supported by NSF under grant ANI-0117840.

which predicts the bandwidth requirement according to current bandwidth usage. The other method is to use time series analysis on the premise that future demand increments are related to past variations. To summarize, most of the existing methods are aimed at nding the most probable cell [7], [25], [13], [18]. Also, there are very limited efforts on estimating a group of probable cells or a cluster of cells without considering the historical records [2], [6], [9]. Oftentimes, the demand for multimedia services is not taken into account, which is critical for efcient resource management. Furthermore, UMP is not well dened, which should consider the characteristics of users mobility and service patterns. In addition, most of the existing solutions do not provide the exibility of prediction and quality demands. Therefore, a framework for UMP is proposed in this paper, which includes the following contributions: A zone concept is proposed to add an additional level of location description for more accurate location prediction because a zone is a subset of a location area (LA). The zone partition is used to differentiate varying future locations of a mobile user depending on its moving direction, thus reducing computation overhead. A new framework of user mobility prole is designed to incorporate service requirements and the mobility model. Based on the valid time of each factor, the user information is categorized into quasi-stationary and dynamic UMP, which corresponds to long-term and short-term information, respectively. By using an order- Markov predictor, the service requirements of a mobile user are predicted based on the most recent records, aiming to minimize the mean square error. An adaptive prediction algorithm is developed to predict a group of cells into which an MT will move by considering historical records, path information, moving direction and speed, cell residence time, and tradeoff of computation overhead. The implementation of the proposed scheme in third generation (3G) wireless systems is discussed with respect to real-time monitoring, measurement, and computation. The performance of the proposed scheme is evaluated by showing its effect on mobility and resource management such as location tracking costs and blocking/dropping probabilities. The rest of this paper is organized as follows. In Section II, a system model with cellular infrastructure is presented. Also in this section, a new concept of zone is introduced for ne granularity of location description. A mobility model, which includes stochastic model, historical records, and predictive trajectory of mobile terminals is also shown. In Section III, the framework of UMP is dened, and it is categorized into quasi-stationary and dynamic UMP, in accordance with long-term and short-term information. The estimation and prediction algorithms of service requirements and future location probabilities are presented in Section IV. A new algorithm is proposed to estimate the service pattern based on mean square error; and an adaptive approach is established for prediction and management of UMP. This scheme is expected to derive location probabilities for a group of cells instead of choosing the most probable cell. The discussion of implementation in real systems and other related issues is presented in Section V. In Section VI, we describe the

simulation model and the parameters used in our experiments. The effectiveness of the proposed schemes and application on mobility and resource management are shown in Section VII, followed by conclusions in Section VIII. II. S YSTEM M ODEL , L OCATION D ESCRIPTION , M OBILITY M ODEL

With regard to different service coverage and infrastructure, such as wireless wide area networks (WWANs) and wireless local area networks (WLANs), a variety of wireless networks exist. In this section, we describe the cellular infrastructure of a wireless network based on which proposed schemes will be implemented. Then we present a new concept, zone partition for a more precise location description, followed by a mobility model, including stochastic behavior, historical records, and predictive future locations. A. System Model Consider a mobile wireless network with a cellular infrastructure; e.g., General Packet Radio System (GPRS), which may be one of several macro-, micro-, and pico-cell systems. This wireless mobile network provides diverse service applications such as voice, audio, data, and video. A typical network is composed of a wired backbone and a number of base stations (BSs). Each BS is in control of a cell, and a group of BSs are managed by a mobile switching center (MSC) in circuit-switch domain and a serving GPRS supporting node (SGSN) in packet-switch domain as shown in Fig. 1. Each BS is responsible for delivering incoming and outgoing calls for the MTs residing in its coverage area. If an MT is moving from one cell to another cell belonging to another MSC, location registration and identity authorization need to be carried out.



Fig. 1. System Architecture.

It is envisioned that in the future, mobile users will be able to roam between wired and wireless networks, with the expectation of the same QoS as they can expect in wired networks. However, when an MT moves into an adjacent cell with an active connection, it must experience handoff or location registration, which may degrade QoS parameters such as connection delays, call dropping and packet losses. Thus, bandwidth needs to be reserved in advance so that mobile users do not have to wait for the release or allocation of channels when they are in new cells. This requires the advance knowledge of service requirements because either over- or under- reserved bandwidth will cause a reduction in system throughput. Thus, it is necessary to predict service requirements prior to the arrival of a request. Here, we extend the shadowing cluster concept, which was introduced in [23]. The basic idea of the shadow cluster is that 2


Zone 6

Zone 1 Zone 2 Zone 0



01 S

Zone 5

Zone 3 Zone 4

Fig. 2. Zone Partition.

There may be varying numbers of zones in real environments, depending on geographical circumstances and network architectures. For the sake of simplicity, we divide an LA into seven zones in this context. If the radius of the coverage area of an MSC is , then zone is a region, which has the same center as the whole area within the MSC with a radius of . The remaining part is then divided uniformly by six, which results in six other zones. This can be implemented through grouping cell IDs based on the coverage area of each cell. For instance, cell 01 and cell 02 may be assigned to zone 3, while cell 07 is assigned to zone 5. When the MT receives broadcast ID from the serving BS via down-link, the MT knows in which cell and zone it is

Location Area

each MT with an active wireless connection will affect the bandwidth reservation in the cells along its moving direction. As this MT travels from one cell to another, the inuenced cells also change over time. The home base station that is serving the MTs ongoing connection may estimate the probability that this MT will migrate to other cells, sending the respective updates with equivalent bandwidth requirements to the cells in the shadow cluster. Usually, in case of two-dimensional topology, hexagons are used to denote the cells; thus, each cell has six neighbors, and the probability of an MT leaving along one side is assumed as . However, this is not true for the MTs with an active connection and a specic destination. Moreover, the original shadow cluster is proposed in the conceptual level without further discussion on how to form such a shadow cluster. Thus, we need to formalize shadow cluster in realistic environments. Due to the process of location registration, which occurs when an MT crosses the boundary of two LAs, the wireless system always knows in which LA the MT is located. However, an LA consists of many cells. In this paper, we propose a zone partition concept to improve the granularity of location description. A zone is a subset of an LA, which is composed of a group of adjacent cells. We expect that the MTs in the same zone demonstrate the same, if not similar movement behavior. For example, in Fig. 2, the coverage of an MSC-A is an LA, i.e., MSC-A handles incoming and outgoing connections of the MTs who move inside this LA. There are two MTs, and , which are currently located in the coverage area of MSC-A and may possibly move into the area of MSC-B and MSC-C, respectively. The service area of each MSC is divided into zones ( ). The estimation of future locations of an MT based on its moving direction is presented in Section IV-C.

currently residing. Note that the zone partition is a geographical sub-layer of a location area for micro- or pico-cell systems, providing a more accurate description of an MTs position than LAs because a zone partition has much less cells than those that are included in an LA. B. Location Description Before we describe detailed mobility model, let us look at location description rst because the ultimate goal of mobility modeling is to predict locations. In fact, the locations of mobile users can be part of the mobility model, whereas traditionally, mobility models are focused on stochastic behaviors of mobile users. Based on our description in Section II-A, locations can be specied at three levels. In other words, a mobile users current position can be represented as follows: Location Area: In cellular networks [3], an MSC controls a group of base stations as shown in Fig. 1. When an MT moves from one MSC to another, location registration process is triggered so that the MTs update location information is described by a new location area ID. This information is available in both home location register (HLR) and visitor location register (VLR). The former is a centralized database that maintains permanent information of mobile users that subscribe services in a system; whereas, the latter is a relatively small database, which keeps the up-to-date locations of those visiting mobile terminals. Therefore, there is no additional memory or cost associated with this information. Zone Partition: The location area gives information of a high level because an LA includes many cells. This granularity is not sufcient to predict future locations. Thus, we use the zone concept proposed in the previous section to take into account an MTs current position as the MTs current position is closely related to its next position due to the continuity of movement. As shown in Fig. 2, if an MT is currently in zone 2, it is likely to move into zone 0,1,3 and even into the coverage area of MSC-C within the next moment. Or, if we know that MT is moving south, it is most likely that it is going to move into zone 3. Therefore, we incorporate an MTs movement direction and position (zone) to provide a more accurate prediction of the MTs mobility pattern than simply using the LA information. The details of collecting position information are beyond the scope of this paper, and they are currently undergoing further research. Cell ID: In order to maintain an active connection for a mobile user, it is most important to know in which cell the mobile user is located. This information is especially useful in resource and mobility management as explained in Section I. The network knows in which cell an MT resides by sending polling messages, which can be acquired without additional cost during call origination/termination, or through location services (LCS) management. The details of implementation are introduced in Section V. Among these three levels of location information, the ultimate goal of this work is to estimate the next cells into which a mobile user will possibly move. The prediction of future LAs is beyond the scope of this paper, and more information can be found in [43]. In this work, we focus on how to use the information at the zone levels and other parameters to predict possible



C. Mobility Model Since mobility models are designed to mimic the movement of mobile users in real life, many parameters need to be considered. Most existing mobility models present the behaviors of mobile objects without considering previous historical records. There are many mobility models [10], [42], depending on parameters or stochastic characteristics. These models can be categorized into different groups based on the following criteria: 1) macro- or micro-mobility scale; 2) 1-D, 2-D, or 3-D dimensions; 3) randomness (since the movement of an MT is random, which depends on many unpredictable factors, it is helpful to describe the movement by varying randomness in different parameters such as direction, speed, and residence time); 4) geographical constraints, which can be very specic for particular scenarios such as indoor, outdoor pedestrian, and vehicles. In this paper, we propose a mobility model that considers stochastic behavior of mobile users, e.g., residence time within a cell, historical records, and predictive location patterns in terms of location probabilities. There are ve components in this model to characterize the movement of a mobile user: The MTs residence time in a cell is represented by a random variable, , which has a Gamma distribution with probability density function (pdf) [4], [26], [27], [30]. Gamma distribution is selected for its exibility because given different parameters, a Gamma distribution can be an Exponential, an Erlang or a Chi-square distribution. The pdf of an MTs residence time with with Gamma distribution in [24] has Laplace transform the mean value and the variance . Then,


Moving Direction Current Direction

Current Cell

Highest Probability Cells

Higher Probability Cells

Low Probability Cells

Fig. 3. Moving Direction.

II sd ( II sd ( s gfI es ih s gfI es

The mean residence time of this distribution (1) is . We assume that the residence time of an MT in each cell is independent throughout this paper. Even considering pdf of residence time enables us to predict the probability that an MT will move out of a cell or a location area; however, it is unlikely to estimate into which cells an MT will move. We need additional parameters for this purpose. The MTs current direction, is collected through realtime monitoring, which can be initiated by the serving BS. The





Moving speed: We also consider the MTs speed in its moving direction [17], [46]. The MT is allowed to move away from its current position in any direction, and variation of the MTs direction based on its previous direction is a uniform distribu. The initial velocity of an MT is tion limited in the range of assumed to be a random variable with Gaussian probability density function truncated in the range of , and the velocity increment is taken to be a uniformly distributed random variable in the range of of the average velocity, . Historical records: In this context, we use historical records to specify a mobile users previous trajectory, e.g., the cells have traversed during the observation time. The historical records are the statistical information accumulated by maintaining a list in each MT or in the servers of location client services (LCS). The details of LCS is presented in Section V. We can trace the historical records of an MT by using a trace records matrix (TRM) of , where is the total number of records and is the total number of cells that an MT has traversed in the period of obser( ; ), vation. The element of the TRM, denotes whether the MT has traversed a cell. The

)p2 U ) x v byVw2 s S

&" acYTbX

q u

U q tre hpX sq S f ig X hVe g f

future cells. It is worth mentioning that much information of mobile users is subscriptive and aggregate information collected and stored in current wireless systems. In the standards such as 3GPP [41], the following parameters are dened for each user. For example, destination address and originating address must be stored in the VLR or SGSN and HLR before the connection is established. In addition, each user is allocated a local mobile station identity (LMSI) by the VLR/SGSN to a given subscriber for internal management of data in the VLR/SGSN. It is also suggested that the previous location area ID and stored location area ID be specied as user proles. The former refers to the identity of the location area from which the subscriber has roamed, while the latter refers to the identity of the location area in which the subscriber is assumed to be located.

MTs current direction is dened as the direction from its previous position to the current position. Here we denote as the moving direction of the MT , which is dened as the degree from the current direction clockwise or counter clockwise; i.e., . Fig. 3 shows the probable cells in the shadow when the current direction is from West cluster for to East. It is reasonable to consider that the cells along the MTs moving direction will have higher location probabilities than those which are not, as shown in Fig. 3. This emphasizes the importance of direction in predicting future cells. The collection of an MTs dynamic information is related to real-time monitoring in wireless networks [11], [28], [32]. The major issue is how to dene the safe regions for monitoring and processing queries with respect to the processing capability of mobile terminals and BSs. If the safe region is too small, then the MTs being monitored must update queries very often since there is no query boundary inside a safe region.

& "  '%$# ! 

& a`Y"  X

BC@A56198$ & 7 5 61 4 3'%$" !  

1) '0(

Predictive future locations: In most of the existing work, the future location is estimated as next cell. However there are several disadvantages for predicting only one cell. First, when an MT moves quickly through in micro-cell networks, the short residence time in a cell does not allow computations in every cell. In other words, there will be tremendous computations along the path of an MTs movement. Second, next cell prediction cannot be integrated with bandwidth reservation schemes very well with the intention of improving QoS. Therefore, we are interested in the probabilities that an MT will be in other cells at time , which is denoted as given the current serving BS of the MT , is . The serving BS of an MT is expected to determine location probabilities and notify other BSs with this estimation. Then,

where is the total number of cells for the estimation. This number is closely related to the scope of location prediction. Theoretically, we can think that it is possible that an MT is likely to move into every cell, does not limit the number of future probable cells. In reality, the prediction is valid for a specic time period. Given the moving speed of an object, the maximum distance that an object can travel within a time period can be determined. Thus, the maximum number of cells that cover the maximum distance can also be decided accordingly. III. T HE U SER M OBILITY P ROFILE (UMP) F RAMEWORK In this section, we present the UMP framework for the management of service patterns and mobility characteristics. The UMP maintains a mobile users information, which is associated with the information of a users velocity, moving direction, current position, and service requirements. In order for wireless networks to support different bandwidth reservation and mobility management strategies, the networks must be able to dynamically and adaptively maintain historical records and predictive locations of mobile users online. The historical records can be collected from the users subscription such as entire information stored in home location register (HLR) for cellular wireless systems or during the origination or termination of a session for the purpose of quality monitoring and billing. On the contrary, future positions must be predicted based on historical records and mobility-related parameters such as current position, direction, and speed [17], [46]. In the proposed framework, the predictive records are referred to as 5

 ys  s  es  d

s fofo s  s o &  " b ys  s  es t`h

& "  oo &  } &  } & "  S & "  s UacY `z W} gfo acY"  ` z acY" c z acY~ `z W} h3acY|z  x y

& a`Y" z  x y

 v ( wub


is given by

p q fgo  q q oo t s ffo gfo gfo oo oo oo ss p ss a o o e ss fr fgo 

. . . . . . .. . . . . . . . . . . . . . .. . if the MT has entered this cell otherwise

lmm mm mmn kj


can be written as

the location probabilities of the cells in the shadow cluster and the estimated service type during an MTs roaming into other cells. (2)
User Mobility Profile

QuasiStationary UMP Network Configuration

Dynamic UMP Service Type Location Probabilities


Time Period Service Description Calling Pattern

Fig. 4. The User Mobility Prole (UMP) Framework.


There are two types of data in the new UMP framework as shown in Fig. 4. The rst type is called quasi-stationary UMP, which represents the MTs information that changes infrequently or that can be obtained from network databases. This includes both subscriptive and historical information. The other type is called dynamic UMP, which changes over time or cannot be obtained from network databases. The quasi-stationary UMP, long-term information, is used to nd the dynamic UMP, short-term information. The following two subsections present the details of the framework. A. Quasi-Stationary UMP An MTs quasi-stationary information includes an MTs current serving system description, calling pattern, and service requirements. In wireless networks, a mobile user receives its mobile services by subscribing to a home network. For example, a user can subscribe to a service package, which is the combination of different service classes. Such information will remain unchanged for a long time until the mobile users explicitly change their services. Moreover, the communication records and location information of mobile users will be updated from time to time during the connection establishment and connection release phases. In other words, this information may remain unchanged for a certain period of time. For instance, each user has an individual record in the HLR of its home network. When a user roams from one location area to another, the location update process is triggered. Accordingly, the mobile users current location at the level of location areas is updated in the HLR and visitor location register (VLR) or SGSN. Therefore, such information can be referred to as quasi-stationary UMP. We denote this long-term information as a quadruple, for an MT, . Here, we consider that each element of is a component of a vector, which indicates a specic character. Thus, are the elements of vectors , , , and , respectively, which are elaborated as follows: Network Conguration . This is a vector that identies different network architectures such as picocell, micro-cell, macro-cell, and satellite systems. Since the ra-

For a particular MT, , the service pattern is denoted by . The PMF can be decided either by the network administrators or by accumulating the mobile users records. The algorithm for estimating will be illustrated in Section IV-B. One of the main tasks of predicting and computing UMP is to determine . The sequence of cells that will be visited by the MT constitutes a random process, with location probabilities, , which is the probability that an MT, , residing in cell , will be in cell at time . The location probabilities, , are calculated by the serving BS based on the MTs historical records, current position, velocity, and moving direction. IV. M EASUREMENT AND P REDICTION A LGORITHMS A block diagram of the algorithms proposed for service and mobility prediction is illustrated in Fig. 5. The objective of these algorithms is to determine dynamic UMP, , using quasistationary UMP, , and other parameters. We start with the account of UMP framework by explaining the parameters used for prediction. Then, we discuss the description and prediction of service requirements. Afterwards, we present the prediction algorithm of location probabilities. A. Flowchart of the Algorithm As shown in Fig. 4, the generic UMP is composed of quasistationary UMP and dynamic UMP, where the former one is obtained during the MTs interaction with the network, such as incoming and outgoing calling requests as shown in box in Fig. 5. First of all, given the network infrastructure by , the cell radius and network conguration are then available. Secis one element in . The direct ond, the time period 6

B. Dynamic UMP The quasi-stationary UMP provides long-term information of mobile users. In order to satisfy real-time prediction requirements, we also need to consider data that changes from time to time such as moving direction, speed, and service type. This type of information is referred to as dynamic UMP. Our objective is to derive or estimate dynamic UMP information based on quasi-stationary UMP. We focus on the predictive records that are derived from the mobile users previous service and mobility information. We assume that locations are independent of services in this context because they can be reected in the network

&" aYce z  x y

x ( & " I 3`R



` & "

& a`Y" z  x  y Y UacY|z  x y s  & "   S d } & " P'03`R  Y 

&  acY" `z  } &" a`Ya z  x y

d {

dius of each cell, the service description, and the signaling formats may be different from system to system, each component is associated with only one type of network. is the number of the registered networks among which an MT is allowed for seamless roaming. At a specic time instant, an MT must be located in a network, e.g., . This information is not required from users; instead, it is recorded in the network entities such as HLR, VLR, and SGSN [41]. Time Period . This set identies different time periods of one day. is the cardinality of the set, i.e., the maximum number of time segments divided within hours. We do not differentiate weekdays and weekends because an MTs movement during a time period on a weekday may be similar to another time period on a weekend [36]. How to divide the time period is more related to trafc load and mobility patterns. In current cellular networks, time segments relating to accounting are categorized into peak time, non-peak time, weekends, weekdays, and are decided by service providers. These time segments sometimes overlap for different service plans. However, each service operator has its proprietary statistics about trafc loads and handoff rates for each local area for network planning [22]. Service Description . This set represents service requirements such as bandwidth requirement and end-to-end delay. is the total number of service patterns allowed in the network. Each element in this set provides the probability mass function (PMF) over a particular sampling space of service types as described in Section III-B. For example, is used to represent the probability mass function (PMF) of an MT, and its corresponding sample space . A mobile terminal may have the same service requirement in different networks if this customer subscribes its service as a xed rate service, or it may need different services in some networks because some networks may provide various services. Calling Pattern , where is the cardinality of the set. Each element is related to the description of calling events, which can be different from one user to another. However, in most of the existing networks, the calling patterns are dened based on a group of mobile users so that computation complexity can be simplied and scalable. According to current service providers, each user has a personal record, which is typically kept for a certain period such as three months or six months, depending on local regulations and operators capability. This data may include the probability distribution of calling arrival time, call holding time, and call origination rate.

conguration. In particular, we examine location probabilities and service type in the future cells because these two metrics are the most often used parameters for mobility and resource management. Let denote the dynamic UMP for the MT, at time . The rst element is the service type that will request, and the second item, , represents the estimation of location probabilities given that the MT, is currently in cell . Location probabilities are used to represent the likelihood of an MTs presence in a cell. For single cell location estimation, there are only two possibilities: 1 or 0, in accordance with whether an MT will be in a cell or not, respectively. Since it is unrealistic to represent an MTs movement by a single random variable, we use location probabilities to describe the result of many factors such as mobility models, changing directions, and geographic conditions. This method is widely used in the research of handoff, location tracking, and mobility management [13], [35], [44]. In this context, we use PMF to describe different service requirements, which are offered to subscribers in terms of service packages, covering the most popular service requirements of mobile users. In other words, each service pattern has a specic PMF with regard to each service type. For a discrete random variable, service type Y, the PMF of over a sample space is dened by [33]


s ffo s  s oo

E s gfo s  s oo

Ves gfo s  es oo d


We also notice that history records mate will result in an estimate PMF,


B. Service Description, Measurement, and Estimation The aim of our algorithm is to minimize the estimation error between the estimated service type and the real value. The future service pattern is estimated based on a priori PMF of service . First, we assume that the service pattern of patterns, an MT can be represented by a random variable . The PMF of this random variable is , which is the probability for value . The sample space of service types is denoted by . For example, there are four samples in , i.e., there are 4 different services which may be requested by an MT, such as voice, data, audio, and video, corresponding to values 1, 2, 3, and 4, respectively. We dene a PMF vector as

where for is one of the values in the sample , and denotes the number of times that value space, occurs in the number historic records. As a result, we can generate an estimate PMF vector, which . We shows the frequency of each service request for denote this PMF as :

We predict the next service requirement by minimizing the mean square error between the prior PMF and the estimate PMF. Considering that the mean square estimation of the random variable is to nd an optimal constant that minimizes ,


where is the probability of each service type for and relating to . Then, we consider an order-L Markov predictor, which assumes that the service can be predicted from the most recent values in the service history, which is

where is the pdf of a continuous random variable the discrete random variable , which follows the PMF, above equation is rewritten as


where is the length of historical records and is the service that the MT, has requested at time . If we think of the users service requirement as a random variable , and is a string representing the sequence of values that takes for any , then the Markov assumption is that Y behaves as follows, for all

z & ffI z `'} II "

where the notation ity that takes the value

denotes the probabil. The rst two lines indicate that the 7

z & q " &&  C|& "  x s h w z ` e z& c"  ' ( #|8`" x `" 


&|8`" x `"  &

& a& "  x %" " 


where is one of the values of . However, this denition cannot indicate the relationship between the history and the selection of an optimal value, showing that an optimal value is selected independently of the previous records of the random variable. When we consider historical records in our estimation, it is necessary to take into account the historical effect on nding an optimal value. In this context, this effect is represented by the estimate PMF for an estimate value. Accordingly, the objective of the estimation is to nd an optimal value that can minimize the difference between the prior PMF vector, in (6) and the empirical PMF, , in (12). Therefore, we dene mean square error between these two vectors, as

s ogfo s z s ffo s  s o oo # & z& c"   e z c"  & `" & " Q s &c" g & e %"  e %R

Az z

U& I |& "  x s ogfo |& "  x s h o & & s a& "  x s h

& "  x q " } & " I & "  x P%u3|& "  x `R & " %

& |s % " s fgo es { b z oo d ( q q & " " a q q w3a&  x s h q z `

&  " x

" qw h ` qw d z cR " q " w ( c S &  "  & b "

information corresponding to a particular timing period is the moving velocity and the pdf of an MTs residence time in a cell, which depend upon trafc conditions, such as congestion in rush hours. Then, the service description is given by , with PMF and its corresponding sample space . Finally, , relating the calling pattern of the MT is described by to the incoming/outgoing call distribution and call holding time. This information will be used for resource allocation in possible future cells. In addition to predicting location probabilities, we estimate the next service request based on the PMF of service patterns as well as historical records shown in box C of Fig. 5. The history of service requirements is known to the network because of billing and service management. Thus, this information does not require extra effort and can be utilized efciently. It is worth noticing that the network does not need to specify bandwidth for each particular MT; rather, it manages resource based on statistical summation of all possible requirements so that the resources are used effectively and the throughput can be maximized [23].

probability depends on only the most recent records, while the latter two lines indicate the assumption of a stationary distribution. Since we do not know the next service type, we can generate an estimate from the current history. We denote the probabil, ity that the next service type takes the value as (10)


  D   R

g g { g fg(I I z z & " s c{ I eY d & acY"T6 " oo " " S " s U& q `YTP gfo &  cYTTW& `YP #& T x

U`P s gfo s & d T s aT 3 " & "  oo "  &("  S & & b " x

&fh#6{ s Gc{ q z %00 & 7 ( 7 " " } &fht ( & " " } s 7 e ` h `'h & " & (" a& T x #t s a 6h P%00} "

& `" 

& " `T#

P & o P "

& " %T z w

and the new esti-



(13) . For , the (14)


B A Cell Radius Network Configuration x Velocity Residence Time x Traffic Condition x PMF of Service Prediction Order Moving Direction Current Position Computing Mobile Probabilities Eq. (4) Dynamic UMP x Service Estimation Eq. (16) C Previous Service Records

Call Arrival/Outgoing x Call Holding Time

QuasiStationary Q UMP x


Fig. 5. Flowchart of Computing User Mobility Prole.

where is the cardinality of . By applying the minimum square error for the estimation, the estimated result of the next service type is:

Therefore, the estimation is determined by both the service PMF and historical records.
Historic Records h1 h 2 hL 1 2 3 U0 Sample Space

C. Prediction Algorithm of Location Probabilities In Section II-C, we introduced a mobility model that describes the movement of an MT by considering the cell residence time, historical records, and future location probabilities. Note that the prediction of location probabilities involves many parameters. In this work, we develop an algorithm which takes several important factors into account, while providing a suboptimal maximum likelihood estimation. C.1 Assumptions and Parameters We assume that the MTs residence time in a cell is represented by a random variable, , which has a Gamma distribution with probability density function (pdf). First of all, we need to assume and measure the following parameters described in Section II-C: The pdf of the residence time is described by Laplace transform, in (1), with a mean of ( ) and variance of . The MTs current position is represented at three levels as described in Section II-B. The zone partition can be obtained by knowing the cell ID. The MTs current direction, , and speed, , are collected through real-time monitoring, which can be initiated by the serving BS as explained in Section V. The initial velocity of an MT is assumed to be a random variable with Gaussian probability density function truncated in the range of . Historical records are available in the form of trace records matrix (TRM) of as in (2). 8

Error Calculator Decision Maker

. . .

1 2 3


Sample Space

. . .

Fig. 6. The Estimation of Service Type.

U ) x yVv 2 s S

& " w`dw

& acY"  X

During the implementation, we consider that an MTs historical records will be updated in VLR/SGSN and HLR during the connection origination stage. Then, we can proceed to estimate the future service type by examining each service type . Each service type in the sample space is selected sequentially and input to the square error calculator together with the historical records as shown in Fig. 6. The results of this calculation are then input into a decision maker to obtain the most probable service type with the minimum error. Finally, is chosen as the estimated the corresponding service type

1) '0(


service pattern. This method guarantees that the estimated service pattern ts well with the known PMF. If an MT is not in the progress of a call and its next probable service request is estimated as , then the bandwidth that needs to be reserved can in (16) for in (31). be determined by substituting

&  '%$" !  


& & I e%" x 6%"  f b eh#& "  x  

. . .

C.2 Prediction Algorithm In a wireless network, the BSs are the entities which can fetch and transfer the users information and provide terminal services. They also allocate bandwidth and maintain the QoS requirements of a connection. In this section, we show how to identify those cells into which an MT will probably move and how to determine their location probabilities. There has been some success of varying degrees on this problem, which is mainly focused on determining the next cell [25]. In [16], [34], it is assumed that the MTs direction is a uniform distribution in a range of or that relies on the signaling measurement without considering road conditions and an MTs historical behavior. The most challenging work is to distinguish those probable cells based on an MTs moving behavior including its velocity and direction. In this work, we consider the following constraints in our prediction: The maximum distance that an MT can travel in the estimation interval: For a certain time period, , the maximum distance of an MTs traveling is , where is the upper limit of moving speed. Therefore, given the cell radius, we can obtain the maximum number of cells that an MT can traverse, i.e., the maximum number of cells to be considered in the prediction. The future zones: By considering the maximum distance, we can determine a group of cells into which an MT can possibly move into. However, this set can be further reduced when we consider moving direction and the MTs current position. The prediction level: Similar to all the prediction-based schemes, our algorithm inevitably involves overhead for computations and communications. Therefore, we introduce a new concept called prediction level to reect the effect of this overhead. The higher the prediction levels, the more complicated the computations, which result in more accurate predictions because more cells are considered. The details of prediction level are described in Section IV-C.2.b.

C.2.a Estimation of Zones. For the sake of simplicity in presentation, we consider a wireless network with hexagonal conguration for the remaining of this paper, although arbitrary shape congurations can also be covered by our solution. We notice that most people travel along a route to a destination, which means that their movement can be traced and predicted. In order to take the MTs direction into consideration, we rst determine the possible zones dened in Section II. In other words, the region of an MTs future position is constrained to a set of specic cells relating to its direction variation. The most commonly used method is to denote the position of an MT by coordinates at a specic time. Here, we use the example of seven zones described in Fig. 2. In Fig. 7, let represent the MT currently in zone , and be the current moving direction of the MT, which is the direction variation derived from the previous direction. This coordinate system is dened with its origin at the current location of the MT; i.e., the MT is always in its origin, and its previous direction is the positive direction of the axis. The axis can be obtained by turning counterclock wise from the axis. Thus, the MTs position can be represented by as shown in Fig. 7. This coordinate system is dynamic in the sense that its origin and axes orientation change over time. Assume an MT is moving from point toward ; thus, its next position may be in zone 1, . In general, the future zone, Case k ( ) can be determined as if if

Case 1 to n-1


where is the total number of zones, and is the angle of each zone. By default, we always consider to be included in the UMP. It is possible that more than one zone is involved in estimating location probabilities since it is difcult to differentiate 9

~ ~X s  a`Y"  X & ( e a e  a`Y"  X fi3 & ( e g g (

& ~ h3a`Y" 

& acY"  X

&& & aacY"  X s a`Y"  c" d

Path database (PD), which is a part of the digital map database. The digital map is obtained by discretizing a map into small segments and each segment has many routes along with their relationship to others [12]. This information is available for many applications, such as nding directions from one location to another. We also assume that an aggregate historical path database is retrievable in the network administration center. Each record in this database is the previous path that the MT, , has traversed. We assume that the movement of an MT have similar patterns such as traveling between home and work sites. Note that the path database is different from trace matrix . The former shows the cells in the sequence of an MTs travels, while the latter may not follow the same order. Another distinction is that the path database maintains an MTs past travel information, which may be considerable than the TRM because TRM is more related to dynamic information in short term. The objective of TRM is to determine future cells given the trajectory of an MT; whereas, the path database objective is to nd the similarity of an MTs movement.

The self-similarity of movement: We assume that an MT has a self-similar moving pattern. Therefore, if a route in a cell has been traversed by an MT before, which is shown in , then we consider this cell to have higher location probabilities, i.e., the MT is more likely to move into this cell. The path information: To be realistic, we take the path information into account. For a given part of the digital map, we assume that it is covered by a number of cells. Each cell has a set of routes. If a cell consists of a route that is continued from the current cell in which an MT resides, then this cell has a higher location probability. In this way, we combine the effect of historical records as well as path information. The prediction of location probabilities involves many parameters, which cannot be simply represented by a mathematical predictor. By considering the above constraints, we will achieve a sub-optimal estimation with maximum likelihood. In our prediction, rst we will nd the most probable cells, based on the MTs 1) current position, 2) moving speed, 3) prediction level, and 4) direction. Then, we will examine each cell in accordance with the last two elements. The cells which satisfy all four will have the highest probabilities.


x v AP2

x v  yVP2 p  p

U f es S d


Zone 2 8 9 Zone 3 Zone 4

A( x(t), x (t) )
Current Direction

in the rst layer of the current cell. The cells adjacent to the rst layer cells form the second layer of the current cell. If the estimated cells cover only the cells of the rst layer, then it is called rst level prediction. Similarly, the second level prediction is associated with both rst and second layer cells. Fig. 8 shows an example of the probable cells with rst and second levels of prediction.


Zone 1

Previous Zone 0 Direction

8 9 Zone 5

Zone 6 Mobile Terminal Previous Direction Current Direction Mutiple Zones

(a) FirstOrder Prediction

(b) SecondOrder Prediction

Fig. 7. Coordinate System with Zone Partition.

Fig. 8. Prediction Levels.

an MTs position around the boundary of zones. Thus, we extend possible zones for Case k as in the following expression: Case


and and and . . .

Case Case

The denition of the prediction level is related to the accuracy of a prediction and can also be used as a QoS constraint, as well. For the rst level of prediction, only six cells are considered; i.e., the computation is not very complicated. On the other hand, if the network should provide location probabilities with second level of prediction, there are eighteen cells, including the rst layer and second layer cells. Correspondingly, the complexity of computation increases, which is demonstrated in Section IVC. Note that some second layer cells may have higher location probabilities than some cells in the rst layer, based on an MTs movement pattern. One example is that an MT traveling on a highway, is more likely to be in the second layer cells along its movement direction than the rst layer cells behind its trajectory or off the highway. C.2.c Calculation of the Number of Cells. Mobile users velocity can be inuenced by circumstances including geographical condition and timing period. For example, in an urban area in which there are many buildings, MTs are forced to travel at a low speed with diverse direction changes. On the contrary, in rural areas, MTs can travel at a higher speed and change direction infrequently. The maximum distance traveled by an MT can be determined by a knowledge of the upper and lower velocity bounds of an MT. We compute the number of cells that an MT could have traveled during the time window . We consider that the MT goes through a cell with the mean residence time in a cell. The average distance, , that an MT may travel along one direction, in terms of number of cells, is obtained by


In the example shown in Fig. 7, is the number of zones, and is the angle of each zone. Although the selected zones can shrink the number of probable cells, it is not sufcient to identify the probable cells. Next, we will further approach those probable cells by considering velocity and residence time distribution. C.2.b Prediction Level. Apparently, the prediction of location probabilities may involve many cells as well as computation complexity, which is incurred mostly from the number of cells to be considered. Basically, the more cells or larger areas considered in the prediction, the better approximation can be reached, requiring more computations. To balance the computation and estimation accuracy, we introduce a new concept called prediction level to describe the inuenced region of the shadow cluster of an MT. According to the concept of shadow cluster, the inuenced cells are a group of cells surrounding the cell in which the MT is residing. Thus, we always start from a current cell, which is the center of a shadow cluster. The vicinity of the current cell can be denoted according to its distance away from the center cell. If a cell is adjacent to the current cell, then it is 10

Note that may be greater or less than the required prediction level at time for a terminal . If the former will not affect the computation of probis the case, the able cells, which means the estimated region covers the area specied by . However, if the latter is the case, then the computed region cannot cover the area that is required by the prediction level. In this scenario, it is necessary to enlarge the

& a`Y"  &" acYw Y &" acYTR & acY"  U S  I C PCQ 3acY"   p &


&" acYT

& '( e d g g c"


7   d ~ X  ~ & "X o  g & f " 'X o d e a`YT0' ( 7   & X ~ X & " X acY" ' o  g f  ~ e  ~ X o  g & f " Va`YT0'  & "X #  ( e d  & X ~ X o  g & f " ~ X o & d e `" e acY"  ' &" acYPb  ~ q 'X


calculation range to meet the requirement of . We de, which is rewritten as note in short, as the number of probable cells in the mobility proles with order at time for Case . The simplest , the number of the probable cells scenario is for Case 1 in (17), , can be calculated by the following formula as there are six cells in the rst layer:

1 2 3

X1 X2 X3 0 0 0 1 0 0

0 0 0 0

... 1 0 ... 0 0 ... 0 ... ... 0 0 ... 1 ... 0 0

XM 0 0 0 1 0 0 0 0 0 0 0 0

Probable Cell
0 0 0 0 1 0 0 0

The second layer includes two parts, one of which is the rst layer cells and the other is the second layer cells which fall into the zone, that is


C.2.d Prediction of Location Probabilities. Thus far, we have determined the total number of probable cells and the set of those cells. Now, we must estimate the location probability for each cell. We observed that for a particular cell, the number of paths or travel routes is nite, i.e., the MT is moving among nite states. The for , which is the probability that an MT, , currently in cell , will be in cell during the timing window , is computed by performing the following procedures: Step 1: Select a value from as the initial point for computing the location probabilities. Step 2: Start from the bottom line of TRM and take the last two non-zero elements of the TRM to make a temporary path as shown in Fig. 9(a). Step 3: Compare the path to the equal or close segment in PD as shown in Fig. 9(b). There may be a set of cells that can be the next cell along with path , which is represented by a set



We describe how to implement the proposed schemes within the existing cellular network specications [1]. Like all prediction related schemes, our scheme is also inevitably involved with overhead for computation and communications, which is worthy of discussion. A. Implementation in Existing Wireless Networks In our proposed framework in Fig. 4, we consider system parameters, such as radius size, time segments, and service description, which are available in the network databases. The implementation of prediction algorithms is completed in the base station which currently serves an MT, and does not affect the entire system design, but is related to the overhead described in Section V-B. The three aspects relating to the implementation 11

( &"  3a`Yb c z W}


~ P6

~ P6

(22) of , the number of cells is doubled For the since there are two possible zones resulting from the moving direction. Notice that the cells included in our illustration are those that fall into the zone partitions and within the average order, or prediction level . The set or sample space of the mobility proles is denoted as , and each cell that belongs to this set is denoted by .


& Y " s @`T4  8 & acY" ` z  } & " `9 8

One additional cell is added to each item in (21) because we consider the worst case scenario to be two incomplete cells falling into the probable zones. Similarly, we can have a general form for calculating the number of probable cells in the mobility proles, , that is

. Each element of this set, is a probable cell in Fig. 9(c), and which provides a possible path, . Step 4: Estimate location probabilities, , by performing the process shown in Fig. 10. This algorithm starts by examining each cell in the set of possible cells, , and the total number of cells in this set is from (22). If a probable cell, , is in the rst order of prediction and its corresponding path, , can be found in the historical path database, , then this cell, has the highest location probabilities. This emphasizes the importance of prediction constraints and user history. Additionally, as the probable cells get further away from the MTs current position, and they are not relevant to the historical paths, the location probabilities decrease. This examination continues until all cells in the possible set are scrutinized. As a result, a sequence of location probabilities is obtained in terms of , where can be solved by applying the following expression:


& `r9 8 "


( e d g ( e d q g g oo & 7 ' q o ( o 7 ( gfo %"7 q 7 e o  o ( o d ( e g ' o ( ( q g 7 o o o 7 q 7 o q( o gf( e 7 0 o ( o d ' o (

I & f d X "Ya&c  ' o ( o f & CacYPd'0 & "X

( 6 ~ 7%  p { d & Y s 5c"  s ffo s acY" `z  } &  oo sd (

&T s %c"  4 Y & s Y cP4 "

&" a`YT

& Y ( " I Va`Y"f d ' o ( #'( s '`Tp & X

& Y ( '( s G&c"  ( &" &" h#acYT3acYP Y & s Y %c"  & & s acY"  s a`Y"  6$u"!`"  & # Y 7 " d 3(7 7 & Y o ( (" ( s d c"  & acY" 



... 1

0 0 0 1 0 0 0 0 0 0 0 0 0

... ...

L2 L1 L

0 1 0 0 0 0 0 0 0 1


1 0 0 0 0 0



Trace \Records Matrix G (a)

0 0 0 1 0 0 0 0

Segment of Path Database (b)

Probable Cell

Fig. 9. Prediction of Location Probabilities.


... ... ...

Aggregate Historic Path

... 1

0 0 0 1 0 0 0 0 0 0 0 0 0


1 0 0 0 0 0





$$ g 03

d 0002 & 0000 1 # s Y %c"  & a`Y" 

)0000 000

& Y " T s %cP



:= initial value of the highest probability := temporary path of TRM := set of probable cells in PD along path := possible path of cell := := location probability at cell given an MT is currently in cell := threshold of probability computation do while do for all if and then else case :

and and

case : case :


case :

Fig. 10. Estimation of Location Probabilities (Step 4).

A.1 Location Information In wireless networks, mobile devices are required to update their locations when they cross the boundaries of two location areas. Therefore, new location area ID will be stored in the MTs home HLR as well as its current VLR. Therefore, the location description at the level of location area is already available in the system, and no extra effort is required to obtain this information. We recognize the effort in the updated specications [1] that the user proles are being required so that the locationdependent services can be provided by the third parties besides the vendors and service operators. The new parameters include current location area ID and target location area ID. The current location area ID indicates the location in which the subscriber is currently served, while the target location area ID indicates the location into which the mobile user intends to 12

A.2 Measurement and Real-Time Monitoring In the proposed framework, collection or measurement of historical records, including moving direction, speed, and current positions, is an important step to predict future locations. The user proles that are quasi-stationary are stored in the HLR and VLR/SGSN because they are considered to be rela-

v 3

$y {

and incorporation within existing networks are location description, real-time monitoring of speed and velocity, and collection of historical records.

end while end for end if end for end while

others : do for all while case :


roam. Note that the LCS (Local Client Service) is a new feature in 3G wireless systems, as specied in Universal Mobile Telecommunications System (UMTS) standards, to provide location information, which can be accessed through a gateway LCS. The proposed user prole framework can be used as an access point to user prole data for the service providers. Of course, the service-specic user preferences may be stored in a local database owned by a service provider. The implementation of this capability is the home subscriber server (HSS) [1]. The second level of our location description is at the zone level, which is a new concept introduced in this paper. The objective of this zone partition is to reduce the granularity of the location description at the LA level, combining information about an MTs current position and moving direction. Therefore, we can reduce the overhead of prediction. Note that, as in the design of LAs, each zone consists of a group of cells, which are represented by cell IDs. Therefore, given a cell ID, we know to which zone a cell belongs, i.e., the zone level information of location description. The remaining issues are whether the system knows an MTs moving direction and its cell ID. The former one involves realtime monitoring, which is discussed in detail in Section V-A.2. As for the collection of cell IDs, there are several ways for wireless systems to collect location information at the level of cell. When an MT initiates a call, it sends a routing request to the serving MSC through its BS, and MSC may convey this request to the HLR for call delivery. As a result, the network knows in which cell the calling MT is located. If the MT moves during the call connection, the updated position of an MT can be obtained through the call release procedure [40]. When an incoming call arrives at an MT, the network rst locates the MT in an MSC that is serving the called MT. Then the MSC pages the BSs in its controlling area in order to nd the serving BS of a called MT. The serving BS responds with a conrmation message to the MSC, indicating that the called MT is in its coverage area. Next, the call connection can be established between the calling and called MTs. Therefore, the called MTs current cell is known to the network [38]. The MTs location can also be obtained through location client services (LCS) management provided by wireless systems [37]. LCS is supported by serving mobile location center (SMLC), which may be call independent. Each SMLC controls a number of location measurement units (LMUs), which intervenes with MTs and BSs via air interface (Type A) or (Type B). The LCS provides the geographical estimation about an MT in terms of universal latitudinal and longitudinal data [39]. Mapping between the MTs position in terms of latitude and longitude and local coordinate system is nished through location client co-ordinate transformation function (LCCTF).

S ` s R ~e qr} jet `d wX Se wxR wvuWrcr` X q s g d wb X ts q z~e q} ne d`t ` s e ts wxwvuW X Scr` q d eg t d d` b wX o | pr S R qt ne d`t wX s lX k ` e t s v e t s ` m{gn de cbaY U zewxyvyxBo` X q wyxwvuiShg` R Xt q us g d b wX q s & ne d`t lX s g ` e mkp de bY U yxwvtuDSs coR ` r X q g d b wX q s & ne d`t lX k ` e s mpgj de bY U yxwvtuDoc` X q g d b S R net d` wX s lX k g mij de cb`aY U ewxwvuighX ` X q g d b ts g V UihdWb e ts d yxwvuWcr` X fe S e d`t wX e s wxwvtpr%et` X yvfet` X wWc` X q xywvusWcr` X q e t fedDcbaY U `X VX g d pUihWb U VWU S TUR

B. Overhead of Prediction Algorithms The overhead that will be incurred by the proposed scheme consists of three parts: 1) buffer space to store historical records; 13

 b Tb hRbg( b(  s (  


tive long-term information compared to parameters that change frequently, such as moving direction and speed. This information can be updated in three processes: location update, call origination, and call termination. During the location update, the current location area in which the mobile user is located, is updated with the serving VLR and the HLR. In the Third Generation Partnership Project (3GPP) specications, three location update schemes are suggested: distance-based, time-based, and movement-based. No matter which scheme is used for wireless networks, the location area ID will be changed upon location registration [3], [4]. For each mobile user, the serving BS and the corresponding VLR/SGSN will update the user proles for any incoming and outgoing calls due to billing purposes. Therefore, the mobile users current position with respect to location areas and cells are maintained in the HLR and VLR/SGSN by mobility application part (MAP) protocol during the call management process. The measurement of directions and velocity is an issue of real-time monitoring. In wireless cellular systems, there have been a few recognized algorithms for mobile velocity and direction estimations [17], [31], [46]. The velocity information can be collected during the handoff process. When an MT moves from one cell to another, the velocity estimation is required to keep handoff delay acceptable. These methods include the estimation of maximum Doppler frequency because of the fading process during the movement [21], [45]. The detailed algorithms about velocity estimation can be found in [17], [31], [45], [46], which is beyond the scope of this paper. According to the 3GPP specications, both the velocity and the location of an MT in a cell can be provided through the LMUs and LCS [39]. Each cell is associated with at least one LMU. Therefore, the location and speed information can be requested through BSs and the LMUs. In [11], the safe region is designed according to the least capable MTs and BSs by assuming that each entity can process the same maximum number of queries. In [28], a variable threshold scheme is proposed to process location-dependent queries so that a base station will not update the user prole if the query results are unlikely to affect the BS records. In [32], the performance of how to efciently process the querying records of mobile terminals is evaluated. Other related work also addresses the scalability issue of monitoring and collecting mobile users information by carefully constructing the databases that maintain the querying records [19]. This work has demonstrated that the BSs, LMUs, and VLRs are able to support discrete information monitoring and collecting. For the continuous monitoring in which the mobile objects are monitored continuously over a time period until the MTs or BSs interrupt, simulation results showed that the existing BSs and MTs are capable of handling even continuous monitoring if the databases and safe regions are designed appropriately. The more details about real-time monitoring and information processing in wireless networks can be found in [11], [19], [28], [32].

2) communication cost to send prediction results to probable cells for bandwidth reservation; and 3) computation time required to perform the algorithms to obtain prediction results. As described in Section III-A, the quasi-stationary, UMP will be updated during the procedures of location update, call origination and termination. This information is stored in the VLR/SGSN and HLR, i.e., the network core entities; therefore, there are no extra requirements for storing quasi-stationary information. On the other hand, the dynamic UMP can be determined according to the information maintained in the serving BS and the MTs. In the proposed framework, each MT keeps a list that stores the cell IDs that it has traversed and the services that it has requested compared to scheme that do not require historical records [2], [25]. This list is refreshed when the MT experiences a handoff, incoming, or outgoing service. According to the specications of cellular handsets, each one can store 50 to 100 records of calls and an additional 50 to 100 frequent calling users. Each of these records can accommodate 180 characters, which results in a total storage of bits to 576,000 bits. Considering each cell ID is 15 bits [5], 50cell IDs will take just 750 bits. The historical records of cell IDs take up to auxiliary storage space in a cellular handset, which is a very small amount. Thus, historical records are considered in many schemes [7], [13], [18]. Note that the historical records can also be stored in the serving base station rather than in the terminals, and the processing capability of a BS is much more powerful than a terminal. Therefore, using historical records will not cause signicant overhead in storage. During the handoff process, this information can be tailored to the signaling exchanged between the mobile terminals and the base stations. Another way to deal with the information change is that the previous BS forwards the information to the current serving BS. We consider that this is a better method because the information can be transmitted via the wired interface between BSs. Therefore, the radio resource used to communicate between the MTs and BSs will not be consumed while delivering user information. We consider three conditions for triggering the update process of location probabilities: the average residence time in a cell, the termination of a service, and the origination of a service. When a mobile user moves to a new cell, a timer is established for this MT. If one of the above three events occurs, the update process of location probabilities is initiated and the timer is reset to zero. For example, if an MT does not have any incoming or outgoing calls after it moves to a cell, the location probabilities will be calculated after the average residence time in a cell according to the stochastic mobility model. If an incoming call arrives before the average residence time has elapsed, the user prole in the HLR and VLR/SGSN will be updated for connection setup and accounting. Therefore, the computation and communication costs can be reduced. Compared to other algorithms introduced in Section I, the proposed scheme will incur extra communication costs for distribution prediction results to other cells. Except for the selective paging algorithm in [2], [23], all other schemes on the mobility prediction are instructed to nd the next cell instead of multiple cells. Let us consider that the maximum number of cells to be

VI. S IMULATION M ODEL In this section, assumptions and specications used in the simulation are described. Instead of assuming in many previous studies, that an MT travels along a highway or one-dimensional environment, we consider random behavior in our simulation, which involves much more complicated computation. We consider two scenarios in our service estimation: uniform and nonuniform distributions. As described in Section III-A, i.e., the service description of quasi-stationary UMP is able to provide the PMF of services as shown in (5). Suppose there are ve types of services supported by the MTs current wireless network, which are audio, video, voice, data, and voice mail. When the uniform PMF is the case, i.e., , the probability of each type of service is equal. As for the non-uniform PMF, we use the same example as in Section III. Each MT is allowed to request any type of service at a particular moment. We will predict the service pattern by using

We also consider the MTs speed and changes in its moving direction [17], [46]. The MT is allowed to move away from its current position in any direction, and variation from its previous direction is a uniform distribution limited in the range of . The initial velocity of an MT is assumed to be a random variable with Gaussian probability density function truncated in the range of and the velocity increment is taken to be a uniformly distributed random variable in the range of of the average velocity, . As for the residence time distribution in (1), the values of is taken with [48]. The most important feature of this simulation is that we use an actual digital map for computing the mobile location probabilities. In the selected segmentation of the MAP as shown in Fig. 11, there is an Interstate Highway I-85 and a toll freeway 400 on which most of the city trafc travels. We deliberately chose this segment, 7km by 5km, because it is a combination of diverse environments that provides a number of choices for a traveling MT, thus generating different location probabilities. If we only consider highways, then the location probability prediction is only related to the next cell in a one-dimensional model. By establishing a grid model, this area is covered by 30 cells covering the area. We nd that the maximum number of routes for each cell is 10. Thus, the maximum number of routes in our simulation is 300. Based on the geographical condition of this area, we generate a path database (PD), which will be used for predicting location probabilities in Section IV-C.2. 14

q u


$  p 

(I (

1 )  bb

U) d ( ( s S

predicted in our algorithm is in (22) for terminal given that the MT is currently in zone , and the prediction level is . Then, the total communication cost of informing other cells is obtained as , where is the communication cost for one cell, i.e., the cost for all algorithms of next-cell prediction. Therefore, the overhead of the proposed scheme is closely dependent on the number of cells in the mobility proles, which is also the number we need to reduce as much as possible. However, this communication cost does not take any radio resources; instead, it uses dedicated wirelines between base stations. With the transmission speed in ber optics at 10Gbps, this communication overhead is trivial. But, we obtain benets in terms of QoS improvement and decreased tracking cost. In fact, this information can also be used for advanced routing algorithms which require reservation and estimation [14], [23]. We can further reduce the overhead by grouping users. Although we describe the prediction scheme based on per-user basis, it is possible that the same results can be applied to different mobile users who follow the same mobility pattern. If a group of mobile users in one cell are moving along the same path within the time interval for records updating, then the location probabilities in the adjacent cells may be the same. However, we need to differentiate two mobile terminals that are in the same cell, but may result in different prediction results. For instance, there are two users that are both traveling on a highway, which is covered by a BS in the downtown area of a city. One of users is a local commuter, and the other user is passing on his way to another city. In other words, they are in the same cell, probably even move moving very close to the same speed, but they have different destinations. The system is not aware of these mobile users destinations, but it is possible to have their records through the HLRs of their home networks. Moreover, in the macrocell systems, the location probabilities and service types can be estimated for adjacent cells only rather than a group of cells. For micro-cell systems, which are now being used in many urban and suburban areas, the estimation for only one cell is not enough as the mobile users may pass through that cell in a short amount of time. The challenge is how to identify these cells, which is being addressed in this paper.

(10) to (16) in Section IV-B.

3 & Y 3 s wc"  & Y " s 5cPp

 & Id %" 

Fig. 11. MAP for Simulation.

The time is quantized in intervals , which is a preset time window. This interval is chosen based on cell residence time for each mobile terminal. In reality, different users may have different moving speeds and different mean residence times. However, for a specic cell, the average velocity can be determined for the mobile users in its coverage area. Accordingly, the cell residence time can be calculated by dividing the cell diameter by the average velocity. The calculation time window is much smaller than the cell residence time except in pico-cell systems. In our simulation, the service request from an MT is generated according to a Bernoulli process in each cell and each moment.

If the called MT is found in the PA, the average paging cost under delay bound , , is computed as follows: (25)

and the average delay,

B. Effect on Resource Management

When the estimation of service type is applied to resource management, we dene the probability vector, , as

where is the probability that an MT, remains in a cell given that the MT initiates a call in cell while this call is a type of service and it is not over by . for , is the probability that a call is still active in cell given that this call is initiated in cell . We also consider the pdf, , which represents the distribution of call holding time of service type . With the above consideration and recalling the denition of of (4) in Section III-B, for an MT, that is initiating a call of service in cell , the probability can be determined by the following expression: (28)

where is the cumulative distribution functions for and for is the probability that the MT will be in cell in (28) which is the product of two probabilities because we assume the calling pattern is independent of an MTs movement. is the probability that the call is not over by time , and is the probability that the MT is still in cell at time given that there are probable cells. For the other cells in the shadow cluster, we dene the cumulative distribution functions for probable cells with service as . Therefore, is determined by (29)



&  oo acY" `z  } fgo    I oo &  fgo acY"  c z W} &    acY" c z  } oo fgo B  x ~ Gw z S   o Ua&Y s `" z  x e x  a& Y s `"yz  x &  acY" z  x UaY s ` `z  aY s `"  c z  aY s `" c z  aY s `a z  x & "  & & S & " Y { UacYb `z W} | e (S Y & "  UaY s %" ee (S & s fgo s ( a`Y" c z  } o o s d aY& `" & aY " & s s %6 | UacY" `z W} e (S o UaY s %" 7e (h#aY s `" ~ R &  & S &  &aY s `" ~  {  d z  x  y & " aY s ` { s gfo s ( o o d aY s `b `z Y & "  { { d & aY s `" ~ c z  & S &  oo &  & s UaY s %" `z  fgo aY s %" `z  DaY s `" ~ c z  #aY s %" z  x &  aY s `" z  x

is a vector with

z & " S I z o { U Q U Q & " S y z & " S s o z U 3 Q z U 3 Q & " S  { zfi tz

is: ones and .. . .. .

I 6

In current personal communication service (PCS) systems, location update and paging are two fundamental operations for locating a mobile terminal (MT). As the demand for wireless services grows rapidly, the signaling trafc caused by location update and paging increases accordingly, which consumes limited available radio resources. Location update is concerned reports current locations of the MTs. In a paging process, the system searches for the MT by sending poll messages to the cells close to the last reported location of the MT at the arrival of an incoming call. Delay time and cost are two key factors in the paging issue. Of the two factors, paging delay is very important as the QoS requirement for multimedia services. Paging cost, which is measured in terms of cells to be polled before the called MT is found, is related to the efciency of bandwidth utilization and should be minimized under delay bound [3], [44]. To evaluate the performance of mobility proles, we use the paging costs and paging delays [2], [15] as metrics to demonstrate the effect of the designed schemes on mobility management. We will also compare the results with the so-called selective paging in [2] because the location probabilities are estiin our simulation. mated. The cell radius is assumed to be The full area of the segmentation map is covered by this type of cells, and we assume there is one BS in each cell and all of them are controlled by an MSC as in Fig. 1. A highest-probabilityrst (HPF) scheme is introduced in which the sequential polling is performed in decreasing order of probabilities to minimize the mean number of cells being searched [35]. of We assume that each LA consists of the same number cells in the system. The worst-case paging delay is considered as delay bound, , in terms of polling cycle. When is equal to , the system should nd the called MT in one polling cycle, requiring all cells within the LA to be polled simultaneously. The paging cost, , which is the number of cells polled to nd the called MT, is equal to . In this case, the average paging delay is at its least, which is one polling cycle, and the average paging cost is at its highest, . On the other hand, when is equal to , the system will poll one cell in each polling cycle and search all cells one by one. Thus, the worst-case scenario occurs when the called MT is found in the last polling cycle, which means the paging delay is at its maximum and equal to polling cycles [44]. However, the average paging cost may be minimized if the cells are searched in decreasing order of location probabilities [35]. We consider the partition of paging areas (PAs) given that , which requires grouping cells within an LA into the smaller PAs under delay bound [35], [44]. Suppose, at a given time, the initial state P is dened as = , where is the location probability of cell to be searched in decreasing order of probability. We use triplets to denote the PAs in which is the sequence number of the PA; is the location probability that the PA and is the number called MT can be found within the of cells contained in this PA. An LA can be divided into PAs because the delay bound is assumed to be . Thus, the worst polling cycles. The system case delay is guaranteed to be searches the PAs one after another until the called MT is found.

A. Effect on Mobility Management

Accordingly, the location probability

of the

is a diagonal matrix as:


PA is: (24)



z t


z& s z ys c{" e} U oo oo P6 s fgo s 6 s gfo s  6 s  6 S

g g (


As a result, each element of in (27) is determined. Accordingly, the bandwidth needed in the next probable cells for an MT, in cell , , can be reserved as:




A. Comparison of Single- and Multi-Cell Prediction In the proposed UMP framework, we consider the following factors in predicting future locations of mobile users: 1) historical records; 2) path information; 3) statistical model; 4) moving direction and velocity; 5) current position of a mobile object; and 6) shadow cluster, i.e., a set of possible cells. All of these factors impact location prediction as described in previous sections. In this section, we evaluate the proposed scheme with regard to the effect on mobility and resource management. In addition, we compare the proposed scheme with previous work on location prediction. As described in Section I, many previous schemes consider only some of the factors that we consider in the proposed UMP framework. This work can be categorized into single-cell prediction [7], [13], [16], [18], [25], [34], [48] and multi-cell prediction [2], [6], [23]. In the former category, each of them considers only two or three factors compared to six factors covered in our proposed scheme. Moreover, location estimation is focused on the next cell instead of a group of cells. Therefore, the impact of other factors is ignored and may also overlook the possibilities in other cells. Especially, for micro-cell systems, single-cell prediction not only can provide advance probabilistic information for resource reservation, but more importantly, the estimation of location probability can be used for location tracking as well. In particular, if there is no new interaction between the system and a mobile user since the last update, multi-cell prediction will have an important inuence on user tracking. Regardless of the reasons causes the interruption of communications, the system can always initiate searching or testing a mobile object based on its proles in mobile environments. Moreover, if the cell radius is very small, it may not be efcient to generate dynamic UMP frequently, e.g., per cell; instead, we can adjust the window time for calculating UMP to reduce computation and communication 16

B. Numerical Results First, we investigate the convergence of mean square error by using (15), which is shown in Fig. 12. For the uniform PMF, the rst convergence point is approximately , while it is approximately for non-uniform PMF. Thus, if an MT subscribes a service package which conforms to a uniform PMF, the next probable service type can be predicted more accurately compared to a non-uniform PMF. Accordingly, the bandwidth requirement can be determined. For example, the bandwidth requirement for each service type may be normalized by voice

We consider the stochastic admission control in our simulation [20], [23]. For each mobile user, the current serving base station periodically exchanges information with its adjacent and non-adjacent neighbors periodically with the equivalent bandwidth requirement. We assume that the time is quantized in slots of length . Also, we assume that new call requests are reported at the beginning of each time slot, and that a decision regarding an admission request is made sometime before the end of the time slot where the request was received. Every base station gathers call connection requests from MTs in its cell, and checks whether its current resources can support the requested equivalent bandwidth. A call is admitted if the sum of the equivalent bandwidth at a link is less than the link capacity, which is also called semi-resource reservation [20]. We use the metrics of handoff call dropping probability for handoff calls and call blocking probability for new arrivals.


& " aY s `a z  x o 3aY s %" z  x & & aY s `" z  x { & " aY s `|z  x


cost. The effect of single-cell prediction and multi-cell prediction are shown in Figs. 15 to 18. For the second category, i.e., multi-cell prediction, it has been very difcult to compare the location probabilities quantitatively. Thus, most of the literature will elaborate the rationale of the proposed schemes and demonstrate the results by the effect on either mobility management or resource management. In [23], the shadow cluster idea was proposed, but there is no further discussion on how to decide such a cluster and how to estimate location probabilities. To nd a group of cells in multicell prediction, in [6], the authors propose a prediction method to nd most likely cluster (MLC) based on directional probabilities. The method of MLC further approaches and reduces the number of cells dened in [23]. The location probabilities, therefore, are dened as the fraction of directional probability in one cell to all the cells through which a mobile object can traverse during the window time. However, there are three major concerns of this method. First, it depends on accurate direction measurements using the Global Positioning System (GPS). This assumption is relaxed in our work by using zone partition. In our algorithm, as long as the direction measurement is accurate to IV-C.2, no cells will be missed in the prediction. This is easier to implement in real systems. Second, this paper does not take user history into consideration. Although user history may not be important for macrocell systems, it is critical to micro- and pico-cell systems because mobile objects are very likely to change their directions frequently. Considering that mobile users usually follow a certain pattern in their movement, this pattern based on directional changes only, is not sufcient for mobility prediction. Third, there was no consideration of path information in [6]. No matter how likely a mobile object can move into a region from the analysis based on current measurements, it must move into an area that allows movement continuity, i.e., a path must exist in the next cell. Therefore, we take the path database into account by comparing available path with the predictive future path. In [2], moving direction, current position, and the grid layout of a cellular architecture are considered for location prediction. Similar to [6], user history and path information are ignored. Location probability is estimated based on cell residence time instead of directional changes. Although it is stated that the location probabilities can be computed, paper [2] is more focused on how to group those possible cells into subsets for paging purposes. The algorithm described in [2] is labeled as Selective Paging in Fig. 15. More details about the paging algorithms and comparison can be found in our paper [44].


0.09 0.08 Mean Square Error of Prediction Mean Square Error of Prediction 50 100 150 200 Number of Records 250 300 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0

0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0 50 100 150 200 Number of Records 250 300

(a) Uniform Fig. 12. Convergence of Mean Square Error.

(b) Non-Uniform

. Then, we estimate service type for an MT using the algorithm in Section IV-B. In our experiments, both uniform PMF and non-uniform PMF are considered. We compute the mean square errors using (14) and (15). Then, we determine the service type according to the scheme in Fig. 6.
0.6 = /3, firstlevel

Location Probabilities


$) f  X

 d ) f X

$) f  X

Then, we predict the probable cells using the algorithm described in Section IV-C. For , and , we rst determine the probable zones using (17) and (18), limiting the probable cells in a particular region. Next, the number of probable cells is computed using (1), (19) and (22). The computed probabilities are shown in Fig. 13 in a decreasing order of location probabilities. It can be seen that the number of probable cells inuenced by shadowing effect depends on the variation of the direction of an MTs movement as well as the velocity and prediction level. For the rst order of prediction, the number of cells is small, indicating that the probability in each cell is higher. For the second order of prediction, the location probabilities decrease with an increase in the number of cells. When the prediction level is the same, the variation of direction is the key factor that affects the number of cells and location probabilities. We notice

u & d ) f 3acY"  X

( D e ' s { { Y { { |s ( ' s b s {
0.5 = /2, firstlevel

service, such as

0.4 = /3, secondlevel




= /2, secondlvel




Number of Probable Cells

Fig. 13. Mobile Probabilities.

that there are fewer high probability cells than low probability cells because an MT is most likely to be moving into a few cells in the next moment rather than uniformly being dispersed in its surrounding cells. There are many ways to evaluate the effectiveness of the prediction algorithm of computing location probabilities [13], [14], [23], [29] as we introduced in Section I. Here, we show the effect of these results on paging issues since it involves both paging costs and paging delays. Paging is the process by which the MSC sends polling message to BSs in its management area to determine the serving cell of the called MT. Paging cost affects network resources because the paging message is sent via down-link channels; thus, it should be reduced as much as possible. Paging delay is part of call delivery delay, and it is related to QoS requirement. Thus, it should also be reduced so that the call connection can be established quickly. Paging costs resulting from location probabilities of rst and second levels of prediction are compared to those of uniform distribution assumed in the existing paging schemes [14], [15]. Note that the number of cells is determined by the prediction level and moving direction presented in Section IV-C.2, which has a strong impact on the performance of the scheme. Increasing the prediction level and change in direction, by including more cells, increases the likelihood of supporting QoS in future possible cells. But, this will cause more computations and communications. Therefore, we compare paging cost and delay with regard to change in direction and prediction levels. An extreme case is that the prediction level is zero, i.e., there is no prediction at all. We conduct the experiments on the following cases: with rst- and second-level prediction; without prediction for the cells in one or two layers surrounding the current cell; with rst- and second-level prediction as shown in Figs 14 and 15. The numerical results of paging costs are given in Fig. 14(a) and Fig. 14(b), in which paging costs are measured by the number of cells to be searched before nding the MT. Since the location probabilities provided in [2] are related to three cells, the paging cost is better than the two-cell prediction, and worse

u $) f 3a&cY"  X

than the 4-cell prediction as shown in Fig. 14(a) and Fig. 15(a). When the variation of moving direction is high, causing more probable changes, the improvement of paging costs is more visible in Fig. 14(a). For example, when , the reduction in paging costs due to the location probabilities is not as large . This means that it is more important to as that of predict location probabilities if the MTs are moving randomly, i.e., the movement of the MTs is not uniformly distributed in the location area. The paging cost of the proposed scheme is very close to the scheme proposed in [2] for the rst level prediction because there are a small number of cells in the shadow cluster. Therefore, the benets of the proposed scheme are not evident compared to the scheme for Selective Paging in [2], which also covers the cells adjacent to the current cell. If the prediction level is higher, the paging costs are signicantly reduced compared to those without prediction. Specically, if MTs are moving very fast and are expected to other cells in a short time, there are more probable cells. As a result, location probability in each cell is smaller. As a result, it is more difcult to locate the MT. Accordingly, the prediction of MTs location probabilities is more effective and more important. Paging delays are compared in Fig. 15(a) and 15(b), which are measured in terms of polling cycles. Each polling cycle is the time from sending a paging request to receiving a response. In Fig. 15(a), the paging delays are greatly reduced in comparison to not using prediction. As the delay constraints increase, the average paging delays increase while the paging costs decrease. We notice that the delays are reduced even more when the delay constraints are greater. Also, the predicted probabilities are more effective on reducing paging delays when the prediction level is higher as shown in Fig. 15(b). This is especially useful for those MTs moving in wide areas, where there are many paths available instead of high-way scenarios. We also conduct simulations of resource management. The predicted service type and location probabilities are used to reserve the equivalent bandwidth for the mobile terminals. In our simulation, we consider different new call ratio as the number of requests of new calls to the total number of call requests. The results shown in Fig. 16 to Fig. 17 are the statistics of 20,000 call requests. If the requested bandwidth cannot be allocated, then a new call will be blocked or the handoff call will be dropped. We consider that handoff calls have higher priority than new calls. Therefore, call dropping/blocking probabilities are related to the new call ratio, which is the fraction of new calls in the total number of connections requested. For each of them, we compare no-reservation, next-cell reservation, two-cell reservation, and 4-cell reservations. In particular, next cell reservation corresponds to single-cell prediction discussed in Section VIIA. In Fig. 16, we can observe that call dropping probabilities increase as call arrival rates increase, indicating that the bandwidth release depends on the call departure. Given the xed call holding time, the more call arrivals, the higher call dropping and blocking probabilities. Meantime, we can see that the call dropping probabilities decrease as we reserve bandwidth in more cells, i.e., better than single-cell prediction. The effect of reservation are obviously on the handoff dropping probabilities

than on the new call blocking probabilities. If major call requests come from handoff, then call dropping probabilities are higher, compared to the case in which the new calls are dominant.
4 x = /2, one layer without prediction 3.5 = /2, firstlevel prediction

Paging Costs

Paging Costs

u q w  X

 q u 0X

Selective Paging


x = /3, one layer without prediction x= /3, firstlevel prediction

1.5 1 1.5 2 2.5 Delay Constraints 3 3.5 4

(a) First-Level Prediction

12 x = /2, two layers without prediction x= /2, secondlevel prediction x = /3, two layers without prediction


x= /3, secondlevel prediction

0 2 4 6 8 Delay Constraints 10 12

(b) Second-Level Prediction Fig. 14. Comparison of Paging Costs.

VIII. C ONCLUSIONS We have explored a fundamental issue of providing high level QoS in wireless networks. Noticing that the key point to QoSbased mobile networks is the knowledge of service requirements and future locations prior to the arrival of mobile objects, we rst proposed a novel framework of UMP, considering many important factors associated with the MTs behavior. Then, we categorized the UMP information into quasi-stationary UMP and dynamic UMP, and discussed how to coordinate them for the prediction of services and possible locations. The service requirement is estimated by using a mean square error method based on the historical records and service probability distributions. Moreover, we introduced the concept of zones and predic18


= /2, one layer without prediction


0.35 Call Dropping Probability 0.3 0.25 0.2 0.15 0.1

No resv Next cell resv 2 cell resv 4 cell resv 8 cell resv

2 Paging Delays

Selective Paging = /3, one layer without prediction

= /3, firstlevel prediction



= /2, firstlevel prediction


0.05 0 0.05 0.1 0.15 0.2 0.25 0.3

1 1

Next cell prediction 1.5 2 2.5 Delay Constraints 3 3.5 4

Call Arrival Rate with 10% New Calls

(a) First-Level Prediction


(a) Handoff Dropping Probabilities

6 5.5 5 Paging Delays 4.5 4 3.5 3 2.5 2 1.5 1 2

x = /2, two layers without prediction

0.02 Call Blocking Probability

x = /3, two layers without prediction

No resv Next cell resv 2 cell resv 4 cell resv 8 cell resv


= /3, secondlevel prediction




x= /2, secondlevel prediction 4 6 Delay Constraints 8 10 12







Arrival Rate with 10% New Calls

(b) New Call Blocking Probabilities (b) Secon-Level Prediction Fig. 16. Dropping and Blocking Probabilities with 10% New Calls. Fig. 15. Comparison of Paging Delays. I.F. Akyildiz, J. McNair, J.S.M. Ho, H. Uzunalio lu, and W. Wang, Mog bility Management in Next-Generation Wireless Systems, Proceedings of the IEEE, Vol. 87, No. 8, pp. 1347-1384, August 1999. [4] I.F. Akyildiz and W. Wang, A Dynamic Location Management Scheme for Next Generation Multi-Tier PCS Systems, IEEE Transactions on Wireless Communications, Vol. 1, no. 1, pp. 2040-2052, January 2002. [5] D.P. Agrawal and Q-A. Zeng, Introduction To Wireless and Mobile Systems, Thomson Brooks/Cole, 2003. [6] A. Aljadhai and T.F. Znati, Predictive Mobility Support for QoS Provisioning in Mobile Wireless Environments, IEEE Journal on Selected Areas in Communications (JSAC), Vol. 19, No. 10, pp. 1915-1931, October 2001. [7] A. Bhattacharya and S. K. Das, LeZi-Update: An Information-Theoretic Approach to Track Mobile Users in PCS Networks, ACM/IEEE MobiCom99, August 1999. [8] E. Cayirci and I.F. Akyildiz, User Mobility Pattern Scheme for Location Update and Paging in Wireless Systems, IEEE Transactions on Mobile Computing, Vol. 1, No. 3, pp. 236-247, July-September 2002. [9] E. Cayirci and I.F. Akyildiz, Optimal Location Area Design to Minimize Registration Signaling Trafc in Wireless Systems, IEEE Transactions on Mobile Computing, Vol. 2, No. 1, pp. 76-85, January-March 2003 [10] T. Camp, J. Boleng, and V. Davies, A Survey of Mobility Modelsfor Ad Hoc Network Research, Wireless Communications & Mobile Computing, Vol. 2, no. 5, pp. 483-502, 2002. [11] Y, Cai and K. A. Hua, An Adaptive Query Management Technique for [3]

tion levels to shrink the region of probable cells. In the proposed prediction algorithms, we took several important factors into account, including direction and velocity of mobile objects, historical records, stochastic model of cell residence time, and path information into account. Therefore, the service requirement and future locations are predicted more accurately compared to previous schemes because this is the rst time that all of these factors are considered. As examples, we provided the simulation results to demonstrate that the proposed algorithms for mobility and resource management are effective in terms of reducing location tracking cost, delays, and call dropping/blocking probabilities.



Generation Partnerships Project: 3G TS 23.002 v3.6.0 (2002-09), Technical Specication Group Services and Systems Aspects: Network Architecture/(Release 1999), September 2002. A. Abutaleb and V.O.K. Li, Paging Strategy Optimization in Personal Communication System, ACM-Baltzer Journal of Wireless Networks (WINET), Vol. 3, pp. 195-204, August 1997.


0.04 0.035 Call Dropping Probability 0.03 0.025 0.02 0.015 0.01 0.005 0

No resv Next cell resv 2 cell resv 4 cell resv 8 cell resv


[21] [22] [23]

0.05 0.1 0.15 0.2 Call Arrival Rate with 10% New Calls 0.25 0.3


(a) Handoff Dropping Probabilities [26]

0.4 0.35 Call Blocking Probability 0.3 0.25 0.2 0.15 0.1 0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 No resv Next cell resv 2 cell resv 4 cell resv 8 cell resv

[27] [28] [29] [30] [31] [32] [33] [34]

Arrival Rate with 90% New Calls

(b) New Call Blocking Probabilities Fig. 17. Dropping and Blocking Probabilities with 90% New Calls.

[35] Efcient Real-Time Monitoring of Spatial Regions in Mobile Environments, Proc. of the 21st IEEEE Intl Performance, Computing, and Communications Conference (IPCCC), pp. 259-266, April 2002. K. Choi and W. Jang, Development of a transit network from a street map database with spatial analysis and dynamic segmentation, Transp Res Part C Emerging Technol, Vol. 8, No. 1, pp. 129-146, January 2000. S. Choi and K.G. Shin, Predictive and Adaptive Bandwidth Reservations for hand-offs in QoS-sensitive Cellular Networks, ACM SIGCOMM1998, pp. 155-166. Vancouver, August 1998. M. Y. Chung and D. K. Sung, Performance Analysis of a Prole Management Scheme for Incall Registration/ Deregistgration in Wireline UPT NetworksPart I: Request-Based Scheme, IEICE Transactions on Communications, Vol. E82-B, No. 5, pp. 686-694, May 1999. D. Gu and S.S. Rappaport, A Dynamic Location Tracking Strategy for Mobile Communication Systems, IEEE VTC2000, vol. 1, pp. 259-263, Ottawa, May 1998. M. Hellebrandt and R. Mathar, Location Tracking of Mobiles in Cellular Radio Networks, IEEE Transactions on Vehicular Technology, vol. 48, no. 3, pp. 1558-1562, September 1999. M. Hellebrandt and R. Mathar and M. Scheibenbogen, Estimation Position and Velocity of Mobiles in a Cellular Radio Network, IEEE Transactions on Vehicular Technology, Vol. 46, No. 1, pp. 65-71, February 1997. J.S.M. Ho and J. Xu, History-Based Location Tracking for Personal Communications Networks, IEEE VTC1998, vol. 1, pp. 244-248, Ottawa, May 1998. C.S. Jensen, A. Friis-Christensen, T.B. Pedersen, D. Pfoser, S. Saltenis, [36] [37] [38]

[12] [13] [14]

[39] [40]

[15] [16] [17] [18] [19]

[41] [42]

and N. Tryfona, Location-Based Services - A Database Perspective, Proc. of the Eighth Scandinavian Research Conference on Geographical Information Science, pp. 59-68, June 2001. G.S. Kuo and P-C. Ko and M-L. Kuo, A Probabilistic Resource Estimation and Semi-Reservation Scheme for Flow-Oriented Multimedia Wireless Networks, IEEE Communications Magazine, Vol. 39, No. 2, pp. 135141, February 2001. K. Kawabata, T. Nakamura, and E. Fukuda, Estimating Velocity Using Diversity Reception, Proc. IEEE Veh. Tech. Conf., vol. 1, 1994, pp. 371374. J. Korhonen, Introduction to 3G Mobile Communications, Artech House, 2001. D.A. Levine, I.F. Akyildiz, and M. Naghshineh, A Resource Estimation and Call Admission Algorithm for Wireless Multimedia Networks Using the Shadow Cluster Concept, IEEE/ACM Transactions on Networking, Vol. 5, No. 5, pp. 1-12, February 1997. A. Leon-Garcia, Probability and Random Processes for Electrical Engineering, Second edition, Addison-Wesley, 1994. T. Liu, P. Bahl and I. Chlamtac, Mobility Modeling, Location Tracking, and Trajectory Prediction in Wireless ATM Networks, IEEE Journal on Selected Areas in Communications (JSAC), Vol. 16, No. 6, pp. 922-936, August 1998. Y-B. Lin, Reducing Location Update Cost in a PCS Network, IEEE/ACM Transactions on Networking, Vol. 5, No. 1, pp. 25-33, February 1997. Y-B. Lin and W-N. Tsai, Location Tracking with Distributed HLRs and Pointer Forwarding, IEEE Transactions on Vehicular Technology, Vol. 47, No. 1 , February 1998. K. Lam, O. Ulusoy, T.S. H. Lee, E. Chan, and G. Li, An Efcient Method for Generating Location Updates for Processing of Location-Dependent Continuous Queries, Proc. of DASFAA2001, pp. 218-225, April 2001. J. Mii` and T. Y. Bun, On Call Level QoS Guarantees under Heterogesc neous User Mobilities in Wireless Multimedia Networks, IEEE GLOBECOM99, Vol. 3, pp. 2730-1769, December 1999. W. Ma and Y. Fang, Two -Level Pointer Forwarding Strategy for Location Management in PCS Networks, IEEE Transactions on Mobile Computing, Vol. 1, No. 1, 32-45, March 2002. R. Narasimhan and D.C. Cox, Speed Estimation in Wireless Systems Using Wavelets, IEEE Transactions on Communications, Vol. 47, pp. 13571364, 1999. D. Pfoser, C.S. Jensen, and Y. Theodoidis, Novel Approaches in Query Processing for Moving Objects, Proc. of Very Large Database (VLDB) Conference 2000, September 2000. S. Ross, A First Course in Probability, Prentice Hall, Sixth Edition, 2002. X. Shen, J.W. Mark and J. Ye, User Mobility Prole Prediction: An Adaptive Fuzzy Inference Approach, ACM-Baltzer Wireless Networks (WINET), No. 6, pp. 363-374, June 2000. C. Rose and R. Yates, Minimizing the Average Cost of Paging Under Delay Constraints, ACM-Baltzer Journal of Wireless Networks (WINET), Vol. 1, pp. 211-219, February 1995. J. Sun and H.C. Lee, Mobility Data Pattern Tracking via Multidimensional Data View Approach, IEEE MilCom97, pp. 1255-1260, November 1997. ETSI TS 101 513 V8.0.1 (2000-11), Digital cellular telecommunications system (Phase 2+); Location Services (LCS); Location services management, (GSM 12.71 version 8.0.1 Release 1999), November 2000. ETSI TS 123 012 V4.0.0 (2001-03), Digital cellular telecommunications system (Phase 2+) (GSM); Universal Mobile Telecommunications System (UMTS); Location Management Procedures, (3GPP TS 23.012 version 4.0.0 Release 4), March 2001. ETSI TS 123 271 V4.1.0 (2001-03), Universal Mobile Telecommunications System (UMTS);Functional stage 2 description of location services, 3GPP TS 23.271 version 4.1.0 Release 4, March 2001. ETSI TS 129 010 V4.0.0 (2001-03), Digital cellular telecommunications system (Phase 2+) (GSM); Universal Mobile Telecommunications System (UMTS); Information element mapping between Mobile Station - Base Station System (MS - BSS) and Base Station System - Mobile-services Switching Center (BSS - MSC); Signaling procedures and the Mobile Application Part (MAP), (3GPP TS 29.010 version 4.0.0 Release 1999), March 2001. 3G TS 23.018 v3.5.0 (2000-06), 3rd Generation Partnership Project; Technical Specication Group Core Network; Basic Call Handling; Technical Realization, June 2000. W. Wang, Modeling and Management of Location and Mobility, Wireless Information Highways, To be published in March 2004.


[43] W. Wang and I.F. Akyildiz, Intersystem Location Update and Paging Schemes for Multitier Wireless Networks, ACM/IEEE MobiCom00, pp. 99-109, August 2000. [44] W. Wang, I.F. Akyildiz, G. St ber, and B-Y. Chung Effective Paging u Schemes with Delay Bounds as QoS Constraints in Wireless Systems, ACM-Baltzer Wireless Networks (WINET), Vol. 7, No. 5, pp. 455-466, September 2001. [45] C. Xiao, Estimating Velocity of Mobiles in EDGE Systems, Proc. 2002 IEEE ICC, pp.3240-3244, April 2002. [46] C. Xiao, K.D. Mann, and J.C. Olivier, Mobile Speed Estimation for TDMA-Based Hierarchical Cellular Systems, IEEE Transactions on Vehicular Technology, Vol. 50, No. 4, pp. 981-991, July 2001. [47] T. Zhang, E.V.D. Berg, J. Chennikara, P. Agrawal, and J-C. Chen and T. Kodama, Local Predictive Resource Reservation for Handoffs in Multimedia Wireless IP Networks, IEEE Journal on Selected Areas in Communications (JSAC), Vol. 19, No. 10, pp. 1931-1941, October 2001. [48] M.M. Zonoozi and P. Dassanayake, User Mobility Modeling and Characterization of Mobility Patterns, IEEE Journal of Selected Areas in Communications, Vol. 15, No. 7, pp. 1239-1252, September 1997.