The Predictive User Mobility Profile Framework for Wireless Multimedia Networks Ian F. Akyildiz
Broadband and Wireless Networking Laboratory Georgia Institute of Technology, Atlanta, GA 30332 Email:
[email protected] Abstract – User Mobility Profile (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. Keywords—Wireless Network, User Mobility Profile, Quality of Service, Mobility and Resource Management.
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 profile (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 efficient 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.
Wenye Wang
Department of Electrical and Computer Engineering North Carolina State University, Raleigh, NC 27695 Email:
[email protected] Call admission control (CAC) and resource management. An efficient 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 traffic 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 benefit 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 user’s 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 user’s records and IDs of most recently visited cells [18]. In [6], the authors proposed a model of mobility profiles that included the trajectory estimation of mobile users and arrival/departure times in each cell along a terminal’s 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
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.
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 finding 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 efficient resource management. Furthermore, UMP is not well defined, which should consider the characteristics of users’ mobility and service patterns. In addition, most of the existing solutions do not provide the flexibility 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 profile 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 fine 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 defined, 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
II. S YSTEM M ODEL , L OCATION D ESCRIPTION , M OBILITY M ODEL
AND
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.
MSC
MSC
Cell
BS
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
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 MT’s position than LAs because a zone partition has much less cells than those that are included in an LA.
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 influenced cells also change over time. The home base station that is serving the MT’s 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 specific 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.
B. Location Description Before we describe detailed mobility model, let us look at location description first 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 specified at three levels. In other words, a mobile user’s 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 MT’s 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 sufficient to predict future locations. Thus, we use the zone concept proposed in the previous section to take into account an MT’s current position as the MT’s 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 MT’s movement direction and position (zone) to provide a more accurate prediction of the MT’s 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
MSC−A N
MSC−B
α Zone 6 W
E
Zone 1 W
E 01 S
Zone 3
Zone 5 07
MSC−C
β
Zone 2 Zone 0
S
N
Zone 4
02
Location Area 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
3
MT’s current direction is defined as the direction from its previous position to the current position. Here we denote as the moving direction of the MT , which is defined 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 MT’s 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 MT’s dynamic information is related to real-time monitoring in wireless networks [11], [28], [32]. The major issue is how to define 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.
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 defined 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 specified as user profiles. 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.
A8BDC : CEB :FD5 AHGJI G 7
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 specific 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 five components in this model to characterize the movement of a mobile user: The MT’s 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 flexibility because given different parameters, a Gamma distribution can be an Exponential, an Erlang or a Chi-square distribution. The pdf of an MT’s residence time with with Gamma distribution in [24] has Laplace transform the mean value and the variance . Then,
!
,
where
#, /- . 021
Current Cell Moving Direction Current Direction
θ =−
Highest Probability Cells
π 2
Higher Probability Cells
Low Probability Cells
Fig. 3. Moving Direction.
Moving speed: We also consider the MT’s 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 MT’s 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 user’s 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
K G
" $ # &%' $#)(+*
π 2
θx=
:>6?;=
@
UVXW
(1)
3465 879 .
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 MT’s current direction, is collected through realtime monitoring, which can be initiated by the serving BS. The
ZY\[
: QPR ST7 F S
[
]_^a` bc IedTI 1f1g1 I hi IedTI 1f1g1 I [
matrix
and
]>^a`
j
the location probabilities of the cells in the shadow cluster and the estimated service type during an MT’s roaming into other cells.
can be written as
lmm ] ] ogofo ] rfss mm e. . . = .p ss mmn .. . . .. .. ss jk ofogo ofgo o ] ^a` offo o I .. .. .. t .. . . . . ]q ]q ogofo ]q p
User Mobility Profile
(2)
Quasi−Stationary UMP
is given by
]>^a`uwv
if the MT has entered this cell otherwise
1
Network Configuration
(3)
Time Period
@
{
Location Probabilities
Calling Pattern
Fig. 4. The User Mobility Profile (UMP) Framework.
There are two types of data in the new UMP framework as shown in Fig. 4. The first type is called quasi-stationary UMP, which represents the MT’s 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 find the dynamic UMP, short-term information. The following two subsections present the details of the framework.
y x z
