Mobile Networks and Applications 1 (1996) 113-121 113

A class of mobile motion prediction algorithms for wireless mobile

computing and communications
George Liu a and Gerald Maguire Jr. b
a System Research Department, Ericsson Radio Systems, S-164 80 Stockholm, Sweden
b Department ofTeleinformatics, Royalfnstitute of Technology, S-16440 Stockholm, Sweden

Abstract. This paper describes a class of novel mobile motion prediction algorithms for supporting global mobile data acces-
sing. Traditionally, mobility and routing management includes functions to passivelykeep track of the location of the users/term-
inals and to maintain connections to the terminals belonging to the system. To maintain uninterrupted high-quality service for
distributed applications, it is important that a mobile systembe more intelligent and can anticipate the change of the location of its
user. We propose an aggressive mobility and routing management scheme, called predictive mobility management9A class of
mobile motion prediction algorithms predicts the "future" location of a mobile user according to the user's movement history, i.e9
previous movement patterns, By combiningthis schemewith mobility agent functions, the serviceand user routing data are actual!y
pre-connected and pre-assigned at the locations to which the user is moving. Thus, the user can immediatelyreceiveserviceor data
with virtually the same efficiencyas at the previous location, i.e., without encountering a large "data structure handover" delay
before serviceor data is available.

1. I n t r o d u c t i o n tion of a mobile user according to the user's m o v e m e n t

history patterns 9 The data or services are pre-connected
Wireless mobile computing [1-4,6,9,11] has been and pre-assigned at the new location before the user
extensively studied in recent years in both the computer moves into the new location. Thus, the user can immedi-
and communication communities 9 However, most of the ately receive service or data with virtually the same effi-
recent studies focus on the network layer communica- ciency as at the previous location.
tion protocols and few of them on the mobility aspects The M M P algorithms are based on the fact that
caused by mobile users behavior. To m a k e intelligent everyone has some degree of regularity in his/her move-
mobile-aware applications and efficient routing man- ment, that is, the m o v e m e n t of people consists of r a n d o m
agement, it is important that a mobile system be more m o v e m e n t and regular movement, and the majority of
intelligent and can anticipate the change of the location mobile users have some regular daily (hourly, weekly,
of its user. On the other hand, service and resource mobi- 9 movement-patterns and follow these patterns m o r e
lity [14,16] introduce new requirements for mobility or less every day of the week (particular hours, particular
management. Mobility is an aspect of user behavior days, ...). A typical example is the "boring professor/
which impacts wireless networks as well as wired ones, researchers" home-office-home pattern 9
The concepts of mobility m a n a g e m e n t have been used in Two basic m o v e m e n t models are proposed, namely,
the cellular mobile networks for m a n y years, but, it was the M o v e m e n t Circle (MC) and M o v e m e n t Track (MT)
primarily designed to support mobile voice communica- models, which efficiently extend the M a r k o v model in
tion. Traditionally, mobility m a n a g e m e n t includes func- mobile applications for modelling the m o v e m e n t of
tions to passively keep track of the location of the users/ mobile users. The M C algorithm is based on the closed
terminals and to maintain connections to the terminals circuit-like (MC) model of user m o v e m e n t behaviors
belonging to the system [7,8]. H o w to adopt mobility and is used for predicting long-term regular movement.
m a n a g e m e n t for efficiently supporting mobile data rout- Such closed circuits, representing a return to a previous
ing, mobile data accessing and mobile computing is still state, are topologically a cycle, The M T algorithm based
an open question. on the M T model is used for predicting the routine ones.
We argue that to maintain uninterrupted high-quality Three classes of matching schemes are used for correla-
service for distributed applications in a mobile envi- tion analysis of MCs or MTs.
ronment, the data routing should not only passively The rest of this paper is divided into the following
reflect the change of terminal location, but also aggres- sections. Section 2 describes the background and pre-
sively anticipate the behavior of mobile users. We pro- vious work. Section 3 proposes the m o v e m e n t models 9
pose an aggressive mobility and routing m a n a g e m e n t Section 4 describes the mobile motion prediction algo-
scheme, called Predictive Mobility M a n a g e m e n t rithms. And finally, in section 5, some simulation
(PMM). A set of Mobile Motion Prediction ( M M P ) results of the mobile motion prediction algorithms are
algorithms [15,17] is used to predict the " f u t u r e " loca- presented.

9 J.C. Baltzer AG, Science Publishers

114 G. Liu, G. MaguireJr. / Mobilemotionprediction algorithms

2. B a c k g r o u n d a n d p r e v i o u s w o r k area or a cell in the service area. A state variable can be

denoted by Sk,, where k<~K and t E T, K is the number
Several user mobility and movement models can be of states in the state space and T is an index set. The state
found in the literature for modelling the mobility behav- space is the set of all possible values and is denoted by I.
ior of mobile users. Among them, a widely used one is A state is called a Stationary State (SS) if a mobile has
the fluid flow model [10] in which it is assumed that stayed in the state for at least a period of time r~. A state
mobile users are uniformly populated and the users car- is called a Transitional State (TS) ira mobile has stayed
rying terminals are moving at an average velocity with in the state for less than a period of time r~. A state is
uniformly distributed moving-direction over [0, 2rr]. called a Boundary State (BS) if the state is at the bound-
With the above assumptions, this model has been effi- ary of the service area.
ciently used in studying the performance of PCS net- In modeling practical movement situations of mobile
works and the impacts of user mobility on the database users, the daily movement of a mobile user can be subdi-
of PCS systems. vided into basically two parts: regular movement and
To study the handoff procedures in mobile systems, random movement. Let Sk,t be state k at time t of the reg-
Hong and Rappaport [5] proposed a mobility model ular movement and X(k, t) be state k at time t of the ran-
which assumes that the speed and direction of a mobile dom parts. Thus, the regular movement process is
terminal are independent random variables. The direc- { S~,t I k <~K, t E T} and the random movement process
tion of a mobile terminal is assumed to be uniformly dis- is {X(k, t) I k ~<K, t E T}. The movement process can be
tributed in the range of [-rr/2, re/2]. It is also assumed denoted by
that when mobile users cross cell boundaries, their speed
{M(k,t)} = {&,t I k<. K , t 9 T}
and direction are independently and randomly regener-
ated. + {X(k,t) I k < K , t c f}. (1)
These models are useful and efficient for analyzing
the performance of mobile systems from the " m a c r o " To model the regular parts of the movement, two
point of view, i.e., focusing on the total population of movement models, namely Movement Circle (MC) and
mobile users. But they are not suitable for predicting the Movement Track (MT) are proposed. Stochastic pro-
movement of mobile users since the characteristics of cesses and Markov chain model are used to model the
each mobile user are not considered individually. To random parts of the movement.
study and predict the movement of mobile users, we take
another approach: modelling the individual behavior of 3.1. Movement Circle (MC) model
the movement of the mobile users. We consider the
movement of individual mobile user to consist of two The MC model is based on the assumption that
parts: a random movement component and regular wherever a user moves from a location, the user will
movement component. In turn, the regular movement eventually return or has returned (backward). Thus,
consists of identical movement patterns of each mobile the movement behavior of mobile users are modeled as
user which represents the special behavior of the move- different circle-like patterns. The states are represented
ment of the user within a defined period of time. We by small' circles identified by numbers which indicate
further use the concepts of stationary, boundary and the Location Areas (LA) corresponding to the states.
transitional states to form and identify movement pat- An MC is defined as a circle which consists o f n (n > 1)
terns. The M T / M C model is proposed to model the reg- unidirectional sequential states with at least 1 station-
ular movement of mobile users. ary state or boundary state. As stated earlier, a bound-
ary state is a state at the boundary of the service area
and a state is called a stationary state if a mobile has
3. T h e p r o p o s e d m o v e m e n t m o d e l s stayed in the state for at least a period of time rs. An
MC is called boundary MC if one or more states in an
We consider a cellular system covering a large geo- MC are boundary states. A BMC has higher priority
graphic area (service area) with many hexagonal cells than an MC with a boundary priority factor /3 (/3 > 1).
with each cell being serviced by a base station. A cell or a Each state in reality may represent an LA or a cell in a
group of cells may form a Location Area (LA) in which cellular system.
the location information is periodically broadcast to the
covered area by the basestation(s). The mobile terminals D e f i n i t i o n o f a m o v e m e n t circle process. An MC
carried by users can themselves identify the current geo- process is defined to be a discrete-parameter and
graphic location by listening to identification informa- discrete-state process { Sk,t ] k <~K, t E T} in which
tion. Each geographic location information (identity there exists Si,t = Sj,t for any t E T,j - i > L (L~>2),
area) corespondents to a state signal received by the where the Si,t, Sj,t are stationary states or boundary
terminal. states. An MC process is denoted by
Let S be the state which is corespondent to an identity {Me(k,t) I k<~K,t C T}.
 G. Liu, G. MaguireJr./MobiIemotionprediction algorithms 115

3,2. Movement-Track ( MT) model t r T, where K is the sample space. Because the daily
m o v e m e n t of a mobile user m a y not always be r a n d o m #1
While the M C model addresses long-term, regular and the m o v e m e n t sequence states are not independent,
user movements, the M T model is a less-constrained ver- that is, the nth- order joint distribution of the m o v e m e n t
sion of the M C and addresses routine movements. As process does not satisfy the following independent condi-
shown in Fig. 1, an M T is an unidirectional itinerary tion:
which begins and ends with a stationary state or a bound-
ary state. F o r example, if a mobile has stayed at location
1 for more than a period of time 7-s, then passed through
V(x; t) = V(xi; ti) = r I P[Sk,,, <.xii,
i=l i=1
location 16, 17, 18 (the time stayed at each location is less
than ~-~)and finally stopped at location 21 for a time per- where x = ( x l , . . . , x n ) , t = ( t l , . . . , t n ) , a M a r k o v pro-
iod longer than ~-s, it forms a m o v e m e n t track. This cess is not completely suitable for modelling the move-
m o v e m e n t track can be recorded as an M T which begins ment processes.
with state 1 and ends with state 21, as shown in Fig. 1 Therefore, we only use the M a r k o v chain model to
(the first MT). Since an M T is an unidirectional itiner- describe the behavior of the r a n d o m parts of the user's
ary, if the mobile has been m o v e d the same way back to movement. We assume that the regular part of the move-
location 1, then it forms another M T , i.e., the third M T ment process consists of only the M C processes and M T
in Fig. 1. An M T is called Boundary M T ( B M T ) if one or processes, that is:
more states in an M T are boundary states. A B M T has
{Mc(k,t) ]k<~K, t E T } + { M r ( k , t ) [k<...K,tr T}
higher priority than an M T with a boundary priority fac-
tor ( > 1). = {sk, ! k K, t c r } . (2)
An M C m a y consist of one or more MTs which begin
and end with the same stationary or boundary states. In So, the remainder of a m o v e m e n t process after the
some special case, an M T can be an M C itself if it begins M C and M T processes have been detected and removed
and ends with the same stationary or boundary states, is a r a n d o m process, that is, applying (2) to (1), we get
such as the second M T in Fig. 1.
{X(k,t) lk<,,K,t e T} = {M(k, t)} - ({M~.(k,t) lk<~K,
D e f i n i t i o n o f a m o v e m e n t t r a c k process. An M T t E T} + {Mr(k,t) lk<~K,t C V}). (3)
process is defined to be a discrete-parameter and
discrete-state process {Sk,t I k<<.K, t r T} in which there We also assume that the remainder process
exists a n u m b e r of L (2 < L~<K) stationary states or {X(k, t) [ k ~<K, t r T} is a M a r k o v process in which the
b o u n d a r y states Sk,~, for any t E T. An M T process is probability distributions for its future development
d e n o t e d b y {MT-(k, t) I k<~K, t E T}. depends only on the present state and not on how the
process arrived in that state #2
3.3. Markov chain model
3.4. MC/MTmodel versus Markov model
The m o v e m e n t of a mobile user can be modeled by a
discrete-parameter and discrete-state stochastic process While the M a r k o v model assumes that state is ran-
if and only if it is assumed that the m o v e m e n t is random, d o m variable and the probability distributions for its
that is, the M(k, t) is a r a n d o m variable for any k c K, future development depend only on the present state, on
the other hand, the M C / M T model assumes that the
m o v e m e n t state has some regular-patterns and the prob-
ability distributions for its future development depend
very much on how and f r o m where the process arrived in
O TransttionalStates 0 StationaryStates O Boundarystates
that state. By exploiting the difference in these assump-
O BoundaryandStationarystates tions and combining the advantages (and minimizing the
disadvantages) of each model we hope to provide a bet-
ter model of the m o v e m e n t of mobile users.
t 2 2 2 2 22 1
The M C / M T model m a y be more efficient for regular
m o v e m e n t or m o v e m e n t with a regular-pattern, for

~1 Actually, no one moves randomly all day and the majority of

mobile users have some regular daily movement-patterns and fol-
low these patterns more or less every day of the week.
#~ Note that this is not strictly true as there may be higher order pat-
terns which we do not look for and thus the remainder processmany
not be completelyrandom. However, for the purposes of this paper
Fig. 1. An example of a movement-track model. we will assumethat these higher order patterns are not significant.
116 G. Liu, G. Maguire Jr. /Mobitemotionprediction algorithms

example, the daily "home-office-home" movement pat- 4.1. Regularity detection algorithms
tern. For the regular movement, M C / M T model may be
very efficient in that no complicated calculation is N o matter how mobile users move, their movement
needed and thus, the speed of the prediction is very fast. history patterns can be basically subdivided into the ran-
But, the disadvantage is that if the length of MTs and dom part and the regularity part. The random informa-
MCs are short, the prediction accuracy may be low. tion can be used for statistical and probabilistic analysis
The M a r k o v model may be more efficient for random using stochastic processes and M a r k o v models for
movement. It can provide detailed calculation of the motion prediction (this will be discussed in next section).
probability distributions for the next possible states with The regularity part can be used to predict the future reg-
a given confidence level. But, it needs processing power ular movement of the users.
to do the complicated calculations and thus, an imple- The Regularity Detection (RD) algorithm detects
mentation may be slow (too slow to be useful when and forms the regular M C and M T patterns from the
implemented on the current location servers or it may movement itineraries of a mobile user. These MC or M T
require too much processing power to make it fast patterns are then saved into an Itinerary-Pattern Base
enough to be useful). (IPB). The IPB is a database in which the itinerary pat-
Therefore, we use the M C / M T model for regular terns are stored. The random parts are also saved for
movement and the M a r k o v model for the stationary and probability analysis. The idea of using the MC and M T
boundary states. To make it little bit easy to understand patterns is to make the movement pattern easy and sim-
the relation between the M C / M T model and the ple to be detected and saved.
M a r k o v model, let us look an example in which we
assume that there are only two stationary states $1, $2, 4.2. The M T / M C detection algorithm
and the rest of the states are transitional states, as shown
in Fig. 2. Thus, it forms two MTs: MT1, MT2 and four In generating itinerary patterns based on the M T /
MCs: MC1, MC2, MC3, MC4 and MC5. M C models, the algorithm carries out a movement track
If we assume that the probability of a movement fol- and a movement circle detection method comprising the
lowing the MT1 from the state $1 is ~, that is, the prob- following steps.
ability of transform from $1 to $2, then the probability of Let Sk,t be state k at time t; let N be the maximum num-
the movement following MCI from the state $1 is 1 - ~. ber of states in a State Queue (SQ); let M be the maximum
The MC2, MC3 and MC4 are identical movement circles number of Movement Circles (MCs) and Movement
with the probability of ~, ( and v respectively. So, Tracks (MTs) in the IPB. Keep the state queue of k
the transition probability from $2 to S1 is states (1 <~k<<,N) in a F I F O order and keep a number j
"y = 1 - (~ + ( + v). The M C / M T model can be viewed of MCs 1 <~j<~M in an L R U replacement order in the
as the extension of the Markov model in the applications IPB. The M C / M T detection algorithm is as follows:
for modelling the movement of mobile users.
BEGIN
1). IF Sk,t is an SS or BS,
4. M o b i l e m o t i o n p r e d i c t i o n a l g o r i t h m s F O R i=k-L to 1,
IF Si,t is an SS or BS (for any t, k-i>L),
The Mobile Motion Prediction (MMP) algorithm then mark [Si,t-tl Si+l,t-t2 9 9 9 Sk,t] a new
consists of regularity detection algorithms and a motion MT, E N D I F ;
prediction algorithm. The motion prediction algorithm IF Si,t == Sk,t (for any t, k-i>L) and Si,t is a
uses the database of regular movement patterns and the SS or BS,
random probability information with constitutional then mark [Si,t_tl Si+l,t_t2... Sk,t] a new
constraints to predict the next movement track o f the MC, E N D I F ;
mobile user. ELSE G O T O END;
ENDIF;
2). IF any state in the new M C / M T is a BS,
transitional states
then mark the new MC " b o u n d a r y priority"
MT 1 t MC
with fl = fl + 1,

!
i

ENDIF;
3). Comparing the new M C / M T with each saved
MC/MTs,
IF # >/cq (matching), then p of saved M C is
MT2
MC4 increased by 1 and calculate the frequency
parameter (F) of the M C / M T in the IPB;
ELSE
Fig: 2. An example o f M C / M T model versus Markov model. Saving the new M C / M T i n t o the IPB with
 G. Liu, G. Maguire Jr./Mobile motion prediction algorithms 117

L R U order; indicates that the current input state does not match the
ENDIF; prediction, or when the M P A is initialized, a motion pre-
4). IF MC, Remove [Si,t-tl Si+l,t-t2...Sk,t] from the diction process is carried out to generate the next predic-
state-queue (only the states on the state-queue not tion.
the ones in the IPB) E N D I F ; When an input state does not match the correspond-
END. ing predicted state, the sequence of input states begin-
ning with the most recent stationary state or boundary
Where L (L = 2, 3, 4, ...) is the minimum length state is compared by the MP to each of the MCs and
(number of states) of an MC. C~l is confidence factor that MTs stored in the IPB. This matching process deter-
is set according to the accuracy requirements or confi- mines the best-matched stored itinerary pattern, which
dence level desired (normally, it is selected to be 0.95, becomes the output of the motion predictor, or selects
0.975, 0.995, ...). SS denotes stationary state and BS the state with the highest probability to be the next state
denotes boundary state. Each MC and M T has an L R U of the movement as the Cout of the prediction by using
parameter p which indicates the priority of the MC in the stochastic and Markov process analysis.
IPB. If one or more states in an M C / M T is a boundary Three classes of matching schemes are used for corre-
state, the M C / M T is called boundary M C / M T with a lation analysis of MCs or MTs. The first-class of match-
boundary priority parameter /~. ing (with a index p,), called state-matching, indicates the
Each movement circle consists of at least one MT. degree to which a sequence of states matches another
Each MC and M T is associated with a L R U priority sequence of states having a similar length:
parameter p, a frequency parameter F, and a boundary ms
priority parameter 3. The priority parameter indicates t~ Ns '
the priority of the M C or M T with respect to other MCs
or MTs in the IPB. The frequency parameter indicates where ms is the number of states in the sequence that
the frequency of the M C or MT. It indicates how many matched, and N~. is the total number of states in the
times the M C or M T re-occurs since it has been saved in sequence.
the IPB. For each new MT, all of the parameters are initi- The second-class of matching, called velocity or time-
alized to zero. F o r each new MC, all of the parameters matching (indexed by rl), indicates the degree of the time
are initialized to the value of its M T which has the high- of a sequence of states matches another time sequence of
est/3. states having a similar length:
This algorithm is actually very simple. First, a /(ti+l - t,h (t,+l - t, hl
" n e w " M T is detected by comparing an incoming M T to ~1=- N,-1
each of the MTs already stored in the IPB. If the new ~-~4=1 [(ti+ 1 -- ti)2 0 (ti+l -- ti)l ]
MT's first-class matching index # matches the index # of where: (ti+l --li)l is the time interval between state i
one of the stored MTs, the priority parameter p of the and state i + 1 in the first sequence; (ti+l -- ti)2 is the time
stored M T is increased by 1. Otherwise, the new M T is interval between state i and state i + 1 in the second
stored in the IPB with the initial values of p = 0, F = 0, sequence; "+-+"and "| are the modulo minus and plus
and/3 = 0. Second, if one or more states in an M T is a operators, where the moduli are 24 for hours and 60 for
boundary state, that M T is called a boundary M T and minutes.
has its boundary priority parameter/3 increased by 1. The third-class of matching, called frequency- match-
ing (indexed by or), indicates the degree that the fre-
4.3. Motion prediction algorithm quency of a first sequence o f states matches a second
sequence of states having a similar length. The frequency
The Motion Prediction Algorithm (MPA) predicts of a sequence of states is calculated by how often the
the next states of movement or movement tracks by "sequence of states" occurs during a defined time per-
using correlation analysis of current movement-itiner- iod. The third class matching-index r is given by the fol-
ary with the M T / M C s in the IPB and with the associated lowing expression:
analysis using the probability information. In general,
the output of the MP algorithm (PDout or Cout)is a future 9 =~ - l,
state or a sequence of future states.
The operation of the M P A includes means for com- where Fi and Fj are the frequencies of the first sequence
paring input states provided to the M P A with predicted and the second sequence; ff = 0 indicates exact match-
states generated by the MPA and methods for matching ing.
input states to IPs ( M T s / M C s ) stored in the IPB and for An example of the calculation of the third class match-
generating predictions. If the comparator indicates that ing index 9 is shown as follows. Referring to Fig. 3,
a prediction is right, i.e., that the current input state which illustrates an example of how a user moves
matches the predicted state, the rest of the predicted IP is through various states (indicated on the vertical axis) as
provided as the output of the MPA. If the comparator a function of time (indicated on the horizontal axis), the
118 G. Liu, G. Maguir e Jr. / M o b i l e m o t i o n p r e d i c t i o n algorithms

The PDout denotes prediction output and Co~t is the

output of stochastic (Markov) processes analysis with
constitutional constraints, r/-matching: ~7> a3; B-
matching: B > a4; c~3and c~4are confidence factors.
Looked at in a different way, the matching process is
advantageously structured like a tree, as depicted in
Fig. 4. As shown in the figure, on the right-hand side is
the regularity prediction output, which is the prediction
output according to the regularity of the movement. On
Fig. 3. An example of calculating the third class matching index q).
the left-hand side is the random and constitutional predi-
cation output (Co,t), which is the results of the constitu-
third class matching index B is determined as follows. tional and stochastic process analysis. The matching tree
The six movement circles depicted in Fig. 3 can be recur- is operated as a filter to select the best suitable predic-
ring with different frequencies F1, F2, where F1 is the fre- tion. If only one stored MC or MT has fulfilled the p-
quency of the two longer movement circles and F2 is the matching requirement with the incoming sequence
frequency of the four shorter movement circles. (P>~al), that one is provided as the prediction by the
The frequency parameter Fe of an MC or an MT is MPA. If more than one MC or MT has fulfilled the p-
given by the following expression: matching requirement, then the second-class matching
n indices are examined. If none of stored MC or MT has
Fk - k=l tk
/'/ 1 fulfilled the p-matching requirement with the incoming
sequence (p < c~1), then the stochastic Markov process
where n = p + 1, p is the priority parameter of the MC analysis with constitutional constraints is used to find
or MT in the IPB. The frequency parameter F~ of a new the state with the highest probability. If only one stored
incoming sequence of states is given by the following MC or MT has fulfilled the r/-matching requirement with
expression: the incoming sequence 07 ~>a3), that one is provided as
1 the prediction by the MPA. If more than one MC or MT
has fulfilled the r/-matching requirement, then the third-
class matching indices are examined. If only one stored
Thus, the third class matching-index B is MC or MT has an F index that matches the F of the
incoming sequence, that one is provided as the predic-
Bk= 1 " t 1..] tion by the MPA. If more than one MC or MT has an F
that matches, then the sequence having the highest p and
The matching process uses explicitly the 3 classes /3, and with highest probability in the constitutional con-
matching methods to select the MC or MT which has the straints is provided as the prediction by the MPA.
highest probability to be reused: For each new incoming Otherwise, the stochastic Markov process analysis with
Sk,t, comparing [Sk-,,t-tnS~-n+~,t-t~+l... Sk,t] with each
MC and MT in the IPB, (Sk-~,t-t, is the latest SS or BS,
n > 0). Step 1). Using p-matching: If only one p-
matched MC or MT (p ~>al), PDo,t = [Sk+l,t+tlSk+2,t+t2
9.. Sk+m,t+tm] of the matched MC or MT, (m > 0); else, if
no p-matched MC or MT, PDo~t = Co,t; else, Step 2).
Using ~ matching: For all the p-matched MCs or MTs
(jointed states), if there is only one MC or MT which is r/-
matched with the incoming sequence of states,
PDo~t = [Sk+l,t+tlSk+2,t+t2 . . . Sk+m,t+trn], (m > 0). Else, if
there is no MC or MT which is ~7-matched with the
incoming sequence of states, find one with the best #-
matched constraint state. Step 3). Using B-matching: for
all the # and ~ matched MC or MT (joined states), if only ,,, c.o,.,tj --o ~ =1 '~ J
one is B-matched with the incoming sequence of states,
PDout = [Sk+l,t+tlSk+e,t+t2 . . . Sk+m,t+tm], (m > 0). Else, if
there is no MC or MT which is B-matched with the
incoming sequence of states, find the one with the best ~-
matched and best #-matched constraint state. If more
than one B-matched MC or MT (joined states), find the
one with the highest p +/3, PDo,t = [Sk+l,t+tlSk+2,t+t2
. . . Sk+m,t+tm], /~ > O. Fig. 4. The matchingtree.
 G. Liu, G. Maguire Jr. / Mobile motion prediction algorithms 119

mobile motion prediction algorithm is the prediction

a). Basic constitutional
constraints: accuracy ratio which is defined as a ratio of the correctly
If Random, P>= 1/8 predicted states out of the total states input to the predic-

Straet~
(1-6, up and down)
tor. The prediction accuracy ratio will approach
(converge to) the probability of correct prediction by the
b). Other constitutional
predictor when the simulated number of predictions is
constraints: very large.
Wall, Another important measure of performance is the
prediction efficiency. The prediction efficiency ratio is
defined as the radio of the prediction accuracy ratio of
the regularity prediction to the regularity factor, that is,
with the random and constitutional predication being
disabled. The prediction efficiency ratio indicates how
efficiently the predictor performs regularity prediction.
Fig. 5. A n example of constitutional constraints.
5.2. Assumptions
constitutional constraints is used to find the state with
the highest probability. In carrying out the simulation, it is assumed that the
The constitutional constraint information used in time intervals between consecutive states were assumed
the matching process is based on the physical construc- to have a Poisson distribution, which was adjusted by a
tion of the communication system, which is known a daily mobility factor M f according to the following rela-
priori #3. If the algorithm is implemented in the mobile tionship:
station, for example, this physical construction informa-
A = ~o x M f , (4)
tion can be provided to the mobile through overhead
messages sent on a control channel. An example of the where )~0 is the unadjusted density of the assumed
basis for the constraint information is illustrated in Poisson distribution.
Fig. 5. For basic constitutional constraints, as shown in To approximate real-world behaviors sufficiently clo-
Fig. 5(a), a mobile user located in a given cell, say cell 0, sely, the Mf takes one of three values over the course of a
can move to any one of the six adjoining cells 1-6 in the working day, depending on the time of day: Mf = 2, for
same plane and adjoining cells that might be above and times between 2000 h and 0600 h; My = 4, for times
below that plane; thus, in the absence of any informa- between 0830 h and 1600 h; and My = 8, for the rest of
tion, each of those eight adjoining cells may be the next the day. This behavior of the mobility factor is believed
state with a probability of 1/8 if the user moves uni- to approximate real-world behaviors sufficiently closely
formly randomly in all directions, that is, we assume the for an initial simulation. The total state/cell covered
directions of the movement of the mobile users are uni- area of the simulation is assumed as in Fig. 6. Fig. 6(A)
formly distributed in [0, 2~r). Fig. 5(b) illustrates an shows the total simulated cell area and Fig. 6(B-E) show
example of a physical configuration in which a door is at the area with different levels of constraints.
one end of a corridor (street or highway) that is also
joined by another corridor. The communication system 5.3. Prediction efficiency
may know apriori that a mobile user in one of the corri-
dors (streets or highways) cannot pass through a wall of To determine the prediction efficiency ratio, we do
the corridor and enter one of the other cells. This infor-
mation can be used to identify the constraint states.

5. P e r f o r m a n c e e v a l u a t i o n

In this section, we present the evaluation results of

the M M P algorithms.

5.1. Performance measures

The most important measure of the performance of

Fig. 6. (A) The total simulated celi area. (B) The area with simple con-
#3 Such constraints could also be derived from m o v e m e n t data and straints (about 34%). (C) The area with 66% constraints. (D) The area
suggested to the network m a n a g e r as possible constitutional con- with constraints level 85%. (E) The area with constraints level 92%.
120 G. L i u , G. M a g u i r e Jr. / M o b i l e m o t i o n p r e d i c t i o n a l g o r i t h m s

not use any constitutional constraints (Co,t --= 0). That example, if an user has 70% randomness in his/her
is, users may be assumed to have their terminals switch- movement, the algorithm can still get about 50% right in
ing-off from a cell and switching-on again in any of the predicting his/her movement. For the class of users
cells in the simulated area, as shown in Fig. 6(A). which has 30% or less randomness in their movement,
The normalized results of a simulation of the M M P the prediction accuracy ratio is more than 75%.
algorithm is shown in Fig. 7, in which the number of
states was 100, the length of the SQ was 500, the size of
the IPB was 500, ,k0 = 0.05, and five weeks was simu- 6. C o n c l u d i n g r e m a r k s
lated.
Fig. 7 shows the relationship between a prediction We have proposed a predictive mobility and routing
ratio and a randomness factor. The randomness factor management scheme using the movement circle and
refers to the proportion of a movement that is due to movement track models as well as the M a r k o v chain
pure uniformly random. For example, the randomness model. A class of mobile motion prediction algorithms
factor of an user equals to 0.5 means that 50% of the which consist of regularity-pattern detection algorithms
user's next movement is purely random, or in the other and a motion prediction algorithm has also been pro-
words, the probability of moving randomly in the next posed for supporting services pre-connection, routing
movement is 0.5. The optimum line is the expected best pre-arrangement and resources pre-allocation. The
(theoretical) result, i.e., if there is a regularity factor X algorithms are very simple and fast in practice. With the
(X = 1 - randomness factor) in movement, then the support of a mobility agent [11,14,16], user's data or net-
prediction ratio is X. The dashed line shows the work service can be pre- connected and assigned at the
mean-squared simulation results. The simulated predic- new location to which the user is moving. The user can
tion efficiency, which is the ratio of the prediction ratio get his/her network service or data accessed with vir-
to the regularity factor, was about 95%. Adding consti- tually the same efficiency as at the previous location.
tutional constraints should only increase the accuracy of Preliminary simulations of the M M P algorithm show
prediction. that the prediction efficiency is about 95%.
This predictive scheme deploys the regular-pattern
5.4. Prediction accuracy ratio detection algorithms to decompose the complicated
daily movement into two parts: regular pattern part and
Fig. 7 also shows the prediction accuracy ratio versus random movement part. The M C / M T model is pro-
randomness factor with basic cellular constraints, that posed and used to predict the regular movement while
is, the mobile terminal's probability of moving from one the random parts of the movement is modelled and pre-
cell to any of its adjacent cells has a uniform distribution dicted by the Markov chain model. The simplicity of the
in a cellular area shown in Fig. 6(A). The line " e " in M C / M T model makes the prediction algorithms very
Fig. 7 shows the prediction accuracy ratio with only cel- efficient. It avoids making complex calculations which
lular pattern constraint, i.e., 1/6, and the line " d " in are unavoidable in stochastic models. Thus, the proces-
Fig. 7 shows the prediction accuracy ratio with about sing power required to process the prediction algorithms
34% constraint patterns, as in Fig. 6(B). As can be seen, is greatly reduced. Furthermore, because the output of
prediction accuracy ratio is over 50% in the interval [0, the prediction using the M C / M T models are movement
0.7] of the randomness factor. This means that, for tracks or circles which consists of a sequence of states,
once it is on the right track, no additional processing is
needed while it continues along this M C or MT.
Because the M C / M T model assumes that the move-
1" , j , i i /
: ~ i ! | ment is regular and the Markov model assumes that the
0.9 i .................................................. .......... ::. . . . . . i........
state is a random variable, the M C / M T model may be

~o7
o.a

0.4 ....... i .... ! ~

: 9...... . . . . . . . . . .

i ....... ......
i. . . . . . . .

~ "'-L'"..
i. . . . . . . . . ::. . . . . . . . . i. . . . . .

::

::
i

:" - .
4 more efficient for regular movement while the Markov
model may be more efficient for random movement. An
interesting issue for future work is to make a comparison
of the efficiency of the M C / M T model with the Markov
model and to find out the "cross-over" point of the ran-
~-o.3 ........ i........ i........ i......... i.......... ..........i......... "-~"~'a i..... >" domness factor.
0.2
a -
........
92:*/~
::. . . . . . . . . .
d * 34 %
::. . . . . . . . . .
:
::. . . . . . . . . .
:
:: . . . . . . . . . .........
:
i. . . . . . . . . .
:
::. . . . . . . . .
":-. ~ .:~
.. : , . . ~ . . ! . >~,~ Another interesting issue remained to be considered
o.1 ......... ~......... ~......... ~ ......... i-':':':'-~mu~ t ~ :. . . . . . . . . . . i. . . . . . . . ~. . . . . . . . . i. ~.'.- ,-
in more details is to combine the P M M functions with
0 c - 66::% l--NO ! ! :: :: :: i " location-dependent caching and prefetching techniques
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Randomness Factor [6], that is, to integrate mobility management into cach-
ing and prefetching policies. The work includes analyz-
Fig. 7. Prediction accuracy ratio versus randomness factor with differ-
ing and simulating of the performance of the location-
ent constraint levels 0%-92% as in Fig. 6(A-F). dependent caching and prefetching and the influence of
 G. Liu, G. Maguir e Jr. / Mobile motion prediction algorithms 121

the accuracy of the PMM functions to the performance [12] K.S. Trivedi, Probability & Statistics with Reliability, Queuing,
and Computer Science Applications, (Prentice-Hall, Englewood
of the caching and prefetching.
Cliffs, N J).
[13] A. Papoulius, Probability, Random Variables, and Stochastic
Processes, Third Ed. (McGraw-Hill, 1991).
Acknowledgement [14] G. Liu and G.Q. Maguire Jr., Efficient mobility management
support for wireiess data services, Proc. 45th IEEE Vehicular
We would like to thank Professor Bj6rn Pehrson, Dr. Technology Conference (VTC'95), Chicago, Illinois (July 26-28,
1995).
Alexander Marlevi, Expert Anders Danne and Professor [15] G. Liu, Description of MMP algorithms, Ericsson Report, T/B
Jens Zander for helpful discussions, comments and their 94:229 (May, 1994).
support of this work. We would also like to thank the [16] G. Liu and G.Q. Maguire Jr., A virtual distributed system
anonymous reviewers of the paper for their valuable architecture for supporting global-distributed mobile computing,
comments and suggestions on the manuscript. Technical Report, ITR 95-01 (December 1994).
[17] A. Marlevi, A. Danne and G. Liu, Method and apparatus for
detecting and predicting of mobile terminals, Ericsson Patent,
No. 027500-969 (October 1994).
References

[1] D. Duchamp, S.K. Feiner and G.Q. Maguire, Jr., Software

technology for wireless mobile computing, IEEE Network George Y. Liu received the B.Sc. and M.Sc.
Magazine (November 1991). degrees, both in electrical engineering, from
[2] R. Hager, A Ktemets, G.Q. Maguire Jr., M.T. Smith and NIT in 1982 and 1986, respectively, MBA from
Frank Reichert, MINT - A mobile internet router, Proc. IEEE CEMP, Belgium, in 1989, and the Ph.D. degree
Vehicular Technology Conference "93, Institute of Electrical and in telecommunication systems from the Royal
Electronics Engineers (May 18-20, 1993). Institute of Technology, Stockholm, in 1996.
[3] J. Ioannidis and G.Q. Maguire Jr., The design and He joined Ericsson in 1990 and was engaged in
implementation of a mobile internetworking architecture, the design and development of GSM mobile
USENIX Winter 1993 Technical Conference (USENIX communication systems. Since 1993, he has
Association, January, 1993) pp. 491-502. been with the System Research Department of
[4] B.R. Badrinath, A. Arup and T.Imielinski, Structuring Ericsson's Core Unit Radio Systems and Technology, where he is
distributed algorithms for mobile hosts, Technical Report, engaged in research on wireless multimedia communication in the
Department of Computer Science, Rutgers University (June, Walkstation and On-The-Move projects. He holds several patents in
1993). the fields of mobile communications. George Liu's current research
[5] D. Hong and S.S. Rappaport, Traffic model and performance interests include mobility management, wireless data networks, wire-
analysis for cellular mobile radio telephone systems with less multimedia communication, mobile computing, mobile-api,
prioritized and non-prioritized handoff procedures, IEEE Trans. mobile-ip and mobile intelligent agent. He is a member of the ACM
Veh. Tech. VT-35 (3) (1986) 77-91. and IEEE.
[6] G. Liu, Exploitation of location-dependent caching and E-mail: george.liu@era-t.ericsson.se
prefetching techniques for supporting mobile computing and
communications, 6th Int. Conf. on Wireless Communications,
Calgary, Canada (July 11-13, 1994).
[7] S. Hoffpauir and J. Parker, Impact of terminal mobility on the Gerald Q. Maguire Jr. received the M.S. and
intelligent network, 6th Int. Conf. on Wireless Communications, Ph.D. degrees in computer science from the
Calgary, Canada (July 1i-13, 1994). University of Utah in 1981 and 1983, respec-
[8] R. Nevoux and S. Tabbane, Mobility features for personal and tively, and received his B.A., magna cum iaude,
terminal UMTS communications, 6th Int. Conf. on Wireless in physics from Indiana University of
Communications, Calgary, Canada (July 11-13, 1994). Pennsylvania, in 1975. He is currently profes-
[9] C. Perkins (ed), Mobility support in IP, version 14, IETF, sor of Computer Communication Systems at
Internet Draft (Jan. 1996). the Department of Teleinformatics, Royal
[10] R. Thomas, H. Giibert and G. Mazziotto, Influence of the institute of Technology, Stockholm, Sweden.
movement of mobile station on the performance of the radio Prof. Maguire's current research interests
cellular network, Proc. o f 3rd Nordic Seminar, Copenhagen (Sept. include mobile computing and communication systems, mobile inter-
1988). networking, and picture archiving and communication systems. He is a
[11] G. Liu, Mobility agent techniques for supporting wireless data member of the ACM and a senior member of the IEEE.
services Ericsson Tech. Report T/B 94:542. E-mail: maguire@it.kth.se