Content uploaded by Seungjae Shin
Author content
All content in this area was uploaded by Seungjae Shin on Jun 11, 2015
Content may be subject to copyright.
A Smart Handover Decision Algorithm Using
Location Prediction for Hierarchical
Macro/Femto-Cell Networks
Byungjin Jeong*, Seungjae Shin*, Ingook Jang*, Nak Woon Sung†, and Hyunsoo Yoon*
*Dept. of Computer Science, Korea Advanced Institute of Science and Technology, Daejeon, Rep. of Korea
†Electronics and Telecommunications Research Institute(ETRI), Daejeon, Rep. of Korea
Email: {bjjeong, sjshin, ikjang}@nslab.kaist.ac.kr, nwsung@etri.re.kr, hyoon@nslab.kaist.ac.kr
Abstract—To reduce the number of unnecessary handover in
hierarchical macro/femto-cell networks, it is necessary to avoid
macro →femto cell handovers of temporary femtocell visitors who
stay in the femtocell for a relatively short time. In this paper, we
propose a smart handover decision algorithm exploiting future
mobility pattern prediction scheme to prevent macro →femto
cell handovers of such temporary femtocell visitors. Our simulation
result shows that the proposed algorithm effectively reduces the
number of unnecessary handovers.
Index Terms—Handover Decision Algorithm, Femtocell, Hi-
erarchical Cellular Networks, Location Prediction, Reducing
Unncessary Handovers
I. INTRODUCTION
Femtocell is a small scale cellular system which has 10 ∼
30 meters as its transmission range. Because the femtocell
provides direct data access without any relaying equipments,
the deployment and operating cost are lower than in tra-
ditional macro only system. By co-deploying many indoor
femtocells with macrocells, the overall network capacity can
be significantly enhanced because the number of concurrent
wireless connections increases. For this advantage, many ISPs
(Internet Service Providers) are hurrying to use hierarchical
networks where a large number of femtocells are deployed
within a macrocell to increase the overall network coverage
and bandwidth with relatively low costs [6].
To fully exploit the advantages offered by hierarchical
macro/femto-cell networks, it is always desirable to maximize
the use of the available femtocell connectivity. In this policy,
when a macrocell includes many femtocells, the number of
handovers largely increases because of frequent macro →
femto cell handovers. However, frequent handovers gives large
load to macrocell base station because the handover process
requires a lot of network resources. Thus a new handover
decision algorithm is required to avoid unnecessary macro →
femto handovers after which the user immediately gets out
from the femtocell. (We refer this type of user as temporary
femtocell visitor.)
Previous researches about handover decision are not proper
way to prevent unnecessary handovers caused by temporary
femtocell visitor because they mainly focus on avoid ping-
pong effects [7]. So, some other handover schemes were
proposed to resolve such temporary femtocell visitor problem
[4], [5]. But, they does not take complex movement patterns
of users into consideration. From this, our goal is to develop a
smart handover decision algorithm which effectively identifies
temporary femtocell visitors and prevents their macro →femto
cell handover under the realistic assumption where the mobile
users move along random movement patterns.
Considering that almost femtocells will be deployed at
indoor environments where people mainly move along aisle
between rooms with a relatively low velocity, we exploit
the next mobility prediction algorithm to identify temporary
femtocell visitors so that the number of unnecessary handovers
can be significantly reduced.
In order to apply location prediction algorithm more sim-
ply, we divide the femtocell area and its surroundings into
grid shaped sub-areas using positioning technology which is
being widely applied for location based data services [2], [3].
Through the movement pattern analysis based on sequential
pattern mining among these sub-areas, we can predict next
sub-area movement patterns when a mobile user approaches
the femtocell. Our idea is to keep macrocell connection rather
than conducting macro →femto handover when the mobile
user may be a temporary femtocell visitor based on next
movement pattern analysis.
We simply explain our handover decision algorithm as
followings: (1) Each mobile terminal transmits its sub-area
movement history to the server periodically. (2) The server
collects the histories and mines the mobility patterns. Then,
mobility rules are extracted. (3) When a mobile terminal comes
in the boundary of the femtocell and predefined handover
conditions are satisfied, the mobile terminal predicts its next
consecutive movements based on its current trajectory and
mobility rules which are broadcasted by the femtocell base
station. If the next consecutive movement sequences are in-
cluded within the coverage of femtocell in enough length, it
eventually performs handover process.
In our extensive simulation, proposed handover decision
algorithm reduces unnecessary handovers up to 40% compared
to the normal handover decision algorithm.
The rest of this paper is organized as follows. In section
II, we present problem definition. In section III, our proposed
scheme is explained. Performance evaluations are presented in
section IV and the paper is concluded in section V.
978-1-4244-8327-3/11/$26.00 ©2011 IEEE
Fig. 1: An example of system model
II. PROBLEM DEFINITION
In this section, we define temporary femtocell visitor using
mathematical notations and present a new handover decision
criterion to prevent unnecessary handover. Firstly, we use the
threshold time Tth to identify temporary femtocell visitor.Tth
can be set differently depending on the administration policy
of each femtocell. If handovered user stays in the femtocell
for more than Tth, we assume that it is appropriate femtocell
user (appropriate handover). Conversely, if handovered user
stays in the femtocell less than Tth, it becomes temporary
femtocell visitor (unnecessary handover). Thus, we define a
new criterion of macro →femto handover as follows:
Sf>S
th and Tc>T
th (1)
where Sfdenotes the received signal strength of femtocell,
Sth denotes the predefined threshold value, Tcis cell residence
time of user. Our scheme is to predict Tcusing future mobility
prediction scheme so that we can perform selective macro →
femto handover not to accept temporary femtocell visitor.
III. PROPOSED SCHEME
A. Preliminary: Location Prediction exploiting Mobility Pat-
tern Mining
Our system consists of user mobility analysis server, fem-
tocell base stations (or access points) and mobile terminals.
As shown in Fig. 1, we divide the cell area and its surround-
ings into grid shaped sub-areas using positioning technology.
In this environment, each mobile terminal can recognize its
precise position with an error tolerance of 5meters by using
indoor localization technique (Wi-Fi, Sensor, Audio Tuner)
[2], [3].
Each mobile terminal saves its consecutive movement histo-
ries and periodically reports them to the user mobility analysis
server. If a mobile terminal stops and stays in a sub-area for a
long time (ex: more than a minute), we assume that its history
ends there and a new movement history is started. The user
mobility analysis server collects the movement histories of
mobile users and mines the mobility patterns from the his-
tories by using generalized mobility pattern mining algorithm
proposed by Yavas et al [1]. From the mined mobility patterns,
the server extracts mobility rules which describe the movement
trends of users among sub-areas. Those rules are periodically
delivered to the femtocell base station.
Algorithm 1, 2 show the pseudo code of the mobility
pattern mining and rule generation algorithm performed by the
mobility analysis server. They are slightly modified versions
of the original ones proposed in [1].
Algorithm 1 UserMobilityPatternMining(H,SPmin,G)
H: All the history of users in the database server
SPmin: Minimum support value
G: Surrounding sub-area graph
1: S1←the candidate sub-area patterns which have a length of one
2: k=1
3: UMP =φ
4: R=φ//Initially the set is empty
5: while Sk=φdo
6: // Skis the candidate length-k sub-area patterns
7: for all s∈Skdo
8: for all history h ∈Hdo
9: if sis a subsequence of hthen
10: s.count =s.count +s.suppInc //increment the support value
11: end if
12: end for
13: end for
14: //choose the candidates which have enough support value
15: Lk={s|s∈Sk,s.count ≥SPmin}
16: UMP =UMP ∪Lk//add these large patterns to set of UMP
17: //Generate next length-(k+1) candidate patterns
18: for all P∈Lk,P=<p
1,p
2,...,p
k>do
19: V={v|vis the neighbor of pk}
20: for all vdo
21: //generate a candidate sub-area patterns
22: S=<p
1,p
2,...,p
k,v >
23: Sk+1 =Sk+1 ∪S
24: end for
25: end for
26: k=k+1
27: end while
28: return UMP
Algorithm 2 GenerationOfMobilityRules(CFmin,UMP)
CFmin: Minimum confidence percentage
UMP: User mobility patterns
1: for all C∈UMP,C=<c
1,c
2,...,c
j>, where j>1do
2: for all ifrom 1 to j-1 do
3: //derive all the possible mobility rules
4: head =<c
1,c
2,...,c
i>
5: tail =<c
i+1,c
i+2,...,c
j>
6: rule =head →tail
7: //calculate the confidence value
8: rule.confidence =(C.count/head.count)·100
9: if rule.conf idence ≥CFmin then
10: R=R∪rule
11: end if
12: end for
13: end for
14: return R
In line 10 of Algorithm 1, suppInc value is calculated as
suppInc =1
1+totDist ,if pattern Bis contained by history A
0,otherwise
(2)
where totDist value denotes the minimum total distance
between two or more sub-areas included by arbitrary history
[9]. For instance, assume that history A=<2, 3, 5, 8, 10>and
candidate pattern B=<3, 10>, then the totDist=2.
Fig. 2 is an illustrative example of Algorithm 1 and 2 with
SPmin=1.33 and CFmin=50%. In this example, the whole
Fig. 2: An example of mobility pattern mining
area is divided to 3X3grid shaped sub-areas. We wrapped
the sub-area pattern like sphere to avoid edge effects, so that
some edge areas are virtually adjacent to each other. Table I
is area movement histories used for input data.
First task is typical movement pattern mining. Using Algo-
rithm 1, the mobility analysis server repeats selecting length-k
large mobility patterns from possible mobility patterns. The
most notable thing of here is that length-k large patterns
are directly used to derive length-(k+1) candidate patterns.
Detailed process is performed as followings:
1) The user mobility analysis server computes support count
values for all length-1 candidate patterns (lines 7 ∼13).
Among them, the server selects length-1 large patterns
where its support count value is greater than or equals to
SPmin (lines 14 ∼16). Table II lists length-1 candidate
patterns and selected large patterns.
2) Length-2 candidate patterns are derived by appending
neighbor sub-area number to each length-1 large pattern
(lines 18 ∼25). Then, support count values are computed
for each length-2 candidate patterns (lines 7 ∼13). The left
parts of Table III lists the 16 possible length-2 candidate
patterns. Among them, <1,3>,<2,3>and <3,1>are
selected as length-2 large patterns because their support
count values are greater than SPmin (lines 14 ∼16).
3) Length-3 candidate patterns are derived in same manner as
described in previous step. They are presented in Table
IV. In this stage, length-3 large pattern does not exist
because every length-3 candidate patterns have support
Fig. 3: Overall protocol of our system
count values less than SPmin. This makes S4(the set of
length-4 candidate patterns) empty. Thus, the while loop
stops because of S4=φ.
4) Eventually, the set of length-k large patterns is returned as
final UMPs (Table V).
Second task is mobility rule generation that will be used
for location prediction. It is conducted by using Algorithm
2. From UMPs, the server generates mobility rules (R) and
calculates their confidence value (line 8). Among all mobility
rules, the server selects final ones where its confidence value is
greater than CFmin=50%. Table VI lists final mobility rules
mined and generated from movement histories in Table I.
In our proposed system, each femtocell base station (access
point) periodically receives the set of mobility rules overlapped
with its coverage area. These mobility rules are broadcasted
through control channels. For example, in OFDMA (Or-
thogonal Frequency Division Multiple Access) based Mobile
WiMAX system, they may be included in FCH (Frame Control
Header) or MBS (Multicast and Broadcast Service) zone.
Therefore the mobile terminal can get the mobility rules of
each femtocell, then predict its next consecutive movement by
comparing current trajectory and received mobility rules. This
prediction technique can be exploited for handover decision.
If the prediction result reveals that the terminal stays in the
femtocell less than Tth, the terminal keeps the connection to
the macrocell so that we can effectively avoid unnecessary
handover. Fig. 3 summarizes our proposed system.
B. Smart Handover Decision Algorithm
In our handover decision algorithm, when the mobile termi-
nal moves into the femtocell area, it firstly checks the Eq. (1)
as the criterion of handover during signal scanning periods.
If received signal strength of femtocell called Sfis higher
than pre-determined threshold Sth, the mobile terminal gets
the information about the sub-area map and user mobility
rules from the control channel. Here, the mobile terminal
tries to predict its next n-consecutive movements among sub-
areas by comparing its actual movement history and received
mobility rules. Handover is occurred only if predicted next
n-movements are within the femtocell area. In other words,
our decision algorithm recognizes that the user is not a
temporary femtocell visitor if his next movement sequence in
the femtocell is longer than pre-determined parameter nwhich
is derived as following:
n=v×Tth
s(3)
Fig. 4: A simple example of prediction
where vis the velocity of mobile terminal, sis the length of
edge of a square shaped sub-area. Thus, a rounded-off integer
nmeans the minimum number of next consecutive sub-areas
which must belongs to the femtocell in the future, for staying
in femtocell longer than Tth.
The movement prediction is done by applying mobility rules
to current movement history of the user. Suppose that the user
has followed a path U=<s1,s2,...,si>up to now, and
the handover criterion is satisfied at sub-area si. The mobile
terminal checks if the head part of rule is contained in U
and ended with si. If it is satisfied, the tail of the mobility
rule becomes a candidate of next possible coming path. Each
possible coming path has a probability value which is the
sum of confidence and support value of the mobility rule. For
instance, a rule <1, 3, 5, 4>→<3, 2, 7>and its confidence
is 0.7and support value is 0.5.If<3, 2, 7>is selected for
next possible coming path of the user, the probability value is
1.2(=0.5+0.7). Eventually, using this movement prediction,
the mobile terminal can obtain all possible coming paths with
their probability values.
Here, the next step is handover decision. For each candidate
coming path, if at least n-consecutive sub-areas are included
in the femtocell area, which is called in-femto paths, the
corresponding probability value is summed up to a value A,
otherwise, summed up to a value B. Finally, if Ais greater
than B, handover is conducted. In short, the system assume
that the terminal will stay in femtocell longer than Tth if the
sum of probabilities of in-femto paths are greater than that of
out-femto paths. The pseudo code of our proposed algorithm
is presented in Algorithm 3.
Now, let’s see an example of our handover decision algo-
rithm depicted in Fig. 4. Assume that a user has followed a
path U=<1, 2, 3, 8>, the information about a set of femto-
cell area F={7,8,9,12,13,14,17,18,19}and four possible
coming paths are given. If n=2,<13, 18, 23, 22>and <13,
14>are in-femto paths. Thus the mobile terminal decides to
conduct macro →femto cell handover because Ais greater
than B(80 + 72.5>60 + 55).
Of course, sometimes, the proposed algorithm can make
wrong decisions. But, considering that almost indoor users
Algorithm 3 SmartHandoverDecision(U,R,n,G,F)
U: Current trajectory of the user, U=<s
1,s
2,...,s
i>
R: Set of mobility rules from control channel
F: In-femto sub-area list from control channel
n: Minimum number of coming sub-areas
G: Surrounding sub-area graph
1: k=0
2: for all rule r:<a
1,a
2,...,a
j>→<a
j+1,a
j+2,...,a
x∈Rdo
3: if <a
1,a
2,...,a
j>is contained by U=<s
1,s
2,...,s
i>and
aj=sithen
4: //Add the rule into the coming path array
5: ComingP aths[k]=(r.confidence +
r.support, aj+1,a
j+2,...,a
x)
6: k=k+1
7: end if
8: end for
9: Index =0
10: A, B =0
11: if ComingP ath.l ength == 0 then
12: Handover =1
13: else
14: while Index < ComingP aths.leng th do
15: inF emto =0
16: if ComingP aths[Index].length < n +1 then
17: if all of next sub-area ∈Fthen
18: A=A+ComingP aths[Index]
19: else
20: B=B+ComingP aths[Index]
21: end if
22: else
23: for all i,1≤i≤ComingP aths[Index].length do
24: if ComingP aths[Index][i]∈Fand it is consecutive then
25: inF emto =inFemto +1
26: end if
27: end for
28: if inF emto ≥nthen
29: A=A+ComingP aths[Index]
30: else
31: B=B+ComingP aths[Index]
32: end if
33: end if
34: end while
35: end if
36: //Make a decision
37: if A≥Bthen
38: Handover =1
39: else
40: Handover =0
41: end if
tend to repeat same movement with high probability due to
partitions or obstacles, the probability of such wrong decision
may be low. In addition, if training data is large enough,
prediction accuracy will be improved so that the performance
of our proposed algorithm is also increased.
IV. PERFORMANCE EVA L U AT I O N
To evaluate the performance of our proposed algorithm, we
perform a simulation with six scenarios as in following table:
F emtocell
Radius
Sub-area
size
The gross
area
Default Average
length of histories
10m3m30m×30m6
5m30m×30m4
20m3m60m×60m13
5m60m×60m8
30m3m90m×90m18
5m90m×90m11
In our simulation, a number of base movement patterns are
firstly generated as random walks in each environment. From
those base patterns, we generated user actual patterns (UAPs)
(a) (b) (c) (d)
Fig. 5: Simulation Results
which are categorized to two types. First type is that follows a
base pattern with randomly added noisy pattern. Second type
is called outlier which does not follow a base pattern [1]. In
this simulation, we set the ratio of the outliers to UAPs to
30% same as in [1]. Then the 70% of UAPs are creates from
base patterns by using corruption mechanism which adds some
noisy pattern to base patterns. We apply a corruption factor
cwhich is the ratio of noise pattern length to base pattern
length.
We generated 10000 UAPs for each environment. From
them, we select the 9000 histories for training process and
remaining 1000 UAPs for simulating handover decision. And
we assume that the velocity of user is 1m/s as in [8]. Note that
the velocity is not constant value in our algorithm as in Eq.
(3). To show the performance of proposed algorithm in worst
case, we set CFmin = 50%. Our simulation parameters are
listed in following table:
cOutlier percentage SPmin CFmin User velocity
0.430% 0.1% 50% 1m/s
Performance metrics are as follows: the number of unnec-
essary handovers, the reduction rate of unnecessary handover
from traditional decision algorithm, and the error probability
of our proposed scheme. Here, the error means the case where
our algorithm may prevent macro →femto cell handovers
although the user is not a temporary femtocell visitor.
Fig. 5(a), (b) show the number of unnecessary handovers of
temporary femtocell visitors for each simulation environments.
In every case, our proposed scheme outperforms traditional
handover which only considers signal strength in decision
process. Fig. 5(c), (d) show the reduction rate and error
probability of proposed algorithm. Reduction rate is up to
around 40%. Conversely error probabilities are always less
than 6%. This result is reasonable in hierarchical networks.
V. C ONCLUSION
In this paper, we proposed a new handover decision algo-
rithm based on the observations that almost femtocells will be
deployed at indoor environments where people tend to move
along specific mobility patterns with a relatively low velocity.
Therefore we divide the cell area and its surroundings into grid
shaped sub-areas using positioning technology, and exploit the
next mobility prediction scheme to identify temporary femto-
cell visitors so that unnecessary handovers can be significantly
prevented.
The simulation results indicate that our proposed handover
decision algorithm reduces unnecessary handovers up to 40%
compared to the normal handover decision algorithm. We
expect that the reduction performance will be better in real
environment due to the limitation of movement pattern of
indoor users.
ACKNOWLEDGEMENT
This work was supported by IT R&D program of
MKE/KEIT [KI002143, Development of IMT-Advanced based
WiBro Platform Technologies].
REFERENCES
[1] G. Yavas, D. Katsaros, O. Ulusoy, and Y. Manolopoulos “A data min-
ing approach for location prediction in mobile environments”, Data &
Knowledge Engineering, vol. 54, pp. 121-146, Aug. 2005
[2] K. Chintalapudi, A. P. Iyer and V. N. Padmanabhan “Indoor Localization
without the Pain”, in proc.ACM Mobicom’10, Sept. 2010
[3] M. Lee, H. Yang, D. Han and C. Yu “Crowd Sourced Radiomap for
Room-Level Place Recognition in Urban Environment”, in proc.IEEE
PerCom Workshop on Smart Environments, SmartE 2010, April. 2010
[4] X. Yan, N. Mani and Y. A. Sekercioglu “A Traveling Distance Prediction
Based Method to Minimize Unnecessary Handovers from Cellular Net-
works to WLANs”, IEEE Communications Letters, vol. 12, no. 1, Jan.
2008
[5] S. Lee, K. Kim, K. Hong, D. Griffith, Y. Kim and N.Golmie “A Proba-
bilistic Call Admission Control Algorithm for WLAN in Heterogeneous
Wireless Environment”, IEEE Trans. on Wireless Communications,vol.
8, no. 5, pp. 1672-1676, April. 2009
[6] H. Claussen, L. T. W. Ho and L. G. Samuel, “An Overview of the Fem-
tocell Concept”, Bell Labs Technical Journal-Next-Generation Wireline
Access Networks, vol. 13, March. 2008
[7] Z. Becvar and P. Mach, “Adaptive Hysteresis Margin for Handover
in Femtocell Networks”, International Conf. on Wireless and Mobile
Communications, August. 2010
[8] W. Lee, E. Kim, J. Kim, I. Lee and C. Lee, “Movement-Aware Vertical
Handoff of WLAN and Mobile WiMAX for Seamless Ubiquitous Ac-
cess”, IEEE Transactions on Consumer Electronics, vol. 53, no. 4, Nov.
2007
[9] D. Gusfield,, “Algorithms on Strings, Trees, and Sequences”, Cambridge
University Press, 1997