Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

An Ideal Base Station Sequence for PatternRecognition Based Handoff in Cellular Networks

Malka N. Halgamuge∗, Student Member, IEEE,Hai Le Vu∗, Member, IEEE,Kotagiri Ramamohanarao†, and Moshe Zukerman∗, Senior Member, IEEE,

∗ Department of Electrical and Electronic EngineeringEmail: malka.nisha,h.vu,m.zukerman@ee.unimelb.edu.au

†Department of Computer Science and Software EngineeringEmail: rao@csse.unimelb.edu.au

ARC Special Research Centre for Ultra-Broadband Information NetworksThe University of Melbourne, VIC 3010, Australia.

Abstract— The significance of having a correct criterion forevaluating handoff methods is that telecommunication providerscan choose the right handoff algorithm in a cost effective way. Incellular or micro cellular environments in cities, users often moveon predetermined paths. As the built environment is not changedoften, this regularity can be exploited in a pattern recognitionbased handoff method. The most suitable sequence of assignedbase stations or the Best Handoff Sequence (BHS) can providethe basis for pattern recognition based handoff methods. Thispaper describes in detail a computationally simple method toestimate BHS which can also be used as a benchmark forcomparing different handoff algorithms. Further, it uses therecent work in Call Quality Signal Level (CQSL) evaluationto compare various well known handoff methods.

I. I NTRODUCTION

Transfer of an ongoing call from one cell to another as auser moves through the coverage area of a cellular system isthe mechanism of the handoff. In wireless cellular systems thehandoff process is expected to be successful, and imperceptibleto users. It is also expected that the need for handoff beinfrequent. In congested inner city type environments withsmall cell sizes, it has become a challenging task to meetthese requirements.

Handoff in current wireless cellular systems is commonlyachieved through hysteresis and threshold based methods. Allsuch methods are centralised and managed by the base stationcontroller assisted by the mobile station and the base station asin Fig. 1. Increased integration in electronic hardware makes itpossible to include many complex features in mobile stations.Therefore, it is timely and useful to develop handoff algorithmsthat can be managed or processed by mobile terminals.

In cellular or micro cellular environments in cities, usersmove on predetermined paths such as roads and sidewalks. Asthe buildings and trees are not changed every day, receivedsignal strengths at a point on such a path will not highlyfluctuate. Considering sample points located on a straight lineperpendicular to the road, it is estimated that received signalstrengths belong to the same distribution [1]. This regularityis not exploited in current handoff methods. In order to usethis regularity, the signal strengths need to be sampled alongsample points in all predetermined paths such as roads.

The most suitable base station assignment at each samplepoint should be determined considering handoff costs andQoS parameters. The most suitable sequence of assigned basestations or the best handoff sequence can provide the basis forpattern recognition based handoff methods.

For example, consider the canonical case with 10 samplepoints involving only two base stations B1 and B2. When theuser is moving from B1 to B2, the ideal or the best handoffsequence could be:{B1, B1, B1, B2, B2, B2, B2, B2, B2, B2}.The sequence entry at each sample indicates the servingbase station at that point. When the cost function for thesequences is known, the ideal or best sequence that optimisesthe cost function can be found. Generally, a handoff sequence

Fig. 1. What is handoff ? To transfer a radio link, to switch anongoing callfrom one base station to another neighboring base station asa mobile usermoves through the coverage area of a cellular system

is evaluated against the cost using the number of handoffs,which is justifiable, and the measure of quality or QoS itprovides. However, the use of simple average signal strength asa measure of quality considered in [1]–[4] may not be the rightchoice. If the received signal strength falls below acceptablelevel, the connection is likely to discontinue, causing degrada-tion of QoS. If there is a handoff sequence with a small numberof assigned base stations having very high signal strengthsand

19

Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

a large number of assigned base stations with signal strengthsjust below an acceptable level, the average signal strengthcanstill indicate a high value of QoS. If it is considered as a part ofthe cost function, it can lead to a sub-optimal sequence as thebest sequence. Unlike small canonical problems consideredinmost research studies, where the user moves from one basestation to another, the exhaustive approach to find the besthandoff sequence is not practicable in realistic scenarioswithmultiple paths and multiple base stations. Once the ideal orbest sequence is known, various pattern recognition methods(for example template matching techniques used in [1]) canbe used to find the pattern prior to each handoff. If the patternrecognition based method used is computationally simple, itcan also have a low delay. Any pattern recognition methoddeveloped should be compared with other existing handoffalgorithms for the handoff delay.

A realistic framework for evaluation and comparison ofhandoffs was presented recently in [5]. It uses a computa-tionally simple benchmark for comparison of various handoffstrategies using a new realistic signal quality measure. Thispaper illustrates in detail the algorithm to obtain the bench-mark sequence and presents a comparison on commonly usedhandoff techniques and a fuzzy rule based handoff methodusing a modified form of the framework presented in [5].

II. L ITERATURE REVIEW

Various network resources are needed for the handoff pro-cess. They include air signaling, network signaling, databaselookup and network configuration [2], [6]. Air signalling isbetween the user and the base station while network sig-nalling is between the base station and other network entitylike mobile switching center. Handoff signaling uses radiobandwidth whether it is using control channels or trafficchannels. Database accesses for registration and authenticationcontributes some handoff cost. Network reconfiguration costsare associated with providing user access to the new basestation and terminating access to the old base station. Eventhough handoff costs are modeled in the literature as a constantcost per handoff due to the difficulty in quantifying the cost,all the above mentioned factors are dependent on the systemdesign and configuration and therefore, influence the handoffcost.

Handoffs can happen between cells, within the cell (betweenchannels) etc. It should be noted that we are also not concernedabout soft handoff [7], where the old base station is releasedit’s connection after a link with the new base station isestablished, as such, handoff is mainly used with CDMA typesystems.

There are several handoff strategies proposed in literature[8]–[12]:

• TheThresholdmethod [8] initiates handoff when averagesignal strength of current base station reduces certaingiven threshold value and signal strength of neighboringbase station is greater than current base station. Properselection of threshold value is very much needed here asit reduces the quality of communication link and result

can be call dropping. This method is recommended byGSM Technical Specification GSM 08.08 [13]

• The Hysteresismethod [11] initiates a handoff only ifsignal strength of the one of neighboring base stations ishigher than certain given hysteresis margin than currentbase station. Advantage of this method is it prevent ping-pong effect, which is defined later, but still this initiatesunnecessary handoffs though current serving base stationsignal strength is sufficiently strong enough.

• The Threshold with Hysteresismethod [10] initiates ahandoff when the signal strength of the current basestation drops below a given threshold and the signalstrength of a neighboring base station is higher by a givenhysteresis margin to that of the current base station. Thismethod is often used in practice with+3dB hysteresis.

It is interesting to observe that theThresholdmethod canbe easily improved by restricting the handoffs to occur underthe given rule, only in the case where the new base station canprovide signal strength stronger than the minimum acceptablelevel. Otherwise, the handoff should not occur as it cannotlead to improvement in QoS. Consequently, theThresholdwith Hysteresismethod can also be extended to allow sucha restriction.

There have been several recent extensions proposed toimprove the hysteresis based strategies. In [14] three valueswere proposed forThreshold with Hysteresisand hysteresismethod. When the current base station is busy (more handoffand new call requests than available resources) the hysteresisis lowered to 2 dB to encourage quicker handoff. When thenew base station is busy, then the hysteresis is increased to10 dB to discourage the handoff. If both base stations havethe similar levels of activity, then a hysteresis at 6 dB is used.This approach has shown right direction, but this method needsto be generalised to include a more adaptive threshold andhysteresis method.Fuzzy Handoff Algorithm (FHA) [3] is a complex schemeand uses a set of prototypes assigned to each cell to calculatethe serving base station. The handoff method based on patternrecognition is the one proposed in [1]. It is practical for acanonical (Manhattan geometry) topology but involves largecomputation when applied to a general network.

It is possible that strong shadowing caused by large obsta-cles found in the line of sight with the serving base stationor a highly mobile user in a boundary region between twobase stations causes handoff from the serving base station tothe neighboring base station only for a short period (generallyfor less than 10 s) until it gets back to the older serving basestation. This effect, called the ping-pong effect, can add manyunnecessary handoffs. Two commonly suggested methods toreduce this effect are:

• increase the hysteresis value• introduce a high averaging length for the signal strength

measure

However, neither of these methods are practical. A highhysteresis value may delay a necessary handoff at a boundary

20

Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

between two cells and a high averaging time may also slow thedynamics of handoff processes to the extent that calls couldbelost. Therefore, finding an appropriate solution to this problemwithout causing delays for necessary handoff is a researchquestion that has the attention of many researchers [15]–[17].If the hysteresis can be varied, and the ping-pong case can beuniquely distinguished from the genuine boundary crossingcase, a solution to this problem can be found.

Although the specification GSM 08.08 [13] considers onlytheThresholdmethod, the commercial providers seems to usea +3dB hysteresis value in addition (i.e., theThreshold withHysteresismethod) to minimise the ping-pong effect.

We consider a cellular mobile network withM base stationsdesignatedB1, B2, ...., BM . Let a sample path be an arbitrarypath in which a mobile user is travelling. Sample points arepoints on the sample path for which the signal strength valuesreceived from base stations are measured. LetSij be the signalstrength at sample pointi received from base stationj. Ahandoff sequence is a sequence of base stations associated withthe sequence of sample points. For a given handoff sequence,let Bi be the element of{B1, B2, ..., BM} assigned to theith

sample point. For every sample path ofN sample points thereexistMN possible handoff sequences. The number of handoffsin a handoff sequence equals to the number of changes inthe base station sequence. For example, the handoff sequence{B1, B1, B2, B3, B3, B3} has two handoffs.

For a given handoff sequence, let us define, for convenience,Si = Si,Bi

. Let Smin be the minimum signal strength belowwhich the signal quality is unacceptable for the user. LetSmax > Smin be the signal strength beyond which themarginal benefit is considered negligible.

In [5] a new measure, Call Quality Signal Level (CQSL),has been defined where the penalty term was deducted fromthe average signal strength of sample points with signalstrength greater thanSmin. We denote these sample pointsas good sample points.

However, the above CQSL measure does not effectivelydistinguish between two sequences with the same averagesignal strength of good sample points, where one has a largenumber of good sample points with a relatively small signalstrength, and another has only few good sample points but witha large signal strength. Because of it, we slightly modify theCQSL measure in this paper by deducting the penalty beforegetting the average as follows.

New CQSL is described as follows:

CQSL(x) =1

N

∑

i∈Ng(x)

Ai(x) − CNb(x)

, (1)

where∀x ∈ Ng(x),

Ai(x) =

{

Si(x) if Si(x) ≤ Smax

Smax otherwise,

N is the number of sample points,Ng(x) = {i|Si(x) ≥Smin}, C is the cost (or the penalty) for an unacceptable

sample point,Nb(x) = (N − |Ng(x)|) is the number ofsamples with signal strength lower thanSmin.

Similar to [5], we can obtain the lower bound as

CQSL(x) ≥

∑

i∈Ng(x) Ai(x)

N−

Smin|Nb(x)| |Ng(x)|

pN2, (2)

wherep is the maximum allowed proportion of sample pointswith signal quality belowSmin, i.e., Nb(x)/N ≤ p.

We consider here the lower bound forCQSL for compari-son of different handoffs. In current practice, service providersdo not associate ‘p’ the maximum bound of the proportionof “bad” sample points withQoS requirement, however, theframework proposed herein provides such a parameter tosupport differentiated services. As every handoff incurs cost,for comparison purposes we also defineλ, the quality perhandoff given by:

λ =CQSL

γ, (3)

where

γ =

∑

l[γ(x(l))]

η. (4)

If a cost of single handoff is estimated asHcost US$, thesignal quality per dollar isλ/Hcost. We will therefore use (3)to compare different handoff methods in section IV.

A pattern recognition method can detect a pattern of servingbase stations that may appear before an eminent handoff. Thefirst step in this process is to determine the best handoffsequence for a given sample path. For example, in a path of100 sample points and 3 serving base stations, there are3100

possible sample paths and one of them is the Best HandoffSequence. Signal level quality and the number of handoffsfor such sequences should be evaluated before selecting theBest Handoff Sequence. As this requires high computationalcomplexity, a cluster based approach is proposed in [5] tofind the ideal or the best handoff sequence. In the next sectionwe describes in detail a computationally simple method toestimate Best Handoff Sequence (BHS) which can also be usedas a benchmark for comparing different handoff algorithms.

III. T HE BEST HANDOFF SEQUENCE(BHS)

Our aim in this section is to obtain the best handoff sequence(BHS) which minimize the number of handoffs, and maintainthe signal strengths≥ Smin at all times [5]. Herein, weprovide an off-line algorithm to find this sequence which willrepresent a benchmark value. A brute force method (exhaustivesearch) is impractical because of a large number of possiblesample pathsMN involved. Due to this reason we use aheuristic method based on a cluster approach specified in [5]tofind the optimalBHS by maximizingCQSL and minimizingthe number of handoffs (γ).

Consider a (N × M ) signal strength matrix,Φ, receivedfrom M = 3 base stations, withN = 8 sample points for aparticular sample path as given in Fig. 2. LetGij , referred toas a cluster, be a set of signal strengths≥ Smin = 15 frombase stationj associated with a group of consecutive sample

21

Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

Fig. 2. Best Handoff Sequence (BHS) algorithm, whenSmin = 15 dB, M = 3, base stations withN = 8, sample points for a particular sample path.

points {i, i + 1, ..., i + Lij − 1}, where1 ≤ Lij = |Gij | ≤N − i + 1.

Let Wij be the average signal strength of the cluster. LetHij be the parameter associated with a clusterGij which isdefined as the weighted valueHij = αLij + (1 − α)Wij ,where α ∈ [0, 1]. The heuristic algorithm maximises signalquality by finding maximum average signal levelWij andat the same time minimises the number of handoffsγ(x) bychoosing longest clusters. In the following we demonstratethealgorithm based on the example given in Fig. 2.

SetBHS = {} and i = 1,Step 1: Find the subsetΦi as a set of the base stations

from which its signal strength in theith row of the matrixΦis ≥ Smin = 15. So according to Fig. 2,Φi = {1, 2}. If Φi

is empty we need to proceed to Step 3.Step 2: Find all the Gij clusters which starts from each

of the element of the setΦ1. In our example, we can identifytwo clustersGi1, Gi2. The algorithm assigns the base station 2associated with largerHij value as the serving base station forthe first three sample points in the path. Therefore, we obtainBHS = {B2, B2, B2}. We proceed to Step 1 withi = 4.

Step 3: With i = 4, Step 1 produces an empty set whichneed to be addressed in this Step. BecauseΦ4 is empty, weskip two rows (4th and 5th) until finding a row in a matrixΦ which contains a signal strength> 15 (in our examplethis is the 6th row). We then proceed again from Step 1by setting i = 6 to find the optimal base station for the6th sample point and onwards. After repeating Step 1 and2 we obtain{B1, B1, B1} as an optimal handoff sequence forthe 6th, 7th and 8th sample points in our example. For theskipped sample points 4th and 5th the algorithm then assignsthe previous serving base station,bold = B2, (no handoff)or the new serving base station,bnew = B1 (handoff) suchthat the average signal strength over all the skipped samplepoints is maximized. In our examplebold = B2 is chosenbecause the average signal strength fromB2 is greater thanthe average signal strength fromB1 over the skipped sample

points 4th and 5th. If after the skipping rows in matrixΦ wecannot continue with Step 1, i.e., there is no new serving basestation, we then simply continue with the old base station overall the skipped sample points. Repeat Steps 1, 2 and 3 untilthe last row of matrixΦ. In our example shown in Fig. 2 weobtainBHS = {B2, B2, B2, B2, B2, B1, B1, B1}.

Recalling our aim of theBHS is to minimize the numberof handoffs and maintain the signal strengths≥ Smin at alltimes, we observe that in the above numerical example, thenumber of handoffsγ(x) = 1.

It should be noted that the Best Handoff SequenceBHScontains indirectly the following features not consideredbyany other handoff method:

• when the strength of a serving base station is well aboveSmin, further reduction in the number of handoffs canbe achieved by adapting aflexibleor variable hysteresis.Clustering all the signal strengths above the minimumacceptable level and then selecting the base station cor-responding to the best cluster as the serving base stationwill help to achieve a nonlinear variable hysteresis.

• unnessary handoffs are avoided in sample points whereall base stations provide signal strengths belowSmin.

IV. SIMULATION RESULTS

Figure 3 shows a realistic scenario in which both new andhandoff calls appear over 24 hours period using the data baseBALI-2 (Stanford University Mobile Activity Traces) [18].BALI-2: Bay Area Location Information (real-time) datasetrecords the mobile users’ moving and calling activities in aday. We extract about 50,000 users from BALI-2 used forour experiments to analyse impact of handoff, in real system.This database was created using the real traffic data in SanFransisco Bay area shown in Fig. 3(a). A more simple dataset is created for our simulations to illustrate the use ofBHSas a benchmark.

Here we compare the different handoff methods introducedin Section II using the quality measures of (2) and (3). We

22

Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

(a) San Francisco Bay Area

0 500 1000 15000

5

10

15

20

25

Time (min)

Num

ber

of U

sers

New usersHandoff users

(b) Ratio of the number of new users: number of handoff users,at a singlecell

Fig. 3. Analysis of real data found in BALI-2: Bay Area Location Information dataset

consider three adjacent cells (M = 3) with 100 m radius.The base stations are located in a plane with the followingcoordinates:(100, 150), (250, 75) and(250, 250) [meters]. Werandomly generateη = 1000 sample paths of mobile users,each withN = 100 where each pair of consecutive pointsare one meter apart. This is a more realistic scheme than thecanonical form used in most studies.

A log-normal propagation model [19] was assumed togenerate signal strengths in each sample point along all thesample paths, i.e.,Sij = K1−K2log(r)+F , whereK1 = 85;K2 = 35 are constants,r is the distance to the base station,and F is Gaussian distributed (N(0, σ2)) representing theshadowing effect. We setσ = 5 dB, shadowing correlationdistance equals 20 m. Average received signal strength in-dicator (RSSI)values of the cell border is set atSmin to15 dB as in [3], andSmax = 1.5Smin. All the samplepaths are straight lines that start from points in the squarearea{(100, 100), (200, 100), (200, 200), (100, 200)}. Their di-rections are randomly chosen between[0, 2π] uniformly. Theparameterα was varied in the best handoff sequence. Here weusep = 0.1.

Figure 4 confirms that the mean number of handoffs forFHA increases with the increase ofσ as reported in [3].

Figure 5(a) compares the number of handoffs of fourhandoff algorithms: Threshold, Hysteresis, Threshold withHysteresis at 3dB, andFHA, with BHS. Figure 5(b) showsthat the performance ofBHS only slightly varies (12.65 to12.7 dB) asα is varied. The best performance is observedwhen α = 1, which justifies the use of cluster length ratherthan the weighted valueHij in Step 2. It can be observedthat the variation is not that significant. Minimum number ofhandoffs needed to guaranteep ≤ 0.1 can be found in Fig5(a) by settingCQSL = 0 for each handoff method. It isclear thatFHA in the present form fails to reach this quality

0 5 10 15 20 25 300

5

10

15

20

25

30

35

Similarity threshold (dB)

Mea

n H

ando

ffs

FHA σ = 3FHA σ = 7FHA σ = 10

Fig. 4. Mean handoffs for 1000 users inFHA algorithm for differentσvalues

level with the low number of handoffs. Similar to the approachtaken by [20] and [1], optimum parameter settings can beobtained from the “knee” of the curves. Clearly the benchmarkBHS with Smin = 15 dB and α = 1 provides the mostefficient parameter setting or the highestλ value. When highnumbers of handoffs can be afforded the Threshold methodwith THO = 14 dB will be as efficient as the above twotraditional handoff methods. Our simulations indicateFHA(similarity threshold at21 dB) is less desirable in comparisonto other methods.

Our results show that there is substantial room for improve-ment in existing handoff algorithms with respect to the signallevel measures as well as the number of required handoffs.Previous studies have concentrated mostly on reducing thenumber of handoffs, neglecting quality of signal. It is thereforeimportant to develop new handoff methods that would take into

23

Proceedings of the International Conference on Information and Automation, December 15-18, 2005, Colombo, Sri Lanka.

0.5 1 1.5 2 2.5−20

−15

−10

−5

0

5

10

15

γ (Mean Handoffs)

λ (d

B)

Best Handoff Sequence (BHS)ThresholdHysteresisThreshold with 3 dB HysteresisFuzzy Handoff Algorithm (FHA)

(a) Comparison of the various handoff algorithms:λ versus mean handoffs(γ), by varying the handoff threshold from 0-30 dB

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 112.65

12.66

12.67

12.68

12.69

12.7

12.71

12.72

Weight Factor (α) of BHS

CQ

SL

(dB

)

(b) The effect of parameterα, for CQSL in the BHS: CQSL versusWeightFactor

Fig. 5. Comparison of the various handoff algorithms with 1000 users

account both signal strength quality and number of handoffsasproposed in this work. The performance of any new proposedalgorithm can be compared with our benchmark solution. Theeffect of the selection ofp on CQSL is also obvious from (2),as theCQSL measure increases with the increase ofp. Wecan callp the probability of call failure due to unavailabilityof a suitable base station ifN is sufficiently large.

V. CONCLUSION

In this paper, we have described a computationally simplebenchmark solution for comparison between existing handoffalgorithms. This ideal or best handoff sequence can be usedas a reference for pattern recognition based handoff. Thismeasure together withCQSL proposed in [5] or the slightlymodified form proposed in this work can be used by telecom-munications providers to choose the best handoff algorithmtooptimise their handoff management functions.

VI. A CKNOWLEDGMENT

This work was supported by the Australian Research Coun-cil (ARC).

REFERENCES

[1] K. D. Wong and D. C. Cox, “A pattern recognition system forhandoffalgorithms,” IEEE J. Select. Areas Commun., vol. 18, no. 7, pp. 1301–1312, July 2000.

[2] K. D. Wong, “Handoff algorithms using pattern recognition,” PhDThesis, Stanford University, Department of Electrical andElectronicEngineering, June 1998.

[3] H. Maturino-Lozoya, D. Munoz-Rodriguez, F. Jaimes-Romero, andH. Tawfik, “Handoff algorithms based on fuzzy classifiers,”IEEE Trans.Veh. Technol., vol. 49, no. 6, pp. 2286–2294, Nov 2000.

[4] K. I. Itoh, S. W. J. S. Shih, and T. Sato, “Performance of handoffalgorithm based on distance and rssi measurements,”IEEE Trans. Veh.Technol., vol. 51, no. 6, pp. 1460–1468, November 2002.

[5] M. N. Halgamuge, H. L. Vu, R. Kotagiri, and M. Zukerman, “Signalbased evaluation of handoff algorithms,”IEEE Commun. Lett., vol. 9,no. 9, pp. 790–792, Sep. 2005.

[6] B. H. Cheung and V. C. M. Leung, “Network configurations for seamlesssupport of cdma soft handoffs between cell clusters,”IEEE J. Select.Areas Commun., vol. 15, no. 7, pp. 1276–1288, September 1997.

[7] R. Prakash and V. V. Veeravalli, “Locally optimal soft handoff algo-rithms,” IEEE Trans. Veh. Technol., vol. 2, no. 3, pp. 347–356, March2003.

[8] S. Moghaddam, V. Tabataba, and A. Falahati, “New handoffinitiationalgorithm (optimum combination of hysteresis and threshold basedmethods),” inProc. IEEE Veh. Technol. Conf., vol. 4, Sept 2000, pp.1567–1574.

[9] G. P. Pollini, “Trends in handover design,”IEEE Communication Mag-azine, vol. 34, no. 3, pp. 82–96, March 1996.

[10] N. Zhang and J. M. Holtzman, “Analysis of handoff algorithms usingboth absolute and relative measurements,”IEEE Trans. Veh. Technol.,vol. 45, no. 1, pp. 174–179, Feb 1996.

[11] R. Vijayan and J. M. Holtzman, “A model for analysing handoffalgorithms,” IEEE Trans. Veh. Technol., vol. 42, no. 3, pp. 351–356,August 1993.

[12] V. V. Veeravalli and O. E. Kelly, “A locally optimal handoff algorithmsfor cellular communication,”IEEE Trans. Veh. Technol., vol. 46, no. 3,pp. 603–609, August 1997.

[13] ETSI, GSM Technical Specification, version 5.12.0 ed., GSM 08.08,France, June 2000.

[14] S. Ohzahata, S. Kimura, and Y. Ebihara, “Adaptive handoff algorithmfor efficient network resource utilization,” in17th Int. Conf. AdvancedInformation Networking and Applications, AINA 2003., March 2003, pp.743–748.

[15] Y. Lin, “Modeling techniques for large-scale pcs networks,” IEEECommun. Mag., vol. 35, no. 2, pp. 102–107, Feb. 1997.

[16] L. K. Rasmussen and I. J. Oppermann, “Ping-pong effectsin linear paral-lel interference cancellation for cdma,”IEEE Trans. Wireless Commun.,vol. 2, no. 2, pp. 357–363, March 2003.

[17] L. K. Rasmussen, “On ping-pong effects in linear interference cancella-tion for cdma,” inSpread Spectrum Techniques and Applications, vol. 2,September 2000.

[18] “Bali-2: Stanford university mobile activity traces,bali - bay area loca-tion information (real-time),” Satanford University, [online] http://www-db.stanford.edu/pleiades/BALI2.html.

[19] M. Gudmundson, “Correlation model for shadow fading inmobile radiosystems,”Electronics Lett., vol. 27, no. 23, pp. 2145–2146, November1991.

[20] A. Sampath and J. M. Holtzman, “Estimation of maximum dopplerfrequency for handoff decisions,” inProc. IEEE Veh. Technol. Conf.,May 1993, pp. 859–862.

24