Environment-Aware Location Estimation in Cellular Networks

Hindawi Publishing Corporation EURASIP Journal on Advances in Signal Processing Volume 2008, Article ID 276456, 9 pages doi:10.1155/2008/276456

Research Article Environment-Aware Location Estimation in Cellular Networks ¨ ¨ Gurkan ¨ ¨ and Tuna Tugcu Onur Turkyılmaz, Fatih Alagoz, Gur, Satellite Networks Research Laboratory (SATLAB), Department of Computer Engineering, Bo˘gazic¸i University, 34342 Bebek, Istanbul, Turkey Correspondence should be addressed to G¨urkan G¨ur, [email protected] Received 23 March 2007; Revised 26 August 2007; Accepted 19 December 2007 Recommended by Rong Zheng We propose a novel mobile positioning algorithm for cellular networks based on the estimation of the radio propagation environment. Since radio propagation characteristics vary in different environments, knowing the environment of the mobile user is essential for accurate Received Signal Strength- (RSS-) based location estimation. The key feature of our method is its capability to estimate the environment of the mobile user using machine learning techniques and to utilize this information for enhancing RSSbased distance calculations. The proposed algorithm, namely, EARBALE, has been evaluated using field measurements collected from a GSM network in diverse geographic locations. Our approach turns out to be significantly beneficial, enhancing estimation accuracy, and thereby enabling high-performance mobile positioning in a practical and cost-effective manner. Additionally, it is computationally light-weight and can be integrated onto any RSS-based algorithm as an enhancement add-on. Copyright © 2008 Onur T¨urkyılmaz et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1.

INTRODUCTION

Location-based services (LBSs) enable personalized services to the mobile subscribers based on their current position.They provide new opportunities for cellular operators as well as application and content providers for the provision of innovative value-added services and creation of new revenue sources. Consequently, mobile positioning in wireless systems has received significant attention in both research and industry over the past few years since it plays a key role in providing LBSs such as wireless emergency services, location tracking services, and location-aware information and advertisement services [1, 2]. Mobile positioning involves a variety of technologies, which are divided into two major categories: network-based and handset-based location estimations. Handset-based positioning methods require a modified handset to calculate its own position, for instance by using a global positioning system (GPS) receiver embedded in the handset [3]. The drawbacks of using handset-based methods are the cost of deploying new handsets, delay for the adoption of LBSs due to slow spread of these new handsets, and the cost of developing a suitable low-power and economical integrated technology for wireless communication systems. On the contrary, using

network-based methods for mobile positioning in wireless communication systems has its advantages. When compared to the handset-based methods, the network-based methods are relatively less complex. However, although they can be used in many situations where GPS-based methods cannot be applied (such as indoor positioning), the mobile positioning in network-based methods is generally less accurate than that in handset-based methods. Thus the improvement of the accuracy for mobile positioning becomes an important issue especially for network-based solutions. So far, a wide variety of localization techniques have been proposed using measurements taken within cellular networks such as received signal strength (RSS), time of arrival (TOA), angle of arrival (AOA), enhanced observed time difference (E-OTD), and GPS method either as stand-alone GPS or assisted GPS (A-GPS) [4–9]. The TOA method is able to locate legacy handsets but requires installation of new network elements called location measurement units (LMUs) at each base transceiver station (BTS). Similarly, the E-OTD method requires installation of an LMU for every 2–5 BTSs and software modification in the handsets [7]. The AOA method requires antenna arrays at BTSs or handsets [10, 11]. In most systems, these arrays are usually arranged at BTSs because, for reasons related to

2

EURASIP Journal on Advances in Signal Processing Table 1: Hata model for three main environments.

Urban area Suburban area Open area

PLurban = 69.55 + 26.16 log( f ) − 13.82 log(ht ) − a(hm ) + [44.9 − 6.55 log(ht )] log(d). a(hm ) = 3.2 log2 (11.75 × hm ) − 4.97. PLsuburban = PLurban − 2log2 ( f /28) − 5.4. PLopen area = PLurban − 4.78log2 ( f ) + 18.33 log( f ) − 40.94.

economy and complexity, a terminal-based solution is not practical. The A-GPS method requires installation of reference GPS receivers beside the integration of GPS receivers into handsets. An RSS-based method basically estimates handset’s coordinates by fusing cell identity (CI), timing advance (TA), and Rx level information. TA value corresponds to the time it takes for a signal to be transmitted from the mobile terminal to the serving BTS and it is used for synchronization purposes in GSM. Rx levels are measurements of the strength (i.e., power) of signals received by the mobile station (MS) from the serving BTS and from up to the six strongest neighbor BTSs in GSM. The power level of a signal received by an MS, or more precisely, the attenuation the signal has experienced, depends on the reciprocal position of the MS and the BTS from which the signal was transmitted. Attenuation values from multiple neighbor BTSs are modeled in the RSS-based method through signal attenuation models such as Okumura-Hata, Longley-Rice, and Walfisch-Bertoni, and used to estimate the location of the MS. Although location estimation algorithms based on signal attenuation may not be the most promising approach for providing LBSs, signal strength is the only common information available among various kinds of mobile networks. Together with the fact that the geographic conditions and the cell layout in metropolitan areas are not the same, new approaches for location estimation algorithms based on signal attenuation have to be investigated [12]. In this paper, we introduce a machine learning-based environment-aware location estimation method, namely, environment-aware RSS-based location estimation (EARBALE), and evaluate its performance. We utilize signal measurements in a 900 MHz GSM network and data compiled from Istanbul, the largest city in Turkey with more than 10 million residents. EARBALE method uses preprocessing and dimensionality reduction via decision tree (DT) on these bulk empirical data and applies artificial neural network(ANN-) [13] based classification for the adoption of the most appropriate wireless channel model. By this way, our trained ANN is capable of identifying the environment of an input measurement as either urban, suburban, or rural. An urban area is an area with an increased density of human-created structures in comparison to the areas surrounding it. A suburban area is an inhabited district located either inside a town or city’s limits or just outside of it. A rural area is a sparsely settled place away from the influence of large cities. According to the identification of the most probable environment, the localization algorithm uses the corresponding Hata propagation model [14] in the triangulation phase. We evaluate and compare EARBALE to the naive RSS-based localization method that uses triangulation with Hata’s generic

urban propagation model. By this way, we investigate the importance of environment estimation for any RSS-based localization algorithm. The remaining part of the paper is organized as follows. In Section 2, we present a brief summary for wireless channel characteristics and Hata’s channel models. In Section 3, we describe our environment-aware algorithm together with its general architecture, data preparation phase, dimensionality reduction phase, training phase, and triangulation phase. Performance evaluation of the proposed algorithm is presented in Section 4 with illustrations and comments on them. Finally, in Section 5, we present our future research directions and draw conclusions. 2.

HATA’S WIRELESS CHANNEL MODELS

Due to physical phenomena such as multipath propagation, scattering, and diffraction, wireless channel modeling is a complex and difficult topic. In the proposed EARBALE scheme, we utilize deterministic Hata models and use them as the basis of the triangulation calculations as discussed in Section 3.3. Hata model is the most popular channel model and it is best suited for large cell coverages (distances up to 100 km) and it can extrapolate predictions up to the 2 GHz band [14]. This model has been proven to be accurate and is used by many computer simulation tools. There are three adoptions of the model to three main environments, which are urban, suburban, and rural areas. Our DT-assisted ANN architecture selects one of these models in accordance with the given input data as explained in the proceeding sections. In Table 1, the analytical representations for the three versions of the Hata model are listed for a large city where (i) (ii) (iii) (iv) (v) (vi) 3.

PL—path loss (dB), f —operating frequency (MHz), a(hm )—correction factor for MS antenna height (dB), d—distance from transmitting BTS (m), ht —transmitting BTS antenna height (m), hm —MS antenna height (m).

THE ENVIRONMENT-AWARE RSS-BASED LOCATION ESTIMATION (EARBALE) METHOD

Since signal propagation characteristics are not the same, or even not similar for different kinds of environments, it is improper to use a single generic signal attenuation model for all cases in the localization process. Even if we are dealing with a metropolitan city, it will very likely have areas that possess suburban or even rural signal propagation characteristics.

Onur T¨urkyılmaz et al.

3

Data collection & preparation

Classification DT

Measurement data from various sites

Estimation

Evaluation

ANN

Data preprocessing (elimination of erroneous and deficient data)

Cell type {urban, suburban, rural}

Decision tree parameters - Minimum split criterion (θ) ANN parameters

- Number of inputs: 4 (reduced from 8 in the DT phase) - Number of neurons in the midlayer: 4 - Training function: Levenberg–Marquardt - Adoption/learning function: gradient descent with momentum weight bias - Layer 1 transfer function: log sigmoid - Layer 2 transfer function: pure linear

Selection of the parameters for the Hata channel model (HCM)

Calculation of the distances via generic & optimal HCM and application of the triangulation method for localization

Performance evaluation of the proposed location estimation system using measurements

Figure 1: EARBALE scheme.

Therefore, if we can estimate the environment type in which the MS in consideration resides, that will provide us with valuable information to be used in the localization process. Figure 1 presents the general architecture of EARBALE, which is explained phase by phase in the following sections. 3.1. Data collection and preparation phase For the location estimation system, raw measurement data have been gathered using an MS and a GPS device connected to a laptop running Ericsson TEMS Investigator software [15] as depicted in Figure 2. The data were collected in various locations of Istanbul with different environmental characteristics. Istanbul has a very complex and formidable geography for location estimation. It has diverse terrain elements such as hills and valleys with dense settlement, the Bosporus Strait that separates two major parts of the city, the Princess Islands in the south, and suburban and rural regions mainly located in the north. These collected data were preprocessed to eliminate the erroneous and deficient data such as missing RSS levels or CIs. In this process, about 80 percent of the data were eliminated. The important issue with the collected experimental data is that the composition and measurement variables are governed by the signaling and network capabilities defined in GSM standards. Additionally, the GPS localization errors are not considered, which implies that the GPS outputs are considered as the exact positions. Data collection was done as a two-party collaborational process (see Figure 3). On the research group side, a party made the drive tests and recorded the coordinates (via GPS) together with the time stamps. The GSM network party recorded the other parameters (TA, Rx values, CIs, etc.) via

SMLC (Serving Mobile Location Center) server sent by our MS together with, again, time stamps. Afterwards, these two sources are fused to get a normalized and proper dataset. In order to eliminate some intrinsic problems (sticking with the serving BTS being the most important one) in continuous call-based localization scheme, we have switched to the trigger-based scheme. The trigger interval is chosen to be 30 seconds in the network side. The 8 input dimensions that will help us identify the environment at the beginning are as follows: TA Rx Rxi Dj

≡ Timing advance value of the MS, ≡ RSS from the serving BTS, ≡ RSS from the ith neighbor BTS, i ∈ {1, . . . , 3}, ≡ Distance from the jth neighbor station to the serving BTS, j ∈ {1, . . . , 3}.

We use data from three out of five or six neighbors that can be provided by the GSM infrastructure. We have eliminated the weaker neighbors as they are usually outliers (NLOS stations) so that we have the first three strongest neighbors and their associated distance values. We also do not use the azimuth values (in contrast to our previous work [16]) as environmental identifiers since they do not have significant contribution as separate inputs. Taking them into consideration as relative arguments with respect to each other in a way may work but for the time being we have not evaluated that possibility. In summary, each proper measurement data is classified into the type of cell it belongs to in order to be used in the dimensionality reduction and the training phases of EARBALE. For this, we classified all of our measurements in the dataset as either urban, suburban, or rural by inspecting them on Google Earth software [17].

4

EURASIP Journal on Advances in Signal Processing

TA, Rx1, Rx2, Rx3, N1, N2, N3, etc.

GPS signals

BTSn

BSC

GSM parameters logging

SMLC server TA, Rx1, Rx2, Rx3, N1, N2, N3, etc.

BTS2

BTS1

GSM parameters logging

Figure 2: Field measurement process.

Server logging

SMLC logs

Data combining

S lo

red d

gs

Raw dataset

Data filtering Filte

Tim info e .

GSM network group

atase

TEM

Time synchronization

t

e Tim . info

Research group

Drive test

Refiltering

Dataset

Figure 3: Data-flow diagram for dataset acquisition.

3.2. ANN-based classification phase Having a clean dataset with about 400 accurate measurements after the preprocessing phase, we move on to the machine learning-based phase of EARBALE. The aim of this phase is to construct an ANN which is capable of identifying the environmental characteristics of the site where the MS resides so that the most suitable channel model for that site is used in the localization process. This section consists of the dimensionality reduction and the training subphases. 3.2.1. Dimensionality reduction subphase We have eight inputs (8-dimensions of data) for the training process that will be used for estimating the environment of a given measurement. These inputs are TA, Rx, Rx1 , Rx2 , Rx3 , d1 , d2 , and d3 . However, since these variables, especially distance values, will imply memory and computational load to the positioning system, we have looked for a way to reduce

the dimensionality to a lower degree. But, while doing that, the classification accuracy should not be left behind. In other words, we have looked for the minimum input subset that can successfully classify the dataset measurements as we use all of the eight dimensions. For that purpose, we have used the DT structure and found out the minimum subset that gives the best classification result by inspecting the relationship between the minimum split criterion value (Q) and the mean square error (MSE) (see Figure 4). Q is an integer value that defines the minimum number of observations (measurements) that impure nodes have to be split in the DT structure. If we choose a very small Q, it indicates an overfitting to the training data (memorizing) and the classification accuracy of the resulting DT will be low on the testing (validation) data. If we choose a very large Q, it indicates an underfitting to the data at hand and the classification error on the testing data will be again high. Between these extreme values, there is a Q interval (or value) that gives the minimum error on the testing data. This interval

Onur T¨urkyılmaz et al.

5

45

1

MSE

40 35

x1

1

x2

2 . . .

. . .

30 xk

25 20

2

Whk

k

50

100 150 200 250 300 350 400 450 500 Q

xn

o

O

h

.. . 0

Woh

. . .

n

Nh

Figure 4: Effect of using the minimum split criterion on MSE. Input layer

Hidden layer

Output layer

Figure 6: Three-layer ANN model. True

d2 < 1534

False TA < 4.5

d1 < 878 True d3 < 1296 True

False

True

False

Urban

Suburban

Rural

False

TA < 3.5

The number of neurons in the input layer is equal to the number of input attributes given to the system, that is, four in our case (TA, d1 , d2 , and d3 ). The one neuron in the output layer gives the output O, that is, the environment. The number of neurons in the hidden layer is four in our case as it gives the best results for our dataset. Our transfer functions for the neurons in the hidden and output layers are, respectively,

Urban

True

False

Urban

Suburban

f1 (y) =



Figure 5: Resultant DT for feature selection (for Q = 110).

1 , 1 + e− y

(1)

where y is the input for the transfer functions. The input of the hth neuron of the hidden layer is NETh =

was found to be [110, 130] for our dataset as can be seen in Figure 4. By selecting the Q value as 110, we construct our DT as shown in Figure 5. The resultant DT structure briefly tells us that by looking at TA and neighbor-serving BTS distances (d1 , d2 , and d3 ), we can classify a given measurement as either urban, suburban, or rural successfully as if we are using all of the eight input dimensions. This enables a simpler and more efficient classification module for EARBALE.

f2 (y) = y,

n 

Whk × xk ,

(2)

k=1

where n is the number of input neurons, Whk is the connection weight between the kth neuron in the input layer and the hth neuron in the hidden layer, and xk is the input of the kth neuron in the input layer. The output of the hth neuron of the hidden layer is OUTh = f1 (NETh ).

(3)

The input of the output neuron is 3.2.2. Training subphase After the dimensionality reduction phase via DT, we move on to the next step, ANN training. ANNs are adjusted, or trained, so that a particular input leads to a specific target output. The network is adjusted, based on a comparison of the output and the target, until the network output matches the target. The structure of the three-layer ANN we use in EARBALE is shown in Figure 6. Multiple-layer networks are quite powerful. A network of two layers, where the first layer is sigmoid and the second layer is linear, can be trained to approximate any function (with a finite number of discontinuities) arbitrarily well [18].

NETo =

Nh 

Woh × OUTh ,

(4)

h=1

where Nh is the number of neurons in the hidden layer and Woh is the weight of connection between the hth neuron in the hidden layer and the only one neuron in the output layer. Then the output of the ANN is calculated as O = f2 (NETo ).

(5)

We have segmented the measurement data into five subsets, picked each subset one by one as validation set, and used the remaining part for training (5-fold cross-validation). Thus

Output (0 = urban, 1 = suburban, 2 = rural)

6

EURASIP Journal on Advances in Signal Processing 2

1

BTS2 (x2 , y2 )

BTS3 (x3 , y3 ) r2

r3 0 0

5

10 15

20

25 30 35 Test data

40

45

50

MS (xm , ym )

55 60 63

r1

Real classes ANN classifications

BTS1 (x1 , y1 )

Figure 7: ANN classification performance.

we have performed five runs with different validation sets and averaged the results. Figure 7 shows the classification performance of one of our trained ANNs whose accuracy is 75.8%. As can be seen, most of the misclassifications occur in the suburban class, which is the transition class. The overall average accuracy of the five ANNs is 74.5%.

Figure 8: Circle intersection formulation using three BTSs.

Then (11) can be rewritten as Hx = b, where

3.3. Estimation (triangulation) phase



In this phase, positions (latitude and longitude values), antenna heights, antenna sector widths, and azimuth angles of serving and neighbor BTSs are selected from the GSM operator’s database in real time. The positioning circles for each estimation are determined using these data together with the signal strength values and the channel model corresponding to the estimated environment in the previous ANN-based classifier phase. With three BTSs, as shown in Figure 8, we calculate the coordinates of an MS using the least-squares method [19]. We use the distance values (r1 , r2 , r3 ) to estimate (xm , ym ) by solving the following equations: 2 2 + ym , r12 = xm

(6)

2

2

r22 = (x2 − xm ) + (y2 − ym ) ,

(7)

2

2

(8)

r32 = (x3 − xm ) + (y3 − ym ) .

H=

⎦

x3 y 3

xm 1 = 2 ym



2 2 2 1 K2 − r2 + r1 . 2 K32 − r32 + r12 (14)

(15)

However, the assumption that there is one single point where all the circles intersect where the MS resides does not usually hold in real-life conditions. In other words, we get more than one intersection point where the MS may be present. In Figure 9, triangulation method is elucidated with BTSs, positioning circles, cell sectors (azimuth angle + beam width), and corresponding intersections. After calculating all the circles, intersection points that are not in the effective area of the serving BTS are eliminated. After this elimination, the coordinates of the mobile terminal are calculated as below:

(9)

1 1 lati = (lat1 + · · · + latn ), n i=1 n

1 1 lone = loni = (lon1 + · · · + lonn ), n i=1 n n

(10)

The above two equations can be written in matrix form as x2 y 2



b=

n

r32 − r12 = x32 − 2x3 xm + y32 − 2y3 ym .

⎣



xm , ym

−1

Similarly, subtracting (6) from (8) gives



x=

x = (H T H) H T b.

late =

r22 − r12 = x22 − 2x2 xm + y22 − 2y2 ym .

⎤





x2 y 2 , x3 y 3

Then the least-squares solution is

Subtracting (6) from (7) gives

⎡

(13)



K22 K32

− r22 − r32

 + r12 , + r12

(11)

where Ki2 = xi2 + yi2 .

(12)

(16)

where (i) (ii) (iii) (iv)

late —latitude of the estimated point, lone —longitude of the estimated point, lati —latitude of the ith of n intersection points, loni —longitude of the ith of n intersection points.

In case of lack of network data due to any problem in GSM network or lack of intersection points, that is, when the minimum operating conditions cannot be satisfied, the system

Onur T¨urkyılmaz et al.

7 1 Serving BTS

0.9

BTS1

x

θ

Percentage of measurements

x x x

x

x

BTS2

α BTS3

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

4

8

12

16

20

24

28

TA

Figure 11: Subscribers’ TA distribution in the real GSM network (measured for 8016 subscribers in July, 2006) and in the dataset. (Dark line: dataset TA distribution, light line: GSM TA distribution.)

θ : Azimuth α : BTS beam width (120◦ ) : Unused intersection point x : Valid intersection point x : Estimated location

Figure 9: Location estimation via circle intersection using RSS values from serving and neighbor BTSs. TA circle x

Hata circle

exact position of the MS determined using GPS. Therefore, when any location estimation algorithm is run on the measurements, the actual position via GPS can be used to evaluate the positioning performance just by calculating the Euclidean distance ε between estimated and actual positions. In other words,

ε = (latGPS − late )2 + (lonGPS − lone )2 . θ

There are also other performance metrics such as maximum error and standard deviation of the localization error. The performance evaluation is described in detail in the following section.

Serving BTS

4. θ: Azimuth x : Estimated location

Figure 10: Singular BTS algorithm used as fallback.

performs a fall-back and uses single BTS- (i.e., serving BTS-) based algorithm (see Figure 10). In this case, location of the MS is estimated as the midpoint between the Hata and TA circles along the azimuth angle. TA value is an integer between 0 and 63 and each step represents an advance of one symbol period, which is approximately 3.69 microseconds. This means that the sensitivity of the TA attribute is about 550 meters. Hence, the radius of a TA circle rTA is found by the following simple formula: rTA = (TA + 1) × 550.

(18)

(17)

3.4. Evaluation phase The dataset consists of the required GSM parameters consisting of timestamps, CIs, TA, Rx values, together with the

EXPERIMENTAL RESULTS AND PERFORMANCE EVALUATION

The general distribution of subscribers’ TA in a location area is shown in Figure 11. These statistics were collected from active calls on a Mobile Switching Center (MSC) in Istanbul in 2006. Since this provides an idea for a realistic distribution of the cell environment property, our experimental data closely follow the same distribution, which is also elucidated in Figure 11. Our experimental data comprise about 400 measurements in various environments. We have segmented the measurement data into five subsets, picked each subset one by one as validation set, and used the remaining part for training (5-fold cross-validation). Thus we have performed five runs with different validation sets and averaged the results. The baseline method, namely, the naive method, used as a benchmark for evaluating EARBALE utilizes triangulation with Hata’s generic urban signal propagation model since it is unaware of the propagation environment types. In this method, positioning circles are determined using only Hata’s default urban signal propagation model utilizing the relevant parameter values such as Rx levels and BTS coordinates. Afterwards, the position of the MS is determined as described

8

EURASIP Journal on Advances in Signal Processing

25

(%)

20 15 10 5 0 200 Naive (%) EARBALE (%)

24 24.06

600 21.6 18.99

16.53 22.46

1000 12.8 11.76

6.67 6.69

1400

1800

4.53

4.27

2.4

5.08

4.54

1.87

1.6 1.34

2200 1.87 1.61

0.8 0.8

2600 0.27 0

0.27 0.53

3000 0.27 0.27

0.27 0

3400 0.27 0

0 0

3800 0.27 0

More

0.27

0.27

0.8

0

0

0

EARBALE Naive (a) 1500

1500

6000

1000

1000

4000

500

500

2000

0

0

Urban

Naive (m) EARBALE (m)

Suburban

Rural

845 726

1033 801

395 399

EARBALE Naive

Urban

Naive (m) EARBALE (m)

Suburban

353 319

781 557

Rural 1029 509

EARBALE Naive (b)

0 Urban Naive (m) EARBALE (m)

Suburban

2200 1895

4762 2637

Rural 4296 2033

EARBALE Naive (c)

(d)

Figure 12: Performance evaluation results for EARBALE and naive methods in each environment: (a) histogram of the localization error distribution, (b) average localization error, (c) standard deviation of localization error, and (d) maximum localization error.

in Section 3.3. It does not involve any intelligent differentiation among different signal propagation environments. In other words, the naive method and the proposed EARBALE scheme are identical except the environment-aware machine learning-based triangulation adopted in EARBALE. In Figure 12, we present the detailed performance evaluation results in terms of error distribution, average error, maximum error, and standard deviation for EARBALE and naive methods. In Figure 12(a), the error distribution for both methods is depicted. The general structure of these distributions is similar for both methods. However, the tail of the distribution, corresponding to higher errors, has been largely eliminated in EARBALE method. In total, EARBALE has reduced the average error from 642 meters to 573 meters (10.75% improvement), standard deviation from 689 meters to 481 meters (30.19% improvement), and maximum error from 4762 meters to 2637 meters (44.6% improvement) compared to naive method. Apart from these numerical improvements, there is also one more important issue to emphasize that higher errors were mostly suppressed. This is critical since many standards such as Enhanced 911 in US and E-112 in Europe enforce more stringent boundaries for localization errors in the higher ranges. When we observe the results in terms of environment types, we see that EARBALE also enhances the localization performance in all criteria. The average error for the urban environment is about the same for both naive and EARBALE methods because naive method already uses Hata’s urban

channel model for all cases since it is unaware of the propagation environment types. However, the ability of EARBALE to distinguish other environment types and mitigate the localization error is especially apparent in the suburban and rural cases. In other words, the most substantial gains are seen to be made for suburban and rural areas in Figures 12(b)–12(d). All metrics have been significantly improved in these cases. Numerically, EARBALE has reduced the average error from 845 meters to 726 meters (14.08% improvement), standard deviation from 781 meters to 557 meters (28.68% improvement), and maximum error from 4762 meters to 2637 meters (44.60% improvement) in suburban environment. Similarly, it has reduced the average error from 1033 meters to 801 meters (19.55% improvement), standard deviation from 1029 meters to 509 meters (50.53% improvement), and maximum error from 4296 meters to 2033 meters (52.67% improvement) in rural environment. This capability can be noted as the main strength of our scheme. 5.

CONCLUSION

Location awareness is imperative to many pervasive computing applications. A fundamental problem in this context is location estimation, which is the estimation of a person’s location from a stream of available data. This ability will allow the facilitation of novel location-aware services, which in turn will provide better and richer service experience to the subscribers of wireless network operators. These network-based

Onur T¨urkyılmaz et al. services, integrating a derived estimate of a mobile device’s position with other contextual information in order to provide added value to the service users, are denoted as LBSs. LBSs provide many business opportunities, most importantly the next major revenue chance for telecom operators services. However, accuracy is a crucial issue for current location estimation systems. Regulatory constraints necessitate location estimation systems with higher accuracy. This better accuracy is also necessary for enhanced and more satisfactory LBSs, which will enable faster return-of-investment and higher revenue generation. Therefore, we have enhanced the positioning accuracy of RSS-based schemes in cellular systems using machine learning techniques. In this paper, we have presented the EARBALE scheme and its performance evaluation results in detail. We have shown that the proposed method enhances the generic algorithm in various criteria, including average error, maximum error, and standard deviation. The key feature of EARBALE is that it can be integrated onto any RSS-based location estimation scheme with minimal cost and impact on the legacy infrastructure. It is relatively lightweight and necessitates only a small amount of preliminary work and measurement activity. For future work, we are planning to investigate integrated approaches with better triangulation methods for our scheme and evaluate EARBALE with larger and more diverse measurement datasets. ACKNOWLEDGMENTS This work has been supported by Oksijen/Vodafone IT Hizmetleri A.S., Turkey, and the State Planning Organization of Turkey under the Next Generation Satellite Networks Project (DPT 03K 1250 250). REFERENCES [1] T. S. Rappaport, J. H. Reed, and B. D. Woerner, “Position location using wireless communications on highways of the future,” IEEE Communications Magazine, vol. 34, no. 10, pp. 33– 41, 1996. [2] I. K. Adusei, K. Kyamakya, and F. Erbas, “Location-based services: advances and challenges,” in Proceedings of the Canadian Conference on Electrical and Computer Engineering (CCECE ’04), vol. 1, pp. 1–7, Niagara Falls, Ontario, Canada, May 2004. [3] C.-D. Wann and M.-H. Lin, “Location estimation with data fusion for wireless location systems,” in Proceedings of IEEE International Conference on Networking, Sensing and Control (ICNSC ’04), vol. 1, pp. 327–332, Taipei, Taiwan, March 2004. [4] C. Drane, M. Macnaughtan, and C. Scott, “Positioning GSM telephones,” IEEE Communications Magazine, vol. 36, no. 4, pp. 46–54, 59, 1998. [5] I. Jami, M. Ali, and R. F. Ormondroyd, “Comparison of methods of locating and tracking cellular mobiles,” in Proceedings of IEE Colloquium on Novel Methods of Location and Tracking of Cellular Mobiles and Their System Applications, pp. 1–6, London, UK, May 1999.

9 [6] Y. Zhao, “Mobile phone location determination and its impact on intelligent transportation systems,” IEEE Transactions on Intelligent Transportation Systems, vol. 1, no. 1, pp. 55–64, 2000. [7] M. A. Spirito, S. P¨oykk¨o, and O. Knuuttila, “Experimental performance of methods to estimate the location of legacy handsets in GSM,” in Proceedings of the 54th IEEE Vehicular Technology Conference (VTC ’01), vol. 4, pp. 2716–2720, Atlantic City, NJ, USA, October 2001. [8] C.-D. Wann and Y.-M. Chen, “Position tracking and velocity estimation for mobile positioning systems,” in Proceedings of the 5th International Symposium on Wireless Personal Multimedia Communications (WPMC ’02), vol. 1, pp. 310–314, Honolulu, Hawaii, USA, October 2002. [9] G. Sun, J. Chen, W. Guo, and K. J. R. Liu, “Signal processing techniques in network-aided positioning: a survey of state-ofthe-art positioning designs,” IEEE Signal Processing Magazine, vol. 22, no. 4, pp. 12–23, 2005. [10] L. Cong and W. Zhuang, “Hybrid TDOA/AOA mobile user location for wideband CDMA cellular systems,” IEEE Transactions on Wireless Communications, vol. 1, no. 3, pp. 439–447, 2002. [11] S. Venkatraman and J. Caffery Jr., “Hybrid TOA/AOA techniques for mobile location in non-line-of-sight environments,” in Proceedings of IEEE Wireless Communications and Networking Conference (WCNC ’04), vol. 1, pp. 274–278, Atlanta, Ga, USA, March 2004. [12] J. K.-Y. Ng, S. K. C. Chan, and K. K. H. Kan, “Providing location estimation within a metropolitan area based on a mobile phone network,” in Proceedings of the 13th International Workshop on Database and Expert Systems Applications (DEXA ’02), pp. 710–715, Aix-en-Provence, France, September 2002. [13] S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice Hall, Upper Saddle River, NJ, USA, 2nd edition, 1998. [14] M. Hata, “Empirical formula for propagation loss in land mobile radio services,” IEEE Transactions on Vehicular Technology, vol. 29, no. 3, pp. 317–325, 1980. [15] Ericsson TEMSTM Investigator, http://www.ericsson.com/solutions/tems/. [16] O. Bayrak, C. Temizyurek, M. Barut, O. Turkyilmaz, and G. G¨ur, “A novel mobile positioning algorithm based on environment estimation,” in Proceedings of the 4th Workshop on Positioning, Navigation and Communication (WPNC ’07), pp. 211–215, Hannover, Germany, March 2007. [17] Google EarthTM Software, http://earth.google.com/. [18] H. Demuth and M. Beale, “Neural network toolbox—for use with MATLAB,” User’s Guide, ver. 4, The MathWork, June 2004. [19] A. H. Sayed, A. Tarighat, and N. Khajehnouri, “Network-based wireless location: challenges faced in developing techniques for accurate wireless location information,” IEEE Signal Processing Magazine, vol. 22, no. 4, pp. 24–40, 2005.