Mobile Netw Appl (2009) 14:677–691

DOI 10.1007/s11036-008-0139-0

Unsupervised Learning for Solving RSS Hardware Variance

Problem in WiFi Localization
Arvin Wen Tsui & Yu-Hsiang Chuang & Hao-Hua Chu

Published online: 16 January 2009

# Springer Science + Business Media, LLC 2008

Abstract Hardware variance can significantly degrade the 1 Introduction

positional accuracy of RSS-based WiFi localization sys-
tems. Although manual adjustment can reduce positional The WiFi location system is a promising technology aimed at
error, this solution is not scalable as the number of new complying with real-world location-based applications. In
WiFi devices increases. We propose an unsupervised comparison with other location systems using GPS, cameras,
learning method to automatically solve the hardware RFID, ultrasound, etc., WiFi location systems have several
variance problem in WiFi localization. This method was advantages in deployment practicability. For examples, WiFi
designed and implemented in a working WiFi positioning works in both indoor and outdoor environments; it also
system and evaluated using different WiFi devices with leverages existing and widely-deployed Wi-Fi networks. As a
diverse RSS signal patterns. Experimental results demon- result, several companies such as Ekahau (http://www.ekahau.
strate that the proposed learning method improves posi- com/), Aeroscout (http://www.aeroscout.com/), Innerwireless
tional accuracy within 100 s of learning time. Pango (http://www.innerwireless.com/) and Skyhook wire-
less (http://www.skyhookwireless.com/) are actively devel-
Keywords localization systems . Wi-Fi network . oping location-based applications for using this technology
unsupervised learning . Wi-Fi device variance in hospitals, warehouses, factories, amusement parks and
other locations. Additionally, Cisco, a key WiFi industry
player, has developed a unique WiFi localization system
A. W. Tsui (http://www.cisco.com/go/location) based on WiFi access
Graduate Institute of Networking and Multimedia,
National Taiwan University,
points that detect signal strengths emitted from client WiFi
No.1, Roosevelt Road, Sec.4, devices such that no software installation is needed on the
Taipei 106, Taiwan WiFi devices.
e-mail: [email protected] Despite its numerous advantages in deployment practi-
A. W. Tsui : Y.-H. Chuang
cability, deployment of current WiFi location systems
Information and Communication Research Laboratories, remains problematic. A major issue is the WiFi hardware
Industrial Technology Research Institute, variance problem: the WiFi device used to train the radio
Rm. 513, Bldg. 51, 195, Sec. 4, Chung Hsing Rd., map during the calibration phase (the training device) may
Chutung, Hsinchu 310, Taiwan
differ from the WiFi devices used during the tracking phase
Y.-H. Chuang (tracking devices). Varying Received Signal Strength (RSS)
e-mail: [email protected]
can degrade the signal patterns between training and
H.-H. Chu (*) tracking devices as well as the positional accuracy of WiFi
Graduate Institute of Networking and Multimedia, location systems. The experiments described later in this
Dept. of Computer Science and Information Engineering, paper indicate that the average positional error may increase
National Taiwan University,
No.1, Roosevelt Road, Sec.4,
by more than 100%. Further, this hardware variance
Taipei 106, Taiwan problem is not limited to differences in the WiFi chipsets
e-mail: [email protected] used by training and tracking devices (e.g., Intel Centrino
678 Mobile Netw Appl (2009) 14:677–691

vs. Lucent chipsets) but also occurs when the same WiFi positional error occurs when RSS values of two WiFi
chipsets are connected to different antenna types and/or devices vary mainly in the offset term.
packaged in different encapsulation materials (e.g., Intel We analyze this hardware variance problem and propose
Centrino chipset in a Sony VAIO laptop PC vs. the same an unsupervised learning approach to automatically and
chipset in a Panasonic laptop PC). Signal patterns are accurately determine a linear transformation function that
affected by both antennas and packaging materials. This can map RSS signal patterns from any unknown tracking
hardware variance problem also applies to the Cisco device to a training device, thus eliminating the need for
location system: positional error is produced when the exhaustive, manual pair-by-pair training. Further, in this
WiFi device used in the training phase emits signals of work we demonstrate that unsupervised learning accurately
different signal strength from WiFi devices used in the and efficiently determines these transformation functions.
tracking phase. To put this work in the appropriate context, current
Of the several proposed solutions for addressing this fingerprint-based positioning systems have two phases: the
hardware variance problem, the most effective is manual off-line training phase during which RSS signals of a
adjustment, proposed by Haeberlen et al. [1]. For RSS training device are used to calibrate a radio map, and the
mapping between training and tracking devices, RSS online estimation phase during which RSS fingerprints
readings are collected from both devices at the same from a tracked device are used for positioning. As shown in
location during the training phase. Experiments conducted Fig. 1, we propose the addition of an intermediate phase,
by the authors and others [1, 2] show that RSS mapping called the online adjustment phase, during which the signal
from a tracking device to a training device exhibits a linear transformation function between the training device and a
relationship. Hence, RSS mapping can be modeled as a new tracked device is learned and determined. Adding this
linear function. The difficulty lies in manually identifying online adjustment phase reduces the positional error caused
the best linear coefficients in a linear function that maps the by the hardware variance between the training device and
RSS signal pattern of the target device to that of the training the tracked device. Note that the online adjustment phase is
device. Although manual adjustment improves the posi- a one-time effort when a new tracked device, identified by
tional accuracy of WiFi localization in hardware variance its unique MAC address, first enters the environment. In
conditions, manually performing all possible combinations subsequent visits of the same tracked device in the
of pair-by-pair training for different WiFi training and environment, the system uses its MAC address to find
tracking devices is overly laborious. Further, the ever- and reuse the previously trained transformation function.
increasing number of new WiFi chipsets, antennas, and Notable contributions of this study include the following:
encapsulation materials makes this manual adjustment several unsupervised learning methods developed in this
approach impractical for real-world deployment. study resolve the WiFi variance problem by accurately and
Ekahau [3] proposes an automated adjustment method in efficiently determining an RSS signal-pattern transformation
which signal transformation functions are learned by function. These methods were implemented in a working
analyzing signal shifts from unknown tracking devices WiFi positioning system, and performance was evaluated in
when they pass through certain easy-to-estimate locations, an actual working environment. The performance of the
such as doorways, where signals exhibit highly distinguish- proposed unsupervised learning method of RSS-based WiFi
able patterns. This method’s primary drawback is that an localization improved positional accuracy by as much as 46%.
unknown tracking device may not immediately enter these Additionally, when the tracking and training WiFi devices
easy-to-estimate locations, or may stay only for a short were identical, applying unsupervised learning to WiFi
time. Kjaergaard et al. [4] propose an improved automated localization did not reduce positional accuracy.
adjustment method that uses the ratio of RSS values from The remainder of this paper is organized as follows.
different access points (rather than the absolute RSS values) Section 2 presents the WiFi hardware variance problem by
in the location estimation to overcome the hardware first demonstrating experimental results on the varying RSS
variance problem. However, the method considers only signal patterns from different WiFi devices, and then the
the ratio term of the linear transformation function in the WiFi hardware variance problem is formulated. Section 3
RSS mapping between two different WiFi devices, but not presents the design and implementation of the proposed
the offset term in the linear function. Hence, a sizable unsupervised learning method to address this WiFi hard-

Offline Training Online Adjustment Online Estimation

(Calibration) Phase Phase (Tracking) Phase

Figure 1 The online adjustment phase in in the WiFi fingerprinting-based localization system
Mobile Netw Appl (2009) 14:677–691 679

ware variance problem while Section 4 describes the Table 1 WiFi client devices and hardware profiles
experimental setup and evaluation of the unsupervised Abbreviations WiFi client devices
learning methods in an actual working environment. In
Section 5 we discuss related work, and in Section 6 we IBM IBM notebook PC with Intel PRO/wireless
draw conclusions and suggest future studies. 2915 ABG
Compaq Compaq WL110 wireless LAN PC card
(Attacted to Fujitsu Tablet PC)
Orinoco Orinoco wireless LAN PC card (attached
2 Rationale to Fujitsu Tablet PC)
HP HP iPAQ RW6828 PDA with built-in WiFi
We first conducted experiments to identify varying RSS
signal patterns between different WiFi devices. The
experimental observations revealed how RSS signal-pattern same location, the three points (x11), (x22), and (x33) are
variations cause positional error in a WiFi localization plotted on the upper left graph of Fig. 3. Each of six plots in
system. We formulate the proposed approach based on Fig. 3 is constructed from 500 sample points.
unsupervised learning to solve the RSS signal-pattern From the RSS trace data shown in Fig. 3, we can
variation problem. observe a linear shift in RSS readings between the two
WiFi devices in all six tested pairs. For example, Fig. 3
2.1 Signal-pattern variations of WiFi client devices shows the Compaq vs. HP RSS mapping relation, which
can be approximated by a line with a slope of 0.93 at
We performed experiments to determine the variation in RSS y-intercept (or “offset”) of 1.20. This linear shift in RSS
signal patterns from different WiFi devices. The test environ- signal patterns was also observed in experiments conducted
ment was the fifth floor of an office building. Figure 2 shows by Kjaergaard et al. [4] and Haeberlen et al. [1].
the floor plan of the test environment, which was 25 m by Figure 4 plots the RSS readings of a Sony VAIO SZ18
47 m and had sixteen WiFi access points (APs). While laptop PC against a Panasonic CFT5 laptop PC at the same
walking the path indicated by the blue dotted line in Fig. 2, location. Both have the same Intel Centrino WiFi chipset
RSS readings for four different WiFi devices were recorded. but use different antennas and packing materials. The RSS
Table 1 shows the hardware profiles of the four WiFi analysis shows that their RSS signal patterns differ with an
devices. approximate linear RSS mapping function of a slope (0.92)
Figure 3 compares the RSS readings of different at an offset (−8.155 dbm).
training/tracking device pairs. Each point (RSS-x, RSS-y)
on the plots represents RSS readings from two different 2.2 Effect of linear signal-pattern shift on the accuracy
devices at the same location and from the same WiFi AP. of a WiFi positioning system
For example, if the Compaq device RSS readings are (x123)
from three WiFi APs (AP123) while the HP device readings Before describing how this linear shift in RSS signal
are (y123) from the same three WiFi APs (AP123) at the patterns affects the positional accuracy of a WiFi position-

Figure 2 Sixteen WiFi APs

(red circles), were distributed
throughout the 25 m×47 m test
environment. The movement of
the tracking device is marked
with a blue dotted line
680 Mobile Netw Appl (2009) 14:677–691

Figure 3 Correlation of several

hardware pairs

ing engine, we examine the characteristics of a general The example in Fig. 5 shows the effect of RSS signal-
RSS-based WiFi location system. Such a system consists of pattern variations on the positional accuracy of RSS-based
two phases. In Phase I, the offline training phase, we WiFi localization. The blue line (▴) indicates the RSS
perform a site survey by using a training device to measure fingerprint of a training device measured at location x. The
RSS signal patterns from different APs at fixed sample green line (▪) shows the RSS fingerprint of the same
points in the environment. These RSS readings are encoded training device measured at location y. The red line (•)
as fingerprints and recorded onto a radio map depicting the indicates the RSS fingerprint of the tracking device at
RSS of APs at different sample location points. Phase II is location x. Since the tracking device differs from the
the online estimation phase, in which the location of training device in RSS signal patterns, the red (•)
tracking devices is calculated in real time by matching fingerprint exhibits a linear shift away from the blue (▴)
sample points on the radio map with the RSS fingerprint fingerprint (Fig. 5). By computing their Euclidean distance,
closest to the tracking device. By considering each RSS the red fingerprint vector is closer to the green fingerprint
fingerprint a vector, the proximity of two RSS fingerprints vector than to the blue fingerprint vector. Thus, the
can be measured using their Euclidean distance [5, 6] or a positioning system mistakenly estimates that the tracking
probabilistic model [1, 7–9]. device is located at y rather than at x.
Mobile Netw Appl (2009) 14:677–691 681

Figure 4 RSS signal patterns

from Sony VAIO and Panasonic
laptop PCs with the same Intel
Centrino WiFi Chip

2.3 Signal-pattern transformation function closer to P (Sc | Location=l, Hardware=Hc) for all li in the
tracking space,
The proposed solution to the above problem is to apply a

Sc ¼ F So ð1Þ
transformation function on the RSS fingerprint of the tracking
device such that the transformed fingerprint (the grey dotted A simple method of determining this transformation function
line in Fig. 5) is shifted closer to the RSS fingerprint of the is to survey the site by collecting RSS samples from devices
training device. Thus, the positioning engine can estimate the Hc and Ho while manually holding them at each location.
tracking location of the tracking device. Using this training dataset as examples of RSS mappings
This hardware variation problem is formally defined as between tracking and training devices, different learning
follows. Denote the training device as Hc (c for calibration) algorithms are applied to learn a signal-pattern transformation
and the tracking device as Ho (o for observed). Since Hc and function. From a learning perspective, this manual procedure
Ho produce different RSS signal patterns, denoted as Sc and of collecting training examples is analogous to manual
So, the probability of training RSS readings at location l using labeling inputs in supervised learning. However, as mentioned
training device Hc, denoted as P (Sc | Location = l, Hardware = in Section 1, this manual data collection is impractical given a
Hc), differs from the one obtained for the tracking device, large number of WiFi device pairs. Therefore, an unsupervised
denoted as P (So | Location=l, Hardware=Ho) . The problem learning algorithm is needed.
is finding an accurate transformation function F such that
applying F to So shifts P (So | Location=l, Hardware=Ho) 2.4 Unsupervised learning
: Tracking device RSS fingerprint at location x
: Training device RSS fingerprint at location x The proposed unsupervised learning system automatically
: Training device RSS fingerprint at location y learns this signal-pattern transformation function at runtime
: Transformed tracking device RSS fingerprint.
for any unknown tracking device. The learning procedure
Received Signal Strength (RSS)

consists of the following two general steps.

- The RSS readings from an unknown tracking device
are first labeled with a rough location estimation using
a correlation ratio computed from the Pearson product-
Transformation
function moment correlation coefficient [10] defined below:
k


P



sic
m Sc sio
m So
i¼1
AP1 AP2 AP3 AP4 r ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffisffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ
P
i
k
2 P k

2
Figure 5 Example of positional estimation error caused by RSS s c
m Sc sio
m So
i
1 i¼1
signal-pattern variation in RSS-based WiFi localization
682 Mobile Netw Appl (2009) 14:677–691

The k is the number of APs (or the dimension of RSS then used as input to the positioning engine, which consults
fingerprint vectors), sc is the RSS fingerprints of the the radio map to find the best matched location point. Since
training device from the radio map, μ is the mean or the parameters to the transformation function have yet to be
expected value of a random variable, and so is the RSS determined at this time, location estimate is rough. In the
fingerprint from the tracking device. The absolute third step, from the rough location estimate, the
value of Pearson correlation ratio has a value range in corresponding RSS fingerprints stored on the radio map
(0–1) where 1 indicates the best linear dependency (or (RSSm) are identified and coupled with the device’s RSS
greatest similarity between two fingerprint vectors), fingerprints (RSSd) to form a training sample (RSSd, RSSm).
and 0 indicates complete linear independency (or least Assuming that n training samples are collected 
over the
0 ; RSS0 ;
d m
similarity). The Pearson correlation ratio is therefore training

period,

the training set consists
 of RSS
used to measure similarities in RSS fingerprints RSS1d ; RSS1m ; RSSn
1d
; RSSn
1
m
g. We propose two learn-
between tracking and training devices. ing methods to train this transformation function – batch
- After labeling the RSS data with rough location learning and iterative learning. In batch learning (repre-
estimates, four learning algorithms including linear sented by the solid path in Fig. 6), the entire training set is
regression, two versions of expectation maximization, input to a learning algorithm at the end of the training
and neural networks, are applied to train the transfor- period to determine the parameters for the transformation
mation function. These four learning algorithms are function. In iterative learning (represented by the dotted
detailed in Section 3. path of Fig. 6), individual training data is input iteratively
to a learning algorithm to determine the parameters for the
The proposed approach differs from the one proposed by
transformation function. Although the initial parameters to
Kjaergaard et al. [4], which computes an RSS ratio (i.e., the
the transformation function are likely to be inaccurate due
RSS reading of one AP divided by the RSS reading from
to the training set’s small size, they will gradually become
another AP) to reduce the effect of linear shift in RSS
more accurate through repetitive training. During the
fingerprint matching. Since the Kjaergaard approach only
training process, users may move around freely: there is
approximates the ratio term but not the offset term in a linear
no need to stand still at certain locations as in the
shift function, approximation error is increased when the
calibration process.
offset term is relatively large. In comparison, the proposed
approach uses the Pearson formula in Eq. (2), which captures
3.1 On-line regression algorithm
both the ratio and the offset terms in a linear relationship.
The linear regression in the first learning algorithm assumes
that the regression model for a tracking device’s RSS
3 Design and implementation
mappings to a training device exhibits a linear relationship,
as shown in Fig. 3:
Our system design is shown in Fig. 6. In the first step, RSS
fingerprints (RSSd) measured from an unknown tracking
RSSc ¼ b þ a  ðRSSo Þ þ ": ð3Þ
device are sent to the localization system server. In the
second step, Eq. (2) is applied to compute the Pearson RSSc and RSSo represent RSS fingerprints from the training
correlation ratio from the RSS fingerprints. This ratio is and tracking devices, (a, b) are coefficients in the

Figure 6 System design

Mobile Netw Appl (2009) 14:677–691 683

transformation function, and ε is the error term. The RSS readings and labeled with rough position estimates from
training dataset is determined by the tracking device’s the Pearson correlation ratio. The network realized using radial
RSS readings and are labeled with rough position estimates basis functions [15] has the following form:
using the Pearson correlation ratio. After collecting a
X
N
sufficient number of data points, least squares analysis is ϕ ðS Þ ¼ wi rðkS
ci kÞ; ð5Þ
applied to find the best (a, b) coefficients with the lowest i¼1
sum of error squares.
where N is the number of neurons and S is the single signal-
strength pattern value observed. The basic function
3.2 EM algorithm rðkS
ci kÞ is Gaussian:

An alternative approach to on-line regression is expectation rðkS
ci kÞ / exp½
bkS
ci k: ð6Þ

maximization (EM). Starting with the initial estimates of
(a, b) coefficients, EM is an iterative procedure to refine the The weights wi are learned using gradient descent:
(a, b) estimates by repeatedly computing expectation and
maximization. wi ðt þ 1Þ ¼ wi ðt Þ þ v½yðt Þ
ϕðS ðt ÞÞrðkS ðt Þ
ci kÞ; ð7Þ
The true location is postulated by combining the original
probabilistic location computation with the Pearson corre- where y(t) is the signal pattern with the highest correlation
lation ratio. Particle filters [11–14] are used for historical coefficient selected from the training data.
moving pattern adjustments with a probability of correla-
tion coefficient for each cell. At each iteration, the 3.4 Extended EM algorithm
estimated linear parameters (a, b) are used in the subse-
quent iteration to transform the RSS fingerprints from the Since both EM and neural networks are limited by local
tracking device. This iterative procedure continues until a optimum, we designed and implemented a fourth method to
convergence criterion is met. extend EM as follows. First, six EM models with different
The convergence criterion is that the result of the linear initial seeds are executed in parallel. From the six EM
mapping function (i.e., the ’y’ in ‘y=ax+b’), stays within models, the results with the highest probability are used as
the range of δ (=1) dbm over a period of time t (=50 s). We the output [16].
choose the result of the linear mapping function (y) over the
parameters (a, b) for faster convergence. Since the RSS
values have limited range between −30 dbm and −90 dbm, 4 Evaluation
training points fall only within a small segment of this
linear function. Therefore, it is more important that we This section describes the experimental procedure and
accurately identify this specific segment rather than the analyzes the performance results of the unsupervised
entire linear function. A more detailed discussion on the learning system in an actual WiFi localization system and
convergence criterion is discussed in Section 4.3. environment.
The output parameters at each iterative step are the linear
coefficients (a, b). The Θ = (a, b) and So ðt Þ are denoted as 4.1 Experimental setup
the tracking device’s RSS readings at time t. The
optimization problem is as follows: As Fig. 2 shows, the test environment was the same as in
(
) the previous experiment for observing signal-pattern varia-

 P Θ; So ðt Þ tions from different WiFi devices. The radio map was
argmax P ΘjSo ðt Þ ¼ argmax
 : ð4Þ
Θ Θ P So ðt Þ constructed as follows. First, the test environment was
divided into 38*69 cells with each cell occupying a
EM
is applied to solve the above problem by estimating 0.69-meter by 0.69-meter space. The position of the
P li jSo ðt Þ; Θðt Þ and maximizing Θ(t+1), where li ∈ L (L tracking devices was estimated using cell granularity. Addi-
for the set of all locations in space) is the latent variable. tionally, to minimize manual calibration, 107 of the 2,622
cells were uniformly selected as the training cells. Thirty RSS
3.3 Neural network algorithm samples were collected from each training cell. The RSS
samples from the other unselected cells were interpolated with
We implemented the neural network learning method to capture those of the training cells. Interpolation was intended to reflect
any non-linear components of the signal-pattern transformation the common practice of minimizing manual calibration in
function. Similar to online regression, the training dataset for deployments over a wide coverage area, even though
the neural network was obtained from the tracking device’s interpolation reduces positional accuracy [17].
684 Mobile Netw Appl (2009) 14:677–691

4.2 Positional accuracy explanation is that the unsupervised system adjusts dynam-
ic changes in environmental factors (for example, humidity
Table 1 shows the four WiFi devices used in the experi- levels, open/closed doors, etc.) that affect signal patterns.
ments. The Compaq device was selected as the training
device for calibrating the radio map. During the online 4.3 Training time
estimation phase, a person carried each of these four WiFi
devices while walking at a constant speed along a path In EM and neural networks, the training time is important
indicated by the blue dotted line in Fig. 2. The RSS for determining a signal-pattern transformation function for
readings along this walking path were collected for use in an unknown tracking device. During the learning phase,
the training dataset for each of the four learning algorithms positional estimates are unreliable.
(that is, online regression, EM, neural network, and The training process is complete when the changes in the
extended EM) to train the signal-pattern transformation linear parameters of the transformation function being
function. trained fall within a limited range. However, the experi-
Figure 7 plots the mean positional error when applying ments reveal that while function parameters rarely con-
each of four learning algorithms to track each of the four verge, the output of transformation functions often do. The
different WiFi devices, and also plots the mean positional reason is that several solutions to function parameters may
error for manual adjustment and non-adjustment. Manual co-exist when the input training dataset is concentrated on a
adjustment re-implements the method developed by small segment of the function where the RSS readings fall
Misikangas et al. [3], which provides near-optimal between −90 dbm and −30 dbm. To address this problem,
performance. Non-adjustment means no transformation the convergence criterion depends on the change in
function was applied for WiFi localization; therefore, its function output rather than the change in the trained input
performance is used as a baseline for measuring improve- parameters.
ments in the four learning algorithms. The convergence criterion uses the near-optimal output
From Fig. 7 we find that: (1) the EM produced the least from manual adjustment (as determined during post-
average positional error of the four tracking devices, and its processing) as a baseline to determine the speed of the
performance approached that of the near-optimal manual proposed runtime learning algorithms (EM or neural
adjustment; (2) except for the Orinoco device, which had an network) which then stabilized at the near-optimal output.
RSS signal pattern similar to that of the Compaq training Specifically, as Fig. 8 shows, the training curves were
device, the learning algorithms all yielded less positional constructed using the near-optimal output differences and
error than non-adjustment. This suggests that applying the the learning algorithm.
unsupervised learning to RSS-based WiFi localization  
output manual adjustment
output learning algorithm ; ð8Þ
effectively reduced positional error due to hardware
variance. (3) Under some training/tracking device pairs, where the tested learning algorithms were EM (dotted line)
EM outperformed the near-optimal manual adjustment. One and neural network (solid line), and the training/tracking

Figure 7 Positional error (in

meters) after applying four dif-
ferent learning algorithms
(online regression, EM, neural
network, and modified EM) to
four different tracking devices
(Orinoco, HP, IBM, and Com-
paq). The Compaq was the
training device
Mobile Netw Appl (2009) 14:677–691 685

Figure 8 Comparison between

convergence times

devices were Compaq/HP. The training times of both the amount and percentage of positional error reduction when
EM and neural network were under 100 s. Although the the Orinoco was used as the training device.
EM training time was slightly longer than that of the neural The experiments show that the IBM device produced the
network, the EM training error was smaller (1.8 dbm vs. strongest RSS readings followed by the Orinoco and
4.2 dbm). Compaq devices (which yielded RSS readings with similar
Whether or not the 100-second training time is sufficient strengths) and finally the HP device, which recorded the
depends on the specific application. For example, 100 s weakest strength. Since the difference in RSS signal
may be acceptable for a location-based museum tour guide patterns between Compaq (the tracking device) and
but may be too long for locating an emergency call. Note Orinoco (the training device) was small, the 0.31 m
that the 100-second training time is a one-time cost when a positional error reduction was also small. Conversely, the
new WiFi device enters an environment. After the system difference in RSS signal patterns was large for Orinoco-
learns the parameters to its transformation function, it stores IBM and Orinoco-HP pairs, as was the positional error
them into a database indexed by the device’s MAC address. reduction (1.66 m and 1.96 m, respectively). Table 2 shows
Subsequent visits from the same WiFi device incur a that, given a larger RSS signal-pattern difference between
database lookup and require no additional learning cost. tracking/training devices, the proposed system generally
achieves higher positional error reduction. However, one
4.4 Case study: the Orinoco as the training device exception is discussed in Section 4.7 below.

After using the Compaq computer as the training device in 4.5 Case study: similarity between tracking/training devices
the above tests, the unsupervised learning system was
tested using other training devices. Table 2 shows the In this case study we investigated whether unsupervised
learning on the same training/tracking device pairs degrades
positional accuracy given that learning and applying
Table 2 Positional error reduction when using orinico as the training transformation functions to the same device pairs is
device unnecessary. Table 3 shows the positional error for Orinoco
Tracked device Positional error Positional error and Compaq training and tracking devices, respectively,
reduction (meters) reduction (%) with and without using unsupervised learning. No degra-
dation of positional accuracy was observed. Surprisingly, a
Compaq 0.31 13.13 slight improvement was observed (i.e., the positional error
IBM 1.66 46.51
was reduced from 2.66–2.40 m for the Orinoco devices and
HP 1.96 44.41
from 2.40–2.08 m for the Compaq devices). A possible
686 Mobile Netw Appl (2009) 14:677–691

Table 3 Positional error reduction in the same training/tracking speed movement often increases positional error. Table 4
device pair with/without learning
shows positional error after applying different learning
Positional error without Positional error with methods to WiFi localization in two cases involving
learning (meters) EM learning (meters) constant-speed movement and variable-speed movement.
The results in Table 4 show that the unsupervised learning
Orinoco 2.66 2.40
method is effective for both constant-speed and variable-
Compaq 2.40 2.08
speed movement.

4.7 Relationship between RSS signal-pattern difference

and positional error
explanation is that, since training was performed 1 day
before testing, the unsupervised system also adjusted to Intuitively, since increased difference in RSS signal
changes in environmental factors (for example, humidity patterns between tracking and training devices produces
levels, open/closed doors, the presence of people, etc.) that larger difference between the tracking device’s RSS
affect WiFi signal patterns. This effect was also observed in fingerprint and the RSS fingerprint on the radio map
previous work [18–21] that applied learning algorithms to constructed by the training devices, the positional error
adjust the radio maps given sensor-detected changes in in WiFi localization should be larger also. Although this
these environmental factors. However, in the future we are relationship between the RSS signal-pattern difference
interested in validating how well unsupervised learning can and positional error seems reasonable, the experiments in
adapt to such changes in the environmental factors. this study proved otherwise.
Figure 10 plots this relationship between the positional
4.6 Case study: variable-speed vs. constant-speed error and RSS signal-pattern variance for several training/
movement tracking device pairs. The x-axis measures the percentage
of the average AP RSS readings between the training and
Since variable-speed movement is common in real-world tracking devices, computed as follows:
scenarios, in this case study we compared the effectiveness
javg ðRSSc
avg ðRSSo Þj
of unsupervised learning for both constant-speed and : ð9Þ
variable-speed movements. The variable-speed trace was avg ðRSSo Þ
collected as follows. Figure 9 shows several walking The y-axis measures the average percentage of increased
segments marked with arrowed lines. Each walking positional error due to device hardware variance, which is
segment has a unique movement speed whereas the speed computed as the difference between the positional error
within each walking segment is constant. without hardware variance (that is, using the same device
Since most WiFi localization systems adopt a motion for tracking and training) and the positional error with
model that assumes constant-speed movement, variable- hardware variance. The results in Fig. 10 do not indicate

Figure 9 Moving testing traces

Mobile Netw Appl (2009) 14:677–691 687

Table 4 Positional error under constant-speed movement and 5 Related work

variable-speed movement

EM Neural Modified Online Many location estimation systems have been developed
network EM regression using Wi-Fi RSS values to estimate locations. These
systems can be categorized into two broad approaches.
Constant-speed 2.46 3.75 3.51 3.50
The first approach is based on the deterministic method [5,
motions
Variable-speed 2.37 3.82 2.48 3.11 6]. Systems following this approach apply deterministic
motions inference, such as triangulation and k-nearest-neighbors
(KNN) search, to estimate the target device’s location. For
example, the RADAR system applies KNN to obtain the
k nearest neighbors and estimates the location of the target
device by averaging the locations of these k nearest
that greater RSS signal-pattern variance increases position- neighbors. The second approach is based on the probabi-
ing error. listic method [7–9, 11]. Seshadri et al. [11] apply Bayesian
Several factors such as AP spatial distribution can inference, which uses multiple probabilistic models and
enlarge or mitigate the effect of RSS signal-pattern variance histograms to enhance the performance of the original
on positional error. An interesting situation in which a large system. It calculates the conditional probabilities over
RSS signal-pattern variance produces negligible positional location-based RSSI. It also adds a motion model to
error is the following. This situation arises under two describe the continuity in human movement such that it
conditions: (1) the RSS readings between the tracking and can lower the oscillatory location estimations in Wi-Fi
training devices differ only in offset but not in ratio, and (2) based localization systems.
the spatial distribution of APs is balanced or evenly A notable WiFi based localization system is from Cisco:
distributed relative to the position of the tracking device. this system tracks all WiFi devices in the environment
Figure 11 shows an example of a balanced distribution of without installing any additional software on the devices.
APs, which are marked in red circles, relative to the WiFi access points are programmed to measure RSS signals
location of a tracking device, which is moving within the emitted by the tracked devices. Then, RSS measurements
blue dotted rectangular area. The distribution is balanced to from multiple access points are aggregated at a localization
the tracking device in that the APs to its right and left are server that determines the locations of all tracked devices.
more or less equal in number and distance. Appendix Similar to other WiFi fingerprint-based localization sys-
describes an analytical model showing how this balanced tems, the Cisco system requires an offline training phase
AP distribution mitigates the effect of RSS signal-pattern during which a tracked device is used to calibrate a radio
variation on positional accuracy. map. Therefore, our unsupervised learning approach is also

Figure 10 RSSI difference vs.

decreased positional accuracy
688 Mobile Netw Appl (2009) 14:677–691

manually obtain the available hardware pairs. They then

assume the existence of an easily distinguishable location
(i.e., a location with a unique RSS fingerprint) such as the
entrance to a room. When an unknown WiFi device moves
through this location, its signal patterns are collected, and a
small set of calibrated-observed variant signal pairs are
produced. Instead of computing the transformation from
this limited information, the best-fit transformation identi-
fied in the manually compiled transformation database is
assigned to this unknown device. However, an easily
distinguishable location is required, which may only exist
in certain environments. Additionally, a new WiFi device
may never pass through these easily distinguishable
locations or may not pass through them early enough to
learn its transformation function. Haeberlen et al. [1] also
explore fully automatic calibration. They suggest EM and
Figure 11 A balanced AP distribution that mitigates the effect of RSS particle filtering to learn coefficients, but do not mention
signal-pattern variation on positional accuracy. The APs are marked in implementation details or experiments. Kjærgaard et al. [4]
red circles. The tracking device is moving within the blue rectangular propose an automatic method using RSS ratios when
area matching closeness in the RSS fingerprints. Although using
the RSS ratios reduces the linear shift effect in RSS
applicable to address the hardware variance problem in the fingerprints caused by hardware variance, it can still
Cisco system. Additionally, since our unsupervised learning produce errors (as described in Section 2) when the offset
algorithms are computed on the server side, it works well component in the linear shift is large.
with the Cisco system in that no software installation is Some WiFi location systems [23] do not use RSS
needed on the WiFi devices. fingerprints for location determination; for these systems,
Hardware variance can be minimized by not relying on therefore, hardware variance does not present a problem.
RSS information to implement WiFi positioning systems. For One such system, the Aeroscout WiFi location system, is
example, Cheng et al. [22] develop a method of ranking (i.e., based on accurately computing the TDOA (time difference
high to low) RSS values instead of relying on absolute RSS of arrival) of WiFi signals. In order to accurately measure
values. Ranking overcomes the hardware variance problem the time differences, specialized and proprietary hardware
because it is not effected by a linear shift in RSS. However, is needed to tightly synchronize the clocks on the WiFi
the ranking approach sacrifices positional accuracy because devices. NearMe [23] also avoids the hardware variance
the detailed data for signal strength levels is not used. problem by using only the received access points for
This hardware variance problem was addressed by proximity determination. Since its goal is to compute not
Misikangas et al. [3], who propose a manual approach fine-grained location but proximity among devices, the
based on pair-wise mapping. In this approach, different detailed RSS information is not needed.
hardware devices are placed at the same position at the
same time to differentiate their signal patterns; their
mapping functions are then derived. Haeberlen et al. [1] 6 Conclusion and future work
propose a similar manual solution. Their experiments
demonstrated a linear RSS mapping function between This work presents an unsupervised learning method for
different WiFi devices. They then used the least-squares solving the hardware variance problem in WiFi localization.
fit method to determine the coefficients in the linear At runtime, the unsupervised learning method automatically
transformation function. However, they did not apply the learns a transformation function for mapping WiFi signal
signal-pattern transformation function in a positioning patterns from an unknown tracking device to the training
engine to demonstrate improved positional accuracy. device under which the radio map was calibrated. Several
Although manual adjustment can achieve good positional learning algorithms, including online regression, EM,
accuracy, its main problem is the required manual labor, neural network and extended EM were designed, imple-
which does not scale well with the size of the environment mented, and evaluated in a working WiFi localization
and the wide array of emerging WiFi devices. system and environment. The experimental results demon-
Misikangas et al. [3] describe an automatic approach for strate that, in WiFi localization, applying a transformation
solving this hardware variance problem. Firstly, they function learned from our unsupervised learning reduces
Mobile Netw Appl (2009) 14:677–691 689

Figure 12 Highest probability

distribution at location u0

position error caused by device hardware variance by an The above localization system is assumed here. Further,
average of 4%. Additional case study experiments show without loss of generality, the following assumptions are
that (1) positional accuracy in the same training/tracking made:
device pairs does not degrade, and (2) unsupervised
1. RSS decays linearly;
learning is effective for both variable-speed and constant-
2. The variances of all pairs are identical.
speed movement. Finally, RSS signal-pattern variance was
shown to be uncorrelated with positioning error. Consider the one-dimensional example in Fig. 12. Two
Several questions remain for future studies in unsuper- access points APa and APb are at either side of a tracking
vised learning. First, this technique could be applicable to device. Suppose the tracking device is at an arbitrary
other RSS-based localization systems such as Zigbee, position u0 on the line from 0 (the leftmost position) to z
WiMax, and GSM, in reducing positional error caused by (the rightmost position).
hardware variance. Second, this unsupervised learning According to the first assumption above, if the RSS
technique could also be applicable in adapting WiFi directly beneath an access point is s, the distributions of
localization to dynamic environment factors affecting WiFi (RSS=s/u0, AP=APa) and (RSS=s/(z-u0), AP=APb) at posi-
signal patterns such as humidity level, human presence, tion u0 are identical, and ua and ub are located at u0. If the
open/closed doors, etc. above two RSS signal patterns are entered into an RSS-based
positioning engine, the estimated location will be u0.
As Fig. 13 shows, if the tracking device differs from the
APPENDIX: Analytical model for the balanced AP training device with a linear RSS mapping function with
distribution slope = 1 as in the first assumption, the RSS distribution is
simply shifted.
A typical location in an RSS fingerprint localization system Although u0 is no longer the most probable location for
is characterized by several (RSS, AP) pairs, where RSS is both APs, the multiplied probability is still the highest. This
not a single value but rather a distribution of signals outcome is demonstrated by comparing the multiplied
collected from the training phase and often modeled by a probability at each position. Since the RSS variances are
Gaussian distribution. While tracking a device, the proba- assumed2 identical, in the p.d.f. of Gaussian distribution
ðx
mÞ
p1ffiffiffiffi e
2s 2 , we need only compare the (x - u) part. Restated,
2
bility of a set of observed (RSS, AP) pairs against a certain s 2p
location is then computed by multiplying all the probabil- the smaller the value, the higher the probability.
ities acquired from the previously modeled Gaussian At u0, after multiplication, the next procedure would be
probability distribution function. The location with the
highest joint probability is the output result. ½u0
ðu0
bÞ2 þ½u0
ðu0 þ bÞ2 ¼ 2b2 :

Figure 13 Highest probability

distribution at location u0 shifted
by hardware difference with
linearity of y=x+b
690 Mobile Netw Appl (2009) 14:677–691

Assume an arbitrary position denoted as (u0+d) in 16. El-Gallad A, El-Hawary M, Sallam A (2001) Swarming of
intelligent particles for solving nonlinear constrained optimization
Fig. 13. After substitution, (x - u)2 becomes
problem. Int J Eng Intell Syst 9(3):155–163 Sept. 2001
17. Krishnan P, Krishnakumar A, Ju W-H, Mallows C, Ganu S (2004)
½ðu0 þ d Þ
ðu0
bÞ2 þ½ðu0 þ d Þ
ðu0 þ bÞ2 A system for LEASE: Location estimation assisted by stationary

 emitters for indoor RF wireless networks, in Proceedings of IEEE
¼ 2 d 2 þ b2 : INFOCOM, pp. 1001–1011
18. Chen Y-C, Chiang J-R, Chu H-H, Huang P, Tsui AW (2005)
Since 2(d2 + b2) > 2b2 for any nonzero d, u0 is the Sensor-assisted Wi-Fi indoor location system for adapting to
location with maximum probability. environmental dynamics, in Proceedings of ACM International
Symposium on Modeling, Analysis and Simulation of Wireless
and Mobile Systems (MSWIM 2005), pp. 118–125
19. Lin T-H, Chen J-P, Chen H-H, Huang P, Chu H-H (2006)
Enabling energy-efficient and quality localization services, in
References Proceedings of the IEEE International Conference on Pervasive
Computer and Communications (PERCOM 2006), Work In
1. Haeberlen A, Flannery E, Ladd AM, Rudys A, Wallach DS, Progress Session, pp. 624–627
Kavraki LE (2004) Practical robust localization over large-scale 20. Chan L-W, Chiang J-R, Chen Y-C, Ke C-N, Hsu J, Chu H-H
802.11 wireless networks, in Proceedings of ACM MOBICOM, (2006) Collaborative localization—enhancing WiFi-based posi-
pp. 70–84 tion estimation with neighborhood links in clusters, in Proceed-
2. Tao P, Rudys A, Ladd AM, Wallach DS (1996) Wireless LAN ings of the International Cconference on Pervasive Computing
location–sensing for security applications, In Proceedings of the (PERVASIVE 2006), pp. 50–66
AAAI Thirteenth National Conference on Artificial Intelligence, 21. You C-W, Chen Y-C, Chiang J-R, Huang P, Chu H-H, Lau S-Y
pp. 11–20 (2006) Sensor-enhanced mobility prediction for energy-efficient
3. Misikangas P, Lekman L (2003) Applications of signal quality localization, in Proceedings of Third Annual IEEE Communica-
observations, Patent Cooperation Treaty (PCT) WO 2004/008796 tions Society Conference on Sensor, Mesh and Ad Hoc Commu-
A1 nications and Networks (SECON 2006), pp. 565–574
4. Kjaergaard M, Munk C (2008) Hyperbolic location fingerprinting: 22. Cheng YC, Chawathe Y, LaMarca A, Krumm J (2005) Accuracy
a calibration-free solution for handling differences in signal Characterization for Metropolitan-scale Wi-Fi Localization, in
strength, In Proceedings of IEEE International Conference on Proceedings of the Third International Conference on Mobile
Pervasive Computing, pp. 110–116 Systems, Applications and Services (MobiSys), pp. 233–245
5. Bahl P, Padmanabhan VN (2000) RADAR: An in-building RF- 23. Krumm J, Hinckley K (2004) The NearMe Wireless Proximity
based user location and tracking system, In Proceedings of IEEE Server, in Proceedings of the Sixth International Conference on
Infocom, pp. 775–784 Ubiquitous Computing, pp. 283–300
6. Bahl P, Padmanabhan VN, Balachandran A (2000) Enhancements
to the RADAR User Location and Tracking System, Microsoft
Research Technical Report: MSR-TR-00-12, February 2000
7. Ladd AM, Bekris KE, Rudys A, Marceau G, Kavraki LE, Wallach
DS (2002) Robotics-based location sensing using wireless Ether-
net, in Proceedings of the Eighth Annual International Conference
on Mobile Computing and Networking (MOBICOM), pp. 227–
238
8. Roos T, Myllymaki P, Tirri H, Misikangas P, Sievanan J (2002) A
probabilistic approach to WLAN user location estimation. Int. j. wirel.
inf. netw. 9(3):155–164, July 2002. doi:10.1023/A:1016003126882
9. Thrun S (2000) Probabilistic algorithms in robotics. AI Magazine
21(4):93–109
10. Pearson K (1900) Mathematical contribution to the theory of
evolution VII: on the correlation of characters not quantitatively
measurable. Philos. Trans. R. Soc. Lond. Ser. A: Math. Phys. Sci.
195:1–47 doi:10.1098/rsta.1900.0022
11. Seshadri V, Zaruba GV, Huber M (2005) A Bayesian sampling
approach to in-door localization of wireless devices using
received signal strength indication, in Proceedings of IEEE
International Conference on Pervasive Computing (PERCOM
2005), pp. 75–84
12. Hightower J, Borriello G (2004) “Particle filters for location
estimation in ubiquitous computing: a case study,” In Proceedings Arvin Wen Tsui is a Ph. D. student at the Graduate Institute of
of Ubiquitous Compuating, pp. 88–106 Networking and Multimedia, National Taiwan University and also
13. Schulz D, Fox D, Hightower J (2003) People tracking with works at the Industrial Technology Research Institute, Taiwan as a
anonymous and ID-sensors using Rao-Blackwellised particle technical deputy manager. He received his B.S. (1994) in Computer
filters, in Proceedings of IJCAI, pp. 921–926 Science from Soochow University and M.S. (1996) in Computer
14. Fox D, Hightower J, Liao L, Schulz D, Borriello G (2003) Science and Information Engineering from National Taiwan Univer-
Bayesian filtering for location estimation, in Proceedings of IEEE sity. From 1996–1998, he was a research assistant at the Institute of
Pervasive Computing, pp. 24–33 Information Science, Academia Sinica. His current research interests
15. Michell T (1993) Machine Learning, McGraw Hill, pp. 239–240 include localization and ubiquitous computing.
Mobile Netw Appl (2009) 14:677–691 691

Hao-Hua Chu is an associate professor at the Graduate Institute of

Yu-Hsiang Chuang received the B.S. and M.S. degrees in informa- Networking and Multimedia and Department of Computer Science and
tion management from National Chung Chen University, Chiayi, Information Engineering, National Taiwan University. He received his B.S.
Taiwan in 2004 and 2006, respectively. He is currently an engineer in (1994) in Computer Science from Cornell University and PhD (1999) in
the Information & Communications Research Laboratories of the Computer Science from University of Illinois at Urbana Champaign. From
Industrial Technology Research Institute of Taiwan, R.O.C. His 1999–2000, he was asenior software engineer at Intel. From 2000–2003, he
research interests include localization over large-scale 802.11 wireless was a project manager at (NTT) DoCoMo USA Labs. His research areas
networks, and wireless sensor networks. are pervasive computing and sensor/wireless network.