MASARYK UNIVERSITY
FACULTY OF INFORMATICS
Localization of mobile devices
using machine learning
BACHELOR'S THESIS
Ales Calabek
Brno, Spring 2020
Declaration
Hereby I declare that this paper is my original authorial work, which
I have worked out on my own. A l l sources, references, and literature
used or excerpted during elaboration of this work are properly cited
and listed in complete reference to the due source.
Aleš Calábek
Advisor: Mgr. Adam Janovský
i
Acknowledgements
Foremost, I would like to express my gratitude to my advisor Adam
Janovský for the advice and guidance he provided throughout the
writing of this thesis. I would also like to thank all my friends and
family for the moral support.
iii
Abstract
This thesis examines the potential of machine learning for G S M positioning
based on received signal strength in urban environments. The
goal was to assess two algorithms, namely Support Vector Machine
and Multilayer Perceptron, and compare their performance to other
localization techniques. The models were evaluated on data collected
in the city of Brno, Czech Republic using the scanner application
Molotras. I show that modelling signal propagation using the studied
algorithms does not provide an advantage over traditional methods
due to excessive noise in the data caused by buildings blocking the
signal's path. Finally, I propose an alternative approach to using neural
networks that works on the principle of fingerprinting.
iv
Keywords
localization, GSM, RSSI, machine learning, Support Vector Machine,
Multilayer Perceptron
v
Contents
1 Introduction 1
2 Radio technology 3
2.1 Wireless Communication 3
2.2 GSM Network 4
2.3 Localization 6
2.3.1 Time of Arrival 6
2.3.2 Received signal strength 7
2.3.3 Angle of Arrival 7
3 Machine learning 11
3.1 Machine learning 11
3.2 Support Vector Machine 12
3.2.1 Support Vector Regression 13
3.3 Artificial Neural Networks 14
3.3.1 Perceptron 14
3.3.2 Multilayer Perceptron 16
4 Dataset 19
4.1 Used technologies 19
4.2 Data collection 20
4.3 Data processing 22
5 Evaluation 27
5.1 Training 27
5.2 Results 29
5.3 Comparison 31
6 Conclusion 35
A Evaluation 43
B Archive structure 45
vii
List of Tables
4.1 Number of received signal strength records per one
measurement. 23
5.1 Fingerprinting in the city centre showed the lowest
localization error as compared to other positioning
methods that otherwise performed comparably. Data
from[l]. 32
5.2 M L models showed higher localization error than the
traditional methods when tested on the village dataset.
Data from [1]. 32
A . l When trained on data from the city centre, both SVR and
M L P showed similar localization error to the baseline
model. 43
A.2 Training on data from the residential areas significantly
impaired the SVR's performance in the city centre. 43
A.3 The implemented models performed worse when
working with only one signal value at a time. Table shows
positioning error when evaluated in the city centre. 44
ix
List of Figures
2.1 Location of the receiver can be found at the intersection of
imaginary spheres formed around each satellite [11]. 7
2.2 Two angles to known locations are sufficient to locate a
mobile device using triangulation [13]. 8
3.1 Multiple hyperplanes can consistently divide an input
space (left). The optimal hyperplane maximizes the
margins between it and the samples on either side (right).
Images from [16]. 12
3.2 The classes cannot always be separated directly (left).
Kernels project the data into a higher dimension where it
may become possible to find a dividing hyperplane
(right). Images from [21]. 13
3.3 Different activation functions are suitable for different
tasks. They also greatly influence prediction accuracy and
computation time. 15
3.4 Neurons in one layer of the M L P connect to all neurons in
the subsequent layer [25]. 16
3.5 During the gradient descent, weights are adjusted in the
direction that reduces error [26]. 17
4.1 I have split the collected measurements into two datasets:
city centre (red) and residential (blue). Includes data
from[l]. 21
4.2 There does not seem to be a clear relationship between
received signal strength and distance. 25
5.1 The M L P showed identical localization accuracy to the
baseline model. 29
5.2 A l l models showed worse performance when working
with only one signal value at a time instead of using data
from all available cells at once. 30
5.3 The positioning error in the village increased by an order
of magnitude. 30
xi
1 Introduction
Mobile phones are a part of everyday life for most people in the developed
world. They provide a variety of convenient utilities, among
others, the location-based services used in mobile navigation and fitness
applications. The majority of people, however, does not realize
that their location can be tracked even with their phone's location
services turned off.
One way of locating devices is to take advantage of its connection to
the cellular network. As signal strength readings include the identity
of the transmitting cell towers, the phone can be located based on its
distance to the towers' positions.
In this thesis, I examined the potential of using machine learning
(ML) for improving the localization accuracy. I studied two M L models,
namely Support Vector Machine and Multilayer Perceptron. The
implementation was done in the Python programming language and
used the scikit-learn and Keras libraries. Both models were evaluated
on data I have collected in the city centre of Brno using the Molotras
scanner application [1]. Finally, I compared my results with traditional
positioning techniques.
The structure of this thesis is as follows: In Chapter 2,1 introduce
the fundamentals of radio signal propagation, the structure of a cellular
network and several localization approaches. Chapter 3 presents
the M L algorithms mentioned above, and Chapter 4 describes the
process of collecting, analyzing and processing the data. I elaborate
on the specifics of training the models, evaluating their performance
and comparing their results in Chapter 5. Finally, Chapter 6 provides
a conclusion and suggests possible future work.
1
2 Radio technology
The primary function of a cellular network is to provide the means for
communication. We can, however, use the properties of the transmitted
signal itself to find the location of a device in the network.
In the first part of this chapter I introduce the basics of signal
propagation. The second part provides information about the structure
and workings of the second-generation cellular network. Finally, the
third part presents several techniques used for localization.
2.1 Wireless Communication
In modern society, people and devices need to communicate with
each other. Common household appliances such as television, radio,
Wi-Fi or mobile phones all connect wirelessly. Wireless technology
enables us to transmit information over long distances by using electromagnetic
(EM) waves without the use of physical point-to-point
links.
Frequency, amplitude and phase are the three fundamental characteristics
of an E M wave. Information can be encoded into the wave by
changing one of the three properties in a process called modulation.
Frequency modulation (FM), in particular, is widely known for its
use in F M radio broadcasting [2].
The E M spectrum encompasses all E M radiation frequencies ranging
from the low-frequency radio waves to the high-frequency gamma
rays. However, not all frequencies are suitable for communication
purposes.
Signals at the very low end of the spectrum, with wavelengths
of thousands of kilometres, can travel halfway around the Earth before
they require any form of amplification, but provide very low
bandwidth. O n the other end of the spectrum, signals can carry an
exceptional amount of information but are prone to absorption by rain
and reflection off obstacles. For the above reasons, communication
technology primarily uses the middle of the E M spectrum—the radio,
microwave, infrared, and visible frequencies [3].
Considering the great impact that obstacles have on wave propagation,
it is worth elaborating on their effects some more.
3
2. RADIO TECHNOLOGY
When a wave collides with an object, part of its energy is transmitted
through, while the rest is reflected back. For instance, reflections
may occur when the receiver lies inside a building and the wave must
pass through the walls.
Diffraction is another phenomenon affecting the wave, causing it to
bend as it travels around an obstacle. This change in direction allows
waves to propagate to places behind objects.
Before reaching the receiver, radio waves may reflect and diffract
many times. Multiple versions of the signal then arrive at different
times and may interfere with each other, making it difficult to decode
the resulting signal. Interference also occurs due to other transmissions
on the same frequency or the presence of noise. Analyzing these
effects is essential in order to understand how waves behave. However,
simulating the propagation of radio signals in a real environment
would be infeasible because it would require precise knowledge about
all objects on the scene [4].
Mathematical models simplify the problem of wave tracing by
estimating how much power the signal loses over the path from transmitter
to receiver. One of the simplest propagation models is Free
Space Path Loss which assumes a direct path of the signal with no atmospheric
attenuation or interference. The path loss is given by the
equation
where d is the distance from the transmitter to the receiver and A is
the wavelength of the radio signal [5] [6] [7].
With the knowledge of radio wave properties and their behaviour,
we can now better understand how cellular networks use wireless
transmission to provide communication services.
The cellular network forms the foundation of modern communication
technology. It enables the transmission of voice, text messages and
Internet data all over the world.
The Global System for Mobile Communications (GSM) Network
is a second-generation (2G) cellular network. It was first deployed
(2.1)
2.2 GSM Network
4
2. RADIO TECHNOLOGY
in 1991 and is still in use today. It comes with many improvements
over the first-generation analogue network—notably, higher capacity
and digital encryption, allowing for more people to communicate
privately.
A n essential part of any cellular network is the Base Transceiver
Station (BTS). A BTS used in the G S M network consists of the radio
transmitter-receiver and its associated antenna. These antennas can be
seen on top of masts or tall buildings, where surrounding structures
do not block their signal.
Still, high-rise buildings in cities prove themselves to be enough
of an obstacle to necessitate higher density of BTS towers to provide
coverage of all areas and to handle the increased traffic. Rural regions,
on the other hand, are usually covered by fewer but more powerful
transmitters placed at high terrain to maximize their range.
It is, however, not possible to prevent overlap of the signals in
adjacent cells which may lead to reduced call quality. The unwanted
interference can be eliminated by using different frequencies in neighbouring
cells and reusing them in cells that are further apart.
Another difficulty that comes with dividing the area into sections
is providing a continuous connection when the mobile moves across
the boundaries. At all times, the mobile scans available radio channels
looking for the most reliable one. When the current serving cell's
signal drops below a critical level, the network arranges a handover,
and the mobile reconnects to a different cell [8] [9] [10].
During the handover, the mobile and the network need an unambiguous
way to refer to the cells. Therefore, every BTS is uniquely
distinguishable by four identification numbers:
• Mobile Country Code (MCC) specifies a country,
• Mobile Network Code (MNC) determines the mobile operator,
• Location Area Code (LAC) refers to a group of cells,
• Cell ID (CID) identifies individual BTSs within that group.
With the knowledge of a cell's identity, its position can be easily learned
by searching a database. If we could use the properties of the received
signal to determine the distance to several known cells, we could find
the location of the mobile receiver.
5
2. RADIO TECHNOLOGY
2.3 Localization
The simplest localization method in a cellular network is Cell Identification
or Cell ID. In Cell ID, the mobile's and the serving cell's
locations are considered to be the same. The accuracy of this method
generally ranges in hundreds or thousands of metres and largely depends
on the dimensions of the area covered by the cell.
Better accuracies can be achieved by using more elaborate approaches.
Time of arrival, received signal strength and angle of arrival
are just three of the radio signal properties that enable us to determine
the location of the receiver.
2.3.1 Time of Arrival
Time of Arrival (TOA) methods make use of the time it takes a signal
to reach the receiver. Distance to the transmitter can be calculated numerically
based on the signal's propagation speed and the time delay.
With known distances to multiple transmitters, the location of the
mobile receiver can be estimated using the method of multilateration.
Multilateration forms imaginary circles around the transmitters'
positions with the calculated distances as radii. These circles, in theory,
intersect at the precise position of the receiver. In practice, however,
the intersecting circles delimit an area, and optimization techniques
have to be used to minimize the positioning error.
Arguably the most well-known system based on TOA is the Global
Positioning System (GPS). Each satellite in the GPS network broadcasts
a report about its position and the time of transmission. With
the data from at least four satellites, the mobile receiver can deduce
its location in three dimensions, as illustrated in Figure 2.1 [11].
While GPS positioning can provide localization accuracies of tens
of centimetres, it comes with several disadvantages. As line of sight
to the satellite is a necessity, the use indoors, between high-rise buildings
and in mountainous terrain is limited. It is also the most powerconsuming
method making it infeasible in smaller devices. Moreover,
additional hardware in the form a GPS chip is required[12].
6
2. RADIO TECHNOLOGY
Figure 2.1: Location of the receiver can be found at the intersection of
imaginary spheres formed around each satellite [11].
2.3.2 Received signal strength
As radio signal travels through space, it loses energy. With the knowledge
of the power radiated by the transmitter Pt and the signal strength
Pr received by the mobile device, the path loss PL can be calculated
using the equation
PL = Pt- Pr. (2.2)
A propagation model, such as FSPL mentioned in section 2.1, can
be applied to estimate the distance that the signal has travelled. The
mobile device then computes its position based on the distance estimates
and cell tower locations using multilateration.
Contrary to the TOA methods, techniques based on Received Signal
Strength work with no additional equipment and come free with the
GSM infrastructure. However, localization accuracy is generally worse
due to variations in signal strength caused by interference, multipath
propagation and path blocking.
2.3.3 Angle of Arrival
Angle of Arrival (AOA) is the principal component in a radar system.
The radar rotates its receiving antenna and notes the changes in received
signal strength. The AOA is determined from the direction of
the antenna at the time it receives the strongest signal. The radar then
uses TOA to calculate the distance to a target and find its location in
three dimensions.
7
2. RADIO TECHNOLOGY
When using A O A to locate mobile devices in a cellular network
without the aid of TOA, at least two angle measurements are necessary.
With two angles to known positions, the mobile station can compute its
location by using the method of triangulation, as shown in Figure 2.2.
One advantage of AOA methods is the reduced effect of interference
achieved by rejecting signals coming from directions other than
those of the transmitter. Still, noise and path blocking limit the localization
accuracy in the same way as with the methods based on
received signal strength [13].
Figure 2.2: Two angles to known locations are sufficient to locate a
mobile device using triangulation [13].
Conclusion
Radio waves enable us to communicate wirelessly. We can encode
information by modulating their properties such as frequency, amplitude
or phase. As radio waves travel through space, they are affected
by obstacles in their way. We use mathematical models to estimate the
loss of signal power on the wave's path from transmitter to receiver.
The GSM network is a cellular network composed of a grid of BTS
towers, each uniquely identifiable. Physical properties of the radio
signal can be used to locate devices in the network.
In this thesis, I focus on positioning methods based on received
signal strength because they require no additional hardware and can
8
2. RADIO TECHNOLOGY
be readily implemented on the existing infrastructure. I explore the
feasibility of improving the traditional techniques by using machine
learning algorithms, which are described in the next chapter.
9
3 Machine learning
Machine learning (ML) systems are part of everyday life. They are
used to recommend videos on Youtube or products on Amazon.
Google uses M L to provide personalized results in its search engine
and to filter spam messages in Gmail's inboxes. Speech recognition
and synthesis enables virtual assistants, such as Apple's Siri
or Amazon's Alexa, to interact with people using natural language.
Machine learning systems can also drive cars, unlock phones with
facial recognition, detect credit card frauds, and much more.
In the first part of this chapter, I introduce the concept of machine
learning and describe one class of M L algorithms. Parts two and three
then provide detailed information about the two algorithms used
in this thesis, namely the Support Vector Machine and Multilayer
Perceptron.
3.1 Machine learning
Machine learning enables computers to perform tasks commonly done
by humans. It provides machines with the ability to learn and improve
from experience without being explicitly programmed to solve a problem.
As a field of study, M L is a part of a broader discipline of Artificial
Intelligence [14].
M L methods can be divided into several categories based on the
type of data they process and the type of problem they intend to solve.
In this thesis, I concentrate on algorithms that fall into the supervised
learning category.
Supervised learning is typically used for classification or regression
problems. During classification, we want to decide to which class an
object belongs. For instance, deciding whether or not an email is spam
is a classification problem. With regression, we want to predict a value
in a continuous range, such as the stock price in a stock market.
In both classification and regression, the goal is to find the connection
between the inputs and the outputs. A learning algorithm
processes the training data and creates a statistical model capturing
this relationship. The quality of the model heavily depends on the
11
3. M A C H I N E LEARNING
quality of the dataset and noisy data with incorrect or missing values
will dramatically reduce the model's accuracy.
Another aspect concerning performance is overfitting. Overfitting
refers to creating a model that fits the training data well and can predict
outputs accurately, but is not able to generalize to data it has not seen
before [15].
Some of the widely adopted supervised learning algorithms include
the Support Vector Machine and Multilayer Perceptron, which I
used in this thesis to locate devices in the G S M network, and which
are described in the next two sections.
3.2 Support Vector Machine
Support Vector Machine (SVM) is a supervised learning algorithm
typically used for classification. It looks at the input data as points
in space and tries to divide the space so that instances of different
categories are separated. New data points are then classified based on
their position relative to the decision boundary, or the hyperplane.
A hyperplane is a subspace of a dimension one less than that of its
ambient space. For example, in two dimensions, the hyperplane is a
line.
Figure 3.1: Multiple hyperplanes can consistently divide an input
space (left). The optimal hyperplane maximizes the margins between
it and the samples on either side (right). Images from [16].
12
3. M A C H I N E LEARNING
Figure 3.1 demonstrates that multiple hyperplanes can consistently
divide an input space. The further a data point is from the decision
boundary the more confident we can be about predicting its class.
Therefore, we want to find the hyperplane with maximum margins
between it and the samples on both sides. The data points closest to
the hyperplane are then called the support vectors.
Nevertheless, a hyperplane cannot always divide the input data
directly, as illustrated by Figure 3.2. The SVM overcomes this problem
by using soft margins and kernels. Soft margins tolerate the misclassification
of some data points and the SVM tries to balance the trade-off between
maximizing the margins and minimizing the mistakes. A kernel
is a function that transforms the input space into a higher dimension
and enables the S V M to model non-linear relationships [17] [18].
3.2.1 Support Vector Regression
The SVM can also be extended to handle regression problems. Instead
of separating the dataset, Support Vector Regression (SVR) tries to find
a hyperplane that fits the data and includes all data points within its
margins. The width of the margins then indicates how much deviation
the models allows [19] [20].
Data in R^3 (separable)
Figure 3.2: The classes cannot always be separated directly (left). Kernels
project the data into a higher dimension where it may become
possible to find a dividing hyperplane (right). Images from [21].
13
3. M A C H I N E LEARNING
3.3 Artificial Neural Networks
The idea of an Artificial Neural Network (ANN) was first introduced
in the 1950s and was inspired by the workings of a human brain.
As in the brain, the output of an A N N depends on the connections
between individual processing units, or neurons. However, contrary
to our brains that work massively in parallel, computers perform
their operations sequentially and cannot simulate large networks in
real-time. Still, A N N s are a powerful class of algorithms capable of
solving many challenging problems, some of which I mentioned at
the beginning of this chapter.
3.3.1 Perceptron
The perceptron is a mathematical model representing a single neuron.
It accepts real number values X[ as inputs and calculates their weighted
sum £ (also called the inner potential) using corresponding weights Wj.
N
l=YJwixi (3.1)
i=Q
The output value of the perceptron y is then obtained by applying
the activation function f to the inner potential.
y = / ( £ ) (3.2)
Activation functions play a critical role in the learning process as
different functions are suitable for different tasks. They can have a
significant influence on prediction accuracy as well as the computation
time. Figure 3.3 shows some of the most commonly used activation
functions that include the linear, sigmoid, tanh or ReLU functions.
The linear function of the form f(x) = cx is one of the simplest
activation functions and is fast to compute. The downside is that it
cannot express non-linear relationships because the combination of
linear functions is again a linear function. It is usually used in the last
layer of the network to predict unbounded values.
The sigmoid function f(x) = 1_^1
e-x maps the outputs to a range
between 0 and 1, which is especially useful for predicting probabilities.
However, for very high or low values of x, there is almost no change
14
3. M A C H I N E LEARNING
(c) tanh (d) ReLU
Figure 3.3: Different activation functions are suitable for different tasks.
They also greatly influence prediction accuracy and computation time.
to the predictions which may lead to the vanishing gradient problem
that causes longer training times of the network.
The hyperbolic tangent or tanh f(x) = 1+^-lx — 1 is similar to
the sigmoid function. It ranges from -1 to 1, meaning the negative
inputs are mapped to negative values, and the zero inputs are mapped
near zero. Along with the sigmoid, tanh is expensive to compute, and
suffers from the vanishing gradient problem as well.
The Rectified Linear Unit or ReLU f(x) = max(0,x) is another
non-linear activation function. At the negative end of its range, it is
also affected by vanishing gradient. Still, it is ideal for use in larger
networks due to its computational efficiency making it possibly the
most used activation function today [22] [23] [24].
15
3. M A C H I N E LEARNING
3.3.2 Multilayer Perceptron
Individual perceptrons can be organized into a structure of interconnected
layers to form one of the simplest neural network models—the
Multilayer Perceptron. The combination of non-linear activations and
the network arrangement makes the M L P capable of approximating
an arbitrary function.
The layers of an M L P are of three types. The first layer accepts the
input data and forwards them into the network, hence the input layer.
It is followed by one or more hidden layers that make up the core of
the network. Perceptrons in the hidden layers use various activation
functions to capture the essence of the data and generalize the relations
between the inputs. In practice, there is no general best number of
hidden layers, and it has to be tailored to the task at hand. Finally, the
output layer is the result of the M L P and is interpreted according to
the specific problem being solved.
As shown in Figure 3.4, every neuron in one layer connects to
all neurons in the previous layer. The backpropagation algorithm then
adjusts the weights associated with these connections, and that is how
the network learns.
Hidden
Figure 3.4: Neurons in one layer of the M L P connect to all neurons in
the subsequent layer [25].
16
3. M A C H I N E LEARNING
Throughout the training, backpropagation uses a lossfunction that
calculates the network's error by comparing its outputs with the correct
solutions in the training dataset. The algorithm then looks for
the combination of weights which minimizes the loss by using the
gradient descent procedure. First, it calculates the partial derivatives of
the loss with respect to weights and looks at the slope of the tangent.
It then adjusts the weights in the direction that reduces the loss, as
illustrated in Figure 3.5. The algorithm repeats this process until the
error converges to a local minimum [27] [28] [29] [30] [31].
The weights alone, however, are not enough to explain how the
algorithm arrives at its solution, and a review of the neural network's
results by human experts is still necessary to detect and correct possible
hidden flaws.
Conclusion
Machine learning enables computers to learn and perform tasks previously
only done by humans. Supervised learning is just one category
of M L algorithms that tries to find connections between inputs and
outputs, and is typically used for classification and regression tasks.
Support Vector Machines classify data by finding a hyperplane
that separates the classes in the sample space. Alternatively, an SVR
can approximate a function by fitting data within the hyperplane's
margins.
17
3. M A C H I N E LEARNING
Multilayer Perceptron is a type of Artificial Neural Network. It
is composed of individual processing units called neurons that are
arranged into interconnected layers. The M L P learns by adjusting these
connections during training using the backpropagation algorithm.
In the next chapter, I continue with how I collected and prepared
the signal strength data that were used to teach these models to locate
mobile devices in the G S M network.
18
4 Dataset
Data play an essential role in the machine learning process. The learning
algorithm requires them in order to create the prediction model,
and the quality of the dataset can significantly influence the model's
performance.
The first part of this chapter introduces some of the utilised tools
and software libraries, the second part describes the process of collecting
the data and the third part focuses on preparation and analysis of
the dataset.
4.1 Used technologies
As a part of this thesis, I have created several Python scripts that experiment
with the M L models and assess their suitability for localization
based on received signal strength. To perform an analysis and process
the collected data, as well as to implement the M L algorithms, I have
used several third-party software tools that are described below.
NumPy
NumPy is a Python library that enables efficient handling of multidimensional
arrays and matrices and provides a variety of mathematical
functions that operate on these arrays. Due to its enhanced performance
compared to the standard Python library, it is well suited for
scientific computations and other applications that have to manage
extensive amounts of data [32].
pandas
•pandas is an open-source tool for data analysis and manipulation written
in Python. It provides two primary data structures—Series and
DataFrames—that are designed for intuitive work with structured
tabular data. Among others, pandas can be used to load raw data
from CSV or Excel files, handling missing values, data aggregation
and database operations such as merging or joining datasets [33].
19
4. DATASET
scikit-learn
scikit-learn is an open-source library built on top of NumPy and pandas
that provides simple and efficient tools for predictive data analysis.
It implements a range of machine learning algorithms, for instance,
Decision Trees, Naive Bayes and Support Vector Machines [34].
Keras
Keras is a high-level API used for fast prototyping with neural networks.
It represents the network as a graph of independent modules that can
be easily combined to create more complex models. Keras is capable of
utilising multiple back-end computation engines such as TensorFlow,
C N T K or Theano. It can also take advantage of the G P U to improve
performance and decrease training times [35].
OpenCelllD
OpenCelllD is a community project that collects GPS positions and
other information about GSM cell towers all over the world. It is freely
accessible and can be used for locating devices or understanding
network coverage [36].
4.2 Data collection
The hardware I used for collecting the data consisted of a single Motorola
Moto G6 Play mobile phone with the Android operating system.
The phone was fitted with an 02 SIM card and equipped with a GPS
unit. The GPS positioning information was used as ground truth during
the experiment.
The signal strength data were obtained using a specialised scanner
application Molotras developed by Mgr. Jakub Martinka as part of his
master's thesis [1]. Once every second, the application requests the
GPS position of the device and gathers signal measurements of up to
seven surrounding cells.
Molotras records the cells' identification numbers and various
properties of the radio signal. However, only the four identifiers—
20
4. DATASET
16.590 16.595 16.600 16.605 16.610 16.615
Longitude
Figure 4.1:1 have split the collected measurements into two datasets:
city centre (red) and residential (blue). Includes data from [1].
M C C , M N C , L A C and Cell ID—along with the received signal
strengths are required to estimate the location of the mobile device.
When the collecting of data is completed, Molotras saves all information
into a CSV file. Every measurement consists of a header line
with the GPS coordinates, followed by one line for each of the scanned
cells.
To collect the measurements, I have ridden a bicycle through the
streets of Brno, Czech republic with the scanner application running.
Cycling was chosen over walking to provide more extensive coverage
of the area.
21
4- DATASET
As signal propagation characteristics vary in different environments,
the focus of this thesis was placed on the urban areas of the city
centre and the surrounding residential blocks. Areas from open environments
that have different propagation properties were removed to
improve the performance of the models. Figure 4.1 shows the final area
under study after filtering the data. In total, I have collected 3.2 MB of
raw data over a period of three days.
Additional data were obtained by including the measurement
taken for the Molotras thesis. One of the datasets also covers the city
centre of Brno and could be readily joined with the rest of the data.
The other dataset was taken in Trenčianska Turná, a village in Slovakia,
that served for testing how well the models generalise for other types
of environments.
4.3 Data processing
The output of the Molotras application was in the form of a single CSV
file with multiple lines per measurement. As the header lines had a
different structure from the rest, the raw data had to be parsed with a
simple script before loading them into a pandas DataFrame.
The parsing script generates an ID number for each measurement
and, along with the GPS location of the sample, attaches it to all records
associated with that measurement. A l l columns are subsequently of
the same type, and the tabular data can be worked with further.
The parsed data were then separated into three subsets according
to the location in which they were collected. The centre dataset covers
the city centre and contains 16,240 measurements. The residential
dataset includes 6,936 measurements from the neighbourhoods north
of the city centre, and the village dataset comprises 2,387 samples from
the Slovakian village.
On average, each measurement contained between five and six
individual signal strength recordings. However, due to imperfections
of the measuring setup, about 10 percent of the records lacked the
cell tower identification and had to be excluded. Table 4.1 provides
a more detailed view of the datasets and the number of records per
measurement.
22
4. DATASET
Table 4.1: Number of received signal strength records per one mea-
surement.
Dataset Centre Residential Village
Records per
measurement
Count % Count % Count %
7 5204 32.80 1421 20.49 376 15.75
6 5790 36.49 2445 35.25 786 32.93
5 3259 20.54 1890 27.25 549 23.00
4 1170 7.37 782 11.28 468 19.60
3 349 2.20 279 4.02 205 8.59
2 42 0.27 84 1.21 0 0.00
1 53 0.33 35 0.50 3 0.13
Sum 15867 100 6936 100 2387 100
The next step was to add the BTS towers' GPS coordinates. Fortunately,
the OpenCelllD project also covers the Czech Republic and
Slovakia, and their database could be conveniently incorporated using
the pandas .merge function that performs a database-style join.
With known positions of the cell towers, distances between the
BTSs and the measurement locations could be determined using the
haversine formula. The haversine formula calculates the distance of
two points on a sphere and can be used to approximate the distance
between two locations on Earth.
Figure 4.2 then plots the calculated distance against received signal
strength across the three datasets. At first glance, there does not seem
to be a clear relationship between the two quantities. Still, the machine
learning models may find a more suitable representation of the data
and extract useful information.
Since the learning algorithms I used in this thesis work with fixedlength
inputs, the measurement lengths had to be equalised. It is,
however, not apparent how many records each measurement should
have. O n the one hand, taking fewer records could lead to a loss
of information, and on the other, taking more would retain records
obtained over a longer distance that are more prone to noise.
23
4. DATASET
-110 -100 -90 -BO -70 -60
received singal strength (dBm)
24
(b) Residential area
4. DATASET
-110 -100 -90 -80 -70 -60 -50
received singal strength (dBm)
(c) Village
Figure 4.2: There does not seem to be a clear relationship between
received signal strength and distance.
By experimenting with different values, taking four records per
measurement showed a good compromise while retaining over 91 percent
of the data since shorter samples were excluded.
The alternative to removing the shorter samples was to generate
the missing values. This approach did not provide any significant
improvements and on the contrary, could potentially introduce a bias
towards the generated values.
For comparison, an additional dataset from the city centre that
considers only the one nearest record was prepared, conceivably modelling
the propagation of radio signal.
Finally, all datasets were split 60/20/20 into training, validation
and testing subsets. The training data served as a reference for teaching
the models, the validation set was used to evaluate the models'
performance during training, and the testing set was put aside for
25
4. DATASET
the final assessment. The next chapter follows with a more detailed
explanation of the training and evaluation.
26
5 Evaluation
The main objective of the models was to transform received signal
strength into distance. The predicted distances were used in multilateration
to find the position of the receiver, and the mean localization
error was used to evaluate the model.
In the first part of this chapter, I discuss the specifics of training the
M L models and tuning hyperparameters of the learning algorithms. In
part two, I present the results and examine the models' performance
on the prepared datasets. Finally, in the third section, I compare the
methods implemented in this thesis with other localization techniques.
5.1 Training
A machine learning model is at its core a mathematical formula. The
model learns by adjusting the formula's parameters in a way that
fits the learning data. There is, however, another set of parameters
affecting the learning process that cannot be derived from the data.
Hyperparameters define higher-level qualities of the model, such
as the number of hidden layers of a neural network, that are usually
set before the training begins. Choosing the right combination of
hyperparameters is done by trying different options and selecting the
one that produces the best performing model.
Besides manually searching for the best candidate solution, the
most straightforward method of optimizing hyperparameters is Grid
Search. Grid Search systematically tests all combinations of hyperparameters
and returns the one that yields the best results.
Despite its simplicity, exploring all possible solutions using Grid
Search becomes increasingly time-consuming as the number of hyperparameters
and their individual choice of values grows. Additionally,
the optimal solution may not even lie on the grid but rather in the
spaces between the tested values [37] [38].
Random Search, in contrast, randomly samples the search space
and has been proven to find higher scoring configurations in a shorter
time than the Grid Search technique [39]. For the above reasons, I
have used Random Search in this thesis to tune the hyperparameters
of the SVR and M L P models.
27
5- EVALUATION
SVR
During the training of the SVR, the considered hyperparameters included
the kernel, the regularization parameter C, the tolerance margin
epsilon, and the kernel coefficient gamma.
The linear and RBF kernels provided the best results among the
tested candidates. I excluded the sigmoid and polynomial kernels from
the final search, due to poor results of the former and long training
times with no significant benefit of the latter.
The C and epsilon parameters were randomly drawn from a uniform
distribution in the range 0-32; gamma came from the range 0-4.
In general, larger values of C and epsilon, in combination with smaller
values of gamma, produced the best results.
MLP
Tuning the hyperparameters of the M L P consisted of adjusting its
topology and trying various activation functions.
I experimented with networks of one to three hidden layers with
the number of neurons ranging between 4 and 192, decreasing in each
consecutive layer. The studied activation functions included the linear,
tanh, ReLU and sigmoid activations.
For training the MLP, I used the Huber loss function [40] and the
Adam [41] optimizer with default parameters. The effects of other
optimizers and learning rates were negligible.
I also examined the option of using Dropout layers and activity
regularization to prevent overfitting. Again, with no significant im-
provements.
Baseline model
I implemented a third, baseline model that estimates the receiver's position
equally distant from each BTS. In such a case, the multilateration
algorithm computes the weighted centroid of the area given by the
cells, assigning each cell the same weight. This model was expected to
provide only rough positioning estimates and served for comparison
to the M L models.
28
5- EVALUATION
5.2 Results
All three implemented models were first trained and evaluated in the
city centre and residential area. The box chart in Figure 5.1 shows a
visual representation of the outcome, and for a more detailed view of
the results, you can see the tables in Appendix A .
The median positioning error in the city centre was around 65 meters
for all models. However, the accuracy of the SVR dropped significantly
when trained on data from the residential parts of the city. The
results in the residential area were more balanced, and the median
error was approximately 114 meters for all models.
Another approach I studied was modelling the signal propagation
function directly. In this case, the models worked with only one signal
value at a time instead of using data from all available cells at once.
Figure 5.2 illustrates that this method was not as effective, and the
median error increased to about 140 meters in the city centre.
SVR
1 1 1 1
Train: Centre Residential Centre Residential
Test: Centre Centre Residential Residential
Figure 5.1: The M L P showed identical localization accuracy to the
baseline model.
29
5- EVALUATION
Figure 5.2: All models showed worse performance when working with
only one signal value at a time instead of using data from all available
cells at once.
Figure 5.3: The positioning error in the village increased by an order
of magnitude.
30
5- EVALUATION
Finally, to test how well the models generalize to different environments,
I evaluated them on the village dataset (see Figure 5.3).
As expected, the results differ by order of magnitude. The median
positioning error was around 700 meters, with the worst estimates
being off by two and a half kilometres.
Discussion
It should be pointed out, that the M L P performed with an identical
precision to the baseline model, and the SVR achieved similar results
at best. Such an outcome implies that the M L models were not able
to filter the noise caused by buildings blocking the signal's path. The
models instead learned to mimic the centroid algorithm, as another
way of minimizing the localization error. Consequently, the better
positioning accuracy in the city centre can be attributed to the higher
density of BTS towers, as compared to the residential area or the
village.
Upon closer inspection of the MLP, the predictions certainly did
not resemble the expected distance values. Moreover, the topology of
the network had little significance, as vastly different configuration
were able to reach the baseline model's performance.
In conclusion, the M L models I tested in this thesis were not able
to take advantage of the multidimensional input and extract useful
information that would make them outperform the baseline model.
Their performance closely imitated the centroid algorithm, and the
localization accuracy could be linked to the density of BTS towers in
the tested area.
5.3 Comparison
Tables 5.1 and 5.2 compare the results of the models I have implemented
to the localization techniques used in the Molotras thesis—
centroid, trilateration and fingerprinting. All methods were evaluated
on the same data from the city centre of Brno and the village Trenčianska
Turná.
The centroid method computes the arithmetic mean of the cells'
coordinates and is comparable to my baseline model. Trilateration uti-
31
5- EVALUATION
Table 5.1: Fingerprinting in the city centre showed the lowest localization
error as compared to other positioning methods that otherwise
performed comparably Data from [1].
Model Median (m) Mean (m) Min (m) Max (m)
SVR 67.26 78.16 0.42 391.30
M L P 65.27 74.64 1.28 316.87
Baseline 65.28 74.63 1.28 316.62
Centroid 66.34 64.77 17.8 128.37
Trilateration 69.70 69.79 13.75 121.99
Fingerprinting 9.23 16.26 2.22 61.37
Table 5.2: M L models showed higher localization error than the traditional
methods when tested on the village dataset. Data from [1].
Model Median (m) Mean (m) Min (m) Max (m)
SVR 716.24 861.50 52.31 2475.42
M L P 696.23 848.56 40.86 2498.73
Baseline 696.30 848.60 40.93 2498.73
Centroid 605.93 691.69 208.17 2172.08
Trilateration 539.11 587.18 154.97 1889.19
Fingerprinting 49.72 90.43 4.89 306.63
lizes the Okumura-Hata signal propagation model to convert received
signal strengths into weights for the multilateration algorithm. Both
approaches showed similar results of 65-70 metres median error in
the city and slightly outperformed the M L models in the village.
By far, the best performance was achieved with the fingerprinting
method. Fingerprinting uses a database of signal strength measurements
taken in the area of interest. It compares the currently observed
signal readings with the database and locates the receiver based on
the most similar reference point. This technique, however, comes with
notable disadvantages as creating and maintaining the database is
32
5- EVALUATION
very time and resource consuming, and has to be done separately for
each location.
Other state-of-the-art studies using received signal strength for
GSM localization show similar results. For instance, the research [42]
achieved 65-metre positioning error in metropolitan areas and showed
that an advantage is gained by scanning cells of all available providers.
The same holds for [43] where Ibrahim et al. demonstrated that fingerprinting
methods reach accuracies around 50 metres even when
only using the associated cell tower information.
33
6 Conclusion
In this thesis, I examined the potential of using machine learning
algorithms in G S M positioning.
I have collected a dataset of over twenty thousand signal strength
measurements in the city of Brno that was used for evaluating the
M L models. The analysis showed a lot of noise in the data caused by
buildings blocking the path of the signal—the same readings could
be obtained from a nearby cell hidden behind a building, as well as
from a more distant cell with a clear line of sight.
The studied models—SVR and MLP—were then expected to learn
the relation between received signal strength and distance while taking
advantage of signal readings from multiple surrounding cells. For
comparison with the M L models, I implemented a third, baseline
model that calculated the centroid of the area given by the cells, and
provided only rough positioning estimates.
Both SVR and M L P achieved similar positioning accuracy with
the median error of 65 metres in the city centre and 115 metres in the
residential area. This outcome, however, closely matched the accuracy
of the baseline model, which achieved identical results.
Further analysis of the models' performance revealed that they
were not able to extract any additional information from the data.
Instead, they learned to imitate the centroid algorithm as another way
of minimizing the localization error. Consequently, by evaluating the
models in different environments, I demonstrated that their accuracy
could be linked to the cell tower density in the tested area.
In conclusion, I showed that the tested M L algorithms do not
provide an advantage over traditional localization techniques when
used to model signal propagation in urban areas.
Possible future research could try to reduce the effects of noise by
measuring signal strengths with a clear line of sight to the transmitting
radio tower. The fitted M L model could then be used for localization
in open environments where buildings do not block the path of the
signal. This technique could be directly compared with the Free Space
Path Loss model.
Another approach would be to let a neural network learn a map of
the area based on signal strength measurements and identities of the
35
6. CONCLUSION
cells. This method would work on the same principle as fingerprinting
techniques and would require extensive coverage of the area for training
the model. The neural network, however, could prove itself to be a
more robust solution that could handle changes in the environment
even after the training phase.
36
Bibliography
1. M A R T I N K A , Jakub. Locating mobile phones using signal strength
measurements. 2020. Available also from: https: / / i s .muni. cz/
th/asrql/. Master's thesis. Masaryk University.
2. POOLE, Ian. What is Frequency Modulation, FM [online] [visited
on 2020-05-25]. Available from: https : / /www . electronics notes
. com / a r t i c l e s / radio / modulation / frequency -
modulation-fm.php.
3. GOLENIEWSKI, Lillian. The Electromagnetic Spectrum and Bandwidth
[online] [visited on 2020-05-25]. Available from: https :
//www.informit.com/articles/article.aspx?p=24687.
4. H E N D E R S O N , Tom. Reflection, Refraction, and Diffraction [online]
[visited on 2020-05-25]. Available from: https : / /www .
physicsclassroom.com/class/waves/Lesson-3/Reflection,-
Refraction,-and-Diffraction.
5. POOLE, Ian. Free Space Path Loss: details & calculator [online] [visited
on 2020-05-25]. Available from: https : //www. electronicsnotes
. com/articles/antennas - propagation / propagation-
overview/free-space-path-loss.php.
6. KOSTANIC, Ivica. Okumura and Hata Macroscopic Propagation Models
[lecture slides]. 2004. Available also from: https : //my. f i t .
edu/~kostanic/RF0
/„5C0
/„20Propagation/01d0
/„5C0
/„20Notes/RF°/o
5C0
/o20Propagation%5Cy„20-y„5Cy„2007-Okumura0
/„5C//
o20and%5Cy„
20Hatayo5Cyo20Macroscopicyo5Cyo20Propagationyo5Cyo20Models .
pdf.
7. A N A N D A M U R U G A N , S.; N A N D H I N I , P. S. Wireless Networks.
Hauppauge, New York: Nova Science Publishers, Inc., 2016. ISBN
978-1634851787.
8. POOLE, Ian. Cellular network basics [online] [visited on 2020-05-
25]. Available from: https : //www . electronics-notes . com/
a r t i c l e s / connectivity / c e l l u l a r - mobile - phone / network -
architecture.php.
37
BIBLIOGRAPHY
9. POOLE, Ian. What is GSM: Primer [online] [visited on 2020-05-
25]. Available from: https : //www . electronics-notes . com/
articles/connectivity/2g-gsm/basics-introduction.php.
10. POOLE, Ian. GSM Network Architecture [online] [visited on 2020-
05-25]. Available from: https://www.electronics-notes.com/
articles/connectivity/2g-gsm/network-architecture.php.
11. GISGEOGRAPHY. Trilateration vs Triangulation - How GPS Receivers
Work [online] [visited on 2020-05-25]. Available from:
https : //gisgeography . c o m / t r i l a t e r a t i o n - t r i a n g u l a t i o n -
gps-
12. POOLE, Ian. What is Satellite Navigation SatNav, GNSS [online]
[visited on 2020-05-25]. Available from: https : / /www .
electronics - notes . c o m / a r t i c l e s / s a t e l l i t e s / s a t e l l i t e -
navigation-satnav/what-is-satnav-gnss-technology.php.
13. BENSKY, Alan. Wireless Positioning Technologies and Applications.
2nd ed. Boston (Mass): Artech House, 2016. ISBN 9781608079513.
14. H E A T H , Nick. What is machine learning? Everything you need to
know [online] [visited on 2020-05-25]. Available from: https :
/ / www . zdnet . com / a r t i c l e / what - i s - machine - learning -
everything-you-need-to-know/.
15. SONI, Devin. Supervised vs. Unsupervised Learning [online]
[visited on 2020-05-25]. Available from: https : / /
towardsdatascience . com / supervised - vs - unsupervised -
learning-14f68e32ea8d.
16. DRAKOS, Georgios. Support Vector Machine vs Logistic Regression
[online] [visited on 2020-05-25]. Available from: https :
//medium.com/Ogeorge.drakos62/support-vector-machine-
vs-logistic-regression-94cc2975433f.
17. C H E N , Lujing. Support Vector Machine — Simply Explained
[online] [visited on 2020-05-25]. Available from: https : / /
towardsdatascience . com/support-vector-machine-simply-
explained-fee28eba5496.
18. NG, Andrew. Support Vector Machines [lecture notes]. 2018. Available
also from: http : //cs229 . Stanford . edu/notes/cs229-
notes3.pdf.
38
BIBLIOGRAPHY
19. BHATTACHARYYA, Indresh. Support Vector Regression Or SVR
[online] [visited on 2020-05-25]. Available from: https : / /
medium.com/coinmonks/support-vector-regression-or-svr-
8eb3acf6d0ff.
20. SHARP, Tom. An Introduction to Support Vector Regression (SVR)
[online] [visited on 2020-05-25]. Available from: https : / /
towardsdatascience . com / an - introduction - to - support -
vector-regression-svr-a3ebcl672c2.
21. K A N D A N , Harish. Understanding the kernel trick, [online] [visited
on 2020-05-25]. Available from: https : //towardsdatascience.
com/understanding-the-kernel-trick-e0bc6112ef78.
22. S H A R M A , Avinash V. Understanding Activation Functions in Neural
Networks [online] [visited on 2020-05-25]. Available from:
https : / / medium . com / the - theory - of - everything /
understanding-activation-functions-in-neural-networks-
9491262884e0.
23. MISSINGLINKAI. 7 Types of Neural Network Activation Functions:
How to Choose? [online] [visited on 2020-05-25]. Available from:
https://missinglink.ai/guides/neural-network-concepts/
7 - types - neural - network - activation - functions - right /
#section2.
24. S H A R M A , Sagar. Activation Functions in Neural Networks [online]
[visited on 2020-05-25]. Available from: https : / /
towardsdatascience . com / activation - functions - neural -
networks-lcbd9f8d91d6.
25. WIKIPEDIA. Artificial neural network [online] [visited on 2020-
05-25]. Available from: https : / / e n . wikipedia . o r g / w i k i /
A r t i f icial_neural_network.
26. H U T S O N , Matthew. AI researchers allege that machine learning is
alchemy [online] [visited on 2020-05-25]. Available from: https:
/ / www . sciencemag . org / news / 2018 / 05 / a i - researchers -
allege-machine-learning-alchemy.
39
BIBLIOGRAPHY
27. H A N S E N , Casper. Neural Networks: Feedforward and Backpropagation
Explained & Optimization [online] [visited on 2020-05-
25]. Available from: https : //mlfroms cratch . com /neural -
networks-explained/*/.
28. K A P U R , Rohan. Rohan & Lenny #1: Neural Networks & The Backpropagation
Algorithm, Explained [online] [visited on 2020-05-
25]. Available from: https : //ayearofai . com/rohan-lenny-
1 - neural - networks - the - backpropagation - algorithm -
explained-abf4609d4f9d.
29. S H A N E , Janelle. Neural networks, explained [online] [visited on
2020-05-25]. Available from: https : //physicsworld . com/a/
neural-networks-explained/.
30. ROJAS, Raul. Neural Networks: A Systematic Introduction. Berlin
Heidelberg: Springer, 1996. ISBN 978-3540605058.
31. KRIESEL, David. A Brief Introduction to Neural Networks. 2007.
Available also from: available°/„20at°/„20http: //www. dkriesel.
com/en/science/neural_networks.
32. O L I P H A N T , Travis. NumPy [online] [visited on 2020-05-25].
Available from: https: //numpy. org/.
33. MCKINNEY, Wes. pandas [online] [visited on 2020-05-25]. Available
from: https: //pandas . pydata. org/.
34. COURNAPEAU, David, scikit-learn [online] [visited on 2020-05-
25]. Available from: h t t p s : / / s c i k i t - l e a r n . o r g / s t a b l e / .
35. CHOLLET, Francois et al. Keras [online] [visited on 2020-05-25].
Available from: https: //keras. i o / .
36. U N W I R E D LABS. OpenCelllD [online] [visited on 2020-05-25].
Available from: https: //opencellid. org/.
37. CAPONETTO, Guilherme. Random Search vs Grid Searchfor hyperparameter
optimization [online] [visited on 2020-05-25]. Available
from: https://towardsdatascience.com/random-searchvs
- grid - search - for - hyperparameter - optimization -
345el422899d.
40
BIBLIOGRAPHY
38. DESCAMPS, Benoit. Tuning Hyperparameters {part II): Random
Search on Spark [online] [visited on 2020-05-25]. Available from:
https : //towardsdatascience . com/hyperparameters - part -
ii-random-search-on-spark-77667e68b606.
39. BERGSTRA, James; BENGIO, Yoshua. Random search for hyperparameter
optimization. Journal of machine learning research. 2012,
vol. 13, no. Feb, pp. 281-305.
40. WIKIPEDIA. Ruber loss [online] [visited on 2020-05-25]. Available
from: https://en.wikipedia.org/wiki/Huber_loss.
41. KINGMA, Diederik P; BA, Jimmy. Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980. 2014.
42. C H E N , Mike Y et al. Practical metropolitan-scale positioning for
gsm phones. In: International Conference on Ubiquitous Computing.
2006, pp. 225-242.
43. IBRAHIM, Mohamed; YOUSSEF, Moustafa. A hidden markov
model for localization using low-end G S M cell phones. In: 2011
IEEE International Conference on Communications (ICC). 2011,
pp. 1-5.
41
A Evaluation
Table A . l : When trained on data from the city centre, both SVR and
M L P showed similar localization error to the baseline model.
Test set Model Median(m) Mean(m) Min(m) Max(m)
SVR 67.26 78.16 0.42 391.31
Centre M L P 65.27 74.64 1.28 316.87
Baseline 65.28 74.63 1.28 316.62
SVR 110.37 127.22 1.19 677.68
Residential M L P 113.85 125.53 2.54 624.85
Baseline 113.75 125.51 2.56 625.21
SVR 716.24 861.50 52.31 2475.42
Village M L P 696.23 848.56 40.86 2498.73
Baseline 696.30 848.60 40.93 2498.73
Table A.2: Training on data from the residential areas significantly
impaired the SVR's performance in the city centre.
Test set Model Median(m) Mean(m) Min(m) Max(m)
SVR 138.05 157.89 2.82 684.31
Residential M L P 114.97 128.09 4.42 569.31
Baseline 114.79 128.09 4.29 569.33
SVR 226.85 225.45 2.51 526.22
Centre M L P 64.57 73.93 0.26 325.74
Baseline 64.55 73.95 0.15 325.71
43
A . EVALUATION
Table A . 3 : The implemented models performed worse when working
with only one signal value at a time. Table shows positioning error
when evaluated in the city centre.
Model Median (m) Mean (m) Min (m) Max (m)
S V R 156.75 165.74 3.79 819.89
M L P 139.20 149.81 7.83 815.00
Baseline 139.69 149.67 9.45 816.79
44
B Archive structure
The electronic archive thesis.zip submitted with this thesis contains the
following files:
• cells/230.csv and cells/231.csv- OpenCelllD database files covering
the Czech Republic and Slovakia respectivelly [36]
• data/cell_data.csv and data/sk_cell_data.csv - signal strength
measurements (Brno + Trenčianska Turná) [1]
• data/my_data.csv - signal strength measurements I have collected
(Brno)
• main.py - main script that runs the experiment, run via. /main.py
• evaluation.py, models.py, parse.py, preprocessing.py and trilateration.py
- Python scripts that execute individual parts of
the experiment
• out.txt - example output of main. py
• requirements.in and requirements.txt - Python requirements
generated by pip-compile
45