MTAT magistritööd – Master's theses
Permanent URI for this collection
Browse
Recent Submissions
Item Prediction of a movie’s box office using pre-release data(Tartu Ülikool, 2020) Bondarenko, Stanislav; Sharma, Rajesh, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutIt’s difficult to overestimate the impact of the film industry in our lives, it expands our knowledge about the world and culture and entertains. Going to the cinema has become an important leisure activity. Moreover, the total worldwide box office in 2018 hit a significant amount of $41B. This is not surprising as only in 2018 there were released 11,911 feature-length films worldwide. The box office generated from cinema ticket sales is the main source of profit for widely released movies. However, not all movies are successful in terms of profit when the cost of production is compared with the total box office. 78% of movies released worldwide are not profitable and 35% of profitable movies earn 80% of the total profit. Seeing the importance of theatrical screenplays and tough competition for the profit made, we want to be able to predict how successful a movie is going to be and whether it is worth taking the risk of investment. Only pre-release available data is used to be able to make a prediction at the earliest stages. We went through several stages typical for data mining and machine learning to obtain possibly the biggest and feature-rich dataset used in box office gross prediction. We use neural networks and gradient boosting machines to be able to predict the absolute box office gross, predict within which range it is likely to be, and whether a movie will be profitable, and the results obtained are very competitive in the domain.Item Measuring Corporate Reputation through Online Social Media: A case study of the Volkswagen Emission Scandal(Tartu Ülikool, 2020) Odeyinka, Olubunmi T; Sharma, Rajesh, juhendaja; Lapitskaya, Darya, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutReputation, a priceless but yet essential factor in a company’s success, influences a company’s customers, employees, and even the public relations with the company. Therefore, most companies consider reputation building as a priority in their daily interaction with their staff and the public. Such companies that successfully build a good reputation, which attracted investors, staff, and goodwill of the public sometimes run into issues or events (i.e. self-inflicted or not) that can result in a significant change in their reputation. The series of events that happen after a reputation tarnishing event includes more related news article releases, comments in message boards, and an increase in the number of people talking about the event on social media. In this thesis, we analyzed the company’s online media reputation (i.e. online media sentiment) and compared it with their financial reputation (i.e. stock price and volume) using the Volkswagen (VW) emission scandal as a case study. We did this by computing the monthly correlation, weekly correlation, rolling correlation, partial correlation, trend analysis, and brought out the context of a discussion with a word clouds and we tested our hypothesis of a relationship between online media sentiment and stock values using the Granger causality test. This analysis was done not just on VW online media sentiment and stocks but also on Ford, Toyota, and Audi as control cases for the period under analysis (i.e. VW scandal period). The result shows that during the heat of the scandal the company that is involved (i.e. VW) is the one that presents their sentiment result to be a measure of their reputation only in the media sources that had their context based on the scandal(i.e. news and Twitter) and with a stronger relationship in the weeks of major scandalous news or events. This relationship is not so for other substitute companies’ media dataset like Ford and Toyota whose media context and sentiment were not influenced by the VW scandal events. While Audi, a subsidiary of VW, and our third control case presented sentiment to be a relative measure of the corporate reputation in the news dataset because that is the only dataset of it that had the VW emissions scandal as one of its main focal points of discussion during the scandal timeline.Item Gender-based segregation in company boards and well-being(Tartu Ülikool, 2020) Kadri, Oluwagbemi Omobolanle; Sharma, Rajesh, juhendaja; Kungas, Peep, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutSegregation is an act of division from a supreme body to smaller groups because of the characteristics of the body. In order terms among humans, we can refer to it as an unwarranted detachment or separation resulting in traits a person possesses; for example, gender, occupation, race, resident, income, religion, age, etc. Analyses based on segregation impact in our society have increased over the years. It has spawned enormous controversial discussions in our modern-day world; and elicited several researchers’ interests to identify the origin of segregation. In this thesis, we investigated if gender and age segregation exist in Estonian companies’ boards and its relationship with the labour market. In addition, we examine if it leads to high credit risks and a negative correlation to the well-being of Estonian society. The key measurement factors for comparison and drawing conclusions are the unemployment rate measured as the labour market, financial key performance indicators measured as credit risk, a well-deprivation index measured as well-being and segregation indexes from ‘SCube’ data model measured as segregation. ‘SCube’ originated from a model created by researchers at the University of Pisa, it uses a data science framework to deal with the problem of social and occupational segregation. Analysis from the ‘SCube’ data-set will be measured with segregation indexes ranging from 0 to 1 in accordance to this range high level of segregation means high value of the segregation index meaning a value close to 1. The Estonian statistics ready-made data-set is used in conjunction with the data-set from ‘SCube’ model to examine and draw conclusions of the occupational segregation problem discussed in this work. In addition, statistical techniques correlation and causal inference are used to determine the relationship and causal effects between segregation and the various factors.Item Systematic Literature Review on EEG-based BCI Applications(Tartu Ülikool, 2020) Värbu, Kaido; Muhammad, Yar, juhendaja; Muhammad, Naveed, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutElectroencephalography (EEG)-based brain-computer interface (BCI) is a system that pro-vides pathway between the brain and external device via interpreting EEG. EEG-based BCI applications have been initially developed for medical reasons such as enabling patients in completely locked-in state to communicate and rehabilitation of stroke patients. Nowadays EEG-based BCI applications gain increasing significance also in the non-medical domain, where applications are being developed also in order to enable healthy persons to be more efficient, collaborate, develop themselves and much more. The applications in non-medical domain include for example applications for smart home control, monitoring concentration, live brain-computer cinema performance and gaming. The objective of the work is to give systematic overview on the literature on EEG-based BCI applications from the period of 2009 until 2019. In the study the trends in the research have been analyzed. The distribution of the research between medical and non-medical domain has been reviewed and further categorization into fields of research within the domains. In the study also, the equipment used for gathering EEG data and signal processing methods have been reviewed. The sys-tematic literature review has been prepared following the PRISMA model. During the pro-cess three well known databases PubMed, Scopus and Web of Science were selected to conduct the publication search. After the initial result, duplicate publications were removed and unique publications further screened and assessed for eligibility. After assessment 202 eligible publications were included in the final analysis. The overall number of articles and conference proceedings has been increasing throughout the years. The amount of research is increasing faster within non-medical domain in comparison to medical domain. The majority of the research has been done in Asia with China contributing the highest number of publications throughout the years. In the study also the overview of the distribution of the EEG based BCI applications among different domains and fields has been given together with techniques and devices used. In the last part of the study current challenges in the field and possibilities for the future have been analyzed.Item Predicting stock price based on media monitoring(Tartu Ülikool, 2020) Aasmaa, Ardi; Sharma, Rajesh, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutUsing automated systems for finding investment ideas becomes more popular every year and the models are getting more complex. In recent years a lot of studies have been conducted that have researched the possibilities of using social media sentiment as input for stock prediction models. However, the results have been contradicting as the problem is complex. In this research, data was collected from Twitter about Standard Poor’s 100 companies over a period of six months. Also, financial data with one minute interval was collected from Alpha Vantage. Five different machine learning algorithms were used to predict maximum profit and maximum loss for the prediction horizon of five trading days. It was investigated whether adding social media based features to financial data based features would improve the results and if so, then tweets from what type of users would give the highest information gain. It was found out that adding social media data as input is beneficial for both, predicting maximum loss and maximum profit. For the explainability part, Shap library was used. As found out, features extracted from financial data were most important. For social media based features, most information was gained from tweets posted by news agencies and by users having relatively few followers.Item Discriminatory Speech on Digital Platform a case study of Twitter (Gender, Race, Politics, Sexuality)(Tartu Ülikool, 2020) Festus, Fortune Ikechukwu; Sharma,Rajesh, juhendaja; Ritter, Christian Simon, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutIn recent years communication via social media has become more personal and available for every individual or group of people irrespective of their interests. This has enable people express their thoughts, ideas and views freely. Though it brings lots of ease to communication, it also gives rise to discriminatory challenges. Online hate is a major example of such challenge.As these social media users grow, so does the impact of hate speech. Despite the magnitude and growth level of research in this field there is a hug gap in understanding the hate speech and how it affects certain aspects of human life’s e.g race,gender,sexuality, politics. This has prompted researchers to apply techniques like social networks analysis to detect hate groups. But in this research we strongly believe that the content of hate matters as well.Thus in this paper we apply sentiment analysis and topic modelling to understand the discuss of hate as it affect race, gender, sexuality, politics. Our result shows that the content plays an important role in preventing and eradicating discrimination of these platforms.Item Algoritmilise mõtlemise oskust arendav arvutivaba programmeerimise õppematerjal II ja III kooliastmele(Tartu Ülikool, 2020) Kokk, Getriin; Palts, Tauno, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutMiddle school curriculum by the government of Estonia supports digital competence, techno-logical and innovative development but does not set exact guidelines for teaching. At the moment, middle school informatics focus mainly on teaching of how to use a computer rather than teaching informatics and programming. Due to curriculum not having exact guidelines for the subject, the objective for this thesis was to create lesson plans and educational materials on education of computational thinking with unplugged programming for the II and III stage of middle school. All materials created for this thesis were tested and assessed by the students on the following topics: captivation; likeability; understandability of the material.Item Real-Time Event Detection System for Mobile Data(Tartu Ülikool, 2020) Mohebbian, Mohammad Mahdi; Hadachi, Amnir, juhendaja; Saluveer, Erki, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutMobile data is one of the prior data sources which can be used for urban study analytics due to the amount of valuable information they contain, such as type of mobile data, time and most importantly data’s coordinates. In order to keep the cell services stable for users all across the country it is crucial for authorities to be aware of unannounced gatherings which can cause traffic overload on cell towers in the area. In this thesis implementation of a enterprise system has been demonstrated for monitoring the behavior of the cell towers under the administration’s authority. The core functionality of this system is detecting ongoing events in different areas on an hourly-basis schedule utilizing multiple statistical approaches for abnormality detection. The output of the event detection section of the system is an approximate estimation of the ongoing event’s location on the map. Current design of the system is aiming to fullfill the downsides of similar approaches for event and crowd detection such as high processing expenses and noncomprehensive resources by using parallel servers, distributing the processing load while keeping the pipline clear for user’s demands, and utilizing Call Detail Records (CDR) data as input resources which gives the advatages of containing the majority of mobile transactions and human behavior in the city.Item Two-Party ECDSA Protocol for Smart-ID(Tartu Ülikool, 2020) Iltšuk, Eduard; Paršovs, Arnis, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutSmart-ID is a digital signature solution based on threshold cryptography where two parties (mobile device and server) collaborate in key generation and signing process. The current solution uses RSA-based two-party signature scheme suggested by Buldas et al. in 2017 paper. This thesis proposes a Smart-ID-like solution based on Two-Party ECDSA protocol suggested by Lindell Yehuda in his 2017 paper. The thesis finds that the suggested ECDSA solution is able to provide the same security features as the current RSA-based Smart-ID solution, but with improved efficiency – more efficient key exchange, smaller signature size and does not require scalable secure storage on the server side. The security proof of the suggested ECDSA solution is not provided. However, the thesis provides a brief security analysis of the solution and the intuition why the suggested solution might be secure. The prototype implementation of the solution is also provided.Item Case Study: Blockchain and E-prescription Process(Tartu Ülikool, 2020) Liba, Martin Johannes; Milani, Fredrik Payman, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutHealthcare is an industry that highly values security, privacy and accessibility of data between patients and healthcare service providers. Prescriptions are an important part of healthcare services that enable patients to get medicine required for their treatment. About a decade ago, prescriptions were mostly handwritten on paper, today most countries use some form of e-prescription systems. E-prescriptions solved many problems that paper prescriptions had, for example pharmacists correctly not understanding what is written on the prescription and dispensing wrong medicine. E-prescription systems still have problems like lack of interoperability between providers and data security and fragmentation issues. Blockchain is one of the new emerging technologies that enables to securely store data using shared ledger, communicate the data between participants of the blockchain network, perform computational tasks via smart contracts and manage assets. These characteristics have potential to impact healthcare sector including e-prescription process. The thesis investigates how Estonian e-prescription process could be redesigned using blockchain technology. For that a case study is conducted to examine the Estonian as-is state of the e-prescription process and propose redesign based on blockchain technology. During the case study, systematic literature review was conducted on the subject to discover blockchain based redesign opportunities. In addition, redesign heuristics were applied on the as-is model. The results of the case study showed that using blockchain as the underlying technology is possible in Estonian case, but it would no bring significant value compared to existing process and Estonia has solved issues that blockchain can improve using different technologies.Item Automating the Release Planning of Mobile Apps by Including App-Reviews(Tartu Ülikool, 2020) Idoko, Onuche Akor; Scott, Ezequiel, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutStakeholders constantly think of the best and sustainable approach in delivering new releases to their customers. Large software companies like Google and Facebook invest huge money in their release planning process. That is because release planning impacts the end-user. One goal in software engineering is to make most or all stakeholders happy. However, start-ups, open-source projects and other small software organizations focused mainly on mobile app development may not have enough resources to invest in their mobile release management; as a result, it is important to plan the releases of mobile apps. The development team makes decisions like who is the release intended for, what functionalities or features should the release have, when should the release happen and how much quality should the release have. In order not to lose customers to competitors, teams must make these decisions carefully. Therefore, it is our strong conviction that with user app-reviews from mobile app stores (e.g., Play Store and App Store), we can automate and optimize the release planning of mobile apps. In this paper, we introduce an approach that automatically plans and optimizes mobile releases for software development teams by combining app-review and issue tracker information.Item Generating Process-based Mobile Applications for the Internet of Things using Automated Planning(Tartu Ülikool, 2020) Kaio, Kelian; Mass, Jakob, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutSmartphone devices are being used by more than half of the world population, and this means more opportunities to create mobile applications that help people with their daily lives. This thesis is looking into mobile apps for the Internet of Things, which are used in areas like smart homes, transportation, and healthcare. However, because of the massive scale of smart devices, supporting all of them is not feasible. Automated planning can help the application adapt to user’s and device’s context and support only those IoT devices which are needed by creating user-specific plans. These plans can be mapped into a business process model so the mobile application could execute them by using a business process engine. The goal of this thesis is to investigate and develop a framework that enables creating dynamic IoT mobile applications, using automated planning and business software management while taking into account user’s preferences and mobile device capabilities. Furthermore, it is analyzed which type of planning algorithm fits best for the motivating scenario. A framework prototype consisting of mobile application and backend is created for the motivating scenario is created as a proof of concept. The performance and scalability of the chosen planning algorithm and the developed prototype are evaluated.Item Splitting User Stories Using Supervised Machine Learning(Tartu Ülikool, 2020) Shahid, Muhammad Bilal; Scott, Ezequiel, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutUser stories are a well-known tool for representing requirements. They define small fragments of the system and help in the developer’s daily work. When we talk about user stories, then splitting them into tasks is common. Many approaches can be used to split a user story into tasks but these all approaches are based on manual working. In this era, where everything is now becoming digitalized. User stories should move to the next phase as well. In this paper, we will implement a novel idea to split a user story into tasks atomically using machine learning. We have used four machine learning algorithms random forest, SVM, KNN, and decision tree (ctree) on three open-source projects from Jira. The dataset we have used for this thesis was imbalanced, so we have used ROSE (randomly over sampling examples) and SMOTE (Synthetic Minority Oversampling Technique) to make a balanced dataset. We have applied machine learning algorithms separately on each project and also all projects combined into one dataset and then made comparisons on the results.Item Genetic effects on gene expression across cell types, tissues and biological contexts(Tartu Ülikool, 2020) Peikova, Kateryna; Alasoo, Kaur, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutThe human body consists of many tissues (e.g. brain, blood, skin or fat) which in turn are made of many different component cell types (e.g. neurons, monocytes, fibroblasts or adipocytes). The identities and functions of different cell types are defined by the different sets of genes that they express. Similarly, genetic differences between individuals can alter gene expression levels and in turn influence one’s risk of developing various complex diseases. Specific genetic variants associated with gene expression levels are referred to as expression quantitative trait loci (eQTLs). While multiple studies have demonstrated that the eQTL effect sizes vary between cell types and tissues, the magnitude of this variation has remained unclear. Although small studies focusing on purified cell types have generally reported large differences in eQTL effect sizes between cell types, the largest analysis of gene expression across 49 human tissues by the GTEx project found a high level of eQTL sharing between tissues. Furthermore, different analytical choices have made it difficult to compare results from different studies. Fortunately, the eQTL Catalogue project has recently released uniformly processed eQTL summary statistics from 19 individual studies. In this thesis, we used the eQTL Catalogue summary statistics to estimate the sharing of eQTLs across up to 46 individual cell types and tissues. Consistent with previous reports, we find high levels of eQTL sharing between tissues. In contrast, there was much less sharing between purified cell types. This suggests that high tissue-level sharing is driven by sharing of cell types between tissues and averaging of effect sizes across many different component cell types. This was further supported by factor analysis, which revealed that eQTL effect sizes in tissues were comprised of multiple shared and cell-type-specific components. Finally we tried use the cell-type-specific eQTL components to interpret complex disease associations, but did not find compelling evidence for specific enrichments. Our results indicate that much larger datasets from purified cell types are needed to completely interpret eQTL signals detected in complex tissues.Item Forecasting Bicycle Demand: Bologna Case Study(Tartu Ülikool, 2020) Davtyan, Taron; Sharma, Rajesh, juhendaja; Bertini, Flavio, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutAlthough there are a large number of academical studies conducted about demand forecasting in docked bike-sharing programs, there is scarce literature on the dockless bikesharing programs and especially in forecasting demand using a deep learning approach. Dockless bike-sharing programs have been growing rapidly during the past few years and having a model that can accurately predict bike usage is becoming essential for bikesharing companies and governmental institutions. This research paper aims to develop a model to forecast the usage of private bicycles with a deep learning approach and fill the research gap mentioned above. For predicting the number of rides, long short-term memory (LSTM) neural networks model was developed. The model was used to predict bike usage for 30-minute and 60-minute intervals. Besides the historical number usage of bikes, the prediction model considers air temperature, precipitation amount, and national holidays. The study results suggest that prediction with the LSTM model gives a more accurate outcome than more widely used machine learning algorithms such as linear regression, Random Forrest, and XGBoost. LSTM model that was developed by this study can be used to predict the utilization of bike lanes, which can be essential for governmental institutions and can also help bike-sharing companies to distribute bikes across the city to provide more convenient experience to the users.Item Twitter sentiment analysis to estimate happiness level(Tartu Ülikool, 2020) Rol, Nikolai; Sharma, Rajesh, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutHappiness is something that people strive for. However, it has always been hard to measure and understand what happiness depends on. This paper investigates if sentiment analysis can be used to estimate how happy people are and if sentiment correlates with socioeconomic factors or with the news. For analysis, text processing techniques were applied to Twitter posts gathered over the period from November 2019 to May 2020. The study shows a weak correlation with socio-economic factors, whereas the strongest relationship was with Health Care Quality. After a closer look into the change in daily sentiment, it was found that certain topics were discussed more than others on the dates with peaks. To investigate this aspect, the correlation analysis between sentiments of Twitter posts and news was made, however, the coefficient appeared to be low. The conclusion is that the result of sentiment analysis over Twitter data does not show a high correlation with socioeconomic factors, but it might have a certain dependency on events, news, or global shocks.Item Using Abstract Harmonic Analysis and the Lie Group Theory for the Study of Parameterized Quantum Circuits(Tartu Ülikool, 2020) Dolzhkov, Evgenii; Theis, Dirk Oliver Jim, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutQuantum computers are believed to be able to significantly outperform classical ones in terms of running time required to solve various problems. Near-term quantum computers, that can already be available in the nearest future, will have fairly limited resources, thus implying additional limitations and challenges. Near-term quantum algorithms are primarily based on parameterized quantum circuits. A parameterized quantum circuit is a quantum circuit, which is run repeatedly, while changing the numerical parameters some of the quantum operations in response to previous measurement results. Parameterized quantum circuits, however, need to be optimized, which can be simplified by endowing them with some mathematical structure, e.g., the ability to take derivatives or compute Fourier transforms. Here we study the possibility of using non-commutative Fourier transforms as a tool to find useful mathematical structure in parameterized quantum circuits.To our knowledge this thesis is the first work, where non-commutative Fourier transforms have been applied to parameterized quantum circuits. Our results include computations and theorems about non-commutative Fourier spectrum on parameterized quantum circuits. The results of this thesis provide a foundation, that opens the door for further study into derivatives and gradients of expectation functions on parameterized quantum circuits via the means of abstract harmonic analysis.Item GDPR and Blockchain Solutions(Tartu Ülikool, 2020) Pavlenkova, Ilona; Milani, Fredrik Payman, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutWithin the last years, there has been an increased attention for the development of Blockchain technologies and solutions. Blockchain represents one of the major headways’ technologies of the past decade, giving large groups of people and companies a possibility to reach agreement on record information without a central authority. Partially this can be achieved because data once stored on a Blockchain is immutable. The General Data Protection Regulation (GDPR) was developed in the European Union (EU) with the aim to standardize the privacy regulations across Europe and introduce the changes on how the personal data should be processed. The GDPR consists of a series of articles and chapters that require, among other things, provision a consent to collect and store data, identification of data controller, provision of encryption, and anonymization of data. This thesis explores the connection between Blockchain and GDPR. In particular, the thesis presents a decision framework that can help developers and decision-makers in their design of the GDPR-compliant blockchain projects. The framework, presented in the thesis, is developed based on Systematic Literature Review of existing studies on the interplay of technology and regulation. The provided findings will help the Blockchain development teams to consider the GDPR-related factors when designing and implementing the blockchain projects.Item Arduino näitel põhinev riistvara programmeerimise valikkursus gümnaasiumiastmele(Tartu Ülikool, 2020) Elberg, Paul; Peets, Alo, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutIn the course of studies for this Master Thesis “An Arduino-Based High School Supplementary Course in Hardware Programming” was completed. This supplementary course has been aimed to teach hardware programming to persons with no prior experience in electronics or programming. The chapters of the supplementary course provide a practical introduction to hardware programming to promote sustained interest in the subject. A key concern throughout this work has been to achieve a material suitable to the entry level abilities; the chapters have been tested on pupils younger than the age group of gymnasium students. The author provides an account of the problems encountered as well as the solutions provided thereto in the course of writing and testing the chapters of the supplementary course. Created materials are available in Estonian at https://courses.cs.ut.ee/t/nooredkoodi/Arduino/ArduinoItem Sample-efficient Online Learning in a Physical Environment(Tartu Ülikool, 2020) Liivak, Martin; Matiisen, Tambet, juhendaja; Paat, Rainer, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituutAutonomous driving has been seen as the next breakthrough in transportation. Autonomous vehicles employ a variety of sensors to understand their surroundings, for example multiple cameras, ultrasound sensors, and LiDARs. In this work, a much smaller scale radio-controlled cars, that only carry a central camera, are used. Their effectiveness as a test-bed for validating autonomous driving methods is evaluated. Multiple neural network architectures were proposed, among which a convolutional neural network was selected as the best candidate. The network was then trained using both supervised learning and online learning, the results of which were then compared. Experiments show that online learning in a physical environment, while costly, is a significant improvement over pure supervised learning. Additionally the radio-controlled cars proved to be a good comparative test-bed for evaluating model performance in an interactive physical environment.