Master in Artificial Intelligence / Μεταπτυχιακό στην Τεχνητή Νοημοσύνη

In this thesis, we explore the application of computer vision techniques for autonomous driving in a video game setting. We focus on the task of predicting the optimal motion the vehicle in the game using only visual information from the screen. To this end, we propose a dataset specifically designed for this task, and evaluate a variety of state-of-the-art classification models on it. Our results show that these models can successfully predict the motion of other vehicles with high accuracy, and are able to run in real-time on the game engine. Furthermore, we also present an analysis of the results and discuss the limitations of our approach. As a future work, we propose to explore other computer vision techniques such as object detection and semantic segmentation to improve the performance of the models and to incorporate more information from the game environment.

In this Master Thesis we aimed to develop a model that will enable the identification of the health condition of vines from images taken by a UAV, with the utilization of image processing and Machine Learning. To achieve this task, we prepared a training dataset containing images of fungal-infected and healthy vine leaves and we tested several Machine Learning and a Convolutional Neural Network models to determine the optimal solution in terms of health condition classification accuracy. We then applied the selected classifier on a set of UAV images captured from five controlled vineyards, obtaining the model’s predictions on the vines’ health condition. The developed model aspires to provide a solution for vine growers who would be able to have a better control of their plants status, allowing quicker identification and cure of potential threats from diseases and safeguarding their vineyard’s yield.

Fast and Accurate SED modelling using Machine Learning is a supervised learning project aiming at exploring efficient representations of spectral energy distribution (SED) fitting codes by leveraging the immense capabilities of machine learning algorithms that have emerged during the last decade. The focus will be on the supervised learning process of two alternative (surrogate) models, a neural network (ANN) and an ensemble regressor (HGB), based on a sufficiently large dataset of simulated galaxy spectra, generated with the CYGNUS (CYprus models for Galaxies and their NUclear Spectra) models. The project is structured into the following chapters: Chapter 1 provides a background of the problem surrounding the effective retrieval of the fundamental physical properties of galaxies by means of studying and modelling the vast and complex cosmological and astrophysical data selected by an ever-growing number of sources. Chapter 2 offers an introduction into the principles of the proposed machine learning models (ANN, HGB), their architecture, and their fundamental functions. The objective is to familiarize the reader with the terminology and the data processing methods referred to at a later stage of the project. Chapter 3 gives a description of surrogate models, along with the justification of the need to employ such models in the case of non-linear models with intractable parameter distributions. As a matter of fact, the project follows the supervised learning of selected SED surrogate models to be eventually used in the inverse process of fast and accurate retrieval of physical parameters. Chapter 4 is dedicated to the specifications and the implementation process of the end-to-end MARGE (Machine learning Algorithm for Radiative transfer of Generated Exoplanets) package. A distinctive part of MARGE is the deep learning functionality (neural network) used to train the SED surrogate model. Chapter 5 outlines the complete training pipeline of this project, starting from the justification of the CYGNUS model. The training data for the proposed machine learning models are made up of one million CYGNUS model simulations of the combined form (input model parameters – output spectra). The most important hyperparameters of the predictive models (ANN, HGB) are optimized on a subset of the data and optimal models are trained on the entire dataset. Finally, regression analysis provides an insight into the appropriateness of both machine learning models, as well as their comparative performance, by means of aggregate measures and individual image plots. Appendix 1 specifies the list of code - and data adjustments that need to be applied to the source files for a functional implementation of the MARGE package. Appendix 2 provides the complete coefficients of determination (R2-scores) achieved in testing by both predictive models.

Recent advancements in artificial intelligence have been significantly inspired by the field of educational data mining, or EDM. According to the needs of the students and instructors, new potential and Possibilities for the advancement of technology-enhanced learning systems have been created via a range of research and put into practice. The use of modern techniques and application strategies by the EDM is crucial in improving the learning environment. The EDM is essential in building the educational setting for scholars along with ameliorating the working conditions for instructors and giving them appropriate information regarding the necessities of the students by looking at both the educational environment and machine learning approaches. In order to reach that goal, the main focus of this thesis will be on the socioeconomic elements that may influence how various institutions, communities, and social backgrounds function at various educational levels. To evaluate student and teacher data and anticipate the critical factors that affect educational effectiveness, a variety of machine learning techniques are applied. Specialized dimensionality reduction techniques are utilized to display the data from diverse locations. Then, regions and countries are categorized based on their common educational qualities using K-Means and Spectral clustering. Two datasets have been used for this analysis. Clustering Results Based on the provided information from the first dataset, a clustering analysis was conducted across various characteristics related to lower education levels. Cluster 1 consists of countries with high working time, teaching time, administrative tasks, additional duties, course planning, amount of lessons, volume of marking, negative impact on mental health, stress, physical health, salary satisfaction, and preparedness for students with disabilities. These countries include the United Kingdom, Sweden, Iceland, Norway, the Netherlands, Croatia, the Czech Republic, and Hungary. Cluster 2 comprises countries with low working time, teaching time, administrative tasks, additional duties, course planning, amount of lessons, volume of marking, negative impact on mental health, stress, physical health, salary satisfaction, and preparedness for students with disabilities. Italy, Cyprus, Romania, Finland, and Montenegro are part of this cluster. Cluster 3 includes the remaining countries, Belgium, Bulgaria, Spain, Czech Republic, Estonia, Croatia, Austria, North Macedonia and Slovenia, which are considered to have medium characteristics across the analyzed dimensions. The second dataset has information for both lower and upper education level. Based on the information provided for the lower education level dataset, here is a summary of the clusters created: Students: Cluster 1 (High Performance): This cluster includes countries like Spain, Belgium, Sweden, Italy, Hungary, and Ireland. These countries exhibit high student performance in terms of academic achievements, participation rates, and overall educational outcomes. Cluster 2 (Moderate Performance): This cluster comprises the United Kingdom and the Netherlands. These countries demonstrate moderate student performance, with relatively average academic achievements and participation rates compared to other countries in the dataset. Cluster 3 (Mixed Performance): This cluster consists of the remaining countries. They exhibit a mix of student performance levels, with some countries showing above-average performance while others have lower performance in terms of academic achievements and participation rates. Teachers: Cluster 1 (Experienced and Well-Qualified): This cluster includes countries such as Latvia, Lithuania, Malta, and Finland. These countries have highly experienced and well-qualified teachers who undergo continuous professional development, resulting in high-quality teaching and effective classroom practices. Cluster 2 (Diverse and Competent): This cluster comprises countries like Iceland, Germany, the United Kingdom, and the Netherlands. These countries have a diverse teacher workforce with a range of competencies and expertise. They prioritize teacher collaboration, innovation, and continuous improvement in their education systems. Cluster 3 (Varied Teacher Profiles): This cluster encompasses the remaining countries. They exhibit a mix of different teacher profiles, including varying levels of experience, qualifications, and professional development opportunities. The characteristics and quality of their teacher workforce vary across these countries.

In a world where personal data can easily be extracted from the simplest of online transactions or actions in general, the need for sufficient data protection protocols, laws and regulations has never been more dire. In past years, the internet was at its infancy which led countries to adapt or establish laws that did not really protect the personal data of its residents from internet attacks, scams etc. The rapid growth of the internet has led countries to expand on their previous laws and regulations or to create new ones. The differences of these laws and regulations is the main scope of this thesis. This thesis will begin with an overview of the previous law in Europe (Data Protection Directive) and will continue with an overview of the law that replaced the DPD, the General Data Protection Regulation while providing a comparison between the two laws. The thesis then will move on with the different state laws in the United States. In both the overview of the European laws and the one of the United States, the format will be the same with minimal to no divergence. There will be some introductory remarks about the law such as the date it was signed into law and then we will list the rights of the data subjects who are subject to this law. We will finish each law by listing the responsibilities and obligations of the controller towards the data subjects. The final part of this thesis will concern the differences between European and US data protection laws and regulations. These differences range from cultural differences regarding who these laws favor most to whether these laws are consent based or not to technical differences such as range of application, lawful basis etc.

This dissertation delves into the complex area of Low Earth Orbit (LEO) satellites, concentrating on their construction, administration, and the issues faced by air attenuation. The survey of literature looks into the history and kinds of satellites, with a close look at important LEO spacecraft such as Micius, Tiangong-2 Space Lab, and others. The research thoroughly addresses LEO satellite design, constellation management, and the function of Optical Ground Stations (OGS). A substantial part of the thesis is devoted to understanding atmospheric attenuation, which includes processes such as beam spreading, absorption, scattering (including MIE and Rayleigh scattering), turbulence, and scintillation. The link budget is examined with the purpose of evaluating the communication characteristics of LEO satellites, and the use of Quantum Cryptography adds a new dimension to secure satellite communication. The methodology section describes the study technique, which includes satellite monitoring, artificial intelligence for classifying passages, and attenuation scripts. The findings are presented and debated, offering insights into the efficacy of the approaches used. The thesis finishes with a complete assessment of the study's findings, stressing major themes and recommending future research directions. The bibliography features a wide variety of sources, while the appendices provide more extensive information, scripts, and supporting materials. This thesis promotes the knowledge of LEO satellites, atmospheric obstacles, and the possible integration of Quantum Cryptography for secure satellite communication.

The issue at hand is the comprehension of race bias and gender selection in the data analysis domain of Technical Intelligence, which is the subject of the original problem analysis in this thesis. We identify the flaws that contribute to bias in Technical Intelligence and assess the repercussions of bias on systems and society. Opportunity and challenge are both apparent and identifiable in the realm of research as a result of technical intelligence

In the era of digital content consumption where the modern webpages are transformed into complex multimedia platforms, often incorporating images and videos, the ability to automatically classify images has become increasingly vital. To monetize content, many webpages embed advertisements in the form of images, blending them with actual content to increase click-through rates. This project aims to tackle the challenge of distinguishing these web advertisements from regular content using image classification techniques powered by Machine Learning (ML) algorithms and Deep Neural Network (DNN) models. The methodology involves, the pre-processing and normalizing of a large dataset of images and advertisements extracted from various websites across different geographic locations. This dataset provides the framework for the subsequent image classification tasks. Secondly, the project investigates the use of diverse machine learning algorithms specifically designed for image classification to differentiate between webpage content and advertisements. Through extensive experimentation, the effectiveness and robustness of the classifier are assessed, offering insights into its potential applications in content moderation, digital marketing analytics, and user experience enhancement.

This thesis investigates the use of machine learning algorithms for wildfire prediction, aiming to improve existing fire risk assessment methods and support preventive actions. Traditional fire prediction models, which often depend solely on meteorological indices, fail to capture complex relationships between environmental, geographic, and human factors that influence wildfire occurrences. This research reviews current prediction approaches, identifies relevant datasets, and evaluates various machine learning algorithms to improve predictive accuracy. Data was collected from five meteorological stations in Cyprus, covering variables such as temperature, humidity, wind speed, and atmospheric pressure. The study included extensive data pre-processing and feature engineering to optimize the models’ performance. The final model, based on Extreme Gradient Boosting demonstrated the ability of machine learning algorithms to produce reliable fire predictions and adapt to varying environmental conditions. Challenges encountered in this study included limited data availability from forest prone areas and the unpredictable nature of human-caused fires, which are not easily modelled with environmental data alone. To improve the model’s performance, future work could focus on collecting more comprehensive data and including more relevant variables such as precipitation. Overall, this work highlights the promise of machine learning techniques in advancing wildfire prediction, ultimately contributing to a more effective fire prevention strategy.

This Master's Thesis focuses on the development of a sophisticated Ground Control Station (GCS) software designed to enhance the management and operational capabilities of unmanned aerial vehicles (UAVs) during computer vision missions. The GCS software integrates advanced technological frameworks to enable real-time monitoring, control, and coordination of drone swarms, providing a comprehensive suite of tools for efficient mission execution. The software architecture incorporates modular components that ensure flexibility and scalability, allowing for the seamless integration of future technologies and adaptation to various mission requirements. Key features include multi-drone management, real-time 3D visualization, geo-fencing, user-friendly interfaces, data logging, predictive maintenance, and a unique mission management system that functions akin to an app store for UAV operations. This thesis demonstrates the implementation and testing of the GCS with custom drone hardware, illustrating its efficacy in improving operational standards, safety, and data management in UAV deployments.

This project focuses on the design and implementation of a mobile application that leverages Artificial Intelligence (AI) techniques, specifically supervised learning, to estimate property valuations. The application allows users to input key building parameters, such as location, size, and features, to predict the estimated cost of a property intended for sale. Utilizing an XGBoost regressor, the model analyzes the provided inputs and applies advanced machine learning algorithms to deliver accurate predictions. The system has undergone extensive testing, including unit, integration, and system testing, with a focus on performance, data handling, and user experience. Key metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and R-squared have been evaluated to validate model accuracy and generalization on unseen data. The application offers an innovative, user-friendly, and efficient tool for property valuation, providing real estate stakeholders with reliable insights to inform decision-making. This work presents significant improvements over traditional valuation methods and demonstrates how AI can streamline processes in the real estate market.

This thesis explores autonomous navigation of Unmanned Aerial Vehicles (UAVs) and a land rover in GPS-denied environments through the integration of SLAM and reinforcement learning techniques. RTAB-Map SLAM is used to create real-time 3D maps, explore lite for autonomous exploration and navigation algorithms for safely navigating the UAV to the target location. The UAV also employs Deep Q-Network (DQN) for discrete action navigation. Additionally, a Twin Delayed Deep Deterministic Policy Gradient (TD3) network is implemented for the land rover, allowing continuous action control. Both systems utilize RGB-D cameras and LiDAR sensors for environmental perception, obstacle avoidance and localization. The performance of these systems is evaluated in simulation.

This research explores the optimization of Evolutionary Algorithms (EAs) using Unity Job System, a framework enabling task parallelization across multiple CPU threads. EAs, widely applied in optimization and artificial intelligence, often face performance bottlenecks due to their computational intensity, especially in real-time or resource-constrained environments such as game development. By leveraging Unity's parallelization capabilities, the study aims to enhance the performance of EAs while addressing their computational challenges. A custom Unity application was developed to compare the performance of EA components - including fitness functions, selection, crossover, and mutation - under single-threaded and multi-threaded implementations. Tests were conducted on functions with varying computational complexities, including Big-O synthetic functions and benchmark functions such as Rastrigin and Rosenbrock. Additional evaluations included the Traveling Salesman Problem and the efficiency of core EA operations. The results demonstrate the potential for significant performance improvements when using Unity's Parallel Job System, particularly for tasks involving extensive data processing. However, limitations such as platform constraints, data preparation overhead, and restricted debugging capabilities are noted. This study provides practical insights for developers and researchers aiming to integrate parallel computing into real-time applications and highlights the potential of Unity as a versatile platform for advancing artificial intelligence in game development.

This thesis explores the Voronoi game, a strategic competition model that combines principles from game theory and computational geometry. In the Voronoi game, players are placed within a bounded space, and territory is determined by proximity, forming regions called Voronoi cells. The objective is to capture the maximum area within these cells, leading to complex strategic behaviors. This study investigates the core mechanics, variations, and applications of the Voronoi game, as well as its implications for broader fields such as network design, facility location, and competitive spatial strategies. Through analytical and computational approaches, we characterize equilibria, evaluate player strategies, and explore multi-round dynamics, providing a comprehensive understanding of the games mathematical and practical relevance.