e-Repositori UPF :: Grau en Enginyeria en Informàtica. Treballs de fi de grau

Enviaments recents

Scalable and privacy-preserving rectangular matrix multiplication with FHE by divide-and-conquer approach
(2025) Obradors Ambros , Ausiàs Pau
Data privacy and security are crucial in today’s world. While encryption secures data at rest and in transit, fully homomorphic encryption (FHE) offers a solution for secure processing by enabling direct computation on encrypted data. However, FHE’s practical application is limited by significant computational overhead, especially in fundamental operations like matrix multiplication. The CKKS scheme, designed for approximate arithmetic, faces challenges with matrix multiplication due to data packing constraints and the need for expensive homomorphic rotations. This dissertation addresses the efficient homomorphic multiplication of rectangular matrices within ciphertexts. It proposes and implements a divideand-conquer strategy, contrasting with the zero-padding approach. The research benchmarks this implementation across various dimensions, offering a framework for improved homomorphic rectangular matrix multiplication.
Estudio e implementacion de Kyber y ́Dilithium
(2025) Prat Bosch, Ariadna
Kyber and Dilithium are encryption and digital signature schemes selected by NIST as post-quantum cryptography standards. Both base their security on the complexity of some lattice problems over polynomial rings, which allows for efficient implementations. This project takes a pedagogical approach to understanding cryptographic systems in the post-quantum era. The goal is to comprehend and implement Kyber and Dilithium from scratch using the computational language SageMath. To optimize polynomial multiplication, we will use the Number Theoretic Transform based on the Chinese Remainder Theorem. During the project, implementation related challenges were encountered and documented, as well as the solutions applied to program a clean demonstration of Kyber and Dilithium.
Procedural Generation Applied to Metroidvania Level Design
(2025) Gasión Medina, Simón
The Metroidvania genre is a growing trend in the video game industry, with a wide range of acclaimed titles. It is characterized by the importance of good level design to correctly guide players in their progression. The task of designing a cohesive and engaging map structure, along with enough content to populate it, leads to infamously lengthy development cycles in both indie and major studios. This work aims to lay the groundwork for introducing procedural generation tools to the workflow of a level designer working in the development of a Metroidvania title, reducing the development time. To that end, a procedural generation algorithm for Metroidvania map structure and content is proposed and partially implemented into the Godot game engine following a set of requirements derived from a preliminary study into level design for Metroidvania video games.
AdvancedStatistics in rust
(2025) Aranich Solís, Marcel
Currently we live in the era of information, which we are generating at an unprecedented pace. But data by itself has no value unless we can extract information from it. Statistics is the branch of mathematics that allows us to analyze and interpret data; and as data becomes more prevalent, the need for statistics also does. Thanks to it we can make data driven decisions, instead of relying on arbitrary or biased choices. In this project we explain the process of the creation of Advanced Statistics, a Rust library aimed at performing most common statistical procedures, such as: computing statistics from theoretical distributions, performing computations on data and performing hypothesis tests.
Estrelles del futur: anàlisi i predicció de talent futbolístic
(2025) Cabot Agustin, Gerard
Aquest treball presenta "Estrelles del Futur", una aplicació web desenvolupada per a l'anàlisi i predicció de talent en futbolistes joves. Mitjançant una arquitectura client-servidor i utilitzant les dades obertes de StatsBomb, el projecte ofereix eines de visualització interactives i un sistema de scouting basat en aprenentatge automàtic. La metodologia principal implementa un model per estimar el potencial màxim que un jugador assolirà en la seva carrera, una variable objectiva generada a través d'una mètrica heurística pròpia basada en la correlació de KPI. La contribució més innovadora és la capacitat que té l'usuari de construir i entrenar els seus propis models predictius, definint el talent segons la seva filosofia de joc. El model final, validat amb una metodologia temporal estricta, demostra una capacitat predictiva significativa i identifica qualitativament jugadors que posteriorment han assolit l'elit mundial, validant la viabilitat de l'eina.
Efficient reinforcement learning with transition-dependent LMDPs and entropy regularization
(2025) Garcies Ramon, Andreu
Reinforcement Learning provides a robust framework for sequential decisionmaking problems, where an agent learns to interact with an environment to maximize long-term reward. While traditional methods involving dynamic programming face performance challenges in large state spaces, the emergence of linearlysolvable Markov Decision Processes (LMDPs) offers a computationally efficient alternative by reformulating the control problem so that the Bellman equation becomes linear. This work explores a novel formulation of LMDPs in which the reward is defined as transition-dependent rather than the conventional statedependent form. This representation is more intuitive for humans who wish to define control problems similarly to how they are specified in the standard nonlinear MDPframework, whileretaining the computational advantages of the linear formulation. Building upon this foundation, we adapt Todorov’s embedding method to transform standard MDPs into this transition-dependent LMDP framework. We also establish a formal equivalence between the transition-dependent formulation and the conventional state-dependent LMDP, demonstrating that both approaches lead to the same optimal control strategies through well-defined transformations. To ensure a fair and consistent comparison between standard MDPs and their linearly solvable version, we integrate entropy regularization into the MDP framework and subsequently extend Todorov’s embedding to accommodate these entropy-regularized MDPs.
Comparing european election narratives on TikTok across EU member states
(2025) León Peidro, Adrià
This study analyses a dataset provided by AI Forensics, which includes TikTok videos related to the 2024 European Elections, focused on contemporary wars and abortion. AI Forensics is a non-profit organization focused on algorithm auditing through investigating opaque algorithms to uncover violations of digital rights, making it an ideal partner for the development of this thesis. The goal was to examine how TikTok videos retrieved in response to specific queries differ across countries, identifying national trends and biases related to sentiment and political discourse. The dataset covers the resulting videos of political search queries on TikTok related to the European elections held in 2024 across five European countries: Spain, France, Germany, Netherlands, and Poland, and includes search queries, order of appearance, video metadata, content creator URLs, video descriptions and the titles. To ensure comparability between the different countries the dataset only covers queries in English for all five countries. This dataset was enriched with various scripts that collected different metrics across the dataset. This enrichment consisted of including engagement data, video captions, political entity mentions and sentiment and subjectivity scores through natural language processing techniques. The analysis revealed several findings. Most of the TikTok videos analyzed contained language associated with positive emotions, particularly in France and the Netherlands. However, no significant differences were found in user subjectivity in the videos across countries. It was found that discussions around contemporary wars and abortion frequently referenced government institutions and political leaders rather than topics like religion or human rights. Political parties mentioned in the videos were predominantly from the right-wing spectrum in all countries except France. Interestingly, the most frequently mentioned political entities were not the most voted European parties, instead national political parties, specially the ones on the extreme of the spectrum, were the most mentioned. Finally, the data analyzed exhibited a correlation between the usage of negative associated language with better engagement results, including the number of likes, video position in the serach results and the amount of followers of the video uploader.
Business plan for custom-made device designs for complex structural heart diseases
(2025) Linacisoro, Telmo
Leaks associated with valvular prostheses and left atrial appendage occlusion devices are linked to an increased risk of cardiac morbidity. Currently, the devices available on the market are offered only in predefined sizes and shapes, without accounting for the unique characteristics of each individual's physiology, creating a significant gap that remains uncovered. To address this issue, the project Custom-Made Device Designs for Complex Structural Heart Diseases is being developed by Dr. Dabit Arzamendi and Dr. Abdel Hakim at Hospital de la Santa Creu i Sant Pau, in collaboration with Prof. Oscar Camara at Universitat Pompeu Fabra and iVascular as an industrial partner. The project employs a personalized medicine approach through predefined design envelopes. Imaging data from computed tomography scans is converted into 3D cardiac models, enabling accurate leak characterization. Multiple device iterations, made of an adjustable core and an anchoring system, are generated through parametric design optimization. These are followed by computational and benchtop testing, and then printed in nitinol. Finally, device implantation is performed percutaneously using catheter systems specifically designed for optimal axial alignment. This thesis presents an exhaustive business strategy to transfer the technology being developed into clinical and commercial applications. It evaluates market opportunities and defines the business model, strategic planning, regulatory pathways, intellectual property management, financial planning, go-to-market strategy, and patient's journey mapping. These ideas have been validated through interviews with domain experts (cardiologists, healthcare investors, startup founders, innovation managers, patent agents, and regulatory consultants, among others) and through the development of the incubator program InnoPau. The analysis demonstrates that custom-made devices are clinically beneficial, technically feasible, and economically viable. The European market alone represents an underserved niche with minimal competition, where strategic growth, both horizontal and vertical, can be achieved through international expansion and intellectual property extrapolation to cover several cardiac diseases. A phased regulatory plan, beginning with CE certification and followed by accelerated FDA approval in the United States, combined with a licensing spin-off or software as a service revenue models, provides controlled risk and financial sustainability, with projections forecasting a break-even point within five years of market entry.
Deployment of artificial intelligence algorithms for medical imaging analysis in a clinical environment
(2024) Tomàs Escudero, Norbert
The rapid evolution of Artificial Intelligence (AI) in medical imaging indicates a transformative shift in hospital environments. However, the full potential of these algorithms remains unrealized due to challenges in integrating them into healthcare facilities. Many hospital workflows, like image processing, and its downstream tasks, such as image segmentation, suffer from being slow, repetitive, and manual, aggravated by the limitations of black-box specialized vendor tools that obstruct innovation. This thesis presents a framework crafted specifically for seamlessly integrating AI algorithms into clinical settings. Designed to be extensible, user-friendly, with no software installation required, and vendor-neutral, the framework uses Orthanc, OHIF, and XNAT—three open-source web-based platforms incorporating these principles. Through the implementation and evaluation of two image segmentation algorithms, the framework’s efficacy and potential is showcased. Deployed and validated at Hospital de la Santa Creu i Sant Pau, this thesis opens the way for broader clinical integration of AI workflows, offering a promising avenue for future algorithm implementations and advancements.
A Conversational agent in PyramidApp: analysis of behaviors, and elements of design for a future integration
(2024) Gutiérrez Ferré, Aldric
This study aimed to evaluate the influence of a Generative Artificial Intelligence (GenAI) agent within the Computer Supported Collaborative Learning (CSCL) platform “PyramidApp”. Using a quasi-experimental within-subject design, the research analyzed the behaviors of the GenAI agent in the PyramidApp environment. The study examined 105 chat room conversations, both assisted and non-assisted by a masked GenAI agent. The findings revealed that the presence of the GenAI agent positively impacted quality aspects such as precision, reduced confusion, and enhanced knowledge building. Additionally, the study determined that the GenAI agent’s ratings correlated positively with students’ ratings, providing student-like feedback on submissions. Furthermore, the analysis indicated that a GenAI chatbot in PyramidApp influenced both individual group member performance, and overall group behavior in online activities. The results offer valuable insights into effective strategies for integrating GenAI into CSCL environments, helping developers consider key design elements for future GenAI integrations.
Estudio del malestar producido por el desplazamiento de la cámara en vídeos 360 y en realidad virtual
(2024) Wei, Yunchao
Existen estudios que demuestran que la navegación en entornos de realidad virtual y la visualización de videos 360 pueden generar malestar (cybersickness) entre los participantes. Según la literatura, la altura de la cámara y la posición de los ojos puede producir este efecto. La presencia de un cuerpo virtual que represente a los participantes en este tipo de entornos puede mitigar este malestar. En este trabajo se presenta un experimento en el que el participante se mueve por un entorno de realidad virtual con diferentes configuraciones en la altura de la cámara y mostrando un cuerpo virtual. El objetivo principal es comprobar si la presencia del cuerpo virtual reduce el malestar provocado por la diferencia entre la altura real del participante y la altura de la cámara virtual. Con este experimento, comprobamos que la experiencia de la visualización de tener un cuerpo virtual era superior a la de no tenerlo cuando la altura de la cámara era superior a la de los propios participantes.
Enginear, aplicación educativa científica inspirada en Duolingo
(2024) Valenzuela Pinazo, Mario
Enginear es una aplicación educativa diseñada para estudiantes de niveles avanzados, como bachillerato o universitarios, enfocada en matemáticas y ciencias generales. Su propósito es proporcionar un medio intuitivo y efectivo para aprender y reforzar conocimientos previamente adquiridos, inspirada en el éxito de plataformas como Duolingo y otras aplicaciones matemáticas. Enginear se centra en la simplicidad y la facilidad de uso, promoviendo la práctica repetitiva como método de aprendizaje. El proceso de desarrollo de Enginear incluyó la realización de encuestas para comprender las necesidades y preferencias de los usuarios. Además, se exploró la integración de Inteligencia Artificial, como la API de Google Generative AI o similares al entorno de la IA de Google Bard (Gemini), para mejorar las funcionalidades de la aplicación. El resultado es una herramienta educativa versátil y adaptable que se ajusta a las necesidades específicas de los estudiantes avanzados en matemáticas y ciencias generales.
Improving water consumption management in Barcelona through data quality enhancement and prediction models
(2024) Ruiz Macià, Edith
The project aims to improve the quality of water consumption data of Barcelona for better management of resources. Aigües de Barcelona has allowed me to continue using their data as an extension of the project done in the AB Data Challenge. I have enhanced the datasets by applying data cleaning and adding new features combining meteorological datasets. I stored all anomalies detected for analysis and classification into different categories. Anomalies can have different sources and serve as potential indicators of a water misuse, leak or errors of the data collection system. I have predicted the missing values caused by the data cleaning and tested the accuracy of five prediction models in order to select the most accurate one for each specific dataset. Finally, I designed a solution in the form of a web application, defined the necessary requirements based on a needs analysis and developed a first version prototype. The app allows users to learn about the insights of datasets with a similar structure to those of Aigües de Barcelona.
Parallel strategies for best-first generalized planning
(2024) Fernández Alburquerque, Alejandro
In recent years, Artificial Intelligence (AI) has become the big trend in computer science, and many related areas are seeing renewed interest. One of these areas is Generalized Planning (GP), which studies the automated synthesis of algorithmicl ike solutions capable of solving multiple classical planning instances. Tightly coupled with the success of AI, there has been a steady increase in computational power thanks to multi-core CPUs and GPUs. This has encouraged active research in parallel programming, which is needed to use the full potential of current hardware. In this work, we explore parallelization strategies for tree-search algorithms. Furthermore, we propose one algorithm to parallelize Best-First Generalized Planning (BFGP), a heuristic search approach to GP. We show that our algorithm can scale linearly with the number of processors, but we also discuss some cases in which pathological behavior can cause performance degradation.
Neural based machine translation from Catalan Sign Language glosses to written catalan
(2024) Chabaud, Luka
This thesis had two objectives: the first one was to experiment with an alternative method to statistical machine translation, which had already been researched for Catalan Sign Language (LSC), using state-of-the-art Neural Networks. More precisely, this work focuses on the yet unexplored translation task from LSC glosses to written Catalan, finding the best possible neural architecture by searching for the best hyperparameter values. The second objective was to apply existing data augmentation techniques for Sign Languages, also developing specific LSC rules, based on its grammar, in order to create a parallel corpus of synthetic data. We then used this augmented data to enhance the training of the models, achieving significant improvements over a baseline without the need of new labeled data.
Gestió de la biodiversitat i patrimoni de l’INCASÒL: creació d’un mercat de CO2 basat en Blockchain..
(2023-12-04) Pedrós Villorbina, Roger
El canvi climàtic és un problema global causat per l'increment de les emissions de gasos d'efecte hivernacle, que estan provocant l'augment de la temperatura global amb conseqüències negatives devastadores. Per aquest motiu institucions de tota índole s’estan mobilitzant dins les seves competències per buscar solucions que mitiguin els problemes seguint els objectius de desenvolupament sostenible proposats per les Nacions Unides. És en aquest punt en què l'Institut Català del Sòl (INCASÒL) s'interessa en els mercats de CO₂, com una manera per millorar la gestió sobre el patrimoni i la biodiversitat, fent ús dels seus actius naturals, com ara terres forestals, agrícoles i de pastura. Dins d'aquesta anàlisi s'inclou una mirada sobre el panorama internacional dels mercats de CO₂, plataformes i empreses associades, així com la realització d’un estudi cas per a INCASÒL i el desenvolupament d’un concepte de prova basat en blockchain. Es proposa una solució que involucra l'ús de la tecnologia blockchain per a la gestió dels boscos, incloent-hi el seu seguiment, absorció en tCO₂ i salut, així com la creació de tokens que demostrin la quantitat de CO₂ retinguda en un bosc, i permetin la seva comercialització. La comercialització dels tokens de CO₂ retinguts per un bosc també té el potencial d'incentivar la gestió responsable dels boscos i la reducció de les emissions de CO₂, al mateix temps que ofereix una forma transparent de demostrar l'impacte ambiental. L'ús de la tecnologia blockchain és particularment adequat en un context en el qual intervenen moltes parts interessades.
CarbonSavings: a blockchain-powered move-to-earn app for citizen’s awareness on greenhouse gas emissions
(2023) Xiang Vico, Hong-ming
Move-to-earn (M2E) is a blockchain-based business model that promotes physical activity through a mobile phone application in exchange for earning digital assets rewards. In this project, the sustainable potential of the model is explored to reduce the impact of transportation emissions in cities, which contributes to increasing the global climate change problem. Our solution proposes the development of a M2E application, called CarbonSavings, embedding a carbon footprint algorithm. Its aim is to reduce greenhouse gas (GHG) emissions produced by individuals and to promote sustainable mobility behaviors while educating users about their carbon footprint impact. The project begins with a market exploration to analyze existing applications and continues with literature research to identify state-of-the-art methods for calculating GHG emissions. The outcomes are used to design a proof-of-concept application with native cryptocurrency rewards, non-fungible tokens (NFTs), or leaderboards as gaming features. The M2E model is implemented to motivate users to use sustainable transportation such as walking or biking over high-emission alternatives like driving private vehicles. Inside the application, users can set their personal goals, track their progress and rewards, or follow their mitigation footprint, by means of the designed algorithm, among other functionalities. The proof-of-concept is developed using Android native development with Kotlin programming language and Jetpack Compose. The blockchain layer is built on the Ethereum blockchain, using Solidity Smart Contracts for business logic. Moreover, a complementary landing page with an open-source carbon footprint calculator is deployed to raise user awareness and promote the application's usage. Finally, Barcelona City has been selected to scope the proof-of-concept, serving as a teaser to showcase the project’s capabilities targeting real users. The results of this project have the potential to contribute to Sustainable Development Goals #11 and #13 from the United Nations 2030 Agenda and empower individuals to take action against climate change, aligned with the Planetary Wellbeing program of the UPF.
Corpus-wise extraction of syntactic structures for data-to-text generators
(2023-11-20) Ricci Comella, Josep
Data-to-text NLG consists in converting structured data as found for instance in the Semantic Web (e.g. DBpedia, Wikidata, etc.) into well-formed text in the target language(s). The input to the system is typically a series of triples “Property(Subject, Object)” that encode a wide variety of knowledge types. There are currently different approaches to converting structured data into text. Lately, the most popular ones are neural machine-learning-based techniques, which lack energy efficiency and content accuracy (hallucinations, omissions). On the other hand, symbolic (rule-based) approaches that are very accurate and thus produce reliable output texts when fed with known Properties, are currently less used, in particular, due to their limitations with new (unseen) Properties when corresponding semantic representations are missed. The project aims at addressing this common problem of symbolic approaches by developing a system that converts a given triple into a plausible syntactic structure, to be further used for the creation of the required delexicalized predicate-argument structures. The main tasks of the project are retrieving a text semantically equivalent to the triple from a large corpus connected with a structured knowledge base, and transforming them into generic syntactic representations.
DragonIce: developing a full-body interaction application to promote and assess prosocial behavior in children
(2023-10-17) Delgado Rueda, Óscar
New technologies, such as full-body Mixed Reality systems, and the affinity that children have for it, have made it possible to develop more engaging and dynamic learning experiences for children with and without autism. Therefore, the objective of this work is to develop a full-body interaction application to promote (first level of the experience) and evaluate (second level of the experience) prosocial behavior in children with and without autism. Through ‘synchronous’ movements from the first level of the experience, prosociality is achieved. About the objectives, more specifically, the first subgoal is to make code improvements to achieve a more stable version and visual improvements to enhance the player's interaction in the first level. These code improvements consist in creating a better structure, distributing the code in different scripts and creating methods that manage the scene efficiently. The second sub-objective is to develop the second level of the game through cooperative design with the multidisciplinary team of the Full-Body Interaction Lab (FuBIntLab) and through co-design sessions with children between 8-10 years of age. The observations made during the usability study show how the system allows children to play satisfactorily in both experiences (first and second level). Therefore, the application can be used by the FuBIntLab to promote and evaluate prosocial behavior in children.
Custom Vulkan Engine to Render Black Holes in Real Time Using Ray-Marching
(2023-10-03) Meseguer Orrit, Roger
This report deals with the Narwhal Engine: an interactive Vulkan-based graphics engine we developed that renders black holes in real time. Our engine is focused on rendering Schwarzschild and Kerr black holes, which are the simplest types without any charge. We adapted an already existing Unity black hole renderer and ported it over to Vulkan. Our implementation improves on the original and manages to render a frame over 15 times faster. It also adds interactivity to the original implementation as it allows users to modify the black hole and camera parameters and have those changes instantly reflected. The visualisation shows different elements of the black hole such as the accretion disk, the Einstein ring or the photon sphere.

Navegar

Examinar

Enviaments recents