Search

Thyroid nodules are very common, 5–15% of which are malignant. Despite the low mortality rate of well-differentiated thyroid cancer, some variants may behave aggressively, making nodule differentiation mandatory. Ultrasound and fine-needle aspiration biopsy are simple, safe, cost-effective and accurate diagnostic tools, but have some potential limits. Recently, machine learning (ML) approaches have been successfully applied to healthcare datasets to predict the outcomes of surgical procedures. The aim of this work is the application of ML to predict tumor histology (HIS), aggressiveness and post-surgical complications in thyroid patients. This retrospective study was conducted at the ENT Division of Eastern Piedmont University, Novara (Italy), and reported data about 1218 patients who underwent surgery between January 2006 and December 2018. For each patient, general information, HIS and outcomes are reported. For each prediction task, we trained ML models on pre-surgery features alone as well as on both pre- and post-surgery data. The ML pipeline included data cleaning, oversampling to deal with unbalanced datasets and exploration of hyper-parameter space for random forest models, testing their stability and ranking feature importance. The main results are (i) the construction of a rich, hand-curated, open dataset including pre- and post-surgery features (ii) the development of accurate yet explainable ML models. Results highlight pre-screening as the most important feature to predict HIS and aggressiveness, and that, in our population, having an out-of-range (Low) fT3 dosage at pre-operative examination is strongly associated with a higher aggressiveness of the disease. Our work shows how ML models can find patterns in thyroid patient data and could support clinicians to refine diagnostic tools and improve their accuracy., Peer Reviewed, Postprint (published version)

Desarrollo de un programa de augmentación de datos para sistemas de processado del habla, Desenvolupament de un programari de augment de dades per sistemes de processament de parla

The Recursive InterNetwork Architecture (RINA) was presented as a clean-slate architecture to solve the current shortcomings of the current Internet architecture. In parallel, the Routenet Graph Neural Network (GNN) is being developed and is capable of predicting networks metrics. Given the Quality of Service (QoS) characteristics of RINA, and specifically of one of its implementations called IRATI, being able to predict the status of the network should allow us to take better decisions in order to assure its QoS requirements. During this thesis, we will analyze if Routenet is a good fit for making predictions in an IRATI environment. We will also develop a proof-of-concept prototype to validate the hypothesis that Routenet can help RINA improve its QoS, as well as demonstrating its capacity to do so from a practical example. The results obtained show that, while more effort needs to be directed to have better predictions, Routenet could predict RINA network metrics accurately and that an integration between the two elements can be very beneficial to the QoS of IRATI when better predictions are achieved. Most importantly, the work in this thesis explores a new field, solves many of the initial steps in the integration which is a crucial step towards opening several new lines research.

© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works., This paper presents a data-driven fault detection method combining structural analysis (SA) and machine learning data-driven algorithms. Given a graphic (or textual) system description and the available inputs/outputs measurements, the structure of analytical redundancy relations (ARRs) between some inputs and outputs can be determined with the aid of the SA of the system. Then, using a machine learning data-driven approach applied to historical data, analytical relations between inputs and outputs can be obtained. Thereby, instead of finding ARRs from physical mathematical model, ARRs are obtained combining SA and data-driven approaches. In this paper, the adaptive network fuzzy inference system (ANFIS) data-driven approach is used to implement the diagnosis system. Once the ANFIS model has been identified, it is reformulated in linear parameter varying (LPV) form. Then, a fault detection scheme based on a LPV Kalman filter and pole placement method is developed. A well-known case study based on a four-tanks system is used for illustrative purposes., This work has been co-financed by the Spanish Research Agency (AEI) through the projects SaCoAV (ref. MINECO PID2020-114244RB-I00) and L-BEST (PID2020-115905RB-C2I)., Peer Reviewed, Postprint (author's final draft)

This project is based on the idea of further improving the implementation of MobilitApp. It is an Android application designed and developed by students of the ETSETB with the collaboration of the ATM of Barcelona. The project started in 2014 with one main objective: to optimise urban mobility. To do so, it uses a Machine Learning model based on collecting data from an accelerometer, a magnetometer and a gyroscope, to train it and be able to detect the transport used by its users. More specifically, it uses a Deep Learning model designed by members of the SISCOM research group of the UPC. The thesis starts at a point where MobilitApp is already able to generate User Info. These are the lines of data with information on how, when and where its users move. All this data is collected in files as a database and automatically the need arises to be able to process all this data in order to reach conclusions at project level. Basically, the aim of this thesis is to create an application capable of reading a database, selecting filters of gender, age, route start dates, route end dates, origin zones, destination zones and transport methods used, and with them create graphs representing all the information. With this, any member involved in the project can draw conclusions from the User Info obtained in a much more agile and user-friendly way., Este proyecto se basa en la idea de seguir mejorando la implementación de MobilitApp. Es una aplicación Android diseñada y desarrollada por estudiantes de la ETSETB junto con la colaboración de la ATM de Barcelona. El proyecto se inició en 2014 con un objetivo principal: optimizar la movilidad urbana. Para ello, se utiliza un modelo de Machine Learning basado en recoger datos de un acelerómetro, un magnetómetro y un giroscopio, para entrenarlo y ser capaz de detectar el transporte utilizado por sus usuarios. Más concretamente, se utiliza un modelo de Deep Learning diseñado por miembros del grupo de investigación SISCOM de la misma UPC. La tesis se inicia en un punto en el que MobilitApp ya es capaz de generar User Info. Estos son las líneas de datos con la información de como, cuando y donde se mueven sus usuarios. Todos estos datos se recogen en archivos a modo de base de datos y automáticamente surge la necesidad de poder procesar todos estos para poder llagar a conclusiones a nivel de proyecto. Básicamente, el objetivo de este trabajo es crear una aplicación capaz de leer una base de datos, seleccionar filtros de género, edad, fechas de inicio de ruta, fechas de final de ruta, zonas de origen, zonas de destino y métodos de transporte utilizados, y con ellos crear gráficos representando toda la información. Con esto, se consigue que cualquier miembro involucrado en el proyecto pueda interpretar grupos de User Info almacenados de una manera mucho más ágil y user-friendly., Aquest projecte es basa en la idea de seguir millorant la implementació de MobilitApp. És una aplicació Android dissenyada y desenvolupada per estudiants de la ETSETB junt amb la col·laboració de la ATM de Barcelona. El projecte es va iniciar al 2014 amb un objectiu principal: optimitzar la mobilitat urbana. Per fer-ho, s?utilitza un model de Machine Learning basat en recollir dades d?un acceleròmetre, un magnetòmetre i un giroscopi, per entrenar- lo i ser capaç de detectar el transport utilitzat pels seus usuaris. Més concretament, s?utilitza un model de Deep Learning dissenyat per els membres del grup d?investigació SISCOM de la mateixa UPC. La tesis s?inicia en un punt en el que MobilitApp ja és capaç de generar User Info. Són les línies de dades amb la informació de com, quan i on es mouen els seus usuaris. Totes aquestes dades es recullen en arxius a mode de base de dades i automàticament sorgeix la necessitat de poder processar totes aquestes per poder arribar a conclusions a nivell de projecte. Bàsicament, l?objectiu d?aquest treball és crear una aplicació capaç de llegir una base de dades i seleccionar filtres de gènere, edat, dates d?inici de ruta, dates de final de ruta, zones origen, zones destí i mètodes de transport utilitzats, i amb ells crear gràfics representant tota la informació recaptada. Amb això, s?aconsegueix que qualsevol membre involucrat en el projecte pugui interpretar els grups de User Info emmagatzemats d?una manera molt més àgil i user-friendly.

During the last few years Artificial Intelligence has gained the public’s attention by being more present in our daily lives. Likewise, industries are taking on this kind of technology with the aim of increasing automation and optimizing processes, ultimately reducing costs. The present study intends to bring these new systems closer to the world of aircraft maintenance, and therefore studies the feasibility of developing an algorithm that can successfully determine whether the images of aircraft given as input show damages or not. Furthermore, the idea is that it should also be capable of differentiating between the most common damage types within the fleet of a commercial airline. As a result, the foundations will be set so that routine aircraft inspections can be cut down from multiple hours to several minutes. For this purpose, Machine Learning techniques will be employed and a code in R will be developed. To be more specific, it will be carried out by means of the TensorFlow and Keras libraries, and RStudio as the IDE. On the other hand, different architectures and ways of stacking layers will be studied, with the focus set on Convolutional Neural Networks (CNN). Several tests will be performed with them in order to see which ones can achieve a better performance and why. The outcome of this project is a model capable of distinguishing between healthy and damaged structures, but not between damage types. This is mainly due to the fact that the higher the number of classes among which the classification has to be done, the more images are required for training the AI. In this case, the database comprises 1268 pictures, which include different types of defects as well as undamaged structure. Once the final model is defined, a simple app, AirScan, is developed to enhance the user interaction. Therefore the first steps are taken for further development and testing stages, in which the system would be tested in a real work environment with completely new images.

This thesis discusses the growing trend of online platforms enabling users worldwide to buy and sell second-hand products, particularly digital devices. The pricing of these items depends on various factors, including their age, condition, and component quality. However, many individuals wish to sell their old electronic devices without the hassle of visiting a physical store for price estimates. This is where Artificial Intelligence (AI), particularly Machine Learning (ML), plays a pivotal role. AI has become increasingly significant across various business sectors, often operating quietly in the background. This thesis explores the idea of employing Machine Learning to provide people with estimated information and insights about their old electronic devices, eliminating the need for external consultations. To achieve this, a substantial amount of data is required. ML relies on models trained with data to make estimations about specific inputs. For instance, to extract various aspects of a digital device, a model trained on similar devices' data is necessary. These training datasets include a target variable, such as the device's price, which serves as the specific information to estimate about a second-hand device. The motivation behind this thesis is to use Machine Learning to develop and evaluate models that estimate information about second-hand market pricing information. To do so, a reliable data source is crucial. Back Market, an online marketplace specializing in refurbished electronic devices, is selected for this purpose.

ML pipelines, as key components of ML systems, shall be developed following quality assurance techniques. Unfortunately, it is often the case in which they present maintainability issues, due to the experimentatal nature of data collection and ML model construction. To address this problem, this work in progress proposes initial metrics to measure the presence of code smells in ML pipelines. These metrics reflect good software engineering practices for code quality of ML pipelines., This work was supported by the “Spanish Ministerio de Ciencia e Innovación” under project / funding scheme PID2020-117191RB-I00 / AEI/10.13039/501100011033., Peer Reviewed, Postprint (author's final draft)

In the ever-evolving world of animation, creating aesthetically pleasing and attractive environments is essential if you want to captivate the audience. Historically, the process of creating and animating scenery has mainly relied on the knowledge and imagination of human artists, requiring a lot of manual labor and many iterations. However, recent advances in computer vision and artificial intelligence have presented new possibilities for automating the generation of animation scenery. In this thesis I will focus on image cartoonization (giving to an image the style of a cartoon), a specific case of image translation. I will explore the state of the art of generative models, test some of them, create my own dataset of real and style images, train it on some technologies and compare the results obtained., En el mundo de la animación, que se encuentra en constante evolución, crear entornos estéticamente agradables y atractivos es esencial si se quiere cautivar al público. Históricamente, el proceso de creación y animación de la escenografía se ha basado principalmente en el conocimiento y la imaginación de los artistas humanos, requiriendo mucho trabajo manual y muchas iteraciones. Sin embargo, los recientes avances en visión por computador e inteligencia artificial han presentado nuevas posibilidades para automatizar la generación de escenarios de animación. En esta tesis me centraré en ‘Image Cartoonization’ (dar a una imagen el estilo de un ‘cartoon’, un dibujo animado), un caso concreto de ‘Image Translation’. Exploraré el estado del arte de los modelos generativos, pondré a prueba algunos de ellos, crearé mi propio ‘dataset’ (base de datos) de imágenes reales y de estilo, lo entrenaré en algunas tecnologías y compararé los resultados obtenidos., En el món de l'animació, que es troba en constant evolució, crear entorns estèticament agradables i atractius és essencial si es vol captivar el públic. Històricament, el procés de creació i animació d’escenaris s'ha basat principalment en el coneixement i la imaginació dels artistes humans, requerint molta feina manual i moltes iteracions. Tanmateix, els avenços recents en visió per computador i intel·ligència artificial han presentat noves possibilitats per automatitzar la generació d'escenaris d'animació. En aquesta tesi em centraré en ‘Image Cartoonization’ (donar a una imatge l'estil d'un ‘cartoon’, d’un dibuix animat), un cas concret d’Image Translation’. Exploraré l'estat de l'art dels models generatius, posaré a prova alguns d'ells, crearé el meu propi ‘dataset’ (base de dades) d'imatges reals i d'estil, l’entrenaré en algunes tecnologies i compararé els resultats obtinguts.

Detecció i classificació d’espècies de peixos a partir d’un conjunt de dades basat en imatges. S’han estudiat i aplicat alguns dels diferents algoritmes o models que s’utilitzen per a aquestes tasques per tal de veure quin d’ells proporciona resultats més precisos. A més, inclou una petita introducció al món del Machine Learning i els resultats amb les espècies de peixos detectades en la zona de l’observatori submarí Obsea., Detección y clasificación de especies de peces a partir de un conjunto de datos basados en imágenes. Se han estudiado y aplicado algunos de los diferentes algoritmos que se utilizan para estas tareas con tal de ver cual de ellos proporciona resultados más precisos. Además, incluye una pequeña introducción al mundo del Machine Learning y los resultados con las especies de peces detectadas en la zona del observatorio submarino Obsea., Fish species detection and classification from a dataset based on images. Some algorithms used for these tasks has been studied and applied in order to see which of them provides more accurate results. Furthermore, it includes a little introduction to Machine Learning world and the results with detected fish species in the area of the marine observatory Obsea.

In federated learning, distributed nodes train a local machine learning model and exchange it through a central aggregator. In real environments, these training nodes are heterogeneous in computing capacity and bandwidth, thus their specific characteristics influence the performance of the federated learning process. We propose for such situations the design of a federated learning server that is able to adapt dynamically to the heterogeneity of the training nodes. In experiments with real devices deployed in a wireless mesh network, we observed that the designed adaptive federated learning server successfully exploited the idle times of the fast nodes by assigning them larger training workloads, which led to a higher global model performance without increasing the training time., This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 871582 — NGIatlantic.eu and was partially supported by the Spanish Government under contracts PID2019-106774RB-C21, PCI2019-111851-2 (LeadingEdge CHIST-ERA), PCI2019-111850-2 (DiPET CHIST-ERA), and by national funds through FCT, Fundação para a Ciência e a Tecnologia, Portugal, under project UIDB/50021/2020. The work of C.-H. Liu was supported in part by the U.S. National Science Foundation (NSF) under Award CNS-2006453 and in part by Mississippi State University under Grant ORED 253551-060702. The work of L. Wei is supported in part by the U.S. National Science Foundation (#2006612 and #2150486)., Peer Reviewed, Postprint (author's final draft)

The problem of naval ship resistance prediction is a critical issue in naval architecture and ship design. It involves estimating the amount of resistance a ship will encounter as it moves through the water under various conditions. This prediction is crucial because it allows ship designers and engineers to determine the power requirements, fuel efficiency, and other characteristics. Computational Fluid Dynamics (CFD) is the most reliedupon prediction technique in hull-form design and analysis for new vessels. It is used to perform fluid flow analyses to help designers extract the highest possible performance in nearly every functional aspect of a naval vessel, including ship resistance. Modern CFD methods are certainly capable of delivering highly accurate results in this domain, but they are fundamentally based on numerical methods that solve the relative equations of mass, energy, and momentum conservation resulting in CPU-intensive mathematical procedures and this hinders their convenient use and integration within the conventional ship design loop. Machine Learning (ML) is rapidly becoming a core technology for scientific computing and its use in CFD is a wide research topic. In this work, we propose data-driven surrogate models, based on stateof-the-art ML algorithms, to speed up the exploration of several hull forms at an early design stage. Results showed that the employed approaches can be a valuable and more efficient solution for the initial qualitative analysis of several hull shapes in order to find the most efficient one for every scope of use., El problema de la predicci´o de la resist`encia dels vaixells ´es un tema cr´ıtic en l’arquitectura naval i el disseny de vaixells. Implica estimar la quantitat de resist`encia que trobar`a un vaixell quan es mou per l’aigua en diverses condicions. Aquesta predicci´o ´es crucial perqu`e permet als dissenyadors i enginyers de vaixells determinar els requisits d’energia, l’efici`encia del combustible i altres caracter´ıstiques. Computational Fluid Dynamics (CFD) ´es la t`ecnica de predicci´o m´es fiable en el disseny i l’an`alisi de la forma del casc per a vaixells nous. S’utilitza per realitzar an`alisis de flux de fluids per ajudar els dissenyadors a extreure el m`axim rendiment possible en gaireb´e tots els aspectes funcionals d’un vaixell naval, inclosa la resist`encia del vaixell. Els m`etodes CFD moderns s´on certament capa¸cos d’oferir resultats altament precisos en aquest domini, per`o es basen fonamentalment en m`etodes num`erics que resolen les equacions relatives de conservaci´o de massa, energia i moment, donant lloc a procediments matem`atics intensius en CPU i aix`o dificulta. el seu ´us convenient i la seva integraci´o dins del bucle de disseny de vaixells convencional. Machine Learning (ML) s’est`a convertint r`apidament en una tecnologia b`asica per a la inform`atica cient´ıfica i el seu ´us a CFD ´es un tema de recerca ampli. En aquest treball, proposem models substituts basats en dades, basats en algorismes ML d’´ultima generaci´o, per accelerar l’exploraci´o de diverses formes de casc en una fase inicial de disseny. Els resultats van mostrar que els enfocaments emprats poden ser una soluci´o valuosa i m´es eficient per a l’an`alisi qualitativa inicial de diverses formes de casc per tal de trobar la m´es eficient per a cada `ambit d’´us

LoRaWAN is a widely used solution in today’s Internet of Things (IoT) applications to connect remote sensor nodes to the Internet. At the same time, microcontroller-based sensor nodes with increased processing capacities are in- creasingly becoming smart nodes applied for performing machine learning tasks. We argue that LoRaWAN has important connectivity limitations to truly unleash the potential of these smart tiny nodes. We position to apply LoRa mesh networks as a communication substrate to enable novel networking capacities for the new scenario of distributed smart IoT nodes., This work has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 957258 — ASSIST-IoT, and was partially supported by the Spanish Government under contract PID2019-106774RB-C21, and the Generalitat de Catalunya as Consolidated Research Group 2021-SGR-01059., Peer Reviewed, Postprint (published version)

Bitcoin is the most popular cryptocurrency that uses a blockchain network to ensure ownership and integrity of its transactions. The high volatility of the cryptocurrency trading market makes the use of Bitcoin interesting as an investment asset since it allows for greater prof- itability than most traditional options. The objective of this thesis is to design, develop and deploy an automated bot that will publish every 60 minutes the predictions for the following 5 hours Bitcoin price. The information will be public and will be displayed in an open Twitter account. The predictions are based on a deep learning model that is trained on past Bitcoin price values and relevant technical analysis indicators. The results obtained in the experiments carried out in this thesis reveal a mean absolute percentage error of 0.73% of adjustment for year 2022 dataset., Bitcoin es la criptomoneda más popular que utiliza una red blockchain para garantizar la propiedad y la integridad de sus transacciones. La alta volatilidad del mercado de compraventa de criptomonedas hace que el uso de Bitcoin sea interesante como activo de inversión, permitiendo una mayor rentabilidad que la mayoría de las opciones tradicionales. El objetivo de esta tesis es diseñar, desarrollar y desplegar un bot automatizado que publicará cada 60 minutos las predicciones para las próximas 5 horas del precio de Bitcoin. La información será pública y se mostrará en una cuenta de Twitter abierta. Las predicciones se basan en un modelo de aprendizaje profundo entrenado con valores del precio de Bitcoin anteriores e indicadores de análisis técnico relevantes. Los resultados obtenidos en los experimentos realizados en esta tesis muestran un error porcentual absoluto medio del 0.73% de ajuste en el conjunto de datos del año 2022., Bitcoin és la criptomoneda més popular que utilitza una xarxa blockchain per garantir la propietat i la integritat de les seves transaccions. L'alta volatilitat del mercat de compravenda de criptomonedes fa que l'ús de Bitcoin sigui interessant com a actiu d'inversió, ja que permet una major rendibilitat que la majoria de les opcions tradicionals. L'objectiu d'aquesta tesi és dissenyar, desenvolupar i desplegar un bot automatitzat que publica cada 60 minuts les prediccions per les properes 5 hores del preu de Bitcoin. La informació serà pública i es mostrarà en un compte de Twitter obert. Les prediccions es basen en un model d'aprenentatge profund que entrenat amb valors de preus de Bitcoin anteriors i indicadors d'anàlisi tècnic rellevants. Els resultats obtinguts en els experiments realitzats en aquesta tesi revelen un error relatiu mitjà del 0.73% d'ajust en el conjunt de dades de l'any 2022.

The rapid expansion in the size of new datasets and available data online has led to significant scaling of the size of neural models. However, training deep learning models and performing hyperparameter tuning is computationally demanding and very time-consuming. Thus, there is a growing need for fast and efficient learning techniques. In this thesis, we explore two meta-learning approaches to provide faster and data-driven systems for multiple tasks. First, we introduce HyperFast, a meta-trained hypernetwork designed for instant classification of tabular data in a single forward pass. HyperFast generates the weights of a task-specific neural network tailored to an unseen dataset that can be directly used for classification inference, removing the need for training a model. We report extensive experiments with OpenML and genomic data, comparing HyperFast to competing tabular data neural networks, traditional ML methods, AutoML systems, and boosting machines such as XGBoost. HyperFast shows highly competitive results, while being significantly faster. Our approach introduces a promising paradigm for fast classification and rapid model deployment. Second, we propose a Compressive Meta-Learning framework as an alternative to traditional compressive learning, wherein random, non-linear features are used to project large-scale databases onto compact information-preserving representations. These database-level summaries are later used to learn parameters of interest from the underlying data distribution, without the need to access the original samples, providing an efficient and privacy-friendly learning framework. However, both the encoding and decoding techniques are typically randomized and data-independent, not benefiting from the underlying structure of the data. In our approach, we meta-learn both the encoding and decoding stages of compressive learning methods in a data-driven fashion by using neural networks. To showcase the potential of the presented Compressive

The advent of 5G has meant providing users with higher data rates, more connected devices and lower latency, implying the need to redefine traditional network management methods. In addition, 6G is expected to provide much more challenging and restrictive performance, enhancing the idea of redefining network management methodologies, aimed at maximising the use of network resources. Network Digital Twinning (NDT) and Optimization-as-a-Service (OaaS) are two key mechanisms, along with the use of artificial intelligence (AI) and machine learning (ML), to assist in network management procedures. On the one hand, the NDT is a digital replica of the physical network, which permits evaluating the impact of concrete configuration settings prior to its real implementation in the system, helping to optimize the overall network performance. On the other hand, OaaS offers ad-hoc optimized recommendations to third-party applications based on the execution of several optimization algorithms. In this work, we propose an open, flexible and scalable OaaS platform, defined in the OPTIMAIX project, to exploit the NDT potential targeted to support automated planning and optimization of networks. Also, an evaluation of the NDT role within this architecture is provided., This work is funded by the Spanish Ministry of Economic Affairs and Digital Transformation and the European Union - NextGenerationEU under projects OPTIMAIX_OaaS (Ref. TSI-063000-2021-34) and OPTIMAIX_NDT (Ref. TSI063000-2021-35)., Peer Reviewed, Postprint (published version)

Reliable anticipation of cyclists' movements is a major challenge for autonomous vehicles in urban environments and can significantly enhance the functionality of intelligent driving systems. In this work, we propose a deep learning approach that combines visual and temporal features to accurately forecast future trajectories of cyclists using on-board camera systems. Specifically, the proposed method employs a sequence-to-sequence architecture with recurrent units and is based on the public Waymo Open Dataset. We conduct various experiments to assess the proposed approach and we demonstrate the benefit of using visual features for cyclist trajectory prediction. Results also show that the proposed approach outperforms widely-used traditional model-based methods in terms of displacement error and intersection-over-union.

The human microbiome has become an area of intense research due to its potential impact on human health. However, the analysis and interpretation of this data have proven to be challenging due to its complexity and high dimensionality. Machine learning (ML) algorithms can process vast amounts of data to uncover informative patterns and relationships within the data, even with limited prior knowledge. Therefore, there has been a rapid growth in the development of software specifically designed for the analysis and interpretation of microbiome data using ML techniques. These software incorporate a wide range of ML algorithms for clustering, classification, regression, or feature selection, to identify microbial patterns and relationships within the data and generate predictive models. This rapid development with a constant need for new developments and integration of new features require efforts into compile, catalog and classify these tools to create infrastructures and services with easy, transparent, and trustable standards. Here we review the state-of-the-art for ML tools applied in human microbiome studies, performed as part of the COST Action ML4Microbiome activities. This scoping review focuses on ML based software and framework resources currently available for the analysis of microbiome data in humans. The aim is to support microbiologists and biomedical scientists to go deeper into specialized resources that integrate ML techniques and facilitate future benchmarking to create standards for the analysis of microbiome data. The software resources are organized based on the type of analysis they were developed for and the ML techniques they implement. A description of each software with examples of usage is provided including comments about pitfalls and lacks in the usage of software based on ML methods in relation to microbiome data that need to be considered by developers and users. This review represents an extensive compilation to date, offering valuable in, Peer Reviewed, Article signat per 48 autors/es Marcos-Zambrano LJ, López-Molina VM, Bakir-Gungor B, Frohme M, Karaduzovic-Hadziabdic K, Klammsteiner T, Ibrahimi E, Lahti L, Loncar-Turukalo T, Dhamo X, Simeon A, Nechyporenko A, Pio G, Przymus P, Sampri A, Trajkovik V, Lacruz-Pleguezuelos B, Aasmets O, Araujo R, Anagnostopoulos I, Aydemir &, Berland M, Calle ML, Ceci M, Duman H, Gündoğdu A, Havulinna AS, Kaka Bra KHN, Kalluci E, Karav S, Lode D, Lopes MB, May P, Nap B, Nedyalkova M, Paciência I, Pasic L, Pujolassos M, Shigdel R, Susín A, Thiele I, Truică C-O, Wilmes P, Yilmaz E, Yousef M, Claesson MJ, Truu J and Carrillo de Santa Pau E, Objectius de Desenvolupament Sostenible::3 - Salut i Benestar, Postprint (published version)

In this paper, we propose the use of nonlinear oblique subspace projections (NObSP) with support vector machines (SVM), as well as an out-of-sample extension of this algorithm to reduce its arithmetic complexity. NObSP was originally developed for least-Squares SVM, and it aims at decomposing the model output into additive components. Each component represents the partial (non)linear contribution of each input regressor, and their interaction effects, on the output. The original NObSP has an arithmetic complexity of O(N3), being N the number of observations. The proposed out-of-sample extension, combined with SVM, reduces the arithmetic complexity of NObSP to O(N*d2), with d the number of support vectors., Peer Reviewed, Postprint (published version)

Background: The rise of machine learning (ML) systems has exacerbated their carbon footprint due to increased capabilities and model sizes. However, there is scarce knowledge on how the carbon footprint of ML models is actually measured, reported, and evaluated. Aims: This paper analyzes the measurement of the carbon footprint of 1,417 ML models and associated datasets on Hugging Face. Hugging Face is the most popular repository for pretrained ML models. We aim to provide insights and recommendations on how to report and optimize the carbon efficiency of ML models. Method: We conduct the first repository mining study on the Hugging Face Hub API on carbon emissions and answer two research questions: (1) how do ML model creators measure and report carbon emissions on Hugging Face Hub?, and (2) what aspects impact the carbon emissions of training ML models? Results: Key findings from the study include a stalled proportion of carbon emissions-reporting models, a slight decrease in reported carbon footprint on Hugging Face over the past 2 years, and a continued dominance of NLP as the main application domain reporting emissions. The study also uncovers correlations between carbon emissions and various attributes, such as model size, dataset size, ML application domains and performance metrics. Conclusions: The results emphasize the need for software measurements to improve energy reporting practices and the promotion of carbon-efficient model development within the Hugging Face community. To address this issue, we propose two classifications: one for categorizing models based on their carbon emission reporting practices and another for their carbon efficiency. With these classification proposals, we aim to encourage transparency and sustainable model development within the ML community., This work is partially supported by the project TED2021- 130923B-I00, funded by MCIN/AEI/10.13039/501100011033 and the European Union Next Generation EU/PRTR., Peer Reviewed, Postprint (author's final draft)

This research paper proposes a novel methodology for evaluating entrepreneurial activity among engineering education alumni using their public CVs as our main source of information. The objective is to go beyond measuring entrepreneurship intentions or mindset through surveys, and instead analyse actual career data to assess the impact of entrepreneurship education. The study utilises shared user data and employs GPT (Generative Pretrained Transformer) models to infer entrepreneurial activity that extends beyond job titles, delving into the specific responsibilities and achievements associated with each position. The analysis shows that the proposed methodology, which uses context enriching to enhance model accuracy, effectively identifies instances of entrepreneurial activity among CVs profiles data. This approach provides a way to evaluate the effectiveness of entrepreneurship and innovation courses. Combining the insights gained through the proposed method with internal data sources would enable institutions to conduct a comprehensive evaluation of program impact on alumni career paths. The study underscores the potential of AI models to facilitate the collection and analysis of data that has traditionally been challenging to access. Moreover, the research highlights the importance of evaluating the long-term impact of entrepreneurship education on alumni career trajectories, a key factor in addressing the growing field of engineering education. Ultimately, this study contributes to the academic discourse on entrepreneurship education by offering a novel approach for assessing the impact of such programs on alumni outcomes, thus enabling institutions to make data-driven decisions to improve program offerings., Peer Reviewed, Objectius de Desenvolupament Sostenible::9 - Indústria, Innovació i Infraestructura, Objectius de Desenvolupament Sostenible::8 - Treball Decent i Creixement Econòmic, Objectius de Desenvolupament Sostenible::4 - Educació de Qualitat, Postprint (published version)

The aim of this project is to improve the MobilitApp application focusing on the Machine Learning model used. MobilitApp is an Android application developed by ETSETB students since 2014, in collaboration with the ATM (Autoritat del Transport Metropolit`a de l?`area de Barcelona) of Barcelona, with the purpose of collect data of accelerometer, magnetometer and gyroscope in order to train a Machine Learning (ML) model able to detect the means of transport used by its users. Specifically, we used a deep learning model designed by members of the SISCOM research group of the UPC. In this thesis we studied an outlier detection algorithm, the selection of the hyper parameters to train the optimal model and finally the feature importance and its relation between the accuracy and the computational cost., El objetivo de este proyecto es mejorar la aplicación MobilitApp centrándose en el modelo de Machine Learning que se utiliza, proponiendo algoritmos usados en la literatura. MobilitApp es una aplicación para Android desarrollada por estudiantes de la ETSETB desde 2014, en colaboración con el ATM (Autoritat del Transport Metropolità de l’Àrea de Barcelona) de Barcelona, con la finalidad de recopilar datos del acelerómetro, magnetómetro y giroscopio para entrenar un modelo de Machine Learning (ML) capaz de detectar los medios de transporte utilizados por sus usuarios. Específicamente, se utilizó un modelo de aprendizaje profundo diseñado por miembros del grupo de investigación SISCOM de la UPC. En esta tesis se estudió un algoritmo de detección de outliers, a selección de los hiperparámetros para entrenar el modelo ´optimo, y finalmente, la importancia de las características y la relación entre la exactitud y el costo computacional. Las líneas de trabajo futuras para MobilitApp incluyen el uso de un conjunto de datos más grande y equilibrado, la automatización de los algoritmos de prueba del modelo y, finalmente, como objetivo final, la integración del código de MobilitApp en el software MOU-TE de ATM, L’objectiu d’aquest projecte és millorar l’aplicació MobilitApp centrant-se en el model de Machine Learning que s’utilitza. MobilitApp és una aplicació per a Android desenvolupada per estudiants de l’ETSETB des de 2014, en col·laboració amb l’ATM (Autoritat del Transport Metropolità de l’Àrea de Barcelona) de Barcelona, amb la finalitat de recopilar dades de l’acceleròmetre, magnetòmetre i giroscopi per entrenar un model de Machine Learning (ML) capaç de detectar els mitjans de transport utilitzats pels seus usuaris. Específicament, vam utilitzar un model d’aprenentatge profund dissenyat per membres del grup de recerca SISCOM de la UPC. En aquesta tesi hem estudiat un algorisme de detecció d’outliers, la selecció dels hiperparàmetres per entrenar el model òptim, i finalment la importància de les característiques i la relació entre l’exactitud i el cost computacional.

In the present work, which contributes to the E-Bike City Project, machine learning and multiple linear regression models are implemented to predict trip generation and trip attraction on cycling demand. The spatial scope is the metropolitan region of Zurich. The input data is obtained from MZMV and EBIS, two mobility surveys which collect tracking data. A modelling framework is developed in R, which trains machine learning models to predict the daily average cycling trips per person of all individuals present in the permanent population, based on socioeconomic variables. The results are then aggregated at the zonal level. These models are compared to traditional linear regression models, which provide lower prediction and explanatory power in the present study. Body-Mass Index, household size and owning either a bike or an e-bike are found to be the most important variables according to Random Forest algorithm. Cycling trip attraction is modelled using spatial variables extracted from OpenStreetMaps data. An MLR model describes the attraction of cycling trips at the zonal level. Statistical significance and large effects are found in variables describing the amount of leisure amenities and very high Pubic Transport accessibility. Trip distribution estimations show larger cycling flows between central and most urbanized areas. The predicted productions and attractions present an over-estimation compared to actual data, although the spatial pattern of trip generation and attraction are reasonably fitting the reality., Outgoing

Aquest projecte es lliura com a Treball Final del Grau d'Enginyeria Informàtica de la Facultat d'Informàtica de Barcelona. Es pretén posar en pràctica els coneixements adquirits durant l'especialitat d'Enginyeria del Software. El projecte consisteix en dissenyar i implementar un Framework en forma d'aplicació web que permeti als empleats del departament de Data Science de l'empresa Grupo Catalana Occidente: Tecnologia y servicios a posar nous models de Machine Learning en producció i monitorar els que ja hi són de forma senzilla i còmoda. L'aplicació automatitzarà la majoria de les tasques a realitzar i guiarà l'usuari en tot moment per a obtenir uns resultats òptims amb una experiència agradable. A més, mostrarà de forma gràfica els aspectes més importants de l'estat dels models en producció., The present work is handed as a senior thesis as a part of the bachelor's degree in Informatics Engineering at Barcelona School of Informatics. It intends to put the acquired knowledge throughout the specialisation of Software Engineering into practice. The project consists of designing and implementing a Framework in the form of a web application that allows the employees of the Data Science department of the Grupo Catalana Occidente: Tecnologia y servicios company to put new Machine Learning models into production and monitor the ones that are already there in a simple and convenient way . The application will automate most of the tasks to be performed and guide the user at all times to obtain optimal results with a pleasant experience. In addition, it will show graphically the most important aspects of the state of the models in production.

The implementation of next generation beyond 5G use cases such as the Digital Twin (DT) requires the transportation of increasingly large amounts of sensor telemetry data. The reduction of the size of these time series shaped data through compression has many benefits, ranging from energy savings in systems where sensors have limited power to transport network bandwidth reduction. Both statistical analysis-based and machine learning (ML)-based approaches have been employed to tackle this task with varying degrees of success. In this paper, we compare two methods for time series data compression: a statistical similarity-based one, and an autoencoder (AE)-based one., This research was funded by European Commission through the HORIZON DESIRE6G project (G.A. 101096466), by the AEI through the IBON project (PID2020-114135RB-I00), and by ICREA institution., Peer Reviewed, Postprint (author's final draft)

In this thesis, we study the problem denoted as Non-Rigid Structure from Motion and tackle the two distinct parts of the problem: motion and shape estimation together with the corresponding temporal segmentation into actions of the body. For the motion estimation, we implement a Single Rotation algorithm (SRA) based on the Weiszfeld algorithm for the median of points in Rn . We use Brand’s method to obtain a full corrective matrix and several estimates for the rotation matrices but find that the first column triplet obtains better results than any of the triplets found by Brand’s method, making SRA pointless. For the shape factor, we make an assumption that shapes lie in a temporal union of subspaces. We perform Sparse Subspace Clustering to jointly reconstruct the shape matrix while computing an affinity matrix that can be used to cluster frames of the input data according to which subspace they belong to., En esta tesis, estudiamos el problema de Non-Rigid Structure from Motion y abordamos dos partes distintas del problema: estimación de movimiento y forma junto con la correspondiente segmentación temporal en acciones del cuerpo no-rígido. Para la estimación del movimiento, implementamos un Single Rotation Averaging (SRA) basado en el algoritmo de Weiszfeld para la mediana de puntos en R^n. Usamos el método de Brand para obtener una matriz correctiva completa y varias estimaciones para las matrices de rotación, pero encontramos que la primera tripleta de columnas obtiene mejores resultados que cualquiera de las tripletas encontrados por el método de Brand, haciendo que SRA sea inútil. Para el factor de forma, asumimos que las formas se encuentran en una unión temporal de subespacios. Realizamos Sparse Subspace Clustering para conjuntamente reconstruir la matriz de forma mientras se calcula una matriz de afinidad que se puede utilizar para agrupar los frames de los datos de entrada según a qué subespacio pertenecen., En aquesta tesi, estudiem el problema de Non-Rigid Structure from Motion i abordem dues parts diferents del problema: estimació de moviment i forma juntament amb la corresponent segmentació temporal en accions del cos no rígid. Per a l'estimació del moviment, implementem un Single Rotation Averaging (SRA) basat en l'algorisme de Weiszfeld per a la mitjana de punts a R^n. Fem servir el mètode de Brand per obtenir una matriu correctiva completa i diverses estimacions per a les matrius de rotació, però trobem que la primera tripleta de columnes obté millors resultats que qualsevol de les tripletes trobades pel mètode de Brand, fent que SRA sigui inútil. Per al factor de forma, assumim que les formes es troben en una unió temporal de subespais. Realitzem Sparse Subspace Clustering per conjuntament reconstruir la matriu, mentre es calcula una matriu d'afinitat que es pot utilitzar per agrupar els frames dades d'entrada segons a quin subespai pertanyen.

En entorns foscos i de difícil accés, l'aplicació del monitoratge acústic passiu (MAP) ha demostrat ser essencial per a una millor comprensió del comportament dels animals i la biodiversitat. La implementació de MAP ha posat en manifest la necessitat de mètodes de aprenentatge automàtic (machine learning, ML) per detectar i classificar senyals en paisatges sonors complexos. Aquesta tesi té com a objectiu trobar les aplicacions i les limitacions d'un conjunt d’algoritmes basats en ML, anomenats soundscape_IR, per intentar detectar i classificar una sèrie de crits d'ecolocalització de diferents espècies de ratpenats i unitats de cant de la balena de Groenlandia (Balaena mysticetus) etiquetades. Els resultats mostren que els models entrenats mitjançant la els algoritmes examinats només es poden utilitzar com a detectors, no com a classificadors, ja que no són específics de les dades en què s'han entrenat. Malgrat el rendiment pobre del model, els algoritmes encara es poden utilitzar per a la detecció de senyals en conjunts de dades extensos, accelerant així el procés d'anàlisi de dades. Paral·lelament, es realitzen una sèrie de proves a l'aigua per avaluar la sensibilitat de dos hidròfons., En entornos oscuros y de difícil acceso, la aplicación del monitoreo acústico pasivo (MAP) ha demostrado ser esencial para una mejor comprensión del comportamiento animal y la biodiversidad. La implementación de MAP ha puesto de relieve la necesidad de métodos de aprendizaje automático (machine learning, ML) para detectar y clasificar señales en paisajes sonoros complejos. Esta tesis tiene como objetivo probar las aplicaciones y limitaciones de una de un conjunto de algoritmos basados en ML, anomenado soundscape_IR, para intentar detectar y clasificar una variedad de gritos de ecolocalización de diferentes especies de murciélagos y unidades de canto etiquetadas de ballenas de Groenlandia (Balaena mysticetus). Los resultados muestran que los modelos entrenados con los algoritmos examinados solo pueden servir como detectores, no clasificadores, ya que no son específicos de los datos con los que fueron entrenados. A pesar del bajo rendimiento del modelo, los algoritmos aún se puede utilizar para la detección de señales en conjuntos de datos extensos, lo que acelera el proceso de análisis de datos. Al mismo tiempo, se realizan una serie de pruebas en el agua para evaluar la sensibilidad de dos hidrófonos., In hard-to-reach and dark environments, the application of passive acoustic monitoring (PAM) has proven to be essential for a better understanding of animal behaviour and biodiversity. The implementation of PAM has highlighted the need for Machine Learning (ML) methods to detect and classify signals in complex soundscapes. This thesis aims to test the applications and limitations of one such ML toolbox, soundscape_IR, by attempting to detect and classify a range of echolocation calls from different bat species and labelled bowhead whale (Balaena mysticetus) song units. Results show that the models trained using the examined toolbox can only serve as detectors, not classifiers, as they are not specific to the data that they were trained on. Despite poor model performances, the toolbox can still be utilised for signal detection in extensive datasets, thereby expediting the data analysis process. Concurrently, a series of in-water tests are performed to assess the sensitivity of two hydrophones.

El aprendizaje automático, o machine learning, ha entrado con fuerza en el ámbito de la investigación, transformando la forma en que los científicos y académicos abordan sus problemas. Se puede considerar como un nuevo pilar en el proceso de la investigación científica, complementando las metodologías tradicionales. En los últimos años, el machine learning ha demostrado ser una herramienta esencial en el análisis de grandes conjuntos de datos, descubrir patrones y extraer conocimientos valiosos que de otra forma se nos escaparían. Además, la programación permite, cada vez más, automatizar procesos que de otra forma serían áreas tediosas y repetitivas, liberando así tiempo para realizar otras actividades de investigación. Las imágenes satelitales y los datos del Sistema de Identificación Automática (AIS, siglas en inglés) de las embarcaciones son dos fuentes valiosas de información en el ámbito de la navegación. Las imágenes radar ofrecen una visión de la ubicación y el movimiento de las embarcaciones independientemente de las condiciones climáticas o de iluminación. Por otro lado, los datos AIS son un sistema de seguimiento cada vez implantado en más buques y navíos de recreo que transmiten datos clave, como la identificación, rumbo, velocidad y medidas de la embarcación, enviados regularmente y recibidos por barcos y puertos mundialmente. Dentro de este marco se busca realizar un trabajo de investigación centrado en estas tecnologías. Dicho trabajo tiene dos partes diferenciadas. En la primera parte creamos un código que cruce datos AIS con las imágenes satelitales cerciorándonos de que realmente las embarcaciones aparezcan y se vean y la creación de etiquetas mediante buffers para alimentar posteriormente al modelo de predicción basado en un aprendizaje supervisado. Todo el proceso de obtener los datos, procesarlos con distintos filtros, crear los buffers y la comprobación de que las etiquetas están realmente bien, ya que si se alimenta erróneamente el algoritmo n

G protein-coupled receptors (GPCRs) are a large superfamily of cell membrane proteins that play an important physiological role as transmitters of extracellular signals. Signal transmission through the cell membrane depends on conformational changes in the transmembrane region of the receptor, which makes the investigation of the dynamics in these regions particularly relevant. Molecular dynamics (MD) simulations provide a wealth of data about the structure, dynamics, and physiological function of biological macromolecules by modelling the interactions between their atomic constituents. In this study, a Recurrent and Convolutional Neural Network (RNN) model, namely Long Short-Term Memory (LSTM), is used to predict the dynamics of two GPCR states and three specific simulations of each one, through their activation path and focussing on specific receptor regions. Active and inactive states of the GPCRs are analysed in six scenarios involving APO, Full Agonist (BI 167107) and Partial Inverse Agonist (carazolol) of the receptor. Four Machine Learning models with increasing complexity in terms of neural network architecture are evaluated, and their results discussed. The best method achieves an overall RMSD lower than 0.139 Å and the transmembrane helices are the regions showing the minimum prediction errors and minimum relative movements of the protein., This work is funded by Spanish PID2019-104551RB-I00 research project and by the PRE2020-092428 Ph.D. training program, through the Ministry of Science and Innovation., Peer Reviewed, Postprint (published version)

En l'actualitat, la gestió dels processos administratius s'ha convertit principalment en una tasca telemàtica. Després de rebre les notificacions, les empreses han de treure'n la informació rellevant. Encara que el contingut de les notificacions pot variar, aquelles que fan referència al mateix tipus de tràmit solen tenir una estructura similar, amb els camps importants identificables pel seu context o ubicació. L'automatització d'aquest procés és essencial per agilitzar les tasques de gestió dels processos administratius. Les companyies asseguradores, en particular, reben un gran volum de notificacions que han de gestionar. Moltes d'aquestes notificacions provenen de processos judicials, per la qual cosa solen tenir un format formal i seguir una estructura específica. Depenent del procés judicial, la gestió de les notificacions pot ser diferent. Aquest projecte té com a objectiu desenvolupar un sistema automàtic capaç de rebre i gestionar documents relacionats amb tràmits judicials en el context de les companyies asseguradores. Aquest sistema ha de ser capaç de classificar els documents en categories específiques i extreure la informació necessària per a l'asseguradora. Considerant els mètodes existents, s'ha optat per implementar un sistema de processament intel·ligent ad-hoc per abordar aquest problema. Es farà servir l'aprenentatge automàtic (Support Vector Machines, SVM) per classificar els diferents tipus de documents, i s'aplicaran mètodes d'extracció de text específics per a cada camp important, amb l'objectiu de maximitzar l'extracció de dades i automatitzar el procés de manera eficient., Currently, the management of administrative processes has become mainly a telematic task. After receiving the notifications, companies must extract the relevant information from them. Although the content of the notifications may vary, those that refer to the same type of procedure usually have a similar structure, with the important fields identifiable by their context or location. The automation of this process is essential to speed up the tasks of managing administrative processes. Insurance companies, in particular, receive a large volume of notifications that they have to manage. Many of these notifications come from court proceedings, so they tend to have a formal format and follow a specific structure. Depending on the judicial process, the handling of notifications may differ. This project aims to develop an automatic system capable of receiving and managing documents related to court procedures in the context of insurance companies. This system must be able to classify documents into specific categories and extract the necessary information for the insurer. Considering the existing methods, it has been chosen to implement an ad-hoc intelligent processing system to address this problem. Machine learning (Support Vector Machines, SVM) will be used to classify the different types of documents, and specific text extraction methods will be applied for each important field, with the aim of maximizing data extraction and automate the process efficiently.

El aprendizaje automático (machine learning) es una tecnología emergente que actualmente está experimentando un auge muy significativo. Es por eso interesante profundizar en un campo de la computación con tanto potencial que se espera que impulse una revolución tecnológica en los próximos años —si no es que lo está haciendo ya—. Este proyecto de fin de grado explora las bases teóricas del aprendizaje automático, más concretamente de las redes neuronales artificiales, y su aplicación en la visión por computador. Utilizando redes neuronales convolucionales se entrenan cuatro modelos de Deep Learning —image classification, object detection en local y object detection online— con el objetivo de conseguir uno capaz de reconocer y localizar una pequeña pelota de forma rápida y precisa. Dicho modelo se implementa en un robot cuadrúpedo, el robot Bittle de Petoi, a través de un algoritmo de control desarrollado en Python y utilizando una Raspberry Pi para ejecutar la inferencia. El objetivo final de este proyecto es conseguir que el robot Bittle avance solo hacia la pelota basándose únicamente en la información que le proporciona el sensor de cámara haciendo uso solamente de técnicas de deep learning, ya que existen otras técnicas como la detección por geometría o por color que consigue resultados similares. Este trabajo pretende ser una ampliación y propuesta de mejora del trabajo final de máster titulado "Estrategia de movimiento basada en Deep Learning y Visión por Ordenador para Robot Cuadrúpedo" realizado por Gabriel Mourad, en el que entrenó también cuatro modelos de clasificación de imágenes con el mismo objetivo

In general, the availability of an accurate machine learning (ML) model plays a particularly important role in the development of new networking solutions and is one of the main drivers for the development of 5G and beyond networking. Although an option is to update the model once inaccurate data is detected, such approach requires high computational effort, specially once the data history is large. In this paper, we propose an approach that combines a traffic prediction model based on Long Short-Term Memory (LSTM) with an analysis module for dynamic connection capacity allocation. Once the model is generated, re-training can be triggered after inaccuracies are detected by the analysis module. Illustrative numerical results show the benefits from the proposed decision-based re-training approach to reduce the number of re-training rounds while maintaining model accuracy., This work was partially funded by MCIN/AEI/ 10.13039/501100011033 grant PID2021-126431OB-I00 (ANEMONE), and Generalitat de Catalunya grant 2021 SGR 00770. This work also was funded by the European Commission Horizon Europe SNS JU DESIRE6G project (G.A. 101096466), from the Spanish MINECO and the European Union though the NextGeneration UNICO5G TIMING (TSI-063000-2021-145), and the AEI IBON (PID2020-114135RB-I00) projects. The authors also acknowledge the support of the French Agence Nationale de la Recherche (ANR), under grant ANR-19-CE-25-0001-01 (ARTIC project)., Peer Reviewed, Postprint (author's final draft)

Network Services automation requires predictable Quality of Service (QoS) performance, measured in terms of throughput, delay and jitter, to allow making proactive decisions. QoS is typically guaranteed by overprovisioning capacity dedicated to the packet connection, which increases costs for customers and network operators, especially when the traffic generated by the users and/or the virtual functions highly varies over the time. This paper presents the PILOT methodology for modeling the performance of packet connections during commissioning testing in terms of throughput, delay and jitter. PILOT runs in a sandbox domain and constructs a scenario where an efficient traffic flow simulation environment, based on the CURSA-SQ model, is used to generate large amounts of data for Machine Learning (ML) model training and validation. The simulation scenario is tuned using real measurements of the connection obtained from a set of active probes., The research leading to these results has received funding from the Smart Networks and Services Joint Undertaking under the European Union's Horizon Europe research and innovation programme under Grant Agreement No. 101095890 (PREDICT6G), from the Spanish MINECO and the European Union though the NextGeneration UNICO5G TIMING (TSI-063000- 2021-145/148/149), and the AEI IBON (PID2020-114135RB-I00) projects and from ICREA., Peer Reviewed, Postprint (author's final draft)

When classifiers face imbalanced class distributions, they often misclassify minority class samples, consequently diminishing the predictive performance of machine learning models. Existing oversampling techniques predominantly rely on the selection of neighboring data via interpolation, with less emphasis on uncovering the intrinsic patterns and relationships within the data. In this research, we present the usefulness of an algorithm named RuLer to deal with the problem of classification with imbalanced data. RuLer is a learning algorithm initially designed to recognize new sound patterns within the context of the performative artistic practice known as live coding. This paper demonstrates that this algorithm, once adapted (Ad-RuLer), has great potential to address the problem of oversampling imbalanced data. An extensive comparison with other mainstream oversampling algorithms (SMOTE, ADASYN, Tomek-links, Borderline-SMOTE, and KmeansSMOTE), using different classifiers (logistic regression, random forest, and XGBoost) is performed on several real-world datasets with different degrees of data imbalance. The experiment results indicate that Ad-RuLer serves as an effective oversampling technique with extensive applicability., Peer Reviewed, Postprint (published version)

The Intent-Based Networking (IBN) paradigm targets at defining high-level abstractions, so network operators can define what are their desired outcomes without specifying how they would be achieved. The latter can be achieved by leveraging network programmability, monitoring and data analytics, as well as the key assurance component. In this paper, we introduce the IBN paradigm and its application to optical networking, highlighting the benefits that Machine Learning (ML) algorithms can provide to IBN. Illustrative examples of intent-based operation are presented for proactive self-configuration and cooperative intent operation., The research leading to these results has received funding from the H2020 B5G-OPEN (G.A. 101016663) and the AEI IBON (PID2020-114135RB-I00) projects and from the ICREA institution., Peer Reviewed, Postprint (author's final draft)

Wind power forecasting plays a crucial role in optimizing the performance and efficiency of wind farms. Traditional machine learning and statistical models for wind power prediction rely heavily on historical power and meteorological observations from wind farms. However, an unresolved challenge in this field is making accurate predictions when there are no observations available. This final thesis describes the work done for Whiffle, a weather forecasting company renowned for its state-of-the-art wind power forecasting system. Whiffle's wind power forecasting system stands out due to its unique and high-resolution weather model. The focus of this project was to augment Whiffle's existing services by integrating data-free forecasting techniques, particu- larly in situations where observational data is unavailable. The research focuses on two main challenges: the absence of real world observations and missing data archives that are relevant for Whiffle's power production forecasts. To overcome the challenge of missing real world obser- vations, Whiffle's unique capability to construct virtual observation datasets is leveraged. This innovative approach allows for the generation of synthetic observations that closely resemble real- world data, enabling the training of forecasting models even in the absence of actual observations. Through extensive experimentation and validation, the effectiveness and reliability of the data- free forecasting techniques are demonstrated. The performance of the models are evaluated using two metrics: Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE). The results consistently demonstrate proportional relationships between MAPE and RMSE, pro- viding confidence in the model's performance evaluation. The utilization of virtual observations yielded favorable outcomes, resulting in only an average increase of 3.8% in MAPE. In addition, this thesis also aims to improve Whiffle's machine learning model used for power produ, La predicció de l'energia eòlica juga un paper crucial en l'optimització del rendiment i l'eficiència dels parcs eòlics. Els models tradicionals d'aprenentatge automàtic i estadístics per a la predicció de l'energia eòlica depenen en gran mesura de les observacions històriques d'energia i meteorològiques dels parcs eòlics. No obstant això, un repte pendent en aquest camp és fer prediccions precises quan no hi ha observacions disponibles. Aquest treball de final de grau descriu la feina realitzada per a Whiffle, una empresa de predicció meteorològica reconeguda pel seu sistema de predicció d'energia eòlica d'última generació. El sistema de predicció d'energia eòlica de Whiffle destaca pel seu model meteorològic únic i d'alta resolució. L'objectiu d'aquest projecte és augmentar els serveis existents de Whiffle mitjançant la integració de tècniques de predicció sense dades, especialment en situacions en què les dades d'observació no estan disponibles. La recerca se centra en dos reptes principals: l'absència d'observacions del món real i les arxius de dades que falten i que són rellevants per a les prediccions de producció d'energia de Whiffle. Per superar el repte de la falta d'observacions del món real, s'aprofita la capacitat única de Whiffle per construir conjunts de dades d'observació virtuals. Aquest enfocament innovador permet generar observacions sintètiques que s’assemblen estretament a les dades del món real, permetent l'entrenament de models de predicció fins i tot en absència d'observacions reals. Mitjançant una experimentació i validació extenses, es demostra l'eficàcia i la fiabilitat de les tècniques de predicció sense dades. El rendiment dels models s'avalua utilitzant dos indicadors: l'Error Percentual Mitjà Absolut (MAPE, per les seves sigles en anglès) i l'Error Quadràtic Mitjà (RMSE, per les seves sigles en anglès). Els resultats mostren de manera consistent relacions proporcionals entre el MAPE i el RMSE, proporcionant confiança en l'avaluació del

Modern optimization problems often involve dealing with conflicting objectives, making them suitable for multi-objective optimization techniques. This research introduces MO-PUGAN, an innovative framework utilizing Generative Adversarial Networks (GANs) for multi-objective optimization in engineering design. Utilizing a training data update method, MO-PUGAN iteratively refines the GAN's training. Comprehensive experimentation reveals MO-PUGAN's ability to learn and mirror the distribution of optimal Pareto fronts, beginning with random populations and progressively enhancing it. In comparison to conventional optimization methods that navigate the design space, our framework focuses on learning the optimal distribution eventually. Results indicate that, while it holds promise, it has yet to surpass traditional techniques like Evolutionary Algorithms in performance. This work underscores the potential of melding machine learning with traditional optimization, offering fresh perspectives and methodologies in the design optimization domain., Los problemas modernos de optimización a menudo implican lidiar con objetivos en conflicto, lo que los hace adecuados para técnicas de optimización multiobjetivo. Esta investigación introduce MO-PUGAN, un marco innovador que utiliza Redes Antagónicas Generativas (GANs por sus siglas en inglés) para la optimización multiobjetivo en diseño de ingeniería. Utilizando un método de Actualización de los datos de entrenamiento, MO-PUGAN refina iterativamente el entrenamiento del GAN. La experimentación exhaustiva revela la habilidad de MO-PUGAN para aprender y reflejar la distribución de frentes de Pareto óptimos, comenzando con poblaciones aleatorias y mejorando progresivamente. En comparación con los métodos convencionales de optimización que navegan por el espacio de diseño, nuestro marco se centra en aprender esta distribución óptima eventualmente. Los resultados indican que, aunque tiene potencial, aún no ha superado técnicas tradicionales como los algoritmos evolutivos en rendimiento. Este trabajo destaca el potencial de fusionar el aprendizaje automático con la optimización tradicional, ofreciendo nuevas perspectivas y metodologías en el dominio de la optimización de diseño., Els problemes moderns d'optimització sovint impliquen tractar amb objectius en conflicte, cosa que els fa adequats per a tècniques d'optimització multiobjectiu. Aquesta recerca introdueix MO-PUGAN, un marc innovador que utilitza Xarxes Generatives Adversaries (GANs per les seves sigles en anglès) per a l'optimització multiobjectiu en disseny d'enginyeria. Utilitzant un mètode d'actualització de les dades d'entrenament, MO-PUGAN refina iterativament l'entrenament del GAN. L'experimentació exhaustiva revela l'habilitat de MO-PUGAN per aprendre i reflectir la distribució de fronts de Pareto òptims, començant amb poblacions aleatòries i millorant progressivament. En comparació amb els mètodes convencionals d'optimització que naveguen per l'espai de disseny, el nostre marc es centra en aprendre aquesta distribució óptima eventualment. Els resultats indiquen que, encara que té potencial, encara no ha superat tècniques tradicionals com els algoritmes evolutius en rendiment. Aquest treball destaca el potencial de fusionar l'aprenentatge automàtic amb l'optimització tradicional, oferint noves perspectives i metodologies en el domini de l'optimització de disseny.

En aquest treball s’analitzaran les expectatives laborals dels estudiants del grau d’Estadística de la Facultat d’Economia i Empresa de la Universitat de Barcelona i de la Facultat d’Estadística i Matemàtiques de la Universitat Politècnica de Catalunya. L’estudi s’ha realitzat a través de les dades obtingudes d’una enquesta enviada telemàticament als mateixos alumnes del grau. L’objectiu principal d’aquest treball és poder explorar els resultats fent ús de tècniques d’anàlisi multivariant, per identificar diferents patrons de comportament i relacions entre els futurs estadístics. Es conclou, després de l’anàlisi que la majoria dels estudiants tenen la intenció de compaginar els estudis amb el món del treball. A més, hi coincideixen en haver triat aquest grau per la bona sortida laboral. També s’ha vist que els estudiants que tenen previst fer pràctiques tenen més probabilitats de trobar la seva primera feina relacionada amb els estudis en un temps menor.

Hoy en día, los Centros de Operaciones de Seguridad (SOC) dependen de herramientas reactivas para la detección de amenazas y la gestión de riesgos que son cada vez menos efectivas en la actual industria de la ciberseguridad, donde las sofisticadas campañas realizadas por los ciberdelincuentes no están siendo observadas por la mayoría de herramientas de detección actuales. Existe una clara evidencia de la rápida evolución de las amenazas de ciberseguridad, con los atacantes desarrollando constantemente nuevas técnicas para evitar ser detectados por los sistemas de seguridad y obtener acceso no autorizado a datos sensibles. En particular, la detección de mutaciones de amenazas de ciberseguridad, en la que los atacantes modifican sus técnicas para evadir la detección, se ha convertido en un reto clave para las organizaciones que buscan proteger sus datos y sistemas. Mediante un análisis exhaustivo de incidentes de ciberseguridad reales y datos de red reales del personal de una universidad regional de España, proponemos una nueva taxonomía y metodología en el ámbito de la ciberseguridad para detectar mutaciones de amenazas dentro de una red, combinando algoritmos de aprendizaje automático con análisis de comportamiento para categorizar las amenazas. Nuestros resultados revelan la probabilidad de que una amenaza sea una mutación de una amenaza conocida y las características que comparten ambas amenazas, incluida una nueva representación de perfiles de comportamiento de usuarios y un análisis ampliado de sus propiedades. Este estudio contribuye al avance de las técnicas de detección y prevención en el ámbito de la ciberseguridad, abriendo el camino a sistemas de defensa más resilientes y adaptativos. Las implicaciones prácticas de nuestra investigación proporcionan a los profesionales de la ciberseguridad conocimientos valiosos para fortalecer sus estrategias de respuesta a incidentes y reforzar la postura general de seguridad de sus organizaciones., Avui en dia, els Centres d'Operacions de Seguretat (SOC) depenen d'eines reactives per a la detecció d'amenaces i la gestió de riscos que són cada cop menys efectives en la indústria actual de la ciberseguretat, on les sofisticades campanyes realitzades pels ciberdelinqüents no estan sent observades per la majoria d'eines de detecció actuals. Hi ha una clara evidència de la ràpida evolució de les amenaces de ciberseguretat, amb els atacants desenvolupant constantment noves tècniques per evitar ser detectats pels sistemes de seguretat i obtenir accés no autoritzat a dades sensibles. En particular, la detecció de mutacions d'amenaces de ciberseguretat, on els atacants modifiquen les seves tècniques per evadir la detecció, s'ha convertit en un repte clau per a les organitzacions que busquen protegir les seves dades i sistemes. Mitjançant una anàlisi exhaustiva d'incidents de ciberseguretat reals i dades de xarxa reals del personal d'una universitat regional d'Espanya, proposem una nova taxonomia i metodologia en l'àmbit de la ciberseguretat per detectar mutacions d'amenaces dins d'una xarxa, combinant algorismes d'aprenentatge automàtic amb anàlisi de comportament per categoritzar les amenaces. Els nostres resultats revelen la probabilitat que una amenaça sigui una mutació d'una amenaça coneguda i les característiques que comparteixen ambdues amenaces, inclosa una nova representació de perfils de comportament d'usuaris i una anàlisi ampliada de les seves propietats. Aquest estudi contribueix a l'avenç de les tècniques de detecció i prevenció en l'àmbit de la ciberseguretat, obrint el camí a sistemes de defensa més resilients i adaptatius. Les implicacions pràctiques de la nostra investigació proporcionen als professionals de la ciberseguretat coneixements valuosos per enfortir les seves estratègies de resposta a incidents i reforçar la postura general de seguretat de les seves organitzacions., Nowadays, Security Operation Centres (SOC) relie on reactive tools for threat detection and risk management which are becoming less effective on the current cybersecurity landscape, where attackers are becoming increasingly sophisticated and targeted campaigns fail to be noticed by most current detection tools. There is clear evidence of the rapid evolution of cybersecurity threats, with attackers constantly developing new techniques to bypass security systems and gain unauthorised access to sensitive data. Particularly, the detection of cybersecurity threat mutations -- where attackers modify their techniques to evade detection -- has emerged as a key challenge for organisations seeking to protect their data and systems. Through an extensive analysis of real-world cybersecurity incidents and real network data from staff of a local university in Spain, we propose a novel taxonomy and methodology in the field of cybersecurity to detect threat mutations within a network by combining machine learning algorithms with behavioural analysis to categorise threats. Our findings reveal the likelihood of a threat being a mutation of a known threat and the features both threats share, including a novel representation of user behaviour profiles and an extended analysis of their properties. This study contributes towards the advancement of detection and prevention techniques in the cybersecurity domain, paving the way for more resilient and adaptive defence systems. The practical implications of our research provide cybersecurity professionals with valuable insights to strengthen their incident response strategies and bolster their organisations' overall security posture.

This thesis is part of the SIA-REMU project, which aims to optimize the management of medical emergency resources using artificial intelligence tools. The specific focus of this work is on employing Reinforcement Learning (RL) to optimize the selection of ambulances dispatched during emergencies. To achieve this, we first create a virtual environment that simulates real-world scenarios, where ambulances are distributed among various ambulance stations. Subsequently, different RL algorithms are trained to select the appropriate ambulance dispatch station, based on the position and priority level of the emergency. The algorithms take into account that once an ambulance is dispatched, it becomes unavailable for a certain period, necessitating careful decisions to not compromise the response to subsequent high-priority emergencies. In this study, we compare various RL algorithms, ranging from simpler approaches to state-of-the-art methods. The attempted algorithms include Q-Learning with linear function approximation, Deep Q-Learning, Double Deep Q-Learning, Soft Actor-Critic, and Safe Model-Based Policy Optimization. By implementing these algorithms, we observe distinct behaviors with different properties. Generally, these behaviors entail selecting ambulances from the closest stations for high-priority emergencies and choosing ambulances from larger stations for less urgent situations.

Creating animation scenarios that are immersive and interesting is critical to attracting viewers' attention in the ever-changing world of animation. Traditionally, the design and animation process of these scenarios has been based above all on the knowledge and imagination of human artists, who have had to do a lot of manual work, methodically for each scene and frame of a footage. However, recent advances in computer vision and artificial intelligence have opened up new opportunities to automate the generation of animation scenarios. This thesis will focus on researching, proposing and implementing a design based on the state of the art in generative models (a branch of machine learning) that, given an artistic style, allows transforming a real video into drawings animated while maintaining temporal coherence. The results of the proposed design will be compared to see which model can allow better immersion and remains more faithful to the desired artistic style, La creación de escenarios de animación que sean inmersivos e interesantes es fundamental para atraer la atención de los espectadores al mundo de la animación, que está en constante cambio. Tradicionalmente, el proceso de diseño y animación de estos escenarios se ha basado sobre todo en el conocimiento y la imaginación de los artistas humanos, que han tenido que realizar mucho trabajo manual, de forma metódica en cada escena y fotograma de un metraje. Sin embargo, los recientes progresos en visión por computador e inteligencia artificial han abierto nuevas oportunidades para automatizar la generación de escenarios de animación. Esta tesis se centrará en investigar, proponer e implementar un diseño basado en el estado del arte en los modelos generativos (una de las ramas del aprendizaje automático) que, dado un estilo artístico, permita transformar un vídeo real a dibujos animados manteniendo la coherencia temporal. Los resultados del diseño propuesto se compararán para ver qué modelo permite una mejor inmersión y se mantiene más fiel al estilo artístico deseado, La creació d'escenaris d'animació que siguin immersius i interessants és fonamental per atraure l'atenció dels espectadors en el món de l'animació, que està en constant canvi. Tradicionalment, el procés de disseny i animació d'aquests escenaris s'ha basat sobretot en el coneixement i la imaginació dels artistes humans, que han hagut de fer molta feina manual, de forma metòdica per a cada escena i fotograma d'un metratge. No obstant això, els progressos recents en visió per computador i intel·ligència artificial han obert noves oportunitats per automatitzar la generació d'escenaris d'animació. Aquesta tesi es centrarà en investigar, proposar i implementar un disseny basat en l'estat de l'art en els models generatius (una de les branques de l'aprenentatge automàtic) que, donat un estil artístic, permeti transformar un vídeo real a dibuixos animats mantenint la coherència temporal. Els resultats del disseny proposat es compararan per veure quin model pot permetre una millor immersió i es manté més fidel a l'estil artístic desitjat.

Cuando se acerca el mes de mayo de cada año, siempre nos preguntamos quién va a ser el ganador de la nueva edición del festival de Eurovisión. Este evento no solo atrae la mirada de los países europeos, sino que pone el foco del talento musical por el resto del mundo. Gracias a mi conocimiento sobre este acontecimiento anual, este proyecto tiene como objetivo poder predecir, mediante la minería de datos, cuál será el ganador, la posición y la puntuación de los representantes de cada país que participarán en la edición de Eurovisión de 2023. Para ello se realizará un estudio de los principales festivales nacionales que más impactan en Eurovisión, como el Benidorm Fest de España y el Melodifestivalen de Suecia, para conocer los gustos musicales de estos países, y así poder inferenciar en los resultados europeos. Mediante información de las apuestas, características de las canciones en Spotify y un análisis de texto de la letra de cada una de ellas, se formarán bases de datos con la finalidad de crear modelos de predicción. Además, también se introducirá un índice (distancia) de amistad entre los países para conocer correctamente sus patrones de votaciones. Finalmente, se logra obtener unas predicciones con bastante precisión y observamos que nuestro mejor modelo para conseguirlo es la red neuronal.

An investigation into the prediction of thermal runaway in lithium-ion batteries subjected to abusive mechanical loading is comparable to a crash or impact test and a crucial safety measure that can prevent catastrophic events. Using an explicit crash simulation method, it is possible to simulate the indentation test model and predict the thermal runaway. However, it is important to note that this approach is associated with a significant time investment. State-of-the-art technologies that involve machine learning for prediction of thermal runaway can be utilised for faster predictions. The research conducted in this thesis focuses on a cell model that simulates an indentation test and a machine learning model that predicts the thermal runaway. This model successfully captures the behavior of an internal short circuit and is verified using experimental data from existing literature. Additionally, a workflow is generated in Altair Hyperstudy to produce data essential for training the machine learning model using an automated process. This data created is based on the design variables of the indenter (impacting body). This facilitates comprehension of potential deformation, damage and related patterns. The machine learning model is created using the Altair Physics AI tool and subsequently trained by the provided dataset. Data, as foundational resource is used to train the machine learning model. These datasets must be available in adequate quantities and of high quality for neural network training to discover the relationship between input and output. The outcomes of this Machine learning model yield an adequate degree of accuracy. The accuracy of these results is highly dependent on the model, quality of data derived from this dataset and the Hyperparameters used for the model training. Insights of the predicted results by ML model are of great use for Design consideration and validation of lithium-ion batteries. The outcomes of conducting the indentation test simula, Outgoing

La sobrequalificació és un factor molt present en els treballadors, es produeix quan disposen d’un nivell educatiu o un conjunt d’habilitats que excedeixen els requisits necessaris per desenvolupar el seu treball actual. Aquest problema pot sorgir a causa de diversos factors, com a canvis en els requisits d'ocupació, la falta d'oportunitats laborals adequades o la cerca d'ocupació en temps de crisi econòmica. Per altra banda, el Machine Learning és una branca de la intel·ligència artificial que se centra en el desenvolupament d'algorismes i models capaços d'aprendre i prendre decisions a partir de dades sense ser programats explícitament. Els algorismes de classificació són una de les principals tècniques utilitzades en l'aprenentatge automàtic per a classificar dades en diferents classes o categories. Aquest projecte té com a objectiu aplicar tècniques de classificació mitjançant l'aprenentatge automàtic (Machine Learning) per a abordar el problema de la sobrequalificació laboral. Per això, s’utilitzen quatre algoritmes de classificació que requereixen tècniques estadístiques per obtenir classificacions i avaluar el seu rendiment, que són: Arbres de decisió, Naive Bayes, Support Vector Machines i Regressió Logística. S’avaluaran els algoritmes esmentats i se seleccionarà el millor model de classificació per les dades del projecte. A partir d’aquest model es buscarà un perfil d’estudiant més probable a ser sobrequalificat.

Els sinistres, en el ram dels automòbils, són uns fets que es produeixen diàriament. Cada dia es produeixen avaries als vehicles o accidents. Tothom té contractada una assegurança obligatòria per cobrir els danys que produeixen a les persones de fora. El grup Catalana Occident és una empresa pionera en el món de les assegurances. Aquesta recopila tots els sinistres que han passat en unes bases de dades mensual. En aquestes bases hi ha una variable que ens indica la fase en la qual es troba el sinistre. Però a vegades ens trobem amb problemes de classificació. Per aquest motiu, l'ús del Machine learning o aprenentatge automàtic amb models de classificació pot ser un recurs per poder solucionar el problema de mala classificació de la fase en la qual es troba el sinistre. Aquests models ens permeten entrenar les dades per aprendre d'elles i obtenir classificacions i prediccions de les bases amb un nivell més alt de precisió. Aquest treball té com a objectiu aplicar tècniques de classificació mitjançant el Machine Learning per a abordar el problema de la mala classificació de les fases del sinistre. Per això, utilitzem tres mètodes de classificació que avaluaran el seu rendiment. Aquests són: Arbres de decisió, Naive Bayes i Regressió Logística. Compararem aquests mètodes i seleccionarem el millor model de classificació per les dades del treball.

Services and use cases in 5G and beyond networks are characterized by strict requirements such as ultra-low latency, increased capacity, and high user mobility. Moreover, these networks must be capable of satisfying these ambitious demands as well as anticipating and adapting to dynamically changing conditions in a quick and feasible manner. This study deals with the problem of determining the best time to readjust the user plane function (UPF) placement and session mapping configuration to avoid quality of service (QoS) degradation in the system due to user mobility. To this aim, we rely on machine learning (ML) techniques to anticipate poor QoS events and decide whether a reconfiguration procedure is required based on a pre-established QoS tolerance threshold. Specifically, an ML-based framework, called intelligent scheduling of the reconfiguration (ISR), is proposed to automate the reconfiguration process. This framework applies supervised ML methods, either regressors or classifiers, to predict the QoS values/status at a given time horizon. The simulation experiments revealed the proposed mechanism’s superiority compared to the established scheduling baseline. The ISR solution could not only keep the system QoS under desired values most of the time but also reduce the number of readjustment events by at least 50% compared to the baselines., This work has been supported by the Agencia Estatal de Investigación of Ministerio de Ciencia e Innovación of Spain underproject PID2019-108713RB-C51/MCIN/AEI/10.13039/501100011033 and through a predoctoral FPI scholarship., Postprint (published version)