Search

The wholesale electricity markets in Europe are undergoing major changes. The pursuit of a major integration for the development of the Internal Energy market is the main driver of this transformation. In this context, the Central Western Europe region implemented for the first time in May 2015 a new and complex mechanism to manage cross-border capacity allocation in the frame of the day-ahead market: the Flow-Based Market Coupling. This paper is the first to consider recent data of the electricity market to develop a predictive model for identifying convergence and congestion situations in the Central Western Europe region. The occurrence of convergence or not is considered as a binary outcome to build a probit model. This model combines meaningful information about the electricity market features of the interconnected countries between 2016 and 2018, including for the first time data on renewable energy sources forecasts for this type of studies. The results of this study identify the major role that Germany and France play in the integration process. The estimated coefficients of the model show that the strong development of solar and wind power in Germany appears as an important driver of congestion in the Central Western Europe. In order to improve the benefits of market integration and the further development of the Internal Energy market, the results of this study encourage policy-makers to promote cooperation and coordination among all the actors involved in the electricity market and to pursue the expansion of the power grid to minimize the occurrence of congestion situations. This project has been funded by the National Distance Education University (UNED), Spain, Applied Economic and Economic History department for open access publishing.

The convective heat transfer in a turbulent boundary layer (TBL) on a flat plate is enhanced using an artificial intelligence approach based on linear genetic algorithms control (LGAC). The actuator is a set of six slot jets in crossflow aligned with the freestream. An open-loop optimal periodic forcing is defined by the carrier frequency, the duty cycle and the phase difference between actuators as control parameters. The control laws are optimised with respect to the unperturbed TBL and to the actuation with a steady jet. The cost function includes the wall convective heat transfer rate and the cost of the actuation. The performance of the controller is assessed by infrared thermography and characterised also with particle image velocimetry measurements. The optimal controller yields a slightly asymmetric flow field. The LGAC algorithm converges to the same frequency and duty cycle for all the actuators. It is noted that such frequency is strikingly equal to the inverse of the characteristic travel time of large-scale turbulent structures advected within the near-wall region. The phase difference between multiple jet actuation has shown to be very relevant and the main driver of flow asymmetry. The results pinpoint the potential of machine learning control in unravelling unexplored controllers within the actuation space. Our study furthermore demonstrates the viability of employing sophisticated measurement techniques together with advanced algorithms in an experimental investigation., Comment: 20 pages, 13 figures

Since the adoption of Convolutional Neural Networks (CNNs) for vision tasks, both hardware and software developments have rushed to make these architectures plausible on constrained hardware. Recently, transformers have shaken the world in Natural Language Processing (NLP) tasks and promise to also be the future of vision Deep Learning (DL). However their strengths are yet to be determined. In this work we: (1) Introduce the concept of a Vision Transformer (ViT), (2) Survey the literature for efficient mobile ViTs architectures, (3) Generate a latency-score comparison of the presented SoTA works, (4) Analyze their effectiveness and viability under a real-world task (facial landmark estimation), (5) Propose a more efficient configuration to a selected model from the SoTA, (6) Discuss ViTs in relation to CNNs. RESUMEN Desde el uso de Redes Neuronales Convolucionales (CNNs) en tareas de Visión, el desarrollo de ambos hardware y software ha avanzando rápidamente para hacer posible su ejecución en dispositivos móviles. Recientemente, los transformers han revolucionado lo que es posible en Procesamiento del Lenguaje Natural (NLP) y ahora prometen ser también el futuro de la visión por computador por aprendizaje profundo (DL). Sin embargo, están aún por determinarse sus puntos fuertes en estas tareas. En este trabajo realizamos lo siguiente: 1. Introducimos el concepto de Vision Transformer (ViT), 2. Revisamos sistemáticamente el estado del arte (SoTA) en arquitecturas ViT para dispositivos móviles, 3. Generamos una comparación de eficiencia de los trabajos del SoTA presentados, 4. Analizamos su efectividad y viabilidad en una tarea del mundo real (detección de landmarks faciales), 5. Proponemos una configuración más eficiente de un modelo del SoTA, 6. Y finalmente discutimos los ViTs en relación a las CNNs.

El trading algorítmico y automatizado está fuera del alcance de los inversores minoristas, debido a su complejidad técnica y a los costes de desarrollo. Algunos brokers ofrecen este servicio pero a un coste inicial y una cuota mensual. Los usuarios dependen de la plataforma y servidores del broker para ejecutar las operaciones, dando al broker demasiado control sobre los datos y actividad de trading del usuario. La solución implementada en este proyecto es una plataforma de trading automatizado FOSS y personalizable que puede ser usada por comunidades de inversores minoristas para implementar sus estrategias de trading y por inversores individuales para automatizar su actividad de trading. La plataforma está diseñada para ser altamente escalable y disponible, desarrollada con un enfoque cloud-first y usando las últimas tecnologías en el campo de los sistemas distribuidos, como Kubernetes y servicios específicos de proveedores de cloud públicos. La implementación específica del bot de trading está fuera del alcance de este proyecto. No obstante, la plataforma está diseñada para ser fácilmente extensible y personalizable, mediante el uso de contenedores docker. Abstract: Algorithmic and automated trading is out of the scope of retail investors, due to the technical complexity of it as well as the costs of development. Some brokers offer this service but at an upfront cost and a monthly fee. Users depend on the broker’s platform and servers to execute the trades, giving the broker too much control over the user’s data and trading activity. The solution implemented in this project is a customizable FOSS automated trading platform that can be used by retail investor communities to implement their trading strategies and by individual investors to automate their trading activity. The platform is designed to be highly scalable and available, developed with a cloud-first approach, and using the latest technologies in the field of distributed systems, such as Kubernetes and public cloud providers’ specific services. The specific implementation of the trading bot is out of the scope of this project. Nonetheless, the platform is designed to be easily extensible and customizable, through the use of docker containers.

Este trabajo de fin de máster presenta un análisis exhaustivo y minucioso del estado del arte actual en la generación de descripciones de images, comúmente denominado subtitulado de imágenes naturales (Natural Image Captioning (NIC)), con especial atención al subtitulado de imágenes médicas (Medical Image Captioning (MIC)), específicamente en la generación automática de informes radiológicos. El trabajo revisa de forma intensiva la literatura existente los métodos que se utilizan en Natural Image Captioning (NIC) y Medical Image Captioning (MIC), con especial atención a estudiar los métodos de NIC que sirven de base para los de MIC. Además, en este trabajo se identifican los avances en ambos campos, discutiendo los retos y limitaciones, y explorando potenciales direcciones de investigación. La tesis subraya la importancia del subtitulado de imágenes, que combina los métodos de visión por computador (Computer Vision (CV)) y procesamiento de lenguaje natural (Natural Language Processing (NLP)) para generar descripciones de imágenes significativas y legibles para los humanos. Además, destaca el potencial de la inteligencia artificial (Artificial Intelligence (AI)) en medicina, en particular en el campo de la radiología, y su capacidad para influir en la atención sanitaria mediante la generación automática de informes, promoviendo la precisión, la fiabilidad, la repetibilidad y la comunicación eficaz entre los profesionales sanitarios. La tesis explora las cuestiones fundamentales que se abordan en el subtitulado de imágenes y los diversos métodos desarrollados en el campo de la MIC, incluidos los enfoques basados en la recuperación, en plantillas, en redes neuronales profundas, en la atención y en los híbridos. Además, este trabajo identifica a los métodos de localización de enfermedades para la generación de informes como un área crítica de investigación en este campo, donde los modelos destinados a la práctica clínica deberían tender a converger. Estos enfoques proporcionan una alternativa más explicable e interpretable a la naturaleza común de caja negra (black-box) de los métodos de aprendizaje profundo (Deep Learning (DL)). De hecho, estudios recientes sugieren que los nuevos modelos de este campo deberían ser capaces de emparejar informes de texto con visualizaciones interactivas y considerar el historial del paciente, mejorando aún más el objetivo de la AI en medicina: proveer de atención personalizada a cada paciente. A través de una metodología propuesta, la tesis realiza una revisión del estado del arte en los métodos empleados en NIC y MIC, enumerando y describiendo los conjuntos de datos disponibles para MIC, así como analizando los métodos de evaluación existentes para comparar los modelos desarrollados de NIC y MIC. Las publicaciones encontradas se analizan y discuten en profundidad, resaltando las limitaciones y posibles implicaciones en la práctica clínica real. La tesis concluye identificando futuras direcciones de investigación, como la integración de la comprensión temporal de las imágenes, la interpretación de varias modalidades radiológicas y la integración de la información del historial del paciente en el proceso de aprendizaje del modelo. También se hace hincapié en la necesidad de disponer de conjuntos de datos diversos, multimodales, multicentros y representativos para mejorar la calidad de los informes generados y estar más cerca de lograr una inteligencia artificial médica general (General Medical Artificial Intelligence (GMAI)). En general, esta tesis contribuye a la comprensión de las aplicaciones en el campo de la radiología, específicamente en la generación automática de informes. Se revisa a fondo la literatura para proporcionar una visión de las relaciones, los avances, desafíos y mejoras potenciales, en NIC y MIC. ABSTRACT This Master’s thesis presents a comprehensive analysis of the current state of the art in Natural Image Captioning (NIC), with a special focus on Medical Image Captioning (MIC) and its application in automatic radiology report generation. The work reviews existing literature for NIC and MIC methods, exhaustively studying NIC methods that serve as foundations for MIC, identifying advancements in both fields, discussing challenges and limitations, and exploring potential research directions. The thesis highlights the importance of image captioning, which combines Computer Vision (CV) and Natural Language Processing (NLP) to generate meaningful and human-readable descriptions of images. It emphasizes the potential of Artificial Intelligence (AI) in medicine, particularly in the field of radiology, and its ability to impact healthcare through automated report generation, promoting accuracy, reliability, repeatability, and effective communication among healthcare professionals. The thesis explores the fundamental questions addressed in image captioning and the various methods developed in the field of MIC, including retrieval-based, template based, neural network-based, attention-based, and hybrid approaches. Additionally, this thesis identifies disease localization methods for report generation as a critical area, where models intended for clinical practice should tend to converge. These approaches provide a more explainable and interpretable alternative to the common black-box nature of Deep Learning (DL) methods. In fact, recent studies suggest that novel models in the field should be able to pair text reports with interactive visualizations and consider the patient’s history, further enhancing the patient-centric objective of AI in medicine. Through a proposed methodology, the thesis conducts a state-of-the-art review in NIC and MIC, enumerating and describing available datasets for MIC as well as analyzing existing evaluation methods to compare NIC and MIC developed models. The publications found are thoroughly analyzed and discussed, highlighting limitations and potential implications in real clinical practice. The thesis concludes by identifying future research directions, such as the integration of temporal understanding of images, the interpretation of various radiology modalities, and the integration of patient history information in the model’s learning process. The need for diverse, multi-modal, multi-centric, and representative datasets is also emphasized to enhance the quality of generated reports and be closer on achieving a General Medical Artificial Intelligence (GMAI). Overall, this thesis contributes to the understanding of AI applications in the radiology field, specifically in automated report generation. It thoroughly reviews the literature to provide insights into the advancements, challenges, and potential improvements, in NIC and MIC.

Mutation testing is a valuable technique for measuring the quality of test suites in terms of detecting faults. However, one of its main drawbacks is its high computational cost. For this purpose, several approaches have been recently proposed to speed-up the mutation testing process by exploiting computational resources in distributed systems. However, bottlenecks have been detected when those techniques are applied in large-scale systems. This work improves the performance of mutation testing using large-scale systems by proposing a new load distribution algorithm, and parallelising different steps of the process. To demonstrate the benefits of our approach, we report on a thorough empirical evaluation, which analyses and compares our proposal with existing solutions executed in large-scale systems. The results show that our proposal outperforms the state-of-the-art distribution algorithms up to 35% in three different scenarios, reaching a reduction of the execution time of—at best—up to 99.66%, This work was supported by the Spanish MINECO/FEDER project under Grants PID2021- 122270OB-I00, TED2021-129381B-C21 and PID2019-108528RBC22, the Comunidad de Madrid project FORTE-CM under Grant S2018/TCS-4314, Project S2018/TCS-4339 (BLOQUES-CM) cofunded by EIE Funds of the European Union and Comunidad de Madrid and the Project HPC-EUROPA3 (INFRAIA-2016-1-730897), with the support of the EC Research Innovation Action under the H2020 Programme

This paper presents a framework for the efficient management of renewable energies to charge a fleet of electric buses (eBuses). Our framework starts with the prediction of clean energy time windows, i.e., periods of time when the production of clean energy exceeds the demand of the country. Then, the optimization phase schedules charging events to reduce the use of non-clean energy to recharge eBuses while passengers are embarking or disembarking. The proposed framework is capable of overcoming the unstable and chaotic nature of wind power generation to operate the fleet without perturbing the quality of service. Our extensive empirical validation with real instances from Ireland suggests that our solutions can significantly reduce non-clean energy consumed on large data sets, This work received funding from the Sustainable Energy Authority of Ireland (SEAI) Research, Development and Demonstration (RDD) 2019 programme under the grant number 19/ RDD/519

Este proyecto tiene como objetivo presentar una plataforma de crowdfunding basada en blockchain que utiliza contratos inteligentes para superar las limitaciones de las plataformas de crowdfunding tradicionales. El objetivo del estudio es investigar el potencial de la tecnología blockchain para mejorar el proceso de crowdfunding y proporcionar un entorno seguro, transparente y de confianza para las transacciones financieras entre múltiples usuarios. Para lograr este objetivo, se realizó una revisión de la literatura para identificar los conceptos clave y las mejores prácticas relacionadas con el crowdfunding, la tecnología blockchain y los contratos inteligentes. Además, se analizaron las plataformas de crowdfunding existentes para identificar sus limitaciones en términos de transparencia, seguridad y confianza. Los resultados de la investigación sugieren que la tecnología blockchain y los contratos inteligentes tienen el potencial de revolucionar la industria del crowdfunding proporcionando un entorno seguro, transparente y descentralizado. El estudio propone un prototipo de plataforma de crowdfunding basada en blockchain que utiliza contratos inteligentes y que se evaluó en términos de viabilidad, usabilidad, seguridad y rendimiento. En general, los resultados indican que la plataforma propuesta puede superar las limitaciones de las plataformas de crowdfunding tradicionales y proporcionar un entorno seguro, transparente y fiable para las transacciones financieras. En conclusión, este estudio contribuye al creciente cuerpo de conocimientos sobre la tecnología blockchain y sus aplicaciones potenciales en la industria financiera, específicamente en el crowdfunding. Abstract: This project aims to present a blockchain-based crowdfunding platform that uses smart contracts to overcome the limitations of traditional crowdfunding platforms. The purpose of the study is to investigate the potential of blockchain technology in improving the crowdfunding process and providing a secure, transparent, and trustworthy environment for financial transactions between multiple users. To achieve this objective, a literature review was conducted to identify the key concepts and best practices related to crowdfunding, blockchain technology, and smart contracts. Additionally, existing crowdfunding platforms were analyzed to identify their limitations in terms of transparency, security, and trust. The research findings suggest that blockchain technology and smart contracts have the potential to revolutionize the crowdfunding industry by providing a secure, transparent, and decentralized environment. The study proposes a prototype of a blockchain-based crowdfunding platform using smart contracts, which was evaluated in terms of its feasibility, usability, security, and performance. Overall, the results indicate that the proposed platform can overcome the limitations of traditional crowdfunding platforms and provide a secure, transparent, and trustworthy environment for financial transactions. In conclusion, this study contributes to the growing body of knowledge on blockchain technology and its potential applications in the financial industry, specifically in crowdfunding.

El objetivo de este trabajo de fin de carrera es realizar una investigación sobre la predicción de interacciones proteína-proteína aprovechando la potencia de las redes neuronales de grafos (GNN). Este estudio explora hiperparámetros y arquitecturas para descubrir su impacto en la precisión de la predicción. Las GNN se utilizan como una valiosa herramienta para captar las características estructurales inherentes a los datos de interacción de proteínas. Para lograr los objetivos de esta investigación, se utiliza un conjunto de datos entre interacción de proteínas para comparar y contrastar el rendimiento de diferentes enfoques de GNN. Mediante una serie de experimentos y un preciso análisis, este estudio evalúa las variaciones en el rendimiento y tiempo de ejecución e intenta identificar los ajustes óptimos de los parámetros para los modelos basados en GNN en el contexto de la predicción de interacciones proteína-proteína. Los resultados empíricos de esta investigación ponen de manifiesto la importancia del ajuste de parámetros en los modelos basados en GNN y en machine learning en general y los factores subyacentes que influyen en la efectividad para predecir las interacciones proteína-proteína. Los conocimientos obtenidos a partir de esta investigación aportan contribuciones al análisis de conjuntos de datos proteína-proteína y ofrecen ideas para mejorar futuros modelos dentro del área. Aunque esta tesis se centra en la comparación de diferentes modelos con distintos hiperparámetros, no pretende realizar avances revolucionarios en el estado del arte actual de la predicción de interacciones proteína-proteína o de las redes neuronales de grafos. Su objetivo es más bien mejorar nuestra comprensión y optimizar la eficiencia de los modelos existentes en este campo. Mediante la exploración de varios modelos e hiperparámetros, este estudio da importancia a los puntos fuertes y las limitaciones de diferentes enfoques para la predicción de interacciones proteína-proteína. Estos hallazgos pueden servir como base para optimizar modelos futuros y pueden ayudar a los investigadores a tomar mejores decisiones en cuanto a la selección de hiperparámetros y arquitecturas apropiadas. Este conocimiento puede mejorar el análisis de conjuntos de datos proteína-proteína al proporcionar una comprensión más clara de los factores que influyen en el rendimiento, la precisión y el tiempo de ejecución de los diferentes modelos. Por lo tanto, los resultados de este estudio contribuyen a la mejora continua de los modelos de predicción de la interacciónes entre proteínas, como las redes neuronales de grafos, mediante el perfeccionamiento de las metodologías de análisis, la optimización de la selección de hiperparámetros y la presentación de algunas ideas sobre las complejidades dentro de las redes de proteínas. ABSTRACT This final thesis aims to investigate the prediction of protein-protein interactions by leveraging the power of Graph Neural Networks (GNNs). This study explores a range of hyperparameters and architectures to uncover their impact on the prediction accuracy. GNNs are used as a valuable tool to capture the structural characteristics inherent in protein interaction data. To achieve the research objectives, a protein interaction dataset is used to compare and contrast the performance of different GNN approaches. Through a series of welldesigned experiments and precise analysis, this study evaluates the variations in performance and tries to identify the optimal parameter settings for GNN-based models in the context of protein-protein interaction prediction. The empirical findings of this research highlight the importance of parameter tuning in GNN-based models and the underlying factors influencing their effectiveness in predicting protein-protein interactions. The insights obtained from this investigation provide contributions to the analysis of protein-protein datasets and offer potential ideas for improving future models within the domain. While this thesis focuses on comparing different models with varying hyperparameters, it does not claim to make groundbreaking advancements in the current state-of-the-art of protein-protein interaction prediction or Graph Neural networks. Rather, its objective is to enhance our understanding and optimize the efficiency of existing models in this field. By exploring various models and hyperparameters, this study sheds light on the strengths and limitations of different approaches for predicting protein-protein interactions. The findings can be used as a foundation for optimizing future models, enabling researchers to make more informed choices regarding the selection of appropriate hyperparameters and architectures. This knowledge can improve the analysis of protein-protein datasets by providing a clearer understanding of the factors influencing the performance, accuracy and time of execution of different models. Therefore, the outcomes of this study contribute to the continuous improvement of protein-protein interaction prediction models such as Graph Neural Networks by refining analysis methodologies, optimizing hyperparameter selection, and presenting some ideas of the complexities within protein networks.

Asynchronous communication practices, including asynchronous video messaging, have become more important in today’s globalized workplace, offering greater flexibility and productivity. In this way, summarizing video messages can further increase employee productivity by allowing them to quickly access key points rather than watching entire videos. However, effectively summarizing these messages remains a challenge. Recent research has shown that Large Language Models (LLMs) can achieve state-ofthe-art performance on summarization tasks, but their size and associated hardware requirements make them expensive to deploy. Therefore, this work explores the use of model compression techniques such as knowledge distillation (Shrink and Fine-Tune approach) and quantization (Dynamic Quantization approach) to create smaller and more efficient LMs while maintaining similar performance. After applying these model compression approaches to the BART and FLAN -T5 models, the study continued to evaluate the performance of these smaller, more efficient LMs. The goal was to determine if they could indeed maintain a comparable level of performance to their larger, more resource-intensive counterparts. Although ROUGE scores remained relatively stable in most cases, a manual assessment showed a significant deterioration in the quality of the summaries when the number of decoder layers decreased. This highlights the limitations of the ROUGE metric and the importance of human evaluations. For future work, the study suggests exploring other compression approaches such as Direct Knowledge Distillation and Quantization Aware Training (QAT), as well as testing compression of other open source models suitable for commercial use, such as OpenLLaMA and Falcon. This study contributes to the ongoing discussion on the use of AI technologies to improve productivity in the workplace and demonstrates the importance of carefully balancing the tradeoffs between model size, computational cost, and summary quality.

Esta tesis explora el desarrollo e implementación de una solución basada en Salesforce para optimizar las operaciones comerciales y mejorar la gestión de clientes de una empresa de venta de productos de limpieza. El análisis exhaustivo del modelo de negocio actual, el plan de marketing detallado y la identificación de fortalezas, debilidades, oportunidades y amenazas son aspectos clave de la investigación. Se presenta el estado del arte y la descripción de la solución elegida, junto con los avances y soluciones existentes en el sector. La investigación se enfoca en el contexto del problema, los requisitos iniciales, la evaluación de alternativas y viabilidad, y las tecnologías y metodologías utilizadas. Se incluye un presupuesto detallado y se establece un plan de migración de datos. Esta tesis proporciona una visión general y completa del desarrollo e implementación de la solución, con énfasis en el análisis, desarrollo iterativo y potencial para trabajos futuros. ABSTRACT This master's thesis explores the development and implementation of a Salesforcebased solution to optimize business operations and improve customer management of a company that sells cleaning products. Key aspects of the research include a comprehensive analysis of the current business model, a detailed marketing plan, and the identification of strengths, weaknesses, opportunities, and threats. The state of the art and the description of the chosen solution, along with the latest advancements and existing solutions in the sector, are presented. The research focuses on the problem's context, initial requirements, alternative evaluation and feasibility, and the technologies and methodologies used. A detailed budget is provided, and a data migration plan is established. This thesis provides a comprehensive overview of the solution's development and implementation, with an emphasis on analysis, iterative development, and potential for future work.

This Master’s thesis explores deep learning experimentation tools and their assessment for optimizing the experimentation process in the field of deep learning. The objective is to identify the strengths and weaknesses of various popular deep learning experimentation tools and provide insights into their effectiveness based on a benchmark study. The thesis begins with an introduction highlighting the importance of experimentation in deep learning and outlining the problems and objectives of the research. It also presents the methods employed for the study, including the theoretical framework covering the fundamentals of deep learning and an overview of relevant tools and frameworks. The focus then shifts to deep learning experimentation tools, where an introduction to these tools is provided, followed by a comprehensive listing of various deep learning experimentation tools available. Noteworthy tools such as Neptune.ai, Weights & Biases, Comet, TensorBoard and Guild.ai are explored in detail. To evaluate and compare the selected tools, a benchmark design is proposed. This includes problem and dataset selection, as well as the design of the deep neural network and hyperparameters for experimentation. The benchmark execution is carried out using the selected tools, and assessment criteria are established to evaluate their performance. The results of the benchmark study are then presented, highlighting the execution and performance of each tool. A comparative analysis is conducted, providing insights into the strengths and weaknesses of each tool. This analysis forms the basis for drawing conclusions about the effectiveness and suitability of the tools for different user profiles, including students, researchers/professors, and professionals. In the final section, the thesis concludes with an analysis of the results obtained and their implications for deep learning experimentation. The contributions of the research are discussed, and future work is outlined to further advance the field of deep learning experimentation. Overall, this Master’s thesis serves as an exploration and assessment of deep learning experimentation tools, offering valuable insights for researchers, practitioners, and decision-makers in the field. The findings contribute to enhancing the understanding and selection of appropriate tools to streamline and optimize the deep learning experimentation process.

La información del ámbito biomédico es fácilmente representable en forma de grafo. Conteniendo de manera compacta y simple información sobre elementos, como fármacos y enfermedades, y sus relaciones. Este trabajo fin de máster se enmarca en el contexto del reposicionamiento de fármacos utilizando grafos como forma de representación de la información del ámbito biomédico disponible, procedente de las bases de datos del proyecto DISNET. Para aprovechar el formato en el que se encuentran los datos se aplica el marco de trabajo que presentan las Graph Neural Networks (GNNs). Se enfoca como un problema de predicción de enlaces, se predice la existencia de enlaces que indiquen la relación de tratamiento terapéutico entre enfermedades y fármacos. Aplicando GNNs sobre el grafo construido se consiguen generar hipótesis de reposicionamiento de fármacos. Se realiza un breve análisis desde la perspectiva biomédica que verifica la validez de estas hipótesis. Además, se explora el impacto que supone añadir características a los nodos del tipo fármaco. Estas se obtienen a través de la aplicación de GNNs para representar vectorialmente la estructura molecular de los fármacos, también en forma de grafo. Dados los buenos resultados reportados en las métricas reportadas y la posterior validación de algunas de las hipótesis, se concluye que, el reposicionamiento de fármacos a través de GNNs tiene potencial para presentar grandes avances en el campo y es un área con especial interés para futuras investigaciones. ABSTRACT Drug repositioning, the discovery of new therapeutic uses for existing drugs, is a technique that has gained relevance in recent years. It greatly simplifies the drug discovery process, reducing the costs and risks associated with de novo development. By using Artificial Intelligence to guide the repositioning process, costs and risks can be further reduced. Information from the biomedical domain is easily represented in the form of a graph. It contains in a compact and simple way information about elements, such as drugs and diseases, and their relations. This master's thesis is framed in the context of drug repositioning using graphs as a way of representing the biomedical information available from the DISNET project databases. To take advantage of the data format, the framework presented by Graph Neural Networks (GNNs) is applied. It is approached as a link prediction problem, predicting the existence of links that indicate the therapeutic treatment relationship between diseases and drugs. By applying GNNs on the constructed graph, drug repurposing hypotheses can be generated. A brief analysis from a biomedical perspective is carried out to verify the validity of these hypotheses. In addition, the impact of adding features to the drug nodes is explored. These are obtained through the application of GNNs to generate vectors that represent the molecular structure of drugs, also in the form of a graph. Given the good results in the reported metrics and the subsequent validation of some of the hypotheses, it is concluded that drug repositioning through GNNs has the potential to present great advances in the field and is an area of particular interest for future research.

La creciente complejidad del software ha llevado a una mayor vulnerabilidad a ciertas amenazas de seguridad que explotan la optimización de los recursos utilizados para ejecutar (y mantener) dicho software. Fuzzing (también conocido como fuzz testing) ha surgido como un enfoque para identificar dichas vulnerabilidades relacionadas con técnicas de optimización. Las CPU modernas utilizan técnicas de optimización como el uso de “pipelines” de instrucciones, donde las instrucciones se procesan etapa por etapa y, por lo tanto, se pueden ejecutar de forma paralela. Sin embargo, la técnica no está exenta de obstáculos, ya que las etapas lentas, como los accesos a memoria a direcciones que no están presentes en la memoria caché, pueden detener momentáneamente la pipeline mientras se ejecutan (los desencadenantes de la desaceleración de una pipeline de instrucciones se denominan "dependencias"). Para hacer frente a la posible pérdida de eficiencia causada por tales eventos, los procesadores emplean lo denominado como "ejecución especulativa". La ejecución especulativa es una técnica mediante la cual un procesador ejecuta instrucciones de manera especulativa mientras se resuelve una dependencia de pipeline. Usando el ejemplo anterior de un acceso a memoria, esto implicaría que el procesador simule la ejecución de tantas instrucciones como pueda mientras se lleva a cabo dicho acceso. Cuando se ha producido el acceso a memoria, el procesador puede confirmar y aplicar los cambios realizados por las instrucciones que simuló si el acceso fue válido. Si, por ejemplo, el proceso que realiza el acceso no tiene permiso para acceder a la dirección de memoria del mismo, entonces el procesador simplemente revertirá la mayoría de los cambios hechos y procederá a manejar la excepción lanzada después de acceder a una dirección ilegal. Sin embargo, los cambios revertidos por una CPU después de abortar la ejecución especulativa no incluyen al estado de las cachés. Por lo tanto, si un agente malicioso pudiera recuperar el contenido de la memoria caché, podría ganar acceso a información que, de lo contrario, se supondría que es inaccesible al analizar las posibles vulnerabilidades de un producto software. Revizor es una herramienta desarrollada para encontrar automáticamente dichas vulnerabilidades esquivas. Originalmente se implementó para hacer fuzzing sobre programas ensamblador generados aleatoriamente de la arquitectura del conjunto de instrucciones x86, con la intención de encontrar posibles fugas microarquitectónicas que podrían explotarse cuando las fallas antes mencionadas están presentes en el código ensamblador de un programa. El proyecto detrás de este trabajo de fin de grado amplía el alcance de Revizor, permitiendo a la herramienta hacer fuzzing con programas de la arquitectura del conjunto de instrucciones RISC-V. Además, proporciona a la herramienta un script de depuración, que facilitará la expansión de la herramienta para sus futuros desarrolladores. ABSTRACT The growing complexity of software has led to an increased presence of vulnerabilities regarding security threats that exploit the optimization of the resources used to run (and maintain) such software. Fuzzing (also referred to as fuzz testing) has emerged as an approach for identifying such vulnerabilities concerning optimization techniques (among others). Modern CPUs make use of optimization techniques such as instruction pipelining, where instructions are processed stage by stage and can therefore be run in a parallel fashion. The technique however is not without holdback, since slow stages, such as memory accesses to addresses that are not present in cache, may momentarily bring the pipeline to a halt while they are being executed (the triggers of the deceleration of an instruction pipeline are called dependencies). To tackle the possible loss of efficiency caused by such events, processors employ what is called "speculative execution". Speculative execution is a technique by which a processor speculatively executes instructions while a pipeline dependency is being solved. Using the previous example of a memory access, this would mean that the processor would simulate the execution of as many instructions as it can while the memory access is taking place. When the memory access has taken place, the processor may commit the changes done by the instructions it simulated if the memory access was valid. If, for example, the process performing the memory access does not have the permission to access the address, then the processor will simply revert most changes and proceed to handle the exception thrown after accessing an illegal address. However, the changes reverted by a CPU after aborting speculative execution do not include the state of the caches. Hence, if a malicious agent were able to retrieve the contents of the cache, they would be able to retrieve information which should otherwise be assumed to be inaccessible when analysing the potential vulnerabilities of a software product. Revizor is a tool developed to automatically find such elusive vulnerabilities. It was originally implemented to fuzz randomly generated assembly programs of the x86 Instruction Set Architecture (or ISA for short), with the intention of finding potential microarchitectural leaks which could be exploited when the aforementioned bugs are present in the assembly code of a program. The work behind this bachelor’s thesis extends the scope of Revizor, allowing it to fuzz programs of the RISC-V ISA. It additionally provides the tool with a debugging script, that will make the expansion of the tool easier for its future developers.

La modelización probabilística de tópicos es una técnica de aprendizaje automático no supervisada que, dada un conjunto de documentos, es capaz de analizar, detectar y agrupar automáticamente palabras que mejor caracterizan tópicos comunes presentes en el conjunto. Sin embargo, muchas veces estamos interesados en saber qué relaciona estos documentos más allá de los patrones más característicos o conjuntos de palabras en el conjunto. En consecuencia, surgió la generación de etiquetas de tópicos, que buscaba generar una etiqueta que caracterizara el conjunto de documentos de forma más interpretable que un grupo de palabras que, a priori, no sabemos qué relación tienen entre sí. Actualmente, se siguen investigando nuevas formas de generar estas etiquetas de temas de manera automática y fácilmente comprensibles. A la vez, recientemente han aparecido Modelos de Lenguaje con fin conversacional, los cuales están entrenados para comprender y generar diálogos entre humanos y máquinas. Estos modelos presentan capacidades más allá de la habilidad de tener una conversación, como por ejemplo ChatGPT, que ha demostrado poder redactar de forma autónoma correos electrónicos o redacciones sobre un tema específico, por ejemplo. Los modelos conversacionales presentan un potencial aparente no solo para ser aplicados en tareas recreativas, sino que también pueden ser útiles para otras tareas, como se indica en la publicación de Sallam "ChatGPT Utility in Healthcare Education, Research, and Practice: Systematic Review on the Promising Perspectives and Valid Concerns" [1], donde el autor analiza 60 publicaciones que hablaban de los beneficios de usar ChatGPT en diferentes tareas, como el análisis eficiente de conjuntos de datos o la generación de código para la investigación en salud. Ante este hecho, el objetivo de este Proyecto Final de Máster es estudiar la capacidad que pueden tener los modelos conversacionales para generar automáticamente y de manera no supervisada etiquetas para tópicos probabilísticos, dado un conjunto de palabras clave representativas del tema, siguiendo una metodología a la cual nos referiremos como Etiquetado Conversacional de Tópicos Probabilísticos (CPTL o Conversational Probabilistic Topic Labelling en inglés). También comparamos el rendimiento de estos modelos conversacionales con el rendimiento de un modelo de lenguaje específico para tareas, entrenado para generar etiquetas de temas. ABSTRACT Probabilistic topic modelling is an unsupervised machine learning technique that, given a set of documents, is capable of scanning, detecting patterns of words and phrases, and automatically grouping words that best characterize a topic. Many times, however, we are interested in knowing what relates these documents beyond the most characteristic patterns or sets of words in the set. Consequently, the generation of topic labels appeared, which sought to generate a label that would characterize the set of documents in a more interpretable way than having a group of words that we, a priori, do not know the relationship they have with each other. Currently, new ways of generating these topic labels that are easily understandable automatically are still being investigated. At the same time, Neural Language Models based on Neural Networks with conversational purpose have recently emerged, which are trained to understand and generate dialogues between humans and machines. These models possess capabilities beyond the ability to engage in conversation, such as ChatGPT, which has demonstrated the ability to autonomously compose emails or write about a specific topic, for example. Conversational models present an apparent potential to not only have recreational applications, but can also be useful for other tasks, as stated in Sallam’s publication "ChatGPT Utility in Healthcare Education, Research, and Practice: Systematic Review on the Promising Perspectives and Valid Concerns" [1], where the author analysed 60 publications that talked about the benefits of using ChatGPT in different tasks, such as efficient analysis of datasets or code generation for health care research. Given this fact, the purpose of this Final Master’s Project is to study the capacity that conversational models may have to automatically and unsupervisedly generate tags for probabilistic topics given a set of keywords representative of the topic, following a methodology which we will refer as Conversational Probabilistic Topic Labelling (CPTL). We also compare the performance of these conversational models with the performance of a task-specific language model trained to generate topic labels.

This MSc Thesis (Master Thesis) deals with the analysis of a multi-agent rescue system, with the aim of understanding its performance and improving its efficiency in emergency situations. This master thesis addresses the problem of facing emergency situations and proposes a solution through simulation and analysis of multi-agent systems in rescue environments. The main objective is to analyze and improve a simulation platform to efficiently train and predict emergency scenarios. To achieve this, different aspects related to agents and multi-agent systems were investigated, as well as the methodologies applied to reach the proposed objectives. Models of synthetic individuals that behave similarly to real human beings were implemented, which allowed a more realistic representation of human interactions and behavior in critical situations. The tests and results obtained showed that the proposed solution, based on the simulation of multi-agent systems, offers an effective approach to deal with emergency situations. The developed algorithms, such as Dijkstra’s algorithm for agent movement in maps with multiple police officers and civilians to rescue, proved to be efficient and provided promising results. Among the main results obtained, it is highlighted that, for situations with a small number of agents, the algorithm based on collaboration between agents achieved better results than the randomized algorithm. However, in scenarios with a high number of agents, it was observed that the randomized approach offered better results in terms of efficiency and performance. These results highlight the importance of defining suitable algorithms adapted to the specific characteristics of each emergency situation. They also emphasize the need to further develop and implement complex rescue algorithms based on team training to further improve the efficiency of rescue operations in emergency situations. In summary, this MSc Thesis addresses the problem of emergency situations by proposing a solution based on simulation and analysis of multi-agent systems. The results obtained show the effectiveness of the proposed solution and underline the importance of defining suitable algorithms to achieve an effective rescue in different emergency scenarios.

We present MESMOC+, an improved version of Max-value Entropy search for Multi-Objective Bayesian optimization with Constraints (MESMOC). MESMOC+ can be used to solve constrained multi-objective problems when the objectives and the constraints are expensive to evaluate. It is based on minimizing the entropy of the solution of the optimization problem in function space (i.e., the Pareto front) to guide the search for the optimum. The cost of MESMOC+ is linear in the number of objectives and constraints. Furthermore, it is often significantly smaller than the cost of alternative methods based on minimizing the entropy of the Pareto set. The reason for this is that it is easier to approximate the required computations in MESMOC+. Moreover, MESMOC+’s acquisition function is expressed as the sum of one acquisition per each black-box (objective or constraint). Therefore, it can be used in a decoupled evaluation setting in which it is chosen not only the next input location to evaluate, but also which black-box to evaluate there. We compare MESMOC+ with related methods in synthetic, benchmark and real optimization problems. These experiments show that MESMOC+ has similar performance to that of state-of-the-art acquisitions based on entropy search, but it is faster to execute and simpler to implement. Moreover, our experiments also show that MESMOC+ is more robust with respect to the number of samples of the Pareto front, The authors acknowledge financial support from Spanish Plan Nacional I + D+i, grant PID2019-106827 GB-I00/ AEI/ 10.13039/501100011033

123 p. La presente tesis incluye 3 contribuciones al área llamada "learning from crowds", que estudia los métodos de aprendizaje basados en datos etiquetados por medio del "crowdsourcing". Estas etiquetas se caracterizan por tener una incertidumbre asociada debido a que la fiabilidad de las personas anotadoras no está garantizada. En primer lugar, se propone un nuevo método de "label aggregation", llamado"domain aware voting", una extensión del popular y simple método "majority voting" que tiene en cuenta la variable descriptiva, obteniendo resultados mejores especialmente cuando hay una mayor escasez de etiquetas. La segunda contribución consiste en la propuesta de un nuevo marco de etiquetado, "candidate labelling", que permite a las personas anotadoras expresar sus dudas acerca de las etiquetas que otorgan,pudiendo otorgar varias etiquetas a cada instancia. Se proponen 2 métodos de "label aggregation"asociados a este tipo de etiquetado, y se muestra, mediante un marco experimental que aúna el etiquetado tradicional y el propuesto, que el "candidate labelling" consigue extraer más información con un mismo número de personas anotadoras. Por último, se desarrolla un modelo de persona anotadora y 2 métodos de aprendizaje adaptados a este nuevo etiquetado, basados en el algoritmo EM, que obtienen mejores resultados en general que los métodos análogos en el marco de etiquetado tradicional. bcam Excelencia Severo Ochoa