La contaminación atmosférica por hollín constituye un desafío crítico de salud pública y ambiental, evidenciando la urgencia de metodologías precisas para su caracterización. Sin embargo, cuantificar hollín en procesos de combustión enfrenta desafíos técnicos fundamentales: la radiación térmica de llamas con alto contenido de hollín resulta de la superposición compleja entre bandas espectrales discretas de gases de combustión y la emisión continua de hollín, dificultando la identificación individual de cada contribuyente mediante métodos convencionales. Este trabajo desarrolla y valida experimentalmente una metodología de modelado directo de transferencia radiativa espectralmente resuelta usando un enfoque Line-by-Line de alta resolución para gases, un modelo para hollín en régimen de Rayleigh, y solución numérica de la ecuación de transferencia radiativa en geometría axisimétrica para sintetizar imágenes infrarrojas multiespectrales comparables con mediciones de una cámara Telops MS-M1K. La validación experimental en llamas de difusión laminares axisimétricas (Yale 60 y Yale 80) demuestra reproducción cuantitativa de mediciones experimentales con errores relativos entre 16-67 % dependiendo del filtro y llama. El análisis de descomposición espectral revela que la separabilidad entre contribuciones de gas y hollín depende críticamente de la selección del filtro y contenido de hollín, identificando F6 como una banda casi pura de CO2, F7 y F8 como bandas sensibles a hollín-H2O y F5 como una banda mixta de todas las especies. El análisis de sensibilidad mediante elasticidades cuantifica cómo la temperatura emerge como la variable dominante universal, permitiendo proponer una estrategia de inversión secuencial que maximiza la separabilidad espectral. La aplicación a combustión de polimetilmetacrilato (PMMA) demuestra la aplicabilidad de la metodología a configuraciones complejas. Este trabajo establece un protocolo sistemático para la síntesis y validación de imágenes multiespectrales sintéticas en llamas con alto hollín, proporcionando herramientas metodológicas para monitoreo de emisiones, optimización de combustión, caracterización de incendios y desarrollo de diagnóstico óptico no intrusivo.

This thesis addresses the feature selection problem as a multi-objective optimization task motivated by the increasing number of features in datasets used in modern machine learning applications. The goal is to identify a reduced subset of features that optimizes the classification accuracy. To this end, an adaptation of the evolutionary algorithm NSGA-II is proposed, integrating filter-based feature selection methods, specifically the chi-square statistical test, incorporated into the initialization, mutation, and a dimensionality reduction process. The proposed approach was evaluated using twenty datasets from different domains, with dimensionalities ranging from 16 to 12,600 features, analyzing the impact of different initial feature selection rates. The results show that the proposed strategies outperform the baseline method on the majority of the evaluated datasets in terms of hypervolume, particularly when using 10\% of initial features, while maintaining computational times comparable to the base version of NSGA-II and achieving competitive results with respect to state-of-the-art approaches, thereby confirming the effectiveness of the proposed adaptation in high-dimensional scenarios.

Workforce shift scheduling is a central problem in human resource management, particularly in contexts characterized by multiple operational constraints and individual preferences. The Employee Shift Scheduling Problem (ESSP) exhibits high combinatorial complexity and inherently conflicting objectives, such as meeting coverage requirements while preserving employee satisfaction. Traditional approaches have typically addressed these aspects through single-objective formulations, which limit the explicit analysis of trade-offs between workforce well-being and operational efficiency. This thesis addresses the ESSP from a bi-objective optimization perspective, explicitly separating the minimization of employee dissatisfaction from the minimization of penalties associated with coverage deviations. To this end, a specialized adaptation of the evolutionary algorithm NSGA-II is proposed, specifically designed to operate within a highly constrained search space. The approach incorporates a hybrid representation based on feasible work sequences, allowing a relevant subset of constraints to be handled implicitly and enabling the design of problem-specific genetic operators. The algorithm includes a feasibility-oriented initialization phase and an automatic parameter tuning process using ParamILS. Experimental validation is conducted on ten benchmark instances from the state of the art, with results compared against an exact method based on integer linear programming implemented in AMPL and solved with Gurobi. Performance is evaluated using the hypervolume indicator and Pareto front analysis. The results show that the proposed approach is capable of generating competitive solution sets, particularly for larger instances, and highlight the significant influence of instance structural characteristics on algorithmic performance.

El presente trabajo considera la extensión de un modelo de lazos córtico–ganglios basales (CBG), conjunto de estructuras corticales y subcorticales relacionadas con la toma de decisiones, a través de la integración de efectos asociados al nivel de dopamina tónica (tipo D1). La dopamina (DA), neurotransmisor asociado con procesos de aprendizaje y memoria, se ha relacionado con efectos en el comportamiento con respecto a la razón entre exploración y explotación. El modelo resultante considera características como la consideración de múltiples lazos paralelos –considerando decisiones en múltiples niveles–, reglas de plasticidad sináptica que describen un aprendizaje dopaminérgico basado en recompensas, y la modulación en los procesos de selección sobre la tendencia a la exploración de nuevas opciones, frente a la explotación de conocimiento previamente adquirido. Para evaluar el comportamiento del modelo con respecto a cambios en los niveles de DA, se simula la ejecución de una tarea de selección forzada de dos opciones, considerando aprendizaje entre selecciones. Los datos obtenidos durante los procesos de selección en la realización de esta tarea demuestran variaciones en el comportamiento, en términos de cuánto se promueve la exploración de nuevas opciones en contra de la explotación de la información aprendida, al modificar los niveles de DA tónica. A pesar de esta modificación sobre el comportamiento y el desempeño del modelo, las pruebas realizadas predicen que las señales internas de aprendizaje no se ven modificadas ante variaciones en los niveles de DA. En conjunto, con el fin de evaluar la aplicabilidad del modelo propuesto como mecanismo de toma de decisiones, y en base a la importancia de la regulación entre exploración y explotación en una plataforma robótica, se describe la estructura de un controlador diseñado para enfrentar una tarea de supervivencia de dos recursos, aplicado sobre el robot MODI (MODular Intelligence). Durante la realización de esta tarea, el robot MODI debe aprender en tiempo real cuáles son las acciones que le permiten aumentar su esperanza de vida. Mediante simulaciones, se prueba que el modelo es utilizable como mecanismo de toma de decisiones, y que variaciones en los niveles de dopamina tónica modifican las habilidades de supervivencia del robot. Los datos obtenidos sugieren la existencia de un nivel de DA tónica constante tal que maximiza la esperanza de vida alcanzada por el robot.

En esta tesis se realiza el desarrollo de un modelo de Reinforcement Learning para optimización secuencial de portafolios, incorporando el riesgo mediante el Expected Shortfall. El modelo aprende políticas de decisiones sobre activos de un portafolio (comprar, vender o mantener) usando información de retornos de distintos tipos de activos (acciones, ETFs y criptomonedas) para distintas líneas temporales, junto con un umbral de decisión compuesta que se adapta de forma dinámica según la volatilidad de los retornos. Basándose en prueba y error, obteniendo como recompensa una métrica de retorno ajustado por riesgo, definida por el retorno marginal obtenido de un conjunto de movimientos respecto a una inversión libre de riesgo, ajustada por una métrica de riesgo explícita.El objetivo principal de la tesis es evaluar si un enfoque de Reinforcement Learning logra detectar patrones dinámicos de mercado y lograr generalizarlos para escenarios fuera de muestra. Comparándolo con otros modelos tanto financieros estadísticos, como modelos de machine learning supervisados y no supervisados. Los resultados muestran que, para enfoques más conservadores sobre el umbral, el modelo de Reinforcement Learning tiene un menor gap entre entrenamiento validación y testeo, siendo sus resultados más permanentes entre fases. Los enfoques más de rentabilidad pura, poseen mayor volatilidad entre fases de validación y testeo, teniendo un mayor riesgo que sus competidores y resultados similares.

Piscirickettsia salmonis is the etiological agent of Piscirickettsiosis, one of the most severe and detrimental diseases affecting the Chilean salmon farming industry. Currently, there are 26 vaccines with provisional registration available for prophylactic use in salmonids. However, none have proven effective in fully controlling epidemic outbreaks under field conditions. Several researchers have identified two predominant genogroups along the Chilean coast, designated as LF-89 and EM-90, which exhibit a variable spatiotemporal distribution in southern Chilean salmon farms in recent years. Recent evidence suggests that the biological mechanisms of both genogroups may interact synergistically during infection, indicating the potential for co-infections in contemporary farms. This scenario further complicates vaccine efficacy, as many are designed to target only a predominant genogroup. Thus, the need arises to identify alternative prophylactic methods to ensure and contribute to the sustainability of the Chilean salmon farming industry. A novel prophylactic approach proposed in this thesis involves the use of recombinant nanoproteins, also known as inclusion bodies, with antigenic functions against Piscirickettsiosis. These inclusion bodies are aggregates of soluble and insoluble proteins with amyloid characteristics, eliminating the need for encapsulation. They are scalable, lyophilizable, and highly stable under adverse temperature and pH conditions. This thesis aims to evaluate the ability of certain antigenic protein inclusion bodies from P. salmonis to induce an adaptive immune response in Atlantic salmon (Salmo salar) and their potential use as immunogens in a vaccine prototype. To achieve this objective, three antigenic sequences with chimeric characteristics from the LF-89 and EM-90 genogroups of P. salmonis were designed and produced as recombinant inclusion bodies in E. coli BL21(DE3). The prototypes were designated for commercial protection purposes as SkipZ, PulseJ, and HopQ. These prototypes were quantified and characterized using techniques such as Western blot. Their uptake and immunostimulatory activity were evaluated in salmon macrophages through in vitro assays. Subsequently, juvenile salmon were vaccinated, and the effects of the SkipZ and HopQ prototypes were analyzed in terms of gene expression (qPCR) and immunoglobulin production (ELISA). The three prototypes achieved production efficiencies of 57 mg/L for PulseJ, 40 mg/L for SkipZ, and 26,8 mg/L for HopQ. These prototypes were internalized by salmon macrophages (RTS-11) with uptake efficiencies of 5% for SkipZ, 26% for PulseJ, and 54% for HopQ. In vitro assays showed activation at both transcriptional and translational levels of molecules associated with antigen presentation and pro-inflammatory markers, with notable activation capacities observed for HopQ and SkipZ at 20 µg/mL and 5 µg/mL, respectively. In vivo assays revealed that both vaccine prototypes, administered at doses of 0.5 mg/kg of fish for SkipZ and 2 mg/kg of fish for HopQ, robustly activated the Th1 response, with a marked increase in IFN-γ and IL-12 expression. This strong upregulation might suppress CD8+ lymphocyte differentiation, potentially reducing the effectiveness of the cellular immune response in fully eliminating the pathogen. However, this effect may contribute to maintaining homeostasis in fish during the exacerbated inflammation induced by the vaccines. Moreover, no significant change was observed in the Th2 response, suggesting that the humoral response was not predominantly activated. In conclusion, the vaccine prototypes based on antigenic sequences of Piscirickettsia salmonis and produced as chimeric inclusion bodies demonstrated potent immunogenicity, capable of inducing a robust adaptive immune response in Atlantic salmon. The activation of the Th1 response by HopQ and SkipZ was particularly noteworthy. However, further studies are needed to evaluate their protective effects in challenge trials with the pathogen, including both genogroups, and to determine their long-term effectiveness.

This study introduces a novel measure for evaluating attribute relevance, specifically designed to accurately identify attributes that are intrinsically related to a phenomenon, while being sensitive to the asymmetry of those relationships and noise conditions. Traditional variable selection techniques, such as filter and wrapper methods, often fall short in capturing these complexities. Our methodology, grounded in decision trees but extendable to other machine learning models, was rigorously evaluated across various data scenarios. The results demonstrate that our measure effectively distinguishes relevant from irrelevant attributes and highlights how relevance is influenced by noise, providing a more nuanced understanding compared to established methods such as Pearson, Spearman, Kendall, MIC, MAS, MEV, GMIC, and 𝑃ℎ𝑖𝑘 This research underscores the importance of phenomenon-centric explainability, reproducibility, and robust attribute relevance evaluation in the development of predictive models. By enhancing both the interpretability and contextual accuracy of models, our approach not only supports more informed decision making but also contributes to a deeper understanding of the underlying mechanisms in diverse application domains, such as biomedical research, financial modeling, astronomy, and others.

A critical element in understanding voice production mechanisms is the characterization of vocal fold collision, which is widely considered a primary etiological factor in the development of common phonotraumatic lesions such as nodules and polyps. This paper describes the development of a transoral, dual-sensor intraglottal/subglottal pressure probe for the simultaneous measurement of vocal fold collision and subglottal pressures during phonation using two miniature sensors positioned 7.6 mm apart at the distal end of a rigid cannula. Proof-of-concept testing was performed using excised whole-mount and hemilarynx human tissue aerodynamically driven into self-sustained oscillation, with systematic variation of the superior–inferior positioning of the vocal fold collision sensor. In the hemilarynx experiment, signals from the pressure sensors were synchronized with an acoustic microphone, a tracheal-surface accelerometer, and two high-speed video cameras recording at 4000 frames per second for top–down and en face imaging of the superior and medial vocal fold surfaces, respectively. As expected, the intraglottal pressure signal exhibited an impulse-like peak when vocal fold contact occurred, followed by a broader peak associated with intraglottal pressure build-up during the de-contacting phase. As subglottal pressure was increased, the peak amplitude of the collision pressure increased and typically reached a value below that of the average subglottal pressure. Results provide important baseline vocal fold collision pressure data with which computational models of voice production can be developed and in vivo measurements can be referenced.