Tesis/Trabajos de Grado

URI permanente para esta colección

https://hdl.handle.net/1992/52298

Incluye documentos como: monografías, reportes, proyectos, prácticas, informes, entre otros; elaborados como requisito de grado para programas de pregrado y posgrado en la Universidad de los Andes.

Navegar

Mostrando 1 - 11 de 11

Acceso abierto
Analítica del aprendizaje sobre plataforma Senecode
(Universidad de los Andes, 2022-12-15) Porras Tascón, Manuel Felipe; Manrique Piramanrique, Rubén Francisco; 1031122840; FLAG research lab
El uso de plataformas de aprendizaje de programación permite por un lado gestionar de manera eficiente las asignaciones y tareas asíncronas de diversos grupos de estudiantes, y por el otro consolidar una fuente relevante de información sobre su desempeño y evolución. En este proyecto, se hace uso de los datos generados por la plataforma Senecode (plataforma apoyo para el curso de la Introducción a la programación) para evaluar un conjunto de hipótesis en relación al desempeño de los estudiantes y la interacción con la plataforma al resolver ejercicios de programación. En conjunto con las notas globales de la materia y una encuesta de percepción de los estudiantes (consolidados en una plataforma denominada "Observatorio"), Senecode es la principal fuente de información sobre los estudiantes que toman la materia. El objetivo de este proyecto es consolidar y unificar las fuentes de información disponibles para analizar las relaciones entre las diferentes interacciones de la plataforma Senecode que realizan los estudiantes y las notas de curso obtenidas.
Acceso abierto
Artificial intelligence techniques applied to cryptocurrency market prediction
(Universidad de los Andes, 2023-05-29) Sánchez Ardila, Juan Diego; Oliveros Forero, Julián Esteban; Manrique Piramanrique, Rubén Francisco; 1031122840; Manrique Piramanrique, Rubén Francisco; Grupo de investigación Flag (https://flaglab.github.io/)
The emerging cryptocurrency market is undergoing rapid evolution and presenting numerous novel technologies aimed at addressing problems and capitalizing on opportunities for societal improvement. Despite substantial growth over the past decade, this market remains immature, resulting in high volatility and risk for all stakeholders involved. The primary issue lies in the market's lack of stability and predictability. The objective of this project is to enhance comprehension of market predictability by employing machine learning techniques. This endeavor aims to assist cryptocurrency investors and stakeholders in making informed decisions by predicting market prices and trends, focusing specifically on Bitcoin, Ethereum, and Cardano. Machine learning models will be trained using historical market data to achieve this goal. Furthermore, the project includes a significant ancillary objective, as relying solely on models can be ineffective. The aim is to develop a robust platform that automatically collects real-time data, incorporates an API for integrating data sources with models, and features a user interface designed to present model results to cryptocurrency stakeholders. Lastly, it is worth noting the intrinsic value of this project, which originates from a shared personal goal among the authors. As an undergraduate degree thesis, this endeavor enables the authors to gain practical knowledge in applied blockchain technology and artificial intelligence, while fostering an integrated approach that consolidates their academic formation.
Acceso abierto
Fine-tuning face anti-spoofing models: exploring the transformative impact of image-level biometric privacy-enhancing techniques
(Universidad de los Andes, 2024-02-01) Avalos López, Fernando Andrés; Manrique Piramanrique, Rubén Francisco; 1031122840
Face Anti-Spoofing (FAS) systems are entrusted with the task of determining whether the content of a facial image is genuine or fake. The performance and overall quality of those systems depend on the data they are fed during their development stage, meaning that high-resolution and diverse facial datasets are commonly used. Even though FAS systems are particularly useful in some situations, they do raise concerns in regards to the privacy of the individuals whose images are used to train them. The straightforward solution is to apply filters to the facial images, at the expense of the systems’ performance, since recognising faces becomes increasingly harder. This is where Biometric Privacy-Enhancing Techniques (B-PETs) come into play, which help to alleviate the adversarial tension between biometric utility and privacy gains. This work concerned itself with assessing the impact of 3 distinct image-level B-PETs in the performance of 3 architecturally distinct FAS systems and found that image-level B-PETs are not fit for finding a valuable trade-off, suggesting that more sophisticated techniques are needed.
Acceso abierto
Graph neural networks vs. Traditional methods for recommending MOOC courses: a comparative analysis
(Universidad de los Andes, 2023-12-04) Pardo Bravo, Santiago; Manrique Piramanrique, Rubén Francisco; 1031122840; Manrique Piramanrique, Rubén Francisco
In recent years, online education has seen remarkable growth, particularly through Massive Open Online Course (MOOC) platforms. These platforms offer a wide range of open courses, making it a problem for users to select one from thousands of courses. Navigating this vast selection led to recommendation systems emerging, enhancing the learning experience by guiding students toward courses aligned with their interests. Addressing this challenge involves three primary approaches. The content-based approach uses the characteristics content of the seen courses to suggest new ones. Collaborative filtering taps into interactions among users with similar preferences, proposing relevant courses. The hybrid approach combines both methods for a comprehensive recommendation system. In a context marked by the rapid expansion of Massive Open Online Courses (MOOCs), there has been a parallel surge in the field of deep learning along with its diverse architectures. Against this backdrop, the focus of this investigation centers on Graph Neural Networks (GNNs), a deep learning architecture explicitly tailored to tackle structured data. In this study, we propose and validate a novel course recommendation model utilizing GNNs on the Xuetang MOOC platform. This research's significance lies in enhancing online course recommendations using GNNs, aiming to elevate user satisfaction and learning effectiveness. Finally, for a comprehensive analysis, we'll compare our model against traditional approaches using the same dataset. This determines GNN's convenience in recommendation precision and relevance against the traditional methods.
Acceso abierto
Implementation of active data Selection algorithms for data choosing in ASV systems
(Universidad de los Andes, 2023-12-04) Jiménez Garizao, Jesús Alberto; Manrique Piramanrique, Rubén Francisco; 1031122840; Facultad de Ingeniería
In the field of voice spoof detection, training countermea- sures (CM) that effectively generalize to various unseen tests is a per- sistent challenge. While strategies such as data augmentation and self- supervised learning have been employed to enhance CM, their limited performance still requires additional approaches. This research focuses on the use of active learning (AL) to select and eliminate training data in CM training, addressing the need to optimize data selection and im- prove model effectiveness. The study proposes several active learning algorithms that offer substantial improvements in detection error rates across multiple datasets in automatic speaker verification (ASV) sys- tems, these are based on the ASVspoof 2019 and the HABLA sets. Thus, these contributions are expected to be valuable for future research and applications in this domain, significantly enhancing model effectiveness.
Acceso abierto
Improving autonomy and natural interaction with a Pepper robot through the evaluation of different large language models
(Universidad de los Andes, 2023-12-04) Romero Colmenares, Juan Andrés; Rojas Becerra, Luccas; Manrique Piramanrique, Rubén Francisco; 1031122840; Manrique Piramanrique, Rubén Francisco
Social Robotics is a field dedicated to exploring robots as interactive social companions and aides. Throughout its evolution, this discipline has centered on creating versatile robots that can tackle a range of tasks, functioning autonomously to aid humans. Nevertheless, achieving this is a complex endeavor, as robots must possess the ability to comprehend their surroundings and accurately interpret human directives. The main challenge confronting both these robots and their developers is the disparity between given instructions and actual behavior within particular contexts. Right now, many of these robots have to be manually programmed for specific tasks. They might be good at those jobs, but they don't do so well when you ask them to do a different activity. This happens due to the lack of cognition in pre-programmed deterministic jobs, where the robot cannot comprehend the reasoning behind its actions.To address this issue, this research aims to improve the autonomy and natural interaction of a Pepper robot by evaluating different large language models (LLMs). By leveraging the capabilities of LLMs, the research aims to develop a system that allows the robot to autonomously follow instructions given in natural language to accomplish general-purpose tasks. The evaluation process involves comparing the performance of commercial and open-source LLMs in generating code instructions for the robot. The generated code will be evaluated based the code execution results and will include a comparison of different prompting strategies and code abstraction levels. The evaluation will be conducted through both automatic and human evaluation processes. The results of this research will contribute to the development of highly effective robots capable of performing various general tasks.
Acceso abierto
Large Language Models como planeadores multiagente en entornos sociales
(Universidad de los Andes, 2024-01-27) Pinzón Roncancio, Juan Sebastian; Manrique Piramanrique, Rubén Francisco; 1031122840; Cardozo Álvarez, Nicolás; Gauthier Umaña, Valerie Elisabeth; Núñez Castro, Haydemar María
Esta investigación explora la integración de Modelos de Lenguaje de Gran Escala (LLMs), como GPT, en la simulación de comportamientos humanos dentro de entornos sociales simulados. Inspirándonos en el trabajo ``Generative Agents" de Stanford, aplicamos LLMs como planeadores en un marco multiagente. Utilizamos la arquitectura de flujo de razonamiento propuesta por Park et al. (2023), que comprende módulos de percepción, planeación, recuperación de memorias, reflexión y actuación. El estudio se centra en entornos simulados de Melting Pot, un Framework de DeepMind para evaluar el Aprendizaje por Refuerzo Multiagente (MARL), destacando dinámicas sociales como la cooperación, competencia y reciprocidad.\\ Nuestro enfoque metodológico incluyó la adaptación de la arquitectura de razonamiento a los escenarios de Melting Pot y la definición de tres conjuntos de escenarios de prueba. Las métricas clave empleadas para evaluar el desempeño de los agentes abarcaron la recompensa por cápita y ciertos indicadores de cooperatividad definidos para el escenario. Los resultados obtenidos ofrecen una visión detallada del desempeño y comportamiento de los agentes, destacando cómo la integración de LLMs en entornos de simulación social puede superar las limitaciones de los enfoques tradicionales de MARL y proporcionar una representación más rica y realista de las interacciones sociales.
Restringido
Modelo predictivo para inversión inmobiliaria, caso Colombia
(Universidad de los Andes, 2024-02-02) Nunes Ariza, Juan Carlos Eduardo; Manrique Piramanrique, Rubén Francisco; 1031122840; Manrique Piramanrique, Rubén Francisco
Este proyecto pretende mejorar la eficiencia al momento de realizar una inversión inmobiliaria en Colombia. En Colombia, los precios de los inmuebles están sujetos a diferentes parámetros como: ubicación, tamaño, antigüedad, sitios de interés cercanos, acabados, etc. Procesar y sopesar esta información suele ser dispendioso, causando que las personas gasten mucho tiempo y dinero en la selección de una buena alternativa para invertir. Este proceso además requiere de un conocimiento de dominio especializado que usualmente recae en agentes inmobiliarios y no en las personas del común. En este proyecto buscamos construir un modelo de predicción de precios e identificación de inversiones rentables. Para la consecución de este modelo se realizó el proceso de anotación de los precios para más de 1.200 predios en Colombia y se entrenaron dos modelos predictivos para calcular cuando una inversión es rentable. Los resultados arrojan un MAE (mean abosulte error) en los precios calculados de más de $878’035.487 para el conjunto de propiedades testeadas. Entonces, desde una perspectiva de seguridad de inversión, el producto conseguido sirve como herramienta para facilitar, agilizar y acompañar el trabajo de agentes inmobiliarios profesionales en Colombia.
Restringido
MusicGen Music generation model as a tool for artistic creation
(Universidad de los Andes, 2024-02-01) Tovar García, Diego Alejandro; Manrique Piramanrique, Rubén Francisco; 1031122840
The current work is an exploration on how to re purpose AI driven technologies to generate music, in a way that prioritizes the artistic endeavour of musical composition. A particular concept, which is hereby addressed, is the idea of agency of decision in the creative process. The study will center on the execution of a spatial intervention where the sound experience will be built from user-given-prompts describing the space they roam. The code for the project can be found at https://github.com/Didage/spatial-music-gen.
Acceso abierto
Recuperación Aumentada Generativa (RAG) para la creación de chatbot de dominio específico
(Universidad de los Andes, 2023-12-04) Cohen Solano, Kevin; Manrique Piramanrique, Rubén Francisco; 1031122840; Manrique Piramanrique, Rubén Francisco; Facultad de Ingeniería
En este proyecto se desarrolló una aplicación web que funciona como chatbot interactivo, en la cual se utilizó la técnica de Recuperación Aumentada Generativa para otorgarle acceso a la información de un dominio específico. En el margen del proyecto, se exploraron diferentes técnicas para mejorar el funcionamiento del RAG, tales como ingeniería de prompts y HyDE. Igualmente, se hizo la comparación entre los modelos grandes de lenguaje GPT 3.5 y GPT 4. Esta aplicación fue puesta a prueba con los potenciales usuarios finales, siendo estos tanto estudiantes como profesores del Instituto Departamental de Bellas Artes. Gracias a esto se obtuvieron variedad de resultados que evidencian la eficiencia de ciertas estrategias, así como del nivel de aceptación final de los usuarios.
Acceso abierto
Studying academic success: A data analytics approach to predict performance of higher education students
(Universidad de los Andes, 2024-01-10) Martínez Osorio, Daniel Felipe; Benítez Amaya, Andrés Felipe; Manrique Piramanrique, Rubén Francisco; 1031122840; Facultad de Ingeniería
The dropout of students in higher education is a concern for universities, as it directly impacts the community and the educational level of future generations. For this reason, a data analytics-based model is proposed to support students in making decisions during the course selection process, aiming to guide them towards completing their degree while maximizing their performance. We have a dataset for three different majors in the Universidad de los Andes, (Systems and Computer Engineering, Industrial Engineering, and Economics), containing historical information about students, the courses they chose each semester in their specific curriculum, and their grades. Based on this data, the model analyzes the completed courses and the ones remaining for each student to fulfill their curriculum requirements. In this way, it creates a student profile that is used to calculate the probability of achieving certain grades in their next semester. Assuming this result, the process is iterated to develop a curriculum plan for the upcoming semesters. This outcome will provide students with a course guide for each semester, increasing their likelihood of achieving better performance in their studies.

Navegar

Examinando Tesis/Trabajos de Grado por Autor "1031122840"

Resultados por página

Opciones de ordenación