Fecha
2025-09
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editorial
Citas
0 citas en 
0 citas en 
Resumen
Esta tesis de master estudia la aplicación de deep neural networks (DNNs) para el diseño de redes cell-free massive MIMO (CF-mMIMO), una topología novedosa destinada a revolucionar las comunicaciones móviles de sexta generación (6G). Más concretamente, el trabajo aborda el diseño de la estrategia de asignación de potencia, es decir, determinar cuánta potencia se asigna a cada usuario con el fin de optimizar una determinada métrica de rendimiento relacionada con la eficiencia espectral. A diferencia de las estrategias clásicas de asignación de potencia, que suelen basarse en métodos de optimización matemática complejos, el enfoque con DNNs establece una relación entre la información de canal a gran escala (vinculada a las posiciones entre usuarios y puntos de acceso (APs)) a los coeficientes de potencia. Es muy destacable que esta función de mapeo es computacionalmente trivial durante la operación en tiempo real de la red móvil, a costa de requerir una fase de entrenamiento costosa que, convenientemente, puede llevarse a cabo como un pre-procesado (i.e., off-line). Los resultados demuestran que la asignación de potencia basada en DNNs funciona a la par de la técnica de referencia (i.e., fractional power allocation (FPA)) para una amplia variedad de configuraciones (p. e., diferentes cargas de red, distintas técnicas de procesamiento, diferentes objetivos de eficiencia espectral).
Además, también se ha propuesto una estrategia basada en aprendizaje por transferencia para generalizar la aplicabilidad de un modelo DNN entrenado para una carga de red específica a otra diferente, sin necesidad de reentrenar el modelo completo.
This master thesis deals with the design of CF-mMIMO networks, a novel topology set to revolutionize Sixth generation (6G) communications, using DNNs. In particular, the work addresses the design of the power allocation policy, that is, determining how much power is assigned to each user so as to optimize a certain performance metric related to the spectral efficiency. In contrast to classical power allocation strategies, which may rest upon complex mathematical optimization procedures, the DNN approach maps the large-scale channel information (related to the positions between users and access points (APs)) to power coefficients. Critically, this mapping operation is computationally trivial during real-time operation at the cost of requiring a costly training phase, which conveniently, can be conducted off-line. Results demonstrate that DNN-based power allocation performs as well as the benchmark technique (i.e., FPA) for a wide variety of configurations (e.g., different network loads, different processing techniques, distinct spectral efficiency objectives). Additionally, a strategy based on transfer learning has also been proposed to generalize the applicability of a DNN model trained for a speciffic network load to different one without having to retrain the full model.
This master thesis deals with the design of CF-mMIMO networks, a novel topology set to revolutionize Sixth generation (6G) communications, using DNNs. In particular, the work addresses the design of the power allocation policy, that is, determining how much power is assigned to each user so as to optimize a certain performance metric related to the spectral efficiency. In contrast to classical power allocation strategies, which may rest upon complex mathematical optimization procedures, the DNN approach maps the large-scale channel information (related to the positions between users and access points (APs)) to power coefficients. Critically, this mapping operation is computationally trivial during real-time operation at the cost of requiring a costly training phase, which conveniently, can be conducted off-line. Results demonstrate that DNN-based power allocation performs as well as the benchmark technique (i.e., FPA) for a wide variety of configurations (e.g., different network loads, different processing techniques, distinct spectral efficiency objectives). Additionally, a strategy based on transfer learning has also been proposed to generalize the applicability of a DNN model trained for a speciffic network load to different one without having to retrain the full model.
Descripción
Categorías UNESCO
Palabras clave
topología cell-free, asignación de potencia, aprendizaje profundo, aprendizaje por transferencia, Cell-free topology, power allocation, deep learning, transfer learning
Citación
Riera Palou, Felip. Trabajo Fin de Máster: "Applications of Deep Learning Techniques to the design of cell-free massive MIMO networks". Universidad Nacional de Educación a Distancia (UNED), 2025
Centro
E.T.S. de Ingeniería Informática