Alberto del Rio
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Alberto del Rio has a total of 98 co-authors. Here are the top co-authors by number of shared publications:
- David Jimenez
Information Processing and Telecommunications Center (IPTC), Universidad Politécnica de Madrid, 28040 Madrid, Spain
14 shared publications - Javier Serrano
Informatic Systems Department, Universidad Politécnica de Madrid, 28031 Madrid, Spain
14 shared publications - Luis Miguel Contreras
Telefnica Innovacin Digital, Madrid, Spain
7 shared publications - Álvaro Llorente
Signals, Systems and Radiocommunications Department, Escuela Técnica Superior de Ingenieros de Telecomunicación (ETSIT), Universidad Politécnica de Madrid, 28040 Madrid, Spain
7 shared publications - Federico Alvarez
Universidad Politecnica de Madrid, Madrid 28040, Spain
5 shared publications - Cristina María Lozano Hernández
Department of Nursing, Physiotherapy and Occupational Therapy, Faculty of Nursing, University of Castilla-La Mancha, Spain
3 shared publications
Latest Publications
- Open Access
- 5 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 10959
Re-Evaluating PSNR Global Weighting Schemes for Quality Assessment in Professional Multigeneration Video Coding Workflows
The transition to IP-based broadcasting and the adoption of UHD resolutions (4K/8K) have made bandwidth optimization a critical challenge in professional video contribution networks. In these environments, successive encoding stages require high-fidelity formats, predominantly 4:2:2, to prevent cumulative degradation. However, legacy objective quality metrics, such as the ITU-recommended 6:1:1 scheme and perceptual color-sensitivity models, apply fixed mathematical coefficients that ignore the physical sample topology of the video signal. Consequently, these static models structurally penalize 4:2:2 workflows by mathematically undervaluing their chrominance payload. To overcome this limitation, this paper proposes a mathematical subsampling-based scheme that dynamically aligns component weights with the actual physical sample density of each chroma profile. To validate this topology-aware framework, ten diverse 4K sequences were evaluated across a three-generation 10-bit HEVC pipeline. Encodings used Main 4:2:2 10 and Main 10 profiles across contribution (20–90 Mbps) and distribution (8–32 Mbps) stages. Evaluated using strictly bounded Bjøntegaard Delta (BD) metrics, results demonstrate that while legacy metrics inaccurately suggest 4:2:0 superiority, the proposed framework reliably captures the advantages of 4:2:2 in 96% of BD-PSNR (>0) and 94% of BD-Rate (<0) cases, yielding indicative average savings of 27.32% in bitrate and +0.53 dB in quality. These findings challenge current industry assessment standards, providing a corrected methodological baseline for optimizing bandwidth allocation in broadcast contribution networks.- 5 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 10959
Open AccessRe-Evaluating PSNR Global Weighting Schemes for Quality Assessment in Professional Multigeneration Video Coding Workflows
The transition to IP-based broadcasting and the adoption of UHD resolutions (4K/8K) have made bandwidth optimization a critical challenge in professional video contribution networks. In these environments, successive encoding stages require high-fidelity formats, predominantly 4:2:2, to prevent cumulative degradation. However, legacy objective quality metrics, such as the ITU-recommended 6:1:1 scheme and perceptual color-sensitivity models, apply fixed mathematical coefficients that ignore the physical sample topology of the video signal. Consequently, these static models structurally penalize 4:2:2 workflows by mathematically undervaluing their chrominance payload. To overcome this limitation, this paper proposes a mathematical subsampling-based scheme that dynamically aligns component weights with the actual physical sample density of each chroma profile. To validate this topology-aware framework, ten diverse 4K sequences were evaluated across a three-generation 10-bit HEVC pipeline. Encodings used Main 4:2:2 10 and Main 10 profiles across contribution (20–90 Mbps) and distribution (8–32 Mbps) stages. Evaluated using strictly bounded Bjøntegaard Delta (BD) metrics, results demonstrate that while legacy metrics inaccurately suggest 4:2:0 superiority, the proposed framework reliably captures the advantages of 4:2:2 in 96% of BD-PSNR (>0) and 94% of BD-Rate (<0) cases, yielding indicative average savings of 27.32% in bitrate and +0.53 dB in quality. These findings challenge current industry assessment standards, providing a corrected methodological baseline for optimizing bandwidth allocation in broadcast contribution networks. - Open Access
- 17 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (18) , 9253
Beyond Hardware Mixers: A Resilient Cloud-Native Architecture for Real-Time Remote Video Production
Professional live video production has traditionally been limited by proprietary hardware mixers and monolithic desktop applications, which impose prohibitive costs and restrict scalable, command-line-based implementation for remote integration (REMI) workflows. To address these limitations, this article presents Voctomix 2.0, a containerized and enhanced live video mixing architecture based on the open-source Voctomix framework. This research introduces a production-ready extension layer that adds dynamic multi-layer overlays, operator-level stream-blanking control with coupled audio muting, and an asynchronous AMQP telemetry framework on top of the upstream mixing engine. The architecture was empirically evaluated in local, Docker, and single-node Kubernetes environments under varying workloads, ranging from 1080p25 to 2160p50 (4K). Results under a continuous four-source 1080p25 workload showed that, in a single-node deployment, containerization and orchestration impose no appreciable processing overhead, with a median host-level CPU utilization near 90% and a stable RAM working set of around 13 GB on standard hardware. The median command-to-output latency within the mixer was 293 ms, an internal switching latency that does not include camera capture or wide-area transport. Forced-failure experiments validated the architecture’s self-healing capability, which autonomously recovered blocked camera signals in approximately 1.8 s while maintaining a continuous live program feed.
- 17 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (18) , 9253
Open AccessBeyond Hardware Mixers: A Resilient Cloud-Native Architecture for Real-Time Remote Video Production
Professional live video production has traditionally been limited by proprietary hardware mixers and monolithic desktop applications, which impose prohibitive costs and restrict scalable, command-line-based implementation for remote integration (REMI) workflows. To address these limitations, this article presents Voctomix 2.0, a containerized and enhanced live video mixing architecture based on the open-source Voctomix framework. This research introduces a production-ready extension layer that adds dynamic multi-layer overlays, operator-level stream-blanking control with coupled audio muting, and an asynchronous AMQP telemetry framework on top of the upstream mixing engine. The architecture was empirically evaluated in local, Docker, and single-node Kubernetes environments under varying workloads, ranging from 1080p25 to 2160p50 (4K). Results under a continuous four-source 1080p25 workload showed that, in a single-node deployment, containerization and orchestration impose no appreciable processing overhead, with a median host-level CPU utilization near 90% and a stable RAM working set of around 13 GB on standard hardware. The median command-to-output latency within the mixer was 293 ms, an internal switching latency that does not include camera capture or wide-area transport. Forced-failure experiments validated the architecture’s self-healing capability, which autonomously recovered blocked camera signals in approximately 1.8 s while maintaining a continuous live program feed.
- Open Access
- 2 September 2026
- Published by Springer Nature
- p. 231-243
Gesture-Driven Remote Robot Mission Activation Through Smart Glasses: An XR Interaction Pipeline for Industrial Assistance
This paper presents the design and early validation of an end-to-end interaction pipeline enabling industrial operators to trigger robotic missions through hand gestures captured by wearable extended reality smart glasses. Developed within the MotivateXR project, the proposed approach integrates on-device visual sensing, embedded machine learning inference, semantic gesture mapping within a Unity application, and secure remote command transmission toward a mobile industrial robot executing real-world tasks. The system was experimentally evaluated in an industrial pilot environment where gestures performed by the operator were detected through the smart glasses’ cameras, interpreted locally, and translated into mission commands transmitted to a robotic platform. The study investigates recognition reliability, end-to-end latency propagation, execution consistency, and robustness across distributed system layers. To assess portability beyond the primary deployment, the gesture interaction layer was additionally integrated into an OpenXR-based XR application running on a standalone headset platform. This secondary scenario evaluates usability for interface navigation and workflow interaction, providing insight into cross-platform applicability of the gesture abstraction layer. Results provide initial evidence of the feasibility of gesture-driven control pipelines in distributed XR-robotics environments, while highlighting design trade-offs, integration constraints, and lessons learned relevant to industrial XR adoption.
- 2 September 2026
- Published by Springer Nature
- p. 231-243
Open AccessGesture-Driven Remote Robot Mission Activation Through Smart Glasses: An XR Interaction Pipeline for Industrial Assistance
This paper presents the design and early validation of an end-to-end interaction pipeline enabling industrial operators to trigger robotic missions through hand gestures captured by wearable extended reality smart glasses. Developed within the MotivateXR project, the proposed approach integrates on-device visual sensing, embedded machine learning inference, semantic gesture mapping within a Unity application, and secure remote command transmission toward a mobile industrial robot executing real-world tasks. The system was experimentally evaluated in an industrial pilot environment where gestures performed by the operator were detected through the smart glasses’ cameras, interpreted locally, and translated into mission commands transmitted to a robotic platform. The study investigates recognition reliability, end-to-end latency propagation, execution consistency, and robustness across distributed system layers. To assess portability beyond the primary deployment, the gesture interaction layer was additionally integrated into an OpenXR-based XR application running on a standalone headset platform. This secondary scenario evaluates usability for interface navigation and workflow interaction, providing insight into cross-platform applicability of the gesture abstraction layer. Results provide initial evidence of the feasibility of gesture-driven control pipelines in distributed XR-robotics environments, while highlighting design trade-offs, integration constraints, and lessons learned relevant to industrial XR adoption.
- Open Access
- 25 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 332
A multimodal GraphRAG engine for semantic knowledge extraction in industrial environments
- 25 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 332
Open AccessA multimodal GraphRAG engine for semantic knowledge extraction in industrial environments
- Open Access
- 4 April 2026
- Published by Elsevier in Computer Communications
- Vol. 252
Adaptive QoE optimization in non-stationary networks: Integrating PPO and digital twins for video streaming
- 4 April 2026
- Published by Elsevier in Computer Communications
- Vol. 252
Open AccessAdaptive QoE optimization in non-stationary networks: Integrating PPO and digital twins for video streaming
- Open Access
- 21 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (20) , 4115
Seamless User-Generated Content Processing for Smart Media: Delivering QoE-Aware Live Media with YOLO-Based Bib Number Recognition
The increasing availability of User-Generated Content during large-scale events is transforming spectators into active co-creators of live narratives while simultaneously introducing challenges in managing heterogeneous sources, ensuring content quality, and orchestrating distributed infrastructures. A trial was conducted to evaluate automated orchestration, media enrichment, and real-time quality assessment in a live sporting scenario. A key innovation of this work is the use of a cloud-native architecture based on Kubernetes, enabling dynamic and scalable integration of smartphone streams and remote production tools into a unified workflow. The system also included advanced cognitive services, such as a Video Quality Probe for estimating perceived visual quality and an AI Engine based on YOLO models for detection and recognition of runners and bib numbers. Together, these components enable a fully automated workflow for live production, combining real-time analysis and quality monitoring, capabilities that previously required manual or offline processing. The results demonstrated consistently high Mean Opinion Score (MOS) values above 3 72.92% of the time, confirming acceptable perceived quality under real network conditions, while the AI Engine achieved strong performance with a Precision of 93.6% and Recall of 80.4%.
- 21 October 2025
- Published by MDPI AG in Electronics
- Vol. 14 (20) , 4115
Open AccessSeamless User-Generated Content Processing for Smart Media: Delivering QoE-Aware Live Media with YOLO-Based Bib Number Recognition
The increasing availability of User-Generated Content during large-scale events is transforming spectators into active co-creators of live narratives while simultaneously introducing challenges in managing heterogeneous sources, ensuring content quality, and orchestrating distributed infrastructures. A trial was conducted to evaluate automated orchestration, media enrichment, and real-time quality assessment in a live sporting scenario. A key innovation of this work is the use of a cloud-native architecture based on Kubernetes, enabling dynamic and scalable integration of smartphone streams and remote production tools into a unified workflow. The system also included advanced cognitive services, such as a Video Quality Probe for estimating perceived visual quality and an AI Engine based on YOLO models for detection and recognition of runners and bib numbers. Together, these components enable a fully automated workflow for live production, combining real-time analysis and quality monitoring, capabilities that previously required manual or offline processing. The results demonstrated consistently high Mean Opinion Score (MOS) values above 3 72.92% of the time, confirming acceptable perceived quality under real network conditions, while the AI Engine achieved strong performance with a Precision of 93.6% and Recall of 80.4%.
- Open Access
- 4 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
An Enhanced Method for Objective QoE Analysis in Adaptive Streaming Services
Evaluating Quality of Experience (QoE) is crucial for multimedia services, as it measures user satisfaction with content delivery. This paper presents an objective method for evaluating QoE in adaptive streaming services, using the commercial tool Video-MOS, which allows real-time monitoring and analysis of multimedia content quality across various platforms and networks. The method aims to provide a precise evaluation that incorporates both technical and subjective factors. This approach integrates multiple factors that influence the overall user perception of adaptive streaming quality, offering greater flexibility and performance, which could contribute to more comprehensive future assessment. The method is validated by a test plan that incorporates a variety of content and scenarios to simulate various network conditions. The results demonstrate the method’s effectiveness in predicting QoE, highlighting the rebuffering frequency as a significant factor. In optimal conditions, specific content types can achieve a QoE score as high as 3.36. Conversely, under unfavorable conditions, the QoE may decrease by up to 1.42 Mean Opinion Score (MOS) points, which represents an 80% reduction from its optimal level. Although the rebuffering frequency has a substantial influence, a long initial buffer can have an even more negative effect on QoE, particularly under adverse conditions. Furthermore, adaptive streaming technologies, such as Dynamic Adaptive Streaming over HTTP (MPEG-DASH) and Adaptive Bitrate Streaming (ABR) are integral to the assessment process.
- 4 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessAn Enhanced Method for Objective QoE Analysis in Adaptive Streaming Services
Evaluating Quality of Experience (QoE) is crucial for multimedia services, as it measures user satisfaction with content delivery. This paper presents an objective method for evaluating QoE in adaptive streaming services, using the commercial tool Video-MOS, which allows real-time monitoring and analysis of multimedia content quality across various platforms and networks. The method aims to provide a precise evaluation that incorporates both technical and subjective factors. This approach integrates multiple factors that influence the overall user perception of adaptive streaming quality, offering greater flexibility and performance, which could contribute to more comprehensive future assessment. The method is validated by a test plan that incorporates a variety of content and scenarios to simulate various network conditions. The results demonstrate the method’s effectiveness in predicting QoE, highlighting the rebuffering frequency as a significant factor. In optimal conditions, specific content types can achieve a QoE score as high as 3.36. Conversely, under unfavorable conditions, the QoE may decrease by up to 1.42 Mean Opinion Score (MOS) points, which represents an 80% reduction from its optimal level. Although the rebuffering frequency has a substantial influence, a long initial buffer can have an even more negative effect on QoE, particularly under adverse conditions. Furthermore, adaptive streaming technologies, such as Dynamic Adaptive Streaming over HTTP (MPEG-DASH) and Adaptive Bitrate Streaming (ABR) are integral to the assessment process.
- Open Access
- 19 June 2025
- Published by JMIR Publications Inc. in JMIR Formative Research
- Vol. 9 (1) , e51831
Cocreating a Mobile Health App Providing Physical Activity Recommendations for Older People Living With Parkinson Disease or Dementia: User-Centered Pilot Study
Background: The project “Personalized Integrated Care Promoting Quality of Life for Older People” aimed to develop an integrated care system based on information and communication technology to support older people living with Parkinson disease or dementia disease. One module focuses on physical activity (PA) recommendations. Objective: The objective of the study is to describe the development process of the PA recommendation system from the behavior-change and technical perspective, followed by its content and satisfaction evaluation. Methods: This study describes the development of the PA recommendations based on the Health Action Process Approach (HAPA). A first pilot assessed the feasibility of the overall PROCare4Life system (previously reported). In a second pilot, users evaluated the content of the PA recommendations during 40 intervention days. In a third pilot, users evaluated their satisfaction with a mobile health satisfaction questionnaire. Results: The PA recommendations focused on different aspects of an adapted version of the HAPA model, while they simultaneously approached 3 activation factors: skills, knowledge, and motivation. The content was generally well-received, with most users rating key sections as excellent or good, particularly “benefits and consequences of PA” (34/43, 79%) and “five golden rules of PA” (34/41, 83%). However, less than a third gave high ratings to “PA guidelines of the WHO” (9/36, 25%) and “practical tips for PA” (10/35, 29%). Regarding satisfaction, at least half of the 237 participants found it easy and good to use, with acceptable time spent and clear instructions. Compared with agreement or neutral evaluations, most disagreed with negative statements about it being time-consuming (111/237, 47%) or boring (99/237, 42%). While 41% (97/237) recommended it and 44% (104/237) felt it helped them understand lifestyle benefits, fewer agreed the recommender system helped them set personal goals (78/237, 33%) or motivated change (88/237, 37%). Users found the recommendations understandable, engaging, and practical, though some aspects, such as motivation and goal setting, received criticism. Challenges in pilot 2, particularly related to setup difficulties and limited participation, led to system modifications in pilot 3 that improved usability and data collection. Conclusions: It has been confirmed that cocreating and iteratively testing the contents on HAPA and approaching the activation factors contributed to increasing the acceptance of the PA. The intervention development was based on user needs and used comparable methodology across user profiles and pilot phases. All in all, users were positive about the content. The research team has identified that digital systems, that provide monitoring functions of the mobile health app and Fitbit wristband are considered advantageous by participants in the cocreation process. Addressing the activation factors can be recommended for researchers and technical developers of other projects. Future adjustments to the design should focus on personalization to encourage the adoption of a healthier lifestyle.
- 19 June 2025
- Published by JMIR Publications Inc. in JMIR Formative Research
- Vol. 9 (1) , e51831
Open AccessCocreating a Mobile Health App Providing Physical Activity Recommendations for Older People Living With Parkinson Disease or Dementia: User-Centered Pilot Study
Background: The project “Personalized Integrated Care Promoting Quality of Life for Older People” aimed to develop an integrated care system based on information and communication technology to support older people living with Parkinson disease or dementia disease. One module focuses on physical activity (PA) recommendations. Objective: The objective of the study is to describe the development process of the PA recommendation system from the behavior-change and technical perspective, followed by its content and satisfaction evaluation. Methods: This study describes the development of the PA recommendations based on the Health Action Process Approach (HAPA). A first pilot assessed the feasibility of the overall PROCare4Life system (previously reported). In a second pilot, users evaluated the content of the PA recommendations during 40 intervention days. In a third pilot, users evaluated their satisfaction with a mobile health satisfaction questionnaire. Results: The PA recommendations focused on different aspects of an adapted version of the HAPA model, while they simultaneously approached 3 activation factors: skills, knowledge, and motivation. The content was generally well-received, with most users rating key sections as excellent or good, particularly “benefits and consequences of PA” (34/43, 79%) and “five golden rules of PA” (34/41, 83%). However, less than a third gave high ratings to “PA guidelines of the WHO” (9/36, 25%) and “practical tips for PA” (10/35, 29%). Regarding satisfaction, at least half of the 237 participants found it easy and good to use, with acceptable time spent and clear instructions. Compared with agreement or neutral evaluations, most disagreed with negative statements about it being time-consuming (111/237, 47%) or boring (99/237, 42%). While 41% (97/237) recommended it and 44% (104/237) felt it helped them understand lifestyle benefits, fewer agreed the recommender system helped them set personal goals (78/237, 33%) or motivated change (88/237, 37%). Users found the recommendations understandable, engaging, and practical, though some aspects, such as motivation and goal setting, received criticism. Challenges in pilot 2, particularly related to setup difficulties and limited participation, led to system modifications in pilot 3 that improved usability and data collection. Conclusions: It has been confirmed that cocreating and iteratively testing the contents on HAPA and approaching the activation factors contributed to increasing the acceptance of the PA. The intervention development was based on user needs and used comparable methodology across user profiles and pilot phases. All in all, users were positive about the content. The research team has identified that digital systems, that provide monitoring functions of the mobile health app and Fitbit wristband are considered advantageous by participants in the cocreation process. Addressing the activation factors can be recommended for researchers and technical developers of other projects. Future adjustments to the design should focus on personalization to encourage the adoption of a healthier lifestyle.
- Open Access
- 17 February 2025
- Published by Universidad Politecnica de Madrid - University Library
Deep Reinforcement Learning Implementation for Video Transmission Optimization in New Generation Networks
This thesis investigates the optimization of multimedia streaming through Deep Reinforcement Learning (DRL) algorithms within advanced network architectures. The study addresses the complex challenges inherent in high-bandwidth, low-latency environments, where fluctuating network conditions and diverse content types require adaptive solutions. Four distinct use cases were developed, three of them targeting specific aspects of Quality of Experience (QoE), while the last one offers a direct comparison on general environments. In the first use case, the Advantage Actor--Critic (A2C) algorithm was employed to optimize the QoE of video streaming services within the 5G-Media platform. The results demonstrated a significant improvement in the Mean Opinion Score (MOS), indicating a higher QoE in comparison to traditional streaming methods. Training the A2C model resulted in a more than 300% improvement in the average reward function. The percentage of poor QoE values (MOS 1-2) reduced from almost 25% to 10%, while the proportion of good quality values (MOS >3.5) increased from 65% to 90%. The second use case extended these principles to multisite gaming streaming optimization, utilizing the Asynchronous Advantage Actor-Critic (A3C) algorithm to manage the complexities of distributed environments on the 5G-EVE platform. The algorithm's ability to handle parallel environments resulted in a more responsive and immersive gaming experience, highlighting its effectiveness over A2C in such contexts. A3C demonstrated an exceptional increase in rewards, by 500% over the training periods. A3C achieved a 20% increase in good blockiness values (>0.65), a 15% reduction in the appearances of block loss artifacts and a bitrate improvement ranging from 5 to 7 Mbps during transmission. The third use case applied the Proximal Policy Optimization (PPO) algorithm to optimize cinematic experiences within the B5Gemini platform. This use case included additional information regarding network status, and PPO was particularly effective in balancing network performance with video quality. The PPO algorithm achieved a 66.87% reduction in block loss, improved the bitrate by 17.18%, and enhanced the reward metric by 73.02%. The last use case showcased a direct evaluation of the A3C and PPO algorithms across general environments. While A3C was more suitable for environments requiring rapid adaptation and parallel processing, PPO demonstrated superior stability and efficiency. In the CartPole scenario, A3C took around 90 seconds to meet the reward threshold, whereas PPO took up to 452 seconds. Nevertheless, PPO outperformed A3C, attaining 10.50% compared to 4.45% for values above the threshold. In the more complex Lunar Lander environment, PPO demonstrated greater stability with fewer fluctuations. The trend on training time is similar as previous environment, reducing on A3C almost half of the time in comparison with PPO. In terms of rewards above threshold, all of them surpassed the 20%. This thesis demonstrates that DRL algorithms are effective tools for multimedia streaming optimization. This research contributes to the field by not only showcasing the practical applications of DRL in real-world scenarios but also by providing a comparative analysis of leading algorithms in varied contexts. RESUMEN Esta tesis investiga la optimización de la transmisión multimedia a través de algoritmos de aprendizaje por refuerzo profundo (DRL) dentro de arquitecturas de red avanzadas. El estudio aborda los desafíos inherentes a los entornos de gran ancho de banda y baja latencia, donde las condiciones de red fluctuantes y los diversos tipos de contenido requieren soluciones adaptativas. Se desarrollaron cuatro casos de uso, tres de ellos orientados a aspectos específicos de la calidad de la experiencia (QoE), mientras que el último ofrece una comparación directa. En el primer caso de uso, se empleó el algoritmo Advantage Actor-Critic (A2C) para optimizar la QoE de la transmisión de video dentro de la plataforma 5G--Media. Los resultados demostraron una mejora significativa en la puntuación de opinión media (MOS). El entrenamiento del modelo A2C resultó en una mejora de más del 300% en la función de recompensa promedio. El porcentaje de valores de QoE deficientes (MOS 1-2) se redujo del 25% al 10%, mientras que los valores de buena calidad (MOS >3.5) aumentaron del 65% al 90%. El segundo caso de uso extendió estas ideas a la optimización de la transmisión de juegos en múltiples lugares, utilizando el algoritmo Asynchronous Advantage Actor-Critic (A3C) para gestionar los entornos distribuidos en la plataforma 5G-EVE. A3C demostró un aumento en las recompensas en un 500% durante los períodos de entrenamiento. Logró un aumento del 20% en los buenos valores de aparición de bloques (> 0.65), una reducción del 15% en la pérdida de bloques y una mejora de la tasa de bits que incrementa desde 5 a 7 Mbps durante la transmisión. El tercer caso de uso aplicó el algoritmo de Optimización de Políticas Proximales (PPO) para optimizar las experiencias cinematográficas dentro de la plataforma B5Gemini. Este caso de uso incluyó información adicional sobre el estado de la red, y PPO fue particularmente eficaz para equilibrar el rendimiento de la red con la calidad del video. El algoritmo PPO logró una reducción del 66.87% en la pérdida de bloques, mejoró la tasa de bits en un 17.18% y mejoró la métrica de recompensa en un 73.02%. El último caso de uso mostró una evaluación directa de los algoritmos A3C y PPO en entornos generales. Si bien A3C era más adecuado para entornos que requieren una adaptación rápida y procesamiento paralelo, PPO demostró una estabilidad y eficiencia superiores. En el escenario CartPole, A3C...
- 17 February 2025
- Published by Universidad Politecnica de Madrid - University Library
Open AccessDeep Reinforcement Learning Implementation for Video Transmission Optimization in New Generation Networks
This thesis investigates the optimization of multimedia streaming through Deep Reinforcement Learning (DRL) algorithms within advanced network architectures. The study addresses the complex challenges inherent in high-bandwidth, low-latency environments, where fluctuating network conditions and diverse content types require adaptive solutions. Four distinct use cases were developed, three of them targeting specific aspects of Quality of Experience (QoE), while the last one offers a direct comparison on general environments. In the first use case, the Advantage Actor--Critic (A2C) algorithm was employed to optimize the QoE of video streaming services within the 5G-Media platform. The results demonstrated a significant improvement in the Mean Opinion Score (MOS), indicating a higher QoE in comparison to traditional streaming methods. Training the A2C model resulted in a more than 300% improvement in the average reward function. The percentage of poor QoE values (MOS 1-2) reduced from almost 25% to 10%, while the proportion of good quality values (MOS >3.5) increased from 65% to 90%. The second use case extended these principles to multisite gaming streaming optimization, utilizing the Asynchronous Advantage Actor-Critic (A3C) algorithm to manage the complexities of distributed environments on the 5G-EVE platform. The algorithm's ability to handle parallel environments resulted in a more responsive and immersive gaming experience, highlighting its effectiveness over A2C in such contexts. A3C demonstrated an exceptional increase in rewards, by 500% over the training periods. A3C achieved a 20% increase in good blockiness values (>0.65), a 15% reduction in the appearances of block loss artifacts and a bitrate improvement ranging from 5 to 7 Mbps during transmission. The third use case applied the Proximal Policy Optimization (PPO) algorithm to optimize cinematic experiences within the B5Gemini platform. This use case included additional information regarding network status, and PPO was particularly effective in balancing network performance with video quality. The PPO algorithm achieved a 66.87% reduction in block loss, improved the bitrate by 17.18%, and enhanced the reward metric by 73.02%. The last use case showcased a direct evaluation of the A3C and PPO algorithms across general environments. While A3C was more suitable for environments requiring rapid adaptation and parallel processing, PPO demonstrated superior stability and efficiency. In the CartPole scenario, A3C took around 90 seconds to meet the reward threshold, whereas PPO took up to 452 seconds. Nevertheless, PPO outperformed A3C, attaining 10.50% compared to 4.45% for values above the threshold. In the more complex Lunar Lander environment, PPO demonstrated greater stability with fewer fluctuations. The trend on training time is similar as previous environment, reducing on A3C almost half of the time in comparison with PPO. In terms of rewards above threshold, all of them surpassed the 20%. This thesis demonstrates that DRL algorithms are effective tools for multimedia streaming optimization. This research contributes to the field by not only showcasing the practical applications of DRL in real-world scenarios but also by providing a comparative analysis of leading algorithms in varied contexts. RESUMEN Esta tesis investiga la optimización de la transmisión multimedia a través de algoritmos de aprendizaje por refuerzo profundo (DRL) dentro de arquitecturas de red avanzadas. El estudio aborda los desafíos inherentes a los entornos de gran ancho de banda y baja latencia, donde las condiciones de red fluctuantes y los diversos tipos de contenido requieren soluciones adaptativas. Se desarrollaron cuatro casos de uso, tres de ellos orientados a aspectos específicos de la calidad de la experiencia (QoE), mientras que el último ofrece una comparación directa. En el primer caso de uso, se empleó el algoritmo Advantage Actor-Critic (A2C) para optimizar la QoE de la transmisión de video dentro de la plataforma 5G--Media. Los resultados demostraron una mejora significativa en la puntuación de opinión media (MOS). El entrenamiento del modelo A2C resultó en una mejora de más del 300% en la función de recompensa promedio. El porcentaje de valores de QoE deficientes (MOS 1-2) se redujo del 25% al 10%, mientras que los valores de buena calidad (MOS >3.5) aumentaron del 65% al 90%. El segundo caso de uso extendió estas ideas a la optimización de la transmisión de juegos en múltiples lugares, utilizando el algoritmo Asynchronous Advantage Actor-Critic (A3C) para gestionar los entornos distribuidos en la plataforma 5G-EVE. A3C demostró un aumento en las recompensas en un 500% durante los períodos de entrenamiento. Logró un aumento del 20% en los buenos valores de aparición de bloques (> 0.65), una reducción del 15% en la pérdida de bloques y una mejora de la tasa de bits que incrementa desde 5 a 7 Mbps durante la transmisión. El tercer caso de uso aplicó el algoritmo de Optimización de Políticas Proximales (PPO) para optimizar las experiencias cinematográficas dentro de la plataforma B5Gemini. Este caso de uso incluyó información adicional sobre el estado de la red, y PPO fue particularmente eficaz para equilibrar el rendimiento de la red con la calidad del video. El algoritmo PPO logró una reducción del 66.87% en la pérdida de bloques, mejoró la tasa de bits en un 17.18% y mejoró la métrica de recompensa en un 73.02%. El último caso de uso mostró una evaluación directa de los algoritmos A3C y PPO en entornos generales. Si bien A3C era más adecuado para entornos que requieren una adaptación rápida y procesamiento paralelo, PPO demostró una estabilidad y eficiencia superiores. En el escenario CartPole, A3C...
- Open Access
Machine Learning-Based Network Anomaly Detection: Design, Implementation, and Evaluation
Background: In the last decade, numerous methods have been proposed to define and detect outliers, particularly in complex environments like networks, where anomalies significantly deviate from normal patterns. Although defining a clear standard is challenging, anomaly detection systems have become essential for network administrators to efficiently identify and resolve irregularities. Methods: This study develops and evaluates a machine learning-based system for network anomaly detection, focusing on point anomalies within network traffic. It employs both unsupervised and supervised learning techniques, including change point detection, clustering, and classification models, to identify anomalies. SHAP values are utilized to enhance model interpretability. Results: Unsupervised models effectively captured temporal patterns, while supervised models, particularly Random Forest (94.3%), demonstrated high accuracy in classifying anomalies, closely approximating the actual anomaly rate. Conclusions: Experimental results indicate that the system can accurately predict network anomalies in advance. Congestion and packet loss were identified as key factors in anomaly detection. This study demonstrates the potential for real-world deployment of the anomaly detection system to validate its scalability.
Open AccessMachine Learning-Based Network Anomaly Detection: Design, Implementation, and Evaluation
Background: In the last decade, numerous methods have been proposed to define and detect outliers, particularly in complex environments like networks, where anomalies significantly deviate from normal patterns. Although defining a clear standard is challenging, anomaly detection systems have become essential for network administrators to efficiently identify and resolve irregularities. Methods: This study develops and evaluates a machine learning-based system for network anomaly detection, focusing on point anomalies within network traffic. It employs both unsupervised and supervised learning techniques, including change point detection, clustering, and classification models, to identify anomalies. SHAP values are utilized to enhance model interpretability. Results: Unsupervised models effectively captured temporal patterns, while supervised models, particularly Random Forest (94.3%), demonstrated high accuracy in classifying anomalies, closely approximating the actual anomaly rate. Conclusions: Experimental results indicate that the system can accurately predict network anomalies in advance. Congestion and packet loss were identified as key factors in anomaly detection. This study demonstrates the potential for real-world deployment of the anomaly detection system to validate its scalability.