Digital Agent-Based Resource Management for Short Video Streaming in Multicast Networks

Abstract

As fifth-generation (5G) networks approach maturity and widespread deployment, both industry and academia are turning their attention to sixth-generation (6G) networks. It is anticipated that 6G networks will support an unprecedented diversity of services with heterogeneous user requirements, accelerating the shift from service-oriented to experience-centric resource management. Among these services, short video streaming has become one of the majority of users’ daily mobile traffic consumptions due to its highly engaging content, but this also leads to substantial traffic increase, especially in densely populated areas. Considering the popularity-based and user similarity-driven recommendation principles in short video platforms, multicast transmission over the air can effectively relieve traffic pressure by delivering the same video data to a group of users with similar characteristics and locations. Quality of experience (QoE), as a subjective performance metric in experience-centric resource management, can reflect the user satisfaction level on multicast short video streaming, which usually consists of rebuffer time, video quality, and video quality variation. To achieve experience-centric resource management, digital agent (DA), as a cutting-edge technology in 6G networks, owns advanced status emulation, data analytics, and decision-making capabilities, which can perceive network dynamics, abstract hidden behavior patterns or QoE models, and solve complex optimization problems. The interesting issue is maximizing user QoE in multicast short video streaming under limited radio and computing resources within dynamic network environments. However, the main technical challenges are: (1) how DAs abstract user swipe behavior patterns for large-timescale resource reservation to enhance resource utilization and improve long-term user QoE; (2) how DAs characterize multicast buffer dynamics for real-time resource allocation to alleviate buffer length overestimation and improve real-time user QoE; (3) how to adaptively select appropriate DA models to assist resource management and timely update them to further improve user QoE.

In this thesis, we develop an efficient DA-based resource management framework to enhance user QoE for multicast short video streaming, including swipe behavior-aware resource reservation, multicast buffer-aware resource allocation, and network dynamics-aware DA management. First, we propose a DA-based resource reservation scheme by considering dynamic user swipe behaviors to enhance resource utilization and large-timescale user QoE. Particularly, user DAs are constructed for individual users, which store users’ historical data for updating multicast groups and abstracting useful information. The swipe probability distributions and recommended video lists are abstracted from user DAs to predict bandwidth and computing resource demands. Parameterized sigmoid functions are leveraged to characterize multicast groups’ user QoE. A joint non-convex bandwidth and computing resource reservation problem is formulated and transformed into a convex piecewise problem by utilizing a tangent function to approximately substitute the concave part. A low-complexity scheduling algorithm is developed to find the optimal resource reservation decisions. Simulation results based on the real-world dataset demonstrate that the proposed scheme outperforms benchmark schemes in terms of user QoE and resource utilization. Second, we propose a DA-based resource allocation scheme by considering multicast buffer dynamics to enhance real-time user QoE. In specific, user statuses emulated by DAs are utilized to estimate the transmission capabilities and watching probability distributions of sub-multicast groups for adaptive segment buffering. The sub-multicast groups’ buffers are aligned to the unique virtual buffers managed by DAs for fine-grained buffer updates. A multicast QoE model consisting of multicast rebuffer time, video quality, and quality variation is developed by considering the mutual influence of segment buffering among sub-multicast groups. A joint optimization problem of segment version selection and slot division is formulated to maximize user QoE. To efficiently solve the problem, a data-model-driven algorithm is proposed by integrating a convex optimization method and a deep reinforcement learning (DRL) algorithm. Simulation results based on the real-world dataset demonstrate that the proposed DA-based resource allocation scheme outperforms benchmark schemes in terms of user QoE improvement. Third, we develop an adaptive DA-based resource management scheme to enhance long-term user QoE. Particularly, DAs consist of user status data and data-based models, which can update multicast groups and abstract user swipe features. An adaptive DA management mechanism for DA data processing model selection and update is developed to adapt to user status dynamics. A fine-grained QoE model is established by considering the impact of resource constraints and DA model accuracy. A joint optimization problem of bandwidth and computing resource management is formulated to maximize long-term user QoE. To efficiently solve this problem, a diffusion-based DRL algorithm is proposed, which utilizes the denoising technique to improve the action exploration capabilities of DRL. Simulation results based on a real-world dataset demonstrate that the proposed adaptive DA-based resource management scheme outperforms benchmark schemes in terms of user QoE, with improvements of 18.4\% and 20.5\% under low and high user dynamics, respectively.

In summary, we have investigated DA-based radio and computing resource management from the perspectives of large-timescale resource reservation, real-time resource allocation, and adaptive DA management. The proposed approaches and theoretical results provide valuable insights and practical guidelines for experience-centric resource management in future 6G networks.