Multi-agent Reinforcement Learning for Live Video Streaming in Vehicular Networks

Live Video Streaming (LVS) services are critical in supporting real-time applications in Internet of Vehicles (IoV) by transmitting real-time generated video content from streaming server to mobile vehicles. However, due to restricted spectrum resources and high mobility of vehicles, LVS suffers from notable performance degradation. Accordingly, our paper investigates the problem of LVS-IoV by synthesizing multicasting and Scalable Video Coding (SVC)-based encoding techniques with the goal of maximizing Quality of Experience (QoE). We resolve the LVS-IoV by decoupling it into three sub-problems: vehicle grouping, quality selection, and resource allocation. Firstly, we propose a K-means-based mechanism for vehicle grouping that considers the geographical distribution and dynamic channels of vehicles. Secondly, we develop the Value Decomposition Network for quality selection which obtains the optimal reward with fast convergence via global value function decomposition and centralized training followed by distributed execution. Lastly, we propose a sub-gradient algorithm to achieve optimal resource allocation within a few iteration steps. We build simulation model and perform extensive evaluation, which demonstrates its superiority compared to other competitive methods.