MobStream: A Reinforcement-Driven Mobile Streaming Methodology Over Multipath QUIC

With the rapid growth of mobile communications, using mobile devices such as smartphones has become a trend. However, the limited uplink of the cellular network constrains the uploading quality. By integrating the link capacity of both cellular networks and WiFi, concurrent multipath transmission, such as multipath QUIC (MP-QUIC), becomes a promising solution for alleviating the uplink bottleneck issue. This paper proposes MobStream, a novel reinforcement learning-driven solution based on MP-QUIC. MobStream maximizes streaming capacity by fully utilizing the bandwidth of both WiFi and cellular networks through MP-QUIC. To address the issue of reduced performance caused by the differences in path quality between WiFi and cellular networks, MobStream incorporates partial reliability and a layered coding scheme, paving the way to adapt to varied network conditions without harming the bandwidth utilization. We formulate the concurrent transmission problem as a stochastic optimization task and demonstrate its solution using reinforcement learning methods. Furthermore, to ensure fairness in transmission among other single-path protocols, we further introduced a fairness factor to the reinforcement learning method. Extensive experiments demonstrate that MobStream outperforms state-of-the-art solutions in terms of bitrate, packet loss, and delay.