Event Details
Video Streaming with Multipath Protocols: Leveraging Learning Techniques
Presenter: Amir Sepahi
Supervisor:
Date: Thu, November 26, 2026
Time: 09:00:00 - 00:00:00
Place: Zoom - see below.
ABSTRACT
Zoom Link: https://uvic.zoom.us/j/87866073641?pwd=TrFQMG0CAVa1dhAZTvrH676DeUD3fW.1
Meeting ID: 878 6607 3641
Passcode: 305248
Abstract:
Modern video streaming applications require high bandwidth, low latency, and reliable connectivity, particularly in mobile environments characterized by rapidly changing network conditions. Multipath transport protocols, such as Multipath QUIC (MPQUIC), enable the simultaneous use of multiple network interfaces, including Wi-Fi and cellular networks. However, efficiently scheduling packets across heterogeneous network paths remains challenging due to variations in bandwidth, latency, packet loss, and application requirements.
This seminar presents a series of learning-based approaches developed to improve the performance and quality of experience (QoE) of video streaming over multipath networks. First, Meta-DAMS introduces a hybrid meta-reinforcement learning framework that combines offline pretraining with online adaptation to enable efficient delay-aware packet scheduling. Building upon this approach, LiveStream Meta-DAMS incorporates video coding characteristics, including group of pictures (GOP), scalable video coding (SVC), and MPEG-DASH, to prioritize critical video packets and improve streaming performance. Next, Fed-MVP introduces a federated reinforcement learning architecture that distributes model training across cloud, edge, and client devices, enabling scalable and adaptive scheduling while keeping fine-grained user data locally. Finally, KESTREL extends this research toward practical deployment through an intent-aware multipath scheduler that combines offline learning with lightweight, real-time scheduling decisions on mobile devices.
Experimental evaluations demonstrate significant improvements in learning efficiency, transmission delay, video quality, and playback continuity compared to conventional multipath scheduling approaches. These contributions highlight the potential of learning-based multipath communication to support reliable, adaptive, and high-quality video streaming in dynamic wireless environments.
