Amir Sepahi
Topic
Video Streaming with Multipath Protocols: Leveraging Learning Techniques
Department of Electrical and Computer Engineering
Date & location
-
Monday, November 30, 2026
-
9:30 A.M.
-
Engineering Computer Science Building
-
Room 467
Reviewers
Supervisory Committee
-
Dr. Lin Cai, Department of Electrical and Computer Engineering, University of Victoria (Supervisor)
-
Dr. Navneet K. Popli, Department of Electrical and Computer Engineering, UVic (Member)
-
Dr. Kui Wu, Department of Computer Science, UVic (Outside Member)
External Examiner
- Dr. Hai Jiang, Department of Electrical and Computer Engineering, University of Alberta
Chair of Oral Examination
-
Prof. Kevin Kerr, Department of Writing, UVic
Abstract
Video streaming needs high bandwidth and low latency over fast-changing wireless links. Multipath protocols such as Multipath QUIC (MPQUIC) use a device’s several interfaces at once, so the packet scheduler shapes delivery performance. This dissertation develops four learning-based multipath schedulers for video streaming, the first three implemented over mpquic-go and evaluated in trace-driven emulation.
First, we propose Meta-DAMS, a delay-aware scheduler whose hybrid meta-reinforcement learning architecture combines offline pretrained meta-models with online fine-tuning based on the Asynchronous Advantage Actor-Critic (A3C) algorithm. After enough episodes it keeps end-to-end delay at or below 50 ms for 98% of packets, against about 88% for a deep Q-network (DQN)-based scheduler, and its online learning phase runs almost three times faster. Second, we present LiveStream Meta-DAMS (LSMeta-DAMS), which tailors the scheduler to live video streaming by exploiting video coding structure, prioritizing I-frame packets, and introducing a modified A3C (MA3C), showing up to 32% improvement in learning, 25% reduction in download time, 15% enhancement in video quality, and 35% reduction in stalling time. Third, we propose Fed-MVP, a federated multi-homing video streaming protocol that decentralizes learning across a cloud–edge hierarchy: meta-models are served by edge servers, fine-tuned on clients with Proximal Policy Optimization (PPO), and aggregated with a Modified Federated Averaging (Mod-FedAvg) algorithm. It achieves up to 61% and 43% faster convergence in the test and training phases, 40% shorter download time, 35% better video quality assessments, and 38% less stalling.
Finally, KESTREL carries this work onto the phone, which turns application needs and user limits into a policy under hard rules that the sender follows for every packet, with all learning offline. We built its picoquic scheduler, trained model, and Android client. When the WiFi rate fell below the stream rate, about 2% of KESTREL’s packets were late against about 41% for minRTT, ECF, and BLEST, because KESTREL also estimates each path’s time to deliver the bytes in flight. After we corrected picoquic’s RTT estimate, the others recovered substantially, but KESTREL still outperformed all benchmarks by having the fewest late packets on average and ensuring the user’s cost rule.