Reinforcement Learning Based Multipath QUIC Scheduler for Multimedia Streaming.

Area of Science:

• Computer Science
• Network Engineering

Background:

• Modern devices utilize multiple network interfaces (cellular, Wi-Fi, Ethernet).
• Multipath TCP (MPTCP) is standard for multipath, but Multipath QUIC (MPQUIC) offers advantages.
• Multipath schedulers critically impact transport performance, with existing options like minRTT and redundant schedulers offering trade-offs.

Purpose of the Study:

• To develop a novel MPQUIC scheduler enhancing multimedia streaming quality.
• To address diverse application requirements (e.g., low latency for web, low jitter for video).
• To improve video chunk download times by considering both delay and throughput.

Main Methods:

• Proposed a Multipath QUIC (MPQUIC) scheduler utilizing Deep Q-Network (DQN) for deep reinforcement learning.
• Incorporated delay and throughput as rewards to optimize video chunk download.
• Developed a chunk manager to provide video information to the scheduler and tuned learning parameters.

Main Results:

• The proposed MPQUIC scheduler was implemented on the Linux kernel and tested using Mininet.
• Evaluation demonstrated that the novel scheduler outperforms legacy schedulers by at least 20%.
• The approach effectively balances delay and throughput for enhanced multimedia streaming.

Conclusions:

• The developed DQN-based MPQUIC scheduler significantly enhances multimedia streaming quality.
• This approach offers a superior alternative to legacy schedulers for applications with specific performance needs.
• The findings highlight the potential of deep reinforcement learning in optimizing network transport protocols.