NSF-BSF: CNS Core: Small: Machine Learning for Real-Time Network Rate Control

点击翻译按钮获取中文摘要

Any time an application uses the Internet, behind the scenes, rate control algorithms decide how quickly to transmit data. These algorithms have a critical task; sending too slowly causes delay or reduced video quality, and sending too quickly causes congestion that impacts both the sender and other users. Rate control is a persistent challenge in networking, as it has to deal with a wide range of dynamic environments while making millisecond-level decisions with limited information. This project is developing new approaches to rate control based on an area of machine learning known as reinforcement learning, leading to potential improvements in performance and functionality.At a high level, the project seeks to develop a fundamental understanding of the use of reinforcement learning for rate control, and apply that understanding to the areas of adaptive bitrate (ABR) video as commonly used in modern web-based video, and to transport layer congestion control, such as the Transmission Control Protocol (TCP). The work will begin with algorithmic foundations by exploring what level of complexity of learning algorithm (ranging from bandit algorithms to deep neural networks) is necessary to achieve high performance. Next, the project will broaden the semantics of inputs and outputs of rate control, including a scavenger rate control protocol and an improved multipath TCP. Finally, the project will use novel automated methods to improve the robustness of rate control protocols in unexpected environments.The results of this project can offer significant performance improvement for deployed protocols, which is of increasing need as modern and emerging applications have ever more demanding network requirements. For example, low latency communication and high quality real-time video are valuable for interactive conferencing, augmented and virtual reality, Internet of Things, edge computing, and more. The project also plans to provide research opportunities for underrepresented groups.The results of this project, including papers and open-source code, will be available at http://pccproject.net and at the code repository, https://github.com/PCCprojectThis award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.