Over the last 20 years, many research projects have used machine learning to create new algorithms that take data over the Internet and try to draw conclusions or make decisions. Unfortunately, although many of these algorithms have performed well in a laboratory setting, the same learned algorithms sometimes underperform when deployed over the real Internet -- or they perform well initially, but decay or produce nonsensical results after some time has passed. These problems may be caused by model mismatch, when laboratory conditions do not match the real world, or by dataset shift, when the environment changes away from the conditions an algorithm was trained to handle, or by the decentralized nature of the Internet where no one person or computer can see how well or how poorly the entire network is working. Whatever the causes, these issues remain vexing obstacles to the practical use of machine learning on the Internet.
This research project will work to discover and validate approaches that allow learned networking algorithms to work reliably over time, in the presence of a varying, unpredictable network and user population that are challenging to simulate faithfully in the lab. The project will explore whether learned algorithms for network video transmission, which accounts for about 75% of Internet usage, can be made robust if the learning happens continually, at the same time and place as deployment, instead of in the lab.
The project will center around a public research experiment, called Puffer: a live video-streaming website run for research purposes. To sample a broad range of network paths that represent the diversity and variability of the Internet, Puffer will stream high-definition broadcast television channels to up to 500 members of the public simultaneously, to be viewed in a Web browser. Puffer will experiment on this live traffic on a continuous basis, teaching itself the best algorithms and parameter values for video streaming and congestion control, and randomizing users to different video-streaming algorithms, including ones contributed by other researchers. If successful, the project will demonstrate (and evaluate over a sustained period of time) whether it is possible to make machine learning robust in this real-life networked setting.
The project website is https://puffer.stanford.edu, which will be maintained with code and data repositories for at least five years.
This 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.
