A New Framework for Improved Video Streaming

Video streaming accounts for 65% of all Internet traffic, but unpredictable network conditions can lead to low-quality video and delays in streaming. To address this, Sagar Patel, a computer science Ph.D. candidate in UC Irvine’s Donald Bren School of Information and Computer Sciences (ICS), has developed a new approach. Working with Sangeetha Abdu Jyothi, his Ph.D. advisor; Junyang Zhang, a computer science major in ICS; and Nina Narodytska, a researcher at VMware, Patel came up with Gelato, a deep reinforcement learning (Deep RL) adaptive bitrate video streaming controller, and Plume, its training framework.

A paper outlining this work, “Practically High Performant Neural Adaptive Video Streaming,” received the Best Paper Award at the ACM International Conference on emerging Networking EXperiments and Technologies (CoNEXT 2024), distinguishing it among 36 accepted and 231 submitted papers.

“Unlike previous solutions that excel only in simulation, Gelato generalizes to the noise and uncertainty of real-world Internet conditions,” says Patel. Deployed on the Puffer live streaming platform, which serves over 300,000 users, Gelato delivers statistically significant improvements in video quality and reduces buffering by up to 75% compared to prior state-of-the-art controllers.

“More than seven years after Deep RL was first applied to adaptive video streaming and over five years after papers identifying its open challenges were generally accepted, we show that many of these challenges can be mitigated by addressing a simple yet overlooked issue: data skew,” explains Patel. While several standard solutions for data skew exist, these solutions fail to account for the Internet’s role in the training process and thus do not provide meaningful improvements. “In contrast, Plume directly targets skew in network traces by automatically identifying critical features, clustering them, and balancing the training data.”

He adds that this work is merely a starting point. “After years of negative results, our approach reaffirms the potential of Deep RL to advance networking,” he says. “As research in general intelligence with Deep RL progresses, our contributions mark an important step toward the coevolution of AI and the networked systems that support it.”

— Shani Murray