Policy Learning For Video Streaming
Lists
Only you see this
Abstract
Facilitating good quality of experience (QoE) for Internet-based video services is a crucial real-world challenge. With remote/hybrid work, education, and telemedicine being here to stay, poor video quality adversely impacts the economy and society at large. The key algorithmic challenge in this context is adaptive bitrate selection (ABR) - continuously adjusting the video bitrate (resolution) to the prevailing traffic conditions. ABR algorithms struggle to maintain high resolutions while avoiding video stalls and long "lags behind live'', and are the subject of extensive attention. In particular, ABR has, in recent years, been approached from different ML perspectives. However, disillusionment with applications of end-to-end deep reinforcement learning (DRL) to ABR have effectively led to abandoning policy learning for ABR altogether in favor of control-theoretic optimization methods. We demonstrate that, through more nuanced policy learning, substantial improvement over the state-of-the-art is achievable. Specifically, we show that applying deep-Q-learning to the output of a supervised predictive model bests alternative approaches. As we believe that the ABR domain is an exciting new playground for policy learning, we release our code for ABR policy learning and experimentation to facilitate further research.
BibTeX
@misc{birka2024policy,
author = {Ofir Birka and Yedid Hoshen and Michael Schapira},
title = {{Policy Learning For Video Streaming}},
year = {2024},
url = {https://openreview.net/forum?id=CW2aryHm95},
howpublished = {OpenReview},
note = {Submitted to ICLR 2024},
}