Candidate: Alex Duanmu
Title: Adaptive Streaming: From Bitrate Maximization to Rate-Distortion Optimization
Date: April 12, 2021
Time: 2:00 PM
Place: REMOTE ATTENDANCE
Supervisor(s): Wang, Zhou
Abstract:
The fundamental conflict between the increasing consumer demand for better quality-of-experience (QoE) and the limited supply of network resources has become significant challenges to modern video streaming systems. State-of-the-art adaptive bitrate (ABR) streaming algorithms tangle bitrate with quality, resulting in inefficient use of network resources and sub-optimal QoE. We aim to develop an information-theoretic framework namely rate-distortion optimized streaming (RDOS) to balance the contrast objectives in streaming applications. Distinct from the existing models, RDOS must operate at any given point along the rate-distortion curve, as specified by a trade-off parameter. We develop a full system implementation of RDOS, including a theoretically-grounded objective QoE model, a recurrent neural network-based control policy, and a large-scale streaming environment to train as well as validate ABR algorithms. Using reinforcement learning, we optimize the policy network with Gated Recurrent Units for rate-distortion performance over a database of training videos and network traces, effectively resolving the long-term dependencies between streaming events. Across an independent test set, the optimized method outperforms existing ABR methods, which is supported by a large-scale subjective quality assessment experiment. To further extend the current work, we will develop encoding profile adaptive ABR algorithms and multi-agent adaptive streaming methodologies.