ganler / ResearchReading

General system research material (not limited to paper) reading notes.
GNU General Public License v3.0
20 stars 1 forks source link

[ICML workshop, 2019] Real-world Video Adaptation with Reinforcement Learning #11

Closed ganler closed 4 years ago

ganler commented 4 years ago

https://openreview.net/pdf?id=SJlCkwN8iV

RL for video streaming(QoE).

ganler commented 4 years ago

Problem: Video gets stuck and rebuffering...

Old approach: Dynamic Streaming over HTTP(DASH)

image

When Video Rate > Capacity, the video gets stuck and gets into a rebuffering period.

Adaptive Bitrate (ABR) Algorithm

Choose available periods adaptively depending on the network and playback buffer condition.

image

Challenge

The future is unknown...

Past throughput observation cannot predict the coming network condition...

User side:

So ABR has to wait for a while to adapt this noisy cloud of future possibilities.

So we need to plan for the future for a better user experience.

Pensieve

Input: Observation@{Bandwidth, Current BitRate, Buffer} Output: Next bitrate.

Contrib:

Previous fixed ABR algorithms

image

All these are based on fixed heuristics based on designers' insights.

=> Simplified inaccurate model leads to suboptimal performance.

MPC: conservative throughput prediction

image

Pensieve

image

Future Questions

image

Others

When doing the evaluation, we can compare the algorithms with the offline optimal result.

Does xxx generalize?