The buffer occupancy is a more reliable signal than bandwidth estimation for ABR algorithms.
Notes
Main Idea:
Utilize capacity estimation when needed (it is not necessary at a steady state).
Able to reduce rebuffer rate by 10-20% compared to Netflix's default ABR algorithm.
Overview and Background:
ABR algorithms balance two goals: maximize video quality and minimize rebuffering events.
One approach is to estimate the future capacity based on prior observations.
Current ABR algorithms estimate capacity and make adjustments such that the rate is more conservative when the buffer is underrunning and more aggressive when the buffer is close to full.
What differentiates this paper:
Here, they will only use the buffer to choose the video rate and then view the capacity estimate when needed.
There are two phases of operation:
Steady-state phase when the buffer is being built up.
Startup phase when the buffer is still growing from empty.
Estimation is not needed in the steady-state phase.
In the baseline algorithm, there is still room for improvement:
The baseline algorithm doesn't address variable bit-rate (VBR) video encoding.
The baseline algorithm is optimized for steady-state phase.
Design of Buffer-Based Algorithm:
Focus on pure buffer-based design: select video rate directly as a function of current buffer level.
In the steady-state phase, choose video rate based only on playback buffer.
In the startup phase, buffer has little information, so use capacity estimate to choose video rate.
BBA-0 -> BBA-1 -> BBA-2:
BBA-0: Basic buffer-based algorithm that maps buffer size to video rate.
BBA-1: Improves BBA-0 by addressing VBR video encoding. Uses "reservoir" to account for VBR fluctuations.
BBA-2: Further improves BBA-1 by optimizing the startup phase using a more aggressive capacity estimation.
