This is my notes reading ICME’22 PAR.

Overview

ABR algorithm can be improved by adding payload awareness. Two main modules used:

1. An online payload-aware throughput predictor, which utilizes a lightweight neural network to predict future bandwidths via past observations, and then calculate the corresponding downloading rates for chunks in different bit rates.
2. A bit rate selector aiming at maximizing the given QoE function

Background

Observed throughput is distinguished for chunks in different bit rates, even if they are downloaded from the same moment.

Estimation error due to ignorance of payload difference will be exacerbated when the gap of bit rate levels increases.

The high volatility of network conditions will also enlarge this estimation deviation.

Detail

3 components

1. The client, running dash.js in web browser.
2. The video server, basically a bunch of sliced videos and a running nginx, nothing fancy.
3. The magic, ABR server powered by neural network.

How it works: client request video chunks from the video server, and send streaming metadata to the ABR server, which then use the NN to predict future throughput and response with a switch request, controlling the video streaming.

In experiment, tc is also used to limit network speed to simulate poor network.

Environment: the LTE dataset and the HSDPA dataset.

$$QoE = \sum^N_{i=1}(Q(i) - \mu W(i)) - \sum^N_{i=2}|Q(i) - Q(i-1)|$$

• $W(i)$: rebuffering time
• $Q(i)$: chunk bit rate

---