1 of 18

Uncertainty-Aware Robust Adaptive Video Streaming with Bayesian Neural Network and Model Prediction Control

Oct. 1, 2021

Nuowen Kan, Chenglin Li, Caiyi Yang, Wenrui Dai, Junni Zou, Hongkai Xiong

@Shanghai Jiao Tong University, P.R. China

2 of 18

Adaptive Video Streaming

2

Streaming a video online

Quality of Experience (QoE)

Video quality

Rebuffering time

Dynamic Streaming over HTTP (DASH)

2. Response: Target video chunk version

Adaptive bitrate (ABR) algorithm

Throughput

Player buffer

Bitrate

ARB control

Goal: Improving

240p

360p

460p

720p

1080p

When we stream a video from the internet, we all have this annoying experience -- the video suddenly gets stuck and starts rebuffering

Adaptive video streaming is a popular solution to address the issue of quality of experience degradation caused by rebuffering and quality fluctuation for users.

We consider a typical adaptive video streaming system, where an original video is temporally divided into multiple chunks at the server. Each chunk has a fixed time duration and is further encoded into several quality versions with different bitrates. At the client, a bitrate controller is employed for requesting an optimal version from these available bitrate versions for each chunk, so as to maximize the user’s QoE under constrained network throughput.

Here, we focus on designing a rate adaptation algorithm for the bitrate controller

3 of 18

Why is adaptive bitrate algorithm challenging?

3

The network throughput dynamics are uncertain and hard to predict accurately.

It is hard to generalize to all the heterogeneous network and user conditions.

Throughput distribution

For example, 4G -> 3G, hall -> toilet, …

Possible dynamic 1

Possible dynamic 2

There are mainly two challenges in designing a good adaptive bitrate algorithm.

One is the uncertain throughput dynamics. At this point, how would the future throughput look like? it goes higher, or it could remain steady for a while and then drop, in fact it can be one of the many possibilities -- past throughput observations cannot accurately predict what will happen exactly . It’s very challenging to select a bitrate that adapts to the real throughput.

Another problem is that the distribution of the network throughput is various in the real world. For example, when we switch from a 4G signal to a 3G signal or when we walk into a toilet, the throughputs distribution that we experienced will change a lot. While most ABR algorithms based on throughput prediction struggle to generalize to a wide range of network environments, especially for data-driven methods.

4 of 18

��Our contribution: BayesMPC

4

Uncertainty-aware throughput prediction
More robust bitrate adaptation over different network conditions
Competitive QoE performance (9.0% gain) over DRL-based ABR algorithm (Pensieve, etc.)

Model Predictive Control

Bayesian Neural Network

(BNN)

(MPC)

In this paper, we propose a Bayesian neural network based robust ABR algorithm, named BayesMPC, with the MPC-based rate adaptation framework. We aim to achieve both superior and robust QoE performance for adaptive video streaming.

First, we adopt a BNN-based model to improve the prediction accuracy of the network throughput and capture the prediction uncertainty.

Second, by using the prediction results, an uncertainty-aware robust MPC strategy is then proposed to improve the robustness of bitrate adaptation. The uncertainty measurement of future prediction at each step can tell us how confident the prediction is, which can significantly reduce the risk of rebuffering events.

Finally, we implement BayesMPC for adaptive video streaming on three real-world network trace datasets. The results show that BayesMPC outperforms the overall QoE performance and the generalization performance across a wide range of network and user conditions.

5 of 18

Previous Related Works

5

Classic control theory-based : pick bitrate based on predicted throughput & buffer occupancy

PANDA [JSAC’14], RobustMPC [SIGCOMM’15], CS2P [SIGCOMM’16]
BBA [SIGCOMM’14], BOLA [INFOCOM’16], Oboe [SIGCOMM’18]

Machine learning (ML)-based: use neural network to directly learn an optimal ABR policy

D-DASH [IEEE TCCN’17], Pensieve [SIGCOMM’17], Hot-DASH [ICNP’18],
Comyco [IEEE JSAC’21] etc.

Inaccurate throughput prediction, suboptimal performance

Only perform well in a specified throughput domain

Before we talk about BayesMPC, let's first look at previous schemes

In general, existing ABR algorithms can be classified into two main categories: classic control theory-based and machine learning-based

The first category determines the bitrates based on the predicted future throughput and buffer occupancy. Because of the challenges of throughput prediction, they always cannot achieve a good performance in practice.

In contrast, recently proposed machine learning-based can achieve outstanding performance, especially the deep reinforcement learning-based, for example, Pensieve from SIGCOMM 2017

They learn an optimal mapping between the dynamic states and bitrate selections with the help of powerful approximation capacity of neural network. However, some drawbacks of these algorithms can be found in comparison to the others. For example, they can only perform well in a specified domain, since their neural network weights can hardly generalize to a new environment that differs from the training data.

6 of 18

How to make a reliable throughput prediction?

Solution 1 : Measure the aleatoric uncertainty

6

Epoch

Stochastic dynamics – aleatoric uncertainty

Probability

Point prediction

Distribution prediction

Prediction

Probability

In fact, classic control theory-based ABR algorithms can always achieve a stable performance in practice. And the prediction accuracy of environment dynamics, such as the network throughput, is a critical ingredient for improving the performance of these algorithms, because even a small prediction bias may significantly influence the performance of later planning.

However, it hard to construct an accuracy dynamic model for the prediction. Due to the limitation that a video client can hardly access to the adequate network information in the chunk downloading period, it is challenging to predict the stochastic future network throughput, we call it the aleatoric uncertainty. In order to capture the aleatoric uncertainty, a probability distribution of the user’s network throughput rather than a single deterministic value point should be considered at each step.

7 of 18

How to make a reliable throughput prediction?

7

Solution 2 : Measure the epistemic uncertainty

Which function matches the real dynamics?

Insufficient data cannot uniquely determine the underlying system exactly.
Epistemic uncertainty about the dynamic function remains .

Epistemic uncertainty

8 of 18

How to make a reliable throughput prediction?

8

aleatoric uncertainty

Assume the probability of network throughput follows

epistemic uncertainty

Introduce the stochasticity to neural network weights

(Bayesian neural network)

9 of 18

Bayesian Neural Network

9

each weight has a fixed value

each weight is assigned a posterior distribution

Sample the weights at each calculation

Neural Network

Bayesian Neural Network

10 of 18

Throughput Prediction with the Bayesian Inference

Predicted distribution

10

Past throughput

Train with variational approximation:

Learn a variational posterior approximation of NN weights

Dataset

True Bayesian posterior

11 of 18

Generalization bound of the throughput prediction

Generalization error

11

The generalization error of the predictor can be minimized with the guarantee of PAC-Bayesian theorem [Henderson 2016].

Bayesian posterior over

Data distribution

Training data

Empirical error

Bayesian prior

over

KL divergence

Function independent of

Enable a bounded generalization error in throughput prediction

12 of 18

Uncertainty-aware robust rate adaptation

12

Confidence region

Past

Future

True throughput

Predicted

Optimized bitrate

Past bitrate

Index of video chunk

Prediction horizon

QoE metric for users

Predicted worst-case

System dynamics

13 of 18

Design: system framework of BayesMPC

Combining classical control with an ML predictor

13

14 of 18

Uncertainty-aware network throughput prediction

14

A confidence region of the throughput prediction can be measured

Calculate the confidence region:

BNN performs better on QoE than point estimate (PE) prediction methods

Confidence region (Shadow)

15 of 18

Comparison of QoE performance on traces from FCC

15

better

QoE metric:

BayesMPC improves the best previous scheme by 9-24%

16 of 18

Comparison of QoE performance on unfamiliar traces

Tested with the linear and log-form QoE metric settings

16

Use the training data of FCC, test with new data from Oboe

better

BayesMPC is always optimal

17 of 18

Summary

17

BayesMPC uses Bayesian neural network to predict network throughput and measure the prediction uncertainties
Based on these uncertainties, BayesMPC can robustly optimize the bitrate
BayesMPC outperforms existing approaches across a wide range of network conditions and QoE preferences

https://github.com/confiwent/BayesMPC

https://main.nuowen.pro/

18 of 18

Q&A

Thank you!