Low-Delay Adaptive Video Streaming Based on Short-Term TCP Throughput Prediction

TL;DR

This paper proposes a new adaptive video streaming algorithm that uses short-term TCP throughput predictions to improve live streaming quality and reduce interruptions, especially in fluctuating wireless networks.

Contribution

It introduces a throughput prediction-based adaptation algorithm for live streaming, utilizing short-term TCP throughput predictions and dynamic accuracy estimation to enhance QoE.

Findings

Simple Moving Average performs best among studied prediction methods.

Underestimation errors follow a truncated normal distribution.

Overestimation errors are best modeled by a Lomax distribution.

Abstract

Recently, HTTP-Based Adaptive Streaming has become the de facto standard for video streaming over the Internet. It allows the client to adapt media characteristics to varying network conditions in order to maximize Quality of Experience (QoE). In the case of live streaming this task becomes particularly challenging. An important factor than might help improving performance is the capability to correctly predict network throughput dynamics on short to medium timescales. It becomes notably difficult in wireless networks that are often subject to continuous throughput fluctuations. In the present work, we develop an adaptation algorithm for HTTP-Based Adaptive Live Streaming that, for each adaptation decision, maximizes a QoE-based utility function depending on the probability of playback interruptions, average video quality, and the amount of video quality fluctuations. To compute the utility function the algorithm leverages throughput predictions, and dynamically estimated prediction accuracy. We are trying to close the gap created by the lack of studies analyzing TCP throughput on short to medium timescales. We study several time series prediction methods and their error distributions. We observe that Simple Moving Average performs best in most cases. We also observe that the relative underestimation error is best represented by a truncated normal distribution, while the relative overestimation error is best represented by a Lomax distribution. Moreover, underestimations and overestimations exhibit a temporal correlation that we use to further improve prediction accuracy. We compare the proposed algorithm with a baseline approach that uses a fixed margin between past throughput and selected media bit rate, and an oracle-based approach that has perfect knowledge over future throughput for a certain time horizon.