A Conditional-Probability-Distribution Model for Bandwidth Estimation with Application in Live Video Streaming

TL;DR

This paper introduces a probabilistic model for bandwidth estimation in live video streaming, using past data to predict future bandwidth distributions, enabling better control over streaming performance targets.

Contribution

It proposes a novel conditional-probability distribution approach to model future bandwidth, improving prediction accuracy and control compared to existing point estimate methods.

Findings

Enhanced bandwidth prediction accuracy

Ability to meet performance targets over time

Improved control in live video streaming applications

Abstract

Experience of live video streaming can be improved if the video uploader has more accurate knowledge about the future available bandwidth. Because with such knowledge, one is able to know what sizes should he encode the frames to be in an ever-changing network. Researchers have developed some algorithms to predict throughputs in the literature, from where some are simple hence practical. However, limitation remains as most current bandwidth prediction methods are predicting a value, or a point estimate, of future bandwidth. Because in many practical scenarios, it is desirable to control the performance to some targets, e.g., video delivery rate over a given target percentage, which cannot be easily achieved via most current methods. In this work, we propose the use of probability distribution to model future bandwidth. Specifically, we model future bandwidth using past data transfer measurements and then derive a probability model for use in the application. This changes the selection of parameters in application into a probabilistic manner such that given target performance can be achieved in the long run. Inside our model, we use the conditional-probability method to correlate past and future bandwidth and hence further improve the estimating performance.