Machine Learning Based Rate Adaptation with Elastic Feature Selection for HTTP-based Streaming
1. Introduction
In this paper, the writer presents Machine Learning-based Adaptive Streaming over HTTP (MLASH) in order to decrease the effect of prediction error and other factors, leading to higher QoE. MLASH can be incorporated with any existing adaptation algorithm and utilize big data characteristics to improve prediction accuracy.
2. Problems
(1) How to select an appropriate rate for each video segment, while taking many conflicting performance metrics into account, such as the video rate, the rebuffer rate or video rate smoothness;
(2) The future bandwidth is hard to predict;
(3) Only one or a few information can be set as input.
3. Design
MLASH is a flexible learning-based rate adaptation framework with two main components: a) model training and b) video rate prediction. The design principle is to exploit the machine learning technique to train a rate classification model for any existing rate adaptation algorithm.
Two advantages:
(1)this classification model can be offline trained using a number of previous requests, or online updated using streaming data to improve prediction accuracy over time, so as to benefit from big data characteristics.
(2)keep the design flexible, and enable any service provider to use its preferable rate labeling algorithm.
Considered features:
(1)Bandwidth: Last segment bandwidth (LSB), Session average bandwidth (SAB), Moving window average bandwidth (WAB) and Variation.
(2)Buffer size: Current buffer length and Maximal buffer length.
(3)Round-trip time (RTT).
(4)Current video rate.
labeling algorithms:
(1)Bandwidth-based rate adaptation;
(2)Buffer-considered rate adaptation;
(3)Smooth rate adaptation.
4. Evaluation
(1) Check whether our machine learning-based approach can improve the performance of bandwidth-based adaptation. Figure 3 shows that MLASH can predict a video rate that is fairly close to the optimal rate chosen based on the true bandwidth.
(2) Check how MLASH performs as it is incorporated with the buffer-considered adaptation algorithm. Figure 4 suggests that MLASH can select a video rate close to the optimal rate chosen based on the true bandwidth.
(3) Evaluate the performance of MLASH with smoothness-based labeling. The result is that MLASH can achieve a similar switching rate, as compared to the traditional smoothness-based algorithm using the true bandwidth information.
5. Conclusion
MLASH can hence elastically utilize comprehensive features, and, more importantly, avoids the difficulty of bandwidth estimation faced by many existing adaptation algorithms. It can make up for the existing algorithms, nevertheless, MLASH does not raise a new algorithm to improve QoE, which means it cannot get a higher performance fundamentally.