Machine Learning Based Rate Adaptation with Elastic Feature Selection for HTTP-based Streaming
In rate adaptation, the error in bandwidth estimation could significantly degage the performance. Therefor, this paper presents Machine Learning-based Adaptive Streaming over HTTP.
The rate adaptation can be modeled as a classification problem in machine learning: given the network environment feature, to determine which rate the client should be requested. Instead of proposing a new adaptation algorithm, the design principle is to exploit the machine learning technique to train a rate classification model for any existing rate adaptation algorithm.
The architecture contains two parts: a) model training and b) video rate prediction. In particular, say that a client has already received k video segments.
- Training: They use 1) a set of historical network-related information measured from the first (k-l) received segments as the feature set and 2) the true best video rate r* of the k-th received segment as the corresponding label to train a classification model.
- Classification: the server feeds the reported features into the model, predicts the best rate of the requested segment
Features:
- Bandwidth-rated features:
- Last segment bandwidth (LSB)
- Session average bandwidth (SAB)
- Moving window average bandwidth (WAB)
- Variation
- buffer size-related feature:
- Current buffer length
- Maximal buffer length
- RTT
- Current video rate
In this paper, the author selects different label method:
- Bandwidth-based rate adaptation
- Buffer-considered rate adaptation
- Smooth rate adaptation
Experiment Settings:
- Trace based experiments: trace provided by S. Basso [MMSys’14]. It includes 10,000 test everyday. Information includes video rate, time required, RTT.
- Rate set: 100, 150, 200, 250, 300, 400, 500, 700, 900, 1200, 1500, 2000, 2500, 3000, 4000, 5000, 6000, 7000, 10000, 20000.
- Video duration: 30s
Results:
- Bandwidth-based rate labeling: a) SAB selects the video rate cannot adapt to network dynamics and is very likely to pick a more conservative rate b) pure buffer-based algorithm does not consider the bandwidth, and could select a rate either much lower or higher than the available bandwidth.
- Buffer-considered rate labeling: a) the trend of result is similar to bandwidth-based rate labeling.
- Smoothness-based rate labeling: a) bandwidth-related feature results in less rate switching than buffer based feature.