New video algorithm with improved streaming efficiency

As Corona 19 prevents going out, the influence of the streaming service Netflix has also led to a runaway. Streaming service providers also make efforts to maintain the service by dropping the bit rate and resolution. In the midst of this, an algorithm that significantly improves communication efficiency without deteriorating the image quality has appeared and attracts attention.

Fugu was developed by a research team such as Francis Yang of Stanford University. The research team built Puffer, a streaming service provided by Stanford University, and developed an algorithm by performing machine learning based on the data obtained from it.

Currently, general streaming uses a buffer-based algorithm BBA (Buffer-Based Algorithm), which aims to play videos seamlessly. It checks how many images are accumulated in the buffer, which is the PC area that temporarily stores data, and adjusts the image quality accordingly. For example, if BBA determines that there are only 5 seconds of video stored on the PC, BBA sends low-quality video to the PC. Conversely, if it is determined that there is a 15 second margin, it sends high-quality video that takes time to communicate. BBA has long been widely used in streaming services because this method can provide the highest possible quality while maintaining smooth video playback.

BBA is a simple, but not sophisticated, old method, so researchers have been looking for efficient communication algorithms. However, the problem was that most of these studies were based on machine learning held in a virtual environment and were not well suited to the real Internet environment.

The research team actually started Purpur, a free streaming service, delivered 38.6 years worth of videos to 63,508 users, and conducted deep learning learning with supervised learning using data obtained in a more realistic environment. The algorithm made in this way predicts the data transmission time ahead of the fugu, and can perform more efficient congestion control than the BBA that only refers to the current buffer situation.

As a result of verifying the effects of five algorithms, including Fugu, in Puffer, the research team said that Fugu showed excellent performance in image quality, resolution, and video pause time. In addition, even in the test of delivering the video with a random algorithm for the viewer, the viewer who saw the video delivered to the puffer said that they enjoyed the video 5 to 9% longer than other algorithms.

The research team says that machine learning has found a way to overcome the difference between reality and simulation, and that it will be quite interesting considering that this will solve many problems. Also, through this study, he learned that in order to create a learning-type algorithm that shows strong performance on the Internet, real-time training with data in a real streaming environment requires an algorithm with a sophisticated structure and conciseness. Explained that education is essential. Related information can be found here .