| Nowadays,the live-video streaming service is widely used in people's daily life.The congestion control algorithm is responsible for the task of adjusting the speed in time according to the network conditions,and it plays an essential role in ensuring the user's experience.However,due to the complex and dynamic real network,the congestion control algorithms are still facing a huge challenge.The academia has introduced machine learning to solve this problem.But the designs of previous research are mainly based on simulator and therefore is difficult to adapt to the real world,which may lead to poor performance.This research relies on the live video data of a partner.Through the online performance and using the online data,it is verified that the problem exists when the rule-based algorithm GCC and the model trained by the simulator are used in the real network.The necessity of training in the real world is confirmed by data analysis and testing the model in the real world.In view of the above motivations,this project takes the online training as the background to research,then design and implement an online training algorithm based on WebRTC.During the actual operation of the live-video streaming service,the network data of the WebRTC side is calculated,integrated and then transmitted to the training server.The training process is conducted without disturbing the live video-streaming service,which makes the live video-streaming service get rid of simulators.This research also analyzes the reason for the poor performance when the reinforcement learning algorithms are used in video applications,and then improve the scheme.For the same RL algorithm,the average difference between the predicted bitrate and the actual bandwidth is reduced by 28.70%,the RTT is reduced by 40.46ms,and the packet loss rate is reduced by 19.97%. |
PDF Full Text Request