| User-perceived Quality of Experience(Qo E)plays an important role in the network video delivery system.Video traffic is the most important type of network traffic in the Internet.There have been a lot of studies on client-side adaptive bitrate algorithms,which have effectively improved the Qo E of single-player.However,when multiple video players share a bottleneck link,adaptive bitrate algorithms just maximize their Qo E independently and congestion control can only provide connection level fairness.Therefore,the current video delivery system has the problem of unfair Qo E among video users.In Qo E fairness optimization of multiple video streams,a distributed video stream fairness scheduling strategy based on federated deep reinforcement learning is designed to solve the problem of low bandwidth utilization caused by unfair bandwidth allocation and difficult convergence of the distributed algorithm in cooperative control of multiple video streams.It predicts a reasonable weight of bandwidth allocation for the current video stream according to its player status and global features provided by the server,and then the congestion control protocol allocates a proportion of available bandwidth for each video stream in the bottleneck link that matches its weight of bandwidth allocation.This strategy trains a local prediction model on each client and periodically performs federated aggregation to generate optimal global scenarios.In addition,in order to reduce the instability of the environment in multi-stream scenarios,constructs global parameters containing the state information of the video system to improve the performance of the distributed scheduling algorithm.Experimental results show that by introducing global parameters,the Qo E fairness and overall Qo E efficiency of the algorithm can be improved by 10% and 8%,respectively.Compared with the latest scheme,the Qo E fairness and overall Qo E efficiency are improved by 10% and 9%,respectively.Figure 17;Table 2;Reference 51... |
