Research On Live Streaming Transmission Simulation System And Algorithm

In recent years,with the further development of mobile Internet and 4G and 5G technologies,streaming media services have become more and more popular.At the same time,streaming media transmission has become more and more important in today's network transmission.In recent years,the academia(MIT,Stanford)and the industry(YouTube,Netflix,iQiyi)have developed a series of algorithms to optimize the QoE(Quality of Experience)in streaming media.With the rise of live broadcast and short video services,compared with the VOD(Video of Demand)era,live streaming has added new requirements for real-time interaction(low latency)on the basis of no rebuffer and high video quality.Therefore,facing the new challenge of delay in live broadcast transmission,many researchers started to evaluate the live streaming service,aiming to improve the quality of experience.Therefore,in order to evaluate and optimize the quality of live streaming services,this thesis starts to develop a live streaming transmission simulation system and design algorithm.First of all,this thesis designs and implements a live streaming transmission simulation system.The simulation system is mainly composed of a upper load end module,a transcoding server module,a CDN(Content Deliver Network)module and a client module.The four modules correspond to the four processes of live streaming,upper load,transcoding,CDN,and playback.This thesis completes the development and implementation of the system based on the demand analysis,the outline design,and detailed analysis.Then,the thesis ensures that the simulation system and the real system are similar,and optimizes the acceleration ratio of the simulation system.In order to facilitate the evaluation and support the various algorithms,it provides many application interface.In order to facilitate analysis and debugging,log and visualization functions are added to the simulation system.Finally,the system finished to be released and open source.Secondly,this thesis finished the evaluation of the classic ABR algorithm in the live streaming scenario.For the ABR algorithm's impact on the quality of experience,This thesis selects four typical classic ABR algorithms from the academia and the industry,selects typical live streaming content scenarios:sports live broadcast,indoor live broadcast,and game live broadcast,and selects for network scenarios:enough network environment,medium network environment,weak network environment,and hybrid network.According to the classic QoE(Quality of Experience)evaluation rules of the academic papers,the thesis borrow it.And the thesis add the live streaming feature:delay.Corresponding to each live broadcast scenario,the performance of the classic ABR algorithm is tested in each network environment.Finally,the performance of the four algorithms is analyzed based on the evaluation results,and the reasons for the poor performance of various algorithms in live broadcast scenarios are summarized.Finally,this thesis finished the training and implementation of ABR algorithm RL-LIVE based on reinforcement learning.In view of the above analysis of the classic ABR algorithms,and inspired by the machine learning in the network,the thesis design the algorithm called RL-LIVE with reinforcement learning for live streaming scenarios.The RL-LIVE is designed based on the A3C neural network structure,optimized the State part and added some rules for performance guarantee on the Pensieve.All the training and implementation are based on the streaming media simulator system.Finally,the performance of RL-LIVE is improved in three live streaming scenarios and four network scenarios.Experiments show that the robustness of RL-LIVE under weak network is better than other algorithms.