Optimization Of DASH Streaming Service Quality In Live Broadcast Scenarios

With the rise of live broadcast platforms such as Yingke,DouYu,and HuYa,how to improve the service quality of live video has become the focus of attention.Dynamic Adaptive Streaming over HTTP(DASH)can adaptively switch bitrate of media segment according to the external environment,providing users with high-quality video viewing experience.However,most of the existing adaptive bitrate(ABR)selection algorithms are designed for on-demand streaming services.The nature of live streaming makes it impossible for client to buffer video of long time to cope with the fluctuation of networks.Meanwhile,users often watch video in moving scenes,which makes the network fluctuate more serious.Therefore,it is of great significance to study the ABR selection algorithm in live broadcast scenarios to improve the quality of streaming services.Quality of Experience(QoE)is a commonly used evaluation method to reflect user's viewing experience.Aimming at optimizing the quality of DASH video service in live broadcast scenarios,this thesis analyzes characteristics of two kinds of QoE(integrated QoE and real-time QoE)from the perspective of data driving,using different approaches to model two QoEs.Afterwards,a live streaming media ABR selection algorithm based on reinforcement learning is proposed,which integrates real-time QoE into the learning of ABR strategy.The main work of this thesis is as follows:(1)An in-depth analysis of real traces is conducted to study the integrated QoE and real-time QoE models.Firstly,the random forest model is used to analyze the influence of each QoE perception factor on the quality of user's experience and model the integrated QoE.Secondly,relevant features are extracted to construct LSTM(Long short-term memory neural network)model to address the continuity of real-time QoE.The experiment results show that the LSTM_QoE model can capture the complex time dependence among features and achieve accurate prediction of real-time QoE.(2)To be adaptive to network status variance in live broadcast scenarios,a ABR selection algorithm RL_LSTM_QoE based on reinforcement learning is proposed.Markov model is applied to describe the process of adaptive bitrate selection and appropriate environment state features are selected as the state space for reinforcement learning,so that the state space can fully reflect the changes of external environment during the playback process.Due to the fact that the reward function plays a critical role in reinforcement learning,the output of the real-time QoE prediction model is used as the reward function of the current action to improve the learning ability of the ABR model.(3)The performance of the proposed algorithm is evaluated on LiveStreaming simulation platform.Firstly,the effects of different neural network models equipped in RL_LSTM_QoE algorithm are compared.The experiment results show that the bitrate selection model based on LSTM neural network obtains higher reward values.Then extensive experiments are conducted to compare the performance of several existing ABR selection algorithms over public mobile network datasets.The results show that the proposed RL_LSTM_QoE algorithm performs much better than the comparison algorithms in terms of intergrated QoE,real-time QoE,and other performance measures.In this thesis,a live streaming ABR selection algorithm based on reinforcement learning is designed based on the characteristics of live streaming media.The final model can select appropriate bitrate according to the changes of the external environment to improve the user experience quality.Meanwhile,the proposed algorithm can learn the code rate selection strategy in different network environments with the expansion of network dataset.