摘要
码率自适应(ABR)算法是提升流媒体服务质量的有效方法,主要分为启发式算法和基于学习的算法两类。传统的启发式算法基于固定的规则,难以应对多变的网络环境,基于深度强化学习的算法映射表达能力较好,但其鲁棒性不佳且可解释性较差。针对上述问题,提出一种基于模仿学习的决策树码率自适应算法ABRTree。针对帧级别直播传输系统设计有效的专家ABR算法,并对专家算法的时序经验数据进行离散化处理。采用分类回归树作为码率决策的基础模型,基于专家算法给出的示例数据,采用DAgger算法进行决策树的训练。在此基础上,通过剪枝操作剔除出现较少的样本,从而提升决策树模型的泛化性。实验结果表明,ABRTree在多种视频场景下均能保证画面质量,同时取得较低的端到端延时和较少的卡顿,相比BBA、HYSA和FrameMPC算法,ABRTree算法的QoE性能可以提升1.0%~29.1%,且决策树模型能够直观表达输入特征与码率决策之间的关系,具有较好的可解释性和映射表达能力。
Adaptive Bitrate(ABR)algorithm is an effective method to improve the quality of streaming media services,mainly divided into heuristic and learning-based algorithms.The traditional heuristic algorithm is based on fixed rules,making it difficult to handle changing network conditions,whereas algorithms based on deep reinforcement learning have strong mapping expression capabilities but lack robustness and interpretability.To address these issues,a decision tree ABR algorithm ABRTree based on imitation learning is proposed.An effective expert ABR algorithm is designed for the frame-level live broadcast transmission system,and the temporal empirical data of the expert algorithm is discretized.The classification regression tree is used as the basic model of rate decisions.The DAgger algorithm is used to train the decision tree based on the example data provided by the expert algorithm.Pruning the sample set reduces its size,thus improving the generalizability of the decision tree model.The experimental results demonstrate that ABRTree can guarantee image quality in various video scenarios and achieve lower end-to-end delay and less stalling.Compared with BBA,HYSA,and FrameMPC algorithms,the ABRTree algorithm can improve the Quality of Experience(QoE)performance by 1.0%to 29.1%,and the decision tree model can intuitively express the relationship between input characteristics and rate decisions,with better interpretability and mapping expression ability.
作者
王博
张远
杨咏蓓
WANG Bo;ZHANG Yuan;YANG Yongbei(School of Information and Communication Engineering,Communication University of China,Beijing 100024,China;State Key Laboratory of Media Convergence and Communication,Communication University of China,Beijing 100024,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2023年第5期206-214,共9页
Computer Engineering
基金
国家自然科学基金(61971382)。
关键词
HTTP自适应流媒体
码率自适应算法
决策树
模仿学习
流媒体直播
HTTP Adaptive Streaming(HAS)
Adaptive Bitrate(ABR)algorithm
decision tree
imitation learning
live streaming
