To design an optimal adaptive video streaming method, video service providers need to consider both the efficiency and the fairness of the Quality of Experience (QoE) of their users. In Reference [8], we proposed a server-side QoE-fair rate adaptation method that considers both efficiency and fairness of the QoE. The server uses Reinforcement Learning (RL) to select a bitrate for each client sharing the same bottleneck link to the server in a way that achieves fairness among concurrent DASH clients and imposes that bitrate by dynamically modifying the client’s Media Presentation Description (MPD) file. In this article, we extend that work to minimize the number of actions the server needs to take to keep the system in its equilibrium state. By incorporating a Recurrent Neural Network, specifically an LSTM model, we modify the server’s training algorithm to achieve improvements in both the quality and the quantity of actions the server takes to guide the client. Performance evaluation of the modified algorithm for clients running both homogeneous and heterogeneous adaptation algorithms showed that the number of server actions dropped by 14% and 22%, respectively, while QoE-fairness improved by at least 6% and 10%, respectively.

中文翻译：

---

使用服务器端强化学习的 QoE-Fair DASH 视频流

为了设计最佳的自适应视频流方法，视频服务提供商需要同时考虑用户体验质量 (QoE) 的效率和公平性。在参考文献 [8] 中，我们提出了一种服务器端 QoE-fair 速率自适应方法，该方法同时考虑了 QoE 的效率和公平性。服务器使用强化学习 (RL) 为与服务器共享相同瓶颈链接的每个客户端选择一个比特率，以实现并发 DASH 客户端之间的公平性，并通过动态修改客户端的媒体呈现描述 (MPD) 文件来强制该比特率。在本文中，我们扩展了这项工作，以最大限度地减少服务器为使系统保持平衡状态所需采取的操作数量。通过结合循环神经网络，特别是 LSTM 模型，我们修改了服务器的训练算法，以提高服务器为引导客户端而采取的操作的质量和数量。对运行同构和异构自适应算法的客户端的改进算法的性能评估表明，服务器动作的数量分别下降了 14% 和 22%，而 QoE 公平性分别提高了至少 6% 和 10%。