Fastconv: Fast Learning Based Adaptive BitRate Algorithm for Video Streaming

Abstract

For video streaming, Adaptive BitRate (ABR) algorithms are usually used to improve end-to-end user’s Quality of Experience (QoE). Many of the state-of-the-art ABR algorithms are based on simplified models, leading to conservative predictions of real situations. To optimize the QoE in complex end-to-end transmission environments, ABR algorithms based on Deep Reinforcement Learning (DRL) has shown a great improvement compared to traditional algorithms. However, the slow convergence of existing DRL-based ABR algorithms limits the QoE performance under dynamic video streaming environments. In this paper, we propose Fastconv, a novel DRL-based ABR algorithm that has a fast convergence speed to ensure a satisfactory QoE performance. Our work can be mainly divided into two parts. First, we preprocess the input data with large fluctuation in order to obtain the steady input and reduce the indeterminacy of convergence. Second, in order to reduce the structural complexity of the neural network itself and the number of parameters, we propose a neural network architecture based on multiplexed convolution kernel. Experiment results based on a real traced mobile dataset have demonstrated that Fastconv outperforms both the traditional and DRL-based ABR algorithms in terms of QoE.