DNN Based Adaptive Video Streaming Using Combination of Supervised Learning and Reinforcement Learning

Abstract

Video streaming has emerged as a major form of entertainment and is more ubiquitous than ever before. However, as per the recent surveys, poor video quality and buffering continue to remain major concerns causing users to abandon streaming video. This is due to the conditional rule-based logic used by state-of-the-art algorithms, which cannot adapt to all the network conditions. In this paper, a Deep Neural Network (DNN) based adaptive streaming system is proposed, which is trained using a combination of supervised learning and reinforcement learning that can adapt to all the network conditions. This method aims to pre-train the model using supervised learning with a labelled data set generated using state-of-the-art rule based algorithm. This pre-trained model will be used as the base model and is trained with reinforcement learning, which aims to maximize quality, minimize buffering and maintain smooth playback. Training can happen on Personal Computer (PC) based server or edge server setup as well as On-Device, which can even be beneficial in providing user personalization based on network throughput collected on the device. It has been shown that this method will give users a superior video streaming experience, and achieve performance improvement of around 30% on QoE over the existing commercial solutions.