ERUDITE: A deep neural network for optimal tuning of adaptive video streaming controllers

Abstract

Adaptive video streaming systems are expected to provide the best user experience to improve service engagement. To this purpose, the video player implements a controller to dynamically choose the most suitable video representation to be downloaded. It is well-known that finding one tuning of the controller's parameters which performs satisfactorily in a wide range of scenarios is very challenging. This paper studies the problem of providing users with (near) optimal Quality of Experience (QoE) for Dynamic Adaptive Streaming over HTTP (DASH) systems. We present ERUDITE, a closed-loop system to optimally tune - at run-time - the adaptive streaming controller's parameters to adapt to changing scenario's parameters. The proposed system is based on a Deep Neural Network (DNN) which continuously provides the streaming controller with estimates of optimal parameters based on measured metrics such as bandwidth samples and overall obtained QoE. The DNN is trained using a dataset that we have built by finding, for thousands of scenarios, the optimal adaptive streaming controller's parameters using a Bayesian optimization algorithm. Results, gathered considering a large number of diverse scenarios, show that ERUDITE is able to provide near optimal performances by reducing impairments due to rebuffering and video level switching.