An inference based end-to-end video streaming application running over openflow networks

As a different approach from the conventional network architecture, Software Defined Networks (SDN), which decouples the control and data plane, is one of the topics that have gain popularity nowadays. In this thesis, Reinforcement Learning based approach for adaptive video streaming systems running over SDN is proposed. In the proposed system, a learning model is developed in order to determine the optimal time to re-route traffic flows and to change the bitrate of the video. The learning model aims to minimize packet loss rate, quality changes and controller cost while dynamically adapts the flow routes and video quality. the performance of this learning based approach is tested by comparing it to the shortest path routing and a greedy approach. The results show that the proposed system significantly outperforms these approaches in terms of Quality of Experience (QoE) and network cost under different network scenarios.