Empirical evaluation of MDP-based DASH player

Dynamic Adaptive Streaming over HTTP (DASH) is one of the most widely used adaptive streaming technique for watching online video content. DASH adapts to the varying network conditions by selecting the appropriate bitrate of the video stream. The bitrate adaptation is typically done by monitoring the playback buffer level and/or the network condition on client side. In this paper we empirically evaluate our JavaScript DASH player in which, Markov Decision Process (MDP) has been considered as the underlying optimization framework. This player uses Q-learning algorithm to learn the model and optimize the Quality of Service (QoS) after multiple streaming sessions. The basic JavaScript DASH player developed by DASH Industry Forum (DASHIF) is used as a benchmarking model in our evaluations. We use Google Chrome´s 3G and 4G network emulators in our experiments and show that our MDP-based DASH player significantly outperforms the DASHIF player which uses buffer control and rate adaptation techniques simultaneously. Using real-time experiments, we show that for similar picture quality we can achieve about 18x fewer deadline misses and 5x fewer quality switches over a 3G network and 32x fewer deadline misses and 1.6x fewer quality switches over 4G.