HASBRAIN: A Machine Learning-based Adaptation Algorithm for HTTP Adaptive Streaming

This thesis presents a new Quality of Experience (QoE) aware adaptation algorithm for HTTP Adaptive Streaming (HAS) based on supervised Machine Learning (ML) techniques named " H TTP A daptive S treaming with B it- R ate change aware A rtificial IN telligence" or just HASBRAIN . With use of the optimal adaptation path, the optimal video quality picking strategy, a ML algorithm is trained and learns to behave in similar optimal fashion. The first contribution of this work is the modification of an existing optimization formulation with the goal of determining the Pareto frontier for QoE aware adaptation algorithms. This formulation allows to evaluate with an aggressiveness switching parameter α the trade-off between average quality and quality switching in the video playout. The results of this evaluation lead to a modified version of a two-step optimization problem formulation with an allowable maximum quality level degradation parameter e while minimizing the number of quality switches. Given a video and a data-rate histogram, the optimization outputs the optimal adaptation path. The optimal adaptation path is the optimal quality level picking strategy at each decision instance in time. The optimal decisions are used as input for training a ML algorithm. These supervised trained ML algorithms consist of an Artificial Neural Network (ANN), a Support Vector Machine (SVM) and a k-Nearest-Neighbors approach. The best performing learning technique, in terms of learning accuracy, is then used as an adaptation logic in a HAS scenario simulated through a Discrete Event Simulation (DES). The performance is evaluated through user-centric evaluation metrics gathered from the DES. These metrics are the average playout quality, the quality switching frequency, the stalling-frequency and -ratio and the buffer level throughout the video playout. The HASBRAIN algorithm's performance is furthermore compared with the performance of well-known threshold-based adaptation algorithms TRDA and KLUDCP within the same evaluation scenario. The ML's adaptation performance behavior allows it to reach a similar average playout quality than the other algorithms while reducing the number of quality switches. But due to wrong adaptation decisions the ML tool is in some scenarios not able to prevent stalling to occur. This work is concluded by the pros and cons of this algorithm compared to the threshold-based algorithms. Furthermore, an outlook for future work is given to further improve the performance through adjustments in the ML algorithms learning.