COCKTAIL: Video streaming QoE optimization with chunk replacement and guided learning

Adaptive bitrate (ABR) algorithms, the de facto standard for video streaming, aim to provide the best user quality of experience (QoE) under fluctuating operating environments. Prior ABR algorithms addressed the QoE maximization problem with a plethora of approximate optimization techniques including model predictive control (MPC), Lyapunov optimization, and deep reinforcement learning (DRL). Even though these algorithms provide adequate performances, most of them primarily focus on choosing optimal bitrates while prohibiting duplicated downloads. We point out that allowing duplicated chunk downloads for replacement can potentially enhance the QoE if they are carefully administered. Moreover, combining ABR algorithms with other techniques may compromise optimization results as each algorithm’s optimization problem does not account for the others. In this paper, we first formulate a novel optimization problem with an expanded decision dimension that encompasses the chunk replacement as well as the bitrate selection in the action space. We then propose COCKTAIL, a DRL-based ABR algorithm that discovers efficient solutions to the new optimization problem by using several learning techniques. Experiments on real-world network traces show that COCKTAIL outperforms state-of-the-art baselines with improvements in average QoE up to 16.4%.