Reinforcement learning algorithms for DASH video streaming

Dynamic Adaptive Streaming over HTTP (DASH) is a video streaming standard developed in 2011; the servers have several copies of every video at different bitrates, leaving the clients complete freedom to choose the bitrate of each segment and adapt to the available bandwidth. The research on client-side strategies to optimize user Quality of Experience (QoE) is ongoing; one of the most promising approaches is based on Reinforcement Learning (RL). RL controllers do not have a pre-set model of the situation, but learn the optimal policy by trial and error. This thesis presents two RL-based algorithms: Offline and Online. The Offline algorithm relies on a training phase to gain information about its environment and refine its policy, while the Online algorithm has a slimmer model and focuses on learning as quickly as possible, allowing immediate deployment and short convergence times without a training phase.