Reinforcement Learning based HTTP Adaptive Streaming Client

At present, HTTP Adaptive Streaming (HAS) is developing into a key technology for video delivery over the Internet. In this delivery strategy, the client proactively requests a quality version of chunked video segments based on its perceived network bandwidth and other relevant factors. However, it is often observed that current HAS client heuristics lack flexibility in vast variable network environments. In this paper, a reinforcement-learning based HAS client is proposed to progressively learn the optimal request strategy by continuously maximizing a pre-defined reward function and to take over the decision process of heuristics. We study several components of the reinforcement learning, including selection of environment variables and definition of Quality of Experience (QoE)-related reward function, and implement it into the HAS client. Experimental results show that comparing with a conventional HAS system, the employed technique is more robust and flexible in the versatile network conditions, resulting in an enhanced QoE at the HAS client.