A Reinforcement Learning Solution for Adaptive Video Streaming

Adaptive streaming is a promising technique for video streaming service to cope with the variability of network load or quality degradation of mobile users' connections. In this paper we propose a stream-control mechanism based on a Reinforcement Learning (RL) paradigm that will gracefully degrade the flow quality experienced by the end-user depending on the network state. Using layered coded videos the client should find the most appropriate quality level for its stream. The proposed mechanism, implemented at the client side, can flexible deal with network dynamics. The adaptive streaming problem could naturally be modeled as a Partial Observable Markov Decision Process because the client has partial information about the network state based on the received throughput, but it cannot be applied on-line during streaming. We propose here a MDP modeling the adaptive streaming problem that could be solved on-line by Q-Learning algorithm. The both models have identical solutions proving the validity of the proposed MDP model.