Structured Reinforcement Learning for Media Streaming at the Wireless Edge

Abstract

When streaming media in an environment with less than ideal resources, simple algorithms do not perform well by not prioritizing specific clients well enough. Therefore, intelligent network sharing must occur to ensure maximum average quality of experience (QoE). We propose formulating this problem of designing intelligent policies as a Constrained Markov decision process. By relaxing the problem by using a Lagrangian formulation we reduce it down into single-client problems. Reinforcement learning can then be applied to control resource allocation to each client by implementing a threshold with respect to each client’s buffer length. By developing a simulation environment and evaluating algorithms on a real world system capable of handling a variable number of clients, we demonstrate improvement in overall average QoE.