Reinforcement learning for dynamic multimedia adaptation

In this paper we present an integration of several user and resource-related factors for the design of dynamic adaptation techniques. Our first contribution is an original reinforcement-learning approach to develop better adaptation agents. Integrated with the content, these agents improve gradually, by taking into account both user's behaviour and the usage context. Our second contribution is to apply this generic approach to solve an ubiquitous streaming problem. Mobile users experience large latencies while accessing streaming media. We propose to adapt the streaming by prefetching and to model this decision problem by using a Markov decision process. We discuss this formal framework and make explicit reference to its relationship with reinforcement learning. We support the benefits of our approach by presenting results from simulations and experiments.