Here's our paper on automated discovery of high level skills (options) in reinforcement learning through MDP abstraction with extension to large state space problems through Deep Unsupervised Action Conditional Video Prediction Network.