RAAM: The Benefits of Robustness in Approximating Aggregated MDPs in Reinforcement Learning

Marek Petrik, Dharmashankar Subramanian

Neural Information Processing Systems 

Neural Information Processing Systems http://nips.cc/