Optimizing for the Future in Non-Stationary MDPs