Non-Stationary Markov Decision Processes a Worst-Case Approach using Model-Based Reinforcement Learning