Most reinforcement learning methods are based upon the key assumption that the transition dynamics and reward functions are fixed, that is, the underlying Markov decision process (MDP) is stationary. However, in many practical real-world applications, this assumption is often violated.