Learning Symbolic Representations for Reinforcement Learning of Non-Markovian Behavior