Probabilistic Subgoal Representations for Hierarchical Reinforcement learning