Hierarchical Average Reward Reinforcement Learning