Hierarchically Optimal Average Reward Reinforcement Learning

Two notions of optimality have been explored in previous work on hierarchical reinforcement learning (HRL): hierarchical optimality, or the optimal policy in the space defined by a task hierarchy, and a weaker local model called recursive optimality. In this paper, we introduce two new average-reward HRL algorithms for finding hierarchically optimal policies. We compare them to our previously reported algorithms for computing recursively optimal policies, using a grid-world taxi problem and a more real-world AGV scheduling problem. The new algorithms are based on a three-part value function decomposition proposed recently by Andre and Russell, which generalizes Dietterich’s MAXQ value function decomposition. A key difference between the algorithms proposed in this paper and our previous work is that there is only a single global gain (average reward), instead of a gain for each subtask. Our results show the new average-reward algorithms have better performance than both the previous recursively optimal counterparts, as well as the corresponding discounted hierarchical optimal algorithms.