Average-Reward Reinforcement Learning for Multichain MDPs: A Hierarchical Decomposition Approach
This research develops model-free reinforcement learning algorithms for average-reward multichain MDPs using a hierarchical decomposition of the state space. The framework introduces three algorithms: one for gain optimality, one for near gain optimality with better transient performance, and one targeting near bias optimality. This provides a structured way to solve complex average-reward problems without needing a model.