Average-Reward Reinforcement Learning for Multichain MDPs: A Hierarchical Decomposition Approach
We study learning optimal policies in average-reward multichain Markov decision processes (MDPs), where the optimal gain may depend on the initial state and recurrence structures vary across policies, creating challenges for reinforcement learning (RL) methods.
Key points
- We propose an asynchronous value-iteration-based RL algorithm that requires no model knowledge beyond the MDP's transition graph and leverages Bather's decomposition to hierarchically partition the state space into communicating subsystems and transient states.
- We show that the algorithm converges to the optimal gain and produces gain-optimal policies after finite time.
- Building on this base algorithm, we develop two further algorithms: one approximately solves the multichain average optimality equations to obtain near gain-optimal policies, and another targets near bias-optimality by approximating the optimal bias function and solving an induced average-reward multichain MDP using the base algorithm.
- To our knowledge, these are the first essentially model-free average-reward RL algorithms for general multichain MDPs without reductions to discounted problems.
Sources (1)
- [1]Average-Reward Reinforcement Learning for Multichain MDPs: A Hierarchical Decomposition ApproacharXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 04:13 PM
We study learning optimal policies in average-reward multichain Markov decision processes (MDPs), where the optimal gain may depend on the initial state and recurrence structures vary across policies, creating challenges for reinforcement learning (RL) methods.
We propose an asynchronous value-iteration-based RL algorithm that requires no model knowledge beyond the MDP's transition graph and leverages Bather's decomposition to hierarchically partition the state space into communicating subsystems and transient states.
Extractive summary: sentences quoted from the sources.
