Identifying Latent State Transition Processes for Individualized Reinforcement Learning

Yuewen Sun, Biwei Huang, Yu Yao, Donghuo Zeng, Xinshuai Dong, Roberto Legaspi, Kazushi Ikeda, Peter Spirtes, Kun Zhang

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: individualized reinforcement learning, latent state transition, identifiability

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In recent years, reinforcement learning (RL) has been increasingly applied to systems that interact with individuals in various domains, such as healthcare, education, and e-commerce. When an RL agent interacts with individuals, individual-specific factors, ranging from personal preferences to physiological nuances, may causally influence state transitions, such as health conditions, learning progress, or user selections. Consequently, different individuals may exhibit different state transition processes. Understanding these individualized state-transition processes is crucial for making individualized policies. In practice, however, identifying these state-transition processes is challenging, especially since individual-specific factors often remain latent. In this paper, we present a practical method that effectively learns these processes from observed state-action trajectories, backed by theoretical guarantees. To our knowledge, this is the first work to provide a theoretical guarantee for identifying the state-transition processes involving latent individual-specific factors. Our experiments on synthetic and real-world datasets demonstrate that our method can effectively identify the latent state-transition processes and help learn individualized RL policies.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8507