Robust Reinforcement Learning for Dynamic Treatment Regimes under Model and Measurement Uncertainty
Robust contextual RL learns adaptive prostate-cancer treatment policies that stay reliable under measurement and model error.
The paper presents a robust reinforcement-learning method for adaptive treatment policies that generalizes across patients under biomarker measurement error and disease-dynamics shifts, using robust contextual Markov decision processes with the patient as hidden context. In a simulator calibrated to clinical data for metastatic prostate cancer, the learned policies substantially outperform an existing adaptive therapy method and Pareto-dominate it across scenarios. Robustness costs almost no performance when the biomarker is accurate and disease evolution matches expectations.
- Policies are learned as robust contextual MDPs.
- Hidden patient context covers generalization and robustness.
- Tested on simulated metastatic prostate cancer adaptive therapy.
- Robust policies Pareto-dominate the existing adaptive method.
- Little performance loss when biomarkers and dynamics match expectations.
From the original article
1 min read at arxiv.orgWe develop a robust reinforcement learning (RL) framework for learning adaptive treatment policies that generalize across patients and remain reliable under errors in the measured biomarker and shifts in the disease dynamics. By modeling the patient as a hidden context, we address generalization and robustness in a unified manner through robust contextual Markov decision processes.
By Chengze Xie · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.