Semi-Supervised Off Policy Reinforcement Learning
This work is significant for medical researchers and practitioners who develop sequential treatment strategies, as it offers a method to leverage readily available unlabeled data to overcome the limitations of small, expensively annotated datasets.
The paper addresses the challenge of small labeled datasets in medical reinforcement learning by proposing a semi-supervised learning (SSL) approach. This method combines a small labeled dataset with true outcomes and a large unlabeled dataset with outcome surrogates to improve the efficiency of Q-learning and doubly robust off-policy value estimation. The proposed method is theoretically proven to be at least as efficient as supervised approaches and robust to imputation model mis-specification.
Reinforcement learning (RL) has shown great success in estimating sequential treatment strategies which take into account patient heterogeneity. However, health-outcome information, which is used as the reward for reinforcement learning methods, is often not well coded but rather embedded in clinical notes. Extracting precise outcome information is a resource intensive task, so most of the available well-annotated cohorts are small. To address this issue, we propose a semi-supervised learning (SSL) approach that efficiently leverages a small sized labeled data with true outcome observed, and a large unlabeled data with outcome surrogates. In particular, we propose a semi-supervised, efficient approach to Q-learning and doubly robust off policy value estimation. Generalizing SSL to sequential treatment regimes brings interesting challenges: 1) Feature distribution for Q-learning is unknown as it includes previous outcomes. 2) The surrogate variables we leverage in the modified SSL framework are predictive of the outcome but not informative to the optimal policy or value function. We provide theoretical results for our Q-function and value function estimators to understand to what degree efficiency can be gained from SSL. Our method is at least as efficient as the supervised approach, and moreover safe as it robust to mis-specification of the imputation models.