Shengpu Tang, Aditya Modi, Michael Sjoding, Jenna Wiens
Standard reinforcement learning aims to find an optimal policy that identifies the best action for each state. However, in healthcare settings, many actions may be near-equivalent with respect to the reward (e.g., survival). We consider an alternative objective -- learning set-valued policies to capture near-equivalent actions that lead to similar cumulative rewards. We propose a model-free, off-policy algorithm based on temporal difference learning and a near-greedy action selection heuristic. We analyze the theoretical properties of the proposed algorithm, providing optimality guarantees and demonstrate our approach on simulated environments and a real clinical task. Empirically, the proposed algorithm exhibits reasonably good convergence properties and discovers meaningful near-equivalent actions. Our work provides theoretical, as well as practical, foundations for clinician-in-the-loop decision making, in which clinicians/patients can incorporate additional knowledge (e.g., side effects and patient preference) to distinguish among near-equivalent actions.