Representation Balancing MDPs for Off-Policy Policy Evaluation
This work addresses a specific challenge in reinforcement learning for domains like healthcare, though it appears incremental by building on prior causal reasoning approaches.
The paper tackles the problem of off-policy policy evaluation in RL by proposing a method to accurately estimate both individual and average policy values, resulting in substantially lower mean squared error in synthetic benchmarks and a HIV treatment simulation.
We study the problem of off-policy policy evaluation (OPPE) in RL. In contrast to prior work, we consider how to estimate both the individual policy value and average policy value accurately. We draw inspiration from recent work in causal reasoning, and propose a new finite sample generalization error bound for value estimates from MDP models. Using this upper bound as an objective, we develop a learning algorithm of an MDP model with a balanced representation, and show that our approach can yield substantially lower MSE in common synthetic benchmarks and a HIV treatment simulation domain.