Deconfounding Reinforcement Learning in Observational Settings
This addresses the challenge of applying RL in real-world observational settings where confounders can bias learning, representing an incremental but novel extension to existing RL methods.
The paper tackles the problem of learning reinforcement learning policies from observational data with unobserved confounders, proposing a deconfounding variant of Actor-Critic methods and showing it outperforms traditional RL in confounded environments.
We propose a general formulation for addressing reinforcement learning (RL) problems in settings with observational data. That is, we consider the problem of learning good policies solely from historical data in which unobserved factors (confounders) affect both observed actions and rewards. Our formulation allows us to extend a representative RL algorithm, the Actor-Critic method, to its deconfounding variant, with the methodology for this extension being easily applied to other RL algorithms. In addition to this, we develop a new benchmark for evaluating deconfounding RL algorithms by modifying the OpenAI Gym environments and the MNIST dataset. Using this benchmark, we demonstrate that the proposed algorithms are superior to traditional RL methods in confounded environments with observational data. To the best of our knowledge, this is the first time that confounders are taken into consideration for addressing full RL problems with observational data. Code is available at https://github.com/CausalRL/DRL.