- Keywords: reinforcement learning, causality, confounding
- Abstract: Learning efficiently a causal model of the environment is a key challenge of model-based RL agents operating in POMDPs. We consider here a scenario where the learning agent has the ability to collect online experiences through direct interactions with the environment (interventional data), but also has access to a large collection of offline experiences, obtained by observing another agent interacting with the environment (observational data). A key ingredient, which makes this situation non-trivial, is that we allow the observed agent to act based on privileged information, hidden from the learning agent. We then ask the following questions: can the online and offline experiences be safely combined for learning a causal transition model ? And can we expect the offline experiences to improve the agent's performances ? To answer these, first we bridge the fields of reinforcement learning and causality, by importing ideas from the well-established causal framework of do-calculus, and expressing model-based reinforcement learning as a causal inference problem. Second, we propose a general yet simple methodology for safely leveraging offline data during learning. In a nutshell, our method relies on learning a latent-based causal transition model that explains both the interventional and observational regimes, and then inferring the standard POMDP transition model via deconfounding using the recovered latent variable. We prove our method is correct and efficient in the sense that it attains better generalization guarantees due to the offline data (in the asymptotic case), and we assess its effectiveness empirically on a series of synthetic toy problems.
- One-sentence Summary: We formulate reinforcement learning as a causal problem, and present a provably efficient method to address a scenario where offline data originating from a confounded policy is available for training.
- Supplementary Material: zip