Abstract
Causal discovery is the challenging task of inferring causal structure fromdata. Motivated by Pearl's Causal Hierarchy (PCH), which tells us that passiveobservations alone are not enough to distinguish correlation from causation,there has been a recent push to incorporate interventions into machine learningresearch. Reinforcement learning provides a convenient framework for such anactive approach to learning. This paper presents CORE, a deep reinforcementlearning-based approach for causal discovery and intervention planning. CORElearns to sequentially reconstruct causal graphs from data while learning toperform informative interventions. Our results demonstrate that COREgeneralizes to unseen graphs and efficiently uncovers causal structures.Furthermore, CORE scales to larger graphs with up to 10 variables andoutperforms existing approaches in structure estimation accuracy and sampleefficiency. All relevant code and supplementary material can be found athttps://github.com/sa-and/CORE