Amortized Active Causal Induction with Deep Reinforcement Learning

Our design policy successfully achieves amortized intervention design on the distribution of the training environment while also generalizing well to distribution shifts in test-time design environments.