Learning by Intervention in Simple Causal Domains

We propose a framework for learning dependencies between variables in an environment with causal relations. We assume that the environment is fully observable and that the underlying causal structure is of a simple nature. We adapt the frameworks of the (epistemic) causal models from [4, 17], and propose a model inspired by action learning [6, 7]. We present two learning methods, using formal and algorithmic approaches. Our learning agents infer dependencies (atomic formulas of Dependence Logic) from observations of interventions on valuations (propositional states), and by doing so efficiently, they obtain insights into how to manipulate their surroundings to achieve goals.