Predictability is not everything: Using invariance to find robust models

Assume that we observe data from a response Y and a set of covariates X under different experimental conditions (or environments). Rather than focusing on the model that is most predictive, it has been suggested to take into account the invariance of a model. This can help us to infer causal structure (Which covariates are causes of Y?) and find models that generalize better (How well does the model perform on an unseen environment?). We discuss an application of this general principle on policy generalization.