B0479
Title: Effect-invariant mechanisms for policy generalization
Authors: Sorawit Saengkyongam - ETH Zürich (Switzerland) [presenting]
Abstract: Policy learning is an important component of many real-world learning systems. A major challenge in policy learning is how to adapt efficiently to unseen environments or tasks. Recently, it has been suggested to exploit invariant conditional distributions to learn models that generalize better to unseen environments. However, assuming invariance of entire conditional distributions (which is called full invariance) may be too strong of an assumption in practice. A relaxation of full invariance called effect-invariance (e-invariance for short) is introduced and proven that it is sufficient, under suitable assumptions, for zero-shot policy generalization. An extension is also discussed that exploits e-invariance when having a small sample from the test environment, enabling few-shot policy generalization. The work does not assume an underlying causal graph or that the data are generated by a structural causal model; instead, testing procedures are developed to test e-invariance directly from data. Empirical results are presented using simulated data and a mobile health intervention dataset to demonstrate the effectiveness of the approach.
