Reproduce “Stable Policy Learning” on OpenResearch

About this paper

Harvey Barnhard, Giacomo Opocher, Rahul Singh

Stable Policy Learning extends policy evaluation beyond expected welfare by treating variation across possible experimental samples as a distinct form of decision risk. It formalizes welfare stability, proves that single-observation sensitivity controls welfare variance and poor-policy tail risk, and shows that policy-vote bagging can reduce this instability while preserving favorable welfare rates.

Key insights

• A learning algorithm’s sensitivity to one training observation provides a general measure of how much its policy welfare can fluctuate across experimental samples.
• Complete bagging limits the welfare effect of replacing one observation to at most an m/n fraction of the underlying policy effect, while concavity implies that averaging can improve expected utility by reducing welfare dispersion.
• The optimal subsample size depends on estimation speed and risk aversion: full-sample learning is rate-optimal in parametric settings, whereas smaller bags can be preferable for slower nonparametric learners or finite samples with high sampling-risk concerns.

Summary by alphaXiv.