Characterization of Efficient Influence Function for Off-Policy Evaluation Under Optimal Policies Journal: arxiv

Haoyu Wei · 2025

Read Characterization of Efficient Influence Function for Off-Policy Evaluation Under Optimal Policies Journal: arxiv by Haoyu Wei (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-13809.txt.

Description

Haoyu Wei studies how to place trustworthy error bars on the value of a best policy learned from data gathered under a different policy. This off-policy problem turns delicate when the best policy is not unique, or when one action wins by a hair, because standard inference tools can break down. Wei examines the problem in Markov decision processes and first works in a simplified sampling setup.

In that setup, the efficient influence function exists only under conditions the paper pins down, and trouble arises when rival best policies pull in different first-order directions. For real trajectory data, Wei derives the best possible accuracy bound and gives an estimator that reaches it when the optimum is unique. A sequential method, NSAVE, trades some sharpness for stability when it is not. Tests on a phone-based health trial round out the work.

Catalog record

CREATORHaoyu Wei
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-13809

Browse the full library · About the Tower of Babel