Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Li, Jingyi, Wu, Peng, Shi, Chengchun
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.25114
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913160291680256
author Li, Jingyi
Wu, Peng
Shi, Chengchun
author_facet Li, Jingyi
Wu, Peng
Shi, Chengchun
contents Reinforcement learning algorithms are generally designed to maximize the expected return across a population. However, a policy that is optimal on average may be suboptimal for certain individuals, leading to potential safety concerns. To address this, we first formalize the notion of individual harm from a counterfactual perspective and define harm as the event in which a chosen action results in a strictly worse outcome than a baseline alternative. We then propose a general two-stage procedure for learning policies that maximize the expected return while accounting for individual harm. We further establish the finite-sample properties of the learned policy, derive an upper bound on its sub-optimality gap, and show that the harm rate remains well-controlled. Numerical experiments on both simulated and real-world datasets demonstrate the effectiveness of the proposed approach.
format Preprint
id arxiv_https___arxiv_org_abs_2605_25114
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Counterfactually Safe Reinforcement Learning
Li, Jingyi
Wu, Peng
Shi, Chengchun
Machine Learning
Reinforcement learning algorithms are generally designed to maximize the expected return across a population. However, a policy that is optimal on average may be suboptimal for certain individuals, leading to potential safety concerns. To address this, we first formalize the notion of individual harm from a counterfactual perspective and define harm as the event in which a chosen action results in a strictly worse outcome than a baseline alternative. We then propose a general two-stage procedure for learning policies that maximize the expected return while accounting for individual harm. We further establish the finite-sample properties of the learned policy, derive an upper bound on its sub-optimality gap, and show that the harm rate remains well-controlled. Numerical experiments on both simulated and real-world datasets demonstrate the effectiveness of the proposed approach.
title Counterfactually Safe Reinforcement Learning
topic Machine Learning
url https://arxiv.org/abs/2605.25114