Towards Optimal Adversarial Robust Reinforcement Learning with Infinity Measurement Error

Haoran Li [email protected], Zicheng Zhang [email protected], Wang Luo [email protected], Congying Han [email protected], Jiayu Lv [email protected], Tiande Guo [email protected], Yudong Hu [email protected] · 2025

Read Towards Optimal Adversarial Robust Reinforcement Learning with Infinity Measurement Error by Haoran Li [email protected], Zicheng Zhang [email protected], Wang Luo [email protected], Congying Han [email protected], Jiayu Lv [email protected], Tiande Guo [email protected], Yudong Hu [email protected] (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-2502-16734.txt.

Introduction

Haoran Li and six co-authors study how to keep learning agents safe under attack in a 2025 paper. Deep reinforcement learning agents can be fooled when an attacker distorts what they observe about their state. Past work even suggested that a best possible robust policy might not always exist, which blocks strict safety demands.

The authors define a new model, the ISA-MDP, in which attacks cannot change the true nature of a state. Inside that model, they prove an optimal robust policy does exist and matches the classic Bellman policy. They show a certain strong error measure is needed to reach it, then build a learning framework around that measure. Tests confirm better performance.

Catalog record

CREATORHaoran Li [email protected], Zicheng Zhang [email protected], Wang Luo [email protected], Congying Han [email protected], Jiayu Lv [email protected], Tiande Guo [email protected], Yudong Hu [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE243 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-16734

Browse the full library · About the Tower of Babel