Saved in:
Bibliographic Details
Main Authors: Chehade, Mohamad Fares El Hajj, Bedi, Amrit Singh, Zhang, Amy, Zhu, Hao
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2408.08812
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910237339942912
author Chehade, Mohamad Fares El Hajj
Bedi, Amrit Singh
Zhang, Amy
Zhu, Hao
author_facet Chehade, Mohamad Fares El Hajj
Bedi, Amrit Singh
Zhang, Amy
Zhu, Hao
contents Deployed reinforcement learning agents often face safety requirements that are specified only after training, such as new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation, where a fixed library of risk-neutral source policies is reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-Time Risk Adaptation via Mixture of Agents), a source-scored composition rule that evaluates each source policy under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence from a reference behavior, and local volatility risks specified at test time. We explicitly characterize TRAM as a surrogate method: it does not solve the full occupancy-control problem of the stitched policy, but admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an LLM alignment setting show that TRAM reduces deployment risk while preserving reward, without requiring any parameter updates at test time.
format Preprint
id arxiv_https___arxiv_org_abs_2408_08812
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle TRAM: Test-Time Risk Adaptation with Mixture of Agents
Chehade, Mohamad Fares El Hajj
Bedi, Amrit Singh
Zhang, Amy
Zhu, Hao
Machine Learning
Deployed reinforcement learning agents often face safety requirements that are specified only after training, such as new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation, where a fixed library of risk-neutral source policies is reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-Time Risk Adaptation via Mixture of Agents), a source-scored composition rule that evaluates each source policy under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence from a reference behavior, and local volatility risks specified at test time. We explicitly characterize TRAM as a surrogate method: it does not solve the full occupancy-control problem of the stitched policy, but admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an LLM alignment setting show that TRAM reduces deployment risk while preserving reward, without requiring any parameter updates at test time.
title TRAM: Test-Time Risk Adaptation with Mixture of Agents
topic Machine Learning
url https://arxiv.org/abs/2408.08812