Actor-Critic Algorithm for Dynamic Expectile and CVaR

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Luo, Yudong, Delage, Erick
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915993680347136
author Luo, Yudong
Delage, Erick
author_facet Luo, Yudong
Delage, Erick
contents Optimizing dynamic risk with stochastic policies is challenging in both policy updates and value learning. The former typically requires transition perturbation, while the latter may rely on model-based approaches. To address these challenges, we propose a surrogate policy gradient without transition perturbation under softmax policy parameterization. We further develop model-free value learning methods for dynamic expectile and conditional value-at-risk by leveraging elicitability. Finally, inspired by Expected SARSA and Expected Policy Gradient, a model-free off-policy actor-critic algorithm is constructed. Empirical results in domains with verifiable risk-averse behavior show that our algorithm can learn risk-averse policy and consistently outperforms other existing methods.
format Preprint
id arxiv_https___arxiv_org_abs_2605_07857
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Actor-Critic Algorithm for Dynamic Expectile and CVaR
Luo, Yudong
Delage, Erick
Machine Learning
Optimizing dynamic risk with stochastic policies is challenging in both policy updates and value learning. The former typically requires transition perturbation, while the latter may rely on model-based approaches. To address these challenges, we propose a surrogate policy gradient without transition perturbation under softmax policy parameterization. We further develop model-free value learning methods for dynamic expectile and conditional value-at-risk by leveraging elicitability. Finally, inspired by Expected SARSA and Expected Policy Gradient, a model-free off-policy actor-critic algorithm is constructed. Empirical results in domains with verifiable risk-averse behavior show that our algorithm can learn risk-averse policy and consistently outperforms other existing methods.
title Actor-Critic Algorithm for Dynamic Expectile and CVaR
topic Machine Learning
url https://arxiv.org/abs/2605.07857