Constraint Reward

Constraint reward methods in reinforcement learning aim to optimize agent behavior while strictly adhering to safety constraints, balancing performance goals with risk mitigation. Current research focuses on developing algorithms that effectively integrate constraint rewards, including dual-expert approaches combining performance and safety objectives, and methods using safety editor policies to modify potentially unsafe actions. These techniques are proving valuable in diverse applications, such as robotics and adversarial attacks on text classifiers, by significantly improving safety and reliability while maintaining high performance. The resulting advancements are driving progress in safe and robust autonomous systems.

Papers

July 2, 2024

Safe CoR: A Dual-Expert Approach to Integrating Imitation Learning and Safe Reinforcement Learning Using Constraint Rewards
Hyeokjin Kwon, Gunmin Lee, Junseo Lee, Songhwai Oh
Imitation Learning Safe Reinforcement Learning Safety Constraint Dual Approach Constraint Reward

May 20, 2024

A Constraint-Enforcing Reward for Adversarial Attacks on Text Classifiers
Tom Roth, Inigo Jauregi Unanue, Alsharif Abuadbba, Massimo Piccardi
Adversarial Attack Adversarial Example Text Classifier Paraphrase Detection Constraint Reward

January 28, 2022

Towards Safe Reinforcement Learning with a Safety Editor Policy
Haonan Yu, Wei Xu, Haichao Zhang
Safe Reinforcement Learning Safety Filter Constraint Reward

Constraint Reward

Papers

Safe CoR: A Dual-Expert Approach to Integrating Imitation Learning and Safe Reinforcement Learning Using Constraint Rewards

A Constraint-Enforcing Reward for Adversarial Attacks on Text Classifiers

Towards Safe Reinforcement Learning with a Safety Editor Policy