Adaptive Safety in Reinforcement Learning via Advanced Lagrangian Optimization

Feihong Zhang, Tianyi Zhang, Maanping Shao

Keywords: safeRL，Constrained optimization，Safety Gym

Abstract:

1. Background
1.1. Introduction
Reinforcement Learning (RL) is a key area in machine learning, achieving success in robotics,
autonomous driving, and games Li (2023). However, real-world applications require Safe RL
to prevent harmful behaviors by integrating safety directly into learning Ray et al. (2019).
1.2. Importance of the Problem
Traditional RL focuses on maximizing rewards without considering safety during learning Achiam et al. (2017). Safe RL methods often use Lagrangian approaches to enforce
constraints but face issues with instability and slow convergence Munos et al. (2016).
1.3. Impact of the Proposed Solution
The proposal aims to develop an optimization algorithm that improves Safe RL by better
handling constraints during training. This aims to achieve faster convergence and better
safety compliance, enabling RL agents in safety-critical environments.

Submission Number: 34