Core Concepts
Safe and robust reinforcement learning involves addressing challenges related to safety, robustness, and ethical considerations in RL applications.
Abstract
The paper explores the dimensions of safe and robust RL, covering algorithmic approaches, environmental factors, and human involvement.
It categorizes research works into different algorithmic approaches enhancing safety and robustness.
Techniques like uncertainty estimation, optimization methodologies, exploration-exploitation trade-offs, and adversarial training are discussed.
The paper introduces a practical checklist for designing safe and robust RL systems, covering aspects of algorithm design, training environment considerations, and ethical guidelines.
Stats
"Robust RL aims to learn a robust optimal policy that accounts for model uncertainty of the transition probability to systematically mitigate the sensitivity of the optimal policy in perturbed environments" - [4]
"Safe RL is the process of learning policies that maximize the expectation of the return in problems in which it is important to ensure reasonable system performance and/or respect safety constraints during the learning and/or deployment processes. Also, the system must have the right objectives (the reward function aligned with the objective of the task) and a mechanism for humans to intervene." - [11]
Quotes
"Safe RL has a mechanism for a human to interfere the agent effectively." - Dylan Hadfield-Menell et al.
"Robust RL aims to learn a robust optimal policy that accounts for model uncertainty of the transition probability to systematically mitigate the sensitivity of the optimal policy in perturbed environments." - Survey Paper