Key Papers in Safe RL
What follows is a list of papers in Safe RL that are worth reading. This is far from comprehensive, but should provide a useful starting point for someone looking to do research in the field.
Table of Contents
1. General Review
Unsolved Problems in ML Safety, Hendrycks et al, 2022.
Concrete Problems in AI Safety, Amodei et al, 2016.
A Comprehensive Survey on Safe Reinforcement Learning, García et al, 2015.
2. Model-free RL
Constrained Policy Optimization, Achiam et al, 2017.
Lyapunov-based Safe Policy Optimization for Continuous Control, Chow et al, 2019
Batch Policy Learning under Constraints, Le et al, 2019
Reward Constrained Policy Optimization, Tessler et all, 2019
Responsive Safety in Reinforcement Learning by PID Lagrangian Methods, Stooke et al, 2020
Projection-based Constrained Policy Optimization, Yang et al, 2020.
3. Model-based RL
Safe Model-based Reinforcement Learning with Stability Guarantees, Berkenkamp et al, 2017
Constrained model predictive control: Stability and optimality, Mayne et al, 2000
Constrained Policy Optimization via Bayesian World Models, As et al, 2022
4. Transfer Learning
Learning to be Safe: Deep RL with a Safety Critic, Srinivasan et al, 2020
5. Ensemble Learning
Gerneralzieing from a Few Environments in Safety-critical Reinforcement Learning, Kenton et al, 2019
Leave No Trace: Learning to Reset for Safe and Autonomous Reinforcement Learning, Eysenbach et al, 2018
6. Human in The Loop
Trial without Error: Towards Safe Reinforcement Learning via Human Intervention, Saunders et al, 2017
7. Curriculum Learning
Safe Reinforcement Learning via Curriculum Induction, Turchetta et al, 2020
8. Risk-sensitive RL
Risk-Sensitive Reinforcement Learning Applied to Control under Constraints, Geibel et al, 2005
Risk-Aware Transfer in Reinforcement Learning using Successor Features, Gimelfarb et al, 2021
Exponential Bellman Equation and Improved Regret Bounds for Risk-Sensitive Reinforcement Learning, Fei et al, 2021
9. Formal Methods
Verifiably Safe Exploration for End-to-End Reinforcement Learning, Hunt et al, 2021. See also Guarantee Safety in Training and Testing for related work.