Research map / Reinforcement learning / Policy gradient and actor-critic
Enforcing Hard Constraints with Soft Barriers: Safe Reinforcement Learning in Unknown Stochastic Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CMDP expected-cost constraints with generative-model-based soft barrier functions for hard safety chance constraints | safe RL in unknown stochastic environments · 2D SDE, Cartpole, Rocket, UAV | empirical safe rate s_e: up to ~10 points vs PPO-L (Cartpole 100% vs 89.3%) and vs FOCOPS* (Cartpole 100% vs 84.2%; UAV 100% vs 91%) | Helped |
| Use PPO-L with per-state expected cost <= 0 instead of cumulative-cost constraint | PPO-L baseline · Cartpole, Rocket | empirical safe rate s_e | No clear effect |
| Adopt a safety-oriented FOCOPS* variant (tighter cost threshold D') relative to standard FOCOPS | FOCOPS baseline · Rocket | empirical safe rate s_e: Rocket 100% vs PPO-L 96.4%, but UAV 91% and Cartpole 84.2% | Mixed |
| Soft barrier approach versus CMDP baselines on total reward | safe RL with soft barrier · 2D SDE, Cartpole, Rocket, UAV | total reward J: Ours higher than PPO-L on Rocket (-143.2 vs -151.4) and UAV (-847.1 vs -895.5), lower on 2D and Cartpole vs FOCOPS*/PPO- | Mixed |
| Generative-model-based soft barrier with practical lower bound 1-eta reported as a safety guarantee | safe RL with soft barrier · 2D SDE, Cartpole, Rocket, UAV | practical safety lower bound 1-eta vs empirical safe rate: 1-eta (86.6% to 97.6%) is below s_e in all examples, consistent with being a lower bound | No clear effect |
Closest papers by meaning
- Safe Reinforcement Learning with Instantaneous Constraints: The Role of Aggressive Exploration (AAAI 2024 · similarity 0.92)
- Probabilistic Shielding for Safe Reinforcement Learning (AAAI 2025 · similarity 0.91)
- Sampling-Based Safe Reinforcement Learning (NeurIPS 2026 · similarity 0.91)
- Near-Optimal Sample Complexity for Online Constrained MDPs (NeurIPS 2025 · similarity 0.90)
- Long-Term Safe Reinforcement Learning with Binary Feedback (AAAI 2024 · similarity 0.90)
- Absolute State-wise Constrained Policy Optimization: High-Probability State-wise Constraints Satisfaction (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.