Research map / Reinforcement learning / Policy gradient and actor-critic
Towards Safe Policy Improvement for Non-Stationary MDPs
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 |
|---|---|---|---|
| account for non-stationarity and use forecasted-performance safety tests instead of assuming stationarity | SPIN versus Baseline · Non-stationary RecoSys and diabetes treatment | safety violation rate | Helped |
| account for non-stationarity rather than ignore it | SPIN versus Baseline · Non-stationary RecoSys | performance gain over the safe policy | Helped |
| account for non-stationarity rather than ignore it | SPIN versus Baseline · Non-stationary diabetes treatment | safety violation rate | Helped |
| use a 75%--25% train-test split (ablation) | SPIN · RecoSys, speeds 0–3 | performance gain over the safe policy: .56, .22, .17, .14 | Mixed |
| use a 25%--75% train-test split (ablation) | SPIN · RecoSys, speeds 0–3 | performance gain over the safe policy: .48, .29, .21, .19 | Mixed |
| optimize mean performance instead of the lower confidence bound (ablation) | SPIN · RecoSys, speeds 0–3 | safety violation percentage: 0.2, 4.9, 6.3, 7.1 | No clear effect |
| remove the non-stationary safety check (ablation) | SPIN · RecoSys, speeds 0–3 | safety violation percentage: 0.0, 4.7, 6.4, 6.6 -> 9.4, 37.6, 40.2, 38.6 | Hurt |
| assume stationarity while retaining a safety procedure (ablation) | SPIN · RecoSys, speeds 1–3 | safety violation percentage: 4.7, 6.4, 6.6 -> 19.8, 15.3, 11.9 | Hurt |
Closest papers by meaning
- Non-Stationary Markov Decision Processes, a Worst-Case Approach using Model-Based Reinforcement Learning (NeurIPS 2019 · similarity 0.88)
- Long-Term Safe Reinforcement Learning with Binary Feedback (AAAI 2024 · similarity 0.87)
- A Provable Approach for End-to-End Safe Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Pessimism Meets Risk: Risk-Sensitive Offline Reinforcement Learning (ICML 2024 · similarity 0.86)
- Probabilistic Shielding for Safe Reinforcement Learning (AAAI 2025 · similarity 0.86)
- Near-Optimal Sample Complexity for Online Constrained MDPs (NeurIPS 2025 · similarity 0.86)
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.