Research map / Reinforcement learning / Imitation learning
SafeDICE: Offline Safe Imitation Learning with Non-Preferred Demonstrations
Youngsoo Jang, Geon-Hyeong Kim, Jongmin Lee, Sungryull Sohn, Byoungjip Kim, Honglak Lee, Moontae Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use SafeDICE instead of BC, DWBC, PPL, DExperts on safety metrics | SafeDICE · RWRL Cartpole, Walker, Quadruped | Average Cost, Cost-Violating Ratio, CVaR 10% Cost | Improved |
| Use SafeDICE on Cartpole for worst-case cost | SafeDICE · RWRL Cartpole | CVaR 10% Cost | No clear effect |
| Use SafeDICE instead of baselines on return | SafeDICE · RWRL tasks | Normalized Return | No clear effect |
| Use SafeDICE instead of baselines | SafeDICE · Safety Gym Goal and Button | safety metrics | Improved |
| Use SafeDICE on Button task | SafeDICE · Safety Gym Button | Return | Hurt |
| Use preference-based PPL instead of BC | BC · RWRL Walker and Quadruped vs Cartpole | safety | Mixed |
| Reduce labeled non-preferred demonstrations |DN| from 50 to 5 (ablation) | SafeDICE and baselines · RWRL Walker | safety metrics | Mixed |
| Vary alpha over 0 to 0.8 instead of theoretically determined alpha (Eq. 15) (ablation) | SafeDICE · RWRL Walker | CVaR cost | Hurt |
| Remove stationary distribution constraint (ablation) | SafeDICE · RWRL Walker | cost violation, CVaR 10% Cost | Hurt |
| Remove labeled non-preferred demonstrations (ablation) | SafeDICE · RWRL Walker | cost violation, CVaR 10% Cost | Hurt |
| Vary ratio of preferred to non-preferred in unlabeled data, from (250, 1000) to (1000, 250) (ablation) | SafeDICE · RWRL Walker | cost violation | No clear effect |
| Tune eta and alpha of DWBC and DExperts over 0.01 to 0.8 (ablation) | DWBC, DExperts · RWRL Walker | safety | Mixed |
Closest papers by meaning
- Simplifying Constraint Inference with Inverse Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- GO-DICE: Goal-Conditioned Option-Aware Offline Imitation Learning via Stationary Distribution Correction Estimation (AAAI 2024 · similarity 0.89)
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.89)
- Learning When to Stop: Selective Imitation Learning Under Arbitrary Dynamics Shift (NeurIPS 2026 · similarity 0.88)
- SPRINQL: Sub-optimal Demonstrations driven Offline Imitation Learning (NeurIPS 2024 · similarity 0.88)
- Offline Safe Reinforcement Learning Using Trajectory Classification (AAAI 2025 · similarity 0.88)
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.
