Research map / Reinforcement learning / Imitation learning
Safe Imitation Learning via Fast Bayesian Reward Inference from Preferences
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 |
|---|---|---|---|
| Combine T-REX ranking loss with self-supervised feature pretraining instead of using self-supervision alone (ablation) | Bayesian REX · Atari games | ground-truth average score: Beam Rider: 180.4 -> 5870.3; Breakout: 360.1 -> 393.1; Enduro: 0.0 -> 487.7; Seaquest: 674.0 -> 606.0; Space Invaders: 3 | Helped |
| Add self-supervised pretraining to ranking-loss pretraining (ablation) | Bayesian REX · Atari games | ground-truth average score: Beam Rider: 3816.7 -> 5870.3; Breakout: 389.9 -> 393.1; Enduro: 472.7 -> 487.7; Seaquest: 675.3 -> 606.0; Space Invaders | Mixed |
| Use Bayesian REX with ranked suboptimal demonstrations rather than Bayesian IRL with suboptimal demonstrations | Bayesian reward inference · 6x6 gridworlds, ranked suboptimal demonstrations | average policy loss: B-IRL: 3.512, 3.319, 2.791, 3.078, 3.158; B-REX: 1.796, 0.393, 0.026, 0.006, 0.006 | Helped |
| Use Bayesian REX instead of Bayesian IRL when both receive only optimal demonstrations | Bayesian reward inference · 6x6 gridworlds, optimal demonstrations | average policy loss: B-IRL: 0.045, 0.034, 0.018, 0.009, 0.006; B-REX: 0.051, 0.045, 0.040, 0.034, 0.034 | Hurt |
| Use Bayesian REX with enough ranked suboptimal demonstrations instead of Bayesian IRL with the same number of optimal demonstrations | Bayesian reward inference · 6x6 gridworlds, ranked suboptimal versus optimal demonstrations | average policy loss: 2 demos: 0.044 vs 1.779; 5 demos: 0.033 vs 0.421; 10 demos: 0.020 vs 0.019; 20 demos: 0.009 vs 0.006; 30 demos: 0.006 vs | Mixed |
| Add one ranked rollout from the MAP policy and one from the No-Op policy | Bayesian REX policy evaluation · Breakout | 0.05-VaR: MAP: -370.2 -> 289.1; No-Op: 1.0 -> -13.5 | Mixed |
Closest papers by meaning
- Bayesian Robust Optimization for Imitation Learning (NeurIPS 2020 · similarity 0.91)
- Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration (ICLR 2025 · similarity 0.88)
- Causal Imitation for Markov Decision Processes: a Partial Identification Approach (NeurIPS 2024 · similarity 0.88)
- Uncertainty-Based Smooth Policy Regularisation for Reinforcement Learning with Few Demonstrations (NeurIPS 2025 · similarity 0.86)
- Limited Preference Aided Imitation Learning from Imperfect Demonstrations (ICML 2024 · similarity 0.86)
- SEABO: A Simple Search-Based Method for Offline Imitation Learning (ICLR 2024 · 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.