Research map / Reinforcement learning / Imitation learning
Learning to Weight Imperfect Demonstrations
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 |
|---|---|---|---|
| weight demonstrations by learned weight (WGAIL) instead of plain GAIL | GAIL with imperfect demonstrations · Mujoco Ant/HalfCheetah/Walker2d/Hopper, Stage 1 and 2 | final x-position: e.g., Ant Stage 2 182.00 vs 113.58 | Helped |
| use weight 1/Dψ instead of learned weight (ablation) | WGAIL · Ant-v2 | final x-position: 155.85 vs 182.00 (Stage 2); 82.54 vs 111.81 (Stage 1) | Hurt |
| multiple demonstrators (3 checkpoints mixed): WGAIL vs GAIL | WGAIL · Ant-v2, HalfCheetah-v2, Walker2d-v2, Hopper-v2 | final x-position: Ant 172.62 vs 122.96; HalfCheetah 187.58 vs 111.63; Walker 14.43 vs 7.16; Hopper 15.56 vs 14.29 | Helped |
| vary ratio of optimal demonstrations (alpha) with WGAIL vs GAIL (ablation) | WGAIL · Ant-v2 Stage 2 | final x-position: large gains at alpha 0.25-0.5, advantage vanishes by alpha 0.75 and both break down at 0 and 0.125 | Mixed |
| weight learning on Atari (PPO-based GAIL) with imperfect demonstrations | WGAIL · Atari BeamRider, Pong, Seaquest, Hero, Qbert | final score: Hero 20042.0 vs 20260.0 (GAIL) and Beamrider/Pong/Seaquest gains; only minor gains in four of five games | Mixed |
| BC-based baseline vs WGAIL on Hopper with fewer demonstrations | BC · Hopper-v2 Stage 2 | final x-position: 0.95 vs 14.54 | Hurt |
| time-index preference ranking (T-REX) instead of learned weighting | T-REX · Ant-v2 Stage 2 | final x-position: 9.25 vs 182.00 (WGAIL) | Hurt |
Closest papers by meaning
- Imitation Learning from Imperfect Demonstration (ICML 2019 · similarity 0.92)
- PN-GAIL: Leveraging Non-optimal Information from Imperfect Demonstrations (ICLR 2025 · similarity 0.91)
- C-GAIL: Stabilizing Generative Adversarial Imitation Learning with Control Theory (NeurIPS 2024 · similarity 0.90)
- Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective (AAAI 2024 · similarity 0.88)
- Adversarial Imitation Learning via Boosting (ICLR 2024 · similarity 0.87)
- Diffusion-Reward Adversarial Imitation Learning (NeurIPS 2024 · similarity 0.87)
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.