Research map / Reinforcement learning / Imitation learning
Discriminator-Weighted Offline Imitation Learning from Suboptimal 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 |
|---|---|---|---|
| Use discriminator outputs as weights for behavioral cloning, with the policy also informing discriminator training | Offline imitation learning from expert and suboptimal demonstrations · D4RL and Adroit tasks, 36-task evaluation | normalized return: 27 out of 36 tasks | Helped |
| Use discriminator-weighted cloning rather than the compared baselines | Expert+random datasets · Settings 1 and 3, 24 tasks | normalized return: 21 out of 24 tasks | Helped |
| Use DWBC instead of BC-exp | Offline imitation learning · ant_exp-rand-90 | normalized return: 11.1$\pm$8.7 -> 25.7$\pm$10.3 | Hurt |
| Use DWBC instead of ORIL | Offline imitation learning · walker2d_mixed-5 | normalized return: 61.3$\pm$5.8 -> 50.7$\pm$8.2 | Hurt |
| Use DWBC instead of BC-exp | Offline imitation learning · door_exp-cloned-30 | normalized return: 31.0$\pm$26.9 -> 5.2$\pm$3.9 | Hurt |
| Use DWBC instead of BC-exp | Offline imitation learning · hammer_exp-cloned-90 | normalized return: 18.5$\pm$7.4 -> 14.2$\pm$8.4 | Hurt |
| Increase the fraction parameter X, leaving fewer well-performing expert transitions in the expert set | BC-exp · Settings 1 and 3 | normalized return | Hurt |
| Use offline policy selection scores from the learned discriminator | Candidate policies evaluated with DWBC discriminator · Hopper, Walker2d, and Pen | policy ranking agreement | Helped |
| Use DWBC rather than BC as the training procedure | Offline imitation learning runtime · 500,000 training steps | runtime | Hurt |
| Use ORIL or BCND rather than BC | Offline imitation learning runtime · 500,000 training steps | runtime: over 7 times | Hurt |
Closest papers by meaning
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.92)
- SEABO: A Simple Search-Based Method for Offline Imitation Learning (ICLR 2024 · similarity 0.92)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.91)
- An Optimal Discriminator Weighted Imitation Perspective for Reinforcement Learning (ICLR 2025 · similarity 0.91)
- SPRINQL: Sub-optimal Demonstrations driven Offline Imitation Learning (NeurIPS 2024 · similarity 0.90)
- Offline Actor-Critic Reinforcement Learning Scales to Large Models (ICML 2024 · 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.