Research map / Reinforcement learning / Imitation learning
Learning a Prior over Intent via Meta-Inverse Reinforcement Learning
Kelvin Xu, Ellis Ratner, Anca Dragan, Sergey Levine, Chelsea Finn
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 |
|---|---|---|---|
| Meta-train a reward initialization for fast adaptation rather than train from scratch | MaxEnt IRL on SUNCG · SUNCG Test (5 demonstrations) | success rate (%): 65.4 -> 77.3 | Improved |
| Meta-train a reward initialization for fast adaptation rather than train from scratch | MaxEnt IRL on SUNCG · SUNCG Unseen Houses (5 demonstrations) | success rate (%): 76.5 -> 82.6 | Improved |
| Optimize the initialization for adaptation instead of following the average task gradient | MaxEnt IRL on SUNCG · SUNCG Test (5 demonstrations) | success rate (%): 60.8 -> 77.3 | Improved |
| Optimize the initialization for adaptation instead of following the average task gradient | MaxEnt IRL on SUNCG · SUNCG Unseen Houses (5 demonstrations) | success rate (%): 73.3 -> 82.6 | Improved |
| Adapt the meta-learned reward initialization using demonstrations (ablation) | MandRIL on SUNCG · SUNCG Test (5 demonstrations) | success rate (%): 20.7 -> 77.3 | Improved |
| Adapt the meta-learned reward initialization using demonstrations (ablation) | MandRIL on SUNCG · SUNCG Unseen Houses (5 demonstrations) | success rate (%): 25.3 -> 82.6 | Improved |
| Use average-gradient supervised pretraining instead of random initialization | MaxEnt IRL on SUNCG · SUNCG Test (5 demonstrations) | success rate (%): 65.4 -> 60.8 | Hurt |
| Use average-gradient supervised pretraining instead of random initialization | MaxEnt IRL on SUNCG · SUNCG Unseen Houses (5 demonstrations) | success rate (%): 76.5 -> 73.3 | Hurt |
| Use an LSTM-based recurrent meta-learner instead of MandRIL | SpriteWorld reward learning · Held-out tasks and held-out tasks with novel sprites | value difference | Hurt |
| Increase the number of demonstrations for learning from scratch | MaxEnt IRL on SpriteWorld · Held-out tasks | value difference: 20 or more demonstrations | Mixed |
| Apply bias transformation (ablation) | SpriteWorld LSTM meta-learner · SpriteWorld | performance | No clear effect |
| Condition the LSTM initial hidden state on image features (ablation) | SpriteWorld LSTM meta-learner · SpriteWorld | performance | No clear effect |
Closest papers by meaning
- Meta-Inverse Reinforcement Learning with Probabilistic Context Variables (NeurIPS 2019 · similarity 0.94)
- Is Inverse Reinforcement Learning Harder than Standard Reinforcement Learning? A Theoretical Perspective (ICML 2024 · similarity 0.91)
- Rethinking Inverse Reinforcement Learning: from Data Alignment to Task Alignment (NeurIPS 2024 · similarity 0.90)
- Recent Advancements in Inverse Reinforcement Learning (AAAI 2024 · similarity 0.90)
- Offline Inverse RL: New Solution Concepts and Provably Efficient Algorithms (ICML 2024 · similarity 0.90)
- Environment Design for Inverse Reinforcement Learning (ICML 2024 · similarity 0.89)
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.
