Research map / Reinforcement learning / Imitation learning
Variational Discriminator Bottleneck: Improving Imitation Learning, Inverse RL, and GANs by Constraining Information Flow
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 |
|---|---|---|---|
| Add the variational discriminator bottleneck to GAIL | GAIL · Humanoid motion imitation | average joint rotation error | Helped |
| Add gradient penalty to VAIL (ablation) | VAIL · Humanoid motion imitation | average joint rotation error: 0.36 ± 0.13 -> 0.46 ± 0.17; 0.40 ± 0.08 -> 0.31 ± 0.02; 0.40 ± 0.21 -> 0.15 ± 0.01; 0.13 ± 0.01 -> 0.10 ± 0.01; 0.34 ± 0 | Mixed |
| Add instance noise to discriminator inputs without the KL constraint (ablation) | GAIL · Humanoid motion imitation | average joint rotation error: 0.74 ± 0.15 -> 0.42 ± 0.02; 0.84 ± 0.05 -> 0.92 ± 0.07; 1.31 ± 0.16 -> 0.96 ± 0.08; 0.17 ± 0.03 -> 0.21 ± 0.05; 1.07 ± 0 | Mixed |
| Add instance noise to the discriminator's last hidden layer without the KL constraint (ablation) | GAIL · Humanoid motion imitation | average joint rotation error: 0.74 ± 0.15 -> 0.67 ± 0.12; 0.84 ± 0.05 -> 0.72 ± 0.04; 1.31 ± 0.16 -> 1.14 ± 0.08; 0.17 ± 0.03 -> 0.14 ± 0.03; 1.07 ± 0 | Mixed |
| Add gradient penalty (ablation) | GAIL · Humanoid motion imitation | average joint rotation error: 0.74 ± 0.15 -> 0.62 ± 0.09; 0.84 ± 0.05 -> 0.69 ± 0.05; 1.31 ± 0.16 -> 0.80 ± 0.32; 0.17 ± 0.03 -> 0.12 ± 0.02; 1.07 ± 0 | Helped |
| Remove the adaptive update of beta and use a fixed beta (ablation) | VAIL · Video imitation | learning performance | Hurt |
| Disable the KL constraint (beta = 0) (ablation) | VAIRL · C-maze and S-maze reward transfer | mean return | Hurt |
| Add gradient penalty to VAIRL (ablation) | VAIRL · S-maze reward transfer | mean return | Helped |
| Apply gradient penalty to both real and fake samples instead of only real samples (ablation) | VDB with gradient penalty · Adversarial learning | performance | Hurt |
| Combine VDB with spectral normalization (ablation) | VGAN · Image generation | training stability | Hurt |
| Use a phase-functioned discriminator (ablation) | VAIL · Humanoid motion imitation | average joint rotation error | Helped |
| Remove the phase variable from the discriminator input (ablation) | VAIL discriminator · Humanoid motion imitation | average joint rotation error | Hurt |
Closest papers by meaning
- Adversarial Soft Advantage Fitting: Imitation Learning without Policy Optimization (NeurIPS 2020 · similarity 0.90)
- Deterministic and Discriminative Imitation (D2-Imitation): Revisiting Adversarial Imitation for Sample Efficiency (AAAI 2022 · similarity 0.88)
- Discriminator-Actor-Critic: Addressing Sample Inefficiency and Reward Bias in Adversarial Imitation Learning (ICLR 2019 · similarity 0.88)
- Domain-Robust Visual Imitation Learning with Mutual Information Constraints (ICLR 2021 · similarity 0.88)
- Diffusing States and Matching Scores: A New Framework for Imitation Learning (ICLR 2025 · similarity 0.87)
- Adversarial Imitation Learning via Boosting (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.