Research map / Reinforcement learning / Imitation learning
Discriminator-Actor-Critic: Addressing Sample Inefficiency and Reward Bias in Adversarial Imitation Learning
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 |
|---|---|---|---|
| Train the actor-critic and discriminator off-policy using replay data | DAC · MuJoCo task suite | environment interactions for comparable reward: average factor of 10 | Helped |
| Learn rewards for absorbing states (ablation) | GAIL-style discriminator reward · Kuka-Reach and Kuka-PushNext | imitation performance: 600 expert trajectories | Helped |
| Use -log(1-D(s,a)) without learning absorbing-state rewards (ablation) | GAIL · Kuka-Reach and Kuka-PushNext | imitation performance: 600 expert trajectories | Hurt |
| Use a fixed untrained discriminator that outputs 0.5 (ablation) | GAIL on Hopper · Hopper | episode reward: around 1000 | Helped |
| Use a strictly positive discriminator reward without absorbing-state handling (ablation) | GAIL · episodic tasks with per-step penalties | imitation performance | Hurt |
| Use a strictly negative discriminator reward (ablation) | GAIL · tasks with a survival bonus | imitation performance | Hurt |
| Learn absorbing-state rewards with an AIRL discriminator (ablation) | AIRL within DAC · survival-bonus environments | policy performance | Helped |
| Compute terminal returns analytically instead of recursively (ablation) | DAC · episodic tasks | training stability | Hurt |
Closest papers by meaning
- Adversarial Soft Advantage Fitting: Imitation Learning without Policy Optimization (NeurIPS 2020 · similarity 0.93)
- Deterministic and Discriminative Imitation (D2-Imitation): Revisiting Adversarial Imitation for Sample Efficiency (AAAI 2022 · similarity 0.92)
- Unlabeled Imperfect Demonstrations in Adversarial Imitation Learning (AAAI 2023 · similarity 0.91)
- Discriminator-Weighted Offline Imitation Learning from Suboptimal Demonstrations (ICML 2022 · similarity 0.91)
- Adversarial Imitation Learning via Boosting (ICLR 2024 · similarity 0.91)
- Adversarial Imitation Learning with Preferences (ICLR 2023 · similarity 0.91)
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.