Research map / Learning theory and algorithms / Sample complexity
Self-Imitation Learning via Generalized Lower Bound Q-learning
Yunhao Tang
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 |
|---|---|---|---|
| Use 5-step generalized SIL instead of 5-step TD3 | TD3 · DMWalkerRun, DMWalkerStand, DMWalkerWalk, DMCheetahRun, Ant, Ant(B), HalfCheetah(B) | performance: 246 ± 49 -> 642 ± 107; 749 ± 150 -> 979 ± 2; 925 ± 10 -> 731 ± 151; 553 ± 92 -> 830 ± 36; 1058 ± 281 -> 4123 ± 364; 2920 | Improved |
| Use 5-step generalized SIL instead of 5-step TD3 | TD3 · HalfCheetah | performance: 6178 ± 151 -> 8246 ± 784 | Hurt |
| Use n=5 generalized SIL rather than n=1 generalized SIL (ablation) | TD3 · DMWalkerRun, DMWalkerStand, DMWalkerWalk, DMCheetahRun, Ant, HalfCheetah, Ant(B), HalfCheetah(B) | performance: 500 ± 138 -> 642 ± 107; 899 ± 55 -> 979 ± 2; 601 ± 108 -> 731 ± 151; 702 ± 72 -> 830 ± 36; 3166 ± 390 -> 4123 ± 364; 104 | Mixed |
| Use 5-step generalized SIL rather than return-based SIL (n=infinity) (ablation) | TD3 · DMWalkerRun, DMWalkerStand, DMWalkerWalk, DMCheetahRun, Ant, HalfCheetah, Ant(B), HalfCheetah(B) | performance: 320 ± 111 -> 642 ± 107; 748 ± 143 -> 979 ± 2; 398 ± 203 -> 731 ± 151; 655 ± 59 -> 830 ± 36; 3787 ± 411 -> 4123 ± 364; 83 | Mixed |
| Remove IS corrections from prioritized replay (ablation) | Generalized SIL n=5 with TD3 · DMWalkerRun, DMWalkerStand, DMWalkerWalk, DMCheetahRun, Ant, HalfCheetah, Ant(B), HalfCheetah(B) | performance: 642 ± 107 -> 675 ± 15; 979 ± 2 -> 947 ± 18; 731 ± 151 -> 622 ± 197; 830 ± 36 -> 597 ± 64; 4123 ± 364 -> 3059 ± 360; 8246 | Mixed |
| Add prioritized sampling (ablation) | Generalized SIL n=5 with TD3 · DMWalkerRun, DMWalkerStand, DMCheetahRun, HalfCheetah, HalfCheetah(B) | performance: 424 ± 127 -> 642 ± 107; 634 ± 184 -> 979 ± 2; 505 ± 182 -> 830 ± 36; 8927 ± 596 -> 8246 ± 784; 2284 ± 85 -> 2619 ± 129 | Improved |
| Add prioritized sampling (ablation) | Generalized SIL n=5 with TD3 · DMWalkerWalk, Ant, Ant(B) | performance: 766 ± 103 -> 731 ± 151; 4358 ± 496 -> 4123 ± 364; 2910 ± 88 -> 2954 ± 54 | Mixed |
| Use generalized SIL n=5 rather than return-based SIL | PPO · Ant and Humanoid delayed-reward tasks | performance | Mixed |
| Apply generalized SIL n=5 instead of plain PPO | PPO · OpenAI Gym delayed-reward tasks | performance | Improved |
| Add generalized SIL n=5 to DDPG | DDPG · Continuous-control tasks | performance | No clear effect |
| Use 5-step generalized SIL rather than n=1 generalized SIL (ablation) | TD3 · Deterministic continuous-control tasks | performance | Improved |
| Use return-based SIL (n=infinity) rather than 5-step generalized SIL (ablation) | TD3 · Deterministic continuous-control tasks | optimization speed | Hurt |
Closest papers by meaning
- Agnostic Interactive Imitation Learning: New Theory and Practical Algorithms (ICML 2024 · similarity 0.86)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.85)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (ICLR 2024 · similarity 0.85)
- Reward-free World Models for Online Imitation Learning (ICML 2025 · similarity 0.85)
- Inverse Factorized Soft Q-Learning for Cooperative Multi-agent Imitation Learning (NeurIPS 2024 · similarity 0.85)
- Enhancing Online Reinforcement Learning with Meta-Learned Objective from Offline Data (AAAI 2025 · similarity 0.85)
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.
