Research map / Continual learning / Online and replay-based
Disentangling Transfer in Continual Reinforcement 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 |
|---|---|---|---|
| carry over the critic from the previous task (ablation) | SAC · CW10 task pairs | forward transfer: 0.20 | Helped |
| carry over the actor from the previous task (ablation) | SAC · CW10 task pairs | forward transfer: 0.06 | Helped |
| carry over the exploration policy (ablation) | SAC · CW10 task pairs | forward transfer: 0.09 | Helped |
| freeze critic body and train only head (ablation) | SAC critic · CW10 task pairs | forward transfer: -1.29 | Hurt |
| carry over actor, critic and exploration together vs separately (ablation) | SAC · CW10 task pairs | forward transfer: 0.35 | Helped |
| carry over the optimizer statistics (ablation) | SAC · CW10 task pairs | forward transfer: 0.02 | No clear effect |
| longer exploration phase (50000 steps vs 5000) when carrying over exploration only (ablation) | SAC · CW10 task pairs | forward transfer: 0.13 -> -0.00 | Hurt |
| add uniform-previous exploration to EWC (ablation) | EWC · CW10 | average performance: 0.63 -> 0.72 | Helped |
| add best-return exploration to EWC (ablation) | EWC · CW20 | forward transfer: -0.14 -> 0.28 | Helped |
| add informed exploration strategies to PackNet (ablation) | PackNet · CW10 | forward transfer: 0.26 -> 0.23 | No clear effect |
| add behavioral cloning of the actor vs Perfect memory/BC of actor and critic in two-task transfer (ablation) | SAC · CW10 task pairs | forward transfer: 0.01; -0.56 | Mixed |
| replace Perfect memory with actor behavioral cloning on long sequences (ablation) | SAC · CW10 | forward transfer: -1.13 -> 0.41 | Helped |
Closest papers by meaning
- Prevalence of Negative Transfer in Continual Reinforcement Learning: Analyses and a Simple Baseline (ICLR 2025 · similarity 0.91)
- Continual Knowledge Adaptation for Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Continual Reinforcement Learning by Planning with Online World Models (ICML 2025 · similarity 0.86)
- Knowledge Retention in Continual Model-Based Reinforcement Learning (ICML 2025 · similarity 0.86)
- Position: Lifetime tuning is incompatible with continual reinforcement learning (ICML 2025 · similarity 0.86)
- Ergodic Risk Measures: Towards a Risk-Aware Foundation for Continual Reinforcement Learning (NeurIPS 2026 · 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.