Research map / Reinforcement learning / Policy gradient and actor-critic
An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay
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 |
|---|---|---|---|
| Replace vanilla TD3 replay/loss with LAP | TD3 · MuJoCo suite | average return | Helped |
| Replace vanilla TD3 loss with PAL | TD3 · MuJoCo suite | average return | Helped |
| Replace PER with LAP | TD3 · Atari games | performance: 9 out of 10 environments; mean performance gain of 53% | Helped |
| Replace PER with PAL | DDQN · Atari games | performance: 6 of the 10 environments | Hurt |
| Replace LAP prioritization with its uniformly sampled PAL equivalent | TD3 · MuJoCo suite | average return | No clear effect |
| Add PER to TD3 | TD3 · MuJoCo suite | average return | No clear effect |
| Use PAL instead of Huber loss (ablation) | TD3 · MuJoCo suite | average return | Helped |
| Add the PAL scaling factor 1/λ (ablation) | TD3 with PAL · MuJoCo suite | average return | Helped |
| Set α ≠ 0 in PAL rather than use the α = 0 Huber-equivalent loss (ablation) | TD3 with PAL · MuJoCo suite | average return | Helped |
| Add scaling 1/λ to the Huber loss (ablation) | TD3 · MuJoCo suite | average return | Helped |
| Use Huber loss rather than PAL (ablation) | TD3 · MuJoCo suite | average return | Hurt |
| Use L1 loss with priority |δ|^α rather than LAP's Huber loss and clipped priority (ablation) | TD3 · MuJoCo suite | average return | Hurt |
Closest papers by meaning
- Actor Prioritized Experience Replay (Abstract Reprint) (AAAI 2024 · similarity 0.91)
- Large Batch Experience Replay (ICML 2022 · similarity 0.91)
- Prioritizing Samples in Reinforcement Learning with Reducible Loss (NeurIPS 2023 · similarity 0.91)
- Reliability-Adjusted Prioritized Experience Replay (ICLR 2026 · similarity 0.90)
- When Does Non-Uniform Replay Matter in Reinforcement Learning? (NeurIPS 2026 · similarity 0.88)
- Prioritized Generative Replay (ICLR 2025 · similarity 0.88)
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.