Research map / Reinforcement learning / Visual and representation RL
Reinforcement Learning with Automated Auxiliary Loss Search
Tairan He, Yuge Zhang, Kan Ren, Minghuan Liu, Che Wang, Weinan Zhang, Yuqing Yang, Dongsheng Li
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 |
|---|---|---|---|
| add searched auxiliary loss A2-winner to SAC vs CURL | SAC · DMC500K Reacher-Easy / Cheetah-Run | episodic reward: 518 -> 613 | Improved |
| add A2-winner vs CURL at 100K steps | SAC · DMC100K Reacher-Easy | episodic reward: 538 -> 778 | Improved |
| add A2-winner vs CURL at 100K steps | SAC · DMC100K Ball in cup-Catch | episodic reward: 769 -> 862 | Improved |
| add A2-winner on Atari with Efficient Rainbow vs CURL | Efficient Rainbow · Atari100k | mean human-normalized score: 0.381 -> 0.568 | Improved |
| add A2-winner on Atari vs DrQ | Efficient Rainbow · Atari100k | median human-normalized score: 0.268 -> 0.317 | Improved |
| use a 6-layer conv encoder instead of the 4-layer encoder used in search (ablation) | A2-winner · image-based DMC | episodic reward | Improved |
| use deeper encoder ranking 6-layer > 4-layer > 2-layer (ablation) | A2-winner · image-based DMC | episodic reward | Improved |
| add A2-winner in 20% masked-state POMDP vs CURL and SAC-DenseMLP (ablation) | SAC · Hopper-Hop and Cheetah-Run (POMDP) | episodic reward | Improved |
| search from scratch on vector tasks (A2-winner-v) instead of reusing A2-winner (ablation) | A2-winner · vector DMC 18 environments | environments beating baselines: 12 -> 16 | Improved |
| MSE loss without negative samples instead of with negatives (ablation) | A2LS loss operator · 3 image DMC500K envs | normalized episodic reward: 0.124 -> 1.000 | Improved |
| use 1-layer DenseMLP instead of 1-layer plain MLP encoder (ablation) | A2LS · vector DMControl100K | normalized episodic reward: 0.919 -> 1.000 | Improved |
| use 4-layer instead of 1-layer DenseMLP encoder (ablation) | A2LS · vector DMControl100K | normalized episodic reward: 1.000 -> 0.813 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CURL — CURL: Contrastive Unsupervised Representations for Reinforcement Learning (ICML 2020) · better on 14 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Return-Based Contrastive Representation Learning for Reinforcement Learning (ICLR 2021 · similarity 0.89)
- Adaptive Auxiliary Task Weighting for Reinforcement Learning (NeurIPS 2019 · similarity 0.88)
- Bootstrapped Representations in Reinforcement Learning (ICML 2023 · similarity 0.88)
- Understanding and Preventing Capacity Loss in Reinforcement Learning (ICLR 2022 · similarity 0.87)
- Retrieval-Augmented Reinforcement Learning (ICML 2022 · similarity 0.86)
- Boosting Reinforcement Learning with Strongly Delayed Feedback Through Auxiliary Short Delays (ICML 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.
