Research map / Reinforcement learning / Policy gradient and actor-critic
Multi-timescale Reinforcement Learning by Value Reconstruction
Zhan Su, Peixi Peng, Xinyu Hu, Cong Li, Yisen Zhao, Zhuojian Li, Yonghong Tian, Fanqi Shen
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 multi-timescale critics with reward decoder and attention fusion to SAC | SAC · DMControl 6 tasks | average score: 206.8 -> 475.2 | Improved |
| add multi-timescale critics with reward decoder and attention fusion to PPO | PPO · DMControl 6 tasks | average score: 167.7 -> 445.7 | Improved |
| add multi-timescale framework to DrQ-v2 | DrQ-v2 · DMControl 6 common tasks, 500K steps | average score: 888.0 -> 918.3 | Improved |
| add multi-timescale framework to DrQ-v2 | DrQ-v2 · DMControl 5 hard tasks, 500K steps | average score: 406.8 -> 528.4 | Improved |
| add multi-timescale framework to DrQ-v2 | DrQ-v2 · DMControl 5 hard tasks, 1M steps | average score: 564.0 -> 686.6 | Improved |
| add multi-timescale framework to DrQ-v2 | DrQ-v2 · CARLA highway (4 weathers) | average episode score: 223.6 -> 343.1 | Improved |
| increase number of critics from N=2 to N=4 (ablation) | Ours · Quadruped Walk 1M | score: 797 -> 830 | Improved |
| increase number of critics from N=4 to N=8 (ablation) | Ours · Walker Run 1M | score: 717 -> 726 | No clear effect |
| vary decoding horizon T among 10, 15, 20 (ablation) | Ours · Quadruped Walk 1M | score: 850 -> 846 -> 857 | No clear effect |
| widen DrQ-v2 critic to match parameter budget (ablation) | DrQ-v2 · Quadruped Walk 1M | score: 732 -> 760 | No clear effect |
| share trunk across critics (compact variant) (ablation) | Ours · Quadruped Walk 1M | score: 846 -> 830 | No clear effect |
| add reconstruction loss, reward loss, attention fusion, and neural decoder (each removed in ablation) (ablation) | Ours · Quadruped Walk and CARLA | learning curve | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Resact — ResAct: Reinforcing Long-term Engagement in Sequential Recommendation with Residual Actor (ICLR 2023) · better on 4 of 4 shared results
- CURL — CURL: Contrastive Unsupervised Representations for Reinforcement Learning (ICML 2020) · better on 1 of 1 shared result
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
- Inverting the Bellman Equation: From $Q$-Values to World Models (NeurIPS 2026 · similarity 0.88)
- Long-Horizon Q-Learning: Accurate Value Learning via n-Step Inequalities (NeurIPS 2026 · similarity 0.88)
- Enhancing Value Function Estimation through First-Order State-Action Dynamics in Offline Reinforcement Learning (ICML 2024 · similarity 0.87)
- A Perspective of Q-value Estimation on Offline-to-Online Reinforcement Learning (AAAI 2024 · similarity 0.87)
- A Unifying View of Linear Function Approximation in Off-Policy RL Through Matrix Splitting and Preconditioning (NeurIPS 2025 · similarity 0.87)
- Towards Global Optimality for Practical Average Reward Reinforcement Learning without Mixing Time Oracles (ICML 2024 · similarity 0.87)
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.
