Research map / Reinforcement learning / Visual and representation RL
RD$^2$: Reward Decomposition with Representation Decomposition
Zichuan Lin, Derek Yang, Li Zhao, Tao Qin, Guangwen Yang, Tie-Yan Liu
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 RD2 to decompose reward from total reward only | RD2 (Ldiv1) · Monster-Treasure | predicted rewards: predicted 1.92 and -1.84 for true 2.00 and -2.00 | Improved |
| RD2 reward decomposition on UpNDown | RD2 · UpNDown | sub-reward: predicted 1.98 and 0.03 for reward 2 | Improved |
| RD2 reward decomposition on Gopher | RD2 · Gopher | sub-reward: predicted 0.61 for reward 0.8 | Improved |
| Add RD2 decomposed rewards to Q-learning versus Rainbow | Q-learning with RD2 · Atari games | sample efficiency | Improved |
| Add RD2 decomposed rewards to Q-learning versus DRDRL | Q-learning with RD2 · Atari games | sample efficiency | Improved |
| Early training phase of RD2 joint training | Q-learning with RD2 · Atari games | learning curve | Hurt |
| Use dynamic (input-dependent) mask | RD2 · Gopher | attention of channel two | Improved |
| Ldiv1 pushes channel masks apart | RD2 · Monster-Treasure | mask values on agent factors | Mixed |
Closest papers by meaning
- Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking (ICLR 2025 · similarity 0.86)
- Episodic Return Decomposition by Difference of Implicitly Assigned Sub-trajectory Reward (AAAI 2024 · similarity 0.86)
- Information-Theoretic Reward Decomposition for Generalizable RLHF (NeurIPS 2025 · similarity 0.86)
- Reward-Aware Proto-Representations in Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Submodular Reinforcement Learning (ICLR 2024 · similarity 0.85)
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 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.
