Research map / Reinforcement learning / Policy gradient and actor-critic
Geometric Coherence Learning for Structuring Value Functions in Plain MDPs
Zuyuan Zhang, Zeyu Fang, Tian Lan
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 logic-order regularization to symmetry-only (ablation) | GCR-RL · E4 Noisy-RPS Chain | final return: 20.7 ± 0.01 -> 21.60 ± 0.20 | Improved |
| add symmetry (equivariance) losses to logic-only (ablation) | GCR-RL · E4 Noisy-RPS Chain | final return: 20.9 ± 0.00 -> 21.60 ± 0.20 | Improved |
| add learned action relabeling (ablation) | GCR-RL · E2 ButtonPermutation | final return: 2.010 ± 0.01 -> 3.700 ± 0.00 | Improved |
| add learned action relabeling (ablation) | GCR-RL · E4 Noisy-RPS Chain | final return: 19.70 ± 0.21 -> 21.60 ± 0.20 | Improved |
| add logic-order to symmetry-only (ablation) | GCR-RL · E1 RotMirror-Grid | final return: −144.3 ± 5.2 -> −136.7 ± 7.1 | Improved |
| add symmetry losses to logic-only (ablation) | GCR-RL · E1 RotMirror-Grid | final return: −153.2 ± 10.1 -> −136.7 ± 7.1 | Improved |
| add logic-order to symmetry-only (ablation) | GCR-RL · Cartpole | final return: 292.6 ± 32.1 -> 321.9 ± 52.7 | Improved |
| add action relabeling (ablation) | GCR-RL · Cartpole | final return: 314.5 ± 13.2 -> 321.9 ± 52.7 | No clear effect |
| use symmetry regularizer weight λeq=0.10 instead of 0 (ablation) | GCR-RL · E4 Noisy-RPS Chain | final reward at 200 epochs: 16.2 ± 1.8 -> 18.7 ± 0.8 | Improved |
| use order weight λiso=0.10 instead of 0 (ablation) | GCR-RL · E4 Noisy-RPS Chain | final reward at 200 epochs: 15.4 ± 2.0 -> 18.7 ± 0.8 | Improved |
| raise λeq from 0.10 to 0.40 (ablation) | GCR-RL · E4 Noisy-RPS Chain | final reward at 200 epochs: 18.7 ± 0.8 -> 17.1 ± 1.5 | Hurt |
| use K=4 transforms instead of 1 (ablation) | GCR-RL · E4 Noisy-RPS Chain | final reward at 200 epochs: 16.6 ± 1.6 -> 18.7 ± 0.8 | Improved |
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.
- Rainbow — Rainbow: Combining Improvements in Deep Reinforcement Learning (AAAI 2018) · better on 9 of 9 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
- $Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training (NeurIPS 2025 · similarity 0.86)
- SiT: Symmetry-invariant Transformers for Generalisation in Reinforcement Learning (ICML 2024 · similarity 0.85)
- Oracle-Efficient Reinforcement Learning for Max Value Ensembles (NeurIPS 2024 · similarity 0.85)
- Learning the Target Network in Function Space (ICML 2024 · similarity 0.84)
- Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces (ICLR 2025 · similarity 0.84)
- Offline RL with Smooth OOD Generalization in Convex Hull and its Neighborhood (ICLR 2025 · similarity 0.84)
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.
