Research map / Reinforcement learning / Visual and representation RL
C-Learning: Learning to Achieve Goals via Recursive Classification
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 |
|---|---|---|---|
| Use off-policy TD C-learning instead of Q-learning with hindsight relabeling | Future state density estimation · Continuous gridworld, off-policy | KL divergence: 14% lower | Helped |
| Use Q-learning with hindsight relabeling instead of MC C-learning or TD C-learning | Future state density estimation · Continuous gridworld, on-policy | prediction error: more than three times worse | Hurt |
| Use TD C-learning instead of MC C-learning | Future state density estimation · Continuous gridworld, off-policy | prediction accuracy | Helped |
| Use Q-learning with hindsight relabeling instead of C-learning | Future state density estimation · Continuous gridworld | future-state density accuracy | Hurt |
| Use a one-step dynamics model instead of MC or TD C-learning | Future-state prediction · Walker2d-v2 | prediction accuracy | Mixed |
| Use MC or TD C-learning instead of a one-step dynamics model at larger discount factors | Future-state prediction · Walker2d-v2 | prediction accuracy | Helped |
| Use Q-learning with hindsight relabeling instead of MC C-learning | Future-state density estimation · Continuous gridworld, on-policy | density normalization: summed to less than one | Hurt |
| Tune the hindsight relabeling ratio rather than use an unsuitable ratio | Q-learning with hindsight relabeling · maze2d-umaze-v0 | final distance: 50% decrease | Helped |
| Use goal-conditioned C-learning instead of prior goal-conditioned Q-learning baselines | Goal-conditioned RL · Continuous control tasks | distance to goal | Mixed |
| Mix TD C-learning with MC C-learning instead of using standard TD C-learning (ablation) | Goal-conditioned C-learning · Sawyer Push and Sawyer Drawer | goal-reaching performance | Helped |
| Mix TD C-learning with MC C-learning instead of using TD C-learning alone (ablation) | Goal-conditioned C-learning · Other tasks | goal-reaching performance | No clear effect |
| Use a 1D classifier output with sigmoid instead of a 2D output with softmax (ablation) | C-learning classifier · Classifier parametrization | performance | No clear effect |
Closest papers by meaning
- C-Planning: An Automatic Curriculum for Learning Goal-Reaching Tasks (ICLR 2022 · similarity 0.89)
- Goal Recognition as Reinforcement Learning (AAAI 2022 · similarity 0.89)
- Learning to Reach Goals via Diffusion (ICML 2024 · similarity 0.88)
- Autonomous Goal Detection and Cessation in Reinforcement Learning: A Case Study on Source Term Estimation (AAAI 2025 · similarity 0.87)
- From Weak Data to Strong Policy: Q-Targets Enable Provable In-Context Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Inverting the Bellman Equation: From $Q$-Values to World Models (NeurIPS 2026 · 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.