Research map / Reinforcement learning / Visual and representation RL
Conditional Mutual Information for Disentangled Representations in Reinforcement Learning
Mhairi Dunion, Trevor McInroe, Kevin Sebastian Luck, Josiah P. Hanna, Stefano V Albrecht
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 CMID to SVEA | SVEA · cartpole_swingup, reversed correlation | zero-shot return: 233.1 ± 25.3 -> 746.0 ± 77.8 | Improved |
| Add CMID to SVEA | SVEA · walker_walk, reversed correlation | zero-shot return: 460.8 ± 50.7 -> 793.5 ± 36.0 | Improved |
| Add CMID to SVEA | SVEA · finger_spin, reversed correlation | zero-shot return: 633.8 ± 122.6 -> 939.5 ± 19.1 | Improved |
| Add CMID to SVEA | SVEA · hopper_stand, reversed correlation | zero-shot return: 686.3 ± 170.8 -> 826.0 ± 15.6 | Improved |
| Add CMID to SVEA | SVEA · cartpole_swingup, uncorrelated variables | zero-shot return: 371.5 ± 21.0 -> 878.8 ± 12.4 | Improved |
| Add CMID to SVEA | SVEA · walker_walk, uncorrelated variables | zero-shot return: 652.4 ± 34.3 -> 815.3 ± 29.9 | Improved |
| Add CMID to SVEA | SVEA · finger_spin, uncorrelated variables | zero-shot return: 680.1 ± 98.4 -> 953.2 ± 16.4 | Improved |
| Add CMID to SVEA | SVEA · hopper_stand, uncorrelated variables | zero-shot return: 526.8 ± 182.2 -> 816.1 ± 37.2 | Improved |
| Use conditional MI instead of unconditional MI (ablation) | SVEA-CMID on cartpole_swingup · Reversed-correlation generalisation and training | return | Improved |
| Condition on the most recent timestep instead of two or three previous timesteps (ablation) | SVEA-CMID on cartpole_swingup · Reversed-correlation generalisation and training | return | Improved |
| Decrease the CMID loss coefficient (ablation) | SVEA-CMID on cartpole_swingup · Reversed-correlation generalisation | return | Hurt |
| Decrease the CMID loss coefficient (ablation) | SVEA-CMID on cartpole_swingup · Training | return | Hurt |
Closest papers by meaning
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.87)
- Disentangled Unsupervised Skill Discovery for Efficient Hierarchical Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.86)
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.86)
- Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- $\alpha$TC-VAE: On the relationship between Disentanglement and Diversity (ICLR 2024 · 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.
