Research map / Reinforcement learning / Visual and representation RL
Skill Learning via Policy Diversity Yields Identifiable Representations for Reinforcement Learning
Patrik Reizinger, Bálint Mucsányi, Siyuan Guo, Benjamin Eysenbach, Bernhard Schölkopf, Wieland Brendel
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 a small fixed set of pre-sampled skills instead of sampling skills from the uniform prior (ablation) | CSF · Ant (MuJoCo) | state coverage and R2 of linear fit to ground-truth state | Hurt |
| increase the number of fixed discrete skills (enough to span the state space) (ablation) | CSF · Ant (MuJoCo) | R2 / zero-shot oracle reward | Improved |
| use a latent dimensionality smaller than the ground-truth state dimensionality (ablation) | CSF · Ant (MuJoCo) | R2 (state identifiability) | Hurt |
| use a smaller latent dimensionality for zero-shot task transfer (ablation) | CSF · Ant (MuJoCo) | oracle reward (zero-shot transfer) | Mixed |
| match latent dimensionality to the ground-truth state dimensionality (ablation) | CSF · Ant (MuJoCo) | R2 (state identifiability) | Improved |
| train CSF from states rather than pixels | CSF · Half Cheetah, Ant, Quadruped (state) vs Quadruped (pixel) | R2 of linear fit to ground-truth state | Improved |
| increase the number of negative samples in the contrastive loss (ablation) | CSF · MuJoCo/DMC environments | state coverage and oracle reward | Mixed |
| add skip-connections to the encoder backbone | CSF · MuJoCo/DMC environments | identifiability | No clear effect |
| use a maximum-entropy skill-conditioned policy | CSF · theory (Lemma) | skill diversity / reward | Hurt |
| optimize I(s;z) instead of I(s,s';z) | CSF · theory | feature geometry | Hurt |
Closest papers by meaning
- Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning (ICLR 2025 · similarity 0.92)
- Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Task Adaptation from Skills: Information Geometry, Disentanglement, and New Objectives for Unsupervised Reinforcement Learning (ICLR 2024 · similarity 0.87)
- Exploration by Learning Diverse Skills through Successor State Representations (NeurIPS 2024 · similarity 0.87)
- Learning Successor Features the Simple Way (NeurIPS 2024 · similarity 0.86)
- Efficient Skill Discovery via Regret-Aware Optimization (ICML 2025 · 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.
