Research map / Probabilistic and scientific modeling / Dynamical systems
Stable Deep Reinforcement Learning via Isotropic Gaussian Representations
Ali Saheb Pasand, Johan Obando-Ceron, Aaron Courville, Pouya Bashivan, Pablo Samuel Castro
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 SIGReg isotropic Gaussian regularization to PQN | PQN · Full Atari (57 games) | games improved / mean AUC improvement: 51 of 57 games improved; mean AUC improvement of 889% and median of 138% | Improved |
| add SIGReg to PQN to keep embedding covariance spread over components | PQN · Atari-10 PCA of embedding covariance | leading-component variance share: PQN: [0.4,0.2] -> [0.9,0.8] vs PQN+SIGReg: [0.3,0.2] -> [0.1,0.1] | Improved |
| replace Gaussian target with Laplacian or Logistic isotropic target distributions (ablation) | PQN + SIGReg · Atari-10 | AUC improvement | Hurt |
| use covariance whitening only (isotropy without Gaussian tails) instead of SIGReg (ablation) | PQN · Atari-10 | AUC improvement | Hurt |
| enforce both symmetry and tail behavior rather than only one (ablation) | PQN + SIGReg · Atari-10 | AUC improvement | Improved |
| add SIGReg on top of baseline first-order optimizer to match Kronecker-factored optimizer | PQN with RAdam · Atari-10 | performance gap to Kronecker-factored optimizer | Improved |
| add SIGReg compared with VICReg, norm regularization, simplicial embeddings, weight orthogonalization | PQN · Atari-10 | average improvement | Improved |
| add SIGReg to PPO | PPO · Full Atari | IQM human-normalized score and AUC | Improved |
| add Kronecker-factored optimization or multi-skip residuals to PPO | PPO · Full Atari | performance | Hurt |
| add SIGReg to continuous control agents | Isaac Gym agents · Isaac Gym (5 seeds) | training stability and variance | Improved |
Closest papers by meaning
- Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Non-Stationary Learning of Neural Networks with Automatic Soft Parameter Reset (NeurIPS 2024 · similarity 0.85)
- MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL (ICLR 2025 · similarity 0.85)
- Understanding and Diagnosing Deep Reinforcement Learning (ICML 2024 · similarity 0.85)
- State Entropy Regularization for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Parseval Regularization for Continual Reinforcement Learning (NeurIPS 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.
