Research map / Reinforcement learning / Policy gradient and actor-critic
Spectral Normalisation for Deep Reinforcement Learning: An Optimisation Perspective
Florin Gogianu, Tudor Berariu, Mihaela C Rosca, Claudia Clopath, Lucian Busoniu, Razvan Pascanu
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 |
|---|---|---|---|
| Apply spectral normalisation to the penultimate layer | C51 · Atari 54 games, random starts | Mean Human Normalised Score: 778.68 -> 1073.18 | Improved |
| Apply spectral normalisation to the penultimate layer | C51 · Atari 54 games, random starts | Median Human Normalised Score: 182.26 -> 248.45 | Improved |
| Apply spectral normalisation to the penultimate layer | DQN-Adam · Atari 54 games, random starts | Mean Human Normalised Score: 358.45 -> 719.95 | Improved |
| Apply spectral normalisation to the penultimate layer | DQN-Adam · Atari 54 games, random starts | Median Human Normalised Score: 119.45 -> 178.18 | Improved |
| Apply spectral normalisation to the penultimate layer | DQN-RMSprop · Atari 54 games, random starts | Mean Human Normalised Score: 357.36 -> 375.37 | Improved |
| Apply spectral normalisation to the penultimate layer | DQN-RMSprop · Atari 54 games, random starts | Median Human Normalised Score: 102.94 -> 105.19 | Improved |
| Apply spectral normalisation to one layer rather than none | DQN across MinAtar games and optimiser settings · MinAtar | MinAtar Normalised Score | Improved |
| Normalise more than half the layers of a depth-5 network (ablation) | DQN · MinAtar | MinAtar Normalised Score | Hurt |
| Apply spectral normalisation to input layers or too many layers (ablation) | DQN · MinAtar | MinAtar Normalised Score | Hurt |
| Replace spectral normalisation with gradient penalty (ablation) | DQN · MinAtar | MinAtar Normalised Score | Hurt |
| Replace spectral normalisation with batch normalisation (ablation) | DQN · MinAtar | MinAtar Normalised Score | Hurt |
| Replace spectral normalisation with divOut output scaling (ablation) | DQN · MinAtar | MinAtar Normalised Score | No clear effect |
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 2 of 2 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
- Towards Deeper Deep Reinforcement Learning with Spectral Normalization (NeurIPS 2021 · similarity 0.91)
- Adapting to Reward Progressivity via Spectral Reinforcement Learning (ICLR 2021 · similarity 0.90)
- Hyperspherical Normalization for Scalable Deep Reinforcement Learning (ICML 2025 · similarity 0.88)
- Spectral Decomposition Representation for Reinforcement Learning (ICLR 2023 · similarity 0.87)
- Overcoming The Spectral Bias of Neural Value Approximation (ICLR 2022 · similarity 0.87)
- Penalizing Infeasible Actions and Reward Scaling in Reinforcement Learning with Offline Data (ICML 2025 · 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.
