Research map / Reinforcement learning / Policy gradient and actor-critic
Discretizing Continuous Action Space for On-Policy Optimization
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 |
|---|---|---|---|
| Discretize each action dimension using a factorized categorical policy instead of a Gaussian policy | PPO · Humanoid | cumulative reward: 3905 ± 502 -> 5119 ± 151 | Helped |
| Discretize each action dimension using a factorized categorical policy instead of a Gaussian policy | TRPO · Ant | cumulative reward: -76 ± 14 -> 2687 ± 556 | Helped |
| Replace the categorical output with the ordinal stick-breaking parameterization (ablation) | PPO + discrete policy · Walker2d | cumulative reward: 3390 ± 190 -> 4249 ± 239 | Helped |
| Replace the categorical output with the ordinal stick-breaking parameterization (ablation) | PPO + discrete policy · Humanoid | cumulative reward: 5119 ± 151 -> 6018 ± 403 | Helped |
| Replace the categorical output with the ordinal stick-breaking parameterization (ablation) | TRPO + discrete policy · Humanoid | cumulative reward: 3908 ± 117 -> 3577 ± 272 | Hurt |
| Use PPO with a discrete policy instead of a Gaussian policy | PPO · Ant | cumulative reward: 4445 ± 194 -> 3256 ± 778 | Hurt |
| Use PPO with an ordinal policy instead of a Gaussian policy | PPO · Ant | cumulative reward: 4445 ± 194 -> 3690 ± 557 | Hurt |
| Increase the number of action bins from a very coarse setting to K=2 (ablation) | Discrete policy · MuJoCo continuous-control tasks | cumulative reward | Helped |
| Use a very large number of bins (around K=50) rather than the reported best range 7≤K≤15 (ablation) | Discrete policy · MuJoCo continuous-control tasks | cumulative reward | Hurt |
| Increase the Gaussian policy network from 64 to 128 hidden units per layer (ablation) | PPO · Humanoid | cumulative reward: 3766 ± 413 -> 3221 ± 535 | Hurt |
| Increase the Gaussian policy network from 32 to 128 hidden units per layer (ablation) | TRPO · Humanoid | cumulative reward: 708 ± 43 -> 501 ± 14 | Hurt |
| Bound the Gaussian policy mean with a tanh output (ablation) | PPO · Walker2d | cumulative reward: 3500 ± 360 -> 3274 ± 251 | Hurt |
Closest papers by meaning
- Distributional Policy Optimization: An Alternative Approach for Continuous Control (NeurIPS 2019 · similarity 0.90)
- Deterministic Policy Gradient Primal-Dual Methods for Continuous-Space Constrained MDPs (AAAI 2025 · similarity 0.88)
- Actor-Free Continuous Control via Structurally Maximizable Q-Functions (NeurIPS 2025 · similarity 0.87)
- On the Convergence of Continuous Single-timescale Actor-critic (ICML 2025 · similarity 0.86)
- Wasserstein Policy Optimization (ICML 2025 · similarity 0.86)
- Identifying Policy Gradient Subspaces (ICLR 2024 · 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.