Research map / Reinforcement learning / Policy gradient and actor-critic
FlowPG: Action-constrained Policy Gradient with Normalizing Flows
Janaka Chathuranga Brahmanage, Jiajing Ling, Akshat Kumar
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 FlowPG's normalizing flow rather than WGAN (ablation) | Reacher action generation · Reacher | valid-action accuracy: 98% -> 99.98% | Improved |
| Use FlowPG's normalizing flow rather than VAE (ablation) | Reacher action generation · Reacher | valid-action accuracy: 83% -> 99.98% | Improved |
| Use HMC rather than rejection sampling to generate feasible training actions (ablation) | FlowPG flow training · Reacher | valid-action sampling success rate: 3.93% -> 100% | Improved |
| Use HMC rather than rejection sampling to generate feasible training actions (ablation) | FlowPG flow training · Half Cheetah | valid-action sampling success rate: 4.7% -> 100% | Improved |
| Replace the standard Gaussian latent prior with a mollified uniform prior (ablation) | FlowPG · Action-constrained continuous control | constraint violations | Improved |
| Use the learned flow mapping instead of QP projection for policy actions | DDPG policy in action-constrained tasks · Reacher and BSS | average return | Improved |
| Use the learned flow mapping instead of Frank-Wolfe policy updates requiring QP solutions | DDPG policy in action-constrained tasks · Continuous control tasks | training speed: 2~3 times | Improved |
| Use FlowPG rather than DDPG+Projection | DDPG on action-constrained tasks · Reacher | average return | Improved |
| Use FlowPG rather than NFWPO | DDPG on action-constrained tasks · Half Cheetah, Hopper, and Walker2d | average return | No clear effect |
| Use FlowPG rather than DDPG+Projection or NFWPO | DDPG on action-constrained tasks · All environments except Walker2d | cumulative constraint violations: an order of magnitude | Improved |
| Use FlowPG rather than the projection-based baselines | DDPG on action-constrained tasks · All environments except Hopper and Walker2d | average magnitude of constraint violations | Improved |
| Use FlowPG rather than DDPG+Projection | DDPG on action-constrained tasks · Continuous control tasks | average return | Improved |
Closest papers by meaning
- Normalizing Flows are Capable Models for Continuous Control (NeurIPS 2025 · similarity 0.89)
- Leveraging Constraint Violation Signals for Action Constrained Reinforcement Learning (AAAI 2025 · similarity 0.89)
- Efficient Action-Constrained Reinforcement Learning via Acceptance-Rejection Method and Augmented MDPs (ICLR 2025 · similarity 0.89)
- Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.86)
- Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning (NeurIPS 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.
