Research map / Optimization / Nonconvex and smooth optimization
Stochastic Policy Gradient Methods: Improved Sample Complexity for Fisher-non-degenerate Policies
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 |
|---|---|---|---|
| Use normalized momentum with implicit gradient transport and a lookahead policy parameter | Policy gradient for Fisher-non-degenerate policies · Fisher-non-degenerate policies | global sample complexity: $\tilde{\mathcal{O}}(\varepsilon^{-3})$ -> $\tilde{\mathcal{O}}(\varepsilon^{-2.5})$ | Helped |
| Use recursive gradient variance reduction with a Hessian-vector correction | Policy gradient for Fisher-non-degenerate policies · Fisher-non-degenerate policies | global sample complexity: $\tilde{\mathcal{O}}(\varepsilon^{-3})$ -> $\tilde{\mathcal{O}}(\varepsilon^{-2})$ | Helped |
| Use N-PG-IGT rather than Vanilla-PG | MuJoCo Humanoid · Humanoid | average reward | Helped |
| Use HARPG or N-HARPG rather than Vanilla-PG with tuned initial step-sizes | MuJoCo Humanoid · Humanoid | average reward | Hurt |
| Use HARPG rather than Vanilla-PG at small initial step-sizes | MuJoCo Reacher · Reacher | system probes to average reward above -11 | Helped |
| Use N-PG-IGT rather than the other tested methods | MuJoCo Reacher · Reacher | robustness to initial step-size | Helped |
| Use HARPG rather than Vanilla-PG with tuned step-sizes | MuJoCo Walker2d and Reacher · Walker2d and Reacher | average reward | Helped |
| Use HARPG rather than Vanilla-PG at small, untuned step-sizes | MuJoCo Reacher and Hopper · Reacher and Hopper | average reward | Helped |
| Use VR-PG or N-VR-PG rather than HARPG or N-HARPG | MuJoCo Hopper and Halfcheetah · Hopper and Halfcheetah | average reward | Mixed |
| Use N-PG-IGT rather than other tested methods | MuJoCo Hopper and Halfcheetah · Hopper and Halfcheetah | convergence speed | Helped |
| Use HARPG rather than Vanilla-PG at small initial step-sizes | MuJoCo Walker2d and Humanoid · Walker2d and Humanoid | convergence speed | Helped |
| Use N-PG-IGT rather than other tested methods | MuJoCo Walker2d · Walker2d | average reward | Helped |
Closest papers by meaning
- An Improved Analysis of (Variance-Reduced) Policy Gradient and Natural Policy Gradient Methods (NeurIPS 2020 · similarity 0.92)
- Convergence Analysis of Policy Gradient Methods with Dynamic Stochasticity (ICML 2025 · similarity 0.91)
- Learning Optimal Deterministic Policies with Stochastic Policy Gradients (ICML 2024 · similarity 0.91)
- Deterministic Policy Gradient Primal-Dual Methods for Continuous-Space Constrained MDPs (AAAI 2025 · similarity 0.90)
- Reusing Trajectories in Policy Gradients Enables Fast Convergence (ICML 2026 · similarity 0.90)
- Accelerated Policy Gradient for s-rectangular Robust MDPs with Large State Spaces (ICML 2024 · similarity 0.88)
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.