Research map / Reinforcement learning / Policy gradient and actor-critic
Adaptive Batch Size for Safe Policy Gradients
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 |
|---|---|---|---|
| G(PO)MDP instead of REINFORCE as gradient estimator | adaptive step and batch size method (Chebyshev bound, δ=0.95) · LQG | Υ (average performance): about 0.72 higher (-10.6085 vs -11.3266) | Helped |
| Empirical range instead of theoretical range in Hoeffding bound (ablation) | adaptive batch size with Hoeffding bound · LQG | Υ (average performance): from -11.914 to -9.8582 | Helped |
| Empirical Bernstein bound instead of Chebyshev bound (ablation) | adaptive batch size method with G(PO)MDP · LQG | Υ (average performance): from -10.6085 (Chebyshev, δ=0.95) to -10.2159 | Helped |
| Empirical range instead of theoretical range in Bernstein bound (ablation) | adaptive batch size with empirical Bernstein bound · LQG | Υ (average performance): from -10.2159 to -9.6623 | Helped |
| Larger δ (allowing more probable worsening updates) (ablation) | adaptive batch size with Chebyshev bound and G(PO)MDP · LQG | Υ (average performance): from -11.836 (δ=0.05) to -10.6085 (δ=0.95) | Helped |
| Theoretical Hoeffding bound instead of Chebyshev bound (ablation) | adaptive batch size with G(PO)MDP · LQG | Υ (average performance): from -10.6085 (Chebyshev, δ=0.95) to -11.914 | Hurt |
| Fixed small batch size combined with large step size (no adaptive batch size) (ablation) | policy gradient updates with fixed step and batch sizes · LQG (σ=0.5 and σ=1) | improvement ratio: e.g. 49.79% at N=100 vs 95.96% at N=10000 with α=1e-3 (σ=0.5) | Hurt |
Closest papers by meaning
- Provably Good Batch Off-Policy Reinforcement Learning Without Great Exploration (NeurIPS 2020 · similarity 0.90)
- Deep Policy Gradient Methods Without Batch Updates, Target Networks, or Replay Buffers (NeurIPS 2024 · similarity 0.88)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.87)
- Learning Optimal Deterministic Policies with Stochastic Policy Gradients (ICML 2024 · similarity 0.87)
- Convergence Analysis of Policy Gradient Methods with Dynamic Stochasticity (ICML 2025 · similarity 0.87)
- Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
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.