Research map / Reinforcement learning / Policy gradient and actor-critic
Non-Uniform Noise-to-Signal Ratio in the REINFORCE Policy-Gradient Estimator
Haoyu Han, Heng Yang
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 |
|---|---|---|---|
| decrease policy covariance sigma^2 relative to initial-state covariance (sigma0 >> sigma) (ablation) | REINFORCE estimator, one-step LQG · double integrator, one-step LQG | noise-to-signal ratio | Hurt |
| policy approaches the optimum during optimization | REINFORCE with GD, SGD, Adam · double integrator LQG, T = 30 | noise-to-signal ratio | Hurt |
| use SGD or Adam instead of exact gradient descent (ablation) | REINFORCE policy optimization · double integrator LQG | convergence stability | Hurt |
| stable versus unstable closed-loop dynamics as horizon T grows (ablation) | REINFORCE variance · four 2D linear systems with varying spectral radius rho(F) | variance: rho(F)=0.912 bounded; rho(F)=1.000 polynomial growth; rho(F)=1.118 exponential growth | Hurt |
| smaller batch size and larger learning rate (ablation) | REINFORCE with Adam · quadratic polynomial system | policy collapse | Hurt |
| use larger batch size (ablation) | REINFORCE with Adam · quadratic polynomial system | policy collapse | Mixed |
| use exact gradient instead of sampled gradient (ablation) | REINFORCE with Adam · quadratic polynomial system | policy collapse | Improved |
| approach optimum with MLP policies | REINFORCE with Adam · LQR and Pendulum with MLP policy | noise-to-signal ratio | Hurt |
| same trend in MuJoCo environments | REINFORCE · Hopper and HalfCheetah | NSR and objective fluctuations | Hurt |
Closest papers by meaning
- Does “Do Differentiable Simulators Give Better Policy Gradients?” Give Better Policy Gradients? (ICLR 2026 · similarity 0.89)
- Statistically Efficient Off-Policy Policy Gradients (ICML 2020 · similarity 0.88)
- Total stochastic gradient algorithms and applications in reinforcement learning (NeurIPS 2018 · similarity 0.88)
- On the Convergence and Sample Efficiency of Variance-Reduced Policy Gradient Method (NeurIPS 2021 · similarity 0.88)
- Implicit Bias of Policy Gradient in Linear Quadratic Control: Extrapolation to Unseen Initial States (ICML 2024 · similarity 0.87)
- Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework (NeurIPS 2026 · 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.
