Research map / Optimization / Nonconvex and smooth optimization
On the Convergence Rates of Policy Gradient Methods
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 weak gradient-mapping domination to analyze projected policy gradient | Projected policy gradient on discounted MDPs · Discounted MDP | convergence rate: O(1/sqrt(k)) -> O(1/k); (1-gamma)^-6 -> (1-gamma)^-5 | Helped |
| Use a geometrically increasing step size eta_(k+1) = eta_k/gamma | Policy mirror descent · Discounted MDP | convergence rate: linear convergence; iteration complexity 1/(1-gamma) ||d_rho(pi*)/rho||_infinity log[2/((1-gamma)epsilon)] | Helped |
| Use constant, arbitrarily large step sizes | Policy mirror descent · Discounted MDP | iteration complexity: 2/((1-gamma)^2 epsilon) | Helped |
| Use squared Euclidean divergence instead of KL divergence | Policy mirror descent · Discounted MDP | iteration complexity: 2/((1-gamma)^2 epsilon) | No clear effect |
| Set the evaluation distribution to the optimal policy's stationary distribution | Policy mirror descent · Discounted MDP with rho = rho* | iteration complexity: 1/(1-gamma) log[2/((1-gamma)epsilon)] | Helped |
| Use repeated simulation of truncated trajectories for Q estimation | Inexact policy mirror descent · Discounted MDP | sample complexity: O~(|S||A|/((1-gamma)^8 epsilon^2) ||d_rho(pi*)/rho||_infinity^3) | Helped |
| Add weak gradient-mapping domination to the proximal-gradient analysis | General composite optimization · Smooth nonconvex composite optimization | convergence rate: O(1/k) | Helped |
Closest papers by meaning
- Convergence and Optimality of Policy Gradient Methods in Weakly Smooth Settings (AAAI 2022 · similarity 0.90)
- On the Global Convergence Rates of Softmax Policy Gradient Methods (ICML 2020 · similarity 0.88)
- Linear Convergence of Natural Policy Gradient Methods with Log-Linear Policies (ICLR 2023 · similarity 0.88)
- Convergence of Policy Mirror Descent Beyond Compatible Function Approximation (ICML 2025 · similarity 0.87)
- Convergence Analysis of Policy Gradient Methods with Dynamic Stochasticity (ICML 2025 · similarity 0.86)
- Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework (NeurIPS 2026 · similarity 0.84)
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.