Research map / Reinforcement learning / Policy gradient and actor-critic
Projections for Approximate Policy Iteration Algorithms
Riad Akrour, Joni Pajarinen, Jan Peters, Gerhard Neumann
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 |
|---|---|---|---|
| add projection-based extra optimization to PPO (PAPI-PPO) | PPO · RSHopper-v1 | initial 100 iteration reward: 68 ± 8 -> 121 ± 8 | Improved |
| add projection-based extra optimization to PPO | PPO · RSHalfCheetah-v1 | initial 100 iteration reward: 30 ± 2 -> 55 ± 4 | Improved |
| add extra optimization steps and projection to TRPO (PAPI-TRPO) | TRPO · RSHopper-v1 | initial reward: 87 ± 11 -> 75 ± 7 | Hurt |
| add extra L∘g optimization steps on top of entropy-constrained TRPO (PAPI-TRPO vs PAPI-0-TRPO) (ablation) | PAPI-0-TRPO · RSHalfCheetah-v1 | initial reward: 20 ± 2 -> 32 ± 4 | Improved |
| entropy-constrained TRPO without extra optimization (PAPI-0-TRPO) vs TRPO (ablation) | TRPO · RSWalker-v1 | best-window reward: 1183 ± 372 -> 1633 ± 342 | Improved |
| PAPI-PPO vs PPO | PPO · RSWalker-v1 | best-window reward: 1016 ± 359 -> 1426 ± 434 | Improved |
| PAPI-TRPO vs TRPO | TRPO · RSHalfCheetah-v1 | best-window reward: 1882 ± 172 -> 2524 ± 113 | Improved |
| PAPI-PPO vs PPO | PPO · RSAnt-v1 | best-window reward: 1871 ± 86 -> 1924 ± 110 | No clear effect |
| optimize L∘g by gradient ascent instead of projected gradient with same projection (ablation) | Projection g · linear problem with entropy constraint (100 runs) | regret and KL to optimal | Improved |
| direct projection-based policy search (ProjectionPS) vs REPS and MORE (ablation) | ProjectionPS · smooth mixture-of-25-Gaussians functions | performance at low sample counts | Improved |
| hard entropy constraint with linearly decreasing beta vs entropy bonus (ablation) | PPO / TRPO · BitFlip | policy return | Improved |
| more optimization steps of L∘g on the policy update objective (ablation) | TRPO / PPO · RoboschoolHopper policy update | objective improvement and KL | Improved |
Closest papers by meaning
- On the Theory of Policy Gradient Methods: Optimality, Approximation, and Distribution Shift (JMLR 2021 · similarity 0.88)
- Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Deterministic Policies for Constrained Reinforcement Learning in Polynomial Time (NeurIPS 2024 · similarity 0.88)
- Polynomial-Time Approximability of Constrained Reinforcement Learning (ICML 2025 · similarity 0.87)
- Natural Policy Gradient as Doubly Smoothed Policy Iteration: A Bellman-Operator Framework (NeurIPS 2026 · similarity 0.87)
- A Unifying View of Linear Function Approximation in Off-Policy RL Through Matrix Splitting and Preconditioning (NeurIPS 2025 · 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.
