Research map / Reinforcement learning / Policy gradient and actor-critic
PPG Reloaded: An Empirical Study on What Matters in Phasic Policy Gradient
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 |
|---|---|---|---|
| increase policy regularization strength as distillation becomes more frequent (T_freq down to 1) (ablation) | PPG · Procgen (16 games) | mean normalized return | Helped |
| distill more frequently with fixed policy regularization (beta_pi = 1) and fixed buffer size (ablation) | PPG · Procgen (16 games) | mean normalized return | Hurt |
| raise beta_pi under already infrequent distillation (T_freq = 32) (ablation) | PPG · Procgen (16 games) | mean normalized return | No clear effect |
| reduce buffer size T_off from 32 to 1 (on-policy data only) (ablation) | PPG · Procgen (16 games) | mean normalized return | Hurt |
| use higher T_off under frequent distillation (T_freq = 1, beta_pi = 16) (ablation) | PPG · Procgen (16 games) | mean normalized return | Helped |
| increase number of parallel environments with on-policy data only (T_off = 1) (ablation) | PPG · Procgen (16 games) | mean normalized return | Helped |
| reduce data diversity (lower T_off) on test performance in generalization setups (ablation) | PPG · Procgen generalization | test return | No clear effect |
| reduce distillation minibatch size with fixed number of updates K_V (lower sample reuse) (ablation) | PPG · Procgen (16 games) | mean normalized return | No clear effect |
| reduce minibatch size to 1/8 (M_V = 128) (ablation) | PPG · Procgen generalization (easy) | mean normalized return | Hurt |
| reduce minibatch size to 1/8 in sample efficiency setup (ablation) | PPG · Procgen sample efficiency (hard) | mean normalized return | Helped |
| scale distillation minibatch size to 1/8 of default (ablation) | PPG · Procgen sample efficiency (hard) | wall-clock time per iteration: 20.4 -> 12.5 | Helped |
| scale distillation minibatch size to 1/8 of default (ablation) | PPG · Procgen generalization (easy) | wall-clock time per iteration: 6.2 -> 3.9 | Helped |
Closest papers by meaning
- Phasic Policy Gradient (ICML 2021 · similarity 0.93)
- SAPG: Split and Aggregate Policy Gradients (ICML 2024 · similarity 0.87)
- Accelerated Policy Gradient: On the Convergence Rates of the Nesterov Momentum for Reinforcement Learning (ICML 2024 · similarity 0.87)
- On the Global Optimality of Policy Gradient Methods in General Utility Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Provable Zero-Shot Generalization in Offline Reinforcement Learning (ICML 2025 · similarity 0.87)
- Convergence Analysis of Policy Gradient Methods with Dynamic Stochasticity (ICML 2025 · 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.