Research map / Reinforcement learning / Policy gradient and actor-critic
GIPO: Gaussian Importance Sampling Policy Optimization
Chengxuan Lu, Zhenquan Zhang, Shukuan Wang, Qunzhi Lin, Baigui Sun, Yang Liu
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 |
|---|---|---|---|
| replace PPO hard clipping with Gaussian log-ratio trust weight (GIPO) | PPO-style actor-critic with stale replay · Meta-World (10 tasks, stale regime) | IQM normalized score: 0.180 -> 0.730 | Improved |
| use symmetric Gaussian weight instead of SAPO smooth weighting | GIPO · Meta-World (10 tasks, stale regime) | IQM normalized score: 0.412 -> 0.730 | Improved |
| use different sigma for positive and negative advantages (advantage-aware) (ablation) | GIPO · Meta-World (10 tasks, stale regime) | IQM normalized score: 0.730 (1.0,1.0) vs 0.543 (1.0,0.5) | No clear effect |
| use advantage-aware sigma on MuJoCo (ablation) | GIPO · MuJoCo (10 tasks) | IQM normalized score: 0.606 -> 0.602 | No clear effect |
| advantage-aware sigma (0.2, 0.1) instead of (0.2, 0.2) (ablation) | GIPO · MuJoCo Hopper | average return: 786.9 -> 1184.1 | Improved |
| increase sigma from 0.5 to 1.0 (weaker damping) (ablation) | GIPO · MuJoCo (10 tasks) | IQM normalized score: 0.606 -> 0.478 | Hurt |
| increase sigma from 0.2 to 1.0 on Ant (ablation) | GIPO · MuJoCo Ant | average return: 977.1 -> 20.4 | Hurt |
| best sigma depends on setting (1.0 on Meta-World stale, 0.5 on MuJoCo) (ablation) | GIPO · Meta-World vs MuJoCo | IQM normalized score: Meta-World best 0.730 at sigma 1.0; MuJoCo 0.478 at sigma 1.0 | Mixed |
| use GIPO instead of PPO on Classic Control | PPO · Classic Control (4 tasks) | IQM normalized score: 0.7908 -> 0.8095 | Improved |
| use GIPO (0.2) instead of PPO on Pendulum | PPO · Pendulum-v1 | final rollout average return: -215.0 -> -276.8 | Hurt |
| use GIPO in fresh data regime | PPO-Clip and SAPO · Meta-World (fresh regime) | episodic return | No clear effect |
| use GIPO with 3 actors per trainer (staler data) | PPO and SAPO · LIBERO (OpenVLA-OFT 7B) | average success rate | Improved |
Closest papers by meaning
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.88)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.84)
- Truncated Gaussian Policy for Debiased Continuous Control (AAAI 2025 · similarity 0.84)
- SeMOPO: Learning High-quality Model and Policy from Low-quality Offline Visual Datasets (ICML 2024 · similarity 0.84)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.84)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · 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.
