Research map / Reinforcement learning / Policy gradient and actor-critic
Model-Ensemble Trust-Region Policy Optimization
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 ME-TRPO instead of model-free deep RL | continuous control benchmarks · Mujoco continuous control tasks | sample complexity: ≈ 100$\times$ less data | Helped |
| Replace BPTT with TRPO for policy optimization (ablation) | vanilla model-based RL with one learned model · continuous control domains | policy performance | Helped |
| Use a model ensemble rather than a single dynamics model (ablation) | TRPO-based model-based RL · continuous control domains | policy performance | Helped |
| Sample each imagined step from a randomly selected ensemble model (ablation) | ME-TRPO with 5 models · continuous control environments | policy performance | Helped |
| Use the mean or median of ensemble predictions instead of random model sampling (ablation) | ME-TRPO with 5 models · Half Cheetah | policy performance | Hurt |
| Average gradients across fixed-model rollouts instead of sampling a different model at each step (ablation) | ME-TRPO with 5 models · continuous control environments | resistance to overfitting | Hurt |
| Validate policy improvement across an ensemble rather than using one model (ablation) | ME-TRPO · continuous control environments | policy performance | Helped |
| Use real-performance oracle validation instead of ensemble validation (ablation) | ME-TRPO · continuous control environments | policy performance | Hurt |
| Stop policy training after a fixed number of updates instead of ensemble validation (ablation) | ME-TRPO · continuous control environments | policy performance | Mixed |
| Use ensemble validation rather than a fixed update cutoff (ablation) | ME-TRPO · continuous control environments | policy performance | Helped |
Closest papers by meaning
- Keep Various Trajectories: Promoting Exploration of Ensemble Policies in Continuous Control (NeurIPS 2023 · similarity 0.90)
- Differentiable Trust Region Layers for Deep Reinforcement Learning (ICLR 2021 · similarity 0.88)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models (NeurIPS 2018 · similarity 0.88)
- When to Trust Your Model: Model-Based Policy Optimization (NeurIPS 2019 · similarity 0.88)
- PWM: Policy Learning with Multi-Task World Models (ICLR 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.