Research map / Reinforcement learning / Policy gradient and actor-critic
Scalable Multi-Objective and Meta Reinforcement Learning via Gradient Estimation
Zhenshuo Zhang, Minxuan Duan, Youran Ye, Hongyang R. Zhang
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 |
|---|---|---|---|
| use affinity-based grouping from first-order gradient estimation instead of random grouping (ablation) | soft modularization on MT10 · Meta-World MT10 | average success rate (%): 58.2 -> 94.0 | Improved |
| use affinity-based grouping instead of gradient-similarity grouping (ablation) | soft modularization on MT10 · Meta-World MT10 | average success rate (%): 69.6 -> 94.0 | Improved |
| group tasks with PolicyGradEx rather than training one multitask policy | soft modularization on MT10 · Meta-World MT10 | average success rate (%): 71.3 -> 94.0 | Improved |
| select meta-training tasks via PolicyGradEx groups (MAML) | MAML on CartPole · CartPole | reward after adaptation: 145.9 -> 159.2 | Improved |
| select meta-training tasks via PolicyGradEx groups instead of gradient-similarity (ablation) | MAML on LunarLander · LunarLander | reward after adaptation: 80.8 -> 82.8 | No clear effect |
| use 3 groups instead of 2 (ablation) | PolicyGradEx · Meta-World MT10 | average success rate (%): 89.5 -> 94.0 | Improved |
| use 4 groups instead of 3 (ablation) | PolicyGradEx · Meta-World MT10 | average success rate (%): 94.0 -> 95.1 | No clear effect |
| increase random projection dimension beyond 400 (ablation) | PolicyGradEx · Meta-World / control | clustering quality | No clear effect |
| increase number of groups k beyond 3 (ablation) | PolicyGradEx · control environments | success rate | No clear effect |
| increase network depth from 2 to 8 MLP layers (ablation) | PolicyGradEx · Meta-World / LunarLander | FLOP speedup: 21x -> 26x | Improved |
| let policy parameters drift farther from meta-initialization (ablation) | first-order approximation · RL environments | approximation error: <2% -> up to 10% | Hurt |
| increase subset size alpha in Meta-World groups (ablation) | multitask policy · Meta-World | generalization error | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PaCo — PaCo: Parameter-Compositional Multi-task Reinforcement Learning (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.91)
- Meta-Reinforcement Learning via Exploratory Task Clustering (AAAI 2024 · similarity 0.90)
- Projection Optimization: A General Framework for Multi-Objective and Multi-Group RLHF (ICML 2025 · similarity 0.90)
- Meta-Reinforcement Learning with Universal Policy Adaptation: Provable Near-Optimality under All-task Optimum Comparator (NeurIPS 2024 · similarity 0.90)
- On Generalization Across Environments In Multi-Objective Reinforcement Learning (ICLR 2025 · similarity 0.90)
- Preference Conditioned Multi-Objective Reinforcement Learning: Decomposed, Diversity-Driven Policy Optimization (NeurIPS 2026 · similarity 0.90)
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.
