Research map / Reinforcement learning / Visual and representation RL
Improving Generalization in Meta Reinforcement Learning using Learned Objectives
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 MetaGenRL's learned objective instead of RL²'s recurrent learning rule | Meta-RL trained on Cheetah & Hopper · Lunar (unseen) | mean return: -479 -> 18 | Helped |
| Use MetaGenRL's learned objective instead of EPG | Meta-RL trained on Lunar & Cheetah · Hopper (unseen) | mean return: 8 -> 2363 | Helped |
| Remove timestep input t from the learned objective (ablation) | MetaGenRL · Cheetah | training convergence speed | Hurt |
| Remove both value-estimate inputs from the learned objective (ablation) | MetaGenRL · Cheetah and Lunar | test-time training progress | Hurt |
| Remove V_{t+1} while retaining V_t (ablation) | MetaGenRL · Cheetah | mean return | Mixed |
| Use 3 inner policy-gradient steps before computing the meta-gradient instead of one (ablation) | MetaGenRL · Hopper and Cheetah | test-time training variance | Hurt |
| Use 5 inner policy-gradient steps before computing the meta-gradient (ablation) | MetaGenRL · Hopper and Cheetah | test-time performance | Hurt |
| Increase the meta-training population size (ablation) | MetaGenRL · Meta-training stability | stability | Helped |
| Train across more distinct environments (ablation) | MetaGenRL · Hopper and other meta-test environments | mean return | No clear effect |
| Use off-policy second-order meta-gradients rather than evolutionary objective search | MetaGenRL versus EPG · Continuous-control meta-training | environment-interaction sample efficiency: 1000000000 environment interactions | Helped |
Closest papers by meaning
- Can Learned Optimization Make Reinforcement Learning Less Difficult? (NeurIPS 2024 · similarity 0.89)
- Discovering Temporally-Aware Reinforcement Learning Algorithms (ICLR 2024 · similarity 0.89)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.88)
- Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity (NeurIPS 2024 · similarity 0.88)
- Meta-Reinforcement Learning with Universal Policy Adaptation: Provable Near-Optimality under All-task Optimum Comparator (NeurIPS 2024 · similarity 0.88)
- MAMBA: an Effective World Model Approach for Meta-Reinforcement Learning (ICLR 2024 · 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.