Research map / Reinforcement learning / Policy gradient and actor-critic
Meta-Gradient Reinforcement Learning
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 |
|---|---|---|---|
| meta-learn γ instead of keeping it fixed | IMPALA · Atari 2600, human starts, γ'=0.99 | median human-normalised score: 144.4% -> 233.2% | Helped |
| meta-learn γ instead of keeping it fixed | IMPALA · Atari 2600, human starts, γ'=0.995 | median human-normalised score: 211.9% -> 267.9% | Helped |
| meta-learn γ instead of keeping it fixed | IMPALA · Atari 2600, no-op starts, γ'=0.99 | median human-normalised score: 191.8% -> 280.9% | Helped |
| meta-learn γ instead of keeping it fixed | IMPALA · Atari 2600, no-op starts, γ'=0.995 | median human-normalised score: 257.1% -> 275.5% | Helped |
| meta-learn λ while keeping γ=0.99 | IMPALA · Atari 2600, human starts | median human-normalised score: 144.4% -> 156.6% | Helped |
| meta-learn λ while keeping γ=0.99 | IMPALA · Atari 2600, no-op starts | median human-normalised score: 191.8% -> 185.5% | Hurt |
| meta-learn λ while keeping γ=0.995 | IMPALA · Atari 2600, human starts | median human-normalised score: 211.9% -> 214.2% | No clear effect |
| meta-learn λ while keeping γ=0.995 | IMPALA · Atari 2600, no-op starts | median human-normalised score: 257.1% -> 246.5% | Hurt |
| meta-learn γ and λ instead of keeping both fixed | IMPALA · Atari 2600, human starts, γ'=0.995 | median human-normalised score: 211.9% -> 292.9% | Helped |
| meta-learn γ and λ instead of keeping both fixed | IMPALA · Atari 2600, no-op starts, γ'=0.995 | median human-normalised score: 257.1% -> 287.6% | Helped |
| add γ embedding as input to policy and value functions (ablation) | Meta-gradient IMPALA adapting γ · Atari 2600, human starts, γ'=0.995 | median human-normalised score: 183% -> 267.9% | Helped |
| increase fixed γ from 0.99 to 0.995 (ablation) | IMPALA · Atari 2600, human starts | median human-normalised score: 144.4% -> 211.9% | Helped |
Closest papers by meaning
- Distributional Meta-Gradient Reinforcement Learning (ICLR 2023 · similarity 0.92)
- Evolving Reinforcement Learning Algorithms (ICLR 2021 · similarity 0.90)
- Meta-Gradient Reinforcement Learning with an Objective Discovered Online (NeurIPS 2020 · similarity 0.89)
- A Theoretical Understanding of Gradient Bias in Meta-Reinforcement Learning (NeurIPS 2022 · similarity 0.89)
- ProMP: Proximal Meta-Policy Search (ICLR 2019 · similarity 0.88)
- Reflect-then-Correct: Rebalancing Task Optimization for Generalizable Meta-Reinforcement Learning via Distributional Value Error Reduction (ICML 2026 · similarity 0.88)
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.