Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Bootstrapped Meta-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 |
|---|---|---|---|
| Replace the STACX meta-objective with BMG policy matching (KL), L=1 | IMPALA actor-critic on Atari ALE · Atari ALE, 57 games | median HNS: ~350% -> ~500% | Helped |
| Add value matching with 0.25 weight to policy matching (ablation) | BMG on Atari ALE · Atari ALE, 57 games | median HNS: ~500% -> ~520% | Helped |
| Extend the bootstrap horizon to L=4 and adjust gradient clipping from .3 to .2 (ablation) | BMG on Atari ALE · Atari ALE, 57 games | median HNS: ~520% -> ~610% | Helped |
| Use RMSProp in the target bootstrap for curvature correction (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance gains: roughly 45% | Helped |
| Use a matching function to control curvature (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance gains: roughly 25% | Helped |
| Increase bootstrap horizon L (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance gains: about 30% | Helped |
| Use joint policy and value matching instead of value matching alone (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance | Helped |
| Use only value matching (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance | Hurt |
| Reverse the KL direction (ablation) | BMG on Atari ALE · Atari ALE, 57 games | performance | Hurt |
| Increase bootstrap horizon L (ablation) | BMG on two-colors actor-critic · Two-colors grid world | total return | Helped |
| Add entropy regularization to the Lth target update (ablation) | BMG on two-colors actor-critic · Two-colors grid world | total return | No clear effect |
| Use the meta-learned update rule for all target bootstrap steps instead of a final unregularized policy-gradient step (ablation) | BMG on two-colors actor-critic · Two-colors grid world | total return | Hurt |
Closest papers by meaning
- Meta-Gradient Reinforcement Learning with an Objective Discovered Online (NeurIPS 2020 · similarity 0.91)
- Distributional Meta-Gradient Reinforcement Learning (ICLR 2023 · similarity 0.89)
- MetaCURE: Meta Reinforcement Learning with Empowerment-Driven Exploration (ICML 2021 · similarity 0.89)
- Skill-based Meta-Reinforcement Learning (ICLR 2022 · similarity 0.89)
- Enhanced Meta Reinforcement Learning via Demonstrations in Sparse Reward Environments (NeurIPS 2022 · similarity 0.89)
- Can Learned Optimization Make Reinforcement Learning Less Difficult? (NeurIPS 2024 · 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.