Research map / Reinforcement learning / Visual and representation RL
VariBAD: Variational Bayes-Adaptive Deep RL via Meta-Learning
Luisa Zintgraf, Sebastian Schulze, Cong Lu, Leo Feng, Maximilian Igl, Kyriacos Shiarlis, Yarin Gal, Katja Hofmann, Shimon Whiteson
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 variBAD instead of RL2 (best baseline) | meta-RL · Meta-World ML1 Push | success rate (%): 87 -> 100 | Improved |
| use variBAD instead of MAML | meta-RL · Meta-World ML1 Reach | success rate (%): 48 -> 100 | Improved |
| use variBAD instead of PEARL | meta-RL · Meta-World ML1 Pick-Place | success rate (%): 28 -> 29 | No clear effect |
| allow second episode to adapt | variBAD · Meta-World ML1 | success rate (%): 100 -> 100 | No clear effect |
| add decoding of both past and future trajectory (vs past only or next step only) (ablation) | variBAD VAE objective · GridWorld, AntGoal, PointRobot | return | Improved |
| decode full trajectory instead of future only (ablation) | variBAD VAE objective · GridWorld | belief quality | Improved |
| use filtering KL (prior is previous posterior) instead of fixed N(0,I) prior (ablation) | variBAD KL term · PointRobot | return | Improved |
| add KL regularisation (ablation) | variBAD KL term · GridWorld, AntGoal, PointRobot | return | Improved |
| do not detach gradient of previous posterior in KL (ablation) | variBAD KL term · GridWorld, AntGoal, PointRobot | return | Improved |
| increase latent dimension from 1-2 to 5 or more (ablation) | variBAD · GridWorld, AntGoal | normalised performance | Improved |
| use very large latent (1000 GridWorld, 300 AntGoal) (ablation) | variBAD · GridWorld, AntGoal | normalised performance | No clear effect |
| use variBAD instead of RL2 on MuJoCo | meta-RL · MuJoCo (5 of 6 environments) | return in first episode | Improved |
Closest papers by meaning
- VariBAD: A Very Good Method for Bayes-Adaptive Deep RL via Meta-Learning (ICLR 2020 · similarity 0.98)
- Learning Task Belief Similarity with Latent Dynamics for Meta-Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Predictive Coding Enhances Meta-RL To Achieve Interpretable Bayes-Optimal Belief Representation Under Partial Observability (NeurIPS 2025 · similarity 0.87)
- Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- DynaMITE-RL: A Dynamic Model for Improved Temporal Meta-Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Bayesian Online Natural Gradient (BONG) (NeurIPS 2024 · similarity 0.85)
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.
