METTA adapts meta-RL agents with two-scale neural memory
Original title: METTA: Memory-Enhanced Test-Time Adaptation for Meta-Reinforcement Learning
On held-out Meta-World ML10/45 environments, METTA reports a success rate over 50% at a low training budget.
The paper proposes Memory-Enhanced Test-Time Adaptation (METTA) for meta-reinforcement learning. A causal neural memory handles within-episode adaptation at linear sequence cost, while compact episode-level summaries support cross-episode task inference without attending over the full history. On unseen Meta-World ML10/45 environments the authors report a success rate over 50%. On long-context sparse-reward PointMaze tasks, METTA is said to outperform transformer-based baselines with more structured exploration.
- METTA splits adaptation into within-episode memory and episode-level summaries.
- Success rate is reported above 50% on held-out Meta-World ML10/45 tests.
- Neural memory is said to enable test-time training on unseen task families.
- PointMaze results are compared with transformer-based baselines.
From the original article
1 min read at arxiv.orgMeta-reinforcement learning aims to train agents that adapt quickly to unseen tasks from limited interaction, but this requires representing task evidence over long interaction histories. Classical recurrent agents compress experience into hidden states, while transformer-based agents retrieve from past trajectories at quadratic memory and computation cost.
By Ali Najar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.