Meta-Reinforcement Learning
Overview
Meta-reinforcement learning studies how an agent can use experience across many related tasks to learn a new task with little additional interaction. Common approaches include gradient-based methods such as MAML, recurrent or context-conditioned policies that infer tasks from interaction histories, and Bayesian or world-model methods that represent task uncertainty and predict outcomes. Researchers also use hierarchical skills, population-based search, and language or reflection signals to support adaptation in settings from robot control to multi-agent games and language agents. A central challenge is making adaptation generalize to unfamiliar task distributions while remaining data-efficient and safe, especially when task context is ambiguous or learned models are unreliable.
Latest papers
44 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.