Stanford CS224R Deep Reinforcement Learning | Spring 2025 | Lecture 13: Meta RL

The podcast discusses Meta-Reinforcement Learning (Meta-RL), contrasting it with Vanilla-RL and multitask reinforcement learning. It addresses how Meta-RL leverages experience from previous tasks to facilitate quicker learning in new tasks, using examples like operating a coffee machine or solving math problems. The lecture distinguishes Meta-RL from transfer learning, emphasizing its optimization for transferability and adaptation with small datasets. It also covers the challenges of exploration versus exploitation in Meta-RL, suggesting solutions like posterior sampling to encourage meaningful task understanding. The podcast touches on the use of Meta-RL in language models for efficient test-time compute and frames Meta-RL as a partially observed Markov decision process (POMDP).