Meta-RL (Meta-Reinforcement Learning) is the application of meta-learning to reinforcement learning — training an agent on a distribution of tasks so that it can rapidly adapt to new, unseen tasks with very little experience, effectively "learning to learn" optimal policies.

Meta-RL Approaches

• Recurrent: Train an RNN policy across task episodes — the hidden state encodes task information (RL², SNAIL).
• Gradient-Based: Use MAML to learn an initialization that adapts quickly to new tasks with few gradient steps.
• Context-Based: Learn a task encoder that infers the task from experience and conditions the policy.
• Hypernetwork: Generate task-specific policy parameters from a meta-learner.

Why It Matters

• Fast Adaptation: Meta-RL agents adapt to new tasks in a few episodes, not thousands.
• Transfer: Captures common structure across tasks — transfers to novel but related tasks.
• Semiconductor: A meta-RL agent could quickly adapt to new process conditions or product recipes.

Meta-RL is learning to learn policies — training an agent that rapidly masters new tasks by leveraging meta-knowledge from many previous tasks.