Chapter 12: Meta-Reinforcement Learning

Humans learn new skills from much fewer data compared to a reinforcement learning agent. Two factors contributing to this are, first, we come with priors in our brains at birth that give us certain capabilities from the get-go; and second, we are able to transfer our knowledge from one skill to another quite efficiently and adapt to new environments fast. Meta-reinforcement learning aims to achieve a similar capability for artificial intelligence agents. In this chapter, we describe what meta-reinforcement learning is, the approaches it uses, and the challenges it faces. Specifically, we cover the following topics:

• Introducing meta-reinforcement learning
• Meta-reinforcement learning with recurrent policies
• Gradient-based meta-reinforcement learning
• Meta-reinforcement learning as partially observed reinforcement learning
• Challenges in meta-reinforcement learning