Abstract
In this lecture, I will introduce model-based reinforcement learning (RL). I will begin by laying the foundations of model-free RL and then define what constitutes a “model” in the context of model-based RL. We will then examine the different ways we can use these “models”, specifically comparing background planning and decision-time planning.
The core of the lecture will focus on decision-time planning strategies within continuous action spaces. I will provide insights into the sources of uncertainty inherent in model-based RL and discuss methods for making decisions under this uncertainty.
This lecture aims to provide a good understanding of decision-time planning in model-based RL and offer insights into the rationale and strategies for making decisions under uncertainty.