Reinforcement Learning (RL) is a framework for defining and solving a learning problem where an animal or agent knows or infers the state of the world and then learns the value of the states and actions that can be taken in them, by receiving a reward signal. Importantly, reinforcement learning provides formal, optimal descriptions of learning first derived from studies of animal behavior and then validated when the formal quantities used in the model were observed in the brain in humans and animals. It is probably one of the most widely used computational approaches in neuroscience.
In the following tutorials, you will learn the core concepts of reinforcement learning and explore some of the most widely used models. In tutorial 1, you will learn about how we learn the value of future states from our experience. In tutorial 2, you will learn about how to make and learn from actions and the explore-exploit dilemma. In tutorial 3, you will explore how we can efficiently learn the future value of actions from experience. Finally, in tutorial 4 you will learn how having a model of the world’s dynamics can help you to learn and act.
Reinforcement learning is a broad framework and it has deep connections to many topics covered in NMA, but the core reinforcement learning approaches define the world as a Markov Decision Problem, which is built on Hidden Dynamics and Optimal Control. Reinforcement learning, more broadly, can be seen as a framework that allows us to bring in many ideas and formalisms from other areas like economics, psychology, computer science, artificial intelligence, etc. to define algorithms or models that can solve large, complex problems with only a simple reward signal.