Introduction

In the previous chapter, we covered the theory behind Reinforcement Learning (RL), explaining topics such as Markov chains and Markov Decision Processes (MDPs), Bellman equations, and a number of techniques we can use to solve MDPs. In this chapter, we will be looking at deep learning methods, all of which will play a primary role in building approximate functions for reinforcement learning. Specifically, we will look at different families of deep neural networks: fully connected, convolutional, and recurrent networks. These algorithms have the key capability of encoding knowledge that's been learned through examples in a compact and effective representation. In RL, they are typically used to approximate the so-called policy functions and value functions, which encode how the RL agent chooses its action, given the current state and the value associated with the current state, respectively. We will study the policy and value functions in the upcoming chapters.

Data...