Foundations of Deep Reinforcement Learning: Theory and Practice in Python

12. Network Architectures

Neural networks are a component of every algorithm we have discussed in this book. However, up to this point, we have never discussed in detail the design of these networks, nor their functionality that is useful when combining neural networks with RL. The purpose of this chapter is to take a closer look at neural network design and training in the context of deep RL.

The chapter begins with a brief introduction to the different neural network families and the types of data they are specialized to process. Then we discuss how to select an appropriate network based on two characteristics of an environment—how observable it is and what the nature of the state space is. To characterize the observability of an environment, ...