Fundamentals of Deep Learning

Chapter 9. Deep Reinforcement Learning

Nicholas Locascio¹

In this chapter, we’ll discuss reinforcement learning, which is a branch of machine learning that deals with learning via interaction and feedback. Reinforcement learning is essential to building agents that can not only perceive and interpret the world, but also take action and interact with it. We will discuss how to incorporate deep neural networks into the framework of reinforcement learning and discuss recent advances and improvements in this field.

Deep Reinforcement Learning Masters Atari Games

The application of deep neural networks to reinforcement learning had a major breakthrough in 2014, when the London startup  DeepMind astonished the machine learning community by unveiling a deep neural network that could learn to play Atari games with superhuman skill. This network, termed a Deep Q-Network (DQN) was the first large-scale successful application of reinforcement learning with deep neural networks. DQN was so remarkable because the same architecture, without any changes, was capable of learning 49 different Atari games, despite each game having different rules, goals, and gameplay structure. To accomplish this feat, DeepMind brought together many traditional ideas in reinforcement learning while also developing a few novel techniques that proved key to DQN’s success. Later in this chapter we will implement DQN, as it is described in the Nature paper “Human-level control through deep reinforcement learning.” ...