Introduction

In the previous chapter, we learned about the Multi-Armed Bandit (MAB) problem – a popular sequential decision-making problem that aims to maximize your reward when playing on the slot machines in a casino. In this chapter, we will combine deep learning techniques with a popular Reinforcement Learning (RL) technique called Q learning. Put simply, Q learning is an RL algorithm that decides the best action to be taken by an agent for maximum rewards. The "Q" in Q learning represents the quality of the action that is used to gain future rewards. In many RL environments, we may not have state transition dynamics (that is, the probability of going from one state to another), or it is too complex to gather state transition dynamics. In these complex RL environments, we can use the Q learning approach to implement RL.

In this chapter, we will start by understanding the very basics of deep learning, such as what a perceptron and a gradient descent are and what...