Imitation Learning and Inverse RL
Learning from demonstration is often called imitation learning. In the imitation learning setting, we have expert demonstrations and train our agent to mimic those expert demonstrations. Learning from demonstrations has many benefits, including helping an agent to learn more quickly. There are several approaches to perform imitation learning, and two of them are supervised imitation learning and Inverse Reinforcement Learning (IRL).
First, we will understand how we can perform imitation learning using supervised learning, and then we will learn about an algorithm called Dataset Aggregation (DAgger). Next, we will learn how to use demonstration data in a DQN using an algorithm called Deep Q Learning from Demonstrations (DQfD).
Moving on, we will learn about IRL and how it differs from reinforcement learning. We will learn about one of the most popular IRL algorithms called maximum entropy IRL. Toward the end of the chapter, we will understand...
