Rewards are a fundamental aspect of reinforcement learning, and the concept is easy to grasp. After all, we partly teach and train others—dogs and children, for instance—with reinforcement through rewards. The concept of implementing rewards or a reward function in a simulation can be somewhat difficult, and prone to a lot of trial and error. This is the reason for waiting until a later and more advanced chapter to talk about rewards, building reward functions, and reward assistance methods such as Curriculum Learning, Backplay, Curiosity Learning, and Imitation Learning / Behavioral Cloning.
Here is a quick summary of the concepts we will cover in this chapter:
- Rewards and reward functions
- Sparsity of rewards
- Curriculum Learning
- Understanding Backplay
- Curiosity Learning
While this is an advanced chapter, it is also an essential...