Summary

In this chapter, we took a closer look at the Unity PPO trainer. This training model, originally developed at OpenAI, is the current advanced model, and was our focus for starting to build more complex training scenarios. We first revisited the GridWorld example to understand what happens when training goes wrong. From there, we looked at some examples of situations where training is performing sub par, and we learned how to fix some of those issues. Then, we learned how an agent can use visual observations as input into our model, providing the data is processed first. We learned that an agent using visual observation required the use of CNN layers to process and extract features from images. After that, we looked at the value of using experience replay in order to further generalize our models. This taught us that experience and memory were valuable to an agent&apos...