Removing the VAE Network: Is It Possible?

[july000]

Hi,
When I run it, I have some questions:
Why is a Variational Autoencoder (VAE) network needed? Is it necessary, or can it be removed?

[shayan223]

A quick, and perhaps over-simplified answer to your question, based on my own understanding:

We need to make a model that creates a simplified/condensed representation of the image of the car, so the PPO agent can make a decision and not take too long to train, as raw images simply contain too much data and are too specific. The VAE is one method to create a latent representation of what the car is seeing, and passes that condensed vector to the PPO for its decision making. One could, for example, swap that out for a CNN or anything else that can learn latent variables regarding an image, but its not likely to work nearly as well if you put raw images into the PPO agent.