In the past, most reinforcement learning agents were developed to solve existing domains. At first, a game was given, for example the snake game, and after a bit of training, the neural network was able to solve the game reasonable well. Unfortunately, this is not exactly needed in reality, but it's more a demonstration of a synthetic benchmark of how to create agents which can learn a policy.
The more interesting problem is called “symbolic grounding” which is a bit harder to understand. According to a paper 1, symbol grounding has to do with the domain modelling. Not the AI agent is under investigation, if he plays a game perfect, but the question is, if the game simulates another game reasonable well.
Question: Is symbol grounding equal to build a physics engine with the help of neural networks?