I don't quite understand

So, the ML-Agents are mostly used for creating AI, before building the game?

What if I want to make a game based on learning, for example, each level AI will learn and update its behavior. Is this possible with the current ML-Agents setup?

 

Is the real-time (in-game build) learning solution planned though? Can't find the roadmap for Machine Learning. If not, still good job, really appreciated your work on ML-Agents.

 

The objective of this is to learn to an agent to do something. For example, hunt the human, kill him, or just simulations. Currently, as far I know, there is not a "real-time learning" on machine learning problems, because learning is just a phase and it is computationally expensive.

 

That is unfortunate, it would be great if, for example, you can train AI each time you make a run in a roguelike game or something like that. This will also lead to a completely different from player to player experience.

 

Yes that is an a limitation of AI right now, but its pretty good as far.

 

Well you can imitate that experience by training multiple Nn models. For each level you can just change the more trained AI and it will basically be better than the previous. Therefore users can think their player is really evolving. However personalizing it cannot be imitate easily. But one can do that with increase the number of trained Nn models.