Everything what you said is mostly correct. Though, I’d disagree on the limitations because that would be based on outdated knowledge.
I’ll address this one by one.
huge amounts of training, especially with something like piloting a plane, potentially exponentially more than tasks like walking and sword fighting
This is correct. However, recent research in deep reinforcement learning kind of already solved quite a number of issues related to long training times.
For example, instead of going full deep reinforcement learning approach where it would make training very long, you could initially train your AI through deep learning approach first with small amount of collected data. This would give a baseline performance, potentially cutting down 10% - 50% for training time. Then you refine this model through deep reinforcement learning, so that it can really learn on how to fly the plane with baseline performance.
Not to mention, researchers still kept increasing deep reinforcement learning training speed through creating new algorithms. I was having a hard time trying to catch up with these researchers and put them into my DataPredict library so that people can train their AIs faster without needing to know the underlying maths. The most recent ones have maths are a headache for me to follow, and so my current list of algorithms that I have are pretty recent, but not too recent.
Additionally, with all that training, the data built up would likely break the new limits on datastore usage
In deep reinforcement learning, we don’t store data to an online storage, but instead gets stored to an AI’s “memories”. What you need is the AI’s neural network’s weights. Those can be stored more easily and doesn’t even need a lot of storage. Believe me, I have done this as many times as before in my own projects.
To be honest, I don’t think it would hurt to just try. After all, deep reinforcement learning AIs does make a game stand out because the other games doesn’t have it and might bring in some curious players.
I can also recommend other libraries if you’re not comfortable with my library.