The question I have is more from interest, as I have seen a few things on the internet about people training an AI using footage of real people playing games.
I was wondering if you could train it more instantaneously using actual user input so that if the young AI encounters exactly that scenario it would do just what the player does but otherwise be quite useless. And then the idea is that you could then duel that AI and reproduce using the new config from how the player beat it the next time and do a sort of faster training cycle that reproduces how the player plays and not necessarily how to best play the game.
I don't have a very good understanding at all of NEAT or AI algorithms at large so I'm sure there's some underlying error in this but I couldn't immediately find an answer online.\
Edit to Answer nbro: 1 - by NEAT config I was meaning the sort of saved state of the neural network I suppose. The array of tolerances or whatever that actually is. 2 - I was just referring to the specificity of the training. That it would just be suited to facing that other agents policy but may not be equipped to face an agent with a differing policy. 3 - The relationship that I was trying to communicate is exactly the one you suggested I think. A neural network that represents some agent's policy. And a player is an actual person at the keyboard though whose actions could be somehow abstracted back into the same format as the non-player agent essentially giving the evolving agent something to play with programmatically.
Thank you for the questions and I apologize if this still isn't clear. Thanks again nbro.
