I suspect you misunderstood Johan. Until AI is super good, for sure the best version possible will be enabled at Normal difficulty with exception of special "niceness" behavior against player.
Aside from that, do a forum search on deep learning as those threads pop up fairly often. My post recent post regarding this and other modern AI approaches: https://forum.paradoxplaza.com/foru...rove-paradox-a-i.1142483/page-2#post-25030111
FWIW the earliest version of Imperator AI used deep neural network for unit AI (reinforcement w/ custom temporal credit assignment algorithm) but I replaced it because of performance issues and doubts whether functional concerns could be handled before release. Making a new game on time isn't nearly the same thing as being able to do research on a static game.
Aside from that, do a forum search on deep learning as those threads pop up fairly often. My post recent post regarding this and other modern AI approaches: https://forum.paradoxplaza.com/foru...rove-paradox-a-i.1142483/page-2#post-25030111
FWIW the earliest version of Imperator AI used deep neural network for unit AI (reinforcement w/ custom temporal credit assignment algorithm) but I replaced it because of performance issues and doubts whether functional concerns could be handled before release. Making a new game on time isn't nearly the same thing as being able to do research on a static game.
Last edited: