Post by Azherae on 2024-01-25 22:52:22
We have not received permission from Otr to share the quoted post.
The most basic thing that would have to be done first:
"Don't design stuff where the player is just reacting to things that are visible on screen without making any true decisions."
Then:
"Don't design stuff that is scripted and does the same things in approximately the same order or timings."
And then most important for Ashes:
"Don't design stuff so that player Synergies are just part of a 'rotation', even if that 'rotation' differs by which party the player is in."
'AI' is quite good at things where you have to recognize patterns and save internal states and events. It will eventually be good at complex stuff. The ones I have can play certain games, even complex state games, fairly well. But those have also taken years to work out.
Basically, AI will always be able to play, but the things listed above drastically change (for now) the amount of effort and time it takes to develop and calibrate one. Even if the game itself isn't super difficult or complicated, requiring people to 'have an intention', then 'observe a semi-random event that affects their intention', then 'take the information from the last 7-10 seconds combined with the outcome of that event, and apply it to their intention', would raise the difficulty of calibrating or training most 'basic AI' enough. It would change the task from 'ok I can just throw a basic Finite State Machine at the problem' up to 'I need thousands of hours of training data before this even starts to do the things I have decided it should do, and I have to set those goals carefully'.
This can be achieved without applying the same complexity to the game's internal 'mob AI', because it isn't a 1v1 style game with equalized parameters or any assumption of fairness.
At this point, the Tracker is just 'a memory assistant/checker', as it should be.