ML can be applied with relative ease to specific AI routines with the current rules. Not sure what that has to do with programming the entire civ4 AI via ML, which obviously makes no sense right now.
My point was that ML can not work unless you are able to close the feedback loop between the parameters under ML optimisation and the output the resulting parameterised algorithms produce, and do so in a way that tracks with how we want the AI to behave in the game. We cannot just pick and choose a subroutine like you suggest and then judge its performance on some overall metric like AI success. You would have to find a heuristic that measures the actual success of e.g. city placement, which is an arbitrary choice that also relies on expert knowledge. And we do not have such a heuristic right now, at least I do not see an obvious one. ML is all about this kind of problem, and it's not something you can delegate to an algorithm. You have to set the constraints of ML yourself.
I think you read on my answer what you wanted to read, instead of my words...
I think you're taking this much too personally. I do not even disagree with your post, my point was that it was too vague and superficial and non-specific to be in any way a useful addition to the discussion. "Have you tried ML?" is just one notch above "have you tried algorithms" in its abstractness.
My concerns were about your proposed AI routine based on players strategy being the best. Why? Players doing that doesn't mean it's the best strategy, neither better than the current one! Then on the next post you change your mind about it because of f1rpo answer, but we have no data to prove any of those statements (even if expert players usually follow the best strategies)
Obviously not, but the "best strategy" is your goalpost and imo not a very useful one. It's not relevant if player strategies are "best strategy", we know enough to say that player strategies are superior to current AI strategies, so it's useful to discuss what we can learn from player strategies to improve the existing AI. That's why I think lofty appeals to having ML solve all our problems are unhelpful distractions. Realistically, the AI is and will be largely an expert system, and the discussion should be if the "expert" knowledge should be contained to Soren Johnson's personal experience from probably MP sessions inside Firaxis, or the collective experience of the much larger community over a much longer time period. Neither is inherently flawless or even better, but I don't think it's any reach at all that this amount of communication and iteration has resulted in some improvement.
My paragraph you are focusing on was intended as an illustration of that fact, and a contrast of the first principles of how common player strategies approach the problem vs how the AI approaches the problem, and even makes mention of the fact that it elides optimisations the AI could make and how those strategies are tailed to a more macro level human perspective.
Lastly, I do not "change my mind" in the next post. My whole goal here was to have a discussion and so obviously part of that is to learn new information and to take that into account when forming my opinion. That's especially true when it comes from f1rpo's perspective because I trust his expertise on the subject (as clearly exceeding mine) and his point was actually substantiated. I am here as much to learn as I am to argue for my point of view.
I find it telling that this eludes you while you are also so defensive of your own post.
That's again a statement without any proven base.
How about you lead with proof, preferably proof of concept, or at least substance? Everything you've said so far are superficial theoretical concepts and abstract ideas. You do not get to ask others for proof when you're not willing to put in any effort of your own. Have you even seen the Civ4 AI code as it is?
But, in any case, (as someone not well familiar with neural networks) I'd worry that an (under-)trained network would frequently make needlessly puzzling decisions or near-optimal decisions when the optimal decision should be obvious.
Yes, I would also be extremely worried about this behaviour of ML trained AI opponents even if their training was feasible.
I've been following the development of AlphaStar (the Google deep mind Starcraft 2 AI) quite closely over the last few years, and it's really interesting to watch both as a fan of Starcraft and artificial intelligence. It was hilarious but also illuminating to see the AI choose strategies that no human ever would, and whose efficacy is still actively debated among the community. Most notably though is that the AI would constantly abuse its benefits of being an AI (perfect control, global awareness, massive multitasking) in ways that made some people question if it is even desirable to play against an opponent that "feels" like that. It's an entirely different experience.
I think that's important because more than in Starcraft, role playing is an important element of the Civ4 AI. Weird but effective may be acceptable (to some) in a competitive RTS, but here we also expect the AI to follow patterns of decision that are parseable to the human mind.
For finding fair starting positions, some kind of intelligent algorithm could indeed be useful. Computing a measure of fairness for a given set of starting locations doesn't seem too challenging, but it's infeasible to do that for every possible assignment of players to potential starting locations. That said, I don't think I'd want to deal with a library for genetic programming or reinforcement learning for this; probably some ad-hoc heuristic search would do.
At least that's a sufficiently self contained algorithm to attempt ML on, yeah.