I'm going to preface this by saying I broadly agree (see my earlier post regarding cost of ML vs need for commercial viability above), but will focus on the more-interesting disagreements or building-on points below.
ML does have some clear advantages in this regard. Namely, it does not require any developer to be elite or even good at the game in order to create an incredibly effective AI agent.
Yes, a designer knows that spears get a bonus against horses. A designer might not necessarily recognize that in practice, the spear unit class is bad/not worth building regardless even as a countermeasure to horses in some cases, because of how that decision interacts with other mechanics/units. A designer is outright unlikely to understand how all the game's nuanced tradeoffs interact and optimize for them. In most cases, it is impractical/unrealistic for a designer to be as good at playing a game as creating it. They are a) different skill sets and b) require massive time investments, independently, to reach high levels of competency. There are some exceptions out there...IIRC the Celeste dev is pretty darned good at it. But for the most part, a dev won't be doing things like routine 200AD liberalism in Civ 4, and will be worse at looking at a set of potential balance tweaks and predicting the consequences to meta gameplay than an expert player.
ML shatters all that with brute force training. It can, and will, find the best choices to win reliably if given enough training. Better than expert players, better than the best humans in the world. It will even find things humans can't, like how it was at least somewhat variable with two factions in StarCraft 2, but spammed the heck out of roach/ravager timing pushes as Zerg vs top tier ladder competition. I don't think there was a person alive, dev or otherwise, that had solid basis to predict that in advance (those pushes almost always won on the spot too of course).
Funny things about this quote:
- In the vast majority of strategy games, including every civ game, the preferences in sentence 1 and 2 are incoherent/mutually exclusive.
- My impression is that your quote is nevertheless an accurate assessment of most players.
If an AI were to "play the same game as the player", the AI would try to win...because the vast majority of players, even if they say they emphasize role play or history or whatever else...nevertheless get upset and complain when they lose, and thus attempt to win/take steps to avoid losing, independent of role playing the game state. Going so far as to reject advice on how to play better and want the game to behave differently instead (it's kinda funny how often people blame losses in games like FTL on RNG with some kind of blind spot for how there are multiple > 100 win streaks on a higher difficulty than they play). But at the end of the day, for every 100 win streak player there are hundreds to tens of thousands of "I died to RNG in sector 4" players, and you need people to buy the game to justify making it in most cases...
This bothers me at a fundamental level. It's one aspect of a "typical player" I can't really empathize with. If an AI is not trying to win, then any "challenge" it presents becomes incidental or the consequence of otherwise-inflated bonuses that compensate whatever it's doing in lieu of that. To me, this directly cuts into the experience, and into the meaning of winning myself. Effectively, my opponents aren't trying, and I know it.
I've been around long enough and seen enough different people play/approach civ and other games to know that I'm not only a minority, but a fairly extreme one. But I just can't lie to myself like that.
I don't envy trying to design with this in mind. To try to create coherent game rules/behaviors for players with objectively incoherent preferences.
See my first response WRT ML. It's not clear to me, at all, that a highly trained ML AI would *necessarily* do any of these things (though worker/settler steal into killing target is most likely, by my estimation). I strongly suspect that if you make predictions about what it would do, you would sometimes be right and more frequently be wrong.
For example, I doubt a highly trained AI would refuse strat resource swap with distant civ, especially if it (correctly) anticipates that it could utilize the resources it gets in return better than the player. Even in PvP with only one winner, there is some room for diplo, and to the extent it's allowed using it in some capacity is optimal. Top tier players trade in games like Dominions 5, despite that it also benefits an opponent, because making yourself + one other player better off relative to a field of 12 improves your chances to win. I find it difficult to believe that a ML AI with a ridiculous amount of training wouldn't notice that trading is at least sometimes beneficial, and I suspect that absent human emotion that it could do it better than the player.
Same thing about "consistently declaring war" based on unit location. One, it would have to have intel/see that and two, it's not necessarily the case, because this AI might already be at war itself, or in tech deficit etc. Though ML AI would probably not be in a tech deficit.
ML AI would trivially outplay anything a designer can possibly produce. That's not the problem. The problem is that it's overkill and, as you say, wouldn't help designers create an experience that the vast majority of players actually want, which is something challenging but beatable that is playing a different game from them without them noticing it too much.
Not right now, anyway. If they can actually get it to the point where you can have difficulties that keep AI bonuses the same but are instead different ability AI agents, and bring the cost of this down low enough, maybe it actually is worth implementing. That's not a thing right now, to put it mildly, but I wouldn't rule it out in 10-20 years for example.